linux/drivers/md/dm-rq.c

597 lines
15 KiB
C
Raw Normal View History

/*
* Copyright (C) 2016 Red Hat, Inc. All rights reserved.
*
* This file is released under the GPL.
*/
#include "dm-core.h"
#include "dm-rq.h"
#include <linux/elevator.h> /* for rq_end_sector() */
#include <linux/blk-mq.h>
#define DM_MSG_PREFIX "core-rq"
/*
* One of these is allocated per request.
*/
struct dm_rq_target_io {
struct mapped_device *md;
struct dm_target *ti;
struct request *orig, *clone;
struct kthread_work work;
blk_status_t error;
union map_info info;
struct dm_stats_aux stats_aux;
unsigned long duration_jiffies;
unsigned n_sectors;
unsigned completed;
};
#define DM_MQ_NR_HW_QUEUES 1
#define DM_MQ_QUEUE_DEPTH 2048
static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
/*
* Request-based DM's mempools' reserved IOs set by the user.
*/
#define RESERVED_REQUEST_BASED_IOS 256
static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
unsigned dm_get_reserved_rq_based_ios(void)
{
return __dm_get_module_param(&reserved_rq_based_ios,
RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
}
EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
static unsigned dm_get_blk_mq_nr_hw_queues(void)
{
return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
}
static unsigned dm_get_blk_mq_queue_depth(void)
{
return __dm_get_module_param(&dm_mq_queue_depth,
DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
}
int dm_request_based(struct mapped_device *md)
{
return queue_is_mq(md->queue);
}
void dm_start_queue(struct request_queue *q)
{
blk_mq_unquiesce_queue(q);
blk_mq_kick_requeue_list(q);
}
void dm_stop_queue(struct request_queue *q)
{
dm: Fix a race condition related to stopping and starting queues Ensure that all ongoing dm_mq_queue_rq() and dm_mq_requeue_request() calls have stopped before setting the "queue stopped" flag. This allows to remove the "queue stopped" test from dm_mq_queue_rq() and dm_mq_requeue_request(). This patch fixes a race condition because dm_mq_queue_rq() is called without holding the queue lock and hence BLK_MQ_S_STOPPED can be set at any time while dm_mq_queue_rq() is in progress. This patch prevents that the following hang occurs sporadically when using dm-mq: INFO: task systemd-udevd:10111 blocked for more than 480 seconds. Call Trace: [<ffffffff8161f397>] schedule+0x37/0x90 [<ffffffff816239ef>] schedule_timeout+0x27f/0x470 [<ffffffff8161e76f>] io_schedule_timeout+0x9f/0x110 [<ffffffff8161fb36>] bit_wait_io+0x16/0x60 [<ffffffff8161f929>] __wait_on_bit_lock+0x49/0xa0 [<ffffffff8114fe69>] __lock_page+0xb9/0xc0 [<ffffffff81165d90>] truncate_inode_pages_range+0x3e0/0x760 [<ffffffff81166120>] truncate_inode_pages+0x10/0x20 [<ffffffff81212a20>] kill_bdev+0x30/0x40 [<ffffffff81213d41>] __blkdev_put+0x71/0x360 [<ffffffff81214079>] blkdev_put+0x49/0x170 [<ffffffff812141c0>] blkdev_close+0x20/0x30 [<ffffffff811d48e8>] __fput+0xe8/0x1f0 [<ffffffff811d4a29>] ____fput+0x9/0x10 [<ffffffff810842d3>] task_work_run+0x83/0xb0 [<ffffffff8106606e>] do_exit+0x3ee/0xc40 [<ffffffff8106694b>] do_group_exit+0x4b/0xc0 [<ffffffff81073d9a>] get_signal+0x2ca/0x940 [<ffffffff8101bf43>] do_signal+0x23/0x660 [<ffffffff810022b3>] exit_to_usermode_loop+0x73/0xb0 [<ffffffff81002cb0>] syscall_return_slowpath+0xb0/0xc0 [<ffffffff81624e33>] entry_SYSCALL_64_fastpath+0xa6/0xa8 Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com> Acked-by: Mike Snitzer <snitzer@redhat.com> Reviewed-by: Hannes Reinecke <hare@suse.com> Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@fb.com>
2016-10-29 00:22:16 +00:00
blk_mq_quiesce_queue(q);
}
/*
* Partial completion handling for request-based dm
*/
static void end_clone_bio(struct bio *clone)
{
struct dm_rq_clone_bio_info *info =
container_of(clone, struct dm_rq_clone_bio_info, clone);
struct dm_rq_target_io *tio = info->tio;
unsigned int nr_bytes = info->orig->bi_iter.bi_size;
blk_status_t error = clone->bi_status;
bool is_last = !clone->bi_next;
bio_put(clone);
if (tio->error)
/*
* An error has already been detected on the request.
* Once error occurred, just let clone->end_io() handle
* the remainder.
*/
return;
else if (error) {
/*
* Don't notice the error to the upper layer yet.
* The error handling decision is made by the target driver,
* when the request is completed.
*/
tio->error = error;
goto exit;
}
/*
* I/O for the bio successfully completed.
* Notice the data completion to the upper layer.
*/
tio->completed += nr_bytes;
/*
* Update the original request.
* Do not use blk_mq_end_request() here, because it may complete
* the original request before the clone, and break the ordering.
*/
if (is_last)
exit:
blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
}
static struct dm_rq_target_io *tio_from_request(struct request *rq)
{
return blk_mq_rq_to_pdu(rq);
}
static void rq_end_stats(struct mapped_device *md, struct request *orig)
{
if (unlikely(dm_stats_used(&md->stats))) {
struct dm_rq_target_io *tio = tio_from_request(orig);
tio->duration_jiffies = jiffies - tio->duration_jiffies;
dm_stats_account_io(&md->stats, rq_data_dir(orig),
blk_rq_pos(orig), tio->n_sectors, true,
tio->duration_jiffies, &tio->stats_aux);
}
}
/*
* Don't touch any member of the md after calling this function because
* the md may be freed in dm_put() at the end of this function.
* Or do dm_get() before calling this function and dm_put() later.
*/
static void rq_completed(struct mapped_device *md)
{
/*
* dm_put() must be at the end of this function. See the comment above
*/
dm_put(md);
}
/*
* Complete the clone and the original request.
* Must be called without clone's queue lock held,
* see end_clone_request() for more details.
*/
static void dm_end_request(struct request *clone, blk_status_t error)
{
struct dm_rq_target_io *tio = clone->end_io_data;
struct mapped_device *md = tio->md;
struct request *rq = tio->orig;
blk_rq_unprep_clone(clone);
tio->ti->type->release_clone_rq(clone, NULL);
rq_end_stats(md, rq);
blk_mq_end_request(rq, error);
rq_completed(md);
}
static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
{
blk-mq: Avoid that requeueing starts stopped queues Since blk_mq_requeue_work() starts stopped queues and since execution of this function can be scheduled after a queue has been stopped it is not possible to stop queues without using an additional state variable to track whether or not the queue has been stopped. Hence modify blk_mq_requeue_work() such that it does not start stopped queues. My conclusion after a review of the blk_mq_stop_hw_queues() and blk_mq_{delay_,}kick_requeue_list() callers is as follows: * In the dm driver starting and stopping queues should only happen if __dm_suspend() or __dm_resume() is called and not if the requeue list is processed. * In the SCSI core queue stopping and starting should only be performed by the scsi_internal_device_block() and scsi_internal_device_unblock() functions but not by any other function. Although the blk_mq_stop_hw_queue() call in scsi_queue_rq() may help to reduce CPU load if a LLD queue is full, figuring out whether or not a queue should be restarted when requeueing a command would require to introduce additional locking in scsi_mq_requeue_cmd() to avoid a race with scsi_internal_device_block(). Avoid this complexity by removing the blk_mq_stop_hw_queue() call from scsi_queue_rq(). * In the NVMe core only the functions that call blk_mq_start_stopped_hw_queues() explicitly should start stopped queues. * A blk_mq_start_stopped_hwqueues() call must be added in the xen-blkfront driver in its blkif_recover() function. Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Roger Pau Monné <roger.pau@citrix.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: James Bottomley <jejb@linux.vnet.ibm.com> Cc: Martin K. Petersen <martin.petersen@oracle.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@fb.com>
2016-10-29 00:20:32 +00:00
blk_mq_delay_kick_requeue_list(q, msecs);
}
void dm_mq_kick_requeue_list(struct mapped_device *md)
{
__dm_mq_kick_requeue_list(md->queue, 0);
}
EXPORT_SYMBOL(dm_mq_kick_requeue_list);
static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
{
blk_mq_requeue_request(rq, false);
__dm_mq_kick_requeue_list(rq->q, msecs);
}
static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
{
struct mapped_device *md = tio->md;
struct request *rq = tio->orig;
unsigned long delay_ms = delay_requeue ? 100 : 0;
rq_end_stats(md, rq);
if (tio->clone) {
blk_rq_unprep_clone(tio->clone);
tio->ti->type->release_clone_rq(tio->clone, NULL);
}
dm_mq_delay_requeue_request(rq, delay_ms);
rq_completed(md);
}
static void dm_done(struct request *clone, blk_status_t error, bool mapped)
{
int r = DM_ENDIO_DONE;
struct dm_rq_target_io *tio = clone->end_io_data;
dm_request_endio_fn rq_end_io = NULL;
if (tio->ti) {
rq_end_io = tio->ti->type->rq_end_io;
if (mapped && rq_end_io)
r = rq_end_io(tio->ti, clone, error, &tio->info);
}
if (unlikely(error == BLK_STS_TARGET)) {
dm: disable DISCARD if the underlying storage no longer supports it Storage devices which report supporting discard commands like WRITE_SAME_16 with unmap, but reject discard commands sent to the storage device. This is a clear storage firmware bug but it doesn't change the fact that should a program cause discards to be sent to a multipath device layered on this buggy storage, all paths can end up failed at the same time from the discards, causing possible I/O loss. The first discard to a path will fail with Illegal Request, Invalid field in cdb, e.g.: kernel: sd 8:0:8:19: [sdfn] tag#0 FAILED Result: hostbyte=DID_OK driverbyte=DRIVER_SENSE kernel: sd 8:0:8:19: [sdfn] tag#0 Sense Key : Illegal Request [current] kernel: sd 8:0:8:19: [sdfn] tag#0 Add. Sense: Invalid field in cdb kernel: sd 8:0:8:19: [sdfn] tag#0 CDB: Write same(16) 93 08 00 00 00 00 00 a0 08 00 00 00 80 00 00 00 kernel: blk_update_request: critical target error, dev sdfn, sector 10487808 The SCSI layer converts this to the BLK_STS_TARGET error number, the sd device disables its support for discard on this path, and because of the BLK_STS_TARGET error multipath fails the discard without failing any path or retrying down a different path. But subsequent discards can cause path failures. Any discards sent to the path which already failed a discard ends up failing with EIO from blk_cloned_rq_check_limits with an "over max size limit" error since the discard limit was set to 0 by the sd driver for the path. As the error is EIO, this now fails the path and multipath tries to send the discard down the next path. This cycle continues as discards are sent until all paths fail. Fix this by training DM core to disable DISCARD if the underlying storage already did so. Also, fix branching in dm_done() and clone_endio() to reflect the mutually exclussive nature of the IO operations in question. Cc: stable@vger.kernel.org Reported-by: David Jeffery <djeffery@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2019-04-03 16:23:11 +00:00
if (req_op(clone) == REQ_OP_DISCARD &&
!clone->q->limits.max_discard_sectors)
disable_discard(tio->md);
else if (req_op(clone) == REQ_OP_WRITE_SAME &&
!clone->q->limits.max_write_same_sectors)
disable_write_same(tio->md);
dm: disable DISCARD if the underlying storage no longer supports it Storage devices which report supporting discard commands like WRITE_SAME_16 with unmap, but reject discard commands sent to the storage device. This is a clear storage firmware bug but it doesn't change the fact that should a program cause discards to be sent to a multipath device layered on this buggy storage, all paths can end up failed at the same time from the discards, causing possible I/O loss. The first discard to a path will fail with Illegal Request, Invalid field in cdb, e.g.: kernel: sd 8:0:8:19: [sdfn] tag#0 FAILED Result: hostbyte=DID_OK driverbyte=DRIVER_SENSE kernel: sd 8:0:8:19: [sdfn] tag#0 Sense Key : Illegal Request [current] kernel: sd 8:0:8:19: [sdfn] tag#0 Add. Sense: Invalid field in cdb kernel: sd 8:0:8:19: [sdfn] tag#0 CDB: Write same(16) 93 08 00 00 00 00 00 a0 08 00 00 00 80 00 00 00 kernel: blk_update_request: critical target error, dev sdfn, sector 10487808 The SCSI layer converts this to the BLK_STS_TARGET error number, the sd device disables its support for discard on this path, and because of the BLK_STS_TARGET error multipath fails the discard without failing any path or retrying down a different path. But subsequent discards can cause path failures. Any discards sent to the path which already failed a discard ends up failing with EIO from blk_cloned_rq_check_limits with an "over max size limit" error since the discard limit was set to 0 by the sd driver for the path. As the error is EIO, this now fails the path and multipath tries to send the discard down the next path. This cycle continues as discards are sent until all paths fail. Fix this by training DM core to disable DISCARD if the underlying storage already did so. Also, fix branching in dm_done() and clone_endio() to reflect the mutually exclussive nature of the IO operations in question. Cc: stable@vger.kernel.org Reported-by: David Jeffery <djeffery@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2019-04-03 16:23:11 +00:00
else if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
!clone->q->limits.max_write_zeroes_sectors)
disable_write_zeroes(tio->md);
}
switch (r) {
case DM_ENDIO_DONE:
/* The target wants to complete the I/O */
dm_end_request(clone, error);
break;
case DM_ENDIO_INCOMPLETE:
/* The target will handle the I/O */
return;
case DM_ENDIO_REQUEUE:
/* The target wants to requeue the I/O */
dm_requeue_original_request(tio, false);
break;
case DM_ENDIO_DELAY_REQUEUE:
/* The target wants to requeue the I/O after a delay */
dm_requeue_original_request(tio, true);
break;
default:
DMWARN("unimplemented target endio return value: %d", r);
BUG();
}
}
/*
* Request completion handler for request-based dm
*/
static void dm_softirq_done(struct request *rq)
{
bool mapped = true;
struct dm_rq_target_io *tio = tio_from_request(rq);
struct request *clone = tio->clone;
if (!clone) {
struct mapped_device *md = tio->md;
rq_end_stats(md, rq);
blk_mq_end_request(rq, tio->error);
rq_completed(md);
return;
}
if (rq->rq_flags & RQF_FAILED)
mapped = false;
dm_done(clone, tio->error, mapped);
}
/*
* Complete the clone and the original request with the error status
* through softirq context.
*/
static void dm_complete_request(struct request *rq, blk_status_t error)
{
struct dm_rq_target_io *tio = tio_from_request(rq);
tio->error = error;
if (likely(!blk_should_fake_timeout(rq->q)))
blk_mq_complete_request(rq);
}
/*
* Complete the not-mapped clone and the original request with the error status
* through softirq context.
* Target's rq_end_io() function isn't called.
* This may be used when the target's clone_and_map_rq() function fails.
*/
static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
{
rq->rq_flags |= RQF_FAILED;
dm_complete_request(rq, error);
}
static void end_clone_request(struct request *clone, blk_status_t error)
{
struct dm_rq_target_io *tio = clone->end_io_data;
dm_complete_request(tio->orig, error);
}
blk-mq: improve DM's blk-mq IO merging via blk_insert_cloned_request feedback blk_insert_cloned_request() is called in the fast path of a dm-rq driver (e.g. blk-mq request-based DM mpath). blk_insert_cloned_request() uses blk_mq_request_bypass_insert() to directly append the request to the blk-mq hctx->dispatch_list of the underlying queue. 1) This way isn't efficient enough because the hctx spinlock is always used. 2) With blk_insert_cloned_request(), we completely bypass underlying queue's elevator and depend on the upper-level dm-rq driver's elevator to schedule IO. But dm-rq currently can't get the underlying queue's dispatch feedback at all. Without knowing whether a request was issued or not (e.g. due to underlying queue being busy) the dm-rq elevator will not be able to provide effective IO merging (as a side-effect of dm-rq currently blindly destaging a request from its elevator only to requeue it after a delay, which kills any opportunity for merging). This obviously causes very bad sequential IO performance. Fix this by updating blk_insert_cloned_request() to use blk_mq_request_direct_issue(). blk_mq_request_direct_issue() allows a request to be issued directly to the underlying queue and returns the dispatch feedback (blk_status_t). If blk_mq_request_direct_issue() returns BLK_SYS_RESOURCE the dm-rq driver will now use DM_MAPIO_REQUEUE to _not_ destage the request. Whereby preserving the opportunity to merge IO. With this, request-based DM's blk-mq sequential IO performance is vastly improved (as much as 3X in mpath/virtio-scsi testing). Signed-off-by: Ming Lei <ming.lei@redhat.com> [blk-mq.c changes heavily influenced by Ming Lei's initial solution, but they were refactored to make them less fragile and easier to read/review] Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-17 16:25:57 +00:00
static blk_status_t dm_dispatch_clone_request(struct request *clone, struct request *rq)
{
blk_status_t r;
if (blk_queue_io_stat(clone->q))
clone->rq_flags |= RQF_IO_STAT;
clone->start_time_ns = ktime_get_ns();
r = blk_insert_cloned_request(clone->q, clone);
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
if (r != BLK_STS_OK && r != BLK_STS_RESOURCE && r != BLK_STS_DEV_RESOURCE)
/* must complete clone in terms of original request */
dm_complete_request(rq, r);
blk-mq: improve DM's blk-mq IO merging via blk_insert_cloned_request feedback blk_insert_cloned_request() is called in the fast path of a dm-rq driver (e.g. blk-mq request-based DM mpath). blk_insert_cloned_request() uses blk_mq_request_bypass_insert() to directly append the request to the blk-mq hctx->dispatch_list of the underlying queue. 1) This way isn't efficient enough because the hctx spinlock is always used. 2) With blk_insert_cloned_request(), we completely bypass underlying queue's elevator and depend on the upper-level dm-rq driver's elevator to schedule IO. But dm-rq currently can't get the underlying queue's dispatch feedback at all. Without knowing whether a request was issued or not (e.g. due to underlying queue being busy) the dm-rq elevator will not be able to provide effective IO merging (as a side-effect of dm-rq currently blindly destaging a request from its elevator only to requeue it after a delay, which kills any opportunity for merging). This obviously causes very bad sequential IO performance. Fix this by updating blk_insert_cloned_request() to use blk_mq_request_direct_issue(). blk_mq_request_direct_issue() allows a request to be issued directly to the underlying queue and returns the dispatch feedback (blk_status_t). If blk_mq_request_direct_issue() returns BLK_SYS_RESOURCE the dm-rq driver will now use DM_MAPIO_REQUEUE to _not_ destage the request. Whereby preserving the opportunity to merge IO. With this, request-based DM's blk-mq sequential IO performance is vastly improved (as much as 3X in mpath/virtio-scsi testing). Signed-off-by: Ming Lei <ming.lei@redhat.com> [blk-mq.c changes heavily influenced by Ming Lei's initial solution, but they were refactored to make them less fragile and easier to read/review] Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-17 16:25:57 +00:00
return r;
}
static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
void *data)
{
struct dm_rq_target_io *tio = data;
struct dm_rq_clone_bio_info *info =
container_of(bio, struct dm_rq_clone_bio_info, clone);
info->orig = bio_orig;
info->tio = tio;
bio->bi_end_io = end_clone_bio;
return 0;
}
static int setup_clone(struct request *clone, struct request *rq,
struct dm_rq_target_io *tio, gfp_t gfp_mask)
{
int r;
r = blk_rq_prep_clone(clone, rq, &tio->md->bs, gfp_mask,
dm_rq_bio_constructor, tio);
if (r)
return r;
clone->end_io = end_clone_request;
clone->end_io_data = tio;
tio->clone = clone;
return 0;
}
static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
struct mapped_device *md)
{
tio->md = md;
tio->ti = NULL;
tio->clone = NULL;
tio->orig = rq;
tio->error = 0;
tio->completed = 0;
/*
* Avoid initializing info for blk-mq; it passes
* target-specific data through info.ptr
* (see: dm_mq_init_request)
*/
if (!md->init_tio_pdu)
memset(&tio->info, 0, sizeof(tio->info));
}
/*
* Returns:
* DM_MAPIO_* : the request has been processed as indicated
* DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
* < 0 : the request was completed due to failure
*/
static int map_request(struct dm_rq_target_io *tio)
{
int r;
struct dm_target *ti = tio->ti;
struct mapped_device *md = tio->md;
struct request *rq = tio->orig;
struct request *clone = NULL;
blk-mq: improve DM's blk-mq IO merging via blk_insert_cloned_request feedback blk_insert_cloned_request() is called in the fast path of a dm-rq driver (e.g. blk-mq request-based DM mpath). blk_insert_cloned_request() uses blk_mq_request_bypass_insert() to directly append the request to the blk-mq hctx->dispatch_list of the underlying queue. 1) This way isn't efficient enough because the hctx spinlock is always used. 2) With blk_insert_cloned_request(), we completely bypass underlying queue's elevator and depend on the upper-level dm-rq driver's elevator to schedule IO. But dm-rq currently can't get the underlying queue's dispatch feedback at all. Without knowing whether a request was issued or not (e.g. due to underlying queue being busy) the dm-rq elevator will not be able to provide effective IO merging (as a side-effect of dm-rq currently blindly destaging a request from its elevator only to requeue it after a delay, which kills any opportunity for merging). This obviously causes very bad sequential IO performance. Fix this by updating blk_insert_cloned_request() to use blk_mq_request_direct_issue(). blk_mq_request_direct_issue() allows a request to be issued directly to the underlying queue and returns the dispatch feedback (blk_status_t). If blk_mq_request_direct_issue() returns BLK_SYS_RESOURCE the dm-rq driver will now use DM_MAPIO_REQUEUE to _not_ destage the request. Whereby preserving the opportunity to merge IO. With this, request-based DM's blk-mq sequential IO performance is vastly improved (as much as 3X in mpath/virtio-scsi testing). Signed-off-by: Ming Lei <ming.lei@redhat.com> [blk-mq.c changes heavily influenced by Ming Lei's initial solution, but they were refactored to make them less fragile and easier to read/review] Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-17 16:25:57 +00:00
blk_status_t ret;
r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
switch (r) {
case DM_MAPIO_SUBMITTED:
/* The target has taken the I/O to submit by itself later */
break;
case DM_MAPIO_REMAPPED:
if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
/* -ENOMEM */
ti->type->release_clone_rq(clone, &tio->info);
return DM_MAPIO_REQUEUE;
}
/* The target has remapped the I/O so dispatch it */
trace_block_rq_remap(clone, disk_devt(dm_disk(md)),
blk_rq_pos(rq));
blk-mq: improve DM's blk-mq IO merging via blk_insert_cloned_request feedback blk_insert_cloned_request() is called in the fast path of a dm-rq driver (e.g. blk-mq request-based DM mpath). blk_insert_cloned_request() uses blk_mq_request_bypass_insert() to directly append the request to the blk-mq hctx->dispatch_list of the underlying queue. 1) This way isn't efficient enough because the hctx spinlock is always used. 2) With blk_insert_cloned_request(), we completely bypass underlying queue's elevator and depend on the upper-level dm-rq driver's elevator to schedule IO. But dm-rq currently can't get the underlying queue's dispatch feedback at all. Without knowing whether a request was issued or not (e.g. due to underlying queue being busy) the dm-rq elevator will not be able to provide effective IO merging (as a side-effect of dm-rq currently blindly destaging a request from its elevator only to requeue it after a delay, which kills any opportunity for merging). This obviously causes very bad sequential IO performance. Fix this by updating blk_insert_cloned_request() to use blk_mq_request_direct_issue(). blk_mq_request_direct_issue() allows a request to be issued directly to the underlying queue and returns the dispatch feedback (blk_status_t). If blk_mq_request_direct_issue() returns BLK_SYS_RESOURCE the dm-rq driver will now use DM_MAPIO_REQUEUE to _not_ destage the request. Whereby preserving the opportunity to merge IO. With this, request-based DM's blk-mq sequential IO performance is vastly improved (as much as 3X in mpath/virtio-scsi testing). Signed-off-by: Ming Lei <ming.lei@redhat.com> [blk-mq.c changes heavily influenced by Ming Lei's initial solution, but they were refactored to make them less fragile and easier to read/review] Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-17 16:25:57 +00:00
ret = dm_dispatch_clone_request(clone, rq);
blk-mq: introduce BLK_STS_DEV_RESOURCE This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-31 03:04:57 +00:00
if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
blk-mq: improve DM's blk-mq IO merging via blk_insert_cloned_request feedback blk_insert_cloned_request() is called in the fast path of a dm-rq driver (e.g. blk-mq request-based DM mpath). blk_insert_cloned_request() uses blk_mq_request_bypass_insert() to directly append the request to the blk-mq hctx->dispatch_list of the underlying queue. 1) This way isn't efficient enough because the hctx spinlock is always used. 2) With blk_insert_cloned_request(), we completely bypass underlying queue's elevator and depend on the upper-level dm-rq driver's elevator to schedule IO. But dm-rq currently can't get the underlying queue's dispatch feedback at all. Without knowing whether a request was issued or not (e.g. due to underlying queue being busy) the dm-rq elevator will not be able to provide effective IO merging (as a side-effect of dm-rq currently blindly destaging a request from its elevator only to requeue it after a delay, which kills any opportunity for merging). This obviously causes very bad sequential IO performance. Fix this by updating blk_insert_cloned_request() to use blk_mq_request_direct_issue(). blk_mq_request_direct_issue() allows a request to be issued directly to the underlying queue and returns the dispatch feedback (blk_status_t). If blk_mq_request_direct_issue() returns BLK_SYS_RESOURCE the dm-rq driver will now use DM_MAPIO_REQUEUE to _not_ destage the request. Whereby preserving the opportunity to merge IO. With this, request-based DM's blk-mq sequential IO performance is vastly improved (as much as 3X in mpath/virtio-scsi testing). Signed-off-by: Ming Lei <ming.lei@redhat.com> [blk-mq.c changes heavily influenced by Ming Lei's initial solution, but they were refactored to make them less fragile and easier to read/review] Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-17 16:25:57 +00:00
blk_rq_unprep_clone(clone);
blk-mq: add callback of .cleanup_rq SCSI maintains its own driver private data hooked off of each SCSI request, and the pridate data won't be freed after scsi_queue_rq() returns BLK_STS_RESOURCE or BLK_STS_DEV_RESOURCE. An upper layer driver (e.g. dm-rq) may need to retry these SCSI requests, before SCSI has fully dispatched them, due to a lower level SCSI driver's resource limitation identified in scsi_queue_rq(). Currently SCSI's per-request private data is leaked when the upper layer driver (dm-rq) frees and then retries these requests in response to BLK_STS_RESOURCE or BLK_STS_DEV_RESOURCE returns from scsi_queue_rq(). This usecase is so specialized that it doesn't warrant training an existing blk-mq interface (e.g. blk_mq_free_request) to allow SCSI to account for freeing its driver private data -- doing so would add an extra branch for handling a special case that all other consumers of SCSI (and blk-mq) won't ever need to worry about. So the most pragmatic way forward is to delegate freeing SCSI driver private data to the upper layer driver (dm-rq). Do so by adding new .cleanup_rq callback and calling a new blk_mq_cleanup_rq() method from dm-rq. A following commit will implement the .cleanup_rq() hook in scsi_mq_ops. Cc: Ewan D. Milne <emilne@redhat.com> Cc: Bart Van Assche <bvanassche@acm.org> Cc: Hannes Reinecke <hare@suse.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: <stable@vger.kernel.org> Fixes: 396eaf21ee17 ("blk-mq: improve DM's blk-mq IO merging via blk_insert_cloned_request feedback") Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-07-25 02:04:59 +00:00
blk_mq_cleanup_rq(clone);
tio->ti->type->release_clone_rq(clone, &tio->info);
blk-mq: improve DM's blk-mq IO merging via blk_insert_cloned_request feedback blk_insert_cloned_request() is called in the fast path of a dm-rq driver (e.g. blk-mq request-based DM mpath). blk_insert_cloned_request() uses blk_mq_request_bypass_insert() to directly append the request to the blk-mq hctx->dispatch_list of the underlying queue. 1) This way isn't efficient enough because the hctx spinlock is always used. 2) With blk_insert_cloned_request(), we completely bypass underlying queue's elevator and depend on the upper-level dm-rq driver's elevator to schedule IO. But dm-rq currently can't get the underlying queue's dispatch feedback at all. Without knowing whether a request was issued or not (e.g. due to underlying queue being busy) the dm-rq elevator will not be able to provide effective IO merging (as a side-effect of dm-rq currently blindly destaging a request from its elevator only to requeue it after a delay, which kills any opportunity for merging). This obviously causes very bad sequential IO performance. Fix this by updating blk_insert_cloned_request() to use blk_mq_request_direct_issue(). blk_mq_request_direct_issue() allows a request to be issued directly to the underlying queue and returns the dispatch feedback (blk_status_t). If blk_mq_request_direct_issue() returns BLK_SYS_RESOURCE the dm-rq driver will now use DM_MAPIO_REQUEUE to _not_ destage the request. Whereby preserving the opportunity to merge IO. With this, request-based DM's blk-mq sequential IO performance is vastly improved (as much as 3X in mpath/virtio-scsi testing). Signed-off-by: Ming Lei <ming.lei@redhat.com> [blk-mq.c changes heavily influenced by Ming Lei's initial solution, but they were refactored to make them less fragile and easier to read/review] Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-17 16:25:57 +00:00
tio->clone = NULL;
return DM_MAPIO_REQUEUE;
blk-mq: improve DM's blk-mq IO merging via blk_insert_cloned_request feedback blk_insert_cloned_request() is called in the fast path of a dm-rq driver (e.g. blk-mq request-based DM mpath). blk_insert_cloned_request() uses blk_mq_request_bypass_insert() to directly append the request to the blk-mq hctx->dispatch_list of the underlying queue. 1) This way isn't efficient enough because the hctx spinlock is always used. 2) With blk_insert_cloned_request(), we completely bypass underlying queue's elevator and depend on the upper-level dm-rq driver's elevator to schedule IO. But dm-rq currently can't get the underlying queue's dispatch feedback at all. Without knowing whether a request was issued or not (e.g. due to underlying queue being busy) the dm-rq elevator will not be able to provide effective IO merging (as a side-effect of dm-rq currently blindly destaging a request from its elevator only to requeue it after a delay, which kills any opportunity for merging). This obviously causes very bad sequential IO performance. Fix this by updating blk_insert_cloned_request() to use blk_mq_request_direct_issue(). blk_mq_request_direct_issue() allows a request to be issued directly to the underlying queue and returns the dispatch feedback (blk_status_t). If blk_mq_request_direct_issue() returns BLK_SYS_RESOURCE the dm-rq driver will now use DM_MAPIO_REQUEUE to _not_ destage the request. Whereby preserving the opportunity to merge IO. With this, request-based DM's blk-mq sequential IO performance is vastly improved (as much as 3X in mpath/virtio-scsi testing). Signed-off-by: Ming Lei <ming.lei@redhat.com> [blk-mq.c changes heavily influenced by Ming Lei's initial solution, but they were refactored to make them less fragile and easier to read/review] Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-01-17 16:25:57 +00:00
}
break;
case DM_MAPIO_REQUEUE:
/* The target wants to requeue the I/O */
break;
case DM_MAPIO_DELAY_REQUEUE:
/* The target wants to requeue the I/O after a delay */
dm_requeue_original_request(tio, true);
break;
case DM_MAPIO_KILL:
/* The target wants to complete the I/O */
dm_kill_unmapped_request(rq, BLK_STS_IOERR);
break;
default:
DMWARN("unimplemented target map return value: %d", r);
BUG();
}
return r;
}
/* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */
ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
{
return sprintf(buf, "%u\n", 0);
}
ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
const char *buf, size_t count)
{
return count;
}
static void dm_start_request(struct mapped_device *md, struct request *orig)
{
blk_mq_start_request(orig);
if (unlikely(dm_stats_used(&md->stats))) {
struct dm_rq_target_io *tio = tio_from_request(orig);
tio->duration_jiffies = jiffies;
tio->n_sectors = blk_rq_sectors(orig);
dm_stats_account_io(&md->stats, rq_data_dir(orig),
blk_rq_pos(orig), tio->n_sectors, false, 0,
&tio->stats_aux);
}
/*
* Hold the md reference here for the in-flight I/O.
* We can't rely on the reference count by device opener,
* because the device may be closed during the request completion
* when all bios are completed.
* See the comment in rq_completed() too.
*/
dm_get(md);
}
static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
unsigned int hctx_idx, unsigned int numa_node)
{
struct mapped_device *md = set->driver_data;
struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
/*
* Must initialize md member of tio, otherwise it won't
* be available in dm_mq_queue_rq.
*/
tio->md = md;
if (md->init_tio_pdu) {
/* target-specific per-io data is immediately after the tio */
tio->info.ptr = tio + 1;
}
return 0;
}
static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
const struct blk_mq_queue_data *bd)
{
struct request *rq = bd->rq;
struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
struct mapped_device *md = tio->md;
struct dm_target *ti = md->immutable_target;
if (unlikely(!ti)) {
int srcu_idx;
struct dm_table *map = dm_get_live_table(md, &srcu_idx);
ti = dm_table_find_target(map, 0);
dm_put_live_table(md, srcu_idx);
}
if (ti->type->busy && ti->type->busy(ti))
return BLK_STS_RESOURCE;
dm_start_request(md, rq);
/* Init tio using md established in .init_request */
init_tio(tio, rq, md);
/*
* Establish tio->ti before calling map_request().
*/
tio->ti = ti;
/* Direct call is fine since .queue_rq allows allocations */
if (map_request(tio) == DM_MAPIO_REQUEUE) {
/* Undo dm_start_request() before requeuing */
rq_end_stats(md, rq);
rq_completed(md);
return BLK_STS_RESOURCE;
}
return BLK_STS_OK;
}
static const struct blk_mq_ops dm_mq_ops = {
.queue_rq = dm_mq_queue_rq,
.complete = dm_softirq_done,
.init_request = dm_mq_init_request,
};
int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
{
struct request_queue *q;
struct dm_target *immutable_tgt;
int err;
md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
if (!md->tag_set)
return -ENOMEM;
md->tag_set->ops = &dm_mq_ops;
md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
md->tag_set->numa_node = md->numa_node_id;
blk-mq: drain I/O when all CPUs in a hctx are offline Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup up queue mapping. Thomas mentioned the following point[1]: "That was the constraint of managed interrupts from the very beginning: The driver/subsystem has to quiesce the interrupt line and the associated queue _before_ it gets shutdown in CPU unplug and not fiddle with it until it's restarted by the core when the CPU is plugged in again." However, current blk-mq implementation doesn't quiesce hw queue before the last CPU in the hctx is shutdown. Even worse, CPUHP_BLK_MQ_DEAD is a cpuhp state handled after the CPU is down, so there isn't any chance to quiesce the hctx before shutting down the CPU. Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs where the last CPU goes away, and wait for completion of in-flight requests. This guarantees that there is no inflight I/O before shutting down the managed IRQ. Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need to wait for completion of in-flight requests from these drivers to avoid a potential dead-lock. It is safe to do this for stacking drivers as those do not use interrupts at all and their I/O completions are triggered by underlying devices I/O completion. [1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/ [hch: different retry mechanism, merged two patches, minor cleanups] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Hannes Reinecke <hare@suse.de> Reviewed-by: Daniel Wagner <dwagner@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-05-29 13:53:15 +00:00
md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_STACKING;
md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
md->tag_set->driver_data = md;
md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
immutable_tgt = dm_table_get_immutable_target(t);
if (immutable_tgt && immutable_tgt->per_io_data_size) {
/* any target-specific per-io data is immediately after the tio */
md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
md->init_tio_pdu = true;
}
err = blk_mq_alloc_tag_set(md->tag_set);
if (err)
goto out_kfree_tag_set;
block: Delay default elevator initialization When elevator_init_mq() is called from blk_mq_init_allocated_queue(), the only information known about the device is the number of hardware queues as the block device scan by the device driver is not completed yet for most drivers. The device type and elevator required features are not set yet, preventing to correctly select the default elevator most suitable for the device. This currently affects all multi-queue zoned block devices which default to the "none" elevator instead of the required "mq-deadline" elevator. These drives currently include host-managed SMR disks connected to a smartpqi HBA and null_blk block devices with zoned mode enabled. Upcoming NVMe Zoned Namespace devices will also be affected. Fix this by adding the boolean elevator_init argument to blk_mq_init_allocated_queue() to control the execution of elevator_init_mq(). Two cases exist: 1) elevator_init = false is used for calls to blk_mq_init_allocated_queue() within blk_mq_init_queue(). In this case, a call to elevator_init_mq() is added to __device_add_disk(), resulting in the delayed initialization of the queue elevator after the device driver finished probing the device information. This effectively allows elevator_init_mq() access to more information about the device. 2) elevator_init = true preserves the current behavior of initializing the elevator directly from blk_mq_init_allocated_queue(). This case is used for the special request based DM devices where the device gendisk is created before the queue initialization and device information (e.g. queue limits) is already known when the queue initialization is executed. Additionally, to make sure that the elevator initialization is never done while requests are in-flight (there should be none when the device driver calls device_add_disk()), freeze and quiesce the device request queue before calling blk_mq_init_sched() in elevator_init_mq(). Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-09-05 09:51:33 +00:00
q = blk_mq_init_allocated_queue(md->tag_set, md->queue, true);
if (IS_ERR(q)) {
err = PTR_ERR(q);
goto out_tag_set;
}
return 0;
out_tag_set:
blk_mq_free_tag_set(md->tag_set);
out_kfree_tag_set:
kfree(md->tag_set);
return err;
}
void dm_mq_cleanup_mapped_device(struct mapped_device *md)
{
if (md->tag_set) {
blk_mq_free_tag_set(md->tag_set);
kfree(md->tag_set);
}
}
module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
/* Unused, but preserved for userspace compatibility */
static bool use_blk_mq = true;
module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");