blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues
Subsystems (2)
v2 cover blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues
Hello!
This patch series aims to fix a regression we've noticed on our test grid when support for multiple HW queues in megaraid_sas driver was added during the 5.10 cycle (103fbf8e4020 scsi: megaraid_sas: Added support for shared host tagset for cpuhotplug). The commit was reverted in the end for other reasons but I believe the fundamental problem still exists for any other similar setup. The problem manifests when the storage card supports multiple hardware queues however storage behind it is slow (single rotating disk in our case) and so using IO scheduler such as BFQ is desirable. See the second patch for details.
Honza
PATCH v2 1/2 Revert "blk-mq, elevator: Count requests per hctx to improve performance"
This reverts commit b445547ec1bbd3e7bf4b1c142550942f70527d95.
Since both mq-deadline and BFQ completely ignore hctx they are passed to their dispatch function and dispatch whatever request they deem fit checking whether any request for a particular hctx is queued is just pointless since we'll very likely get a request from a different hctx anyway. In the following commit we'll deal with lock contention in these IO schedulers in presence of multiple HW queues in a different way.
--- block/bfq-iosched.c | 5 ----- block/blk-mq.c | 1 - block/mq-deadline.c | 6 ------ include/linux/blk-mq.h | 4 ---- 4 files changed, 16 deletions(-)
block/bfq-iosched.c +0 -5
@@ -4640,9 +4640,6 @@ static bool bfq_has_work(struct blk_mq_hw_ctx *hctx) {struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; -if (!atomic_read (&hctx->elevator_queued)) -return false; -/* * Avoiding lock: a race on bfqd->busy_queues should cause at * most a call to dispatch for nothing
@@ -5557,7 +5554,6 @@ static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx, rq = list_first_entry(list, struct request, queuelist);list_del_init (&rq->queuelist);bfq_insert_request (hctx, rq, at_head); -atomic_inc (&hctx->elevator_queued); } }
@@ -5925,7 +5921,6 @@ static void bfq_finish_requeue_request(struct request *rq)bfq_completed_request (bfqq, bfqd);bfq_finish_requeue_request_body (bfqq); -atomic_dec (&rq->mq_hctx->elevator_queued);spin_unlock_irqrestore (&bfqd->lock, flags); }else {
block/blk-mq.c +0 -1
@@ -2653,7 +2653,6 @@ blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, goto free_hctx;atomic_set (&hctx->nr_active,0 ); -atomic_set (&hctx->elevator_queued,0 );if (node ==NUMA_NO_NODE ) node = set->numa_node; hctx->numa_node = node;
block/mq-deadline.c +0 -6
@@ -386,8 +386,6 @@ static struct request *dd_dispatch_request(struct blk_mq_hw_ctx *hctx)spin_lock (&dd->lock); rq =__dd_dispatch_request (dd);spin_unlock (&dd->lock); -if (rq) -atomic_dec (&rq->mq_hctx->elevator_queued);return rq; }
@@ -535,7 +533,6 @@ static void dd_insert_requests(struct blk_mq_hw_ctx *hctx, rq = list_first_entry(list, struct request, queuelist);list_del_init (&rq->queuelist);dd_insert_request (hctx, rq, at_head); -atomic_inc (&hctx->elevator_queued); }spin_unlock (&dd->lock); }
@@ -582,9 +579,6 @@ static bool dd_has_work(struct blk_mq_hw_ctx *hctx) {struct deadline_data *dd = hctx->queue->elevator->elevator_data; -if (!atomic_read (&hctx->elevator_queued)) -return false; -return !list_empty_careful (&dd->dispatch) || !list_empty_careful (&dd->fifo_list[0 ]) || !list_empty_careful (&dd->fifo_list[1 ]);
include/linux/blk-mq.h +0 -4
@@ -140,10 +140,6 @@ struct blk_mq_hw_ctx {* shared across request queues. */ atomic_t nr_active; -/** -* @elevator_queued: Number of queued requests on hctx. -*/ -atomic_t elevator_queued;/** @cpuhp_online: List to store request if CPU is going to die */ struct hlist_node cpuhp_online;
Signature · 1 line
2.26.2
PATCH v2 2/2 blk-mq: Improve performance of non-mq IO schedulers with multiple HW queues
In reply to Jan KaraCurrently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention.
Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler:
clients MQ SQ MQ-Patched
Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%*
Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%*
Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied.
You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance.
--- block/blk-mq.c | 62 ++++++++++++++++++++++++++++++++++------ block/kyber-iosched.c | 1 + include/linux/elevator.h | 2 ++ 3 files changed, 56 insertions(+), 9 deletions(-)
block/blk-mq.c +53 -9
@@ -63,15 +63,20 @@ static int blk_mq_poll_stats_bkt(const struct request *rq)return bucket; } +/* Check if there are requests queued in hctx lists. */ +static bool blk_mq_hctx_has_queued_rq (struct blk_mq_hw_ctx *hctx) +{ +return !list_empty_careful (&hctx->dispatch) || +sbitmap_any_bit_set (&hctx->ctx_map); +} +/* * Check if any of the ctx, dispatch list or elevator * have pending work in this hardware queue. */ static bool blk_mq_hctx_has_pending (struct blk_mq_hw_ctx *hctx) { -return !list_empty_careful (&hctx->dispatch) || -sbitmap_any_bit_set (&hctx->ctx_map) || -blk_mq_sched_has_work (hctx); +return blk_mq_hctx_has_queued_rq (hctx) ||blk_mq_sched_has_work (hctx); }/*
@@ -1646,6 +1651,31 @@ void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) }EXPORT_SYMBOL (blk_mq_run_hw_queue); +static struct blk_mq_hw_ctx *blk_mq_get_sqsched_hctx (struct request_queue *q) +{ +struct elevator_queue *e = q->elevator; +struct blk_mq_hw_ctx *hctx; + +/* +* The queue has multiple hardware queues but uses IO scheduler that +* does not respect hardware queues when dispatching? This is not a +* great setup but it can be sensible when we have a single rotational +* disk behind a raid card. Just don't bother with multiple HW queues +* and dispatch from hctx for the current CPU since running multiple +* queues just causes lock contention inside the scheduler and +* pointless cache bouncing because the hctx is not respected by the IO +* scheduler's dispatch function anyway. +*/ +if (q->nr_hw_queues >1 && e && e->type->ops.dispatch_request && + !(e->type->elevator_features &ELEVATOR_F_MQ_AWARE )) { + hctx =blk_mq_map_queue_type (q,HCTX_TYPE_DEFAULT , +raw_smp_processor_id ()); +if (!blk_mq_hctx_stopped (hctx)) +return hctx; + } +return NULL ; +} +/** * blk_mq_run_hw_queues - Run all hardware queues in a request queue. * @q: Pointer to the request queue to run.
@@ -1653,14 +1683,21 @@ EXPORT_SYMBOL(blk_mq_run_hw_queue);*/ void blk_mq_run_hw_queues (struct request_queue *q,bool async) { -struct blk_mq_hw_ctx *hctx; +struct blk_mq_hw_ctx *hctx, *sq_hctx;int i; + sq_hctx =blk_mq_get_sqsched_hctx (q);queue_for_each_hw_ctx (q, hctx, i) {if (blk_mq_hctx_stopped (hctx))continue ; - -blk_mq_run_hw_queue (hctx, async); +/* +* Dispatch from this hctx either if there's no hctx preferred +* by IO scheduler or if it has requests that bypass the +* scheduler. +*/ +if (!sq_hctx ||blk_mq_hctx_has_queued_rq (hctx) || + sq_hctx == hctx) +blk_mq_run_hw_queue (hctx, async); } }EXPORT_SYMBOL (blk_mq_run_hw_queues);
@@ -1672,14 +1709,21 @@ EXPORT_SYMBOL(blk_mq_run_hw_queues);*/ void blk_mq_delay_run_hw_queues (struct request_queue *q,unsigned long msecs) { -struct blk_mq_hw_ctx *hctx; +struct blk_mq_hw_ctx *hctx, *sq_hctx;int i; + sq_hctx =blk_mq_get_sqsched_hctx (q);queue_for_each_hw_ctx (q, hctx, i) {if (blk_mq_hctx_stopped (hctx))continue ; - -blk_mq_delay_run_hw_queue (hctx, msecs); +/* +* Dispatch from this hctx either if there's no hctx preferred +* by IO scheduler or if it has requests that bypass the +* scheduler. +*/ +if (!sq_hctx ||blk_mq_hctx_has_queued_rq (hctx) || + sq_hctx == hctx) +blk_mq_delay_run_hw_queue (hctx, msecs); } }EXPORT_SYMBOL (blk_mq_delay_run_hw_queues);
block/kyber-iosched.c +1 -0
@@ -1029,6 +1029,7 @@ static struct elevator_type kyber_sched = {#endif .elevator_attrs = kyber_sched_attrs, .elevator_name = "kyber", + .elevator_features = ELEVATOR_F_MQ_AWARE, .elevator_owner = THIS_MODULE, };
include/linux/elevator.h +2 -0
@@ -172,6 +172,8 @@ extern struct request *elv_rb_find(struct rb_root *, sector_t);/* Supports zoned block devices sequential write constraint */ #define ELEVATOR_F_ZBD_SEQ_WRITE (1U << 0) +/* Supports scheduling on multiple hardware queues */ +#define ELEVATOR_F_MQ_AWARE (1U << 1)#endif /* CONFIG_BLOCK */ #endif
Signature · 1 line
2.26.2
Quoted 44 lines · Jan Kara
Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for a queue with multiple HW queues, the performance it rather bad. The problem is that these IO schedulers use queue-wide locking and their dispatch function does not respect the hctx it is passed in and returns any request it finds appropriate. Thus locality of request access is broken and dispatch from multiple CPUs just contends on IO scheduler locks. For these IO schedulers there's little point in dispatching from multiple CPUs. Instead dispatch always only from a single CPU to limit contention.
Below is a comparison of dbench runs on XFS filesystem where the storage is a raid card with 64 HW queues and to it attached a single rotating disk. BFQ is used as IO scheduler:
clients MQ SQ MQ-Patched
Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%*Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%* Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%* Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%* Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%* Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%* Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%* Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%*Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is the same kernel with megaraid_sas.host_tagset_enable=0 so that the card advertises just a single HW queue. MQ-Patched is a kernel with this patch applied.
You can see multiple hardware queues heavily hurt performance in combination with BFQ. The patch restores the performance.
Signed-off-by Jan Kara <[email protected]>--- block/blk-mq.c | 62 ++++++++++++++++++++++++++++++++++------ block/kyber-iosched.c | 1 + include/linux/elevator.h | 2 ++ 3 files changed, 56 insertions(+), 9 deletions(-)block/blk-mq.c +0
-0
block/blk-mq.c static int blk_mq_poll_stats_bkt(const struct request *rq) · line 72
Quoted 10 lines
block/blk-mq.c +7
-0@@ -63,15 +63,20 @@ static int blk_mq_poll_stats_bkt(const struct request *rq)return bucket; } +/* Check if there are requests queued in hctx lists. */ +static bool blk_mq_hctx_has_queued_rq (struct blk_mq_hw_ctx *hctx) +{ +return !list_empty_careful (&hctx->dispatch) || +sbitmap_any_bit_set (&hctx->ctx_map); +} +
blk_mq_hctx_mark_pending() is only called in case of none scheduler, so looks not necessary to check hctx->ctx_map in blk_mq_hctx_has_queued_rq() which is supposed to be used when real io scheduler is attached to MQ queue.
Thanks,
Ming
Quoted 12 lines · Ming Lei
block/blk-mq.c static int blk_mq_poll_stats_bkt(const struct request *rq) · line 72
Quoted 7 lines · Jan Kara
block/blk-mq.c +7
-0+/* Check if there are requests queued in hctx lists. */ +static bool blk_mq_hctx_has_queued_rq (struct blk_mq_hw_ctx *hctx) +{ +return !list_empty_careful (&hctx->dispatch) || +sbitmap_any_bit_set (&hctx->ctx_map); +} +
blk_mq_hctx_mark_pending()is only called in case of none scheduler, so looks not necessary to checkhctx->ctx_mapinblk_mq_hctx_has_queued_rq()which is supposed to be used when real io scheduler is attached to MQ queue.
Yes, I know. I just wanted to make the code less fragile... In particular I was somewhat uneasy that we'd rely on the implicit behavior that blk_mq_get_sqsched_hctx() can return non-NULL only if sbitmap_any_bit_set() is not needed. But maybe we could structure the code like:
sq_hctx = NULL;
if (blk_mq_has_sqsched(q))
sq_hctx = blk_mq_get_sq_hctx(q);
queue_for_each_hw_ctx(q, hctx, i) {
...
if (!sq_hctx || sq_hctx == hctx ||
!list_empty_careful(&hctx->dispatch))
... run ...
}Because then it is kind of obvious that sq_hctx is set only if there's IO scheduler for the queue and thus ctx_map is unused. What do you think?
Honza
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Quoted 14 lines · Jan Kara
Quoted 12 lines · Ming Lei
block/blk-mq.c static int blk_mq_poll_stats_bkt(const struct request *rq) · line 72
Quoted 7 lines · Jan Kara
block/blk-mq.c +7
-0+/* Check if there are requests queued in hctx lists. */ +static bool blk_mq_hctx_has_queued_rq (struct blk_mq_hw_ctx *hctx) +{ +return !list_empty_careful (&hctx->dispatch) || +sbitmap_any_bit_set (&hctx->ctx_map); +} +
blk_mq_hctx_mark_pending()is only called in case of none scheduler, so looks not necessary to checkhctx->ctx_mapinblk_mq_hctx_has_queued_rq()which is supposed to be used when real io scheduler is attached to MQ queue.
Quoted 4 lines
Yes, I know. I just wanted to make the code less fragile... In particular I was somewhat uneasy that we'd rely on the implicit behavior that
blk_mq_get_sqsched_hctx()can return non-NULL only ifsbitmap_any_bit_set()is not needed. But maybe we could structure the code like:
BTW, I mentioned the point because sbitmap_any_bit_set(hctx->ctx_map) may take some CPU cycle in case that nr_cpu_ids is big.
Quoted 11 lines
sq_hctx = NULL;
if (blk_mq_has_sqsched(q))
sq_hctx = blk_mq_get_sq_hctx(q);
queue_for_each_hw_ctx(q, hctx, i) {
...
if (!sq_hctx || sq_hctx == hctx ||
!list_empty_careful(&hctx->dispatch))
... run ...
}Quoted 2 lines
Because then it is kind of obvious that
sq_hctxis set only if there's IO scheduler for the queue and thusctx_mapis unused. What do you think?
IMO, the above is more readable and efficient.
Thanks,
Ming