linux/block/blk-mq.c

4960 lines
124 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0
/*
* Block multiqueue core code
*
* Copyright (C) 2013-2014 Jens Axboe
* Copyright (C) 2013-2014 Christoph Hellwig
*/
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/backing-dev.h>
#include <linux/bio.h>
#include <linux/blkdev.h>
#include <linux/blk-integrity.h>
#include <linux/kmemleak.h>
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
#include <linux/mm.h>
#include <linux/init.h>
#include <linux/slab.h>
#include <linux/workqueue.h>
#include <linux/smp.h>
#include <linux/interrupt.h>
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
#include <linux/llist.h>
#include <linux/cpu.h>
#include <linux/cache.h>
#include <linux/sched/topology.h>
#include <linux/sched/signal.h>
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
#include <linux/delay.h>
#include <linux/crash_dump.h>
#include <linux/prefetch.h>
block: Inline encryption support for blk-mq We must have some way of letting a storage device driver know what encryption context it should use for en/decrypting a request. However, it's the upper layers (like the filesystem/fscrypt) that know about and manages encryption contexts. As such, when the upper layer submits a bio to the block layer, and this bio eventually reaches a device driver with support for inline encryption, the device driver will need to have been told the encryption context for that bio. We want to communicate the encryption context from the upper layer to the storage device along with the bio, when the bio is submitted to the block layer. To do this, we add a struct bio_crypt_ctx to struct bio, which can represent an encryption context (note that we can't use the bi_private field in struct bio to do this because that field does not function to pass information across layers in the storage stack). We also introduce various functions to manipulate the bio_crypt_ctx and make the bio/request merging logic aware of the bio_crypt_ctx. We also make changes to blk-mq to make it handle bios with encryption contexts. blk-mq can merge many bios into the same request. These bios need to have contiguous data unit numbers (the necessary changes to blk-merge are also made to ensure this) - as such, it suffices to keep the data unit number of just the first bio, since that's all a storage driver needs to infer the data unit number to use for each data block in each bio in a request. blk-mq keeps track of the encryption context to be used for all the bios in a request with the request's rq_crypt_ctx. When the first bio is added to an empty request, blk-mq will program the encryption context of that bio into the request_queue's keyslot manager, and store the returned keyslot in the request's rq_crypt_ctx. All the functions to operate on encryption contexts are in blk-crypto.c. Upper layers only need to call bio_crypt_set_ctx with the encryption key, algorithm and data_unit_num; they don't have to worry about getting a keyslot for each encryption context, as blk-mq/blk-crypto handles that. Blk-crypto also makes it possible for request-based layered devices like dm-rq to make use of inline encryption hardware by cloning the rq_crypt_ctx and programming a keyslot in the new request_queue when necessary. Note that any user of the block layer can submit bios with an encryption context, such as filesystems, device-mapper targets, etc. Signed-off-by: Satya Tangirala <satyat@google.com> Reviewed-by: Eric Biggers <ebiggers@google.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-14 00:37:18 +00:00
#include <linux/blk-crypto.h>
#include <linux/part_stat.h>
#include <linux/sched/isolation.h>
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
#include <trace/events/block.h>
#include <linux/t10-pi.h>
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
#include "blk.h"
#include "blk-mq.h"
#include "blk-mq-debugfs.h"
#include "blk-pm.h"
#include "blk-stat.h"
#include "blk-mq-sched.h"
#include "blk-rq-qos.h"
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);
static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
static void blk_mq_request_bypass_insert(struct request *rq,
blk_insert_t flags);
static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
struct list_head *list);
static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
struct io_comp_batch *iob, unsigned int flags);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
/*
* Check if any of the ctx, dispatch list or elevator
* have pending work in this hardware queue.
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
*/
static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
return !list_empty_careful(&hctx->dispatch) ||
sbitmap_any_bit_set(&hctx->ctx_map) ||
blk_mq_sched_has_work(hctx);
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
/*
* Mark this ctx as having pending work in this hardware queue
*/
static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
struct blk_mq_ctx *ctx)
{
const int bit = ctx->index_hw[hctx->type];
if (!sbitmap_test_bit(&hctx->ctx_map, bit))
sbitmap_set_bit(&hctx->ctx_map, bit);
}
static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
struct blk_mq_ctx *ctx)
{
const int bit = ctx->index_hw[hctx->type];
sbitmap_clear_bit(&hctx->ctx_map, bit);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
struct mq_inflight {
struct block_device *part;
unsigned int inflight[2];
};
static bool blk_mq_check_inflight(struct request *rq, void *priv)
{
struct mq_inflight *mi = priv;
if (rq->part && blk_do_io_stat(rq) &&
(!bdev_is_partition(mi->part) || rq->part == mi->part) &&
block: fix inflight statistics of part0 The inflight of partition 0 doesn't include inflight IOs to all sub-partitions, since currently mq calculates inflight of specific partition by simply camparing the value of the partition pointer. Thus the following case is possible: $ cat /sys/block/vda/inflight        0        0 $ cat /sys/block/vda/vda1/inflight        0      128 While single queue device (on a previous version, e.g. v3.10) has no this issue: $cat /sys/block/sda/sda3/inflight 0 33 $cat /sys/block/sda/inflight 0 33 Partition 0 should be specially handled since it represents the whole disk. This issue is introduced since commit bf0ddaba65dd ("blk-mq: fix sysfs inflight counter"). Besides, this patch can also fix the inflight statistics of part 0 in /proc/diskstats. Before this patch, the inflight statistics of part 0 doesn't include that of sub partitions. (I have marked the 'inflight' field with asterisk.) $cat /proc/diskstats 259 0 nvme0n1 45974469 0 367814768 6445794 1 0 1 0 *0* 111062 6445794 0 0 0 0 0 0 259 2 nvme0n1p1 45974058 0 367797952 6445727 0 0 0 0 *33* 111001 6445727 0 0 0 0 0 0 This is introduced since commit f299b7c7a9de ("blk-mq: provide internal in-flight variant"). Fixes: bf0ddaba65dd ("blk-mq: fix sysfs inflight counter") Fixes: f299b7c7a9de ("blk-mq: provide internal in-flight variant") Signed-off-by: Jeffle Xu <jefflexu@linux.alibaba.com> Reviewed-by: Christoph Hellwig <hch@lst.de> [axboe: adapt for 5.11 partition change] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-12-02 11:11:45 +00:00
blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
mi->inflight[rq_data_dir(rq)]++;
return true;
}
unsigned int blk_mq_in_flight(struct request_queue *q,
struct block_device *part)
{
struct mq_inflight mi = { .part = part };
blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
return mi.inflight[0] + mi.inflight[1];
}
void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
unsigned int inflight[2])
{
struct mq_inflight mi = { .part = part };
blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
inflight[0] = mi.inflight[0];
inflight[1] = mi.inflight[1];
}
void blk_freeze_queue_start(struct request_queue *q)
{
blk-mq: fix hang caused by freeze/unfreeze sequence The following is a description of a hang in blk_mq_freeze_queue_wait(). The hang happens on attempt to freeze a queue while another task does queue unfreeze. The root cause is an incorrect sequence of percpu_ref_resurrect() and percpu_ref_kill() and as a result those two can be swapped: CPU#0 CPU#1 ---------------- ----------------- q1 = blk_mq_init_queue(shared_tags) q2 = blk_mq_init_queue(shared_tags): blk_mq_add_queue_tag_set(shared_tags): blk_mq_update_tag_set_depth(shared_tags): list_for_each_entry() blk_mq_freeze_queue(q1) > percpu_ref_kill() > blk_mq_freeze_queue_wait() blk_cleanup_queue(q1) blk_mq_freeze_queue(q1) > percpu_ref_kill() ^^^^^^ freeze_depth can't guarantee the order blk_mq_unfreeze_queue() > percpu_ref_resurrect() > blk_mq_freeze_queue_wait() ^^^^^^ Hang here!!!! This wrong sequence raises kernel warning: percpu_ref_kill_and_confirm called more than once on blk_queue_usage_counter_release! WARNING: CPU: 0 PID: 11854 at lib/percpu-refcount.c:336 percpu_ref_kill_and_confirm+0x99/0xb0 But the most unpleasant effect is a hang of a blk_mq_freeze_queue_wait(), which waits for a zero of a q_usage_counter, which never happens because percpu-ref was reinited (instead of being killed) and stays in PERCPU state forever. How to reproduce: - "insmod null_blk.ko shared_tags=1 nr_devices=0 queue_mode=2" - cpu0: python Script.py 0; taskset the corresponding process running on cpu0 - cpu1: python Script.py 1; taskset the corresponding process running on cpu1 Script.py: ------ #!/usr/bin/python3 import os import sys while True: on = "echo 1 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] off = "echo 0 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] os.system(on) os.system(off) ------ This bug was first reported and fixed by Roman, previous discussion: [1] Message id: 1443287365-4244-7-git-send-email-akinobu.mita@gmail.com [2] Message id: 1443563240-29306-6-git-send-email-tj@kernel.org [3] https://patchwork.kernel.org/patch/9268199/ Reviewed-by: Hannes Reinecke <hare@suse.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Roman Pen <roman.penyaev@profitbricks.com> Signed-off-by: Bob Liu <bob.liu@oracle.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-05-21 03:25:55 +00:00
mutex_lock(&q->mq_freeze_lock);
if (++q->mq_freeze_depth == 1) {
block: generic request_queue reference counting Allow pmem, and other synchronous/bio-based block drivers, to fallback on a per-cpu reference count managed by the core for tracking queue live/dead state. The existing per-cpu reference count for the blk_mq case is promoted to be used in all block i/o scenarios. This involves initializing it by default, waiting for it to drop to zero at exit, and holding a live reference over the invocation of q->make_request_fn() in generic_make_request(). The blk_mq code continues to take its own reference per blk_mq request and retains the ability to freeze the queue, but the check that the queue is frozen is moved to generic_make_request(). This fixes crash signatures like the following: BUG: unable to handle kernel paging request at ffff880140000000 [..] Call Trace: [<ffffffff8145e8bf>] ? copy_user_handle_tail+0x5f/0x70 [<ffffffffa004e1e0>] pmem_do_bvec.isra.11+0x70/0xf0 [nd_pmem] [<ffffffffa004e331>] pmem_make_request+0xd1/0x200 [nd_pmem] [<ffffffff811c3162>] ? mempool_alloc+0x72/0x1a0 [<ffffffff8141f8b6>] generic_make_request+0xd6/0x110 [<ffffffff8141f966>] submit_bio+0x76/0x170 [<ffffffff81286dff>] submit_bh_wbc+0x12f/0x160 [<ffffffff81286e62>] submit_bh+0x12/0x20 [<ffffffff813395bd>] jbd2_write_superblock+0x8d/0x170 [<ffffffff8133974d>] jbd2_mark_journal_empty+0x5d/0x90 [<ffffffff813399cb>] jbd2_journal_destroy+0x24b/0x270 [<ffffffff810bc4ca>] ? put_pwq_unlocked+0x2a/0x30 [<ffffffff810bc6f5>] ? destroy_workqueue+0x225/0x250 [<ffffffff81303494>] ext4_put_super+0x64/0x360 [<ffffffff8124ab1a>] generic_shutdown_super+0x6a/0xf0 Cc: Jens Axboe <axboe@kernel.dk> Cc: Keith Busch <keith.busch@intel.com> Cc: Ross Zwisler <ross.zwisler@linux.intel.com> Suggested-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: Ross Zwisler <ross.zwisler@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-10-21 17:20:12 +00:00
percpu_ref_kill(&q->q_usage_counter);
blk-mq: fix hang caused by freeze/unfreeze sequence The following is a description of a hang in blk_mq_freeze_queue_wait(). The hang happens on attempt to freeze a queue while another task does queue unfreeze. The root cause is an incorrect sequence of percpu_ref_resurrect() and percpu_ref_kill() and as a result those two can be swapped: CPU#0 CPU#1 ---------------- ----------------- q1 = blk_mq_init_queue(shared_tags) q2 = blk_mq_init_queue(shared_tags): blk_mq_add_queue_tag_set(shared_tags): blk_mq_update_tag_set_depth(shared_tags): list_for_each_entry() blk_mq_freeze_queue(q1) > percpu_ref_kill() > blk_mq_freeze_queue_wait() blk_cleanup_queue(q1) blk_mq_freeze_queue(q1) > percpu_ref_kill() ^^^^^^ freeze_depth can't guarantee the order blk_mq_unfreeze_queue() > percpu_ref_resurrect() > blk_mq_freeze_queue_wait() ^^^^^^ Hang here!!!! This wrong sequence raises kernel warning: percpu_ref_kill_and_confirm called more than once on blk_queue_usage_counter_release! WARNING: CPU: 0 PID: 11854 at lib/percpu-refcount.c:336 percpu_ref_kill_and_confirm+0x99/0xb0 But the most unpleasant effect is a hang of a blk_mq_freeze_queue_wait(), which waits for a zero of a q_usage_counter, which never happens because percpu-ref was reinited (instead of being killed) and stays in PERCPU state forever. How to reproduce: - "insmod null_blk.ko shared_tags=1 nr_devices=0 queue_mode=2" - cpu0: python Script.py 0; taskset the corresponding process running on cpu0 - cpu1: python Script.py 1; taskset the corresponding process running on cpu1 Script.py: ------ #!/usr/bin/python3 import os import sys while True: on = "echo 1 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] off = "echo 0 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] os.system(on) os.system(off) ------ This bug was first reported and fixed by Roman, previous discussion: [1] Message id: 1443287365-4244-7-git-send-email-akinobu.mita@gmail.com [2] Message id: 1443563240-29306-6-git-send-email-tj@kernel.org [3] https://patchwork.kernel.org/patch/9268199/ Reviewed-by: Hannes Reinecke <hare@suse.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Roman Pen <roman.penyaev@profitbricks.com> Signed-off-by: Bob Liu <bob.liu@oracle.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-05-21 03:25:55 +00:00
mutex_unlock(&q->mq_freeze_lock);
if (queue_is_mq(q))
blk_mq_run_hw_queues(q, false);
blk-mq: fix hang caused by freeze/unfreeze sequence The following is a description of a hang in blk_mq_freeze_queue_wait(). The hang happens on attempt to freeze a queue while another task does queue unfreeze. The root cause is an incorrect sequence of percpu_ref_resurrect() and percpu_ref_kill() and as a result those two can be swapped: CPU#0 CPU#1 ---------------- ----------------- q1 = blk_mq_init_queue(shared_tags) q2 = blk_mq_init_queue(shared_tags): blk_mq_add_queue_tag_set(shared_tags): blk_mq_update_tag_set_depth(shared_tags): list_for_each_entry() blk_mq_freeze_queue(q1) > percpu_ref_kill() > blk_mq_freeze_queue_wait() blk_cleanup_queue(q1) blk_mq_freeze_queue(q1) > percpu_ref_kill() ^^^^^^ freeze_depth can't guarantee the order blk_mq_unfreeze_queue() > percpu_ref_resurrect() > blk_mq_freeze_queue_wait() ^^^^^^ Hang here!!!! This wrong sequence raises kernel warning: percpu_ref_kill_and_confirm called more than once on blk_queue_usage_counter_release! WARNING: CPU: 0 PID: 11854 at lib/percpu-refcount.c:336 percpu_ref_kill_and_confirm+0x99/0xb0 But the most unpleasant effect is a hang of a blk_mq_freeze_queue_wait(), which waits for a zero of a q_usage_counter, which never happens because percpu-ref was reinited (instead of being killed) and stays in PERCPU state forever. How to reproduce: - "insmod null_blk.ko shared_tags=1 nr_devices=0 queue_mode=2" - cpu0: python Script.py 0; taskset the corresponding process running on cpu0 - cpu1: python Script.py 1; taskset the corresponding process running on cpu1 Script.py: ------ #!/usr/bin/python3 import os import sys while True: on = "echo 1 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] off = "echo 0 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] os.system(on) os.system(off) ------ This bug was first reported and fixed by Roman, previous discussion: [1] Message id: 1443287365-4244-7-git-send-email-akinobu.mita@gmail.com [2] Message id: 1443563240-29306-6-git-send-email-tj@kernel.org [3] https://patchwork.kernel.org/patch/9268199/ Reviewed-by: Hannes Reinecke <hare@suse.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Roman Pen <roman.penyaev@profitbricks.com> Signed-off-by: Bob Liu <bob.liu@oracle.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-05-21 03:25:55 +00:00
} else {
mutex_unlock(&q->mq_freeze_lock);
}
}
EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
void blk_mq_freeze_queue_wait(struct request_queue *q)
{
block: generic request_queue reference counting Allow pmem, and other synchronous/bio-based block drivers, to fallback on a per-cpu reference count managed by the core for tracking queue live/dead state. The existing per-cpu reference count for the blk_mq case is promoted to be used in all block i/o scenarios. This involves initializing it by default, waiting for it to drop to zero at exit, and holding a live reference over the invocation of q->make_request_fn() in generic_make_request(). The blk_mq code continues to take its own reference per blk_mq request and retains the ability to freeze the queue, but the check that the queue is frozen is moved to generic_make_request(). This fixes crash signatures like the following: BUG: unable to handle kernel paging request at ffff880140000000 [..] Call Trace: [<ffffffff8145e8bf>] ? copy_user_handle_tail+0x5f/0x70 [<ffffffffa004e1e0>] pmem_do_bvec.isra.11+0x70/0xf0 [nd_pmem] [<ffffffffa004e331>] pmem_make_request+0xd1/0x200 [nd_pmem] [<ffffffff811c3162>] ? mempool_alloc+0x72/0x1a0 [<ffffffff8141f8b6>] generic_make_request+0xd6/0x110 [<ffffffff8141f966>] submit_bio+0x76/0x170 [<ffffffff81286dff>] submit_bh_wbc+0x12f/0x160 [<ffffffff81286e62>] submit_bh+0x12/0x20 [<ffffffff813395bd>] jbd2_write_superblock+0x8d/0x170 [<ffffffff8133974d>] jbd2_mark_journal_empty+0x5d/0x90 [<ffffffff813399cb>] jbd2_journal_destroy+0x24b/0x270 [<ffffffff810bc4ca>] ? put_pwq_unlocked+0x2a/0x30 [<ffffffff810bc6f5>] ? destroy_workqueue+0x225/0x250 [<ffffffff81303494>] ext4_put_super+0x64/0x360 [<ffffffff8124ab1a>] generic_shutdown_super+0x6a/0xf0 Cc: Jens Axboe <axboe@kernel.dk> Cc: Keith Busch <keith.busch@intel.com> Cc: Ross Zwisler <ross.zwisler@linux.intel.com> Suggested-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: Ross Zwisler <ross.zwisler@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-10-21 17:20:12 +00:00
wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
}
EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
unsigned long timeout)
{
return wait_event_timeout(q->mq_freeze_wq,
percpu_ref_is_zero(&q->q_usage_counter),
timeout);
}
EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
/*
* Guarantee no request is in use, so we can change any data structure of
* the queue afterward.
*/
block: generic request_queue reference counting Allow pmem, and other synchronous/bio-based block drivers, to fallback on a per-cpu reference count managed by the core for tracking queue live/dead state. The existing per-cpu reference count for the blk_mq case is promoted to be used in all block i/o scenarios. This involves initializing it by default, waiting for it to drop to zero at exit, and holding a live reference over the invocation of q->make_request_fn() in generic_make_request(). The blk_mq code continues to take its own reference per blk_mq request and retains the ability to freeze the queue, but the check that the queue is frozen is moved to generic_make_request(). This fixes crash signatures like the following: BUG: unable to handle kernel paging request at ffff880140000000 [..] Call Trace: [<ffffffff8145e8bf>] ? copy_user_handle_tail+0x5f/0x70 [<ffffffffa004e1e0>] pmem_do_bvec.isra.11+0x70/0xf0 [nd_pmem] [<ffffffffa004e331>] pmem_make_request+0xd1/0x200 [nd_pmem] [<ffffffff811c3162>] ? mempool_alloc+0x72/0x1a0 [<ffffffff8141f8b6>] generic_make_request+0xd6/0x110 [<ffffffff8141f966>] submit_bio+0x76/0x170 [<ffffffff81286dff>] submit_bh_wbc+0x12f/0x160 [<ffffffff81286e62>] submit_bh+0x12/0x20 [<ffffffff813395bd>] jbd2_write_superblock+0x8d/0x170 [<ffffffff8133974d>] jbd2_mark_journal_empty+0x5d/0x90 [<ffffffff813399cb>] jbd2_journal_destroy+0x24b/0x270 [<ffffffff810bc4ca>] ? put_pwq_unlocked+0x2a/0x30 [<ffffffff810bc6f5>] ? destroy_workqueue+0x225/0x250 [<ffffffff81303494>] ext4_put_super+0x64/0x360 [<ffffffff8124ab1a>] generic_shutdown_super+0x6a/0xf0 Cc: Jens Axboe <axboe@kernel.dk> Cc: Keith Busch <keith.busch@intel.com> Cc: Ross Zwisler <ross.zwisler@linux.intel.com> Suggested-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: Ross Zwisler <ross.zwisler@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-10-21 17:20:12 +00:00
void blk_freeze_queue(struct request_queue *q)
{
block: generic request_queue reference counting Allow pmem, and other synchronous/bio-based block drivers, to fallback on a per-cpu reference count managed by the core for tracking queue live/dead state. The existing per-cpu reference count for the blk_mq case is promoted to be used in all block i/o scenarios. This involves initializing it by default, waiting for it to drop to zero at exit, and holding a live reference over the invocation of q->make_request_fn() in generic_make_request(). The blk_mq code continues to take its own reference per blk_mq request and retains the ability to freeze the queue, but the check that the queue is frozen is moved to generic_make_request(). This fixes crash signatures like the following: BUG: unable to handle kernel paging request at ffff880140000000 [..] Call Trace: [<ffffffff8145e8bf>] ? copy_user_handle_tail+0x5f/0x70 [<ffffffffa004e1e0>] pmem_do_bvec.isra.11+0x70/0xf0 [nd_pmem] [<ffffffffa004e331>] pmem_make_request+0xd1/0x200 [nd_pmem] [<ffffffff811c3162>] ? mempool_alloc+0x72/0x1a0 [<ffffffff8141f8b6>] generic_make_request+0xd6/0x110 [<ffffffff8141f966>] submit_bio+0x76/0x170 [<ffffffff81286dff>] submit_bh_wbc+0x12f/0x160 [<ffffffff81286e62>] submit_bh+0x12/0x20 [<ffffffff813395bd>] jbd2_write_superblock+0x8d/0x170 [<ffffffff8133974d>] jbd2_mark_journal_empty+0x5d/0x90 [<ffffffff813399cb>] jbd2_journal_destroy+0x24b/0x270 [<ffffffff810bc4ca>] ? put_pwq_unlocked+0x2a/0x30 [<ffffffff810bc6f5>] ? destroy_workqueue+0x225/0x250 [<ffffffff81303494>] ext4_put_super+0x64/0x360 [<ffffffff8124ab1a>] generic_shutdown_super+0x6a/0xf0 Cc: Jens Axboe <axboe@kernel.dk> Cc: Keith Busch <keith.busch@intel.com> Cc: Ross Zwisler <ross.zwisler@linux.intel.com> Suggested-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: Ross Zwisler <ross.zwisler@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-10-21 17:20:12 +00:00
/*
* In the !blk_mq case we are only calling this to kill the
* q_usage_counter, otherwise this increases the freeze depth
* and waits for it to return to zero. For this reason there is
* no blk_unfreeze_queue(), and blk_freeze_queue() is not
* exported to drivers as the only user for unfreeze is blk_mq.
*/
blk_freeze_queue_start(q);
blk_mq_freeze_queue_wait(q);
}
block: generic request_queue reference counting Allow pmem, and other synchronous/bio-based block drivers, to fallback on a per-cpu reference count managed by the core for tracking queue live/dead state. The existing per-cpu reference count for the blk_mq case is promoted to be used in all block i/o scenarios. This involves initializing it by default, waiting for it to drop to zero at exit, and holding a live reference over the invocation of q->make_request_fn() in generic_make_request(). The blk_mq code continues to take its own reference per blk_mq request and retains the ability to freeze the queue, but the check that the queue is frozen is moved to generic_make_request(). This fixes crash signatures like the following: BUG: unable to handle kernel paging request at ffff880140000000 [..] Call Trace: [<ffffffff8145e8bf>] ? copy_user_handle_tail+0x5f/0x70 [<ffffffffa004e1e0>] pmem_do_bvec.isra.11+0x70/0xf0 [nd_pmem] [<ffffffffa004e331>] pmem_make_request+0xd1/0x200 [nd_pmem] [<ffffffff811c3162>] ? mempool_alloc+0x72/0x1a0 [<ffffffff8141f8b6>] generic_make_request+0xd6/0x110 [<ffffffff8141f966>] submit_bio+0x76/0x170 [<ffffffff81286dff>] submit_bh_wbc+0x12f/0x160 [<ffffffff81286e62>] submit_bh+0x12/0x20 [<ffffffff813395bd>] jbd2_write_superblock+0x8d/0x170 [<ffffffff8133974d>] jbd2_mark_journal_empty+0x5d/0x90 [<ffffffff813399cb>] jbd2_journal_destroy+0x24b/0x270 [<ffffffff810bc4ca>] ? put_pwq_unlocked+0x2a/0x30 [<ffffffff810bc6f5>] ? destroy_workqueue+0x225/0x250 [<ffffffff81303494>] ext4_put_super+0x64/0x360 [<ffffffff8124ab1a>] generic_shutdown_super+0x6a/0xf0 Cc: Jens Axboe <axboe@kernel.dk> Cc: Keith Busch <keith.busch@intel.com> Cc: Ross Zwisler <ross.zwisler@linux.intel.com> Suggested-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: Ross Zwisler <ross.zwisler@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-10-21 17:20:12 +00:00
void blk_mq_freeze_queue(struct request_queue *q)
{
/*
* ...just an alias to keep freeze and unfreeze actions balanced
* in the blk_mq_* namespace
*/
blk_freeze_queue(q);
}
EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
blk-mq: fix hang caused by freeze/unfreeze sequence The following is a description of a hang in blk_mq_freeze_queue_wait(). The hang happens on attempt to freeze a queue while another task does queue unfreeze. The root cause is an incorrect sequence of percpu_ref_resurrect() and percpu_ref_kill() and as a result those two can be swapped: CPU#0 CPU#1 ---------------- ----------------- q1 = blk_mq_init_queue(shared_tags) q2 = blk_mq_init_queue(shared_tags): blk_mq_add_queue_tag_set(shared_tags): blk_mq_update_tag_set_depth(shared_tags): list_for_each_entry() blk_mq_freeze_queue(q1) > percpu_ref_kill() > blk_mq_freeze_queue_wait() blk_cleanup_queue(q1) blk_mq_freeze_queue(q1) > percpu_ref_kill() ^^^^^^ freeze_depth can't guarantee the order blk_mq_unfreeze_queue() > percpu_ref_resurrect() > blk_mq_freeze_queue_wait() ^^^^^^ Hang here!!!! This wrong sequence raises kernel warning: percpu_ref_kill_and_confirm called more than once on blk_queue_usage_counter_release! WARNING: CPU: 0 PID: 11854 at lib/percpu-refcount.c:336 percpu_ref_kill_and_confirm+0x99/0xb0 But the most unpleasant effect is a hang of a blk_mq_freeze_queue_wait(), which waits for a zero of a q_usage_counter, which never happens because percpu-ref was reinited (instead of being killed) and stays in PERCPU state forever. How to reproduce: - "insmod null_blk.ko shared_tags=1 nr_devices=0 queue_mode=2" - cpu0: python Script.py 0; taskset the corresponding process running on cpu0 - cpu1: python Script.py 1; taskset the corresponding process running on cpu1 Script.py: ------ #!/usr/bin/python3 import os import sys while True: on = "echo 1 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] off = "echo 0 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] os.system(on) os.system(off) ------ This bug was first reported and fixed by Roman, previous discussion: [1] Message id: 1443287365-4244-7-git-send-email-akinobu.mita@gmail.com [2] Message id: 1443563240-29306-6-git-send-email-tj@kernel.org [3] https://patchwork.kernel.org/patch/9268199/ Reviewed-by: Hannes Reinecke <hare@suse.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Roman Pen <roman.penyaev@profitbricks.com> Signed-off-by: Bob Liu <bob.liu@oracle.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-05-21 03:25:55 +00:00
mutex_lock(&q->mq_freeze_lock);
if (force_atomic)
q->q_usage_counter.data->force_atomic = true;
blk-mq: fix hang caused by freeze/unfreeze sequence The following is a description of a hang in blk_mq_freeze_queue_wait(). The hang happens on attempt to freeze a queue while another task does queue unfreeze. The root cause is an incorrect sequence of percpu_ref_resurrect() and percpu_ref_kill() and as a result those two can be swapped: CPU#0 CPU#1 ---------------- ----------------- q1 = blk_mq_init_queue(shared_tags) q2 = blk_mq_init_queue(shared_tags): blk_mq_add_queue_tag_set(shared_tags): blk_mq_update_tag_set_depth(shared_tags): list_for_each_entry() blk_mq_freeze_queue(q1) > percpu_ref_kill() > blk_mq_freeze_queue_wait() blk_cleanup_queue(q1) blk_mq_freeze_queue(q1) > percpu_ref_kill() ^^^^^^ freeze_depth can't guarantee the order blk_mq_unfreeze_queue() > percpu_ref_resurrect() > blk_mq_freeze_queue_wait() ^^^^^^ Hang here!!!! This wrong sequence raises kernel warning: percpu_ref_kill_and_confirm called more than once on blk_queue_usage_counter_release! WARNING: CPU: 0 PID: 11854 at lib/percpu-refcount.c:336 percpu_ref_kill_and_confirm+0x99/0xb0 But the most unpleasant effect is a hang of a blk_mq_freeze_queue_wait(), which waits for a zero of a q_usage_counter, which never happens because percpu-ref was reinited (instead of being killed) and stays in PERCPU state forever. How to reproduce: - "insmod null_blk.ko shared_tags=1 nr_devices=0 queue_mode=2" - cpu0: python Script.py 0; taskset the corresponding process running on cpu0 - cpu1: python Script.py 1; taskset the corresponding process running on cpu1 Script.py: ------ #!/usr/bin/python3 import os import sys while True: on = "echo 1 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] off = "echo 0 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] os.system(on) os.system(off) ------ This bug was first reported and fixed by Roman, previous discussion: [1] Message id: 1443287365-4244-7-git-send-email-akinobu.mita@gmail.com [2] Message id: 1443563240-29306-6-git-send-email-tj@kernel.org [3] https://patchwork.kernel.org/patch/9268199/ Reviewed-by: Hannes Reinecke <hare@suse.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Roman Pen <roman.penyaev@profitbricks.com> Signed-off-by: Bob Liu <bob.liu@oracle.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-05-21 03:25:55 +00:00
q->mq_freeze_depth--;
WARN_ON_ONCE(q->mq_freeze_depth < 0);
if (!q->mq_freeze_depth) {
percpu_ref_resurrect(&q->q_usage_counter);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
wake_up_all(&q->mq_freeze_wq);
}
blk-mq: fix hang caused by freeze/unfreeze sequence The following is a description of a hang in blk_mq_freeze_queue_wait(). The hang happens on attempt to freeze a queue while another task does queue unfreeze. The root cause is an incorrect sequence of percpu_ref_resurrect() and percpu_ref_kill() and as a result those two can be swapped: CPU#0 CPU#1 ---------------- ----------------- q1 = blk_mq_init_queue(shared_tags) q2 = blk_mq_init_queue(shared_tags): blk_mq_add_queue_tag_set(shared_tags): blk_mq_update_tag_set_depth(shared_tags): list_for_each_entry() blk_mq_freeze_queue(q1) > percpu_ref_kill() > blk_mq_freeze_queue_wait() blk_cleanup_queue(q1) blk_mq_freeze_queue(q1) > percpu_ref_kill() ^^^^^^ freeze_depth can't guarantee the order blk_mq_unfreeze_queue() > percpu_ref_resurrect() > blk_mq_freeze_queue_wait() ^^^^^^ Hang here!!!! This wrong sequence raises kernel warning: percpu_ref_kill_and_confirm called more than once on blk_queue_usage_counter_release! WARNING: CPU: 0 PID: 11854 at lib/percpu-refcount.c:336 percpu_ref_kill_and_confirm+0x99/0xb0 But the most unpleasant effect is a hang of a blk_mq_freeze_queue_wait(), which waits for a zero of a q_usage_counter, which never happens because percpu-ref was reinited (instead of being killed) and stays in PERCPU state forever. How to reproduce: - "insmod null_blk.ko shared_tags=1 nr_devices=0 queue_mode=2" - cpu0: python Script.py 0; taskset the corresponding process running on cpu0 - cpu1: python Script.py 1; taskset the corresponding process running on cpu1 Script.py: ------ #!/usr/bin/python3 import os import sys while True: on = "echo 1 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] off = "echo 0 > /sys/kernel/config/nullb/%s/power" % sys.argv[1] os.system(on) os.system(off) ------ This bug was first reported and fixed by Roman, previous discussion: [1] Message id: 1443287365-4244-7-git-send-email-akinobu.mita@gmail.com [2] Message id: 1443563240-29306-6-git-send-email-tj@kernel.org [3] https://patchwork.kernel.org/patch/9268199/ Reviewed-by: Hannes Reinecke <hare@suse.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Roman Pen <roman.penyaev@profitbricks.com> Signed-off-by: Bob Liu <bob.liu@oracle.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-05-21 03:25:55 +00:00
mutex_unlock(&q->mq_freeze_lock);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
void blk_mq_unfreeze_queue(struct request_queue *q)
{
__blk_mq_unfreeze_queue(q, false);
}
EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
/*
* FIXME: replace the scsi_internal_device_*block_nowait() calls in the
* mpt3sas driver such that this function can be removed.
*/
void blk_mq_quiesce_queue_nowait(struct request_queue *q)
{
unsigned long flags;
spin_lock_irqsave(&q->queue_lock, flags);
if (!q->quiesce_depth++)
blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
spin_unlock_irqrestore(&q->queue_lock, flags);
}
EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
/**
* blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
* @set: tag_set to wait on
*
* Note: it is driver's responsibility for making sure that quiesce has
* been started on or more of the request_queues of the tag_set. This
* function only waits for the quiesce on those request_queues that had
* the quiesce flag set using blk_mq_quiesce_queue_nowait.
*/
void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
{
if (set->flags & BLK_MQ_F_BLOCKING)
synchronize_srcu(set->srcu);
else
synchronize_rcu();
}
EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
/**
* blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
* @q: request queue.
*
* Note: this function does not prevent that the struct request end_io()
* callback function is invoked. Once this function is returned, we make
* sure no dispatch can happen until the queue is unquiesced via
* blk_mq_unquiesce_queue().
*/
void blk_mq_quiesce_queue(struct request_queue *q)
{
blk_mq_quiesce_queue_nowait(q);
/* nothing to wait for non-mq queues */
if (queue_is_mq(q))
blk_mq_wait_quiesce_done(q->tag_set);
}
EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
/*
* blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
* @q: request queue.
*
* This function recovers queue into the state before quiescing
* which is done by blk_mq_quiesce_queue.
*/
void blk_mq_unquiesce_queue(struct request_queue *q)
{
unsigned long flags;
bool run_queue = false;
spin_lock_irqsave(&q->queue_lock, flags);
if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
;
} else if (!--q->quiesce_depth) {
blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
run_queue = true;
}
spin_unlock_irqrestore(&q->queue_lock, flags);
/* dispatch requests which are inserted during quiescing */
if (run_queue)
blk_mq_run_hw_queues(q, true);
}
EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
{
struct request_queue *q;
mutex_lock(&set->tag_list_lock);
list_for_each_entry(q, &set->tag_list, tag_set_list) {
if (!blk_queue_skip_tagset_quiesce(q))
blk_mq_quiesce_queue_nowait(q);
}
blk_mq_wait_quiesce_done(set);
mutex_unlock(&set->tag_list_lock);
}
EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
{
struct request_queue *q;
mutex_lock(&set->tag_list_lock);
list_for_each_entry(q, &set->tag_list, tag_set_list) {
if (!blk_queue_skip_tagset_quiesce(q))
blk_mq_unquiesce_queue(q);
}
mutex_unlock(&set->tag_list_lock);
}
EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
void blk_mq_wake_waiters(struct request_queue *q)
{
struct blk_mq_hw_ctx *hctx;
unsigned long i;
queue_for_each_hw_ctx(q, hctx, i)
if (blk_mq_hw_queue_mapped(hctx))
blk_mq_tag_wakeup_all(hctx->tags, true);
}
void blk_rq_init(struct request_queue *q, struct request *rq)
{
memset(rq, 0, sizeof(*rq));
INIT_LIST_HEAD(&rq->queuelist);
rq->q = q;
rq->__sector = (sector_t) -1;
INIT_HLIST_NODE(&rq->hash);
RB_CLEAR_NODE(&rq->rb_node);
rq->tag = BLK_MQ_NO_TAG;
rq->internal_tag = BLK_MQ_NO_TAG;
rq->start_time_ns = blk_time_get_ns();
rq->part = NULL;
blk_crypto_rq_set_defaults(rq);
}
EXPORT_SYMBOL(blk_rq_init);
/* Set start and alloc time when the allocated request is actually used */
static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
{
if (blk_mq_need_time_stamp(rq))
rq->start_time_ns = blk_time_get_ns();
else
rq->start_time_ns = 0;
#ifdef CONFIG_BLK_RQ_ALLOC_TIME
if (blk_queue_rq_alloc_time(rq->q))
rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
else
rq->alloc_time_ns = 0;
#endif
}
static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
struct blk_mq_tags *tags, unsigned int tag)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct blk_mq_ctx *ctx = data->ctx;
struct blk_mq_hw_ctx *hctx = data->hctx;
struct request_queue *q = data->q;
struct request *rq = tags->static_rqs[tag];
rq->q = q;
rq->mq_ctx = ctx;
rq->mq_hctx = hctx;
rq->cmd_flags = data->cmd_flags;
if (data->flags & BLK_MQ_REQ_PM)
data->rq_flags |= RQF_PM;
if (blk_queue_io_stat(q))
data->rq_flags |= RQF_IO_STAT;
rq->rq_flags = data->rq_flags;
if (data->rq_flags & RQF_SCHED_TAGS) {
rq->tag = BLK_MQ_NO_TAG;
rq->internal_tag = tag;
} else {
rq->tag = tag;
rq->internal_tag = BLK_MQ_NO_TAG;
}
rq->timeout = 0;
rq->part = NULL;
rq->io_start_time_ns = 0;
rq->stats_sectors = 0;
rq->nr_phys_segments = 0;
#if defined(CONFIG_BLK_DEV_INTEGRITY)
rq->nr_integrity_segments = 0;
#endif
rq->end_io = NULL;
rq->end_io_data = NULL;
blk_crypto_rq_set_defaults(rq);
INIT_LIST_HEAD(&rq->queuelist);
/* tag was already set */
WRITE_ONCE(rq->deadline, 0);
req_ref_set(rq, 1);
if (rq->rq_flags & RQF_USE_SCHED) {
struct elevator_queue *e = data->q->elevator;
INIT_HLIST_NODE(&rq->hash);
RB_CLEAR_NODE(&rq->rb_node);
if (e->type->ops.prepare_request)
e->type->ops.prepare_request(rq);
}
return rq;
}
static inline struct request *
__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
{
unsigned int tag, tag_offset;
struct blk_mq_tags *tags;
struct request *rq;
unsigned long tag_mask;
int i, nr = 0;
tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
if (unlikely(!tag_mask))
return NULL;
tags = blk_mq_tags_from_data(data);
for (i = 0; tag_mask; i++) {
if (!(tag_mask & (1UL << i)))
continue;
tag = tag_offset + i;
prefetch(tags->static_rqs[tag]);
tag_mask &= ~(1UL << i);
rq = blk_mq_rq_ctx_init(data, tags, tag);
rq_list_add(data->cached_rq, rq);
nr++;
}
if (!(data->rq_flags & RQF_SCHED_TAGS))
blk_mq_add_active_requests(data->hctx, nr);
/* caller already holds a reference, add for remainder */
percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
data->nr_tags -= nr;
return rq_list_pop(data->cached_rq);
}
static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
{
struct request_queue *q = data->q;
u64 alloc_time_ns = 0;
struct request *rq;
unsigned int tag;
/* alloc_time includes depth and tag waits */
if (blk_queue_rq_alloc_time(q))
alloc_time_ns = blk_time_get_ns();
if (data->cmd_flags & REQ_NOWAIT)
data->flags |= BLK_MQ_REQ_NOWAIT;
retry:
data->ctx = blk_mq_get_ctx(q);
data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
if (q->elevator) {
/*
* All requests use scheduler tags when an I/O scheduler is
* enabled for the queue.
*/
data->rq_flags |= RQF_SCHED_TAGS;
/*
* Flush/passthrough requests are special and go directly to the
* dispatch list.
*/
if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
!blk_op_is_passthrough(data->cmd_flags)) {
struct elevator_mq_ops *ops = &q->elevator->type->ops;
WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
data->rq_flags |= RQF_USE_SCHED;
if (ops->limit_depth)
ops->limit_depth(data->cmd_flags, data);
}
} else {
blk_mq_tag_busy(data->hctx);
}
if (data->flags & BLK_MQ_REQ_RESERVED)
data->rq_flags |= RQF_RESV;
/*
* Try batched alloc if we want more than 1 tag.
*/
if (data->nr_tags > 1) {
rq = __blk_mq_alloc_requests_batch(data);
if (rq) {
blk_mq_rq_time_init(rq, alloc_time_ns);
return rq;
}
data->nr_tags = 1;
}
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
/*
* Waiting allocations only fail because of an inactive hctx. In that
* case just retry the hctx assignment and tag allocation as CPU hotplug
* should have migrated us to an online CPU by now.
*/
tag = blk_mq_get_tag(data);
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
if (tag == BLK_MQ_NO_TAG) {
if (data->flags & BLK_MQ_REQ_NOWAIT)
return NULL;
/*
* Give up the CPU and sleep for a random short time to
* ensure that thread using a realtime scheduling class
* are migrated off the CPU, and thus off the hctx that
* is going away.
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
*/
msleep(3);
goto retry;
}
if (!(data->rq_flags & RQF_SCHED_TAGS))
blk_mq_inc_active_requests(data->hctx);
rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
blk_mq_rq_time_init(rq, alloc_time_ns);
return rq;
}
static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
struct blk_plug *plug,
blk_opf_t opf,
blk_mq_req_flags_t flags)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct blk_mq_alloc_data data = {
.q = q,
.flags = flags,
.cmd_flags = opf,
.nr_tags = plug->nr_ios,
.cached_rq = &plug->cached_rq,
};
struct request *rq;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
if (blk_queue_enter(q, flags))
return NULL;
plug->nr_ios = 1;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
rq = __blk_mq_alloc_requests(&data);
if (unlikely(!rq))
blk_queue_exit(q);
return rq;
}
static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
blk_opf_t opf,
blk_mq_req_flags_t flags)
{
struct blk_plug *plug = current->plug;
struct request *rq;
if (!plug)
return NULL;
if (rq_list_empty(plug->cached_rq)) {
if (plug->nr_ios == 1)
return NULL;
rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
if (!rq)
return NULL;
} else {
rq = rq_list_peek(&plug->cached_rq);
if (!rq || rq->q != q)
return NULL;
if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
return NULL;
if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
return NULL;
plug->cached_rq = rq_list_next(rq);
blk_mq_rq_time_init(rq, 0);
}
rq->cmd_flags = opf;
INIT_LIST_HEAD(&rq->queuelist);
return rq;
}
struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
blk_mq_req_flags_t flags)
{
struct request *rq;
rq = blk_mq_alloc_cached_request(q, opf, flags);
if (!rq) {
struct blk_mq_alloc_data data = {
.q = q,
.flags = flags,
.cmd_flags = opf,
.nr_tags = 1,
};
int ret;
ret = blk_queue_enter(q, flags);
if (ret)
return ERR_PTR(ret);
rq = __blk_mq_alloc_requests(&data);
if (!rq)
goto out_queue_exit;
}
rq->__data_len = 0;
rq->__sector = (sector_t) -1;
rq->bio = rq->biotail = NULL;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
return rq;
out_queue_exit:
blk_queue_exit(q);
return ERR_PTR(-EWOULDBLOCK);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
EXPORT_SYMBOL(blk_mq_alloc_request);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
{
struct blk_mq_alloc_data data = {
.q = q,
.flags = flags,
.cmd_flags = opf,
.nr_tags = 1,
};
u64 alloc_time_ns = 0;
struct request *rq;
unsigned int cpu;
unsigned int tag;
int ret;
/* alloc_time includes depth and tag waits */
if (blk_queue_rq_alloc_time(q))
alloc_time_ns = blk_time_get_ns();
/*
* If the tag allocator sleeps we could get an allocation for a
* different hardware context. No need to complicate the low level
* allocator for this for the rare use case of a command tied to
* a specific queue.
*/
if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
return ERR_PTR(-EINVAL);
if (hctx_idx >= q->nr_hw_queues)
return ERR_PTR(-EIO);
block, scsi: Make SCSI quiesce and resume work reliably The contexts from which a SCSI device can be quiesced or resumed are: * Writing into /sys/class/scsi_device/*/device/state. * SCSI parallel (SPI) domain validation. * The SCSI device power management methods. See also scsi_bus_pm_ops. It is essential during suspend and resume that neither the filesystem state nor the filesystem metadata in RAM changes. This is why while the hibernation image is being written or restored that SCSI devices are quiesced. The SCSI core quiesces devices through scsi_device_quiesce() and scsi_device_resume(). In the SDEV_QUIESCE state execution of non-preempt requests is deferred. This is realized by returning BLKPREP_DEFER from inside scsi_prep_state_check() for quiesced SCSI devices. Avoid that a full queue prevents power management requests to be submitted by deferring allocation of non-preempt requests for devices in the quiesced state. This patch has been tested by running the following commands and by verifying that after each resume the fio job was still running: for ((i=0; i<10; i++)); do ( cd /sys/block/md0/md && while true; do [ "$(<sync_action)" = "idle" ] && echo check > sync_action sleep 1 done ) & pids=($!) for d in /sys/class/block/sd*[a-z]; do bdev=${d#/sys/class/block/} hcil=$(readlink "$d/device") hcil=${hcil#../../../} echo 4 > "$d/queue/nr_requests" echo 1 > "/sys/class/scsi_device/$hcil/device/queue_depth" fio --name="$bdev" --filename="/dev/$bdev" --buffered=0 --bs=512 \ --rw=randread --ioengine=libaio --numjobs=4 --iodepth=16 \ --iodepth_batch=1 --thread --loops=$((2**31)) & pids+=($!) done sleep 1 echo "$(date) Hibernating ..." >>hibernate-test-log.txt systemctl hibernate sleep 10 kill "${pids[@]}" echo idle > /sys/block/md0/md/sync_action wait echo "$(date) Done." >>hibernate-test-log.txt done Reported-by: Oleksandr Natalenko <oleksandr@natalenko.name> References: "I/O hangs after resuming from suspend-to-ram" (https://marc.info/?l=linux-block&m=150340235201348). Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com> Reviewed-by: Hannes Reinecke <hare@suse.com> Tested-by: Martin Steigerwald <martin@lichtvoll.de> Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Cc: Martin K. Petersen <martin.petersen@oracle.com> Cc: Ming Lei <ming.lei@redhat.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Johannes Thumshirn <jthumshirn@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2017-11-09 18:49:58 +00:00
ret = blk_queue_enter(q, flags);
if (ret)
return ERR_PTR(ret);
/*
* Check if the hardware context is actually mapped to anything.
* If not tell the caller that it should skip this queue.
*/
ret = -EXDEV;
data.hctx = xa_load(&q->hctx_table, hctx_idx);
if (!blk_mq_hw_queue_mapped(data.hctx))
goto out_queue_exit;
cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
if (cpu >= nr_cpu_ids)
goto out_queue_exit;
data.ctx = __blk_mq_get_ctx(q, cpu);
if (q->elevator)
data.rq_flags |= RQF_SCHED_TAGS;
else
blk_mq_tag_busy(data.hctx);
if (flags & BLK_MQ_REQ_RESERVED)
data.rq_flags |= RQF_RESV;
ret = -EWOULDBLOCK;
tag = blk_mq_get_tag(&data);
if (tag == BLK_MQ_NO_TAG)
goto out_queue_exit;
if (!(data.rq_flags & RQF_SCHED_TAGS))
blk_mq_inc_active_requests(data.hctx);
rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
blk_mq_rq_time_init(rq, alloc_time_ns);
rq->__data_len = 0;
rq->__sector = (sector_t) -1;
rq->bio = rq->biotail = NULL;
return rq;
out_queue_exit:
blk_queue_exit(q);
return ERR_PTR(ret);
}
EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
blk-mq: release scheduler resource when request completes Chuck reported [1] an IO hang problem on NFS exports that reside on SATA devices and bisected to commit 615939a2ae73 ("blk-mq: defer to the normal submission path for post-flush requests"). We analysed the IO hang problem, found there are two postflush requests waiting for each other. The first postflush request completed the REQ_FSEQ_DATA sequence, so go to the REQ_FSEQ_POSTFLUSH sequence and added in the flush pending list, but failed to blk_kick_flush() because of the second postflush request which is inflight waiting in scheduler queue. The second postflush waiting in scheduler queue can't be dispatched because the first postflush hasn't released scheduler resource even though it has completed by itself. Fix it by releasing scheduler resource when the first postflush request completed, so the second postflush can be dispatched and completed, then make blk_kick_flush() succeed. While at it, remove the check for e->ops.finish_request, as all schedulers set that. Reaffirm this requirement by adding a WARN_ON_ONCE() at scheduler registration time, just like we do for insert_requests and dispatch_request. [1] https://lore.kernel.org/all/7A57C7AE-A51A-4254-888B-FE15CA21F9E9@oracle.com/ Link: https://lore.kernel.org/linux-block/20230819031206.2744005-1-chengming.zhou@linux.dev/ Reported-by: kernel test robot <oliver.sang@intel.com> Closes: https://lore.kernel.org/oe-lkp/202308172100.8ce4b853-oliver.sang@intel.com Fixes: 615939a2ae73 ("blk-mq: defer to the normal submission path for post-flush requests") Reported-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com> Tested-by: Chuck Lever <chuck.lever@oracle.com> Link: https://lore.kernel.org/r/20230813152325.3017343-1-chengming.zhou@linux.dev [axboe: folded in incremental fix and added tags] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-08-13 15:23:25 +00:00
static void blk_mq_finish_request(struct request *rq)
{
struct request_queue *q = rq->q;
blk_zone_finish_request(rq);
blk-mq: release scheduler resource when request completes Chuck reported [1] an IO hang problem on NFS exports that reside on SATA devices and bisected to commit 615939a2ae73 ("blk-mq: defer to the normal submission path for post-flush requests"). We analysed the IO hang problem, found there are two postflush requests waiting for each other. The first postflush request completed the REQ_FSEQ_DATA sequence, so go to the REQ_FSEQ_POSTFLUSH sequence and added in the flush pending list, but failed to blk_kick_flush() because of the second postflush request which is inflight waiting in scheduler queue. The second postflush waiting in scheduler queue can't be dispatched because the first postflush hasn't released scheduler resource even though it has completed by itself. Fix it by releasing scheduler resource when the first postflush request completed, so the second postflush can be dispatched and completed, then make blk_kick_flush() succeed. While at it, remove the check for e->ops.finish_request, as all schedulers set that. Reaffirm this requirement by adding a WARN_ON_ONCE() at scheduler registration time, just like we do for insert_requests and dispatch_request. [1] https://lore.kernel.org/all/7A57C7AE-A51A-4254-888B-FE15CA21F9E9@oracle.com/ Link: https://lore.kernel.org/linux-block/20230819031206.2744005-1-chengming.zhou@linux.dev/ Reported-by: kernel test robot <oliver.sang@intel.com> Closes: https://lore.kernel.org/oe-lkp/202308172100.8ce4b853-oliver.sang@intel.com Fixes: 615939a2ae73 ("blk-mq: defer to the normal submission path for post-flush requests") Reported-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com> Tested-by: Chuck Lever <chuck.lever@oracle.com> Link: https://lore.kernel.org/r/20230813152325.3017343-1-chengming.zhou@linux.dev [axboe: folded in incremental fix and added tags] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-08-13 15:23:25 +00:00
if (rq->rq_flags & RQF_USE_SCHED) {
q->elevator->type->ops.finish_request(rq);
/*
* For postflush request that may need to be
* completed twice, we should clear this flag
* to avoid double finish_request() on the rq.
*/
rq->rq_flags &= ~RQF_USE_SCHED;
}
}
static void __blk_mq_free_request(struct request *rq)
{
struct request_queue *q = rq->q;
struct blk_mq_ctx *ctx = rq->mq_ctx;
struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
const int sched_tag = rq->internal_tag;
block: Inline encryption support for blk-mq We must have some way of letting a storage device driver know what encryption context it should use for en/decrypting a request. However, it's the upper layers (like the filesystem/fscrypt) that know about and manages encryption contexts. As such, when the upper layer submits a bio to the block layer, and this bio eventually reaches a device driver with support for inline encryption, the device driver will need to have been told the encryption context for that bio. We want to communicate the encryption context from the upper layer to the storage device along with the bio, when the bio is submitted to the block layer. To do this, we add a struct bio_crypt_ctx to struct bio, which can represent an encryption context (note that we can't use the bi_private field in struct bio to do this because that field does not function to pass information across layers in the storage stack). We also introduce various functions to manipulate the bio_crypt_ctx and make the bio/request merging logic aware of the bio_crypt_ctx. We also make changes to blk-mq to make it handle bios with encryption contexts. blk-mq can merge many bios into the same request. These bios need to have contiguous data unit numbers (the necessary changes to blk-merge are also made to ensure this) - as such, it suffices to keep the data unit number of just the first bio, since that's all a storage driver needs to infer the data unit number to use for each data block in each bio in a request. blk-mq keeps track of the encryption context to be used for all the bios in a request with the request's rq_crypt_ctx. When the first bio is added to an empty request, blk-mq will program the encryption context of that bio into the request_queue's keyslot manager, and store the returned keyslot in the request's rq_crypt_ctx. All the functions to operate on encryption contexts are in blk-crypto.c. Upper layers only need to call bio_crypt_set_ctx with the encryption key, algorithm and data_unit_num; they don't have to worry about getting a keyslot for each encryption context, as blk-mq/blk-crypto handles that. Blk-crypto also makes it possible for request-based layered devices like dm-rq to make use of inline encryption hardware by cloning the rq_crypt_ctx and programming a keyslot in the new request_queue when necessary. Note that any user of the block layer can submit bios with an encryption context, such as filesystems, device-mapper targets, etc. Signed-off-by: Satya Tangirala <satyat@google.com> Reviewed-by: Eric Biggers <ebiggers@google.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-14 00:37:18 +00:00
blk_crypto_free_request(rq);
blk_pm_mark_last_busy(rq);
rq->mq_hctx = NULL;
if (rq->tag != BLK_MQ_NO_TAG) {
blk_mq_dec_active_requests(hctx);
blk_mq_put_tag(hctx->tags, ctx, rq->tag);
}
if (sched_tag != BLK_MQ_NO_TAG)
blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
blk_mq_sched_restart(hctx);
blk_queue_exit(q);
}
void blk_mq_free_request(struct request *rq)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct request_queue *q = rq->q;
blk-mq: release scheduler resource when request completes Chuck reported [1] an IO hang problem on NFS exports that reside on SATA devices and bisected to commit 615939a2ae73 ("blk-mq: defer to the normal submission path for post-flush requests"). We analysed the IO hang problem, found there are two postflush requests waiting for each other. The first postflush request completed the REQ_FSEQ_DATA sequence, so go to the REQ_FSEQ_POSTFLUSH sequence and added in the flush pending list, but failed to blk_kick_flush() because of the second postflush request which is inflight waiting in scheduler queue. The second postflush waiting in scheduler queue can't be dispatched because the first postflush hasn't released scheduler resource even though it has completed by itself. Fix it by releasing scheduler resource when the first postflush request completed, so the second postflush can be dispatched and completed, then make blk_kick_flush() succeed. While at it, remove the check for e->ops.finish_request, as all schedulers set that. Reaffirm this requirement by adding a WARN_ON_ONCE() at scheduler registration time, just like we do for insert_requests and dispatch_request. [1] https://lore.kernel.org/all/7A57C7AE-A51A-4254-888B-FE15CA21F9E9@oracle.com/ Link: https://lore.kernel.org/linux-block/20230819031206.2744005-1-chengming.zhou@linux.dev/ Reported-by: kernel test robot <oliver.sang@intel.com> Closes: https://lore.kernel.org/oe-lkp/202308172100.8ce4b853-oliver.sang@intel.com Fixes: 615939a2ae73 ("blk-mq: defer to the normal submission path for post-flush requests") Reported-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com> Tested-by: Chuck Lever <chuck.lever@oracle.com> Link: https://lore.kernel.org/r/20230813152325.3017343-1-chengming.zhou@linux.dev [axboe: folded in incremental fix and added tags] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-08-13 15:23:25 +00:00
blk_mq_finish_request(rq);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
laptop_io_completion(q->disk->bdi);
rq_qos_done(q, rq);
WRITE_ONCE(rq->state, MQ_RQ_IDLE);
if (req_ref_put_and_test(rq))
__blk_mq_free_request(rq);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
EXPORT_SYMBOL_GPL(blk_mq_free_request);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
void blk_mq_free_plug_rqs(struct blk_plug *plug)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct request *rq;
while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
blk_mq_free_request(rq);
}
void blk_dump_rq_flags(struct request *rq, char *msg)
{
printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
rq->q->disk ? rq->q->disk->disk_name : "?",
(__force unsigned long long) rq->cmd_flags);
printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
(unsigned long long)blk_rq_pos(rq),
blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
printk(KERN_INFO " bio %p, biotail %p, len %u\n",
rq->bio, rq->biotail, blk_rq_bytes(rq));
}
EXPORT_SYMBOL(blk_dump_rq_flags);
static void blk_account_io_completion(struct request *req, unsigned int bytes)
{
if (req->part && blk_do_io_stat(req)) {
const int sgrp = op_stat_group(req_op(req));
part_stat_lock();
part_stat_add(req->part, sectors[sgrp], bytes >> 9);
part_stat_unlock();
}
}
static void blk_print_req_error(struct request *req, blk_status_t status)
{
printk_ratelimited(KERN_ERR
"%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
"phys_seg %u prio class %u\n",
blk_status_to_str(status),
req->q->disk ? req->q->disk->disk_name : "?",
blk_rq_pos(req), (__force u32)req_op(req),
blk_op_str(req_op(req)),
(__force u32)(req->cmd_flags & ~REQ_OP_MASK),
req->nr_phys_segments,
IOPRIO_PRIO_CLASS(req->ioprio));
}
/*
* Fully end IO on a request. Does not support partial completions, or
* errors.
*/
static void blk_complete_request(struct request *req)
{
const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
int total_bytes = blk_rq_bytes(req);
struct bio *bio = req->bio;
trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
if (!bio)
return;
if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
blk_integrity_complete(req, total_bytes);
/*
* Upper layers may call blk_crypto_evict_key() anytime after the last
* bio_endio(). Therefore, the keyslot must be released before that.
*/
blk_crypto_rq_put_keyslot(req);
blk_account_io_completion(req, total_bytes);
do {
struct bio *next = bio->bi_next;
/* Completion has already been traced */
bio_clear_flag(bio, BIO_TRACE_COMPLETION);
blk_zone_update_request_bio(req, bio);
if (!is_flush)
bio_endio(bio);
bio = next;
} while (bio);
/*
* Reset counters so that the request stacking driver
* can find how many bytes remain in the request
* later.
*/
if (!req->end_io) {
req->bio = NULL;
req->__data_len = 0;
}
}
/**
* blk_update_request - Complete multiple bytes without completing the request
* @req: the request being processed
* @error: block status code
* @nr_bytes: number of bytes to complete for @req
*
* Description:
* Ends I/O on a number of bytes attached to @req, but doesn't complete
* the request structure even if @req doesn't have leftover.
* If @req has leftover, sets it up for the next range of segments.
*
* Passing the result of blk_rq_bytes() as @nr_bytes guarantees
* %false return from this function.
*
* Note:
* The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
* except in the consistency check at the end of this function.
*
* Return:
* %false - this request doesn't have any more data
* %true - this request has more data
**/
bool blk_update_request(struct request *req, blk_status_t error,
unsigned int nr_bytes)
{
bool is_flush = req->rq_flags & RQF_FLUSH_SEQ;
bool quiet = req->rq_flags & RQF_QUIET;
int total_bytes;
trace_block_rq_complete(req, error, nr_bytes);
if (!req->bio)
return false;
if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
error == BLK_STS_OK)
blk_integrity_complete(req, nr_bytes);
/*
* Upper layers may call blk_crypto_evict_key() anytime after the last
* bio_endio(). Therefore, the keyslot must be released before that.
*/
if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
__blk_crypto_rq_put_keyslot(req);
if (unlikely(error && !blk_rq_is_passthrough(req) && !quiet) &&
!test_bit(GD_DEAD, &req->q->disk->state)) {
blk_print_req_error(req, error);
block: introduce block_rq_error tracepoint Currently, rasdaemon uses the existing tracepoint block_rq_complete and filters out non-error cases in order to capture block disk errors. But there are a few problems with this approach: 1. Even kernel trace filter could do the filtering work, there is still some overhead after we enable this tracepoint. 2. The filter is merely based on errno, which does not align with kernel logic to check the errors for print_req_error(). 3. block_rq_complete only provides dev major and minor to identify the block device, it is not convenient to use in user-space. So introduce a new tracepoint block_rq_error just for the error case. With this patch, rasdaemon could switch to block_rq_error. Since the new tracepoint has the similar implementation with block_rq_complete, so move the existing code from TRACE_EVENT block_rq_complete() into new event class block_rq_completion(). Then add event for block_rq_complete and block_rq_err respectively from the newly created event class per the suggestion from Chaitanya Kulkarni. Cc: Jens Axboe <axboe@kernel.dk> Cc: Christoph Hellwig <hch@infradead.org> Reviewed-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com> Signed-off-by: Chaitanya Kulkarni <kch@nvidia.com> Signed-off-by: Yang Shi <shy828301@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Link: https://lore.kernel.org/r/20220210225222.260069-1-shy828301@gmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-02-10 22:52:22 +00:00
trace_block_rq_error(req, error, nr_bytes);
}
blk_account_io_completion(req, nr_bytes);
total_bytes = 0;
while (req->bio) {
struct bio *bio = req->bio;
unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
if (unlikely(error))
bio->bi_status = error;
if (bio_bytes == bio->bi_iter.bi_size) {
req->bio = bio->bi_next;
block: Implement zone append emulation Given that zone write plugging manages all writes to zones of a zoned block device and tracks the write pointer position of all zones that are not full nor empty, emulating zone append operations using regular writes can be implemented generically, without relying on the underlying device driver to implement such emulation. This is needed for devices that do not natively support the zone append command (e.g. SMR hard-disks). A device may request zone append emulation by setting its max_zone_append_sectors queue limit to 0. For such device, the function blk_zone_wplug_prepare_bio() changes zone append BIOs into non-mergeable regular write BIOs. Modified zone append BIOs are flagged with the new BIO flag BIO_EMULATES_ZONE_APPEND. This flag is checked on completion of the BIO in blk_zone_write_plug_bio_endio() to restore the original REQ_OP_ZONE_APPEND operation code of the BIO. The block layer internal inline helper function bio_is_zone_append() is added to test if a BIO is either a native zone append operation (REQ_OP_ZONE_APPEND operation code) or if it is flagged with BIO_EMULATES_ZONE_APPEND. Given that both native and emulated zone append BIO completion handling should be similar, The functions blk_update_request() and blk_zone_complete_request_bio() are modified to use bio_is_zone_append() to execute blk_zone_update_request_bio() for both native and emulated zone append operations. This commit contains contributions from Christoph Hellwig <hch@lst.de>. Signed-off-by: Damien Le Moal <dlemoal@kernel.org> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Tested-by: Hans Holmberg <hans.holmberg@wdc.com> Tested-by: Dennis Maisenbacher <dennis.maisenbacher@wdc.com> Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com> Link: https://lore.kernel.org/r/20240408014128.205141-11-dlemoal@kernel.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2024-04-08 01:41:10 +00:00
} else if (bio_is_zone_append(bio) && error == BLK_STS_OK) {
/*
* Partial zone append completions cannot be supported
* as the BIO fragments may end up not being written
* sequentially.
*/
bio->bi_status = BLK_STS_IOERR;
}
/* Completion has already been traced */
bio_clear_flag(bio, BIO_TRACE_COMPLETION);
if (unlikely(quiet))
bio_set_flag(bio, BIO_QUIET);
bio_advance(bio, bio_bytes);
/* Don't actually finish bio if it's part of flush sequence */
if (!bio->bi_iter.bi_size) {
blk_zone_update_request_bio(req, bio);
if (!is_flush)
bio_endio(bio);
}
total_bytes += bio_bytes;
nr_bytes -= bio_bytes;
if (!nr_bytes)
break;
}
/*
* completely done
*/
if (!req->bio) {
/*
* Reset counters so that the request stacking driver
* can find how many bytes remain in the request
* later.
*/
req->__data_len = 0;
return false;
}
req->__data_len -= total_bytes;
/* update sector only for requests with clear definition of sector */
if (!blk_rq_is_passthrough(req))
req->__sector += total_bytes >> 9;
/* mixed attributes always follow the first bio */
if (req->rq_flags & RQF_MIXED_MERGE) {
req->cmd_flags &= ~REQ_FAILFAST_MASK;
req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
}
if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
/*
* If total number of sectors is less than the first segment
* size, something has gone terribly wrong.
*/
if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
blk_dump_rq_flags(req, "request botched");
req->__data_len = blk_rq_cur_bytes(req);
}
/* recalculate the number of segments */
req->nr_phys_segments = blk_recalc_rq_segments(req);
}
return true;
}
EXPORT_SYMBOL_GPL(blk_update_request);
static inline void blk_account_io_done(struct request *req, u64 now)
{
trace_block_io_done(req);
/*
* Account IO completion. flush_rq isn't accounted as a
* normal IO on queueing nor completion. Accounting the
* containing request is enough.
*/
if (blk_do_io_stat(req) && req->part &&
!(req->rq_flags & RQF_FLUSH_SEQ)) {
const int sgrp = op_stat_group(req_op(req));
part_stat_lock();
update_io_ticks(req->part, jiffies, true);
part_stat_inc(req->part, ios[sgrp]);
part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
block: support to account io_ticks precisely Currently, io_ticks is accounted based on sampling, specifically update_io_ticks() will always account io_ticks by 1 jiffies from bdev_start_io_acct()/blk_account_io_start(), and the result can be inaccurate, for example(HZ is 250): Test script: fio -filename=/dev/sda -bs=4k -rw=write -direct=1 -name=test -thinktime=4ms Test result: util is about 90%, while the disk is really idle. This behaviour is introduced by commit 5b18b5a73760 ("block: delete part_round_stats and switch to less precise counting"), however, there was a key point that is missed that this patch also improve performance a lot: Before the commit: part_round_stats: if (part->stamp != now) stats |= 1; part_in_flight() -> there can be lots of task here in 1 jiffies. part_round_stats_single() __part_stat_add() part->stamp = now; After the commit: update_io_ticks: stamp = part->bd_stamp; if (time_after(now, stamp)) if (try_cmpxchg()) __part_stat_add() -> only one task can reach here in 1 jiffies. Hence in order to account io_ticks precisely, we only need to know if there are IO inflight at most once in one jiffies. Noted that for rq-based device, iterating tags should not be used here because 'tags->lock' is grabbed in blk_mq_find_and_get_req(), hence part_stat_lock_inc/dec() and part_in_flight() is used to trace inflight. The additional overhead is quite little: - per cpu add/dec for each IO for rq-based device; - per cpu sum for each jiffies; And it's verified by null-blk that there are no performance degration under heavy IO pressure. Fixes: 5b18b5a73760 ("block: delete part_round_stats and switch to less precise counting") Signed-off-by: Yu Kuai <yukuai3@huawei.com> Link: https://lore.kernel.org/r/20240509123717.3223892-2-yukuai1@huaweicloud.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2024-05-09 12:37:16 +00:00
part_stat_local_dec(req->part,
in_flight[op_is_write(req_op(req))]);
part_stat_unlock();
}
}
static inline void blk_account_io_start(struct request *req)
{
trace_block_io_start(req);
if (blk_do_io_stat(req)) {
/*
* All non-passthrough requests are created from a bio with one
* exception: when a flush command that is part of a flush sequence
* generated by the state machine in blk-flush.c is cloned onto the
* lower device by dm-multipath we can get here without a bio.
*/
if (req->bio)
req->part = req->bio->bi_bdev;
else
req->part = req->q->disk->part0;
part_stat_lock();
update_io_ticks(req->part, jiffies, false);
block: support to account io_ticks precisely Currently, io_ticks is accounted based on sampling, specifically update_io_ticks() will always account io_ticks by 1 jiffies from bdev_start_io_acct()/blk_account_io_start(), and the result can be inaccurate, for example(HZ is 250): Test script: fio -filename=/dev/sda -bs=4k -rw=write -direct=1 -name=test -thinktime=4ms Test result: util is about 90%, while the disk is really idle. This behaviour is introduced by commit 5b18b5a73760 ("block: delete part_round_stats and switch to less precise counting"), however, there was a key point that is missed that this patch also improve performance a lot: Before the commit: part_round_stats: if (part->stamp != now) stats |= 1; part_in_flight() -> there can be lots of task here in 1 jiffies. part_round_stats_single() __part_stat_add() part->stamp = now; After the commit: update_io_ticks: stamp = part->bd_stamp; if (time_after(now, stamp)) if (try_cmpxchg()) __part_stat_add() -> only one task can reach here in 1 jiffies. Hence in order to account io_ticks precisely, we only need to know if there are IO inflight at most once in one jiffies. Noted that for rq-based device, iterating tags should not be used here because 'tags->lock' is grabbed in blk_mq_find_and_get_req(), hence part_stat_lock_inc/dec() and part_in_flight() is used to trace inflight. The additional overhead is quite little: - per cpu add/dec for each IO for rq-based device; - per cpu sum for each jiffies; And it's verified by null-blk that there are no performance degration under heavy IO pressure. Fixes: 5b18b5a73760 ("block: delete part_round_stats and switch to less precise counting") Signed-off-by: Yu Kuai <yukuai3@huawei.com> Link: https://lore.kernel.org/r/20240509123717.3223892-2-yukuai1@huaweicloud.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2024-05-09 12:37:16 +00:00
part_stat_local_inc(req->part,
in_flight[op_is_write(req_op(req))]);
part_stat_unlock();
}
}
static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
if (rq->rq_flags & RQF_STATS)
blk_stat_add(rq, now);
blk_mq_sched_completed_request(rq, now);
blk_account_io_done(rq, now);
}
inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
{
if (blk_mq_need_time_stamp(rq))
__blk_mq_end_request_acct(rq, blk_time_get_ns());
blk-mq: release scheduler resource when request completes Chuck reported [1] an IO hang problem on NFS exports that reside on SATA devices and bisected to commit 615939a2ae73 ("blk-mq: defer to the normal submission path for post-flush requests"). We analysed the IO hang problem, found there are two postflush requests waiting for each other. The first postflush request completed the REQ_FSEQ_DATA sequence, so go to the REQ_FSEQ_POSTFLUSH sequence and added in the flush pending list, but failed to blk_kick_flush() because of the second postflush request which is inflight waiting in scheduler queue. The second postflush waiting in scheduler queue can't be dispatched because the first postflush hasn't released scheduler resource even though it has completed by itself. Fix it by releasing scheduler resource when the first postflush request completed, so the second postflush can be dispatched and completed, then make blk_kick_flush() succeed. While at it, remove the check for e->ops.finish_request, as all schedulers set that. Reaffirm this requirement by adding a WARN_ON_ONCE() at scheduler registration time, just like we do for insert_requests and dispatch_request. [1] https://lore.kernel.org/all/7A57C7AE-A51A-4254-888B-FE15CA21F9E9@oracle.com/ Link: https://lore.kernel.org/linux-block/20230819031206.2744005-1-chengming.zhou@linux.dev/ Reported-by: kernel test robot <oliver.sang@intel.com> Closes: https://lore.kernel.org/oe-lkp/202308172100.8ce4b853-oliver.sang@intel.com Fixes: 615939a2ae73 ("blk-mq: defer to the normal submission path for post-flush requests") Reported-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com> Tested-by: Chuck Lever <chuck.lever@oracle.com> Link: https://lore.kernel.org/r/20230813152325.3017343-1-chengming.zhou@linux.dev [axboe: folded in incremental fix and added tags] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-08-13 15:23:25 +00:00
blk_mq_finish_request(rq);
if (rq->end_io) {
rq_qos_done(rq->q, rq);
if (rq->end_io(rq, error) == RQ_END_IO_FREE)
blk_mq_free_request(rq);
} else {
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
blk_mq_free_request(rq);
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
EXPORT_SYMBOL(__blk_mq_end_request);
void blk_mq_end_request(struct request *rq, blk_status_t error)
{
if (blk_update_request(rq, error, blk_rq_bytes(rq)))
BUG();
__blk_mq_end_request(rq, error);
}
EXPORT_SYMBOL(blk_mq_end_request);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
#define TAG_COMP_BATCH 32
static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
int *tag_array, int nr_tags)
{
struct request_queue *q = hctx->queue;
blk_mq_sub_active_requests(hctx, nr_tags);
blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
percpu_ref_put_many(&q->q_usage_counter, nr_tags);
}
void blk_mq_end_request_batch(struct io_comp_batch *iob)
{
int tags[TAG_COMP_BATCH], nr_tags = 0;
struct blk_mq_hw_ctx *cur_hctx = NULL;
struct request *rq;
u64 now = 0;
if (iob->need_ts)
now = blk_time_get_ns();
while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
prefetch(rq->bio);
prefetch(rq->rq_next);
blk_complete_request(rq);
if (iob->need_ts)
__blk_mq_end_request_acct(rq, now);
blk-mq: release scheduler resource when request completes Chuck reported [1] an IO hang problem on NFS exports that reside on SATA devices and bisected to commit 615939a2ae73 ("blk-mq: defer to the normal submission path for post-flush requests"). We analysed the IO hang problem, found there are two postflush requests waiting for each other. The first postflush request completed the REQ_FSEQ_DATA sequence, so go to the REQ_FSEQ_POSTFLUSH sequence and added in the flush pending list, but failed to blk_kick_flush() because of the second postflush request which is inflight waiting in scheduler queue. The second postflush waiting in scheduler queue can't be dispatched because the first postflush hasn't released scheduler resource even though it has completed by itself. Fix it by releasing scheduler resource when the first postflush request completed, so the second postflush can be dispatched and completed, then make blk_kick_flush() succeed. While at it, remove the check for e->ops.finish_request, as all schedulers set that. Reaffirm this requirement by adding a WARN_ON_ONCE() at scheduler registration time, just like we do for insert_requests and dispatch_request. [1] https://lore.kernel.org/all/7A57C7AE-A51A-4254-888B-FE15CA21F9E9@oracle.com/ Link: https://lore.kernel.org/linux-block/20230819031206.2744005-1-chengming.zhou@linux.dev/ Reported-by: kernel test robot <oliver.sang@intel.com> Closes: https://lore.kernel.org/oe-lkp/202308172100.8ce4b853-oliver.sang@intel.com Fixes: 615939a2ae73 ("blk-mq: defer to the normal submission path for post-flush requests") Reported-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com> Tested-by: Chuck Lever <chuck.lever@oracle.com> Link: https://lore.kernel.org/r/20230813152325.3017343-1-chengming.zhou@linux.dev [axboe: folded in incremental fix and added tags] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-08-13 15:23:25 +00:00
blk_mq_finish_request(rq);
rq_qos_done(rq->q, rq);
/*
* If end_io handler returns NONE, then it still has
* ownership of the request.
*/
if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
continue;
WRITE_ONCE(rq->state, MQ_RQ_IDLE);
if (!req_ref_put_and_test(rq))
continue;
blk_crypto_free_request(rq);
blk_pm_mark_last_busy(rq);
if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
if (cur_hctx)
blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
nr_tags = 0;
cur_hctx = rq->mq_hctx;
}
tags[nr_tags++] = rq->tag;
}
if (nr_tags)
blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
}
EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
static void blk_complete_reqs(struct llist_head *list)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct llist_node *entry = llist_reverse_order(llist_del_all(list));
struct request *rq, *next;
llist_for_each_entry_safe(rq, next, entry, ipi_list)
rq->q->mq_ops->complete(rq);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
static __latent_entropy void blk_done_softirq(struct softirq_action *h)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
}
static int blk_softirq_cpu_dead(unsigned int cpu)
{
blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
return 0;
}
static void __blk_mq_complete_request_remote(void *data)
{
__raise_softirq_irqoff(BLOCK_SOFTIRQ);
}
static inline bool blk_mq_complete_need_ipi(struct request *rq)
{
int cpu = raw_smp_processor_id();
if (!IS_ENABLED(CONFIG_SMP) ||
!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
return false;
/*
* With force threaded interrupts enabled, raising softirq from an SMP
* function call will always result in waking the ksoftirqd thread.
* This is probably worse than completing the request on a different
* cache domain.
*/
if (force_irqthreads())
return false;
block/blk-mq: Don't complete locally if capacities are different The logic in blk_mq_complete_need_ipi() assumes SMP systems where all CPUs have equal compute capacities and only LLC cache can make a different on perceived performance. But this assumption falls apart on HMP systems where LLC is shared, but the CPUs have different capacities. Staying local then can have a big performance impact if the IO request was done from a CPU with higher capacity but the interrupt is serviced on a lower capacity CPU. Use the new cpus_equal_capacity() function to check if we need to send an IPI. Without the patch I see the BLOCK softirq always running on little cores (where the hardirq is serviced). With it I can see it running on all cores. This was noticed after the topology change [1] where now on a big.LITTLE we truly get that the LLC is shared between all cores where as in the past it was being misrepresented for historical reasons. The logic exposed a missing dependency on capacities for such systems where there can be a big performance difference between the CPUs. This of course introduced a noticeable change in behavior depending on how the topology is presented. Leading to regressions in some workloads as the performance of the BLOCK softirq on littles can be noticeably worse on some platforms. Worth noting that we could have checked for capacities being greater than or equal instead for equality. This will lead to favouring higher performance always. But opted for equality instead to match the performance of the requester without making an assumption that can lead to power trade-offs which these systems tend to be sensitive about. If the requester would like to run faster, it's better to rely on the scheduler to give the IO requester via some facility to run on a faster core; and then if the interrupt triggered on a CPU with different capacity we'll make sure to match the performance the requester is supposed to run at. [1] https://lpc.events/event/16/contributions/1342/attachments/962/1883/LPC-2022-Android-MC-Phantom-Domains.pdf Signed-off-by: Qais Yousef <qyousef@layalina.io> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Link: https://lore.kernel.org/r/20240223155749.2958009-3-qyousef@layalina.io Signed-off-by: Jens Axboe <axboe@kernel.dk>
2024-02-23 15:57:49 +00:00
/* same CPU or cache domain and capacity? Complete locally */
if (cpu == rq->mq_ctx->cpu ||
(!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
block/blk-mq: Don't complete locally if capacities are different The logic in blk_mq_complete_need_ipi() assumes SMP systems where all CPUs have equal compute capacities and only LLC cache can make a different on perceived performance. But this assumption falls apart on HMP systems where LLC is shared, but the CPUs have different capacities. Staying local then can have a big performance impact if the IO request was done from a CPU with higher capacity but the interrupt is serviced on a lower capacity CPU. Use the new cpus_equal_capacity() function to check if we need to send an IPI. Without the patch I see the BLOCK softirq always running on little cores (where the hardirq is serviced). With it I can see it running on all cores. This was noticed after the topology change [1] where now on a big.LITTLE we truly get that the LLC is shared between all cores where as in the past it was being misrepresented for historical reasons. The logic exposed a missing dependency on capacities for such systems where there can be a big performance difference between the CPUs. This of course introduced a noticeable change in behavior depending on how the topology is presented. Leading to regressions in some workloads as the performance of the BLOCK softirq on littles can be noticeably worse on some platforms. Worth noting that we could have checked for capacities being greater than or equal instead for equality. This will lead to favouring higher performance always. But opted for equality instead to match the performance of the requester without making an assumption that can lead to power trade-offs which these systems tend to be sensitive about. If the requester would like to run faster, it's better to rely on the scheduler to give the IO requester via some facility to run on a faster core; and then if the interrupt triggered on a CPU with different capacity we'll make sure to match the performance the requester is supposed to run at. [1] https://lpc.events/event/16/contributions/1342/attachments/962/1883/LPC-2022-Android-MC-Phantom-Domains.pdf Signed-off-by: Qais Yousef <qyousef@layalina.io> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Link: https://lore.kernel.org/r/20240223155749.2958009-3-qyousef@layalina.io Signed-off-by: Jens Axboe <axboe@kernel.dk>
2024-02-23 15:57:49 +00:00
cpus_share_cache(cpu, rq->mq_ctx->cpu) &&
cpus_equal_capacity(cpu, rq->mq_ctx->cpu)))
return false;
/* don't try to IPI to an offline CPU */
return cpu_online(rq->mq_ctx->cpu);
}
static void blk_mq_complete_send_ipi(struct request *rq)
{
unsigned int cpu;
cpu = rq->mq_ctx->cpu;
if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
}
static void blk_mq_raise_softirq(struct request *rq)
{
struct llist_head *list;
preempt_disable();
list = this_cpu_ptr(&blk_cpu_done);
if (llist_add(&rq->ipi_list, list))
raise_softirq(BLOCK_SOFTIRQ);
preempt_enable();
}
bool blk_mq_complete_request_remote(struct request *rq)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
/*
* For request which hctx has only one ctx mapping,
* or a polled request, always complete locally,
* it's pointless to redirect the completion.
*/
if ((rq->mq_hctx->nr_ctx == 1 &&
rq->mq_ctx->cpu == raw_smp_processor_id()) ||
rq->cmd_flags & REQ_POLLED)
return false;
if (blk_mq_complete_need_ipi(rq)) {
blk_mq_complete_send_ipi(rq);
return true;
}
if (rq->q->nr_hw_queues == 1) {
blk_mq_raise_softirq(rq);
return true;
}
return false;
}
EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
/**
* blk_mq_complete_request - end I/O on a request
* @rq: the request being processed
*
* Description:
* Complete a request by scheduling the ->complete_rq operation.
**/
void blk_mq_complete_request(struct request *rq)
{
if (!blk_mq_complete_request_remote(rq))
rq->q->mq_ops->complete(rq);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
EXPORT_SYMBOL(blk_mq_complete_request);
/**
* blk_mq_start_request - Start processing a request
* @rq: Pointer to request to be started
*
* Function used by device drivers to notify the block layer that a request
* is going to be processed now, so blk layer can do proper initializations
* such as starting the timeout timer.
*/
void blk_mq_start_request(struct request *rq)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct request_queue *q = rq->q;
trace_block_rq_issue(rq);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
!blk_rq_is_passthrough(rq)) {
rq->io_start_time_ns = blk_time_get_ns();
rq->stats_sectors = blk_rq_sectors(rq);
rq->rq_flags |= RQF_STATS;
rq_qos_issue(q, rq);
}
blk-mq: replace timeout synchronization with a RCU and generation based scheme Currently, blk-mq timeout path synchronizes against the usual issue/completion path using a complex scheme involving atomic bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence rules. Unfortunately, it contains quite a few holes. There's a complex dancing around REQ_ATOM_STARTED and REQ_ATOM_COMPLETE between issue/completion and timeout paths; however, they don't have a synchronization point across request recycle instances and it isn't clear what the barriers add. blk_mq_check_expired() can easily read STARTED from N-2'th iteration, deadline from N-1'th, blk_mark_rq_complete() against Nth instance. In fact, it's pretty easy to make blk_mq_check_expired() terminate a later instance of a request. If we induce 5 sec delay before time_after_eq() test in blk_mq_check_expired(), shorten the timeout to 2s, and issue back-to-back large IOs, blk-mq starts timing out requests spuriously pretty quickly. Nothing actually timed out. It just made the call on a recycle instance of a request and then terminated a later instance long after the original instance finished. The scenario isn't theoretical either. This patch replaces the broken synchronization mechanism with a RCU and generation number based one. 1. Each request has a u64 generation + state value, which can be updated only by the request owner. Whenever a request becomes in-flight, the generation number gets bumped up too. This provides the basis for the timeout path to distinguish different recycle instances of the request. Also, marking a request in-flight and setting its deadline are protected with a seqcount so that the timeout path can fetch both values coherently. 2. The timeout path fetches the generation, state and deadline. If the verdict is timeout, it records the generation into a dedicated request abortion field and does RCU wait. 3. The completion path is also protected by RCU (from the previous patch) and checks whether the current generation number and state match the abortion field. If so, it skips completion. 4. The timeout path, after RCU wait, scans requests again and terminates the ones whose generation and state still match the ones requested for abortion. By now, the timeout path knows that either the generation number and state changed if it lost the race or the completion will yield to it and can safely timeout the request. While it's more lines of code, it's conceptually simpler, doesn't depend on direct use of subtle memory ordering or coherence, and hopefully doesn't terminate the wrong instance. While this change makes REQ_ATOM_COMPLETE synchronization unnecessary between issue/complete and timeout paths, REQ_ATOM_COMPLETE isn't removed yet as it's still used in other places. Future patches will move all state tracking to the new mechanism and remove all bitops in the hot paths. Note that this patch adds a comment explaining a race condition in BLK_EH_RESET_TIMER path. The race has always been there and this patch doesn't change it. It's just documenting the existing race. v2: - Fixed BLK_EH_RESET_TIMER handling as pointed out by Jianchao. - s/request->gstate_seqc/request->gstate_seq/ as suggested by Peter. - READ_ONCE() added in blk_mq_rq_update_state() as suggested by Peter. v3: - Fixed possible extended seqcount / u64_stats_sync read looping spotted by Peter. - MQ_RQ_IDLE was incorrectly being set in complete_request instead of free_request. Fixed. v4: - Rebased on top of hctx_lock() refactoring patch. - Added comment explaining the use of hctx_lock() in completion path. v5: - Added comments requested by Bart. - Note the addition of BLK_EH_RESET_TIMER race condition in the commit message. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: "jianchao.wang" <jianchao.w.wang@oracle.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <Bart.VanAssche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-09 16:29:48 +00:00
WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
blk-mq: replace timeout synchronization with a RCU and generation based scheme Currently, blk-mq timeout path synchronizes against the usual issue/completion path using a complex scheme involving atomic bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence rules. Unfortunately, it contains quite a few holes. There's a complex dancing around REQ_ATOM_STARTED and REQ_ATOM_COMPLETE between issue/completion and timeout paths; however, they don't have a synchronization point across request recycle instances and it isn't clear what the barriers add. blk_mq_check_expired() can easily read STARTED from N-2'th iteration, deadline from N-1'th, blk_mark_rq_complete() against Nth instance. In fact, it's pretty easy to make blk_mq_check_expired() terminate a later instance of a request. If we induce 5 sec delay before time_after_eq() test in blk_mq_check_expired(), shorten the timeout to 2s, and issue back-to-back large IOs, blk-mq starts timing out requests spuriously pretty quickly. Nothing actually timed out. It just made the call on a recycle instance of a request and then terminated a later instance long after the original instance finished. The scenario isn't theoretical either. This patch replaces the broken synchronization mechanism with a RCU and generation number based one. 1. Each request has a u64 generation + state value, which can be updated only by the request owner. Whenever a request becomes in-flight, the generation number gets bumped up too. This provides the basis for the timeout path to distinguish different recycle instances of the request. Also, marking a request in-flight and setting its deadline are protected with a seqcount so that the timeout path can fetch both values coherently. 2. The timeout path fetches the generation, state and deadline. If the verdict is timeout, it records the generation into a dedicated request abortion field and does RCU wait. 3. The completion path is also protected by RCU (from the previous patch) and checks whether the current generation number and state match the abortion field. If so, it skips completion. 4. The timeout path, after RCU wait, scans requests again and terminates the ones whose generation and state still match the ones requested for abortion. By now, the timeout path knows that either the generation number and state changed if it lost the race or the completion will yield to it and can safely timeout the request. While it's more lines of code, it's conceptually simpler, doesn't depend on direct use of subtle memory ordering or coherence, and hopefully doesn't terminate the wrong instance. While this change makes REQ_ATOM_COMPLETE synchronization unnecessary between issue/complete and timeout paths, REQ_ATOM_COMPLETE isn't removed yet as it's still used in other places. Future patches will move all state tracking to the new mechanism and remove all bitops in the hot paths. Note that this patch adds a comment explaining a race condition in BLK_EH_RESET_TIMER path. The race has always been there and this patch doesn't change it. It's just documenting the existing race. v2: - Fixed BLK_EH_RESET_TIMER handling as pointed out by Jianchao. - s/request->gstate_seqc/request->gstate_seq/ as suggested by Peter. - READ_ONCE() added in blk_mq_rq_update_state() as suggested by Peter. v3: - Fixed possible extended seqcount / u64_stats_sync read looping spotted by Peter. - MQ_RQ_IDLE was incorrectly being set in complete_request instead of free_request. Fixed. v4: - Rebased on top of hctx_lock() refactoring patch. - Added comment explaining the use of hctx_lock() in completion path. v5: - Added comments requested by Bart. - Note the addition of BLK_EH_RESET_TIMER race condition in the commit message. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: "jianchao.wang" <jianchao.w.wang@oracle.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <Bart.VanAssche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-09 16:29:48 +00:00
blk_add_timer(rq);
WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
rq->mq_hctx->tags->rqs[rq->tag] = rq;
if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
blk_integrity_prepare(rq);
if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
EXPORT_SYMBOL(blk_mq_start_request);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
/*
* Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
* queues. This is important for md arrays to benefit from merging
* requests.
*/
static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
{
if (plug->multiple_queues)
return BLK_MAX_REQUEST_COUNT * 2;
return BLK_MAX_REQUEST_COUNT;
}
static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
{
struct request *last = rq_list_peek(&plug->mq_list);
if (!plug->rq_count) {
trace_block_plug(rq->q);
} else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
(!blk_queue_nomerges(rq->q) &&
blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
blk_mq_flush_plug_list(plug, false);
last = NULL;
trace_block_plug(rq->q);
}
if (!plug->multiple_queues && last && last->q != rq->q)
plug->multiple_queues = true;
/*
* Any request allocated from sched tags can't be issued to
* ->queue_rqs() directly
*/
if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
plug->has_elevator = true;
rq->rq_next = NULL;
rq_list_add(&plug->mq_list, rq);
plug->rq_count++;
}
/**
* blk_execute_rq_nowait - insert a request to I/O scheduler for execution
* @rq: request to insert
* @at_head: insert request at head or tail of queue
*
* Description:
* Insert a fully prepared request at the back of the I/O scheduler queue
* for execution. Don't wait for completion.
*
* Note:
* This function will invoke @done directly if the queue is dead.
*/
void blk_execute_rq_nowait(struct request *rq, bool at_head)
{
struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
WARN_ON(irqs_disabled());
WARN_ON(!blk_rq_is_passthrough(rq));
blk_account_io_start(rq);
if (current->plug && !at_head) {
blk_add_rq_to_plug(current->plug, rq);
return;
}
blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
block: Improve performance for BLK_MQ_F_BLOCKING drivers blk_mq_run_queue() runs the queue asynchronously if BLK_MQ_F_BLOCKING has been set. This is suboptimal since running the queue asynchronously is slower than running the queue synchronously. This patch modifies blk_mq_run_queue() as follows if BLK_MQ_F_BLOCKING has been set: - Run the queue synchronously if it is allowed to sleep. - Run the queue asynchronously if it is not allowed to sleep. Additionally, blk_mq_run_hw_queue(hctx, false) calls are modified into blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING) if the caller may be invoked from atomic context. The following caller chains have been reviewed: blk_mq_run_hw_queue(hctx, false) blk_mq_get_tag() /* may sleep, hence the functions it calls may also sleep */ blk_execute_rq() /* may sleep */ blk_mq_run_hw_queues(q, async=false) blk_freeze_queue_start() /* may sleep */ blk_mq_requeue_work() /* may sleep */ scsi_kick_queue() scsi_requeue_run_queue() /* may sleep */ scsi_run_host_queues() scsi_ioctl_reset() /* may sleep */ blk_mq_insert_requests(hctx, ctx, list, run_queue_async=false) blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list(plug, from_schedule=false) __blk_flush_plug(plug, from_schedule=false) blk_add_rq_to_plug() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_plug_issue_direct() blk_mq_flush_plug_list() /* see above */ blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list() /* see above */ blk_mq_try_issue_directly() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_try_issue_list_directly(hctx, list) blk_mq_insert_requests() /* see above */ Cc: Christoph Hellwig <hch@lst.de> Cc: Ming Lei <ming.lei@redhat.com> Signed-off-by: Bart Van Assche <bvanassche@acm.org> Link: https://lore.kernel.org/r/20230721172731.955724-4-bvanassche@acm.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-07-21 17:27:30 +00:00
blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
}
EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
struct blk_rq_wait {
struct completion done;
blk_status_t ret;
};
static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
{
struct blk_rq_wait *wait = rq->end_io_data;
wait->ret = ret;
complete(&wait->done);
return RQ_END_IO_NONE;
}
bool blk_rq_is_poll(struct request *rq)
{
if (!rq->mq_hctx)
return false;
if (rq->mq_hctx->type != HCTX_TYPE_POLL)
return false;
return true;
}
EXPORT_SYMBOL_GPL(blk_rq_is_poll);
static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
{
do {
blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
cond_resched();
} while (!completion_done(wait));
}
/**
* blk_execute_rq - insert a request into queue for execution
* @rq: request to insert
* @at_head: insert request at head or tail of queue
*
* Description:
* Insert a fully prepared request at the back of the I/O scheduler queue
* for execution and wait for completion.
* Return: The blk_status_t result provided to blk_mq_end_request().
*/
blk_status_t blk_execute_rq(struct request *rq, bool at_head)
{
struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
struct blk_rq_wait wait = {
.done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
};
WARN_ON(irqs_disabled());
WARN_ON(!blk_rq_is_passthrough(rq));
rq->end_io_data = &wait;
rq->end_io = blk_end_sync_rq;
blk_account_io_start(rq);
blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
blk_mq_run_hw_queue(hctx, false);
if (blk_rq_is_poll(rq))
blk_rq_poll_completion(rq, &wait.done);
else
blk_wait_io(&wait.done);
return wait.ret;
}
EXPORT_SYMBOL(blk_execute_rq);
static void __blk_mq_requeue_request(struct request *rq)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct request_queue *q = rq->q;
blk_mq_put_driver_tag(rq);
trace_block_rq_requeue(rq);
rq_qos_requeue(q, rq);
if (blk_mq_request_started(rq)) {
WRITE_ONCE(rq->state, MQ_RQ_IDLE);
rq->rq_flags &= ~RQF_TIMED_OUT;
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
{
struct request_queue *q = rq->q;
unsigned long flags;
__blk_mq_requeue_request(rq);
/* this request will be re-inserted to io scheduler queue */
blk_mq_sched_requeue_request(rq);
spin_lock_irqsave(&q->requeue_lock, flags);
list_add_tail(&rq->queuelist, &q->requeue_list);
spin_unlock_irqrestore(&q->requeue_lock, flags);
if (kick_requeue_list)
blk_mq_kick_requeue_list(q);
}
EXPORT_SYMBOL(blk_mq_requeue_request);
static void blk_mq_requeue_work(struct work_struct *work)
{
struct request_queue *q =
container_of(work, struct request_queue, requeue_work.work);
LIST_HEAD(rq_list);
LIST_HEAD(flush_list);
struct request *rq;
spin_lock_irq(&q->requeue_lock);
list_splice_init(&q->requeue_list, &rq_list);
list_splice_init(&q->flush_list, &flush_list);
spin_unlock_irq(&q->requeue_lock);
while (!list_empty(&rq_list)) {
rq = list_entry(rq_list.next, struct request, queuelist);
blk-mq: insert rq with DONTPREP to hctx dispatch list when requeue When requeue, if RQF_DONTPREP, rq has contained some driver specific data, so insert it to hctx dispatch list to avoid any merge. Take scsi as example, here is the trace event log (no io scheduler, because RQF_STARTED would prevent merging), kworker/0:1H-339 [000] ...1 2037.209289: block_rq_insert: 8,0 R 4096 () 32768 + 8 [kworker/0:1H] scsi_inert_test-1987 [000] .... 2037.220465: block_bio_queue: 8,0 R 32776 + 8 [scsi_inert_test] scsi_inert_test-1987 [000] ...2 2037.220466: block_bio_backmerge: 8,0 R 32776 + 8 [scsi_inert_test] kworker/0:1H-339 [000] .... 2047.220913: block_rq_issue: 8,0 R 8192 () 32768 + 16 [kworker/0:1H] scsi_inert_test-1996 [000] ..s1 2047.221007: block_rq_complete: 8,0 R () 32768 + 8 [0] scsi_inert_test-1996 [000] .Ns1 2047.221045: block_rq_requeue: 8,0 R () 32776 + 8 [0] kworker/0:1H-339 [000] ...1 2047.221054: block_rq_insert: 8,0 R 4096 () 32776 + 8 [kworker/0:1H] kworker/0:1H-339 [000] ...1 2047.221056: block_rq_issue: 8,0 R 4096 () 32776 + 8 [kworker/0:1H] scsi_inert_test-1986 [000] ..s1 2047.221119: block_rq_complete: 8,0 R () 32776 + 8 [0] (32768 + 8) was requeued by scsi_queue_insert and had RQF_DONTPREP. Then it was merged with (32776 + 8) and issued. Due to RQF_DONTPREP, the sdb only contained the part of (32768 + 8), then only that part was completed. The lucky thing was that scsi_io_completion detected it and requeued the remaining part. So we didn't get corrupted data. However, the requeue of (32776 + 8) is not expected. Suggested-by: Jens Axboe <axboe@kernel.dk> Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-02-12 01:56:25 +00:00
/*
* If RQF_DONTPREP ist set, the request has been started by the
* driver already and might have driver-specific data allocated
* already. Insert it into the hctx dispatch list to avoid
* block layer merges for the request.
blk-mq: insert rq with DONTPREP to hctx dispatch list when requeue When requeue, if RQF_DONTPREP, rq has contained some driver specific data, so insert it to hctx dispatch list to avoid any merge. Take scsi as example, here is the trace event log (no io scheduler, because RQF_STARTED would prevent merging), kworker/0:1H-339 [000] ...1 2037.209289: block_rq_insert: 8,0 R 4096 () 32768 + 8 [kworker/0:1H] scsi_inert_test-1987 [000] .... 2037.220465: block_bio_queue: 8,0 R 32776 + 8 [scsi_inert_test] scsi_inert_test-1987 [000] ...2 2037.220466: block_bio_backmerge: 8,0 R 32776 + 8 [scsi_inert_test] kworker/0:1H-339 [000] .... 2047.220913: block_rq_issue: 8,0 R 8192 () 32768 + 16 [kworker/0:1H] scsi_inert_test-1996 [000] ..s1 2047.221007: block_rq_complete: 8,0 R () 32768 + 8 [0] scsi_inert_test-1996 [000] .Ns1 2047.221045: block_rq_requeue: 8,0 R () 32776 + 8 [0] kworker/0:1H-339 [000] ...1 2047.221054: block_rq_insert: 8,0 R 4096 () 32776 + 8 [kworker/0:1H] kworker/0:1H-339 [000] ...1 2047.221056: block_rq_issue: 8,0 R 4096 () 32776 + 8 [kworker/0:1H] scsi_inert_test-1986 [000] ..s1 2047.221119: block_rq_complete: 8,0 R () 32776 + 8 [0] (32768 + 8) was requeued by scsi_queue_insert and had RQF_DONTPREP. Then it was merged with (32776 + 8) and issued. Due to RQF_DONTPREP, the sdb only contained the part of (32768 + 8), then only that part was completed. The lucky thing was that scsi_io_completion detected it and requeued the remaining part. So we didn't get corrupted data. However, the requeue of (32776 + 8) is not expected. Suggested-by: Jens Axboe <axboe@kernel.dk> Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-02-12 01:56:25 +00:00
*/
if (rq->rq_flags & RQF_DONTPREP) {
list_del_init(&rq->queuelist);
blk_mq_request_bypass_insert(rq, 0);
} else {
list_del_init(&rq->queuelist);
blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
}
}
while (!list_empty(&flush_list)) {
rq = list_entry(flush_list.next, struct request, queuelist);
list_del_init(&rq->queuelist);
blk_mq_insert_request(rq, 0);
}
blk-mq: Avoid that requeueing starts stopped queues Since blk_mq_requeue_work() starts stopped queues and since execution of this function can be scheduled after a queue has been stopped it is not possible to stop queues without using an additional state variable to track whether or not the queue has been stopped. Hence modify blk_mq_requeue_work() such that it does not start stopped queues. My conclusion after a review of the blk_mq_stop_hw_queues() and blk_mq_{delay_,}kick_requeue_list() callers is as follows: * In the dm driver starting and stopping queues should only happen if __dm_suspend() or __dm_resume() is called and not if the requeue list is processed. * In the SCSI core queue stopping and starting should only be performed by the scsi_internal_device_block() and scsi_internal_device_unblock() functions but not by any other function. Although the blk_mq_stop_hw_queue() call in scsi_queue_rq() may help to reduce CPU load if a LLD queue is full, figuring out whether or not a queue should be restarted when requeueing a command would require to introduce additional locking in scsi_mq_requeue_cmd() to avoid a race with scsi_internal_device_block(). Avoid this complexity by removing the blk_mq_stop_hw_queue() call from scsi_queue_rq(). * In the NVMe core only the functions that call blk_mq_start_stopped_hw_queues() explicitly should start stopped queues. * A blk_mq_start_stopped_hwqueues() call must be added in the xen-blkfront driver in its blkif_recover() function. Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Roger Pau Monné <roger.pau@citrix.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: James Bottomley <jejb@linux.vnet.ibm.com> Cc: Martin K. Petersen <martin.petersen@oracle.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@fb.com>
2016-10-29 00:20:32 +00:00
blk_mq_run_hw_queues(q, false);
}
void blk_mq_kick_requeue_list(struct request_queue *q)
{
kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
}
EXPORT_SYMBOL(blk_mq_kick_requeue_list);
void blk_mq_delay_kick_requeue_list(struct request_queue *q,
unsigned long msecs)
{
kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
msecs_to_jiffies(msecs));
}
EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
static bool blk_is_flush_data_rq(struct request *rq)
{
return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
}
static bool blk_mq_rq_inflight(struct request *rq, void *priv)
{
/*
* If we find a request that isn't idle we know the queue is busy
* as it's checked in the iter.
* Return false to stop the iteration.
*
* In case of queue quiesce, if one flush data request is completed,
* don't count it as inflight given the flush sequence is suspended,
* and the original flush data request is invisible to driver, just
* like other pending requests because of quiesce
*/
if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
blk_is_flush_data_rq(rq) &&
blk_mq_request_completed(rq))) {
bool *busy = priv;
*busy = true;
return false;
}
return true;
}
bool blk_mq_queue_inflight(struct request_queue *q)
{
bool busy = false;
blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
return busy;
}
EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
static void blk_mq_rq_timed_out(struct request *req)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
req->rq_flags |= RQF_TIMED_OUT;
if (req->q->mq_ops->timeout) {
enum blk_eh_timer_return ret;
ret = req->q->mq_ops->timeout(req);
if (ret == BLK_EH_DONE)
return;
WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
}
blk_add_timer(req);
}
struct blk_expired_data {
bool has_timedout_rq;
unsigned long next;
unsigned long timeout_start;
};
static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
{
unsigned long deadline;
if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
return false;
if (rq->rq_flags & RQF_TIMED_OUT)
return false;
blk-mq: attempt to fix atomic flag memory ordering Attempt to untangle the ordering in blk-mq. The patch introducing the single smp_mb__before_atomic() is obviously broken in that it doesn't clearly specify a pairing barrier and an obtained guarantee. The comment is further misleading in that it hints that the deadline store and the COMPLETE store also need to be ordered, but AFAICT there is no such dependency. However what does appear to be important is the clear happening _after_ the store, and that worked by pure accident. This clarifies blk_mq_start_request() -- we should not get there with STARTING set -- this simplifies the code and makes the barrier usage sane (the old code could be read to allow not having _any_ atomic after the barrier, in which case the barrier hasn't got anything to order). We then also introduce the missing pairing barrier for it. Also down-grade the barrier to smp_wmb(), this is cheaper for PowerPC/ARM and doesn't cost anything extra on x86. And it documents the STARTING vs COMPLETE ordering. Although I've not been entirely successful in reverse engineering the blk-mq state machine so there might still be more funnies around timeout vs requeue. If I got anything wrong, feel free to educate me by adding comments to clarify things ;-) Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Will Deacon <will.deacon@arm.com> Cc: Ming Lei <tom.leiming@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Andrea Parri <parri.andrea@gmail.com> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Bart Van Assche <bart.vanassche@wdc.com> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Fixes: 538b75341835 ("blk-mq: request deadline must be visible before marking rq as started") Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2017-09-06 08:00:22 +00:00
deadline = READ_ONCE(rq->deadline);
if (time_after_eq(expired->timeout_start, deadline))
return true;
blk-mq: attempt to fix atomic flag memory ordering Attempt to untangle the ordering in blk-mq. The patch introducing the single smp_mb__before_atomic() is obviously broken in that it doesn't clearly specify a pairing barrier and an obtained guarantee. The comment is further misleading in that it hints that the deadline store and the COMPLETE store also need to be ordered, but AFAICT there is no such dependency. However what does appear to be important is the clear happening _after_ the store, and that worked by pure accident. This clarifies blk_mq_start_request() -- we should not get there with STARTING set -- this simplifies the code and makes the barrier usage sane (the old code could be read to allow not having _any_ atomic after the barrier, in which case the barrier hasn't got anything to order). We then also introduce the missing pairing barrier for it. Also down-grade the barrier to smp_wmb(), this is cheaper for PowerPC/ARM and doesn't cost anything extra on x86. And it documents the STARTING vs COMPLETE ordering. Although I've not been entirely successful in reverse engineering the blk-mq state machine so there might still be more funnies around timeout vs requeue. If I got anything wrong, feel free to educate me by adding comments to clarify things ;-) Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Will Deacon <will.deacon@arm.com> Cc: Ming Lei <tom.leiming@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Andrea Parri <parri.andrea@gmail.com> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Bart Van Assche <bart.vanassche@wdc.com> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Fixes: 538b75341835 ("blk-mq: request deadline must be visible before marking rq as started") Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2017-09-06 08:00:22 +00:00
if (expired->next == 0)
expired->next = deadline;
else if (time_after(expired->next, deadline))
expired->next = deadline;
return false;
}
void blk_mq_put_rq_ref(struct request *rq)
{
if (is_flush_rq(rq)) {
if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
blk_mq_free_request(rq);
} else if (req_ref_put_and_test(rq)) {
__blk_mq_free_request(rq);
}
}
static bool blk_mq_check_expired(struct request *rq, void *priv)
blk-mq: replace timeout synchronization with a RCU and generation based scheme Currently, blk-mq timeout path synchronizes against the usual issue/completion path using a complex scheme involving atomic bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence rules. Unfortunately, it contains quite a few holes. There's a complex dancing around REQ_ATOM_STARTED and REQ_ATOM_COMPLETE between issue/completion and timeout paths; however, they don't have a synchronization point across request recycle instances and it isn't clear what the barriers add. blk_mq_check_expired() can easily read STARTED from N-2'th iteration, deadline from N-1'th, blk_mark_rq_complete() against Nth instance. In fact, it's pretty easy to make blk_mq_check_expired() terminate a later instance of a request. If we induce 5 sec delay before time_after_eq() test in blk_mq_check_expired(), shorten the timeout to 2s, and issue back-to-back large IOs, blk-mq starts timing out requests spuriously pretty quickly. Nothing actually timed out. It just made the call on a recycle instance of a request and then terminated a later instance long after the original instance finished. The scenario isn't theoretical either. This patch replaces the broken synchronization mechanism with a RCU and generation number based one. 1. Each request has a u64 generation + state value, which can be updated only by the request owner. Whenever a request becomes in-flight, the generation number gets bumped up too. This provides the basis for the timeout path to distinguish different recycle instances of the request. Also, marking a request in-flight and setting its deadline are protected with a seqcount so that the timeout path can fetch both values coherently. 2. The timeout path fetches the generation, state and deadline. If the verdict is timeout, it records the generation into a dedicated request abortion field and does RCU wait. 3. The completion path is also protected by RCU (from the previous patch) and checks whether the current generation number and state match the abortion field. If so, it skips completion. 4. The timeout path, after RCU wait, scans requests again and terminates the ones whose generation and state still match the ones requested for abortion. By now, the timeout path knows that either the generation number and state changed if it lost the race or the completion will yield to it and can safely timeout the request. While it's more lines of code, it's conceptually simpler, doesn't depend on direct use of subtle memory ordering or coherence, and hopefully doesn't terminate the wrong instance. While this change makes REQ_ATOM_COMPLETE synchronization unnecessary between issue/complete and timeout paths, REQ_ATOM_COMPLETE isn't removed yet as it's still used in other places. Future patches will move all state tracking to the new mechanism and remove all bitops in the hot paths. Note that this patch adds a comment explaining a race condition in BLK_EH_RESET_TIMER path. The race has always been there and this patch doesn't change it. It's just documenting the existing race. v2: - Fixed BLK_EH_RESET_TIMER handling as pointed out by Jianchao. - s/request->gstate_seqc/request->gstate_seq/ as suggested by Peter. - READ_ONCE() added in blk_mq_rq_update_state() as suggested by Peter. v3: - Fixed possible extended seqcount / u64_stats_sync read looping spotted by Peter. - MQ_RQ_IDLE was incorrectly being set in complete_request instead of free_request. Fixed. v4: - Rebased on top of hctx_lock() refactoring patch. - Added comment explaining the use of hctx_lock() in completion path. v5: - Added comments requested by Bart. - Note the addition of BLK_EH_RESET_TIMER race condition in the commit message. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: "jianchao.wang" <jianchao.w.wang@oracle.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <Bart.VanAssche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-09 16:29:48 +00:00
{
struct blk_expired_data *expired = priv;
/*
* blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
* be reallocated underneath the timeout handler's processing, then
* the expire check is reliable. If the request is not expired, then
* it was completed and reallocated as a new request after returning
* from blk_mq_check_expired().
blk-mq: replace timeout synchronization with a RCU and generation based scheme Currently, blk-mq timeout path synchronizes against the usual issue/completion path using a complex scheme involving atomic bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence rules. Unfortunately, it contains quite a few holes. There's a complex dancing around REQ_ATOM_STARTED and REQ_ATOM_COMPLETE between issue/completion and timeout paths; however, they don't have a synchronization point across request recycle instances and it isn't clear what the barriers add. blk_mq_check_expired() can easily read STARTED from N-2'th iteration, deadline from N-1'th, blk_mark_rq_complete() against Nth instance. In fact, it's pretty easy to make blk_mq_check_expired() terminate a later instance of a request. If we induce 5 sec delay before time_after_eq() test in blk_mq_check_expired(), shorten the timeout to 2s, and issue back-to-back large IOs, blk-mq starts timing out requests spuriously pretty quickly. Nothing actually timed out. It just made the call on a recycle instance of a request and then terminated a later instance long after the original instance finished. The scenario isn't theoretical either. This patch replaces the broken synchronization mechanism with a RCU and generation number based one. 1. Each request has a u64 generation + state value, which can be updated only by the request owner. Whenever a request becomes in-flight, the generation number gets bumped up too. This provides the basis for the timeout path to distinguish different recycle instances of the request. Also, marking a request in-flight and setting its deadline are protected with a seqcount so that the timeout path can fetch both values coherently. 2. The timeout path fetches the generation, state and deadline. If the verdict is timeout, it records the generation into a dedicated request abortion field and does RCU wait. 3. The completion path is also protected by RCU (from the previous patch) and checks whether the current generation number and state match the abortion field. If so, it skips completion. 4. The timeout path, after RCU wait, scans requests again and terminates the ones whose generation and state still match the ones requested for abortion. By now, the timeout path knows that either the generation number and state changed if it lost the race or the completion will yield to it and can safely timeout the request. While it's more lines of code, it's conceptually simpler, doesn't depend on direct use of subtle memory ordering or coherence, and hopefully doesn't terminate the wrong instance. While this change makes REQ_ATOM_COMPLETE synchronization unnecessary between issue/complete and timeout paths, REQ_ATOM_COMPLETE isn't removed yet as it's still used in other places. Future patches will move all state tracking to the new mechanism and remove all bitops in the hot paths. Note that this patch adds a comment explaining a race condition in BLK_EH_RESET_TIMER path. The race has always been there and this patch doesn't change it. It's just documenting the existing race. v2: - Fixed BLK_EH_RESET_TIMER handling as pointed out by Jianchao. - s/request->gstate_seqc/request->gstate_seq/ as suggested by Peter. - READ_ONCE() added in blk_mq_rq_update_state() as suggested by Peter. v3: - Fixed possible extended seqcount / u64_stats_sync read looping spotted by Peter. - MQ_RQ_IDLE was incorrectly being set in complete_request instead of free_request. Fixed. v4: - Rebased on top of hctx_lock() refactoring patch. - Added comment explaining the use of hctx_lock() in completion path. v5: - Added comments requested by Bart. - Note the addition of BLK_EH_RESET_TIMER race condition in the commit message. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: "jianchao.wang" <jianchao.w.wang@oracle.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <Bart.VanAssche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-09 16:29:48 +00:00
*/
if (blk_mq_req_expired(rq, expired)) {
expired->has_timedout_rq = true;
return false;
}
return true;
}
static bool blk_mq_handle_expired(struct request *rq, void *priv)
{
struct blk_expired_data *expired = priv;
if (blk_mq_req_expired(rq, expired))
blk_mq_rq_timed_out(rq);
return true;
blk-mq: replace timeout synchronization with a RCU and generation based scheme Currently, blk-mq timeout path synchronizes against the usual issue/completion path using a complex scheme involving atomic bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence rules. Unfortunately, it contains quite a few holes. There's a complex dancing around REQ_ATOM_STARTED and REQ_ATOM_COMPLETE between issue/completion and timeout paths; however, they don't have a synchronization point across request recycle instances and it isn't clear what the barriers add. blk_mq_check_expired() can easily read STARTED from N-2'th iteration, deadline from N-1'th, blk_mark_rq_complete() against Nth instance. In fact, it's pretty easy to make blk_mq_check_expired() terminate a later instance of a request. If we induce 5 sec delay before time_after_eq() test in blk_mq_check_expired(), shorten the timeout to 2s, and issue back-to-back large IOs, blk-mq starts timing out requests spuriously pretty quickly. Nothing actually timed out. It just made the call on a recycle instance of a request and then terminated a later instance long after the original instance finished. The scenario isn't theoretical either. This patch replaces the broken synchronization mechanism with a RCU and generation number based one. 1. Each request has a u64 generation + state value, which can be updated only by the request owner. Whenever a request becomes in-flight, the generation number gets bumped up too. This provides the basis for the timeout path to distinguish different recycle instances of the request. Also, marking a request in-flight and setting its deadline are protected with a seqcount so that the timeout path can fetch both values coherently. 2. The timeout path fetches the generation, state and deadline. If the verdict is timeout, it records the generation into a dedicated request abortion field and does RCU wait. 3. The completion path is also protected by RCU (from the previous patch) and checks whether the current generation number and state match the abortion field. If so, it skips completion. 4. The timeout path, after RCU wait, scans requests again and terminates the ones whose generation and state still match the ones requested for abortion. By now, the timeout path knows that either the generation number and state changed if it lost the race or the completion will yield to it and can safely timeout the request. While it's more lines of code, it's conceptually simpler, doesn't depend on direct use of subtle memory ordering or coherence, and hopefully doesn't terminate the wrong instance. While this change makes REQ_ATOM_COMPLETE synchronization unnecessary between issue/complete and timeout paths, REQ_ATOM_COMPLETE isn't removed yet as it's still used in other places. Future patches will move all state tracking to the new mechanism and remove all bitops in the hot paths. Note that this patch adds a comment explaining a race condition in BLK_EH_RESET_TIMER path. The race has always been there and this patch doesn't change it. It's just documenting the existing race. v2: - Fixed BLK_EH_RESET_TIMER handling as pointed out by Jianchao. - s/request->gstate_seqc/request->gstate_seq/ as suggested by Peter. - READ_ONCE() added in blk_mq_rq_update_state() as suggested by Peter. v3: - Fixed possible extended seqcount / u64_stats_sync read looping spotted by Peter. - MQ_RQ_IDLE was incorrectly being set in complete_request instead of free_request. Fixed. v4: - Rebased on top of hctx_lock() refactoring patch. - Added comment explaining the use of hctx_lock() in completion path. v5: - Added comments requested by Bart. - Note the addition of BLK_EH_RESET_TIMER race condition in the commit message. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: "jianchao.wang" <jianchao.w.wang@oracle.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <Bart.VanAssche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-09 16:29:48 +00:00
}
static void blk_mq_timeout_work(struct work_struct *work)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct request_queue *q =
container_of(work, struct request_queue, timeout_work);
struct blk_expired_data expired = {
.timeout_start = jiffies,
};
blk-mq: replace timeout synchronization with a RCU and generation based scheme Currently, blk-mq timeout path synchronizes against the usual issue/completion path using a complex scheme involving atomic bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence rules. Unfortunately, it contains quite a few holes. There's a complex dancing around REQ_ATOM_STARTED and REQ_ATOM_COMPLETE between issue/completion and timeout paths; however, they don't have a synchronization point across request recycle instances and it isn't clear what the barriers add. blk_mq_check_expired() can easily read STARTED from N-2'th iteration, deadline from N-1'th, blk_mark_rq_complete() against Nth instance. In fact, it's pretty easy to make blk_mq_check_expired() terminate a later instance of a request. If we induce 5 sec delay before time_after_eq() test in blk_mq_check_expired(), shorten the timeout to 2s, and issue back-to-back large IOs, blk-mq starts timing out requests spuriously pretty quickly. Nothing actually timed out. It just made the call on a recycle instance of a request and then terminated a later instance long after the original instance finished. The scenario isn't theoretical either. This patch replaces the broken synchronization mechanism with a RCU and generation number based one. 1. Each request has a u64 generation + state value, which can be updated only by the request owner. Whenever a request becomes in-flight, the generation number gets bumped up too. This provides the basis for the timeout path to distinguish different recycle instances of the request. Also, marking a request in-flight and setting its deadline are protected with a seqcount so that the timeout path can fetch both values coherently. 2. The timeout path fetches the generation, state and deadline. If the verdict is timeout, it records the generation into a dedicated request abortion field and does RCU wait. 3. The completion path is also protected by RCU (from the previous patch) and checks whether the current generation number and state match the abortion field. If so, it skips completion. 4. The timeout path, after RCU wait, scans requests again and terminates the ones whose generation and state still match the ones requested for abortion. By now, the timeout path knows that either the generation number and state changed if it lost the race or the completion will yield to it and can safely timeout the request. While it's more lines of code, it's conceptually simpler, doesn't depend on direct use of subtle memory ordering or coherence, and hopefully doesn't terminate the wrong instance. While this change makes REQ_ATOM_COMPLETE synchronization unnecessary between issue/complete and timeout paths, REQ_ATOM_COMPLETE isn't removed yet as it's still used in other places. Future patches will move all state tracking to the new mechanism and remove all bitops in the hot paths. Note that this patch adds a comment explaining a race condition in BLK_EH_RESET_TIMER path. The race has always been there and this patch doesn't change it. It's just documenting the existing race. v2: - Fixed BLK_EH_RESET_TIMER handling as pointed out by Jianchao. - s/request->gstate_seqc/request->gstate_seq/ as suggested by Peter. - READ_ONCE() added in blk_mq_rq_update_state() as suggested by Peter. v3: - Fixed possible extended seqcount / u64_stats_sync read looping spotted by Peter. - MQ_RQ_IDLE was incorrectly being set in complete_request instead of free_request. Fixed. v4: - Rebased on top of hctx_lock() refactoring patch. - Added comment explaining the use of hctx_lock() in completion path. v5: - Added comments requested by Bart. - Note the addition of BLK_EH_RESET_TIMER race condition in the commit message. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: "jianchao.wang" <jianchao.w.wang@oracle.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <Bart.VanAssche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-09 16:29:48 +00:00
struct blk_mq_hw_ctx *hctx;
unsigned long i;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
blk-mq: Allow timeouts to run while queue is freezing In case a submitted request gets stuck for some reason, the block layer can prevent the request starvation by starting the scheduled timeout work. If this stuck request occurs at the same time another thread has started a queue freeze, the blk_mq_timeout_work will not be able to acquire the queue reference and will return silently, thus not issuing the timeout. But since the request is already holding a q_usage_counter reference and is unable to complete, it will never release its reference, preventing the queue from completing the freeze started by first thread. This puts the request_queue in a hung state, forever waiting for the freeze completion. This was observed while running IO to a NVMe device at the same time we toggled the CPU hotplug code. Eventually, once a request got stuck requiring a timeout during a queue freeze, we saw the CPU Hotplug notification code get stuck inside blk_mq_freeze_queue_wait, as shown in the trace below. [c000000deaf13690] [c000000deaf13738] 0xc000000deaf13738 (unreliable) [c000000deaf13860] [c000000000015ce8] __switch_to+0x1f8/0x350 [c000000deaf138b0] [c000000000ade0e4] __schedule+0x314/0x990 [c000000deaf13940] [c000000000ade7a8] schedule+0x48/0xc0 [c000000deaf13970] [c0000000005492a4] blk_mq_freeze_queue_wait+0x74/0x110 [c000000deaf139e0] [c00000000054b6a8] blk_mq_queue_reinit_notify+0x1a8/0x2e0 [c000000deaf13a40] [c0000000000e7878] notifier_call_chain+0x98/0x100 [c000000deaf13a90] [c0000000000b8e08] cpu_notify_nofail+0x48/0xa0 [c000000deaf13ac0] [c0000000000b92f0] _cpu_down+0x2a0/0x400 [c000000deaf13b90] [c0000000000b94a8] cpu_down+0x58/0xa0 [c000000deaf13bc0] [c0000000006d5dcc] cpu_subsys_offline+0x2c/0x50 [c000000deaf13bf0] [c0000000006cd244] device_offline+0x104/0x140 [c000000deaf13c30] [c0000000006cd40c] online_store+0x6c/0xc0 [c000000deaf13c80] [c0000000006c8c78] dev_attr_store+0x68/0xa0 [c000000deaf13cc0] [c0000000003974d0] sysfs_kf_write+0x80/0xb0 [c000000deaf13d00] [c0000000003963e8] kernfs_fop_write+0x188/0x200 [c000000deaf13d50] [c0000000002e0f6c] __vfs_write+0x6c/0xe0 [c000000deaf13d90] [c0000000002e1ca0] vfs_write+0xc0/0x230 [c000000deaf13de0] [c0000000002e2cdc] SyS_write+0x6c/0x110 [c000000deaf13e30] [c000000000009204] system_call+0x38/0xb4 The fix is to allow the timeout work to execute in the window between dropping the initial refcount reference and the release of the last reference, which actually marks the freeze completion. This can be achieved with percpu_refcount_tryget, which does not require the counter to be alive. This way the timeout work can do it's job and terminate a stuck request even during a freeze, returning its reference and avoiding the deadlock. Allowing the timeout to run is just a part of the fix, since for some devices, we might get stuck again inside the device driver's timeout handler, should it attempt to allocate a new request in that path - which is a quite common action for Abort commands, which need to be sent after a timeout. In NVMe, for instance, we call blk_mq_alloc_request from inside the timeout handler, which will fail during a freeze, since it also tries to acquire a queue reference. I considered a similar change to blk_mq_alloc_request as a generic solution for further device driver hangs, but we can't do that, since it would allow new requests to disturb the freeze process. I thought about creating a new function in the block layer to support unfreezable requests for these occasions, but after working on it for a while, I feel like this should be handled in a per-driver basis. I'm now experimenting with changes to the NVMe timeout path, but I'm open to suggestions of ways to make this generic. Signed-off-by: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com> Cc: Brian King <brking@linux.vnet.ibm.com> Cc: Keith Busch <keith.busch@intel.com> Cc: linux-nvme@lists.infradead.org Cc: linux-block@vger.kernel.org Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@fb.com>
2016-08-01 14:23:39 +00:00
/* A deadlock might occur if a request is stuck requiring a
* timeout at the same time a queue freeze is waiting
* completion, since the timeout code would not be able to
* acquire the queue reference here.
*
* That's why we don't use blk_queue_enter here; instead, we use
* percpu_ref_tryget directly, because we need to be able to
* obtain a reference even in the short window between the queue
* starting to freeze, by dropping the first reference in
* blk_freeze_queue_start, and the moment the last request is
blk-mq: Allow timeouts to run while queue is freezing In case a submitted request gets stuck for some reason, the block layer can prevent the request starvation by starting the scheduled timeout work. If this stuck request occurs at the same time another thread has started a queue freeze, the blk_mq_timeout_work will not be able to acquire the queue reference and will return silently, thus not issuing the timeout. But since the request is already holding a q_usage_counter reference and is unable to complete, it will never release its reference, preventing the queue from completing the freeze started by first thread. This puts the request_queue in a hung state, forever waiting for the freeze completion. This was observed while running IO to a NVMe device at the same time we toggled the CPU hotplug code. Eventually, once a request got stuck requiring a timeout during a queue freeze, we saw the CPU Hotplug notification code get stuck inside blk_mq_freeze_queue_wait, as shown in the trace below. [c000000deaf13690] [c000000deaf13738] 0xc000000deaf13738 (unreliable) [c000000deaf13860] [c000000000015ce8] __switch_to+0x1f8/0x350 [c000000deaf138b0] [c000000000ade0e4] __schedule+0x314/0x990 [c000000deaf13940] [c000000000ade7a8] schedule+0x48/0xc0 [c000000deaf13970] [c0000000005492a4] blk_mq_freeze_queue_wait+0x74/0x110 [c000000deaf139e0] [c00000000054b6a8] blk_mq_queue_reinit_notify+0x1a8/0x2e0 [c000000deaf13a40] [c0000000000e7878] notifier_call_chain+0x98/0x100 [c000000deaf13a90] [c0000000000b8e08] cpu_notify_nofail+0x48/0xa0 [c000000deaf13ac0] [c0000000000b92f0] _cpu_down+0x2a0/0x400 [c000000deaf13b90] [c0000000000b94a8] cpu_down+0x58/0xa0 [c000000deaf13bc0] [c0000000006d5dcc] cpu_subsys_offline+0x2c/0x50 [c000000deaf13bf0] [c0000000006cd244] device_offline+0x104/0x140 [c000000deaf13c30] [c0000000006cd40c] online_store+0x6c/0xc0 [c000000deaf13c80] [c0000000006c8c78] dev_attr_store+0x68/0xa0 [c000000deaf13cc0] [c0000000003974d0] sysfs_kf_write+0x80/0xb0 [c000000deaf13d00] [c0000000003963e8] kernfs_fop_write+0x188/0x200 [c000000deaf13d50] [c0000000002e0f6c] __vfs_write+0x6c/0xe0 [c000000deaf13d90] [c0000000002e1ca0] vfs_write+0xc0/0x230 [c000000deaf13de0] [c0000000002e2cdc] SyS_write+0x6c/0x110 [c000000deaf13e30] [c000000000009204] system_call+0x38/0xb4 The fix is to allow the timeout work to execute in the window between dropping the initial refcount reference and the release of the last reference, which actually marks the freeze completion. This can be achieved with percpu_refcount_tryget, which does not require the counter to be alive. This way the timeout work can do it's job and terminate a stuck request even during a freeze, returning its reference and avoiding the deadlock. Allowing the timeout to run is just a part of the fix, since for some devices, we might get stuck again inside the device driver's timeout handler, should it attempt to allocate a new request in that path - which is a quite common action for Abort commands, which need to be sent after a timeout. In NVMe, for instance, we call blk_mq_alloc_request from inside the timeout handler, which will fail during a freeze, since it also tries to acquire a queue reference. I considered a similar change to blk_mq_alloc_request as a generic solution for further device driver hangs, but we can't do that, since it would allow new requests to disturb the freeze process. I thought about creating a new function in the block layer to support unfreezable requests for these occasions, but after working on it for a while, I feel like this should be handled in a per-driver basis. I'm now experimenting with changes to the NVMe timeout path, but I'm open to suggestions of ways to make this generic. Signed-off-by: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com> Cc: Brian King <brking@linux.vnet.ibm.com> Cc: Keith Busch <keith.busch@intel.com> Cc: linux-nvme@lists.infradead.org Cc: linux-block@vger.kernel.org Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@fb.com>
2016-08-01 14:23:39 +00:00
* consumed, marked by the instant q_usage_counter reaches
* zero.
*/
if (!percpu_ref_tryget(&q->q_usage_counter))
return;
/* check if there is any timed-out request */
blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
if (expired.has_timedout_rq) {
/*
* Before walking tags, we must ensure any submit started
* before the current time has finished. Since the submit
* uses srcu or rcu, wait for a synchronization point to
* ensure all running submits have finished
*/
blk_mq_wait_quiesce_done(q->tag_set);
expired.next = 0;
blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
if (expired.next != 0) {
mod_timer(&q->timeout, expired.next);
} else {
/*
* Request timeouts are handled as a forward rolling timer. If
* we end up here it means that no requests are pending and
* also that no request has been pending for a while. Mark
* each hctx as idle.
*/
queue_for_each_hw_ctx(q, hctx, i) {
/* the hctx may be unmapped, so check it here */
if (blk_mq_hw_queue_mapped(hctx))
blk_mq_tag_idle(hctx);
}
}
blk_queue_exit(q);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
struct flush_busy_ctx_data {
struct blk_mq_hw_ctx *hctx;
struct list_head *list;
};
static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
{
struct flush_busy_ctx_data *flush_data = data;
struct blk_mq_hw_ctx *hctx = flush_data->hctx;
struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
enum hctx_type type = hctx->type;
spin_lock(&ctx->lock);
list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
sbitmap_clear_bit(sb, bitnr);
spin_unlock(&ctx->lock);
return true;
}
/*
* Process software queues that have been marked busy, splicing them
* to the for-dispatch
*/
void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
{
struct flush_busy_ctx_data data = {
.hctx = hctx,
.list = list,
};
sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
}
EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
struct dispatch_rq_data {
struct blk_mq_hw_ctx *hctx;
struct request *rq;
};
static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
void *data)
{
struct dispatch_rq_data *dispatch_data = data;
struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
enum hctx_type type = hctx->type;
spin_lock(&ctx->lock);
if (!list_empty(&ctx->rq_lists[type])) {
dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
list_del_init(&dispatch_data->rq->queuelist);
if (list_empty(&ctx->rq_lists[type]))
sbitmap_clear_bit(sb, bitnr);
}
spin_unlock(&ctx->lock);
return !dispatch_data->rq;
}
struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
struct blk_mq_ctx *start)
{
unsigned off = start ? start->index_hw[hctx->type] : 0;
struct dispatch_rq_data data = {
.hctx = hctx,
.rq = NULL,
};
__sbitmap_for_each_set(&hctx->ctx_map, off,
dispatch_rq_from_ctx, &data);
return data.rq;
}
bool __blk_mq_alloc_driver_tag(struct request *rq)
{
struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
int tag;
blk_mq_tag_busy(rq->mq_hctx);
if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
bt = &rq->mq_hctx->tags->breserved_tags;
tag_offset = 0;
} else {
if (!hctx_may_queue(rq->mq_hctx, bt))
return false;
}
tag = __sbitmap_queue_get(bt);
if (tag == BLK_MQ_NO_TAG)
return false;
rq->tag = tag + tag_offset;
blk_mq_inc_active_requests(rq->mq_hctx);
return true;
}
static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
int flags, void *key)
{
struct blk_mq_hw_ctx *hctx;
hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
spin_lock(&hctx->dispatch_wait_lock);
if (!list_empty(&wait->entry)) {
struct sbitmap_queue *sbq;
list_del_init(&wait->entry);
sbq = &hctx->tags->bitmap_tags;
atomic_dec(&sbq->ws_active);
}
spin_unlock(&hctx->dispatch_wait_lock);
blk_mq_run_hw_queue(hctx, true);
return 1;
}
/*
* Mark us waiting for a tag. For shared tags, this involves hooking us into
* the tag wakeups. For non-shared tags, we can simply mark us needing a
* restart. For both cases, take care to check the condition again after
* marking us as waiting.
*/
static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
struct request *rq)
{
struct sbitmap_queue *sbq;
struct wait_queue_head *wq;
wait_queue_entry_t *wait;
bool ret;
if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
!(blk_mq_is_shared_tags(hctx->flags))) {
blk_mq_sched_mark_restart_hctx(hctx);
/*
* It's possible that a tag was freed in the window between the
* allocation failure and adding the hardware queue to the wait
* queue.
*
* Don't clear RESTART here, someone else could have set it.
* At most this will cost an extra queue run.
*/
return blk_mq_get_driver_tag(rq);
}
wait = &hctx->dispatch_wait;
if (!list_empty_careful(&wait->entry))
return false;
if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
sbq = &hctx->tags->breserved_tags;
else
sbq = &hctx->tags->bitmap_tags;
wq = &bt_wait_ptr(sbq, hctx)->wait;
spin_lock_irq(&wq->lock);
spin_lock(&hctx->dispatch_wait_lock);
if (!list_empty(&wait->entry)) {
spin_unlock(&hctx->dispatch_wait_lock);
spin_unlock_irq(&wq->lock);
return false;
}
atomic_inc(&sbq->ws_active);
wait->flags &= ~WQ_FLAG_EXCLUSIVE;
__add_wait_queue(wq, wait);
/*
* Add one explicit barrier since blk_mq_get_driver_tag() may
* not imply barrier in case of failure.
*
* Order adding us to wait queue and allocating driver tag.
*
* The pair is the one implied in sbitmap_queue_wake_up() which
* orders clearing sbitmap tag bits and waitqueue_active() in
* __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
*
* Otherwise, re-order of adding wait queue and getting driver tag
* may cause __sbitmap_queue_wake_up() to wake up nothing because
* the waitqueue_active() may not observe us in wait queue.
*/
smp_mb();
/*
* It's possible that a tag was freed in the window between the
* allocation failure and adding the hardware queue to the wait
* queue.
*/
ret = blk_mq_get_driver_tag(rq);
if (!ret) {
spin_unlock(&hctx->dispatch_wait_lock);
spin_unlock_irq(&wq->lock);
return false;
}
/*
* We got a tag, remove ourselves from the wait queue to ensure
* someone else gets the wakeup.
*/
list_del_init(&wait->entry);
atomic_dec(&sbq->ws_active);
spin_unlock(&hctx->dispatch_wait_lock);
spin_unlock_irq(&wq->lock);
return true;
}
#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
/*
* Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
* - EWMA is one simple way to compute running average value
* - weight(7/8 and 1/8) is applied so that it can decrease exponentially
* - take 4 as factor for avoiding to get too small(0) result, and this
* factor doesn't matter because EWMA decreases exponentially
*/
static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
{
unsigned int ewma;
ewma = hctx->dispatch_busy;
if (!ewma && !busy)
return;
ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
if (busy)
ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
hctx->dispatch_busy = ewma;
}
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
static void blk_mq_handle_dev_resource(struct request *rq,
struct list_head *list)
{
list_add(&rq->queuelist, list);
__blk_mq_requeue_request(rq);
}
enum prep_dispatch {
PREP_DISPATCH_OK,
PREP_DISPATCH_NO_TAG,
PREP_DISPATCH_NO_BUDGET,
};
static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
bool need_budget)
{
struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
int budget_token = -1;
if (need_budget) {
budget_token = blk_mq_get_dispatch_budget(rq->q);
if (budget_token < 0) {
blk_mq_put_driver_tag(rq);
return PREP_DISPATCH_NO_BUDGET;
}
blk_mq_set_rq_budget_token(rq, budget_token);
}
if (!blk_mq_get_driver_tag(rq)) {
/*
* The initial allocation attempt failed, so we need to
* rerun the hardware queue when a tag is freed. The
* waitqueue takes care of that. If the queue is run
* before we add this entry back on the dispatch list,
* we'll re-run it below.
*/
if (!blk_mq_mark_tag_wait(hctx, rq)) {
/*
* All budgets not got from this function will be put
* together during handling partial dispatch
*/
if (need_budget)
blk_mq_put_dispatch_budget(rq->q, budget_token);
return PREP_DISPATCH_NO_TAG;
}
}
return PREP_DISPATCH_OK;
}
/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
static void blk_mq_release_budgets(struct request_queue *q,
struct list_head *list)
{
struct request *rq;
list_for_each_entry(rq, list, queuelist) {
int budget_token = blk_mq_get_rq_budget_token(rq);
if (budget_token >= 0)
blk_mq_put_dispatch_budget(q, budget_token);
}
}
/*
* blk_mq_commit_rqs will notify driver using bd->last that there is no
* more requests. (See comment in struct blk_mq_ops for commit_rqs for
* details)
* Attention, we should explicitly call this in unusual cases:
* 1) did not queue everything initially scheduled to queue
* 2) the last attempt to queue a request failed
*/
static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
bool from_schedule)
{
if (hctx->queue->mq_ops->commit_rqs && queued) {
trace_block_unplug(hctx->queue, queued, !from_schedule);
hctx->queue->mq_ops->commit_rqs(hctx);
}
}
/*
* Returns true if we did some work AND can potentially do more.
*/
bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
unsigned int nr_budgets)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
enum prep_dispatch prep;
struct request_queue *q = hctx->queue;
struct request *rq;
int queued;
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
blk_status_t ret = BLK_STS_OK;
bool needs_resource = false;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
if (list_empty(list))
return false;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
/*
* Now process all the entries, sending them to the driver.
*/
queued = 0;
do {
struct blk_mq_queue_data bd;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
rq = list_first_entry(list, struct request, queuelist);
WARN_ON_ONCE(hctx != rq->mq_hctx);
prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
if (prep != PREP_DISPATCH_OK)
break;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
list_del_init(&rq->queuelist);
bd.rq = rq;
bd.last = list_empty(list);
/*
* once the request is queued to lld, no need to cover the
* budget any more
*/
if (nr_budgets)
nr_budgets--;
ret = q->mq_ops->queue_rq(hctx, &bd);
switch (ret) {
case BLK_STS_OK:
queued++;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
break;
case BLK_STS_RESOURCE:
needs_resource = true;
fallthrough;
case BLK_STS_DEV_RESOURCE:
blk_mq_handle_dev_resource(rq, list);
goto out;
default:
blk_mq_end_request(rq, ret);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
} while (!list_empty(list));
out:
/* If we didn't flush the entire list, we could have told the driver
* there was more coming, but that turned out to be a lie.
*/
if (!list_empty(list) || ret != BLK_STS_OK)
blk_mq_commit_rqs(hctx, queued, false);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
/*
* Any items that need requeuing? Stuff them into hctx->dispatch,
* that is where we will continue on next queue run.
*/
if (!list_empty(list)) {
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
bool needs_restart;
/* For non-shared tags, the RESTART check will suffice */
bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
blk_mq_is_shared_tags(hctx->flags));
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
if (nr_budgets)
blk_mq_release_budgets(q, list);
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
spin_lock(&hctx->lock);
list_splice_tail_init(list, &hctx->dispatch);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
spin_unlock(&hctx->lock);
blk-mq: order adding requests to hctx->dispatch and checking SCHED_RESTART SCHED_RESTART code path is relied to re-run queue for dispatch requests in hctx->dispatch. Meantime the SCHED_RSTART flag is checked when adding requests to hctx->dispatch. memory barriers have to be used for ordering the following two pair of OPs: 1) adding requests to hctx->dispatch and checking SCHED_RESTART in blk_mq_dispatch_rq_list() 2) clearing SCHED_RESTART and checking if there is request in hctx->dispatch in blk_mq_sched_restart(). Without the added memory barrier, either: 1) blk_mq_sched_restart() may miss requests added to hctx->dispatch meantime blk_mq_dispatch_rq_list() observes SCHED_RESTART, and not run queue in dispatch side or 2) blk_mq_dispatch_rq_list still sees SCHED_RESTART, and not run queue in dispatch side, meantime checking if there is request in hctx->dispatch from blk_mq_sched_restart() is missed. IO hang in ltp/fs_fill test is reported by kernel test robot: https://lkml.org/lkml/2020/7/26/77 Turns out it is caused by the above out-of-order OPs. And the IO hang can't be observed any more after applying this patch. Fixes: bd166ef183c2 ("blk-mq-sched: add framework for MQ capable IO schedulers") Reported-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <bvanassche@acm.org> Cc: Christoph Hellwig <hch@lst.de> Cc: David Jeffery <djeffery@redhat.com> Cc: <stable@vger.kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-08-17 10:01:15 +00:00
/*
* Order adding requests to hctx->dispatch and checking
* SCHED_RESTART flag. The pair of this smp_mb() is the one
* in blk_mq_sched_restart(). Avoid restart code path to
* miss the new added requests to hctx->dispatch, meantime
* SCHED_RESTART is observed here.
*/
smp_mb();
/*
* If SCHED_RESTART was set by the caller of this function and
* it is no longer set that means that it was cleared by another
* thread and hence that a queue rerun is needed.
*
* If 'no_tag' is set, that means that we failed getting
* a driver tag with an I/O scheduler attached. If our dispatch
* waitqueue is no longer active, ensure that we run the queue
* AFTER adding our entries back to the list.
*
* If no I/O scheduler has been configured it is possible that
* the hardware queue got stopped and restarted before requests
* were pushed back onto the dispatch list. Rerun the queue to
* avoid starvation. Notes:
* - blk_mq_run_hw_queue() checks whether or not a queue has
* been stopped before rerunning a queue.
* - Some but not all block drivers stop a queue before
* returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
* and dm-rq.
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
*
* If driver returns BLK_STS_RESOURCE and SCHED_RESTART
* bit is set, run queue after a delay to avoid IO stalls
* that could otherwise occur if the queue is idle. We'll do
* similar if we couldn't get budget or couldn't lock a zone
* and SCHED_RESTART is set.
*/
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
needs_restart = blk_mq_sched_needs_restart(hctx);
if (prep == PREP_DISPATCH_NO_BUDGET)
needs_resource = true;
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
if (!needs_restart ||
(no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
blk_mq_run_hw_queue(hctx, true);
else if (needs_resource)
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
blk_mq_update_dispatch_busy(hctx, true);
return false;
}
blk_mq_update_dispatch_busy(hctx, false);
return true;
}
static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
{
int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
if (cpu >= nr_cpu_ids)
cpu = cpumask_first(hctx->cpumask);
return cpu;
}
/*
* ->next_cpu is always calculated from hctx->cpumask, so simply use
* it for speeding up the check
*/
static bool blk_mq_hctx_empty_cpumask(struct blk_mq_hw_ctx *hctx)
{
return hctx->next_cpu >= nr_cpu_ids;
}
/*
* It'd be great if the workqueue API had a way to pass
* in a mask and had some smarts for more clever placement.
* For now we just round-robin here, switching for every
* BLK_MQ_CPU_WORK_BATCH queued items.
*/
static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
{
bool tried = false;
int next_cpu = hctx->next_cpu;
/* Switch to unbound if no allowable CPUs in this hctx */
if (hctx->queue->nr_hw_queues == 1 || blk_mq_hctx_empty_cpumask(hctx))
return WORK_CPU_UNBOUND;
if (--hctx->next_cpu_batch <= 0) {
select_cpu:
next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
cpu_online_mask);
if (next_cpu >= nr_cpu_ids)
next_cpu = blk_mq_first_mapped_cpu(hctx);
hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
}
/*
* Do unbound schedule if we can't find a online CPU for this hctx,
* and it should only happen in the path of handling CPU DEAD.
*/
if (!cpu_online(next_cpu)) {
if (!tried) {
tried = true;
goto select_cpu;
}
/*
* Make sure to re-select CPU next time once after CPUs
* in hctx->cpumask become online again.
*/
hctx->next_cpu = next_cpu;
hctx->next_cpu_batch = 1;
return WORK_CPU_UNBOUND;
}
hctx->next_cpu = next_cpu;
return next_cpu;
}
/**
* blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
* @hctx: Pointer to the hardware queue to run.
* @msecs: Milliseconds of delay to wait before running the queue.
*
* Run a hardware queue asynchronously with a delay of @msecs.
*/
void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
if (unlikely(blk_mq_hctx_stopped(hctx)))
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
return;
kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
msecs_to_jiffies(msecs));
}
EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
/**
* blk_mq_run_hw_queue - Start to run a hardware queue.
* @hctx: Pointer to the hardware queue to run.
* @async: If we want to run the queue asynchronously.
*
* Check if the request queue is not in a quiesced state and if there are
* pending requests to be sent. If this is true, run the queue to send requests
* to hardware.
*/
void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
{
bool need_run;
/*
* We can't run the queue inline with interrupts disabled.
*/
WARN_ON_ONCE(!async && in_interrupt());
block: Improve performance for BLK_MQ_F_BLOCKING drivers blk_mq_run_queue() runs the queue asynchronously if BLK_MQ_F_BLOCKING has been set. This is suboptimal since running the queue asynchronously is slower than running the queue synchronously. This patch modifies blk_mq_run_queue() as follows if BLK_MQ_F_BLOCKING has been set: - Run the queue synchronously if it is allowed to sleep. - Run the queue asynchronously if it is not allowed to sleep. Additionally, blk_mq_run_hw_queue(hctx, false) calls are modified into blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING) if the caller may be invoked from atomic context. The following caller chains have been reviewed: blk_mq_run_hw_queue(hctx, false) blk_mq_get_tag() /* may sleep, hence the functions it calls may also sleep */ blk_execute_rq() /* may sleep */ blk_mq_run_hw_queues(q, async=false) blk_freeze_queue_start() /* may sleep */ blk_mq_requeue_work() /* may sleep */ scsi_kick_queue() scsi_requeue_run_queue() /* may sleep */ scsi_run_host_queues() scsi_ioctl_reset() /* may sleep */ blk_mq_insert_requests(hctx, ctx, list, run_queue_async=false) blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list(plug, from_schedule=false) __blk_flush_plug(plug, from_schedule=false) blk_add_rq_to_plug() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_plug_issue_direct() blk_mq_flush_plug_list() /* see above */ blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list() /* see above */ blk_mq_try_issue_directly() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_try_issue_list_directly(hctx, list) blk_mq_insert_requests() /* see above */ Cc: Christoph Hellwig <hch@lst.de> Cc: Ming Lei <ming.lei@redhat.com> Signed-off-by: Bart Van Assche <bvanassche@acm.org> Link: https://lore.kernel.org/r/20230721172731.955724-4-bvanassche@acm.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-07-21 17:27:30 +00:00
might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
/*
* When queue is quiesced, we may be switching io scheduler, or
* updating nr_hw_queues, or other things, and we can't run queue
* any more, even __blk_mq_hctx_has_pending() can't be called safely.
*
* And queue will be rerun in blk_mq_unquiesce_queue() if it is
* quiesced.
*/
__blk_mq_run_dispatch_ops(hctx->queue, false,
need_run = !blk_queue_quiesced(hctx->queue) &&
blk_mq_hctx_has_pending(hctx));
if (!need_run)
return;
block: Improve performance for BLK_MQ_F_BLOCKING drivers blk_mq_run_queue() runs the queue asynchronously if BLK_MQ_F_BLOCKING has been set. This is suboptimal since running the queue asynchronously is slower than running the queue synchronously. This patch modifies blk_mq_run_queue() as follows if BLK_MQ_F_BLOCKING has been set: - Run the queue synchronously if it is allowed to sleep. - Run the queue asynchronously if it is not allowed to sleep. Additionally, blk_mq_run_hw_queue(hctx, false) calls are modified into blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING) if the caller may be invoked from atomic context. The following caller chains have been reviewed: blk_mq_run_hw_queue(hctx, false) blk_mq_get_tag() /* may sleep, hence the functions it calls may also sleep */ blk_execute_rq() /* may sleep */ blk_mq_run_hw_queues(q, async=false) blk_freeze_queue_start() /* may sleep */ blk_mq_requeue_work() /* may sleep */ scsi_kick_queue() scsi_requeue_run_queue() /* may sleep */ scsi_run_host_queues() scsi_ioctl_reset() /* may sleep */ blk_mq_insert_requests(hctx, ctx, list, run_queue_async=false) blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list(plug, from_schedule=false) __blk_flush_plug(plug, from_schedule=false) blk_add_rq_to_plug() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_plug_issue_direct() blk_mq_flush_plug_list() /* see above */ blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list() /* see above */ blk_mq_try_issue_directly() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_try_issue_list_directly(hctx, list) blk_mq_insert_requests() /* see above */ Cc: Christoph Hellwig <hch@lst.de> Cc: Ming Lei <ming.lei@redhat.com> Signed-off-by: Bart Van Assche <bvanassche@acm.org> Link: https://lore.kernel.org/r/20230721172731.955724-4-bvanassche@acm.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-07-21 17:27:30 +00:00
if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
blk_mq_delay_run_hw_queue(hctx, 0);
return;
}
blk_mq_run_dispatch_ops(hctx->queue,
blk_mq_sched_dispatch_requests(hctx));
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
EXPORT_SYMBOL(blk_mq_run_hw_queue);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention. Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler: clients MQ SQ MQ-Patched Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%* Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%* Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied. You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-11 16:47:17 +00:00
/*
* Return prefered queue to dispatch from (if any) for non-mq aware IO
* scheduler.
*/
static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
{
struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention. Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler: clients MQ SQ MQ-Patched Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%* Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%* Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied. You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-11 16:47:17 +00:00
/*
* If the IO scheduler does not respect hardware queues when
* dispatching, we just don't bother with multiple HW queues and
* dispatch from hctx for the current CPU since running multiple queues
* just causes lock contention inside the scheduler and pointless cache
* bouncing.
*/
struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention. Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler: clients MQ SQ MQ-Patched Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%* Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%* Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied. You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-11 16:47:17 +00:00
if (!blk_mq_hctx_stopped(hctx))
return hctx;
return NULL;
}
/**
* blk_mq_run_hw_queues - Run all hardware queues in a request queue.
* @q: Pointer to the request queue to run.
* @async: If we want to run the queue asynchronously.
*/
void blk_mq_run_hw_queues(struct request_queue *q, bool async)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention. Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler: clients MQ SQ MQ-Patched Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%* Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%* Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied. You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-11 16:47:17 +00:00
struct blk_mq_hw_ctx *hctx, *sq_hctx;
unsigned long i;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention. Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler: clients MQ SQ MQ-Patched Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%* Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%* Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied. You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-11 16:47:17 +00:00
sq_hctx = NULL;
if (blk_queue_sq_sched(q))
blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention. Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler: clients MQ SQ MQ-Patched Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%* Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%* Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied. You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-11 16:47:17 +00:00
sq_hctx = blk_mq_get_sq_hctx(q);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
queue_for_each_hw_ctx(q, hctx, i) {
if (blk_mq_hctx_stopped(hctx))
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
continue;
blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention. Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler: clients MQ SQ MQ-Patched Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%* Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%* Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied. You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-11 16:47:17 +00:00
/*
* Dispatch from this hctx either if there's no hctx preferred
* by IO scheduler or if it has requests that bypass the
* scheduler.
*/
if (!sq_hctx || sq_hctx == hctx ||
!list_empty_careful(&hctx->dispatch))
blk_mq_run_hw_queue(hctx, async);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
}
EXPORT_SYMBOL(blk_mq_run_hw_queues);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
/**
* blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
* @q: Pointer to the request queue to run.
* @msecs: Milliseconds of delay to wait before running the queues.
*/
void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
{
blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention. Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler: clients MQ SQ MQ-Patched Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%* Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%* Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied. You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-11 16:47:17 +00:00
struct blk_mq_hw_ctx *hctx, *sq_hctx;
unsigned long i;
blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention. Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler: clients MQ SQ MQ-Patched Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%* Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%* Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied. You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-11 16:47:17 +00:00
sq_hctx = NULL;
if (blk_queue_sq_sched(q))
blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention. Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler: clients MQ SQ MQ-Patched Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%* Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%* Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied. You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-11 16:47:17 +00:00
sq_hctx = blk_mq_get_sq_hctx(q);
queue_for_each_hw_ctx(q, hctx, i) {
if (blk_mq_hctx_stopped(hctx))
continue;
/*
* If there is already a run_work pending, leave the
* pending delay untouched. Otherwise, a hctx can stall
* if another hctx is re-delaying the other's work
* before the work executes.
*/
if (delayed_work_pending(&hctx->run_work))
continue;
blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention. Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler: clients MQ SQ MQ-Patched Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%* Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%* Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied. You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance. Signed-off-by: Jan Kara <jack@suse.cz> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-11 16:47:17 +00:00
/*
* Dispatch from this hctx either if there's no hctx preferred
* by IO scheduler or if it has requests that bypass the
* scheduler.
*/
if (!sq_hctx || sq_hctx == hctx ||
!list_empty_careful(&hctx->dispatch))
blk_mq_delay_run_hw_queue(hctx, msecs);
}
}
EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
/*
* This function is often used for pausing .queue_rq() by driver when
* there isn't enough resource or some conditions aren't satisfied, and
* BLK_STS_RESOURCE is usually returned.
*
* We do not guarantee that dispatch can be drained or blocked
* after blk_mq_stop_hw_queue() returns. Please use
* blk_mq_quiesce_queue() for that requirement.
*/
void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
{
cancel_delayed_work(&hctx->run_work);
set_bit(BLK_MQ_S_STOPPED, &hctx->state);
}
EXPORT_SYMBOL(blk_mq_stop_hw_queue);
/*
* This function is often used for pausing .queue_rq() by driver when
* there isn't enough resource or some conditions aren't satisfied, and
* BLK_STS_RESOURCE is usually returned.
*
* We do not guarantee that dispatch can be drained or blocked
* after blk_mq_stop_hw_queues() returns. Please use
* blk_mq_quiesce_queue() for that requirement.
*/
void blk_mq_stop_hw_queues(struct request_queue *q)
{
struct blk_mq_hw_ctx *hctx;
unsigned long i;
queue_for_each_hw_ctx(q, hctx, i)
blk_mq_stop_hw_queue(hctx);
}
EXPORT_SYMBOL(blk_mq_stop_hw_queues);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
{
clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
block: Improve performance for BLK_MQ_F_BLOCKING drivers blk_mq_run_queue() runs the queue asynchronously if BLK_MQ_F_BLOCKING has been set. This is suboptimal since running the queue asynchronously is slower than running the queue synchronously. This patch modifies blk_mq_run_queue() as follows if BLK_MQ_F_BLOCKING has been set: - Run the queue synchronously if it is allowed to sleep. - Run the queue asynchronously if it is not allowed to sleep. Additionally, blk_mq_run_hw_queue(hctx, false) calls are modified into blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING) if the caller may be invoked from atomic context. The following caller chains have been reviewed: blk_mq_run_hw_queue(hctx, false) blk_mq_get_tag() /* may sleep, hence the functions it calls may also sleep */ blk_execute_rq() /* may sleep */ blk_mq_run_hw_queues(q, async=false) blk_freeze_queue_start() /* may sleep */ blk_mq_requeue_work() /* may sleep */ scsi_kick_queue() scsi_requeue_run_queue() /* may sleep */ scsi_run_host_queues() scsi_ioctl_reset() /* may sleep */ blk_mq_insert_requests(hctx, ctx, list, run_queue_async=false) blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list(plug, from_schedule=false) __blk_flush_plug(plug, from_schedule=false) blk_add_rq_to_plug() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_plug_issue_direct() blk_mq_flush_plug_list() /* see above */ blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list() /* see above */ blk_mq_try_issue_directly() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_try_issue_list_directly(hctx, list) blk_mq_insert_requests() /* see above */ Cc: Christoph Hellwig <hch@lst.de> Cc: Ming Lei <ming.lei@redhat.com> Signed-off-by: Bart Van Assche <bvanassche@acm.org> Link: https://lore.kernel.org/r/20230721172731.955724-4-bvanassche@acm.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-07-21 17:27:30 +00:00
blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
EXPORT_SYMBOL(blk_mq_start_hw_queue);
void blk_mq_start_hw_queues(struct request_queue *q)
{
struct blk_mq_hw_ctx *hctx;
unsigned long i;
queue_for_each_hw_ctx(q, hctx, i)
blk_mq_start_hw_queue(hctx);
}
EXPORT_SYMBOL(blk_mq_start_hw_queues);
void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
{
if (!blk_mq_hctx_stopped(hctx))
return;
clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
blk_mq_run_hw_queue(hctx, async);
}
EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct blk_mq_hw_ctx *hctx;
unsigned long i;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
queue_for_each_hw_ctx(q, hctx, i)
block: Improve performance for BLK_MQ_F_BLOCKING drivers blk_mq_run_queue() runs the queue asynchronously if BLK_MQ_F_BLOCKING has been set. This is suboptimal since running the queue asynchronously is slower than running the queue synchronously. This patch modifies blk_mq_run_queue() as follows if BLK_MQ_F_BLOCKING has been set: - Run the queue synchronously if it is allowed to sleep. - Run the queue asynchronously if it is not allowed to sleep. Additionally, blk_mq_run_hw_queue(hctx, false) calls are modified into blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING) if the caller may be invoked from atomic context. The following caller chains have been reviewed: blk_mq_run_hw_queue(hctx, false) blk_mq_get_tag() /* may sleep, hence the functions it calls may also sleep */ blk_execute_rq() /* may sleep */ blk_mq_run_hw_queues(q, async=false) blk_freeze_queue_start() /* may sleep */ blk_mq_requeue_work() /* may sleep */ scsi_kick_queue() scsi_requeue_run_queue() /* may sleep */ scsi_run_host_queues() scsi_ioctl_reset() /* may sleep */ blk_mq_insert_requests(hctx, ctx, list, run_queue_async=false) blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list(plug, from_schedule=false) __blk_flush_plug(plug, from_schedule=false) blk_add_rq_to_plug() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_plug_issue_direct() blk_mq_flush_plug_list() /* see above */ blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list() /* see above */ blk_mq_try_issue_directly() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_try_issue_list_directly(hctx, list) blk_mq_insert_requests() /* see above */ Cc: Christoph Hellwig <hch@lst.de> Cc: Ming Lei <ming.lei@redhat.com> Signed-off-by: Bart Van Assche <bvanassche@acm.org> Link: https://lore.kernel.org/r/20230721172731.955724-4-bvanassche@acm.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-07-21 17:27:30 +00:00
blk_mq_start_stopped_hw_queue(hctx, async ||
(hctx->flags & BLK_MQ_F_BLOCKING));
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
static void blk_mq_run_work_fn(struct work_struct *work)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct blk_mq_hw_ctx *hctx =
container_of(work, struct blk_mq_hw_ctx, run_work.work);
blk_mq_run_dispatch_ops(hctx->queue,
blk_mq_sched_dispatch_requests(hctx));
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
/**
* blk_mq_request_bypass_insert - Insert a request at dispatch list.
* @rq: Pointer to request to be inserted.
* @flags: BLK_MQ_INSERT_*
*
block: directly insert blk-mq request from blk_insert_cloned_request() A NULL pointer crash was reported for the case of having the BFQ IO scheduler attached to the underlying blk-mq paths of a DM multipath device. The crash occured in blk_mq_sched_insert_request()'s call to e->type->ops.mq.insert_requests(). Paolo Valente correctly summarized why the crash occured with: "the call chain (dm_mq_queue_rq -> map_request -> setup_clone -> blk_rq_prep_clone) creates a cloned request without invoking e->type->ops.mq.prepare_request for the target elevator e. The cloned request is therefore not initialized for the scheduler, but it is however inserted into the scheduler by blk_mq_sched_insert_request." All said, a request-based DM multipath device's IO scheduler should be the only one used -- when the original requests are issued to the underlying paths as cloned requests they are inserted directly in the underlying dispatch queue(s) rather than through an additional elevator. But commit bd166ef18 ("blk-mq-sched: add framework for MQ capable IO schedulers") switched blk_insert_cloned_request() from using blk_mq_insert_request() to blk_mq_sched_insert_request(). Which incorrectly added elevator machinery into a call chain that isn't supposed to have any. To fix this introduce a blk-mq private blk_mq_request_bypass_insert() that blk_insert_cloned_request() calls to insert the request without involving any elevator that may be attached to the cloned request's request_queue. Fixes: bd166ef183c2 ("blk-mq-sched: add framework for MQ capable IO schedulers") Cc: stable@vger.kernel.org Reported-by: Bart Van Assche <Bart.VanAssche@wdc.com> Tested-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2017-09-11 22:43:57 +00:00
* Should only be used carefully, when the caller knows we want to
* bypass a potential IO scheduler on the target device.
*/
static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
block: directly insert blk-mq request from blk_insert_cloned_request() A NULL pointer crash was reported for the case of having the BFQ IO scheduler attached to the underlying blk-mq paths of a DM multipath device. The crash occured in blk_mq_sched_insert_request()'s call to e->type->ops.mq.insert_requests(). Paolo Valente correctly summarized why the crash occured with: "the call chain (dm_mq_queue_rq -> map_request -> setup_clone -> blk_rq_prep_clone) creates a cloned request without invoking e->type->ops.mq.prepare_request for the target elevator e. The cloned request is therefore not initialized for the scheduler, but it is however inserted into the scheduler by blk_mq_sched_insert_request." All said, a request-based DM multipath device's IO scheduler should be the only one used -- when the original requests are issued to the underlying paths as cloned requests they are inserted directly in the underlying dispatch queue(s) rather than through an additional elevator. But commit bd166ef18 ("blk-mq-sched: add framework for MQ capable IO schedulers") switched blk_insert_cloned_request() from using blk_mq_insert_request() to blk_mq_sched_insert_request(). Which incorrectly added elevator machinery into a call chain that isn't supposed to have any. To fix this introduce a blk-mq private blk_mq_request_bypass_insert() that blk_insert_cloned_request() calls to insert the request without involving any elevator that may be attached to the cloned request's request_queue. Fixes: bd166ef183c2 ("blk-mq-sched: add framework for MQ capable IO schedulers") Cc: stable@vger.kernel.org Reported-by: Bart Van Assche <Bart.VanAssche@wdc.com> Tested-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2017-09-11 22:43:57 +00:00
{
struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
block: directly insert blk-mq request from blk_insert_cloned_request() A NULL pointer crash was reported for the case of having the BFQ IO scheduler attached to the underlying blk-mq paths of a DM multipath device. The crash occured in blk_mq_sched_insert_request()'s call to e->type->ops.mq.insert_requests(). Paolo Valente correctly summarized why the crash occured with: "the call chain (dm_mq_queue_rq -> map_request -> setup_clone -> blk_rq_prep_clone) creates a cloned request without invoking e->type->ops.mq.prepare_request for the target elevator e. The cloned request is therefore not initialized for the scheduler, but it is however inserted into the scheduler by blk_mq_sched_insert_request." All said, a request-based DM multipath device's IO scheduler should be the only one used -- when the original requests are issued to the underlying paths as cloned requests they are inserted directly in the underlying dispatch queue(s) rather than through an additional elevator. But commit bd166ef18 ("blk-mq-sched: add framework for MQ capable IO schedulers") switched blk_insert_cloned_request() from using blk_mq_insert_request() to blk_mq_sched_insert_request(). Which incorrectly added elevator machinery into a call chain that isn't supposed to have any. To fix this introduce a blk-mq private blk_mq_request_bypass_insert() that blk_insert_cloned_request() calls to insert the request without involving any elevator that may be attached to the cloned request's request_queue. Fixes: bd166ef183c2 ("blk-mq-sched: add framework for MQ capable IO schedulers") Cc: stable@vger.kernel.org Reported-by: Bart Van Assche <Bart.VanAssche@wdc.com> Tested-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2017-09-11 22:43:57 +00:00
spin_lock(&hctx->lock);
if (flags & BLK_MQ_INSERT_AT_HEAD)
list_add(&rq->queuelist, &hctx->dispatch);
else
list_add_tail(&rq->queuelist, &hctx->dispatch);
block: directly insert blk-mq request from blk_insert_cloned_request() A NULL pointer crash was reported for the case of having the BFQ IO scheduler attached to the underlying blk-mq paths of a DM multipath device. The crash occured in blk_mq_sched_insert_request()'s call to e->type->ops.mq.insert_requests(). Paolo Valente correctly summarized why the crash occured with: "the call chain (dm_mq_queue_rq -> map_request -> setup_clone -> blk_rq_prep_clone) creates a cloned request without invoking e->type->ops.mq.prepare_request for the target elevator e. The cloned request is therefore not initialized for the scheduler, but it is however inserted into the scheduler by blk_mq_sched_insert_request." All said, a request-based DM multipath device's IO scheduler should be the only one used -- when the original requests are issued to the underlying paths as cloned requests they are inserted directly in the underlying dispatch queue(s) rather than through an additional elevator. But commit bd166ef18 ("blk-mq-sched: add framework for MQ capable IO schedulers") switched blk_insert_cloned_request() from using blk_mq_insert_request() to blk_mq_sched_insert_request(). Which incorrectly added elevator machinery into a call chain that isn't supposed to have any. To fix this introduce a blk-mq private blk_mq_request_bypass_insert() that blk_insert_cloned_request() calls to insert the request without involving any elevator that may be attached to the cloned request's request_queue. Fixes: bd166ef183c2 ("blk-mq-sched: add framework for MQ capable IO schedulers") Cc: stable@vger.kernel.org Reported-by: Bart Van Assche <Bart.VanAssche@wdc.com> Tested-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2017-09-11 22:43:57 +00:00
spin_unlock(&hctx->lock);
}
static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
struct blk_mq_ctx *ctx, struct list_head *list,
bool run_queue_async)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct request *rq;
enum hctx_type type = hctx->type;
/*
* Try to issue requests directly if the hw queue isn't busy to save an
* extra enqueue & dequeue to the sw queue.
*/
if (!hctx->dispatch_busy && !run_queue_async) {
blk_mq_run_dispatch_ops(hctx->queue,
blk_mq_try_issue_list_directly(hctx, list));
if (list_empty(list))
goto out;
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
/*
* preemption doesn't flush plug list, so it's possible ctx->cpu is
* offline now
*/
list_for_each_entry(rq, list, queuelist) {
BUG_ON(rq->mq_ctx != ctx);
trace_block_rq_insert(rq);
block: Improve performance for BLK_MQ_F_BLOCKING drivers blk_mq_run_queue() runs the queue asynchronously if BLK_MQ_F_BLOCKING has been set. This is suboptimal since running the queue asynchronously is slower than running the queue synchronously. This patch modifies blk_mq_run_queue() as follows if BLK_MQ_F_BLOCKING has been set: - Run the queue synchronously if it is allowed to sleep. - Run the queue asynchronously if it is not allowed to sleep. Additionally, blk_mq_run_hw_queue(hctx, false) calls are modified into blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING) if the caller may be invoked from atomic context. The following caller chains have been reviewed: blk_mq_run_hw_queue(hctx, false) blk_mq_get_tag() /* may sleep, hence the functions it calls may also sleep */ blk_execute_rq() /* may sleep */ blk_mq_run_hw_queues(q, async=false) blk_freeze_queue_start() /* may sleep */ blk_mq_requeue_work() /* may sleep */ scsi_kick_queue() scsi_requeue_run_queue() /* may sleep */ scsi_run_host_queues() scsi_ioctl_reset() /* may sleep */ blk_mq_insert_requests(hctx, ctx, list, run_queue_async=false) blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list(plug, from_schedule=false) __blk_flush_plug(plug, from_schedule=false) blk_add_rq_to_plug() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_plug_issue_direct() blk_mq_flush_plug_list() /* see above */ blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list() /* see above */ blk_mq_try_issue_directly() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_try_issue_list_directly(hctx, list) blk_mq_insert_requests() /* see above */ Cc: Christoph Hellwig <hch@lst.de> Cc: Ming Lei <ming.lei@redhat.com> Signed-off-by: Bart Van Assche <bvanassche@acm.org> Link: https://lore.kernel.org/r/20230721172731.955724-4-bvanassche@acm.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-07-21 17:27:30 +00:00
if (rq->cmd_flags & REQ_NOWAIT)
run_queue_async = true;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
spin_lock(&ctx->lock);
list_splice_tail_init(list, &ctx->rq_lists[type]);
blk_mq_hctx_mark_pending(hctx, ctx);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
spin_unlock(&ctx->lock);
out:
blk_mq_run_hw_queue(hctx, run_queue_async);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
{
struct request_queue *q = rq->q;
struct blk_mq_ctx *ctx = rq->mq_ctx;
struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
if (blk_rq_is_passthrough(rq)) {
/*
* Passthrough request have to be added to hctx->dispatch
* directly. The device may be in a situation where it can't
* handle FS request, and always returns BLK_STS_RESOURCE for
* them, which gets them added to hctx->dispatch.
*
* If a passthrough request is required to unblock the queues,
* and it is added to the scheduler queue, there is no chance to
* dispatch it given we prioritize requests in hctx->dispatch.
*/
blk_mq_request_bypass_insert(rq, flags);
} else if (req_op(rq) == REQ_OP_FLUSH) {
/*
* Firstly normal IO request is inserted to scheduler queue or
* sw queue, meantime we add flush request to dispatch queue(
* hctx->dispatch) directly and there is at most one in-flight
* flush request for each hw queue, so it doesn't matter to add
* flush request to tail or front of the dispatch queue.
*
* Secondly in case of NCQ, flush request belongs to non-NCQ
* command, and queueing it will fail when there is any
* in-flight normal IO request(NCQ command). When adding flush
* rq to the front of hctx->dispatch, it is easier to introduce
* extra time to flush rq's latency because of S_SCHED_RESTART
* compared with adding to the tail of dispatch queue, then
* chance of flush merge is increased, and less flush requests
* will be issued to controller. It is observed that ~10% time
* is saved in blktests block/004 on disk attached to AHCI/NCQ
* drive when adding flush rq to the front of hctx->dispatch.
*
* Simply queue flush rq to the front of hctx->dispatch so that
* intensive flush workloads can benefit in case of NCQ HW.
*/
blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
} else if (q->elevator) {
LIST_HEAD(list);
WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
list_add(&rq->queuelist, &list);
q->elevator->type->ops.insert_requests(hctx, &list, flags);
} else {
trace_block_rq_insert(rq);
spin_lock(&ctx->lock);
if (flags & BLK_MQ_INSERT_AT_HEAD)
list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
else
list_add_tail(&rq->queuelist,
&ctx->rq_lists[hctx->type]);
blk_mq_hctx_mark_pending(hctx, ctx);
spin_unlock(&ctx->lock);
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
unsigned int nr_segs)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
int err;
if (bio->bi_opf & REQ_RAHEAD)
rq->cmd_flags |= REQ_FAILFAST_MASK;
rq->__sector = bio->bi_iter.bi_sector;
rq->write_hint = bio->bi_write_hint;
blk_rq_bio_prep(rq, bio, nr_segs);
/* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
WARN_ON_ONCE(err);
blk_account_io_start(rq);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
struct request *rq, bool last)
{
struct request_queue *q = rq->q;
struct blk_mq_queue_data bd = {
.rq = rq,
.last = last,
};
blk_status_t ret;
/*
* For OK queue, we are done. For error, caller may kill it.
* Any other error (busy), just add it to our list as we
* previously would have done.
*/
ret = q->mq_ops->queue_rq(hctx, &bd);
switch (ret) {
case BLK_STS_OK:
blk_mq_update_dispatch_busy(hctx, false);
break;
case BLK_STS_RESOURCE:
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
case BLK_STS_DEV_RESOURCE:
blk_mq_update_dispatch_busy(hctx, true);
__blk_mq_requeue_request(rq);
break;
default:
blk_mq_update_dispatch_busy(hctx, false);
break;
}
return ret;
}
static bool blk_mq_get_budget_and_tag(struct request *rq)
{
int budget_token;
blk-mq: fix direct issue If queue is stopped, we shouldn't dispatch request into driver and hardware, unfortunately the check is removed in bd166ef183c2(blk-mq-sched: add framework for MQ capable IO schedulers). This patch fixes the issue by moving the check back into __blk_mq_try_issue_directly(). This patch fixes request use-after-free[1][2] during canceling requets of NVMe in nvme_dev_disable(), which can be triggered easily during NVMe reset & remove test. [1] oops kernel log when CONFIG_BLK_DEV_INTEGRITY is on [ 103.412969] BUG: unable to handle kernel NULL pointer dereference at 000000000000000a [ 103.412980] IP: bio_integrity_advance+0x48/0xf0 [ 103.412981] PGD 275a88067 [ 103.412981] P4D 275a88067 [ 103.412982] PUD 276c43067 [ 103.412983] PMD 0 [ 103.412984] [ 103.412986] Oops: 0000 [#1] SMP [ 103.412989] Modules linked in: vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel pcbc aesni_intel crypto_simd cryptd ipmi_ssif iTCO_wdt iTCO_vendor_support mxm_wmi glue_helper dcdbas ipmi_si mei_me pcspkr mei sg ipmi_devintf lpc_ich ipmi_msghandler shpchp acpi_power_meter wmi nfsd auth_rpcgss nfs_acl lockd grace sunrpc ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm crc32c_intel nvme ahci nvme_core libahci libata tg3 i2c_core megaraid_sas ptp pps_core dm_mirror dm_region_hash dm_log dm_mod [ 103.413035] CPU: 0 PID: 102 Comm: kworker/0:2 Not tainted 4.11.0+ #1 [ 103.413036] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.2.5 09/06/2016 [ 103.413041] Workqueue: events nvme_remove_dead_ctrl_work [nvme] [ 103.413043] task: ffff9cc8775c8000 task.stack: ffffc033c252c000 [ 103.413045] RIP: 0010:bio_integrity_advance+0x48/0xf0 [ 103.413046] RSP: 0018:ffffc033c252fc10 EFLAGS: 00010202 [ 103.413048] RAX: 0000000000000000 RBX: ffff9cc8720a8cc0 RCX: ffff9cca72958240 [ 103.413049] RDX: ffff9cca72958000 RSI: 0000000000000008 RDI: ffff9cc872537f00 [ 103.413049] RBP: ffffc033c252fc28 R08: 0000000000000000 R09: ffffffffb963a0d5 [ 103.413050] R10: 000000000000063e R11: 0000000000000000 R12: ffff9cc8720a8d18 [ 103.413051] R13: 0000000000001000 R14: ffff9cc872682e00 R15: 00000000fffffffb [ 103.413053] FS: 0000000000000000(0000) GS:ffff9cc877c00000(0000) knlGS:0000000000000000 [ 103.413054] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 103.413055] CR2: 000000000000000a CR3: 0000000276c41000 CR4: 00000000001406f0 [ 103.413056] Call Trace: [ 103.413063] bio_advance+0x2a/0xe0 [ 103.413067] blk_update_request+0x76/0x330 [ 103.413072] blk_mq_end_request+0x1a/0x70 [ 103.413074] blk_mq_dispatch_rq_list+0x370/0x410 [ 103.413076] ? blk_mq_flush_busy_ctxs+0x94/0xe0 [ 103.413080] blk_mq_sched_dispatch_requests+0x173/0x1a0 [ 103.413083] __blk_mq_run_hw_queue+0x8e/0xa0 [ 103.413085] __blk_mq_delay_run_hw_queue+0x9d/0xa0 [ 103.413088] blk_mq_start_hw_queue+0x17/0x20 [ 103.413090] blk_mq_start_hw_queues+0x32/0x50 [ 103.413095] nvme_kill_queues+0x54/0x80 [nvme_core] [ 103.413097] nvme_remove_dead_ctrl_work+0x1f/0x40 [nvme] [ 103.413103] process_one_work+0x149/0x360 [ 103.413105] worker_thread+0x4d/0x3c0 [ 103.413109] kthread+0x109/0x140 [ 103.413111] ? rescuer_thread+0x380/0x380 [ 103.413113] ? kthread_park+0x60/0x60 [ 103.413120] ret_from_fork+0x2c/0x40 [ 103.413121] Code: 08 4c 8b 63 50 48 8b 80 80 00 00 00 48 8b 90 d0 03 00 00 31 c0 48 83 ba 40 02 00 00 00 48 8d 8a 40 02 00 00 48 0f 45 c1 c1 ee 09 <0f> b6 48 0a 0f b6 40 09 41 89 f5 83 e9 09 41 d3 ed 44 0f af e8 [ 103.413145] RIP: bio_integrity_advance+0x48/0xf0 RSP: ffffc033c252fc10 [ 103.413146] CR2: 000000000000000a [ 103.413157] ---[ end trace cd6875d16eb5a11e ]--- [ 103.455368] Kernel panic - not syncing: Fatal exception [ 103.459826] Kernel Offset: 0x37600000 from 0xffffffff81000000 (relocation range: 0xffffffff80000000-0xffffffffbfffffff) [ 103.850916] ---[ end Kernel panic - not syncing: Fatal exception [ 103.857637] sched: Unexpected reschedule of offline CPU#1! [ 103.863762] ------------[ cut here ]------------ [2] kernel hang in blk_mq_freeze_queue_wait() when CONFIG_BLK_DEV_INTEGRITY is off [ 247.129825] INFO: task nvme-test:1772 blocked for more than 120 seconds. [ 247.137311] Not tainted 4.12.0-rc2.upstream+ #4 [ 247.142954] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 247.151704] Call Trace: [ 247.154445] __schedule+0x28a/0x880 [ 247.158341] schedule+0x36/0x80 [ 247.161850] blk_mq_freeze_queue_wait+0x4b/0xb0 [ 247.166913] ? remove_wait_queue+0x60/0x60 [ 247.171485] blk_freeze_queue+0x1a/0x20 [ 247.175770] blk_cleanup_queue+0x7f/0x140 [ 247.180252] nvme_ns_remove+0xa3/0xb0 [nvme_core] [ 247.185503] nvme_remove_namespaces+0x32/0x50 [nvme_core] [ 247.191532] nvme_uninit_ctrl+0x2d/0xa0 [nvme_core] [ 247.196977] nvme_remove+0x70/0x110 [nvme] [ 247.201545] pci_device_remove+0x39/0xc0 [ 247.205927] device_release_driver_internal+0x141/0x200 [ 247.211761] device_release_driver+0x12/0x20 [ 247.216531] pci_stop_bus_device+0x8c/0xa0 [ 247.221104] pci_stop_and_remove_bus_device_locked+0x1a/0x30 [ 247.227420] remove_store+0x7c/0x90 [ 247.231320] dev_attr_store+0x18/0x30 [ 247.235409] sysfs_kf_write+0x3a/0x50 [ 247.239497] kernfs_fop_write+0xff/0x180 [ 247.243867] __vfs_write+0x37/0x160 [ 247.247757] ? selinux_file_permission+0xe5/0x120 [ 247.253011] ? security_file_permission+0x3b/0xc0 [ 247.258260] vfs_write+0xb2/0x1b0 [ 247.261964] ? syscall_trace_enter+0x1d0/0x2b0 [ 247.266924] SyS_write+0x55/0xc0 [ 247.270540] do_syscall_64+0x67/0x150 [ 247.274636] entry_SYSCALL64_slow_path+0x25/0x25 [ 247.279794] RIP: 0033:0x7f5c96740840 [ 247.283785] RSP: 002b:00007ffd00e87ee8 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 247.292238] RAX: ffffffffffffffda RBX: 0000000000000002 RCX: 00007f5c96740840 [ 247.300194] RDX: 0000000000000002 RSI: 00007f5c97060000 RDI: 0000000000000001 [ 247.308159] RBP: 00007f5c97060000 R08: 000000000000000a R09: 00007f5c97059740 [ 247.316123] R10: 0000000000000001 R11: 0000000000000246 R12: 00007f5c96a14400 [ 247.324087] R13: 0000000000000002 R14: 0000000000000001 R15: 0000000000000000 [ 370.016340] INFO: task nvme-test:1772 blocked for more than 120 seconds. Fixes: 12d70958a2e8(blk-mq: don't fail allocating driver tag for stopped hw queue) Cc: stable@vger.kernel.org Signed-off-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Bart Van Assche <Bart.VanAssche@sandisk.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2017-06-06 15:22:00 +00:00
budget_token = blk_mq_get_dispatch_budget(rq->q);
if (budget_token < 0)
return false;
blk_mq_set_rq_budget_token(rq, budget_token);
if (!blk_mq_get_driver_tag(rq)) {
blk_mq_put_dispatch_budget(rq->q, budget_token);
return false;
}
return true;
}
/**
* blk_mq_try_issue_directly - Try to send a request directly to device driver.
* @hctx: Pointer of the associated hardware queue.
* @rq: Pointer to request to be sent.
*
* If the device has enough resources to accept a new request now, send the
* request directly to device driver. Else, insert at hctx->dispatch queue, so
* we can try send it another time in the future. Requests inserted at this
* queue have higher priority.
*/
static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
struct request *rq)
{
blk_status_t ret;
if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
blk_mq_insert_request(rq, 0);
return;
}
if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
blk_mq_insert_request(rq, 0);
block: Improve performance for BLK_MQ_F_BLOCKING drivers blk_mq_run_queue() runs the queue asynchronously if BLK_MQ_F_BLOCKING has been set. This is suboptimal since running the queue asynchronously is slower than running the queue synchronously. This patch modifies blk_mq_run_queue() as follows if BLK_MQ_F_BLOCKING has been set: - Run the queue synchronously if it is allowed to sleep. - Run the queue asynchronously if it is not allowed to sleep. Additionally, blk_mq_run_hw_queue(hctx, false) calls are modified into blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING) if the caller may be invoked from atomic context. The following caller chains have been reviewed: blk_mq_run_hw_queue(hctx, false) blk_mq_get_tag() /* may sleep, hence the functions it calls may also sleep */ blk_execute_rq() /* may sleep */ blk_mq_run_hw_queues(q, async=false) blk_freeze_queue_start() /* may sleep */ blk_mq_requeue_work() /* may sleep */ scsi_kick_queue() scsi_requeue_run_queue() /* may sleep */ scsi_run_host_queues() scsi_ioctl_reset() /* may sleep */ blk_mq_insert_requests(hctx, ctx, list, run_queue_async=false) blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list(plug, from_schedule=false) __blk_flush_plug(plug, from_schedule=false) blk_add_rq_to_plug() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_plug_issue_direct() blk_mq_flush_plug_list() /* see above */ blk_mq_dispatch_plug_list(plug, from_sched=false) blk_mq_flush_plug_list() /* see above */ blk_mq_try_issue_directly() blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */ blk_mq_try_issue_list_directly(hctx, list) blk_mq_insert_requests() /* see above */ Cc: Christoph Hellwig <hch@lst.de> Cc: Ming Lei <ming.lei@redhat.com> Signed-off-by: Bart Van Assche <bvanassche@acm.org> Link: https://lore.kernel.org/r/20230721172731.955724-4-bvanassche@acm.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-07-21 17:27:30 +00:00
blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
return;
}
ret = __blk_mq_issue_directly(hctx, rq, true);
switch (ret) {
case BLK_STS_OK:
break;
case BLK_STS_RESOURCE:
case BLK_STS_DEV_RESOURCE:
blk_mq_request_bypass_insert(rq, 0);
blk_mq_run_hw_queue(hctx, false);
break;
default:
blk_mq_end_request(rq, ret);
break;
}
}
static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
{
struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
blk_mq_insert_request(rq, 0);
return BLK_STS_OK;
}
if (!blk_mq_get_budget_and_tag(rq))
return BLK_STS_RESOURCE;
return __blk_mq_issue_directly(hctx, rq, last);
}
static void blk_mq_plug_issue_direct(struct blk_plug *plug)
{
struct blk_mq_hw_ctx *hctx = NULL;
struct request *rq;
int queued = 0;
blk_status_t ret = BLK_STS_OK;
while ((rq = rq_list_pop(&plug->mq_list))) {
bool last = rq_list_empty(plug->mq_list);
if (hctx != rq->mq_hctx) {
if (hctx) {
blk_mq_commit_rqs(hctx, queued, false);
queued = 0;
}
hctx = rq->mq_hctx;
}
ret = blk_mq_request_issue_directly(rq, last);
switch (ret) {
case BLK_STS_OK:
queued++;
break;
case BLK_STS_RESOURCE:
case BLK_STS_DEV_RESOURCE:
blk_mq_request_bypass_insert(rq, 0);
blk_mq_run_hw_queue(hctx, false);
goto out;
default:
blk_mq_end_request(rq, ret);
break;
}
}
out:
if (ret != BLK_STS_OK)
blk_mq_commit_rqs(hctx, queued, false);
}
static void __blk_mq_flush_plug_list(struct request_queue *q,
struct blk_plug *plug)
{
if (blk_queue_quiesced(q))
return;
q->mq_ops->queue_rqs(&plug->mq_list);
}
static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
{
struct blk_mq_hw_ctx *this_hctx = NULL;
struct blk_mq_ctx *this_ctx = NULL;
struct request *requeue_list = NULL;
struct request **requeue_lastp = &requeue_list;
unsigned int depth = 0;
bool is_passthrough = false;
LIST_HEAD(list);
do {
struct request *rq = rq_list_pop(&plug->mq_list);
if (!this_hctx) {
this_hctx = rq->mq_hctx;
this_ctx = rq->mq_ctx;
is_passthrough = blk_rq_is_passthrough(rq);
} else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
is_passthrough != blk_rq_is_passthrough(rq)) {
rq_list_add_tail(&requeue_lastp, rq);
continue;
}
list_add(&rq->queuelist, &list);
depth++;
} while (!rq_list_empty(plug->mq_list));
plug->mq_list = requeue_list;
trace_block_unplug(this_hctx->queue, depth, !from_sched);
percpu_ref_get(&this_hctx->queue->q_usage_counter);
/* passthrough requests should never be issued to the I/O scheduler */
if (is_passthrough) {
spin_lock(&this_hctx->lock);
list_splice_tail_init(&list, &this_hctx->dispatch);
spin_unlock(&this_hctx->lock);
blk_mq_run_hw_queue(this_hctx, from_sched);
} else if (this_hctx->queue->elevator) {
this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
&list, 0);
blk_mq_run_hw_queue(this_hctx, from_sched);
} else {
blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
}
percpu_ref_put(&this_hctx->queue->q_usage_counter);
}
void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
{
struct request *rq;
/*
* We may have been called recursively midway through handling
* plug->mq_list via a schedule() in the driver's queue_rq() callback.
* To avoid mq_list changing under our feet, clear rq_count early and
* bail out specifically if rq_count is 0 rather than checking
* whether the mq_list is empty.
*/
if (plug->rq_count == 0)
return;
plug->rq_count = 0;
if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
struct request_queue *q;
rq = rq_list_peek(&plug->mq_list);
q = rq->q;
/*
* Peek first request and see if we have a ->queue_rqs() hook.
* If we do, we can dispatch the whole plug list in one go. We
* already know at this point that all requests belong to the
* same queue, caller must ensure that's the case.
*/
if (q->mq_ops->queue_rqs) {
blk_mq_run_dispatch_ops(q,
__blk_mq_flush_plug_list(q, plug));
if (rq_list_empty(plug->mq_list))
return;
}
blk_mq_run_dispatch_ops(q,
blk_mq_plug_issue_direct(plug));
if (rq_list_empty(plug->mq_list))
return;
}
do {
blk_mq_dispatch_plug_list(plug, from_schedule);
} while (!rq_list_empty(plug->mq_list));
}
static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
struct list_head *list)
{
int queued = 0;
blk_status_t ret = BLK_STS_OK;
while (!list_empty(list)) {
struct request *rq = list_first_entry(list, struct request,
queuelist);
list_del_init(&rq->queuelist);
ret = blk_mq_request_issue_directly(rq, list_empty(list));
switch (ret) {
case BLK_STS_OK:
queued++;
break;
case BLK_STS_RESOURCE:
case BLK_STS_DEV_RESOURCE:
blk_mq_request_bypass_insert(rq, 0);
if (list_empty(list))
blk_mq_run_hw_queue(hctx, false);
goto out;
default:
blk_mq_end_request(rq, ret);
break;
}
}
out:
if (ret != BLK_STS_OK)
blk_mq_commit_rqs(hctx, queued, false);
}
static bool blk_mq_attempt_bio_merge(struct request_queue *q,
struct bio *bio, unsigned int nr_segs)
{
if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
if (blk_attempt_plug_merge(q, bio, nr_segs))
return true;
if (blk_mq_sched_bio_merge(q, bio, nr_segs))
return true;
}
return false;
}
static struct request *blk_mq_get_new_requests(struct request_queue *q,
struct blk_plug *plug,
struct bio *bio,
unsigned int nsegs)
{
struct blk_mq_alloc_data data = {
.q = q,
.nr_tags = 1,
.cmd_flags = bio->bi_opf,
};
struct request *rq;
rq_qos_throttle(q, bio);
if (plug) {
data.nr_tags = plug->nr_ios;
plug->nr_ios = 1;
data.cached_rq = &plug->cached_rq;
}
rq = __blk_mq_alloc_requests(&data);
if (rq)
return rq;
rq_qos_cleanup(q, bio);
if (bio->bi_opf & REQ_NOWAIT)
bio_wouldblock_error(bio);
return NULL;
}
/*
* Check if there is a suitable cached request and return it.
*/
static struct request *blk_mq_peek_cached_request(struct blk_plug *plug,
struct request_queue *q, blk_opf_t opf)
{
enum hctx_type type = blk_mq_get_hctx_type(opf);
struct request *rq;
if (!plug)
return NULL;
rq = rq_list_peek(&plug->cached_rq);
if (!rq || rq->q != q)
return NULL;
if (type != rq->mq_hctx->type &&
(type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT))
return NULL;
if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
return NULL;
return rq;
}
static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
struct bio *bio)
{
WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
/*
* If any qos ->throttle() end up blocking, we will have flushed the
* plug and hence killed the cached_rq list as well. Pop this entry
* before we throttle.
*/
plug->cached_rq = rq_list_next(rq);
rq_qos_throttle(rq->q, bio);
blk_mq_rq_time_init(rq, 0);
rq->cmd_flags = bio->bi_opf;
INIT_LIST_HEAD(&rq->queuelist);
}
static bool bio_unaligned(const struct bio *bio, struct request_queue *q)
{
unsigned int bs_mask = queue_logical_block_size(q) - 1;
/* .bi_sector of any zero sized bio need to be initialized */
if ((bio->bi_iter.bi_size & bs_mask) ||
((bio->bi_iter.bi_sector << SECTOR_SHIFT) & bs_mask))
return true;
return false;
}
/**
* blk_mq_submit_bio - Create and send a request to block device.
* @bio: Bio pointer.
*
* Builds up a request structure from @q and @bio and send to the device. The
* request may not be queued directly to hardware if:
* * This request can be merged with another one
* * We want to place request at plug queue for possible future merging
* * There is an IO scheduler active at this queue
*
* It will not queue the request if there is an error with the bio, or at the
* request creation.
*/
void blk_mq_submit_bio(struct bio *bio)
{
struct request_queue *q = bdev_get_queue(bio->bi_bdev);
struct blk_plug *plug = current->plug;
const int is_sync = op_is_sync(bio->bi_opf);
struct blk_mq_hw_ctx *hctx;
unsigned int nr_segs = 1;
struct request *rq;
block: Inline encryption support for blk-mq We must have some way of letting a storage device driver know what encryption context it should use for en/decrypting a request. However, it's the upper layers (like the filesystem/fscrypt) that know about and manages encryption contexts. As such, when the upper layer submits a bio to the block layer, and this bio eventually reaches a device driver with support for inline encryption, the device driver will need to have been told the encryption context for that bio. We want to communicate the encryption context from the upper layer to the storage device along with the bio, when the bio is submitted to the block layer. To do this, we add a struct bio_crypt_ctx to struct bio, which can represent an encryption context (note that we can't use the bi_private field in struct bio to do this because that field does not function to pass information across layers in the storage stack). We also introduce various functions to manipulate the bio_crypt_ctx and make the bio/request merging logic aware of the bio_crypt_ctx. We also make changes to blk-mq to make it handle bios with encryption contexts. blk-mq can merge many bios into the same request. These bios need to have contiguous data unit numbers (the necessary changes to blk-merge are also made to ensure this) - as such, it suffices to keep the data unit number of just the first bio, since that's all a storage driver needs to infer the data unit number to use for each data block in each bio in a request. blk-mq keeps track of the encryption context to be used for all the bios in a request with the request's rq_crypt_ctx. When the first bio is added to an empty request, blk-mq will program the encryption context of that bio into the request_queue's keyslot manager, and store the returned keyslot in the request's rq_crypt_ctx. All the functions to operate on encryption contexts are in blk-crypto.c. Upper layers only need to call bio_crypt_set_ctx with the encryption key, algorithm and data_unit_num; they don't have to worry about getting a keyslot for each encryption context, as blk-mq/blk-crypto handles that. Blk-crypto also makes it possible for request-based layered devices like dm-rq to make use of inline encryption hardware by cloning the rq_crypt_ctx and programming a keyslot in the new request_queue when necessary. Note that any user of the block layer can submit bios with an encryption context, such as filesystems, device-mapper targets, etc. Signed-off-by: Satya Tangirala <satyat@google.com> Reviewed-by: Eric Biggers <ebiggers@google.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-14 00:37:18 +00:00
blk_status_t ret;
block: Introduce zone write plugging Zone write plugging implements a per-zone "plug" for write operations to control the submission and execution order of write operations to sequential write required zones of a zoned block device. Per-zone plugging guarantees that at any time there is at most only one write request per zone being executed. This mechanism is intended to replace zone write locking which implements a similar per-zone write throttling at the scheduler level, but is implemented only by mq-deadline. Unlike zone write locking which operates on requests, zone write plugging operates on BIOs. A zone write plug is simply a BIO list that is atomically manipulated using a spinlock and a kblockd submission work. A write BIO to a zone is "plugged" to delay its execution if a write BIO for the same zone was already issued, that is, if a write request for the same zone is being executed. The next plugged BIO is unplugged and issued once the write request completes. This mechanism allows to: - Untangle zone write ordering from block IO schedulers. This allows removing the restriction on using mq-deadline for writing to zoned block devices. Any block IO scheduler, including "none" can be used. - Zone write plugging operates on BIOs instead of requests. Plugged BIOs waiting for execution thus do not hold scheduling tags and thus are not preventing other BIOs from executing (reads or writes to other zones). Depending on the workload, this can significantly improve the device use (higher queue depth operation) and performance. - Both blk-mq (request based) zoned devices and BIO-based zoned devices (e.g. device mapper) can use zone write plugging. It is mandatory for the former but optional for the latter. BIO-based drivers can use zone write plugging to implement write ordering guarantees, or the drivers can implement their own if needed. - The code is less invasive in the block layer and is mostly limited to blk-zoned.c with some small changes in blk-mq.c, blk-merge.c and bio.c. Zone write plugging is implemented using struct blk_zone_wplug. This structure includes a spinlock, a BIO list and a work structure to handle the submission of plugged BIOs. Zone write plugs structures are managed using a per-disk hash table. Plugging of zone write BIOs is done using the function blk_zone_write_plug_bio() which returns false if a BIO execution does not need to be delayed and true otherwise. This function is called from blk_mq_submit_bio() after a BIO is split to avoid large BIOs spanning multiple zones which would cause mishandling of zone write plugs. This ichange enables by default zone write plugging for any mq request-based block device. BIO-based device drivers can also use zone write plugging by expliclty calling blk_zone_write_plug_bio() in their ->submit_bio method. For such devices, the driver must ensure that a BIO passed to blk_zone_write_plug_bio() is already split and not straddling zone boundaries. Only write and write zeroes BIOs are plugged. Zone write plugging does not introduce any significant overhead for other operations. A BIO that is being handled through zone write plugging is flagged using the new BIO flag BIO_ZONE_WRITE_PLUGGING. A request handling a BIO flagged with this new flag is flagged with the new RQF_ZONE_WRITE_PLUGGING flag. The completion of BIOs and requests flagged trigger respectively calls to the functions blk_zone_write_bio_endio() and blk_zone_write_complete_request(). The latter function is used to trigger submission of the next plugged BIO using the zone plug work. blk_zone_write_bio_endio() does the same for BIO-based devices. This ensures that at any time, at most one request (blk-mq devices) or one BIO (BIO-based devices) is being executed for any zone. The handling of zone write plugs using a per-zone plug spinlock maximizes parallelism and device usage by allowing multiple zones to be writen simultaneously without lock contention. Zone write plugging ignores flush BIOs without data. Hovever, any flush BIO that has data is always plugged so that the write part of the flush sequence is serialized with other regular writes. Given that any BIO handled through zone write plugging will be the only BIO in flight for the target zone when it is executed, the unplugging and submission of a BIO will have no chance of successfully merging with plugged requests or requests in the scheduler. To overcome this potential performance degradation, blk_mq_submit_bio() calls the function blk_zone_write_plug_attempt_merge() to try to merge other plugged BIOs with the one just unplugged and submitted. Successful merging is signaled using blk_zone_write_plug_bio_merged(), called from bio_attempt_back_merge(). Furthermore, to avoid recalculating the number of segments of plugged BIOs to attempt merging, the number of segments of a plugged BIO is saved using the new struct bio field __bi_nr_segments. To avoid growing the size of struct bio, this field is added as a union with the bio_cookie field. This is safe to do as polling is always disabled for plugged BIOs. When BIOs are plugged in a zone write plug, the device request queue usage counter is always incremented. This reference is kept and reused for blk-mq devices when the plugged BIO is unplugged and submitted again using submit_bio_noacct_nocheck(). For this case, the unplugged BIO is already flagged with BIO_ZONE_WRITE_PLUGGING and blk_mq_submit_bio() proceeds directly to allocating a new request for the BIO, re-using the usage reference count taken when the BIO was plugged. This extra reference count is dropped in blk_zone_write_plug_attempt_merge() for any plugged BIO that is successfully merged. Given that BIO-based devices will not take this path, the extra reference is dropped after a plugged BIO is unplugged and submitted. Zone write plugs are dynamically allocated and managed using a hash table (an array of struct hlist_head) with RCU protection. A zone write plug is allocated when a write BIO is received for the zone and not freed until the zone is fully written, reset or finished. To detect when a zone write plug can be freed, the write state of each zone is tracked using a write pointer offset which corresponds to the offset of a zone write pointer relative to the zone start. Write operations always increment this write pointer offset. Zone reset operations set it to 0 and zone finish operations set it to the zone size. If a write error happens, the wp_offset value of a zone write plug may become incorrect and out of sync with the device managed write pointer. This is handled using the zone write plug flag BLK_ZONE_WPLUG_ERROR. The function blk_zone_wplug_handle_error() is called from the new disk zone write plug work when this flag is set. This function executes a report zone to update the zone write pointer offset to the current value as indicated by the device. The disk zone write plug work is scheduled whenever a BIO flagged with BIO_ZONE_WRITE_PLUGGING completes with an error or when bio_zone_wplug_prepare_bio() detects an unaligned write. Once scheduled, the disk zone write plugs work keeps running until all zone errors are handled. To match the new data structures used for zoned disks, the function disk_free_zone_bitmaps() is renamed to the more generic disk_free_zone_resources(). The function disk_init_zone_resources() is also introduced to initialize zone write plugs resources when a gendisk is allocated. In order to guarantee that the user can simultaneously write up to a number of zones equal to a device max active zone limit or max open zone limit, zone write plugs are allocated using a mempool sized to the maximum of these 2 device limits. For a device that does not have active and open zone limits, 128 is used as the default mempool size. If a change to the device active and open zone limits is detected, the disk mempool is resized when blk_revalidate_disk_zones() is executed. This commit contains contributions from Christoph Hellwig <hch@lst.de>. Signed-off-by: Damien Le Moal <dlemoal@kernel.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Tested-by: Hans Holmberg <hans.holmberg@wdc.com> Tested-by: Dennis Maisenbacher <dennis.maisenbacher@wdc.com> Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Link: https://lore.kernel.org/r/20240408014128.205141-8-dlemoal@kernel.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2024-04-08 01:41:07 +00:00
/*
* If the plug has a cached request for this queue, try to use it.
*/
rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf);
/*
* A BIO that was released from a zone write plug has already been
* through the preparation in this function, already holds a reference
* on the queue usage counter, and is the only write BIO in-flight for
* the target zone. Go straight to preparing a request for it.
*/
if (bio_zone_write_plugging(bio)) {
nr_segs = bio->__bi_nr_segments;
if (rq)
blk_queue_exit(q);
goto new_request;
}
bio = blk_queue_bounce(bio, q);
/*
* The cached request already holds a q_usage_counter reference and we
* don't have to acquire a new one if we use it.
*/
if (!rq) {
if (unlikely(bio_queue_enter(bio)))
return;
}
/*
* Device reconfiguration may change logical block size, so alignment
* check has to be done with queue usage counter held
*/
if (unlikely(bio_unaligned(bio, q))) {
bio_io_error(bio);
goto queue_exit;
}
if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
if (!bio)
goto queue_exit;
}
if (!bio_integrity_prep(bio))
goto queue_exit;
if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
goto queue_exit;
block: Introduce zone write plugging Zone write plugging implements a per-zone "plug" for write operations to control the submission and execution order of write operations to sequential write required zones of a zoned block device. Per-zone plugging guarantees that at any time there is at most only one write request per zone being executed. This mechanism is intended to replace zone write locking which implements a similar per-zone write throttling at the scheduler level, but is implemented only by mq-deadline. Unlike zone write locking which operates on requests, zone write plugging operates on BIOs. A zone write plug is simply a BIO list that is atomically manipulated using a spinlock and a kblockd submission work. A write BIO to a zone is "plugged" to delay its execution if a write BIO for the same zone was already issued, that is, if a write request for the same zone is being executed. The next plugged BIO is unplugged and issued once the write request completes. This mechanism allows to: - Untangle zone write ordering from block IO schedulers. This allows removing the restriction on using mq-deadline for writing to zoned block devices. Any block IO scheduler, including "none" can be used. - Zone write plugging operates on BIOs instead of requests. Plugged BIOs waiting for execution thus do not hold scheduling tags and thus are not preventing other BIOs from executing (reads or writes to other zones). Depending on the workload, this can significantly improve the device use (higher queue depth operation) and performance. - Both blk-mq (request based) zoned devices and BIO-based zoned devices (e.g. device mapper) can use zone write plugging. It is mandatory for the former but optional for the latter. BIO-based drivers can use zone write plugging to implement write ordering guarantees, or the drivers can implement their own if needed. - The code is less invasive in the block layer and is mostly limited to blk-zoned.c with some small changes in blk-mq.c, blk-merge.c and bio.c. Zone write plugging is implemented using struct blk_zone_wplug. This structure includes a spinlock, a BIO list and a work structure to handle the submission of plugged BIOs. Zone write plugs structures are managed using a per-disk hash table. Plugging of zone write BIOs is done using the function blk_zone_write_plug_bio() which returns false if a BIO execution does not need to be delayed and true otherwise. This function is called from blk_mq_submit_bio() after a BIO is split to avoid large BIOs spanning multiple zones which would cause mishandling of zone write plugs. This ichange enables by default zone write plugging for any mq request-based block device. BIO-based device drivers can also use zone write plugging by expliclty calling blk_zone_write_plug_bio() in their ->submit_bio method. For such devices, the driver must ensure that a BIO passed to blk_zone_write_plug_bio() is already split and not straddling zone boundaries. Only write and write zeroes BIOs are plugged. Zone write plugging does not introduce any significant overhead for other operations. A BIO that is being handled through zone write plugging is flagged using the new BIO flag BIO_ZONE_WRITE_PLUGGING. A request handling a BIO flagged with this new flag is flagged with the new RQF_ZONE_WRITE_PLUGGING flag. The completion of BIOs and requests flagged trigger respectively calls to the functions blk_zone_write_bio_endio() and blk_zone_write_complete_request(). The latter function is used to trigger submission of the next plugged BIO using the zone plug work. blk_zone_write_bio_endio() does the same for BIO-based devices. This ensures that at any time, at most one request (blk-mq devices) or one BIO (BIO-based devices) is being executed for any zone. The handling of zone write plugs using a per-zone plug spinlock maximizes parallelism and device usage by allowing multiple zones to be writen simultaneously without lock contention. Zone write plugging ignores flush BIOs without data. Hovever, any flush BIO that has data is always plugged so that the write part of the flush sequence is serialized with other regular writes. Given that any BIO handled through zone write plugging will be the only BIO in flight for the target zone when it is executed, the unplugging and submission of a BIO will have no chance of successfully merging with plugged requests or requests in the scheduler. To overcome this potential performance degradation, blk_mq_submit_bio() calls the function blk_zone_write_plug_attempt_merge() to try to merge other plugged BIOs with the one just unplugged and submitted. Successful merging is signaled using blk_zone_write_plug_bio_merged(), called from bio_attempt_back_merge(). Furthermore, to avoid recalculating the number of segments of plugged BIOs to attempt merging, the number of segments of a plugged BIO is saved using the new struct bio field __bi_nr_segments. To avoid growing the size of struct bio, this field is added as a union with the bio_cookie field. This is safe to do as polling is always disabled for plugged BIOs. When BIOs are plugged in a zone write plug, the device request queue usage counter is always incremented. This reference is kept and reused for blk-mq devices when the plugged BIO is unplugged and submitted again using submit_bio_noacct_nocheck(). For this case, the unplugged BIO is already flagged with BIO_ZONE_WRITE_PLUGGING and blk_mq_submit_bio() proceeds directly to allocating a new request for the BIO, re-using the usage reference count taken when the BIO was plugged. This extra reference count is dropped in blk_zone_write_plug_attempt_merge() for any plugged BIO that is successfully merged. Given that BIO-based devices will not take this path, the extra reference is dropped after a plugged BIO is unplugged and submitted. Zone write plugs are dynamically allocated and managed using a hash table (an array of struct hlist_head) with RCU protection. A zone write plug is allocated when a write BIO is received for the zone and not freed until the zone is fully written, reset or finished. To detect when a zone write plug can be freed, the write state of each zone is tracked using a write pointer offset which corresponds to the offset of a zone write pointer relative to the zone start. Write operations always increment this write pointer offset. Zone reset operations set it to 0 and zone finish operations set it to the zone size. If a write error happens, the wp_offset value of a zone write plug may become incorrect and out of sync with the device managed write pointer. This is handled using the zone write plug flag BLK_ZONE_WPLUG_ERROR. The function blk_zone_wplug_handle_error() is called from the new disk zone write plug work when this flag is set. This function executes a report zone to update the zone write pointer offset to the current value as indicated by the device. The disk zone write plug work is scheduled whenever a BIO flagged with BIO_ZONE_WRITE_PLUGGING completes with an error or when bio_zone_wplug_prepare_bio() detects an unaligned write. Once scheduled, the disk zone write plugs work keeps running until all zone errors are handled. To match the new data structures used for zoned disks, the function disk_free_zone_bitmaps() is renamed to the more generic disk_free_zone_resources(). The function disk_init_zone_resources() is also introduced to initialize zone write plugs resources when a gendisk is allocated. In order to guarantee that the user can simultaneously write up to a number of zones equal to a device max active zone limit or max open zone limit, zone write plugs are allocated using a mempool sized to the maximum of these 2 device limits. For a device that does not have active and open zone limits, 128 is used as the default mempool size. If a change to the device active and open zone limits is detected, the disk mempool is resized when blk_revalidate_disk_zones() is executed. This commit contains contributions from Christoph Hellwig <hch@lst.de>. Signed-off-by: Damien Le Moal <dlemoal@kernel.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Tested-by: Hans Holmberg <hans.holmberg@wdc.com> Tested-by: Dennis Maisenbacher <dennis.maisenbacher@wdc.com> Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Link: https://lore.kernel.org/r/20240408014128.205141-8-dlemoal@kernel.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2024-04-08 01:41:07 +00:00
if (blk_queue_is_zoned(q) && blk_zone_plug_bio(bio, nr_segs))
goto queue_exit;
new_request:
if (!rq) {
rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
if (unlikely(!rq))
goto queue_exit;
} else {
blk_mq_use_cached_rq(rq, plug, bio);
}
block: hook up writeback throttling Enable throttling of buffered writeback to make it a lot more smooth, and has way less impact on other system activity. Background writeback should be, by definition, background activity. The fact that we flush huge bundles of it at the time means that it potentially has heavy impacts on foreground workloads, which isn't ideal. We can't easily limit the sizes of writes that we do, since that would impact file system layout in the presence of delayed allocation. So just throttle back buffered writeback, unless someone is waiting for it. The algorithm for when to throttle takes its inspiration in the CoDel networking scheduling algorithm. Like CoDel, blk-wb monitors the minimum latencies of requests over a window of time. In that window of time, if the minimum latency of any request exceeds a given target, then a scale count is incremented and the queue depth is shrunk. The next monitoring window is shrunk accordingly. Unlike CoDel, if we hit a window that exhibits good behavior, then we simply increment the scale count and re-calculate the limits for that scale value. This prevents us from oscillating between a close-to-ideal value and max all the time, instead remaining in the windows where we get good behavior. Unlike CoDel, blk-wb allows the scale count to to negative. This happens if we primarily have writes going on. Unlike positive scale counts, this doesn't change the size of the monitoring window. When the heavy writers finish, blk-bw quickly snaps back to it's stable state of a zero scale count. The patch registers a sysfs entry, 'wb_lat_usec'. This sets the latency target to me met. It defaults to 2 msec for non-rotational storage, and 75 msec for rotational storage. Setting this value to '0' disables blk-wb. Generally, a user would not have to touch this setting. We don't enable WBT on devices that are managed with CFQ, and have a non-root block cgroup attached. If we have a proportional share setup on this particular disk, then the wbt throttling will interfere with that. We don't have a strong need for wbt for that case, since we will rely on CFQ doing that for us. Signed-off-by: Jens Axboe <axboe@fb.com>
2016-11-09 19:38:14 +00:00
trace_block_getrq(bio);
rq_qos_track(q, rq, bio);
blk_mq_bio_to_request(rq, bio, nr_segs);
ret = blk_crypto_rq_get_keyslot(rq);
block: Inline encryption support for blk-mq We must have some way of letting a storage device driver know what encryption context it should use for en/decrypting a request. However, it's the upper layers (like the filesystem/fscrypt) that know about and manages encryption contexts. As such, when the upper layer submits a bio to the block layer, and this bio eventually reaches a device driver with support for inline encryption, the device driver will need to have been told the encryption context for that bio. We want to communicate the encryption context from the upper layer to the storage device along with the bio, when the bio is submitted to the block layer. To do this, we add a struct bio_crypt_ctx to struct bio, which can represent an encryption context (note that we can't use the bi_private field in struct bio to do this because that field does not function to pass information across layers in the storage stack). We also introduce various functions to manipulate the bio_crypt_ctx and make the bio/request merging logic aware of the bio_crypt_ctx. We also make changes to blk-mq to make it handle bios with encryption contexts. blk-mq can merge many bios into the same request. These bios need to have contiguous data unit numbers (the necessary changes to blk-merge are also made to ensure this) - as such, it suffices to keep the data unit number of just the first bio, since that's all a storage driver needs to infer the data unit number to use for each data block in each bio in a request. blk-mq keeps track of the encryption context to be used for all the bios in a request with the request's rq_crypt_ctx. When the first bio is added to an empty request, blk-mq will program the encryption context of that bio into the request_queue's keyslot manager, and store the returned keyslot in the request's rq_crypt_ctx. All the functions to operate on encryption contexts are in blk-crypto.c. Upper layers only need to call bio_crypt_set_ctx with the encryption key, algorithm and data_unit_num; they don't have to worry about getting a keyslot for each encryption context, as blk-mq/blk-crypto handles that. Blk-crypto also makes it possible for request-based layered devices like dm-rq to make use of inline encryption hardware by cloning the rq_crypt_ctx and programming a keyslot in the new request_queue when necessary. Note that any user of the block layer can submit bios with an encryption context, such as filesystems, device-mapper targets, etc. Signed-off-by: Satya Tangirala <satyat@google.com> Reviewed-by: Eric Biggers <ebiggers@google.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-14 00:37:18 +00:00
if (ret != BLK_STS_OK) {
bio->bi_status = ret;
bio_endio(bio);
blk_mq_free_request(rq);
return;
block: Inline encryption support for blk-mq We must have some way of letting a storage device driver know what encryption context it should use for en/decrypting a request. However, it's the upper layers (like the filesystem/fscrypt) that know about and manages encryption contexts. As such, when the upper layer submits a bio to the block layer, and this bio eventually reaches a device driver with support for inline encryption, the device driver will need to have been told the encryption context for that bio. We want to communicate the encryption context from the upper layer to the storage device along with the bio, when the bio is submitted to the block layer. To do this, we add a struct bio_crypt_ctx to struct bio, which can represent an encryption context (note that we can't use the bi_private field in struct bio to do this because that field does not function to pass information across layers in the storage stack). We also introduce various functions to manipulate the bio_crypt_ctx and make the bio/request merging logic aware of the bio_crypt_ctx. We also make changes to blk-mq to make it handle bios with encryption contexts. blk-mq can merge many bios into the same request. These bios need to have contiguous data unit numbers (the necessary changes to blk-merge are also made to ensure this) - as such, it suffices to keep the data unit number of just the first bio, since that's all a storage driver needs to infer the data unit number to use for each data block in each bio in a request. blk-mq keeps track of the encryption context to be used for all the bios in a request with the request's rq_crypt_ctx. When the first bio is added to an empty request, blk-mq will program the encryption context of that bio into the request_queue's keyslot manager, and store the returned keyslot in the request's rq_crypt_ctx. All the functions to operate on encryption contexts are in blk-crypto.c. Upper layers only need to call bio_crypt_set_ctx with the encryption key, algorithm and data_unit_num; they don't have to worry about getting a keyslot for each encryption context, as blk-mq/blk-crypto handles that. Blk-crypto also makes it possible for request-based layered devices like dm-rq to make use of inline encryption hardware by cloning the rq_crypt_ctx and programming a keyslot in the new request_queue when necessary. Note that any user of the block layer can submit bios with an encryption context, such as filesystems, device-mapper targets, etc. Signed-off-by: Satya Tangirala <satyat@google.com> Reviewed-by: Eric Biggers <ebiggers@google.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-14 00:37:18 +00:00
}
if (bio_zone_write_plugging(bio))
blk_zone_write_plug_init_request(rq);
if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
return;
if (plug) {
blk_add_rq_to_plug(plug, rq);
return;
}
hctx = rq->mq_hctx;
if ((rq->rq_flags & RQF_USE_SCHED) ||
(hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
blk_mq_insert_request(rq, 0);
blk_mq_run_hw_queue(hctx, true);
} else {
blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
}
return;
queue_exit:
/*
* Don't drop the queue reference if we were trying to use a cached
* request and thus didn't acquire one.
*/
if (!rq)
blk_queue_exit(q);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
#ifdef CONFIG_BLK_MQ_STACKING
/**
* blk_insert_cloned_request - Helper for stacking drivers to submit a request
* @rq: the request being queued
*/
blk_status_t blk_insert_cloned_request(struct request *rq)
{
struct request_queue *q = rq->q;
unsigned int max_sectors = blk_queue_get_max_sectors(rq);
unsigned int max_segments = blk_rq_get_max_segments(rq);
blk_status_t ret;
if (blk_rq_sectors(rq) > max_sectors) {
/*
* SCSI device does not have a good way to return if
* Write Same/Zero is actually supported. If a device rejects
* a non-read/write command (discard, write same,etc.) the
* low-level device driver will set the relevant queue limit to
* 0 to prevent blk-lib from issuing more of the offending
* operations. Commands queued prior to the queue limit being
* reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
* errors being propagated to upper layers.
*/
if (max_sectors == 0)
return BLK_STS_NOTSUPP;
printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
__func__, blk_rq_sectors(rq), max_sectors);
return BLK_STS_IOERR;
}
/*
* The queue settings related to segment counting may differ from the
* original queue.
*/
rq->nr_phys_segments = blk_recalc_rq_segments(rq);
if (rq->nr_phys_segments > max_segments) {
printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
__func__, rq->nr_phys_segments, max_segments);
return BLK_STS_IOERR;
}
if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
return BLK_STS_IOERR;
ret = blk_crypto_rq_get_keyslot(rq);
if (ret != BLK_STS_OK)
return ret;
blk_account_io_start(rq);
/*
* Since we have a scheduler attached on the top device,
* bypass a potential scheduler on the bottom device for
* insert.
*/
blk_mq_run_dispatch_ops(q,
ret = blk_mq_request_issue_directly(rq, true));
if (ret)
blk_account_io_done(rq, blk_time_get_ns());
return ret;
}
EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
/**
* blk_rq_unprep_clone - Helper function to free all bios in a cloned request
* @rq: the clone request to be cleaned up
*
* Description:
* Free all bios in @rq for a cloned request.
*/
void blk_rq_unprep_clone(struct request *rq)
{
struct bio *bio;
while ((bio = rq->bio) != NULL) {
rq->bio = bio->bi_next;
bio_put(bio);
}
}
EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
/**
* blk_rq_prep_clone - Helper function to setup clone request
* @rq: the request to be setup
* @rq_src: original request to be cloned
* @bs: bio_set that bios for clone are allocated from
* @gfp_mask: memory allocation mask for bio
* @bio_ctr: setup function to be called for each clone bio.
* Returns %0 for success, non %0 for failure.
* @data: private data to be passed to @bio_ctr
*
* Description:
* Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
* Also, pages which the original bios are pointing to are not copied
* and the cloned bios just point same pages.
* So cloned bios must be completed before original bios, which means
* the caller must complete @rq before @rq_src.
*/
int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
struct bio_set *bs, gfp_t gfp_mask,
int (*bio_ctr)(struct bio *, struct bio *, void *),
void *data)
{
struct bio *bio, *bio_src;
if (!bs)
bs = &fs_bio_set;
__rq_for_each_bio(bio_src, rq_src) {
bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
bs);
if (!bio)
goto free_and_out;
if (bio_ctr && bio_ctr(bio, bio_src, data))
goto free_and_out;
if (rq->bio) {
rq->biotail->bi_next = bio;
rq->biotail = bio;
} else {
rq->bio = rq->biotail = bio;
}
bio = NULL;
}
/* Copy attributes of the original request to the clone request. */
rq->__sector = blk_rq_pos(rq_src);
rq->__data_len = blk_rq_bytes(rq_src);
if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
rq->special_vec = rq_src->special_vec;
}
rq->nr_phys_segments = rq_src->nr_phys_segments;
rq->ioprio = rq_src->ioprio;
rq->write_hint = rq_src->write_hint;
if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
goto free_and_out;
return 0;
free_and_out:
if (bio)
bio_put(bio);
blk_rq_unprep_clone(rq);
return -ENOMEM;
}
EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
#endif /* CONFIG_BLK_MQ_STACKING */
/*
* Steal bios from a request and add them to a bio list.
* The request must not have been partially completed before.
*/
void blk_steal_bios(struct bio_list *list, struct request *rq)
{
if (rq->bio) {
if (list->tail)
list->tail->bi_next = rq->bio;
else
list->head = rq->bio;
list->tail = rq->biotail;
rq->bio = NULL;
rq->biotail = NULL;
}
rq->__data_len = 0;
}
EXPORT_SYMBOL_GPL(blk_steal_bios);
static size_t order_to_size(unsigned int order)
{
return (size_t)PAGE_SIZE << order;
}
/* called before freeing request pool in @tags */
static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
struct blk_mq_tags *tags)
{
struct page *page;
unsigned long flags;
/*
* There is no need to clear mapping if driver tags is not initialized
* or the mapping belongs to the driver tags.
*/
if (!drv_tags || drv_tags == tags)
return;
list_for_each_entry(page, &tags->page_list, lru) {
unsigned long start = (unsigned long)page_address(page);
unsigned long end = start + order_to_size(page->private);
int i;
for (i = 0; i < drv_tags->nr_tags; i++) {
struct request *rq = drv_tags->rqs[i];
unsigned long rq_addr = (unsigned long)rq;
if (rq_addr >= start && rq_addr < end) {
WARN_ON_ONCE(req_ref_read(rq) != 0);
cmpxchg(&drv_tags->rqs[i], rq, NULL);
}
}
}
/*
* Wait until all pending iteration is done.
*
* Request reference is cleared and it is guaranteed to be observed
* after the ->lock is released.
*/
spin_lock_irqsave(&drv_tags->lock, flags);
spin_unlock_irqrestore(&drv_tags->lock, flags);
}
void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
unsigned int hctx_idx)
{
struct blk_mq_tags *drv_tags;
struct page *page;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
if (list_empty(&tags->page_list))
return;
if (blk_mq_is_shared_tags(set->flags))
drv_tags = set->shared_tags;
else
drv_tags = set->tags[hctx_idx];
if (tags->static_rqs && set->ops->exit_request) {
int i;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
for (i = 0; i < tags->nr_tags; i++) {
struct request *rq = tags->static_rqs[i];
if (!rq)
continue;
set->ops->exit_request(set, rq, hctx_idx);
tags->static_rqs[i] = NULL;
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
blk_mq_clear_rq_mapping(drv_tags, tags);
while (!list_empty(&tags->page_list)) {
page = list_first_entry(&tags->page_list, struct page, lru);
list_del_init(&page->lru);
/*
* Remove kmemleak object previously allocated in
* blk_mq_alloc_rqs().
*/
kmemleak_free(page_address(page));
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
__free_pages(page, page->private);
}
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
void blk_mq_free_rq_map(struct blk_mq_tags *tags)
{
kfree(tags->rqs);
tags->rqs = NULL;
kfree(tags->static_rqs);
tags->static_rqs = NULL;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
blk_mq_free_tags(tags);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
unsigned int hctx_idx)
{
int i;
for (i = 0; i < set->nr_maps; i++) {
unsigned int start = set->map[i].queue_offset;
unsigned int end = start + set->map[i].nr_queues;
if (hctx_idx >= start && hctx_idx < end)
break;
}
if (i >= set->nr_maps)
i = HCTX_TYPE_DEFAULT;
return i;
}
static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
unsigned int hctx_idx)
{
enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
}
static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
unsigned int hctx_idx,
unsigned int nr_tags,
unsigned int reserved_tags)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
int node = blk_mq_get_hctx_node(set, hctx_idx);
struct blk_mq_tags *tags;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
if (node == NUMA_NO_NODE)
node = set->numa_node;
tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
if (!tags)
return NULL;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
treewide: kzalloc_node() -> kcalloc_node() The kzalloc_node() function has a 2-factor argument form, kcalloc_node(). This patch replaces cases of: kzalloc_node(a * b, gfp, node) with: kcalloc_node(a * b, gfp, node) as well as handling cases of: kzalloc_node(a * b * c, gfp, node) with: kzalloc_node(array3_size(a, b, c), gfp, node) as it's slightly less ugly than: kcalloc_node(array_size(a, b), c, gfp, node) This does, however, attempt to ignore constant size factors like: kzalloc_node(4 * 1024, gfp, node) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kzalloc_node( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kzalloc_node( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kzalloc_node( - sizeof(u8) * (COUNT) + COUNT , ...) | kzalloc_node( - sizeof(__u8) * (COUNT) + COUNT , ...) | kzalloc_node( - sizeof(char) * (COUNT) + COUNT , ...) | kzalloc_node( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kzalloc_node( - sizeof(u8) * COUNT + COUNT , ...) | kzalloc_node( - sizeof(__u8) * COUNT + COUNT , ...) | kzalloc_node( - sizeof(char) * COUNT + COUNT , ...) | kzalloc_node( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kzalloc_node + kcalloc_node ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kzalloc_node( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc_node( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc_node( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc_node( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc_node( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc_node( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc_node( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc_node( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kzalloc_node( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc_node( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc_node( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc_node( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc_node( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kzalloc_node( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kzalloc_node( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kzalloc_node(C1 * C2 * C3, ...) | kzalloc_node( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kzalloc_node( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kzalloc_node( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kzalloc_node( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kzalloc_node(sizeof(THING) * C2, ...) | kzalloc_node(sizeof(TYPE) * C2, ...) | kzalloc_node(C1 * C2 * C3, ...) | kzalloc_node(C1 * C2, ...) | - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kzalloc_node + kcalloc_node ( - (E1) * E2 + E1, E2 , ...) | - kzalloc_node + kcalloc_node ( - (E1) * (E2) + E1, E2 , ...) | - kzalloc_node + kcalloc_node ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-12 21:04:20 +00:00
tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
blk-mq: Avoid memory reclaim when remapping queues While stressing memory and IO at the same time we changed SMT settings, we were able to consistently trigger deadlocks in the mm system, which froze the entire machine. I think that under memory stress conditions, the large allocations performed by blk_mq_init_rq_map may trigger a reclaim, which stalls waiting on the block layer remmaping completion, thus deadlocking the system. The trace below was collected after the machine stalled, waiting for the hotplug event completion. The simplest fix for this is to make allocations in this path non-reclaimable, with GFP_NOIO. With this patch, We couldn't hit the issue anymore. This should apply on top of Jens's for-next branch cleanly. Changes since v1: - Use GFP_NOIO instead of GFP_NOWAIT. Call Trace: [c000000f0160aaf0] [c000000f0160ab50] 0xc000000f0160ab50 (unreliable) [c000000f0160acc0] [c000000000016624] __switch_to+0x2e4/0x430 [c000000f0160ad20] [c000000000b1a880] __schedule+0x310/0x9b0 [c000000f0160ae00] [c000000000b1af68] schedule+0x48/0xc0 [c000000f0160ae30] [c000000000b1b4b0] schedule_preempt_disabled+0x20/0x30 [c000000f0160ae50] [c000000000b1d4fc] __mutex_lock_slowpath+0xec/0x1f0 [c000000f0160aed0] [c000000000b1d678] mutex_lock+0x78/0xa0 [c000000f0160af00] [d000000019413cac] xfs_reclaim_inodes_ag+0x33c/0x380 [xfs] [c000000f0160b0b0] [d000000019415164] xfs_reclaim_inodes_nr+0x54/0x70 [xfs] [c000000f0160b0f0] [d0000000194297f8] xfs_fs_free_cached_objects+0x38/0x60 [xfs] [c000000f0160b120] [c0000000003172c8] super_cache_scan+0x1f8/0x210 [c000000f0160b190] [c00000000026301c] shrink_slab.part.13+0x21c/0x4c0 [c000000f0160b2d0] [c000000000268088] shrink_zone+0x2d8/0x3c0 [c000000f0160b380] [c00000000026834c] do_try_to_free_pages+0x1dc/0x520 [c000000f0160b450] [c00000000026876c] try_to_free_pages+0xdc/0x250 [c000000f0160b4e0] [c000000000251978] __alloc_pages_nodemask+0x868/0x10d0 [c000000f0160b6f0] [c000000000567030] blk_mq_init_rq_map+0x160/0x380 [c000000f0160b7a0] [c00000000056758c] blk_mq_map_swqueue+0x33c/0x360 [c000000f0160b820] [c000000000567904] blk_mq_queue_reinit+0x64/0xb0 [c000000f0160b850] [c00000000056a16c] blk_mq_queue_reinit_notify+0x19c/0x250 [c000000f0160b8a0] [c0000000000f5d38] notifier_call_chain+0x98/0x100 [c000000f0160b8f0] [c0000000000c5fb0] __cpu_notify+0x70/0xe0 [c000000f0160b930] [c0000000000c63c4] notify_prepare+0x44/0xb0 [c000000f0160b9b0] [c0000000000c52f4] cpuhp_invoke_callback+0x84/0x250 [c000000f0160ba10] [c0000000000c570c] cpuhp_up_callbacks+0x5c/0x120 [c000000f0160ba60] [c0000000000c7cb8] _cpu_up+0xf8/0x1d0 [c000000f0160bac0] [c0000000000c7eb0] do_cpu_up+0x120/0x150 [c000000f0160bb40] [c0000000006fe024] cpu_subsys_online+0x64/0xe0 [c000000f0160bb90] [c0000000006f5124] device_online+0xb4/0x120 [c000000f0160bbd0] [c0000000006f5244] online_store+0xb4/0xc0 [c000000f0160bc20] [c0000000006f0a68] dev_attr_store+0x68/0xa0 [c000000f0160bc60] [c0000000003ccc30] sysfs_kf_write+0x80/0xb0 [c000000f0160bca0] [c0000000003cbabc] kernfs_fop_write+0x17c/0x250 [c000000f0160bcf0] [c00000000030fe6c] __vfs_write+0x6c/0x1e0 [c000000f0160bd90] [c000000000311490] vfs_write+0xd0/0x270 [c000000f0160bde0] [c0000000003131fc] SyS_write+0x6c/0x110 [c000000f0160be30] [c000000000009204] system_call+0x38/0xec Signed-off-by: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com> Cc: Brian King <brking@linux.vnet.ibm.com> Cc: Douglas Miller <dougmill@linux.vnet.ibm.com> Cc: linux-block@vger.kernel.org Cc: linux-scsi@vger.kernel.org Signed-off-by: Jens Axboe <axboe@fb.com>
2016-12-06 15:31:44 +00:00
GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
node);
if (!tags->rqs)
goto err_free_tags;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
treewide: kzalloc_node() -> kcalloc_node() The kzalloc_node() function has a 2-factor argument form, kcalloc_node(). This patch replaces cases of: kzalloc_node(a * b, gfp, node) with: kcalloc_node(a * b, gfp, node) as well as handling cases of: kzalloc_node(a * b * c, gfp, node) with: kzalloc_node(array3_size(a, b, c), gfp, node) as it's slightly less ugly than: kcalloc_node(array_size(a, b), c, gfp, node) This does, however, attempt to ignore constant size factors like: kzalloc_node(4 * 1024, gfp, node) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kzalloc_node( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kzalloc_node( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kzalloc_node( - sizeof(u8) * (COUNT) + COUNT , ...) | kzalloc_node( - sizeof(__u8) * (COUNT) + COUNT , ...) | kzalloc_node( - sizeof(char) * (COUNT) + COUNT , ...) | kzalloc_node( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kzalloc_node( - sizeof(u8) * COUNT + COUNT , ...) | kzalloc_node( - sizeof(__u8) * COUNT + COUNT , ...) | kzalloc_node( - sizeof(char) * COUNT + COUNT , ...) | kzalloc_node( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kzalloc_node + kcalloc_node ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kzalloc_node( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc_node( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc_node( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc_node( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc_node( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc_node( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc_node( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc_node( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kzalloc_node( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc_node( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc_node( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc_node( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc_node( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kzalloc_node( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kzalloc_node( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc_node( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kzalloc_node(C1 * C2 * C3, ...) | kzalloc_node( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kzalloc_node( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kzalloc_node( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kzalloc_node( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kzalloc_node(sizeof(THING) * C2, ...) | kzalloc_node(sizeof(TYPE) * C2, ...) | kzalloc_node(C1 * C2 * C3, ...) | kzalloc_node(C1 * C2, ...) | - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kzalloc_node + kcalloc_node ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kzalloc_node + kcalloc_node ( - (E1) * E2 + E1, E2 , ...) | - kzalloc_node + kcalloc_node ( - (E1) * (E2) + E1, E2 , ...) | - kzalloc_node + kcalloc_node ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-12 21:04:20 +00:00
tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
node);
if (!tags->static_rqs)
goto err_free_rqs;
return tags;
err_free_rqs:
kfree(tags->rqs);
err_free_tags:
blk_mq_free_tags(tags);
return NULL;
}
blk-mq: replace timeout synchronization with a RCU and generation based scheme Currently, blk-mq timeout path synchronizes against the usual issue/completion path using a complex scheme involving atomic bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence rules. Unfortunately, it contains quite a few holes. There's a complex dancing around REQ_ATOM_STARTED and REQ_ATOM_COMPLETE between issue/completion and timeout paths; however, they don't have a synchronization point across request recycle instances and it isn't clear what the barriers add. blk_mq_check_expired() can easily read STARTED from N-2'th iteration, deadline from N-1'th, blk_mark_rq_complete() against Nth instance. In fact, it's pretty easy to make blk_mq_check_expired() terminate a later instance of a request. If we induce 5 sec delay before time_after_eq() test in blk_mq_check_expired(), shorten the timeout to 2s, and issue back-to-back large IOs, blk-mq starts timing out requests spuriously pretty quickly. Nothing actually timed out. It just made the call on a recycle instance of a request and then terminated a later instance long after the original instance finished. The scenario isn't theoretical either. This patch replaces the broken synchronization mechanism with a RCU and generation number based one. 1. Each request has a u64 generation + state value, which can be updated only by the request owner. Whenever a request becomes in-flight, the generation number gets bumped up too. This provides the basis for the timeout path to distinguish different recycle instances of the request. Also, marking a request in-flight and setting its deadline are protected with a seqcount so that the timeout path can fetch both values coherently. 2. The timeout path fetches the generation, state and deadline. If the verdict is timeout, it records the generation into a dedicated request abortion field and does RCU wait. 3. The completion path is also protected by RCU (from the previous patch) and checks whether the current generation number and state match the abortion field. If so, it skips completion. 4. The timeout path, after RCU wait, scans requests again and terminates the ones whose generation and state still match the ones requested for abortion. By now, the timeout path knows that either the generation number and state changed if it lost the race or the completion will yield to it and can safely timeout the request. While it's more lines of code, it's conceptually simpler, doesn't depend on direct use of subtle memory ordering or coherence, and hopefully doesn't terminate the wrong instance. While this change makes REQ_ATOM_COMPLETE synchronization unnecessary between issue/complete and timeout paths, REQ_ATOM_COMPLETE isn't removed yet as it's still used in other places. Future patches will move all state tracking to the new mechanism and remove all bitops in the hot paths. Note that this patch adds a comment explaining a race condition in BLK_EH_RESET_TIMER path. The race has always been there and this patch doesn't change it. It's just documenting the existing race. v2: - Fixed BLK_EH_RESET_TIMER handling as pointed out by Jianchao. - s/request->gstate_seqc/request->gstate_seq/ as suggested by Peter. - READ_ONCE() added in blk_mq_rq_update_state() as suggested by Peter. v3: - Fixed possible extended seqcount / u64_stats_sync read looping spotted by Peter. - MQ_RQ_IDLE was incorrectly being set in complete_request instead of free_request. Fixed. v4: - Rebased on top of hctx_lock() refactoring patch. - Added comment explaining the use of hctx_lock() in completion path. v5: - Added comments requested by Bart. - Note the addition of BLK_EH_RESET_TIMER race condition in the commit message. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: "jianchao.wang" <jianchao.w.wang@oracle.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <Bart.VanAssche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-09 16:29:48 +00:00
static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
unsigned int hctx_idx, int node)
{
int ret;
if (set->ops->init_request) {
ret = set->ops->init_request(set, rq, hctx_idx, node);
if (ret)
return ret;
}
WRITE_ONCE(rq->state, MQ_RQ_IDLE);
blk-mq: replace timeout synchronization with a RCU and generation based scheme Currently, blk-mq timeout path synchronizes against the usual issue/completion path using a complex scheme involving atomic bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence rules. Unfortunately, it contains quite a few holes. There's a complex dancing around REQ_ATOM_STARTED and REQ_ATOM_COMPLETE between issue/completion and timeout paths; however, they don't have a synchronization point across request recycle instances and it isn't clear what the barriers add. blk_mq_check_expired() can easily read STARTED from N-2'th iteration, deadline from N-1'th, blk_mark_rq_complete() against Nth instance. In fact, it's pretty easy to make blk_mq_check_expired() terminate a later instance of a request. If we induce 5 sec delay before time_after_eq() test in blk_mq_check_expired(), shorten the timeout to 2s, and issue back-to-back large IOs, blk-mq starts timing out requests spuriously pretty quickly. Nothing actually timed out. It just made the call on a recycle instance of a request and then terminated a later instance long after the original instance finished. The scenario isn't theoretical either. This patch replaces the broken synchronization mechanism with a RCU and generation number based one. 1. Each request has a u64 generation + state value, which can be updated only by the request owner. Whenever a request becomes in-flight, the generation number gets bumped up too. This provides the basis for the timeout path to distinguish different recycle instances of the request. Also, marking a request in-flight and setting its deadline are protected with a seqcount so that the timeout path can fetch both values coherently. 2. The timeout path fetches the generation, state and deadline. If the verdict is timeout, it records the generation into a dedicated request abortion field and does RCU wait. 3. The completion path is also protected by RCU (from the previous patch) and checks whether the current generation number and state match the abortion field. If so, it skips completion. 4. The timeout path, after RCU wait, scans requests again and terminates the ones whose generation and state still match the ones requested for abortion. By now, the timeout path knows that either the generation number and state changed if it lost the race or the completion will yield to it and can safely timeout the request. While it's more lines of code, it's conceptually simpler, doesn't depend on direct use of subtle memory ordering or coherence, and hopefully doesn't terminate the wrong instance. While this change makes REQ_ATOM_COMPLETE synchronization unnecessary between issue/complete and timeout paths, REQ_ATOM_COMPLETE isn't removed yet as it's still used in other places. Future patches will move all state tracking to the new mechanism and remove all bitops in the hot paths. Note that this patch adds a comment explaining a race condition in BLK_EH_RESET_TIMER path. The race has always been there and this patch doesn't change it. It's just documenting the existing race. v2: - Fixed BLK_EH_RESET_TIMER handling as pointed out by Jianchao. - s/request->gstate_seqc/request->gstate_seq/ as suggested by Peter. - READ_ONCE() added in blk_mq_rq_update_state() as suggested by Peter. v3: - Fixed possible extended seqcount / u64_stats_sync read looping spotted by Peter. - MQ_RQ_IDLE was incorrectly being set in complete_request instead of free_request. Fixed. v4: - Rebased on top of hctx_lock() refactoring patch. - Added comment explaining the use of hctx_lock() in completion path. v5: - Added comments requested by Bart. - Note the addition of BLK_EH_RESET_TIMER race condition in the commit message. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: "jianchao.wang" <jianchao.w.wang@oracle.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <Bart.VanAssche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-09 16:29:48 +00:00
return 0;
}
static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
struct blk_mq_tags *tags,
unsigned int hctx_idx, unsigned int depth)
{
unsigned int i, j, entries_per_page, max_order = 4;
int node = blk_mq_get_hctx_node(set, hctx_idx);
size_t rq_size, left;
if (node == NUMA_NO_NODE)
node = set->numa_node;
INIT_LIST_HEAD(&tags->page_list);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
/*
* rq_size is the size of the request plus driver payload, rounded
* to the cacheline size
*/
rq_size = round_up(sizeof(struct request) + set->cmd_size,
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
cache_line_size());
left = rq_size * depth;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
for (i = 0; i < depth; ) {
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
int this_order = max_order;
struct page *page;
int to_do;
void *p;
while (this_order && left < order_to_size(this_order - 1))
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
this_order--;
do {
page = alloc_pages_node(node,
blk-mq: Avoid memory reclaim when remapping queues While stressing memory and IO at the same time we changed SMT settings, we were able to consistently trigger deadlocks in the mm system, which froze the entire machine. I think that under memory stress conditions, the large allocations performed by blk_mq_init_rq_map may trigger a reclaim, which stalls waiting on the block layer remmaping completion, thus deadlocking the system. The trace below was collected after the machine stalled, waiting for the hotplug event completion. The simplest fix for this is to make allocations in this path non-reclaimable, with GFP_NOIO. With this patch, We couldn't hit the issue anymore. This should apply on top of Jens's for-next branch cleanly. Changes since v1: - Use GFP_NOIO instead of GFP_NOWAIT. Call Trace: [c000000f0160aaf0] [c000000f0160ab50] 0xc000000f0160ab50 (unreliable) [c000000f0160acc0] [c000000000016624] __switch_to+0x2e4/0x430 [c000000f0160ad20] [c000000000b1a880] __schedule+0x310/0x9b0 [c000000f0160ae00] [c000000000b1af68] schedule+0x48/0xc0 [c000000f0160ae30] [c000000000b1b4b0] schedule_preempt_disabled+0x20/0x30 [c000000f0160ae50] [c000000000b1d4fc] __mutex_lock_slowpath+0xec/0x1f0 [c000000f0160aed0] [c000000000b1d678] mutex_lock+0x78/0xa0 [c000000f0160af00] [d000000019413cac] xfs_reclaim_inodes_ag+0x33c/0x380 [xfs] [c000000f0160b0b0] [d000000019415164] xfs_reclaim_inodes_nr+0x54/0x70 [xfs] [c000000f0160b0f0] [d0000000194297f8] xfs_fs_free_cached_objects+0x38/0x60 [xfs] [c000000f0160b120] [c0000000003172c8] super_cache_scan+0x1f8/0x210 [c000000f0160b190] [c00000000026301c] shrink_slab.part.13+0x21c/0x4c0 [c000000f0160b2d0] [c000000000268088] shrink_zone+0x2d8/0x3c0 [c000000f0160b380] [c00000000026834c] do_try_to_free_pages+0x1dc/0x520 [c000000f0160b450] [c00000000026876c] try_to_free_pages+0xdc/0x250 [c000000f0160b4e0] [c000000000251978] __alloc_pages_nodemask+0x868/0x10d0 [c000000f0160b6f0] [c000000000567030] blk_mq_init_rq_map+0x160/0x380 [c000000f0160b7a0] [c00000000056758c] blk_mq_map_swqueue+0x33c/0x360 [c000000f0160b820] [c000000000567904] blk_mq_queue_reinit+0x64/0xb0 [c000000f0160b850] [c00000000056a16c] blk_mq_queue_reinit_notify+0x19c/0x250 [c000000f0160b8a0] [c0000000000f5d38] notifier_call_chain+0x98/0x100 [c000000f0160b8f0] [c0000000000c5fb0] __cpu_notify+0x70/0xe0 [c000000f0160b930] [c0000000000c63c4] notify_prepare+0x44/0xb0 [c000000f0160b9b0] [c0000000000c52f4] cpuhp_invoke_callback+0x84/0x250 [c000000f0160ba10] [c0000000000c570c] cpuhp_up_callbacks+0x5c/0x120 [c000000f0160ba60] [c0000000000c7cb8] _cpu_up+0xf8/0x1d0 [c000000f0160bac0] [c0000000000c7eb0] do_cpu_up+0x120/0x150 [c000000f0160bb40] [c0000000006fe024] cpu_subsys_online+0x64/0xe0 [c000000f0160bb90] [c0000000006f5124] device_online+0xb4/0x120 [c000000f0160bbd0] [c0000000006f5244] online_store+0xb4/0xc0 [c000000f0160bc20] [c0000000006f0a68] dev_attr_store+0x68/0xa0 [c000000f0160bc60] [c0000000003ccc30] sysfs_kf_write+0x80/0xb0 [c000000f0160bca0] [c0000000003cbabc] kernfs_fop_write+0x17c/0x250 [c000000f0160bcf0] [c00000000030fe6c] __vfs_write+0x6c/0x1e0 [c000000f0160bd90] [c000000000311490] vfs_write+0xd0/0x270 [c000000f0160bde0] [c0000000003131fc] SyS_write+0x6c/0x110 [c000000f0160be30] [c000000000009204] system_call+0x38/0xec Signed-off-by: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com> Cc: Brian King <brking@linux.vnet.ibm.com> Cc: Douglas Miller <dougmill@linux.vnet.ibm.com> Cc: linux-block@vger.kernel.org Cc: linux-scsi@vger.kernel.org Signed-off-by: Jens Axboe <axboe@fb.com>
2016-12-06 15:31:44 +00:00
GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
this_order);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
if (page)
break;
if (!this_order--)
break;
if (order_to_size(this_order) < rq_size)
break;
} while (1);
if (!page)
goto fail;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
page->private = this_order;
list_add_tail(&page->lru, &tags->page_list);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
p = page_address(page);
/*
* Allow kmemleak to scan these pages as they contain pointers
* to additional allocations like via ops->init_request().
*/
blk-mq: Avoid memory reclaim when remapping queues While stressing memory and IO at the same time we changed SMT settings, we were able to consistently trigger deadlocks in the mm system, which froze the entire machine. I think that under memory stress conditions, the large allocations performed by blk_mq_init_rq_map may trigger a reclaim, which stalls waiting on the block layer remmaping completion, thus deadlocking the system. The trace below was collected after the machine stalled, waiting for the hotplug event completion. The simplest fix for this is to make allocations in this path non-reclaimable, with GFP_NOIO. With this patch, We couldn't hit the issue anymore. This should apply on top of Jens's for-next branch cleanly. Changes since v1: - Use GFP_NOIO instead of GFP_NOWAIT. Call Trace: [c000000f0160aaf0] [c000000f0160ab50] 0xc000000f0160ab50 (unreliable) [c000000f0160acc0] [c000000000016624] __switch_to+0x2e4/0x430 [c000000f0160ad20] [c000000000b1a880] __schedule+0x310/0x9b0 [c000000f0160ae00] [c000000000b1af68] schedule+0x48/0xc0 [c000000f0160ae30] [c000000000b1b4b0] schedule_preempt_disabled+0x20/0x30 [c000000f0160ae50] [c000000000b1d4fc] __mutex_lock_slowpath+0xec/0x1f0 [c000000f0160aed0] [c000000000b1d678] mutex_lock+0x78/0xa0 [c000000f0160af00] [d000000019413cac] xfs_reclaim_inodes_ag+0x33c/0x380 [xfs] [c000000f0160b0b0] [d000000019415164] xfs_reclaim_inodes_nr+0x54/0x70 [xfs] [c000000f0160b0f0] [d0000000194297f8] xfs_fs_free_cached_objects+0x38/0x60 [xfs] [c000000f0160b120] [c0000000003172c8] super_cache_scan+0x1f8/0x210 [c000000f0160b190] [c00000000026301c] shrink_slab.part.13+0x21c/0x4c0 [c000000f0160b2d0] [c000000000268088] shrink_zone+0x2d8/0x3c0 [c000000f0160b380] [c00000000026834c] do_try_to_free_pages+0x1dc/0x520 [c000000f0160b450] [c00000000026876c] try_to_free_pages+0xdc/0x250 [c000000f0160b4e0] [c000000000251978] __alloc_pages_nodemask+0x868/0x10d0 [c000000f0160b6f0] [c000000000567030] blk_mq_init_rq_map+0x160/0x380 [c000000f0160b7a0] [c00000000056758c] blk_mq_map_swqueue+0x33c/0x360 [c000000f0160b820] [c000000000567904] blk_mq_queue_reinit+0x64/0xb0 [c000000f0160b850] [c00000000056a16c] blk_mq_queue_reinit_notify+0x19c/0x250 [c000000f0160b8a0] [c0000000000f5d38] notifier_call_chain+0x98/0x100 [c000000f0160b8f0] [c0000000000c5fb0] __cpu_notify+0x70/0xe0 [c000000f0160b930] [c0000000000c63c4] notify_prepare+0x44/0xb0 [c000000f0160b9b0] [c0000000000c52f4] cpuhp_invoke_callback+0x84/0x250 [c000000f0160ba10] [c0000000000c570c] cpuhp_up_callbacks+0x5c/0x120 [c000000f0160ba60] [c0000000000c7cb8] _cpu_up+0xf8/0x1d0 [c000000f0160bac0] [c0000000000c7eb0] do_cpu_up+0x120/0x150 [c000000f0160bb40] [c0000000006fe024] cpu_subsys_online+0x64/0xe0 [c000000f0160bb90] [c0000000006f5124] device_online+0xb4/0x120 [c000000f0160bbd0] [c0000000006f5244] online_store+0xb4/0xc0 [c000000f0160bc20] [c0000000006f0a68] dev_attr_store+0x68/0xa0 [c000000f0160bc60] [c0000000003ccc30] sysfs_kf_write+0x80/0xb0 [c000000f0160bca0] [c0000000003cbabc] kernfs_fop_write+0x17c/0x250 [c000000f0160bcf0] [c00000000030fe6c] __vfs_write+0x6c/0x1e0 [c000000f0160bd90] [c000000000311490] vfs_write+0xd0/0x270 [c000000f0160bde0] [c0000000003131fc] SyS_write+0x6c/0x110 [c000000f0160be30] [c000000000009204] system_call+0x38/0xec Signed-off-by: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com> Cc: Brian King <brking@linux.vnet.ibm.com> Cc: Douglas Miller <dougmill@linux.vnet.ibm.com> Cc: linux-block@vger.kernel.org Cc: linux-scsi@vger.kernel.org Signed-off-by: Jens Axboe <axboe@fb.com>
2016-12-06 15:31:44 +00:00
kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
entries_per_page = order_to_size(this_order) / rq_size;
to_do = min(entries_per_page, depth - i);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
left -= to_do * rq_size;
for (j = 0; j < to_do; j++) {
struct request *rq = p;
tags->static_rqs[i] = rq;
blk-mq: replace timeout synchronization with a RCU and generation based scheme Currently, blk-mq timeout path synchronizes against the usual issue/completion path using a complex scheme involving atomic bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence rules. Unfortunately, it contains quite a few holes. There's a complex dancing around REQ_ATOM_STARTED and REQ_ATOM_COMPLETE between issue/completion and timeout paths; however, they don't have a synchronization point across request recycle instances and it isn't clear what the barriers add. blk_mq_check_expired() can easily read STARTED from N-2'th iteration, deadline from N-1'th, blk_mark_rq_complete() against Nth instance. In fact, it's pretty easy to make blk_mq_check_expired() terminate a later instance of a request. If we induce 5 sec delay before time_after_eq() test in blk_mq_check_expired(), shorten the timeout to 2s, and issue back-to-back large IOs, blk-mq starts timing out requests spuriously pretty quickly. Nothing actually timed out. It just made the call on a recycle instance of a request and then terminated a later instance long after the original instance finished. The scenario isn't theoretical either. This patch replaces the broken synchronization mechanism with a RCU and generation number based one. 1. Each request has a u64 generation + state value, which can be updated only by the request owner. Whenever a request becomes in-flight, the generation number gets bumped up too. This provides the basis for the timeout path to distinguish different recycle instances of the request. Also, marking a request in-flight and setting its deadline are protected with a seqcount so that the timeout path can fetch both values coherently. 2. The timeout path fetches the generation, state and deadline. If the verdict is timeout, it records the generation into a dedicated request abortion field and does RCU wait. 3. The completion path is also protected by RCU (from the previous patch) and checks whether the current generation number and state match the abortion field. If so, it skips completion. 4. The timeout path, after RCU wait, scans requests again and terminates the ones whose generation and state still match the ones requested for abortion. By now, the timeout path knows that either the generation number and state changed if it lost the race or the completion will yield to it and can safely timeout the request. While it's more lines of code, it's conceptually simpler, doesn't depend on direct use of subtle memory ordering or coherence, and hopefully doesn't terminate the wrong instance. While this change makes REQ_ATOM_COMPLETE synchronization unnecessary between issue/complete and timeout paths, REQ_ATOM_COMPLETE isn't removed yet as it's still used in other places. Future patches will move all state tracking to the new mechanism and remove all bitops in the hot paths. Note that this patch adds a comment explaining a race condition in BLK_EH_RESET_TIMER path. The race has always been there and this patch doesn't change it. It's just documenting the existing race. v2: - Fixed BLK_EH_RESET_TIMER handling as pointed out by Jianchao. - s/request->gstate_seqc/request->gstate_seq/ as suggested by Peter. - READ_ONCE() added in blk_mq_rq_update_state() as suggested by Peter. v3: - Fixed possible extended seqcount / u64_stats_sync read looping spotted by Peter. - MQ_RQ_IDLE was incorrectly being set in complete_request instead of free_request. Fixed. v4: - Rebased on top of hctx_lock() refactoring patch. - Added comment explaining the use of hctx_lock() in completion path. v5: - Added comments requested by Bart. - Note the addition of BLK_EH_RESET_TIMER race condition in the commit message. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: "jianchao.wang" <jianchao.w.wang@oracle.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Bart Van Assche <Bart.VanAssche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-09 16:29:48 +00:00
if (blk_mq_init_request(set, rq, hctx_idx, node)) {
tags->static_rqs[i] = NULL;
goto fail;
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
p += rq_size;
i++;
}
}
return 0;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
fail:
blk_mq_free_rqs(set, tags, hctx_idx);
return -ENOMEM;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
struct rq_iter_data {
struct blk_mq_hw_ctx *hctx;
bool has_rq;
};
static bool blk_mq_has_request(struct request *rq, void *data)
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
{
struct rq_iter_data *iter_data = data;
if (rq->mq_hctx != iter_data->hctx)
return true;
iter_data->has_rq = true;
return false;
}
static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
{
struct blk_mq_tags *tags = hctx->sched_tags ?
hctx->sched_tags : hctx->tags;
struct rq_iter_data data = {
.hctx = hctx,
};
blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
return data.has_rq;
}
static bool blk_mq_hctx_has_online_cpu(struct blk_mq_hw_ctx *hctx,
unsigned int this_cpu)
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
{
enum hctx_type type = hctx->type;
int cpu;
/*
* hctx->cpumask has to rule out isolated CPUs, but userspace still
* might submit IOs on these isolated CPUs, so use the queue map to
* check if all CPUs mapped to this hctx are offline
*/
for_each_online_cpu(cpu) {
struct blk_mq_hw_ctx *h = blk_mq_map_queue_type(hctx->queue,
type, cpu);
if (h != hctx)
continue;
/* this hctx has at least one online CPU */
if (this_cpu != cpu)
return true;
}
return false;
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
}
static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
{
struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
struct blk_mq_hw_ctx, cpuhp_online);
if (blk_mq_hctx_has_online_cpu(hctx, cpu))
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
return 0;
/*
* Prevent new request from being allocated on the current hctx.
*
* The smp_mb__after_atomic() Pairs with the implied barrier in
* test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
* seen once we return from the tag allocator.
*/
set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
smp_mb__after_atomic();
/*
* Try to grab a reference to the queue and wait for any outstanding
* requests. If we could not grab a reference the queue has been
* frozen and there are no requests.
*/
if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
while (blk_mq_hctx_has_requests(hctx))
msleep(5);
percpu_ref_put(&hctx->queue->q_usage_counter);
}
return 0;
}
/*
* Check if one CPU is mapped to the specified hctx
*
* Isolated CPUs have been ruled out from hctx->cpumask, which is supposed
* to be used for scheduling kworker only. For other usage, please call this
* helper for checking if one CPU belongs to the specified hctx
*/
static bool blk_mq_cpu_mapped_to_hctx(unsigned int cpu,
const struct blk_mq_hw_ctx *hctx)
{
struct blk_mq_hw_ctx *mapped_hctx = blk_mq_map_queue_type(hctx->queue,
hctx->type, cpu);
return mapped_hctx == hctx;
}
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
{
struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
struct blk_mq_hw_ctx, cpuhp_online);
if (blk_mq_cpu_mapped_to_hctx(cpu, hctx))
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
return 0;
}
/*
* 'cpu' is going away. splice any existing rq_list entries from this
* software queue to the hw queue dispatch list, and ensure that it
* gets run.
*/
static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
{
struct blk_mq_hw_ctx *hctx;
struct blk_mq_ctx *ctx;
LIST_HEAD(tmp);
enum hctx_type type;
hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
if (!blk_mq_cpu_mapped_to_hctx(cpu, hctx))
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
return 0;
ctx = __blk_mq_get_ctx(hctx->queue, cpu);
type = hctx->type;
spin_lock(&ctx->lock);
if (!list_empty(&ctx->rq_lists[type])) {
list_splice_init(&ctx->rq_lists[type], &tmp);
blk_mq_hctx_clear_pending(hctx, ctx);
}
spin_unlock(&ctx->lock);
if (list_empty(&tmp))
return 0;
spin_lock(&hctx->lock);
list_splice_tail_init(&tmp, &hctx->dispatch);
spin_unlock(&hctx->lock);
blk_mq_run_hw_queue(hctx, true);
return 0;
}
static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
{
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
if (!(hctx->flags & BLK_MQ_F_STACKING))
cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
&hctx->cpuhp_online);
cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
&hctx->cpuhp_dead);
}
/*
* Before freeing hw queue, clearing the flush request reference in
* tags->rqs[] for avoiding potential UAF.
*/
static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
unsigned int queue_depth, struct request *flush_rq)
{
int i;
unsigned long flags;
/* The hw queue may not be mapped yet */
if (!tags)
return;
WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
for (i = 0; i < queue_depth; i++)
cmpxchg(&tags->rqs[i], flush_rq, NULL);
/*
* Wait until all pending iteration is done.
*
* Request reference is cleared and it is guaranteed to be observed
* after the ->lock is released.
*/
spin_lock_irqsave(&tags->lock, flags);
spin_unlock_irqrestore(&tags->lock, flags);
}
/* hctx->ctxs will be freed in queue's release handler */
static void blk_mq_exit_hctx(struct request_queue *q,
struct blk_mq_tag_set *set,
struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
{
struct request *flush_rq = hctx->fq->flush_rq;
blk-mq: fix kernel oops in blk_mq_tag_idle() HW queues may be unmapped in some cases, such as blk_mq_update_nr_hw_queues(), then we need to check it before calling blk_mq_tag_idle(), otherwise the following kernel oops can be triggered, so fix it by checking if the hw queue is unmapped since it doesn't make sense to idle the tags any more after hw queues are unmapped. [ 440.771298] Workqueue: nvme-wq nvme_rdma_del_ctrl_work [nvme_rdma] [ 440.779104] task: ffff894bae755ee0 ti: ffff893bf9bc8000 task.ti: ffff893bf9bc8000 [ 440.788359] RIP: 0010:[<ffffffffb730e2b4>] [<ffffffffb730e2b4>] __blk_mq_tag_idle+0x24/0x40 [ 440.798697] RSP: 0018:ffff893bf9bcbd10 EFLAGS: 00010286 [ 440.805538] RAX: 0000000000000000 RBX: ffff895bb131dc00 RCX: 000000000000011f [ 440.814426] RDX: 00000000ffffffff RSI: 0000000000000120 RDI: ffff895bb131dc00 [ 440.823301] RBP: ffff893bf9bcbd10 R08: 000000000001b860 R09: 4a51d361c00c0000 [ 440.832193] R10: b5907f32b4cc7003 R11: ffffd6cabfb57000 R12: ffff894bafd1e008 [ 440.841091] R13: 0000000000000001 R14: ffff895baf770000 R15: 0000000000000080 [ 440.849988] FS: 0000000000000000(0000) GS:ffff894bbdcc0000(0000) knlGS:0000000000000000 [ 440.859955] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 440.867274] CR2: 0000000000000008 CR3: 000000103d098000 CR4: 00000000001407e0 [ 440.876169] Call Trace: [ 440.879818] [<ffffffffb7309d68>] blk_mq_exit_hctx+0xd8/0xe0 [ 440.887051] [<ffffffffb730dc40>] blk_mq_free_queue+0xf0/0x160 [ 440.894465] [<ffffffffb72ff679>] blk_cleanup_queue+0xd9/0x150 [ 440.901881] [<ffffffffc08a802b>] nvme_ns_remove+0x5b/0xb0 [nvme_core] [ 440.910068] [<ffffffffc08a811b>] nvme_remove_namespaces+0x3b/0x60 [nvme_core] [ 440.919026] [<ffffffffc08b817b>] __nvme_rdma_remove_ctrl+0x2b/0xb0 [nvme_rdma] [ 440.928079] [<ffffffffc08b8237>] nvme_rdma_del_ctrl_work+0x17/0x20 [nvme_rdma] [ 440.937126] [<ffffffffb70ab58a>] process_one_work+0x17a/0x440 [ 440.944517] [<ffffffffb70ac3a8>] worker_thread+0x278/0x3c0 [ 440.951607] [<ffffffffb70ac130>] ? manage_workers.isra.24+0x2a0/0x2a0 [ 440.959760] [<ffffffffb70b352f>] kthread+0xcf/0xe0 [ 440.966055] [<ffffffffb70b3460>] ? insert_kthread_work+0x40/0x40 [ 440.973715] [<ffffffffb76d8658>] ret_from_fork+0x58/0x90 [ 440.980586] [<ffffffffb70b3460>] ? insert_kthread_work+0x40/0x40 [ 440.988229] Code: 5b 41 5c 5d c3 66 90 0f 1f 44 00 00 48 8b 87 20 01 00 00 f0 0f ba 77 40 01 19 d2 85 d2 75 08 c3 0f 1f 80 00 00 00 00 55 48 89 e5 <f0> ff 48 08 48 8d 78 10 e8 7f 0f 05 00 5d c3 0f 1f 00 66 2e 0f [ 441.011620] RIP [<ffffffffb730e2b4>] __blk_mq_tag_idle+0x24/0x40 [ 441.019301] RSP <ffff893bf9bcbd10> [ 441.024052] CR2: 0000000000000008 Reported-by: Zhang Yi <yizhan@redhat.com> Tested-by: Zhang Yi <yizhan@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-09 13:28:29 +00:00
if (blk_mq_hw_queue_mapped(hctx))
blk_mq_tag_idle(hctx);
if (blk_queue_init_done(q))
blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
set->queue_depth, flush_rq);
if (set->ops->exit_request)
set->ops->exit_request(set, flush_rq, hctx_idx);
if (set->ops->exit_hctx)
set->ops->exit_hctx(hctx, hctx_idx);
blk_mq_remove_cpuhp(hctx);
xa_erase(&q->hctx_table, hctx_idx);
spin_lock(&q->unused_hctx_lock);
list_add(&hctx->hctx_list, &q->unused_hctx_list);
spin_unlock(&q->unused_hctx_lock);
}
static void blk_mq_exit_hw_queues(struct request_queue *q,
struct blk_mq_tag_set *set, int nr_queue)
{
struct blk_mq_hw_ctx *hctx;
unsigned long i;
queue_for_each_hw_ctx(q, hctx, i) {
if (i == nr_queue)
break;
blk_mq_exit_hctx(q, set, hctx, i);
}
}
static int blk_mq_init_hctx(struct request_queue *q,
struct blk_mq_tag_set *set,
struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
hctx->queue_num = hctx_idx;
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
if (!(hctx->flags & BLK_MQ_F_STACKING))
cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
&hctx->cpuhp_online);
cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
hctx->tags = set->tags[hctx_idx];
if (set->ops->init_hctx &&
set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
goto unregister_cpu_notifier;
if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
hctx->numa_node))
goto exit_hctx;
if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
goto exit_flush_rq;
return 0;
exit_flush_rq:
if (set->ops->exit_request)
set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
exit_hctx:
if (set->ops->exit_hctx)
set->ops->exit_hctx(hctx, hctx_idx);
unregister_cpu_notifier:
blk_mq_remove_cpuhp(hctx);
return -1;
}
static struct blk_mq_hw_ctx *
blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
int node)
{
struct blk_mq_hw_ctx *hctx;
gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
if (!hctx)
goto fail_alloc_hctx;
if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
goto free_hctx;
atomic_set(&hctx->nr_active, 0);
if (node == NUMA_NO_NODE)
node = set->numa_node;
hctx->numa_node = node;
INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
spin_lock_init(&hctx->lock);
INIT_LIST_HEAD(&hctx->dispatch);
hctx->queue = q;
hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
INIT_LIST_HEAD(&hctx->hctx_list);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
/*
* Allocate space for all possible cpus to avoid allocation at
* runtime
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
*/
hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
gfp, node);
if (!hctx->ctxs)
goto free_cpumask;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
gfp, node, false, false))
goto free_ctxs;
hctx->nr_ctx = 0;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
spin_lock_init(&hctx->dispatch_wait_lock);
init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
if (!hctx->fq)
goto free_bitmap;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
blk_mq_hctx_kobj_init(hctx);
return hctx;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
free_bitmap:
sbitmap_free(&hctx->ctx_map);
free_ctxs:
kfree(hctx->ctxs);
free_cpumask:
free_cpumask_var(hctx->cpumask);
free_hctx:
kfree(hctx);
fail_alloc_hctx:
return NULL;
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
static void blk_mq_init_cpu_queues(struct request_queue *q,
unsigned int nr_hw_queues)
{
struct blk_mq_tag_set *set = q->tag_set;
unsigned int i, j;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
for_each_possible_cpu(i) {
struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
struct blk_mq_hw_ctx *hctx;
int k;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
__ctx->cpu = i;
spin_lock_init(&__ctx->lock);
for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
INIT_LIST_HEAD(&__ctx->rq_lists[k]);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
__ctx->queue = q;
/*
* Set local node, IFF we have more than one hw queue. If
* not, we remain on the home node of the device
*/
for (j = 0; j < set->nr_maps; j++) {
hctx = blk_mq_map_queue_type(q, j, i);
if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
hctx->numa_node = cpu_to_node(i);
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
}
struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
unsigned int hctx_idx,
unsigned int depth)
{
struct blk_mq_tags *tags;
int ret;
tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
if (!tags)
return NULL;
ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
if (ret) {
blk_mq_free_rq_map(tags);
return NULL;
}
return tags;
}
static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
int hctx_idx)
{
if (blk_mq_is_shared_tags(set->flags)) {
set->tags[hctx_idx] = set->shared_tags;
return true;
}
set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
set->queue_depth);
return set->tags[hctx_idx];
}
void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
struct blk_mq_tags *tags,
unsigned int hctx_idx)
{
if (tags) {
blk_mq_free_rqs(set, tags, hctx_idx);
blk_mq_free_rq_map(tags);
}
}
static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
unsigned int hctx_idx)
{
if (!blk_mq_is_shared_tags(set->flags))
blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
set->tags[hctx_idx] = NULL;
}
static void blk_mq_map_swqueue(struct request_queue *q)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
unsigned int j, hctx_idx;
unsigned long i;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
struct blk_mq_hw_ctx *hctx;
struct blk_mq_ctx *ctx;
struct blk_mq_tag_set *set = q->tag_set;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
queue_for_each_hw_ctx(q, hctx, i) {
cpumask_clear(hctx->cpumask);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
hctx->nr_ctx = 0;
hctx->dispatch_from = NULL;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
/*
* Map software to hardware queues.
*
* If the cpu isn't present, the cpu is mapped to first hctx.
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
*/
for_each_possible_cpu(i) {
ctx = per_cpu_ptr(q->queue_ctx, i);
for (j = 0; j < set->nr_maps; j++) {
if (!set->map[j].nr_queues) {
ctx->hctxs[j] = blk_mq_map_queue_type(q,
HCTX_TYPE_DEFAULT, i);
continue;
}
block: alloc map and request for new hardware queue Alloc new map and request for new hardware queue when increse hardware queue count. Before this patch, it will show a warning for each new hardware queue, but it's not enough, these hctx have no maps and reqeust, when a bio was mapped to these hardware queue, it will trigger kernel panic when get request from these hctx. Test environment: * A NVMe disk supports 128 io queues * 96 cpus in system A corner case can always trigger this panic, there are 96 io queues allocated for HCTX_TYPE_DEFAULT type, the corresponding kernel log: nvme nvme0: 96/0/0 default/read/poll queues. Now we set nvme write queues to 96, then nvme will alloc others(32) queues for read, but blk_mq_update_nr_hw_queues does not alloc map and request for these new added io queues. So when process read nvme disk, it will trigger kernel panic when get request from these hardware context. Reproduce script: nr=$(expr `cat /sys/block/nvme0n1/device/queue_count` - 1) echo $nr > /sys/module/nvme/parameters/write_queues echo 1 > /sys/block/nvme0n1/device/reset_controller dd if=/dev/nvme0n1 of=/dev/null bs=4K count=1 [ 8040.805626] ------------[ cut here ]------------ [ 8040.805627] WARNING: CPU: 82 PID: 12921 at block/blk-mq.c:2578 blk_mq_map_swqueue+0x2b6/0x2c0 [ 8040.805627] Modules linked in: nvme nvme_core nf_conntrack_netlink xt_addrtype br_netfilter overlay xt_CHECKSUM xt_MASQUERADE xt_conntrack ipt_REJECT nft_counter nf_nat_tftp nf_conntrack_tftp nft_masq nf_tables_set nft_fib_inet nft_f ib_ipv4 nft_fib_ipv6 nft_fib nft_reject_inet nf_reject_ipv4 nf_reject_ipv6 nft_reject nft_ct nft_chain_nat nf_nat nf_conntrack tun bridge nf_defrag_ipv6 nf_defrag_ipv4 stp llc ip6_tables ip_tables nft_compat rfkill ip_set nf_tables nfne tlink sunrpc intel_rapl_msr intel_rapl_common skx_edac nfit libnvdimm x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass ipmi_ssif crct10dif_pclmul crc32_pclmul iTCO_wdt iTCO_vendor_support ghash_clmulni_intel intel_ cstate intel_uncore raid0 joydev intel_rapl_perf ipmi_si pcspkr mei_me ioatdma sg ipmi_devintf mei i2c_i801 dca lpc_ich ipmi_msghandler acpi_power_meter acpi_pad xfs libcrc32c sd_mod ast i2c_algo_bit drm_vram_helper drm_ttm_helper ttm d rm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops [ 8040.805637] ahci drm i40e libahci crc32c_intel libata t10_pi wmi dm_mirror dm_region_hash dm_log dm_mod [last unloaded: nvme_core] [ 8040.805640] CPU: 82 PID: 12921 Comm: kworker/u194:2 Kdump: loaded Tainted: G W 5.6.0-rc5.78317c+ #2 [ 8040.805640] Hardware name: Inspur SA5212M5/YZMB-00882-104, BIOS 4.0.9 08/27/2019 [ 8040.805641] Workqueue: nvme-reset-wq nvme_reset_work [nvme] [ 8040.805642] RIP: 0010:blk_mq_map_swqueue+0x2b6/0x2c0 [ 8040.805643] Code: 00 00 00 00 00 41 83 c5 01 44 39 6d 50 77 b8 5b 5d 41 5c 41 5d 41 5e 41 5f c3 48 8b bb 98 00 00 00 89 d6 e8 8c 81 03 00 eb 83 <0f> 0b e9 52 ff ff ff 0f 1f 00 0f 1f 44 00 00 41 57 48 89 f1 41 56 [ 8040.805643] RSP: 0018:ffffba590d2e7d48 EFLAGS: 00010246 [ 8040.805643] RAX: 0000000000000000 RBX: ffff9f013e1ba800 RCX: 000000000000003d [ 8040.805644] RDX: ffff9f00ffff6000 RSI: 0000000000000003 RDI: ffff9ed200246d90 [ 8040.805644] RBP: ffff9f00f6a79860 R08: 0000000000000000 R09: 000000000000003d [ 8040.805645] R10: 0000000000000001 R11: ffff9f0138c3d000 R12: ffff9f00fb3a9008 [ 8040.805645] R13: 000000000000007f R14: ffffffff96822660 R15: 000000000000005f [ 8040.805645] FS: 0000000000000000(0000) GS:ffff9f013fa80000(0000) knlGS:0000000000000000 [ 8040.805646] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 8040.805646] CR2: 00007f7f397fa6f8 CR3: 0000003d8240a002 CR4: 00000000007606e0 [ 8040.805647] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 8040.805647] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 8040.805647] PKRU: 55555554 [ 8040.805647] Call Trace: [ 8040.805649] blk_mq_update_nr_hw_queues+0x31b/0x390 [ 8040.805650] nvme_reset_work+0xb4b/0xeab [nvme] [ 8040.805651] process_one_work+0x1a7/0x370 [ 8040.805652] worker_thread+0x1c9/0x380 [ 8040.805653] ? max_active_store+0x80/0x80 [ 8040.805655] kthread+0x112/0x130 [ 8040.805656] ? __kthread_parkme+0x70/0x70 [ 8040.805657] ret_from_fork+0x35/0x40 [ 8040.805658] ---[ end trace b5f13b1e73ccb5d3 ]--- [ 8229.365135] BUG: kernel NULL pointer dereference, address: 0000000000000004 [ 8229.365165] #PF: supervisor read access in kernel mode [ 8229.365178] #PF: error_code(0x0000) - not-present page [ 8229.365191] PGD 0 P4D 0 [ 8229.365201] Oops: 0000 [#1] SMP PTI [ 8229.365212] CPU: 77 PID: 13024 Comm: dd Kdump: loaded Tainted: G W 5.6.0-rc5.78317c+ #2 [ 8229.365232] Hardware name: Inspur SA5212M5/YZMB-00882-104, BIOS 4.0.9 08/27/2019 [ 8229.365253] RIP: 0010:blk_mq_get_tag+0x227/0x250 [ 8229.365265] Code: 44 24 04 44 01 e0 48 8b 74 24 38 65 48 33 34 25 28 00 00 00 75 33 48 83 c4 40 5b 5d 41 5c 41 5d 41 5e c3 48 8d 68 10 4c 89 ef <44> 8b 60 04 48 89 ee e8 dd f9 ff ff 83 f8 ff 75 c8 e9 67 fe ff ff [ 8229.365304] RSP: 0018:ffffba590e977970 EFLAGS: 00010246 [ 8229.365317] RAX: 0000000000000000 RBX: ffff9f00f6a79860 RCX: ffffba590e977998 [ 8229.365333] RDX: 0000000000000000 RSI: ffff9f012039b140 RDI: ffffba590e977a38 [ 8229.365349] RBP: 0000000000000010 R08: ffffda58ff94e190 R09: ffffda58ff94e198 [ 8229.365365] R10: 0000000000000011 R11: ffff9f00f6a79860 R12: 0000000000000000 [ 8229.365381] R13: ffffba590e977a38 R14: ffff9f012039b140 R15: 0000000000000001 [ 8229.365397] FS: 00007f481c230580(0000) GS:ffff9f013f940000(0000) knlGS:0000000000000000 [ 8229.365415] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 8229.365428] CR2: 0000000000000004 CR3: 0000005f35e26004 CR4: 00000000007606e0 [ 8229.365444] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 8229.365460] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 8229.365476] PKRU: 55555554 [ 8229.365484] Call Trace: [ 8229.365498] ? finish_wait+0x80/0x80 [ 8229.365512] blk_mq_get_request+0xcb/0x3f0 [ 8229.365525] blk_mq_make_request+0x143/0x5d0 [ 8229.365538] generic_make_request+0xcf/0x310 [ 8229.365553] ? scan_shadow_nodes+0x30/0x30 [ 8229.365564] submit_bio+0x3c/0x150 [ 8229.365576] mpage_readpages+0x163/0x1a0 [ 8229.365588] ? blkdev_direct_IO+0x490/0x490 [ 8229.365601] read_pages+0x6b/0x190 [ 8229.365612] __do_page_cache_readahead+0x1c1/0x1e0 [ 8229.365626] ondemand_readahead+0x182/0x2f0 [ 8229.365639] generic_file_buffered_read+0x590/0xab0 [ 8229.365655] new_sync_read+0x12a/0x1c0 [ 8229.365666] vfs_read+0x8a/0x140 [ 8229.365676] ksys_read+0x59/0xd0 [ 8229.365688] do_syscall_64+0x55/0x1d0 [ 8229.365700] entry_SYSCALL_64_after_hwframe+0x44/0xa9 Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Weiping Zhang <zhangweiping@didiglobal.com> Tested-by: Weiping Zhang <zhangweiping@didiglobal.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-07 13:04:08 +00:00
hctx_idx = set->map[j].mq_map[i];
/* unmapped hw queue can be remapped after CPU topo changed */
if (!set->tags[hctx_idx] &&
!__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
block: alloc map and request for new hardware queue Alloc new map and request for new hardware queue when increse hardware queue count. Before this patch, it will show a warning for each new hardware queue, but it's not enough, these hctx have no maps and reqeust, when a bio was mapped to these hardware queue, it will trigger kernel panic when get request from these hctx. Test environment: * A NVMe disk supports 128 io queues * 96 cpus in system A corner case can always trigger this panic, there are 96 io queues allocated for HCTX_TYPE_DEFAULT type, the corresponding kernel log: nvme nvme0: 96/0/0 default/read/poll queues. Now we set nvme write queues to 96, then nvme will alloc others(32) queues for read, but blk_mq_update_nr_hw_queues does not alloc map and request for these new added io queues. So when process read nvme disk, it will trigger kernel panic when get request from these hardware context. Reproduce script: nr=$(expr `cat /sys/block/nvme0n1/device/queue_count` - 1) echo $nr > /sys/module/nvme/parameters/write_queues echo 1 > /sys/block/nvme0n1/device/reset_controller dd if=/dev/nvme0n1 of=/dev/null bs=4K count=1 [ 8040.805626] ------------[ cut here ]------------ [ 8040.805627] WARNING: CPU: 82 PID: 12921 at block/blk-mq.c:2578 blk_mq_map_swqueue+0x2b6/0x2c0 [ 8040.805627] Modules linked in: nvme nvme_core nf_conntrack_netlink xt_addrtype br_netfilter overlay xt_CHECKSUM xt_MASQUERADE xt_conntrack ipt_REJECT nft_counter nf_nat_tftp nf_conntrack_tftp nft_masq nf_tables_set nft_fib_inet nft_f ib_ipv4 nft_fib_ipv6 nft_fib nft_reject_inet nf_reject_ipv4 nf_reject_ipv6 nft_reject nft_ct nft_chain_nat nf_nat nf_conntrack tun bridge nf_defrag_ipv6 nf_defrag_ipv4 stp llc ip6_tables ip_tables nft_compat rfkill ip_set nf_tables nfne tlink sunrpc intel_rapl_msr intel_rapl_common skx_edac nfit libnvdimm x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass ipmi_ssif crct10dif_pclmul crc32_pclmul iTCO_wdt iTCO_vendor_support ghash_clmulni_intel intel_ cstate intel_uncore raid0 joydev intel_rapl_perf ipmi_si pcspkr mei_me ioatdma sg ipmi_devintf mei i2c_i801 dca lpc_ich ipmi_msghandler acpi_power_meter acpi_pad xfs libcrc32c sd_mod ast i2c_algo_bit drm_vram_helper drm_ttm_helper ttm d rm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops [ 8040.805637] ahci drm i40e libahci crc32c_intel libata t10_pi wmi dm_mirror dm_region_hash dm_log dm_mod [last unloaded: nvme_core] [ 8040.805640] CPU: 82 PID: 12921 Comm: kworker/u194:2 Kdump: loaded Tainted: G W 5.6.0-rc5.78317c+ #2 [ 8040.805640] Hardware name: Inspur SA5212M5/YZMB-00882-104, BIOS 4.0.9 08/27/2019 [ 8040.805641] Workqueue: nvme-reset-wq nvme_reset_work [nvme] [ 8040.805642] RIP: 0010:blk_mq_map_swqueue+0x2b6/0x2c0 [ 8040.805643] Code: 00 00 00 00 00 41 83 c5 01 44 39 6d 50 77 b8 5b 5d 41 5c 41 5d 41 5e 41 5f c3 48 8b bb 98 00 00 00 89 d6 e8 8c 81 03 00 eb 83 <0f> 0b e9 52 ff ff ff 0f 1f 00 0f 1f 44 00 00 41 57 48 89 f1 41 56 [ 8040.805643] RSP: 0018:ffffba590d2e7d48 EFLAGS: 00010246 [ 8040.805643] RAX: 0000000000000000 RBX: ffff9f013e1ba800 RCX: 000000000000003d [ 8040.805644] RDX: ffff9f00ffff6000 RSI: 0000000000000003 RDI: ffff9ed200246d90 [ 8040.805644] RBP: ffff9f00f6a79860 R08: 0000000000000000 R09: 000000000000003d [ 8040.805645] R10: 0000000000000001 R11: ffff9f0138c3d000 R12: ffff9f00fb3a9008 [ 8040.805645] R13: 000000000000007f R14: ffffffff96822660 R15: 000000000000005f [ 8040.805645] FS: 0000000000000000(0000) GS:ffff9f013fa80000(0000) knlGS:0000000000000000 [ 8040.805646] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 8040.805646] CR2: 00007f7f397fa6f8 CR3: 0000003d8240a002 CR4: 00000000007606e0 [ 8040.805647] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 8040.805647] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 8040.805647] PKRU: 55555554 [ 8040.805647] Call Trace: [ 8040.805649] blk_mq_update_nr_hw_queues+0x31b/0x390 [ 8040.805650] nvme_reset_work+0xb4b/0xeab [nvme] [ 8040.805651] process_one_work+0x1a7/0x370 [ 8040.805652] worker_thread+0x1c9/0x380 [ 8040.805653] ? max_active_store+0x80/0x80 [ 8040.805655] kthread+0x112/0x130 [ 8040.805656] ? __kthread_parkme+0x70/0x70 [ 8040.805657] ret_from_fork+0x35/0x40 [ 8040.805658] ---[ end trace b5f13b1e73ccb5d3 ]--- [ 8229.365135] BUG: kernel NULL pointer dereference, address: 0000000000000004 [ 8229.365165] #PF: supervisor read access in kernel mode [ 8229.365178] #PF: error_code(0x0000) - not-present page [ 8229.365191] PGD 0 P4D 0 [ 8229.365201] Oops: 0000 [#1] SMP PTI [ 8229.365212] CPU: 77 PID: 13024 Comm: dd Kdump: loaded Tainted: G W 5.6.0-rc5.78317c+ #2 [ 8229.365232] Hardware name: Inspur SA5212M5/YZMB-00882-104, BIOS 4.0.9 08/27/2019 [ 8229.365253] RIP: 0010:blk_mq_get_tag+0x227/0x250 [ 8229.365265] Code: 44 24 04 44 01 e0 48 8b 74 24 38 65 48 33 34 25 28 00 00 00 75 33 48 83 c4 40 5b 5d 41 5c 41 5d 41 5e c3 48 8d 68 10 4c 89 ef <44> 8b 60 04 48 89 ee e8 dd f9 ff ff 83 f8 ff 75 c8 e9 67 fe ff ff [ 8229.365304] RSP: 0018:ffffba590e977970 EFLAGS: 00010246 [ 8229.365317] RAX: 0000000000000000 RBX: ffff9f00f6a79860 RCX: ffffba590e977998 [ 8229.365333] RDX: 0000000000000000 RSI: ffff9f012039b140 RDI: ffffba590e977a38 [ 8229.365349] RBP: 0000000000000010 R08: ffffda58ff94e190 R09: ffffda58ff94e198 [ 8229.365365] R10: 0000000000000011 R11: ffff9f00f6a79860 R12: 0000000000000000 [ 8229.365381] R13: ffffba590e977a38 R14: ffff9f012039b140 R15: 0000000000000001 [ 8229.365397] FS: 00007f481c230580(0000) GS:ffff9f013f940000(0000) knlGS:0000000000000000 [ 8229.365415] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 8229.365428] CR2: 0000000000000004 CR3: 0000005f35e26004 CR4: 00000000007606e0 [ 8229.365444] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 8229.365460] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 8229.365476] PKRU: 55555554 [ 8229.365484] Call Trace: [ 8229.365498] ? finish_wait+0x80/0x80 [ 8229.365512] blk_mq_get_request+0xcb/0x3f0 [ 8229.365525] blk_mq_make_request+0x143/0x5d0 [ 8229.365538] generic_make_request+0xcf/0x310 [ 8229.365553] ? scan_shadow_nodes+0x30/0x30 [ 8229.365564] submit_bio+0x3c/0x150 [ 8229.365576] mpage_readpages+0x163/0x1a0 [ 8229.365588] ? blkdev_direct_IO+0x490/0x490 [ 8229.365601] read_pages+0x6b/0x190 [ 8229.365612] __do_page_cache_readahead+0x1c1/0x1e0 [ 8229.365626] ondemand_readahead+0x182/0x2f0 [ 8229.365639] generic_file_buffered_read+0x590/0xab0 [ 8229.365655] new_sync_read+0x12a/0x1c0 [ 8229.365666] vfs_read+0x8a/0x140 [ 8229.365676] ksys_read+0x59/0xd0 [ 8229.365688] do_syscall_64+0x55/0x1d0 [ 8229.365700] entry_SYSCALL_64_after_hwframe+0x44/0xa9 Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Weiping Zhang <zhangweiping@didiglobal.com> Tested-by: Weiping Zhang <zhangweiping@didiglobal.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-07 13:04:08 +00:00
/*
* If tags initialization fail for some hctx,
* that hctx won't be brought online. In this
* case, remap the current ctx to hctx[0] which
* is guaranteed to always have tags allocated
*/
set->map[j].mq_map[i] = 0;
}
hctx = blk_mq_map_queue_type(q, j, i);
ctx->hctxs[j] = hctx;
/*
* If the CPU is already set in the mask, then we've
* mapped this one already. This can happen if
* devices share queues across queue maps.
*/
if (cpumask_test_cpu(i, hctx->cpumask))
continue;
cpumask_set_cpu(i, hctx->cpumask);
hctx->type = j;
ctx->index_hw[hctx->type] = hctx->nr_ctx;
hctx->ctxs[hctx->nr_ctx++] = ctx;
/*
* If the nr_ctx type overflows, we have exceeded the
* amount of sw queues we can support.
*/
BUG_ON(!hctx->nr_ctx);
}
for (; j < HCTX_MAX_TYPES; j++)
ctx->hctxs[j] = blk_mq_map_queue_type(q,
HCTX_TYPE_DEFAULT, i);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
queue_for_each_hw_ctx(q, hctx, i) {
int cpu;
/*
* If no software queues are mapped to this hardware queue,
* disable it and free the request entries.
*/
if (!hctx->nr_ctx) {
/* Never unmap queue 0. We need it as a
* fallback in case of a new remap fails
* allocation
*/
if (i)
__blk_mq_free_map_and_rqs(set, i);
hctx->tags = NULL;
continue;
}
hctx->tags = set->tags[i];
WARN_ON(!hctx->tags);
/*
* Set the map size to the number of mapped software queues.
* This is more accurate and more efficient than looping
* over all possibly mapped software queues.
*/
sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
/*
* Rule out isolated CPUs from hctx->cpumask to avoid
* running block kworker on isolated CPUs
*/
for_each_cpu(cpu, hctx->cpumask) {
if (cpu_is_isolated(cpu))
cpumask_clear_cpu(cpu, hctx->cpumask);
}
/*
* Initialize batch roundrobin counts
*/
hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
/*
* Caller needs to ensure that we're either frozen/quiesced, or that
* the queue isn't live yet.
*/
static void queue_set_hctx_shared(struct request_queue *q, bool shared)
{
struct blk_mq_hw_ctx *hctx;
unsigned long i;
queue_for_each_hw_ctx(q, hctx, i) {
if (shared) {
hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
} else {
blk_mq_tag_idle(hctx);
hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
}
}
}
static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
bool shared)
{
struct request_queue *q;
lockdep_assert_held(&set->tag_list_lock);
list_for_each_entry(q, &set->tag_list, tag_set_list) {
blk_mq_freeze_queue(q);
queue_set_hctx_shared(q, shared);
blk_mq_unfreeze_queue(q);
}
}
static void blk_mq_del_queue_tag_set(struct request_queue *q)
{
struct blk_mq_tag_set *set = q->tag_set;
mutex_lock(&set->tag_list_lock);
list_del(&q->tag_set_list);
if (list_is_singular(&set->tag_list)) {
/* just transitioned to unshared */
set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
/* update existing queue */
blk_mq_update_tag_set_shared(set, false);
}
mutex_unlock(&set->tag_list_lock);
blk-mq: reinit q->tag_set_list entry only after grace period It is not allowed to reinit q->tag_set_list list entry while RCU grace period has not completed yet, otherwise the following soft lockup in blk_mq_sched_restart() happens: [ 1064.252652] watchdog: BUG: soft lockup - CPU#12 stuck for 23s! [fio:9270] [ 1064.254445] task: ffff99b912e8b900 task.stack: ffffa6d54c758000 [ 1064.254613] RIP: 0010:blk_mq_sched_restart+0x96/0x150 [ 1064.256510] Call Trace: [ 1064.256664] <IRQ> [ 1064.256824] blk_mq_free_request+0xea/0x100 [ 1064.256987] msg_io_conf+0x59/0xd0 [ibnbd_client] [ 1064.257175] complete_rdma_req+0xf2/0x230 [ibtrs_client] [ 1064.257340] ? ibtrs_post_recv_empty+0x4d/0x70 [ibtrs_core] [ 1064.257502] ibtrs_clt_rdma_done+0xd1/0x1e0 [ibtrs_client] [ 1064.257669] ib_create_qp+0x321/0x380 [ib_core] [ 1064.257841] ib_process_cq_direct+0xbd/0x120 [ib_core] [ 1064.258007] irq_poll_softirq+0xb7/0xe0 [ 1064.258165] __do_softirq+0x106/0x2a2 [ 1064.258328] irq_exit+0x92/0xa0 [ 1064.258509] do_IRQ+0x4a/0xd0 [ 1064.258660] common_interrupt+0x7a/0x7a [ 1064.258818] </IRQ> Meanwhile another context frees other queue but with the same set of shared tags: [ 1288.201183] INFO: task bash:5910 blocked for more than 180 seconds. [ 1288.201833] bash D 0 5910 5820 0x00000000 [ 1288.202016] Call Trace: [ 1288.202315] schedule+0x32/0x80 [ 1288.202462] schedule_timeout+0x1e5/0x380 [ 1288.203838] wait_for_completion+0xb0/0x120 [ 1288.204137] __wait_rcu_gp+0x125/0x160 [ 1288.204287] synchronize_sched+0x6e/0x80 [ 1288.204770] blk_mq_free_queue+0x74/0xe0 [ 1288.204922] blk_cleanup_queue+0xc7/0x110 [ 1288.205073] ibnbd_clt_unmap_device+0x1bc/0x280 [ibnbd_client] [ 1288.205389] ibnbd_clt_unmap_dev_store+0x169/0x1f0 [ibnbd_client] [ 1288.205548] kernfs_fop_write+0x109/0x180 [ 1288.206328] vfs_write+0xb3/0x1a0 [ 1288.206476] SyS_write+0x52/0xc0 [ 1288.206624] do_syscall_64+0x68/0x1d0 [ 1288.206774] entry_SYSCALL_64_after_hwframe+0x3d/0xa2 What happened is the following: 1. There are several MQ queues with shared tags. 2. One queue is about to be freed and now task is in blk_mq_del_queue_tag_set(). 3. Other CPU is in blk_mq_sched_restart() and loops over all queues in tag list in order to find hctx to restart. Because linked list entry was modified in blk_mq_del_queue_tag_set() without proper waiting for a grace period, blk_mq_sched_restart() never ends, spining in list_for_each_entry_rcu_rr(), thus soft lockup. Fix is simple: reinit list entry after an RCU grace period elapsed. Fixes: Fixes: 705cda97ee3a ("blk-mq: Make it safe to use RCU to iterate over blk_mq_tag_set.tag_list") Cc: stable@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-block@vger.kernel.org Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com> Signed-off-by: Roman Pen <roman.penyaev@profitbricks.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-06-10 20:38:24 +00:00
INIT_LIST_HEAD(&q->tag_set_list);
}
static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
struct request_queue *q)
{
mutex_lock(&set->tag_list_lock);
/*
* Check to see if we're transitioning to shared (from 1 to 2 queues).
*/
if (!list_empty(&set->tag_list) &&
!(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
/* update existing queue */
blk_mq_update_tag_set_shared(set, true);
}
if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
queue_set_hctx_shared(q, true);
list_add_tail(&q->tag_set_list, &set->tag_list);
mutex_unlock(&set->tag_list_lock);
}
/* All allocations will be freed in release handler of q->mq_kobj */
static int blk_mq_alloc_ctxs(struct request_queue *q)
{
struct blk_mq_ctxs *ctxs;
int cpu;
ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
if (!ctxs)
return -ENOMEM;
ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
if (!ctxs->queue_ctx)
goto fail;
for_each_possible_cpu(cpu) {
struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
ctx->ctxs = ctxs;
}
q->mq_kobj = &ctxs->kobj;
q->queue_ctx = ctxs->queue_ctx;
return 0;
fail:
kfree(ctxs);
return -ENOMEM;
}
/*
* It is the actual release handler for mq, but we do it from
* request queue's release handler for avoiding use-after-free
* and headache because q->mq_kobj shouldn't have been introduced,
* but we can't group ctx/kctx kobj without it.
*/
void blk_mq_release(struct request_queue *q)
{
struct blk_mq_hw_ctx *hctx, *next;
unsigned long i;
queue_for_each_hw_ctx(q, hctx, i)
WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
/* all hctx are in .unused_hctx_list now */
list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
list_del_init(&hctx->hctx_list);
kobject_put(&hctx->kobj);
}
xa_destroy(&q->hctx_table);
/*
* release .mq_kobj and sw queue's kobject now because
* both share lifetime with request queue.
*/
blk_mq_sysfs_deinit(q);
}
static bool blk_mq_can_poll(struct blk_mq_tag_set *set)
{
return set->nr_maps > HCTX_TYPE_POLL &&
set->map[HCTX_TYPE_POLL].nr_queues;
}
struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set,
struct queue_limits *lim, void *queuedata)
{
struct queue_limits default_lim = { };
struct request_queue *q;
int ret;
if (!lim)
lim = &default_lim;
lim->features |= BLK_FEAT_IO_STAT | BLK_FEAT_NOWAIT;
if (blk_mq_can_poll(set))
lim->features |= BLK_FEAT_POLL;
q = blk_alloc_queue(lim, set->numa_node);
if (IS_ERR(q))
return q;
q->queuedata = queuedata;
ret = blk_mq_init_allocated_queue(set, q);
if (ret) {
blk_put_queue(q);
return ERR_PTR(ret);
}
return q;
}
EXPORT_SYMBOL(blk_mq_alloc_queue);
/**
* blk_mq_destroy_queue - shutdown a request queue
* @q: request queue to shutdown
*
* This shuts down a request queue allocated by blk_mq_alloc_queue(). All future
* requests will be failed with -ENODEV. The caller is responsible for dropping
* the reference from blk_mq_alloc_queue() by calling blk_put_queue().
*
* Context: can sleep
*/
void blk_mq_destroy_queue(struct request_queue *q)
{
WARN_ON_ONCE(!queue_is_mq(q));
WARN_ON_ONCE(blk_queue_registered(q));
might_sleep();
blk_queue_flag_set(QUEUE_FLAG_DYING, q);
blk_queue_start_drain(q);
blk_mq_freeze_queue_wait(q);
blk_sync_queue(q);
blk_mq_cancel_work_sync(q);
blk_mq_exit_queue(q);
}
EXPORT_SYMBOL(blk_mq_destroy_queue);
struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set,
struct queue_limits *lim, void *queuedata,
struct lock_class_key *lkclass)
{
struct request_queue *q;
struct gendisk *disk;
q = blk_mq_alloc_queue(set, lim, queuedata);
if (IS_ERR(q))
return ERR_CAST(q);
disk = __alloc_disk_node(q, set->numa_node, lkclass);
if (!disk) {
blk_mq_destroy_queue(q);
blk_put_queue(q);
return ERR_PTR(-ENOMEM);
}
set_bit(GD_OWNS_QUEUE, &disk->state);
return disk;
}
EXPORT_SYMBOL(__blk_mq_alloc_disk);
struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
struct lock_class_key *lkclass)
{
struct gendisk *disk;
if (!blk_get_queue(q))
return NULL;
disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
if (!disk)
blk_put_queue(q);
return disk;
}
EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
struct blk_mq_tag_set *set, struct request_queue *q,
int hctx_idx, int node)
{
struct blk_mq_hw_ctx *hctx = NULL, *tmp;
/* reuse dead hctx first */
spin_lock(&q->unused_hctx_lock);
list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
if (tmp->numa_node == node) {
hctx = tmp;
break;
}
}
if (hctx)
list_del_init(&hctx->hctx_list);
spin_unlock(&q->unused_hctx_lock);
if (!hctx)
hctx = blk_mq_alloc_hctx(q, set, node);
if (!hctx)
goto fail;
if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
goto free_hctx;
return hctx;
free_hctx:
kobject_put(&hctx->kobj);
fail:
return NULL;
}
static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
struct request_queue *q)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct blk_mq_hw_ctx *hctx;
unsigned long i, j;
/* protect against switching io scheduler */
mutex_lock(&q->sysfs_lock);
for (i = 0; i < set->nr_hw_queues; i++) {
int old_node;
int node = blk_mq_get_hctx_node(set, i);
struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
if (old_hctx) {
old_node = old_hctx->numa_node;
blk_mq_exit_hctx(q, set, old_hctx, i);
}
if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
if (!old_hctx)
break;
pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
node, old_node);
hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
WARN_ON_ONCE(!hctx);
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
/*
* Increasing nr_hw_queues fails. Free the newly allocated
* hctxs and keep the previous q->nr_hw_queues.
*/
if (i != set->nr_hw_queues) {
j = q->nr_hw_queues;
} else {
j = i;
q->nr_hw_queues = set->nr_hw_queues;
}
xa_for_each_start(&q->hctx_table, j, hctx, j)
blk_mq_exit_hctx(q, set, hctx, j);
mutex_unlock(&q->sysfs_lock);
}
int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
struct request_queue *q)
{
/* mark the queue as mq asap */
q->mq_ops = set->ops;
if (blk_mq_alloc_ctxs(q))
goto err_exit;
blk-mq: initialize mq kobjects in blk_mq_init_allocated_queue() Both q->mq_kobj and sw queues' kobjects should have been initialized once, instead of doing that each add_disk context. Also this patch removes clearing of ctx in blk_mq_init_cpu_queues() because percpu allocator fills zero to allocated variable. This patch fixes one issue[1] reported from Omar. [1] kernel wearning when doing unbind/bind on one scsi-mq device [ 19.347924] kobject (ffff8800791ea0b8): tried to init an initialized object, something is seriously wrong. [ 19.349781] CPU: 1 PID: 84 Comm: kworker/u8:1 Not tainted 4.10.0-rc7-00210-g53f39eeaa263 #34 [ 19.350686] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.1-20161122_114906-anatol 04/01/2014 [ 19.350920] Workqueue: events_unbound async_run_entry_fn [ 19.350920] Call Trace: [ 19.350920] dump_stack+0x63/0x83 [ 19.350920] kobject_init+0x77/0x90 [ 19.350920] blk_mq_register_dev+0x40/0x130 [ 19.350920] blk_register_queue+0xb6/0x190 [ 19.350920] device_add_disk+0x1ec/0x4b0 [ 19.350920] sd_probe_async+0x10d/0x1c0 [sd_mod] [ 19.350920] async_run_entry_fn+0x48/0x150 [ 19.350920] process_one_work+0x1d0/0x480 [ 19.350920] worker_thread+0x48/0x4e0 [ 19.350920] kthread+0x101/0x140 [ 19.350920] ? process_one_work+0x480/0x480 [ 19.350920] ? kthread_create_on_node+0x60/0x60 [ 19.350920] ret_from_fork+0x2c/0x40 Cc: Omar Sandoval <osandov@osandov.com> Signed-off-by: Ming Lei <tom.leiming@gmail.com> Tested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Jens Axboe <axboe@fb.com>
2017-02-22 10:13:59 +00:00
/* init q->mq_kobj and sw queues' kobjects */
blk_mq_sysfs_init(q);
INIT_LIST_HEAD(&q->unused_hctx_list);
spin_lock_init(&q->unused_hctx_lock);
xa_init(&q->hctx_table);
blk_mq_realloc_hw_ctxs(set, q);
if (!q->nr_hw_queues)
goto err_hctxs;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
q->tag_set = set;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
INIT_LIST_HEAD(&q->flush_list);
INIT_LIST_HEAD(&q->requeue_list);
spin_lock_init(&q->requeue_lock);
q->nr_requests = set->queue_depth;
blk_mq_init_cpu_queues(q, set->nr_hw_queues);
blk_mq_add_queue_tag_set(set, q);
blk_mq_map_swqueue(q);
return 0;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
err_hctxs:
blk-mq: Fix kmemleak in blk_mq_init_allocated_queue There is a kmemleak caused by modprobe null_blk.ko unreferenced object 0xffff8881acb1f000 (size 1024): comm "modprobe", pid 836, jiffies 4294971190 (age 27.068s) hex dump (first 32 bytes): 00 00 00 00 ad 4e ad de ff ff ff ff 00 00 00 00 .....N.......... ff ff ff ff ff ff ff ff 00 53 99 9e ff ff ff ff .........S...... backtrace: [<000000004a10c249>] kmalloc_node_trace+0x22/0x60 [<00000000648f7950>] blk_mq_alloc_and_init_hctx+0x289/0x350 [<00000000af06de0e>] blk_mq_realloc_hw_ctxs+0x2fe/0x3d0 [<00000000e00c1872>] blk_mq_init_allocated_queue+0x48c/0x1440 [<00000000d16b4e68>] __blk_mq_alloc_disk+0xc8/0x1c0 [<00000000d10c98c3>] 0xffffffffc450d69d [<00000000b9299f48>] 0xffffffffc4538392 [<0000000061c39ed6>] do_one_initcall+0xd0/0x4f0 [<00000000b389383b>] do_init_module+0x1a4/0x680 [<0000000087cf3542>] load_module+0x6249/0x7110 [<00000000beba61b8>] __do_sys_finit_module+0x140/0x200 [<00000000fdcfff51>] do_syscall_64+0x35/0x80 [<000000003c0f1f71>] entry_SYSCALL_64_after_hwframe+0x46/0xb0 That is because q->ma_ops is set to NULL before blk_release_queue is called. blk_mq_init_queue_data blk_mq_init_allocated_queue blk_mq_realloc_hw_ctxs for (i = 0; i < set->nr_hw_queues; i++) { old_hctx = xa_load(&q->hctx_table, i); if (!blk_mq_alloc_and_init_hctx(.., i, ..)) [1] if (!old_hctx) break; xa_for_each_start(&q->hctx_table, j, hctx, j) blk_mq_exit_hctx(q, set, hctx, j); [2] if (!q->nr_hw_queues) [3] goto err_hctxs; err_exit: q->mq_ops = NULL; [4] blk_put_queue blk_release_queue if (queue_is_mq(q)) [5] blk_mq_release(q); [1]: blk_mq_alloc_and_init_hctx failed at i != 0. [2]: The hctxs allocated by [1] are moved to q->unused_hctx_list and will be cleaned up in blk_mq_release. [3]: q->nr_hw_queues is 0. [4]: Set q->mq_ops to NULL. [5]: queue_is_mq returns false due to [4]. And blk_mq_release will not be called. The hctxs in q->unused_hctx_list are leaked. To fix it, call blk_release_queue in exception path. Fixes: 2f8f1336a48b ("blk-mq: always free hctx after request queue is freed") Signed-off-by: Yuan Can <yuancan@huawei.com> Signed-off-by: Chen Jun <chenjun102@huawei.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Link: https://lore.kernel.org/r/20221031031242.94107-1-chenjun102@huawei.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-10-31 03:12:42 +00:00
blk_mq_release(q);
err_exit:
q->mq_ops = NULL;
return -ENOMEM;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
EXPORT_SYMBOL(blk_mq_init_allocated_queue);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
blk-mq: free hw queue's resource in hctx's release handler Once blk_cleanup_queue() returns, tags shouldn't be used any more, because blk_mq_free_tag_set() may be called. Commit 45a9c9d909b2 ("blk-mq: Fix a use-after-free") fixes this issue exactly. However, that commit introduces another issue. Before 45a9c9d909b2, we are allowed to run queue during cleaning up queue if the queue's kobj refcount is held. After that commit, queue can't be run during queue cleaning up, otherwise oops can be triggered easily because some fields of hctx are freed by blk_mq_free_queue() in blk_cleanup_queue(). We have invented ways for addressing this kind of issue before, such as: 8dc765d438f1 ("SCSI: fix queue cleanup race before queue initialization is done") c2856ae2f315 ("blk-mq: quiesce queue before freeing queue") But still can't cover all cases, recently James reports another such kind of issue: https://marc.info/?l=linux-scsi&m=155389088124782&w=2 This issue can be quite hard to address by previous way, given scsi_run_queue() may run requeues for other LUNs. Fixes the above issue by freeing hctx's resources in its release handler, and this way is safe becasue tags isn't needed for freeing such hctx resource. This approach follows typical design pattern wrt. kobject's release handler. Cc: Dongli Zhang <dongli.zhang@oracle.com> Cc: James Smart <james.smart@broadcom.com> Cc: Bart Van Assche <bart.vanassche@wdc.com> Cc: linux-scsi@vger.kernel.org, Cc: Martin K . Petersen <martin.petersen@oracle.com>, Cc: Christoph Hellwig <hch@lst.de>, Cc: James E . J . Bottomley <jejb@linux.vnet.ibm.com>, Reported-by: James Smart <james.smart@broadcom.com> Fixes: 45a9c9d909b2 ("blk-mq: Fix a use-after-free") Cc: stable@vger.kernel.org Reviewed-by: Hannes Reinecke <hare@suse.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: James Smart <james.smart@broadcom.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-30 01:52:25 +00:00
/* tags can _not_ be used after returning from blk_mq_exit_queue */
void blk_mq_exit_queue(struct request_queue *q)
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
{
struct blk_mq_tag_set *set = q->tag_set;
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
/* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
/* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
blk_mq_del_queue_tag_set(q);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
}
static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
{
int i;
if (blk_mq_is_shared_tags(set->flags)) {
set->shared_tags = blk_mq_alloc_map_and_rqs(set,
BLK_MQ_NO_HCTX_IDX,
set->queue_depth);
if (!set->shared_tags)
return -ENOMEM;
}
blk-mq: add cond_resched() in __blk_mq_alloc_rq_maps() We found blk_mq_alloc_rq_maps() takes more time in kernel space when testing nvme device hot-plugging. The test and anlysis as below. Debug code, 1, blk_mq_alloc_rq_maps(): u64 start, end; depth = set->queue_depth; start = ktime_get_ns(); pr_err("[%d:%s switch:%ld,%ld] queue depth %d, nr_hw_queues %d\n", current->pid, current->comm, current->nvcsw, current->nivcsw, set->queue_depth, set->nr_hw_queues); do { err = __blk_mq_alloc_rq_maps(set); if (!err) break; set->queue_depth >>= 1; if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { err = -ENOMEM; break; } } while (set->queue_depth); end = ktime_get_ns(); pr_err("[%d:%s switch:%ld,%ld] all hw queues init cost time %lld ns\n", current->pid, current->comm, current->nvcsw, current->nivcsw, end - start); 2, __blk_mq_alloc_rq_maps(): u64 start, end; for (i = 0; i < set->nr_hw_queues; i++) { start = ktime_get_ns(); if (!__blk_mq_alloc_rq_map(set, i)) goto out_unwind; end = ktime_get_ns(); pr_err("hw queue %d init cost time %lld ns\n", i, end - start); } Test nvme hot-plugging with above debug code, we found it totally cost more than 3ms in kernel space without being scheduled out when alloc rqs for all 16 hw queues with depth 1023, each hw queue cost about 140-250us. The cost time will be increased with hw queue number and queue depth increasing. And in an extreme case, if __blk_mq_alloc_rq_maps() returns -ENOMEM, it will try "queue_depth >>= 1", more time will be consumed. [ 428.428771] nvme nvme0: pci function 10000:01:00.0 [ 428.428798] nvme 10000:01:00.0: enabling device (0000 -> 0002) [ 428.428806] pcieport 10000:00:00.0: can't derive routing for PCI INT A [ 428.428809] nvme 10000:01:00.0: PCI INT A: no GSI [ 432.593374] [4688:kworker/u33:8 switch:663,2] queue depth 30, nr_hw_queues 1 [ 432.593404] hw queue 0 init cost time 22883 ns [ 432.593408] [4688:kworker/u33:8 switch:663,2] all hw queues init cost time 35960 ns [ 432.595953] nvme nvme0: 16/0/0 default/read/poll queues [ 432.595958] [4688:kworker/u33:8 switch:700,2] queue depth 1023, nr_hw_queues 16 [ 432.596203] hw queue 0 init cost time 242630 ns [ 432.596441] hw queue 1 init cost time 235913 ns [ 432.596659] hw queue 2 init cost time 216461 ns [ 432.596877] hw queue 3 init cost time 215851 ns [ 432.597107] hw queue 4 init cost time 228406 ns [ 432.597336] hw queue 5 init cost time 227298 ns [ 432.597564] hw queue 6 init cost time 224633 ns [ 432.597785] hw queue 7 init cost time 219954 ns [ 432.597937] hw queue 8 init cost time 150930 ns [ 432.598082] hw queue 9 init cost time 143496 ns [ 432.598231] hw queue 10 init cost time 147261 ns [ 432.598397] hw queue 11 init cost time 164522 ns [ 432.598542] hw queue 12 init cost time 143401 ns [ 432.598692] hw queue 13 init cost time 148934 ns [ 432.598841] hw queue 14 init cost time 147194 ns [ 432.598991] hw queue 15 init cost time 148942 ns [ 432.598993] [4688:kworker/u33:8 switch:700,2] all hw queues init cost time 3035099 ns [ 432.602611] nvme0n1: p1 So use this patch to trigger schedule between each hw queue init, to avoid other threads getting stuck. It is not in atomic context when executing __blk_mq_alloc_rq_maps(), so it is safe to call cond_resched(). Signed-off-by: Xianting Tian <tian.xianting@h3c.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-09-26 02:39:47 +00:00
for (i = 0; i < set->nr_hw_queues; i++) {
if (!__blk_mq_alloc_map_and_rqs(set, i))
goto out_unwind;
blk-mq: add cond_resched() in __blk_mq_alloc_rq_maps() We found blk_mq_alloc_rq_maps() takes more time in kernel space when testing nvme device hot-plugging. The test and anlysis as below. Debug code, 1, blk_mq_alloc_rq_maps(): u64 start, end; depth = set->queue_depth; start = ktime_get_ns(); pr_err("[%d:%s switch:%ld,%ld] queue depth %d, nr_hw_queues %d\n", current->pid, current->comm, current->nvcsw, current->nivcsw, set->queue_depth, set->nr_hw_queues); do { err = __blk_mq_alloc_rq_maps(set); if (!err) break; set->queue_depth >>= 1; if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { err = -ENOMEM; break; } } while (set->queue_depth); end = ktime_get_ns(); pr_err("[%d:%s switch:%ld,%ld] all hw queues init cost time %lld ns\n", current->pid, current->comm, current->nvcsw, current->nivcsw, end - start); 2, __blk_mq_alloc_rq_maps(): u64 start, end; for (i = 0; i < set->nr_hw_queues; i++) { start = ktime_get_ns(); if (!__blk_mq_alloc_rq_map(set, i)) goto out_unwind; end = ktime_get_ns(); pr_err("hw queue %d init cost time %lld ns\n", i, end - start); } Test nvme hot-plugging with above debug code, we found it totally cost more than 3ms in kernel space without being scheduled out when alloc rqs for all 16 hw queues with depth 1023, each hw queue cost about 140-250us. The cost time will be increased with hw queue number and queue depth increasing. And in an extreme case, if __blk_mq_alloc_rq_maps() returns -ENOMEM, it will try "queue_depth >>= 1", more time will be consumed. [ 428.428771] nvme nvme0: pci function 10000:01:00.0 [ 428.428798] nvme 10000:01:00.0: enabling device (0000 -> 0002) [ 428.428806] pcieport 10000:00:00.0: can't derive routing for PCI INT A [ 428.428809] nvme 10000:01:00.0: PCI INT A: no GSI [ 432.593374] [4688:kworker/u33:8 switch:663,2] queue depth 30, nr_hw_queues 1 [ 432.593404] hw queue 0 init cost time 22883 ns [ 432.593408] [4688:kworker/u33:8 switch:663,2] all hw queues init cost time 35960 ns [ 432.595953] nvme nvme0: 16/0/0 default/read/poll queues [ 432.595958] [4688:kworker/u33:8 switch:700,2] queue depth 1023, nr_hw_queues 16 [ 432.596203] hw queue 0 init cost time 242630 ns [ 432.596441] hw queue 1 init cost time 235913 ns [ 432.596659] hw queue 2 init cost time 216461 ns [ 432.596877] hw queue 3 init cost time 215851 ns [ 432.597107] hw queue 4 init cost time 228406 ns [ 432.597336] hw queue 5 init cost time 227298 ns [ 432.597564] hw queue 6 init cost time 224633 ns [ 432.597785] hw queue 7 init cost time 219954 ns [ 432.597937] hw queue 8 init cost time 150930 ns [ 432.598082] hw queue 9 init cost time 143496 ns [ 432.598231] hw queue 10 init cost time 147261 ns [ 432.598397] hw queue 11 init cost time 164522 ns [ 432.598542] hw queue 12 init cost time 143401 ns [ 432.598692] hw queue 13 init cost time 148934 ns [ 432.598841] hw queue 14 init cost time 147194 ns [ 432.598991] hw queue 15 init cost time 148942 ns [ 432.598993] [4688:kworker/u33:8 switch:700,2] all hw queues init cost time 3035099 ns [ 432.602611] nvme0n1: p1 So use this patch to trigger schedule between each hw queue init, to avoid other threads getting stuck. It is not in atomic context when executing __blk_mq_alloc_rq_maps(), so it is safe to call cond_resched(). Signed-off-by: Xianting Tian <tian.xianting@h3c.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-09-26 02:39:47 +00:00
cond_resched();
}
return 0;
out_unwind:
while (--i >= 0)
__blk_mq_free_map_and_rqs(set, i);
if (blk_mq_is_shared_tags(set->flags)) {
blk_mq_free_map_and_rqs(set, set->shared_tags,
BLK_MQ_NO_HCTX_IDX);
}
return -ENOMEM;
}
/*
* Allocate the request maps associated with this tag_set. Note that this
* may reduce the depth asked for, if memory is tight. set->queue_depth
* will be updated to reflect the allocated depth.
*/
static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
{
unsigned int depth;
int err;
depth = set->queue_depth;
do {
err = __blk_mq_alloc_rq_maps(set);
if (!err)
break;
set->queue_depth >>= 1;
if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
err = -ENOMEM;
break;
}
} while (set->queue_depth);
if (!set->queue_depth || err) {
pr_err("blk-mq: failed to allocate request map\n");
return -ENOMEM;
}
if (depth != set->queue_depth)
pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
depth, set->queue_depth);
return 0;
}
static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
{
/*
* blk_mq_map_queues() and multiple .map_queues() implementations
* expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
* number of hardware queues.
*/
if (set->nr_maps == 1)
set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
if (set->ops->map_queues) {
int i;
blk-mq: avoid to map CPU into stale hw queue blk_mq_pci_map_queues() may not map one CPU into any hw queue, but its previous map isn't cleared yet, and may point to one stale hw queue index. This patch fixes the following issue by clearing the mapping table before setting it up in blk_mq_pci_map_queues(). This patches fixes this following issue reported by Zhang Yi: [ 101.202734] BUG: unable to handle kernel NULL pointer dereference at 0000000094d3013f [ 101.211487] IP: blk_mq_map_swqueue+0xbc/0x200 [ 101.216346] PGD 0 P4D 0 [ 101.219171] Oops: 0000 [#1] SMP [ 101.222674] Modules linked in: sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel intel_cstate intel_uncore mxm_wmi intel_rapl_perf iTCO_wdt ipmi_si ipmi_devintf pcspkr iTCO_vendor_support sg dcdbas ipmi_msghandler wmi mei_me lpc_ich shpchp mei acpi_power_meter dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci crc32c_intel libata tg3 nvme nvme_core megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 101.284881] CPU: 0 PID: 504 Comm: kworker/u25:5 Not tainted 4.15.0-rc2 #1 [ 101.292455] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 101.301001] Workqueue: nvme-wq nvme_reset_work [nvme] [ 101.306636] task: 00000000f2c53190 task.stack: 000000002da874f9 [ 101.313241] RIP: 0010:blk_mq_map_swqueue+0xbc/0x200 [ 101.318681] RSP: 0018:ffffc9000234fd70 EFLAGS: 00010282 [ 101.324511] RAX: ffff88047ffc9480 RBX: ffff88047e130850 RCX: 0000000000000000 [ 101.332471] RDX: ffffe8ffffd40580 RSI: ffff88047e509b40 RDI: ffff88046f37a008 [ 101.340432] RBP: 000000000000000b R08: ffff88046f37a008 R09: 0000000011f94280 [ 101.348392] R10: ffff88047ffd4d00 R11: 0000000000000000 R12: ffff88046f37a008 [ 101.356353] R13: ffff88047e130f38 R14: 000000000000000b R15: ffff88046f37a558 [ 101.364314] FS: 0000000000000000(0000) GS:ffff880277c00000(0000) knlGS:0000000000000000 [ 101.373342] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 101.379753] CR2: 0000000000000098 CR3: 000000047f409004 CR4: 00000000001606f0 [ 101.387714] Call Trace: [ 101.390445] blk_mq_update_nr_hw_queues+0xbf/0x130 [ 101.395791] nvme_reset_work+0x6f4/0xc06 [nvme] [ 101.400848] ? pick_next_task_fair+0x290/0x5f0 [ 101.405807] ? __switch_to+0x1f5/0x430 [ 101.409988] ? put_prev_entity+0x2f/0xd0 [ 101.414365] process_one_work+0x141/0x340 [ 101.418836] worker_thread+0x47/0x3e0 [ 101.422921] kthread+0xf5/0x130 [ 101.426424] ? rescuer_thread+0x380/0x380 [ 101.430896] ? kthread_associate_blkcg+0x90/0x90 [ 101.436048] ret_from_fork+0x1f/0x30 [ 101.440034] Code: 48 83 3c ca 00 0f 84 2b 01 00 00 48 63 cd 48 8b 93 10 01 00 00 8b 0c 88 48 8b 83 20 01 00 00 4a 03 14 f5 60 04 af 81 48 8b 0c c8 <48> 8b 81 98 00 00 00 f0 4c 0f ab 30 8b 81 f8 00 00 00 89 42 44 [ 101.461116] RIP: blk_mq_map_swqueue+0xbc/0x200 RSP: ffffc9000234fd70 [ 101.468205] CR2: 0000000000000098 [ 101.471907] ---[ end trace 5fe710f98228a3ca ]--- [ 101.482489] Kernel panic - not syncing: Fatal exception [ 101.488505] Kernel Offset: disabled [ 101.497752] ---[ end Kernel panic - not syncing: Fatal exception Reviewed-by: Christoph Hellwig <hch@lst.de> Suggested-by: Christoph Hellwig <hch@lst.de> Reported-by: Yi Zhang <yi.zhang@redhat.com> Tested-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-06 08:27:39 +00:00
/*
* transport .map_queues is usually done in the following
* way:
*
* for (queue = 0; queue < set->nr_hw_queues; queue++) {
* mask = get_cpu_mask(queue)
* for_each_cpu(cpu, mask)
* set->map[x].mq_map[cpu] = queue;
blk-mq: avoid to map CPU into stale hw queue blk_mq_pci_map_queues() may not map one CPU into any hw queue, but its previous map isn't cleared yet, and may point to one stale hw queue index. This patch fixes the following issue by clearing the mapping table before setting it up in blk_mq_pci_map_queues(). This patches fixes this following issue reported by Zhang Yi: [ 101.202734] BUG: unable to handle kernel NULL pointer dereference at 0000000094d3013f [ 101.211487] IP: blk_mq_map_swqueue+0xbc/0x200 [ 101.216346] PGD 0 P4D 0 [ 101.219171] Oops: 0000 [#1] SMP [ 101.222674] Modules linked in: sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel intel_cstate intel_uncore mxm_wmi intel_rapl_perf iTCO_wdt ipmi_si ipmi_devintf pcspkr iTCO_vendor_support sg dcdbas ipmi_msghandler wmi mei_me lpc_ich shpchp mei acpi_power_meter dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci crc32c_intel libata tg3 nvme nvme_core megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 101.284881] CPU: 0 PID: 504 Comm: kworker/u25:5 Not tainted 4.15.0-rc2 #1 [ 101.292455] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 101.301001] Workqueue: nvme-wq nvme_reset_work [nvme] [ 101.306636] task: 00000000f2c53190 task.stack: 000000002da874f9 [ 101.313241] RIP: 0010:blk_mq_map_swqueue+0xbc/0x200 [ 101.318681] RSP: 0018:ffffc9000234fd70 EFLAGS: 00010282 [ 101.324511] RAX: ffff88047ffc9480 RBX: ffff88047e130850 RCX: 0000000000000000 [ 101.332471] RDX: ffffe8ffffd40580 RSI: ffff88047e509b40 RDI: ffff88046f37a008 [ 101.340432] RBP: 000000000000000b R08: ffff88046f37a008 R09: 0000000011f94280 [ 101.348392] R10: ffff88047ffd4d00 R11: 0000000000000000 R12: ffff88046f37a008 [ 101.356353] R13: ffff88047e130f38 R14: 000000000000000b R15: ffff88046f37a558 [ 101.364314] FS: 0000000000000000(0000) GS:ffff880277c00000(0000) knlGS:0000000000000000 [ 101.373342] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 101.379753] CR2: 0000000000000098 CR3: 000000047f409004 CR4: 00000000001606f0 [ 101.387714] Call Trace: [ 101.390445] blk_mq_update_nr_hw_queues+0xbf/0x130 [ 101.395791] nvme_reset_work+0x6f4/0xc06 [nvme] [ 101.400848] ? pick_next_task_fair+0x290/0x5f0 [ 101.405807] ? __switch_to+0x1f5/0x430 [ 101.409988] ? put_prev_entity+0x2f/0xd0 [ 101.414365] process_one_work+0x141/0x340 [ 101.418836] worker_thread+0x47/0x3e0 [ 101.422921] kthread+0xf5/0x130 [ 101.426424] ? rescuer_thread+0x380/0x380 [ 101.430896] ? kthread_associate_blkcg+0x90/0x90 [ 101.436048] ret_from_fork+0x1f/0x30 [ 101.440034] Code: 48 83 3c ca 00 0f 84 2b 01 00 00 48 63 cd 48 8b 93 10 01 00 00 8b 0c 88 48 8b 83 20 01 00 00 4a 03 14 f5 60 04 af 81 48 8b 0c c8 <48> 8b 81 98 00 00 00 f0 4c 0f ab 30 8b 81 f8 00 00 00 89 42 44 [ 101.461116] RIP: blk_mq_map_swqueue+0xbc/0x200 RSP: ffffc9000234fd70 [ 101.468205] CR2: 0000000000000098 [ 101.471907] ---[ end trace 5fe710f98228a3ca ]--- [ 101.482489] Kernel panic - not syncing: Fatal exception [ 101.488505] Kernel Offset: disabled [ 101.497752] ---[ end Kernel panic - not syncing: Fatal exception Reviewed-by: Christoph Hellwig <hch@lst.de> Suggested-by: Christoph Hellwig <hch@lst.de> Reported-by: Yi Zhang <yi.zhang@redhat.com> Tested-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-06 08:27:39 +00:00
* }
*
* When we need to remap, the table has to be cleared for
* killing stale mapping since one CPU may not be mapped
* to any hw queue.
*/
for (i = 0; i < set->nr_maps; i++)
blk_mq_clear_mq_map(&set->map[i]);
blk-mq: avoid to map CPU into stale hw queue blk_mq_pci_map_queues() may not map one CPU into any hw queue, but its previous map isn't cleared yet, and may point to one stale hw queue index. This patch fixes the following issue by clearing the mapping table before setting it up in blk_mq_pci_map_queues(). This patches fixes this following issue reported by Zhang Yi: [ 101.202734] BUG: unable to handle kernel NULL pointer dereference at 0000000094d3013f [ 101.211487] IP: blk_mq_map_swqueue+0xbc/0x200 [ 101.216346] PGD 0 P4D 0 [ 101.219171] Oops: 0000 [#1] SMP [ 101.222674] Modules linked in: sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel intel_cstate intel_uncore mxm_wmi intel_rapl_perf iTCO_wdt ipmi_si ipmi_devintf pcspkr iTCO_vendor_support sg dcdbas ipmi_msghandler wmi mei_me lpc_ich shpchp mei acpi_power_meter dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci crc32c_intel libata tg3 nvme nvme_core megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 101.284881] CPU: 0 PID: 504 Comm: kworker/u25:5 Not tainted 4.15.0-rc2 #1 [ 101.292455] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 101.301001] Workqueue: nvme-wq nvme_reset_work [nvme] [ 101.306636] task: 00000000f2c53190 task.stack: 000000002da874f9 [ 101.313241] RIP: 0010:blk_mq_map_swqueue+0xbc/0x200 [ 101.318681] RSP: 0018:ffffc9000234fd70 EFLAGS: 00010282 [ 101.324511] RAX: ffff88047ffc9480 RBX: ffff88047e130850 RCX: 0000000000000000 [ 101.332471] RDX: ffffe8ffffd40580 RSI: ffff88047e509b40 RDI: ffff88046f37a008 [ 101.340432] RBP: 000000000000000b R08: ffff88046f37a008 R09: 0000000011f94280 [ 101.348392] R10: ffff88047ffd4d00 R11: 0000000000000000 R12: ffff88046f37a008 [ 101.356353] R13: ffff88047e130f38 R14: 000000000000000b R15: ffff88046f37a558 [ 101.364314] FS: 0000000000000000(0000) GS:ffff880277c00000(0000) knlGS:0000000000000000 [ 101.373342] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 101.379753] CR2: 0000000000000098 CR3: 000000047f409004 CR4: 00000000001606f0 [ 101.387714] Call Trace: [ 101.390445] blk_mq_update_nr_hw_queues+0xbf/0x130 [ 101.395791] nvme_reset_work+0x6f4/0xc06 [nvme] [ 101.400848] ? pick_next_task_fair+0x290/0x5f0 [ 101.405807] ? __switch_to+0x1f5/0x430 [ 101.409988] ? put_prev_entity+0x2f/0xd0 [ 101.414365] process_one_work+0x141/0x340 [ 101.418836] worker_thread+0x47/0x3e0 [ 101.422921] kthread+0xf5/0x130 [ 101.426424] ? rescuer_thread+0x380/0x380 [ 101.430896] ? kthread_associate_blkcg+0x90/0x90 [ 101.436048] ret_from_fork+0x1f/0x30 [ 101.440034] Code: 48 83 3c ca 00 0f 84 2b 01 00 00 48 63 cd 48 8b 93 10 01 00 00 8b 0c 88 48 8b 83 20 01 00 00 4a 03 14 f5 60 04 af 81 48 8b 0c c8 <48> 8b 81 98 00 00 00 f0 4c 0f ab 30 8b 81 f8 00 00 00 89 42 44 [ 101.461116] RIP: blk_mq_map_swqueue+0xbc/0x200 RSP: ffffc9000234fd70 [ 101.468205] CR2: 0000000000000098 [ 101.471907] ---[ end trace 5fe710f98228a3ca ]--- [ 101.482489] Kernel panic - not syncing: Fatal exception [ 101.488505] Kernel Offset: disabled [ 101.497752] ---[ end Kernel panic - not syncing: Fatal exception Reviewed-by: Christoph Hellwig <hch@lst.de> Suggested-by: Christoph Hellwig <hch@lst.de> Reported-by: Yi Zhang <yi.zhang@redhat.com> Tested-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-06 08:27:39 +00:00
set->ops->map_queues(set);
} else {
BUG_ON(set->nr_maps > 1);
blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
}
}
static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
int new_nr_hw_queues)
{
struct blk_mq_tags **new_tags;
int i;
if (set->nr_hw_queues >= new_nr_hw_queues)
goto done;
new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
GFP_KERNEL, set->numa_node);
if (!new_tags)
return -ENOMEM;
if (set->tags)
memcpy(new_tags, set->tags, set->nr_hw_queues *
sizeof(*set->tags));
kfree(set->tags);
set->tags = new_tags;
for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
if (!__blk_mq_alloc_map_and_rqs(set, i)) {
while (--i >= set->nr_hw_queues)
__blk_mq_free_map_and_rqs(set, i);
return -ENOMEM;
}
cond_resched();
}
done:
set->nr_hw_queues = new_nr_hw_queues;
return 0;
}
/*
* Alloc a tag set to be associated with one or more request queues.
* May fail with EINVAL for various error conditions. May adjust the
* requested depth down, if it's too large. In that case, the set
* value will be stored in set->queue_depth.
*/
int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
{
int i, ret;
BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
if (!set->nr_hw_queues)
return -EINVAL;
if (!set->queue_depth)
return -EINVAL;
if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
return -EINVAL;
if (!set->ops->queue_rq)
return -EINVAL;
if (!set->ops->get_budget ^ !set->ops->put_budget)
return -EINVAL;
if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
pr_info("blk-mq: reduced tag depth to %u\n",
BLK_MQ_MAX_DEPTH);
set->queue_depth = BLK_MQ_MAX_DEPTH;
}
if (!set->nr_maps)
set->nr_maps = 1;
else if (set->nr_maps > HCTX_MAX_TYPES)
return -EINVAL;
/*
* If a crashdump is active, then we are potentially in a very
* memory constrained environment. Limit us to 64 tags to prevent
* using too much memory.
*/
if (is_kdump_kernel())
set->queue_depth = min(64U, set->queue_depth);
/*
* There is no use for more h/w queues than cpus if we just have
* a single map
*/
if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
set->nr_hw_queues = nr_cpu_ids;
if (set->flags & BLK_MQ_F_BLOCKING) {
set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
if (!set->srcu)
return -ENOMEM;
ret = init_srcu_struct(set->srcu);
if (ret)
goto out_free_srcu;
}
ret = -ENOMEM;
set->tags = kcalloc_node(set->nr_hw_queues,
sizeof(struct blk_mq_tags *), GFP_KERNEL,
set->numa_node);
if (!set->tags)
goto out_cleanup_srcu;
for (i = 0; i < set->nr_maps; i++) {
set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
sizeof(set->map[i].mq_map[0]),
GFP_KERNEL, set->numa_node);
if (!set->map[i].mq_map)
goto out_free_mq_map;
set->map[i].nr_queues = set->nr_hw_queues;
}
blk_mq_update_queue_map(set);
ret = blk_mq_alloc_set_map_and_rqs(set);
if (ret)
goto out_free_mq_map;
mutex_init(&set->tag_list_lock);
INIT_LIST_HEAD(&set->tag_list);
return 0;
out_free_mq_map:
for (i = 0; i < set->nr_maps; i++) {
kfree(set->map[i].mq_map);
set->map[i].mq_map = NULL;
}
kfree(set->tags);
set->tags = NULL;
out_cleanup_srcu:
if (set->flags & BLK_MQ_F_BLOCKING)
cleanup_srcu_struct(set->srcu);
out_free_srcu:
if (set->flags & BLK_MQ_F_BLOCKING)
kfree(set->srcu);
return ret;
}
EXPORT_SYMBOL(blk_mq_alloc_tag_set);
/* allocate and initialize a tagset for a simple single-queue device */
int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
const struct blk_mq_ops *ops, unsigned int queue_depth,
unsigned int set_flags)
{
memset(set, 0, sizeof(*set));
set->ops = ops;
set->nr_hw_queues = 1;
set->nr_maps = 1;
set->queue_depth = queue_depth;
set->numa_node = NUMA_NO_NODE;
set->flags = set_flags;
return blk_mq_alloc_tag_set(set);
}
EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
{
int i, j;
for (i = 0; i < set->nr_hw_queues; i++)
__blk_mq_free_map_and_rqs(set, i);
if (blk_mq_is_shared_tags(set->flags)) {
blk_mq_free_map_and_rqs(set, set->shared_tags,
BLK_MQ_NO_HCTX_IDX);
}
blk-mq: Facilitate a shared sbitmap per tagset Some SCSI HBAs (such as HPSA, megaraid, mpt3sas, hisi_sas_v3 ..) support multiple reply queues with single hostwide tags. In addition, these drivers want to use interrupt assignment in pci_alloc_irq_vectors(PCI_IRQ_AFFINITY). However, as discussed in [0], CPU hotplug may cause in-flight IO completion to not be serviced when an interrupt is shutdown. That problem is solved in commit bf0beec0607d ("blk-mq: drain I/O when all CPUs in a hctx are offline"). However, to take advantage of that blk-mq feature, the HBA HW queuess are required to be mapped to that of the blk-mq hctx's; to do that, the HBA HW queues need to be exposed to the upper layer. In making that transition, the per-SCSI command request tags are no longer unique per Scsi host - they are just unique per hctx. As such, the HBA LLDD would have to generate this tag internally, which has a certain performance overhead. However another problem is that blk-mq assumes the host may accept (Scsi_host.can_queue * #hw queue) commands. In commit 6eb045e092ef ("scsi: core: avoid host-wide host_busy counter for scsi_mq"), the Scsi host busy counter was removed, which would stop the LLDD being sent more than .can_queue commands; however, it should still be ensured that the block layer does not issue more than .can_queue commands to the Scsi host. To solve this problem, introduce a shared sbitmap per blk_mq_tag_set, which may be requested at init time. New flag BLK_MQ_F_TAG_HCTX_SHARED should be set when requesting the tagset to indicate whether the shared sbitmap should be used. Even when BLK_MQ_F_TAG_HCTX_SHARED is set, a full set of tags and requests are still allocated per hctx; the reason for this is that if tags and requests were only allocated for a single hctx - like hctx0 - it may break block drivers which expect a request be associated with a specific hctx, i.e. not always hctx0. This will introduce extra memory usage. This change is based on work originally from Ming Lei in [1] and from Bart's suggestion in [2]. [0] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [1] https://lore.kernel.org/linux-block/20190531022801.10003-1-ming.lei@redhat.com/ [2] https://lore.kernel.org/linux-block/ff77beff-5fd9-9f05-12b6-826922bace1f@huawei.com/T/#m3db0a602f095cbcbff27e9c884d6b4ae826144be Signed-off-by: John Garry <john.garry@huawei.com> Tested-by: Don Brace<don.brace@microsemi.com> #SCSI resv cmds patches used Tested-by: Douglas Gilbert <dgilbert@interlog.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-08-19 15:20:24 +00:00
for (j = 0; j < set->nr_maps; j++) {
kfree(set->map[j].mq_map);
set->map[j].mq_map = NULL;
}
kfree(set->tags);
set->tags = NULL;
if (set->flags & BLK_MQ_F_BLOCKING) {
cleanup_srcu_struct(set->srcu);
kfree(set->srcu);
}
}
EXPORT_SYMBOL(blk_mq_free_tag_set);
int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
{
struct blk_mq_tag_set *set = q->tag_set;
struct blk_mq_hw_ctx *hctx;
int ret;
unsigned long i;
if (WARN_ON_ONCE(!q->mq_freeze_depth))
return -EINVAL;
if (!set)
return -EINVAL;
if (q->nr_requests == nr)
return 0;
blk_mq_quiesce_queue(q);
ret = 0;
queue_for_each_hw_ctx(q, hctx, i) {
if (!hctx->tags)
continue;
/*
* If we're using an MQ scheduler, just update the scheduler
* queue depth. This is similar to what the old code would do.
*/
if (hctx->sched_tags) {
ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
nr, true);
} else {
ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
false);
}
if (ret)
break;
if (q->elevator && q->elevator->type->ops.depth_updated)
q->elevator->type->ops.depth_updated(hctx);
}
2021-05-13 12:00:58 +00:00
if (!ret) {
q->nr_requests = nr;
if (blk_mq_is_shared_tags(set->flags)) {
if (q->elevator)
blk_mq_tag_update_sched_shared_tags(q);
else
blk_mq_tag_resize_shared_tags(set, nr);
}
2021-05-13 12:00:58 +00:00
}
blk_mq_unquiesce_queue(q);
return ret;
}
/*
* request_queue and elevator_type pair.
* It is just used by __blk_mq_update_nr_hw_queues to cache
* the elevator_type associated with a request_queue.
*/
struct blk_mq_qe_pair {
struct list_head node;
struct request_queue *q;
struct elevator_type *type;
};
/*
* Cache the elevator_type in qe pair list and switch the
* io scheduler to 'none'
*/
static bool blk_mq_elv_switch_none(struct list_head *head,
struct request_queue *q)
{
struct blk_mq_qe_pair *qe;
qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
if (!qe)
return false;
/* q->elevator needs protection from ->sysfs_lock */
mutex_lock(&q->sysfs_lock);
/* the check has to be done with holding sysfs_lock */
if (!q->elevator) {
kfree(qe);
goto unlock;
}
INIT_LIST_HEAD(&qe->node);
qe->q = q;
qe->type = q->elevator->type;
/* keep a reference to the elevator module as we'll switch back */
__elevator_get(qe->type);
list_add(&qe->node, head);
elevator_disable(q);
unlock:
mutex_unlock(&q->sysfs_lock);
return true;
}
static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
struct request_queue *q)
{
struct blk_mq_qe_pair *qe;
list_for_each_entry(qe, head, node)
if (qe->q == q)
return qe;
return NULL;
}
static void blk_mq_elv_switch_back(struct list_head *head,
struct request_queue *q)
{
struct blk_mq_qe_pair *qe;
struct elevator_type *t;
qe = blk_lookup_qe_pair(head, q);
if (!qe)
return;
t = qe->type;
list_del(&qe->node);
kfree(qe);
mutex_lock(&q->sysfs_lock);
elevator_switch(q, t);
block: fix up elevator_type refcounting The current reference management logic of io scheduler modules contains refcnt problems. For example, blk_mq_init_sched may fail before or after the calling of e->ops.init_sched. If it fails before the calling, it does nothing to the reference to the io scheduler module. But if it fails after the calling, it releases the reference by calling kobject_put(&eq->kobj). As the callers of blk_mq_init_sched can't know exactly where the failure happens, they can't handle the reference to the io scheduler module properly: releasing the reference on failure results in double-release if blk_mq_init_sched has released it, and not releasing the reference results in ghost reference if blk_mq_init_sched did not release it either. The same problem also exists in io schedulers' init_sched implementations. We can address the problem by adding releasing statements to the error handling procedures of blk_mq_init_sched and init_sched implementations. But that is counterintuitive and requires modifications to existing io schedulers. Instead, We make elevator_alloc get the io scheduler module references that will be released by elevator_release. And then, we match each elevator_get with an elevator_put. Therefore, each reference to an io scheduler module explicitly has its own getter and releaser, and we no longer need to worry about the refcnt problems. The bugs and the patch can be validated with tools here: https://github.com/nickyc975/linux_elv_refcnt_bug.git [hch: split out a few bits into separate patches, use a non-try module_get in elevator_alloc] Signed-off-by: Jinlong Chen <nickyc975@zju.edu.cn> Signed-off-by: Christoph Hellwig <hch@lst.de> Link: https://lore.kernel.org/r/20221020064819.1469928-5-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-10-20 06:48:19 +00:00
/* drop the reference acquired in blk_mq_elv_switch_none */
elevator_put(t);
mutex_unlock(&q->sysfs_lock);
}
static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
int nr_hw_queues)
{
struct request_queue *q;
LIST_HEAD(head);
int prev_nr_hw_queues = set->nr_hw_queues;
int i;
lockdep_assert_held(&set->tag_list_lock);
if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
nr_hw_queues = nr_cpu_ids;
if (nr_hw_queues < 1)
return;
if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
return;
list_for_each_entry(q, &set->tag_list, tag_set_list)
blk_mq_freeze_queue(q);
/*
* Switch IO scheduler to 'none', cleaning up the data associated
* with the previous scheduler. We will switch back once we are done
* updating the new sw to hw queue mappings.
*/
list_for_each_entry(q, &set->tag_list, tag_set_list)
if (!blk_mq_elv_switch_none(&head, q))
goto switch_back;
list_for_each_entry(q, &set->tag_list, tag_set_list) {
blk_mq_debugfs_unregister_hctxs(q);
blk_mq_sysfs_unregister_hctxs(q);
}
if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
goto reregister;
fallback:
block: reset mapping if failed to update hardware queue count When we increase hardware queue count, blk_mq_update_queue_map will reset the mapping between cpu and hardware queue base on the hardware queue count(set->nr_hw_queues). The mapping cannot be reset if it encounters error in blk_mq_realloc_hw_ctxs, but the fallback flow will continue using it, then blk_mq_map_swqueue will touch a invalid memory, because the mapping points to a wrong hctx. blktest block/030: null_blk: module loaded Increasing nr_hw_queues to 8 fails, fallback to 1 ================================================================== BUG: KASAN: null-ptr-deref in blk_mq_map_swqueue+0x2f2/0x830 Read of size 8 at addr 0000000000000128 by task nproc/8541 CPU: 5 PID: 8541 Comm: nproc Not tainted 5.7.0-rc4-dbg+ #3 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.13.0-0-gf21b5a4-rebuilt.opensuse.org 04/01/2014 Call Trace: dump_stack+0xa5/0xe6 __kasan_report.cold+0x65/0xbb kasan_report+0x45/0x60 check_memory_region+0x15e/0x1c0 __kasan_check_read+0x15/0x20 blk_mq_map_swqueue+0x2f2/0x830 __blk_mq_update_nr_hw_queues+0x3df/0x690 blk_mq_update_nr_hw_queues+0x32/0x50 nullb_device_submit_queues_store+0xde/0x160 [null_blk] configfs_write_file+0x1c4/0x250 [configfs] __vfs_write+0x4c/0x90 vfs_write+0x14b/0x2d0 ksys_write+0xdd/0x180 __x64_sys_write+0x47/0x50 do_syscall_64+0x6f/0x310 entry_SYSCALL_64_after_hwframe+0x49/0xb3 Signed-off-by: Weiping Zhang <zhangweiping@didiglobal.com> Tested-by: Bart van Assche <bvanassche@acm.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-13 00:44:05 +00:00
blk_mq_update_queue_map(set);
list_for_each_entry(q, &set->tag_list, tag_set_list) {
struct queue_limits lim;
blk_mq_realloc_hw_ctxs(set, q);
if (q->nr_hw_queues != set->nr_hw_queues) {
blk-mq: don't free tags if the tag_set is used by other device in queue initialztion We got UAF report on v5.10 as follows: [ 1446.674930] ================================================================== [ 1446.675970] BUG: KASAN: use-after-free in blk_mq_get_driver_tag+0x9a4/0xa90 [ 1446.676902] Read of size 8 at addr ffff8880185afd10 by task kworker/1:2/12348 [ 1446.677851] [ 1446.678073] CPU: 1 PID: 12348 Comm: kworker/1:2 Not tainted 5.10.0-10177-gc9c81b1e346a #2 [ 1446.679168] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014 [ 1446.680692] Workqueue: kthrotld blk_throtl_dispatch_work_fn [ 1446.681448] Call Trace: [ 1446.681800] dump_stack+0x9b/0xce [ 1446.682916] print_address_description.constprop.6+0x3e/0x60 [ 1446.685999] kasan_report.cold.9+0x22/0x3a [ 1446.687186] blk_mq_get_driver_tag+0x9a4/0xa90 [ 1446.687785] blk_mq_dispatch_rq_list+0x21a/0x1d40 [ 1446.692576] __blk_mq_do_dispatch_sched+0x394/0x830 [ 1446.695758] __blk_mq_sched_dispatch_requests+0x398/0x4f0 [ 1446.698279] blk_mq_sched_dispatch_requests+0xdf/0x140 [ 1446.698967] __blk_mq_run_hw_queue+0xc0/0x270 [ 1446.699561] __blk_mq_delay_run_hw_queue+0x4cc/0x550 [ 1446.701407] blk_mq_run_hw_queue+0x13b/0x2b0 [ 1446.702593] blk_mq_sched_insert_requests+0x1de/0x390 [ 1446.703309] blk_mq_flush_plug_list+0x4b4/0x760 [ 1446.705408] blk_flush_plug_list+0x2c5/0x480 [ 1446.708471] blk_finish_plug+0x55/0xa0 [ 1446.708980] blk_throtl_dispatch_work_fn+0x23b/0x2e0 [ 1446.711236] process_one_work+0x6d4/0xfe0 [ 1446.711778] worker_thread+0x91/0xc80 [ 1446.713400] kthread+0x32d/0x3f0 [ 1446.714362] ret_from_fork+0x1f/0x30 [ 1446.714846] [ 1446.715062] Allocated by task 1: [ 1446.715509] kasan_save_stack+0x19/0x40 [ 1446.716026] __kasan_kmalloc.constprop.1+0xc1/0xd0 [ 1446.716673] blk_mq_init_tags+0x6d/0x330 [ 1446.717207] blk_mq_alloc_rq_map+0x50/0x1c0 [ 1446.717769] __blk_mq_alloc_map_and_request+0xe5/0x320 [ 1446.718459] blk_mq_alloc_tag_set+0x679/0xdc0 [ 1446.719050] scsi_add_host_with_dma.cold.3+0xa0/0x5db [ 1446.719736] virtscsi_probe+0x7bf/0xbd0 [ 1446.720265] virtio_dev_probe+0x402/0x6c0 [ 1446.720808] really_probe+0x276/0xde0 [ 1446.721320] driver_probe_device+0x267/0x3d0 [ 1446.721892] device_driver_attach+0xfe/0x140 [ 1446.722491] __driver_attach+0x13a/0x2c0 [ 1446.723037] bus_for_each_dev+0x146/0x1c0 [ 1446.723603] bus_add_driver+0x3fc/0x680 [ 1446.724145] driver_register+0x1c0/0x400 [ 1446.724693] init+0xa2/0xe8 [ 1446.725091] do_one_initcall+0x9e/0x310 [ 1446.725626] kernel_init_freeable+0xc56/0xcb9 [ 1446.726231] kernel_init+0x11/0x198 [ 1446.726714] ret_from_fork+0x1f/0x30 [ 1446.727212] [ 1446.727433] Freed by task 26992: [ 1446.727882] kasan_save_stack+0x19/0x40 [ 1446.728420] kasan_set_track+0x1c/0x30 [ 1446.728943] kasan_set_free_info+0x1b/0x30 [ 1446.729517] __kasan_slab_free+0x111/0x160 [ 1446.730084] kfree+0xb8/0x520 [ 1446.730507] blk_mq_free_map_and_requests+0x10b/0x1b0 [ 1446.731206] blk_mq_realloc_hw_ctxs+0x8cb/0x15b0 [ 1446.731844] blk_mq_init_allocated_queue+0x374/0x1380 [ 1446.732540] blk_mq_init_queue_data+0x7f/0xd0 [ 1446.733155] scsi_mq_alloc_queue+0x45/0x170 [ 1446.733730] scsi_alloc_sdev+0x73c/0xb20 [ 1446.734281] scsi_probe_and_add_lun+0x9a6/0x2d90 [ 1446.734916] __scsi_scan_target+0x208/0xc50 [ 1446.735500] scsi_scan_channel.part.3+0x113/0x170 [ 1446.736149] scsi_scan_host_selected+0x25a/0x360 [ 1446.736783] store_scan+0x290/0x2d0 [ 1446.737275] dev_attr_store+0x55/0x80 [ 1446.737782] sysfs_kf_write+0x132/0x190 [ 1446.738313] kernfs_fop_write_iter+0x319/0x4b0 [ 1446.738921] new_sync_write+0x40e/0x5c0 [ 1446.739429] vfs_write+0x519/0x720 [ 1446.739877] ksys_write+0xf8/0x1f0 [ 1446.740332] do_syscall_64+0x2d/0x40 [ 1446.740802] entry_SYSCALL_64_after_hwframe+0x44/0xa9 [ 1446.741462] [ 1446.741670] The buggy address belongs to the object at ffff8880185afd00 [ 1446.741670] which belongs to the cache kmalloc-256 of size 256 [ 1446.743276] The buggy address is located 16 bytes inside of [ 1446.743276] 256-byte region [ffff8880185afd00, ffff8880185afe00) [ 1446.744765] The buggy address belongs to the page: [ 1446.745416] page:ffffea0000616b00 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x185ac [ 1446.746694] head:ffffea0000616b00 order:2 compound_mapcount:0 compound_pincount:0 [ 1446.747719] flags: 0x1fffff80010200(slab|head) [ 1446.748337] raw: 001fffff80010200 ffffea00006a3208 ffffea000061bf08 ffff88801004f240 [ 1446.749404] raw: 0000000000000000 0000000000100010 00000001ffffffff 0000000000000000 [ 1446.750455] page dumped because: kasan: bad access detected [ 1446.751227] [ 1446.751445] Memory state around the buggy address: [ 1446.752102] ffff8880185afc00: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [ 1446.753090] ffff8880185afc80: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [ 1446.754079] >ffff8880185afd00: fa fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 1446.755065] ^ [ 1446.755589] ffff8880185afd80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 1446.756574] ffff8880185afe00: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [ 1446.757566] ================================================================== Flag 'BLK_MQ_F_TAG_QUEUE_SHARED' will be set if the second device on the same host initializes it's queue successfully. However, if the second device failed to allocate memory in blk_mq_alloc_and_init_hctx() from blk_mq_realloc_hw_ctxs() from blk_mq_init_allocated_queue(), __blk_mq_free_map_and_rqs() will be called on error path, and if 'BLK_MQ_TAG_HCTX_SHARED' is not set, 'tag_set->tags' will be freed while it's still used by the first device. To fix this issue we move release newly allocated hardware context from blk_mq_realloc_hw_ctxs to __blk_mq_update_nr_hw_queues. As there is needn't to release hardware context in blk_mq_init_allocated_queue. Fixes: 868f2f0b7206 ("blk-mq: dynamic h/w context count") Signed-off-by: Ye Bin <yebin10@huawei.com> Signed-off-by: Yu Kuai <yukuai3@huawei.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Link: https://lore.kernel.org/r/20211108074019.1058843-1-yebin10@huawei.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-08 07:40:19 +00:00
int i = prev_nr_hw_queues;
pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
nr_hw_queues, prev_nr_hw_queues);
blk-mq: don't free tags if the tag_set is used by other device in queue initialztion We got UAF report on v5.10 as follows: [ 1446.674930] ================================================================== [ 1446.675970] BUG: KASAN: use-after-free in blk_mq_get_driver_tag+0x9a4/0xa90 [ 1446.676902] Read of size 8 at addr ffff8880185afd10 by task kworker/1:2/12348 [ 1446.677851] [ 1446.678073] CPU: 1 PID: 12348 Comm: kworker/1:2 Not tainted 5.10.0-10177-gc9c81b1e346a #2 [ 1446.679168] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014 [ 1446.680692] Workqueue: kthrotld blk_throtl_dispatch_work_fn [ 1446.681448] Call Trace: [ 1446.681800] dump_stack+0x9b/0xce [ 1446.682916] print_address_description.constprop.6+0x3e/0x60 [ 1446.685999] kasan_report.cold.9+0x22/0x3a [ 1446.687186] blk_mq_get_driver_tag+0x9a4/0xa90 [ 1446.687785] blk_mq_dispatch_rq_list+0x21a/0x1d40 [ 1446.692576] __blk_mq_do_dispatch_sched+0x394/0x830 [ 1446.695758] __blk_mq_sched_dispatch_requests+0x398/0x4f0 [ 1446.698279] blk_mq_sched_dispatch_requests+0xdf/0x140 [ 1446.698967] __blk_mq_run_hw_queue+0xc0/0x270 [ 1446.699561] __blk_mq_delay_run_hw_queue+0x4cc/0x550 [ 1446.701407] blk_mq_run_hw_queue+0x13b/0x2b0 [ 1446.702593] blk_mq_sched_insert_requests+0x1de/0x390 [ 1446.703309] blk_mq_flush_plug_list+0x4b4/0x760 [ 1446.705408] blk_flush_plug_list+0x2c5/0x480 [ 1446.708471] blk_finish_plug+0x55/0xa0 [ 1446.708980] blk_throtl_dispatch_work_fn+0x23b/0x2e0 [ 1446.711236] process_one_work+0x6d4/0xfe0 [ 1446.711778] worker_thread+0x91/0xc80 [ 1446.713400] kthread+0x32d/0x3f0 [ 1446.714362] ret_from_fork+0x1f/0x30 [ 1446.714846] [ 1446.715062] Allocated by task 1: [ 1446.715509] kasan_save_stack+0x19/0x40 [ 1446.716026] __kasan_kmalloc.constprop.1+0xc1/0xd0 [ 1446.716673] blk_mq_init_tags+0x6d/0x330 [ 1446.717207] blk_mq_alloc_rq_map+0x50/0x1c0 [ 1446.717769] __blk_mq_alloc_map_and_request+0xe5/0x320 [ 1446.718459] blk_mq_alloc_tag_set+0x679/0xdc0 [ 1446.719050] scsi_add_host_with_dma.cold.3+0xa0/0x5db [ 1446.719736] virtscsi_probe+0x7bf/0xbd0 [ 1446.720265] virtio_dev_probe+0x402/0x6c0 [ 1446.720808] really_probe+0x276/0xde0 [ 1446.721320] driver_probe_device+0x267/0x3d0 [ 1446.721892] device_driver_attach+0xfe/0x140 [ 1446.722491] __driver_attach+0x13a/0x2c0 [ 1446.723037] bus_for_each_dev+0x146/0x1c0 [ 1446.723603] bus_add_driver+0x3fc/0x680 [ 1446.724145] driver_register+0x1c0/0x400 [ 1446.724693] init+0xa2/0xe8 [ 1446.725091] do_one_initcall+0x9e/0x310 [ 1446.725626] kernel_init_freeable+0xc56/0xcb9 [ 1446.726231] kernel_init+0x11/0x198 [ 1446.726714] ret_from_fork+0x1f/0x30 [ 1446.727212] [ 1446.727433] Freed by task 26992: [ 1446.727882] kasan_save_stack+0x19/0x40 [ 1446.728420] kasan_set_track+0x1c/0x30 [ 1446.728943] kasan_set_free_info+0x1b/0x30 [ 1446.729517] __kasan_slab_free+0x111/0x160 [ 1446.730084] kfree+0xb8/0x520 [ 1446.730507] blk_mq_free_map_and_requests+0x10b/0x1b0 [ 1446.731206] blk_mq_realloc_hw_ctxs+0x8cb/0x15b0 [ 1446.731844] blk_mq_init_allocated_queue+0x374/0x1380 [ 1446.732540] blk_mq_init_queue_data+0x7f/0xd0 [ 1446.733155] scsi_mq_alloc_queue+0x45/0x170 [ 1446.733730] scsi_alloc_sdev+0x73c/0xb20 [ 1446.734281] scsi_probe_and_add_lun+0x9a6/0x2d90 [ 1446.734916] __scsi_scan_target+0x208/0xc50 [ 1446.735500] scsi_scan_channel.part.3+0x113/0x170 [ 1446.736149] scsi_scan_host_selected+0x25a/0x360 [ 1446.736783] store_scan+0x290/0x2d0 [ 1446.737275] dev_attr_store+0x55/0x80 [ 1446.737782] sysfs_kf_write+0x132/0x190 [ 1446.738313] kernfs_fop_write_iter+0x319/0x4b0 [ 1446.738921] new_sync_write+0x40e/0x5c0 [ 1446.739429] vfs_write+0x519/0x720 [ 1446.739877] ksys_write+0xf8/0x1f0 [ 1446.740332] do_syscall_64+0x2d/0x40 [ 1446.740802] entry_SYSCALL_64_after_hwframe+0x44/0xa9 [ 1446.741462] [ 1446.741670] The buggy address belongs to the object at ffff8880185afd00 [ 1446.741670] which belongs to the cache kmalloc-256 of size 256 [ 1446.743276] The buggy address is located 16 bytes inside of [ 1446.743276] 256-byte region [ffff8880185afd00, ffff8880185afe00) [ 1446.744765] The buggy address belongs to the page: [ 1446.745416] page:ffffea0000616b00 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x185ac [ 1446.746694] head:ffffea0000616b00 order:2 compound_mapcount:0 compound_pincount:0 [ 1446.747719] flags: 0x1fffff80010200(slab|head) [ 1446.748337] raw: 001fffff80010200 ffffea00006a3208 ffffea000061bf08 ffff88801004f240 [ 1446.749404] raw: 0000000000000000 0000000000100010 00000001ffffffff 0000000000000000 [ 1446.750455] page dumped because: kasan: bad access detected [ 1446.751227] [ 1446.751445] Memory state around the buggy address: [ 1446.752102] ffff8880185afc00: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [ 1446.753090] ffff8880185afc80: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [ 1446.754079] >ffff8880185afd00: fa fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 1446.755065] ^ [ 1446.755589] ffff8880185afd80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 1446.756574] ffff8880185afe00: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [ 1446.757566] ================================================================== Flag 'BLK_MQ_F_TAG_QUEUE_SHARED' will be set if the second device on the same host initializes it's queue successfully. However, if the second device failed to allocate memory in blk_mq_alloc_and_init_hctx() from blk_mq_realloc_hw_ctxs() from blk_mq_init_allocated_queue(), __blk_mq_free_map_and_rqs() will be called on error path, and if 'BLK_MQ_TAG_HCTX_SHARED' is not set, 'tag_set->tags' will be freed while it's still used by the first device. To fix this issue we move release newly allocated hardware context from blk_mq_realloc_hw_ctxs to __blk_mq_update_nr_hw_queues. As there is needn't to release hardware context in blk_mq_init_allocated_queue. Fixes: 868f2f0b7206 ("blk-mq: dynamic h/w context count") Signed-off-by: Ye Bin <yebin10@huawei.com> Signed-off-by: Yu Kuai <yukuai3@huawei.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Link: https://lore.kernel.org/r/20211108074019.1058843-1-yebin10@huawei.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-08 07:40:19 +00:00
for (; i < set->nr_hw_queues; i++)
__blk_mq_free_map_and_rqs(set, i);
set->nr_hw_queues = prev_nr_hw_queues;
goto fallback;
}
lim = queue_limits_start_update(q);
if (blk_mq_can_poll(set))
lim.features |= BLK_FEAT_POLL;
else
lim.features &= ~BLK_FEAT_POLL;
if (queue_limits_commit_update(q, &lim) < 0)
pr_warn("updating the poll flag failed\n");
blk_mq_map_swqueue(q);
}
reregister:
list_for_each_entry(q, &set->tag_list, tag_set_list) {
blk_mq_sysfs_register_hctxs(q);
blk_mq_debugfs_register_hctxs(q);
}
switch_back:
list_for_each_entry(q, &set->tag_list, tag_set_list)
blk_mq_elv_switch_back(&head, q);
list_for_each_entry(q, &set->tag_list, tag_set_list)
blk_mq_unfreeze_queue(q);
/* Free the excess tags when nr_hw_queues shrink. */
for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
__blk_mq_free_map_and_rqs(set, i);
}
void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
{
mutex_lock(&set->tag_list_lock);
__blk_mq_update_nr_hw_queues(set, nr_hw_queues);
mutex_unlock(&set->tag_list_lock);
}
EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
struct io_comp_batch *iob, unsigned int flags)
{
long state = get_current_state();
int ret;
do {
ret = q->mq_ops->poll(hctx, iob);
if (ret > 0) {
__set_current_state(TASK_RUNNING);
return ret;
}
if (signal_pending_state(state, current))
__set_current_state(TASK_RUNNING);
if (task_is_running(current))
return 1;
if (ret < 0 || (flags & BLK_POLL_ONESHOT))
break;
cpu_relax();
} while (!need_resched());
__set_current_state(TASK_RUNNING);
return 0;
}
int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
struct io_comp_batch *iob, unsigned int flags)
{
struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
return blk_hctx_poll(q, hctx, iob, flags);
}
int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
unsigned int poll_flags)
{
struct request_queue *q = rq->q;
int ret;
if (!blk_rq_is_poll(rq))
return 0;
if (!percpu_ref_tryget(&q->q_usage_counter))
return 0;
ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
blk_queue_exit(q);
return ret;
}
EXPORT_SYMBOL_GPL(blk_rq_poll);
unsigned int blk_mq_rq_cpu(struct request *rq)
{
return rq->mq_ctx->cpu;
}
EXPORT_SYMBOL(blk_mq_rq_cpu);
blk-mq: cancel blk-mq dispatch work in both blk_cleanup_queue and disk_release() For avoiding to slow down queue destroy, we don't call blk_mq_quiesce_queue() in blk_cleanup_queue(), instead of delaying to cancel dispatch work in blk_release_queue(). However, this way has caused kernel oops[1], reported by Changhui. The log shows that scsi_device can be freed before running blk_release_queue(), which is expected too since scsi_device is released after the scsi disk is closed and the scsi_device is removed. Fixes the issue by canceling blk-mq dispatch work in both blk_cleanup_queue() and disk_release(): 1) when disk_release() is run, the disk has been closed, and any sync dispatch activities have been done, so canceling dispatch work is enough to quiesce filesystem I/O dispatch activity. 2) in blk_cleanup_queue(), we only focus on passthrough request, and passthrough request is always explicitly allocated & freed by its caller, so once queue is frozen, all sync dispatch activity for passthrough request has been done, then it is enough to just cancel dispatch work for avoiding any dispatch activity. [1] kernel panic log [12622.769416] BUG: kernel NULL pointer dereference, address: 0000000000000300 [12622.777186] #PF: supervisor read access in kernel mode [12622.782918] #PF: error_code(0x0000) - not-present page [12622.788649] PGD 0 P4D 0 [12622.791474] Oops: 0000 [#1] PREEMPT SMP PTI [12622.796138] CPU: 10 PID: 744 Comm: kworker/10:1H Kdump: loaded Not tainted 5.15.0+ #1 [12622.804877] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 1.5.4 10/002/2015 [12622.813321] Workqueue: kblockd blk_mq_run_work_fn [12622.818572] RIP: 0010:sbitmap_get+0x75/0x190 [12622.823336] Code: 85 80 00 00 00 41 8b 57 08 85 d2 0f 84 b1 00 00 00 45 31 e4 48 63 cd 48 8d 1c 49 48 c1 e3 06 49 03 5f 10 4c 8d 6b 40 83 f0 01 <48> 8b 33 44 89 f2 4c 89 ef 0f b6 c8 e8 fa f3 ff ff 83 f8 ff 75 58 [12622.844290] RSP: 0018:ffffb00a446dbd40 EFLAGS: 00010202 [12622.850120] RAX: 0000000000000001 RBX: 0000000000000300 RCX: 0000000000000004 [12622.858082] RDX: 0000000000000006 RSI: 0000000000000082 RDI: ffffa0b7a2dfe030 [12622.866042] RBP: 0000000000000004 R08: 0000000000000001 R09: ffffa0b742721334 [12622.874003] R10: 0000000000000008 R11: 0000000000000008 R12: 0000000000000000 [12622.881964] R13: 0000000000000340 R14: 0000000000000000 R15: ffffa0b7a2dfe030 [12622.889926] FS: 0000000000000000(0000) GS:ffffa0baafb40000(0000) knlGS:0000000000000000 [12622.898956] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [12622.905367] CR2: 0000000000000300 CR3: 0000000641210001 CR4: 00000000001706e0 [12622.913328] Call Trace: [12622.916055] <TASK> [12622.918394] scsi_mq_get_budget+0x1a/0x110 [12622.922969] __blk_mq_do_dispatch_sched+0x1d4/0x320 [12622.928404] ? pick_next_task_fair+0x39/0x390 [12622.933268] __blk_mq_sched_dispatch_requests+0xf4/0x140 [12622.939194] blk_mq_sched_dispatch_requests+0x30/0x60 [12622.944829] __blk_mq_run_hw_queue+0x30/0xa0 [12622.949593] process_one_work+0x1e8/0x3c0 [12622.954059] worker_thread+0x50/0x3b0 [12622.958144] ? rescuer_thread+0x370/0x370 [12622.962616] kthread+0x158/0x180 [12622.966218] ? set_kthread_struct+0x40/0x40 [12622.970884] ret_from_fork+0x22/0x30 [12622.974875] </TASK> [12622.977309] Modules linked in: scsi_debug rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache netfs sunrpc dm_multipath intel_rapl_msr intel_rapl_common dell_wmi_descriptor sb_edac rfkill video x86_pkg_temp_thermal intel_powerclamp dcdbas coretemp kvm_intel kvm mgag200 irqbypass i2c_algo_bit rapl drm_kms_helper ipmi_ssif intel_cstate intel_uncore syscopyarea sysfillrect sysimgblt fb_sys_fops pcspkr cec mei_me lpc_ich mei ipmi_si ipmi_devintf ipmi_msghandler acpi_power_meter drm fuse xfs libcrc32c sr_mod cdrom sd_mod t10_pi sg ixgbe ahci libahci crct10dif_pclmul crc32_pclmul crc32c_intel libata megaraid_sas ghash_clmulni_intel tg3 wdat_wdt mdio dca wmi dm_mirror dm_region_hash dm_log dm_mod [last unloaded: scsi_debug] Reported-by: ChanghuiZhong <czhong@redhat.com> Cc: Christoph Hellwig <hch@lst.de> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Bart Van Assche <bvanassche@acm.org> Cc: linux-scsi@vger.kernel.org Signed-off-by: Ming Lei <ming.lei@redhat.com> Link: https://lore.kernel.org/r/20211116014343.610501-1-ming.lei@redhat.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-16 01:43:43 +00:00
void blk_mq_cancel_work_sync(struct request_queue *q)
{
struct blk_mq_hw_ctx *hctx;
unsigned long i;
blk-mq: cancel blk-mq dispatch work in both blk_cleanup_queue and disk_release() For avoiding to slow down queue destroy, we don't call blk_mq_quiesce_queue() in blk_cleanup_queue(), instead of delaying to cancel dispatch work in blk_release_queue(). However, this way has caused kernel oops[1], reported by Changhui. The log shows that scsi_device can be freed before running blk_release_queue(), which is expected too since scsi_device is released after the scsi disk is closed and the scsi_device is removed. Fixes the issue by canceling blk-mq dispatch work in both blk_cleanup_queue() and disk_release(): 1) when disk_release() is run, the disk has been closed, and any sync dispatch activities have been done, so canceling dispatch work is enough to quiesce filesystem I/O dispatch activity. 2) in blk_cleanup_queue(), we only focus on passthrough request, and passthrough request is always explicitly allocated & freed by its caller, so once queue is frozen, all sync dispatch activity for passthrough request has been done, then it is enough to just cancel dispatch work for avoiding any dispatch activity. [1] kernel panic log [12622.769416] BUG: kernel NULL pointer dereference, address: 0000000000000300 [12622.777186] #PF: supervisor read access in kernel mode [12622.782918] #PF: error_code(0x0000) - not-present page [12622.788649] PGD 0 P4D 0 [12622.791474] Oops: 0000 [#1] PREEMPT SMP PTI [12622.796138] CPU: 10 PID: 744 Comm: kworker/10:1H Kdump: loaded Not tainted 5.15.0+ #1 [12622.804877] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 1.5.4 10/002/2015 [12622.813321] Workqueue: kblockd blk_mq_run_work_fn [12622.818572] RIP: 0010:sbitmap_get+0x75/0x190 [12622.823336] Code: 85 80 00 00 00 41 8b 57 08 85 d2 0f 84 b1 00 00 00 45 31 e4 48 63 cd 48 8d 1c 49 48 c1 e3 06 49 03 5f 10 4c 8d 6b 40 83 f0 01 <48> 8b 33 44 89 f2 4c 89 ef 0f b6 c8 e8 fa f3 ff ff 83 f8 ff 75 58 [12622.844290] RSP: 0018:ffffb00a446dbd40 EFLAGS: 00010202 [12622.850120] RAX: 0000000000000001 RBX: 0000000000000300 RCX: 0000000000000004 [12622.858082] RDX: 0000000000000006 RSI: 0000000000000082 RDI: ffffa0b7a2dfe030 [12622.866042] RBP: 0000000000000004 R08: 0000000000000001 R09: ffffa0b742721334 [12622.874003] R10: 0000000000000008 R11: 0000000000000008 R12: 0000000000000000 [12622.881964] R13: 0000000000000340 R14: 0000000000000000 R15: ffffa0b7a2dfe030 [12622.889926] FS: 0000000000000000(0000) GS:ffffa0baafb40000(0000) knlGS:0000000000000000 [12622.898956] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [12622.905367] CR2: 0000000000000300 CR3: 0000000641210001 CR4: 00000000001706e0 [12622.913328] Call Trace: [12622.916055] <TASK> [12622.918394] scsi_mq_get_budget+0x1a/0x110 [12622.922969] __blk_mq_do_dispatch_sched+0x1d4/0x320 [12622.928404] ? pick_next_task_fair+0x39/0x390 [12622.933268] __blk_mq_sched_dispatch_requests+0xf4/0x140 [12622.939194] blk_mq_sched_dispatch_requests+0x30/0x60 [12622.944829] __blk_mq_run_hw_queue+0x30/0xa0 [12622.949593] process_one_work+0x1e8/0x3c0 [12622.954059] worker_thread+0x50/0x3b0 [12622.958144] ? rescuer_thread+0x370/0x370 [12622.962616] kthread+0x158/0x180 [12622.966218] ? set_kthread_struct+0x40/0x40 [12622.970884] ret_from_fork+0x22/0x30 [12622.974875] </TASK> [12622.977309] Modules linked in: scsi_debug rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache netfs sunrpc dm_multipath intel_rapl_msr intel_rapl_common dell_wmi_descriptor sb_edac rfkill video x86_pkg_temp_thermal intel_powerclamp dcdbas coretemp kvm_intel kvm mgag200 irqbypass i2c_algo_bit rapl drm_kms_helper ipmi_ssif intel_cstate intel_uncore syscopyarea sysfillrect sysimgblt fb_sys_fops pcspkr cec mei_me lpc_ich mei ipmi_si ipmi_devintf ipmi_msghandler acpi_power_meter drm fuse xfs libcrc32c sr_mod cdrom sd_mod t10_pi sg ixgbe ahci libahci crct10dif_pclmul crc32_pclmul crc32c_intel libata megaraid_sas ghash_clmulni_intel tg3 wdat_wdt mdio dca wmi dm_mirror dm_region_hash dm_log dm_mod [last unloaded: scsi_debug] Reported-by: ChanghuiZhong <czhong@redhat.com> Cc: Christoph Hellwig <hch@lst.de> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Bart Van Assche <bvanassche@acm.org> Cc: linux-scsi@vger.kernel.org Signed-off-by: Ming Lei <ming.lei@redhat.com> Link: https://lore.kernel.org/r/20211116014343.610501-1-ming.lei@redhat.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-16 01:43:43 +00:00
cancel_delayed_work_sync(&q->requeue_work);
blk-mq: cancel blk-mq dispatch work in both blk_cleanup_queue and disk_release() For avoiding to slow down queue destroy, we don't call blk_mq_quiesce_queue() in blk_cleanup_queue(), instead of delaying to cancel dispatch work in blk_release_queue(). However, this way has caused kernel oops[1], reported by Changhui. The log shows that scsi_device can be freed before running blk_release_queue(), which is expected too since scsi_device is released after the scsi disk is closed and the scsi_device is removed. Fixes the issue by canceling blk-mq dispatch work in both blk_cleanup_queue() and disk_release(): 1) when disk_release() is run, the disk has been closed, and any sync dispatch activities have been done, so canceling dispatch work is enough to quiesce filesystem I/O dispatch activity. 2) in blk_cleanup_queue(), we only focus on passthrough request, and passthrough request is always explicitly allocated & freed by its caller, so once queue is frozen, all sync dispatch activity for passthrough request has been done, then it is enough to just cancel dispatch work for avoiding any dispatch activity. [1] kernel panic log [12622.769416] BUG: kernel NULL pointer dereference, address: 0000000000000300 [12622.777186] #PF: supervisor read access in kernel mode [12622.782918] #PF: error_code(0x0000) - not-present page [12622.788649] PGD 0 P4D 0 [12622.791474] Oops: 0000 [#1] PREEMPT SMP PTI [12622.796138] CPU: 10 PID: 744 Comm: kworker/10:1H Kdump: loaded Not tainted 5.15.0+ #1 [12622.804877] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 1.5.4 10/002/2015 [12622.813321] Workqueue: kblockd blk_mq_run_work_fn [12622.818572] RIP: 0010:sbitmap_get+0x75/0x190 [12622.823336] Code: 85 80 00 00 00 41 8b 57 08 85 d2 0f 84 b1 00 00 00 45 31 e4 48 63 cd 48 8d 1c 49 48 c1 e3 06 49 03 5f 10 4c 8d 6b 40 83 f0 01 <48> 8b 33 44 89 f2 4c 89 ef 0f b6 c8 e8 fa f3 ff ff 83 f8 ff 75 58 [12622.844290] RSP: 0018:ffffb00a446dbd40 EFLAGS: 00010202 [12622.850120] RAX: 0000000000000001 RBX: 0000000000000300 RCX: 0000000000000004 [12622.858082] RDX: 0000000000000006 RSI: 0000000000000082 RDI: ffffa0b7a2dfe030 [12622.866042] RBP: 0000000000000004 R08: 0000000000000001 R09: ffffa0b742721334 [12622.874003] R10: 0000000000000008 R11: 0000000000000008 R12: 0000000000000000 [12622.881964] R13: 0000000000000340 R14: 0000000000000000 R15: ffffa0b7a2dfe030 [12622.889926] FS: 0000000000000000(0000) GS:ffffa0baafb40000(0000) knlGS:0000000000000000 [12622.898956] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [12622.905367] CR2: 0000000000000300 CR3: 0000000641210001 CR4: 00000000001706e0 [12622.913328] Call Trace: [12622.916055] <TASK> [12622.918394] scsi_mq_get_budget+0x1a/0x110 [12622.922969] __blk_mq_do_dispatch_sched+0x1d4/0x320 [12622.928404] ? pick_next_task_fair+0x39/0x390 [12622.933268] __blk_mq_sched_dispatch_requests+0xf4/0x140 [12622.939194] blk_mq_sched_dispatch_requests+0x30/0x60 [12622.944829] __blk_mq_run_hw_queue+0x30/0xa0 [12622.949593] process_one_work+0x1e8/0x3c0 [12622.954059] worker_thread+0x50/0x3b0 [12622.958144] ? rescuer_thread+0x370/0x370 [12622.962616] kthread+0x158/0x180 [12622.966218] ? set_kthread_struct+0x40/0x40 [12622.970884] ret_from_fork+0x22/0x30 [12622.974875] </TASK> [12622.977309] Modules linked in: scsi_debug rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache netfs sunrpc dm_multipath intel_rapl_msr intel_rapl_common dell_wmi_descriptor sb_edac rfkill video x86_pkg_temp_thermal intel_powerclamp dcdbas coretemp kvm_intel kvm mgag200 irqbypass i2c_algo_bit rapl drm_kms_helper ipmi_ssif intel_cstate intel_uncore syscopyarea sysfillrect sysimgblt fb_sys_fops pcspkr cec mei_me lpc_ich mei ipmi_si ipmi_devintf ipmi_msghandler acpi_power_meter drm fuse xfs libcrc32c sr_mod cdrom sd_mod t10_pi sg ixgbe ahci libahci crct10dif_pclmul crc32_pclmul crc32c_intel libata megaraid_sas ghash_clmulni_intel tg3 wdat_wdt mdio dca wmi dm_mirror dm_region_hash dm_log dm_mod [last unloaded: scsi_debug] Reported-by: ChanghuiZhong <czhong@redhat.com> Cc: Christoph Hellwig <hch@lst.de> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Bart Van Assche <bvanassche@acm.org> Cc: linux-scsi@vger.kernel.org Signed-off-by: Ming Lei <ming.lei@redhat.com> Link: https://lore.kernel.org/r/20211116014343.610501-1-ming.lei@redhat.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-16 01:43:43 +00:00
queue_for_each_hw_ctx(q, hctx, i)
cancel_delayed_work_sync(&hctx->run_work);
blk-mq: cancel blk-mq dispatch work in both blk_cleanup_queue and disk_release() For avoiding to slow down queue destroy, we don't call blk_mq_quiesce_queue() in blk_cleanup_queue(), instead of delaying to cancel dispatch work in blk_release_queue(). However, this way has caused kernel oops[1], reported by Changhui. The log shows that scsi_device can be freed before running blk_release_queue(), which is expected too since scsi_device is released after the scsi disk is closed and the scsi_device is removed. Fixes the issue by canceling blk-mq dispatch work in both blk_cleanup_queue() and disk_release(): 1) when disk_release() is run, the disk has been closed, and any sync dispatch activities have been done, so canceling dispatch work is enough to quiesce filesystem I/O dispatch activity. 2) in blk_cleanup_queue(), we only focus on passthrough request, and passthrough request is always explicitly allocated & freed by its caller, so once queue is frozen, all sync dispatch activity for passthrough request has been done, then it is enough to just cancel dispatch work for avoiding any dispatch activity. [1] kernel panic log [12622.769416] BUG: kernel NULL pointer dereference, address: 0000000000000300 [12622.777186] #PF: supervisor read access in kernel mode [12622.782918] #PF: error_code(0x0000) - not-present page [12622.788649] PGD 0 P4D 0 [12622.791474] Oops: 0000 [#1] PREEMPT SMP PTI [12622.796138] CPU: 10 PID: 744 Comm: kworker/10:1H Kdump: loaded Not tainted 5.15.0+ #1 [12622.804877] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 1.5.4 10/002/2015 [12622.813321] Workqueue: kblockd blk_mq_run_work_fn [12622.818572] RIP: 0010:sbitmap_get+0x75/0x190 [12622.823336] Code: 85 80 00 00 00 41 8b 57 08 85 d2 0f 84 b1 00 00 00 45 31 e4 48 63 cd 48 8d 1c 49 48 c1 e3 06 49 03 5f 10 4c 8d 6b 40 83 f0 01 <48> 8b 33 44 89 f2 4c 89 ef 0f b6 c8 e8 fa f3 ff ff 83 f8 ff 75 58 [12622.844290] RSP: 0018:ffffb00a446dbd40 EFLAGS: 00010202 [12622.850120] RAX: 0000000000000001 RBX: 0000000000000300 RCX: 0000000000000004 [12622.858082] RDX: 0000000000000006 RSI: 0000000000000082 RDI: ffffa0b7a2dfe030 [12622.866042] RBP: 0000000000000004 R08: 0000000000000001 R09: ffffa0b742721334 [12622.874003] R10: 0000000000000008 R11: 0000000000000008 R12: 0000000000000000 [12622.881964] R13: 0000000000000340 R14: 0000000000000000 R15: ffffa0b7a2dfe030 [12622.889926] FS: 0000000000000000(0000) GS:ffffa0baafb40000(0000) knlGS:0000000000000000 [12622.898956] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [12622.905367] CR2: 0000000000000300 CR3: 0000000641210001 CR4: 00000000001706e0 [12622.913328] Call Trace: [12622.916055] <TASK> [12622.918394] scsi_mq_get_budget+0x1a/0x110 [12622.922969] __blk_mq_do_dispatch_sched+0x1d4/0x320 [12622.928404] ? pick_next_task_fair+0x39/0x390 [12622.933268] __blk_mq_sched_dispatch_requests+0xf4/0x140 [12622.939194] blk_mq_sched_dispatch_requests+0x30/0x60 [12622.944829] __blk_mq_run_hw_queue+0x30/0xa0 [12622.949593] process_one_work+0x1e8/0x3c0 [12622.954059] worker_thread+0x50/0x3b0 [12622.958144] ? rescuer_thread+0x370/0x370 [12622.962616] kthread+0x158/0x180 [12622.966218] ? set_kthread_struct+0x40/0x40 [12622.970884] ret_from_fork+0x22/0x30 [12622.974875] </TASK> [12622.977309] Modules linked in: scsi_debug rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache netfs sunrpc dm_multipath intel_rapl_msr intel_rapl_common dell_wmi_descriptor sb_edac rfkill video x86_pkg_temp_thermal intel_powerclamp dcdbas coretemp kvm_intel kvm mgag200 irqbypass i2c_algo_bit rapl drm_kms_helper ipmi_ssif intel_cstate intel_uncore syscopyarea sysfillrect sysimgblt fb_sys_fops pcspkr cec mei_me lpc_ich mei ipmi_si ipmi_devintf ipmi_msghandler acpi_power_meter drm fuse xfs libcrc32c sr_mod cdrom sd_mod t10_pi sg ixgbe ahci libahci crct10dif_pclmul crc32_pclmul crc32c_intel libata megaraid_sas ghash_clmulni_intel tg3 wdat_wdt mdio dca wmi dm_mirror dm_region_hash dm_log dm_mod [last unloaded: scsi_debug] Reported-by: ChanghuiZhong <czhong@redhat.com> Cc: Christoph Hellwig <hch@lst.de> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Bart Van Assche <bvanassche@acm.org> Cc: linux-scsi@vger.kernel.org Signed-off-by: Ming Lei <ming.lei@redhat.com> Link: https://lore.kernel.org/r/20211116014343.610501-1-ming.lei@redhat.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-11-16 01:43:43 +00:00
}
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
static int __init blk_mq_init(void)
{
int i;
for_each_possible_cpu(i)
init_llist_head(&per_cpu(blk_cpu_done, i));
for_each_possible_cpu(i)
INIT_CSD(&per_cpu(blk_cpu_csd, i),
__blk_mq_complete_request_remote, NULL);
open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
"block/softirq:dead", NULL,
blk_softirq_cpu_dead);
cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
blk_mq_hctx_notify_dead);
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
blk_mq_hctx_notify_online,
blk_mq_hctx_notify_offline);
blk-mq: new multi-queue block IO queueing mechanism Linux currently has two models for block devices: - The classic request_fn based approach, where drivers use struct request units for IO. The block layer provides various helper functionalities to let drivers share code, things like tag management, timeout handling, queueing, etc. - The "stacked" approach, where a driver squeezes in between the block layer and IO submitter. Since this bypasses the IO stack, driver generally have to manage everything themselves. With drivers being written for new high IOPS devices, the classic request_fn based driver doesn't work well enough. The design dates back to when both SMP and high IOPS was rare. It has problems with scaling to bigger machines, and runs into scaling issues even on smaller machines when you have IOPS in the hundreds of thousands per device. The stacked approach is then most often selected as the model for the driver. But this means that everybody has to re-invent everything, and along with that we get all the problems again that the shared approach solved. This commit introduces blk-mq, block multi queue support. The design is centered around per-cpu queues for queueing IO, which then funnel down into x number of hardware submission queues. We might have a 1:1 mapping between the two, or it might be an N:M mapping. That all depends on what the hardware supports. blk-mq provides various helper functions, which include: - Scalable support for request tagging. Most devices need to be able to uniquely identify a request both in the driver and to the hardware. The tagging uses per-cpu caches for freed tags, to enable cache hot reuse. - Timeout handling without tracking request on a per-device basis. Basically the driver should be able to get a notification, if a request happens to fail. - Optional support for non 1:1 mappings between issue and submission queues. blk-mq can redirect IO completions to the desired location. - Support for per-request payloads. Drivers almost always need to associate a request structure with some driver private command structure. Drivers can tell blk-mq this at init time, and then any request handed to the driver will have the required size of memory associated with it. - Support for merging of IO, and plugging. The stacked model gets neither of these. Even for high IOPS devices, merging sequential IO reduces per-command overhead and thus increases bandwidth. For now, this is provided as a potential 3rd queueing model, with the hope being that, as it matures, it can replace both the classic and stacked model. That would get us back to having just 1 real model for block devices, leaving the stacked approach to dm/md devices (as it was originally intended). Contributions in this patch from the following people: Shaohua Li <shli@fusionio.com> Alexander Gordeev <agordeev@redhat.com> Christoph Hellwig <hch@infradead.org> Mike Christie <michaelc@cs.wisc.edu> Matias Bjorling <m@bjorling.me> Jeff Moyer <jmoyer@redhat.com> Acked-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2013-10-24 08:20:05 +00:00
return 0;
}
subsys_initcall(blk_mq_init);