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);
43 * Check if any of the ctx's have pending work in this hardware queue
45 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
47 return sbitmap_any_bit_set(&hctx->ctx_map) ||
48 !list_empty_careful(&hctx->dispatch) ||
49 blk_mq_sched_has_work(hctx);
53 * Mark this ctx as having pending work in this hardware queue
55 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
56 struct blk_mq_ctx *ctx)
58 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
59 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
62 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
63 struct blk_mq_ctx *ctx)
65 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
68 void blk_mq_freeze_queue_start(struct request_queue *q)
72 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
73 if (freeze_depth == 1) {
74 percpu_ref_kill(&q->q_usage_counter);
75 blk_mq_run_hw_queues(q, false);
78 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
80 void blk_mq_freeze_queue_wait(struct request_queue *q)
82 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
84 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
86 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
87 unsigned long timeout)
89 return wait_event_timeout(q->mq_freeze_wq,
90 percpu_ref_is_zero(&q->q_usage_counter),
93 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
96 * Guarantee no request is in use, so we can change any data structure of
97 * the queue afterward.
99 void blk_freeze_queue(struct request_queue *q)
102 * In the !blk_mq case we are only calling this to kill the
103 * q_usage_counter, otherwise this increases the freeze depth
104 * and waits for it to return to zero. For this reason there is
105 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
106 * exported to drivers as the only user for unfreeze is blk_mq.
108 blk_mq_freeze_queue_start(q);
109 blk_mq_freeze_queue_wait(q);
112 void blk_mq_freeze_queue(struct request_queue *q)
115 * ...just an alias to keep freeze and unfreeze actions balanced
116 * in the blk_mq_* namespace
120 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
122 void blk_mq_unfreeze_queue(struct request_queue *q)
126 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
127 WARN_ON_ONCE(freeze_depth < 0);
129 percpu_ref_reinit(&q->q_usage_counter);
130 wake_up_all(&q->mq_freeze_wq);
133 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
136 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
139 * Note: this function does not prevent that the struct request end_io()
140 * callback function is invoked. Additionally, it is not prevented that
141 * new queue_rq() calls occur unless the queue has been stopped first.
143 void blk_mq_quiesce_queue(struct request_queue *q)
145 struct blk_mq_hw_ctx *hctx;
149 blk_mq_stop_hw_queues(q);
151 queue_for_each_hw_ctx(q, hctx, i) {
152 if (hctx->flags & BLK_MQ_F_BLOCKING)
153 synchronize_srcu(&hctx->queue_rq_srcu);
160 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
162 void blk_mq_wake_waiters(struct request_queue *q)
164 struct blk_mq_hw_ctx *hctx;
167 queue_for_each_hw_ctx(q, hctx, i)
168 if (blk_mq_hw_queue_mapped(hctx))
169 blk_mq_tag_wakeup_all(hctx->tags, true);
172 * If we are called because the queue has now been marked as
173 * dying, we need to ensure that processes currently waiting on
174 * the queue are notified as well.
176 wake_up_all(&q->mq_freeze_wq);
179 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
181 return blk_mq_has_free_tags(hctx->tags);
183 EXPORT_SYMBOL(blk_mq_can_queue);
185 void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
186 struct request *rq, unsigned int op)
188 INIT_LIST_HEAD(&rq->queuelist);
189 /* csd/requeue_work/fifo_time is initialized before use */
193 if (blk_queue_io_stat(q))
194 rq->rq_flags |= RQF_IO_STAT;
195 /* do not touch atomic flags, it needs atomic ops against the timer */
197 INIT_HLIST_NODE(&rq->hash);
198 RB_CLEAR_NODE(&rq->rb_node);
201 rq->start_time = jiffies;
202 #ifdef CONFIG_BLK_CGROUP
204 set_start_time_ns(rq);
205 rq->io_start_time_ns = 0;
207 rq->nr_phys_segments = 0;
208 #if defined(CONFIG_BLK_DEV_INTEGRITY)
209 rq->nr_integrity_segments = 0;
212 /* tag was already set */
216 INIT_LIST_HEAD(&rq->timeout_list);
220 rq->end_io_data = NULL;
223 ctx->rq_dispatched[op_is_sync(op)]++;
225 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init);
227 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
233 tag = blk_mq_get_tag(data);
234 if (tag != BLK_MQ_TAG_FAIL) {
235 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
237 rq = tags->static_rqs[tag];
239 if (data->flags & BLK_MQ_REQ_INTERNAL) {
241 rq->internal_tag = tag;
243 if (blk_mq_tag_busy(data->hctx)) {
244 rq->rq_flags = RQF_MQ_INFLIGHT;
245 atomic_inc(&data->hctx->nr_active);
248 rq->internal_tag = -1;
249 data->hctx->tags->rqs[rq->tag] = rq;
252 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
258 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
260 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
263 struct blk_mq_alloc_data alloc_data = { .flags = flags };
267 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
271 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
273 blk_mq_put_ctx(alloc_data.ctx);
277 return ERR_PTR(-EWOULDBLOCK);
280 rq->__sector = (sector_t) -1;
281 rq->bio = rq->biotail = NULL;
284 EXPORT_SYMBOL(blk_mq_alloc_request);
286 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
287 unsigned int flags, unsigned int hctx_idx)
289 struct blk_mq_alloc_data alloc_data = { .flags = flags };
295 * If the tag allocator sleeps we could get an allocation for a
296 * different hardware context. No need to complicate the low level
297 * allocator for this for the rare use case of a command tied to
300 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
301 return ERR_PTR(-EINVAL);
303 if (hctx_idx >= q->nr_hw_queues)
304 return ERR_PTR(-EIO);
306 ret = blk_queue_enter(q, true);
311 * Check if the hardware context is actually mapped to anything.
312 * If not tell the caller that it should skip this queue.
314 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
315 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
317 return ERR_PTR(-EXDEV);
319 cpu = cpumask_first(alloc_data.hctx->cpumask);
320 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
322 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
327 return ERR_PTR(-EWOULDBLOCK);
331 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
333 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
336 const int sched_tag = rq->internal_tag;
337 struct request_queue *q = rq->q;
339 if (rq->rq_flags & RQF_MQ_INFLIGHT)
340 atomic_dec(&hctx->nr_active);
342 wbt_done(q->rq_wb, &rq->issue_stat);
345 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
346 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
348 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
350 blk_mq_sched_completed_request(hctx, rq);
351 blk_mq_sched_restart_queues(hctx);
355 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
358 struct blk_mq_ctx *ctx = rq->mq_ctx;
360 ctx->rq_completed[rq_is_sync(rq)]++;
361 __blk_mq_finish_request(hctx, ctx, rq);
364 void blk_mq_finish_request(struct request *rq)
366 blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
369 void blk_mq_free_request(struct request *rq)
371 blk_mq_sched_put_request(rq);
373 EXPORT_SYMBOL_GPL(blk_mq_free_request);
375 inline void __blk_mq_end_request(struct request *rq, int error)
377 blk_account_io_done(rq);
380 wbt_done(rq->q->rq_wb, &rq->issue_stat);
381 rq->end_io(rq, error);
383 if (unlikely(blk_bidi_rq(rq)))
384 blk_mq_free_request(rq->next_rq);
385 blk_mq_free_request(rq);
388 EXPORT_SYMBOL(__blk_mq_end_request);
390 void blk_mq_end_request(struct request *rq, int error)
392 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
394 __blk_mq_end_request(rq, error);
396 EXPORT_SYMBOL(blk_mq_end_request);
398 static void __blk_mq_complete_request_remote(void *data)
400 struct request *rq = data;
402 rq->q->softirq_done_fn(rq);
405 static void blk_mq_ipi_complete_request(struct request *rq)
407 struct blk_mq_ctx *ctx = rq->mq_ctx;
411 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
412 rq->q->softirq_done_fn(rq);
417 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
418 shared = cpus_share_cache(cpu, ctx->cpu);
420 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
421 rq->csd.func = __blk_mq_complete_request_remote;
424 smp_call_function_single_async(ctx->cpu, &rq->csd);
426 rq->q->softirq_done_fn(rq);
431 static void blk_mq_stat_add(struct request *rq)
433 if (rq->rq_flags & RQF_STATS) {
435 * We could rq->mq_ctx here, but there's less of a risk
436 * of races if we have the completion event add the stats
437 * to the local software queue.
439 struct blk_mq_ctx *ctx;
441 ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id());
442 blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq);
446 static void __blk_mq_complete_request(struct request *rq)
448 struct request_queue *q = rq->q;
452 if (!q->softirq_done_fn)
453 blk_mq_end_request(rq, rq->errors);
455 blk_mq_ipi_complete_request(rq);
459 * blk_mq_complete_request - end I/O on a request
460 * @rq: the request being processed
463 * Ends all I/O on a request. It does not handle partial completions.
464 * The actual completion happens out-of-order, through a IPI handler.
466 void blk_mq_complete_request(struct request *rq, int error)
468 struct request_queue *q = rq->q;
470 if (unlikely(blk_should_fake_timeout(q)))
472 if (!blk_mark_rq_complete(rq)) {
474 __blk_mq_complete_request(rq);
477 EXPORT_SYMBOL(blk_mq_complete_request);
479 int blk_mq_request_started(struct request *rq)
481 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
483 EXPORT_SYMBOL_GPL(blk_mq_request_started);
485 void blk_mq_start_request(struct request *rq)
487 struct request_queue *q = rq->q;
489 blk_mq_sched_started_request(rq);
491 trace_block_rq_issue(q, rq);
493 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
494 blk_stat_set_issue_time(&rq->issue_stat);
495 rq->rq_flags |= RQF_STATS;
496 wbt_issue(q->rq_wb, &rq->issue_stat);
502 * Ensure that ->deadline is visible before set the started
503 * flag and clear the completed flag.
505 smp_mb__before_atomic();
508 * Mark us as started and clear complete. Complete might have been
509 * set if requeue raced with timeout, which then marked it as
510 * complete. So be sure to clear complete again when we start
511 * the request, otherwise we'll ignore the completion event.
513 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
514 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
515 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
516 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
518 if (q->dma_drain_size && blk_rq_bytes(rq)) {
520 * Make sure space for the drain appears. We know we can do
521 * this because max_hw_segments has been adjusted to be one
522 * fewer than the device can handle.
524 rq->nr_phys_segments++;
527 EXPORT_SYMBOL(blk_mq_start_request);
529 static void __blk_mq_requeue_request(struct request *rq)
531 struct request_queue *q = rq->q;
533 trace_block_rq_requeue(q, rq);
534 wbt_requeue(q->rq_wb, &rq->issue_stat);
535 blk_mq_sched_requeue_request(rq);
537 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
538 if (q->dma_drain_size && blk_rq_bytes(rq))
539 rq->nr_phys_segments--;
543 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
545 __blk_mq_requeue_request(rq);
547 BUG_ON(blk_queued_rq(rq));
548 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
550 EXPORT_SYMBOL(blk_mq_requeue_request);
552 static void blk_mq_requeue_work(struct work_struct *work)
554 struct request_queue *q =
555 container_of(work, struct request_queue, requeue_work.work);
557 struct request *rq, *next;
560 spin_lock_irqsave(&q->requeue_lock, flags);
561 list_splice_init(&q->requeue_list, &rq_list);
562 spin_unlock_irqrestore(&q->requeue_lock, flags);
564 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
565 if (!(rq->rq_flags & RQF_SOFTBARRIER))
568 rq->rq_flags &= ~RQF_SOFTBARRIER;
569 list_del_init(&rq->queuelist);
570 blk_mq_sched_insert_request(rq, true, false, false, true);
573 while (!list_empty(&rq_list)) {
574 rq = list_entry(rq_list.next, struct request, queuelist);
575 list_del_init(&rq->queuelist);
576 blk_mq_sched_insert_request(rq, false, false, false, true);
579 blk_mq_run_hw_queues(q, false);
582 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
583 bool kick_requeue_list)
585 struct request_queue *q = rq->q;
589 * We abuse this flag that is otherwise used by the I/O scheduler to
590 * request head insertation from the workqueue.
592 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
594 spin_lock_irqsave(&q->requeue_lock, flags);
596 rq->rq_flags |= RQF_SOFTBARRIER;
597 list_add(&rq->queuelist, &q->requeue_list);
599 list_add_tail(&rq->queuelist, &q->requeue_list);
601 spin_unlock_irqrestore(&q->requeue_lock, flags);
603 if (kick_requeue_list)
604 blk_mq_kick_requeue_list(q);
606 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
608 void blk_mq_kick_requeue_list(struct request_queue *q)
610 kblockd_schedule_delayed_work(&q->requeue_work, 0);
612 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
614 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
617 kblockd_schedule_delayed_work(&q->requeue_work,
618 msecs_to_jiffies(msecs));
620 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
622 void blk_mq_abort_requeue_list(struct request_queue *q)
627 spin_lock_irqsave(&q->requeue_lock, flags);
628 list_splice_init(&q->requeue_list, &rq_list);
629 spin_unlock_irqrestore(&q->requeue_lock, flags);
631 while (!list_empty(&rq_list)) {
634 rq = list_first_entry(&rq_list, struct request, queuelist);
635 list_del_init(&rq->queuelist);
637 blk_mq_end_request(rq, rq->errors);
640 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
642 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
644 if (tag < tags->nr_tags) {
645 prefetch(tags->rqs[tag]);
646 return tags->rqs[tag];
651 EXPORT_SYMBOL(blk_mq_tag_to_rq);
653 struct blk_mq_timeout_data {
655 unsigned int next_set;
658 void blk_mq_rq_timed_out(struct request *req, bool reserved)
660 const struct blk_mq_ops *ops = req->q->mq_ops;
661 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
664 * We know that complete is set at this point. If STARTED isn't set
665 * anymore, then the request isn't active and the "timeout" should
666 * just be ignored. This can happen due to the bitflag ordering.
667 * Timeout first checks if STARTED is set, and if it is, assumes
668 * the request is active. But if we race with completion, then
669 * we both flags will get cleared. So check here again, and ignore
670 * a timeout event with a request that isn't active.
672 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
676 ret = ops->timeout(req, reserved);
680 __blk_mq_complete_request(req);
682 case BLK_EH_RESET_TIMER:
684 blk_clear_rq_complete(req);
686 case BLK_EH_NOT_HANDLED:
689 printk(KERN_ERR "block: bad eh return: %d\n", ret);
694 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
695 struct request *rq, void *priv, bool reserved)
697 struct blk_mq_timeout_data *data = priv;
699 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
702 if (time_after_eq(jiffies, rq->deadline)) {
703 if (!blk_mark_rq_complete(rq))
704 blk_mq_rq_timed_out(rq, reserved);
705 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
706 data->next = rq->deadline;
711 static void blk_mq_timeout_work(struct work_struct *work)
713 struct request_queue *q =
714 container_of(work, struct request_queue, timeout_work);
715 struct blk_mq_timeout_data data = {
721 /* A deadlock might occur if a request is stuck requiring a
722 * timeout at the same time a queue freeze is waiting
723 * completion, since the timeout code would not be able to
724 * acquire the queue reference here.
726 * That's why we don't use blk_queue_enter here; instead, we use
727 * percpu_ref_tryget directly, because we need to be able to
728 * obtain a reference even in the short window between the queue
729 * starting to freeze, by dropping the first reference in
730 * blk_mq_freeze_queue_start, and the moment the last request is
731 * consumed, marked by the instant q_usage_counter reaches
734 if (!percpu_ref_tryget(&q->q_usage_counter))
737 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
740 data.next = blk_rq_timeout(round_jiffies_up(data.next));
741 mod_timer(&q->timeout, data.next);
743 struct blk_mq_hw_ctx *hctx;
745 queue_for_each_hw_ctx(q, hctx, i) {
746 /* the hctx may be unmapped, so check it here */
747 if (blk_mq_hw_queue_mapped(hctx))
748 blk_mq_tag_idle(hctx);
755 * Reverse check our software queue for entries that we could potentially
756 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
757 * too much time checking for merges.
759 static bool blk_mq_attempt_merge(struct request_queue *q,
760 struct blk_mq_ctx *ctx, struct bio *bio)
765 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
771 if (!blk_rq_merge_ok(rq, bio))
774 switch (blk_try_merge(rq, bio)) {
775 case ELEVATOR_BACK_MERGE:
776 if (blk_mq_sched_allow_merge(q, rq, bio))
777 merged = bio_attempt_back_merge(q, rq, bio);
779 case ELEVATOR_FRONT_MERGE:
780 if (blk_mq_sched_allow_merge(q, rq, bio))
781 merged = bio_attempt_front_merge(q, rq, bio);
783 case ELEVATOR_DISCARD_MERGE:
784 merged = bio_attempt_discard_merge(q, rq, bio);
798 struct flush_busy_ctx_data {
799 struct blk_mq_hw_ctx *hctx;
800 struct list_head *list;
803 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
805 struct flush_busy_ctx_data *flush_data = data;
806 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
807 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
809 sbitmap_clear_bit(sb, bitnr);
810 spin_lock(&ctx->lock);
811 list_splice_tail_init(&ctx->rq_list, flush_data->list);
812 spin_unlock(&ctx->lock);
817 * Process software queues that have been marked busy, splicing them
818 * to the for-dispatch
820 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
822 struct flush_busy_ctx_data data = {
827 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
829 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
831 static inline unsigned int queued_to_index(unsigned int queued)
836 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
839 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
842 struct blk_mq_alloc_data data = {
844 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
845 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
851 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
852 data.flags |= BLK_MQ_REQ_RESERVED;
854 rq->tag = blk_mq_get_tag(&data);
856 if (blk_mq_tag_busy(data.hctx)) {
857 rq->rq_flags |= RQF_MQ_INFLIGHT;
858 atomic_inc(&data.hctx->nr_active);
860 data.hctx->tags->rqs[rq->tag] = rq;
866 return rq->tag != -1;
869 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
872 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
875 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
876 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
877 atomic_dec(&hctx->nr_active);
881 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
884 if (rq->tag == -1 || rq->internal_tag == -1)
887 __blk_mq_put_driver_tag(hctx, rq);
890 static void blk_mq_put_driver_tag(struct request *rq)
892 struct blk_mq_hw_ctx *hctx;
894 if (rq->tag == -1 || rq->internal_tag == -1)
897 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
898 __blk_mq_put_driver_tag(hctx, rq);
902 * If we fail getting a driver tag because all the driver tags are already
903 * assigned and on the dispatch list, BUT the first entry does not have a
904 * tag, then we could deadlock. For that case, move entries with assigned
905 * driver tags to the front, leaving the set of tagged requests in the
906 * same order, and the untagged set in the same order.
908 static bool reorder_tags_to_front(struct list_head *list)
910 struct request *rq, *tmp, *first = NULL;
912 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
916 list_move(&rq->queuelist, list);
922 return first != NULL;
925 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
928 struct blk_mq_hw_ctx *hctx;
930 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
932 list_del(&wait->task_list);
933 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
934 blk_mq_run_hw_queue(hctx, true);
938 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
940 struct sbq_wait_state *ws;
943 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
944 * The thread which wins the race to grab this bit adds the hardware
945 * queue to the wait queue.
947 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
948 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
951 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
952 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
955 * As soon as this returns, it's no longer safe to fiddle with
956 * hctx->dispatch_wait, since a completion can wake up the wait queue
957 * and unlock the bit.
959 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
963 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
965 struct blk_mq_hw_ctx *hctx;
967 LIST_HEAD(driver_list);
968 struct list_head *dptr;
969 int errors, queued, ret = BLK_MQ_RQ_QUEUE_OK;
971 if (list_empty(list))
975 * Start off with dptr being NULL, so we start the first request
976 * immediately, even if we have more pending.
981 * Now process all the entries, sending them to the driver.
985 struct blk_mq_queue_data bd;
987 rq = list_first_entry(list, struct request, queuelist);
988 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
989 if (!queued && reorder_tags_to_front(list))
993 * The initial allocation attempt failed, so we need to
994 * rerun the hardware queue when a tag is freed.
996 if (blk_mq_dispatch_wait_add(hctx)) {
998 * It's possible that a tag was freed in the
999 * window between the allocation failure and
1000 * adding the hardware queue to the wait queue.
1002 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1009 list_del_init(&rq->queuelist);
1015 * Flag last if we have no more requests, or if we have more
1016 * but can't assign a driver tag to it.
1018 if (list_empty(list))
1021 struct request *nxt;
1023 nxt = list_first_entry(list, struct request, queuelist);
1024 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1027 ret = q->mq_ops->queue_rq(hctx, &bd);
1029 case BLK_MQ_RQ_QUEUE_OK:
1032 case BLK_MQ_RQ_QUEUE_BUSY:
1033 blk_mq_put_driver_tag_hctx(hctx, rq);
1034 list_add(&rq->queuelist, list);
1035 __blk_mq_requeue_request(rq);
1038 pr_err("blk-mq: bad return on queue: %d\n", ret);
1039 case BLK_MQ_RQ_QUEUE_ERROR:
1042 blk_mq_end_request(rq, rq->errors);
1046 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
1050 * We've done the first request. If we have more than 1
1051 * left in the list, set dptr to defer issue.
1053 if (!dptr && list->next != list->prev)
1054 dptr = &driver_list;
1055 } while (!list_empty(list));
1057 hctx->dispatched[queued_to_index(queued)]++;
1060 * Any items that need requeuing? Stuff them into hctx->dispatch,
1061 * that is where we will continue on next queue run.
1063 if (!list_empty(list)) {
1065 * If we got a driver tag for the next request already,
1068 rq = list_first_entry(list, struct request, queuelist);
1069 blk_mq_put_driver_tag(rq);
1071 spin_lock(&hctx->lock);
1072 list_splice_init(list, &hctx->dispatch);
1073 spin_unlock(&hctx->lock);
1076 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1077 * it's possible the queue is stopped and restarted again
1078 * before this. Queue restart will dispatch requests. And since
1079 * requests in rq_list aren't added into hctx->dispatch yet,
1080 * the requests in rq_list might get lost.
1082 * blk_mq_run_hw_queue() already checks the STOPPED bit
1084 * If RESTART or TAG_WAITING is set, then let completion restart
1085 * the queue instead of potentially looping here.
1087 if (!blk_mq_sched_needs_restart(hctx) &&
1088 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1089 blk_mq_run_hw_queue(hctx, true);
1092 return (queued + errors) != 0;
1095 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1099 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1100 cpu_online(hctx->next_cpu));
1102 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1104 blk_mq_sched_dispatch_requests(hctx);
1107 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1108 blk_mq_sched_dispatch_requests(hctx);
1109 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1114 * It'd be great if the workqueue API had a way to pass
1115 * in a mask and had some smarts for more clever placement.
1116 * For now we just round-robin here, switching for every
1117 * BLK_MQ_CPU_WORK_BATCH queued items.
1119 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1121 if (hctx->queue->nr_hw_queues == 1)
1122 return WORK_CPU_UNBOUND;
1124 if (--hctx->next_cpu_batch <= 0) {
1127 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1128 if (next_cpu >= nr_cpu_ids)
1129 next_cpu = cpumask_first(hctx->cpumask);
1131 hctx->next_cpu = next_cpu;
1132 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1135 return hctx->next_cpu;
1138 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1140 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1141 !blk_mq_hw_queue_mapped(hctx)))
1144 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1145 int cpu = get_cpu();
1146 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1147 __blk_mq_run_hw_queue(hctx);
1155 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
1158 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1160 struct blk_mq_hw_ctx *hctx;
1163 queue_for_each_hw_ctx(q, hctx, i) {
1164 if (!blk_mq_hctx_has_pending(hctx) ||
1165 blk_mq_hctx_stopped(hctx))
1168 blk_mq_run_hw_queue(hctx, async);
1171 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1174 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1175 * @q: request queue.
1177 * The caller is responsible for serializing this function against
1178 * blk_mq_{start,stop}_hw_queue().
1180 bool blk_mq_queue_stopped(struct request_queue *q)
1182 struct blk_mq_hw_ctx *hctx;
1185 queue_for_each_hw_ctx(q, hctx, i)
1186 if (blk_mq_hctx_stopped(hctx))
1191 EXPORT_SYMBOL(blk_mq_queue_stopped);
1193 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1195 cancel_work(&hctx->run_work);
1196 cancel_delayed_work(&hctx->delay_work);
1197 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1199 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1201 void blk_mq_stop_hw_queues(struct request_queue *q)
1203 struct blk_mq_hw_ctx *hctx;
1206 queue_for_each_hw_ctx(q, hctx, i)
1207 blk_mq_stop_hw_queue(hctx);
1209 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1211 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1213 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1215 blk_mq_run_hw_queue(hctx, false);
1217 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1219 void blk_mq_start_hw_queues(struct request_queue *q)
1221 struct blk_mq_hw_ctx *hctx;
1224 queue_for_each_hw_ctx(q, hctx, i)
1225 blk_mq_start_hw_queue(hctx);
1227 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1229 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1231 if (!blk_mq_hctx_stopped(hctx))
1234 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1235 blk_mq_run_hw_queue(hctx, async);
1237 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1239 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1241 struct blk_mq_hw_ctx *hctx;
1244 queue_for_each_hw_ctx(q, hctx, i)
1245 blk_mq_start_stopped_hw_queue(hctx, async);
1247 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1249 static void blk_mq_run_work_fn(struct work_struct *work)
1251 struct blk_mq_hw_ctx *hctx;
1253 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1255 __blk_mq_run_hw_queue(hctx);
1258 static void blk_mq_delay_work_fn(struct work_struct *work)
1260 struct blk_mq_hw_ctx *hctx;
1262 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1264 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1265 __blk_mq_run_hw_queue(hctx);
1268 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1270 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1273 blk_mq_stop_hw_queue(hctx);
1274 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1275 &hctx->delay_work, msecs_to_jiffies(msecs));
1277 EXPORT_SYMBOL(blk_mq_delay_queue);
1279 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1283 struct blk_mq_ctx *ctx = rq->mq_ctx;
1285 trace_block_rq_insert(hctx->queue, rq);
1288 list_add(&rq->queuelist, &ctx->rq_list);
1290 list_add_tail(&rq->queuelist, &ctx->rq_list);
1293 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1296 struct blk_mq_ctx *ctx = rq->mq_ctx;
1298 __blk_mq_insert_req_list(hctx, rq, at_head);
1299 blk_mq_hctx_mark_pending(hctx, ctx);
1302 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1303 struct list_head *list)
1307 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1310 spin_lock(&ctx->lock);
1311 while (!list_empty(list)) {
1314 rq = list_first_entry(list, struct request, queuelist);
1315 BUG_ON(rq->mq_ctx != ctx);
1316 list_del_init(&rq->queuelist);
1317 __blk_mq_insert_req_list(hctx, rq, false);
1319 blk_mq_hctx_mark_pending(hctx, ctx);
1320 spin_unlock(&ctx->lock);
1323 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1325 struct request *rqa = container_of(a, struct request, queuelist);
1326 struct request *rqb = container_of(b, struct request, queuelist);
1328 return !(rqa->mq_ctx < rqb->mq_ctx ||
1329 (rqa->mq_ctx == rqb->mq_ctx &&
1330 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1333 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1335 struct blk_mq_ctx *this_ctx;
1336 struct request_queue *this_q;
1339 LIST_HEAD(ctx_list);
1342 list_splice_init(&plug->mq_list, &list);
1344 list_sort(NULL, &list, plug_ctx_cmp);
1350 while (!list_empty(&list)) {
1351 rq = list_entry_rq(list.next);
1352 list_del_init(&rq->queuelist);
1354 if (rq->mq_ctx != this_ctx) {
1356 trace_block_unplug(this_q, depth, from_schedule);
1357 blk_mq_sched_insert_requests(this_q, this_ctx,
1362 this_ctx = rq->mq_ctx;
1368 list_add_tail(&rq->queuelist, &ctx_list);
1372 * If 'this_ctx' is set, we know we have entries to complete
1373 * on 'ctx_list'. Do those.
1376 trace_block_unplug(this_q, depth, from_schedule);
1377 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1382 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1384 init_request_from_bio(rq, bio);
1386 blk_account_io_start(rq, true);
1389 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1391 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1392 !blk_queue_nomerges(hctx->queue);
1395 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1396 struct blk_mq_ctx *ctx,
1397 struct request *rq, struct bio *bio)
1399 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1400 blk_mq_bio_to_request(rq, bio);
1401 spin_lock(&ctx->lock);
1403 __blk_mq_insert_request(hctx, rq, false);
1404 spin_unlock(&ctx->lock);
1407 struct request_queue *q = hctx->queue;
1409 spin_lock(&ctx->lock);
1410 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1411 blk_mq_bio_to_request(rq, bio);
1415 spin_unlock(&ctx->lock);
1416 __blk_mq_finish_request(hctx, ctx, rq);
1421 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1424 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1426 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1429 static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie,
1432 struct request_queue *q = rq->q;
1433 struct blk_mq_queue_data bd = {
1438 struct blk_mq_hw_ctx *hctx;
1439 blk_qc_t new_cookie;
1445 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1448 new_cookie = request_to_qc_t(hctx, rq);
1451 * For OK queue, we are done. For error, kill it. Any other
1452 * error (busy), just add it to our list as we previously
1455 ret = q->mq_ops->queue_rq(hctx, &bd);
1456 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1457 *cookie = new_cookie;
1461 __blk_mq_requeue_request(rq);
1463 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1464 *cookie = BLK_QC_T_NONE;
1466 blk_mq_end_request(rq, rq->errors);
1471 blk_mq_sched_insert_request(rq, false, true, false, may_sleep);
1475 * Multiple hardware queue variant. This will not use per-process plugs,
1476 * but will attempt to bypass the hctx queueing if we can go straight to
1477 * hardware for SYNC IO.
1479 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1481 const int is_sync = op_is_sync(bio->bi_opf);
1482 const int is_flush_fua = op_is_flush(bio->bi_opf);
1483 struct blk_mq_alloc_data data = { .flags = 0 };
1485 unsigned int request_count = 0, srcu_idx;
1486 struct blk_plug *plug;
1487 struct request *same_queue_rq = NULL;
1489 unsigned int wb_acct;
1491 blk_queue_bounce(q, &bio);
1493 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1495 return BLK_QC_T_NONE;
1498 blk_queue_split(q, &bio, q->bio_split);
1500 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1501 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1502 return BLK_QC_T_NONE;
1504 if (blk_mq_sched_bio_merge(q, bio))
1505 return BLK_QC_T_NONE;
1507 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1509 trace_block_getrq(q, bio, bio->bi_opf);
1511 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1512 if (unlikely(!rq)) {
1513 __wbt_done(q->rq_wb, wb_acct);
1514 return BLK_QC_T_NONE;
1517 wbt_track(&rq->issue_stat, wb_acct);
1519 cookie = request_to_qc_t(data.hctx, rq);
1521 if (unlikely(is_flush_fua)) {
1524 blk_mq_bio_to_request(rq, bio);
1525 blk_insert_flush(rq);
1529 plug = current->plug;
1531 * If the driver supports defer issued based on 'last', then
1532 * queue it up like normal since we can potentially save some
1535 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1536 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1537 struct request *old_rq = NULL;
1539 blk_mq_bio_to_request(rq, bio);
1542 * We do limited plugging. If the bio can be merged, do that.
1543 * Otherwise the existing request in the plug list will be
1544 * issued. So the plug list will have one request at most
1548 * The plug list might get flushed before this. If that
1549 * happens, same_queue_rq is invalid and plug list is
1552 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1553 old_rq = same_queue_rq;
1554 list_del_init(&old_rq->queuelist);
1556 list_add_tail(&rq->queuelist, &plug->mq_list);
1557 } else /* is_sync */
1559 blk_mq_put_ctx(data.ctx);
1563 if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) {
1565 blk_mq_try_issue_directly(old_rq, &cookie, false);
1568 srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu);
1569 blk_mq_try_issue_directly(old_rq, &cookie, true);
1570 srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx);
1577 blk_mq_put_ctx(data.ctx);
1578 blk_mq_bio_to_request(rq, bio);
1579 blk_mq_sched_insert_request(rq, false, true,
1580 !is_sync || is_flush_fua, true);
1583 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1585 * For a SYNC request, send it to the hardware immediately. For
1586 * an ASYNC request, just ensure that we run it later on. The
1587 * latter allows for merging opportunities and more efficient
1591 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1593 blk_mq_put_ctx(data.ctx);
1599 * Single hardware queue variant. This will attempt to use any per-process
1600 * plug for merging and IO deferral.
1602 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1604 const int is_sync = op_is_sync(bio->bi_opf);
1605 const int is_flush_fua = op_is_flush(bio->bi_opf);
1606 struct blk_plug *plug;
1607 unsigned int request_count = 0;
1608 struct blk_mq_alloc_data data = { .flags = 0 };
1611 unsigned int wb_acct;
1613 blk_queue_bounce(q, &bio);
1615 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1617 return BLK_QC_T_NONE;
1620 blk_queue_split(q, &bio, q->bio_split);
1622 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1623 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1624 return BLK_QC_T_NONE;
1626 request_count = blk_plug_queued_count(q);
1628 if (blk_mq_sched_bio_merge(q, bio))
1629 return BLK_QC_T_NONE;
1631 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1633 trace_block_getrq(q, bio, bio->bi_opf);
1635 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1636 if (unlikely(!rq)) {
1637 __wbt_done(q->rq_wb, wb_acct);
1638 return BLK_QC_T_NONE;
1641 wbt_track(&rq->issue_stat, wb_acct);
1643 cookie = request_to_qc_t(data.hctx, rq);
1645 if (unlikely(is_flush_fua)) {
1648 blk_mq_bio_to_request(rq, bio);
1649 blk_insert_flush(rq);
1654 * A task plug currently exists. Since this is completely lockless,
1655 * utilize that to temporarily store requests until the task is
1656 * either done or scheduled away.
1658 plug = current->plug;
1660 struct request *last = NULL;
1662 blk_mq_bio_to_request(rq, bio);
1665 * @request_count may become stale because of schedule
1666 * out, so check the list again.
1668 if (list_empty(&plug->mq_list))
1671 trace_block_plug(q);
1673 last = list_entry_rq(plug->mq_list.prev);
1675 blk_mq_put_ctx(data.ctx);
1677 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1678 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1679 blk_flush_plug_list(plug, false);
1680 trace_block_plug(q);
1683 list_add_tail(&rq->queuelist, &plug->mq_list);
1689 blk_mq_put_ctx(data.ctx);
1690 blk_mq_bio_to_request(rq, bio);
1691 blk_mq_sched_insert_request(rq, false, true,
1692 !is_sync || is_flush_fua, true);
1695 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1697 * For a SYNC request, send it to the hardware immediately. For
1698 * an ASYNC request, just ensure that we run it later on. The
1699 * latter allows for merging opportunities and more efficient
1703 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1706 blk_mq_put_ctx(data.ctx);
1711 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1712 unsigned int hctx_idx)
1716 if (tags->rqs && set->ops->exit_request) {
1719 for (i = 0; i < tags->nr_tags; i++) {
1720 struct request *rq = tags->static_rqs[i];
1724 set->ops->exit_request(set->driver_data, rq,
1726 tags->static_rqs[i] = NULL;
1730 while (!list_empty(&tags->page_list)) {
1731 page = list_first_entry(&tags->page_list, struct page, lru);
1732 list_del_init(&page->lru);
1734 * Remove kmemleak object previously allocated in
1735 * blk_mq_init_rq_map().
1737 kmemleak_free(page_address(page));
1738 __free_pages(page, page->private);
1742 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1746 kfree(tags->static_rqs);
1747 tags->static_rqs = NULL;
1749 blk_mq_free_tags(tags);
1752 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1753 unsigned int hctx_idx,
1754 unsigned int nr_tags,
1755 unsigned int reserved_tags)
1757 struct blk_mq_tags *tags;
1760 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1761 if (node == NUMA_NO_NODE)
1762 node = set->numa_node;
1764 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1765 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1769 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1770 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1773 blk_mq_free_tags(tags);
1777 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1778 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1780 if (!tags->static_rqs) {
1782 blk_mq_free_tags(tags);
1789 static size_t order_to_size(unsigned int order)
1791 return (size_t)PAGE_SIZE << order;
1794 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1795 unsigned int hctx_idx, unsigned int depth)
1797 unsigned int i, j, entries_per_page, max_order = 4;
1798 size_t rq_size, left;
1801 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1802 if (node == NUMA_NO_NODE)
1803 node = set->numa_node;
1805 INIT_LIST_HEAD(&tags->page_list);
1808 * rq_size is the size of the request plus driver payload, rounded
1809 * to the cacheline size
1811 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1813 left = rq_size * depth;
1815 for (i = 0; i < depth; ) {
1816 int this_order = max_order;
1821 while (this_order && left < order_to_size(this_order - 1))
1825 page = alloc_pages_node(node,
1826 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1832 if (order_to_size(this_order) < rq_size)
1839 page->private = this_order;
1840 list_add_tail(&page->lru, &tags->page_list);
1842 p = page_address(page);
1844 * Allow kmemleak to scan these pages as they contain pointers
1845 * to additional allocations like via ops->init_request().
1847 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1848 entries_per_page = order_to_size(this_order) / rq_size;
1849 to_do = min(entries_per_page, depth - i);
1850 left -= to_do * rq_size;
1851 for (j = 0; j < to_do; j++) {
1852 struct request *rq = p;
1854 tags->static_rqs[i] = rq;
1855 if (set->ops->init_request) {
1856 if (set->ops->init_request(set->driver_data,
1859 tags->static_rqs[i] = NULL;
1871 blk_mq_free_rqs(set, tags, hctx_idx);
1876 * 'cpu' is going away. splice any existing rq_list entries from this
1877 * software queue to the hw queue dispatch list, and ensure that it
1880 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1882 struct blk_mq_hw_ctx *hctx;
1883 struct blk_mq_ctx *ctx;
1886 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1887 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1889 spin_lock(&ctx->lock);
1890 if (!list_empty(&ctx->rq_list)) {
1891 list_splice_init(&ctx->rq_list, &tmp);
1892 blk_mq_hctx_clear_pending(hctx, ctx);
1894 spin_unlock(&ctx->lock);
1896 if (list_empty(&tmp))
1899 spin_lock(&hctx->lock);
1900 list_splice_tail_init(&tmp, &hctx->dispatch);
1901 spin_unlock(&hctx->lock);
1903 blk_mq_run_hw_queue(hctx, true);
1907 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1909 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1913 /* hctx->ctxs will be freed in queue's release handler */
1914 static void blk_mq_exit_hctx(struct request_queue *q,
1915 struct blk_mq_tag_set *set,
1916 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1918 unsigned flush_start_tag = set->queue_depth;
1920 blk_mq_tag_idle(hctx);
1922 if (set->ops->exit_request)
1923 set->ops->exit_request(set->driver_data,
1924 hctx->fq->flush_rq, hctx_idx,
1925 flush_start_tag + hctx_idx);
1927 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1929 if (set->ops->exit_hctx)
1930 set->ops->exit_hctx(hctx, hctx_idx);
1932 if (hctx->flags & BLK_MQ_F_BLOCKING)
1933 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1935 blk_mq_remove_cpuhp(hctx);
1936 blk_free_flush_queue(hctx->fq);
1937 sbitmap_free(&hctx->ctx_map);
1940 static void blk_mq_exit_hw_queues(struct request_queue *q,
1941 struct blk_mq_tag_set *set, int nr_queue)
1943 struct blk_mq_hw_ctx *hctx;
1946 queue_for_each_hw_ctx(q, hctx, i) {
1949 blk_mq_exit_hctx(q, set, hctx, i);
1953 static int blk_mq_init_hctx(struct request_queue *q,
1954 struct blk_mq_tag_set *set,
1955 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1958 unsigned flush_start_tag = set->queue_depth;
1960 node = hctx->numa_node;
1961 if (node == NUMA_NO_NODE)
1962 node = hctx->numa_node = set->numa_node;
1964 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1965 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1966 spin_lock_init(&hctx->lock);
1967 INIT_LIST_HEAD(&hctx->dispatch);
1969 hctx->queue_num = hctx_idx;
1970 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1972 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1974 hctx->tags = set->tags[hctx_idx];
1977 * Allocate space for all possible cpus to avoid allocation at
1980 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1983 goto unregister_cpu_notifier;
1985 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1991 if (set->ops->init_hctx &&
1992 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1995 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1998 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2000 goto sched_exit_hctx;
2002 if (set->ops->init_request &&
2003 set->ops->init_request(set->driver_data,
2004 hctx->fq->flush_rq, hctx_idx,
2005 flush_start_tag + hctx_idx, node))
2008 if (hctx->flags & BLK_MQ_F_BLOCKING)
2009 init_srcu_struct(&hctx->queue_rq_srcu);
2016 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2018 if (set->ops->exit_hctx)
2019 set->ops->exit_hctx(hctx, hctx_idx);
2021 sbitmap_free(&hctx->ctx_map);
2024 unregister_cpu_notifier:
2025 blk_mq_remove_cpuhp(hctx);
2029 static void blk_mq_init_cpu_queues(struct request_queue *q,
2030 unsigned int nr_hw_queues)
2034 for_each_possible_cpu(i) {
2035 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2036 struct blk_mq_hw_ctx *hctx;
2039 spin_lock_init(&__ctx->lock);
2040 INIT_LIST_HEAD(&__ctx->rq_list);
2042 blk_stat_init(&__ctx->stat[BLK_STAT_READ]);
2043 blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]);
2045 /* If the cpu isn't online, the cpu is mapped to first hctx */
2049 hctx = blk_mq_map_queue(q, i);
2052 * Set local node, IFF we have more than one hw queue. If
2053 * not, we remain on the home node of the device
2055 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2056 hctx->numa_node = local_memory_node(cpu_to_node(i));
2060 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2064 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2065 set->queue_depth, set->reserved_tags);
2066 if (!set->tags[hctx_idx])
2069 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2074 blk_mq_free_rq_map(set->tags[hctx_idx]);
2075 set->tags[hctx_idx] = NULL;
2079 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2080 unsigned int hctx_idx)
2082 if (set->tags[hctx_idx]) {
2083 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2084 blk_mq_free_rq_map(set->tags[hctx_idx]);
2085 set->tags[hctx_idx] = NULL;
2089 static void blk_mq_map_swqueue(struct request_queue *q,
2090 const struct cpumask *online_mask)
2092 unsigned int i, hctx_idx;
2093 struct blk_mq_hw_ctx *hctx;
2094 struct blk_mq_ctx *ctx;
2095 struct blk_mq_tag_set *set = q->tag_set;
2098 * Avoid others reading imcomplete hctx->cpumask through sysfs
2100 mutex_lock(&q->sysfs_lock);
2102 queue_for_each_hw_ctx(q, hctx, i) {
2103 cpumask_clear(hctx->cpumask);
2108 * Map software to hardware queues
2110 for_each_possible_cpu(i) {
2111 /* If the cpu isn't online, the cpu is mapped to first hctx */
2112 if (!cpumask_test_cpu(i, online_mask))
2115 hctx_idx = q->mq_map[i];
2116 /* unmapped hw queue can be remapped after CPU topo changed */
2117 if (!set->tags[hctx_idx] &&
2118 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2120 * If tags initialization fail for some hctx,
2121 * that hctx won't be brought online. In this
2122 * case, remap the current ctx to hctx[0] which
2123 * is guaranteed to always have tags allocated
2128 ctx = per_cpu_ptr(q->queue_ctx, i);
2129 hctx = blk_mq_map_queue(q, i);
2131 cpumask_set_cpu(i, hctx->cpumask);
2132 ctx->index_hw = hctx->nr_ctx;
2133 hctx->ctxs[hctx->nr_ctx++] = ctx;
2136 mutex_unlock(&q->sysfs_lock);
2138 queue_for_each_hw_ctx(q, hctx, i) {
2140 * If no software queues are mapped to this hardware queue,
2141 * disable it and free the request entries.
2143 if (!hctx->nr_ctx) {
2144 /* Never unmap queue 0. We need it as a
2145 * fallback in case of a new remap fails
2148 if (i && set->tags[i])
2149 blk_mq_free_map_and_requests(set, i);
2155 hctx->tags = set->tags[i];
2156 WARN_ON(!hctx->tags);
2159 * Set the map size to the number of mapped software queues.
2160 * This is more accurate and more efficient than looping
2161 * over all possibly mapped software queues.
2163 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2166 * Initialize batch roundrobin counts
2168 hctx->next_cpu = cpumask_first(hctx->cpumask);
2169 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2173 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2175 struct blk_mq_hw_ctx *hctx;
2178 queue_for_each_hw_ctx(q, hctx, i) {
2180 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2182 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2186 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2188 struct request_queue *q;
2190 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2191 blk_mq_freeze_queue(q);
2192 queue_set_hctx_shared(q, shared);
2193 blk_mq_unfreeze_queue(q);
2197 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2199 struct blk_mq_tag_set *set = q->tag_set;
2201 mutex_lock(&set->tag_list_lock);
2202 list_del_init(&q->tag_set_list);
2203 if (list_is_singular(&set->tag_list)) {
2204 /* just transitioned to unshared */
2205 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2206 /* update existing queue */
2207 blk_mq_update_tag_set_depth(set, false);
2209 mutex_unlock(&set->tag_list_lock);
2212 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2213 struct request_queue *q)
2217 mutex_lock(&set->tag_list_lock);
2219 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2220 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2221 set->flags |= BLK_MQ_F_TAG_SHARED;
2222 /* update existing queue */
2223 blk_mq_update_tag_set_depth(set, true);
2225 if (set->flags & BLK_MQ_F_TAG_SHARED)
2226 queue_set_hctx_shared(q, true);
2227 list_add_tail(&q->tag_set_list, &set->tag_list);
2229 mutex_unlock(&set->tag_list_lock);
2233 * It is the actual release handler for mq, but we do it from
2234 * request queue's release handler for avoiding use-after-free
2235 * and headache because q->mq_kobj shouldn't have been introduced,
2236 * but we can't group ctx/kctx kobj without it.
2238 void blk_mq_release(struct request_queue *q)
2240 struct blk_mq_hw_ctx *hctx;
2243 /* hctx kobj stays in hctx */
2244 queue_for_each_hw_ctx(q, hctx, i) {
2247 kobject_put(&hctx->kobj);
2252 kfree(q->queue_hw_ctx);
2255 * release .mq_kobj and sw queue's kobject now because
2256 * both share lifetime with request queue.
2258 blk_mq_sysfs_deinit(q);
2260 free_percpu(q->queue_ctx);
2263 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2265 struct request_queue *uninit_q, *q;
2267 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2269 return ERR_PTR(-ENOMEM);
2271 q = blk_mq_init_allocated_queue(set, uninit_q);
2273 blk_cleanup_queue(uninit_q);
2277 EXPORT_SYMBOL(blk_mq_init_queue);
2279 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2280 struct request_queue *q)
2283 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2285 blk_mq_sysfs_unregister(q);
2286 for (i = 0; i < set->nr_hw_queues; i++) {
2292 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2293 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2298 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2305 atomic_set(&hctxs[i]->nr_active, 0);
2306 hctxs[i]->numa_node = node;
2307 hctxs[i]->queue_num = i;
2309 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2310 free_cpumask_var(hctxs[i]->cpumask);
2315 blk_mq_hctx_kobj_init(hctxs[i]);
2317 for (j = i; j < q->nr_hw_queues; j++) {
2318 struct blk_mq_hw_ctx *hctx = hctxs[j];
2322 blk_mq_free_map_and_requests(set, j);
2323 blk_mq_exit_hctx(q, set, hctx, j);
2324 kobject_put(&hctx->kobj);
2329 q->nr_hw_queues = i;
2330 blk_mq_sysfs_register(q);
2333 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2334 struct request_queue *q)
2336 /* mark the queue as mq asap */
2337 q->mq_ops = set->ops;
2339 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2343 /* init q->mq_kobj and sw queues' kobjects */
2344 blk_mq_sysfs_init(q);
2346 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2347 GFP_KERNEL, set->numa_node);
2348 if (!q->queue_hw_ctx)
2351 q->mq_map = set->mq_map;
2353 blk_mq_realloc_hw_ctxs(set, q);
2354 if (!q->nr_hw_queues)
2357 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2358 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2360 q->nr_queues = nr_cpu_ids;
2362 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2364 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2365 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2367 q->sg_reserved_size = INT_MAX;
2369 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2370 INIT_LIST_HEAD(&q->requeue_list);
2371 spin_lock_init(&q->requeue_lock);
2373 if (q->nr_hw_queues > 1)
2374 blk_queue_make_request(q, blk_mq_make_request);
2376 blk_queue_make_request(q, blk_sq_make_request);
2379 * Do this after blk_queue_make_request() overrides it...
2381 q->nr_requests = set->queue_depth;
2384 * Default to classic polling
2388 if (set->ops->complete)
2389 blk_queue_softirq_done(q, set->ops->complete);
2391 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2394 mutex_lock(&all_q_mutex);
2396 list_add_tail(&q->all_q_node, &all_q_list);
2397 blk_mq_add_queue_tag_set(set, q);
2398 blk_mq_map_swqueue(q, cpu_online_mask);
2400 mutex_unlock(&all_q_mutex);
2403 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2406 ret = blk_mq_sched_init(q);
2408 return ERR_PTR(ret);
2414 kfree(q->queue_hw_ctx);
2416 free_percpu(q->queue_ctx);
2419 return ERR_PTR(-ENOMEM);
2421 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2423 void blk_mq_free_queue(struct request_queue *q)
2425 struct blk_mq_tag_set *set = q->tag_set;
2427 mutex_lock(&all_q_mutex);
2428 list_del_init(&q->all_q_node);
2429 mutex_unlock(&all_q_mutex);
2433 blk_mq_del_queue_tag_set(q);
2435 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2438 /* Basically redo blk_mq_init_queue with queue frozen */
2439 static void blk_mq_queue_reinit(struct request_queue *q,
2440 const struct cpumask *online_mask)
2442 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2444 blk_mq_sysfs_unregister(q);
2447 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2448 * we should change hctx numa_node according to new topology (this
2449 * involves free and re-allocate memory, worthy doing?)
2452 blk_mq_map_swqueue(q, online_mask);
2454 blk_mq_sysfs_register(q);
2458 * New online cpumask which is going to be set in this hotplug event.
2459 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2460 * one-by-one and dynamically allocating this could result in a failure.
2462 static struct cpumask cpuhp_online_new;
2464 static void blk_mq_queue_reinit_work(void)
2466 struct request_queue *q;
2468 mutex_lock(&all_q_mutex);
2470 * We need to freeze and reinit all existing queues. Freezing
2471 * involves synchronous wait for an RCU grace period and doing it
2472 * one by one may take a long time. Start freezing all queues in
2473 * one swoop and then wait for the completions so that freezing can
2474 * take place in parallel.
2476 list_for_each_entry(q, &all_q_list, all_q_node)
2477 blk_mq_freeze_queue_start(q);
2478 list_for_each_entry(q, &all_q_list, all_q_node)
2479 blk_mq_freeze_queue_wait(q);
2481 list_for_each_entry(q, &all_q_list, all_q_node)
2482 blk_mq_queue_reinit(q, &cpuhp_online_new);
2484 list_for_each_entry(q, &all_q_list, all_q_node)
2485 blk_mq_unfreeze_queue(q);
2487 mutex_unlock(&all_q_mutex);
2490 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2492 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2493 blk_mq_queue_reinit_work();
2498 * Before hotadded cpu starts handling requests, new mappings must be
2499 * established. Otherwise, these requests in hw queue might never be
2502 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2503 * for CPU0, and ctx1 for CPU1).
2505 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2506 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2508 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2509 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2510 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2513 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2515 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2516 cpumask_set_cpu(cpu, &cpuhp_online_new);
2517 blk_mq_queue_reinit_work();
2521 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2525 for (i = 0; i < set->nr_hw_queues; i++)
2526 if (!__blk_mq_alloc_rq_map(set, i))
2533 blk_mq_free_rq_map(set->tags[i]);
2539 * Allocate the request maps associated with this tag_set. Note that this
2540 * may reduce the depth asked for, if memory is tight. set->queue_depth
2541 * will be updated to reflect the allocated depth.
2543 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2548 depth = set->queue_depth;
2550 err = __blk_mq_alloc_rq_maps(set);
2554 set->queue_depth >>= 1;
2555 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2559 } while (set->queue_depth);
2561 if (!set->queue_depth || err) {
2562 pr_err("blk-mq: failed to allocate request map\n");
2566 if (depth != set->queue_depth)
2567 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2568 depth, set->queue_depth);
2574 * Alloc a tag set to be associated with one or more request queues.
2575 * May fail with EINVAL for various error conditions. May adjust the
2576 * requested depth down, if if it too large. In that case, the set
2577 * value will be stored in set->queue_depth.
2579 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2583 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2585 if (!set->nr_hw_queues)
2587 if (!set->queue_depth)
2589 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2592 if (!set->ops->queue_rq)
2595 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2596 pr_info("blk-mq: reduced tag depth to %u\n",
2598 set->queue_depth = BLK_MQ_MAX_DEPTH;
2602 * If a crashdump is active, then we are potentially in a very
2603 * memory constrained environment. Limit us to 1 queue and
2604 * 64 tags to prevent using too much memory.
2606 if (is_kdump_kernel()) {
2607 set->nr_hw_queues = 1;
2608 set->queue_depth = min(64U, set->queue_depth);
2611 * There is no use for more h/w queues than cpus.
2613 if (set->nr_hw_queues > nr_cpu_ids)
2614 set->nr_hw_queues = nr_cpu_ids;
2616 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2617 GFP_KERNEL, set->numa_node);
2622 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2623 GFP_KERNEL, set->numa_node);
2627 if (set->ops->map_queues)
2628 ret = set->ops->map_queues(set);
2630 ret = blk_mq_map_queues(set);
2632 goto out_free_mq_map;
2634 ret = blk_mq_alloc_rq_maps(set);
2636 goto out_free_mq_map;
2638 mutex_init(&set->tag_list_lock);
2639 INIT_LIST_HEAD(&set->tag_list);
2651 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2653 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2657 for (i = 0; i < nr_cpu_ids; i++)
2658 blk_mq_free_map_and_requests(set, i);
2666 EXPORT_SYMBOL(blk_mq_free_tag_set);
2668 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2670 struct blk_mq_tag_set *set = q->tag_set;
2671 struct blk_mq_hw_ctx *hctx;
2677 blk_mq_freeze_queue(q);
2678 blk_mq_quiesce_queue(q);
2681 queue_for_each_hw_ctx(q, hctx, i) {
2685 * If we're using an MQ scheduler, just update the scheduler
2686 * queue depth. This is similar to what the old code would do.
2688 if (!hctx->sched_tags) {
2689 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2690 min(nr, set->queue_depth),
2693 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2701 q->nr_requests = nr;
2703 blk_mq_unfreeze_queue(q);
2704 blk_mq_start_stopped_hw_queues(q, true);
2709 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2711 struct request_queue *q;
2713 if (nr_hw_queues > nr_cpu_ids)
2714 nr_hw_queues = nr_cpu_ids;
2715 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2718 list_for_each_entry(q, &set->tag_list, tag_set_list)
2719 blk_mq_freeze_queue(q);
2721 set->nr_hw_queues = nr_hw_queues;
2722 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2723 blk_mq_realloc_hw_ctxs(set, q);
2726 * Manually set the make_request_fn as blk_queue_make_request
2727 * resets a lot of the queue settings.
2729 if (q->nr_hw_queues > 1)
2730 q->make_request_fn = blk_mq_make_request;
2732 q->make_request_fn = blk_sq_make_request;
2734 blk_mq_queue_reinit(q, cpu_online_mask);
2737 list_for_each_entry(q, &set->tag_list, tag_set_list)
2738 blk_mq_unfreeze_queue(q);
2740 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2742 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2743 struct blk_mq_hw_ctx *hctx,
2746 struct blk_rq_stat stat[2];
2747 unsigned long ret = 0;
2750 * If stats collection isn't on, don't sleep but turn it on for
2753 if (!blk_stat_enable(q))
2757 * We don't have to do this once per IO, should optimize this
2758 * to just use the current window of stats until it changes
2760 memset(&stat, 0, sizeof(stat));
2761 blk_hctx_stat_get(hctx, stat);
2764 * As an optimistic guess, use half of the mean service time
2765 * for this type of request. We can (and should) make this smarter.
2766 * For instance, if the completion latencies are tight, we can
2767 * get closer than just half the mean. This is especially
2768 * important on devices where the completion latencies are longer
2771 if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples)
2772 ret = (stat[BLK_STAT_READ].mean + 1) / 2;
2773 else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples)
2774 ret = (stat[BLK_STAT_WRITE].mean + 1) / 2;
2779 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2780 struct blk_mq_hw_ctx *hctx,
2783 struct hrtimer_sleeper hs;
2784 enum hrtimer_mode mode;
2788 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2794 * -1: don't ever hybrid sleep
2795 * 0: use half of prev avg
2796 * >0: use this specific value
2798 if (q->poll_nsec == -1)
2800 else if (q->poll_nsec > 0)
2801 nsecs = q->poll_nsec;
2803 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2808 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2811 * This will be replaced with the stats tracking code, using
2812 * 'avg_completion_time / 2' as the pre-sleep target.
2816 mode = HRTIMER_MODE_REL;
2817 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2818 hrtimer_set_expires(&hs.timer, kt);
2820 hrtimer_init_sleeper(&hs, current);
2822 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2824 set_current_state(TASK_UNINTERRUPTIBLE);
2825 hrtimer_start_expires(&hs.timer, mode);
2828 hrtimer_cancel(&hs.timer);
2829 mode = HRTIMER_MODE_ABS;
2830 } while (hs.task && !signal_pending(current));
2832 __set_current_state(TASK_RUNNING);
2833 destroy_hrtimer_on_stack(&hs.timer);
2837 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2839 struct request_queue *q = hctx->queue;
2843 * If we sleep, have the caller restart the poll loop to reset
2844 * the state. Like for the other success return cases, the
2845 * caller is responsible for checking if the IO completed. If
2846 * the IO isn't complete, we'll get called again and will go
2847 * straight to the busy poll loop.
2849 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2852 hctx->poll_considered++;
2854 state = current->state;
2855 while (!need_resched()) {
2858 hctx->poll_invoked++;
2860 ret = q->mq_ops->poll(hctx, rq->tag);
2862 hctx->poll_success++;
2863 set_current_state(TASK_RUNNING);
2867 if (signal_pending_state(state, current))
2868 set_current_state(TASK_RUNNING);
2870 if (current->state == TASK_RUNNING)
2880 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2882 struct blk_mq_hw_ctx *hctx;
2883 struct blk_plug *plug;
2886 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2887 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2890 plug = current->plug;
2892 blk_flush_plug_list(plug, false);
2894 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2895 if (!blk_qc_t_is_internal(cookie))
2896 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2898 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2900 return __blk_mq_poll(hctx, rq);
2902 EXPORT_SYMBOL_GPL(blk_mq_poll);
2904 void blk_mq_disable_hotplug(void)
2906 mutex_lock(&all_q_mutex);
2909 void blk_mq_enable_hotplug(void)
2911 mutex_unlock(&all_q_mutex);
2914 static int __init blk_mq_init(void)
2916 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2917 blk_mq_hctx_notify_dead);
2919 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2920 blk_mq_queue_reinit_prepare,
2921 blk_mq_queue_reinit_dead);
2924 subsys_initcall(blk_mq_init);