dm: add unstriped target

This device mapper "unstriped" target remaps and unstripes I/O so it
is issued solely on a single drive in a HW RAID0 or dm-striped target.

In a 4 drive HW RAID0 the striped target exposes 1/4th of the LBA range
as a virtual drive.  Each I/O to that virtual drive will only be issued
to the 1 drive that was selected of the 4 drives in the HW RAID0.

This unstriped target is most useful for Intel NVMe drives that have
multiple cores but that do not have firmware control to pin separate LBA
ranges to each discrete cpu core.

Signed-off-by: Scott Bauer <scott.bauer@intel.com>
Signed-off-by: Heinz Mauelshagen <heinzm@redhat.com>
Acked-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>

Documentation/device-mapper/unstriped.txt

@@ -0,0 +1,124 @@
+Introduction
+============
+
+The device-mapper "unstriped" target provides a transparent mechanism to
+unstripe a device-mapper "striped" target to access the underlying disks
+without having to touch the true backing block-device.  It can also be
+used to unstripe a hardware RAID-0 to access backing disks.
+
+Parameters:
+<number of stripes> <chunk size> <stripe #> <dev_path> <offset>
+
+<number of stripes>
+        The number of stripes in the RAID 0.
+
+<chunk size>
+	The amount of 512B sectors in the chunk striping.
+
+<dev_path>
+	The block device you wish to unstripe.
+
+<stripe #>
+        The stripe number within the device that corresponds to physical
+        drive you wish to unstripe.  This must be 0 indexed.
+
+
+Why use this module?
+====================
+
+An example of undoing an existing dm-stripe
+-------------------------------------------
+
+This small bash script will setup 4 loop devices and use the existing
+striped target to combine the 4 devices into one.  It then will use
+the unstriped target ontop of the striped device to access the
+individual backing loop devices.  We write data to the newly exposed
+unstriped devices and verify the data written matches the correct
+underlying device on the striped array.
+
+#!/bin/bash
+
+MEMBER_SIZE=$((128 * 1024 * 1024))
+NUM=4
+SEQ_END=$((${NUM}-1))
+CHUNK=256
+BS=4096
+
+RAID_SIZE=$((${MEMBER_SIZE}*${NUM}/512))
+DM_PARMS="0 ${RAID_SIZE} striped ${NUM} ${CHUNK}"
+COUNT=$((${MEMBER_SIZE} / ${BS}))
+
+for i in $(seq 0 ${SEQ_END}); do
+  dd if=/dev/zero of=member-${i} bs=${MEMBER_SIZE} count=1 oflag=direct
+  losetup /dev/loop${i} member-${i}
+  DM_PARMS+=" /dev/loop${i} 0"
+done
+
+echo $DM_PARMS | dmsetup create raid0
+for i in $(seq 0 ${SEQ_END}); do
+  echo "0 1 unstriped ${NUM} ${CHUNK} ${i} /dev/mapper/raid0 0" | dmsetup create set-${i}
+done;
+
+for i in $(seq 0 ${SEQ_END}); do
+  dd if=/dev/urandom of=/dev/mapper/set-${i} bs=${BS} count=${COUNT} oflag=direct
+  diff /dev/mapper/set-${i} member-${i}
+done;
+
+for i in $(seq 0 ${SEQ_END}); do
+  dmsetup remove set-${i}
+done
+
+dmsetup remove raid0
+
+for i in $(seq 0 ${SEQ_END}); do
+  losetup -d /dev/loop${i}
+  rm -f member-${i}
+done
+
+Another example
+---------------
+
+Intel NVMe drives contain two cores on the physical device.
+Each core of the drive has segregated access to its LBA range.
+The current LBA model has a RAID 0 128k chunk on each core, resulting
+in a 256k stripe across the two cores:
+
+   Core 0:       Core 1:
+  __________    __________
+  | LBA 512|    | LBA 768|
+  | LBA 0  |    | LBA 256|
+  ----------    ----------
+
+The purpose of this unstriping is to provide better QoS in noisy
+neighbor environments. When two partitions are created on the
+aggregate drive without this unstriping, reads on one partition
+can affect writes on another partition.  This is because the partitions
+are striped across the two cores.  When we unstripe this hardware RAID 0
+and make partitions on each new exposed device the two partitions are now
+physically separated.
+
+With the dm-unstriped target we're able to segregate an fio script that
+has read and write jobs that are independent of each other.  Compared to
+when we run the test on a combined drive with partitions, we were able
+to get a 92% reduction in read latency using this device mapper target.
+
+
+Example dmsetup usage
+=====================
+
+unstriped ontop of Intel NVMe device that has 2 cores
+-----------------------------------------------------
+dmsetup create nvmset0 --table '0 512 unstriped 2 256 0 /dev/nvme0n1 0'
+dmsetup create nvmset1 --table '0 512 unstriped 2 256 1 /dev/nvme0n1 0'
+
+There will now be two devices that expose Intel NVMe core 0 and 1
+respectively:
+/dev/mapper/nvmset0
+/dev/mapper/nvmset1
+
+unstriped ontop of striped with 4 drives using 128K chunk size
+--------------------------------------------------------------
+dmsetup create raid_disk0 --table '0 512 unstriped 4 256 0 /dev/mapper/striped 0'
+dmsetup create raid_disk1 --table '0 512 unstriped 4 256 1 /dev/mapper/striped 0'
+dmsetup create raid_disk2 --table '0 512 unstriped 4 256 2 /dev/mapper/striped 0'
+dmsetup create raid_disk3 --table '0 512 unstriped 4 256 3 /dev/mapper/striped 0'

drivers/md/Kconfig

@@ -269,6 +269,13 @@ config DM_BIO_PRISON
 
 source "drivers/md/persistent-data/Kconfig"
 
+config DM_UNSTRIPED
+       tristate "Unstriped target"
+       depends on BLK_DEV_DM
+       ---help---
+	  Unstripes I/O so it is issued solely on a single drive in a HW
+	  RAID0 or dm-striped target.
+
 config DM_CRYPT
 	tristate "Crypt target support"
 	depends on BLK_DEV_DM

drivers/md/Makefile

@@ -43,6 +43,7 @@ obj-$(CONFIG_BCACHE)		+= bcache/
 obj-$(CONFIG_BLK_DEV_MD)	+= md-mod.o
 obj-$(CONFIG_BLK_DEV_DM)	+= dm-mod.o
 obj-$(CONFIG_BLK_DEV_DM_BUILTIN) += dm-builtin.o
+obj-$(CONFIG_DM_UNSTRIPED)	+= dm-unstripe.o
 obj-$(CONFIG_DM_BUFIO)		+= dm-bufio.o
 obj-$(CONFIG_DM_BIO_PRISON)	+= dm-bio-prison.o
 obj-$(CONFIG_DM_CRYPT)		+= dm-crypt.o

drivers/md/dm-unstripe.c

@@ -0,0 +1,225 @@
+/*
+ * Copyright (C) 2017 Intel Corporation.
+ *
+ * This file is released under the GPL.
+ */
+
+#include "dm.h"
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/blkdev.h>
+#include <linux/bio.h>
+#include <linux/slab.h>
+#include <linux/bitops.h>
+#include <linux/device-mapper.h>
+
+struct unstripe_c {
+	struct dm_dev *dev;
+	sector_t physical_start;
+
+	uint32_t stripes;
+
+	uint32_t unstripe;
+	sector_t unstripe_width;
+	sector_t unstripe_offset;
+
+	uint32_t chunk_size;
+	u8 chunk_shift;
+};
+
+#define DM_MSG_PREFIX "unstriped"
+
+static void cleanup_unstripe(struct unstripe_c *uc, struct dm_target *ti)
+{
+	if (uc->dev)
+		dm_put_device(ti, uc->dev);
+	kfree(uc);
+}
+
+/*
+ * Contruct an unstriped mapping.
+ * <number of stripes> <chunk size> <stripe #> <dev_path> <offset>
+ */
+static int unstripe_ctr(struct dm_target *ti, unsigned int argc, char **argv)
+{
+	struct unstripe_c *uc;
+	sector_t width, tmp_len;
+	unsigned long long start;
+	char dummy;
+
+	if (argc != 5) {
+		ti->error = "Invalid number of arguments";
+		return -EINVAL;
+	}
+
+	uc = kzalloc(sizeof(*uc), GFP_KERNEL);
+	if (!uc) {
+		ti->error = "Memory allocation for unstriped context failed";
+		return -ENOMEM;
+	}
+
+	if (kstrtouint(argv[0], 10, &uc->stripes) || !uc->stripes) {
+		ti->error = "Invalid stripe count";
+		goto err;
+	}
+
+	if (kstrtouint(argv[1], 10, &uc->chunk_size) || !uc->chunk_size) {
+		ti->error = "Invalid chunk_size";
+		goto err;
+	}
+
+	// FIXME: must support non power of 2 chunk_size, dm-stripe.c does
+	if (!is_power_of_2(uc->chunk_size)) {
+		ti->error = "Non power of 2 chunk_size is not supported yet";
+		goto err;
+	}
+
+	if (kstrtouint(argv[2], 10, &uc->unstripe)) {
+		ti->error = "Invalid stripe number";
+		goto err;
+	}
+
+	if (uc->unstripe > uc->stripes && uc->stripes > 1) {
+		ti->error = "Please provide stripe between [0, # of stripes]";
+		goto err;
+	}
+
+	if (dm_get_device(ti, argv[3], dm_table_get_mode(ti->table), &uc->dev)) {
+		ti->error = "Couldn't get striped device";
+		goto err;
+	}
+
+	if (sscanf(argv[4], "%llu%c", &start, &dummy) != 1) {
+		ti->error = "Invalid striped device offset";
+		goto err;
+	}
+	uc->physical_start = start;
+
+	uc->unstripe_offset = uc->unstripe * uc->chunk_size;
+	uc->unstripe_width = (uc->stripes - 1) * uc->chunk_size;
+	uc->chunk_shift = fls(uc->chunk_size) - 1;
+
+	width = ti->len;
+	if (sector_div(width, uc->stripes)) {
+		ti->error = "Target length not divisible by number of stripes";
+		goto err;
+	}
+
+	tmp_len = width;
+	if (sector_div(tmp_len, uc->chunk_size)) {
+		ti->error = "Target length not divisible by chunk size";
+		goto err;
+	}
+
+	if (dm_set_target_max_io_len(ti, uc->chunk_size)) {
+		ti->error = "Failed to set max io len";
+		goto err;
+	}
+
+	ti->private = uc;
+	return 0;
+err:
+	cleanup_unstripe(uc, ti);
+	return -EINVAL;
+}
+
+static void unstripe_dtr(struct dm_target *ti)
+{
+	struct unstripe_c *uc = ti->private;
+
+	cleanup_unstripe(uc, ti);
+}
+
+static sector_t map_to_core(struct dm_target *ti, struct bio *bio)
+{
+	struct unstripe_c *uc = ti->private;
+	sector_t sector = bio->bi_iter.bi_sector;
+
+	/* Shift us up to the right "row" on the stripe */
+	sector += uc->unstripe_width * (sector >> uc->chunk_shift);
+
+	/* Account for what stripe we're operating on */
+	sector += uc->unstripe_offset;
+
+	return sector;
+}
+
+static int unstripe_map(struct dm_target *ti, struct bio *bio)
+{
+	struct unstripe_c *uc = ti->private;
+
+	bio_set_dev(bio, uc->dev->bdev);
+	bio->bi_iter.bi_sector = map_to_core(ti, bio) + uc->physical_start;
+
+	return DM_MAPIO_REMAPPED;
+}
+
+static void unstripe_status(struct dm_target *ti, status_type_t type,
+			    unsigned int status_flags, char *result, unsigned int maxlen)
+{
+	struct unstripe_c *uc = ti->private;
+	unsigned int sz = 0;
+
+	switch (type) {
+	case STATUSTYPE_INFO:
+		break;
+
+	case STATUSTYPE_TABLE:
+		DMEMIT("%d %llu %d %s %llu",
+		       uc->stripes, (unsigned long long)uc->chunk_size, uc->unstripe,
+		       uc->dev->name, (unsigned long long)uc->physical_start);
+		break;
+	}
+}
+
+static int unstripe_iterate_devices(struct dm_target *ti,
+				    iterate_devices_callout_fn fn, void *data)
+{
+	struct unstripe_c *uc = ti->private;
+
+	return fn(ti, uc->dev, uc->physical_start, ti->len, data);
+}
+
+static void unstripe_io_hints(struct dm_target *ti,
+			       struct queue_limits *limits)
+{
+	struct unstripe_c *uc = ti->private;
+
+	limits->chunk_sectors = uc->chunk_size;
+}
+
+static struct target_type unstripe_target = {
+	.name = "unstriped",
+	.version = {1, 0, 0},
+	.module = THIS_MODULE,
+	.ctr = unstripe_ctr,
+	.dtr = unstripe_dtr,
+	.map = unstripe_map,
+	.status = unstripe_status,
+	.iterate_devices = unstripe_iterate_devices,
+	.io_hints = unstripe_io_hints,
+};
+
+static int __init dm_unstripe_init(void)
+{
+	int r;
+
+	r = dm_register_target(&unstripe_target);
+	if (r < 0)
+		DMERR("target registration failed");
+
+	return r;
+}
+
+static void __exit dm_unstripe_exit(void)
+{
+	dm_unregister_target(&unstripe_target);
+}
+
+module_init(dm_unstripe_init);
+module_exit(dm_unstripe_exit);
+
+MODULE_DESCRIPTION(DM_NAME " unstriped target");
+MODULE_AUTHOR("Scott Bauer <scott.bauer@intel.com>");
+MODULE_LICENSE("GPL");