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-debugfs.h"
35 #include "blk-mq-tag.h"
38 #include "blk-mq-sched.h"
40 static void blk_mq_poll_stats_start(struct request_queue *q);
41 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
43 static int blk_mq_poll_stats_bkt(const struct request *rq)
45 int ddir, bytes, bucket;
47 ddir = rq_data_dir(rq);
48 bytes = blk_rq_bytes(rq);
50 bucket = ddir + 2*(ilog2(bytes) - 9);
54 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
55 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
61 * Check if any of the ctx's have pending work in this hardware queue
63 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
65 return sbitmap_any_bit_set(&hctx->ctx_map) ||
66 !list_empty_careful(&hctx->dispatch) ||
67 blk_mq_sched_has_work(hctx);
71 * Mark this ctx as having pending work in this hardware queue
73 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
74 struct blk_mq_ctx *ctx)
76 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
77 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
80 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
81 struct blk_mq_ctx *ctx)
83 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
86 void blk_freeze_queue_start(struct request_queue *q)
90 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
91 if (freeze_depth == 1) {
92 percpu_ref_kill(&q->q_usage_counter);
93 blk_mq_run_hw_queues(q, false);
96 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
98 void blk_mq_freeze_queue_wait(struct request_queue *q)
100 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
102 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
104 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
105 unsigned long timeout)
107 return wait_event_timeout(q->mq_freeze_wq,
108 percpu_ref_is_zero(&q->q_usage_counter),
111 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
114 * Guarantee no request is in use, so we can change any data structure of
115 * the queue afterward.
117 void blk_freeze_queue(struct request_queue *q)
120 * In the !blk_mq case we are only calling this to kill the
121 * q_usage_counter, otherwise this increases the freeze depth
122 * and waits for it to return to zero. For this reason there is
123 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
124 * exported to drivers as the only user for unfreeze is blk_mq.
126 blk_freeze_queue_start(q);
127 blk_mq_freeze_queue_wait(q);
130 void blk_mq_freeze_queue(struct request_queue *q)
133 * ...just an alias to keep freeze and unfreeze actions balanced
134 * in the blk_mq_* namespace
138 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
140 void blk_mq_unfreeze_queue(struct request_queue *q)
144 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
145 WARN_ON_ONCE(freeze_depth < 0);
147 percpu_ref_reinit(&q->q_usage_counter);
148 wake_up_all(&q->mq_freeze_wq);
151 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
154 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
155 * mpt3sas driver such that this function can be removed.
157 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
161 spin_lock_irqsave(q->queue_lock, flags);
162 queue_flag_set(QUEUE_FLAG_QUIESCED, q);
163 spin_unlock_irqrestore(q->queue_lock, flags);
165 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
168 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
171 * Note: this function does not prevent that the struct request end_io()
172 * callback function is invoked. Once this function is returned, we make
173 * sure no dispatch can happen until the queue is unquiesced via
174 * blk_mq_unquiesce_queue().
176 void blk_mq_quiesce_queue(struct request_queue *q)
178 struct blk_mq_hw_ctx *hctx;
182 blk_mq_quiesce_queue_nowait(q);
184 queue_for_each_hw_ctx(q, hctx, i) {
185 if (hctx->flags & BLK_MQ_F_BLOCKING)
186 synchronize_srcu(hctx->queue_rq_srcu);
193 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
196 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
199 * This function recovers queue into the state before quiescing
200 * which is done by blk_mq_quiesce_queue.
202 void blk_mq_unquiesce_queue(struct request_queue *q)
206 spin_lock_irqsave(q->queue_lock, flags);
207 queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
208 spin_unlock_irqrestore(q->queue_lock, flags);
210 /* dispatch requests which are inserted during quiescing */
211 blk_mq_run_hw_queues(q, true);
213 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
215 void blk_mq_wake_waiters(struct request_queue *q)
217 struct blk_mq_hw_ctx *hctx;
220 queue_for_each_hw_ctx(q, hctx, i)
221 if (blk_mq_hw_queue_mapped(hctx))
222 blk_mq_tag_wakeup_all(hctx->tags, true);
225 * If we are called because the queue has now been marked as
226 * dying, we need to ensure that processes currently waiting on
227 * the queue are notified as well.
229 wake_up_all(&q->mq_freeze_wq);
232 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
234 return blk_mq_has_free_tags(hctx->tags);
236 EXPORT_SYMBOL(blk_mq_can_queue);
238 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
239 unsigned int tag, unsigned int op)
241 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
242 struct request *rq = tags->static_rqs[tag];
246 if (data->flags & BLK_MQ_REQ_INTERNAL) {
248 rq->internal_tag = tag;
250 if (blk_mq_tag_busy(data->hctx)) {
251 rq->rq_flags = RQF_MQ_INFLIGHT;
252 atomic_inc(&data->hctx->nr_active);
255 rq->internal_tag = -1;
256 data->hctx->tags->rqs[rq->tag] = rq;
259 INIT_LIST_HEAD(&rq->queuelist);
260 /* csd/requeue_work/fifo_time is initialized before use */
262 rq->mq_ctx = data->ctx;
264 if (blk_queue_io_stat(data->q))
265 rq->rq_flags |= RQF_IO_STAT;
266 /* do not touch atomic flags, it needs atomic ops against the timer */
268 INIT_HLIST_NODE(&rq->hash);
269 RB_CLEAR_NODE(&rq->rb_node);
272 rq->start_time = jiffies;
273 #ifdef CONFIG_BLK_CGROUP
275 set_start_time_ns(rq);
276 rq->io_start_time_ns = 0;
278 rq->nr_phys_segments = 0;
279 #if defined(CONFIG_BLK_DEV_INTEGRITY)
280 rq->nr_integrity_segments = 0;
283 /* tag was already set */
286 INIT_LIST_HEAD(&rq->timeout_list);
290 rq->end_io_data = NULL;
293 data->ctx->rq_dispatched[op_is_sync(op)]++;
297 static struct request *blk_mq_get_request(struct request_queue *q,
298 struct bio *bio, unsigned int op,
299 struct blk_mq_alloc_data *data)
301 struct elevator_queue *e = q->elevator;
305 blk_queue_enter_live(q);
307 if (likely(!data->ctx))
308 data->ctx = blk_mq_get_ctx(q);
309 if (likely(!data->hctx))
310 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
312 data->flags |= BLK_MQ_REQ_NOWAIT;
315 data->flags |= BLK_MQ_REQ_INTERNAL;
318 * Flush requests are special and go directly to the
321 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
322 e->type->ops.mq.limit_depth(op, data);
325 tag = blk_mq_get_tag(data);
326 if (tag == BLK_MQ_TAG_FAIL) {
331 rq = blk_mq_rq_ctx_init(data, tag, op);
332 if (!op_is_flush(op)) {
334 if (e && e->type->ops.mq.prepare_request) {
335 if (e->type->icq_cache && rq_ioc(bio))
336 blk_mq_sched_assign_ioc(rq, bio);
338 e->type->ops.mq.prepare_request(rq, bio);
339 rq->rq_flags |= RQF_ELVPRIV;
342 data->hctx->queued++;
346 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
349 struct blk_mq_alloc_data alloc_data = { .flags = flags };
353 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
357 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
359 blk_mq_put_ctx(alloc_data.ctx);
363 return ERR_PTR(-EWOULDBLOCK);
366 rq->__sector = (sector_t) -1;
367 rq->bio = rq->biotail = NULL;
370 EXPORT_SYMBOL(blk_mq_alloc_request);
372 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
373 unsigned int op, unsigned int flags, unsigned int hctx_idx)
375 struct blk_mq_alloc_data alloc_data = { .flags = flags };
381 * If the tag allocator sleeps we could get an allocation for a
382 * different hardware context. No need to complicate the low level
383 * allocator for this for the rare use case of a command tied to
386 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
387 return ERR_PTR(-EINVAL);
389 if (hctx_idx >= q->nr_hw_queues)
390 return ERR_PTR(-EIO);
392 ret = blk_queue_enter(q, true);
397 * Check if the hardware context is actually mapped to anything.
398 * If not tell the caller that it should skip this queue.
400 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
401 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
403 return ERR_PTR(-EXDEV);
405 cpu = cpumask_first(alloc_data.hctx->cpumask);
406 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
408 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
413 return ERR_PTR(-EWOULDBLOCK);
417 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
419 void blk_mq_free_request(struct request *rq)
421 struct request_queue *q = rq->q;
422 struct elevator_queue *e = q->elevator;
423 struct blk_mq_ctx *ctx = rq->mq_ctx;
424 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
425 const int sched_tag = rq->internal_tag;
427 if (rq->rq_flags & RQF_ELVPRIV) {
428 if (e && e->type->ops.mq.finish_request)
429 e->type->ops.mq.finish_request(rq);
431 put_io_context(rq->elv.icq->ioc);
436 ctx->rq_completed[rq_is_sync(rq)]++;
437 if (rq->rq_flags & RQF_MQ_INFLIGHT)
438 atomic_dec(&hctx->nr_active);
440 wbt_done(q->rq_wb, &rq->issue_stat);
442 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
443 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
445 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
447 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
448 blk_mq_sched_restart(hctx);
451 EXPORT_SYMBOL_GPL(blk_mq_free_request);
453 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
455 blk_account_io_done(rq);
458 wbt_done(rq->q->rq_wb, &rq->issue_stat);
459 rq->end_io(rq, error);
461 if (unlikely(blk_bidi_rq(rq)))
462 blk_mq_free_request(rq->next_rq);
463 blk_mq_free_request(rq);
466 EXPORT_SYMBOL(__blk_mq_end_request);
468 void blk_mq_end_request(struct request *rq, blk_status_t error)
470 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
472 __blk_mq_end_request(rq, error);
474 EXPORT_SYMBOL(blk_mq_end_request);
476 static void __blk_mq_complete_request_remote(void *data)
478 struct request *rq = data;
480 rq->q->softirq_done_fn(rq);
483 static void __blk_mq_complete_request(struct request *rq)
485 struct blk_mq_ctx *ctx = rq->mq_ctx;
489 if (rq->internal_tag != -1)
490 blk_mq_sched_completed_request(rq);
491 if (rq->rq_flags & RQF_STATS) {
492 blk_mq_poll_stats_start(rq->q);
496 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
497 rq->q->softirq_done_fn(rq);
502 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
503 shared = cpus_share_cache(cpu, ctx->cpu);
505 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
506 rq->csd.func = __blk_mq_complete_request_remote;
509 smp_call_function_single_async(ctx->cpu, &rq->csd);
511 rq->q->softirq_done_fn(rq);
517 * blk_mq_complete_request - end I/O on a request
518 * @rq: the request being processed
521 * Ends all I/O on a request. It does not handle partial completions.
522 * The actual completion happens out-of-order, through a IPI handler.
524 void blk_mq_complete_request(struct request *rq)
526 struct request_queue *q = rq->q;
528 if (unlikely(blk_should_fake_timeout(q)))
530 if (!blk_mark_rq_complete(rq))
531 __blk_mq_complete_request(rq);
533 EXPORT_SYMBOL(blk_mq_complete_request);
535 int blk_mq_request_started(struct request *rq)
537 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
539 EXPORT_SYMBOL_GPL(blk_mq_request_started);
541 void blk_mq_start_request(struct request *rq)
543 struct request_queue *q = rq->q;
545 blk_mq_sched_started_request(rq);
547 trace_block_rq_issue(q, rq);
549 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
550 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
551 rq->rq_flags |= RQF_STATS;
552 wbt_issue(q->rq_wb, &rq->issue_stat);
558 * Ensure that ->deadline is visible before set the started
559 * flag and clear the completed flag.
561 smp_mb__before_atomic();
564 * Mark us as started and clear complete. Complete might have been
565 * set if requeue raced with timeout, which then marked it as
566 * complete. So be sure to clear complete again when we start
567 * the request, otherwise we'll ignore the completion event.
569 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
570 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
571 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
572 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
574 if (q->dma_drain_size && blk_rq_bytes(rq)) {
576 * Make sure space for the drain appears. We know we can do
577 * this because max_hw_segments has been adjusted to be one
578 * fewer than the device can handle.
580 rq->nr_phys_segments++;
583 EXPORT_SYMBOL(blk_mq_start_request);
586 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
587 * flag isn't set yet, so there may be race with timeout handler,
588 * but given rq->deadline is just set in .queue_rq() under
589 * this situation, the race won't be possible in reality because
590 * rq->timeout should be set as big enough to cover the window
591 * between blk_mq_start_request() called from .queue_rq() and
592 * clearing REQ_ATOM_STARTED here.
594 static void __blk_mq_requeue_request(struct request *rq)
596 struct request_queue *q = rq->q;
598 trace_block_rq_requeue(q, rq);
599 wbt_requeue(q->rq_wb, &rq->issue_stat);
600 blk_mq_sched_requeue_request(rq);
602 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
603 if (q->dma_drain_size && blk_rq_bytes(rq))
604 rq->nr_phys_segments--;
608 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
610 __blk_mq_requeue_request(rq);
612 BUG_ON(blk_queued_rq(rq));
613 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
615 EXPORT_SYMBOL(blk_mq_requeue_request);
617 static void blk_mq_requeue_work(struct work_struct *work)
619 struct request_queue *q =
620 container_of(work, struct request_queue, requeue_work.work);
622 struct request *rq, *next;
625 spin_lock_irqsave(&q->requeue_lock, flags);
626 list_splice_init(&q->requeue_list, &rq_list);
627 spin_unlock_irqrestore(&q->requeue_lock, flags);
629 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
630 if (!(rq->rq_flags & RQF_SOFTBARRIER))
633 rq->rq_flags &= ~RQF_SOFTBARRIER;
634 list_del_init(&rq->queuelist);
635 blk_mq_sched_insert_request(rq, true, false, false, true);
638 while (!list_empty(&rq_list)) {
639 rq = list_entry(rq_list.next, struct request, queuelist);
640 list_del_init(&rq->queuelist);
641 blk_mq_sched_insert_request(rq, false, false, false, true);
644 blk_mq_run_hw_queues(q, false);
647 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
648 bool kick_requeue_list)
650 struct request_queue *q = rq->q;
654 * We abuse this flag that is otherwise used by the I/O scheduler to
655 * request head insertation from the workqueue.
657 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
659 spin_lock_irqsave(&q->requeue_lock, flags);
661 rq->rq_flags |= RQF_SOFTBARRIER;
662 list_add(&rq->queuelist, &q->requeue_list);
664 list_add_tail(&rq->queuelist, &q->requeue_list);
666 spin_unlock_irqrestore(&q->requeue_lock, flags);
668 if (kick_requeue_list)
669 blk_mq_kick_requeue_list(q);
671 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
673 void blk_mq_kick_requeue_list(struct request_queue *q)
675 kblockd_schedule_delayed_work(&q->requeue_work, 0);
677 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
679 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
682 kblockd_schedule_delayed_work(&q->requeue_work,
683 msecs_to_jiffies(msecs));
685 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
687 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
689 if (tag < tags->nr_tags) {
690 prefetch(tags->rqs[tag]);
691 return tags->rqs[tag];
696 EXPORT_SYMBOL(blk_mq_tag_to_rq);
698 struct blk_mq_timeout_data {
700 unsigned int next_set;
703 void blk_mq_rq_timed_out(struct request *req, bool reserved)
705 const struct blk_mq_ops *ops = req->q->mq_ops;
706 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
709 * We know that complete is set at this point. If STARTED isn't set
710 * anymore, then the request isn't active and the "timeout" should
711 * just be ignored. This can happen due to the bitflag ordering.
712 * Timeout first checks if STARTED is set, and if it is, assumes
713 * the request is active. But if we race with completion, then
714 * both flags will get cleared. So check here again, and ignore
715 * a timeout event with a request that isn't active.
717 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
721 ret = ops->timeout(req, reserved);
725 __blk_mq_complete_request(req);
727 case BLK_EH_RESET_TIMER:
729 blk_clear_rq_complete(req);
731 case BLK_EH_NOT_HANDLED:
734 printk(KERN_ERR "block: bad eh return: %d\n", ret);
739 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
740 struct request *rq, void *priv, bool reserved)
742 struct blk_mq_timeout_data *data = priv;
744 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
748 * The rq being checked may have been freed and reallocated
749 * out already here, we avoid this race by checking rq->deadline
750 * and REQ_ATOM_COMPLETE flag together:
752 * - if rq->deadline is observed as new value because of
753 * reusing, the rq won't be timed out because of timing.
754 * - if rq->deadline is observed as previous value,
755 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
756 * because we put a barrier between setting rq->deadline
757 * and clearing the flag in blk_mq_start_request(), so
758 * this rq won't be timed out too.
760 if (time_after_eq(jiffies, rq->deadline)) {
761 if (!blk_mark_rq_complete(rq))
762 blk_mq_rq_timed_out(rq, reserved);
763 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
764 data->next = rq->deadline;
769 static void blk_mq_timeout_work(struct work_struct *work)
771 struct request_queue *q =
772 container_of(work, struct request_queue, timeout_work);
773 struct blk_mq_timeout_data data = {
779 /* A deadlock might occur if a request is stuck requiring a
780 * timeout at the same time a queue freeze is waiting
781 * completion, since the timeout code would not be able to
782 * acquire the queue reference here.
784 * That's why we don't use blk_queue_enter here; instead, we use
785 * percpu_ref_tryget directly, because we need to be able to
786 * obtain a reference even in the short window between the queue
787 * starting to freeze, by dropping the first reference in
788 * blk_freeze_queue_start, and the moment the last request is
789 * consumed, marked by the instant q_usage_counter reaches
792 if (!percpu_ref_tryget(&q->q_usage_counter))
795 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
798 data.next = blk_rq_timeout(round_jiffies_up(data.next));
799 mod_timer(&q->timeout, data.next);
801 struct blk_mq_hw_ctx *hctx;
803 queue_for_each_hw_ctx(q, hctx, i) {
804 /* the hctx may be unmapped, so check it here */
805 if (blk_mq_hw_queue_mapped(hctx))
806 blk_mq_tag_idle(hctx);
812 struct flush_busy_ctx_data {
813 struct blk_mq_hw_ctx *hctx;
814 struct list_head *list;
817 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
819 struct flush_busy_ctx_data *flush_data = data;
820 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
821 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
823 sbitmap_clear_bit(sb, bitnr);
824 spin_lock(&ctx->lock);
825 list_splice_tail_init(&ctx->rq_list, flush_data->list);
826 spin_unlock(&ctx->lock);
831 * Process software queues that have been marked busy, splicing them
832 * to the for-dispatch
834 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
836 struct flush_busy_ctx_data data = {
841 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
843 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
845 static inline unsigned int queued_to_index(unsigned int queued)
850 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
853 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
856 struct blk_mq_alloc_data data = {
858 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
859 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
862 might_sleep_if(wait);
867 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
868 data.flags |= BLK_MQ_REQ_RESERVED;
870 rq->tag = blk_mq_get_tag(&data);
872 if (blk_mq_tag_busy(data.hctx)) {
873 rq->rq_flags |= RQF_MQ_INFLIGHT;
874 atomic_inc(&data.hctx->nr_active);
876 data.hctx->tags->rqs[rq->tag] = rq;
882 return rq->tag != -1;
885 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
888 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
891 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
892 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
893 atomic_dec(&hctx->nr_active);
897 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
900 if (rq->tag == -1 || rq->internal_tag == -1)
903 __blk_mq_put_driver_tag(hctx, rq);
906 static void blk_mq_put_driver_tag(struct request *rq)
908 struct blk_mq_hw_ctx *hctx;
910 if (rq->tag == -1 || rq->internal_tag == -1)
913 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
914 __blk_mq_put_driver_tag(hctx, rq);
918 * If we fail getting a driver tag because all the driver tags are already
919 * assigned and on the dispatch list, BUT the first entry does not have a
920 * tag, then we could deadlock. For that case, move entries with assigned
921 * driver tags to the front, leaving the set of tagged requests in the
922 * same order, and the untagged set in the same order.
924 static bool reorder_tags_to_front(struct list_head *list)
926 struct request *rq, *tmp, *first = NULL;
928 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
932 list_move(&rq->queuelist, list);
938 return first != NULL;
941 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
944 struct blk_mq_hw_ctx *hctx;
946 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
948 list_del(&wait->entry);
949 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
950 blk_mq_run_hw_queue(hctx, true);
954 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
956 struct sbq_wait_state *ws;
959 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
960 * The thread which wins the race to grab this bit adds the hardware
961 * queue to the wait queue.
963 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
964 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
967 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
968 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
971 * As soon as this returns, it's no longer safe to fiddle with
972 * hctx->dispatch_wait, since a completion can wake up the wait queue
973 * and unlock the bit.
975 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
979 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
981 struct blk_mq_hw_ctx *hctx;
985 if (list_empty(list))
989 * Now process all the entries, sending them to the driver.
993 struct blk_mq_queue_data bd;
996 rq = list_first_entry(list, struct request, queuelist);
997 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
998 if (!queued && reorder_tags_to_front(list))
1002 * The initial allocation attempt failed, so we need to
1003 * rerun the hardware queue when a tag is freed.
1005 if (!blk_mq_dispatch_wait_add(hctx))
1009 * It's possible that a tag was freed in the window
1010 * between the allocation failure and adding the
1011 * hardware queue to the wait queue.
1013 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1017 list_del_init(&rq->queuelist);
1022 * Flag last if we have no more requests, or if we have more
1023 * but can't assign a driver tag to it.
1025 if (list_empty(list))
1028 struct request *nxt;
1030 nxt = list_first_entry(list, struct request, queuelist);
1031 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1034 ret = q->mq_ops->queue_rq(hctx, &bd);
1035 if (ret == BLK_STS_RESOURCE) {
1036 blk_mq_put_driver_tag_hctx(hctx, rq);
1037 list_add(&rq->queuelist, list);
1038 __blk_mq_requeue_request(rq);
1042 if (unlikely(ret != BLK_STS_OK)) {
1044 blk_mq_end_request(rq, BLK_STS_IOERR);
1049 } while (!list_empty(list));
1051 hctx->dispatched[queued_to_index(queued)]++;
1054 * Any items that need requeuing? Stuff them into hctx->dispatch,
1055 * that is where we will continue on next queue run.
1057 if (!list_empty(list)) {
1059 * If an I/O scheduler has been configured and we got a driver
1060 * tag for the next request already, free it again.
1062 rq = list_first_entry(list, struct request, queuelist);
1063 blk_mq_put_driver_tag(rq);
1065 spin_lock(&hctx->lock);
1066 list_splice_init(list, &hctx->dispatch);
1067 spin_unlock(&hctx->lock);
1070 * If SCHED_RESTART was set by the caller of this function and
1071 * it is no longer set that means that it was cleared by another
1072 * thread and hence that a queue rerun is needed.
1074 * If TAG_WAITING is set that means that an I/O scheduler has
1075 * been configured and another thread is waiting for a driver
1076 * tag. To guarantee fairness, do not rerun this hardware queue
1077 * but let the other thread grab the driver tag.
1079 * If no I/O scheduler has been configured it is possible that
1080 * the hardware queue got stopped and restarted before requests
1081 * were pushed back onto the dispatch list. Rerun the queue to
1082 * avoid starvation. Notes:
1083 * - blk_mq_run_hw_queue() checks whether or not a queue has
1084 * been stopped before rerunning a queue.
1085 * - Some but not all block drivers stop a queue before
1086 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1089 if (!blk_mq_sched_needs_restart(hctx) &&
1090 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1091 blk_mq_run_hw_queue(hctx, true);
1094 return (queued + errors) != 0;
1097 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1101 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1102 cpu_online(hctx->next_cpu));
1104 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1106 blk_mq_sched_dispatch_requests(hctx);
1111 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1112 blk_mq_sched_dispatch_requests(hctx);
1113 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1118 * It'd be great if the workqueue API had a way to pass
1119 * in a mask and had some smarts for more clever placement.
1120 * For now we just round-robin here, switching for every
1121 * BLK_MQ_CPU_WORK_BATCH queued items.
1123 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1125 if (hctx->queue->nr_hw_queues == 1)
1126 return WORK_CPU_UNBOUND;
1128 if (--hctx->next_cpu_batch <= 0) {
1131 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1132 if (next_cpu >= nr_cpu_ids)
1133 next_cpu = cpumask_first(hctx->cpumask);
1135 hctx->next_cpu = next_cpu;
1136 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1139 return hctx->next_cpu;
1142 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1143 unsigned long msecs)
1145 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1148 if (unlikely(blk_mq_hctx_stopped(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);
1162 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1164 msecs_to_jiffies(msecs));
1167 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1169 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1171 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1173 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1175 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1177 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1179 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1181 struct blk_mq_hw_ctx *hctx;
1184 queue_for_each_hw_ctx(q, hctx, i) {
1185 if (!blk_mq_hctx_has_pending(hctx) ||
1186 blk_mq_hctx_stopped(hctx))
1189 blk_mq_run_hw_queue(hctx, async);
1192 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1195 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1196 * @q: request queue.
1198 * The caller is responsible for serializing this function against
1199 * blk_mq_{start,stop}_hw_queue().
1201 bool blk_mq_queue_stopped(struct request_queue *q)
1203 struct blk_mq_hw_ctx *hctx;
1206 queue_for_each_hw_ctx(q, hctx, i)
1207 if (blk_mq_hctx_stopped(hctx))
1212 EXPORT_SYMBOL(blk_mq_queue_stopped);
1215 * This function is often used for pausing .queue_rq() by driver when
1216 * there isn't enough resource or some conditions aren't satisfied, and
1217 * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1219 * We do not guarantee that dispatch can be drained or blocked
1220 * after blk_mq_stop_hw_queue() returns. Please use
1221 * blk_mq_quiesce_queue() for that requirement.
1223 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1225 cancel_delayed_work(&hctx->run_work);
1227 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1229 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1232 * This function is often used for pausing .queue_rq() by driver when
1233 * there isn't enough resource or some conditions aren't satisfied, and
1234 * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1236 * We do not guarantee that dispatch can be drained or blocked
1237 * after blk_mq_stop_hw_queues() returns. Please use
1238 * blk_mq_quiesce_queue() for that requirement.
1240 void blk_mq_stop_hw_queues(struct request_queue *q)
1242 struct blk_mq_hw_ctx *hctx;
1245 queue_for_each_hw_ctx(q, hctx, i)
1246 blk_mq_stop_hw_queue(hctx);
1248 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1250 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1252 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1254 blk_mq_run_hw_queue(hctx, false);
1256 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1258 void blk_mq_start_hw_queues(struct request_queue *q)
1260 struct blk_mq_hw_ctx *hctx;
1263 queue_for_each_hw_ctx(q, hctx, i)
1264 blk_mq_start_hw_queue(hctx);
1266 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1268 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1270 if (!blk_mq_hctx_stopped(hctx))
1273 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1274 blk_mq_run_hw_queue(hctx, async);
1276 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1278 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1280 struct blk_mq_hw_ctx *hctx;
1283 queue_for_each_hw_ctx(q, hctx, i)
1284 blk_mq_start_stopped_hw_queue(hctx, async);
1286 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1288 static void blk_mq_run_work_fn(struct work_struct *work)
1290 struct blk_mq_hw_ctx *hctx;
1292 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1295 * If we are stopped, don't run the queue. The exception is if
1296 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1297 * the STOPPED bit and run it.
1299 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1300 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1303 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1304 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1307 __blk_mq_run_hw_queue(hctx);
1311 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1313 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1317 * Stop the hw queue, then modify currently delayed work.
1318 * This should prevent us from running the queue prematurely.
1319 * Mark the queue as auto-clearing STOPPED when it runs.
1321 blk_mq_stop_hw_queue(hctx);
1322 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1323 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1325 msecs_to_jiffies(msecs));
1327 EXPORT_SYMBOL(blk_mq_delay_queue);
1329 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1333 struct blk_mq_ctx *ctx = rq->mq_ctx;
1335 lockdep_assert_held(&ctx->lock);
1337 trace_block_rq_insert(hctx->queue, rq);
1340 list_add(&rq->queuelist, &ctx->rq_list);
1342 list_add_tail(&rq->queuelist, &ctx->rq_list);
1345 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1348 struct blk_mq_ctx *ctx = rq->mq_ctx;
1350 lockdep_assert_held(&ctx->lock);
1352 __blk_mq_insert_req_list(hctx, rq, at_head);
1353 blk_mq_hctx_mark_pending(hctx, ctx);
1356 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1357 struct list_head *list)
1361 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1364 spin_lock(&ctx->lock);
1365 while (!list_empty(list)) {
1368 rq = list_first_entry(list, struct request, queuelist);
1369 BUG_ON(rq->mq_ctx != ctx);
1370 list_del_init(&rq->queuelist);
1371 __blk_mq_insert_req_list(hctx, rq, false);
1373 blk_mq_hctx_mark_pending(hctx, ctx);
1374 spin_unlock(&ctx->lock);
1377 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1379 struct request *rqa = container_of(a, struct request, queuelist);
1380 struct request *rqb = container_of(b, struct request, queuelist);
1382 return !(rqa->mq_ctx < rqb->mq_ctx ||
1383 (rqa->mq_ctx == rqb->mq_ctx &&
1384 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1387 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1389 struct blk_mq_ctx *this_ctx;
1390 struct request_queue *this_q;
1393 LIST_HEAD(ctx_list);
1396 list_splice_init(&plug->mq_list, &list);
1398 list_sort(NULL, &list, plug_ctx_cmp);
1404 while (!list_empty(&list)) {
1405 rq = list_entry_rq(list.next);
1406 list_del_init(&rq->queuelist);
1408 if (rq->mq_ctx != this_ctx) {
1410 trace_block_unplug(this_q, depth, from_schedule);
1411 blk_mq_sched_insert_requests(this_q, this_ctx,
1416 this_ctx = rq->mq_ctx;
1422 list_add_tail(&rq->queuelist, &ctx_list);
1426 * If 'this_ctx' is set, we know we have entries to complete
1427 * on 'ctx_list'. Do those.
1430 trace_block_unplug(this_q, depth, from_schedule);
1431 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1436 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1438 blk_init_request_from_bio(rq, bio);
1440 blk_account_io_start(rq, true);
1443 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1445 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1446 !blk_queue_nomerges(hctx->queue);
1449 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1450 struct blk_mq_ctx *ctx,
1453 spin_lock(&ctx->lock);
1454 __blk_mq_insert_request(hctx, rq, false);
1455 spin_unlock(&ctx->lock);
1458 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1461 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1463 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1466 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1468 blk_qc_t *cookie, bool may_sleep)
1470 struct request_queue *q = rq->q;
1471 struct blk_mq_queue_data bd = {
1475 blk_qc_t new_cookie;
1477 bool run_queue = true;
1479 /* RCU or SRCU read lock is needed before checking quiesced flag */
1480 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1488 if (!blk_mq_get_driver_tag(rq, NULL, false))
1491 new_cookie = request_to_qc_t(hctx, rq);
1494 * For OK queue, we are done. For error, kill it. Any other
1495 * error (busy), just add it to our list as we previously
1498 ret = q->mq_ops->queue_rq(hctx, &bd);
1501 *cookie = new_cookie;
1503 case BLK_STS_RESOURCE:
1504 __blk_mq_requeue_request(rq);
1507 *cookie = BLK_QC_T_NONE;
1508 blk_mq_end_request(rq, ret);
1513 blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1516 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1517 struct request *rq, blk_qc_t *cookie)
1519 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1521 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1524 unsigned int srcu_idx;
1528 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1529 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1530 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1534 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1536 const int is_sync = op_is_sync(bio->bi_opf);
1537 const int is_flush_fua = op_is_flush(bio->bi_opf);
1538 struct blk_mq_alloc_data data = { .flags = 0 };
1540 unsigned int request_count = 0;
1541 struct blk_plug *plug;
1542 struct request *same_queue_rq = NULL;
1544 unsigned int wb_acct;
1546 blk_queue_bounce(q, &bio);
1548 blk_queue_split(q, &bio);
1550 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1552 return BLK_QC_T_NONE;
1555 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1556 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1557 return BLK_QC_T_NONE;
1559 if (blk_mq_sched_bio_merge(q, bio))
1560 return BLK_QC_T_NONE;
1562 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1564 trace_block_getrq(q, bio, bio->bi_opf);
1566 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1567 if (unlikely(!rq)) {
1568 __wbt_done(q->rq_wb, wb_acct);
1569 if (bio->bi_opf & REQ_NOWAIT)
1570 bio_wouldblock_error(bio);
1571 return BLK_QC_T_NONE;
1574 wbt_track(&rq->issue_stat, wb_acct);
1576 cookie = request_to_qc_t(data.hctx, rq);
1578 plug = current->plug;
1579 if (unlikely(is_flush_fua)) {
1580 blk_mq_put_ctx(data.ctx);
1581 blk_mq_bio_to_request(rq, bio);
1583 blk_mq_sched_insert_request(rq, false, true, true,
1586 blk_insert_flush(rq);
1587 blk_mq_run_hw_queue(data.hctx, true);
1589 } else if (plug && q->nr_hw_queues == 1) {
1590 struct request *last = NULL;
1592 blk_mq_put_ctx(data.ctx);
1593 blk_mq_bio_to_request(rq, bio);
1596 * @request_count may become stale because of schedule
1597 * out, so check the list again.
1599 if (list_empty(&plug->mq_list))
1601 else if (blk_queue_nomerges(q))
1602 request_count = blk_plug_queued_count(q);
1605 trace_block_plug(q);
1607 last = list_entry_rq(plug->mq_list.prev);
1609 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1610 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1611 blk_flush_plug_list(plug, false);
1612 trace_block_plug(q);
1615 list_add_tail(&rq->queuelist, &plug->mq_list);
1616 } else if (plug && !blk_queue_nomerges(q)) {
1617 blk_mq_bio_to_request(rq, bio);
1620 * We do limited plugging. If the bio can be merged, do that.
1621 * Otherwise the existing request in the plug list will be
1622 * issued. So the plug list will have one request at most
1623 * The plug list might get flushed before this. If that happens,
1624 * the plug list is empty, and same_queue_rq is invalid.
1626 if (list_empty(&plug->mq_list))
1627 same_queue_rq = NULL;
1629 list_del_init(&same_queue_rq->queuelist);
1630 list_add_tail(&rq->queuelist, &plug->mq_list);
1632 blk_mq_put_ctx(data.ctx);
1634 if (same_queue_rq) {
1635 data.hctx = blk_mq_map_queue(q,
1636 same_queue_rq->mq_ctx->cpu);
1637 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1640 } else if (q->nr_hw_queues > 1 && is_sync) {
1641 blk_mq_put_ctx(data.ctx);
1642 blk_mq_bio_to_request(rq, bio);
1643 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1644 } else if (q->elevator) {
1645 blk_mq_put_ctx(data.ctx);
1646 blk_mq_bio_to_request(rq, bio);
1647 blk_mq_sched_insert_request(rq, false, true, true, true);
1649 blk_mq_put_ctx(data.ctx);
1650 blk_mq_bio_to_request(rq, bio);
1651 blk_mq_queue_io(data.hctx, data.ctx, rq);
1652 blk_mq_run_hw_queue(data.hctx, true);
1658 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1659 unsigned int hctx_idx)
1663 if (tags->rqs && set->ops->exit_request) {
1666 for (i = 0; i < tags->nr_tags; i++) {
1667 struct request *rq = tags->static_rqs[i];
1671 set->ops->exit_request(set, rq, hctx_idx);
1672 tags->static_rqs[i] = NULL;
1676 while (!list_empty(&tags->page_list)) {
1677 page = list_first_entry(&tags->page_list, struct page, lru);
1678 list_del_init(&page->lru);
1680 * Remove kmemleak object previously allocated in
1681 * blk_mq_init_rq_map().
1683 kmemleak_free(page_address(page));
1684 __free_pages(page, page->private);
1688 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1692 kfree(tags->static_rqs);
1693 tags->static_rqs = NULL;
1695 blk_mq_free_tags(tags);
1698 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1699 unsigned int hctx_idx,
1700 unsigned int nr_tags,
1701 unsigned int reserved_tags)
1703 struct blk_mq_tags *tags;
1706 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1707 if (node == NUMA_NO_NODE)
1708 node = set->numa_node;
1710 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1711 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1715 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1716 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1719 blk_mq_free_tags(tags);
1723 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1724 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1726 if (!tags->static_rqs) {
1728 blk_mq_free_tags(tags);
1735 static size_t order_to_size(unsigned int order)
1737 return (size_t)PAGE_SIZE << order;
1740 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1741 unsigned int hctx_idx, unsigned int depth)
1743 unsigned int i, j, entries_per_page, max_order = 4;
1744 size_t rq_size, left;
1747 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1748 if (node == NUMA_NO_NODE)
1749 node = set->numa_node;
1751 INIT_LIST_HEAD(&tags->page_list);
1754 * rq_size is the size of the request plus driver payload, rounded
1755 * to the cacheline size
1757 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1759 left = rq_size * depth;
1761 for (i = 0; i < depth; ) {
1762 int this_order = max_order;
1767 while (this_order && left < order_to_size(this_order - 1))
1771 page = alloc_pages_node(node,
1772 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1778 if (order_to_size(this_order) < rq_size)
1785 page->private = this_order;
1786 list_add_tail(&page->lru, &tags->page_list);
1788 p = page_address(page);
1790 * Allow kmemleak to scan these pages as they contain pointers
1791 * to additional allocations like via ops->init_request().
1793 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1794 entries_per_page = order_to_size(this_order) / rq_size;
1795 to_do = min(entries_per_page, depth - i);
1796 left -= to_do * rq_size;
1797 for (j = 0; j < to_do; j++) {
1798 struct request *rq = p;
1800 tags->static_rqs[i] = rq;
1801 if (set->ops->init_request) {
1802 if (set->ops->init_request(set, rq, hctx_idx,
1804 tags->static_rqs[i] = NULL;
1816 blk_mq_free_rqs(set, tags, hctx_idx);
1821 * 'cpu' is going away. splice any existing rq_list entries from this
1822 * software queue to the hw queue dispatch list, and ensure that it
1825 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1827 struct blk_mq_hw_ctx *hctx;
1828 struct blk_mq_ctx *ctx;
1831 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1832 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1834 spin_lock(&ctx->lock);
1835 if (!list_empty(&ctx->rq_list)) {
1836 list_splice_init(&ctx->rq_list, &tmp);
1837 blk_mq_hctx_clear_pending(hctx, ctx);
1839 spin_unlock(&ctx->lock);
1841 if (list_empty(&tmp))
1844 spin_lock(&hctx->lock);
1845 list_splice_tail_init(&tmp, &hctx->dispatch);
1846 spin_unlock(&hctx->lock);
1848 blk_mq_run_hw_queue(hctx, true);
1852 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1854 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1858 /* hctx->ctxs will be freed in queue's release handler */
1859 static void blk_mq_exit_hctx(struct request_queue *q,
1860 struct blk_mq_tag_set *set,
1861 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1863 blk_mq_debugfs_unregister_hctx(hctx);
1865 blk_mq_tag_idle(hctx);
1867 if (set->ops->exit_request)
1868 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1870 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1872 if (set->ops->exit_hctx)
1873 set->ops->exit_hctx(hctx, hctx_idx);
1875 if (hctx->flags & BLK_MQ_F_BLOCKING)
1876 cleanup_srcu_struct(hctx->queue_rq_srcu);
1878 blk_mq_remove_cpuhp(hctx);
1879 blk_free_flush_queue(hctx->fq);
1880 sbitmap_free(&hctx->ctx_map);
1883 static void blk_mq_exit_hw_queues(struct request_queue *q,
1884 struct blk_mq_tag_set *set, int nr_queue)
1886 struct blk_mq_hw_ctx *hctx;
1889 queue_for_each_hw_ctx(q, hctx, i) {
1892 blk_mq_exit_hctx(q, set, hctx, i);
1896 static int blk_mq_init_hctx(struct request_queue *q,
1897 struct blk_mq_tag_set *set,
1898 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1902 node = hctx->numa_node;
1903 if (node == NUMA_NO_NODE)
1904 node = hctx->numa_node = set->numa_node;
1906 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1907 spin_lock_init(&hctx->lock);
1908 INIT_LIST_HEAD(&hctx->dispatch);
1910 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1912 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1914 hctx->tags = set->tags[hctx_idx];
1917 * Allocate space for all possible cpus to avoid allocation at
1920 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1923 goto unregister_cpu_notifier;
1925 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1931 if (set->ops->init_hctx &&
1932 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1935 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1938 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1940 goto sched_exit_hctx;
1942 if (set->ops->init_request &&
1943 set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1947 if (hctx->flags & BLK_MQ_F_BLOCKING)
1948 init_srcu_struct(hctx->queue_rq_srcu);
1950 blk_mq_debugfs_register_hctx(q, hctx);
1957 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1959 if (set->ops->exit_hctx)
1960 set->ops->exit_hctx(hctx, hctx_idx);
1962 sbitmap_free(&hctx->ctx_map);
1965 unregister_cpu_notifier:
1966 blk_mq_remove_cpuhp(hctx);
1970 static void blk_mq_init_cpu_queues(struct request_queue *q,
1971 unsigned int nr_hw_queues)
1975 for_each_possible_cpu(i) {
1976 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1977 struct blk_mq_hw_ctx *hctx;
1980 spin_lock_init(&__ctx->lock);
1981 INIT_LIST_HEAD(&__ctx->rq_list);
1984 /* If the cpu isn't present, the cpu is mapped to first hctx */
1985 if (!cpu_present(i))
1988 hctx = blk_mq_map_queue(q, i);
1991 * Set local node, IFF we have more than one hw queue. If
1992 * not, we remain on the home node of the device
1994 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1995 hctx->numa_node = local_memory_node(cpu_to_node(i));
1999 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2003 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2004 set->queue_depth, set->reserved_tags);
2005 if (!set->tags[hctx_idx])
2008 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2013 blk_mq_free_rq_map(set->tags[hctx_idx]);
2014 set->tags[hctx_idx] = NULL;
2018 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2019 unsigned int hctx_idx)
2021 if (set->tags[hctx_idx]) {
2022 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2023 blk_mq_free_rq_map(set->tags[hctx_idx]);
2024 set->tags[hctx_idx] = NULL;
2028 static void blk_mq_map_swqueue(struct request_queue *q)
2030 unsigned int i, hctx_idx;
2031 struct blk_mq_hw_ctx *hctx;
2032 struct blk_mq_ctx *ctx;
2033 struct blk_mq_tag_set *set = q->tag_set;
2036 * Avoid others reading imcomplete hctx->cpumask through sysfs
2038 mutex_lock(&q->sysfs_lock);
2040 queue_for_each_hw_ctx(q, hctx, i) {
2041 cpumask_clear(hctx->cpumask);
2046 * Map software to hardware queues.
2048 * If the cpu isn't present, the cpu is mapped to first hctx.
2050 for_each_present_cpu(i) {
2051 hctx_idx = q->mq_map[i];
2052 /* unmapped hw queue can be remapped after CPU topo changed */
2053 if (!set->tags[hctx_idx] &&
2054 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2056 * If tags initialization fail for some hctx,
2057 * that hctx won't be brought online. In this
2058 * case, remap the current ctx to hctx[0] which
2059 * is guaranteed to always have tags allocated
2064 ctx = per_cpu_ptr(q->queue_ctx, i);
2065 hctx = blk_mq_map_queue(q, i);
2067 cpumask_set_cpu(i, hctx->cpumask);
2068 ctx->index_hw = hctx->nr_ctx;
2069 hctx->ctxs[hctx->nr_ctx++] = ctx;
2072 mutex_unlock(&q->sysfs_lock);
2074 queue_for_each_hw_ctx(q, hctx, i) {
2076 * If no software queues are mapped to this hardware queue,
2077 * disable it and free the request entries.
2079 if (!hctx->nr_ctx) {
2080 /* Never unmap queue 0. We need it as a
2081 * fallback in case of a new remap fails
2084 if (i && set->tags[i])
2085 blk_mq_free_map_and_requests(set, i);
2091 hctx->tags = set->tags[i];
2092 WARN_ON(!hctx->tags);
2095 * Set the map size to the number of mapped software queues.
2096 * This is more accurate and more efficient than looping
2097 * over all possibly mapped software queues.
2099 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2102 * Initialize batch roundrobin counts
2104 hctx->next_cpu = cpumask_first(hctx->cpumask);
2105 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2110 * Caller needs to ensure that we're either frozen/quiesced, or that
2111 * the queue isn't live yet.
2113 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2115 struct blk_mq_hw_ctx *hctx;
2118 queue_for_each_hw_ctx(q, hctx, i) {
2120 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2121 atomic_inc(&q->shared_hctx_restart);
2122 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2124 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2125 atomic_dec(&q->shared_hctx_restart);
2126 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2131 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2134 struct request_queue *q;
2136 lockdep_assert_held(&set->tag_list_lock);
2138 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2139 blk_mq_freeze_queue(q);
2140 queue_set_hctx_shared(q, shared);
2141 blk_mq_unfreeze_queue(q);
2145 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2147 struct blk_mq_tag_set *set = q->tag_set;
2149 mutex_lock(&set->tag_list_lock);
2150 list_del_rcu(&q->tag_set_list);
2151 INIT_LIST_HEAD(&q->tag_set_list);
2152 if (list_is_singular(&set->tag_list)) {
2153 /* just transitioned to unshared */
2154 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2155 /* update existing queue */
2156 blk_mq_update_tag_set_depth(set, false);
2158 mutex_unlock(&set->tag_list_lock);
2163 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2164 struct request_queue *q)
2168 mutex_lock(&set->tag_list_lock);
2170 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2171 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2172 set->flags |= BLK_MQ_F_TAG_SHARED;
2173 /* update existing queue */
2174 blk_mq_update_tag_set_depth(set, true);
2176 if (set->flags & BLK_MQ_F_TAG_SHARED)
2177 queue_set_hctx_shared(q, true);
2178 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2180 mutex_unlock(&set->tag_list_lock);
2184 * It is the actual release handler for mq, but we do it from
2185 * request queue's release handler for avoiding use-after-free
2186 * and headache because q->mq_kobj shouldn't have been introduced,
2187 * but we can't group ctx/kctx kobj without it.
2189 void blk_mq_release(struct request_queue *q)
2191 struct blk_mq_hw_ctx *hctx;
2194 /* hctx kobj stays in hctx */
2195 queue_for_each_hw_ctx(q, hctx, i) {
2198 kobject_put(&hctx->kobj);
2203 kfree(q->queue_hw_ctx);
2206 * release .mq_kobj and sw queue's kobject now because
2207 * both share lifetime with request queue.
2209 blk_mq_sysfs_deinit(q);
2211 free_percpu(q->queue_ctx);
2214 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2216 struct request_queue *uninit_q, *q;
2218 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2220 return ERR_PTR(-ENOMEM);
2222 q = blk_mq_init_allocated_queue(set, uninit_q);
2224 blk_cleanup_queue(uninit_q);
2228 EXPORT_SYMBOL(blk_mq_init_queue);
2230 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2232 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2234 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2235 __alignof__(struct blk_mq_hw_ctx)) !=
2236 sizeof(struct blk_mq_hw_ctx));
2238 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2239 hw_ctx_size += sizeof(struct srcu_struct);
2244 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2245 struct request_queue *q)
2248 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2250 blk_mq_sysfs_unregister(q);
2251 for (i = 0; i < set->nr_hw_queues; i++) {
2257 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2258 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2263 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2270 atomic_set(&hctxs[i]->nr_active, 0);
2271 hctxs[i]->numa_node = node;
2272 hctxs[i]->queue_num = i;
2274 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2275 free_cpumask_var(hctxs[i]->cpumask);
2280 blk_mq_hctx_kobj_init(hctxs[i]);
2282 for (j = i; j < q->nr_hw_queues; j++) {
2283 struct blk_mq_hw_ctx *hctx = hctxs[j];
2287 blk_mq_free_map_and_requests(set, j);
2288 blk_mq_exit_hctx(q, set, hctx, j);
2289 kobject_put(&hctx->kobj);
2294 q->nr_hw_queues = i;
2295 blk_mq_sysfs_register(q);
2298 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2299 struct request_queue *q)
2301 /* mark the queue as mq asap */
2302 q->mq_ops = set->ops;
2304 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2305 blk_mq_poll_stats_bkt,
2306 BLK_MQ_POLL_STATS_BKTS, q);
2310 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2314 /* init q->mq_kobj and sw queues' kobjects */
2315 blk_mq_sysfs_init(q);
2317 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2318 GFP_KERNEL, set->numa_node);
2319 if (!q->queue_hw_ctx)
2322 q->mq_map = set->mq_map;
2324 blk_mq_realloc_hw_ctxs(set, q);
2325 if (!q->nr_hw_queues)
2328 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2329 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2331 q->nr_queues = nr_cpu_ids;
2333 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2335 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2336 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2338 q->sg_reserved_size = INT_MAX;
2340 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2341 INIT_LIST_HEAD(&q->requeue_list);
2342 spin_lock_init(&q->requeue_lock);
2344 blk_queue_make_request(q, blk_mq_make_request);
2347 * Do this after blk_queue_make_request() overrides it...
2349 q->nr_requests = set->queue_depth;
2352 * Default to classic polling
2356 if (set->ops->complete)
2357 blk_queue_softirq_done(q, set->ops->complete);
2359 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2360 blk_mq_add_queue_tag_set(set, q);
2361 blk_mq_map_swqueue(q);
2363 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2366 ret = blk_mq_sched_init(q);
2368 return ERR_PTR(ret);
2374 kfree(q->queue_hw_ctx);
2376 free_percpu(q->queue_ctx);
2379 return ERR_PTR(-ENOMEM);
2381 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2383 void blk_mq_free_queue(struct request_queue *q)
2385 struct blk_mq_tag_set *set = q->tag_set;
2387 blk_mq_del_queue_tag_set(q);
2388 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2391 /* Basically redo blk_mq_init_queue with queue frozen */
2392 static void blk_mq_queue_reinit(struct request_queue *q)
2394 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2396 blk_mq_debugfs_unregister_hctxs(q);
2397 blk_mq_sysfs_unregister(q);
2400 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2401 * we should change hctx numa_node according to new topology (this
2402 * involves free and re-allocate memory, worthy doing?)
2405 blk_mq_map_swqueue(q);
2407 blk_mq_sysfs_register(q);
2408 blk_mq_debugfs_register_hctxs(q);
2411 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2415 for (i = 0; i < set->nr_hw_queues; i++)
2416 if (!__blk_mq_alloc_rq_map(set, i))
2423 blk_mq_free_rq_map(set->tags[i]);
2429 * Allocate the request maps associated with this tag_set. Note that this
2430 * may reduce the depth asked for, if memory is tight. set->queue_depth
2431 * will be updated to reflect the allocated depth.
2433 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2438 depth = set->queue_depth;
2440 err = __blk_mq_alloc_rq_maps(set);
2444 set->queue_depth >>= 1;
2445 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2449 } while (set->queue_depth);
2451 if (!set->queue_depth || err) {
2452 pr_err("blk-mq: failed to allocate request map\n");
2456 if (depth != set->queue_depth)
2457 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2458 depth, set->queue_depth);
2463 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2465 if (set->ops->map_queues)
2466 return set->ops->map_queues(set);
2468 return blk_mq_map_queues(set);
2472 * Alloc a tag set to be associated with one or more request queues.
2473 * May fail with EINVAL for various error conditions. May adjust the
2474 * requested depth down, if if it too large. In that case, the set
2475 * value will be stored in set->queue_depth.
2477 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2481 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2483 if (!set->nr_hw_queues)
2485 if (!set->queue_depth)
2487 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2490 if (!set->ops->queue_rq)
2493 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2494 pr_info("blk-mq: reduced tag depth to %u\n",
2496 set->queue_depth = BLK_MQ_MAX_DEPTH;
2500 * If a crashdump is active, then we are potentially in a very
2501 * memory constrained environment. Limit us to 1 queue and
2502 * 64 tags to prevent using too much memory.
2504 if (is_kdump_kernel()) {
2505 set->nr_hw_queues = 1;
2506 set->queue_depth = min(64U, set->queue_depth);
2509 * There is no use for more h/w queues than cpus.
2511 if (set->nr_hw_queues > nr_cpu_ids)
2512 set->nr_hw_queues = nr_cpu_ids;
2514 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2515 GFP_KERNEL, set->numa_node);
2520 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2521 GFP_KERNEL, set->numa_node);
2525 ret = blk_mq_update_queue_map(set);
2527 goto out_free_mq_map;
2529 ret = blk_mq_alloc_rq_maps(set);
2531 goto out_free_mq_map;
2533 mutex_init(&set->tag_list_lock);
2534 INIT_LIST_HEAD(&set->tag_list);
2546 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2548 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2552 for (i = 0; i < nr_cpu_ids; i++)
2553 blk_mq_free_map_and_requests(set, i);
2561 EXPORT_SYMBOL(blk_mq_free_tag_set);
2563 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2565 struct blk_mq_tag_set *set = q->tag_set;
2566 struct blk_mq_hw_ctx *hctx;
2572 blk_mq_freeze_queue(q);
2575 queue_for_each_hw_ctx(q, hctx, i) {
2579 * If we're using an MQ scheduler, just update the scheduler
2580 * queue depth. This is similar to what the old code would do.
2582 if (!hctx->sched_tags) {
2583 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2584 min(nr, set->queue_depth),
2587 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2595 q->nr_requests = nr;
2597 blk_mq_unfreeze_queue(q);
2602 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2605 struct request_queue *q;
2607 lockdep_assert_held(&set->tag_list_lock);
2609 if (nr_hw_queues > nr_cpu_ids)
2610 nr_hw_queues = nr_cpu_ids;
2611 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2614 list_for_each_entry(q, &set->tag_list, tag_set_list)
2615 blk_mq_freeze_queue(q);
2617 set->nr_hw_queues = nr_hw_queues;
2618 blk_mq_update_queue_map(set);
2619 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2620 blk_mq_realloc_hw_ctxs(set, q);
2621 blk_mq_queue_reinit(q);
2624 list_for_each_entry(q, &set->tag_list, tag_set_list)
2625 blk_mq_unfreeze_queue(q);
2628 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2630 mutex_lock(&set->tag_list_lock);
2631 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2632 mutex_unlock(&set->tag_list_lock);
2634 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2636 /* Enable polling stats and return whether they were already enabled. */
2637 static bool blk_poll_stats_enable(struct request_queue *q)
2639 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2640 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2642 blk_stat_add_callback(q, q->poll_cb);
2646 static void blk_mq_poll_stats_start(struct request_queue *q)
2649 * We don't arm the callback if polling stats are not enabled or the
2650 * callback is already active.
2652 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2653 blk_stat_is_active(q->poll_cb))
2656 blk_stat_activate_msecs(q->poll_cb, 100);
2659 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2661 struct request_queue *q = cb->data;
2664 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2665 if (cb->stat[bucket].nr_samples)
2666 q->poll_stat[bucket] = cb->stat[bucket];
2670 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2671 struct blk_mq_hw_ctx *hctx,
2674 unsigned long ret = 0;
2678 * If stats collection isn't on, don't sleep but turn it on for
2681 if (!blk_poll_stats_enable(q))
2685 * As an optimistic guess, use half of the mean service time
2686 * for this type of request. We can (and should) make this smarter.
2687 * For instance, if the completion latencies are tight, we can
2688 * get closer than just half the mean. This is especially
2689 * important on devices where the completion latencies are longer
2690 * than ~10 usec. We do use the stats for the relevant IO size
2691 * if available which does lead to better estimates.
2693 bucket = blk_mq_poll_stats_bkt(rq);
2697 if (q->poll_stat[bucket].nr_samples)
2698 ret = (q->poll_stat[bucket].mean + 1) / 2;
2703 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2704 struct blk_mq_hw_ctx *hctx,
2707 struct hrtimer_sleeper hs;
2708 enum hrtimer_mode mode;
2712 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2718 * -1: don't ever hybrid sleep
2719 * 0: use half of prev avg
2720 * >0: use this specific value
2722 if (q->poll_nsec == -1)
2724 else if (q->poll_nsec > 0)
2725 nsecs = q->poll_nsec;
2727 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2732 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2735 * This will be replaced with the stats tracking code, using
2736 * 'avg_completion_time / 2' as the pre-sleep target.
2740 mode = HRTIMER_MODE_REL;
2741 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2742 hrtimer_set_expires(&hs.timer, kt);
2744 hrtimer_init_sleeper(&hs, current);
2746 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2748 set_current_state(TASK_UNINTERRUPTIBLE);
2749 hrtimer_start_expires(&hs.timer, mode);
2752 hrtimer_cancel(&hs.timer);
2753 mode = HRTIMER_MODE_ABS;
2754 } while (hs.task && !signal_pending(current));
2756 __set_current_state(TASK_RUNNING);
2757 destroy_hrtimer_on_stack(&hs.timer);
2761 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2763 struct request_queue *q = hctx->queue;
2767 * If we sleep, have the caller restart the poll loop to reset
2768 * the state. Like for the other success return cases, the
2769 * caller is responsible for checking if the IO completed. If
2770 * the IO isn't complete, we'll get called again and will go
2771 * straight to the busy poll loop.
2773 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2776 hctx->poll_considered++;
2778 state = current->state;
2779 while (!need_resched()) {
2782 hctx->poll_invoked++;
2784 ret = q->mq_ops->poll(hctx, rq->tag);
2786 hctx->poll_success++;
2787 set_current_state(TASK_RUNNING);
2791 if (signal_pending_state(state, current))
2792 set_current_state(TASK_RUNNING);
2794 if (current->state == TASK_RUNNING)
2804 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2806 struct blk_mq_hw_ctx *hctx;
2807 struct blk_plug *plug;
2810 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2811 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2814 plug = current->plug;
2816 blk_flush_plug_list(plug, false);
2818 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2819 if (!blk_qc_t_is_internal(cookie))
2820 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2822 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2824 * With scheduling, if the request has completed, we'll
2825 * get a NULL return here, as we clear the sched tag when
2826 * that happens. The request still remains valid, like always,
2827 * so we should be safe with just the NULL check.
2833 return __blk_mq_poll(hctx, rq);
2835 EXPORT_SYMBOL_GPL(blk_mq_poll);
2837 static int __init blk_mq_init(void)
2839 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2840 blk_mq_hctx_notify_dead);
2843 subsys_initcall(blk_mq_init);