blk-mq: support batching dispatch from scheduler
Subsystems (2)
v5 cover blk-mq: support batching dispatch from scheduler
Hi Jens,
More and more drivers want to get batching requests queued from block layer, such as mmc[1], and tcp based storage drivers[2]. Also current in-tree users have virtio-scsi, virtio-blk and nvme.
For none, we already support batching dispatch.
But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue.
Tries to start the support by dequeuing more requests from scheduler if budget is enough and device isn't busy.
Simple fio test over virtio-scsi shows IO can get improved by 5~10%.
Baolin has tested previous versions and found performance on MMC can be improved.
Patch 1 ~ 4 are improvement and cleanup, which can't applied without supporting batching dispatch.
Patch 5 ~ 6 starts to support batching dispatch from scheduler.
[1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r
[2] https://lore.kernel.org/linux-block/[email protected]/
V5:
- code style changes suggested by Damien
V4:
- fix releasing budgets and avoids IO hang(5/6)
- dispatch more batches if the device can accept more(6/6)
- verified by running more tests
V3:
- add reviewed-by tag
- fix one typo
- fix one budget leak issue in case that .queue_rq returned *_RESOURCE in 5/6
V2:
- remove 'got_budget' from blk_mq_dispatch_rq_list
- drop patch for getting driver tag & handling partial dispatch
Ming Lei (6):
blk-mq: pass request queue into get/put budget callback
blk-mq: pass hctx to blk_mq_dispatch_rq_list
blk-mq: move getting driver tag and budget into one helper
blk-mq: remove dead check from blk_mq_dispatch_rq_list
blk-mq: pass obtained budget count to blk_mq_dispatch_rq_list
blk-mq: support batching dispatch in case of io scheduler
block/blk-mq-sched.c | 116 +++++++++++++++++++++++++++++++++++----- block/blk-mq.c | 110 +++++++++++++++++++++++-------------- block/blk-mq.h | 15 +++--- drivers/scsi/scsi_lib.c | 8 ++- include/linux/blk-mq.h | 4 +- 5 files changed, 182 insertions(+), 71 deletions(-)
Signature · 1 line
2.25.2
PATCH v5 1/6 blk-mq: pass request queue into get/put budget callback
blk-mq budget is abstract from scsi's device queue depth, and it is always per-request-queue instead of hctx.
It can be quite absurd to get a budget from one hctx, then dequeue a request from scheduler queue, and this request may not belong to this hctx, at least for bfq and deadline.
So fix the mess and always pass request queue to get/put budget callback.
--- block/blk-mq-sched.c | 8 ++++---- block/blk-mq.c | 8 ++++---- block/blk-mq.h | 12 ++++-------- drivers/scsi/scsi_lib.c | 8 +++----- include/linux/blk-mq.h | 4 ++-- 5 files changed, 17 insertions(+), 23 deletions(-)
block/blk-mq-sched.c +4 -4
@@ -108,12 +108,12 @@ static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)break ; } -if (!blk_mq_get_dispatch_budget (hctx)) +if (!blk_mq_get_dispatch_budget (q))break ; rq = e->type->ops.dispatch_request (hctx);if (!rq) { -blk_mq_put_dispatch_budget (hctx); +blk_mq_put_dispatch_budget (q);/* * We're releasing without dispatching. Holding the * budget could have blocked any "hctx"s with the
@@ -173,12 +173,12 @@ static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)if (!sbitmap_any_bit_set (&hctx->ctx_map))break ; -if (!blk_mq_get_dispatch_budget (hctx)) +if (!blk_mq_get_dispatch_budget (q))break ; rq =blk_mq_dequeue_from_ctx (hctx, ctx);if (!rq) { -blk_mq_put_dispatch_budget (hctx); +blk_mq_put_dispatch_budget (q);/* * We're releasing without dispatching. Holding the * budget could have blocked any "hctx"s with the
block/blk-mq.c +4 -4
@@ -1262,7 +1262,7 @@ bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, rq = list_first_entry(list, struct request, queuelist); hctx = rq->mq_hctx; -if (!got_budget && !blk_mq_get_dispatch_budget (hctx)) { +if (!got_budget && !blk_mq_get_dispatch_budget (q)) {blk_mq_put_driver_tag (rq); no_budget_avail = true;break ;
@@ -1277,7 +1277,7 @@ bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,* we'll re-run it below. */ if (!blk_mq_mark_tag_wait (hctx, rq)) { -blk_mq_put_dispatch_budget (hctx); +blk_mq_put_dispatch_budget (q);/* * For non-shared tags, the RESTART check * will suffice.
@@ -1925,11 +1925,11 @@ static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,if (q->elevator && !bypass_insert) goto insert; -if (!blk_mq_get_dispatch_budget (hctx)) +if (!blk_mq_get_dispatch_budget (q)) goto insert;if (!blk_mq_get_driver_tag (rq)) { -blk_mq_put_dispatch_budget (hctx); +blk_mq_put_dispatch_budget (q); goto insert; }
block/blk-mq.h +4 -8
@@ -180,20 +180,16 @@ unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part);void blk_mq_in_flight_rw (struct request_queue *q,struct hd_struct *part,unsigned int inflight[2 ]); -static inline void blk_mq_put_dispatch_budget (struct blk_mq_hw_ctx *hctx) +static inline void blk_mq_put_dispatch_budget (struct request_queue *q) { -struct request_queue *q = hctx->queue; -if (q->mq_ops->put_budget) - q->mq_ops->put_budget (hctx); + q->mq_ops->put_budget (q); } -static inline bool blk_mq_get_dispatch_budget (struct blk_mq_hw_ctx *hctx) +static inline bool blk_mq_get_dispatch_budget (struct request_queue *q) { -struct request_queue *q = hctx->queue; -if (q->mq_ops->get_budget) -return q->mq_ops->get_budget (hctx); +return q->mq_ops->get_budget (q);return true; }
drivers/scsi/scsi_lib.c +3 -5
@@ -1624,17 +1624,15 @@ static void scsi_mq_done(struct scsi_cmnd *cmd)clear_bit (SCMD_STATE_COMPLETE , &cmd->state); } -static void scsi_mq_put_budget (struct blk_mq_hw_ctx *hctx) +static void scsi_mq_put_budget (struct request_queue *q) { -struct request_queue *q = hctx->queue;struct scsi_device *sdev = q->queuedata;atomic_dec (&sdev->device_busy); } -static bool scsi_mq_get_budget (struct blk_mq_hw_ctx *hctx) +static bool scsi_mq_get_budget (struct request_queue *q) { -struct request_queue *q = hctx->queue;struct scsi_device *sdev = q->queuedata;return scsi_dev_queue_ready (q, sdev);
@@ -1701,7 +1699,7 @@ static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,if (scsi_target (sdev)->can_queue >0 )atomic_dec (&scsi_target (sdev)->target_busy); out_put_budget: -scsi_mq_put_budget (hctx); +scsi_mq_put_budget (q);switch (ret) {case BLK_STS_OK :break ;
include/linux/blk-mq.h +2 -2
@@ -270,8 +270,8 @@ struct blk_mq_queue_data {typedef blk_status_t (queue_rq_fn )(struct blk_mq_hw_ctx *,const struct blk_mq_queue_data *);typedef void (commit_rqs_fn )(struct blk_mq_hw_ctx *); -typedef bool (get_budget_fn )(struct blk_mq_hw_ctx *); -typedef void (put_budget_fn )(struct blk_mq_hw_ctx *); +typedef bool (get_budget_fn )(struct request_queue *); +typedef void (put_budget_fn )(struct request_queue *);typedef enum blk_eh_timer_return (timeout_fn )(struct request *,bool );typedef int (init_hctx_fn )(struct blk_mq_hw_ctx *,void *,unsigned int );typedef void (exit_hctx_fn )(struct blk_mq_hw_ctx *,unsigned int );
Signature · 1 line
2.25.2
PATCH v5 2/6 blk-mq: pass hctx to blk_mq_dispatch_rq_list
In reply to Ming LeiAll requests in the 'list' of blk_mq_dispatch_rq_list belong to same hctx, so it is better to pass hctx instead of request queue, because blk-mq's dispatch target is hctx instead of request queue.
--- block/blk-mq-sched.c | 14 ++++++-------- block/blk-mq.c | 6 +++--- block/blk-mq.h | 2 +- 3 files changed, 10 insertions(+), 12 deletions(-)
block/blk-mq-sched.c +6 -8
@@ -96,10 +96,9 @@ static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)struct elevator_queue *e = q->elevator;LIST_HEAD (rq_list);int ret =0 ; +struct request *rq; do { - struct request *rq; - if (e->type->ops.has_work && !e->type->ops.has_work(hctx)) break;
@@ -131,7 +130,7 @@ static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)* in blk_mq_dispatch_rq_list(). */ list_add (&rq->queuelist, &rq_list); - }while (blk_mq_dispatch_rq_list (q, &rq_list, true)); + }while (blk_mq_dispatch_rq_list (rq->mq_hctx, &rq_list, true));return ret; }
@@ -161,10 +160,9 @@ static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)LIST_HEAD (rq_list);struct blk_mq_ctx *ctx =READ_ONCE (hctx->dispatch_from);int ret =0 ; +struct request *rq; do { - struct request *rq; - if (!list_empty_careful(&hctx->dispatch)) { ret = -EAGAIN; break;
@@ -200,7 +198,7 @@ static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)/* round robin for fair dispatch */ ctx =blk_mq_next_ctx (hctx, rq->mq_ctx); - }while (blk_mq_dispatch_rq_list (q, &rq_list, true)); + }while (blk_mq_dispatch_rq_list (rq->mq_hctx, &rq_list, true));WRITE_ONCE (hctx->dispatch_from, ctx);return ret;
@@ -240,7 +238,7 @@ static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)*/ if (!list_empty (&rq_list)) {blk_mq_sched_mark_restart_hctx (hctx); -if (blk_mq_dispatch_rq_list (q, &rq_list, false)) { +if (blk_mq_dispatch_rq_list (hctx, &rq_list, false)) {if (has_sched_dispatch) ret =blk_mq_do_dispatch_sched (hctx);else
@@ -253,7 +251,7 @@ static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) ret =blk_mq_do_dispatch_ctx (hctx); }else {blk_mq_flush_busy_ctxs (hctx, &rq_list); -blk_mq_dispatch_rq_list (q, &rq_list, false); +blk_mq_dispatch_rq_list (hctx, &rq_list, false); }return ret;
block/blk-mq.c +3 -3
@@ -1236,10 +1236,10 @@ static void blk_mq_handle_zone_resource(struct request *rq,/* * Returns true if we did some work AND can potentially do more. */ -bool blk_mq_dispatch_rq_list (struct request_queue *q,struct list_head *list, +bool blk_mq_dispatch_rq_list (struct blk_mq_hw_ctx *hctx,struct list_head *list,bool got_budget) { -struct blk_mq_hw_ctx *hctx; +struct request_queue *q = hctx->queue;struct request *rq, *nxt;bool no_tag = false;int errors, queued;
@@ -1261,7 +1261,7 @@ bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, rq = list_first_entry(list, struct request, queuelist); - hctx = rq->mq_hctx; +WARN_ON_ONCE (hctx != rq->mq_hctx);if (!got_budget && !blk_mq_get_dispatch_budget (q)) {blk_mq_put_driver_tag (rq); no_budget_avail = true;
block/blk-mq.h +1 -1
@@ -40,7 +40,7 @@ struct blk_mq_ctx {void blk_mq_exit_queue (struct request_queue *q);int blk_mq_update_nr_requests (struct request_queue *q,unsigned int nr);void blk_mq_wake_waiters (struct request_queue *q); -bool blk_mq_dispatch_rq_list (struct request_queue *,struct list_head *,bool ); +bool blk_mq_dispatch_rq_list (struct blk_mq_hw_ctx *hctx,struct list_head *,bool );void blk_mq_add_to_requeue_list (struct request *rq,bool at_head,bool kick_requeue_list);void blk_mq_flush_busy_ctxs (struct blk_mq_hw_ctx *hctx,struct list_head *list);
Signature · 1 line
2.25.2
PATCH v5 3/6 blk-mq: move getting driver tag and budget into one helper
In reply to Ming LeiMove code for getting driver tag and budget into one helper, so blk_mq_dispatch_rq_list gets a bit simplified, and easier to read.
Meantime move updating of 'no_tag' and 'no_budget_available' into the branch for handling partial dispatch because that is exactly consumer of the two local variables.
Also rename the parameter of 'got_budget' as 'ask_budget'.
No functional change.
--- block/blk-mq.c | 66 ++++++++++++++++++++++++++++++-------------------- 1 file changed, 40 insertions(+), 26 deletions(-)
block/blk-mq.c +40 -26
@@ -1233,18 +1233,50 @@ static void blk_mq_handle_zone_resource(struct request *rq,__blk_mq_requeue_request (rq); } +enum prep_dispatch { +PREP_DISPATCH_OK , +PREP_DISPATCH_NO_TAG , +PREP_DISPATCH_NO_BUDGET , +}; + +static enum prep_dispatch blk_mq_prep_dispatch_rq (struct request *rq, +bool need_budget) +{ +struct blk_mq_hw_ctx *hctx = rq->mq_hctx; + +if (need_budget && !blk_mq_get_dispatch_budget (rq->q)) { +blk_mq_put_driver_tag (rq); +return PREP_DISPATCH_NO_BUDGET ; + } + +if (!blk_mq_get_driver_tag (rq)) { +/* +* The initial allocation attempt failed, so we need to +* rerun the hardware queue when a tag is freed. The +* waitqueue takes care of that. If the queue is run +* before we add this entry back on the dispatch list, +* we'll re-run it below. +*/ +if (!blk_mq_mark_tag_wait (hctx, rq)) { +blk_mq_put_dispatch_budget (rq->q); +return PREP_DISPATCH_NO_TAG ; + } + } + +return PREP_DISPATCH_OK ; +} +/* * Returns true if we did some work AND can potentially do more. */ bool blk_mq_dispatch_rq_list (struct blk_mq_hw_ctx *hctx,struct list_head *list,bool got_budget) { +enum prep_dispatch prep;struct request_queue *q = hctx->queue;struct request *rq, *nxt; -bool no_tag = false;int errors, queued;blk_status_t ret =BLK_STS_OK ; -bool no_budget_avail = false;LIST_HEAD (zone_list);if (list_empty (list))
@@ -1262,31 +1294,9 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, rq = list_first_entry(list, struct request, queuelist);WARN_ON_ONCE (hctx != rq->mq_hctx); -if (!got_budget && !blk_mq_get_dispatch_budget (q)) { -blk_mq_put_driver_tag (rq); - no_budget_avail = true; + prep =blk_mq_prep_dispatch_rq (rq, !got_budget); +if (prep !=PREP_DISPATCH_OK )break ; - } - -if (!blk_mq_get_driver_tag (rq)) { -/* -* The initial allocation attempt failed, so we need to -* rerun the hardware queue when a tag is freed. The -* waitqueue takes care of that. If the queue is run -* before we add this entry back on the dispatch list, -* we'll re-run it below. -*/ -if (!blk_mq_mark_tag_wait (hctx, rq)) { -blk_mq_put_dispatch_budget (q); -/* -* For non-shared tags, the RESTART check -* will suffice. -*/ -if (hctx->flags &BLK_MQ_F_TAG_SHARED ) - no_tag = true; -break ; - } - }list_del_init (&rq->queuelist);
@@ -1339,6 +1349,10 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,*/ if (!list_empty (list)) {bool needs_restart; +/* For non-shared tags, the RESTART check will suffice */ +bool no_tag = prep ==PREP_DISPATCH_NO_TAG && + (hctx->flags &BLK_MQ_F_TAG_SHARED ); +bool no_budget_avail = prep ==PREP_DISPATCH_NO_BUDGET ;/* * If we didn't flush the entire list, we could have told
Signature · 1 line
2.25.2
PATCH v5 4/6 blk-mq: remove dead check from blk_mq_dispatch_rq_list
In reply to Ming LeiWhen BLK_STS_RESOURCE or BLK_STS_DEV_RESOURCE is returned from .queue_rq, the 'list' variable always holds this rq which isn't queued to LLD successfully.
So blk_mq_dispatch_rq_list() always returns false from the branch of '!list_empty(list)'.
No functional change.
--- block/blk-mq.c | 7 ------- 1 file changed, 7 deletions(-)
block/blk-mq.c +0 -7
@@ -1404,13 +1404,6 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, }else blk_mq_update_dispatch_busy (hctx, false); -/* -* If the host/device is unable to accept more work, inform the -* caller of that. -*/ -if (ret ==BLK_STS_RESOURCE || ret ==BLK_STS_DEV_RESOURCE ) -return false; -return (queued + errors) !=0 ; }
Signature · 1 line
2.25.2
PATCH v5 5/6 blk-mq: pass obtained budget count to blk_mq_dispatch_rq_list
In reply to Ming LeiPass obtained budget count to blk_mq_dispatch_rq_list(), and prepare for supporting fully batching submission.
With the obtained budget count, it is easier to put extra budgets in case of .queue_rq failure.
Meantime remove the old 'got_budget' parameter.
--- block/blk-mq-sched.c | 8 ++++---- block/blk-mq.c | 31 +++++++++++++++++++++++++++---- block/blk-mq.h | 3 ++- 3 files changed, 33 insertions(+), 9 deletions(-)
block/blk-mq-sched.c +4 -4
@@ -130,7 +130,7 @@ static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)* in blk_mq_dispatch_rq_list(). */ list_add (&rq->queuelist, &rq_list); - }while (blk_mq_dispatch_rq_list (rq->mq_hctx, &rq_list, true)); + }while (blk_mq_dispatch_rq_list (rq->mq_hctx, &rq_list,1 ));return ret; }
@@ -198,7 +198,7 @@ static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)/* round robin for fair dispatch */ ctx =blk_mq_next_ctx (hctx, rq->mq_ctx); - }while (blk_mq_dispatch_rq_list (rq->mq_hctx, &rq_list, true)); + }while (blk_mq_dispatch_rq_list (rq->mq_hctx, &rq_list,1 ));WRITE_ONCE (hctx->dispatch_from, ctx);return ret;
@@ -238,7 +238,7 @@ static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)*/ if (!list_empty (&rq_list)) {blk_mq_sched_mark_restart_hctx (hctx); -if (blk_mq_dispatch_rq_list (hctx, &rq_list, false)) { +if (blk_mq_dispatch_rq_list (hctx, &rq_list,0 )) {if (has_sched_dispatch) ret =blk_mq_do_dispatch_sched (hctx);else
@@ -251,7 +251,7 @@ static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) ret =blk_mq_do_dispatch_ctx (hctx); }else {blk_mq_flush_busy_ctxs (hctx, &rq_list); -blk_mq_dispatch_rq_list (hctx, &rq_list, false); +blk_mq_dispatch_rq_list (hctx, &rq_list,0 ); }return ret;
block/blk-mq.c +27 -4
@@ -1258,7 +1258,12 @@ static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,* we'll re-run it below. */ if (!blk_mq_mark_tag_wait (hctx, rq)) { -blk_mq_put_dispatch_budget (rq->q); +/* +* All budgets not got from this function will be put +* together during handling partial dispatch +*/ +if (need_budget) +blk_mq_put_dispatch_budget (rq->q);return PREP_DISPATCH_NO_TAG ; } }
@@ -1266,11 +1271,21 @@ static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,return PREP_DISPATCH_OK ; } +/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ +static void blk_mq_release_budgets (struct request_queue *q, +unsigned int nr_budgets) +{ +int i; + +for (i =0 ; i < nr_budgets; i++) +blk_mq_put_dispatch_budget (q); +} +/* * Returns true if we did some work AND can potentially do more. */ bool blk_mq_dispatch_rq_list (struct blk_mq_hw_ctx *hctx,struct list_head *list, -bool got_budget) +unsigned int nr_budgets) {enum prep_dispatch prep;struct request_queue *q = hctx->queue;
@@ -1282,7 +1297,7 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,if (list_empty (list))return false; -WARN_ON (!list_is_singular (list) && got_budget); +WARN_ON (!list_is_singular (list) && nr_budgets);/* * Now process all the entries, sending them to the driver.
@@ -1294,7 +1309,7 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, rq = list_first_entry(list, struct request, queuelist);WARN_ON_ONCE (hctx != rq->mq_hctx); - prep =blk_mq_prep_dispatch_rq (rq, !got_budget); + prep =blk_mq_prep_dispatch_rq (rq, !nr_budgets);if (prep !=PREP_DISPATCH_OK )break ;
@@ -1313,6 +1328,12 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, bd.last = !blk_mq_get_driver_tag (nxt); } +/* +* once the request is queued to lld, no need to cover the +* budget any more +*/ +if (nr_budgets) + nr_budgets--; ret = q->mq_ops->queue_rq (hctx, &bd);if (ret ==BLK_STS_RESOURCE || ret ==BLK_STS_DEV_RESOURCE ) {blk_mq_handle_dev_resource (rq, list);
@@ -1354,6 +1375,8 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, (hctx->flags &BLK_MQ_F_TAG_SHARED );bool no_budget_avail = prep ==PREP_DISPATCH_NO_BUDGET ; +blk_mq_release_budgets (q, nr_budgets); +/* * If we didn't flush the entire list, we could have told * the driver there was more coming, but that turned out to
block/blk-mq.h +2 -1
@@ -40,7 +40,8 @@ struct blk_mq_ctx {void blk_mq_exit_queue (struct request_queue *q);int blk_mq_update_nr_requests (struct request_queue *q,unsigned int nr);void blk_mq_wake_waiters (struct request_queue *q); -bool blk_mq_dispatch_rq_list (struct blk_mq_hw_ctx *hctx,struct list_head *,bool ); +bool blk_mq_dispatch_rq_list (struct blk_mq_hw_ctx *hctx,struct list_head *, +unsigned int );void blk_mq_add_to_requeue_list (struct request *rq,bool at_head,bool kick_requeue_list);void blk_mq_flush_busy_ctxs (struct blk_mq_hw_ctx *hctx,struct list_head *list);
Signature · 1 line
2.25.2
PATCH v5 6/6 blk-mq: support batching dispatch in case of io scheduler
In reply to Ming LeiMore and more drivers want to get batching requests queued from block layer, such as mmc, and tcp based storage drivers. Also current in-tree users have virtio-scsi, virtio-blk and nvme.
For none, we already support batching dispatch.
But for io scheduler, every time we just take one request from scheduler and pass the single request to blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue.
Try to support batching dispatch for io scheduler by starting with the following simple approach:
1) still make sure we can get budget before dequeueing request
2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy we fackback to non-batching dispatch, otherwise dequeue as many as possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list().
Wrt. 2), we use similar policy for none, and turns out that SCSI SSD performance got improved much.
In future, maybe we can develop more intelligent algorithem for batching dispatch.
Baolin has tested this patch and found that MMC performance is improved[3].
[1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r
[2] https://lore.kernel.org/linux-block/[email protected]/
[3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/
--- block/blk-mq-sched.c | 96 ++++++++++++++++++++++++++++++++++++++++++-- block/blk-mq.c | 2 - 2 files changed, 93 insertions(+), 5 deletions(-)
block/blk-mq-sched.c +93 -3
@@ -7,6 +7,7 @@#include <linux/kernel.h> #include <linux/module.h> #include <linux/blk-mq.h> +#include <linux/list_sort.h> #include <trace/events/block.h>
@@ -80,6 +81,74 @@ void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)blk_mq_run_hw_queue (hctx, true); } +/* +* We know bfq and deadline apply single scheduler queue instead of multi +* queue. However, the two are often used on single queue devices, also +* the current @hctx should affect the real device status most of times +* because of locality principle. +* +* So use current hctx->dispatch_busy directly for figuring out batching +* dispatch count. +*/ +static unsigned int blk_mq_sched_get_batching_nr (struct blk_mq_hw_ctx *hctx) +{ +if (hctx->dispatch_busy) +return 1 ; +return hctx->queue->nr_requests; +} + +static int sched_rq_cmp (void *priv,struct list_head *a,struct list_head *b) +{ + struct request *rqa = container_of(a, struct request, queuelist); + struct request *rqb = container_of(b, struct request, queuelist); + +return rqa->mq_hctx > rqb->mq_hctx; +} + +static inline bool blk_mq_do_dispatch_rq_lists (struct blk_mq_hw_ctx *hctx, +struct list_head *lists,bool multi_hctxs,unsigned count) +{ +bool ret; + +if (!count) +return false; + +if (likely (!multi_hctxs)) +return blk_mq_dispatch_rq_list (hctx, lists, count); + +/* +* Requests from different hctx may be dequeued from some scheduler, +* such as bfq and deadline. +* +* Sort the requests in the list according to their hctx, dispatch +* batching requests from same hctx +*/ +list_sort (NULL , lists, sched_rq_cmp); + + ret = false; +while (!list_empty (lists)) { +LIST_HEAD (list); +struct request *new, *rq =list_first_entry (lists, + struct request, queuelist); +unsigned cnt =0 ; + +list_for_each_entry (new, lists, queuelist) { +if (new->mq_hctx != rq->mq_hctx) +break ; + cnt++; + } + +if (new->mq_hctx == rq->mq_hctx) +list_splice_tail_init (lists, &list); +else +list_cut_before (&list, lists, &new->queuelist); + + ret =blk_mq_dispatch_rq_list (rq->mq_hctx, &list, cnt); + } + +return ret; +} +#define BLK_MQ_BUDGET_DELAY 3/* ms units */ /*
@@ -97,7 +166,15 @@ static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)LIST_HEAD (rq_list);int ret =0 ;struct request *rq; - +int cnt; +unsigned int max_dispatch; +bool multi_hctxs, run_queue; + + again: +/* prepare one batch for dispatch */ + cnt =0 ; + max_dispatch =blk_mq_sched_get_batching_nr (hctx); + multi_hctxs = run_queue = false; do { if (e->type->ops.has_work && !e->type->ops.has_work(hctx)) break;
@@ -120,7 +197,7 @@ static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)* no guarantee anyone will kick the queue. Kick it * ourselves. */ -blk_mq_delay_run_hw_queues (q,BLK_MQ_BUDGET_DELAY ); + run_queue = true;break ; }
@@ -130,7 +207,20 @@ static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)* in blk_mq_dispatch_rq_list(). */ list_add (&rq->queuelist, &rq_list); - }while (blk_mq_dispatch_rq_list (rq->mq_hctx, &rq_list,1 )); + cnt++; + +if (rq->mq_hctx != hctx && !multi_hctxs) + multi_hctxs = true; + }while (cnt < max_dispatch); + +/* dispatch more if we may do more */ +if (blk_mq_do_dispatch_rq_lists (hctx, &rq_list, multi_hctxs, cnt) && + !ret) + goto again; + +/* in-flight request's completion can rerun queue */ +if (!cnt && run_queue) +blk_mq_delay_run_hw_queues (q,BLK_MQ_BUDGET_DELAY );return ret; }
block/blk-mq.c +0 -2
@@ -1297,8 +1297,6 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,if (list_empty (list))return false; -WARN_ON (!list_is_singular (list) && nr_budgets); -/* * Now process all the entries, sending them to the driver. */
Signature · 1 line
2.25.2
Quoted 41 lines
Hi Jens,
More and more drivers want to get batching requests queued from block layer, such as mmc[1], and tcp based storage drivers[2]. Also current in-tree users have virtio-scsi, virtio-blk and nvme.
For none, we already support batching dispatch.
But for io scheduler, every time we just take one request from scheduler and pass the single request to
blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue.Tries to start the support by dequeuing more requests from scheduler if budget is enough and device isn't busy.
Simple fio test over virtio-scsi shows IO can get improved by 5~10%.
Baolin has tested previous versions and found performance on MMC can be improved.
Patch 1 ~ 4 are improvement and cleanup, which can't applied without supporting batching dispatch.
Patch 5 ~ 6 starts to support batching dispatch from scheduler.
[1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r
[2] https://lore.kernel.org/linux-block/[email protected]/V5:
- code style changes suggested by DamienV4:
- fix releasing budgets and avoids IO hang(5/6)
- dispatch more batches if the device can accept more(6/6)
- verified by running more tests
Quoted 8 lines
V3:
- add reviewed-by tag
- fix one typo
- fix one budget leak issue in case that .queue_rq returned*_RESOURCEin 5/6V2:
- remove 'got_budget' fromblk_mq_dispatch_rq_list
- drop patch for getting driver tag & handling partial dispatch
Hey Ming,
What ever happened to this one?
Quoted 46 lines · Sagi Grimberg
Quoted 46 lines
Hi Jens,
More and more drivers want to get batching requests queued from block layer, such as mmc[1], and tcp based storage drivers[2]. Also current in-tree users have virtio-scsi, virtio-blk and nvme.
For none, we already support batching dispatch.
But for io scheduler, every time we just take one request from scheduler and pass the single request to
blk_mq_dispatch_rq_list(). This way makes batching dispatch not possible when io scheduler is applied. One reason is that we don't want to hurt sequential IO performance, becasue IO merge chance is reduced if more requests are dequeued from scheduler queue.Tries to start the support by dequeuing more requests from scheduler if budget is enough and device isn't busy.
Simple fio test over virtio-scsi shows IO can get improved by 5~10%.
Baolin has tested previous versions and found performance on MMC can be improved.
Patch 1 ~ 4 are improvement and cleanup, which can't applied without supporting batching dispatch.
Patch 5 ~ 6 starts to support batching dispatch from scheduler.
[1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r
[2] https://lore.kernel.org/linux-block/[email protected]/V5:
- code style changes suggested by DamienV4:
- fix releasing budgets and avoids IO hang(5/6)
- dispatch more batches if the device can accept more(6/6)
- verified by running more testsV3:
- add reviewed-by tag
- fix one typo
- fix one budget leak issue in case that .queue_rq returned*_RESOURCEin 5/6
Quoted 7 lines
Quoted 3 lines
V2:
- remove 'got_budget' fromblk_mq_dispatch_rq_list
- drop patch for getting driver tag & handling partial dispatchHey Ming,
What ever happened to this one?
Hi Sagi,
Looks the 1st patch can't be applied cleanly against for-5.9/block, I will send V6 later.
thanks,
Ming