linux/block/blk-iolatency.c

1085 lines
30 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
/*
* Block rq-qos base io controller
*
* This works similar to wbt with a few exceptions
*
* - It's bio based, so the latency covers the whole block layer in addition to
* the actual io.
* - We will throttle all IO that comes in here if we need to.
* - We use the mean latency over the 100ms window. This is because writes can
* be particularly fast, which could give us a false sense of the impact of
* other workloads on our protected workload.
* - By default there's no throttling, we set the queue_depth to UINT_MAX so
* that we can have as many outstanding bio's as we're allowed to. Only at
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
* throttle time do we pay attention to the actual queue depth.
*
* The hierarchy works like the cpu controller does, we track the latency at
* every configured node, and each configured node has it's own independent
* queue depth. This means that we only care about our latency targets at the
* peer level. Some group at the bottom of the hierarchy isn't going to affect
* a group at the end of some other path if we're only configred at leaf level.
*
* Consider the following
*
* root blkg
* / \
* fast (target=5ms) slow (target=10ms)
* / \ / \
* a b normal(15ms) unloved
*
* "a" and "b" have no target, but their combined io under "fast" cannot exceed
* an average latency of 5ms. If it does then we will throttle the "slow"
* group. In the case of "normal", if it exceeds its 15ms target, we will
* throttle "unloved", but nobody else.
*
* In this example "fast", "slow", and "normal" will be the only groups actually
* accounting their io latencies. We have to walk up the heirarchy to the root
* on every submit and complete so we can do the appropriate stat recording and
* adjust the queue depth of ourselves if needed.
*
* There are 2 ways we throttle IO.
*
* 1) Queue depth throttling. As we throttle down we will adjust the maximum
* number of IO's we're allowed to have in flight. This starts at (u64)-1 down
* to 1. If the group is only ever submitting IO for itself then this is the
* only way we throttle.
*
* 2) Induced delay throttling. This is for the case that a group is generating
* IO that has to be issued by the root cg to avoid priority inversion. So think
* REQ_META or REQ_SWAP. If we are already at qd == 1 and we're getting a lot
* of work done for us on behalf of the root cg and are being asked to scale
* down more then we induce a latency at userspace return. We accumulate the
* total amount of time we need to be punished by doing
*
* total_time += min_lat_nsec - actual_io_completion
*
* and then at throttle time will do
*
* throttle_time = min(total_time, NSEC_PER_SEC)
*
* This induced delay will throttle back the activity that is generating the
* root cg issued io's, wethere that's some metadata intensive operation or the
* group is using so much memory that it is pushing us into swap.
*
* Copyright (C) 2018 Josef Bacik
*/
#include <linux/kernel.h>
#include <linux/blk_types.h>
#include <linux/backing-dev.h>
#include <linux/module.h>
#include <linux/timer.h>
#include <linux/memcontrol.h>
#include <linux/sched/loadavg.h>
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
#include <linux/sched/signal.h>
#include <trace/events/block.h>
blk-iolatency: fix IO hang due to negative inflight counter Our test reported the following stack, and vmcore showed that ->inflight counter is -1. [ffffc9003fcc38d0] __schedule at ffffffff8173d95d [ffffc9003fcc3958] schedule at ffffffff8173de26 [ffffc9003fcc3970] io_schedule at ffffffff810bb6b6 [ffffc9003fcc3988] blkcg_iolatency_throttle at ffffffff813911cb [ffffc9003fcc3a20] rq_qos_throttle at ffffffff813847f3 [ffffc9003fcc3a48] blk_mq_make_request at ffffffff8137468a [ffffc9003fcc3b08] generic_make_request at ffffffff81368b49 [ffffc9003fcc3b68] submit_bio at ffffffff81368d7d [ffffc9003fcc3bb8] ext4_io_submit at ffffffffa031be00 [ext4] [ffffc9003fcc3c00] ext4_writepages at ffffffffa03163de [ext4] [ffffc9003fcc3d68] do_writepages at ffffffff811c49ae [ffffc9003fcc3d78] __filemap_fdatawrite_range at ffffffff811b6188 [ffffc9003fcc3e30] filemap_write_and_wait_range at ffffffff811b6301 [ffffc9003fcc3e60] ext4_sync_file at ffffffffa030cee8 [ext4] [ffffc9003fcc3ea8] vfs_fsync_range at ffffffff8128594b [ffffc9003fcc3ee8] do_fsync at ffffffff81285abd [ffffc9003fcc3f18] sys_fsync at ffffffff81285d50 [ffffc9003fcc3f28] do_syscall_64 at ffffffff81003c04 [ffffc9003fcc3f50] entry_SYSCALL_64_after_swapgs at ffffffff81742b8e The ->inflight counter may be negative (-1) if 1) blk-iolatency was disabled when the IO was issued, 2) blk-iolatency was enabled before this IO reached its endio, 3) the ->inflight counter is decreased from 0 to -1 in endio() In fact the hang can be easily reproduced by the below script, H=/sys/fs/cgroup/unified/ P=/sys/fs/cgroup/unified/test echo "+io" > $H/cgroup.subtree_control mkdir -p $P echo $$ > $P/cgroup.procs xfs_io -f -d -c "pwrite 0 4k" /dev/sdg echo "`cat /sys/block/sdg/dev` target=1000000" > $P/io.latency xfs_io -f -d -c "pwrite 0 4k" /dev/sdg This fixes the problem by freezing the queue so that while enabling/disabling iolatency, there is no inflight rq running. Note that quiesce_queue is not needed as this only updating iolatency configuration about which dispatching request_queue doesn't care. Signed-off-by: Liu Bo <bo.liu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-25 00:12:47 +00:00
#include <linux/blk-mq.h>
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
#include "blk-rq-qos.h"
#include "blk-stat.h"
#include "blk-cgroup.h"
#include "blk.h"
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
#define DEFAULT_SCALE_COOKIE 1000000U
static struct blkcg_policy blkcg_policy_iolatency;
struct iolatency_grp;
struct blk_iolatency {
struct rq_qos rqos;
struct timer_list timer;
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
/*
* ->enabled is the master enable switch gating the throttling logic and
* inflight tracking. The number of cgroups which have iolat enabled is
* tracked in ->enable_cnt, and ->enable is flipped on/off accordingly
* from ->enable_work with the request_queue frozen. For details, See
* blkiolatency_enable_work_fn().
*/
bool enabled;
atomic_t enable_cnt;
struct work_struct enable_work;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
};
static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
{
return container_of(rqos, struct blk_iolatency, rqos);
}
struct child_latency_info {
spinlock_t lock;
/* Last time we adjusted the scale of everybody. */
u64 last_scale_event;
/* The latency that we missed. */
u64 scale_lat;
/* Total io's from all of our children for the last summation. */
u64 nr_samples;
/* The guy who actually changed the latency numbers. */
struct iolatency_grp *scale_grp;
/* Cookie to tell if we need to scale up or down. */
atomic_t scale_cookie;
};
struct percentile_stats {
u64 total;
u64 missed;
};
struct latency_stat {
union {
struct percentile_stats ps;
struct blk_rq_stat rqs;
};
};
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
struct iolatency_grp {
struct blkg_policy_data pd;
struct latency_stat __percpu *stats;
struct latency_stat cur_stat;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
struct blk_iolatency *blkiolat;
unsigned int max_depth;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
struct rq_wait rq_wait;
atomic64_t window_start;
atomic_t scale_cookie;
u64 min_lat_nsec;
u64 cur_win_nsec;
/* total running average of our io latency. */
u64 lat_avg;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
/* Our current number of IO's for the last summation. */
u64 nr_samples;
bool ssd;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
struct child_latency_info child_lat;
};
#define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
#define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
/*
* These are the constants used to fake the fixed-point moving average
* calculation just like load average. The call to calc_load() folds
* (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg. The sampling
* window size is bucketed to try to approximately calculate average
* latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
* elapse immediately. Note, windows only elapse with IO activity. Idle
* periods extend the most recent window.
*/
#define BLKIOLATENCY_NR_EXP_FACTORS 5
#define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
(BLKIOLATENCY_NR_EXP_FACTORS - 1))
static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
2045, // exp(1/600) - 600 samples
2039, // exp(1/240) - 240 samples
2031, // exp(1/120) - 120 samples
2023, // exp(1/80) - 80 samples
2014, // exp(1/60) - 60 samples
};
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
{
return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
}
static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
{
return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
}
static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
{
return pd_to_blkg(&iolat->pd);
}
static inline void latency_stat_init(struct iolatency_grp *iolat,
struct latency_stat *stat)
{
if (iolat->ssd) {
stat->ps.total = 0;
stat->ps.missed = 0;
} else
blk_rq_stat_init(&stat->rqs);
}
static inline void latency_stat_sum(struct iolatency_grp *iolat,
struct latency_stat *sum,
struct latency_stat *stat)
{
if (iolat->ssd) {
sum->ps.total += stat->ps.total;
sum->ps.missed += stat->ps.missed;
} else
blk_rq_stat_sum(&sum->rqs, &stat->rqs);
}
static inline void latency_stat_record_time(struct iolatency_grp *iolat,
u64 req_time)
{
struct latency_stat *stat = get_cpu_ptr(iolat->stats);
if (iolat->ssd) {
if (req_time >= iolat->min_lat_nsec)
stat->ps.missed++;
stat->ps.total++;
} else
blk_rq_stat_add(&stat->rqs, req_time);
put_cpu_ptr(stat);
}
static inline bool latency_sum_ok(struct iolatency_grp *iolat,
struct latency_stat *stat)
{
if (iolat->ssd) {
u64 thresh = div64_u64(stat->ps.total, 10);
thresh = max(thresh, 1ULL);
return stat->ps.missed < thresh;
}
return stat->rqs.mean <= iolat->min_lat_nsec;
}
static inline u64 latency_stat_samples(struct iolatency_grp *iolat,
struct latency_stat *stat)
{
if (iolat->ssd)
return stat->ps.total;
return stat->rqs.nr_samples;
}
static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat,
struct latency_stat *stat)
{
int exp_idx;
if (iolat->ssd)
return;
/*
* calc_load() takes in a number stored in fixed point representation.
* Because we are using this for IO time in ns, the values stored
* are significantly larger than the FIXED_1 denominator (2048).
* Therefore, rounding errors in the calculation are negligible and
* can be ignored.
*/
exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
div64_u64(iolat->cur_win_nsec,
BLKIOLATENCY_EXP_BUCKET_SIZE));
iolat->lat_avg = calc_load(iolat->lat_avg,
iolatency_exp_factors[exp_idx],
stat->rqs.mean);
}
static void iolat_cleanup_cb(struct rq_wait *rqw, void *private_data)
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
{
atomic_dec(&rqw->inflight);
wake_up(&rqw->wait);
}
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
static bool iolat_acquire_inflight(struct rq_wait *rqw, void *private_data)
{
struct iolatency_grp *iolat = private_data;
return rq_wait_inc_below(rqw, iolat->max_depth);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
}
static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
struct iolatency_grp *iolat,
bool issue_as_root,
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
bool use_memdelay)
{
struct rq_wait *rqw = &iolat->rq_wait;
unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
if (use_delay)
blkcg_schedule_throttle(rqos->disk, use_memdelay);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
/*
* To avoid priority inversions we want to just take a slot if we are
* issuing as root. If we're being killed off there's no point in
* delaying things, we may have been killed by OOM so throttling may
* make recovery take even longer, so just let the IO's through so the
* task can go away.
*/
if (issue_as_root || fatal_signal_pending(current)) {
atomic_inc(&rqw->inflight);
return;
}
rq_qos_wait(rqw, iolat, iolat_acquire_inflight, iolat_cleanup_cb);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
}
#define SCALE_DOWN_FACTOR 2
#define SCALE_UP_FACTOR 4
static inline unsigned long scale_amount(unsigned long qd, bool up)
{
return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
}
/*
* We scale the qd down faster than we scale up, so we need to use this helper
* to adjust the scale_cookie accordingly so we don't prematurely get
* scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
*
* Each group has their own local copy of the last scale cookie they saw, so if
* the global scale cookie goes up or down they know which way they need to go
* based on their last knowledge of it.
*/
static void scale_cookie_change(struct blk_iolatency *blkiolat,
struct child_latency_info *lat_info,
bool up)
{
unsigned long qd = blkiolat->rqos.disk->queue->nr_requests;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
unsigned long scale = scale_amount(qd, up);
unsigned long old = atomic_read(&lat_info->scale_cookie);
unsigned long max_scale = qd << 1;
unsigned long diff = 0;
if (old < DEFAULT_SCALE_COOKIE)
diff = DEFAULT_SCALE_COOKIE - old;
if (up) {
if (scale + old > DEFAULT_SCALE_COOKIE)
atomic_set(&lat_info->scale_cookie,
DEFAULT_SCALE_COOKIE);
else if (diff > qd)
atomic_inc(&lat_info->scale_cookie);
else
atomic_add(scale, &lat_info->scale_cookie);
} else {
/*
* We don't want to dig a hole so deep that it takes us hours to
* dig out of it. Just enough that we don't throttle/unthrottle
* with jagged workloads but can still unthrottle once pressure
* has sufficiently dissipated.
*/
if (diff > qd) {
if (diff < max_scale)
atomic_dec(&lat_info->scale_cookie);
} else {
atomic_sub(scale, &lat_info->scale_cookie);
}
}
}
/*
* Change the queue depth of the iolatency_grp. We add 1/16th of the
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
* queue depth at a time so we don't get wild swings and hopefully dial in to
* fairer distribution of the overall queue depth. We halve the queue depth
* at a time so we can scale down queue depth quickly from default unlimited
* to target.
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
*/
static void scale_change(struct iolatency_grp *iolat, bool up)
{
unsigned long qd = iolat->blkiolat->rqos.disk->queue->nr_requests;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
unsigned long scale = scale_amount(qd, up);
unsigned long old = iolat->max_depth;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
if (old > qd)
old = qd;
if (up) {
if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
return;
if (old < qd) {
old += scale;
old = min(old, qd);
iolat->max_depth = old;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
wake_up_all(&iolat->rq_wait.wait);
}
} else {
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
old >>= 1;
iolat->max_depth = max(old, 1UL);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
}
}
/* Check our parent and see if the scale cookie has changed. */
static void check_scale_change(struct iolatency_grp *iolat)
{
struct iolatency_grp *parent;
struct child_latency_info *lat_info;
unsigned int cur_cookie;
unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
u64 scale_lat;
int direction = 0;
parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
if (!parent)
return;
lat_info = &parent->child_lat;
cur_cookie = atomic_read(&lat_info->scale_cookie);
scale_lat = READ_ONCE(lat_info->scale_lat);
if (cur_cookie < our_cookie)
direction = -1;
else if (cur_cookie > our_cookie)
direction = 1;
else
return;
if (!atomic_try_cmpxchg(&iolat->scale_cookie, &our_cookie, cur_cookie)) {
/* Somebody beat us to the punch, just bail. */
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
return;
}
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
if (direction < 0 && iolat->min_lat_nsec) {
u64 samples_thresh;
if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
return;
/*
* Sometimes high priority groups are their own worst enemy, so
* instead of taking it out on some poor other group that did 5%
* or less of the IO's for the last summation just skip this
* scale down event.
*/
samples_thresh = lat_info->nr_samples * 5;
samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
if (iolat->nr_samples <= samples_thresh)
return;
}
/* We're as low as we can go. */
if (iolat->max_depth == 1 && direction < 0) {
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
blkcg_use_delay(lat_to_blkg(iolat));
return;
}
/* We're back to the default cookie, unthrottle all the things. */
if (cur_cookie == DEFAULT_SCALE_COOKIE) {
blkcg_clear_delay(lat_to_blkg(iolat));
iolat->max_depth = UINT_MAX;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
wake_up_all(&iolat->rq_wait.wait);
return;
}
scale_change(iolat, direction > 0);
}
static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
{
struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
struct blkcg_gq *blkg = bio->bi_blkg;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
bool issue_as_root = bio_issue_as_root_blkg(bio);
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
if (!blkiolat->enabled)
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
return;
while (blkg && blkg->parent) {
struct iolatency_grp *iolat = blkg_to_lat(blkg);
if (!iolat) {
blkg = blkg->parent;
continue;
}
check_scale_change(iolat);
__blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
(bio->bi_opf & REQ_SWAP) == REQ_SWAP);
blkg = blkg->parent;
}
if (!timer_pending(&blkiolat->timer))
mod_timer(&blkiolat->timer, jiffies + HZ);
}
static void iolatency_record_time(struct iolatency_grp *iolat,
struct bio_issue *issue, u64 now,
bool issue_as_root)
{
u64 start = bio_issue_time(issue);
u64 req_time;
/*
* Have to do this so we are truncated to the correct time that our
* issue is truncated to.
*/
now = __bio_issue_time(now);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
if (now <= start)
return;
req_time = now - start;
/*
* We don't want to count issue_as_root bio's in the cgroups latency
* statistics as it could skew the numbers downwards.
*/
if (unlikely(issue_as_root && iolat->max_depth != UINT_MAX)) {
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
u64 sub = iolat->min_lat_nsec;
if (req_time < sub)
blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
return;
}
latency_stat_record_time(iolat, req_time);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
}
#define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
#define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
{
struct blkcg_gq *blkg = lat_to_blkg(iolat);
struct iolatency_grp *parent;
struct child_latency_info *lat_info;
struct latency_stat stat;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
unsigned long flags;
int cpu;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
latency_stat_init(iolat, &stat);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
preempt_disable();
for_each_online_cpu(cpu) {
struct latency_stat *s;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
s = per_cpu_ptr(iolat->stats, cpu);
latency_stat_sum(iolat, &stat, s);
latency_stat_init(iolat, s);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
}
preempt_enable();
parent = blkg_to_lat(blkg->parent);
if (!parent)
return;
lat_info = &parent->child_lat;
iolat_update_total_lat_avg(iolat, &stat);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
/* Everything is ok and we don't need to adjust the scale. */
if (latency_sum_ok(iolat, &stat) &&
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
return;
/* Somebody beat us to the punch, just bail. */
spin_lock_irqsave(&lat_info->lock, flags);
latency_stat_sum(iolat, &iolat->cur_stat, &stat);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
lat_info->nr_samples -= iolat->nr_samples;
lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
if ((lat_info->last_scale_event >= now ||
now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
goto out;
if (latency_sum_ok(iolat, &iolat->cur_stat) &&
latency_sum_ok(iolat, &stat)) {
if (latency_stat_samples(iolat, &iolat->cur_stat) <
BLKIOLATENCY_MIN_GOOD_SAMPLES)
goto out;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
if (lat_info->scale_grp == iolat) {
lat_info->last_scale_event = now;
scale_cookie_change(iolat->blkiolat, lat_info, true);
}
} else if (lat_info->scale_lat == 0 ||
lat_info->scale_lat >= iolat->min_lat_nsec) {
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
lat_info->last_scale_event = now;
if (!lat_info->scale_grp ||
lat_info->scale_lat > iolat->min_lat_nsec) {
WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
lat_info->scale_grp = iolat;
}
scale_cookie_change(iolat->blkiolat, lat_info, false);
}
latency_stat_init(iolat, &iolat->cur_stat);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
out:
spin_unlock_irqrestore(&lat_info->lock, flags);
}
static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
{
struct blkcg_gq *blkg;
struct rq_wait *rqw;
struct iolatency_grp *iolat;
u64 window_start;
u64 now;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
bool issue_as_root = bio_issue_as_root_blkg(bio);
int inflight = 0;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
blkg = bio->bi_blkg;
block: fix rq-qos breakage from skipping rq_qos_done_bio() a647a524a467 ("block: don't call rq_qos_ops->done_bio if the bio isn't tracked") made bio_endio() skip rq_qos_done_bio() if BIO_TRACKED is not set. While this fixed a potential oops, it also broke blk-iocost by skipping the done_bio callback for merged bios. Before, whether a bio goes through rq_qos_throttle() or rq_qos_merge(), rq_qos_done_bio() would be called on the bio on completion with BIO_TRACKED distinguishing the former from the latter. rq_qos_done_bio() is not called for bios which wenth through rq_qos_merge(). This royally confuses blk-iocost as the merged bios never finish and are considered perpetually in-flight. One reliably reproducible failure mode is an intermediate cgroup geting stuck active preventing its children from being activated due to the leaf-only rule, leading to loss of control. The following is from resctl-bench protection scenario which emulates isolating a web server like workload from a memory bomb run on an iocost configuration which should yield a reasonable level of protection. # cat /sys/block/nvme2n1/device/model Samsung SSD 970 PRO 512GB # cat /sys/fs/cgroup/io.cost.model 259:0 ctrl=user model=linear rbps=834913556 rseqiops=93622 rrandiops=102913 wbps=618985353 wseqiops=72325 wrandiops=71025 # cat /sys/fs/cgroup/io.cost.qos 259:0 enable=1 ctrl=user rpct=95.00 rlat=18776 wpct=95.00 wlat=8897 min=60.00 max=100.00 # resctl-bench -m 29.6G -r out.json run protection::scenario=mem-hog,loops=1 ... Memory Hog Summary ================== IO Latency: R p50=242u:336u/2.5m p90=794u:1.4m/7.5m p99=2.7m:8.0m/62.5m max=8.0m:36.4m/350m W p50=221u:323u/1.5m p90=709u:1.2m/5.5m p99=1.5m:2.5m/9.5m max=6.9m:35.9m/350m Isolation and Request Latency Impact Distributions: min p01 p05 p10 p25 p50 p75 p90 p95 p99 max mean stdev isol% 15.90 15.90 15.90 40.05 57.24 59.07 60.01 74.63 74.63 90.35 90.35 58.12 15.82 lat-imp% 0 0 0 0 0 4.55 14.68 15.54 233.5 548.1 548.1 53.88 143.6 Result: isol=58.12:15.82% lat_imp=53.88%:143.6 work_csv=100.0% missing=3.96% The isolation result of 58.12% is close to what this device would show without any IO control. Fix it by introducing a new flag BIO_QOS_MERGED to mark merged bios and calling rq_qos_done_bio() on them too. For consistency and clarity, rename BIO_TRACKED to BIO_QOS_THROTTLED. The flag checks are moved into rq_qos_done_bio() so that it's next to the code paths that set the flags. With the patch applied, the above same benchmark shows: # resctl-bench -m 29.6G -r out.json run protection::scenario=mem-hog,loops=1 ... Memory Hog Summary ================== IO Latency: R p50=123u:84.4u/985u p90=322u:256u/2.5m p99=1.6m:1.4m/9.5m max=11.1m:36.0m/350m W p50=429u:274u/995u p90=1.7m:1.3m/4.5m p99=3.4m:2.7m/11.5m max=7.9m:5.9m/26.5m Isolation and Request Latency Impact Distributions: min p01 p05 p10 p25 p50 p75 p90 p95 p99 max mean stdev isol% 84.91 84.91 89.51 90.73 92.31 94.49 96.36 98.04 98.71 100.0 100.0 94.42 2.81 lat-imp% 0 0 0 0 0 2.81 5.73 11.11 13.92 17.53 22.61 4.10 4.68 Result: isol=94.42:2.81% lat_imp=4.10%:4.68 work_csv=58.34% missing=0% Signed-off-by: Tejun Heo <tj@kernel.org> Fixes: a647a524a467 ("block: don't call rq_qos_ops->done_bio if the bio isn't tracked") Cc: stable@vger.kernel.org # v5.15+ Cc: Ming Lei <ming.lei@redhat.com> Cc: Yu Kuai <yukuai3@huawei.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Link: https://lore.kernel.org/r/Yi7rdrzQEHjJLGKB@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-03-14 07:15:02 +00:00
if (!blkg || !bio_flagged(bio, BIO_QOS_THROTTLED))
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
return;
iolat = blkg_to_lat(bio->bi_blkg);
if (!iolat)
return;
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
if (!iolat->blkiolat->enabled)
blk-iolatency: fix IO hang due to negative inflight counter Our test reported the following stack, and vmcore showed that ->inflight counter is -1. [ffffc9003fcc38d0] __schedule at ffffffff8173d95d [ffffc9003fcc3958] schedule at ffffffff8173de26 [ffffc9003fcc3970] io_schedule at ffffffff810bb6b6 [ffffc9003fcc3988] blkcg_iolatency_throttle at ffffffff813911cb [ffffc9003fcc3a20] rq_qos_throttle at ffffffff813847f3 [ffffc9003fcc3a48] blk_mq_make_request at ffffffff8137468a [ffffc9003fcc3b08] generic_make_request at ffffffff81368b49 [ffffc9003fcc3b68] submit_bio at ffffffff81368d7d [ffffc9003fcc3bb8] ext4_io_submit at ffffffffa031be00 [ext4] [ffffc9003fcc3c00] ext4_writepages at ffffffffa03163de [ext4] [ffffc9003fcc3d68] do_writepages at ffffffff811c49ae [ffffc9003fcc3d78] __filemap_fdatawrite_range at ffffffff811b6188 [ffffc9003fcc3e30] filemap_write_and_wait_range at ffffffff811b6301 [ffffc9003fcc3e60] ext4_sync_file at ffffffffa030cee8 [ext4] [ffffc9003fcc3ea8] vfs_fsync_range at ffffffff8128594b [ffffc9003fcc3ee8] do_fsync at ffffffff81285abd [ffffc9003fcc3f18] sys_fsync at ffffffff81285d50 [ffffc9003fcc3f28] do_syscall_64 at ffffffff81003c04 [ffffc9003fcc3f50] entry_SYSCALL_64_after_swapgs at ffffffff81742b8e The ->inflight counter may be negative (-1) if 1) blk-iolatency was disabled when the IO was issued, 2) blk-iolatency was enabled before this IO reached its endio, 3) the ->inflight counter is decreased from 0 to -1 in endio() In fact the hang can be easily reproduced by the below script, H=/sys/fs/cgroup/unified/ P=/sys/fs/cgroup/unified/test echo "+io" > $H/cgroup.subtree_control mkdir -p $P echo $$ > $P/cgroup.procs xfs_io -f -d -c "pwrite 0 4k" /dev/sdg echo "`cat /sys/block/sdg/dev` target=1000000" > $P/io.latency xfs_io -f -d -c "pwrite 0 4k" /dev/sdg This fixes the problem by freezing the queue so that while enabling/disabling iolatency, there is no inflight rq running. Note that quiesce_queue is not needed as this only updating iolatency configuration about which dispatching request_queue doesn't care. Signed-off-by: Liu Bo <bo.liu@linux.alibaba.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-25 00:12:47 +00:00
return;
now = ktime_to_ns(ktime_get());
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
while (blkg && blkg->parent) {
iolat = blkg_to_lat(blkg);
if (!iolat) {
blkg = blkg->parent;
continue;
}
rqw = &iolat->rq_wait;
inflight = atomic_dec_return(&rqw->inflight);
WARN_ON_ONCE(inflight < 0);
blk-iolatency: fix STS_AGAIN handling The iolatency controller is based on rq_qos. It increments on rq_qos_throttle() and decrements on either rq_qos_cleanup() or rq_qos_done_bio(). a3fb01ba5af0 fixes the double accounting issue where blk_mq_make_request() may call both rq_qos_cleanup() and rq_qos_done_bio() on REQ_NO_WAIT. So checking STS_AGAIN prevents the double decrement. The above works upstream as the only way we can get STS_AGAIN is from blk_mq_get_request() failing. The STS_AGAIN handling isn't a real problem as bio_endio() skipping only happens on reserved tag allocation failures which can only be caused by driver bugs and already triggers WARN. However, the fix creates a not so great dependency on how STS_AGAIN can be propagated. Internally, we (Facebook) carry a patch that kills read ahead if a cgroup is io congested or a fatal signal is pending. This combined with chained bios progagate their bi_status to the parent is not already set can can cause the parent bio to not clean up properly even though it was successful. This consequently leaks the inflight counter and can hang all IOs under that blkg. To nip the adverse interaction early, this removes the rq_qos_cleanup() callback in iolatency in favor of cleaning up always on the rq_qos_done_bio() path. Fixes: a3fb01ba5af0 ("blk-iolatency: only account submitted bios") Debugged-by: Tejun Heo <tj@kernel.org> Debugged-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Dennis Zhou <dennis@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-07-05 21:09:09 +00:00
/*
* If bi_status is BLK_STS_AGAIN, the bio wasn't actually
* submitted, so do not account for it.
*/
if (iolat->min_lat_nsec && bio->bi_status != BLK_STS_AGAIN) {
iolatency_record_time(iolat, &bio->bi_issue, now,
issue_as_root);
window_start = atomic64_read(&iolat->window_start);
if (now > window_start &&
(now - window_start) >= iolat->cur_win_nsec) {
if (atomic64_try_cmpxchg(&iolat->window_start,
&window_start, now))
blk-iolatency: fix STS_AGAIN handling The iolatency controller is based on rq_qos. It increments on rq_qos_throttle() and decrements on either rq_qos_cleanup() or rq_qos_done_bio(). a3fb01ba5af0 fixes the double accounting issue where blk_mq_make_request() may call both rq_qos_cleanup() and rq_qos_done_bio() on REQ_NO_WAIT. So checking STS_AGAIN prevents the double decrement. The above works upstream as the only way we can get STS_AGAIN is from blk_mq_get_request() failing. The STS_AGAIN handling isn't a real problem as bio_endio() skipping only happens on reserved tag allocation failures which can only be caused by driver bugs and already triggers WARN. However, the fix creates a not so great dependency on how STS_AGAIN can be propagated. Internally, we (Facebook) carry a patch that kills read ahead if a cgroup is io congested or a fatal signal is pending. This combined with chained bios progagate their bi_status to the parent is not already set can can cause the parent bio to not clean up properly even though it was successful. This consequently leaks the inflight counter and can hang all IOs under that blkg. To nip the adverse interaction early, this removes the rq_qos_cleanup() callback in iolatency in favor of cleaning up always on the rq_qos_done_bio() path. Fixes: a3fb01ba5af0 ("blk-iolatency: only account submitted bios") Debugged-by: Tejun Heo <tj@kernel.org> Debugged-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Dennis Zhou <dennis@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-07-05 21:09:09 +00:00
iolatency_check_latencies(iolat, now);
}
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
}
wake_up(&rqw->wait);
blkg = blkg->parent;
}
}
static void blkcg_iolatency_exit(struct rq_qos *rqos)
{
struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
timer_shutdown_sync(&blkiolat->timer);
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
flush_work(&blkiolat->enable_work);
blkcg_deactivate_policy(rqos->disk, &blkcg_policy_iolatency);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
kfree(blkiolat);
}
static const struct rq_qos_ops blkcg_iolatency_ops = {
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
.throttle = blkcg_iolatency_throttle,
.done_bio = blkcg_iolatency_done_bio,
.exit = blkcg_iolatency_exit,
};
static void blkiolatency_timer_fn(struct timer_list *t)
{
struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
struct blkcg_gq *blkg;
struct cgroup_subsys_state *pos_css;
u64 now = ktime_to_ns(ktime_get());
rcu_read_lock();
blkg_for_each_descendant_pre(blkg, pos_css,
blkiolat->rqos.disk->queue->root_blkg) {
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
struct iolatency_grp *iolat;
struct child_latency_info *lat_info;
unsigned long flags;
u64 cookie;
/*
* We could be exiting, don't access the pd unless we have a
* ref on the blkg.
*/
if (!blkg_tryget(blkg))
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
continue;
iolat = blkg_to_lat(blkg);
if (!iolat)
goto next;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
lat_info = &iolat->child_lat;
cookie = atomic_read(&lat_info->scale_cookie);
if (cookie >= DEFAULT_SCALE_COOKIE)
goto next;
spin_lock_irqsave(&lat_info->lock, flags);
if (lat_info->last_scale_event >= now)
goto next_lock;
/*
* We scaled down but don't have a scale_grp, scale up and carry
* on.
*/
if (lat_info->scale_grp == NULL) {
scale_cookie_change(iolat->blkiolat, lat_info, true);
goto next_lock;
}
/*
* It's been 5 seconds since our last scale event, clear the
* scale grp in case the group that needed the scale down isn't
* doing any IO currently.
*/
if (now - lat_info->last_scale_event >=
((u64)NSEC_PER_SEC * 5))
lat_info->scale_grp = NULL;
next_lock:
spin_unlock_irqrestore(&lat_info->lock, flags);
next:
blkg_put(blkg);
}
rcu_read_unlock();
}
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
/**
* blkiolatency_enable_work_fn - Enable or disable iolatency on the device
* @work: enable_work of the blk_iolatency of interest
*
* iolatency needs to keep track of the number of in-flight IOs per cgroup. This
* is relatively expensive as it involves walking up the hierarchy twice for
* every IO. Thus, if iolatency is not enabled in any cgroup for the device, we
* want to disable the in-flight tracking.
*
* We have to make sure that the counting is balanced - we don't want to leak
* the in-flight counts by disabling accounting in the completion path while IOs
* are in flight. This is achieved by ensuring that no IO is in flight by
* freezing the queue while flipping ->enabled. As this requires a sleepable
* context, ->enabled flipping is punted to this work function.
*/
static void blkiolatency_enable_work_fn(struct work_struct *work)
{
struct blk_iolatency *blkiolat = container_of(work, struct blk_iolatency,
enable_work);
bool enabled;
/*
* There can only be one instance of this function running for @blkiolat
* and it's guaranteed to be executed at least once after the latest
* ->enabled_cnt modification. Acting on the latest ->enable_cnt is
* sufficient.
*
* Also, we know @blkiolat is safe to access as ->enable_work is flushed
* in blkcg_iolatency_exit().
*/
enabled = atomic_read(&blkiolat->enable_cnt);
if (enabled != blkiolat->enabled) {
blk_mq_freeze_queue(blkiolat->rqos.disk->queue);
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
blkiolat->enabled = enabled;
blk_mq_unfreeze_queue(blkiolat->rqos.disk->queue);
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
}
}
static int blk_iolatency_init(struct gendisk *disk)
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
{
struct blk_iolatency *blkiolat;
int ret;
blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
if (!blkiolat)
return -ENOMEM;
ret = rq_qos_add(&blkiolat->rqos, disk, RQ_QOS_LATENCY,
&blkcg_iolatency_ops);
if (ret)
goto err_free;
ret = blkcg_activate_policy(disk, &blkcg_policy_iolatency);
if (ret)
goto err_qos_del;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
INIT_WORK(&blkiolat->enable_work, blkiolatency_enable_work_fn);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
return 0;
err_qos_del:
rq_qos_del(&blkiolat->rqos);
err_free:
kfree(blkiolat);
return ret;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
}
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
{
struct iolatency_grp *iolat = blkg_to_lat(blkg);
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
struct blk_iolatency *blkiolat = iolat->blkiolat;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
u64 oldval = iolat->min_lat_nsec;
iolat->min_lat_nsec = val;
iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
BLKIOLATENCY_MAX_WIN_SIZE);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
if (!oldval && val) {
if (atomic_inc_return(&blkiolat->enable_cnt) == 1)
schedule_work(&blkiolat->enable_work);
}
if (oldval && !val) {
blkcg_clear_delay(blkg);
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
if (atomic_dec_return(&blkiolat->enable_cnt) == 0)
schedule_work(&blkiolat->enable_work);
}
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
}
static void iolatency_clear_scaling(struct blkcg_gq *blkg)
{
if (blkg->parent) {
struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
struct child_latency_info *lat_info;
if (!iolat)
return;
lat_info = &iolat->child_lat;
spin_lock(&lat_info->lock);
atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
lat_info->last_scale_event = 0;
lat_info->scale_grp = NULL;
lat_info->scale_lat = 0;
spin_unlock(&lat_info->lock);
}
}
static int blk_iolatency_try_init(struct blkg_conf_ctx *ctx)
{
static DEFINE_MUTEX(init_mutex);
int ret;
ret = blkg_conf_open_bdev(ctx);
if (ret)
return ret;
/*
* blk_iolatency_init() may fail after rq_qos_add() succeeds which can
* confuse iolat_rq_qos() test. Make the test and init atomic.
*/
mutex_lock(&init_mutex);
if (!iolat_rq_qos(ctx->bdev->bd_queue))
ret = blk_iolatency_init(ctx->bdev->bd_disk);
mutex_unlock(&init_mutex);
return ret;
}
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
size_t nbytes, loff_t off)
{
struct blkcg *blkcg = css_to_blkcg(of_css(of));
struct blkcg_gq *blkg;
struct blkg_conf_ctx ctx;
struct iolatency_grp *iolat;
char *p, *tok;
u64 lat_val = 0;
u64 oldval;
int ret;
blkcg: Restructure blkg_conf_prep() and friends We want to support lazy init of rq-qos policies so that iolatency is enabled lazily on configuration instead of gendisk initialization. The way blkg config helpers are structured now is a bit awkward for that. Let's restructure: * blkcg_conf_open_bdev() is renamed to blkg_conf_open_bdev(). The blkcg_ prefix was used because the bdev opening step is blkg-independent. However, the distinction is too subtle and confuses more than helps. Let's switch to blkg prefix so that it's consistent with the type and other helper names. * struct blkg_conf_ctx now remembers the original input string and is always initialized by the new blkg_conf_init(). * blkg_conf_open_bdev() is updated to take a pointer to blkg_conf_ctx like blkg_conf_prep() and can be called multiple times safely. Instead of modifying the double pointer to input string directly, blkg_conf_open_bdev() now sets blkg_conf_ctx->body. * blkg_conf_finish() is renamed to blkg_conf_exit() for symmetry and now must be called on all blkg_conf_ctx's which were initialized with blkg_conf_init(). Combined, this allows the users to either open the bdev first or do it altogether with blkg_conf_prep() which will help implementing lazy init of rq-qos policies. blkg_conf_init/exit() will also be used implement synchronization against device removal. This is necessary because iolat / iocost are configured through cgroupfs instead of one of the files under /sys/block/DEVICE. As cgroupfs operations aren't synchronized with block layer, the lazy init and other configuration operations may race against device removal. This patch makes blkg_conf_init/exit() used consistently for all cgroup-orginating configurations making them a good place to implement explicit synchronization. Users are updated accordingly. No behavior change is intended by this patch. v2: bfq wasn't updated in v1 causing a build error. Fixed. v3: Update the description to include future use of blkg_conf_init/exit() as synchronization points. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Yu Kuai <yukuai1@huaweicloud.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Link: https://lore.kernel.org/r/20230413000649.115785-3-tj@kernel.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-04-13 00:06:47 +00:00
blkg_conf_init(&ctx, buf);
ret = blk_iolatency_try_init(&ctx);
if (ret)
goto out;
blkcg: Restructure blkg_conf_prep() and friends We want to support lazy init of rq-qos policies so that iolatency is enabled lazily on configuration instead of gendisk initialization. The way blkg config helpers are structured now is a bit awkward for that. Let's restructure: * blkcg_conf_open_bdev() is renamed to blkg_conf_open_bdev(). The blkcg_ prefix was used because the bdev opening step is blkg-independent. However, the distinction is too subtle and confuses more than helps. Let's switch to blkg prefix so that it's consistent with the type and other helper names. * struct blkg_conf_ctx now remembers the original input string and is always initialized by the new blkg_conf_init(). * blkg_conf_open_bdev() is updated to take a pointer to blkg_conf_ctx like blkg_conf_prep() and can be called multiple times safely. Instead of modifying the double pointer to input string directly, blkg_conf_open_bdev() now sets blkg_conf_ctx->body. * blkg_conf_finish() is renamed to blkg_conf_exit() for symmetry and now must be called on all blkg_conf_ctx's which were initialized with blkg_conf_init(). Combined, this allows the users to either open the bdev first or do it altogether with blkg_conf_prep() which will help implementing lazy init of rq-qos policies. blkg_conf_init/exit() will also be used implement synchronization against device removal. This is necessary because iolat / iocost are configured through cgroupfs instead of one of the files under /sys/block/DEVICE. As cgroupfs operations aren't synchronized with block layer, the lazy init and other configuration operations may race against device removal. This patch makes blkg_conf_init/exit() used consistently for all cgroup-orginating configurations making them a good place to implement explicit synchronization. Users are updated accordingly. No behavior change is intended by this patch. v2: bfq wasn't updated in v1 causing a build error. Fixed. v3: Update the description to include future use of blkg_conf_init/exit() as synchronization points. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Yu Kuai <yukuai1@huaweicloud.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Link: https://lore.kernel.org/r/20230413000649.115785-3-tj@kernel.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-04-13 00:06:47 +00:00
ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, &ctx);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
if (ret)
blkcg: Restructure blkg_conf_prep() and friends We want to support lazy init of rq-qos policies so that iolatency is enabled lazily on configuration instead of gendisk initialization. The way blkg config helpers are structured now is a bit awkward for that. Let's restructure: * blkcg_conf_open_bdev() is renamed to blkg_conf_open_bdev(). The blkcg_ prefix was used because the bdev opening step is blkg-independent. However, the distinction is too subtle and confuses more than helps. Let's switch to blkg prefix so that it's consistent with the type and other helper names. * struct blkg_conf_ctx now remembers the original input string and is always initialized by the new blkg_conf_init(). * blkg_conf_open_bdev() is updated to take a pointer to blkg_conf_ctx like blkg_conf_prep() and can be called multiple times safely. Instead of modifying the double pointer to input string directly, blkg_conf_open_bdev() now sets blkg_conf_ctx->body. * blkg_conf_finish() is renamed to blkg_conf_exit() for symmetry and now must be called on all blkg_conf_ctx's which were initialized with blkg_conf_init(). Combined, this allows the users to either open the bdev first or do it altogether with blkg_conf_prep() which will help implementing lazy init of rq-qos policies. blkg_conf_init/exit() will also be used implement synchronization against device removal. This is necessary because iolat / iocost are configured through cgroupfs instead of one of the files under /sys/block/DEVICE. As cgroupfs operations aren't synchronized with block layer, the lazy init and other configuration operations may race against device removal. This patch makes blkg_conf_init/exit() used consistently for all cgroup-orginating configurations making them a good place to implement explicit synchronization. Users are updated accordingly. No behavior change is intended by this patch. v2: bfq wasn't updated in v1 causing a build error. Fixed. v3: Update the description to include future use of blkg_conf_init/exit() as synchronization points. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Yu Kuai <yukuai1@huaweicloud.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Link: https://lore.kernel.org/r/20230413000649.115785-3-tj@kernel.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-04-13 00:06:47 +00:00
goto out;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
iolat = blkg_to_lat(ctx.blkg);
p = ctx.body;
ret = -EINVAL;
while ((tok = strsep(&p, " "))) {
char key[16];
char val[21]; /* 18446744073709551616 */
if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
goto out;
if (!strcmp(key, "target")) {
u64 v;
if (!strcmp(val, "max"))
lat_val = 0;
else if (sscanf(val, "%llu", &v) == 1)
lat_val = v * NSEC_PER_USEC;
else
goto out;
} else {
goto out;
}
}
/* Walk up the tree to see if our new val is lower than it should be. */
blkg = ctx.blkg;
oldval = iolat->min_lat_nsec;
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
iolatency_set_min_lat_nsec(blkg, lat_val);
if (oldval != iolat->min_lat_nsec)
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
iolatency_clear_scaling(blkg);
ret = 0;
out:
blkcg: Restructure blkg_conf_prep() and friends We want to support lazy init of rq-qos policies so that iolatency is enabled lazily on configuration instead of gendisk initialization. The way blkg config helpers are structured now is a bit awkward for that. Let's restructure: * blkcg_conf_open_bdev() is renamed to blkg_conf_open_bdev(). The blkcg_ prefix was used because the bdev opening step is blkg-independent. However, the distinction is too subtle and confuses more than helps. Let's switch to blkg prefix so that it's consistent with the type and other helper names. * struct blkg_conf_ctx now remembers the original input string and is always initialized by the new blkg_conf_init(). * blkg_conf_open_bdev() is updated to take a pointer to blkg_conf_ctx like blkg_conf_prep() and can be called multiple times safely. Instead of modifying the double pointer to input string directly, blkg_conf_open_bdev() now sets blkg_conf_ctx->body. * blkg_conf_finish() is renamed to blkg_conf_exit() for symmetry and now must be called on all blkg_conf_ctx's which were initialized with blkg_conf_init(). Combined, this allows the users to either open the bdev first or do it altogether with blkg_conf_prep() which will help implementing lazy init of rq-qos policies. blkg_conf_init/exit() will also be used implement synchronization against device removal. This is necessary because iolat / iocost are configured through cgroupfs instead of one of the files under /sys/block/DEVICE. As cgroupfs operations aren't synchronized with block layer, the lazy init and other configuration operations may race against device removal. This patch makes blkg_conf_init/exit() used consistently for all cgroup-orginating configurations making them a good place to implement explicit synchronization. Users are updated accordingly. No behavior change is intended by this patch. v2: bfq wasn't updated in v1 causing a build error. Fixed. v3: Update the description to include future use of blkg_conf_init/exit() as synchronization points. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Yu Kuai <yukuai1@huaweicloud.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Link: https://lore.kernel.org/r/20230413000649.115785-3-tj@kernel.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2023-04-13 00:06:47 +00:00
blkg_conf_exit(&ctx);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
return ret ?: nbytes;
}
static u64 iolatency_prfill_limit(struct seq_file *sf,
struct blkg_policy_data *pd, int off)
{
struct iolatency_grp *iolat = pd_to_lat(pd);
const char *dname = blkg_dev_name(pd->blkg);
if (!dname || !iolat->min_lat_nsec)
return 0;
seq_printf(sf, "%s target=%llu\n",
dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
return 0;
}
static int iolatency_print_limit(struct seq_file *sf, void *v)
{
blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
iolatency_prfill_limit,
&blkcg_policy_iolatency, seq_cft(sf)->private, false);
return 0;
}
static void iolatency_ssd_stat(struct iolatency_grp *iolat, struct seq_file *s)
{
struct latency_stat stat;
int cpu;
latency_stat_init(iolat, &stat);
preempt_disable();
for_each_online_cpu(cpu) {
struct latency_stat *s;
s = per_cpu_ptr(iolat->stats, cpu);
latency_stat_sum(iolat, &stat, s);
}
preempt_enable();
if (iolat->max_depth == UINT_MAX)
seq_printf(s, " missed=%llu total=%llu depth=max",
(unsigned long long)stat.ps.missed,
(unsigned long long)stat.ps.total);
else
seq_printf(s, " missed=%llu total=%llu depth=%u",
(unsigned long long)stat.ps.missed,
(unsigned long long)stat.ps.total,
iolat->max_depth);
}
static void iolatency_pd_stat(struct blkg_policy_data *pd, struct seq_file *s)
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
{
struct iolatency_grp *iolat = pd_to_lat(pd);
unsigned long long avg_lat;
unsigned long long cur_win;
if (!blkcg_debug_stats)
return;
if (iolat->ssd)
return iolatency_ssd_stat(iolat, s);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
if (iolat->max_depth == UINT_MAX)
seq_printf(s, " depth=max avg_lat=%llu win=%llu",
avg_lat, cur_win);
else
seq_printf(s, " depth=%u avg_lat=%llu win=%llu",
iolat->max_depth, avg_lat, cur_win);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
}
static struct blkg_policy_data *iolatency_pd_alloc(struct gendisk *disk,
struct blkcg *blkcg, gfp_t gfp)
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
{
struct iolatency_grp *iolat;
iolat = kzalloc_node(sizeof(*iolat), gfp, disk->node_id);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
if (!iolat)
return NULL;
iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
__alignof__(struct latency_stat), gfp);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
if (!iolat->stats) {
kfree(iolat);
return NULL;
}
return &iolat->pd;
}
static void iolatency_pd_init(struct blkg_policy_data *pd)
{
struct iolatency_grp *iolat = pd_to_lat(pd);
struct blkcg_gq *blkg = lat_to_blkg(iolat);
struct rq_qos *rqos = iolat_rq_qos(blkg->q);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
u64 now = ktime_to_ns(ktime_get());
int cpu;
if (blk_queue_nonrot(blkg->q))
iolat->ssd = true;
else
iolat->ssd = false;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
for_each_possible_cpu(cpu) {
struct latency_stat *stat;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
stat = per_cpu_ptr(iolat->stats, cpu);
latency_stat_init(iolat, stat);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
}
latency_stat_init(iolat, &iolat->cur_stat);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
rq_wait_init(&iolat->rq_wait);
spin_lock_init(&iolat->child_lat.lock);
iolat->max_depth = UINT_MAX;
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
iolat->blkiolat = blkiolat;
iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
atomic64_set(&iolat->window_start, now);
/*
* We init things in list order, so the pd for the parent may not be
* init'ed yet for whatever reason.
*/
if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
atomic_set(&iolat->scale_cookie,
atomic_read(&parent->child_lat.scale_cookie));
} else {
atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
}
atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
}
static void iolatency_pd_offline(struct blkg_policy_data *pd)
{
struct iolatency_grp *iolat = pd_to_lat(pd);
struct blkcg_gq *blkg = lat_to_blkg(iolat);
blk-iolatency: Fix inflight count imbalances and IO hangs on offline iolatency needs to track the number of inflight IOs per cgroup. As this tracking can be expensive, it is disabled when no cgroup has iolatency configured for the device. To ensure that the inflight counters stay balanced, iolatency_set_limit() freezes the request_queue while manipulating the enabled counter, which ensures that no IO is in flight and thus all counters are zero. Unfortunately, iolatency_set_limit() isn't the only place where the enabled counter is manipulated. iolatency_pd_offline() can also dec the counter and trigger disabling. As this disabling happens without freezing the q, this can easily happen while some IOs are in flight and thus leak the counts. This can be easily demonstrated by turning on iolatency on an one empty cgroup while IOs are in flight in other cgroups and then removing the cgroup. Note that iolatency shouldn't have been enabled elsewhere in the system to ensure that removing the cgroup disables iolatency for the whole device. The following keeps flipping on and off iolatency on sda: echo +io > /sys/fs/cgroup/cgroup.subtree_control while true; do mkdir -p /sys/fs/cgroup/test echo '8:0 target=100000' > /sys/fs/cgroup/test/io.latency sleep 1 rmdir /sys/fs/cgroup/test sleep 1 done and there's concurrent fio generating direct rand reads: fio --name test --filename=/dev/sda --direct=1 --rw=randread \ --runtime=600 --time_based --iodepth=256 --numjobs=4 --bs=4k while monitoring with the following drgn script: while True: for css in css_for_each_descendant_pre(prog['blkcg_root'].css.address_of_()): for pos in hlist_for_each(container_of(css, 'struct blkcg', 'css').blkg_list): blkg = container_of(pos, 'struct blkcg_gq', 'blkcg_node') pd = blkg.pd[prog['blkcg_policy_iolatency'].plid] if pd.value_() == 0: continue iolat = container_of(pd, 'struct iolatency_grp', 'pd') inflight = iolat.rq_wait.inflight.counter.value_() if inflight: print(f'inflight={inflight} {disk_name(blkg.q.disk).decode("utf-8")} ' f'{cgroup_path(css.cgroup).decode("utf-8")}') time.sleep(1) The monitoring output looks like the following: inflight=1 sda /user.slice inflight=1 sda /user.slice ... inflight=14 sda /user.slice inflight=13 sda /user.slice inflight=17 sda /user.slice inflight=15 sda /user.slice inflight=18 sda /user.slice inflight=17 sda /user.slice inflight=20 sda /user.slice inflight=19 sda /user.slice <- fio stopped, inflight stuck at 19 inflight=19 sda /user.slice inflight=19 sda /user.slice If a cgroup with stuck inflight ends up getting throttled, the throttled IOs will never get issued as there's no completion event to wake it up leading to an indefinite hang. This patch fixes the bug by unifying enable handling into a work item which is automatically kicked off from iolatency_set_min_lat_nsec() which is called from both iolatency_set_limit() and iolatency_pd_offline() paths. Punting to a work item is necessary as iolatency_pd_offline() is called under spinlocks while freezing a request_queue requires a sleepable context. This also simplifies the code reducing LOC sans the comments and avoids the unnecessary freezes which were happening whenever a cgroup's latency target is newly set or cleared. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Josef Bacik <josef@toxicpanda.com> Cc: Liu Bo <bo.liu@linux.alibaba.com> Fixes: 8c772a9bfc7c ("blk-iolatency: fix IO hang due to negative inflight counter") Cc: stable@vger.kernel.org # v5.0+ Link: https://lore.kernel.org/r/Yn9ScX6Nx2qIiQQi@slm.duckdns.org Signed-off-by: Jens Axboe <axboe@kernel.dk>
2022-05-14 06:55:45 +00:00
iolatency_set_min_lat_nsec(blkg, 0);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
iolatency_clear_scaling(blkg);
}
static void iolatency_pd_free(struct blkg_policy_data *pd)
{
struct iolatency_grp *iolat = pd_to_lat(pd);
free_percpu(iolat->stats);
kfree(iolat);
}
static struct cftype iolatency_files[] = {
{
.name = "latency",
.flags = CFTYPE_NOT_ON_ROOT,
.seq_show = iolatency_print_limit,
.write = iolatency_set_limit,
},
{}
};
static struct blkcg_policy blkcg_policy_iolatency = {
.dfl_cftypes = iolatency_files,
.pd_alloc_fn = iolatency_pd_alloc,
.pd_init_fn = iolatency_pd_init,
.pd_offline_fn = iolatency_pd_offline,
.pd_free_fn = iolatency_pd_free,
.pd_stat_fn = iolatency_pd_stat,
};
static int __init iolatency_init(void)
{
return blkcg_policy_register(&blkcg_policy_iolatency);
}
static void __exit iolatency_exit(void)
{
blkcg_policy_unregister(&blkcg_policy_iolatency);
block: introduce blk-iolatency io controller Current IO controllers for the block layer are less than ideal for our use case. The io.max controller is great at hard limiting, but it is not work conserving. This patch introduces io.latency. You provide a latency target for your group and we monitor the io in short windows to make sure we are not exceeding those latency targets. This makes use of the rq-qos infrastructure and works much like the wbt stuff. There are a few differences from wbt - It's bio based, so the latency covers the whole block layer in addition to the actual io. - We will throttle all IO types that comes in here if we need to. - We use the mean latency over the 100ms window. This is because writes can be particularly fast, which could give us a false sense of the impact of other workloads on our protected workload. - By default there's no throttling, we set the queue_depth to INT_MAX so that we can have as many outstanding bio's as we're allowed to. Only at throttle time do we pay attention to the actual queue depth. - We backcharge cgroups for root cg issued IO and induce artificial delays in order to deal with cases like metadata only or swap heavy workloads. In testing this has worked out relatively well. Protected workloads will throttle noisy workloads down to 1 io at time if they are doing normal IO on their own, or induce up to a 1 second delay per syscall if they are doing a lot of root issued IO (metadata/swap IO). Our testing has revolved mostly around our production web servers where we have hhvm (the web server application) in a protected group and everything else in another group. We see slightly higher requests per second (RPS) on the test tier vs the control tier, and much more stable RPS across all machines in the test tier vs the control tier. Another test we run is a slow memory allocator in the unprotected group. Before this would eventually push us into swap and cause the whole box to die and not recover at all. With these patches we see slight RPS drops (usually 10-15%) before the memory consumer is properly killed and things recover within seconds. Signed-off-by: Josef Bacik <jbacik@fb.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-03 15:15:01 +00:00
}
module_init(iolatency_init);
module_exit(iolatency_exit);