1 // SPDX-License-Identifier: GPL-2.0
3 * blk-mq scheduling framework
5 * Copyright (C) 2016 Jens Axboe
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/blk-mq.h>
10 #include <linux/list_sort.h>
12 #include <trace/events/block.h>
16 #include "blk-mq-debugfs.h"
17 #include "blk-mq-sched.h"
18 #include "blk-mq-tag.h"
21 void blk_mq_sched_assign_ioc(struct request *rq)
23 struct request_queue *q = rq->q;
24 struct io_context *ioc;
28 * May not have an IO context if it's a passthrough request
30 ioc = current->io_context;
34 spin_lock_irq(&q->queue_lock);
35 icq = ioc_lookup_icq(ioc, q);
36 spin_unlock_irq(&q->queue_lock);
39 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
43 get_io_context(icq->ioc);
48 * Mark a hardware queue as needing a restart. For shared queues, maintain
49 * a count of how many hardware queues are marked for restart.
51 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
53 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
56 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
58 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
60 void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
62 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
64 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
67 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
68 * in blk_mq_run_hw_queue(). Its pair is the barrier in
69 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
70 * meantime new request added to hctx->dispatch is missed to check in
71 * blk_mq_run_hw_queue().
75 blk_mq_run_hw_queue(hctx, true);
78 static int sched_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
80 struct request *rqa = container_of(a, struct request, queuelist);
81 struct request *rqb = container_of(b, struct request, queuelist);
83 return rqa->mq_hctx > rqb->mq_hctx;
86 static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
88 struct blk_mq_hw_ctx *hctx =
89 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
92 unsigned int count = 0;
94 list_for_each_entry(rq, rq_list, queuelist) {
95 if (rq->mq_hctx != hctx) {
96 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
101 list_splice_tail_init(rq_list, &hctx_list);
104 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
107 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
110 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
111 * its queue by itself in its completion handler, so we don't need to
112 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
114 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
115 * be run again. This is necessary to avoid starving flushes.
117 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
119 struct request_queue *q = hctx->queue;
120 struct elevator_queue *e = q->elevator;
121 bool multi_hctxs = false, run_queue = false;
122 bool dispatched = false, busy = false;
123 unsigned int max_dispatch;
127 if (hctx->dispatch_busy)
130 max_dispatch = hctx->queue->nr_requests;
135 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
138 if (!list_empty_careful(&hctx->dispatch)) {
143 if (!blk_mq_get_dispatch_budget(q))
146 rq = e->type->ops.dispatch_request(hctx);
148 blk_mq_put_dispatch_budget(q);
150 * We're releasing without dispatching. Holding the
151 * budget could have blocked any "hctx"s with the
152 * same queue and if we didn't dispatch then there's
153 * no guarantee anyone will kick the queue. Kick it
161 * Now this rq owns the budget which has to be released
162 * if this rq won't be queued to driver via .queue_rq()
163 * in blk_mq_dispatch_rq_list().
165 list_add_tail(&rq->queuelist, &rq_list);
166 if (rq->mq_hctx != hctx)
168 } while (++count < max_dispatch);
172 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
173 } else if (multi_hctxs) {
175 * Requests from different hctx may be dequeued from some
176 * schedulers, such as bfq and deadline.
178 * Sort the requests in the list according to their hctx,
179 * dispatch batching requests from same hctx at a time.
181 list_sort(NULL, &rq_list, sched_rq_cmp);
183 dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
184 } while (!list_empty(&rq_list));
186 dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count);
194 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
199 ret = __blk_mq_do_dispatch_sched(hctx);
205 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
206 struct blk_mq_ctx *ctx)
208 unsigned short idx = ctx->index_hw[hctx->type];
210 if (++idx == hctx->nr_ctx)
213 return hctx->ctxs[idx];
217 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
218 * its queue by itself in its completion handler, so we don't need to
219 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
221 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
222 * be run again. This is necessary to avoid starving flushes.
224 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
226 struct request_queue *q = hctx->queue;
228 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
233 if (!list_empty_careful(&hctx->dispatch)) {
238 if (!sbitmap_any_bit_set(&hctx->ctx_map))
241 if (!blk_mq_get_dispatch_budget(q))
244 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
246 blk_mq_put_dispatch_budget(q);
248 * We're releasing without dispatching. Holding the
249 * budget could have blocked any "hctx"s with the
250 * same queue and if we didn't dispatch then there's
251 * no guarantee anyone will kick the queue. Kick it
254 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
259 * Now this rq owns the budget which has to be released
260 * if this rq won't be queued to driver via .queue_rq()
261 * in blk_mq_dispatch_rq_list().
263 list_add(&rq->queuelist, &rq_list);
265 /* round robin for fair dispatch */
266 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
268 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
270 WRITE_ONCE(hctx->dispatch_from, ctx);
274 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
276 struct request_queue *q = hctx->queue;
277 struct elevator_queue *e = q->elevator;
278 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
283 * If we have previous entries on our dispatch list, grab them first for
284 * more fair dispatch.
286 if (!list_empty_careful(&hctx->dispatch)) {
287 spin_lock(&hctx->lock);
288 if (!list_empty(&hctx->dispatch))
289 list_splice_init(&hctx->dispatch, &rq_list);
290 spin_unlock(&hctx->lock);
294 * Only ask the scheduler for requests, if we didn't have residual
295 * requests from the dispatch list. This is to avoid the case where
296 * we only ever dispatch a fraction of the requests available because
297 * of low device queue depth. Once we pull requests out of the IO
298 * scheduler, we can no longer merge or sort them. So it's best to
299 * leave them there for as long as we can. Mark the hw queue as
300 * needing a restart in that case.
302 * We want to dispatch from the scheduler if there was nothing
303 * on the dispatch list or we were able to dispatch from the
306 if (!list_empty(&rq_list)) {
307 blk_mq_sched_mark_restart_hctx(hctx);
308 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
309 if (has_sched_dispatch)
310 ret = blk_mq_do_dispatch_sched(hctx);
312 ret = blk_mq_do_dispatch_ctx(hctx);
314 } else if (has_sched_dispatch) {
315 ret = blk_mq_do_dispatch_sched(hctx);
316 } else if (hctx->dispatch_busy) {
317 /* dequeue request one by one from sw queue if queue is busy */
318 ret = blk_mq_do_dispatch_ctx(hctx);
320 blk_mq_flush_busy_ctxs(hctx, &rq_list);
321 blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
327 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
329 struct request_queue *q = hctx->queue;
331 /* RCU or SRCU read lock is needed before checking quiesced flag */
332 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
338 * A return of -EAGAIN is an indication that hctx->dispatch is not
339 * empty and we must run again in order to avoid starving flushes.
341 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
342 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
343 blk_mq_run_hw_queue(hctx, true);
347 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
348 unsigned int nr_segs)
350 struct elevator_queue *e = q->elevator;
351 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
352 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
356 if (e && e->type->ops.bio_merge)
357 return e->type->ops.bio_merge(hctx, bio, nr_segs);
360 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) ||
361 list_empty_careful(&ctx->rq_lists[type]))
364 /* default per sw-queue merge */
365 spin_lock(&ctx->lock);
367 * Reverse check our software queue for entries that we could
368 * potentially merge with. Currently includes a hand-wavy stop
369 * count of 8, to not spend too much time checking for merges.
371 if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
376 spin_unlock(&ctx->lock);
381 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
383 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
385 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
387 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
391 * dispatch flush and passthrough rq directly
393 * passthrough request has to be added to hctx->dispatch directly.
394 * For some reason, device may be in one situation which can't
395 * handle FS request, so STS_RESOURCE is always returned and the
396 * FS request will be added to hctx->dispatch. However passthrough
397 * request may be required at that time for fixing the problem. If
398 * passthrough request is added to scheduler queue, there isn't any
399 * chance to dispatch it given we prioritize requests in hctx->dispatch.
401 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
407 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
408 bool run_queue, bool async)
410 struct request_queue *q = rq->q;
411 struct elevator_queue *e = q->elevator;
412 struct blk_mq_ctx *ctx = rq->mq_ctx;
413 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
415 WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG));
417 if (blk_mq_sched_bypass_insert(hctx, rq)) {
419 * Firstly normal IO request is inserted to scheduler queue or
420 * sw queue, meantime we add flush request to dispatch queue(
421 * hctx->dispatch) directly and there is at most one in-flight
422 * flush request for each hw queue, so it doesn't matter to add
423 * flush request to tail or front of the dispatch queue.
425 * Secondly in case of NCQ, flush request belongs to non-NCQ
426 * command, and queueing it will fail when there is any
427 * in-flight normal IO request(NCQ command). When adding flush
428 * rq to the front of hctx->dispatch, it is easier to introduce
429 * extra time to flush rq's latency because of S_SCHED_RESTART
430 * compared with adding to the tail of dispatch queue, then
431 * chance of flush merge is increased, and less flush requests
432 * will be issued to controller. It is observed that ~10% time
433 * is saved in blktests block/004 on disk attached to AHCI/NCQ
434 * drive when adding flush rq to the front of hctx->dispatch.
436 * Simply queue flush rq to the front of hctx->dispatch so that
437 * intensive flush workloads can benefit in case of NCQ HW.
439 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
440 blk_mq_request_bypass_insert(rq, at_head, false);
444 if (e && e->type->ops.insert_requests) {
447 list_add(&rq->queuelist, &list);
448 e->type->ops.insert_requests(hctx, &list, at_head);
450 spin_lock(&ctx->lock);
451 __blk_mq_insert_request(hctx, rq, at_head);
452 spin_unlock(&ctx->lock);
457 blk_mq_run_hw_queue(hctx, async);
460 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
461 struct blk_mq_ctx *ctx,
462 struct list_head *list, bool run_queue_async)
464 struct elevator_queue *e;
465 struct request_queue *q = hctx->queue;
468 * blk_mq_sched_insert_requests() is called from flush plug
469 * context only, and hold one usage counter to prevent queue
470 * from being released.
472 percpu_ref_get(&q->q_usage_counter);
474 e = hctx->queue->elevator;
475 if (e && e->type->ops.insert_requests)
476 e->type->ops.insert_requests(hctx, list, false);
479 * try to issue requests directly if the hw queue isn't
480 * busy in case of 'none' scheduler, and this way may save
481 * us one extra enqueue & dequeue to sw queue.
483 if (!hctx->dispatch_busy && !e && !run_queue_async) {
484 blk_mq_try_issue_list_directly(hctx, list);
485 if (list_empty(list))
488 blk_mq_insert_requests(hctx, ctx, list);
491 blk_mq_run_hw_queue(hctx, run_queue_async);
493 percpu_ref_put(&q->q_usage_counter);
496 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
497 struct blk_mq_hw_ctx *hctx,
498 unsigned int hctx_idx)
500 unsigned int flags = set->flags & ~BLK_MQ_F_TAG_HCTX_SHARED;
502 if (hctx->sched_tags) {
503 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
504 blk_mq_free_rq_map(hctx->sched_tags, flags);
505 hctx->sched_tags = NULL;
509 static int blk_mq_sched_alloc_tags(struct request_queue *q,
510 struct blk_mq_hw_ctx *hctx,
511 unsigned int hctx_idx)
513 struct blk_mq_tag_set *set = q->tag_set;
514 /* Clear HCTX_SHARED so tags are init'ed */
515 unsigned int flags = set->flags & ~BLK_MQ_F_TAG_HCTX_SHARED;
518 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
519 set->reserved_tags, flags);
520 if (!hctx->sched_tags)
523 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
525 blk_mq_sched_free_tags(set, hctx, hctx_idx);
530 /* called in queue's release handler, tagset has gone away */
531 static void blk_mq_sched_tags_teardown(struct request_queue *q)
533 struct blk_mq_hw_ctx *hctx;
536 queue_for_each_hw_ctx(q, hctx, i) {
537 /* Clear HCTX_SHARED so tags are freed */
538 unsigned int flags = hctx->flags & ~BLK_MQ_F_TAG_HCTX_SHARED;
540 if (hctx->sched_tags) {
541 blk_mq_free_rq_map(hctx->sched_tags, flags);
542 hctx->sched_tags = NULL;
547 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
549 struct blk_mq_hw_ctx *hctx;
550 struct elevator_queue *eq;
556 q->nr_requests = q->tag_set->queue_depth;
561 * Default to double of smaller one between hw queue_depth and 128,
562 * since we don't split into sync/async like the old code did.
563 * Additionally, this is a per-hw queue depth.
565 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
568 queue_for_each_hw_ctx(q, hctx, i) {
569 ret = blk_mq_sched_alloc_tags(q, hctx, i);
574 ret = e->ops.init_sched(q, e);
578 blk_mq_debugfs_register_sched(q);
580 queue_for_each_hw_ctx(q, hctx, i) {
581 if (e->ops.init_hctx) {
582 ret = e->ops.init_hctx(hctx, i);
585 blk_mq_sched_free_requests(q);
586 blk_mq_exit_sched(q, eq);
587 kobject_put(&eq->kobj);
591 blk_mq_debugfs_register_sched_hctx(q, hctx);
597 blk_mq_sched_free_requests(q);
598 blk_mq_sched_tags_teardown(q);
604 * called in either blk_queue_cleanup or elevator_switch, tagset
605 * is required for freeing requests
607 void blk_mq_sched_free_requests(struct request_queue *q)
609 struct blk_mq_hw_ctx *hctx;
612 queue_for_each_hw_ctx(q, hctx, i) {
613 if (hctx->sched_tags)
614 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
618 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
620 struct blk_mq_hw_ctx *hctx;
623 queue_for_each_hw_ctx(q, hctx, i) {
624 blk_mq_debugfs_unregister_sched_hctx(hctx);
625 if (e->type->ops.exit_hctx && hctx->sched_data) {
626 e->type->ops.exit_hctx(hctx, i);
627 hctx->sched_data = NULL;
630 blk_mq_debugfs_unregister_sched(q);
631 if (e->type->ops.exit_sched)
632 e->type->ops.exit_sched(e);
633 blk_mq_sched_tags_teardown(q);