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/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
35 #include "blk-mq-sched.h"
37 static DEFINE_MUTEX(all_q_mutex);
38 static LIST_HEAD(all_q_list);
41 * Check if any of the ctx's have pending work in this hardware queue
43 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
45 return sbitmap_any_bit_set(&hctx->ctx_map) ||
46 !list_empty_careful(&hctx->dispatch) ||
47 blk_mq_sched_has_work(hctx);
51 * Mark this ctx as having pending work in this hardware queue
53 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
54 struct blk_mq_ctx *ctx)
56 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
57 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
60 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
61 struct blk_mq_ctx *ctx)
63 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
66 void blk_mq_freeze_queue_start(struct request_queue *q)
70 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
71 if (freeze_depth == 1) {
72 percpu_ref_kill(&q->q_usage_counter);
73 blk_mq_run_hw_queues(q, false);
76 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
78 static void blk_mq_freeze_queue_wait(struct request_queue *q)
80 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
84 * Guarantee no request is in use, so we can change any data structure of
85 * the queue afterward.
87 void blk_freeze_queue(struct request_queue *q)
90 * In the !blk_mq case we are only calling this to kill the
91 * q_usage_counter, otherwise this increases the freeze depth
92 * and waits for it to return to zero. For this reason there is
93 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
94 * exported to drivers as the only user for unfreeze is blk_mq.
96 blk_mq_freeze_queue_start(q);
97 blk_mq_freeze_queue_wait(q);
100 void blk_mq_freeze_queue(struct request_queue *q)
103 * ...just an alias to keep freeze and unfreeze actions balanced
104 * in the blk_mq_* namespace
108 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
110 void blk_mq_unfreeze_queue(struct request_queue *q)
114 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
115 WARN_ON_ONCE(freeze_depth < 0);
117 percpu_ref_reinit(&q->q_usage_counter);
118 wake_up_all(&q->mq_freeze_wq);
121 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
124 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
127 * Note: this function does not prevent that the struct request end_io()
128 * callback function is invoked. Additionally, it is not prevented that
129 * new queue_rq() calls occur unless the queue has been stopped first.
131 void blk_mq_quiesce_queue(struct request_queue *q)
133 struct blk_mq_hw_ctx *hctx;
137 blk_mq_stop_hw_queues(q);
139 queue_for_each_hw_ctx(q, hctx, i) {
140 if (hctx->flags & BLK_MQ_F_BLOCKING)
141 synchronize_srcu(&hctx->queue_rq_srcu);
148 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
150 void blk_mq_wake_waiters(struct request_queue *q)
152 struct blk_mq_hw_ctx *hctx;
155 queue_for_each_hw_ctx(q, hctx, i)
156 if (blk_mq_hw_queue_mapped(hctx))
157 blk_mq_tag_wakeup_all(hctx->tags, true);
160 * If we are called because the queue has now been marked as
161 * dying, we need to ensure that processes currently waiting on
162 * the queue are notified as well.
164 wake_up_all(&q->mq_freeze_wq);
167 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
169 return blk_mq_has_free_tags(hctx->tags);
171 EXPORT_SYMBOL(blk_mq_can_queue);
173 void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
174 struct request *rq, unsigned int op)
176 INIT_LIST_HEAD(&rq->queuelist);
177 /* csd/requeue_work/fifo_time is initialized before use */
181 if (blk_queue_io_stat(q))
182 rq->rq_flags |= RQF_IO_STAT;
183 /* do not touch atomic flags, it needs atomic ops against the timer */
185 INIT_HLIST_NODE(&rq->hash);
186 RB_CLEAR_NODE(&rq->rb_node);
189 rq->start_time = jiffies;
190 #ifdef CONFIG_BLK_CGROUP
192 set_start_time_ns(rq);
193 rq->io_start_time_ns = 0;
195 rq->nr_phys_segments = 0;
196 #if defined(CONFIG_BLK_DEV_INTEGRITY)
197 rq->nr_integrity_segments = 0;
200 /* tag was already set */
210 INIT_LIST_HEAD(&rq->timeout_list);
214 rq->end_io_data = NULL;
217 ctx->rq_dispatched[op_is_sync(op)]++;
219 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init);
221 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
227 tag = blk_mq_get_tag(data);
228 if (tag != BLK_MQ_TAG_FAIL) {
229 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
231 rq = tags->static_rqs[tag];
233 if (blk_mq_tag_busy(data->hctx)) {
234 rq->rq_flags = RQF_MQ_INFLIGHT;
235 atomic_inc(&data->hctx->nr_active);
238 if (data->flags & BLK_MQ_REQ_INTERNAL) {
240 rq->internal_tag = tag;
243 rq->internal_tag = -1;
246 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
252 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
254 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
257 struct blk_mq_alloc_data alloc_data;
261 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
265 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
267 blk_mq_put_ctx(alloc_data.ctx);
271 return ERR_PTR(-EWOULDBLOCK);
274 rq->__sector = (sector_t) -1;
275 rq->bio = rq->biotail = NULL;
278 EXPORT_SYMBOL(blk_mq_alloc_request);
280 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
281 unsigned int flags, unsigned int hctx_idx)
283 struct blk_mq_hw_ctx *hctx;
284 struct blk_mq_ctx *ctx;
286 struct blk_mq_alloc_data alloc_data;
290 * If the tag allocator sleeps we could get an allocation for a
291 * different hardware context. No need to complicate the low level
292 * allocator for this for the rare use case of a command tied to
295 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
296 return ERR_PTR(-EINVAL);
298 if (hctx_idx >= q->nr_hw_queues)
299 return ERR_PTR(-EIO);
301 ret = blk_queue_enter(q, true);
306 * Check if the hardware context is actually mapped to anything.
307 * If not tell the caller that it should skip this queue.
309 hctx = q->queue_hw_ctx[hctx_idx];
310 if (!blk_mq_hw_queue_mapped(hctx)) {
314 ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
316 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
317 rq = __blk_mq_alloc_request(&alloc_data, rw);
329 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
331 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
334 const int sched_tag = rq->internal_tag;
335 struct request_queue *q = rq->q;
337 if (rq->rq_flags & RQF_MQ_INFLIGHT)
338 atomic_dec(&hctx->nr_active);
340 wbt_done(q->rq_wb, &rq->issue_stat);
343 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
344 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
346 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
348 blk_mq_sched_completed_request(hctx, rq);
352 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
355 struct blk_mq_ctx *ctx = rq->mq_ctx;
357 ctx->rq_completed[rq_is_sync(rq)]++;
358 __blk_mq_finish_request(hctx, ctx, rq);
361 void blk_mq_finish_request(struct request *rq)
363 blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
366 void blk_mq_free_request(struct request *rq)
368 blk_mq_sched_put_request(rq);
370 EXPORT_SYMBOL_GPL(blk_mq_free_request);
372 inline void __blk_mq_end_request(struct request *rq, int error)
374 blk_account_io_done(rq);
377 wbt_done(rq->q->rq_wb, &rq->issue_stat);
378 rq->end_io(rq, error);
380 if (unlikely(blk_bidi_rq(rq)))
381 blk_mq_free_request(rq->next_rq);
382 blk_mq_free_request(rq);
385 EXPORT_SYMBOL(__blk_mq_end_request);
387 void blk_mq_end_request(struct request *rq, int error)
389 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
391 __blk_mq_end_request(rq, error);
393 EXPORT_SYMBOL(blk_mq_end_request);
395 static void __blk_mq_complete_request_remote(void *data)
397 struct request *rq = data;
399 rq->q->softirq_done_fn(rq);
402 static void blk_mq_ipi_complete_request(struct request *rq)
404 struct blk_mq_ctx *ctx = rq->mq_ctx;
408 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
409 rq->q->softirq_done_fn(rq);
414 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
415 shared = cpus_share_cache(cpu, ctx->cpu);
417 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
418 rq->csd.func = __blk_mq_complete_request_remote;
421 smp_call_function_single_async(ctx->cpu, &rq->csd);
423 rq->q->softirq_done_fn(rq);
428 static void blk_mq_stat_add(struct request *rq)
430 if (rq->rq_flags & RQF_STATS) {
432 * We could rq->mq_ctx here, but there's less of a risk
433 * of races if we have the completion event add the stats
434 * to the local software queue.
436 struct blk_mq_ctx *ctx;
438 ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id());
439 blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq);
443 static void __blk_mq_complete_request(struct request *rq)
445 struct request_queue *q = rq->q;
449 if (!q->softirq_done_fn)
450 blk_mq_end_request(rq, rq->errors);
452 blk_mq_ipi_complete_request(rq);
456 * blk_mq_complete_request - end I/O on a request
457 * @rq: the request being processed
460 * Ends all I/O on a request. It does not handle partial completions.
461 * The actual completion happens out-of-order, through a IPI handler.
463 void blk_mq_complete_request(struct request *rq, int error)
465 struct request_queue *q = rq->q;
467 if (unlikely(blk_should_fake_timeout(q)))
469 if (!blk_mark_rq_complete(rq)) {
471 __blk_mq_complete_request(rq);
474 EXPORT_SYMBOL(blk_mq_complete_request);
476 int blk_mq_request_started(struct request *rq)
478 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
480 EXPORT_SYMBOL_GPL(blk_mq_request_started);
482 void blk_mq_start_request(struct request *rq)
484 struct request_queue *q = rq->q;
486 blk_mq_sched_started_request(rq);
488 trace_block_rq_issue(q, rq);
490 rq->resid_len = blk_rq_bytes(rq);
491 if (unlikely(blk_bidi_rq(rq)))
492 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
494 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
495 blk_stat_set_issue_time(&rq->issue_stat);
496 rq->rq_flags |= RQF_STATS;
497 wbt_issue(q->rq_wb, &rq->issue_stat);
503 * Ensure that ->deadline is visible before set the started
504 * flag and clear the completed flag.
506 smp_mb__before_atomic();
509 * Mark us as started and clear complete. Complete might have been
510 * set if requeue raced with timeout, which then marked it as
511 * complete. So be sure to clear complete again when we start
512 * the request, otherwise we'll ignore the completion event.
514 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
515 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
516 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
517 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
519 if (q->dma_drain_size && blk_rq_bytes(rq)) {
521 * Make sure space for the drain appears. We know we can do
522 * this because max_hw_segments has been adjusted to be one
523 * fewer than the device can handle.
525 rq->nr_phys_segments++;
528 EXPORT_SYMBOL(blk_mq_start_request);
530 static void __blk_mq_requeue_request(struct request *rq)
532 struct request_queue *q = rq->q;
534 trace_block_rq_requeue(q, rq);
535 wbt_requeue(q->rq_wb, &rq->issue_stat);
536 blk_mq_sched_requeue_request(rq);
538 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
539 if (q->dma_drain_size && blk_rq_bytes(rq))
540 rq->nr_phys_segments--;
544 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
546 __blk_mq_requeue_request(rq);
548 BUG_ON(blk_queued_rq(rq));
549 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
551 EXPORT_SYMBOL(blk_mq_requeue_request);
553 static void blk_mq_requeue_work(struct work_struct *work)
555 struct request_queue *q =
556 container_of(work, struct request_queue, requeue_work.work);
558 struct request *rq, *next;
561 spin_lock_irqsave(&q->requeue_lock, flags);
562 list_splice_init(&q->requeue_list, &rq_list);
563 spin_unlock_irqrestore(&q->requeue_lock, flags);
565 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
566 if (!(rq->rq_flags & RQF_SOFTBARRIER))
569 rq->rq_flags &= ~RQF_SOFTBARRIER;
570 list_del_init(&rq->queuelist);
571 blk_mq_sched_insert_request(rq, true, false, false);
574 while (!list_empty(&rq_list)) {
575 rq = list_entry(rq_list.next, struct request, queuelist);
576 list_del_init(&rq->queuelist);
577 blk_mq_sched_insert_request(rq, false, false, false);
580 blk_mq_run_hw_queues(q, false);
583 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
584 bool kick_requeue_list)
586 struct request_queue *q = rq->q;
590 * We abuse this flag that is otherwise used by the I/O scheduler to
591 * request head insertation from the workqueue.
593 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
595 spin_lock_irqsave(&q->requeue_lock, flags);
597 rq->rq_flags |= RQF_SOFTBARRIER;
598 list_add(&rq->queuelist, &q->requeue_list);
600 list_add_tail(&rq->queuelist, &q->requeue_list);
602 spin_unlock_irqrestore(&q->requeue_lock, flags);
604 if (kick_requeue_list)
605 blk_mq_kick_requeue_list(q);
607 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
609 void blk_mq_kick_requeue_list(struct request_queue *q)
611 kblockd_schedule_delayed_work(&q->requeue_work, 0);
613 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
615 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
618 kblockd_schedule_delayed_work(&q->requeue_work,
619 msecs_to_jiffies(msecs));
621 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
623 void blk_mq_abort_requeue_list(struct request_queue *q)
628 spin_lock_irqsave(&q->requeue_lock, flags);
629 list_splice_init(&q->requeue_list, &rq_list);
630 spin_unlock_irqrestore(&q->requeue_lock, flags);
632 while (!list_empty(&rq_list)) {
635 rq = list_first_entry(&rq_list, struct request, queuelist);
636 list_del_init(&rq->queuelist);
638 blk_mq_end_request(rq, rq->errors);
641 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
643 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
645 if (tag < tags->nr_tags) {
646 prefetch(tags->rqs[tag]);
647 return tags->rqs[tag];
652 EXPORT_SYMBOL(blk_mq_tag_to_rq);
654 struct blk_mq_timeout_data {
656 unsigned int next_set;
659 void blk_mq_rq_timed_out(struct request *req, bool reserved)
661 const struct blk_mq_ops *ops = req->q->mq_ops;
662 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
665 * We know that complete is set at this point. If STARTED isn't set
666 * anymore, then the request isn't active and the "timeout" should
667 * just be ignored. This can happen due to the bitflag ordering.
668 * Timeout first checks if STARTED is set, and if it is, assumes
669 * the request is active. But if we race with completion, then
670 * we both flags will get cleared. So check here again, and ignore
671 * a timeout event with a request that isn't active.
673 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
677 ret = ops->timeout(req, reserved);
681 __blk_mq_complete_request(req);
683 case BLK_EH_RESET_TIMER:
685 blk_clear_rq_complete(req);
687 case BLK_EH_NOT_HANDLED:
690 printk(KERN_ERR "block: bad eh return: %d\n", ret);
695 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
696 struct request *rq, void *priv, bool reserved)
698 struct blk_mq_timeout_data *data = priv;
700 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
702 * If a request wasn't started before the queue was
703 * marked dying, kill it here or it'll go unnoticed.
705 if (unlikely(blk_queue_dying(rq->q))) {
707 blk_mq_end_request(rq, rq->errors);
712 if (time_after_eq(jiffies, rq->deadline)) {
713 if (!blk_mark_rq_complete(rq))
714 blk_mq_rq_timed_out(rq, reserved);
715 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
716 data->next = rq->deadline;
721 static void blk_mq_timeout_work(struct work_struct *work)
723 struct request_queue *q =
724 container_of(work, struct request_queue, timeout_work);
725 struct blk_mq_timeout_data data = {
731 /* A deadlock might occur if a request is stuck requiring a
732 * timeout at the same time a queue freeze is waiting
733 * completion, since the timeout code would not be able to
734 * acquire the queue reference here.
736 * That's why we don't use blk_queue_enter here; instead, we use
737 * percpu_ref_tryget directly, because we need to be able to
738 * obtain a reference even in the short window between the queue
739 * starting to freeze, by dropping the first reference in
740 * blk_mq_freeze_queue_start, and the moment the last request is
741 * consumed, marked by the instant q_usage_counter reaches
744 if (!percpu_ref_tryget(&q->q_usage_counter))
747 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
750 data.next = blk_rq_timeout(round_jiffies_up(data.next));
751 mod_timer(&q->timeout, data.next);
753 struct blk_mq_hw_ctx *hctx;
755 queue_for_each_hw_ctx(q, hctx, i) {
756 /* the hctx may be unmapped, so check it here */
757 if (blk_mq_hw_queue_mapped(hctx))
758 blk_mq_tag_idle(hctx);
765 * Reverse check our software queue for entries that we could potentially
766 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
767 * too much time checking for merges.
769 static bool blk_mq_attempt_merge(struct request_queue *q,
770 struct blk_mq_ctx *ctx, struct bio *bio)
775 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
781 if (!blk_rq_merge_ok(rq, bio))
784 el_ret = blk_try_merge(rq, bio);
785 if (el_ret == ELEVATOR_NO_MERGE)
788 if (!blk_mq_sched_allow_merge(q, rq, bio))
791 if (el_ret == ELEVATOR_BACK_MERGE) {
792 if (bio_attempt_back_merge(q, rq, bio)) {
797 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
798 if (bio_attempt_front_merge(q, rq, bio)) {
809 struct flush_busy_ctx_data {
810 struct blk_mq_hw_ctx *hctx;
811 struct list_head *list;
814 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
816 struct flush_busy_ctx_data *flush_data = data;
817 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
818 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
820 sbitmap_clear_bit(sb, bitnr);
821 spin_lock(&ctx->lock);
822 list_splice_tail_init(&ctx->rq_list, flush_data->list);
823 spin_unlock(&ctx->lock);
828 * Process software queues that have been marked busy, splicing them
829 * to the for-dispatch
831 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
833 struct flush_busy_ctx_data data = {
838 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
840 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
842 static inline unsigned int queued_to_index(unsigned int queued)
847 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
850 static bool blk_mq_get_driver_tag(struct request *rq,
851 struct blk_mq_hw_ctx **hctx, bool wait)
853 struct blk_mq_alloc_data data = {
856 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
857 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
860 if (blk_mq_hctx_stopped(data.hctx))
870 rq->tag = blk_mq_get_tag(&data);
872 data.hctx->tags->rqs[rq->tag] = rq;
880 * If we fail getting a driver tag because all the driver tags are already
881 * assigned and on the dispatch list, BUT the first entry does not have a
882 * tag, then we could deadlock. For that case, move entries with assigned
883 * driver tags to the front, leaving the set of tagged requests in the
884 * same order, and the untagged set in the same order.
886 static bool reorder_tags_to_front(struct list_head *list)
888 struct request *rq, *tmp, *first = NULL;
890 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
894 list_move(&rq->queuelist, list);
900 return first != NULL;
903 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list)
905 struct request_queue *q = hctx->queue;
907 LIST_HEAD(driver_list);
908 struct list_head *dptr;
909 int queued, ret = BLK_MQ_RQ_QUEUE_OK;
912 * Start off with dptr being NULL, so we start the first request
913 * immediately, even if we have more pending.
918 * Now process all the entries, sending them to the driver.
921 while (!list_empty(list)) {
922 struct blk_mq_queue_data bd;
924 rq = list_first_entry(list, struct request, queuelist);
925 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
926 if (!queued && reorder_tags_to_front(list))
928 blk_mq_sched_mark_restart(hctx);
931 list_del_init(&rq->queuelist);
935 bd.last = list_empty(list);
937 ret = q->mq_ops->queue_rq(hctx, &bd);
939 case BLK_MQ_RQ_QUEUE_OK:
942 case BLK_MQ_RQ_QUEUE_BUSY:
943 list_add(&rq->queuelist, list);
944 __blk_mq_requeue_request(rq);
947 pr_err("blk-mq: bad return on queue: %d\n", ret);
948 case BLK_MQ_RQ_QUEUE_ERROR:
950 blk_mq_end_request(rq, rq->errors);
954 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
958 * We've done the first request. If we have more than 1
959 * left in the list, set dptr to defer issue.
961 if (!dptr && list->next != list->prev)
965 hctx->dispatched[queued_to_index(queued)]++;
968 * Any items that need requeuing? Stuff them into hctx->dispatch,
969 * that is where we will continue on next queue run.
971 if (!list_empty(list)) {
972 spin_lock(&hctx->lock);
973 list_splice(list, &hctx->dispatch);
974 spin_unlock(&hctx->lock);
977 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
978 * it's possible the queue is stopped and restarted again
979 * before this. Queue restart will dispatch requests. And since
980 * requests in rq_list aren't added into hctx->dispatch yet,
981 * the requests in rq_list might get lost.
983 * blk_mq_run_hw_queue() already checks the STOPPED bit
985 * If RESTART is set, then let completion restart the queue
986 * instead of potentially looping here.
988 if (!blk_mq_sched_needs_restart(hctx))
989 blk_mq_run_hw_queue(hctx, true);
992 return ret != BLK_MQ_RQ_QUEUE_BUSY;
995 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
999 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1000 cpu_online(hctx->next_cpu));
1002 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1004 blk_mq_sched_dispatch_requests(hctx);
1007 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1008 blk_mq_sched_dispatch_requests(hctx);
1009 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1014 * It'd be great if the workqueue API had a way to pass
1015 * in a mask and had some smarts for more clever placement.
1016 * For now we just round-robin here, switching for every
1017 * BLK_MQ_CPU_WORK_BATCH queued items.
1019 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1021 if (hctx->queue->nr_hw_queues == 1)
1022 return WORK_CPU_UNBOUND;
1024 if (--hctx->next_cpu_batch <= 0) {
1027 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1028 if (next_cpu >= nr_cpu_ids)
1029 next_cpu = cpumask_first(hctx->cpumask);
1031 hctx->next_cpu = next_cpu;
1032 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1035 return hctx->next_cpu;
1038 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1040 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1041 !blk_mq_hw_queue_mapped(hctx)))
1044 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1045 int cpu = get_cpu();
1046 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1047 __blk_mq_run_hw_queue(hctx);
1055 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
1058 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1060 struct blk_mq_hw_ctx *hctx;
1063 queue_for_each_hw_ctx(q, hctx, i) {
1064 if (!blk_mq_hctx_has_pending(hctx) ||
1065 blk_mq_hctx_stopped(hctx))
1068 blk_mq_run_hw_queue(hctx, async);
1071 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1074 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1075 * @q: request queue.
1077 * The caller is responsible for serializing this function against
1078 * blk_mq_{start,stop}_hw_queue().
1080 bool blk_mq_queue_stopped(struct request_queue *q)
1082 struct blk_mq_hw_ctx *hctx;
1085 queue_for_each_hw_ctx(q, hctx, i)
1086 if (blk_mq_hctx_stopped(hctx))
1091 EXPORT_SYMBOL(blk_mq_queue_stopped);
1093 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1095 cancel_work(&hctx->run_work);
1096 cancel_delayed_work(&hctx->delay_work);
1097 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1099 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1101 void blk_mq_stop_hw_queues(struct request_queue *q)
1103 struct blk_mq_hw_ctx *hctx;
1106 queue_for_each_hw_ctx(q, hctx, i)
1107 blk_mq_stop_hw_queue(hctx);
1109 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1111 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1113 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1115 blk_mq_run_hw_queue(hctx, false);
1117 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1119 void blk_mq_start_hw_queues(struct request_queue *q)
1121 struct blk_mq_hw_ctx *hctx;
1124 queue_for_each_hw_ctx(q, hctx, i)
1125 blk_mq_start_hw_queue(hctx);
1127 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1129 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1131 if (!blk_mq_hctx_stopped(hctx))
1134 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1135 blk_mq_run_hw_queue(hctx, async);
1137 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1139 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1141 struct blk_mq_hw_ctx *hctx;
1144 queue_for_each_hw_ctx(q, hctx, i)
1145 blk_mq_start_stopped_hw_queue(hctx, async);
1147 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1149 static void blk_mq_run_work_fn(struct work_struct *work)
1151 struct blk_mq_hw_ctx *hctx;
1153 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1155 __blk_mq_run_hw_queue(hctx);
1158 static void blk_mq_delay_work_fn(struct work_struct *work)
1160 struct blk_mq_hw_ctx *hctx;
1162 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1164 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1165 __blk_mq_run_hw_queue(hctx);
1168 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1170 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1173 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1174 &hctx->delay_work, msecs_to_jiffies(msecs));
1176 EXPORT_SYMBOL(blk_mq_delay_queue);
1178 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1182 struct blk_mq_ctx *ctx = rq->mq_ctx;
1184 trace_block_rq_insert(hctx->queue, rq);
1187 list_add(&rq->queuelist, &ctx->rq_list);
1189 list_add_tail(&rq->queuelist, &ctx->rq_list);
1192 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1195 struct blk_mq_ctx *ctx = rq->mq_ctx;
1197 __blk_mq_insert_req_list(hctx, rq, at_head);
1198 blk_mq_hctx_mark_pending(hctx, ctx);
1201 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1202 struct list_head *list)
1206 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1209 spin_lock(&ctx->lock);
1210 while (!list_empty(list)) {
1213 rq = list_first_entry(list, struct request, queuelist);
1214 BUG_ON(rq->mq_ctx != ctx);
1215 list_del_init(&rq->queuelist);
1216 __blk_mq_insert_req_list(hctx, rq, false);
1218 blk_mq_hctx_mark_pending(hctx, ctx);
1219 spin_unlock(&ctx->lock);
1222 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1224 struct request *rqa = container_of(a, struct request, queuelist);
1225 struct request *rqb = container_of(b, struct request, queuelist);
1227 return !(rqa->mq_ctx < rqb->mq_ctx ||
1228 (rqa->mq_ctx == rqb->mq_ctx &&
1229 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1232 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1234 struct blk_mq_ctx *this_ctx;
1235 struct request_queue *this_q;
1238 LIST_HEAD(ctx_list);
1241 list_splice_init(&plug->mq_list, &list);
1243 list_sort(NULL, &list, plug_ctx_cmp);
1249 while (!list_empty(&list)) {
1250 rq = list_entry_rq(list.next);
1251 list_del_init(&rq->queuelist);
1253 if (rq->mq_ctx != this_ctx) {
1255 trace_block_unplug(this_q, depth, from_schedule);
1256 blk_mq_sched_insert_requests(this_q, this_ctx,
1261 this_ctx = rq->mq_ctx;
1267 list_add_tail(&rq->queuelist, &ctx_list);
1271 * If 'this_ctx' is set, we know we have entries to complete
1272 * on 'ctx_list'. Do those.
1275 trace_block_unplug(this_q, depth, from_schedule);
1276 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1281 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1283 init_request_from_bio(rq, bio);
1285 blk_account_io_start(rq, true);
1288 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1290 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1291 !blk_queue_nomerges(hctx->queue);
1294 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1295 struct blk_mq_ctx *ctx,
1296 struct request *rq, struct bio *bio)
1298 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1299 blk_mq_bio_to_request(rq, bio);
1300 spin_lock(&ctx->lock);
1302 __blk_mq_insert_request(hctx, rq, false);
1303 spin_unlock(&ctx->lock);
1306 struct request_queue *q = hctx->queue;
1308 spin_lock(&ctx->lock);
1309 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1310 blk_mq_bio_to_request(rq, bio);
1314 spin_unlock(&ctx->lock);
1315 __blk_mq_finish_request(hctx, ctx, rq);
1320 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1323 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1325 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1328 static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie)
1330 struct request_queue *q = rq->q;
1331 struct blk_mq_queue_data bd = {
1336 struct blk_mq_hw_ctx *hctx;
1337 blk_qc_t new_cookie;
1343 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1346 new_cookie = request_to_qc_t(hctx, rq);
1349 * For OK queue, we are done. For error, kill it. Any other
1350 * error (busy), just add it to our list as we previously
1353 ret = q->mq_ops->queue_rq(hctx, &bd);
1354 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1355 *cookie = new_cookie;
1359 __blk_mq_requeue_request(rq);
1361 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1362 *cookie = BLK_QC_T_NONE;
1364 blk_mq_end_request(rq, rq->errors);
1369 blk_mq_sched_insert_request(rq, false, true, true);
1373 * Multiple hardware queue variant. This will not use per-process plugs,
1374 * but will attempt to bypass the hctx queueing if we can go straight to
1375 * hardware for SYNC IO.
1377 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1379 const int is_sync = op_is_sync(bio->bi_opf);
1380 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1381 struct blk_mq_alloc_data data;
1383 unsigned int request_count = 0, srcu_idx;
1384 struct blk_plug *plug;
1385 struct request *same_queue_rq = NULL;
1387 unsigned int wb_acct;
1389 blk_queue_bounce(q, &bio);
1391 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1393 return BLK_QC_T_NONE;
1396 blk_queue_split(q, &bio, q->bio_split);
1398 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1399 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1400 return BLK_QC_T_NONE;
1402 if (blk_mq_sched_bio_merge(q, bio))
1403 return BLK_QC_T_NONE;
1405 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1407 trace_block_getrq(q, bio, bio->bi_opf);
1409 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1410 if (unlikely(!rq)) {
1411 __wbt_done(q->rq_wb, wb_acct);
1412 return BLK_QC_T_NONE;
1415 wbt_track(&rq->issue_stat, wb_acct);
1417 cookie = request_to_qc_t(data.hctx, rq);
1419 if (unlikely(is_flush_fua)) {
1420 blk_mq_bio_to_request(rq, bio);
1421 blk_mq_get_driver_tag(rq, NULL, true);
1422 blk_insert_flush(rq);
1426 plug = current->plug;
1428 * If the driver supports defer issued based on 'last', then
1429 * queue it up like normal since we can potentially save some
1432 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1433 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1434 struct request *old_rq = NULL;
1436 blk_mq_bio_to_request(rq, bio);
1439 * We do limited plugging. If the bio can be merged, do that.
1440 * Otherwise the existing request in the plug list will be
1441 * issued. So the plug list will have one request at most
1445 * The plug list might get flushed before this. If that
1446 * happens, same_queue_rq is invalid and plug list is
1449 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1450 old_rq = same_queue_rq;
1451 list_del_init(&old_rq->queuelist);
1453 list_add_tail(&rq->queuelist, &plug->mq_list);
1454 } else /* is_sync */
1456 blk_mq_put_ctx(data.ctx);
1460 if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) {
1462 blk_mq_try_issue_directly(old_rq, &cookie);
1465 srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu);
1466 blk_mq_try_issue_directly(old_rq, &cookie);
1467 srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx);
1473 blk_mq_put_ctx(data.ctx);
1474 blk_mq_bio_to_request(rq, bio);
1475 blk_mq_sched_insert_request(rq, false, true, true);
1478 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1480 * For a SYNC request, send it to the hardware immediately. For
1481 * an ASYNC request, just ensure that we run it later on. The
1482 * latter allows for merging opportunities and more efficient
1486 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1488 blk_mq_put_ctx(data.ctx);
1494 * Single hardware queue variant. This will attempt to use any per-process
1495 * plug for merging and IO deferral.
1497 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1499 const int is_sync = op_is_sync(bio->bi_opf);
1500 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1501 struct blk_plug *plug;
1502 unsigned int request_count = 0;
1503 struct blk_mq_alloc_data data;
1506 unsigned int wb_acct;
1508 blk_queue_bounce(q, &bio);
1510 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1512 return BLK_QC_T_NONE;
1515 blk_queue_split(q, &bio, q->bio_split);
1517 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1518 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1519 return BLK_QC_T_NONE;
1521 request_count = blk_plug_queued_count(q);
1523 if (blk_mq_sched_bio_merge(q, bio))
1524 return BLK_QC_T_NONE;
1526 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1528 trace_block_getrq(q, bio, bio->bi_opf);
1530 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1531 if (unlikely(!rq)) {
1532 __wbt_done(q->rq_wb, wb_acct);
1533 return BLK_QC_T_NONE;
1536 wbt_track(&rq->issue_stat, wb_acct);
1538 cookie = request_to_qc_t(data.hctx, rq);
1540 if (unlikely(is_flush_fua)) {
1541 blk_mq_bio_to_request(rq, bio);
1542 blk_mq_get_driver_tag(rq, NULL, true);
1543 blk_insert_flush(rq);
1548 * A task plug currently exists. Since this is completely lockless,
1549 * utilize that to temporarily store requests until the task is
1550 * either done or scheduled away.
1552 plug = current->plug;
1554 struct request *last = NULL;
1556 blk_mq_bio_to_request(rq, bio);
1559 * @request_count may become stale because of schedule
1560 * out, so check the list again.
1562 if (list_empty(&plug->mq_list))
1565 trace_block_plug(q);
1567 last = list_entry_rq(plug->mq_list.prev);
1569 blk_mq_put_ctx(data.ctx);
1571 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1572 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1573 blk_flush_plug_list(plug, false);
1574 trace_block_plug(q);
1577 list_add_tail(&rq->queuelist, &plug->mq_list);
1582 blk_mq_put_ctx(data.ctx);
1583 blk_mq_bio_to_request(rq, bio);
1584 blk_mq_sched_insert_request(rq, false, true, true);
1587 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1589 * For a SYNC request, send it to the hardware immediately. For
1590 * an ASYNC request, just ensure that we run it later on. The
1591 * latter allows for merging opportunities and more efficient
1595 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1598 blk_mq_put_ctx(data.ctx);
1603 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1604 unsigned int hctx_idx)
1608 if (tags->rqs && set->ops->exit_request) {
1611 for (i = 0; i < tags->nr_tags; i++) {
1612 struct request *rq = tags->static_rqs[i];
1616 set->ops->exit_request(set->driver_data, rq,
1618 tags->static_rqs[i] = NULL;
1622 while (!list_empty(&tags->page_list)) {
1623 page = list_first_entry(&tags->page_list, struct page, lru);
1624 list_del_init(&page->lru);
1626 * Remove kmemleak object previously allocated in
1627 * blk_mq_init_rq_map().
1629 kmemleak_free(page_address(page));
1630 __free_pages(page, page->private);
1634 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1638 kfree(tags->static_rqs);
1639 tags->static_rqs = NULL;
1641 blk_mq_free_tags(tags);
1644 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1645 unsigned int hctx_idx,
1646 unsigned int nr_tags,
1647 unsigned int reserved_tags)
1649 struct blk_mq_tags *tags;
1651 tags = blk_mq_init_tags(nr_tags, reserved_tags,
1653 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1657 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1658 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1661 blk_mq_free_tags(tags);
1665 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1666 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1668 if (!tags->static_rqs) {
1670 blk_mq_free_tags(tags);
1677 static size_t order_to_size(unsigned int order)
1679 return (size_t)PAGE_SIZE << order;
1682 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1683 unsigned int hctx_idx, unsigned int depth)
1685 unsigned int i, j, entries_per_page, max_order = 4;
1686 size_t rq_size, left;
1688 INIT_LIST_HEAD(&tags->page_list);
1691 * rq_size is the size of the request plus driver payload, rounded
1692 * to the cacheline size
1694 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1696 left = rq_size * depth;
1698 for (i = 0; i < depth; ) {
1699 int this_order = max_order;
1704 while (this_order && left < order_to_size(this_order - 1))
1708 page = alloc_pages_node(set->numa_node,
1709 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1715 if (order_to_size(this_order) < rq_size)
1722 page->private = this_order;
1723 list_add_tail(&page->lru, &tags->page_list);
1725 p = page_address(page);
1727 * Allow kmemleak to scan these pages as they contain pointers
1728 * to additional allocations like via ops->init_request().
1730 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1731 entries_per_page = order_to_size(this_order) / rq_size;
1732 to_do = min(entries_per_page, depth - i);
1733 left -= to_do * rq_size;
1734 for (j = 0; j < to_do; j++) {
1735 struct request *rq = p;
1737 tags->static_rqs[i] = rq;
1738 if (set->ops->init_request) {
1739 if (set->ops->init_request(set->driver_data,
1742 tags->static_rqs[i] = NULL;
1754 blk_mq_free_rqs(set, tags, hctx_idx);
1759 * 'cpu' is going away. splice any existing rq_list entries from this
1760 * software queue to the hw queue dispatch list, and ensure that it
1763 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1765 struct blk_mq_hw_ctx *hctx;
1766 struct blk_mq_ctx *ctx;
1769 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1770 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1772 spin_lock(&ctx->lock);
1773 if (!list_empty(&ctx->rq_list)) {
1774 list_splice_init(&ctx->rq_list, &tmp);
1775 blk_mq_hctx_clear_pending(hctx, ctx);
1777 spin_unlock(&ctx->lock);
1779 if (list_empty(&tmp))
1782 spin_lock(&hctx->lock);
1783 list_splice_tail_init(&tmp, &hctx->dispatch);
1784 spin_unlock(&hctx->lock);
1786 blk_mq_run_hw_queue(hctx, true);
1790 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1792 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1796 /* hctx->ctxs will be freed in queue's release handler */
1797 static void blk_mq_exit_hctx(struct request_queue *q,
1798 struct blk_mq_tag_set *set,
1799 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1801 unsigned flush_start_tag = set->queue_depth;
1803 blk_mq_tag_idle(hctx);
1805 if (set->ops->exit_request)
1806 set->ops->exit_request(set->driver_data,
1807 hctx->fq->flush_rq, hctx_idx,
1808 flush_start_tag + hctx_idx);
1810 if (set->ops->exit_hctx)
1811 set->ops->exit_hctx(hctx, hctx_idx);
1813 if (hctx->flags & BLK_MQ_F_BLOCKING)
1814 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1816 blk_mq_remove_cpuhp(hctx);
1817 blk_free_flush_queue(hctx->fq);
1818 sbitmap_free(&hctx->ctx_map);
1821 static void blk_mq_exit_hw_queues(struct request_queue *q,
1822 struct blk_mq_tag_set *set, int nr_queue)
1824 struct blk_mq_hw_ctx *hctx;
1827 queue_for_each_hw_ctx(q, hctx, i) {
1830 blk_mq_exit_hctx(q, set, hctx, i);
1834 static void blk_mq_free_hw_queues(struct request_queue *q,
1835 struct blk_mq_tag_set *set)
1837 struct blk_mq_hw_ctx *hctx;
1840 queue_for_each_hw_ctx(q, hctx, i)
1841 free_cpumask_var(hctx->cpumask);
1844 static int blk_mq_init_hctx(struct request_queue *q,
1845 struct blk_mq_tag_set *set,
1846 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1849 unsigned flush_start_tag = set->queue_depth;
1851 node = hctx->numa_node;
1852 if (node == NUMA_NO_NODE)
1853 node = hctx->numa_node = set->numa_node;
1855 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1856 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1857 spin_lock_init(&hctx->lock);
1858 INIT_LIST_HEAD(&hctx->dispatch);
1860 hctx->queue_num = hctx_idx;
1861 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1863 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1865 hctx->tags = set->tags[hctx_idx];
1868 * Allocate space for all possible cpus to avoid allocation at
1871 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1874 goto unregister_cpu_notifier;
1876 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1882 if (set->ops->init_hctx &&
1883 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1886 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1890 if (set->ops->init_request &&
1891 set->ops->init_request(set->driver_data,
1892 hctx->fq->flush_rq, hctx_idx,
1893 flush_start_tag + hctx_idx, node))
1896 if (hctx->flags & BLK_MQ_F_BLOCKING)
1897 init_srcu_struct(&hctx->queue_rq_srcu);
1904 if (set->ops->exit_hctx)
1905 set->ops->exit_hctx(hctx, hctx_idx);
1907 sbitmap_free(&hctx->ctx_map);
1910 unregister_cpu_notifier:
1911 blk_mq_remove_cpuhp(hctx);
1915 static void blk_mq_init_cpu_queues(struct request_queue *q,
1916 unsigned int nr_hw_queues)
1920 for_each_possible_cpu(i) {
1921 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1922 struct blk_mq_hw_ctx *hctx;
1924 memset(__ctx, 0, sizeof(*__ctx));
1926 spin_lock_init(&__ctx->lock);
1927 INIT_LIST_HEAD(&__ctx->rq_list);
1929 blk_stat_init(&__ctx->stat[BLK_STAT_READ]);
1930 blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]);
1932 /* If the cpu isn't online, the cpu is mapped to first hctx */
1936 hctx = blk_mq_map_queue(q, i);
1939 * Set local node, IFF we have more than one hw queue. If
1940 * not, we remain on the home node of the device
1942 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1943 hctx->numa_node = local_memory_node(cpu_to_node(i));
1947 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1951 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1952 set->queue_depth, set->reserved_tags);
1953 if (!set->tags[hctx_idx])
1956 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
1961 blk_mq_free_rq_map(set->tags[hctx_idx]);
1962 set->tags[hctx_idx] = NULL;
1966 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
1967 unsigned int hctx_idx)
1969 if (set->tags[hctx_idx]) {
1970 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
1971 blk_mq_free_rq_map(set->tags[hctx_idx]);
1972 set->tags[hctx_idx] = NULL;
1976 static void blk_mq_map_swqueue(struct request_queue *q,
1977 const struct cpumask *online_mask)
1979 unsigned int i, hctx_idx;
1980 struct blk_mq_hw_ctx *hctx;
1981 struct blk_mq_ctx *ctx;
1982 struct blk_mq_tag_set *set = q->tag_set;
1985 * Avoid others reading imcomplete hctx->cpumask through sysfs
1987 mutex_lock(&q->sysfs_lock);
1989 queue_for_each_hw_ctx(q, hctx, i) {
1990 cpumask_clear(hctx->cpumask);
1995 * Map software to hardware queues
1997 for_each_possible_cpu(i) {
1998 /* If the cpu isn't online, the cpu is mapped to first hctx */
1999 if (!cpumask_test_cpu(i, online_mask))
2002 hctx_idx = q->mq_map[i];
2003 /* unmapped hw queue can be remapped after CPU topo changed */
2004 if (!set->tags[hctx_idx] &&
2005 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2007 * If tags initialization fail for some hctx,
2008 * that hctx won't be brought online. In this
2009 * case, remap the current ctx to hctx[0] which
2010 * is guaranteed to always have tags allocated
2015 ctx = per_cpu_ptr(q->queue_ctx, i);
2016 hctx = blk_mq_map_queue(q, i);
2018 cpumask_set_cpu(i, hctx->cpumask);
2019 ctx->index_hw = hctx->nr_ctx;
2020 hctx->ctxs[hctx->nr_ctx++] = ctx;
2023 mutex_unlock(&q->sysfs_lock);
2025 queue_for_each_hw_ctx(q, hctx, i) {
2027 * If no software queues are mapped to this hardware queue,
2028 * disable it and free the request entries.
2030 if (!hctx->nr_ctx) {
2031 /* Never unmap queue 0. We need it as a
2032 * fallback in case of a new remap fails
2035 if (i && set->tags[i])
2036 blk_mq_free_map_and_requests(set, i);
2042 hctx->tags = set->tags[i];
2043 WARN_ON(!hctx->tags);
2046 * Set the map size to the number of mapped software queues.
2047 * This is more accurate and more efficient than looping
2048 * over all possibly mapped software queues.
2050 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2053 * Initialize batch roundrobin counts
2055 hctx->next_cpu = cpumask_first(hctx->cpumask);
2056 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2060 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2062 struct blk_mq_hw_ctx *hctx;
2065 queue_for_each_hw_ctx(q, hctx, i) {
2067 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2069 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2073 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2075 struct request_queue *q;
2077 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2078 blk_mq_freeze_queue(q);
2079 queue_set_hctx_shared(q, shared);
2080 blk_mq_unfreeze_queue(q);
2084 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2086 struct blk_mq_tag_set *set = q->tag_set;
2088 mutex_lock(&set->tag_list_lock);
2089 list_del_init(&q->tag_set_list);
2090 if (list_is_singular(&set->tag_list)) {
2091 /* just transitioned to unshared */
2092 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2093 /* update existing queue */
2094 blk_mq_update_tag_set_depth(set, false);
2096 mutex_unlock(&set->tag_list_lock);
2099 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2100 struct request_queue *q)
2104 mutex_lock(&set->tag_list_lock);
2106 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2107 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2108 set->flags |= BLK_MQ_F_TAG_SHARED;
2109 /* update existing queue */
2110 blk_mq_update_tag_set_depth(set, true);
2112 if (set->flags & BLK_MQ_F_TAG_SHARED)
2113 queue_set_hctx_shared(q, true);
2114 list_add_tail(&q->tag_set_list, &set->tag_list);
2116 mutex_unlock(&set->tag_list_lock);
2120 * It is the actual release handler for mq, but we do it from
2121 * request queue's release handler for avoiding use-after-free
2122 * and headache because q->mq_kobj shouldn't have been introduced,
2123 * but we can't group ctx/kctx kobj without it.
2125 void blk_mq_release(struct request_queue *q)
2127 struct blk_mq_hw_ctx *hctx;
2130 blk_mq_sched_teardown(q);
2132 /* hctx kobj stays in hctx */
2133 queue_for_each_hw_ctx(q, hctx, i) {
2142 kfree(q->queue_hw_ctx);
2144 /* ctx kobj stays in queue_ctx */
2145 free_percpu(q->queue_ctx);
2148 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2150 struct request_queue *uninit_q, *q;
2152 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2154 return ERR_PTR(-ENOMEM);
2156 q = blk_mq_init_allocated_queue(set, uninit_q);
2158 blk_cleanup_queue(uninit_q);
2162 EXPORT_SYMBOL(blk_mq_init_queue);
2164 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2165 struct request_queue *q)
2168 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2170 blk_mq_sysfs_unregister(q);
2171 for (i = 0; i < set->nr_hw_queues; i++) {
2177 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2178 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2183 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2190 atomic_set(&hctxs[i]->nr_active, 0);
2191 hctxs[i]->numa_node = node;
2192 hctxs[i]->queue_num = i;
2194 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2195 free_cpumask_var(hctxs[i]->cpumask);
2200 blk_mq_hctx_kobj_init(hctxs[i]);
2202 for (j = i; j < q->nr_hw_queues; j++) {
2203 struct blk_mq_hw_ctx *hctx = hctxs[j];
2207 blk_mq_free_map_and_requests(set, j);
2208 blk_mq_exit_hctx(q, set, hctx, j);
2209 free_cpumask_var(hctx->cpumask);
2210 kobject_put(&hctx->kobj);
2217 q->nr_hw_queues = i;
2218 blk_mq_sysfs_register(q);
2221 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2222 struct request_queue *q)
2224 /* mark the queue as mq asap */
2225 q->mq_ops = set->ops;
2227 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2231 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2232 GFP_KERNEL, set->numa_node);
2233 if (!q->queue_hw_ctx)
2236 q->mq_map = set->mq_map;
2238 blk_mq_realloc_hw_ctxs(set, q);
2239 if (!q->nr_hw_queues)
2242 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2243 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2245 q->nr_queues = nr_cpu_ids;
2247 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2249 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2250 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2252 q->sg_reserved_size = INT_MAX;
2254 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2255 INIT_LIST_HEAD(&q->requeue_list);
2256 spin_lock_init(&q->requeue_lock);
2258 if (q->nr_hw_queues > 1)
2259 blk_queue_make_request(q, blk_mq_make_request);
2261 blk_queue_make_request(q, blk_sq_make_request);
2264 * Do this after blk_queue_make_request() overrides it...
2266 q->nr_requests = set->queue_depth;
2269 * Default to classic polling
2273 if (set->ops->complete)
2274 blk_queue_softirq_done(q, set->ops->complete);
2276 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2279 mutex_lock(&all_q_mutex);
2281 list_add_tail(&q->all_q_node, &all_q_list);
2282 blk_mq_add_queue_tag_set(set, q);
2283 blk_mq_map_swqueue(q, cpu_online_mask);
2285 mutex_unlock(&all_q_mutex);
2291 kfree(q->queue_hw_ctx);
2293 free_percpu(q->queue_ctx);
2296 return ERR_PTR(-ENOMEM);
2298 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2300 void blk_mq_free_queue(struct request_queue *q)
2302 struct blk_mq_tag_set *set = q->tag_set;
2304 mutex_lock(&all_q_mutex);
2305 list_del_init(&q->all_q_node);
2306 mutex_unlock(&all_q_mutex);
2310 blk_mq_del_queue_tag_set(q);
2312 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2313 blk_mq_free_hw_queues(q, set);
2316 /* Basically redo blk_mq_init_queue with queue frozen */
2317 static void blk_mq_queue_reinit(struct request_queue *q,
2318 const struct cpumask *online_mask)
2320 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2322 blk_mq_sysfs_unregister(q);
2325 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2326 * we should change hctx numa_node according to new topology (this
2327 * involves free and re-allocate memory, worthy doing?)
2330 blk_mq_map_swqueue(q, online_mask);
2332 blk_mq_sysfs_register(q);
2336 * New online cpumask which is going to be set in this hotplug event.
2337 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2338 * one-by-one and dynamically allocating this could result in a failure.
2340 static struct cpumask cpuhp_online_new;
2342 static void blk_mq_queue_reinit_work(void)
2344 struct request_queue *q;
2346 mutex_lock(&all_q_mutex);
2348 * We need to freeze and reinit all existing queues. Freezing
2349 * involves synchronous wait for an RCU grace period and doing it
2350 * one by one may take a long time. Start freezing all queues in
2351 * one swoop and then wait for the completions so that freezing can
2352 * take place in parallel.
2354 list_for_each_entry(q, &all_q_list, all_q_node)
2355 blk_mq_freeze_queue_start(q);
2356 list_for_each_entry(q, &all_q_list, all_q_node)
2357 blk_mq_freeze_queue_wait(q);
2359 list_for_each_entry(q, &all_q_list, all_q_node)
2360 blk_mq_queue_reinit(q, &cpuhp_online_new);
2362 list_for_each_entry(q, &all_q_list, all_q_node)
2363 blk_mq_unfreeze_queue(q);
2365 mutex_unlock(&all_q_mutex);
2368 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2370 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2371 blk_mq_queue_reinit_work();
2376 * Before hotadded cpu starts handling requests, new mappings must be
2377 * established. Otherwise, these requests in hw queue might never be
2380 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2381 * for CPU0, and ctx1 for CPU1).
2383 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2384 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2386 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2387 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2388 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2391 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2393 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2394 cpumask_set_cpu(cpu, &cpuhp_online_new);
2395 blk_mq_queue_reinit_work();
2399 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2403 for (i = 0; i < set->nr_hw_queues; i++)
2404 if (!__blk_mq_alloc_rq_map(set, i))
2411 blk_mq_free_rq_map(set->tags[i]);
2417 * Allocate the request maps associated with this tag_set. Note that this
2418 * may reduce the depth asked for, if memory is tight. set->queue_depth
2419 * will be updated to reflect the allocated depth.
2421 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2426 depth = set->queue_depth;
2428 err = __blk_mq_alloc_rq_maps(set);
2432 set->queue_depth >>= 1;
2433 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2437 } while (set->queue_depth);
2439 if (!set->queue_depth || err) {
2440 pr_err("blk-mq: failed to allocate request map\n");
2444 if (depth != set->queue_depth)
2445 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2446 depth, set->queue_depth);
2452 * Alloc a tag set to be associated with one or more request queues.
2453 * May fail with EINVAL for various error conditions. May adjust the
2454 * requested depth down, if if it too large. In that case, the set
2455 * value will be stored in set->queue_depth.
2457 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2461 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2463 if (!set->nr_hw_queues)
2465 if (!set->queue_depth)
2467 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2470 if (!set->ops->queue_rq)
2473 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2474 pr_info("blk-mq: reduced tag depth to %u\n",
2476 set->queue_depth = BLK_MQ_MAX_DEPTH;
2480 * If a crashdump is active, then we are potentially in a very
2481 * memory constrained environment. Limit us to 1 queue and
2482 * 64 tags to prevent using too much memory.
2484 if (is_kdump_kernel()) {
2485 set->nr_hw_queues = 1;
2486 set->queue_depth = min(64U, set->queue_depth);
2489 * There is no use for more h/w queues than cpus.
2491 if (set->nr_hw_queues > nr_cpu_ids)
2492 set->nr_hw_queues = nr_cpu_ids;
2494 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2495 GFP_KERNEL, set->numa_node);
2500 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2501 GFP_KERNEL, set->numa_node);
2505 if (set->ops->map_queues)
2506 ret = set->ops->map_queues(set);
2508 ret = blk_mq_map_queues(set);
2510 goto out_free_mq_map;
2512 ret = blk_mq_alloc_rq_maps(set);
2514 goto out_free_mq_map;
2516 mutex_init(&set->tag_list_lock);
2517 INIT_LIST_HEAD(&set->tag_list);
2529 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2531 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2535 for (i = 0; i < nr_cpu_ids; i++)
2536 blk_mq_free_map_and_requests(set, i);
2544 EXPORT_SYMBOL(blk_mq_free_tag_set);
2546 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2548 struct blk_mq_tag_set *set = q->tag_set;
2549 struct blk_mq_hw_ctx *hctx;
2556 queue_for_each_hw_ctx(q, hctx, i) {
2560 * If we're using an MQ scheduler, just update the scheduler
2561 * queue depth. This is similar to what the old code would do.
2563 if (!hctx->sched_tags)
2564 ret = blk_mq_tag_update_depth(hctx->tags,
2565 min(nr, set->queue_depth));
2567 ret = blk_mq_tag_update_depth(hctx->sched_tags, nr);
2573 q->nr_requests = nr;
2578 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2580 struct request_queue *q;
2582 if (nr_hw_queues > nr_cpu_ids)
2583 nr_hw_queues = nr_cpu_ids;
2584 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2587 list_for_each_entry(q, &set->tag_list, tag_set_list)
2588 blk_mq_freeze_queue(q);
2590 set->nr_hw_queues = nr_hw_queues;
2591 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2592 blk_mq_realloc_hw_ctxs(set, q);
2594 if (q->nr_hw_queues > 1)
2595 blk_queue_make_request(q, blk_mq_make_request);
2597 blk_queue_make_request(q, blk_sq_make_request);
2599 blk_mq_queue_reinit(q, cpu_online_mask);
2602 list_for_each_entry(q, &set->tag_list, tag_set_list)
2603 blk_mq_unfreeze_queue(q);
2605 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2607 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2608 struct blk_mq_hw_ctx *hctx,
2611 struct blk_rq_stat stat[2];
2612 unsigned long ret = 0;
2615 * If stats collection isn't on, don't sleep but turn it on for
2618 if (!blk_stat_enable(q))
2622 * We don't have to do this once per IO, should optimize this
2623 * to just use the current window of stats until it changes
2625 memset(&stat, 0, sizeof(stat));
2626 blk_hctx_stat_get(hctx, stat);
2629 * As an optimistic guess, use half of the mean service time
2630 * for this type of request. We can (and should) make this smarter.
2631 * For instance, if the completion latencies are tight, we can
2632 * get closer than just half the mean. This is especially
2633 * important on devices where the completion latencies are longer
2636 if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples)
2637 ret = (stat[BLK_STAT_READ].mean + 1) / 2;
2638 else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples)
2639 ret = (stat[BLK_STAT_WRITE].mean + 1) / 2;
2644 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2645 struct blk_mq_hw_ctx *hctx,
2648 struct hrtimer_sleeper hs;
2649 enum hrtimer_mode mode;
2653 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2659 * -1: don't ever hybrid sleep
2660 * 0: use half of prev avg
2661 * >0: use this specific value
2663 if (q->poll_nsec == -1)
2665 else if (q->poll_nsec > 0)
2666 nsecs = q->poll_nsec;
2668 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2673 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2676 * This will be replaced with the stats tracking code, using
2677 * 'avg_completion_time / 2' as the pre-sleep target.
2681 mode = HRTIMER_MODE_REL;
2682 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2683 hrtimer_set_expires(&hs.timer, kt);
2685 hrtimer_init_sleeper(&hs, current);
2687 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2689 set_current_state(TASK_UNINTERRUPTIBLE);
2690 hrtimer_start_expires(&hs.timer, mode);
2693 hrtimer_cancel(&hs.timer);
2694 mode = HRTIMER_MODE_ABS;
2695 } while (hs.task && !signal_pending(current));
2697 __set_current_state(TASK_RUNNING);
2698 destroy_hrtimer_on_stack(&hs.timer);
2702 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2704 struct request_queue *q = hctx->queue;
2708 * If we sleep, have the caller restart the poll loop to reset
2709 * the state. Like for the other success return cases, the
2710 * caller is responsible for checking if the IO completed. If
2711 * the IO isn't complete, we'll get called again and will go
2712 * straight to the busy poll loop.
2714 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2717 hctx->poll_considered++;
2719 state = current->state;
2720 while (!need_resched()) {
2723 hctx->poll_invoked++;
2725 ret = q->mq_ops->poll(hctx, rq->tag);
2727 hctx->poll_success++;
2728 set_current_state(TASK_RUNNING);
2732 if (signal_pending_state(state, current))
2733 set_current_state(TASK_RUNNING);
2735 if (current->state == TASK_RUNNING)
2745 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2747 struct blk_mq_hw_ctx *hctx;
2748 struct blk_plug *plug;
2751 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2752 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2755 plug = current->plug;
2757 blk_flush_plug_list(plug, false);
2759 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2760 if (!blk_qc_t_is_internal(cookie))
2761 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2763 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2765 return __blk_mq_poll(hctx, rq);
2767 EXPORT_SYMBOL_GPL(blk_mq_poll);
2769 void blk_mq_disable_hotplug(void)
2771 mutex_lock(&all_q_mutex);
2774 void blk_mq_enable_hotplug(void)
2776 mutex_unlock(&all_q_mutex);
2779 static int __init blk_mq_init(void)
2781 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2782 blk_mq_hctx_notify_dead);
2784 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2785 blk_mq_queue_reinit_prepare,
2786 blk_mq_queue_reinit_dead);
2789 subsys_initcall(blk_mq_init);