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 DEFINE_MUTEX(all_q_mutex);
41 static LIST_HEAD(all_q_list);
43 static void blk_mq_poll_stats_start(struct request_queue *q);
44 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
45 static void __blk_mq_stop_hw_queues(struct request_queue *q, bool sync);
47 static int blk_mq_poll_stats_bkt(const struct request *rq)
49 int ddir, bytes, bucket;
51 ddir = rq_data_dir(rq);
52 bytes = blk_rq_bytes(rq);
54 bucket = ddir + 2*(ilog2(bytes) - 9);
58 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
59 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
65 * Check if any of the ctx's have pending work in this hardware queue
67 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
69 return sbitmap_any_bit_set(&hctx->ctx_map) ||
70 !list_empty_careful(&hctx->dispatch) ||
71 blk_mq_sched_has_work(hctx);
75 * Mark this ctx as having pending work in this hardware queue
77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
78 struct blk_mq_ctx *ctx)
80 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
81 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
84 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
85 struct blk_mq_ctx *ctx)
87 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
90 void blk_freeze_queue_start(struct request_queue *q)
94 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
95 if (freeze_depth == 1) {
96 percpu_ref_kill(&q->q_usage_counter);
97 blk_mq_run_hw_queues(q, false);
100 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
102 void blk_mq_freeze_queue_wait(struct request_queue *q)
104 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
106 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
108 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
109 unsigned long timeout)
111 return wait_event_timeout(q->mq_freeze_wq,
112 percpu_ref_is_zero(&q->q_usage_counter),
115 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
118 * Guarantee no request is in use, so we can change any data structure of
119 * the queue afterward.
121 void blk_freeze_queue(struct request_queue *q)
124 * In the !blk_mq case we are only calling this to kill the
125 * q_usage_counter, otherwise this increases the freeze depth
126 * and waits for it to return to zero. For this reason there is
127 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
128 * exported to drivers as the only user for unfreeze is blk_mq.
130 blk_freeze_queue_start(q);
131 blk_mq_freeze_queue_wait(q);
134 void blk_mq_freeze_queue(struct request_queue *q)
137 * ...just an alias to keep freeze and unfreeze actions balanced
138 * in the blk_mq_* namespace
142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
144 void blk_mq_unfreeze_queue(struct request_queue *q)
148 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
149 WARN_ON_ONCE(freeze_depth < 0);
151 percpu_ref_reinit(&q->q_usage_counter);
152 wake_up_all(&q->mq_freeze_wq);
155 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
158 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
161 * Note: this function does not prevent that the struct request end_io()
162 * callback function is invoked. Additionally, it is not prevented that
163 * new queue_rq() calls occur unless the queue has been stopped first.
165 void blk_mq_quiesce_queue(struct request_queue *q)
167 struct blk_mq_hw_ctx *hctx;
171 __blk_mq_stop_hw_queues(q, true);
173 queue_for_each_hw_ctx(q, hctx, i) {
174 if (hctx->flags & BLK_MQ_F_BLOCKING)
175 synchronize_srcu(&hctx->queue_rq_srcu);
182 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
184 void blk_mq_wake_waiters(struct request_queue *q)
186 struct blk_mq_hw_ctx *hctx;
189 queue_for_each_hw_ctx(q, hctx, i)
190 if (blk_mq_hw_queue_mapped(hctx))
191 blk_mq_tag_wakeup_all(hctx->tags, true);
194 * If we are called because the queue has now been marked as
195 * dying, we need to ensure that processes currently waiting on
196 * the queue are notified as well.
198 wake_up_all(&q->mq_freeze_wq);
201 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
203 return blk_mq_has_free_tags(hctx->tags);
205 EXPORT_SYMBOL(blk_mq_can_queue);
207 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
208 struct request *rq, unsigned int op)
210 INIT_LIST_HEAD(&rq->queuelist);
211 /* csd/requeue_work/fifo_time is initialized before use */
215 if (blk_queue_io_stat(q))
216 rq->rq_flags |= RQF_IO_STAT;
217 /* do not touch atomic flags, it needs atomic ops against the timer */
219 INIT_HLIST_NODE(&rq->hash);
220 RB_CLEAR_NODE(&rq->rb_node);
223 rq->start_time = jiffies;
224 #ifdef CONFIG_BLK_CGROUP
226 set_start_time_ns(rq);
227 rq->io_start_time_ns = 0;
229 rq->nr_phys_segments = 0;
230 #if defined(CONFIG_BLK_DEV_INTEGRITY)
231 rq->nr_integrity_segments = 0;
234 /* tag was already set */
237 INIT_LIST_HEAD(&rq->timeout_list);
241 rq->end_io_data = NULL;
244 ctx->rq_dispatched[op_is_sync(op)]++;
247 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
253 tag = blk_mq_get_tag(data);
254 if (tag != BLK_MQ_TAG_FAIL) {
255 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
257 rq = tags->static_rqs[tag];
259 if (data->flags & BLK_MQ_REQ_INTERNAL) {
261 rq->internal_tag = tag;
263 if (blk_mq_tag_busy(data->hctx)) {
264 rq->rq_flags = RQF_MQ_INFLIGHT;
265 atomic_inc(&data->hctx->nr_active);
268 rq->internal_tag = -1;
269 data->hctx->tags->rqs[rq->tag] = rq;
272 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
278 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
280 static struct request *blk_mq_get_request(struct request_queue *q,
281 struct bio *bio, unsigned int op,
282 struct blk_mq_alloc_data *data)
284 struct elevator_queue *e = q->elevator;
287 blk_queue_enter_live(q);
289 if (likely(!data->ctx))
290 data->ctx = blk_mq_get_ctx(q);
291 if (likely(!data->hctx))
292 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
295 data->flags |= BLK_MQ_REQ_INTERNAL;
298 * Flush requests are special and go directly to the
301 if (!op_is_flush(op) && e->type->ops.mq.get_request) {
302 rq = e->type->ops.mq.get_request(q, op, data);
304 rq->rq_flags |= RQF_QUEUED;
306 rq = __blk_mq_alloc_request(data, op);
308 rq = __blk_mq_alloc_request(data, op);
312 if (!op_is_flush(op)) {
314 if (e && e->type->icq_cache)
315 blk_mq_sched_assign_ioc(q, rq, bio);
317 data->hctx->queued++;
325 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
328 struct blk_mq_alloc_data alloc_data = { .flags = flags };
332 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
336 rq = blk_mq_get_request(q, NULL, rw, &alloc_data);
338 blk_mq_put_ctx(alloc_data.ctx);
342 return ERR_PTR(-EWOULDBLOCK);
345 rq->__sector = (sector_t) -1;
346 rq->bio = rq->biotail = NULL;
349 EXPORT_SYMBOL(blk_mq_alloc_request);
351 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
352 unsigned int flags, unsigned int hctx_idx)
354 struct blk_mq_alloc_data alloc_data = { .flags = flags };
360 * If the tag allocator sleeps we could get an allocation for a
361 * different hardware context. No need to complicate the low level
362 * allocator for this for the rare use case of a command tied to
365 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
366 return ERR_PTR(-EINVAL);
368 if (hctx_idx >= q->nr_hw_queues)
369 return ERR_PTR(-EIO);
371 ret = blk_queue_enter(q, true);
376 * Check if the hardware context is actually mapped to anything.
377 * If not tell the caller that it should skip this queue.
379 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
380 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
382 return ERR_PTR(-EXDEV);
384 cpu = cpumask_first(alloc_data.hctx->cpumask);
385 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
387 rq = blk_mq_get_request(q, NULL, rw, &alloc_data);
392 return ERR_PTR(-EWOULDBLOCK);
396 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
398 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
401 const int sched_tag = rq->internal_tag;
402 struct request_queue *q = rq->q;
404 if (rq->rq_flags & RQF_MQ_INFLIGHT)
405 atomic_dec(&hctx->nr_active);
407 wbt_done(q->rq_wb, &rq->issue_stat);
410 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
411 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
413 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
415 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
416 blk_mq_sched_restart(hctx);
420 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
423 struct blk_mq_ctx *ctx = rq->mq_ctx;
425 ctx->rq_completed[rq_is_sync(rq)]++;
426 __blk_mq_finish_request(hctx, ctx, rq);
429 void blk_mq_finish_request(struct request *rq)
431 blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
433 EXPORT_SYMBOL_GPL(blk_mq_finish_request);
435 void blk_mq_free_request(struct request *rq)
437 struct request_queue *q = rq->q;
438 struct elevator_queue *e = q->elevator;
440 if (rq->rq_flags & RQF_ELVPRIV) {
441 blk_mq_sched_put_rq_priv(rq->q, rq);
443 put_io_context(rq->elv.icq->ioc);
448 if ((rq->rq_flags & RQF_QUEUED) && e && e->type->ops.mq.put_request)
449 e->type->ops.mq.put_request(rq);
451 blk_mq_finish_request(rq);
453 EXPORT_SYMBOL_GPL(blk_mq_free_request);
455 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
457 blk_account_io_done(rq);
460 wbt_done(rq->q->rq_wb, &rq->issue_stat);
461 rq->end_io(rq, error);
463 if (unlikely(blk_bidi_rq(rq)))
464 blk_mq_free_request(rq->next_rq);
465 blk_mq_free_request(rq);
468 EXPORT_SYMBOL(__blk_mq_end_request);
470 void blk_mq_end_request(struct request *rq, blk_status_t error)
472 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
474 __blk_mq_end_request(rq, error);
476 EXPORT_SYMBOL(blk_mq_end_request);
478 static void __blk_mq_complete_request_remote(void *data)
480 struct request *rq = data;
482 rq->q->softirq_done_fn(rq);
485 static void __blk_mq_complete_request(struct request *rq)
487 struct blk_mq_ctx *ctx = rq->mq_ctx;
491 if (rq->internal_tag != -1)
492 blk_mq_sched_completed_request(rq);
493 if (rq->rq_flags & RQF_STATS) {
494 blk_mq_poll_stats_start(rq->q);
498 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
499 rq->q->softirq_done_fn(rq);
504 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
505 shared = cpus_share_cache(cpu, ctx->cpu);
507 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
508 rq->csd.func = __blk_mq_complete_request_remote;
511 smp_call_function_single_async(ctx->cpu, &rq->csd);
513 rq->q->softirq_done_fn(rq);
519 * blk_mq_complete_request - end I/O on a request
520 * @rq: the request being processed
523 * Ends all I/O on a request. It does not handle partial completions.
524 * The actual completion happens out-of-order, through a IPI handler.
526 void blk_mq_complete_request(struct request *rq)
528 struct request_queue *q = rq->q;
530 if (unlikely(blk_should_fake_timeout(q)))
532 if (!blk_mark_rq_complete(rq))
533 __blk_mq_complete_request(rq);
535 EXPORT_SYMBOL(blk_mq_complete_request);
537 int blk_mq_request_started(struct request *rq)
539 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
541 EXPORT_SYMBOL_GPL(blk_mq_request_started);
543 void blk_mq_start_request(struct request *rq)
545 struct request_queue *q = rq->q;
547 blk_mq_sched_started_request(rq);
549 trace_block_rq_issue(q, rq);
551 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
552 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
553 rq->rq_flags |= RQF_STATS;
554 wbt_issue(q->rq_wb, &rq->issue_stat);
560 * Ensure that ->deadline is visible before set the started
561 * flag and clear the completed flag.
563 smp_mb__before_atomic();
566 * Mark us as started and clear complete. Complete might have been
567 * set if requeue raced with timeout, which then marked it as
568 * complete. So be sure to clear complete again when we start
569 * the request, otherwise we'll ignore the completion event.
571 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
572 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
573 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
574 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
576 if (q->dma_drain_size && blk_rq_bytes(rq)) {
578 * Make sure space for the drain appears. We know we can do
579 * this because max_hw_segments has been adjusted to be one
580 * fewer than the device can handle.
582 rq->nr_phys_segments++;
585 EXPORT_SYMBOL(blk_mq_start_request);
588 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
589 * flag isn't set yet, so there may be race with timeout handler,
590 * but given rq->deadline is just set in .queue_rq() under
591 * this situation, the race won't be possible in reality because
592 * rq->timeout should be set as big enough to cover the window
593 * between blk_mq_start_request() called from .queue_rq() and
594 * clearing REQ_ATOM_STARTED here.
596 static void __blk_mq_requeue_request(struct request *rq)
598 struct request_queue *q = rq->q;
600 trace_block_rq_requeue(q, rq);
601 wbt_requeue(q->rq_wb, &rq->issue_stat);
602 blk_mq_sched_requeue_request(rq);
604 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
605 if (q->dma_drain_size && blk_rq_bytes(rq))
606 rq->nr_phys_segments--;
610 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
612 __blk_mq_requeue_request(rq);
614 BUG_ON(blk_queued_rq(rq));
615 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
617 EXPORT_SYMBOL(blk_mq_requeue_request);
619 static void blk_mq_requeue_work(struct work_struct *work)
621 struct request_queue *q =
622 container_of(work, struct request_queue, requeue_work.work);
624 struct request *rq, *next;
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 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
632 if (!(rq->rq_flags & RQF_SOFTBARRIER))
635 rq->rq_flags &= ~RQF_SOFTBARRIER;
636 list_del_init(&rq->queuelist);
637 blk_mq_sched_insert_request(rq, true, false, false, true);
640 while (!list_empty(&rq_list)) {
641 rq = list_entry(rq_list.next, struct request, queuelist);
642 list_del_init(&rq->queuelist);
643 blk_mq_sched_insert_request(rq, false, false, false, true);
646 blk_mq_run_hw_queues(q, false);
649 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
650 bool kick_requeue_list)
652 struct request_queue *q = rq->q;
656 * We abuse this flag that is otherwise used by the I/O scheduler to
657 * request head insertation from the workqueue.
659 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
661 spin_lock_irqsave(&q->requeue_lock, flags);
663 rq->rq_flags |= RQF_SOFTBARRIER;
664 list_add(&rq->queuelist, &q->requeue_list);
666 list_add_tail(&rq->queuelist, &q->requeue_list);
668 spin_unlock_irqrestore(&q->requeue_lock, flags);
670 if (kick_requeue_list)
671 blk_mq_kick_requeue_list(q);
673 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
675 void blk_mq_kick_requeue_list(struct request_queue *q)
677 kblockd_schedule_delayed_work(&q->requeue_work, 0);
679 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
681 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
684 kblockd_schedule_delayed_work(&q->requeue_work,
685 msecs_to_jiffies(msecs));
687 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
689 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
691 if (tag < tags->nr_tags) {
692 prefetch(tags->rqs[tag]);
693 return tags->rqs[tag];
698 EXPORT_SYMBOL(blk_mq_tag_to_rq);
700 struct blk_mq_timeout_data {
702 unsigned int next_set;
705 void blk_mq_rq_timed_out(struct request *req, bool reserved)
707 const struct blk_mq_ops *ops = req->q->mq_ops;
708 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
711 * We know that complete is set at this point. If STARTED isn't set
712 * anymore, then the request isn't active and the "timeout" should
713 * just be ignored. This can happen due to the bitflag ordering.
714 * Timeout first checks if STARTED is set, and if it is, assumes
715 * the request is active. But if we race with completion, then
716 * both flags will get cleared. So check here again, and ignore
717 * a timeout event with a request that isn't active.
719 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
723 ret = ops->timeout(req, reserved);
727 __blk_mq_complete_request(req);
729 case BLK_EH_RESET_TIMER:
731 blk_clear_rq_complete(req);
733 case BLK_EH_NOT_HANDLED:
736 printk(KERN_ERR "block: bad eh return: %d\n", ret);
741 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
742 struct request *rq, void *priv, bool reserved)
744 struct blk_mq_timeout_data *data = priv;
746 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
750 * The rq being checked may have been freed and reallocated
751 * out already here, we avoid this race by checking rq->deadline
752 * and REQ_ATOM_COMPLETE flag together:
754 * - if rq->deadline is observed as new value because of
755 * reusing, the rq won't be timed out because of timing.
756 * - if rq->deadline is observed as previous value,
757 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
758 * because we put a barrier between setting rq->deadline
759 * and clearing the flag in blk_mq_start_request(), so
760 * this rq won't be timed out too.
762 if (time_after_eq(jiffies, rq->deadline)) {
763 if (!blk_mark_rq_complete(rq))
764 blk_mq_rq_timed_out(rq, reserved);
765 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
766 data->next = rq->deadline;
771 static void blk_mq_timeout_work(struct work_struct *work)
773 struct request_queue *q =
774 container_of(work, struct request_queue, timeout_work);
775 struct blk_mq_timeout_data data = {
781 /* A deadlock might occur if a request is stuck requiring a
782 * timeout at the same time a queue freeze is waiting
783 * completion, since the timeout code would not be able to
784 * acquire the queue reference here.
786 * That's why we don't use blk_queue_enter here; instead, we use
787 * percpu_ref_tryget directly, because we need to be able to
788 * obtain a reference even in the short window between the queue
789 * starting to freeze, by dropping the first reference in
790 * blk_freeze_queue_start, and the moment the last request is
791 * consumed, marked by the instant q_usage_counter reaches
794 if (!percpu_ref_tryget(&q->q_usage_counter))
797 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
800 data.next = blk_rq_timeout(round_jiffies_up(data.next));
801 mod_timer(&q->timeout, data.next);
803 struct blk_mq_hw_ctx *hctx;
805 queue_for_each_hw_ctx(q, hctx, i) {
806 /* the hctx may be unmapped, so check it here */
807 if (blk_mq_hw_queue_mapped(hctx))
808 blk_mq_tag_idle(hctx);
814 struct flush_busy_ctx_data {
815 struct blk_mq_hw_ctx *hctx;
816 struct list_head *list;
819 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
821 struct flush_busy_ctx_data *flush_data = data;
822 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
823 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
825 sbitmap_clear_bit(sb, bitnr);
826 spin_lock(&ctx->lock);
827 list_splice_tail_init(&ctx->rq_list, flush_data->list);
828 spin_unlock(&ctx->lock);
833 * Process software queues that have been marked busy, splicing them
834 * to the for-dispatch
836 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
838 struct flush_busy_ctx_data data = {
843 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
845 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
847 static inline unsigned int queued_to_index(unsigned int queued)
852 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
855 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
858 struct blk_mq_alloc_data data = {
860 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
861 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
864 might_sleep_if(wait);
869 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
870 data.flags |= BLK_MQ_REQ_RESERVED;
872 rq->tag = blk_mq_get_tag(&data);
874 if (blk_mq_tag_busy(data.hctx)) {
875 rq->rq_flags |= RQF_MQ_INFLIGHT;
876 atomic_inc(&data.hctx->nr_active);
878 data.hctx->tags->rqs[rq->tag] = rq;
884 return rq->tag != -1;
887 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
890 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
893 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
894 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
895 atomic_dec(&hctx->nr_active);
899 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
902 if (rq->tag == -1 || rq->internal_tag == -1)
905 __blk_mq_put_driver_tag(hctx, rq);
908 static void blk_mq_put_driver_tag(struct request *rq)
910 struct blk_mq_hw_ctx *hctx;
912 if (rq->tag == -1 || rq->internal_tag == -1)
915 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
916 __blk_mq_put_driver_tag(hctx, rq);
920 * If we fail getting a driver tag because all the driver tags are already
921 * assigned and on the dispatch list, BUT the first entry does not have a
922 * tag, then we could deadlock. For that case, move entries with assigned
923 * driver tags to the front, leaving the set of tagged requests in the
924 * same order, and the untagged set in the same order.
926 static bool reorder_tags_to_front(struct list_head *list)
928 struct request *rq, *tmp, *first = NULL;
930 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
934 list_move(&rq->queuelist, list);
940 return first != NULL;
943 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
946 struct blk_mq_hw_ctx *hctx;
948 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
950 list_del(&wait->task_list);
951 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
952 blk_mq_run_hw_queue(hctx, true);
956 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
958 struct sbq_wait_state *ws;
961 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
962 * The thread which wins the race to grab this bit adds the hardware
963 * queue to the wait queue.
965 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
966 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
969 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
970 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
973 * As soon as this returns, it's no longer safe to fiddle with
974 * hctx->dispatch_wait, since a completion can wake up the wait queue
975 * and unlock the bit.
977 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
981 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
983 struct blk_mq_hw_ctx *hctx;
987 if (list_empty(list))
991 * Now process all the entries, sending them to the driver.
995 struct blk_mq_queue_data bd;
998 rq = list_first_entry(list, struct request, queuelist);
999 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1000 if (!queued && reorder_tags_to_front(list))
1004 * The initial allocation attempt failed, so we need to
1005 * rerun the hardware queue when a tag is freed.
1007 if (!blk_mq_dispatch_wait_add(hctx))
1011 * It's possible that a tag was freed in the window
1012 * between the allocation failure and adding the
1013 * hardware queue to the wait queue.
1015 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1019 list_del_init(&rq->queuelist);
1024 * Flag last if we have no more requests, or if we have more
1025 * but can't assign a driver tag to it.
1027 if (list_empty(list))
1030 struct request *nxt;
1032 nxt = list_first_entry(list, struct request, queuelist);
1033 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1036 ret = q->mq_ops->queue_rq(hctx, &bd);
1037 if (ret == BLK_STS_RESOURCE) {
1038 blk_mq_put_driver_tag_hctx(hctx, rq);
1039 list_add(&rq->queuelist, list);
1040 __blk_mq_requeue_request(rq);
1044 if (unlikely(ret != BLK_STS_OK)) {
1046 blk_mq_end_request(rq, BLK_STS_IOERR);
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_STS_RESOURCE. 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);
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);
1214 static void __blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx, bool sync)
1217 cancel_delayed_work_sync(&hctx->run_work);
1219 cancel_delayed_work(&hctx->run_work);
1221 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1224 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1226 __blk_mq_stop_hw_queue(hctx, false);
1228 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1230 static void __blk_mq_stop_hw_queues(struct request_queue *q, bool sync)
1232 struct blk_mq_hw_ctx *hctx;
1235 queue_for_each_hw_ctx(q, hctx, i)
1236 __blk_mq_stop_hw_queue(hctx, sync);
1239 void blk_mq_stop_hw_queues(struct request_queue *q)
1241 __blk_mq_stop_hw_queues(q, false);
1243 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1245 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1247 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1249 blk_mq_run_hw_queue(hctx, false);
1251 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1253 void blk_mq_start_hw_queues(struct request_queue *q)
1255 struct blk_mq_hw_ctx *hctx;
1258 queue_for_each_hw_ctx(q, hctx, i)
1259 blk_mq_start_hw_queue(hctx);
1261 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1263 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1265 if (!blk_mq_hctx_stopped(hctx))
1268 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1269 blk_mq_run_hw_queue(hctx, async);
1271 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1273 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1275 struct blk_mq_hw_ctx *hctx;
1278 queue_for_each_hw_ctx(q, hctx, i)
1279 blk_mq_start_stopped_hw_queue(hctx, async);
1281 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1283 static void blk_mq_run_work_fn(struct work_struct *work)
1285 struct blk_mq_hw_ctx *hctx;
1287 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1290 * If we are stopped, don't run the queue. The exception is if
1291 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1292 * the STOPPED bit and run it.
1294 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1295 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1298 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1299 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1302 __blk_mq_run_hw_queue(hctx);
1306 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1308 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1312 * Stop the hw queue, then modify currently delayed work.
1313 * This should prevent us from running the queue prematurely.
1314 * Mark the queue as auto-clearing STOPPED when it runs.
1316 blk_mq_stop_hw_queue(hctx);
1317 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1318 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1320 msecs_to_jiffies(msecs));
1322 EXPORT_SYMBOL(blk_mq_delay_queue);
1324 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1328 struct blk_mq_ctx *ctx = rq->mq_ctx;
1330 trace_block_rq_insert(hctx->queue, rq);
1333 list_add(&rq->queuelist, &ctx->rq_list);
1335 list_add_tail(&rq->queuelist, &ctx->rq_list);
1338 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1341 struct blk_mq_ctx *ctx = rq->mq_ctx;
1343 __blk_mq_insert_req_list(hctx, rq, at_head);
1344 blk_mq_hctx_mark_pending(hctx, ctx);
1347 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1348 struct list_head *list)
1352 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1355 spin_lock(&ctx->lock);
1356 while (!list_empty(list)) {
1359 rq = list_first_entry(list, struct request, queuelist);
1360 BUG_ON(rq->mq_ctx != ctx);
1361 list_del_init(&rq->queuelist);
1362 __blk_mq_insert_req_list(hctx, rq, false);
1364 blk_mq_hctx_mark_pending(hctx, ctx);
1365 spin_unlock(&ctx->lock);
1368 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1370 struct request *rqa = container_of(a, struct request, queuelist);
1371 struct request *rqb = container_of(b, struct request, queuelist);
1373 return !(rqa->mq_ctx < rqb->mq_ctx ||
1374 (rqa->mq_ctx == rqb->mq_ctx &&
1375 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1378 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1380 struct blk_mq_ctx *this_ctx;
1381 struct request_queue *this_q;
1384 LIST_HEAD(ctx_list);
1387 list_splice_init(&plug->mq_list, &list);
1389 list_sort(NULL, &list, plug_ctx_cmp);
1395 while (!list_empty(&list)) {
1396 rq = list_entry_rq(list.next);
1397 list_del_init(&rq->queuelist);
1399 if (rq->mq_ctx != this_ctx) {
1401 trace_block_unplug(this_q, depth, from_schedule);
1402 blk_mq_sched_insert_requests(this_q, this_ctx,
1407 this_ctx = rq->mq_ctx;
1413 list_add_tail(&rq->queuelist, &ctx_list);
1417 * If 'this_ctx' is set, we know we have entries to complete
1418 * on 'ctx_list'. Do those.
1421 trace_block_unplug(this_q, depth, from_schedule);
1422 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1427 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1429 blk_init_request_from_bio(rq, bio);
1431 blk_account_io_start(rq, true);
1434 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1436 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1437 !blk_queue_nomerges(hctx->queue);
1440 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1441 struct blk_mq_ctx *ctx,
1444 spin_lock(&ctx->lock);
1445 __blk_mq_insert_request(hctx, rq, false);
1446 spin_unlock(&ctx->lock);
1449 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1452 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1454 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1457 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1459 blk_qc_t *cookie, bool may_sleep)
1461 struct request_queue *q = rq->q;
1462 struct blk_mq_queue_data bd = {
1466 blk_qc_t new_cookie;
1468 bool run_queue = true;
1470 if (blk_mq_hctx_stopped(hctx)) {
1478 if (!blk_mq_get_driver_tag(rq, NULL, 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);
1491 *cookie = new_cookie;
1493 case BLK_STS_RESOURCE:
1494 __blk_mq_requeue_request(rq);
1497 *cookie = BLK_QC_T_NONE;
1498 blk_mq_end_request(rq, ret);
1503 blk_mq_sched_insert_request(rq, false, run_queue, 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(hctx, rq, cookie, false);
1514 unsigned int srcu_idx;
1518 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1519 __blk_mq_try_issue_directly(hctx, 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 blk_queue_split(q, &bio, q->bio_split);
1540 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1542 return BLK_QC_T_NONE;
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_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_put_ctx(data.ctx);
1569 blk_mq_bio_to_request(rq, bio);
1571 blk_mq_sched_insert_request(rq, false, true, true,
1574 blk_insert_flush(rq);
1575 blk_mq_run_hw_queue(data.hctx, true);
1577 } else if (plug && q->nr_hw_queues == 1) {
1578 struct request *last = NULL;
1580 blk_mq_put_ctx(data.ctx);
1581 blk_mq_bio_to_request(rq, bio);
1584 * @request_count may become stale because of schedule
1585 * out, so check the list again.
1587 if (list_empty(&plug->mq_list))
1589 else if (blk_queue_nomerges(q))
1590 request_count = blk_plug_queued_count(q);
1593 trace_block_plug(q);
1595 last = list_entry_rq(plug->mq_list.prev);
1597 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1598 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1599 blk_flush_plug_list(plug, false);
1600 trace_block_plug(q);
1603 list_add_tail(&rq->queuelist, &plug->mq_list);
1604 } else if (plug && !blk_queue_nomerges(q)) {
1605 blk_mq_bio_to_request(rq, bio);
1608 * We do limited plugging. If the bio can be merged, do that.
1609 * Otherwise the existing request in the plug list will be
1610 * issued. So the plug list will have one request at most
1611 * The plug list might get flushed before this. If that happens,
1612 * the plug list is empty, and same_queue_rq is invalid.
1614 if (list_empty(&plug->mq_list))
1615 same_queue_rq = NULL;
1617 list_del_init(&same_queue_rq->queuelist);
1618 list_add_tail(&rq->queuelist, &plug->mq_list);
1620 blk_mq_put_ctx(data.ctx);
1622 if (same_queue_rq) {
1623 data.hctx = blk_mq_map_queue(q,
1624 same_queue_rq->mq_ctx->cpu);
1625 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1628 } else if (q->nr_hw_queues > 1 && is_sync) {
1629 blk_mq_put_ctx(data.ctx);
1630 blk_mq_bio_to_request(rq, bio);
1631 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1632 } else if (q->elevator) {
1633 blk_mq_put_ctx(data.ctx);
1634 blk_mq_bio_to_request(rq, bio);
1635 blk_mq_sched_insert_request(rq, false, true, true, true);
1637 blk_mq_put_ctx(data.ctx);
1638 blk_mq_bio_to_request(rq, bio);
1639 blk_mq_queue_io(data.hctx, data.ctx, rq);
1640 blk_mq_run_hw_queue(data.hctx, true);
1646 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1647 unsigned int hctx_idx)
1651 if (tags->rqs && set->ops->exit_request) {
1654 for (i = 0; i < tags->nr_tags; i++) {
1655 struct request *rq = tags->static_rqs[i];
1659 set->ops->exit_request(set, rq, hctx_idx);
1660 tags->static_rqs[i] = NULL;
1664 while (!list_empty(&tags->page_list)) {
1665 page = list_first_entry(&tags->page_list, struct page, lru);
1666 list_del_init(&page->lru);
1668 * Remove kmemleak object previously allocated in
1669 * blk_mq_init_rq_map().
1671 kmemleak_free(page_address(page));
1672 __free_pages(page, page->private);
1676 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1680 kfree(tags->static_rqs);
1681 tags->static_rqs = NULL;
1683 blk_mq_free_tags(tags);
1686 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1687 unsigned int hctx_idx,
1688 unsigned int nr_tags,
1689 unsigned int reserved_tags)
1691 struct blk_mq_tags *tags;
1694 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1695 if (node == NUMA_NO_NODE)
1696 node = set->numa_node;
1698 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1699 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1703 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1704 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1707 blk_mq_free_tags(tags);
1711 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1712 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1714 if (!tags->static_rqs) {
1716 blk_mq_free_tags(tags);
1723 static size_t order_to_size(unsigned int order)
1725 return (size_t)PAGE_SIZE << order;
1728 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1729 unsigned int hctx_idx, unsigned int depth)
1731 unsigned int i, j, entries_per_page, max_order = 4;
1732 size_t rq_size, left;
1735 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1736 if (node == NUMA_NO_NODE)
1737 node = set->numa_node;
1739 INIT_LIST_HEAD(&tags->page_list);
1742 * rq_size is the size of the request plus driver payload, rounded
1743 * to the cacheline size
1745 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1747 left = rq_size * depth;
1749 for (i = 0; i < depth; ) {
1750 int this_order = max_order;
1755 while (this_order && left < order_to_size(this_order - 1))
1759 page = alloc_pages_node(node,
1760 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1766 if (order_to_size(this_order) < rq_size)
1773 page->private = this_order;
1774 list_add_tail(&page->lru, &tags->page_list);
1776 p = page_address(page);
1778 * Allow kmemleak to scan these pages as they contain pointers
1779 * to additional allocations like via ops->init_request().
1781 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1782 entries_per_page = order_to_size(this_order) / rq_size;
1783 to_do = min(entries_per_page, depth - i);
1784 left -= to_do * rq_size;
1785 for (j = 0; j < to_do; j++) {
1786 struct request *rq = p;
1788 tags->static_rqs[i] = rq;
1789 if (set->ops->init_request) {
1790 if (set->ops->init_request(set, rq, hctx_idx,
1792 tags->static_rqs[i] = NULL;
1804 blk_mq_free_rqs(set, tags, hctx_idx);
1809 * 'cpu' is going away. splice any existing rq_list entries from this
1810 * software queue to the hw queue dispatch list, and ensure that it
1813 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1815 struct blk_mq_hw_ctx *hctx;
1816 struct blk_mq_ctx *ctx;
1819 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1820 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1822 spin_lock(&ctx->lock);
1823 if (!list_empty(&ctx->rq_list)) {
1824 list_splice_init(&ctx->rq_list, &tmp);
1825 blk_mq_hctx_clear_pending(hctx, ctx);
1827 spin_unlock(&ctx->lock);
1829 if (list_empty(&tmp))
1832 spin_lock(&hctx->lock);
1833 list_splice_tail_init(&tmp, &hctx->dispatch);
1834 spin_unlock(&hctx->lock);
1836 blk_mq_run_hw_queue(hctx, true);
1840 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1842 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1846 /* hctx->ctxs will be freed in queue's release handler */
1847 static void blk_mq_exit_hctx(struct request_queue *q,
1848 struct blk_mq_tag_set *set,
1849 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1851 blk_mq_debugfs_unregister_hctx(hctx);
1853 blk_mq_tag_idle(hctx);
1855 if (set->ops->exit_request)
1856 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1858 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1860 if (set->ops->exit_hctx)
1861 set->ops->exit_hctx(hctx, hctx_idx);
1863 if (hctx->flags & BLK_MQ_F_BLOCKING)
1864 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1866 blk_mq_remove_cpuhp(hctx);
1867 blk_free_flush_queue(hctx->fq);
1868 sbitmap_free(&hctx->ctx_map);
1871 static void blk_mq_exit_hw_queues(struct request_queue *q,
1872 struct blk_mq_tag_set *set, int nr_queue)
1874 struct blk_mq_hw_ctx *hctx;
1877 queue_for_each_hw_ctx(q, hctx, i) {
1880 blk_mq_exit_hctx(q, set, hctx, i);
1884 static int blk_mq_init_hctx(struct request_queue *q,
1885 struct blk_mq_tag_set *set,
1886 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1890 node = hctx->numa_node;
1891 if (node == NUMA_NO_NODE)
1892 node = hctx->numa_node = set->numa_node;
1894 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1895 spin_lock_init(&hctx->lock);
1896 INIT_LIST_HEAD(&hctx->dispatch);
1898 hctx->queue_num = hctx_idx;
1899 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1901 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1903 hctx->tags = set->tags[hctx_idx];
1906 * Allocate space for all possible cpus to avoid allocation at
1909 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1912 goto unregister_cpu_notifier;
1914 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1920 if (set->ops->init_hctx &&
1921 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1924 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1927 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1929 goto sched_exit_hctx;
1931 if (set->ops->init_request &&
1932 set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1936 if (hctx->flags & BLK_MQ_F_BLOCKING)
1937 init_srcu_struct(&hctx->queue_rq_srcu);
1939 blk_mq_debugfs_register_hctx(q, hctx);
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_mq_poll_stats_bkt,
2274 BLK_MQ_POLL_STATS_BKTS, q);
2278 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2282 /* init q->mq_kobj and sw queues' kobjects */
2283 blk_mq_sysfs_init(q);
2285 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2286 GFP_KERNEL, set->numa_node);
2287 if (!q->queue_hw_ctx)
2290 q->mq_map = set->mq_map;
2292 blk_mq_realloc_hw_ctxs(set, q);
2293 if (!q->nr_hw_queues)
2296 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2297 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2299 q->nr_queues = nr_cpu_ids;
2301 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2303 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2304 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2306 q->sg_reserved_size = INT_MAX;
2308 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2309 INIT_LIST_HEAD(&q->requeue_list);
2310 spin_lock_init(&q->requeue_lock);
2312 blk_queue_make_request(q, blk_mq_make_request);
2315 * Do this after blk_queue_make_request() overrides it...
2317 q->nr_requests = set->queue_depth;
2320 * Default to classic polling
2324 if (set->ops->complete)
2325 blk_queue_softirq_done(q, set->ops->complete);
2327 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2330 mutex_lock(&all_q_mutex);
2332 list_add_tail(&q->all_q_node, &all_q_list);
2333 blk_mq_add_queue_tag_set(set, q);
2334 blk_mq_map_swqueue(q, cpu_online_mask);
2336 mutex_unlock(&all_q_mutex);
2339 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2342 ret = blk_mq_sched_init(q);
2344 return ERR_PTR(ret);
2350 kfree(q->queue_hw_ctx);
2352 free_percpu(q->queue_ctx);
2355 return ERR_PTR(-ENOMEM);
2357 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2359 void blk_mq_free_queue(struct request_queue *q)
2361 struct blk_mq_tag_set *set = q->tag_set;
2363 mutex_lock(&all_q_mutex);
2364 list_del_init(&q->all_q_node);
2365 mutex_unlock(&all_q_mutex);
2367 blk_mq_del_queue_tag_set(q);
2369 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2372 /* Basically redo blk_mq_init_queue with queue frozen */
2373 static void blk_mq_queue_reinit(struct request_queue *q,
2374 const struct cpumask *online_mask)
2376 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2378 blk_mq_debugfs_unregister_hctxs(q);
2379 blk_mq_sysfs_unregister(q);
2382 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2383 * we should change hctx numa_node according to new topology (this
2384 * involves free and re-allocate memory, worthy doing?)
2387 blk_mq_map_swqueue(q, online_mask);
2389 blk_mq_sysfs_register(q);
2390 blk_mq_debugfs_register_hctxs(q);
2394 * New online cpumask which is going to be set in this hotplug event.
2395 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2396 * one-by-one and dynamically allocating this could result in a failure.
2398 static struct cpumask cpuhp_online_new;
2400 static void blk_mq_queue_reinit_work(void)
2402 struct request_queue *q;
2404 mutex_lock(&all_q_mutex);
2406 * We need to freeze and reinit all existing queues. Freezing
2407 * involves synchronous wait for an RCU grace period and doing it
2408 * one by one may take a long time. Start freezing all queues in
2409 * one swoop and then wait for the completions so that freezing can
2410 * take place in parallel.
2412 list_for_each_entry(q, &all_q_list, all_q_node)
2413 blk_freeze_queue_start(q);
2414 list_for_each_entry(q, &all_q_list, all_q_node)
2415 blk_mq_freeze_queue_wait(q);
2417 list_for_each_entry(q, &all_q_list, all_q_node)
2418 blk_mq_queue_reinit(q, &cpuhp_online_new);
2420 list_for_each_entry(q, &all_q_list, all_q_node)
2421 blk_mq_unfreeze_queue(q);
2423 mutex_unlock(&all_q_mutex);
2426 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2428 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2429 blk_mq_queue_reinit_work();
2434 * Before hotadded cpu starts handling requests, new mappings must be
2435 * established. Otherwise, these requests in hw queue might never be
2438 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2439 * for CPU0, and ctx1 for CPU1).
2441 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2442 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2444 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2445 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2446 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2449 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2451 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2452 cpumask_set_cpu(cpu, &cpuhp_online_new);
2453 blk_mq_queue_reinit_work();
2457 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2461 for (i = 0; i < set->nr_hw_queues; i++)
2462 if (!__blk_mq_alloc_rq_map(set, i))
2469 blk_mq_free_rq_map(set->tags[i]);
2475 * Allocate the request maps associated with this tag_set. Note that this
2476 * may reduce the depth asked for, if memory is tight. set->queue_depth
2477 * will be updated to reflect the allocated depth.
2479 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2484 depth = set->queue_depth;
2486 err = __blk_mq_alloc_rq_maps(set);
2490 set->queue_depth >>= 1;
2491 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2495 } while (set->queue_depth);
2497 if (!set->queue_depth || err) {
2498 pr_err("blk-mq: failed to allocate request map\n");
2502 if (depth != set->queue_depth)
2503 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2504 depth, set->queue_depth);
2509 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2511 if (set->ops->map_queues)
2512 return set->ops->map_queues(set);
2514 return blk_mq_map_queues(set);
2518 * Alloc a tag set to be associated with one or more request queues.
2519 * May fail with EINVAL for various error conditions. May adjust the
2520 * requested depth down, if if it too large. In that case, the set
2521 * value will be stored in set->queue_depth.
2523 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2527 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2529 if (!set->nr_hw_queues)
2531 if (!set->queue_depth)
2533 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2536 if (!set->ops->queue_rq)
2539 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2540 pr_info("blk-mq: reduced tag depth to %u\n",
2542 set->queue_depth = BLK_MQ_MAX_DEPTH;
2546 * If a crashdump is active, then we are potentially in a very
2547 * memory constrained environment. Limit us to 1 queue and
2548 * 64 tags to prevent using too much memory.
2550 if (is_kdump_kernel()) {
2551 set->nr_hw_queues = 1;
2552 set->queue_depth = min(64U, set->queue_depth);
2555 * There is no use for more h/w queues than cpus.
2557 if (set->nr_hw_queues > nr_cpu_ids)
2558 set->nr_hw_queues = nr_cpu_ids;
2560 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2561 GFP_KERNEL, set->numa_node);
2566 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2567 GFP_KERNEL, set->numa_node);
2571 ret = blk_mq_update_queue_map(set);
2573 goto out_free_mq_map;
2575 ret = blk_mq_alloc_rq_maps(set);
2577 goto out_free_mq_map;
2579 mutex_init(&set->tag_list_lock);
2580 INIT_LIST_HEAD(&set->tag_list);
2592 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2594 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2598 for (i = 0; i < nr_cpu_ids; i++)
2599 blk_mq_free_map_and_requests(set, i);
2607 EXPORT_SYMBOL(blk_mq_free_tag_set);
2609 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2611 struct blk_mq_tag_set *set = q->tag_set;
2612 struct blk_mq_hw_ctx *hctx;
2618 blk_mq_freeze_queue(q);
2621 queue_for_each_hw_ctx(q, hctx, i) {
2625 * If we're using an MQ scheduler, just update the scheduler
2626 * queue depth. This is similar to what the old code would do.
2628 if (!hctx->sched_tags) {
2629 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2630 min(nr, set->queue_depth),
2633 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2641 q->nr_requests = nr;
2643 blk_mq_unfreeze_queue(q);
2648 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2651 struct request_queue *q;
2653 lockdep_assert_held(&set->tag_list_lock);
2655 if (nr_hw_queues > nr_cpu_ids)
2656 nr_hw_queues = nr_cpu_ids;
2657 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2660 list_for_each_entry(q, &set->tag_list, tag_set_list)
2661 blk_mq_freeze_queue(q);
2663 set->nr_hw_queues = nr_hw_queues;
2664 blk_mq_update_queue_map(set);
2665 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2666 blk_mq_realloc_hw_ctxs(set, q);
2667 blk_mq_queue_reinit(q, cpu_online_mask);
2670 list_for_each_entry(q, &set->tag_list, tag_set_list)
2671 blk_mq_unfreeze_queue(q);
2674 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2676 mutex_lock(&set->tag_list_lock);
2677 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2678 mutex_unlock(&set->tag_list_lock);
2680 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2682 /* Enable polling stats and return whether they were already enabled. */
2683 static bool blk_poll_stats_enable(struct request_queue *q)
2685 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2686 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2688 blk_stat_add_callback(q, q->poll_cb);
2692 static void blk_mq_poll_stats_start(struct request_queue *q)
2695 * We don't arm the callback if polling stats are not enabled or the
2696 * callback is already active.
2698 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2699 blk_stat_is_active(q->poll_cb))
2702 blk_stat_activate_msecs(q->poll_cb, 100);
2705 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2707 struct request_queue *q = cb->data;
2710 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2711 if (cb->stat[bucket].nr_samples)
2712 q->poll_stat[bucket] = cb->stat[bucket];
2716 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2717 struct blk_mq_hw_ctx *hctx,
2720 unsigned long ret = 0;
2724 * If stats collection isn't on, don't sleep but turn it on for
2727 if (!blk_poll_stats_enable(q))
2731 * As an optimistic guess, use half of the mean service time
2732 * for this type of request. We can (and should) make this smarter.
2733 * For instance, if the completion latencies are tight, we can
2734 * get closer than just half the mean. This is especially
2735 * important on devices where the completion latencies are longer
2736 * than ~10 usec. We do use the stats for the relevant IO size
2737 * if available which does lead to better estimates.
2739 bucket = blk_mq_poll_stats_bkt(rq);
2743 if (q->poll_stat[bucket].nr_samples)
2744 ret = (q->poll_stat[bucket].mean + 1) / 2;
2749 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2750 struct blk_mq_hw_ctx *hctx,
2753 struct hrtimer_sleeper hs;
2754 enum hrtimer_mode mode;
2758 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2764 * -1: don't ever hybrid sleep
2765 * 0: use half of prev avg
2766 * >0: use this specific value
2768 if (q->poll_nsec == -1)
2770 else if (q->poll_nsec > 0)
2771 nsecs = q->poll_nsec;
2773 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2778 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2781 * This will be replaced with the stats tracking code, using
2782 * 'avg_completion_time / 2' as the pre-sleep target.
2786 mode = HRTIMER_MODE_REL;
2787 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2788 hrtimer_set_expires(&hs.timer, kt);
2790 hrtimer_init_sleeper(&hs, current);
2792 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2794 set_current_state(TASK_UNINTERRUPTIBLE);
2795 hrtimer_start_expires(&hs.timer, mode);
2798 hrtimer_cancel(&hs.timer);
2799 mode = HRTIMER_MODE_ABS;
2800 } while (hs.task && !signal_pending(current));
2802 __set_current_state(TASK_RUNNING);
2803 destroy_hrtimer_on_stack(&hs.timer);
2807 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2809 struct request_queue *q = hctx->queue;
2813 * If we sleep, have the caller restart the poll loop to reset
2814 * the state. Like for the other success return cases, the
2815 * caller is responsible for checking if the IO completed. If
2816 * the IO isn't complete, we'll get called again and will go
2817 * straight to the busy poll loop.
2819 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2822 hctx->poll_considered++;
2824 state = current->state;
2825 while (!need_resched()) {
2828 hctx->poll_invoked++;
2830 ret = q->mq_ops->poll(hctx, rq->tag);
2832 hctx->poll_success++;
2833 set_current_state(TASK_RUNNING);
2837 if (signal_pending_state(state, current))
2838 set_current_state(TASK_RUNNING);
2840 if (current->state == TASK_RUNNING)
2850 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2852 struct blk_mq_hw_ctx *hctx;
2853 struct blk_plug *plug;
2856 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2857 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2860 plug = current->plug;
2862 blk_flush_plug_list(plug, false);
2864 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2865 if (!blk_qc_t_is_internal(cookie))
2866 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2868 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2870 * With scheduling, if the request has completed, we'll
2871 * get a NULL return here, as we clear the sched tag when
2872 * that happens. The request still remains valid, like always,
2873 * so we should be safe with just the NULL check.
2879 return __blk_mq_poll(hctx, rq);
2881 EXPORT_SYMBOL_GPL(blk_mq_poll);
2883 void blk_mq_disable_hotplug(void)
2885 mutex_lock(&all_q_mutex);
2888 void blk_mq_enable_hotplug(void)
2890 mutex_unlock(&all_q_mutex);
2893 static int __init blk_mq_init(void)
2895 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2896 blk_mq_hctx_notify_dead);
2898 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2899 blk_mq_queue_reinit_prepare,
2900 blk_mq_queue_reinit_dead);
2903 subsys_initcall(blk_mq_init);