2 * blk-mq scheduling framework
4 * Copyright (C) 2016 Jens Axboe
6 #include <linux/kernel.h>
7 #include <linux/module.h>
8 #include <linux/blk-mq.h>
10 #include <trace/events/block.h>
14 #include "blk-mq-debugfs.h"
15 #include "blk-mq-sched.h"
16 #include "blk-mq-tag.h"
19 void blk_mq_sched_free_hctx_data(struct request_queue *q,
20 void (*exit)(struct blk_mq_hw_ctx *))
22 struct blk_mq_hw_ctx *hctx;
25 queue_for_each_hw_ctx(q, hctx, i) {
26 if (exit && hctx->sched_data)
28 kfree(hctx->sched_data);
29 hctx->sched_data = NULL;
32 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
34 void blk_mq_sched_assign_ioc(struct request *rq, struct bio *bio)
36 struct request_queue *q = rq->q;
37 struct io_context *ioc = rq_ioc(bio);
40 spin_lock_irq(q->queue_lock);
41 icq = ioc_lookup_icq(ioc, q);
42 spin_unlock_irq(q->queue_lock);
45 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
49 get_io_context(icq->ioc);
54 * Mark a hardware queue as needing a restart. For shared queues, maintain
55 * a count of how many hardware queues are marked for restart.
57 static void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
59 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
62 if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
63 struct request_queue *q = hctx->queue;
65 if (!test_and_set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
66 atomic_inc(&q->shared_hctx_restart);
68 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
71 static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
73 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
76 if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
77 struct request_queue *q = hctx->queue;
79 if (test_and_clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
80 atomic_dec(&q->shared_hctx_restart);
82 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
84 return blk_mq_run_hw_queue(hctx, true);
88 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
89 * its queue by itself in its completion handler, so we don't need to
90 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
92 static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
94 struct request_queue *q = hctx->queue;
95 struct elevator_queue *e = q->elevator;
101 if (e->type->ops.mq.has_work &&
102 !e->type->ops.mq.has_work(hctx))
105 if (!blk_mq_get_dispatch_budget(hctx))
108 rq = e->type->ops.mq.dispatch_request(hctx);
110 blk_mq_put_dispatch_budget(hctx);
115 * Now this rq owns the budget which has to be released
116 * if this rq won't be queued to driver via .queue_rq()
117 * in blk_mq_dispatch_rq_list().
119 list_add(&rq->queuelist, &rq_list);
120 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
123 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
124 struct blk_mq_ctx *ctx)
126 unsigned idx = ctx->index_hw;
128 if (++idx == hctx->nr_ctx)
131 return hctx->ctxs[idx];
135 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
136 * its queue by itself in its completion handler, so we don't need to
137 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
139 static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
141 struct request_queue *q = hctx->queue;
143 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
148 if (!sbitmap_any_bit_set(&hctx->ctx_map))
151 if (!blk_mq_get_dispatch_budget(hctx))
154 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
156 blk_mq_put_dispatch_budget(hctx);
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(&rq->queuelist, &rq_list);
167 /* round robin for fair dispatch */
168 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
170 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
172 WRITE_ONCE(hctx->dispatch_from, ctx);
175 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
177 struct request_queue *q = hctx->queue;
178 struct elevator_queue *e = q->elevator;
179 const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
182 /* RCU or SRCU read lock is needed before checking quiesced flag */
183 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
189 * If we have previous entries on our dispatch list, grab them first for
190 * more fair dispatch.
192 if (!list_empty_careful(&hctx->dispatch)) {
193 spin_lock(&hctx->lock);
194 if (!list_empty(&hctx->dispatch))
195 list_splice_init(&hctx->dispatch, &rq_list);
196 spin_unlock(&hctx->lock);
200 * Only ask the scheduler for requests, if we didn't have residual
201 * requests from the dispatch list. This is to avoid the case where
202 * we only ever dispatch a fraction of the requests available because
203 * of low device queue depth. Once we pull requests out of the IO
204 * scheduler, we can no longer merge or sort them. So it's best to
205 * leave them there for as long as we can. Mark the hw queue as
206 * needing a restart in that case.
208 * We want to dispatch from the scheduler if there was nothing
209 * on the dispatch list or we were able to dispatch from the
212 if (!list_empty(&rq_list)) {
213 blk_mq_sched_mark_restart_hctx(hctx);
214 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
215 if (has_sched_dispatch)
216 blk_mq_do_dispatch_sched(hctx);
218 blk_mq_do_dispatch_ctx(hctx);
220 } else if (has_sched_dispatch) {
221 blk_mq_do_dispatch_sched(hctx);
222 } else if (q->mq_ops->get_budget) {
224 * If we need to get budget before queuing request, we
225 * dequeue request one by one from sw queue for avoiding
226 * to mess up I/O merge when dispatch runs out of resource.
228 * TODO: get more budgets, and dequeue more requests in
231 blk_mq_do_dispatch_ctx(hctx);
233 blk_mq_flush_busy_ctxs(hctx, &rq_list);
234 blk_mq_dispatch_rq_list(q, &rq_list, false);
238 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
239 struct request **merged_request)
243 switch (elv_merge(q, &rq, bio)) {
244 case ELEVATOR_BACK_MERGE:
245 if (!blk_mq_sched_allow_merge(q, rq, bio))
247 if (!bio_attempt_back_merge(q, rq, bio))
249 *merged_request = attempt_back_merge(q, rq);
250 if (!*merged_request)
251 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
253 case ELEVATOR_FRONT_MERGE:
254 if (!blk_mq_sched_allow_merge(q, rq, bio))
256 if (!bio_attempt_front_merge(q, rq, bio))
258 *merged_request = attempt_front_merge(q, rq);
259 if (!*merged_request)
260 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
266 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
269 * Reverse check our software queue for entries that we could potentially
270 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
271 * too much time checking for merges.
273 static bool blk_mq_attempt_merge(struct request_queue *q,
274 struct blk_mq_ctx *ctx, struct bio *bio)
279 lockdep_assert_held(&ctx->lock);
281 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
287 if (!blk_rq_merge_ok(rq, bio))
290 switch (blk_try_merge(rq, bio)) {
291 case ELEVATOR_BACK_MERGE:
292 if (blk_mq_sched_allow_merge(q, rq, bio))
293 merged = bio_attempt_back_merge(q, rq, bio);
295 case ELEVATOR_FRONT_MERGE:
296 if (blk_mq_sched_allow_merge(q, rq, bio))
297 merged = bio_attempt_front_merge(q, rq, bio);
299 case ELEVATOR_DISCARD_MERGE:
300 merged = bio_attempt_discard_merge(q, rq, bio);
314 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
316 struct elevator_queue *e = q->elevator;
317 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
318 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
321 if (e && e->type->ops.mq.bio_merge) {
323 return e->type->ops.mq.bio_merge(hctx, bio);
326 if (hctx->flags & BLK_MQ_F_SHOULD_MERGE) {
327 /* default per sw-queue merge */
328 spin_lock(&ctx->lock);
329 ret = blk_mq_attempt_merge(q, ctx, bio);
330 spin_unlock(&ctx->lock);
337 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
339 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
341 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
343 void blk_mq_sched_request_inserted(struct request *rq)
345 trace_block_rq_insert(rq->q, rq);
347 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
349 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
353 /* dispatch flush rq directly */
354 if (rq->rq_flags & RQF_FLUSH_SEQ) {
355 spin_lock(&hctx->lock);
356 list_add(&rq->queuelist, &hctx->dispatch);
357 spin_unlock(&hctx->lock);
362 rq->rq_flags |= RQF_SORTED;
368 * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
370 * @skip: the list element that will not be examined. Iteration starts at
372 * @head: head of the list to examine. This list must have at least one
373 * element, namely @skip.
374 * @member: name of the list_head structure within typeof(*pos).
376 #define list_for_each_entry_rcu_rr(pos, skip, head, member) \
377 for ((pos) = (skip); \
378 (pos = (pos)->member.next != (head) ? list_entry_rcu( \
379 (pos)->member.next, typeof(*pos), member) : \
380 list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
384 * Called after a driver tag has been freed to check whether a hctx needs to
385 * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
386 * queues in a round-robin fashion if the tag set of @hctx is shared with other
389 void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
391 struct blk_mq_tags *const tags = hctx->tags;
392 struct blk_mq_tag_set *const set = hctx->queue->tag_set;
393 struct request_queue *const queue = hctx->queue, *q;
394 struct blk_mq_hw_ctx *hctx2;
397 if (set->flags & BLK_MQ_F_TAG_SHARED) {
399 * If this is 0, then we know that no hardware queues
400 * have RESTART marked. We're done.
402 if (!atomic_read(&queue->shared_hctx_restart))
406 list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
408 queue_for_each_hw_ctx(q, hctx2, i)
409 if (hctx2->tags == tags &&
410 blk_mq_sched_restart_hctx(hctx2))
413 j = hctx->queue_num + 1;
414 for (i = 0; i < queue->nr_hw_queues; i++, j++) {
415 if (j == queue->nr_hw_queues)
417 hctx2 = queue->queue_hw_ctx[j];
418 if (hctx2->tags == tags &&
419 blk_mq_sched_restart_hctx(hctx2))
425 blk_mq_sched_restart_hctx(hctx);
429 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
430 bool run_queue, bool async)
432 struct request_queue *q = rq->q;
433 struct elevator_queue *e = q->elevator;
434 struct blk_mq_ctx *ctx = rq->mq_ctx;
435 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
437 /* flush rq in flush machinery need to be dispatched directly */
438 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
439 blk_insert_flush(rq);
443 WARN_ON(e && (rq->tag != -1));
445 if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
448 if (e && e->type->ops.mq.insert_requests) {
451 list_add(&rq->queuelist, &list);
452 e->type->ops.mq.insert_requests(hctx, &list, at_head);
454 spin_lock(&ctx->lock);
455 __blk_mq_insert_request(hctx, rq, at_head);
456 spin_unlock(&ctx->lock);
461 blk_mq_run_hw_queue(hctx, async);
464 void blk_mq_sched_insert_requests(struct request_queue *q,
465 struct blk_mq_ctx *ctx,
466 struct list_head *list, bool run_queue_async)
468 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
469 struct elevator_queue *e = hctx->queue->elevator;
471 if (e && e->type->ops.mq.insert_requests)
472 e->type->ops.mq.insert_requests(hctx, list, false);
474 blk_mq_insert_requests(hctx, ctx, list);
476 blk_mq_run_hw_queue(hctx, run_queue_async);
479 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
480 struct blk_mq_hw_ctx *hctx,
481 unsigned int hctx_idx)
483 if (hctx->sched_tags) {
484 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
485 blk_mq_free_rq_map(hctx->sched_tags);
486 hctx->sched_tags = NULL;
490 static int blk_mq_sched_alloc_tags(struct request_queue *q,
491 struct blk_mq_hw_ctx *hctx,
492 unsigned int hctx_idx)
494 struct blk_mq_tag_set *set = q->tag_set;
497 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
499 if (!hctx->sched_tags)
502 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
504 blk_mq_sched_free_tags(set, hctx, hctx_idx);
509 static void blk_mq_sched_tags_teardown(struct request_queue *q)
511 struct blk_mq_tag_set *set = q->tag_set;
512 struct blk_mq_hw_ctx *hctx;
515 queue_for_each_hw_ctx(q, hctx, i)
516 blk_mq_sched_free_tags(set, hctx, i);
519 int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
520 unsigned int hctx_idx)
522 struct elevator_queue *e = q->elevator;
528 ret = blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
532 if (e->type->ops.mq.init_hctx) {
533 ret = e->type->ops.mq.init_hctx(hctx, hctx_idx);
535 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
540 blk_mq_debugfs_register_sched_hctx(q, hctx);
545 void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
546 unsigned int hctx_idx)
548 struct elevator_queue *e = q->elevator;
553 blk_mq_debugfs_unregister_sched_hctx(hctx);
555 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
556 e->type->ops.mq.exit_hctx(hctx, hctx_idx);
557 hctx->sched_data = NULL;
560 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
563 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
565 struct blk_mq_hw_ctx *hctx;
566 struct elevator_queue *eq;
576 * Default to double of smaller one between hw queue_depth and 128,
577 * since we don't split into sync/async like the old code did.
578 * Additionally, this is a per-hw queue depth.
580 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
583 queue_for_each_hw_ctx(q, hctx, i) {
584 ret = blk_mq_sched_alloc_tags(q, hctx, i);
589 ret = e->ops.mq.init_sched(q, e);
593 blk_mq_debugfs_register_sched(q);
595 queue_for_each_hw_ctx(q, hctx, i) {
596 if (e->ops.mq.init_hctx) {
597 ret = e->ops.mq.init_hctx(hctx, i);
600 blk_mq_exit_sched(q, eq);
601 kobject_put(&eq->kobj);
605 blk_mq_debugfs_register_sched_hctx(q, hctx);
611 blk_mq_sched_tags_teardown(q);
616 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
618 struct blk_mq_hw_ctx *hctx;
621 queue_for_each_hw_ctx(q, hctx, i) {
622 blk_mq_debugfs_unregister_sched_hctx(hctx);
623 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
624 e->type->ops.mq.exit_hctx(hctx, i);
625 hctx->sched_data = NULL;
628 blk_mq_debugfs_unregister_sched(q);
629 if (e->type->ops.mq.exit_sched)
630 e->type->ops.mq.exit_sched(e);
631 blk_mq_sched_tags_teardown(q);
635 int blk_mq_sched_init(struct request_queue *q)
639 mutex_lock(&q->sysfs_lock);
640 ret = elevator_init(q, NULL);
641 mutex_unlock(&q->sysfs_lock);