2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-debugfs.h"
35 #include "blk-mq-tag.h"
38 #include "blk-mq-sched.h"
40 static DEFINE_MUTEX(all_q_mutex);
41 static LIST_HEAD(all_q_list);
43 static void blk_mq_poll_stats_start(struct request_queue *q);
44 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
45 static void __blk_mq_stop_hw_queues(struct request_queue *q, bool sync);
47 static int blk_mq_poll_stats_bkt(const struct request *rq)
49 int ddir, bytes, bucket;
51 ddir = rq_data_dir(rq);
52 bytes = blk_rq_bytes(rq);
54 bucket = ddir + 2*(ilog2(bytes) - 9);
58 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
59 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
65 * Check if any of the ctx's have pending work in this hardware queue
67 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
69 return sbitmap_any_bit_set(&hctx->ctx_map) ||
70 !list_empty_careful(&hctx->dispatch) ||
71 blk_mq_sched_has_work(hctx);
75 * Mark this ctx as having pending work in this hardware queue
77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
78 struct blk_mq_ctx *ctx)
80 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
81 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
84 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
85 struct blk_mq_ctx *ctx)
87 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
90 void blk_freeze_queue_start(struct request_queue *q)
94 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
95 if (freeze_depth == 1) {
96 percpu_ref_kill(&q->q_usage_counter);
97 blk_mq_run_hw_queues(q, false);
100 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
102 void blk_mq_freeze_queue_wait(struct request_queue *q)
104 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
106 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
108 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
109 unsigned long timeout)
111 return wait_event_timeout(q->mq_freeze_wq,
112 percpu_ref_is_zero(&q->q_usage_counter),
115 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
118 * Guarantee no request is in use, so we can change any data structure of
119 * the queue afterward.
121 void blk_freeze_queue(struct request_queue *q)
124 * In the !blk_mq case we are only calling this to kill the
125 * q_usage_counter, otherwise this increases the freeze depth
126 * and waits for it to return to zero. For this reason there is
127 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
128 * exported to drivers as the only user for unfreeze is blk_mq.
130 blk_freeze_queue_start(q);
131 blk_mq_freeze_queue_wait(q);
134 void blk_mq_freeze_queue(struct request_queue *q)
137 * ...just an alias to keep freeze and unfreeze actions balanced
138 * in the blk_mq_* namespace
142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
144 void blk_mq_unfreeze_queue(struct request_queue *q)
148 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
149 WARN_ON_ONCE(freeze_depth < 0);
151 percpu_ref_reinit(&q->q_usage_counter);
152 wake_up_all(&q->mq_freeze_wq);
155 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
158 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
161 * Note: this function does not prevent that the struct request end_io()
162 * callback function is invoked. Additionally, it is not prevented that
163 * new queue_rq() calls occur unless the queue has been stopped first.
165 void blk_mq_quiesce_queue(struct request_queue *q)
167 struct blk_mq_hw_ctx *hctx;
171 __blk_mq_stop_hw_queues(q, true);
173 queue_for_each_hw_ctx(q, hctx, i) {
174 if (hctx->flags & BLK_MQ_F_BLOCKING)
175 synchronize_srcu(&hctx->queue_rq_srcu);
182 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
184 void blk_mq_wake_waiters(struct request_queue *q)
186 struct blk_mq_hw_ctx *hctx;
189 queue_for_each_hw_ctx(q, hctx, i)
190 if (blk_mq_hw_queue_mapped(hctx))
191 blk_mq_tag_wakeup_all(hctx->tags, true);
194 * If we are called because the queue has now been marked as
195 * dying, we need to ensure that processes currently waiting on
196 * the queue are notified as well.
198 wake_up_all(&q->mq_freeze_wq);
201 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
203 return blk_mq_has_free_tags(hctx->tags);
205 EXPORT_SYMBOL(blk_mq_can_queue);
207 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
208 struct request *rq, unsigned int op)
210 INIT_LIST_HEAD(&rq->queuelist);
211 /* csd/requeue_work/fifo_time is initialized before use */
215 if (blk_queue_io_stat(q))
216 rq->rq_flags |= RQF_IO_STAT;
217 /* do not touch atomic flags, it needs atomic ops against the timer */
219 INIT_HLIST_NODE(&rq->hash);
220 RB_CLEAR_NODE(&rq->rb_node);
223 rq->start_time = jiffies;
224 #ifdef CONFIG_BLK_CGROUP
226 set_start_time_ns(rq);
227 rq->io_start_time_ns = 0;
229 rq->nr_phys_segments = 0;
230 #if defined(CONFIG_BLK_DEV_INTEGRITY)
231 rq->nr_integrity_segments = 0;
234 /* tag was already set */
237 INIT_LIST_HEAD(&rq->timeout_list);
241 rq->end_io_data = NULL;
244 ctx->rq_dispatched[op_is_sync(op)]++;
247 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
253 tag = blk_mq_get_tag(data);
254 if (tag != BLK_MQ_TAG_FAIL) {
255 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
257 rq = tags->static_rqs[tag];
259 if (data->flags & BLK_MQ_REQ_INTERNAL) {
261 rq->internal_tag = tag;
263 if (blk_mq_tag_busy(data->hctx)) {
264 rq->rq_flags = RQF_MQ_INFLIGHT;
265 atomic_inc(&data->hctx->nr_active);
268 rq->internal_tag = -1;
269 data->hctx->tags->rqs[rq->tag] = rq;
272 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
278 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
280 static struct request *blk_mq_get_request(struct request_queue *q,
281 struct bio *bio, unsigned int op,
282 struct blk_mq_alloc_data *data)
284 struct elevator_queue *e = q->elevator;
287 blk_queue_enter_live(q);
289 if (likely(!data->ctx))
290 data->ctx = blk_mq_get_ctx(q);
291 if (likely(!data->hctx))
292 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
295 data->flags |= BLK_MQ_REQ_INTERNAL;
298 * Flush requests are special and go directly to the
301 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
302 e->type->ops.mq.limit_depth(op, data);
305 rq = __blk_mq_alloc_request(data, op);
311 if (!op_is_flush(op)) {
313 if (e && e->type->ops.mq.prepare_request) {
314 if (e->type->icq_cache && rq_ioc(bio))
315 blk_mq_sched_assign_ioc(rq, bio);
317 e->type->ops.mq.prepare_request(rq, bio);
318 rq->rq_flags |= RQF_ELVPRIV;
321 data->hctx->queued++;
325 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
328 struct blk_mq_alloc_data alloc_data = { .flags = flags };
332 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
336 rq = blk_mq_get_request(q, NULL, rw, &alloc_data);
338 blk_mq_put_ctx(alloc_data.ctx);
342 return ERR_PTR(-EWOULDBLOCK);
345 rq->__sector = (sector_t) -1;
346 rq->bio = rq->biotail = NULL;
349 EXPORT_SYMBOL(blk_mq_alloc_request);
351 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
352 unsigned int flags, unsigned int hctx_idx)
354 struct blk_mq_alloc_data alloc_data = { .flags = flags };
360 * If the tag allocator sleeps we could get an allocation for a
361 * different hardware context. No need to complicate the low level
362 * allocator for this for the rare use case of a command tied to
365 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
366 return ERR_PTR(-EINVAL);
368 if (hctx_idx >= q->nr_hw_queues)
369 return ERR_PTR(-EIO);
371 ret = blk_queue_enter(q, true);
376 * Check if the hardware context is actually mapped to anything.
377 * If not tell the caller that it should skip this queue.
379 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
380 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
382 return ERR_PTR(-EXDEV);
384 cpu = cpumask_first(alloc_data.hctx->cpumask);
385 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
387 rq = blk_mq_get_request(q, NULL, rw, &alloc_data);
392 return ERR_PTR(-EWOULDBLOCK);
396 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
398 void blk_mq_free_request(struct request *rq)
400 struct request_queue *q = rq->q;
401 struct elevator_queue *e = q->elevator;
402 struct blk_mq_ctx *ctx = rq->mq_ctx;
403 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
404 const int sched_tag = rq->internal_tag;
406 if (rq->rq_flags & RQF_ELVPRIV) {
407 if (e && e->type->ops.mq.finish_request)
408 e->type->ops.mq.finish_request(rq);
410 put_io_context(rq->elv.icq->ioc);
415 ctx->rq_completed[rq_is_sync(rq)]++;
416 if (rq->rq_flags & RQF_MQ_INFLIGHT)
417 atomic_dec(&hctx->nr_active);
419 wbt_done(q->rq_wb, &rq->issue_stat);
422 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
423 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
425 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
427 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
428 blk_mq_sched_restart(hctx);
431 EXPORT_SYMBOL_GPL(blk_mq_free_request);
433 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
435 blk_account_io_done(rq);
438 wbt_done(rq->q->rq_wb, &rq->issue_stat);
439 rq->end_io(rq, error);
441 if (unlikely(blk_bidi_rq(rq)))
442 blk_mq_free_request(rq->next_rq);
443 blk_mq_free_request(rq);
446 EXPORT_SYMBOL(__blk_mq_end_request);
448 void blk_mq_end_request(struct request *rq, blk_status_t error)
450 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
452 __blk_mq_end_request(rq, error);
454 EXPORT_SYMBOL(blk_mq_end_request);
456 static void __blk_mq_complete_request_remote(void *data)
458 struct request *rq = data;
460 rq->q->softirq_done_fn(rq);
463 static void __blk_mq_complete_request(struct request *rq)
465 struct blk_mq_ctx *ctx = rq->mq_ctx;
469 if (rq->internal_tag != -1)
470 blk_mq_sched_completed_request(rq);
471 if (rq->rq_flags & RQF_STATS) {
472 blk_mq_poll_stats_start(rq->q);
476 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
477 rq->q->softirq_done_fn(rq);
482 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
483 shared = cpus_share_cache(cpu, ctx->cpu);
485 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
486 rq->csd.func = __blk_mq_complete_request_remote;
489 smp_call_function_single_async(ctx->cpu, &rq->csd);
491 rq->q->softirq_done_fn(rq);
497 * blk_mq_complete_request - end I/O on a request
498 * @rq: the request being processed
501 * Ends all I/O on a request. It does not handle partial completions.
502 * The actual completion happens out-of-order, through a IPI handler.
504 void blk_mq_complete_request(struct request *rq)
506 struct request_queue *q = rq->q;
508 if (unlikely(blk_should_fake_timeout(q)))
510 if (!blk_mark_rq_complete(rq))
511 __blk_mq_complete_request(rq);
513 EXPORT_SYMBOL(blk_mq_complete_request);
515 int blk_mq_request_started(struct request *rq)
517 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
519 EXPORT_SYMBOL_GPL(blk_mq_request_started);
521 void blk_mq_start_request(struct request *rq)
523 struct request_queue *q = rq->q;
525 blk_mq_sched_started_request(rq);
527 trace_block_rq_issue(q, rq);
529 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
530 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
531 rq->rq_flags |= RQF_STATS;
532 wbt_issue(q->rq_wb, &rq->issue_stat);
538 * Ensure that ->deadline is visible before set the started
539 * flag and clear the completed flag.
541 smp_mb__before_atomic();
544 * Mark us as started and clear complete. Complete might have been
545 * set if requeue raced with timeout, which then marked it as
546 * complete. So be sure to clear complete again when we start
547 * the request, otherwise we'll ignore the completion event.
549 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
550 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
551 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
552 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
554 if (q->dma_drain_size && blk_rq_bytes(rq)) {
556 * Make sure space for the drain appears. We know we can do
557 * this because max_hw_segments has been adjusted to be one
558 * fewer than the device can handle.
560 rq->nr_phys_segments++;
563 EXPORT_SYMBOL(blk_mq_start_request);
566 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
567 * flag isn't set yet, so there may be race with timeout handler,
568 * but given rq->deadline is just set in .queue_rq() under
569 * this situation, the race won't be possible in reality because
570 * rq->timeout should be set as big enough to cover the window
571 * between blk_mq_start_request() called from .queue_rq() and
572 * clearing REQ_ATOM_STARTED here.
574 static void __blk_mq_requeue_request(struct request *rq)
576 struct request_queue *q = rq->q;
578 trace_block_rq_requeue(q, rq);
579 wbt_requeue(q->rq_wb, &rq->issue_stat);
580 blk_mq_sched_requeue_request(rq);
582 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
583 if (q->dma_drain_size && blk_rq_bytes(rq))
584 rq->nr_phys_segments--;
588 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
590 __blk_mq_requeue_request(rq);
592 BUG_ON(blk_queued_rq(rq));
593 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
595 EXPORT_SYMBOL(blk_mq_requeue_request);
597 static void blk_mq_requeue_work(struct work_struct *work)
599 struct request_queue *q =
600 container_of(work, struct request_queue, requeue_work.work);
602 struct request *rq, *next;
605 spin_lock_irqsave(&q->requeue_lock, flags);
606 list_splice_init(&q->requeue_list, &rq_list);
607 spin_unlock_irqrestore(&q->requeue_lock, flags);
609 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
610 if (!(rq->rq_flags & RQF_SOFTBARRIER))
613 rq->rq_flags &= ~RQF_SOFTBARRIER;
614 list_del_init(&rq->queuelist);
615 blk_mq_sched_insert_request(rq, true, false, false, true);
618 while (!list_empty(&rq_list)) {
619 rq = list_entry(rq_list.next, struct request, queuelist);
620 list_del_init(&rq->queuelist);
621 blk_mq_sched_insert_request(rq, false, false, false, true);
624 blk_mq_run_hw_queues(q, false);
627 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
628 bool kick_requeue_list)
630 struct request_queue *q = rq->q;
634 * We abuse this flag that is otherwise used by the I/O scheduler to
635 * request head insertation from the workqueue.
637 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
639 spin_lock_irqsave(&q->requeue_lock, flags);
641 rq->rq_flags |= RQF_SOFTBARRIER;
642 list_add(&rq->queuelist, &q->requeue_list);
644 list_add_tail(&rq->queuelist, &q->requeue_list);
646 spin_unlock_irqrestore(&q->requeue_lock, flags);
648 if (kick_requeue_list)
649 blk_mq_kick_requeue_list(q);
651 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
653 void blk_mq_kick_requeue_list(struct request_queue *q)
655 kblockd_schedule_delayed_work(&q->requeue_work, 0);
657 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
659 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
662 kblockd_schedule_delayed_work(&q->requeue_work,
663 msecs_to_jiffies(msecs));
665 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
667 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
669 if (tag < tags->nr_tags) {
670 prefetch(tags->rqs[tag]);
671 return tags->rqs[tag];
676 EXPORT_SYMBOL(blk_mq_tag_to_rq);
678 struct blk_mq_timeout_data {
680 unsigned int next_set;
683 void blk_mq_rq_timed_out(struct request *req, bool reserved)
685 const struct blk_mq_ops *ops = req->q->mq_ops;
686 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
689 * We know that complete is set at this point. If STARTED isn't set
690 * anymore, then the request isn't active and the "timeout" should
691 * just be ignored. This can happen due to the bitflag ordering.
692 * Timeout first checks if STARTED is set, and if it is, assumes
693 * the request is active. But if we race with completion, then
694 * both flags will get cleared. So check here again, and ignore
695 * a timeout event with a request that isn't active.
697 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
701 ret = ops->timeout(req, reserved);
705 __blk_mq_complete_request(req);
707 case BLK_EH_RESET_TIMER:
709 blk_clear_rq_complete(req);
711 case BLK_EH_NOT_HANDLED:
714 printk(KERN_ERR "block: bad eh return: %d\n", ret);
719 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
720 struct request *rq, void *priv, bool reserved)
722 struct blk_mq_timeout_data *data = priv;
724 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
728 * The rq being checked may have been freed and reallocated
729 * out already here, we avoid this race by checking rq->deadline
730 * and REQ_ATOM_COMPLETE flag together:
732 * - if rq->deadline is observed as new value because of
733 * reusing, the rq won't be timed out because of timing.
734 * - if rq->deadline is observed as previous value,
735 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
736 * because we put a barrier between setting rq->deadline
737 * and clearing the flag in blk_mq_start_request(), so
738 * this rq won't be timed out too.
740 if (time_after_eq(jiffies, rq->deadline)) {
741 if (!blk_mark_rq_complete(rq))
742 blk_mq_rq_timed_out(rq, reserved);
743 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
744 data->next = rq->deadline;
749 static void blk_mq_timeout_work(struct work_struct *work)
751 struct request_queue *q =
752 container_of(work, struct request_queue, timeout_work);
753 struct blk_mq_timeout_data data = {
759 /* A deadlock might occur if a request is stuck requiring a
760 * timeout at the same time a queue freeze is waiting
761 * completion, since the timeout code would not be able to
762 * acquire the queue reference here.
764 * That's why we don't use blk_queue_enter here; instead, we use
765 * percpu_ref_tryget directly, because we need to be able to
766 * obtain a reference even in the short window between the queue
767 * starting to freeze, by dropping the first reference in
768 * blk_freeze_queue_start, and the moment the last request is
769 * consumed, marked by the instant q_usage_counter reaches
772 if (!percpu_ref_tryget(&q->q_usage_counter))
775 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
778 data.next = blk_rq_timeout(round_jiffies_up(data.next));
779 mod_timer(&q->timeout, data.next);
781 struct blk_mq_hw_ctx *hctx;
783 queue_for_each_hw_ctx(q, hctx, i) {
784 /* the hctx may be unmapped, so check it here */
785 if (blk_mq_hw_queue_mapped(hctx))
786 blk_mq_tag_idle(hctx);
792 struct flush_busy_ctx_data {
793 struct blk_mq_hw_ctx *hctx;
794 struct list_head *list;
797 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
799 struct flush_busy_ctx_data *flush_data = data;
800 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
801 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
803 sbitmap_clear_bit(sb, bitnr);
804 spin_lock(&ctx->lock);
805 list_splice_tail_init(&ctx->rq_list, flush_data->list);
806 spin_unlock(&ctx->lock);
811 * Process software queues that have been marked busy, splicing them
812 * to the for-dispatch
814 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
816 struct flush_busy_ctx_data data = {
821 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
823 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
825 static inline unsigned int queued_to_index(unsigned int queued)
830 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
833 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
836 struct blk_mq_alloc_data data = {
838 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
839 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
842 might_sleep_if(wait);
847 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
848 data.flags |= BLK_MQ_REQ_RESERVED;
850 rq->tag = blk_mq_get_tag(&data);
852 if (blk_mq_tag_busy(data.hctx)) {
853 rq->rq_flags |= RQF_MQ_INFLIGHT;
854 atomic_inc(&data.hctx->nr_active);
856 data.hctx->tags->rqs[rq->tag] = rq;
862 return rq->tag != -1;
865 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
868 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
871 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
872 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
873 atomic_dec(&hctx->nr_active);
877 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
880 if (rq->tag == -1 || rq->internal_tag == -1)
883 __blk_mq_put_driver_tag(hctx, rq);
886 static void blk_mq_put_driver_tag(struct request *rq)
888 struct blk_mq_hw_ctx *hctx;
890 if (rq->tag == -1 || rq->internal_tag == -1)
893 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
894 __blk_mq_put_driver_tag(hctx, rq);
898 * If we fail getting a driver tag because all the driver tags are already
899 * assigned and on the dispatch list, BUT the first entry does not have a
900 * tag, then we could deadlock. For that case, move entries with assigned
901 * driver tags to the front, leaving the set of tagged requests in the
902 * same order, and the untagged set in the same order.
904 static bool reorder_tags_to_front(struct list_head *list)
906 struct request *rq, *tmp, *first = NULL;
908 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
912 list_move(&rq->queuelist, list);
918 return first != NULL;
921 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
924 struct blk_mq_hw_ctx *hctx;
926 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
928 list_del(&wait->task_list);
929 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
930 blk_mq_run_hw_queue(hctx, true);
934 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
936 struct sbq_wait_state *ws;
939 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
940 * The thread which wins the race to grab this bit adds the hardware
941 * queue to the wait queue.
943 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
944 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
947 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
948 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
951 * As soon as this returns, it's no longer safe to fiddle with
952 * hctx->dispatch_wait, since a completion can wake up the wait queue
953 * and unlock the bit.
955 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
959 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
961 struct blk_mq_hw_ctx *hctx;
965 if (list_empty(list))
969 * Now process all the entries, sending them to the driver.
973 struct blk_mq_queue_data bd;
976 rq = list_first_entry(list, struct request, queuelist);
977 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
978 if (!queued && reorder_tags_to_front(list))
982 * The initial allocation attempt failed, so we need to
983 * rerun the hardware queue when a tag is freed.
985 if (!blk_mq_dispatch_wait_add(hctx))
989 * It's possible that a tag was freed in the window
990 * between the allocation failure and adding the
991 * hardware queue to the wait queue.
993 if (!blk_mq_get_driver_tag(rq, &hctx, false))
997 list_del_init(&rq->queuelist);
1002 * Flag last if we have no more requests, or if we have more
1003 * but can't assign a driver tag to it.
1005 if (list_empty(list))
1008 struct request *nxt;
1010 nxt = list_first_entry(list, struct request, queuelist);
1011 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1014 ret = q->mq_ops->queue_rq(hctx, &bd);
1015 if (ret == BLK_STS_RESOURCE) {
1016 blk_mq_put_driver_tag_hctx(hctx, rq);
1017 list_add(&rq->queuelist, list);
1018 __blk_mq_requeue_request(rq);
1022 if (unlikely(ret != BLK_STS_OK)) {
1024 blk_mq_end_request(rq, BLK_STS_IOERR);
1029 } while (!list_empty(list));
1031 hctx->dispatched[queued_to_index(queued)]++;
1034 * Any items that need requeuing? Stuff them into hctx->dispatch,
1035 * that is where we will continue on next queue run.
1037 if (!list_empty(list)) {
1039 * If an I/O scheduler has been configured and we got a driver
1040 * tag for the next request already, free it again.
1042 rq = list_first_entry(list, struct request, queuelist);
1043 blk_mq_put_driver_tag(rq);
1045 spin_lock(&hctx->lock);
1046 list_splice_init(list, &hctx->dispatch);
1047 spin_unlock(&hctx->lock);
1050 * If SCHED_RESTART was set by the caller of this function and
1051 * it is no longer set that means that it was cleared by another
1052 * thread and hence that a queue rerun is needed.
1054 * If TAG_WAITING is set that means that an I/O scheduler has
1055 * been configured and another thread is waiting for a driver
1056 * tag. To guarantee fairness, do not rerun this hardware queue
1057 * but let the other thread grab the driver tag.
1059 * If no I/O scheduler has been configured it is possible that
1060 * the hardware queue got stopped and restarted before requests
1061 * were pushed back onto the dispatch list. Rerun the queue to
1062 * avoid starvation. Notes:
1063 * - blk_mq_run_hw_queue() checks whether or not a queue has
1064 * been stopped before rerunning a queue.
1065 * - Some but not all block drivers stop a queue before
1066 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1069 if (!blk_mq_sched_needs_restart(hctx) &&
1070 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1071 blk_mq_run_hw_queue(hctx, true);
1074 return (queued + errors) != 0;
1077 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1081 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1082 cpu_online(hctx->next_cpu));
1084 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1086 blk_mq_sched_dispatch_requests(hctx);
1091 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1092 blk_mq_sched_dispatch_requests(hctx);
1093 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1098 * It'd be great if the workqueue API had a way to pass
1099 * in a mask and had some smarts for more clever placement.
1100 * For now we just round-robin here, switching for every
1101 * BLK_MQ_CPU_WORK_BATCH queued items.
1103 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1105 if (hctx->queue->nr_hw_queues == 1)
1106 return WORK_CPU_UNBOUND;
1108 if (--hctx->next_cpu_batch <= 0) {
1111 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1112 if (next_cpu >= nr_cpu_ids)
1113 next_cpu = cpumask_first(hctx->cpumask);
1115 hctx->next_cpu = next_cpu;
1116 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1119 return hctx->next_cpu;
1122 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1123 unsigned long msecs)
1125 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1126 !blk_mq_hw_queue_mapped(hctx)))
1129 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1130 int cpu = get_cpu();
1131 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1132 __blk_mq_run_hw_queue(hctx);
1140 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1142 msecs_to_jiffies(msecs));
1145 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1147 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1149 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1151 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1153 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1155 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1157 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1159 struct blk_mq_hw_ctx *hctx;
1162 queue_for_each_hw_ctx(q, hctx, i) {
1163 if (!blk_mq_hctx_has_pending(hctx) ||
1164 blk_mq_hctx_stopped(hctx))
1167 blk_mq_run_hw_queue(hctx, async);
1170 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1173 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1174 * @q: request queue.
1176 * The caller is responsible for serializing this function against
1177 * blk_mq_{start,stop}_hw_queue().
1179 bool blk_mq_queue_stopped(struct request_queue *q)
1181 struct blk_mq_hw_ctx *hctx;
1184 queue_for_each_hw_ctx(q, hctx, i)
1185 if (blk_mq_hctx_stopped(hctx))
1190 EXPORT_SYMBOL(blk_mq_queue_stopped);
1192 static void __blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx, bool sync)
1195 cancel_delayed_work_sync(&hctx->run_work);
1197 cancel_delayed_work(&hctx->run_work);
1199 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1202 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1204 __blk_mq_stop_hw_queue(hctx, false);
1206 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1208 static void __blk_mq_stop_hw_queues(struct request_queue *q, bool sync)
1210 struct blk_mq_hw_ctx *hctx;
1213 queue_for_each_hw_ctx(q, hctx, i)
1214 __blk_mq_stop_hw_queue(hctx, sync);
1217 void blk_mq_stop_hw_queues(struct request_queue *q)
1219 __blk_mq_stop_hw_queues(q, false);
1221 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1223 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1225 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1227 blk_mq_run_hw_queue(hctx, false);
1229 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1231 void blk_mq_start_hw_queues(struct request_queue *q)
1233 struct blk_mq_hw_ctx *hctx;
1236 queue_for_each_hw_ctx(q, hctx, i)
1237 blk_mq_start_hw_queue(hctx);
1239 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1241 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1243 if (!blk_mq_hctx_stopped(hctx))
1246 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1247 blk_mq_run_hw_queue(hctx, async);
1249 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1251 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1253 struct blk_mq_hw_ctx *hctx;
1256 queue_for_each_hw_ctx(q, hctx, i)
1257 blk_mq_start_stopped_hw_queue(hctx, async);
1259 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1261 static void blk_mq_run_work_fn(struct work_struct *work)
1263 struct blk_mq_hw_ctx *hctx;
1265 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1268 * If we are stopped, don't run the queue. The exception is if
1269 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1270 * the STOPPED bit and run it.
1272 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1273 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1276 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1277 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1280 __blk_mq_run_hw_queue(hctx);
1284 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1286 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1290 * Stop the hw queue, then modify currently delayed work.
1291 * This should prevent us from running the queue prematurely.
1292 * Mark the queue as auto-clearing STOPPED when it runs.
1294 blk_mq_stop_hw_queue(hctx);
1295 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1296 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1298 msecs_to_jiffies(msecs));
1300 EXPORT_SYMBOL(blk_mq_delay_queue);
1302 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1306 struct blk_mq_ctx *ctx = rq->mq_ctx;
1308 trace_block_rq_insert(hctx->queue, rq);
1311 list_add(&rq->queuelist, &ctx->rq_list);
1313 list_add_tail(&rq->queuelist, &ctx->rq_list);
1316 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1319 struct blk_mq_ctx *ctx = rq->mq_ctx;
1321 __blk_mq_insert_req_list(hctx, rq, at_head);
1322 blk_mq_hctx_mark_pending(hctx, ctx);
1325 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1326 struct list_head *list)
1330 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1333 spin_lock(&ctx->lock);
1334 while (!list_empty(list)) {
1337 rq = list_first_entry(list, struct request, queuelist);
1338 BUG_ON(rq->mq_ctx != ctx);
1339 list_del_init(&rq->queuelist);
1340 __blk_mq_insert_req_list(hctx, rq, false);
1342 blk_mq_hctx_mark_pending(hctx, ctx);
1343 spin_unlock(&ctx->lock);
1346 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1348 struct request *rqa = container_of(a, struct request, queuelist);
1349 struct request *rqb = container_of(b, struct request, queuelist);
1351 return !(rqa->mq_ctx < rqb->mq_ctx ||
1352 (rqa->mq_ctx == rqb->mq_ctx &&
1353 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1356 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1358 struct blk_mq_ctx *this_ctx;
1359 struct request_queue *this_q;
1362 LIST_HEAD(ctx_list);
1365 list_splice_init(&plug->mq_list, &list);
1367 list_sort(NULL, &list, plug_ctx_cmp);
1373 while (!list_empty(&list)) {
1374 rq = list_entry_rq(list.next);
1375 list_del_init(&rq->queuelist);
1377 if (rq->mq_ctx != this_ctx) {
1379 trace_block_unplug(this_q, depth, from_schedule);
1380 blk_mq_sched_insert_requests(this_q, this_ctx,
1385 this_ctx = rq->mq_ctx;
1391 list_add_tail(&rq->queuelist, &ctx_list);
1395 * If 'this_ctx' is set, we know we have entries to complete
1396 * on 'ctx_list'. Do those.
1399 trace_block_unplug(this_q, depth, from_schedule);
1400 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1405 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1407 blk_init_request_from_bio(rq, bio);
1409 blk_account_io_start(rq, true);
1412 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1414 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1415 !blk_queue_nomerges(hctx->queue);
1418 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1419 struct blk_mq_ctx *ctx,
1422 spin_lock(&ctx->lock);
1423 __blk_mq_insert_request(hctx, rq, false);
1424 spin_unlock(&ctx->lock);
1427 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1430 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1432 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1435 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1437 blk_qc_t *cookie, bool may_sleep)
1439 struct request_queue *q = rq->q;
1440 struct blk_mq_queue_data bd = {
1444 blk_qc_t new_cookie;
1446 bool run_queue = true;
1448 if (blk_mq_hctx_stopped(hctx)) {
1456 if (!blk_mq_get_driver_tag(rq, NULL, false))
1459 new_cookie = request_to_qc_t(hctx, rq);
1462 * For OK queue, we are done. For error, kill it. Any other
1463 * error (busy), just add it to our list as we previously
1466 ret = q->mq_ops->queue_rq(hctx, &bd);
1469 *cookie = new_cookie;
1471 case BLK_STS_RESOURCE:
1472 __blk_mq_requeue_request(rq);
1475 *cookie = BLK_QC_T_NONE;
1476 blk_mq_end_request(rq, ret);
1481 blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1484 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1485 struct request *rq, blk_qc_t *cookie)
1487 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1489 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1492 unsigned int srcu_idx;
1496 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1497 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1498 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1502 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1504 const int is_sync = op_is_sync(bio->bi_opf);
1505 const int is_flush_fua = op_is_flush(bio->bi_opf);
1506 struct blk_mq_alloc_data data = { .flags = 0 };
1508 unsigned int request_count = 0;
1509 struct blk_plug *plug;
1510 struct request *same_queue_rq = NULL;
1512 unsigned int wb_acct;
1514 blk_queue_bounce(q, &bio);
1516 blk_queue_split(q, &bio, q->bio_split);
1518 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1520 return BLK_QC_T_NONE;
1523 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1524 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1525 return BLK_QC_T_NONE;
1527 if (blk_mq_sched_bio_merge(q, bio))
1528 return BLK_QC_T_NONE;
1530 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1532 trace_block_getrq(q, bio, bio->bi_opf);
1534 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1535 if (unlikely(!rq)) {
1536 __wbt_done(q->rq_wb, wb_acct);
1537 return BLK_QC_T_NONE;
1540 wbt_track(&rq->issue_stat, wb_acct);
1542 cookie = request_to_qc_t(data.hctx, rq);
1544 plug = current->plug;
1545 if (unlikely(is_flush_fua)) {
1546 blk_mq_put_ctx(data.ctx);
1547 blk_mq_bio_to_request(rq, bio);
1549 blk_mq_sched_insert_request(rq, false, true, true,
1552 blk_insert_flush(rq);
1553 blk_mq_run_hw_queue(data.hctx, true);
1555 } else if (plug && q->nr_hw_queues == 1) {
1556 struct request *last = NULL;
1558 blk_mq_put_ctx(data.ctx);
1559 blk_mq_bio_to_request(rq, bio);
1562 * @request_count may become stale because of schedule
1563 * out, so check the list again.
1565 if (list_empty(&plug->mq_list))
1567 else if (blk_queue_nomerges(q))
1568 request_count = blk_plug_queued_count(q);
1571 trace_block_plug(q);
1573 last = list_entry_rq(plug->mq_list.prev);
1575 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1576 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1577 blk_flush_plug_list(plug, false);
1578 trace_block_plug(q);
1581 list_add_tail(&rq->queuelist, &plug->mq_list);
1582 } else if (plug && !blk_queue_nomerges(q)) {
1583 blk_mq_bio_to_request(rq, bio);
1586 * We do limited plugging. If the bio can be merged, do that.
1587 * Otherwise the existing request in the plug list will be
1588 * issued. So the plug list will have one request at most
1589 * The plug list might get flushed before this. If that happens,
1590 * the plug list is empty, and same_queue_rq is invalid.
1592 if (list_empty(&plug->mq_list))
1593 same_queue_rq = NULL;
1595 list_del_init(&same_queue_rq->queuelist);
1596 list_add_tail(&rq->queuelist, &plug->mq_list);
1598 blk_mq_put_ctx(data.ctx);
1600 if (same_queue_rq) {
1601 data.hctx = blk_mq_map_queue(q,
1602 same_queue_rq->mq_ctx->cpu);
1603 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1606 } else if (q->nr_hw_queues > 1 && is_sync) {
1607 blk_mq_put_ctx(data.ctx);
1608 blk_mq_bio_to_request(rq, bio);
1609 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1610 } else if (q->elevator) {
1611 blk_mq_put_ctx(data.ctx);
1612 blk_mq_bio_to_request(rq, bio);
1613 blk_mq_sched_insert_request(rq, false, true, true, true);
1615 blk_mq_put_ctx(data.ctx);
1616 blk_mq_bio_to_request(rq, bio);
1617 blk_mq_queue_io(data.hctx, data.ctx, rq);
1618 blk_mq_run_hw_queue(data.hctx, true);
1624 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1625 unsigned int hctx_idx)
1629 if (tags->rqs && set->ops->exit_request) {
1632 for (i = 0; i < tags->nr_tags; i++) {
1633 struct request *rq = tags->static_rqs[i];
1637 set->ops->exit_request(set, rq, hctx_idx);
1638 tags->static_rqs[i] = NULL;
1642 while (!list_empty(&tags->page_list)) {
1643 page = list_first_entry(&tags->page_list, struct page, lru);
1644 list_del_init(&page->lru);
1646 * Remove kmemleak object previously allocated in
1647 * blk_mq_init_rq_map().
1649 kmemleak_free(page_address(page));
1650 __free_pages(page, page->private);
1654 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1658 kfree(tags->static_rqs);
1659 tags->static_rqs = NULL;
1661 blk_mq_free_tags(tags);
1664 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1665 unsigned int hctx_idx,
1666 unsigned int nr_tags,
1667 unsigned int reserved_tags)
1669 struct blk_mq_tags *tags;
1672 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1673 if (node == NUMA_NO_NODE)
1674 node = set->numa_node;
1676 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1677 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1681 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1682 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1685 blk_mq_free_tags(tags);
1689 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1690 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1692 if (!tags->static_rqs) {
1694 blk_mq_free_tags(tags);
1701 static size_t order_to_size(unsigned int order)
1703 return (size_t)PAGE_SIZE << order;
1706 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1707 unsigned int hctx_idx, unsigned int depth)
1709 unsigned int i, j, entries_per_page, max_order = 4;
1710 size_t rq_size, left;
1713 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1714 if (node == NUMA_NO_NODE)
1715 node = set->numa_node;
1717 INIT_LIST_HEAD(&tags->page_list);
1720 * rq_size is the size of the request plus driver payload, rounded
1721 * to the cacheline size
1723 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1725 left = rq_size * depth;
1727 for (i = 0; i < depth; ) {
1728 int this_order = max_order;
1733 while (this_order && left < order_to_size(this_order - 1))
1737 page = alloc_pages_node(node,
1738 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1744 if (order_to_size(this_order) < rq_size)
1751 page->private = this_order;
1752 list_add_tail(&page->lru, &tags->page_list);
1754 p = page_address(page);
1756 * Allow kmemleak to scan these pages as they contain pointers
1757 * to additional allocations like via ops->init_request().
1759 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1760 entries_per_page = order_to_size(this_order) / rq_size;
1761 to_do = min(entries_per_page, depth - i);
1762 left -= to_do * rq_size;
1763 for (j = 0; j < to_do; j++) {
1764 struct request *rq = p;
1766 tags->static_rqs[i] = rq;
1767 if (set->ops->init_request) {
1768 if (set->ops->init_request(set, rq, hctx_idx,
1770 tags->static_rqs[i] = NULL;
1782 blk_mq_free_rqs(set, tags, hctx_idx);
1787 * 'cpu' is going away. splice any existing rq_list entries from this
1788 * software queue to the hw queue dispatch list, and ensure that it
1791 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1793 struct blk_mq_hw_ctx *hctx;
1794 struct blk_mq_ctx *ctx;
1797 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1798 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1800 spin_lock(&ctx->lock);
1801 if (!list_empty(&ctx->rq_list)) {
1802 list_splice_init(&ctx->rq_list, &tmp);
1803 blk_mq_hctx_clear_pending(hctx, ctx);
1805 spin_unlock(&ctx->lock);
1807 if (list_empty(&tmp))
1810 spin_lock(&hctx->lock);
1811 list_splice_tail_init(&tmp, &hctx->dispatch);
1812 spin_unlock(&hctx->lock);
1814 blk_mq_run_hw_queue(hctx, true);
1818 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1820 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1824 /* hctx->ctxs will be freed in queue's release handler */
1825 static void blk_mq_exit_hctx(struct request_queue *q,
1826 struct blk_mq_tag_set *set,
1827 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1829 blk_mq_debugfs_unregister_hctx(hctx);
1831 blk_mq_tag_idle(hctx);
1833 if (set->ops->exit_request)
1834 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1836 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1838 if (set->ops->exit_hctx)
1839 set->ops->exit_hctx(hctx, hctx_idx);
1841 if (hctx->flags & BLK_MQ_F_BLOCKING)
1842 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1844 blk_mq_remove_cpuhp(hctx);
1845 blk_free_flush_queue(hctx->fq);
1846 sbitmap_free(&hctx->ctx_map);
1849 static void blk_mq_exit_hw_queues(struct request_queue *q,
1850 struct blk_mq_tag_set *set, int nr_queue)
1852 struct blk_mq_hw_ctx *hctx;
1855 queue_for_each_hw_ctx(q, hctx, i) {
1858 blk_mq_exit_hctx(q, set, hctx, i);
1862 static int blk_mq_init_hctx(struct request_queue *q,
1863 struct blk_mq_tag_set *set,
1864 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1868 node = hctx->numa_node;
1869 if (node == NUMA_NO_NODE)
1870 node = hctx->numa_node = set->numa_node;
1872 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1873 spin_lock_init(&hctx->lock);
1874 INIT_LIST_HEAD(&hctx->dispatch);
1876 hctx->queue_num = hctx_idx;
1877 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1879 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1881 hctx->tags = set->tags[hctx_idx];
1884 * Allocate space for all possible cpus to avoid allocation at
1887 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1890 goto unregister_cpu_notifier;
1892 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1898 if (set->ops->init_hctx &&
1899 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1902 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1905 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1907 goto sched_exit_hctx;
1909 if (set->ops->init_request &&
1910 set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1914 if (hctx->flags & BLK_MQ_F_BLOCKING)
1915 init_srcu_struct(&hctx->queue_rq_srcu);
1917 blk_mq_debugfs_register_hctx(q, hctx);
1924 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1926 if (set->ops->exit_hctx)
1927 set->ops->exit_hctx(hctx, hctx_idx);
1929 sbitmap_free(&hctx->ctx_map);
1932 unregister_cpu_notifier:
1933 blk_mq_remove_cpuhp(hctx);
1937 static void blk_mq_init_cpu_queues(struct request_queue *q,
1938 unsigned int nr_hw_queues)
1942 for_each_possible_cpu(i) {
1943 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1944 struct blk_mq_hw_ctx *hctx;
1947 spin_lock_init(&__ctx->lock);
1948 INIT_LIST_HEAD(&__ctx->rq_list);
1951 /* If the cpu isn't online, the cpu is mapped to first hctx */
1955 hctx = blk_mq_map_queue(q, i);
1958 * Set local node, IFF we have more than one hw queue. If
1959 * not, we remain on the home node of the device
1961 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1962 hctx->numa_node = local_memory_node(cpu_to_node(i));
1966 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1970 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1971 set->queue_depth, set->reserved_tags);
1972 if (!set->tags[hctx_idx])
1975 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
1980 blk_mq_free_rq_map(set->tags[hctx_idx]);
1981 set->tags[hctx_idx] = NULL;
1985 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
1986 unsigned int hctx_idx)
1988 if (set->tags[hctx_idx]) {
1989 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
1990 blk_mq_free_rq_map(set->tags[hctx_idx]);
1991 set->tags[hctx_idx] = NULL;
1995 static void blk_mq_map_swqueue(struct request_queue *q,
1996 const struct cpumask *online_mask)
1998 unsigned int i, hctx_idx;
1999 struct blk_mq_hw_ctx *hctx;
2000 struct blk_mq_ctx *ctx;
2001 struct blk_mq_tag_set *set = q->tag_set;
2004 * Avoid others reading imcomplete hctx->cpumask through sysfs
2006 mutex_lock(&q->sysfs_lock);
2008 queue_for_each_hw_ctx(q, hctx, i) {
2009 cpumask_clear(hctx->cpumask);
2014 * Map software to hardware queues
2016 for_each_possible_cpu(i) {
2017 /* If the cpu isn't online, the cpu is mapped to first hctx */
2018 if (!cpumask_test_cpu(i, online_mask))
2021 hctx_idx = q->mq_map[i];
2022 /* unmapped hw queue can be remapped after CPU topo changed */
2023 if (!set->tags[hctx_idx] &&
2024 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2026 * If tags initialization fail for some hctx,
2027 * that hctx won't be brought online. In this
2028 * case, remap the current ctx to hctx[0] which
2029 * is guaranteed to always have tags allocated
2034 ctx = per_cpu_ptr(q->queue_ctx, i);
2035 hctx = blk_mq_map_queue(q, i);
2037 cpumask_set_cpu(i, hctx->cpumask);
2038 ctx->index_hw = hctx->nr_ctx;
2039 hctx->ctxs[hctx->nr_ctx++] = ctx;
2042 mutex_unlock(&q->sysfs_lock);
2044 queue_for_each_hw_ctx(q, hctx, i) {
2046 * If no software queues are mapped to this hardware queue,
2047 * disable it and free the request entries.
2049 if (!hctx->nr_ctx) {
2050 /* Never unmap queue 0. We need it as a
2051 * fallback in case of a new remap fails
2054 if (i && set->tags[i])
2055 blk_mq_free_map_and_requests(set, i);
2061 hctx->tags = set->tags[i];
2062 WARN_ON(!hctx->tags);
2065 * Set the map size to the number of mapped software queues.
2066 * This is more accurate and more efficient than looping
2067 * over all possibly mapped software queues.
2069 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2072 * Initialize batch roundrobin counts
2074 hctx->next_cpu = cpumask_first(hctx->cpumask);
2075 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2079 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2081 struct blk_mq_hw_ctx *hctx;
2084 queue_for_each_hw_ctx(q, hctx, i) {
2086 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2088 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2092 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2094 struct request_queue *q;
2096 lockdep_assert_held(&set->tag_list_lock);
2098 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2099 blk_mq_freeze_queue(q);
2100 queue_set_hctx_shared(q, shared);
2101 blk_mq_unfreeze_queue(q);
2105 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2107 struct blk_mq_tag_set *set = q->tag_set;
2109 mutex_lock(&set->tag_list_lock);
2110 list_del_rcu(&q->tag_set_list);
2111 INIT_LIST_HEAD(&q->tag_set_list);
2112 if (list_is_singular(&set->tag_list)) {
2113 /* just transitioned to unshared */
2114 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2115 /* update existing queue */
2116 blk_mq_update_tag_set_depth(set, false);
2118 mutex_unlock(&set->tag_list_lock);
2123 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2124 struct request_queue *q)
2128 mutex_lock(&set->tag_list_lock);
2130 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2131 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2132 set->flags |= BLK_MQ_F_TAG_SHARED;
2133 /* update existing queue */
2134 blk_mq_update_tag_set_depth(set, true);
2136 if (set->flags & BLK_MQ_F_TAG_SHARED)
2137 queue_set_hctx_shared(q, true);
2138 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2140 mutex_unlock(&set->tag_list_lock);
2144 * It is the actual release handler for mq, but we do it from
2145 * request queue's release handler for avoiding use-after-free
2146 * and headache because q->mq_kobj shouldn't have been introduced,
2147 * but we can't group ctx/kctx kobj without it.
2149 void blk_mq_release(struct request_queue *q)
2151 struct blk_mq_hw_ctx *hctx;
2154 /* hctx kobj stays in hctx */
2155 queue_for_each_hw_ctx(q, hctx, i) {
2158 kobject_put(&hctx->kobj);
2163 kfree(q->queue_hw_ctx);
2166 * release .mq_kobj and sw queue's kobject now because
2167 * both share lifetime with request queue.
2169 blk_mq_sysfs_deinit(q);
2171 free_percpu(q->queue_ctx);
2174 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2176 struct request_queue *uninit_q, *q;
2178 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2180 return ERR_PTR(-ENOMEM);
2182 q = blk_mq_init_allocated_queue(set, uninit_q);
2184 blk_cleanup_queue(uninit_q);
2188 EXPORT_SYMBOL(blk_mq_init_queue);
2190 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2191 struct request_queue *q)
2194 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2196 blk_mq_sysfs_unregister(q);
2197 for (i = 0; i < set->nr_hw_queues; i++) {
2203 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2204 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2209 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2216 atomic_set(&hctxs[i]->nr_active, 0);
2217 hctxs[i]->numa_node = node;
2218 hctxs[i]->queue_num = i;
2220 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2221 free_cpumask_var(hctxs[i]->cpumask);
2226 blk_mq_hctx_kobj_init(hctxs[i]);
2228 for (j = i; j < q->nr_hw_queues; j++) {
2229 struct blk_mq_hw_ctx *hctx = hctxs[j];
2233 blk_mq_free_map_and_requests(set, j);
2234 blk_mq_exit_hctx(q, set, hctx, j);
2235 kobject_put(&hctx->kobj);
2240 q->nr_hw_queues = i;
2241 blk_mq_sysfs_register(q);
2244 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2245 struct request_queue *q)
2247 /* mark the queue as mq asap */
2248 q->mq_ops = set->ops;
2250 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2251 blk_mq_poll_stats_bkt,
2252 BLK_MQ_POLL_STATS_BKTS, q);
2256 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2260 /* init q->mq_kobj and sw queues' kobjects */
2261 blk_mq_sysfs_init(q);
2263 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2264 GFP_KERNEL, set->numa_node);
2265 if (!q->queue_hw_ctx)
2268 q->mq_map = set->mq_map;
2270 blk_mq_realloc_hw_ctxs(set, q);
2271 if (!q->nr_hw_queues)
2274 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2275 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2277 q->nr_queues = nr_cpu_ids;
2279 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2281 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2282 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2284 q->sg_reserved_size = INT_MAX;
2286 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2287 INIT_LIST_HEAD(&q->requeue_list);
2288 spin_lock_init(&q->requeue_lock);
2290 blk_queue_make_request(q, blk_mq_make_request);
2293 * Do this after blk_queue_make_request() overrides it...
2295 q->nr_requests = set->queue_depth;
2298 * Default to classic polling
2302 if (set->ops->complete)
2303 blk_queue_softirq_done(q, set->ops->complete);
2305 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2308 mutex_lock(&all_q_mutex);
2310 list_add_tail(&q->all_q_node, &all_q_list);
2311 blk_mq_add_queue_tag_set(set, q);
2312 blk_mq_map_swqueue(q, cpu_online_mask);
2314 mutex_unlock(&all_q_mutex);
2317 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2320 ret = blk_mq_sched_init(q);
2322 return ERR_PTR(ret);
2328 kfree(q->queue_hw_ctx);
2330 free_percpu(q->queue_ctx);
2333 return ERR_PTR(-ENOMEM);
2335 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2337 void blk_mq_free_queue(struct request_queue *q)
2339 struct blk_mq_tag_set *set = q->tag_set;
2341 mutex_lock(&all_q_mutex);
2342 list_del_init(&q->all_q_node);
2343 mutex_unlock(&all_q_mutex);
2345 blk_mq_del_queue_tag_set(q);
2347 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2350 /* Basically redo blk_mq_init_queue with queue frozen */
2351 static void blk_mq_queue_reinit(struct request_queue *q,
2352 const struct cpumask *online_mask)
2354 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2356 blk_mq_debugfs_unregister_hctxs(q);
2357 blk_mq_sysfs_unregister(q);
2360 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2361 * we should change hctx numa_node according to new topology (this
2362 * involves free and re-allocate memory, worthy doing?)
2365 blk_mq_map_swqueue(q, online_mask);
2367 blk_mq_sysfs_register(q);
2368 blk_mq_debugfs_register_hctxs(q);
2372 * New online cpumask which is going to be set in this hotplug event.
2373 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2374 * one-by-one and dynamically allocating this could result in a failure.
2376 static struct cpumask cpuhp_online_new;
2378 static void blk_mq_queue_reinit_work(void)
2380 struct request_queue *q;
2382 mutex_lock(&all_q_mutex);
2384 * We need to freeze and reinit all existing queues. Freezing
2385 * involves synchronous wait for an RCU grace period and doing it
2386 * one by one may take a long time. Start freezing all queues in
2387 * one swoop and then wait for the completions so that freezing can
2388 * take place in parallel.
2390 list_for_each_entry(q, &all_q_list, all_q_node)
2391 blk_freeze_queue_start(q);
2392 list_for_each_entry(q, &all_q_list, all_q_node)
2393 blk_mq_freeze_queue_wait(q);
2395 list_for_each_entry(q, &all_q_list, all_q_node)
2396 blk_mq_queue_reinit(q, &cpuhp_online_new);
2398 list_for_each_entry(q, &all_q_list, all_q_node)
2399 blk_mq_unfreeze_queue(q);
2401 mutex_unlock(&all_q_mutex);
2404 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2406 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2407 blk_mq_queue_reinit_work();
2412 * Before hotadded cpu starts handling requests, new mappings must be
2413 * established. Otherwise, these requests in hw queue might never be
2416 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2417 * for CPU0, and ctx1 for CPU1).
2419 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2420 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2422 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2423 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2424 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2427 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2429 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2430 cpumask_set_cpu(cpu, &cpuhp_online_new);
2431 blk_mq_queue_reinit_work();
2435 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2439 for (i = 0; i < set->nr_hw_queues; i++)
2440 if (!__blk_mq_alloc_rq_map(set, i))
2447 blk_mq_free_rq_map(set->tags[i]);
2453 * Allocate the request maps associated with this tag_set. Note that this
2454 * may reduce the depth asked for, if memory is tight. set->queue_depth
2455 * will be updated to reflect the allocated depth.
2457 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2462 depth = set->queue_depth;
2464 err = __blk_mq_alloc_rq_maps(set);
2468 set->queue_depth >>= 1;
2469 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2473 } while (set->queue_depth);
2475 if (!set->queue_depth || err) {
2476 pr_err("blk-mq: failed to allocate request map\n");
2480 if (depth != set->queue_depth)
2481 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2482 depth, set->queue_depth);
2487 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2489 if (set->ops->map_queues)
2490 return set->ops->map_queues(set);
2492 return blk_mq_map_queues(set);
2496 * Alloc a tag set to be associated with one or more request queues.
2497 * May fail with EINVAL for various error conditions. May adjust the
2498 * requested depth down, if if it too large. In that case, the set
2499 * value will be stored in set->queue_depth.
2501 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2505 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2507 if (!set->nr_hw_queues)
2509 if (!set->queue_depth)
2511 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2514 if (!set->ops->queue_rq)
2517 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2518 pr_info("blk-mq: reduced tag depth to %u\n",
2520 set->queue_depth = BLK_MQ_MAX_DEPTH;
2524 * If a crashdump is active, then we are potentially in a very
2525 * memory constrained environment. Limit us to 1 queue and
2526 * 64 tags to prevent using too much memory.
2528 if (is_kdump_kernel()) {
2529 set->nr_hw_queues = 1;
2530 set->queue_depth = min(64U, set->queue_depth);
2533 * There is no use for more h/w queues than cpus.
2535 if (set->nr_hw_queues > nr_cpu_ids)
2536 set->nr_hw_queues = nr_cpu_ids;
2538 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2539 GFP_KERNEL, set->numa_node);
2544 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2545 GFP_KERNEL, set->numa_node);
2549 ret = blk_mq_update_queue_map(set);
2551 goto out_free_mq_map;
2553 ret = blk_mq_alloc_rq_maps(set);
2555 goto out_free_mq_map;
2557 mutex_init(&set->tag_list_lock);
2558 INIT_LIST_HEAD(&set->tag_list);
2570 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2572 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2576 for (i = 0; i < nr_cpu_ids; i++)
2577 blk_mq_free_map_and_requests(set, i);
2585 EXPORT_SYMBOL(blk_mq_free_tag_set);
2587 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2589 struct blk_mq_tag_set *set = q->tag_set;
2590 struct blk_mq_hw_ctx *hctx;
2596 blk_mq_freeze_queue(q);
2599 queue_for_each_hw_ctx(q, hctx, i) {
2603 * If we're using an MQ scheduler, just update the scheduler
2604 * queue depth. This is similar to what the old code would do.
2606 if (!hctx->sched_tags) {
2607 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2608 min(nr, set->queue_depth),
2611 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2619 q->nr_requests = nr;
2621 blk_mq_unfreeze_queue(q);
2626 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2629 struct request_queue *q;
2631 lockdep_assert_held(&set->tag_list_lock);
2633 if (nr_hw_queues > nr_cpu_ids)
2634 nr_hw_queues = nr_cpu_ids;
2635 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2638 list_for_each_entry(q, &set->tag_list, tag_set_list)
2639 blk_mq_freeze_queue(q);
2641 set->nr_hw_queues = nr_hw_queues;
2642 blk_mq_update_queue_map(set);
2643 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2644 blk_mq_realloc_hw_ctxs(set, q);
2645 blk_mq_queue_reinit(q, cpu_online_mask);
2648 list_for_each_entry(q, &set->tag_list, tag_set_list)
2649 blk_mq_unfreeze_queue(q);
2652 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2654 mutex_lock(&set->tag_list_lock);
2655 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2656 mutex_unlock(&set->tag_list_lock);
2658 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2660 /* Enable polling stats and return whether they were already enabled. */
2661 static bool blk_poll_stats_enable(struct request_queue *q)
2663 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2664 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2666 blk_stat_add_callback(q, q->poll_cb);
2670 static void blk_mq_poll_stats_start(struct request_queue *q)
2673 * We don't arm the callback if polling stats are not enabled or the
2674 * callback is already active.
2676 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2677 blk_stat_is_active(q->poll_cb))
2680 blk_stat_activate_msecs(q->poll_cb, 100);
2683 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2685 struct request_queue *q = cb->data;
2688 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2689 if (cb->stat[bucket].nr_samples)
2690 q->poll_stat[bucket] = cb->stat[bucket];
2694 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2695 struct blk_mq_hw_ctx *hctx,
2698 unsigned long ret = 0;
2702 * If stats collection isn't on, don't sleep but turn it on for
2705 if (!blk_poll_stats_enable(q))
2709 * As an optimistic guess, use half of the mean service time
2710 * for this type of request. We can (and should) make this smarter.
2711 * For instance, if the completion latencies are tight, we can
2712 * get closer than just half the mean. This is especially
2713 * important on devices where the completion latencies are longer
2714 * than ~10 usec. We do use the stats for the relevant IO size
2715 * if available which does lead to better estimates.
2717 bucket = blk_mq_poll_stats_bkt(rq);
2721 if (q->poll_stat[bucket].nr_samples)
2722 ret = (q->poll_stat[bucket].mean + 1) / 2;
2727 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2728 struct blk_mq_hw_ctx *hctx,
2731 struct hrtimer_sleeper hs;
2732 enum hrtimer_mode mode;
2736 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2742 * -1: don't ever hybrid sleep
2743 * 0: use half of prev avg
2744 * >0: use this specific value
2746 if (q->poll_nsec == -1)
2748 else if (q->poll_nsec > 0)
2749 nsecs = q->poll_nsec;
2751 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2756 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2759 * This will be replaced with the stats tracking code, using
2760 * 'avg_completion_time / 2' as the pre-sleep target.
2764 mode = HRTIMER_MODE_REL;
2765 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2766 hrtimer_set_expires(&hs.timer, kt);
2768 hrtimer_init_sleeper(&hs, current);
2770 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2772 set_current_state(TASK_UNINTERRUPTIBLE);
2773 hrtimer_start_expires(&hs.timer, mode);
2776 hrtimer_cancel(&hs.timer);
2777 mode = HRTIMER_MODE_ABS;
2778 } while (hs.task && !signal_pending(current));
2780 __set_current_state(TASK_RUNNING);
2781 destroy_hrtimer_on_stack(&hs.timer);
2785 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2787 struct request_queue *q = hctx->queue;
2791 * If we sleep, have the caller restart the poll loop to reset
2792 * the state. Like for the other success return cases, the
2793 * caller is responsible for checking if the IO completed. If
2794 * the IO isn't complete, we'll get called again and will go
2795 * straight to the busy poll loop.
2797 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2800 hctx->poll_considered++;
2802 state = current->state;
2803 while (!need_resched()) {
2806 hctx->poll_invoked++;
2808 ret = q->mq_ops->poll(hctx, rq->tag);
2810 hctx->poll_success++;
2811 set_current_state(TASK_RUNNING);
2815 if (signal_pending_state(state, current))
2816 set_current_state(TASK_RUNNING);
2818 if (current->state == TASK_RUNNING)
2828 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2830 struct blk_mq_hw_ctx *hctx;
2831 struct blk_plug *plug;
2834 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2835 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2838 plug = current->plug;
2840 blk_flush_plug_list(plug, false);
2842 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2843 if (!blk_qc_t_is_internal(cookie))
2844 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2846 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2848 * With scheduling, if the request has completed, we'll
2849 * get a NULL return here, as we clear the sched tag when
2850 * that happens. The request still remains valid, like always,
2851 * so we should be safe with just the NULL check.
2857 return __blk_mq_poll(hctx, rq);
2859 EXPORT_SYMBOL_GPL(blk_mq_poll);
2861 void blk_mq_disable_hotplug(void)
2863 mutex_lock(&all_q_mutex);
2866 void blk_mq_enable_hotplug(void)
2868 mutex_unlock(&all_q_mutex);
2871 static int __init blk_mq_init(void)
2873 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2874 blk_mq_hctx_notify_dead);
2876 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2877 blk_mq_queue_reinit_prepare,
2878 blk_mq_queue_reinit_dead);
2881 subsys_initcall(blk_mq_init);