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_sched_completed_request(hctx, rq);
354 blk_mq_sched_restart_queues(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);
372 void blk_mq_free_request(struct request *rq)
374 blk_mq_sched_put_request(rq);
376 EXPORT_SYMBOL_GPL(blk_mq_free_request);
378 inline void __blk_mq_end_request(struct request *rq, int error)
380 blk_account_io_done(rq);
383 wbt_done(rq->q->rq_wb, &rq->issue_stat);
384 rq->end_io(rq, error);
386 if (unlikely(blk_bidi_rq(rq)))
387 blk_mq_free_request(rq->next_rq);
388 blk_mq_free_request(rq);
391 EXPORT_SYMBOL(__blk_mq_end_request);
393 void blk_mq_end_request(struct request *rq, int error)
395 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
397 __blk_mq_end_request(rq, error);
399 EXPORT_SYMBOL(blk_mq_end_request);
401 static void __blk_mq_complete_request_remote(void *data)
403 struct request *rq = data;
405 rq->q->softirq_done_fn(rq);
408 static void blk_mq_ipi_complete_request(struct request *rq)
410 struct blk_mq_ctx *ctx = rq->mq_ctx;
414 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
415 rq->q->softirq_done_fn(rq);
420 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
421 shared = cpus_share_cache(cpu, ctx->cpu);
423 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
424 rq->csd.func = __blk_mq_complete_request_remote;
427 smp_call_function_single_async(ctx->cpu, &rq->csd);
429 rq->q->softirq_done_fn(rq);
434 static void blk_mq_stat_add(struct request *rq)
436 if (rq->rq_flags & RQF_STATS) {
437 blk_mq_poll_stats_start(rq->q);
442 static void __blk_mq_complete_request(struct request *rq)
444 struct request_queue *q = rq->q;
448 if (!q->softirq_done_fn)
449 blk_mq_end_request(rq, rq->errors);
451 blk_mq_ipi_complete_request(rq);
455 * blk_mq_complete_request - end I/O on a request
456 * @rq: the request being processed
459 * Ends all I/O on a request. It does not handle partial completions.
460 * The actual completion happens out-of-order, through a IPI handler.
462 void blk_mq_complete_request(struct request *rq, int error)
464 struct request_queue *q = rq->q;
466 if (unlikely(blk_should_fake_timeout(q)))
468 if (!blk_mark_rq_complete(rq)) {
470 __blk_mq_complete_request(rq);
473 EXPORT_SYMBOL(blk_mq_complete_request);
475 int blk_mq_request_started(struct request *rq)
477 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
479 EXPORT_SYMBOL_GPL(blk_mq_request_started);
481 void blk_mq_start_request(struct request *rq)
483 struct request_queue *q = rq->q;
485 blk_mq_sched_started_request(rq);
487 trace_block_rq_issue(q, rq);
489 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
490 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
491 rq->rq_flags |= RQF_STATS;
492 wbt_issue(q->rq_wb, &rq->issue_stat);
498 * Ensure that ->deadline is visible before set the started
499 * flag and clear the completed flag.
501 smp_mb__before_atomic();
504 * Mark us as started and clear complete. Complete might have been
505 * set if requeue raced with timeout, which then marked it as
506 * complete. So be sure to clear complete again when we start
507 * the request, otherwise we'll ignore the completion event.
509 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
510 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
511 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
512 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
514 if (q->dma_drain_size && blk_rq_bytes(rq)) {
516 * Make sure space for the drain appears. We know we can do
517 * this because max_hw_segments has been adjusted to be one
518 * fewer than the device can handle.
520 rq->nr_phys_segments++;
523 EXPORT_SYMBOL(blk_mq_start_request);
526 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
527 * flag isn't set yet, so there may be race with timeout handler,
528 * but given rq->deadline is just set in .queue_rq() under
529 * this situation, the race won't be possible in reality because
530 * rq->timeout should be set as big enough to cover the window
531 * between blk_mq_start_request() called from .queue_rq() and
532 * clearing REQ_ATOM_STARTED here.
534 static void __blk_mq_requeue_request(struct request *rq)
536 struct request_queue *q = rq->q;
538 trace_block_rq_requeue(q, rq);
539 wbt_requeue(q->rq_wb, &rq->issue_stat);
540 blk_mq_sched_requeue_request(rq);
542 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
543 if (q->dma_drain_size && blk_rq_bytes(rq))
544 rq->nr_phys_segments--;
548 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
550 __blk_mq_requeue_request(rq);
552 BUG_ON(blk_queued_rq(rq));
553 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
555 EXPORT_SYMBOL(blk_mq_requeue_request);
557 static void blk_mq_requeue_work(struct work_struct *work)
559 struct request_queue *q =
560 container_of(work, struct request_queue, requeue_work.work);
562 struct request *rq, *next;
565 spin_lock_irqsave(&q->requeue_lock, flags);
566 list_splice_init(&q->requeue_list, &rq_list);
567 spin_unlock_irqrestore(&q->requeue_lock, flags);
569 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
570 if (!(rq->rq_flags & RQF_SOFTBARRIER))
573 rq->rq_flags &= ~RQF_SOFTBARRIER;
574 list_del_init(&rq->queuelist);
575 blk_mq_sched_insert_request(rq, true, false, false, true);
578 while (!list_empty(&rq_list)) {
579 rq = list_entry(rq_list.next, struct request, queuelist);
580 list_del_init(&rq->queuelist);
581 blk_mq_sched_insert_request(rq, false, false, false, true);
584 blk_mq_run_hw_queues(q, false);
587 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
588 bool kick_requeue_list)
590 struct request_queue *q = rq->q;
594 * We abuse this flag that is otherwise used by the I/O scheduler to
595 * request head insertation from the workqueue.
597 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
599 spin_lock_irqsave(&q->requeue_lock, flags);
601 rq->rq_flags |= RQF_SOFTBARRIER;
602 list_add(&rq->queuelist, &q->requeue_list);
604 list_add_tail(&rq->queuelist, &q->requeue_list);
606 spin_unlock_irqrestore(&q->requeue_lock, flags);
608 if (kick_requeue_list)
609 blk_mq_kick_requeue_list(q);
611 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
613 void blk_mq_kick_requeue_list(struct request_queue *q)
615 kblockd_schedule_delayed_work(&q->requeue_work, 0);
617 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
619 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
622 kblockd_schedule_delayed_work(&q->requeue_work,
623 msecs_to_jiffies(msecs));
625 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
627 void blk_mq_abort_requeue_list(struct request_queue *q)
632 spin_lock_irqsave(&q->requeue_lock, flags);
633 list_splice_init(&q->requeue_list, &rq_list);
634 spin_unlock_irqrestore(&q->requeue_lock, flags);
636 while (!list_empty(&rq_list)) {
639 rq = list_first_entry(&rq_list, struct request, queuelist);
640 list_del_init(&rq->queuelist);
642 blk_mq_end_request(rq, rq->errors);
645 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
647 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
649 if (tag < tags->nr_tags) {
650 prefetch(tags->rqs[tag]);
651 return tags->rqs[tag];
656 EXPORT_SYMBOL(blk_mq_tag_to_rq);
658 struct blk_mq_timeout_data {
660 unsigned int next_set;
663 void blk_mq_rq_timed_out(struct request *req, bool reserved)
665 const struct blk_mq_ops *ops = req->q->mq_ops;
666 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
669 * We know that complete is set at this point. If STARTED isn't set
670 * anymore, then the request isn't active and the "timeout" should
671 * just be ignored. This can happen due to the bitflag ordering.
672 * Timeout first checks if STARTED is set, and if it is, assumes
673 * the request is active. But if we race with completion, then
674 * both flags will get cleared. So check here again, and ignore
675 * a timeout event with a request that isn't active.
677 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
681 ret = ops->timeout(req, reserved);
685 __blk_mq_complete_request(req);
687 case BLK_EH_RESET_TIMER:
689 blk_clear_rq_complete(req);
691 case BLK_EH_NOT_HANDLED:
694 printk(KERN_ERR "block: bad eh return: %d\n", ret);
699 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
700 struct request *rq, void *priv, bool reserved)
702 struct blk_mq_timeout_data *data = priv;
704 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
708 * The rq being checked may have been freed and reallocated
709 * out already here, we avoid this race by checking rq->deadline
710 * and REQ_ATOM_COMPLETE flag together:
712 * - if rq->deadline is observed as new value because of
713 * reusing, the rq won't be timed out because of timing.
714 * - if rq->deadline is observed as previous value,
715 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
716 * because we put a barrier between setting rq->deadline
717 * and clearing the flag in blk_mq_start_request(), so
718 * this rq won't be timed out too.
720 if (time_after_eq(jiffies, rq->deadline)) {
721 if (!blk_mark_rq_complete(rq))
722 blk_mq_rq_timed_out(rq, reserved);
723 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
724 data->next = rq->deadline;
729 static void blk_mq_timeout_work(struct work_struct *work)
731 struct request_queue *q =
732 container_of(work, struct request_queue, timeout_work);
733 struct blk_mq_timeout_data data = {
739 /* A deadlock might occur if a request is stuck requiring a
740 * timeout at the same time a queue freeze is waiting
741 * completion, since the timeout code would not be able to
742 * acquire the queue reference here.
744 * That's why we don't use blk_queue_enter here; instead, we use
745 * percpu_ref_tryget directly, because we need to be able to
746 * obtain a reference even in the short window between the queue
747 * starting to freeze, by dropping the first reference in
748 * blk_freeze_queue_start, and the moment the last request is
749 * consumed, marked by the instant q_usage_counter reaches
752 if (!percpu_ref_tryget(&q->q_usage_counter))
755 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
758 data.next = blk_rq_timeout(round_jiffies_up(data.next));
759 mod_timer(&q->timeout, data.next);
761 struct blk_mq_hw_ctx *hctx;
763 queue_for_each_hw_ctx(q, hctx, i) {
764 /* the hctx may be unmapped, so check it here */
765 if (blk_mq_hw_queue_mapped(hctx))
766 blk_mq_tag_idle(hctx);
773 * Reverse check our software queue for entries that we could potentially
774 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
775 * too much time checking for merges.
777 static bool blk_mq_attempt_merge(struct request_queue *q,
778 struct blk_mq_ctx *ctx, struct bio *bio)
783 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
789 if (!blk_rq_merge_ok(rq, bio))
792 switch (blk_try_merge(rq, bio)) {
793 case ELEVATOR_BACK_MERGE:
794 if (blk_mq_sched_allow_merge(q, rq, bio))
795 merged = bio_attempt_back_merge(q, rq, bio);
797 case ELEVATOR_FRONT_MERGE:
798 if (blk_mq_sched_allow_merge(q, rq, bio))
799 merged = bio_attempt_front_merge(q, rq, bio);
801 case ELEVATOR_DISCARD_MERGE:
802 merged = bio_attempt_discard_merge(q, rq, bio);
816 struct flush_busy_ctx_data {
817 struct blk_mq_hw_ctx *hctx;
818 struct list_head *list;
821 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
823 struct flush_busy_ctx_data *flush_data = data;
824 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
825 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
827 sbitmap_clear_bit(sb, bitnr);
828 spin_lock(&ctx->lock);
829 list_splice_tail_init(&ctx->rq_list, flush_data->list);
830 spin_unlock(&ctx->lock);
835 * Process software queues that have been marked busy, splicing them
836 * to the for-dispatch
838 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
840 struct flush_busy_ctx_data data = {
845 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
847 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
849 static inline unsigned int queued_to_index(unsigned int queued)
854 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
857 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
860 struct blk_mq_alloc_data data = {
862 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
863 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
873 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
874 data.flags |= BLK_MQ_REQ_RESERVED;
876 rq->tag = blk_mq_get_tag(&data);
878 if (blk_mq_tag_busy(data.hctx)) {
879 rq->rq_flags |= RQF_MQ_INFLIGHT;
880 atomic_inc(&data.hctx->nr_active);
882 data.hctx->tags->rqs[rq->tag] = rq;
889 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
892 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
895 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
896 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
897 atomic_dec(&hctx->nr_active);
901 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
904 if (rq->tag == -1 || rq->internal_tag == -1)
907 __blk_mq_put_driver_tag(hctx, rq);
910 static void blk_mq_put_driver_tag(struct request *rq)
912 struct blk_mq_hw_ctx *hctx;
914 if (rq->tag == -1 || rq->internal_tag == -1)
917 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
918 __blk_mq_put_driver_tag(hctx, rq);
922 * If we fail getting a driver tag because all the driver tags are already
923 * assigned and on the dispatch list, BUT the first entry does not have a
924 * tag, then we could deadlock. For that case, move entries with assigned
925 * driver tags to the front, leaving the set of tagged requests in the
926 * same order, and the untagged set in the same order.
928 static bool reorder_tags_to_front(struct list_head *list)
930 struct request *rq, *tmp, *first = NULL;
932 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
936 list_move(&rq->queuelist, list);
942 return first != NULL;
945 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
948 struct blk_mq_hw_ctx *hctx;
950 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
952 list_del(&wait->task_list);
953 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
954 blk_mq_run_hw_queue(hctx, true);
958 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
960 struct sbq_wait_state *ws;
963 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
964 * The thread which wins the race to grab this bit adds the hardware
965 * queue to the wait queue.
967 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
968 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
971 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
972 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
975 * As soon as this returns, it's no longer safe to fiddle with
976 * hctx->dispatch_wait, since a completion can wake up the wait queue
977 * and unlock the bit.
979 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
983 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list)
985 struct request_queue *q = hctx->queue;
987 int errors, queued, ret = BLK_MQ_RQ_QUEUE_OK;
990 * Now process all the entries, sending them to the driver.
993 while (!list_empty(list)) {
994 struct blk_mq_queue_data bd;
996 rq = list_first_entry(list, struct request, queuelist);
997 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
998 if (!queued && reorder_tags_to_front(list))
1002 * The initial allocation attempt failed, so we need to
1003 * rerun the hardware queue when a tag is freed.
1005 if (blk_mq_dispatch_wait_add(hctx)) {
1007 * It's possible that a tag was freed in the
1008 * window between the allocation failure and
1009 * adding the hardware queue to the wait queue.
1011 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1018 list_del_init(&rq->queuelist);
1023 * Flag last if we have no more requests, or if we have more
1024 * but can't assign a driver tag to it.
1026 if (list_empty(list))
1029 struct request *nxt;
1031 nxt = list_first_entry(list, struct request, queuelist);
1032 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1035 ret = q->mq_ops->queue_rq(hctx, &bd);
1037 case BLK_MQ_RQ_QUEUE_OK:
1040 case BLK_MQ_RQ_QUEUE_BUSY:
1041 blk_mq_put_driver_tag_hctx(hctx, rq);
1042 list_add(&rq->queuelist, list);
1043 __blk_mq_requeue_request(rq);
1046 pr_err("blk-mq: bad return on queue: %d\n", ret);
1047 case BLK_MQ_RQ_QUEUE_ERROR:
1050 blk_mq_end_request(rq, rq->errors);
1054 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
1058 hctx->dispatched[queued_to_index(queued)]++;
1061 * Any items that need requeuing? Stuff them into hctx->dispatch,
1062 * that is where we will continue on next queue run.
1064 if (!list_empty(list)) {
1066 * If we got a driver tag for the next request already,
1069 rq = list_first_entry(list, struct request, queuelist);
1070 blk_mq_put_driver_tag(rq);
1072 spin_lock(&hctx->lock);
1073 list_splice_init(list, &hctx->dispatch);
1074 spin_unlock(&hctx->lock);
1077 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1078 * it's possible the queue is stopped and restarted again
1079 * before this. Queue restart will dispatch requests. And since
1080 * requests in rq_list aren't added into hctx->dispatch yet,
1081 * the requests in rq_list might get lost.
1083 * blk_mq_run_hw_queue() already checks the STOPPED bit
1085 * If RESTART or TAG_WAITING is set, then let completion restart
1086 * the queue instead of potentially looping here.
1088 if (!blk_mq_sched_needs_restart(hctx) &&
1089 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1090 blk_mq_run_hw_queue(hctx, true);
1093 return (queued + errors) != 0;
1096 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1100 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1101 cpu_online(hctx->next_cpu));
1103 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1105 blk_mq_sched_dispatch_requests(hctx);
1110 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1111 blk_mq_sched_dispatch_requests(hctx);
1112 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1117 * It'd be great if the workqueue API had a way to pass
1118 * in a mask and had some smarts for more clever placement.
1119 * For now we just round-robin here, switching for every
1120 * BLK_MQ_CPU_WORK_BATCH queued items.
1122 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1124 if (hctx->queue->nr_hw_queues == 1)
1125 return WORK_CPU_UNBOUND;
1127 if (--hctx->next_cpu_batch <= 0) {
1130 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1131 if (next_cpu >= nr_cpu_ids)
1132 next_cpu = cpumask_first(hctx->cpumask);
1134 hctx->next_cpu = next_cpu;
1135 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1138 return hctx->next_cpu;
1141 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1143 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1144 !blk_mq_hw_queue_mapped(hctx)))
1147 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1148 int cpu = get_cpu();
1149 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1150 __blk_mq_run_hw_queue(hctx);
1158 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
1161 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1163 struct blk_mq_hw_ctx *hctx;
1166 queue_for_each_hw_ctx(q, hctx, i) {
1167 if (!blk_mq_hctx_has_pending(hctx) ||
1168 blk_mq_hctx_stopped(hctx))
1171 blk_mq_run_hw_queue(hctx, async);
1174 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1177 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1178 * @q: request queue.
1180 * The caller is responsible for serializing this function against
1181 * blk_mq_{start,stop}_hw_queue().
1183 bool blk_mq_queue_stopped(struct request_queue *q)
1185 struct blk_mq_hw_ctx *hctx;
1188 queue_for_each_hw_ctx(q, hctx, i)
1189 if (blk_mq_hctx_stopped(hctx))
1194 EXPORT_SYMBOL(blk_mq_queue_stopped);
1196 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1198 cancel_work(&hctx->run_work);
1199 cancel_delayed_work(&hctx->delay_work);
1200 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1202 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1204 void blk_mq_stop_hw_queues(struct request_queue *q)
1206 struct blk_mq_hw_ctx *hctx;
1209 queue_for_each_hw_ctx(q, hctx, i)
1210 blk_mq_stop_hw_queue(hctx);
1212 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1214 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1216 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1218 blk_mq_run_hw_queue(hctx, false);
1220 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1222 void blk_mq_start_hw_queues(struct request_queue *q)
1224 struct blk_mq_hw_ctx *hctx;
1227 queue_for_each_hw_ctx(q, hctx, i)
1228 blk_mq_start_hw_queue(hctx);
1230 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1232 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1234 if (!blk_mq_hctx_stopped(hctx))
1237 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1238 blk_mq_run_hw_queue(hctx, async);
1240 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1242 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1244 struct blk_mq_hw_ctx *hctx;
1247 queue_for_each_hw_ctx(q, hctx, i)
1248 blk_mq_start_stopped_hw_queue(hctx, async);
1250 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1252 static void blk_mq_run_work_fn(struct work_struct *work)
1254 struct blk_mq_hw_ctx *hctx;
1256 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1258 __blk_mq_run_hw_queue(hctx);
1261 static void blk_mq_delay_work_fn(struct work_struct *work)
1263 struct blk_mq_hw_ctx *hctx;
1265 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1267 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1268 __blk_mq_run_hw_queue(hctx);
1271 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1273 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1276 blk_mq_stop_hw_queue(hctx);
1277 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1278 &hctx->delay_work, msecs_to_jiffies(msecs));
1280 EXPORT_SYMBOL(blk_mq_delay_queue);
1282 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1286 struct blk_mq_ctx *ctx = rq->mq_ctx;
1288 trace_block_rq_insert(hctx->queue, rq);
1291 list_add(&rq->queuelist, &ctx->rq_list);
1293 list_add_tail(&rq->queuelist, &ctx->rq_list);
1296 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1299 struct blk_mq_ctx *ctx = rq->mq_ctx;
1301 __blk_mq_insert_req_list(hctx, rq, at_head);
1302 blk_mq_hctx_mark_pending(hctx, ctx);
1305 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1306 struct list_head *list)
1310 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1313 spin_lock(&ctx->lock);
1314 while (!list_empty(list)) {
1317 rq = list_first_entry(list, struct request, queuelist);
1318 BUG_ON(rq->mq_ctx != ctx);
1319 list_del_init(&rq->queuelist);
1320 __blk_mq_insert_req_list(hctx, rq, false);
1322 blk_mq_hctx_mark_pending(hctx, ctx);
1323 spin_unlock(&ctx->lock);
1326 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1328 struct request *rqa = container_of(a, struct request, queuelist);
1329 struct request *rqb = container_of(b, struct request, queuelist);
1331 return !(rqa->mq_ctx < rqb->mq_ctx ||
1332 (rqa->mq_ctx == rqb->mq_ctx &&
1333 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1336 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1338 struct blk_mq_ctx *this_ctx;
1339 struct request_queue *this_q;
1342 LIST_HEAD(ctx_list);
1345 list_splice_init(&plug->mq_list, &list);
1347 list_sort(NULL, &list, plug_ctx_cmp);
1353 while (!list_empty(&list)) {
1354 rq = list_entry_rq(list.next);
1355 list_del_init(&rq->queuelist);
1357 if (rq->mq_ctx != this_ctx) {
1359 trace_block_unplug(this_q, depth, from_schedule);
1360 blk_mq_sched_insert_requests(this_q, this_ctx,
1365 this_ctx = rq->mq_ctx;
1371 list_add_tail(&rq->queuelist, &ctx_list);
1375 * If 'this_ctx' is set, we know we have entries to complete
1376 * on 'ctx_list'. Do those.
1379 trace_block_unplug(this_q, depth, from_schedule);
1380 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1385 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1387 init_request_from_bio(rq, bio);
1389 blk_account_io_start(rq, true);
1392 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1394 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1395 !blk_queue_nomerges(hctx->queue);
1398 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1399 struct blk_mq_ctx *ctx,
1400 struct request *rq, struct bio *bio)
1402 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1403 blk_mq_bio_to_request(rq, bio);
1404 spin_lock(&ctx->lock);
1406 __blk_mq_insert_request(hctx, rq, false);
1407 spin_unlock(&ctx->lock);
1410 struct request_queue *q = hctx->queue;
1412 spin_lock(&ctx->lock);
1413 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1414 blk_mq_bio_to_request(rq, bio);
1418 spin_unlock(&ctx->lock);
1419 __blk_mq_finish_request(hctx, ctx, rq);
1424 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1427 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1429 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1432 static void __blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie,
1435 struct request_queue *q = rq->q;
1436 struct blk_mq_queue_data bd = {
1440 struct blk_mq_hw_ctx *hctx;
1441 blk_qc_t new_cookie;
1447 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1450 new_cookie = request_to_qc_t(hctx, rq);
1453 * For OK queue, we are done. For error, kill it. Any other
1454 * error (busy), just add it to our list as we previously
1457 ret = q->mq_ops->queue_rq(hctx, &bd);
1458 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1459 *cookie = new_cookie;
1463 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1464 *cookie = BLK_QC_T_NONE;
1466 blk_mq_end_request(rq, rq->errors);
1470 __blk_mq_requeue_request(rq);
1472 blk_mq_sched_insert_request(rq, false, true, false, may_sleep);
1475 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1476 struct request *rq, blk_qc_t *cookie)
1478 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1480 __blk_mq_try_issue_directly(rq, cookie, false);
1483 unsigned int srcu_idx;
1487 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1488 __blk_mq_try_issue_directly(rq, cookie, true);
1489 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1493 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1495 const int is_sync = op_is_sync(bio->bi_opf);
1496 const int is_flush_fua = op_is_flush(bio->bi_opf);
1497 struct blk_mq_alloc_data data = { .flags = 0 };
1499 unsigned int request_count = 0;
1500 struct blk_plug *plug;
1501 struct request *same_queue_rq = NULL;
1503 unsigned int wb_acct;
1505 blk_queue_bounce(q, &bio);
1507 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1509 return BLK_QC_T_NONE;
1512 blk_queue_split(q, &bio, q->bio_split);
1514 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1515 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1516 return BLK_QC_T_NONE;
1518 if (blk_mq_sched_bio_merge(q, bio))
1519 return BLK_QC_T_NONE;
1521 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1523 trace_block_getrq(q, bio, bio->bi_opf);
1525 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1526 if (unlikely(!rq)) {
1527 __wbt_done(q->rq_wb, wb_acct);
1528 return BLK_QC_T_NONE;
1531 wbt_track(&rq->issue_stat, wb_acct);
1533 cookie = request_to_qc_t(data.hctx, rq);
1535 plug = current->plug;
1536 if (unlikely(is_flush_fua)) {
1537 blk_mq_bio_to_request(rq, bio);
1539 blk_mq_sched_insert_request(rq, false, true, true,
1542 blk_insert_flush(rq);
1543 blk_mq_run_hw_queue(data.hctx, true);
1545 } else if (plug && q->nr_hw_queues == 1) {
1546 struct request *last = NULL;
1548 blk_mq_bio_to_request(rq, bio);
1551 * @request_count may become stale because of schedule
1552 * out, so check the list again.
1554 if (list_empty(&plug->mq_list))
1556 else if (blk_queue_nomerges(q))
1557 request_count = blk_plug_queued_count(q);
1560 trace_block_plug(q);
1562 last = list_entry_rq(plug->mq_list.prev);
1564 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1565 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1566 blk_flush_plug_list(plug, false);
1567 trace_block_plug(q);
1570 list_add_tail(&rq->queuelist, &plug->mq_list);
1571 } else if (plug && !blk_queue_nomerges(q)) {
1572 blk_mq_bio_to_request(rq, bio);
1575 * We do limited plugging. If the bio can be merged, do that.
1576 * Otherwise the existing request in the plug list will be
1577 * issued. So the plug list will have one request at most
1578 * The plug list might get flushed before this. If that happens,
1579 * the plug list is empty, and same_queue_rq is invalid.
1581 if (list_empty(&plug->mq_list))
1582 same_queue_rq = NULL;
1584 list_del_init(&same_queue_rq->queuelist);
1585 list_add_tail(&rq->queuelist, &plug->mq_list);
1587 blk_mq_put_ctx(data.ctx);
1590 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1594 } else if (q->nr_hw_queues > 1 && is_sync) {
1595 blk_mq_put_ctx(data.ctx);
1596 blk_mq_bio_to_request(rq, bio);
1597 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1599 } else if (q->elevator) {
1600 blk_mq_bio_to_request(rq, bio);
1601 blk_mq_sched_insert_request(rq, false, true, true, true);
1602 } else if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio))
1603 blk_mq_run_hw_queue(data.hctx, true);
1605 blk_mq_put_ctx(data.ctx);
1609 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1610 unsigned int hctx_idx)
1614 if (tags->rqs && set->ops->exit_request) {
1617 for (i = 0; i < tags->nr_tags; i++) {
1618 struct request *rq = tags->static_rqs[i];
1622 set->ops->exit_request(set->driver_data, rq,
1624 tags->static_rqs[i] = NULL;
1628 while (!list_empty(&tags->page_list)) {
1629 page = list_first_entry(&tags->page_list, struct page, lru);
1630 list_del_init(&page->lru);
1632 * Remove kmemleak object previously allocated in
1633 * blk_mq_init_rq_map().
1635 kmemleak_free(page_address(page));
1636 __free_pages(page, page->private);
1640 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1644 kfree(tags->static_rqs);
1645 tags->static_rqs = NULL;
1647 blk_mq_free_tags(tags);
1650 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1651 unsigned int hctx_idx,
1652 unsigned int nr_tags,
1653 unsigned int reserved_tags)
1655 struct blk_mq_tags *tags;
1658 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1659 if (node == NUMA_NO_NODE)
1660 node = set->numa_node;
1662 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1663 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1667 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1668 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1671 blk_mq_free_tags(tags);
1675 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1676 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1678 if (!tags->static_rqs) {
1680 blk_mq_free_tags(tags);
1687 static size_t order_to_size(unsigned int order)
1689 return (size_t)PAGE_SIZE << order;
1692 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1693 unsigned int hctx_idx, unsigned int depth)
1695 unsigned int i, j, entries_per_page, max_order = 4;
1696 size_t rq_size, left;
1699 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1700 if (node == NUMA_NO_NODE)
1701 node = set->numa_node;
1703 INIT_LIST_HEAD(&tags->page_list);
1706 * rq_size is the size of the request plus driver payload, rounded
1707 * to the cacheline size
1709 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1711 left = rq_size * depth;
1713 for (i = 0; i < depth; ) {
1714 int this_order = max_order;
1719 while (this_order && left < order_to_size(this_order - 1))
1723 page = alloc_pages_node(node,
1724 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1730 if (order_to_size(this_order) < rq_size)
1737 page->private = this_order;
1738 list_add_tail(&page->lru, &tags->page_list);
1740 p = page_address(page);
1742 * Allow kmemleak to scan these pages as they contain pointers
1743 * to additional allocations like via ops->init_request().
1745 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1746 entries_per_page = order_to_size(this_order) / rq_size;
1747 to_do = min(entries_per_page, depth - i);
1748 left -= to_do * rq_size;
1749 for (j = 0; j < to_do; j++) {
1750 struct request *rq = p;
1752 tags->static_rqs[i] = rq;
1753 if (set->ops->init_request) {
1754 if (set->ops->init_request(set->driver_data,
1757 tags->static_rqs[i] = NULL;
1769 blk_mq_free_rqs(set, tags, hctx_idx);
1774 * 'cpu' is going away. splice any existing rq_list entries from this
1775 * software queue to the hw queue dispatch list, and ensure that it
1778 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1780 struct blk_mq_hw_ctx *hctx;
1781 struct blk_mq_ctx *ctx;
1784 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1785 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1787 spin_lock(&ctx->lock);
1788 if (!list_empty(&ctx->rq_list)) {
1789 list_splice_init(&ctx->rq_list, &tmp);
1790 blk_mq_hctx_clear_pending(hctx, ctx);
1792 spin_unlock(&ctx->lock);
1794 if (list_empty(&tmp))
1797 spin_lock(&hctx->lock);
1798 list_splice_tail_init(&tmp, &hctx->dispatch);
1799 spin_unlock(&hctx->lock);
1801 blk_mq_run_hw_queue(hctx, true);
1805 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1807 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1811 /* hctx->ctxs will be freed in queue's release handler */
1812 static void blk_mq_exit_hctx(struct request_queue *q,
1813 struct blk_mq_tag_set *set,
1814 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1816 unsigned flush_start_tag = set->queue_depth;
1818 blk_mq_tag_idle(hctx);
1820 if (set->ops->exit_request)
1821 set->ops->exit_request(set->driver_data,
1822 hctx->fq->flush_rq, hctx_idx,
1823 flush_start_tag + hctx_idx);
1825 if (set->ops->exit_hctx)
1826 set->ops->exit_hctx(hctx, hctx_idx);
1828 if (hctx->flags & BLK_MQ_F_BLOCKING)
1829 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1831 blk_mq_remove_cpuhp(hctx);
1832 blk_free_flush_queue(hctx->fq);
1833 sbitmap_free(&hctx->ctx_map);
1836 static void blk_mq_exit_hw_queues(struct request_queue *q,
1837 struct blk_mq_tag_set *set, int nr_queue)
1839 struct blk_mq_hw_ctx *hctx;
1842 queue_for_each_hw_ctx(q, hctx, i) {
1845 blk_mq_exit_hctx(q, set, hctx, i);
1849 static int blk_mq_init_hctx(struct request_queue *q,
1850 struct blk_mq_tag_set *set,
1851 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1854 unsigned flush_start_tag = set->queue_depth;
1856 node = hctx->numa_node;
1857 if (node == NUMA_NO_NODE)
1858 node = hctx->numa_node = set->numa_node;
1860 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1861 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1862 spin_lock_init(&hctx->lock);
1863 INIT_LIST_HEAD(&hctx->dispatch);
1865 hctx->queue_num = hctx_idx;
1866 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1868 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1870 hctx->tags = set->tags[hctx_idx];
1873 * Allocate space for all possible cpus to avoid allocation at
1876 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1879 goto unregister_cpu_notifier;
1881 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1887 if (set->ops->init_hctx &&
1888 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1891 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1895 if (set->ops->init_request &&
1896 set->ops->init_request(set->driver_data,
1897 hctx->fq->flush_rq, hctx_idx,
1898 flush_start_tag + hctx_idx, node))
1901 if (hctx->flags & BLK_MQ_F_BLOCKING)
1902 init_srcu_struct(&hctx->queue_rq_srcu);
1909 if (set->ops->exit_hctx)
1910 set->ops->exit_hctx(hctx, hctx_idx);
1912 sbitmap_free(&hctx->ctx_map);
1915 unregister_cpu_notifier:
1916 blk_mq_remove_cpuhp(hctx);
1920 static void blk_mq_init_cpu_queues(struct request_queue *q,
1921 unsigned int nr_hw_queues)
1925 for_each_possible_cpu(i) {
1926 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1927 struct blk_mq_hw_ctx *hctx;
1930 spin_lock_init(&__ctx->lock);
1931 INIT_LIST_HEAD(&__ctx->rq_list);
1934 /* If the cpu isn't online, the cpu is mapped to first hctx */
1938 hctx = blk_mq_map_queue(q, i);
1941 * Set local node, IFF we have more than one hw queue. If
1942 * not, we remain on the home node of the device
1944 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1945 hctx->numa_node = local_memory_node(cpu_to_node(i));
1949 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1953 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1954 set->queue_depth, set->reserved_tags);
1955 if (!set->tags[hctx_idx])
1958 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
1963 blk_mq_free_rq_map(set->tags[hctx_idx]);
1964 set->tags[hctx_idx] = NULL;
1968 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
1969 unsigned int hctx_idx)
1971 if (set->tags[hctx_idx]) {
1972 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
1973 blk_mq_free_rq_map(set->tags[hctx_idx]);
1974 set->tags[hctx_idx] = NULL;
1978 static void blk_mq_map_swqueue(struct request_queue *q,
1979 const struct cpumask *online_mask)
1981 unsigned int i, hctx_idx;
1982 struct blk_mq_hw_ctx *hctx;
1983 struct blk_mq_ctx *ctx;
1984 struct blk_mq_tag_set *set = q->tag_set;
1987 * Avoid others reading imcomplete hctx->cpumask through sysfs
1989 mutex_lock(&q->sysfs_lock);
1991 queue_for_each_hw_ctx(q, hctx, i) {
1992 cpumask_clear(hctx->cpumask);
1997 * Map software to hardware queues
1999 for_each_possible_cpu(i) {
2000 /* If the cpu isn't online, the cpu is mapped to first hctx */
2001 if (!cpumask_test_cpu(i, online_mask))
2004 hctx_idx = q->mq_map[i];
2005 /* unmapped hw queue can be remapped after CPU topo changed */
2006 if (!set->tags[hctx_idx] &&
2007 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2009 * If tags initialization fail for some hctx,
2010 * that hctx won't be brought online. In this
2011 * case, remap the current ctx to hctx[0] which
2012 * is guaranteed to always have tags allocated
2017 ctx = per_cpu_ptr(q->queue_ctx, i);
2018 hctx = blk_mq_map_queue(q, i);
2020 cpumask_set_cpu(i, hctx->cpumask);
2021 ctx->index_hw = hctx->nr_ctx;
2022 hctx->ctxs[hctx->nr_ctx++] = ctx;
2025 mutex_unlock(&q->sysfs_lock);
2027 queue_for_each_hw_ctx(q, hctx, i) {
2029 * If no software queues are mapped to this hardware queue,
2030 * disable it and free the request entries.
2032 if (!hctx->nr_ctx) {
2033 /* Never unmap queue 0. We need it as a
2034 * fallback in case of a new remap fails
2037 if (i && set->tags[i])
2038 blk_mq_free_map_and_requests(set, i);
2044 hctx->tags = set->tags[i];
2045 WARN_ON(!hctx->tags);
2048 * Set the map size to the number of mapped software queues.
2049 * This is more accurate and more efficient than looping
2050 * over all possibly mapped software queues.
2052 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2055 * Initialize batch roundrobin counts
2057 hctx->next_cpu = cpumask_first(hctx->cpumask);
2058 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2062 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2064 struct blk_mq_hw_ctx *hctx;
2067 queue_for_each_hw_ctx(q, hctx, i) {
2069 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2071 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2075 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2077 struct request_queue *q;
2079 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2080 blk_mq_freeze_queue(q);
2081 queue_set_hctx_shared(q, shared);
2082 blk_mq_unfreeze_queue(q);
2086 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2088 struct blk_mq_tag_set *set = q->tag_set;
2090 mutex_lock(&set->tag_list_lock);
2091 list_del_init(&q->tag_set_list);
2092 if (list_is_singular(&set->tag_list)) {
2093 /* just transitioned to unshared */
2094 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2095 /* update existing queue */
2096 blk_mq_update_tag_set_depth(set, false);
2098 mutex_unlock(&set->tag_list_lock);
2101 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2102 struct request_queue *q)
2106 mutex_lock(&set->tag_list_lock);
2108 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2109 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2110 set->flags |= BLK_MQ_F_TAG_SHARED;
2111 /* update existing queue */
2112 blk_mq_update_tag_set_depth(set, true);
2114 if (set->flags & BLK_MQ_F_TAG_SHARED)
2115 queue_set_hctx_shared(q, true);
2116 list_add_tail(&q->tag_set_list, &set->tag_list);
2118 mutex_unlock(&set->tag_list_lock);
2122 * It is the actual release handler for mq, but we do it from
2123 * request queue's release handler for avoiding use-after-free
2124 * and headache because q->mq_kobj shouldn't have been introduced,
2125 * but we can't group ctx/kctx kobj without it.
2127 void blk_mq_release(struct request_queue *q)
2129 struct blk_mq_hw_ctx *hctx;
2132 blk_mq_sched_teardown(q);
2134 /* hctx kobj stays in hctx */
2135 queue_for_each_hw_ctx(q, hctx, i) {
2138 kobject_put(&hctx->kobj);
2143 kfree(q->queue_hw_ctx);
2146 * release .mq_kobj and sw queue's kobject now because
2147 * both share lifetime with request queue.
2149 blk_mq_sysfs_deinit(q);
2151 free_percpu(q->queue_ctx);
2154 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2156 struct request_queue *uninit_q, *q;
2158 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2160 return ERR_PTR(-ENOMEM);
2162 q = blk_mq_init_allocated_queue(set, uninit_q);
2164 blk_cleanup_queue(uninit_q);
2168 EXPORT_SYMBOL(blk_mq_init_queue);
2170 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2171 struct request_queue *q)
2174 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2176 blk_mq_sysfs_unregister(q);
2177 for (i = 0; i < set->nr_hw_queues; i++) {
2183 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2184 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2189 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2196 atomic_set(&hctxs[i]->nr_active, 0);
2197 hctxs[i]->numa_node = node;
2198 hctxs[i]->queue_num = i;
2200 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2201 free_cpumask_var(hctxs[i]->cpumask);
2206 blk_mq_hctx_kobj_init(hctxs[i]);
2208 for (j = i; j < q->nr_hw_queues; j++) {
2209 struct blk_mq_hw_ctx *hctx = hctxs[j];
2213 blk_mq_free_map_and_requests(set, j);
2214 blk_mq_exit_hctx(q, set, hctx, j);
2215 kobject_put(&hctx->kobj);
2220 q->nr_hw_queues = i;
2221 blk_mq_sysfs_register(q);
2224 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2225 struct request_queue *q)
2227 /* mark the queue as mq asap */
2228 q->mq_ops = set->ops;
2230 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2231 blk_stat_rq_ddir, 2, q);
2235 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2239 /* init q->mq_kobj and sw queues' kobjects */
2240 blk_mq_sysfs_init(q);
2242 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2243 GFP_KERNEL, set->numa_node);
2244 if (!q->queue_hw_ctx)
2247 q->mq_map = set->mq_map;
2249 blk_mq_realloc_hw_ctxs(set, q);
2250 if (!q->nr_hw_queues)
2253 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2254 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2256 q->nr_queues = nr_cpu_ids;
2258 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2260 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2261 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2263 q->sg_reserved_size = INT_MAX;
2265 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2266 INIT_LIST_HEAD(&q->requeue_list);
2267 spin_lock_init(&q->requeue_lock);
2269 blk_queue_make_request(q, blk_mq_make_request);
2272 * Do this after blk_queue_make_request() overrides it...
2274 q->nr_requests = set->queue_depth;
2277 * Default to classic polling
2281 if (set->ops->complete)
2282 blk_queue_softirq_done(q, set->ops->complete);
2284 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2287 mutex_lock(&all_q_mutex);
2289 list_add_tail(&q->all_q_node, &all_q_list);
2290 blk_mq_add_queue_tag_set(set, q);
2291 blk_mq_map_swqueue(q, cpu_online_mask);
2293 mutex_unlock(&all_q_mutex);
2296 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2299 ret = blk_mq_sched_init(q);
2301 return ERR_PTR(ret);
2307 kfree(q->queue_hw_ctx);
2309 free_percpu(q->queue_ctx);
2312 return ERR_PTR(-ENOMEM);
2314 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2316 void blk_mq_free_queue(struct request_queue *q)
2318 struct blk_mq_tag_set *set = q->tag_set;
2320 mutex_lock(&all_q_mutex);
2321 list_del_init(&q->all_q_node);
2322 mutex_unlock(&all_q_mutex);
2324 blk_mq_del_queue_tag_set(q);
2326 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2329 /* Basically redo blk_mq_init_queue with queue frozen */
2330 static void blk_mq_queue_reinit(struct request_queue *q,
2331 const struct cpumask *online_mask)
2333 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2335 blk_mq_sysfs_unregister(q);
2338 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2339 * we should change hctx numa_node according to new topology (this
2340 * involves free and re-allocate memory, worthy doing?)
2343 blk_mq_map_swqueue(q, online_mask);
2345 blk_mq_sysfs_register(q);
2349 * New online cpumask which is going to be set in this hotplug event.
2350 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2351 * one-by-one and dynamically allocating this could result in a failure.
2353 static struct cpumask cpuhp_online_new;
2355 static void blk_mq_queue_reinit_work(void)
2357 struct request_queue *q;
2359 mutex_lock(&all_q_mutex);
2361 * We need to freeze and reinit all existing queues. Freezing
2362 * involves synchronous wait for an RCU grace period and doing it
2363 * one by one may take a long time. Start freezing all queues in
2364 * one swoop and then wait for the completions so that freezing can
2365 * take place in parallel.
2367 list_for_each_entry(q, &all_q_list, all_q_node)
2368 blk_freeze_queue_start(q);
2369 list_for_each_entry(q, &all_q_list, all_q_node)
2370 blk_mq_freeze_queue_wait(q);
2372 list_for_each_entry(q, &all_q_list, all_q_node)
2373 blk_mq_queue_reinit(q, &cpuhp_online_new);
2375 list_for_each_entry(q, &all_q_list, all_q_node)
2376 blk_mq_unfreeze_queue(q);
2378 mutex_unlock(&all_q_mutex);
2381 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2383 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2384 blk_mq_queue_reinit_work();
2389 * Before hotadded cpu starts handling requests, new mappings must be
2390 * established. Otherwise, these requests in hw queue might never be
2393 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2394 * for CPU0, and ctx1 for CPU1).
2396 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2397 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2399 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2400 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2401 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2404 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2406 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2407 cpumask_set_cpu(cpu, &cpuhp_online_new);
2408 blk_mq_queue_reinit_work();
2412 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2416 for (i = 0; i < set->nr_hw_queues; i++)
2417 if (!__blk_mq_alloc_rq_map(set, i))
2424 blk_mq_free_rq_map(set->tags[i]);
2430 * Allocate the request maps associated with this tag_set. Note that this
2431 * may reduce the depth asked for, if memory is tight. set->queue_depth
2432 * will be updated to reflect the allocated depth.
2434 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2439 depth = set->queue_depth;
2441 err = __blk_mq_alloc_rq_maps(set);
2445 set->queue_depth >>= 1;
2446 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2450 } while (set->queue_depth);
2452 if (!set->queue_depth || err) {
2453 pr_err("blk-mq: failed to allocate request map\n");
2457 if (depth != set->queue_depth)
2458 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2459 depth, set->queue_depth);
2465 * Alloc a tag set to be associated with one or more request queues.
2466 * May fail with EINVAL for various error conditions. May adjust the
2467 * requested depth down, if if it too large. In that case, the set
2468 * value will be stored in set->queue_depth.
2470 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2474 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2476 if (!set->nr_hw_queues)
2478 if (!set->queue_depth)
2480 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2483 if (!set->ops->queue_rq)
2486 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2487 pr_info("blk-mq: reduced tag depth to %u\n",
2489 set->queue_depth = BLK_MQ_MAX_DEPTH;
2493 * If a crashdump is active, then we are potentially in a very
2494 * memory constrained environment. Limit us to 1 queue and
2495 * 64 tags to prevent using too much memory.
2497 if (is_kdump_kernel()) {
2498 set->nr_hw_queues = 1;
2499 set->queue_depth = min(64U, set->queue_depth);
2502 * There is no use for more h/w queues than cpus.
2504 if (set->nr_hw_queues > nr_cpu_ids)
2505 set->nr_hw_queues = nr_cpu_ids;
2507 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2508 GFP_KERNEL, set->numa_node);
2513 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2514 GFP_KERNEL, set->numa_node);
2518 if (set->ops->map_queues)
2519 ret = set->ops->map_queues(set);
2521 ret = blk_mq_map_queues(set);
2523 goto out_free_mq_map;
2525 ret = blk_mq_alloc_rq_maps(set);
2527 goto out_free_mq_map;
2529 mutex_init(&set->tag_list_lock);
2530 INIT_LIST_HEAD(&set->tag_list);
2542 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2544 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2548 for (i = 0; i < nr_cpu_ids; i++)
2549 blk_mq_free_map_and_requests(set, i);
2557 EXPORT_SYMBOL(blk_mq_free_tag_set);
2559 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2561 struct blk_mq_tag_set *set = q->tag_set;
2562 struct blk_mq_hw_ctx *hctx;
2568 blk_mq_freeze_queue(q);
2569 blk_mq_quiesce_queue(q);
2572 queue_for_each_hw_ctx(q, hctx, i) {
2576 * If we're using an MQ scheduler, just update the scheduler
2577 * queue depth. This is similar to what the old code would do.
2579 if (!hctx->sched_tags) {
2580 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2581 min(nr, set->queue_depth),
2584 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2592 q->nr_requests = nr;
2594 blk_mq_unfreeze_queue(q);
2595 blk_mq_start_stopped_hw_queues(q, true);
2600 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2602 struct request_queue *q;
2604 if (nr_hw_queues > nr_cpu_ids)
2605 nr_hw_queues = nr_cpu_ids;
2606 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2609 list_for_each_entry(q, &set->tag_list, tag_set_list)
2610 blk_mq_freeze_queue(q);
2612 set->nr_hw_queues = nr_hw_queues;
2613 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2614 blk_mq_realloc_hw_ctxs(set, q);
2615 blk_mq_queue_reinit(q, cpu_online_mask);
2618 list_for_each_entry(q, &set->tag_list, tag_set_list)
2619 blk_mq_unfreeze_queue(q);
2621 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2623 /* Enable polling stats and return whether they were already enabled. */
2624 static bool blk_poll_stats_enable(struct request_queue *q)
2626 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2627 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2629 blk_stat_add_callback(q, q->poll_cb);
2633 static void blk_mq_poll_stats_start(struct request_queue *q)
2636 * We don't arm the callback if polling stats are not enabled or the
2637 * callback is already active.
2639 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2640 blk_stat_is_active(q->poll_cb))
2643 blk_stat_activate_msecs(q->poll_cb, 100);
2646 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2648 struct request_queue *q = cb->data;
2650 if (cb->stat[READ].nr_samples)
2651 q->poll_stat[READ] = cb->stat[READ];
2652 if (cb->stat[WRITE].nr_samples)
2653 q->poll_stat[WRITE] = cb->stat[WRITE];
2656 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2657 struct blk_mq_hw_ctx *hctx,
2660 unsigned long ret = 0;
2663 * If stats collection isn't on, don't sleep but turn it on for
2666 if (!blk_poll_stats_enable(q))
2670 * As an optimistic guess, use half of the mean service time
2671 * for this type of request. We can (and should) make this smarter.
2672 * For instance, if the completion latencies are tight, we can
2673 * get closer than just half the mean. This is especially
2674 * important on devices where the completion latencies are longer
2677 if (req_op(rq) == REQ_OP_READ && q->poll_stat[READ].nr_samples)
2678 ret = (q->poll_stat[READ].mean + 1) / 2;
2679 else if (req_op(rq) == REQ_OP_WRITE && q->poll_stat[WRITE].nr_samples)
2680 ret = (q->poll_stat[WRITE].mean + 1) / 2;
2685 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2686 struct blk_mq_hw_ctx *hctx,
2689 struct hrtimer_sleeper hs;
2690 enum hrtimer_mode mode;
2694 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2700 * -1: don't ever hybrid sleep
2701 * 0: use half of prev avg
2702 * >0: use this specific value
2704 if (q->poll_nsec == -1)
2706 else if (q->poll_nsec > 0)
2707 nsecs = q->poll_nsec;
2709 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2714 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2717 * This will be replaced with the stats tracking code, using
2718 * 'avg_completion_time / 2' as the pre-sleep target.
2722 mode = HRTIMER_MODE_REL;
2723 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2724 hrtimer_set_expires(&hs.timer, kt);
2726 hrtimer_init_sleeper(&hs, current);
2728 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2730 set_current_state(TASK_UNINTERRUPTIBLE);
2731 hrtimer_start_expires(&hs.timer, mode);
2734 hrtimer_cancel(&hs.timer);
2735 mode = HRTIMER_MODE_ABS;
2736 } while (hs.task && !signal_pending(current));
2738 __set_current_state(TASK_RUNNING);
2739 destroy_hrtimer_on_stack(&hs.timer);
2743 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2745 struct request_queue *q = hctx->queue;
2749 * If we sleep, have the caller restart the poll loop to reset
2750 * the state. Like for the other success return cases, the
2751 * caller is responsible for checking if the IO completed. If
2752 * the IO isn't complete, we'll get called again and will go
2753 * straight to the busy poll loop.
2755 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2758 hctx->poll_considered++;
2760 state = current->state;
2761 while (!need_resched()) {
2764 hctx->poll_invoked++;
2766 ret = q->mq_ops->poll(hctx, rq->tag);
2768 hctx->poll_success++;
2769 set_current_state(TASK_RUNNING);
2773 if (signal_pending_state(state, current))
2774 set_current_state(TASK_RUNNING);
2776 if (current->state == TASK_RUNNING)
2786 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2788 struct blk_mq_hw_ctx *hctx;
2789 struct blk_plug *plug;
2792 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2793 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2796 plug = current->plug;
2798 blk_flush_plug_list(plug, false);
2800 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2801 if (!blk_qc_t_is_internal(cookie))
2802 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2804 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2806 return __blk_mq_poll(hctx, rq);
2808 EXPORT_SYMBOL_GPL(blk_mq_poll);
2810 void blk_mq_disable_hotplug(void)
2812 mutex_lock(&all_q_mutex);
2815 void blk_mq_enable_hotplug(void)
2817 mutex_unlock(&all_q_mutex);
2820 static int __init blk_mq_init(void)
2822 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2823 blk_mq_hctx_notify_dead);
2825 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2826 blk_mq_queue_reinit_prepare,
2827 blk_mq_queue_reinit_dead);
2830 subsys_initcall(blk_mq_init);