Improve performance for BLK_MQ_F_BLOCKING drivers
Subsystems (2)
v3 cover Improve performance for BLK_MQ_F_BLOCKING drivers
Hi Jens,
Request queues are run asynchronously if BLK_MQ_F_BLOCKING has been set. This results in suboptimal performance. This patch series improves performance if BLK_MQ_F_BLOCKING has been set by running request queues synchronously whenever possible.
Please consider this patch series for the next merge window.
Thanks,
Bart.
Changes compared to v2:
- Improved the descriptions of patches 1/3 and 2/3.
- Added a comment in patch 2/3.
- In patch 3/3, included a change for blk_execute_rq_nowait() since I noticed that this function may be called from a context where it is not allowed to sleep (nvme_uring_cmd_io() for NVMe over TCP).
Changes compared to v1:
- Fixed support for the combination BLK_MQ_F_BLOCKING + BLIST_SINGLELUN.
Bart Van Assche (3):
scsi: Inline scsi_kick_queue()
scsi: Remove a blk_mq_run_hw_queues() call
block: Improve performance for BLK_MQ_F_BLOCKING drivers
block/blk-mq.c | 17 +++++++++++------ drivers/scsi/scsi_lib.c | 12 ++++-------- 2 files changed, 15 insertions(+), 14 deletions(-)
PATCH v3 1/3 scsi: Inline scsi_kick_queue()
Inline scsi_kick_queue() to prepare for modifying the second argument passed to blk_mq_run_hw_queues().
--- drivers/scsi/scsi_lib.c | 9 ++------- 1 file changed, 2 insertions(+), 7 deletions(-)
drivers/scsi/scsi_lib.c +2 -7
@@ -300,11 +300,6 @@ void scsi_device_unbusy(struct scsi_device *sdev, struct scsi_cmnd *cmd) cmd->budget_token =-1 ; } -static void scsi_kick_queue (struct request_queue *q) -{ -blk_mq_run_hw_queues (q, false); -} -/* * Kick the queue of SCSI device @sdev if @sdev != current_sdev. Called with * interrupts disabled.
@@ -340,7 +335,7 @@ static void scsi_single_lun_run(struct scsi_device *current_sdev)* but in most cases, we will be first. Ideally, each LU on the * target would get some limited time or requests on the target. */ -scsi_kick_queue (current_sdev->request_queue); +blk_mq_run_hw_queues (current_sdev->request_queue, false);spin_lock_irqsave (shost->host_lock, flags);if (!starget->starget_sdev_user)
@@ -427,7 +422,7 @@ static void scsi_starved_list_run(struct Scsi_Host *shost)continue ;spin_unlock_irqrestore (shost->host_lock, flags); -scsi_kick_queue (slq); +blk_mq_run_hw_queues (slq, false);blk_put_queue (slq);spin_lock_irqsave (shost->host_lock, flags);
Re: [PATCH v3 1/3] scsi: Inline scsi_kick_queue()
Bart,
Quoted 2 lines
Inline
scsi_kick_queue()to prepare for modifying the second argument passed toblk_mq_run_hw_queues().
Signature · 1 line
Martin K. Petersen Oracle Linux Engineering
PATCH v3 2/3 scsi: Remove a blk_mq_run_hw_queues() call
In reply to Bart Van Asscheblk_mq_kick_requeue_list() calls blk_mq_run_hw_queues() asynchronously. Leave out the direct blk_mq_run_hw_queues() call. This patch causes scsi_run_queue() to call blk_mq_run_hw_queues() asynchronously instead of synchronously. Since scsi_run_queue() is not called from the hot I/O submission path, this patch does not affect the hot path.
This patch prepares for allowing blk_mq_run_hw_queue() to sleep if BLK_MQ_F_BLOCKING has been set. scsi_run_queue() may be called from atomic context and must not sleep. Hence the removal of the blk_mq_run_hw_queues(q, false) call. See also scsi_unblock_requests().
--- drivers/scsi/scsi_lib.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/scsi/scsi_lib.c +1 -1
@@ -447,8 +447,8 @@ static void scsi_run_queue(struct request_queue *q)if (!list_empty (&sdev->host->starved_list))scsi_starved_list_run (sdev->host); +/* Note: blk_mq_kick_requeue_list() runs the queue asynchronously. */ blk_mq_kick_requeue_list (q); -blk_mq_run_hw_queues (q, false); }void scsi_requeue_run_queue (struct work_struct *work)
Re: [PATCH v3 2/3] scsi: Remove a blk_mq_run_hw_queues() call
Bart,
Quoted 10 lines
blk_mq_kick_requeue_list()callsblk_mq_run_hw_queues()asynchronously. Leave out the directblk_mq_run_hw_queues()call. This patch causesscsi_run_queue()to callblk_mq_run_hw_queues()asynchronously instead of synchronously. Sincescsi_run_queue()is not called from the hot I/O submission path, this patch does not affect the hot path.This patch prepares for allowing
blk_mq_run_hw_queue()to sleep ifBLK_MQ_F_BLOCKINGhas been set.scsi_run_queue()may be called from atomic context and must not sleep. Hence the removal of theblk_mq_run_hw_queues(q, false)call. See alsoscsi_unblock_requests().
Signature · 1 line
Martin K. Petersen Oracle Linux Engineering
PATCH v3 3/3 block: Improve performance for BLK_MQ_F_BLOCKING drivers
In reply to Bart Van Asscheblk_mq_run_queue() runs the queue asynchronously if BLK_MQ_F_BLOCKING has been set. This is suboptimal since running the queue asynchronously is slower than running the queue synchronously. This patch modifies blk_mq_run_queue() as follows if BLK_MQ_F_BLOCKING has been set:
- Run the queue synchronously if it is allowed to sleep.
- Run the queue asynchronously if it is not allowed to sleep. Additionally, blk_mq_run_hw_queue(hctx, false) calls are modified into blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING) if the caller may be invoked from atomic context.
The following caller chains have been reviewed:
blk_mq_run_hw_queue(hctx, false)
blk_mq_get_tag() /* may sleep, hence the functions it calls may also sleep */
blk_execute_rq() /* may sleep */
blk_mq_run_hw_queues(q, async=false)
blk_freeze_queue_start() /* may sleep */
blk_mq_requeue_work() /* may sleep */
scsi_kick_queue()
scsi_requeue_run_queue() /* may sleep */ scsi_run_host_queues()
scsi_ioctl_reset() /* may sleep */
blk_mq_insert_requests(hctx, ctx, list, run_queue_async=false)
blk_mq_dispatch_plug_list(plug, from_sched=false)
blk_mq_flush_plug_list(plug, from_schedule=false)
__blk_flush_plug(plug, from_schedule=false) blk_add_rq_to_plug()
blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */
blk_mq_plug_issue_direct()
blk_mq_flush_plug_list() /* see above */
blk_mq_dispatch_plug_list(plug, from_sched=false)
blk_mq_flush_plug_list() /* see above */
blk_mq_try_issue_directly()
blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */
blk_mq_try_issue_list_directly(hctx, list)
blk_mq_insert_requests() /* see above */
--- block/blk-mq.c | 16 ++++++++++------ drivers/scsi/scsi_lib.c | 3 ++- 2 files changed, 12 insertions(+), 7 deletions(-)
block/blk-mq.c +10 -6
@@ -1323,7 +1323,7 @@ void blk_execute_rq_nowait(struct request *rq, bool at_head) }blk_mq_insert_request (rq, at_head ?BLK_MQ_INSERT_AT_HEAD :0 ); -blk_mq_run_hw_queue (hctx, false); +blk_mq_run_hw_queue (hctx, hctx->flags &BLK_MQ_F_BLOCKING ); }EXPORT_SYMBOL_GPL (blk_execute_rq_nowait);
@@ -2222,6 +2222,8 @@ void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)*/ WARN_ON_ONCE (!async &&in_interrupt ()); +might_sleep_if (!async && hctx->flags &BLK_MQ_F_BLOCKING ); +/* * When queue is quiesced, we may be switching io scheduler, or * updating nr_hw_queues, or other things, and we can't run queue
@@ -2237,8 +2239,7 @@ void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)if (!need_run)return ; -if (async || (hctx->flags &BLK_MQ_F_BLOCKING ) || - !cpumask_test_cpu (raw_smp_processor_id (), hctx->cpumask)) { +if (async || !cpumask_test_cpu (raw_smp_processor_id (), hctx->cpumask)) {blk_mq_delay_run_hw_queue (hctx,0 );return ; }
@@ -2373,7 +2374,7 @@ void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) {clear_bit (BLK_MQ_S_STOPPED , &hctx->state); -blk_mq_run_hw_queue (hctx, false); +blk_mq_run_hw_queue (hctx, hctx->flags &BLK_MQ_F_BLOCKING ); }EXPORT_SYMBOL (blk_mq_start_hw_queue);
@@ -2403,7 +2404,8 @@ void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)unsigned long i;queue_for_each_hw_ctx (q, hctx, i) -blk_mq_start_stopped_hw_queue (hctx, async); +blk_mq_start_stopped_hw_queue (hctx, async || + (hctx->flags &BLK_MQ_F_BLOCKING )); }EXPORT_SYMBOL (blk_mq_start_stopped_hw_queues);
@@ -2461,6 +2463,8 @@ static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,list_for_each_entry (rq, list, queuelist) {BUG_ON (rq->mq_ctx != ctx);trace_block_rq_insert (rq); +if (rq->cmd_flags &REQ_NOWAIT ) + run_queue_async = true; }spin_lock (&ctx->lock);
@@ -2621,7 +2625,7 @@ static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,if ((rq->rq_flags &RQF_USE_SCHED ) || !blk_mq_get_budget_and_tag (rq)) {blk_mq_insert_request (rq,0 ); -blk_mq_run_hw_queue (hctx, false); +blk_mq_run_hw_queue (hctx, rq->cmd_flags &REQ_NOWAIT );return ; }
drivers/scsi/scsi_lib.c +2 -1
@@ -335,7 +335,8 @@ static void scsi_single_lun_run(struct scsi_device *current_sdev)* but in most cases, we will be first. Ideally, each LU on the * target would get some limited time or requests on the target. */ -blk_mq_run_hw_queues (current_sdev->request_queue, false); +blk_mq_run_hw_queues (current_sdev->request_queue, + shost->queuecommand_may_block);spin_lock_irqsave (shost->host_lock, flags);if (!starget->starget_sdev_user)
Quoted 8 lines · Bart Van Assche
Request queues are run asynchronously if
BLK_MQ_F_BLOCKINGhas been set. This results in suboptimal performance. This patch series improves performance ifBLK_MQ_F_BLOCKINGhas been set by running request queues synchronously whenever possible.Please consider this patch series for the next merge window.
[...]
Applied, thanks!
[1/3] scsi: Inline scsi_kick_queue()
commit: b5ca9acff553874aaf1faf176e076cbd7cc4aa0e
[2/3] scsi: Remove a blk_mq_run_hw_queues() call
commit: d42e2e3448a99c41c8489766eeb732d8d741d5be
[3/3] block: Improve performance for BLK_MQ_F_BLOCKING drivers
commit: 65a558f66c308251e256317957b75d1e643c33c3Best regards,
Signature · 1 line
Jens Axboe