2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-tag.h"
37 #include "blk-mq-sched.h"
39 static DEFINE_MUTEX(all_q_mutex);
40 static LIST_HEAD(all_q_list);
42 static void blk_mq_poll_stats_start(struct request_queue *q);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
46 * Check if any of the ctx's have pending work in this hardware queue
48 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
50 return sbitmap_any_bit_set(&hctx->ctx_map) ||
51 !list_empty_careful(&hctx->dispatch) ||
52 blk_mq_sched_has_work(hctx);
56 * Mark this ctx as having pending work in this hardware queue
58 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
59 struct blk_mq_ctx *ctx)
61 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
62 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
65 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
66 struct blk_mq_ctx *ctx)
68 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
71 void blk_freeze_queue_start(struct request_queue *q)
75 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
76 if (freeze_depth == 1) {
77 percpu_ref_kill(&q->q_usage_counter);
78 blk_mq_run_hw_queues(q, false);
81 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
83 void blk_mq_freeze_queue_wait(struct request_queue *q)
85 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
87 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
89 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
90 unsigned long timeout)
92 return wait_event_timeout(q->mq_freeze_wq,
93 percpu_ref_is_zero(&q->q_usage_counter),
96 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
102 void blk_freeze_queue(struct request_queue *q)
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
111 blk_freeze_queue_start(q);
112 blk_mq_freeze_queue_wait(q);
115 void blk_mq_freeze_queue(struct request_queue *q)
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
125 void blk_mq_unfreeze_queue(struct request_queue *q)
129 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
130 WARN_ON_ONCE(freeze_depth < 0);
132 percpu_ref_reinit(&q->q_usage_counter);
133 wake_up_all(&q->mq_freeze_wq);
136 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
139 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
142 * Note: this function does not prevent that the struct request end_io()
143 * callback function is invoked. Additionally, it is not prevented that
144 * new queue_rq() calls occur unless the queue has been stopped first.
146 void blk_mq_quiesce_queue(struct request_queue *q)
148 struct blk_mq_hw_ctx *hctx;
152 blk_mq_stop_hw_queues(q);
154 queue_for_each_hw_ctx(q, hctx, i) {
155 if (hctx->flags & BLK_MQ_F_BLOCKING)
156 synchronize_srcu(&hctx->queue_rq_srcu);
163 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
165 void blk_mq_wake_waiters(struct request_queue *q)
167 struct blk_mq_hw_ctx *hctx;
170 queue_for_each_hw_ctx(q, hctx, i)
171 if (blk_mq_hw_queue_mapped(hctx))
172 blk_mq_tag_wakeup_all(hctx->tags, true);
175 * If we are called because the queue has now been marked as
176 * dying, we need to ensure that processes currently waiting on
177 * the queue are notified as well.
179 wake_up_all(&q->mq_freeze_wq);
182 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
184 return blk_mq_has_free_tags(hctx->tags);
186 EXPORT_SYMBOL(blk_mq_can_queue);
188 void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
189 struct request *rq, unsigned int op)
191 INIT_LIST_HEAD(&rq->queuelist);
192 /* csd/requeue_work/fifo_time is initialized before use */
196 if (blk_queue_io_stat(q))
197 rq->rq_flags |= RQF_IO_STAT;
198 /* do not touch atomic flags, it needs atomic ops against the timer */
200 INIT_HLIST_NODE(&rq->hash);
201 RB_CLEAR_NODE(&rq->rb_node);
204 rq->start_time = jiffies;
205 #ifdef CONFIG_BLK_CGROUP
207 set_start_time_ns(rq);
208 rq->io_start_time_ns = 0;
210 rq->nr_phys_segments = 0;
211 #if defined(CONFIG_BLK_DEV_INTEGRITY)
212 rq->nr_integrity_segments = 0;
215 /* tag was already set */
219 INIT_LIST_HEAD(&rq->timeout_list);
223 rq->end_io_data = NULL;
226 ctx->rq_dispatched[op_is_sync(op)]++;
228 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init);
230 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
236 tag = blk_mq_get_tag(data);
237 if (tag != BLK_MQ_TAG_FAIL) {
238 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
240 rq = tags->static_rqs[tag];
242 if (data->flags & BLK_MQ_REQ_INTERNAL) {
244 rq->internal_tag = tag;
246 if (blk_mq_tag_busy(data->hctx)) {
247 rq->rq_flags = RQF_MQ_INFLIGHT;
248 atomic_inc(&data->hctx->nr_active);
251 rq->internal_tag = -1;
252 data->hctx->tags->rqs[rq->tag] = rq;
255 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
261 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
263 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
266 struct blk_mq_alloc_data alloc_data = { .flags = flags };
270 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
274 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
276 blk_mq_put_ctx(alloc_data.ctx);
280 return ERR_PTR(-EWOULDBLOCK);
283 rq->__sector = (sector_t) -1;
284 rq->bio = rq->biotail = NULL;
287 EXPORT_SYMBOL(blk_mq_alloc_request);
289 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
290 unsigned int flags, unsigned int hctx_idx)
292 struct blk_mq_alloc_data alloc_data = { .flags = flags };
298 * If the tag allocator sleeps we could get an allocation for a
299 * different hardware context. No need to complicate the low level
300 * allocator for this for the rare use case of a command tied to
303 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
304 return ERR_PTR(-EINVAL);
306 if (hctx_idx >= q->nr_hw_queues)
307 return ERR_PTR(-EIO);
309 ret = blk_queue_enter(q, true);
314 * Check if the hardware context is actually mapped to anything.
315 * If not tell the caller that it should skip this queue.
317 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
318 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
320 return ERR_PTR(-EXDEV);
322 cpu = cpumask_first(alloc_data.hctx->cpumask);
323 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
325 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
330 return ERR_PTR(-EWOULDBLOCK);
334 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
336 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
339 const int sched_tag = rq->internal_tag;
340 struct request_queue *q = rq->q;
342 if (rq->rq_flags & RQF_MQ_INFLIGHT)
343 atomic_dec(&hctx->nr_active);
345 wbt_done(q->rq_wb, &rq->issue_stat);
348 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
349 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
351 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
353 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
354 blk_mq_sched_restart(hctx);
358 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
361 struct blk_mq_ctx *ctx = rq->mq_ctx;
363 ctx->rq_completed[rq_is_sync(rq)]++;
364 __blk_mq_finish_request(hctx, ctx, rq);
367 void blk_mq_finish_request(struct request *rq)
369 blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
371 EXPORT_SYMBOL_GPL(blk_mq_finish_request);
373 void blk_mq_free_request(struct request *rq)
375 blk_mq_sched_put_request(rq);
377 EXPORT_SYMBOL_GPL(blk_mq_free_request);
379 inline void __blk_mq_end_request(struct request *rq, int error)
381 blk_account_io_done(rq);
384 wbt_done(rq->q->rq_wb, &rq->issue_stat);
385 rq->end_io(rq, error);
387 if (unlikely(blk_bidi_rq(rq)))
388 blk_mq_free_request(rq->next_rq);
389 blk_mq_free_request(rq);
392 EXPORT_SYMBOL(__blk_mq_end_request);
394 void blk_mq_end_request(struct request *rq, int error)
396 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
398 __blk_mq_end_request(rq, error);
400 EXPORT_SYMBOL(blk_mq_end_request);
402 static void __blk_mq_complete_request_remote(void *data)
404 struct request *rq = data;
406 rq->q->softirq_done_fn(rq);
409 static void blk_mq_ipi_complete_request(struct request *rq)
411 struct blk_mq_ctx *ctx = rq->mq_ctx;
415 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
416 rq->q->softirq_done_fn(rq);
421 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
422 shared = cpus_share_cache(cpu, ctx->cpu);
424 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
425 rq->csd.func = __blk_mq_complete_request_remote;
428 smp_call_function_single_async(ctx->cpu, &rq->csd);
430 rq->q->softirq_done_fn(rq);
435 static void blk_mq_stat_add(struct request *rq)
437 if (rq->rq_flags & RQF_STATS) {
438 blk_mq_poll_stats_start(rq->q);
443 static void __blk_mq_complete_request(struct request *rq)
445 if (rq->internal_tag != -1)
446 blk_mq_sched_completed_request(rq);
448 blk_mq_ipi_complete_request(rq);
452 * blk_mq_complete_request - end I/O on a request
453 * @rq: the request being processed
456 * Ends all I/O on a request. It does not handle partial completions.
457 * The actual completion happens out-of-order, through a IPI handler.
459 void blk_mq_complete_request(struct request *rq)
461 struct request_queue *q = rq->q;
463 if (unlikely(blk_should_fake_timeout(q)))
465 if (!blk_mark_rq_complete(rq))
466 __blk_mq_complete_request(rq);
468 EXPORT_SYMBOL(blk_mq_complete_request);
470 int blk_mq_request_started(struct request *rq)
472 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
474 EXPORT_SYMBOL_GPL(blk_mq_request_started);
476 void blk_mq_start_request(struct request *rq)
478 struct request_queue *q = rq->q;
480 blk_mq_sched_started_request(rq);
482 trace_block_rq_issue(q, rq);
484 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
485 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
486 rq->rq_flags |= RQF_STATS;
487 wbt_issue(q->rq_wb, &rq->issue_stat);
493 * Ensure that ->deadline is visible before set the started
494 * flag and clear the completed flag.
496 smp_mb__before_atomic();
499 * Mark us as started and clear complete. Complete might have been
500 * set if requeue raced with timeout, which then marked it as
501 * complete. So be sure to clear complete again when we start
502 * the request, otherwise we'll ignore the completion event.
504 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
505 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
506 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
507 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
509 if (q->dma_drain_size && blk_rq_bytes(rq)) {
511 * Make sure space for the drain appears. We know we can do
512 * this because max_hw_segments has been adjusted to be one
513 * fewer than the device can handle.
515 rq->nr_phys_segments++;
518 EXPORT_SYMBOL(blk_mq_start_request);
521 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
522 * flag isn't set yet, so there may be race with timeout handler,
523 * but given rq->deadline is just set in .queue_rq() under
524 * this situation, the race won't be possible in reality because
525 * rq->timeout should be set as big enough to cover the window
526 * between blk_mq_start_request() called from .queue_rq() and
527 * clearing REQ_ATOM_STARTED here.
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 * 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))
703 * The rq being checked may have been freed and reallocated
704 * out already here, we avoid this race by checking rq->deadline
705 * and REQ_ATOM_COMPLETE flag together:
707 * - if rq->deadline is observed as new value because of
708 * reusing, the rq won't be timed out because of timing.
709 * - if rq->deadline is observed as previous value,
710 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
711 * because we put a barrier between setting rq->deadline
712 * and clearing the flag in blk_mq_start_request(), so
713 * this rq won't be timed out too.
715 if (time_after_eq(jiffies, rq->deadline)) {
716 if (!blk_mark_rq_complete(rq))
717 blk_mq_rq_timed_out(rq, reserved);
718 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
719 data->next = rq->deadline;
724 static void blk_mq_timeout_work(struct work_struct *work)
726 struct request_queue *q =
727 container_of(work, struct request_queue, timeout_work);
728 struct blk_mq_timeout_data data = {
734 /* A deadlock might occur if a request is stuck requiring a
735 * timeout at the same time a queue freeze is waiting
736 * completion, since the timeout code would not be able to
737 * acquire the queue reference here.
739 * That's why we don't use blk_queue_enter here; instead, we use
740 * percpu_ref_tryget directly, because we need to be able to
741 * obtain a reference even in the short window between the queue
742 * starting to freeze, by dropping the first reference in
743 * blk_freeze_queue_start, and the moment the last request is
744 * consumed, marked by the instant q_usage_counter reaches
747 if (!percpu_ref_tryget(&q->q_usage_counter))
750 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
753 data.next = blk_rq_timeout(round_jiffies_up(data.next));
754 mod_timer(&q->timeout, data.next);
756 struct blk_mq_hw_ctx *hctx;
758 queue_for_each_hw_ctx(q, hctx, i) {
759 /* the hctx may be unmapped, so check it here */
760 if (blk_mq_hw_queue_mapped(hctx))
761 blk_mq_tag_idle(hctx);
768 * Reverse check our software queue for entries that we could potentially
769 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
770 * too much time checking for merges.
772 static bool blk_mq_attempt_merge(struct request_queue *q,
773 struct blk_mq_ctx *ctx, struct bio *bio)
778 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
784 if (!blk_rq_merge_ok(rq, bio))
787 switch (blk_try_merge(rq, bio)) {
788 case ELEVATOR_BACK_MERGE:
789 if (blk_mq_sched_allow_merge(q, rq, bio))
790 merged = bio_attempt_back_merge(q, rq, bio);
792 case ELEVATOR_FRONT_MERGE:
793 if (blk_mq_sched_allow_merge(q, rq, bio))
794 merged = bio_attempt_front_merge(q, rq, bio);
796 case ELEVATOR_DISCARD_MERGE:
797 merged = bio_attempt_discard_merge(q, rq, bio);
811 struct flush_busy_ctx_data {
812 struct blk_mq_hw_ctx *hctx;
813 struct list_head *list;
816 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
818 struct flush_busy_ctx_data *flush_data = data;
819 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
820 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
822 sbitmap_clear_bit(sb, bitnr);
823 spin_lock(&ctx->lock);
824 list_splice_tail_init(&ctx->rq_list, flush_data->list);
825 spin_unlock(&ctx->lock);
830 * Process software queues that have been marked busy, splicing them
831 * to the for-dispatch
833 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
835 struct flush_busy_ctx_data data = {
840 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
842 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
844 static inline unsigned int queued_to_index(unsigned int queued)
849 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
852 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
855 struct blk_mq_alloc_data data = {
857 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
858 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
864 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
865 data.flags |= BLK_MQ_REQ_RESERVED;
867 rq->tag = blk_mq_get_tag(&data);
869 if (blk_mq_tag_busy(data.hctx)) {
870 rq->rq_flags |= RQF_MQ_INFLIGHT;
871 atomic_inc(&data.hctx->nr_active);
873 data.hctx->tags->rqs[rq->tag] = rq;
879 return rq->tag != -1;
882 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
885 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
888 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
889 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
890 atomic_dec(&hctx->nr_active);
894 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
897 if (rq->tag == -1 || rq->internal_tag == -1)
900 __blk_mq_put_driver_tag(hctx, rq);
903 static void blk_mq_put_driver_tag(struct request *rq)
905 struct blk_mq_hw_ctx *hctx;
907 if (rq->tag == -1 || rq->internal_tag == -1)
910 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
911 __blk_mq_put_driver_tag(hctx, rq);
915 * If we fail getting a driver tag because all the driver tags are already
916 * assigned and on the dispatch list, BUT the first entry does not have a
917 * tag, then we could deadlock. For that case, move entries with assigned
918 * driver tags to the front, leaving the set of tagged requests in the
919 * same order, and the untagged set in the same order.
921 static bool reorder_tags_to_front(struct list_head *list)
923 struct request *rq, *tmp, *first = NULL;
925 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
929 list_move(&rq->queuelist, list);
935 return first != NULL;
938 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
941 struct blk_mq_hw_ctx *hctx;
943 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
945 list_del(&wait->task_list);
946 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
947 blk_mq_run_hw_queue(hctx, true);
951 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
953 struct sbq_wait_state *ws;
956 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
957 * The thread which wins the race to grab this bit adds the hardware
958 * queue to the wait queue.
960 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
961 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
964 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
965 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
968 * As soon as this returns, it's no longer safe to fiddle with
969 * hctx->dispatch_wait, since a completion can wake up the wait queue
970 * and unlock the bit.
972 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
976 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
978 struct blk_mq_hw_ctx *hctx;
980 int errors, queued, ret = BLK_MQ_RQ_QUEUE_OK;
982 if (list_empty(list))
986 * Now process all the entries, sending them to the driver.
990 struct blk_mq_queue_data bd;
992 rq = list_first_entry(list, struct request, queuelist);
993 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
994 if (!queued && reorder_tags_to_front(list))
998 * The initial allocation attempt failed, so we need to
999 * rerun the hardware queue when a tag is freed.
1001 if (!blk_mq_dispatch_wait_add(hctx))
1005 * It's possible that a tag was freed in the window
1006 * between the allocation failure and adding the
1007 * hardware queue to the wait queue.
1009 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1013 list_del_init(&rq->queuelist);
1018 * Flag last if we have no more requests, or if we have more
1019 * but can't assign a driver tag to it.
1021 if (list_empty(list))
1024 struct request *nxt;
1026 nxt = list_first_entry(list, struct request, queuelist);
1027 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1030 ret = q->mq_ops->queue_rq(hctx, &bd);
1032 case BLK_MQ_RQ_QUEUE_OK:
1035 case BLK_MQ_RQ_QUEUE_BUSY:
1036 blk_mq_put_driver_tag_hctx(hctx, rq);
1037 list_add(&rq->queuelist, list);
1038 __blk_mq_requeue_request(rq);
1041 pr_err("blk-mq: bad return on queue: %d\n", ret);
1042 case BLK_MQ_RQ_QUEUE_ERROR:
1045 blk_mq_end_request(rq, rq->errors);
1049 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
1051 } while (!list_empty(list));
1053 hctx->dispatched[queued_to_index(queued)]++;
1056 * Any items that need requeuing? Stuff them into hctx->dispatch,
1057 * that is where we will continue on next queue run.
1059 if (!list_empty(list)) {
1061 * If an I/O scheduler has been configured and we got a driver
1062 * tag for the next request already, free it again.
1064 rq = list_first_entry(list, struct request, queuelist);
1065 blk_mq_put_driver_tag(rq);
1067 spin_lock(&hctx->lock);
1068 list_splice_init(list, &hctx->dispatch);
1069 spin_unlock(&hctx->lock);
1072 * If SCHED_RESTART was set by the caller of this function and
1073 * it is no longer set that means that it was cleared by another
1074 * thread and hence that a queue rerun is needed.
1076 * If TAG_WAITING is set that means that an I/O scheduler has
1077 * been configured and another thread is waiting for a driver
1078 * tag. To guarantee fairness, do not rerun this hardware queue
1079 * but let the other thread grab the driver tag.
1081 * If no I/O scheduler has been configured it is possible that
1082 * the hardware queue got stopped and restarted before requests
1083 * were pushed back onto the dispatch list. Rerun the queue to
1084 * avoid starvation. Notes:
1085 * - blk_mq_run_hw_queue() checks whether or not a queue has
1086 * been stopped before rerunning a queue.
1087 * - Some but not all block drivers stop a queue before
1088 * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq
1091 if (!blk_mq_sched_needs_restart(hctx) &&
1092 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1093 blk_mq_run_hw_queue(hctx, true);
1096 return (queued + errors) != 0;
1099 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1103 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1104 cpu_online(hctx->next_cpu));
1106 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1108 blk_mq_sched_dispatch_requests(hctx);
1113 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1114 blk_mq_sched_dispatch_requests(hctx);
1115 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1120 * It'd be great if the workqueue API had a way to pass
1121 * in a mask and had some smarts for more clever placement.
1122 * For now we just round-robin here, switching for every
1123 * BLK_MQ_CPU_WORK_BATCH queued items.
1125 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1127 if (hctx->queue->nr_hw_queues == 1)
1128 return WORK_CPU_UNBOUND;
1130 if (--hctx->next_cpu_batch <= 0) {
1133 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1134 if (next_cpu >= nr_cpu_ids)
1135 next_cpu = cpumask_first(hctx->cpumask);
1137 hctx->next_cpu = next_cpu;
1138 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1141 return hctx->next_cpu;
1144 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1145 unsigned long msecs)
1147 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1148 !blk_mq_hw_queue_mapped(hctx)))
1151 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1152 int cpu = get_cpu();
1153 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1154 __blk_mq_run_hw_queue(hctx);
1163 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx),
1166 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1167 &hctx->delayed_run_work,
1168 msecs_to_jiffies(msecs));
1171 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1173 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1175 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1177 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1179 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1181 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1183 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1185 struct blk_mq_hw_ctx *hctx;
1188 queue_for_each_hw_ctx(q, hctx, i) {
1189 if (!blk_mq_hctx_has_pending(hctx) ||
1190 blk_mq_hctx_stopped(hctx))
1193 blk_mq_run_hw_queue(hctx, async);
1196 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1199 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1200 * @q: request queue.
1202 * The caller is responsible for serializing this function against
1203 * blk_mq_{start,stop}_hw_queue().
1205 bool blk_mq_queue_stopped(struct request_queue *q)
1207 struct blk_mq_hw_ctx *hctx;
1210 queue_for_each_hw_ctx(q, hctx, i)
1211 if (blk_mq_hctx_stopped(hctx))
1216 EXPORT_SYMBOL(blk_mq_queue_stopped);
1218 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1220 cancel_work(&hctx->run_work);
1221 cancel_delayed_work(&hctx->delay_work);
1222 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1224 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1226 void blk_mq_stop_hw_queues(struct request_queue *q)
1228 struct blk_mq_hw_ctx *hctx;
1231 queue_for_each_hw_ctx(q, hctx, i)
1232 blk_mq_stop_hw_queue(hctx);
1234 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1236 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1238 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1240 blk_mq_run_hw_queue(hctx, false);
1242 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1244 void blk_mq_start_hw_queues(struct request_queue *q)
1246 struct blk_mq_hw_ctx *hctx;
1249 queue_for_each_hw_ctx(q, hctx, i)
1250 blk_mq_start_hw_queue(hctx);
1252 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1254 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1256 if (!blk_mq_hctx_stopped(hctx))
1259 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1260 blk_mq_run_hw_queue(hctx, async);
1262 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1264 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1266 struct blk_mq_hw_ctx *hctx;
1269 queue_for_each_hw_ctx(q, hctx, i)
1270 blk_mq_start_stopped_hw_queue(hctx, async);
1272 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1274 static void blk_mq_run_work_fn(struct work_struct *work)
1276 struct blk_mq_hw_ctx *hctx;
1278 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1280 __blk_mq_run_hw_queue(hctx);
1283 static void blk_mq_delayed_run_work_fn(struct work_struct *work)
1285 struct blk_mq_hw_ctx *hctx;
1287 hctx = container_of(work, struct blk_mq_hw_ctx, delayed_run_work.work);
1289 __blk_mq_run_hw_queue(hctx);
1292 static void blk_mq_delay_work_fn(struct work_struct *work)
1294 struct blk_mq_hw_ctx *hctx;
1296 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1298 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1299 __blk_mq_run_hw_queue(hctx);
1302 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1304 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1307 blk_mq_stop_hw_queue(hctx);
1308 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1309 &hctx->delay_work, msecs_to_jiffies(msecs));
1311 EXPORT_SYMBOL(blk_mq_delay_queue);
1313 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1317 struct blk_mq_ctx *ctx = rq->mq_ctx;
1319 trace_block_rq_insert(hctx->queue, rq);
1322 list_add(&rq->queuelist, &ctx->rq_list);
1324 list_add_tail(&rq->queuelist, &ctx->rq_list);
1327 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1330 struct blk_mq_ctx *ctx = rq->mq_ctx;
1332 __blk_mq_insert_req_list(hctx, rq, at_head);
1333 blk_mq_hctx_mark_pending(hctx, ctx);
1336 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1337 struct list_head *list)
1341 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1344 spin_lock(&ctx->lock);
1345 while (!list_empty(list)) {
1348 rq = list_first_entry(list, struct request, queuelist);
1349 BUG_ON(rq->mq_ctx != ctx);
1350 list_del_init(&rq->queuelist);
1351 __blk_mq_insert_req_list(hctx, rq, false);
1353 blk_mq_hctx_mark_pending(hctx, ctx);
1354 spin_unlock(&ctx->lock);
1357 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1359 struct request *rqa = container_of(a, struct request, queuelist);
1360 struct request *rqb = container_of(b, struct request, queuelist);
1362 return !(rqa->mq_ctx < rqb->mq_ctx ||
1363 (rqa->mq_ctx == rqb->mq_ctx &&
1364 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1367 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1369 struct blk_mq_ctx *this_ctx;
1370 struct request_queue *this_q;
1373 LIST_HEAD(ctx_list);
1376 list_splice_init(&plug->mq_list, &list);
1378 list_sort(NULL, &list, plug_ctx_cmp);
1384 while (!list_empty(&list)) {
1385 rq = list_entry_rq(list.next);
1386 list_del_init(&rq->queuelist);
1388 if (rq->mq_ctx != this_ctx) {
1390 trace_block_unplug(this_q, depth, from_schedule);
1391 blk_mq_sched_insert_requests(this_q, this_ctx,
1396 this_ctx = rq->mq_ctx;
1402 list_add_tail(&rq->queuelist, &ctx_list);
1406 * If 'this_ctx' is set, we know we have entries to complete
1407 * on 'ctx_list'. Do those.
1410 trace_block_unplug(this_q, depth, from_schedule);
1411 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1416 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1418 blk_init_request_from_bio(rq, bio);
1420 blk_account_io_start(rq, true);
1423 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1425 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1426 !blk_queue_nomerges(hctx->queue);
1429 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1430 struct blk_mq_ctx *ctx,
1431 struct request *rq, struct bio *bio)
1433 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1434 blk_mq_bio_to_request(rq, bio);
1435 spin_lock(&ctx->lock);
1437 __blk_mq_insert_request(hctx, rq, false);
1438 spin_unlock(&ctx->lock);
1441 struct request_queue *q = hctx->queue;
1443 spin_lock(&ctx->lock);
1444 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1445 blk_mq_bio_to_request(rq, bio);
1449 spin_unlock(&ctx->lock);
1450 __blk_mq_finish_request(hctx, ctx, rq);
1455 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1458 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1460 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1463 static void __blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie,
1466 struct request_queue *q = rq->q;
1467 struct blk_mq_queue_data bd = {
1471 struct blk_mq_hw_ctx *hctx;
1472 blk_qc_t new_cookie;
1478 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1481 new_cookie = request_to_qc_t(hctx, rq);
1484 * For OK queue, we are done. For error, kill it. Any other
1485 * error (busy), just add it to our list as we previously
1488 ret = q->mq_ops->queue_rq(hctx, &bd);
1489 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1490 *cookie = new_cookie;
1494 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1495 *cookie = BLK_QC_T_NONE;
1497 blk_mq_end_request(rq, rq->errors);
1501 __blk_mq_requeue_request(rq);
1503 blk_mq_sched_insert_request(rq, false, true, false, may_sleep);
1506 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1507 struct request *rq, blk_qc_t *cookie)
1509 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1511 __blk_mq_try_issue_directly(rq, cookie, false);
1514 unsigned int srcu_idx;
1518 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1519 __blk_mq_try_issue_directly(rq, cookie, true);
1520 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1524 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1526 const int is_sync = op_is_sync(bio->bi_opf);
1527 const int is_flush_fua = op_is_flush(bio->bi_opf);
1528 struct blk_mq_alloc_data data = { .flags = 0 };
1530 unsigned int request_count = 0;
1531 struct blk_plug *plug;
1532 struct request *same_queue_rq = NULL;
1534 unsigned int wb_acct;
1536 blk_queue_bounce(q, &bio);
1538 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1540 return BLK_QC_T_NONE;
1543 blk_queue_split(q, &bio, q->bio_split);
1545 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1546 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1547 return BLK_QC_T_NONE;
1549 if (blk_mq_sched_bio_merge(q, bio))
1550 return BLK_QC_T_NONE;
1552 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1554 trace_block_getrq(q, bio, bio->bi_opf);
1556 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1557 if (unlikely(!rq)) {
1558 __wbt_done(q->rq_wb, wb_acct);
1559 return BLK_QC_T_NONE;
1562 wbt_track(&rq->issue_stat, wb_acct);
1564 cookie = request_to_qc_t(data.hctx, rq);
1566 plug = current->plug;
1567 if (unlikely(is_flush_fua)) {
1568 blk_mq_bio_to_request(rq, bio);
1570 blk_mq_sched_insert_request(rq, false, true, true,
1573 blk_insert_flush(rq);
1574 blk_mq_run_hw_queue(data.hctx, true);
1576 } else if (plug && q->nr_hw_queues == 1) {
1577 struct request *last = NULL;
1579 blk_mq_bio_to_request(rq, bio);
1582 * @request_count may become stale because of schedule
1583 * out, so check the list again.
1585 if (list_empty(&plug->mq_list))
1587 else if (blk_queue_nomerges(q))
1588 request_count = blk_plug_queued_count(q);
1591 trace_block_plug(q);
1593 last = list_entry_rq(plug->mq_list.prev);
1595 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1596 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1597 blk_flush_plug_list(plug, false);
1598 trace_block_plug(q);
1601 list_add_tail(&rq->queuelist, &plug->mq_list);
1602 } else if (plug && !blk_queue_nomerges(q)) {
1603 blk_mq_bio_to_request(rq, bio);
1606 * We do limited plugging. If the bio can be merged, do that.
1607 * Otherwise the existing request in the plug list will be
1608 * issued. So the plug list will have one request at most
1609 * The plug list might get flushed before this. If that happens,
1610 * the plug list is empty, and same_queue_rq is invalid.
1612 if (list_empty(&plug->mq_list))
1613 same_queue_rq = NULL;
1615 list_del_init(&same_queue_rq->queuelist);
1616 list_add_tail(&rq->queuelist, &plug->mq_list);
1618 blk_mq_put_ctx(data.ctx);
1621 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1625 } else if (q->nr_hw_queues > 1 && is_sync) {
1626 blk_mq_put_ctx(data.ctx);
1627 blk_mq_bio_to_request(rq, bio);
1628 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1630 } else if (q->elevator) {
1631 blk_mq_bio_to_request(rq, bio);
1632 blk_mq_sched_insert_request(rq, false, true, true, true);
1633 } else if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio))
1634 blk_mq_run_hw_queue(data.hctx, true);
1636 blk_mq_put_ctx(data.ctx);
1640 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1641 unsigned int hctx_idx)
1645 if (tags->rqs && set->ops->exit_request) {
1648 for (i = 0; i < tags->nr_tags; i++) {
1649 struct request *rq = tags->static_rqs[i];
1653 set->ops->exit_request(set->driver_data, rq,
1655 tags->static_rqs[i] = NULL;
1659 while (!list_empty(&tags->page_list)) {
1660 page = list_first_entry(&tags->page_list, struct page, lru);
1661 list_del_init(&page->lru);
1663 * Remove kmemleak object previously allocated in
1664 * blk_mq_init_rq_map().
1666 kmemleak_free(page_address(page));
1667 __free_pages(page, page->private);
1671 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1675 kfree(tags->static_rqs);
1676 tags->static_rqs = NULL;
1678 blk_mq_free_tags(tags);
1681 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1682 unsigned int hctx_idx,
1683 unsigned int nr_tags,
1684 unsigned int reserved_tags)
1686 struct blk_mq_tags *tags;
1689 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1690 if (node == NUMA_NO_NODE)
1691 node = set->numa_node;
1693 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1694 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1698 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1699 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1702 blk_mq_free_tags(tags);
1706 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1707 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1709 if (!tags->static_rqs) {
1711 blk_mq_free_tags(tags);
1718 static size_t order_to_size(unsigned int order)
1720 return (size_t)PAGE_SIZE << order;
1723 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1724 unsigned int hctx_idx, unsigned int depth)
1726 unsigned int i, j, entries_per_page, max_order = 4;
1727 size_t rq_size, left;
1730 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1731 if (node == NUMA_NO_NODE)
1732 node = set->numa_node;
1734 INIT_LIST_HEAD(&tags->page_list);
1737 * rq_size is the size of the request plus driver payload, rounded
1738 * to the cacheline size
1740 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1742 left = rq_size * depth;
1744 for (i = 0; i < depth; ) {
1745 int this_order = max_order;
1750 while (this_order && left < order_to_size(this_order - 1))
1754 page = alloc_pages_node(node,
1755 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1761 if (order_to_size(this_order) < rq_size)
1768 page->private = this_order;
1769 list_add_tail(&page->lru, &tags->page_list);
1771 p = page_address(page);
1773 * Allow kmemleak to scan these pages as they contain pointers
1774 * to additional allocations like via ops->init_request().
1776 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1777 entries_per_page = order_to_size(this_order) / rq_size;
1778 to_do = min(entries_per_page, depth - i);
1779 left -= to_do * rq_size;
1780 for (j = 0; j < to_do; j++) {
1781 struct request *rq = p;
1783 tags->static_rqs[i] = rq;
1784 if (set->ops->init_request) {
1785 if (set->ops->init_request(set->driver_data,
1788 tags->static_rqs[i] = NULL;
1800 blk_mq_free_rqs(set, tags, hctx_idx);
1805 * 'cpu' is going away. splice any existing rq_list entries from this
1806 * software queue to the hw queue dispatch list, and ensure that it
1809 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1811 struct blk_mq_hw_ctx *hctx;
1812 struct blk_mq_ctx *ctx;
1815 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1816 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1818 spin_lock(&ctx->lock);
1819 if (!list_empty(&ctx->rq_list)) {
1820 list_splice_init(&ctx->rq_list, &tmp);
1821 blk_mq_hctx_clear_pending(hctx, ctx);
1823 spin_unlock(&ctx->lock);
1825 if (list_empty(&tmp))
1828 spin_lock(&hctx->lock);
1829 list_splice_tail_init(&tmp, &hctx->dispatch);
1830 spin_unlock(&hctx->lock);
1832 blk_mq_run_hw_queue(hctx, true);
1836 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1838 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1842 /* hctx->ctxs will be freed in queue's release handler */
1843 static void blk_mq_exit_hctx(struct request_queue *q,
1844 struct blk_mq_tag_set *set,
1845 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1847 unsigned flush_start_tag = set->queue_depth;
1849 blk_mq_tag_idle(hctx);
1851 if (set->ops->exit_request)
1852 set->ops->exit_request(set->driver_data,
1853 hctx->fq->flush_rq, hctx_idx,
1854 flush_start_tag + hctx_idx);
1856 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1858 if (set->ops->exit_hctx)
1859 set->ops->exit_hctx(hctx, hctx_idx);
1861 if (hctx->flags & BLK_MQ_F_BLOCKING)
1862 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1864 blk_mq_remove_cpuhp(hctx);
1865 blk_free_flush_queue(hctx->fq);
1866 sbitmap_free(&hctx->ctx_map);
1869 static void blk_mq_exit_hw_queues(struct request_queue *q,
1870 struct blk_mq_tag_set *set, int nr_queue)
1872 struct blk_mq_hw_ctx *hctx;
1875 queue_for_each_hw_ctx(q, hctx, i) {
1878 blk_mq_exit_hctx(q, set, hctx, i);
1882 static int blk_mq_init_hctx(struct request_queue *q,
1883 struct blk_mq_tag_set *set,
1884 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1887 unsigned flush_start_tag = set->queue_depth;
1889 node = hctx->numa_node;
1890 if (node == NUMA_NO_NODE)
1891 node = hctx->numa_node = set->numa_node;
1893 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1894 INIT_DELAYED_WORK(&hctx->delayed_run_work, blk_mq_delayed_run_work_fn);
1895 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1896 spin_lock_init(&hctx->lock);
1897 INIT_LIST_HEAD(&hctx->dispatch);
1899 hctx->queue_num = hctx_idx;
1900 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1902 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1904 hctx->tags = set->tags[hctx_idx];
1907 * Allocate space for all possible cpus to avoid allocation at
1910 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1913 goto unregister_cpu_notifier;
1915 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1921 if (set->ops->init_hctx &&
1922 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1925 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1928 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1930 goto sched_exit_hctx;
1932 if (set->ops->init_request &&
1933 set->ops->init_request(set->driver_data,
1934 hctx->fq->flush_rq, hctx_idx,
1935 flush_start_tag + hctx_idx, node))
1938 if (hctx->flags & BLK_MQ_F_BLOCKING)
1939 init_srcu_struct(&hctx->queue_rq_srcu);
1946 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1948 if (set->ops->exit_hctx)
1949 set->ops->exit_hctx(hctx, hctx_idx);
1951 sbitmap_free(&hctx->ctx_map);
1954 unregister_cpu_notifier:
1955 blk_mq_remove_cpuhp(hctx);
1959 static void blk_mq_init_cpu_queues(struct request_queue *q,
1960 unsigned int nr_hw_queues)
1964 for_each_possible_cpu(i) {
1965 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1966 struct blk_mq_hw_ctx *hctx;
1969 spin_lock_init(&__ctx->lock);
1970 INIT_LIST_HEAD(&__ctx->rq_list);
1973 /* If the cpu isn't online, the cpu is mapped to first hctx */
1977 hctx = blk_mq_map_queue(q, i);
1980 * Set local node, IFF we have more than one hw queue. If
1981 * not, we remain on the home node of the device
1983 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1984 hctx->numa_node = local_memory_node(cpu_to_node(i));
1988 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1992 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1993 set->queue_depth, set->reserved_tags);
1994 if (!set->tags[hctx_idx])
1997 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2002 blk_mq_free_rq_map(set->tags[hctx_idx]);
2003 set->tags[hctx_idx] = NULL;
2007 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2008 unsigned int hctx_idx)
2010 if (set->tags[hctx_idx]) {
2011 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2012 blk_mq_free_rq_map(set->tags[hctx_idx]);
2013 set->tags[hctx_idx] = NULL;
2017 static void blk_mq_map_swqueue(struct request_queue *q,
2018 const struct cpumask *online_mask)
2020 unsigned int i, hctx_idx;
2021 struct blk_mq_hw_ctx *hctx;
2022 struct blk_mq_ctx *ctx;
2023 struct blk_mq_tag_set *set = q->tag_set;
2026 * Avoid others reading imcomplete hctx->cpumask through sysfs
2028 mutex_lock(&q->sysfs_lock);
2030 queue_for_each_hw_ctx(q, hctx, i) {
2031 cpumask_clear(hctx->cpumask);
2036 * Map software to hardware queues
2038 for_each_possible_cpu(i) {
2039 /* If the cpu isn't online, the cpu is mapped to first hctx */
2040 if (!cpumask_test_cpu(i, online_mask))
2043 hctx_idx = q->mq_map[i];
2044 /* unmapped hw queue can be remapped after CPU topo changed */
2045 if (!set->tags[hctx_idx] &&
2046 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2048 * If tags initialization fail for some hctx,
2049 * that hctx won't be brought online. In this
2050 * case, remap the current ctx to hctx[0] which
2051 * is guaranteed to always have tags allocated
2056 ctx = per_cpu_ptr(q->queue_ctx, i);
2057 hctx = blk_mq_map_queue(q, i);
2059 cpumask_set_cpu(i, hctx->cpumask);
2060 ctx->index_hw = hctx->nr_ctx;
2061 hctx->ctxs[hctx->nr_ctx++] = ctx;
2064 mutex_unlock(&q->sysfs_lock);
2066 queue_for_each_hw_ctx(q, hctx, i) {
2068 * If no software queues are mapped to this hardware queue,
2069 * disable it and free the request entries.
2071 if (!hctx->nr_ctx) {
2072 /* Never unmap queue 0. We need it as a
2073 * fallback in case of a new remap fails
2076 if (i && set->tags[i])
2077 blk_mq_free_map_and_requests(set, i);
2083 hctx->tags = set->tags[i];
2084 WARN_ON(!hctx->tags);
2087 * Set the map size to the number of mapped software queues.
2088 * This is more accurate and more efficient than looping
2089 * over all possibly mapped software queues.
2091 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2094 * Initialize batch roundrobin counts
2096 hctx->next_cpu = cpumask_first(hctx->cpumask);
2097 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2101 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2103 struct blk_mq_hw_ctx *hctx;
2106 queue_for_each_hw_ctx(q, hctx, i) {
2108 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2110 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2114 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2116 struct request_queue *q;
2118 lockdep_assert_held(&set->tag_list_lock);
2120 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2121 blk_mq_freeze_queue(q);
2122 queue_set_hctx_shared(q, shared);
2123 blk_mq_unfreeze_queue(q);
2127 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2129 struct blk_mq_tag_set *set = q->tag_set;
2131 mutex_lock(&set->tag_list_lock);
2132 list_del_rcu(&q->tag_set_list);
2133 INIT_LIST_HEAD(&q->tag_set_list);
2134 if (list_is_singular(&set->tag_list)) {
2135 /* just transitioned to unshared */
2136 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2137 /* update existing queue */
2138 blk_mq_update_tag_set_depth(set, false);
2140 mutex_unlock(&set->tag_list_lock);
2145 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2146 struct request_queue *q)
2150 mutex_lock(&set->tag_list_lock);
2152 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2153 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2154 set->flags |= BLK_MQ_F_TAG_SHARED;
2155 /* update existing queue */
2156 blk_mq_update_tag_set_depth(set, true);
2158 if (set->flags & BLK_MQ_F_TAG_SHARED)
2159 queue_set_hctx_shared(q, true);
2160 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2162 mutex_unlock(&set->tag_list_lock);
2166 * It is the actual release handler for mq, but we do it from
2167 * request queue's release handler for avoiding use-after-free
2168 * and headache because q->mq_kobj shouldn't have been introduced,
2169 * but we can't group ctx/kctx kobj without it.
2171 void blk_mq_release(struct request_queue *q)
2173 struct blk_mq_hw_ctx *hctx;
2176 /* hctx kobj stays in hctx */
2177 queue_for_each_hw_ctx(q, hctx, i) {
2180 kobject_put(&hctx->kobj);
2185 kfree(q->queue_hw_ctx);
2188 * release .mq_kobj and sw queue's kobject now because
2189 * both share lifetime with request queue.
2191 blk_mq_sysfs_deinit(q);
2193 free_percpu(q->queue_ctx);
2196 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2198 struct request_queue *uninit_q, *q;
2200 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2202 return ERR_PTR(-ENOMEM);
2204 q = blk_mq_init_allocated_queue(set, uninit_q);
2206 blk_cleanup_queue(uninit_q);
2210 EXPORT_SYMBOL(blk_mq_init_queue);
2212 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2213 struct request_queue *q)
2216 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2218 blk_mq_sysfs_unregister(q);
2219 for (i = 0; i < set->nr_hw_queues; i++) {
2225 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2226 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2231 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2238 atomic_set(&hctxs[i]->nr_active, 0);
2239 hctxs[i]->numa_node = node;
2240 hctxs[i]->queue_num = i;
2242 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2243 free_cpumask_var(hctxs[i]->cpumask);
2248 blk_mq_hctx_kobj_init(hctxs[i]);
2250 for (j = i; j < q->nr_hw_queues; j++) {
2251 struct blk_mq_hw_ctx *hctx = hctxs[j];
2255 blk_mq_free_map_and_requests(set, j);
2256 blk_mq_exit_hctx(q, set, hctx, j);
2257 kobject_put(&hctx->kobj);
2262 q->nr_hw_queues = i;
2263 blk_mq_sysfs_register(q);
2266 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2267 struct request_queue *q)
2269 /* mark the queue as mq asap */
2270 q->mq_ops = set->ops;
2272 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2273 blk_stat_rq_ddir, 2, q);
2277 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2281 /* init q->mq_kobj and sw queues' kobjects */
2282 blk_mq_sysfs_init(q);
2284 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2285 GFP_KERNEL, set->numa_node);
2286 if (!q->queue_hw_ctx)
2289 q->mq_map = set->mq_map;
2291 blk_mq_realloc_hw_ctxs(set, q);
2292 if (!q->nr_hw_queues)
2295 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2296 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2298 q->nr_queues = nr_cpu_ids;
2300 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2302 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2303 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2305 q->sg_reserved_size = INT_MAX;
2307 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2308 INIT_LIST_HEAD(&q->requeue_list);
2309 spin_lock_init(&q->requeue_lock);
2311 blk_queue_make_request(q, blk_mq_make_request);
2314 * Do this after blk_queue_make_request() overrides it...
2316 q->nr_requests = set->queue_depth;
2319 * Default to classic polling
2323 if (set->ops->complete)
2324 blk_queue_softirq_done(q, set->ops->complete);
2326 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2329 mutex_lock(&all_q_mutex);
2331 list_add_tail(&q->all_q_node, &all_q_list);
2332 blk_mq_add_queue_tag_set(set, q);
2333 blk_mq_map_swqueue(q, cpu_online_mask);
2335 mutex_unlock(&all_q_mutex);
2338 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2341 ret = blk_mq_sched_init(q);
2343 return ERR_PTR(ret);
2349 kfree(q->queue_hw_ctx);
2351 free_percpu(q->queue_ctx);
2354 return ERR_PTR(-ENOMEM);
2356 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2358 void blk_mq_free_queue(struct request_queue *q)
2360 struct blk_mq_tag_set *set = q->tag_set;
2362 mutex_lock(&all_q_mutex);
2363 list_del_init(&q->all_q_node);
2364 mutex_unlock(&all_q_mutex);
2366 blk_mq_del_queue_tag_set(q);
2368 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2371 /* Basically redo blk_mq_init_queue with queue frozen */
2372 static void blk_mq_queue_reinit(struct request_queue *q,
2373 const struct cpumask *online_mask)
2375 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2377 blk_mq_sysfs_unregister(q);
2380 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2381 * we should change hctx numa_node according to new topology (this
2382 * involves free and re-allocate memory, worthy doing?)
2385 blk_mq_map_swqueue(q, online_mask);
2387 blk_mq_sysfs_register(q);
2391 * New online cpumask which is going to be set in this hotplug event.
2392 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2393 * one-by-one and dynamically allocating this could result in a failure.
2395 static struct cpumask cpuhp_online_new;
2397 static void blk_mq_queue_reinit_work(void)
2399 struct request_queue *q;
2401 mutex_lock(&all_q_mutex);
2403 * We need to freeze and reinit all existing queues. Freezing
2404 * involves synchronous wait for an RCU grace period and doing it
2405 * one by one may take a long time. Start freezing all queues in
2406 * one swoop and then wait for the completions so that freezing can
2407 * take place in parallel.
2409 list_for_each_entry(q, &all_q_list, all_q_node)
2410 blk_freeze_queue_start(q);
2411 list_for_each_entry(q, &all_q_list, all_q_node)
2412 blk_mq_freeze_queue_wait(q);
2414 list_for_each_entry(q, &all_q_list, all_q_node)
2415 blk_mq_queue_reinit(q, &cpuhp_online_new);
2417 list_for_each_entry(q, &all_q_list, all_q_node)
2418 blk_mq_unfreeze_queue(q);
2420 mutex_unlock(&all_q_mutex);
2423 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2425 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2426 blk_mq_queue_reinit_work();
2431 * Before hotadded cpu starts handling requests, new mappings must be
2432 * established. Otherwise, these requests in hw queue might never be
2435 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2436 * for CPU0, and ctx1 for CPU1).
2438 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2439 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2441 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2442 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2443 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2446 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2448 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2449 cpumask_set_cpu(cpu, &cpuhp_online_new);
2450 blk_mq_queue_reinit_work();
2454 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2458 for (i = 0; i < set->nr_hw_queues; i++)
2459 if (!__blk_mq_alloc_rq_map(set, i))
2466 blk_mq_free_rq_map(set->tags[i]);
2472 * Allocate the request maps associated with this tag_set. Note that this
2473 * may reduce the depth asked for, if memory is tight. set->queue_depth
2474 * will be updated to reflect the allocated depth.
2476 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2481 depth = set->queue_depth;
2483 err = __blk_mq_alloc_rq_maps(set);
2487 set->queue_depth >>= 1;
2488 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2492 } while (set->queue_depth);
2494 if (!set->queue_depth || err) {
2495 pr_err("blk-mq: failed to allocate request map\n");
2499 if (depth != set->queue_depth)
2500 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2501 depth, set->queue_depth);
2506 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2508 if (set->ops->map_queues)
2509 return set->ops->map_queues(set);
2511 return blk_mq_map_queues(set);
2515 * Alloc a tag set to be associated with one or more request queues.
2516 * May fail with EINVAL for various error conditions. May adjust the
2517 * requested depth down, if if it too large. In that case, the set
2518 * value will be stored in set->queue_depth.
2520 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2524 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2526 if (!set->nr_hw_queues)
2528 if (!set->queue_depth)
2530 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2533 if (!set->ops->queue_rq)
2536 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2537 pr_info("blk-mq: reduced tag depth to %u\n",
2539 set->queue_depth = BLK_MQ_MAX_DEPTH;
2543 * If a crashdump is active, then we are potentially in a very
2544 * memory constrained environment. Limit us to 1 queue and
2545 * 64 tags to prevent using too much memory.
2547 if (is_kdump_kernel()) {
2548 set->nr_hw_queues = 1;
2549 set->queue_depth = min(64U, set->queue_depth);
2552 * There is no use for more h/w queues than cpus.
2554 if (set->nr_hw_queues > nr_cpu_ids)
2555 set->nr_hw_queues = nr_cpu_ids;
2557 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2558 GFP_KERNEL, set->numa_node);
2563 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2564 GFP_KERNEL, set->numa_node);
2568 ret = blk_mq_update_queue_map(set);
2570 goto out_free_mq_map;
2572 ret = blk_mq_alloc_rq_maps(set);
2574 goto out_free_mq_map;
2576 mutex_init(&set->tag_list_lock);
2577 INIT_LIST_HEAD(&set->tag_list);
2589 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2591 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2595 for (i = 0; i < nr_cpu_ids; i++)
2596 blk_mq_free_map_and_requests(set, i);
2604 EXPORT_SYMBOL(blk_mq_free_tag_set);
2606 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2608 struct blk_mq_tag_set *set = q->tag_set;
2609 struct blk_mq_hw_ctx *hctx;
2615 blk_mq_freeze_queue(q);
2616 blk_mq_quiesce_queue(q);
2619 queue_for_each_hw_ctx(q, hctx, i) {
2623 * If we're using an MQ scheduler, just update the scheduler
2624 * queue depth. This is similar to what the old code would do.
2626 if (!hctx->sched_tags) {
2627 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2628 min(nr, set->queue_depth),
2631 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2639 q->nr_requests = nr;
2641 blk_mq_unfreeze_queue(q);
2642 blk_mq_start_stopped_hw_queues(q, true);
2647 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2649 struct request_queue *q;
2651 lockdep_assert_held(&set->tag_list_lock);
2653 if (nr_hw_queues > nr_cpu_ids)
2654 nr_hw_queues = nr_cpu_ids;
2655 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2658 list_for_each_entry(q, &set->tag_list, tag_set_list)
2659 blk_mq_freeze_queue(q);
2661 set->nr_hw_queues = nr_hw_queues;
2662 blk_mq_update_queue_map(set);
2663 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2664 blk_mq_realloc_hw_ctxs(set, q);
2665 blk_mq_queue_reinit(q, cpu_online_mask);
2668 list_for_each_entry(q, &set->tag_list, tag_set_list)
2669 blk_mq_unfreeze_queue(q);
2671 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2673 /* Enable polling stats and return whether they were already enabled. */
2674 static bool blk_poll_stats_enable(struct request_queue *q)
2676 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2677 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2679 blk_stat_add_callback(q, q->poll_cb);
2683 static void blk_mq_poll_stats_start(struct request_queue *q)
2686 * We don't arm the callback if polling stats are not enabled or the
2687 * callback is already active.
2689 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2690 blk_stat_is_active(q->poll_cb))
2693 blk_stat_activate_msecs(q->poll_cb, 100);
2696 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2698 struct request_queue *q = cb->data;
2700 if (cb->stat[READ].nr_samples)
2701 q->poll_stat[READ] = cb->stat[READ];
2702 if (cb->stat[WRITE].nr_samples)
2703 q->poll_stat[WRITE] = cb->stat[WRITE];
2706 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2707 struct blk_mq_hw_ctx *hctx,
2710 unsigned long ret = 0;
2713 * If stats collection isn't on, don't sleep but turn it on for
2716 if (!blk_poll_stats_enable(q))
2720 * As an optimistic guess, use half of the mean service time
2721 * for this type of request. We can (and should) make this smarter.
2722 * For instance, if the completion latencies are tight, we can
2723 * get closer than just half the mean. This is especially
2724 * important on devices where the completion latencies are longer
2727 if (req_op(rq) == REQ_OP_READ && q->poll_stat[READ].nr_samples)
2728 ret = (q->poll_stat[READ].mean + 1) / 2;
2729 else if (req_op(rq) == REQ_OP_WRITE && q->poll_stat[WRITE].nr_samples)
2730 ret = (q->poll_stat[WRITE].mean + 1) / 2;
2735 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2736 struct blk_mq_hw_ctx *hctx,
2739 struct hrtimer_sleeper hs;
2740 enum hrtimer_mode mode;
2744 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2750 * -1: don't ever hybrid sleep
2751 * 0: use half of prev avg
2752 * >0: use this specific value
2754 if (q->poll_nsec == -1)
2756 else if (q->poll_nsec > 0)
2757 nsecs = q->poll_nsec;
2759 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2764 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2767 * This will be replaced with the stats tracking code, using
2768 * 'avg_completion_time / 2' as the pre-sleep target.
2772 mode = HRTIMER_MODE_REL;
2773 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2774 hrtimer_set_expires(&hs.timer, kt);
2776 hrtimer_init_sleeper(&hs, current);
2778 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2780 set_current_state(TASK_UNINTERRUPTIBLE);
2781 hrtimer_start_expires(&hs.timer, mode);
2784 hrtimer_cancel(&hs.timer);
2785 mode = HRTIMER_MODE_ABS;
2786 } while (hs.task && !signal_pending(current));
2788 __set_current_state(TASK_RUNNING);
2789 destroy_hrtimer_on_stack(&hs.timer);
2793 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2795 struct request_queue *q = hctx->queue;
2799 * If we sleep, have the caller restart the poll loop to reset
2800 * the state. Like for the other success return cases, the
2801 * caller is responsible for checking if the IO completed. If
2802 * the IO isn't complete, we'll get called again and will go
2803 * straight to the busy poll loop.
2805 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2808 hctx->poll_considered++;
2810 state = current->state;
2811 while (!need_resched()) {
2814 hctx->poll_invoked++;
2816 ret = q->mq_ops->poll(hctx, rq->tag);
2818 hctx->poll_success++;
2819 set_current_state(TASK_RUNNING);
2823 if (signal_pending_state(state, current))
2824 set_current_state(TASK_RUNNING);
2826 if (current->state == TASK_RUNNING)
2836 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2838 struct blk_mq_hw_ctx *hctx;
2839 struct blk_plug *plug;
2842 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2843 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2846 plug = current->plug;
2848 blk_flush_plug_list(plug, false);
2850 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2851 if (!blk_qc_t_is_internal(cookie))
2852 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2854 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2856 return __blk_mq_poll(hctx, rq);
2858 EXPORT_SYMBOL_GPL(blk_mq_poll);
2860 void blk_mq_disable_hotplug(void)
2862 mutex_lock(&all_q_mutex);
2865 void blk_mq_enable_hotplug(void)
2867 mutex_unlock(&all_q_mutex);
2870 static int __init blk_mq_init(void)
2872 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2873 blk_mq_hctx_notify_dead);
2875 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2876 blk_mq_queue_reinit_prepare,
2877 blk_mq_queue_reinit_dead);
2880 subsys_initcall(blk_mq_init);