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/list_sort.h>
11 #include <trace/events/block.h>
15 #include "blk-mq-debugfs.h"
16 #include "blk-mq-sched.h"
20 * Mark a hardware queue as needing a restart.
22 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
24 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
27 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
29 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
31 void __blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
33 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
36 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
37 * in blk_mq_run_hw_queue(). Its pair is the barrier in
38 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
39 * meantime new request added to hctx->dispatch is missed to check in
40 * blk_mq_run_hw_queue().
44 blk_mq_run_hw_queue(hctx, true);
47 static int sched_rq_cmp(void *priv, const struct list_head *a,
48 const struct list_head *b)
50 struct request *rqa = container_of(a, struct request, queuelist);
51 struct request *rqb = container_of(b, struct request, queuelist);
53 return rqa->mq_hctx > rqb->mq_hctx;
56 static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
58 struct blk_mq_hw_ctx *hctx =
59 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
63 list_for_each_entry(rq, rq_list, queuelist) {
64 if (rq->mq_hctx != hctx) {
65 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
69 list_splice_tail_init(rq_list, &hctx_list);
72 return blk_mq_dispatch_rq_list(hctx, &hctx_list, false);
75 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
78 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
79 * its queue by itself in its completion handler, so we don't need to
80 * restart queue if .get_budget() fails to get the budget.
82 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
83 * be run again. This is necessary to avoid starving flushes.
85 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
87 struct request_queue *q = hctx->queue;
88 struct elevator_queue *e = q->elevator;
89 bool multi_hctxs = false, run_queue = false;
90 bool dispatched = false, busy = false;
91 unsigned int max_dispatch;
95 if (hctx->dispatch_busy)
98 max_dispatch = hctx->queue->nr_requests;
104 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
107 if (!list_empty_careful(&hctx->dispatch)) {
112 budget_token = blk_mq_get_dispatch_budget(q);
113 if (budget_token < 0)
116 rq = e->type->ops.dispatch_request(hctx);
118 blk_mq_put_dispatch_budget(q, budget_token);
120 * We're releasing without dispatching. Holding the
121 * budget could have blocked any "hctx"s with the
122 * same queue and if we didn't dispatch then there's
123 * no guarantee anyone will kick the queue. Kick it
130 blk_mq_set_rq_budget_token(rq, budget_token);
133 * Now this rq owns the budget which has to be released
134 * if this rq won't be queued to driver via .queue_rq()
135 * in blk_mq_dispatch_rq_list().
137 list_add_tail(&rq->queuelist, &rq_list);
139 if (rq->mq_hctx != hctx)
143 * If we cannot get tag for the request, stop dequeueing
144 * requests from the IO scheduler. We are unlikely to be able
145 * to submit them anyway and it creates false impression for
146 * scheduling heuristics that the device can take more IO.
148 if (!blk_mq_get_driver_tag(rq))
150 } while (count < max_dispatch);
154 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
155 } else if (multi_hctxs) {
157 * Requests from different hctx may be dequeued from some
158 * schedulers, such as bfq and deadline.
160 * Sort the requests in the list according to their hctx,
161 * dispatch batching requests from same hctx at a time.
163 list_sort(NULL, &rq_list, sched_rq_cmp);
165 dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
166 } while (!list_empty(&rq_list));
168 dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, false);
176 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
178 unsigned long end = jiffies + HZ;
182 ret = __blk_mq_do_dispatch_sched(hctx);
185 if (need_resched() || time_is_before_jiffies(end)) {
186 blk_mq_delay_run_hw_queue(hctx, 0);
194 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
195 struct blk_mq_ctx *ctx)
197 unsigned short idx = ctx->index_hw[hctx->type];
199 if (++idx == hctx->nr_ctx)
202 return hctx->ctxs[idx];
206 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
207 * its queue by itself in its completion handler, so we don't need to
208 * restart queue if .get_budget() fails to get the budget.
210 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
211 * be run again. This is necessary to avoid starving flushes.
213 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
215 struct request_queue *q = hctx->queue;
217 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
224 if (!list_empty_careful(&hctx->dispatch)) {
229 if (!sbitmap_any_bit_set(&hctx->ctx_map))
232 budget_token = blk_mq_get_dispatch_budget(q);
233 if (budget_token < 0)
236 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
238 blk_mq_put_dispatch_budget(q, budget_token);
240 * We're releasing without dispatching. Holding the
241 * budget could have blocked any "hctx"s with the
242 * same queue and if we didn't dispatch then there's
243 * no guarantee anyone will kick the queue. Kick it
246 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
250 blk_mq_set_rq_budget_token(rq, budget_token);
253 * Now this rq owns the budget which has to be released
254 * if this rq won't be queued to driver via .queue_rq()
255 * in blk_mq_dispatch_rq_list().
257 list_add(&rq->queuelist, &rq_list);
259 /* round robin for fair dispatch */
260 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
262 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, false));
264 WRITE_ONCE(hctx->dispatch_from, ctx);
268 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
270 bool need_dispatch = false;
274 * If we have previous entries on our dispatch list, grab them first for
275 * more fair dispatch.
277 if (!list_empty_careful(&hctx->dispatch)) {
278 spin_lock(&hctx->lock);
279 if (!list_empty(&hctx->dispatch))
280 list_splice_init(&hctx->dispatch, &rq_list);
281 spin_unlock(&hctx->lock);
285 * Only ask the scheduler for requests, if we didn't have residual
286 * requests from the dispatch list. This is to avoid the case where
287 * we only ever dispatch a fraction of the requests available because
288 * of low device queue depth. Once we pull requests out of the IO
289 * scheduler, we can no longer merge or sort them. So it's best to
290 * leave them there for as long as we can. Mark the hw queue as
291 * needing a restart in that case.
293 * We want to dispatch from the scheduler if there was nothing
294 * on the dispatch list or we were able to dispatch from the
297 if (!list_empty(&rq_list)) {
298 blk_mq_sched_mark_restart_hctx(hctx);
299 if (!blk_mq_dispatch_rq_list(hctx, &rq_list, true))
301 need_dispatch = true;
303 need_dispatch = hctx->dispatch_busy;
306 if (hctx->queue->elevator)
307 return blk_mq_do_dispatch_sched(hctx);
309 /* dequeue request one by one from sw queue if queue is busy */
311 return blk_mq_do_dispatch_ctx(hctx);
312 blk_mq_flush_busy_ctxs(hctx, &rq_list);
313 blk_mq_dispatch_rq_list(hctx, &rq_list, true);
317 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
319 struct request_queue *q = hctx->queue;
321 /* RCU or SRCU read lock is needed before checking quiesced flag */
322 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
326 * A return of -EAGAIN is an indication that hctx->dispatch is not
327 * empty and we must run again in order to avoid starving flushes.
329 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
330 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
331 blk_mq_run_hw_queue(hctx, true);
335 bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
336 unsigned int nr_segs)
338 struct elevator_queue *e = q->elevator;
339 struct blk_mq_ctx *ctx;
340 struct blk_mq_hw_ctx *hctx;
344 if (e && e->type->ops.bio_merge) {
345 ret = e->type->ops.bio_merge(q, bio, nr_segs);
349 ctx = blk_mq_get_ctx(q);
350 hctx = blk_mq_map_queue(bio->bi_opf, ctx);
352 if (list_empty_careful(&ctx->rq_lists[type]))
355 /* default per sw-queue merge */
356 spin_lock(&ctx->lock);
358 * Reverse check our software queue for entries that we could
359 * potentially merge with. Currently includes a hand-wavy stop
360 * count of 8, to not spend too much time checking for merges.
362 if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs))
365 spin_unlock(&ctx->lock);
370 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
371 struct list_head *free)
373 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
375 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
377 static int blk_mq_sched_alloc_map_and_rqs(struct request_queue *q,
378 struct blk_mq_hw_ctx *hctx,
379 unsigned int hctx_idx)
381 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
382 hctx->sched_tags = q->sched_shared_tags;
386 hctx->sched_tags = blk_mq_alloc_map_and_rqs(q->tag_set, hctx_idx,
389 if (!hctx->sched_tags)
394 static void blk_mq_exit_sched_shared_tags(struct request_queue *queue)
396 blk_mq_free_rq_map(queue->sched_shared_tags);
397 queue->sched_shared_tags = NULL;
400 /* called in queue's release handler, tagset has gone away */
401 static void blk_mq_sched_tags_teardown(struct request_queue *q, unsigned int flags)
403 struct blk_mq_hw_ctx *hctx;
406 queue_for_each_hw_ctx(q, hctx, i) {
407 if (hctx->sched_tags) {
408 if (!blk_mq_is_shared_tags(flags))
409 blk_mq_free_rq_map(hctx->sched_tags);
410 hctx->sched_tags = NULL;
414 if (blk_mq_is_shared_tags(flags))
415 blk_mq_exit_sched_shared_tags(q);
418 static int blk_mq_init_sched_shared_tags(struct request_queue *queue)
420 struct blk_mq_tag_set *set = queue->tag_set;
423 * Set initial depth at max so that we don't need to reallocate for
424 * updating nr_requests.
426 queue->sched_shared_tags = blk_mq_alloc_map_and_rqs(set,
429 if (!queue->sched_shared_tags)
432 blk_mq_tag_update_sched_shared_tags(queue);
437 void blk_mq_sched_reg_debugfs(struct request_queue *q)
439 struct blk_mq_hw_ctx *hctx;
442 mutex_lock(&q->debugfs_mutex);
443 blk_mq_debugfs_register_sched(q);
444 queue_for_each_hw_ctx(q, hctx, i)
445 blk_mq_debugfs_register_sched_hctx(q, hctx);
446 mutex_unlock(&q->debugfs_mutex);
449 void blk_mq_sched_unreg_debugfs(struct request_queue *q)
451 struct blk_mq_hw_ctx *hctx;
454 mutex_lock(&q->debugfs_mutex);
455 queue_for_each_hw_ctx(q, hctx, i)
456 blk_mq_debugfs_unregister_sched_hctx(hctx);
457 blk_mq_debugfs_unregister_sched(q);
458 mutex_unlock(&q->debugfs_mutex);
461 /* caller must have a reference to @e, will grab another one if successful */
462 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
464 unsigned int flags = q->tag_set->flags;
465 struct blk_mq_hw_ctx *hctx;
466 struct elevator_queue *eq;
471 * Default to double of smaller one between hw queue_depth and 128,
472 * since we don't split into sync/async like the old code did.
473 * Additionally, this is a per-hw queue depth.
475 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
478 if (blk_mq_is_shared_tags(flags)) {
479 ret = blk_mq_init_sched_shared_tags(q);
484 queue_for_each_hw_ctx(q, hctx, i) {
485 ret = blk_mq_sched_alloc_map_and_rqs(q, hctx, i);
487 goto err_free_map_and_rqs;
490 ret = e->ops.init_sched(q, e);
492 goto err_free_map_and_rqs;
494 queue_for_each_hw_ctx(q, hctx, i) {
495 if (e->ops.init_hctx) {
496 ret = e->ops.init_hctx(hctx, i);
499 blk_mq_sched_free_rqs(q);
500 blk_mq_exit_sched(q, eq);
501 kobject_put(&eq->kobj);
508 err_free_map_and_rqs:
509 blk_mq_sched_free_rqs(q);
510 blk_mq_sched_tags_teardown(q, flags);
517 * called in either blk_queue_cleanup or elevator_switch, tagset
518 * is required for freeing requests
520 void blk_mq_sched_free_rqs(struct request_queue *q)
522 struct blk_mq_hw_ctx *hctx;
525 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
526 blk_mq_free_rqs(q->tag_set, q->sched_shared_tags,
529 queue_for_each_hw_ctx(q, hctx, i) {
530 if (hctx->sched_tags)
531 blk_mq_free_rqs(q->tag_set,
532 hctx->sched_tags, i);
537 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
539 struct blk_mq_hw_ctx *hctx;
541 unsigned int flags = 0;
543 queue_for_each_hw_ctx(q, hctx, i) {
544 if (e->type->ops.exit_hctx && hctx->sched_data) {
545 e->type->ops.exit_hctx(hctx, i);
546 hctx->sched_data = NULL;
551 if (e->type->ops.exit_sched)
552 e->type->ops.exit_sched(e);
553 blk_mq_sched_tags_teardown(q, flags);
554 set_bit(ELEVATOR_FLAG_DYING, &q->elevator->flags);