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-tag.h"
37 #include "blk-mq-sched.h"
39 static DEFINE_MUTEX(all_q_mutex);
40 static LIST_HEAD(all_q_list);
42 static void blk_mq_poll_stats_start(struct request_queue *q);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
46 * Check if any of the ctx's have pending work in this hardware queue
48 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
50 return sbitmap_any_bit_set(&hctx->ctx_map) ||
51 !list_empty_careful(&hctx->dispatch) ||
52 blk_mq_sched_has_work(hctx);
56 * Mark this ctx as having pending work in this hardware queue
58 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
59 struct blk_mq_ctx *ctx)
61 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
62 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
65 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
66 struct blk_mq_ctx *ctx)
68 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
71 void blk_freeze_queue_start(struct request_queue *q)
75 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
76 if (freeze_depth == 1) {
77 percpu_ref_kill(&q->q_usage_counter);
78 blk_mq_run_hw_queues(q, false);
81 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
83 void blk_mq_freeze_queue_wait(struct request_queue *q)
85 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
87 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
89 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
90 unsigned long timeout)
92 return wait_event_timeout(q->mq_freeze_wq,
93 percpu_ref_is_zero(&q->q_usage_counter),
96 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
102 void blk_freeze_queue(struct request_queue *q)
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
111 blk_freeze_queue_start(q);
112 blk_mq_freeze_queue_wait(q);
115 void blk_mq_freeze_queue(struct request_queue *q)
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
125 void blk_mq_unfreeze_queue(struct request_queue *q)
129 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
130 WARN_ON_ONCE(freeze_depth < 0);
132 percpu_ref_reinit(&q->q_usage_counter);
133 wake_up_all(&q->mq_freeze_wq);
136 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
139 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
142 * Note: this function does not prevent that the struct request end_io()
143 * callback function is invoked. Additionally, it is not prevented that
144 * new queue_rq() calls occur unless the queue has been stopped first.
146 void blk_mq_quiesce_queue(struct request_queue *q)
148 struct blk_mq_hw_ctx *hctx;
152 blk_mq_stop_hw_queues(q);
154 queue_for_each_hw_ctx(q, hctx, i) {
155 if (hctx->flags & BLK_MQ_F_BLOCKING)
156 synchronize_srcu(&hctx->queue_rq_srcu);
163 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
165 void blk_mq_wake_waiters(struct request_queue *q)
167 struct blk_mq_hw_ctx *hctx;
170 queue_for_each_hw_ctx(q, hctx, i)
171 if (blk_mq_hw_queue_mapped(hctx))
172 blk_mq_tag_wakeup_all(hctx->tags, true);
175 * If we are called because the queue has now been marked as
176 * dying, we need to ensure that processes currently waiting on
177 * the queue are notified as well.
179 wake_up_all(&q->mq_freeze_wq);
182 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
184 return blk_mq_has_free_tags(hctx->tags);
186 EXPORT_SYMBOL(blk_mq_can_queue);
188 void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
189 struct request *rq, unsigned int op)
191 INIT_LIST_HEAD(&rq->queuelist);
192 /* csd/requeue_work/fifo_time is initialized before use */
196 if (blk_queue_io_stat(q))
197 rq->rq_flags |= RQF_IO_STAT;
198 /* do not touch atomic flags, it needs atomic ops against the timer */
200 INIT_HLIST_NODE(&rq->hash);
201 RB_CLEAR_NODE(&rq->rb_node);
204 rq->start_time = jiffies;
205 #ifdef CONFIG_BLK_CGROUP
207 set_start_time_ns(rq);
208 rq->io_start_time_ns = 0;
210 rq->nr_phys_segments = 0;
211 #if defined(CONFIG_BLK_DEV_INTEGRITY)
212 rq->nr_integrity_segments = 0;
215 /* tag was already set */
219 INIT_LIST_HEAD(&rq->timeout_list);
223 rq->end_io_data = NULL;
226 ctx->rq_dispatched[op_is_sync(op)]++;
228 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init);
230 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
236 tag = blk_mq_get_tag(data);
237 if (tag != BLK_MQ_TAG_FAIL) {
238 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
240 rq = tags->static_rqs[tag];
242 if (data->flags & BLK_MQ_REQ_INTERNAL) {
244 rq->internal_tag = tag;
246 if (blk_mq_tag_busy(data->hctx)) {
247 rq->rq_flags = RQF_MQ_INFLIGHT;
248 atomic_inc(&data->hctx->nr_active);
251 rq->internal_tag = -1;
252 data->hctx->tags->rqs[rq->tag] = rq;
255 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
261 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
263 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
266 struct blk_mq_alloc_data alloc_data = { .flags = flags };
270 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
274 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
276 blk_mq_put_ctx(alloc_data.ctx);
280 return ERR_PTR(-EWOULDBLOCK);
283 rq->__sector = (sector_t) -1;
284 rq->bio = rq->biotail = NULL;
287 EXPORT_SYMBOL(blk_mq_alloc_request);
289 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
290 unsigned int flags, unsigned int hctx_idx)
292 struct blk_mq_alloc_data alloc_data = { .flags = flags };
298 * If the tag allocator sleeps we could get an allocation for a
299 * different hardware context. No need to complicate the low level
300 * allocator for this for the rare use case of a command tied to
303 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
304 return ERR_PTR(-EINVAL);
306 if (hctx_idx >= q->nr_hw_queues)
307 return ERR_PTR(-EIO);
309 ret = blk_queue_enter(q, true);
314 * Check if the hardware context is actually mapped to anything.
315 * If not tell the caller that it should skip this queue.
317 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
318 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
320 return ERR_PTR(-EXDEV);
322 cpu = cpumask_first(alloc_data.hctx->cpumask);
323 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
325 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
330 return ERR_PTR(-EWOULDBLOCK);
334 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
336 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
339 const int sched_tag = rq->internal_tag;
340 struct request_queue *q = rq->q;
342 if (rq->rq_flags & RQF_MQ_INFLIGHT)
343 atomic_dec(&hctx->nr_active);
345 wbt_done(q->rq_wb, &rq->issue_stat);
348 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
349 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
351 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
353 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
354 blk_mq_sched_restart(hctx);
358 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
361 struct blk_mq_ctx *ctx = rq->mq_ctx;
363 ctx->rq_completed[rq_is_sync(rq)]++;
364 __blk_mq_finish_request(hctx, ctx, rq);
367 void blk_mq_finish_request(struct request *rq)
369 blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
371 EXPORT_SYMBOL_GPL(blk_mq_finish_request);
373 void blk_mq_free_request(struct request *rq)
375 blk_mq_sched_put_request(rq);
377 EXPORT_SYMBOL_GPL(blk_mq_free_request);
379 inline void __blk_mq_end_request(struct request *rq, int error)
381 blk_account_io_done(rq);
384 wbt_done(rq->q->rq_wb, &rq->issue_stat);
385 rq->end_io(rq, error);
387 if (unlikely(blk_bidi_rq(rq)))
388 blk_mq_free_request(rq->next_rq);
389 blk_mq_free_request(rq);
392 EXPORT_SYMBOL(__blk_mq_end_request);
394 void blk_mq_end_request(struct request *rq, int error)
396 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
398 __blk_mq_end_request(rq, error);
400 EXPORT_SYMBOL(blk_mq_end_request);
402 static void __blk_mq_complete_request_remote(void *data)
404 struct request *rq = data;
406 rq->q->softirq_done_fn(rq);
409 static void __blk_mq_complete_request(struct request *rq)
411 struct blk_mq_ctx *ctx = rq->mq_ctx;
415 if (rq->internal_tag != -1)
416 blk_mq_sched_completed_request(rq);
417 if (rq->rq_flags & RQF_STATS) {
418 blk_mq_poll_stats_start(rq->q);
422 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
423 rq->q->softirq_done_fn(rq);
428 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
429 shared = cpus_share_cache(cpu, ctx->cpu);
431 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
432 rq->csd.func = __blk_mq_complete_request_remote;
435 smp_call_function_single_async(ctx->cpu, &rq->csd);
437 rq->q->softirq_done_fn(rq);
443 * blk_mq_complete_request - end I/O on a request
444 * @rq: the request being processed
447 * Ends all I/O on a request. It does not handle partial completions.
448 * The actual completion happens out-of-order, through a IPI handler.
450 void blk_mq_complete_request(struct request *rq)
452 struct request_queue *q = rq->q;
454 if (unlikely(blk_should_fake_timeout(q)))
456 if (!blk_mark_rq_complete(rq))
457 __blk_mq_complete_request(rq);
459 EXPORT_SYMBOL(blk_mq_complete_request);
461 int blk_mq_request_started(struct request *rq)
463 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
465 EXPORT_SYMBOL_GPL(blk_mq_request_started);
467 void blk_mq_start_request(struct request *rq)
469 struct request_queue *q = rq->q;
471 blk_mq_sched_started_request(rq);
473 trace_block_rq_issue(q, rq);
475 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
476 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
477 rq->rq_flags |= RQF_STATS;
478 wbt_issue(q->rq_wb, &rq->issue_stat);
484 * Ensure that ->deadline is visible before set the started
485 * flag and clear the completed flag.
487 smp_mb__before_atomic();
490 * Mark us as started and clear complete. Complete might have been
491 * set if requeue raced with timeout, which then marked it as
492 * complete. So be sure to clear complete again when we start
493 * the request, otherwise we'll ignore the completion event.
495 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
496 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
497 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
498 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
500 if (q->dma_drain_size && blk_rq_bytes(rq)) {
502 * Make sure space for the drain appears. We know we can do
503 * this because max_hw_segments has been adjusted to be one
504 * fewer than the device can handle.
506 rq->nr_phys_segments++;
509 EXPORT_SYMBOL(blk_mq_start_request);
512 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
513 * flag isn't set yet, so there may be race with timeout handler,
514 * but given rq->deadline is just set in .queue_rq() under
515 * this situation, the race won't be possible in reality because
516 * rq->timeout should be set as big enough to cover the window
517 * between blk_mq_start_request() called from .queue_rq() and
518 * clearing REQ_ATOM_STARTED here.
520 static void __blk_mq_requeue_request(struct request *rq)
522 struct request_queue *q = rq->q;
524 trace_block_rq_requeue(q, rq);
525 wbt_requeue(q->rq_wb, &rq->issue_stat);
526 blk_mq_sched_requeue_request(rq);
528 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
529 if (q->dma_drain_size && blk_rq_bytes(rq))
530 rq->nr_phys_segments--;
534 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
536 __blk_mq_requeue_request(rq);
538 BUG_ON(blk_queued_rq(rq));
539 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
541 EXPORT_SYMBOL(blk_mq_requeue_request);
543 static void blk_mq_requeue_work(struct work_struct *work)
545 struct request_queue *q =
546 container_of(work, struct request_queue, requeue_work.work);
548 struct request *rq, *next;
551 spin_lock_irqsave(&q->requeue_lock, flags);
552 list_splice_init(&q->requeue_list, &rq_list);
553 spin_unlock_irqrestore(&q->requeue_lock, flags);
555 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
556 if (!(rq->rq_flags & RQF_SOFTBARRIER))
559 rq->rq_flags &= ~RQF_SOFTBARRIER;
560 list_del_init(&rq->queuelist);
561 blk_mq_sched_insert_request(rq, true, false, false, true);
564 while (!list_empty(&rq_list)) {
565 rq = list_entry(rq_list.next, struct request, queuelist);
566 list_del_init(&rq->queuelist);
567 blk_mq_sched_insert_request(rq, false, false, false, true);
570 blk_mq_run_hw_queues(q, false);
573 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
574 bool kick_requeue_list)
576 struct request_queue *q = rq->q;
580 * We abuse this flag that is otherwise used by the I/O scheduler to
581 * request head insertation from the workqueue.
583 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
585 spin_lock_irqsave(&q->requeue_lock, flags);
587 rq->rq_flags |= RQF_SOFTBARRIER;
588 list_add(&rq->queuelist, &q->requeue_list);
590 list_add_tail(&rq->queuelist, &q->requeue_list);
592 spin_unlock_irqrestore(&q->requeue_lock, flags);
594 if (kick_requeue_list)
595 blk_mq_kick_requeue_list(q);
597 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
599 void blk_mq_kick_requeue_list(struct request_queue *q)
601 kblockd_schedule_delayed_work(&q->requeue_work, 0);
603 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
605 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
608 kblockd_schedule_delayed_work(&q->requeue_work,
609 msecs_to_jiffies(msecs));
611 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
613 void blk_mq_abort_requeue_list(struct request_queue *q)
618 spin_lock_irqsave(&q->requeue_lock, flags);
619 list_splice_init(&q->requeue_list, &rq_list);
620 spin_unlock_irqrestore(&q->requeue_lock, flags);
622 while (!list_empty(&rq_list)) {
625 rq = list_first_entry(&rq_list, struct request, queuelist);
626 list_del_init(&rq->queuelist);
628 blk_mq_end_request(rq, rq->errors);
631 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
633 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
635 if (tag < tags->nr_tags) {
636 prefetch(tags->rqs[tag]);
637 return tags->rqs[tag];
642 EXPORT_SYMBOL(blk_mq_tag_to_rq);
644 struct blk_mq_timeout_data {
646 unsigned int next_set;
649 void blk_mq_rq_timed_out(struct request *req, bool reserved)
651 const struct blk_mq_ops *ops = req->q->mq_ops;
652 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
655 * We know that complete is set at this point. If STARTED isn't set
656 * anymore, then the request isn't active and the "timeout" should
657 * just be ignored. This can happen due to the bitflag ordering.
658 * Timeout first checks if STARTED is set, and if it is, assumes
659 * the request is active. But if we race with completion, then
660 * both flags will get cleared. So check here again, and ignore
661 * a timeout event with a request that isn't active.
663 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
667 ret = ops->timeout(req, reserved);
671 __blk_mq_complete_request(req);
673 case BLK_EH_RESET_TIMER:
675 blk_clear_rq_complete(req);
677 case BLK_EH_NOT_HANDLED:
680 printk(KERN_ERR "block: bad eh return: %d\n", ret);
685 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
686 struct request *rq, void *priv, bool reserved)
688 struct blk_mq_timeout_data *data = priv;
690 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
694 * The rq being checked may have been freed and reallocated
695 * out already here, we avoid this race by checking rq->deadline
696 * and REQ_ATOM_COMPLETE flag together:
698 * - if rq->deadline is observed as new value because of
699 * reusing, the rq won't be timed out because of timing.
700 * - if rq->deadline is observed as previous value,
701 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
702 * because we put a barrier between setting rq->deadline
703 * and clearing the flag in blk_mq_start_request(), so
704 * this rq won't be timed out too.
706 if (time_after_eq(jiffies, rq->deadline)) {
707 if (!blk_mark_rq_complete(rq))
708 blk_mq_rq_timed_out(rq, reserved);
709 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
710 data->next = rq->deadline;
715 static void blk_mq_timeout_work(struct work_struct *work)
717 struct request_queue *q =
718 container_of(work, struct request_queue, timeout_work);
719 struct blk_mq_timeout_data data = {
725 /* A deadlock might occur if a request is stuck requiring a
726 * timeout at the same time a queue freeze is waiting
727 * completion, since the timeout code would not be able to
728 * acquire the queue reference here.
730 * That's why we don't use blk_queue_enter here; instead, we use
731 * percpu_ref_tryget directly, because we need to be able to
732 * obtain a reference even in the short window between the queue
733 * starting to freeze, by dropping the first reference in
734 * blk_freeze_queue_start, and the moment the last request is
735 * consumed, marked by the instant q_usage_counter reaches
738 if (!percpu_ref_tryget(&q->q_usage_counter))
741 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
744 data.next = blk_rq_timeout(round_jiffies_up(data.next));
745 mod_timer(&q->timeout, data.next);
747 struct blk_mq_hw_ctx *hctx;
749 queue_for_each_hw_ctx(q, hctx, i) {
750 /* the hctx may be unmapped, so check it here */
751 if (blk_mq_hw_queue_mapped(hctx))
752 blk_mq_tag_idle(hctx);
759 * Reverse check our software queue for entries that we could potentially
760 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
761 * too much time checking for merges.
763 static bool blk_mq_attempt_merge(struct request_queue *q,
764 struct blk_mq_ctx *ctx, struct bio *bio)
769 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
775 if (!blk_rq_merge_ok(rq, bio))
778 switch (blk_try_merge(rq, bio)) {
779 case ELEVATOR_BACK_MERGE:
780 if (blk_mq_sched_allow_merge(q, rq, bio))
781 merged = bio_attempt_back_merge(q, rq, bio);
783 case ELEVATOR_FRONT_MERGE:
784 if (blk_mq_sched_allow_merge(q, rq, bio))
785 merged = bio_attempt_front_merge(q, rq, bio);
787 case ELEVATOR_DISCARD_MERGE:
788 merged = bio_attempt_discard_merge(q, rq, bio);
802 struct flush_busy_ctx_data {
803 struct blk_mq_hw_ctx *hctx;
804 struct list_head *list;
807 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
809 struct flush_busy_ctx_data *flush_data = data;
810 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
811 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
813 sbitmap_clear_bit(sb, bitnr);
814 spin_lock(&ctx->lock);
815 list_splice_tail_init(&ctx->rq_list, flush_data->list);
816 spin_unlock(&ctx->lock);
821 * Process software queues that have been marked busy, splicing them
822 * to the for-dispatch
824 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
826 struct flush_busy_ctx_data data = {
831 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
833 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
835 static inline unsigned int queued_to_index(unsigned int queued)
840 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
843 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
846 struct blk_mq_alloc_data data = {
848 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
849 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
855 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
856 data.flags |= BLK_MQ_REQ_RESERVED;
858 rq->tag = blk_mq_get_tag(&data);
860 if (blk_mq_tag_busy(data.hctx)) {
861 rq->rq_flags |= RQF_MQ_INFLIGHT;
862 atomic_inc(&data.hctx->nr_active);
864 data.hctx->tags->rqs[rq->tag] = rq;
870 return rq->tag != -1;
873 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
876 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
879 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
880 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
881 atomic_dec(&hctx->nr_active);
885 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
888 if (rq->tag == -1 || rq->internal_tag == -1)
891 __blk_mq_put_driver_tag(hctx, rq);
894 static void blk_mq_put_driver_tag(struct request *rq)
896 struct blk_mq_hw_ctx *hctx;
898 if (rq->tag == -1 || rq->internal_tag == -1)
901 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
902 __blk_mq_put_driver_tag(hctx, rq);
906 * If we fail getting a driver tag because all the driver tags are already
907 * assigned and on the dispatch list, BUT the first entry does not have a
908 * tag, then we could deadlock. For that case, move entries with assigned
909 * driver tags to the front, leaving the set of tagged requests in the
910 * same order, and the untagged set in the same order.
912 static bool reorder_tags_to_front(struct list_head *list)
914 struct request *rq, *tmp, *first = NULL;
916 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
920 list_move(&rq->queuelist, list);
926 return first != NULL;
929 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
932 struct blk_mq_hw_ctx *hctx;
934 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
936 list_del(&wait->task_list);
937 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
938 blk_mq_run_hw_queue(hctx, true);
942 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
944 struct sbq_wait_state *ws;
947 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
948 * The thread which wins the race to grab this bit adds the hardware
949 * queue to the wait queue.
951 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
952 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
955 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
956 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
959 * As soon as this returns, it's no longer safe to fiddle with
960 * hctx->dispatch_wait, since a completion can wake up the wait queue
961 * and unlock the bit.
963 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
967 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
969 struct blk_mq_hw_ctx *hctx;
971 int errors, queued, ret = BLK_MQ_RQ_QUEUE_OK;
973 if (list_empty(list))
977 * Now process all the entries, sending them to the driver.
981 struct blk_mq_queue_data bd;
983 rq = list_first_entry(list, struct request, queuelist);
984 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
985 if (!queued && reorder_tags_to_front(list))
989 * The initial allocation attempt failed, so we need to
990 * rerun the hardware queue when a tag is freed.
992 if (!blk_mq_dispatch_wait_add(hctx))
996 * It's possible that a tag was freed in the window
997 * between the allocation failure and adding the
998 * hardware queue to the wait queue.
1000 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1004 list_del_init(&rq->queuelist);
1009 * Flag last if we have no more requests, or if we have more
1010 * but can't assign a driver tag to it.
1012 if (list_empty(list))
1015 struct request *nxt;
1017 nxt = list_first_entry(list, struct request, queuelist);
1018 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1021 ret = q->mq_ops->queue_rq(hctx, &bd);
1023 case BLK_MQ_RQ_QUEUE_OK:
1026 case BLK_MQ_RQ_QUEUE_BUSY:
1027 blk_mq_put_driver_tag_hctx(hctx, rq);
1028 list_add(&rq->queuelist, list);
1029 __blk_mq_requeue_request(rq);
1032 pr_err("blk-mq: bad return on queue: %d\n", ret);
1033 case BLK_MQ_RQ_QUEUE_ERROR:
1036 blk_mq_end_request(rq, rq->errors);
1040 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
1042 } while (!list_empty(list));
1044 hctx->dispatched[queued_to_index(queued)]++;
1047 * Any items that need requeuing? Stuff them into hctx->dispatch,
1048 * that is where we will continue on next queue run.
1050 if (!list_empty(list)) {
1052 * If an I/O scheduler has been configured and we got a driver
1053 * tag for the next request already, free it again.
1055 rq = list_first_entry(list, struct request, queuelist);
1056 blk_mq_put_driver_tag(rq);
1058 spin_lock(&hctx->lock);
1059 list_splice_init(list, &hctx->dispatch);
1060 spin_unlock(&hctx->lock);
1063 * If SCHED_RESTART was set by the caller of this function and
1064 * it is no longer set that means that it was cleared by another
1065 * thread and hence that a queue rerun is needed.
1067 * If TAG_WAITING is set that means that an I/O scheduler has
1068 * been configured and another thread is waiting for a driver
1069 * tag. To guarantee fairness, do not rerun this hardware queue
1070 * but let the other thread grab the driver tag.
1072 * If no I/O scheduler has been configured it is possible that
1073 * the hardware queue got stopped and restarted before requests
1074 * were pushed back onto the dispatch list. Rerun the queue to
1075 * avoid starvation. Notes:
1076 * - blk_mq_run_hw_queue() checks whether or not a queue has
1077 * been stopped before rerunning a queue.
1078 * - Some but not all block drivers stop a queue before
1079 * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq
1082 if (!blk_mq_sched_needs_restart(hctx) &&
1083 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1084 blk_mq_run_hw_queue(hctx, true);
1087 return (queued + errors) != 0;
1090 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1094 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1095 cpu_online(hctx->next_cpu));
1097 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1099 blk_mq_sched_dispatch_requests(hctx);
1104 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1105 blk_mq_sched_dispatch_requests(hctx);
1106 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1111 * It'd be great if the workqueue API had a way to pass
1112 * in a mask and had some smarts for more clever placement.
1113 * For now we just round-robin here, switching for every
1114 * BLK_MQ_CPU_WORK_BATCH queued items.
1116 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1118 if (hctx->queue->nr_hw_queues == 1)
1119 return WORK_CPU_UNBOUND;
1121 if (--hctx->next_cpu_batch <= 0) {
1124 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1125 if (next_cpu >= nr_cpu_ids)
1126 next_cpu = cpumask_first(hctx->cpumask);
1128 hctx->next_cpu = next_cpu;
1129 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1132 return hctx->next_cpu;
1135 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1136 unsigned long msecs)
1138 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1139 !blk_mq_hw_queue_mapped(hctx)))
1142 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1143 int cpu = get_cpu();
1144 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1145 __blk_mq_run_hw_queue(hctx);
1154 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx),
1157 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1158 &hctx->delayed_run_work,
1159 msecs_to_jiffies(msecs));
1162 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1164 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1166 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1168 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1170 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1172 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1174 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1176 struct blk_mq_hw_ctx *hctx;
1179 queue_for_each_hw_ctx(q, hctx, i) {
1180 if (!blk_mq_hctx_has_pending(hctx) ||
1181 blk_mq_hctx_stopped(hctx))
1184 blk_mq_run_hw_queue(hctx, async);
1187 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1190 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1191 * @q: request queue.
1193 * The caller is responsible for serializing this function against
1194 * blk_mq_{start,stop}_hw_queue().
1196 bool blk_mq_queue_stopped(struct request_queue *q)
1198 struct blk_mq_hw_ctx *hctx;
1201 queue_for_each_hw_ctx(q, hctx, i)
1202 if (blk_mq_hctx_stopped(hctx))
1207 EXPORT_SYMBOL(blk_mq_queue_stopped);
1209 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1211 cancel_work(&hctx->run_work);
1212 cancel_delayed_work(&hctx->delay_work);
1213 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1215 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1217 void blk_mq_stop_hw_queues(struct request_queue *q)
1219 struct blk_mq_hw_ctx *hctx;
1222 queue_for_each_hw_ctx(q, hctx, i)
1223 blk_mq_stop_hw_queue(hctx);
1225 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1227 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1229 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1231 blk_mq_run_hw_queue(hctx, false);
1233 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1235 void blk_mq_start_hw_queues(struct request_queue *q)
1237 struct blk_mq_hw_ctx *hctx;
1240 queue_for_each_hw_ctx(q, hctx, i)
1241 blk_mq_start_hw_queue(hctx);
1243 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1245 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1247 if (!blk_mq_hctx_stopped(hctx))
1250 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1251 blk_mq_run_hw_queue(hctx, async);
1253 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1255 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1257 struct blk_mq_hw_ctx *hctx;
1260 queue_for_each_hw_ctx(q, hctx, i)
1261 blk_mq_start_stopped_hw_queue(hctx, async);
1263 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1265 static void blk_mq_run_work_fn(struct work_struct *work)
1267 struct blk_mq_hw_ctx *hctx;
1269 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1271 __blk_mq_run_hw_queue(hctx);
1274 static void blk_mq_delayed_run_work_fn(struct work_struct *work)
1276 struct blk_mq_hw_ctx *hctx;
1278 hctx = container_of(work, struct blk_mq_hw_ctx, delayed_run_work.work);
1280 __blk_mq_run_hw_queue(hctx);
1283 static void blk_mq_delay_work_fn(struct work_struct *work)
1285 struct blk_mq_hw_ctx *hctx;
1287 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1289 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1290 __blk_mq_run_hw_queue(hctx);
1293 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1295 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1298 blk_mq_stop_hw_queue(hctx);
1299 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1300 &hctx->delay_work, msecs_to_jiffies(msecs));
1302 EXPORT_SYMBOL(blk_mq_delay_queue);
1304 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1308 struct blk_mq_ctx *ctx = rq->mq_ctx;
1310 trace_block_rq_insert(hctx->queue, rq);
1313 list_add(&rq->queuelist, &ctx->rq_list);
1315 list_add_tail(&rq->queuelist, &ctx->rq_list);
1318 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1321 struct blk_mq_ctx *ctx = rq->mq_ctx;
1323 __blk_mq_insert_req_list(hctx, rq, at_head);
1324 blk_mq_hctx_mark_pending(hctx, ctx);
1327 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1328 struct list_head *list)
1332 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1335 spin_lock(&ctx->lock);
1336 while (!list_empty(list)) {
1339 rq = list_first_entry(list, struct request, queuelist);
1340 BUG_ON(rq->mq_ctx != ctx);
1341 list_del_init(&rq->queuelist);
1342 __blk_mq_insert_req_list(hctx, rq, false);
1344 blk_mq_hctx_mark_pending(hctx, ctx);
1345 spin_unlock(&ctx->lock);
1348 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1350 struct request *rqa = container_of(a, struct request, queuelist);
1351 struct request *rqb = container_of(b, struct request, queuelist);
1353 return !(rqa->mq_ctx < rqb->mq_ctx ||
1354 (rqa->mq_ctx == rqb->mq_ctx &&
1355 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1358 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1360 struct blk_mq_ctx *this_ctx;
1361 struct request_queue *this_q;
1364 LIST_HEAD(ctx_list);
1367 list_splice_init(&plug->mq_list, &list);
1369 list_sort(NULL, &list, plug_ctx_cmp);
1375 while (!list_empty(&list)) {
1376 rq = list_entry_rq(list.next);
1377 list_del_init(&rq->queuelist);
1379 if (rq->mq_ctx != this_ctx) {
1381 trace_block_unplug(this_q, depth, from_schedule);
1382 blk_mq_sched_insert_requests(this_q, this_ctx,
1387 this_ctx = rq->mq_ctx;
1393 list_add_tail(&rq->queuelist, &ctx_list);
1397 * If 'this_ctx' is set, we know we have entries to complete
1398 * on 'ctx_list'. Do those.
1401 trace_block_unplug(this_q, depth, from_schedule);
1402 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1407 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1409 blk_init_request_from_bio(rq, bio);
1411 blk_account_io_start(rq, true);
1414 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1416 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1417 !blk_queue_nomerges(hctx->queue);
1420 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1421 struct blk_mq_ctx *ctx,
1422 struct request *rq, struct bio *bio)
1424 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1425 blk_mq_bio_to_request(rq, bio);
1426 spin_lock(&ctx->lock);
1428 __blk_mq_insert_request(hctx, rq, false);
1429 spin_unlock(&ctx->lock);
1432 struct request_queue *q = hctx->queue;
1434 spin_lock(&ctx->lock);
1435 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1436 blk_mq_bio_to_request(rq, bio);
1440 spin_unlock(&ctx->lock);
1441 __blk_mq_finish_request(hctx, ctx, rq);
1446 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1449 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1451 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1454 static void __blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie,
1457 struct request_queue *q = rq->q;
1458 struct blk_mq_queue_data bd = {
1462 struct blk_mq_hw_ctx *hctx;
1463 blk_qc_t new_cookie;
1469 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1472 new_cookie = request_to_qc_t(hctx, rq);
1475 * For OK queue, we are done. For error, kill it. Any other
1476 * error (busy), just add it to our list as we previously
1479 ret = q->mq_ops->queue_rq(hctx, &bd);
1480 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1481 *cookie = new_cookie;
1485 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1486 *cookie = BLK_QC_T_NONE;
1488 blk_mq_end_request(rq, rq->errors);
1492 __blk_mq_requeue_request(rq);
1494 blk_mq_sched_insert_request(rq, false, true, false, may_sleep);
1497 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1498 struct request *rq, blk_qc_t *cookie)
1500 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1502 __blk_mq_try_issue_directly(rq, cookie, false);
1505 unsigned int srcu_idx;
1509 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1510 __blk_mq_try_issue_directly(rq, cookie, true);
1511 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1515 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1517 const int is_sync = op_is_sync(bio->bi_opf);
1518 const int is_flush_fua = op_is_flush(bio->bi_opf);
1519 struct blk_mq_alloc_data data = { .flags = 0 };
1521 unsigned int request_count = 0;
1522 struct blk_plug *plug;
1523 struct request *same_queue_rq = NULL;
1525 unsigned int wb_acct;
1527 blk_queue_bounce(q, &bio);
1529 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1531 return BLK_QC_T_NONE;
1534 blk_queue_split(q, &bio, q->bio_split);
1536 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1537 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1538 return BLK_QC_T_NONE;
1540 if (blk_mq_sched_bio_merge(q, bio))
1541 return BLK_QC_T_NONE;
1543 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1545 trace_block_getrq(q, bio, bio->bi_opf);
1547 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1548 if (unlikely(!rq)) {
1549 __wbt_done(q->rq_wb, wb_acct);
1550 return BLK_QC_T_NONE;
1553 wbt_track(&rq->issue_stat, wb_acct);
1555 cookie = request_to_qc_t(data.hctx, rq);
1557 plug = current->plug;
1558 if (unlikely(is_flush_fua)) {
1559 blk_mq_bio_to_request(rq, bio);
1561 blk_mq_sched_insert_request(rq, false, true, true,
1564 blk_insert_flush(rq);
1565 blk_mq_run_hw_queue(data.hctx, true);
1567 } else if (plug && q->nr_hw_queues == 1) {
1568 struct request *last = NULL;
1570 blk_mq_bio_to_request(rq, bio);
1573 * @request_count may become stale because of schedule
1574 * out, so check the list again.
1576 if (list_empty(&plug->mq_list))
1578 else if (blk_queue_nomerges(q))
1579 request_count = blk_plug_queued_count(q);
1582 trace_block_plug(q);
1584 last = list_entry_rq(plug->mq_list.prev);
1586 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1587 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1588 blk_flush_plug_list(plug, false);
1589 trace_block_plug(q);
1592 list_add_tail(&rq->queuelist, &plug->mq_list);
1593 } else if (plug && !blk_queue_nomerges(q)) {
1594 blk_mq_bio_to_request(rq, bio);
1597 * We do limited plugging. If the bio can be merged, do that.
1598 * Otherwise the existing request in the plug list will be
1599 * issued. So the plug list will have one request at most
1600 * The plug list might get flushed before this. If that happens,
1601 * the plug list is empty, and same_queue_rq is invalid.
1603 if (list_empty(&plug->mq_list))
1604 same_queue_rq = NULL;
1606 list_del_init(&same_queue_rq->queuelist);
1607 list_add_tail(&rq->queuelist, &plug->mq_list);
1609 blk_mq_put_ctx(data.ctx);
1612 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1616 } else if (q->nr_hw_queues > 1 && is_sync) {
1617 blk_mq_put_ctx(data.ctx);
1618 blk_mq_bio_to_request(rq, bio);
1619 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1621 } else if (q->elevator) {
1622 blk_mq_bio_to_request(rq, bio);
1623 blk_mq_sched_insert_request(rq, false, true, true, true);
1624 } else if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio))
1625 blk_mq_run_hw_queue(data.hctx, true);
1627 blk_mq_put_ctx(data.ctx);
1631 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1632 unsigned int hctx_idx)
1636 if (tags->rqs && set->ops->exit_request) {
1639 for (i = 0; i < tags->nr_tags; i++) {
1640 struct request *rq = tags->static_rqs[i];
1644 set->ops->exit_request(set->driver_data, rq,
1646 tags->static_rqs[i] = NULL;
1650 while (!list_empty(&tags->page_list)) {
1651 page = list_first_entry(&tags->page_list, struct page, lru);
1652 list_del_init(&page->lru);
1654 * Remove kmemleak object previously allocated in
1655 * blk_mq_init_rq_map().
1657 kmemleak_free(page_address(page));
1658 __free_pages(page, page->private);
1662 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1666 kfree(tags->static_rqs);
1667 tags->static_rqs = NULL;
1669 blk_mq_free_tags(tags);
1672 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1673 unsigned int hctx_idx,
1674 unsigned int nr_tags,
1675 unsigned int reserved_tags)
1677 struct blk_mq_tags *tags;
1680 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1681 if (node == NUMA_NO_NODE)
1682 node = set->numa_node;
1684 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1685 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1689 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1690 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1693 blk_mq_free_tags(tags);
1697 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1698 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1700 if (!tags->static_rqs) {
1702 blk_mq_free_tags(tags);
1709 static size_t order_to_size(unsigned int order)
1711 return (size_t)PAGE_SIZE << order;
1714 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1715 unsigned int hctx_idx, unsigned int depth)
1717 unsigned int i, j, entries_per_page, max_order = 4;
1718 size_t rq_size, left;
1721 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1722 if (node == NUMA_NO_NODE)
1723 node = set->numa_node;
1725 INIT_LIST_HEAD(&tags->page_list);
1728 * rq_size is the size of the request plus driver payload, rounded
1729 * to the cacheline size
1731 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1733 left = rq_size * depth;
1735 for (i = 0; i < depth; ) {
1736 int this_order = max_order;
1741 while (this_order && left < order_to_size(this_order - 1))
1745 page = alloc_pages_node(node,
1746 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1752 if (order_to_size(this_order) < rq_size)
1759 page->private = this_order;
1760 list_add_tail(&page->lru, &tags->page_list);
1762 p = page_address(page);
1764 * Allow kmemleak to scan these pages as they contain pointers
1765 * to additional allocations like via ops->init_request().
1767 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1768 entries_per_page = order_to_size(this_order) / rq_size;
1769 to_do = min(entries_per_page, depth - i);
1770 left -= to_do * rq_size;
1771 for (j = 0; j < to_do; j++) {
1772 struct request *rq = p;
1774 tags->static_rqs[i] = rq;
1775 if (set->ops->init_request) {
1776 if (set->ops->init_request(set->driver_data,
1779 tags->static_rqs[i] = NULL;
1791 blk_mq_free_rqs(set, tags, hctx_idx);
1796 * 'cpu' is going away. splice any existing rq_list entries from this
1797 * software queue to the hw queue dispatch list, and ensure that it
1800 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1802 struct blk_mq_hw_ctx *hctx;
1803 struct blk_mq_ctx *ctx;
1806 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1807 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1809 spin_lock(&ctx->lock);
1810 if (!list_empty(&ctx->rq_list)) {
1811 list_splice_init(&ctx->rq_list, &tmp);
1812 blk_mq_hctx_clear_pending(hctx, ctx);
1814 spin_unlock(&ctx->lock);
1816 if (list_empty(&tmp))
1819 spin_lock(&hctx->lock);
1820 list_splice_tail_init(&tmp, &hctx->dispatch);
1821 spin_unlock(&hctx->lock);
1823 blk_mq_run_hw_queue(hctx, true);
1827 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1829 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1833 /* hctx->ctxs will be freed in queue's release handler */
1834 static void blk_mq_exit_hctx(struct request_queue *q,
1835 struct blk_mq_tag_set *set,
1836 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1838 unsigned flush_start_tag = set->queue_depth;
1840 blk_mq_tag_idle(hctx);
1842 if (set->ops->exit_request)
1843 set->ops->exit_request(set->driver_data,
1844 hctx->fq->flush_rq, hctx_idx,
1845 flush_start_tag + hctx_idx);
1847 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1849 if (set->ops->exit_hctx)
1850 set->ops->exit_hctx(hctx, hctx_idx);
1852 if (hctx->flags & BLK_MQ_F_BLOCKING)
1853 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1855 blk_mq_remove_cpuhp(hctx);
1856 blk_free_flush_queue(hctx->fq);
1857 sbitmap_free(&hctx->ctx_map);
1860 static void blk_mq_exit_hw_queues(struct request_queue *q,
1861 struct blk_mq_tag_set *set, int nr_queue)
1863 struct blk_mq_hw_ctx *hctx;
1866 queue_for_each_hw_ctx(q, hctx, i) {
1869 blk_mq_exit_hctx(q, set, hctx, i);
1873 static int blk_mq_init_hctx(struct request_queue *q,
1874 struct blk_mq_tag_set *set,
1875 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1878 unsigned flush_start_tag = set->queue_depth;
1880 node = hctx->numa_node;
1881 if (node == NUMA_NO_NODE)
1882 node = hctx->numa_node = set->numa_node;
1884 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1885 INIT_DELAYED_WORK(&hctx->delayed_run_work, blk_mq_delayed_run_work_fn);
1886 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1887 spin_lock_init(&hctx->lock);
1888 INIT_LIST_HEAD(&hctx->dispatch);
1890 hctx->queue_num = hctx_idx;
1891 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1893 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1895 hctx->tags = set->tags[hctx_idx];
1898 * Allocate space for all possible cpus to avoid allocation at
1901 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1904 goto unregister_cpu_notifier;
1906 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1912 if (set->ops->init_hctx &&
1913 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1916 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1919 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1921 goto sched_exit_hctx;
1923 if (set->ops->init_request &&
1924 set->ops->init_request(set->driver_data,
1925 hctx->fq->flush_rq, hctx_idx,
1926 flush_start_tag + hctx_idx, node))
1929 if (hctx->flags & BLK_MQ_F_BLOCKING)
1930 init_srcu_struct(&hctx->queue_rq_srcu);
1937 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1939 if (set->ops->exit_hctx)
1940 set->ops->exit_hctx(hctx, hctx_idx);
1942 sbitmap_free(&hctx->ctx_map);
1945 unregister_cpu_notifier:
1946 blk_mq_remove_cpuhp(hctx);
1950 static void blk_mq_init_cpu_queues(struct request_queue *q,
1951 unsigned int nr_hw_queues)
1955 for_each_possible_cpu(i) {
1956 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1957 struct blk_mq_hw_ctx *hctx;
1960 spin_lock_init(&__ctx->lock);
1961 INIT_LIST_HEAD(&__ctx->rq_list);
1964 /* If the cpu isn't online, the cpu is mapped to first hctx */
1968 hctx = blk_mq_map_queue(q, i);
1971 * Set local node, IFF we have more than one hw queue. If
1972 * not, we remain on the home node of the device
1974 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1975 hctx->numa_node = local_memory_node(cpu_to_node(i));
1979 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1983 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1984 set->queue_depth, set->reserved_tags);
1985 if (!set->tags[hctx_idx])
1988 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
1993 blk_mq_free_rq_map(set->tags[hctx_idx]);
1994 set->tags[hctx_idx] = NULL;
1998 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
1999 unsigned int hctx_idx)
2001 if (set->tags[hctx_idx]) {
2002 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2003 blk_mq_free_rq_map(set->tags[hctx_idx]);
2004 set->tags[hctx_idx] = NULL;
2008 static void blk_mq_map_swqueue(struct request_queue *q,
2009 const struct cpumask *online_mask)
2011 unsigned int i, hctx_idx;
2012 struct blk_mq_hw_ctx *hctx;
2013 struct blk_mq_ctx *ctx;
2014 struct blk_mq_tag_set *set = q->tag_set;
2017 * Avoid others reading imcomplete hctx->cpumask through sysfs
2019 mutex_lock(&q->sysfs_lock);
2021 queue_for_each_hw_ctx(q, hctx, i) {
2022 cpumask_clear(hctx->cpumask);
2027 * Map software to hardware queues
2029 for_each_possible_cpu(i) {
2030 /* If the cpu isn't online, the cpu is mapped to first hctx */
2031 if (!cpumask_test_cpu(i, online_mask))
2034 hctx_idx = q->mq_map[i];
2035 /* unmapped hw queue can be remapped after CPU topo changed */
2036 if (!set->tags[hctx_idx] &&
2037 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2039 * If tags initialization fail for some hctx,
2040 * that hctx won't be brought online. In this
2041 * case, remap the current ctx to hctx[0] which
2042 * is guaranteed to always have tags allocated
2047 ctx = per_cpu_ptr(q->queue_ctx, i);
2048 hctx = blk_mq_map_queue(q, i);
2050 cpumask_set_cpu(i, hctx->cpumask);
2051 ctx->index_hw = hctx->nr_ctx;
2052 hctx->ctxs[hctx->nr_ctx++] = ctx;
2055 mutex_unlock(&q->sysfs_lock);
2057 queue_for_each_hw_ctx(q, hctx, i) {
2059 * If no software queues are mapped to this hardware queue,
2060 * disable it and free the request entries.
2062 if (!hctx->nr_ctx) {
2063 /* Never unmap queue 0. We need it as a
2064 * fallback in case of a new remap fails
2067 if (i && set->tags[i])
2068 blk_mq_free_map_and_requests(set, i);
2074 hctx->tags = set->tags[i];
2075 WARN_ON(!hctx->tags);
2078 * Set the map size to the number of mapped software queues.
2079 * This is more accurate and more efficient than looping
2080 * over all possibly mapped software queues.
2082 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2085 * Initialize batch roundrobin counts
2087 hctx->next_cpu = cpumask_first(hctx->cpumask);
2088 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2092 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2094 struct blk_mq_hw_ctx *hctx;
2097 queue_for_each_hw_ctx(q, hctx, i) {
2099 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2101 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2105 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2107 struct request_queue *q;
2109 lockdep_assert_held(&set->tag_list_lock);
2111 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2112 blk_mq_freeze_queue(q);
2113 queue_set_hctx_shared(q, shared);
2114 blk_mq_unfreeze_queue(q);
2118 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2120 struct blk_mq_tag_set *set = q->tag_set;
2122 mutex_lock(&set->tag_list_lock);
2123 list_del_rcu(&q->tag_set_list);
2124 INIT_LIST_HEAD(&q->tag_set_list);
2125 if (list_is_singular(&set->tag_list)) {
2126 /* just transitioned to unshared */
2127 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2128 /* update existing queue */
2129 blk_mq_update_tag_set_depth(set, false);
2131 mutex_unlock(&set->tag_list_lock);
2136 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2137 struct request_queue *q)
2141 mutex_lock(&set->tag_list_lock);
2143 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2144 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2145 set->flags |= BLK_MQ_F_TAG_SHARED;
2146 /* update existing queue */
2147 blk_mq_update_tag_set_depth(set, true);
2149 if (set->flags & BLK_MQ_F_TAG_SHARED)
2150 queue_set_hctx_shared(q, true);
2151 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2153 mutex_unlock(&set->tag_list_lock);
2157 * It is the actual release handler for mq, but we do it from
2158 * request queue's release handler for avoiding use-after-free
2159 * and headache because q->mq_kobj shouldn't have been introduced,
2160 * but we can't group ctx/kctx kobj without it.
2162 void blk_mq_release(struct request_queue *q)
2164 struct blk_mq_hw_ctx *hctx;
2167 /* hctx kobj stays in hctx */
2168 queue_for_each_hw_ctx(q, hctx, i) {
2171 kobject_put(&hctx->kobj);
2176 kfree(q->queue_hw_ctx);
2179 * release .mq_kobj and sw queue's kobject now because
2180 * both share lifetime with request queue.
2182 blk_mq_sysfs_deinit(q);
2184 free_percpu(q->queue_ctx);
2187 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2189 struct request_queue *uninit_q, *q;
2191 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2193 return ERR_PTR(-ENOMEM);
2195 q = blk_mq_init_allocated_queue(set, uninit_q);
2197 blk_cleanup_queue(uninit_q);
2201 EXPORT_SYMBOL(blk_mq_init_queue);
2203 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2204 struct request_queue *q)
2207 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2209 blk_mq_sysfs_unregister(q);
2210 for (i = 0; i < set->nr_hw_queues; i++) {
2216 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2217 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2222 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2229 atomic_set(&hctxs[i]->nr_active, 0);
2230 hctxs[i]->numa_node = node;
2231 hctxs[i]->queue_num = i;
2233 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2234 free_cpumask_var(hctxs[i]->cpumask);
2239 blk_mq_hctx_kobj_init(hctxs[i]);
2241 for (j = i; j < q->nr_hw_queues; j++) {
2242 struct blk_mq_hw_ctx *hctx = hctxs[j];
2246 blk_mq_free_map_and_requests(set, j);
2247 blk_mq_exit_hctx(q, set, hctx, j);
2248 kobject_put(&hctx->kobj);
2253 q->nr_hw_queues = i;
2254 blk_mq_sysfs_register(q);
2257 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2258 struct request_queue *q)
2260 /* mark the queue as mq asap */
2261 q->mq_ops = set->ops;
2263 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2264 blk_stat_rq_ddir, 2, q);
2268 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2272 /* init q->mq_kobj and sw queues' kobjects */
2273 blk_mq_sysfs_init(q);
2275 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2276 GFP_KERNEL, set->numa_node);
2277 if (!q->queue_hw_ctx)
2280 q->mq_map = set->mq_map;
2282 blk_mq_realloc_hw_ctxs(set, q);
2283 if (!q->nr_hw_queues)
2286 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2287 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2289 q->nr_queues = nr_cpu_ids;
2291 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2293 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2294 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2296 q->sg_reserved_size = INT_MAX;
2298 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2299 INIT_LIST_HEAD(&q->requeue_list);
2300 spin_lock_init(&q->requeue_lock);
2302 blk_queue_make_request(q, blk_mq_make_request);
2305 * Do this after blk_queue_make_request() overrides it...
2307 q->nr_requests = set->queue_depth;
2310 * Default to classic polling
2314 if (set->ops->complete)
2315 blk_queue_softirq_done(q, set->ops->complete);
2317 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2320 mutex_lock(&all_q_mutex);
2322 list_add_tail(&q->all_q_node, &all_q_list);
2323 blk_mq_add_queue_tag_set(set, q);
2324 blk_mq_map_swqueue(q, cpu_online_mask);
2326 mutex_unlock(&all_q_mutex);
2329 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2332 ret = blk_mq_sched_init(q);
2334 return ERR_PTR(ret);
2340 kfree(q->queue_hw_ctx);
2342 free_percpu(q->queue_ctx);
2345 return ERR_PTR(-ENOMEM);
2347 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2349 void blk_mq_free_queue(struct request_queue *q)
2351 struct blk_mq_tag_set *set = q->tag_set;
2353 mutex_lock(&all_q_mutex);
2354 list_del_init(&q->all_q_node);
2355 mutex_unlock(&all_q_mutex);
2357 blk_mq_del_queue_tag_set(q);
2359 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2362 /* Basically redo blk_mq_init_queue with queue frozen */
2363 static void blk_mq_queue_reinit(struct request_queue *q,
2364 const struct cpumask *online_mask)
2366 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2368 blk_mq_sysfs_unregister(q);
2371 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2372 * we should change hctx numa_node according to new topology (this
2373 * involves free and re-allocate memory, worthy doing?)
2376 blk_mq_map_swqueue(q, online_mask);
2378 blk_mq_sysfs_register(q);
2382 * New online cpumask which is going to be set in this hotplug event.
2383 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2384 * one-by-one and dynamically allocating this could result in a failure.
2386 static struct cpumask cpuhp_online_new;
2388 static void blk_mq_queue_reinit_work(void)
2390 struct request_queue *q;
2392 mutex_lock(&all_q_mutex);
2394 * We need to freeze and reinit all existing queues. Freezing
2395 * involves synchronous wait for an RCU grace period and doing it
2396 * one by one may take a long time. Start freezing all queues in
2397 * one swoop and then wait for the completions so that freezing can
2398 * take place in parallel.
2400 list_for_each_entry(q, &all_q_list, all_q_node)
2401 blk_freeze_queue_start(q);
2402 list_for_each_entry(q, &all_q_list, all_q_node)
2403 blk_mq_freeze_queue_wait(q);
2405 list_for_each_entry(q, &all_q_list, all_q_node)
2406 blk_mq_queue_reinit(q, &cpuhp_online_new);
2408 list_for_each_entry(q, &all_q_list, all_q_node)
2409 blk_mq_unfreeze_queue(q);
2411 mutex_unlock(&all_q_mutex);
2414 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2416 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2417 blk_mq_queue_reinit_work();
2422 * Before hotadded cpu starts handling requests, new mappings must be
2423 * established. Otherwise, these requests in hw queue might never be
2426 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2427 * for CPU0, and ctx1 for CPU1).
2429 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2430 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2432 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2433 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2434 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2437 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2439 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2440 cpumask_set_cpu(cpu, &cpuhp_online_new);
2441 blk_mq_queue_reinit_work();
2445 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2449 for (i = 0; i < set->nr_hw_queues; i++)
2450 if (!__blk_mq_alloc_rq_map(set, i))
2457 blk_mq_free_rq_map(set->tags[i]);
2463 * Allocate the request maps associated with this tag_set. Note that this
2464 * may reduce the depth asked for, if memory is tight. set->queue_depth
2465 * will be updated to reflect the allocated depth.
2467 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2472 depth = set->queue_depth;
2474 err = __blk_mq_alloc_rq_maps(set);
2478 set->queue_depth >>= 1;
2479 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2483 } while (set->queue_depth);
2485 if (!set->queue_depth || err) {
2486 pr_err("blk-mq: failed to allocate request map\n");
2490 if (depth != set->queue_depth)
2491 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2492 depth, set->queue_depth);
2497 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2499 if (set->ops->map_queues)
2500 return set->ops->map_queues(set);
2502 return blk_mq_map_queues(set);
2506 * Alloc a tag set to be associated with one or more request queues.
2507 * May fail with EINVAL for various error conditions. May adjust the
2508 * requested depth down, if if it too large. In that case, the set
2509 * value will be stored in set->queue_depth.
2511 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2515 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2517 if (!set->nr_hw_queues)
2519 if (!set->queue_depth)
2521 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2524 if (!set->ops->queue_rq)
2527 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2528 pr_info("blk-mq: reduced tag depth to %u\n",
2530 set->queue_depth = BLK_MQ_MAX_DEPTH;
2534 * If a crashdump is active, then we are potentially in a very
2535 * memory constrained environment. Limit us to 1 queue and
2536 * 64 tags to prevent using too much memory.
2538 if (is_kdump_kernel()) {
2539 set->nr_hw_queues = 1;
2540 set->queue_depth = min(64U, set->queue_depth);
2543 * There is no use for more h/w queues than cpus.
2545 if (set->nr_hw_queues > nr_cpu_ids)
2546 set->nr_hw_queues = nr_cpu_ids;
2548 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2549 GFP_KERNEL, set->numa_node);
2554 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2555 GFP_KERNEL, set->numa_node);
2559 ret = blk_mq_update_queue_map(set);
2561 goto out_free_mq_map;
2563 ret = blk_mq_alloc_rq_maps(set);
2565 goto out_free_mq_map;
2567 mutex_init(&set->tag_list_lock);
2568 INIT_LIST_HEAD(&set->tag_list);
2580 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2582 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2586 for (i = 0; i < nr_cpu_ids; i++)
2587 blk_mq_free_map_and_requests(set, i);
2595 EXPORT_SYMBOL(blk_mq_free_tag_set);
2597 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2599 struct blk_mq_tag_set *set = q->tag_set;
2600 struct blk_mq_hw_ctx *hctx;
2606 blk_mq_freeze_queue(q);
2607 blk_mq_quiesce_queue(q);
2610 queue_for_each_hw_ctx(q, hctx, i) {
2614 * If we're using an MQ scheduler, just update the scheduler
2615 * queue depth. This is similar to what the old code would do.
2617 if (!hctx->sched_tags) {
2618 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2619 min(nr, set->queue_depth),
2622 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2630 q->nr_requests = nr;
2632 blk_mq_unfreeze_queue(q);
2633 blk_mq_start_stopped_hw_queues(q, true);
2638 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2640 struct request_queue *q;
2642 lockdep_assert_held(&set->tag_list_lock);
2644 if (nr_hw_queues > nr_cpu_ids)
2645 nr_hw_queues = nr_cpu_ids;
2646 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2649 list_for_each_entry(q, &set->tag_list, tag_set_list)
2650 blk_mq_freeze_queue(q);
2652 set->nr_hw_queues = nr_hw_queues;
2653 blk_mq_update_queue_map(set);
2654 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2655 blk_mq_realloc_hw_ctxs(set, q);
2656 blk_mq_queue_reinit(q, cpu_online_mask);
2659 list_for_each_entry(q, &set->tag_list, tag_set_list)
2660 blk_mq_unfreeze_queue(q);
2662 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2664 /* Enable polling stats and return whether they were already enabled. */
2665 static bool blk_poll_stats_enable(struct request_queue *q)
2667 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2668 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2670 blk_stat_add_callback(q, q->poll_cb);
2674 static void blk_mq_poll_stats_start(struct request_queue *q)
2677 * We don't arm the callback if polling stats are not enabled or the
2678 * callback is already active.
2680 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2681 blk_stat_is_active(q->poll_cb))
2684 blk_stat_activate_msecs(q->poll_cb, 100);
2687 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2689 struct request_queue *q = cb->data;
2691 if (cb->stat[READ].nr_samples)
2692 q->poll_stat[READ] = cb->stat[READ];
2693 if (cb->stat[WRITE].nr_samples)
2694 q->poll_stat[WRITE] = cb->stat[WRITE];
2697 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2698 struct blk_mq_hw_ctx *hctx,
2701 unsigned long ret = 0;
2704 * If stats collection isn't on, don't sleep but turn it on for
2707 if (!blk_poll_stats_enable(q))
2711 * As an optimistic guess, use half of the mean service time
2712 * for this type of request. We can (and should) make this smarter.
2713 * For instance, if the completion latencies are tight, we can
2714 * get closer than just half the mean. This is especially
2715 * important on devices where the completion latencies are longer
2718 if (req_op(rq) == REQ_OP_READ && q->poll_stat[READ].nr_samples)
2719 ret = (q->poll_stat[READ].mean + 1) / 2;
2720 else if (req_op(rq) == REQ_OP_WRITE && q->poll_stat[WRITE].nr_samples)
2721 ret = (q->poll_stat[WRITE].mean + 1) / 2;
2726 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2727 struct blk_mq_hw_ctx *hctx,
2730 struct hrtimer_sleeper hs;
2731 enum hrtimer_mode mode;
2735 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2741 * -1: don't ever hybrid sleep
2742 * 0: use half of prev avg
2743 * >0: use this specific value
2745 if (q->poll_nsec == -1)
2747 else if (q->poll_nsec > 0)
2748 nsecs = q->poll_nsec;
2750 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2755 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2758 * This will be replaced with the stats tracking code, using
2759 * 'avg_completion_time / 2' as the pre-sleep target.
2763 mode = HRTIMER_MODE_REL;
2764 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2765 hrtimer_set_expires(&hs.timer, kt);
2767 hrtimer_init_sleeper(&hs, current);
2769 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2771 set_current_state(TASK_UNINTERRUPTIBLE);
2772 hrtimer_start_expires(&hs.timer, mode);
2775 hrtimer_cancel(&hs.timer);
2776 mode = HRTIMER_MODE_ABS;
2777 } while (hs.task && !signal_pending(current));
2779 __set_current_state(TASK_RUNNING);
2780 destroy_hrtimer_on_stack(&hs.timer);
2784 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2786 struct request_queue *q = hctx->queue;
2790 * If we sleep, have the caller restart the poll loop to reset
2791 * the state. Like for the other success return cases, the
2792 * caller is responsible for checking if the IO completed. If
2793 * the IO isn't complete, we'll get called again and will go
2794 * straight to the busy poll loop.
2796 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2799 hctx->poll_considered++;
2801 state = current->state;
2802 while (!need_resched()) {
2805 hctx->poll_invoked++;
2807 ret = q->mq_ops->poll(hctx, rq->tag);
2809 hctx->poll_success++;
2810 set_current_state(TASK_RUNNING);
2814 if (signal_pending_state(state, current))
2815 set_current_state(TASK_RUNNING);
2817 if (current->state == TASK_RUNNING)
2827 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2829 struct blk_mq_hw_ctx *hctx;
2830 struct blk_plug *plug;
2833 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2834 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2837 plug = current->plug;
2839 blk_flush_plug_list(plug, false);
2841 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2842 if (!blk_qc_t_is_internal(cookie))
2843 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2845 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2847 return __blk_mq_poll(hctx, rq);
2849 EXPORT_SYMBOL_GPL(blk_mq_poll);
2851 void blk_mq_disable_hotplug(void)
2853 mutex_lock(&all_q_mutex);
2856 void blk_mq_enable_hotplug(void)
2858 mutex_unlock(&all_q_mutex);
2861 static int __init blk_mq_init(void)
2863 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2864 blk_mq_hctx_notify_dead);
2866 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2867 blk_mq_queue_reinit_prepare,
2868 blk_mq_queue_reinit_dead);
2871 subsys_initcall(blk_mq_init);