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/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
35 #include "blk-mq-sched.h"
37 static DEFINE_MUTEX(all_q_mutex);
38 static LIST_HEAD(all_q_list);
41 * Check if any of the ctx's have pending work in this hardware queue
43 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
45 return sbitmap_any_bit_set(&hctx->ctx_map) ||
46 !list_empty_careful(&hctx->dispatch) ||
47 blk_mq_sched_has_work(hctx);
51 * Mark this ctx as having pending work in this hardware queue
53 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
54 struct blk_mq_ctx *ctx)
56 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
57 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
60 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
61 struct blk_mq_ctx *ctx)
63 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
66 void blk_mq_freeze_queue_start(struct request_queue *q)
70 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
71 if (freeze_depth == 1) {
72 percpu_ref_kill(&q->q_usage_counter);
73 blk_mq_run_hw_queues(q, false);
76 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
78 static void blk_mq_freeze_queue_wait(struct request_queue *q)
80 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
84 * Guarantee no request is in use, so we can change any data structure of
85 * the queue afterward.
87 void blk_freeze_queue(struct request_queue *q)
90 * In the !blk_mq case we are only calling this to kill the
91 * q_usage_counter, otherwise this increases the freeze depth
92 * and waits for it to return to zero. For this reason there is
93 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
94 * exported to drivers as the only user for unfreeze is blk_mq.
96 blk_mq_freeze_queue_start(q);
97 blk_mq_freeze_queue_wait(q);
100 void blk_mq_freeze_queue(struct request_queue *q)
103 * ...just an alias to keep freeze and unfreeze actions balanced
104 * in the blk_mq_* namespace
108 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
110 void blk_mq_unfreeze_queue(struct request_queue *q)
114 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
115 WARN_ON_ONCE(freeze_depth < 0);
117 percpu_ref_reinit(&q->q_usage_counter);
118 wake_up_all(&q->mq_freeze_wq);
121 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
124 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
127 * Note: this function does not prevent that the struct request end_io()
128 * callback function is invoked. Additionally, it is not prevented that
129 * new queue_rq() calls occur unless the queue has been stopped first.
131 void blk_mq_quiesce_queue(struct request_queue *q)
133 struct blk_mq_hw_ctx *hctx;
137 blk_mq_stop_hw_queues(q);
139 queue_for_each_hw_ctx(q, hctx, i) {
140 if (hctx->flags & BLK_MQ_F_BLOCKING)
141 synchronize_srcu(&hctx->queue_rq_srcu);
148 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
150 void blk_mq_wake_waiters(struct request_queue *q)
152 struct blk_mq_hw_ctx *hctx;
155 queue_for_each_hw_ctx(q, hctx, i)
156 if (blk_mq_hw_queue_mapped(hctx))
157 blk_mq_tag_wakeup_all(hctx->tags, true);
160 * If we are called because the queue has now been marked as
161 * dying, we need to ensure that processes currently waiting on
162 * the queue are notified as well.
164 wake_up_all(&q->mq_freeze_wq);
167 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
169 return blk_mq_has_free_tags(hctx->tags);
171 EXPORT_SYMBOL(blk_mq_can_queue);
173 void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
174 struct request *rq, unsigned int op)
176 INIT_LIST_HEAD(&rq->queuelist);
177 /* csd/requeue_work/fifo_time is initialized before use */
181 if (blk_queue_io_stat(q))
182 rq->rq_flags |= RQF_IO_STAT;
183 /* do not touch atomic flags, it needs atomic ops against the timer */
185 INIT_HLIST_NODE(&rq->hash);
186 RB_CLEAR_NODE(&rq->rb_node);
189 rq->start_time = jiffies;
190 #ifdef CONFIG_BLK_CGROUP
192 set_start_time_ns(rq);
193 rq->io_start_time_ns = 0;
195 rq->nr_phys_segments = 0;
196 #if defined(CONFIG_BLK_DEV_INTEGRITY)
197 rq->nr_integrity_segments = 0;
200 /* tag was already set */
204 INIT_LIST_HEAD(&rq->timeout_list);
208 rq->end_io_data = NULL;
211 ctx->rq_dispatched[op_is_sync(op)]++;
213 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init);
215 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
221 tag = blk_mq_get_tag(data);
222 if (tag != BLK_MQ_TAG_FAIL) {
223 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
225 rq = tags->static_rqs[tag];
227 if (data->flags & BLK_MQ_REQ_INTERNAL) {
229 rq->internal_tag = tag;
231 if (blk_mq_tag_busy(data->hctx)) {
232 rq->rq_flags = RQF_MQ_INFLIGHT;
233 atomic_inc(&data->hctx->nr_active);
236 rq->internal_tag = -1;
237 data->hctx->tags->rqs[rq->tag] = rq;
240 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
246 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
248 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
251 struct blk_mq_alloc_data alloc_data = { .flags = flags };
255 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
259 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
261 blk_mq_put_ctx(alloc_data.ctx);
265 return ERR_PTR(-EWOULDBLOCK);
268 rq->__sector = (sector_t) -1;
269 rq->bio = rq->biotail = NULL;
272 EXPORT_SYMBOL(blk_mq_alloc_request);
274 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
275 unsigned int flags, unsigned int hctx_idx)
277 struct blk_mq_alloc_data alloc_data = { .flags = flags };
283 * If the tag allocator sleeps we could get an allocation for a
284 * different hardware context. No need to complicate the low level
285 * allocator for this for the rare use case of a command tied to
288 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
289 return ERR_PTR(-EINVAL);
291 if (hctx_idx >= q->nr_hw_queues)
292 return ERR_PTR(-EIO);
294 ret = blk_queue_enter(q, true);
299 * Check if the hardware context is actually mapped to anything.
300 * If not tell the caller that it should skip this queue.
302 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
303 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
305 return ERR_PTR(-EXDEV);
307 cpu = cpumask_first(alloc_data.hctx->cpumask);
308 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
310 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
312 blk_mq_put_ctx(alloc_data.ctx);
316 return ERR_PTR(-EWOULDBLOCK);
320 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
322 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
325 const int sched_tag = rq->internal_tag;
326 struct request_queue *q = rq->q;
328 if (rq->rq_flags & RQF_MQ_INFLIGHT)
329 atomic_dec(&hctx->nr_active);
331 wbt_done(q->rq_wb, &rq->issue_stat);
334 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
335 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
337 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
339 blk_mq_sched_completed_request(hctx, rq);
340 blk_mq_sched_restart_queues(hctx);
344 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
347 struct blk_mq_ctx *ctx = rq->mq_ctx;
349 ctx->rq_completed[rq_is_sync(rq)]++;
350 __blk_mq_finish_request(hctx, ctx, rq);
353 void blk_mq_finish_request(struct request *rq)
355 blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
358 void blk_mq_free_request(struct request *rq)
360 blk_mq_sched_put_request(rq);
362 EXPORT_SYMBOL_GPL(blk_mq_free_request);
364 inline void __blk_mq_end_request(struct request *rq, int error)
366 blk_account_io_done(rq);
369 wbt_done(rq->q->rq_wb, &rq->issue_stat);
370 rq->end_io(rq, error);
372 if (unlikely(blk_bidi_rq(rq)))
373 blk_mq_free_request(rq->next_rq);
374 blk_mq_free_request(rq);
377 EXPORT_SYMBOL(__blk_mq_end_request);
379 void blk_mq_end_request(struct request *rq, int error)
381 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
383 __blk_mq_end_request(rq, error);
385 EXPORT_SYMBOL(blk_mq_end_request);
387 static void __blk_mq_complete_request_remote(void *data)
389 struct request *rq = data;
391 rq->q->softirq_done_fn(rq);
394 static void blk_mq_ipi_complete_request(struct request *rq)
396 struct blk_mq_ctx *ctx = rq->mq_ctx;
400 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
401 rq->q->softirq_done_fn(rq);
406 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
407 shared = cpus_share_cache(cpu, ctx->cpu);
409 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
410 rq->csd.func = __blk_mq_complete_request_remote;
413 smp_call_function_single_async(ctx->cpu, &rq->csd);
415 rq->q->softirq_done_fn(rq);
420 static void blk_mq_stat_add(struct request *rq)
422 if (rq->rq_flags & RQF_STATS) {
424 * We could rq->mq_ctx here, but there's less of a risk
425 * of races if we have the completion event add the stats
426 * to the local software queue.
428 struct blk_mq_ctx *ctx;
430 ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id());
431 blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq);
435 static void __blk_mq_complete_request(struct request *rq)
437 struct request_queue *q = rq->q;
441 if (!q->softirq_done_fn)
442 blk_mq_end_request(rq, rq->errors);
444 blk_mq_ipi_complete_request(rq);
448 * blk_mq_complete_request - end I/O on a request
449 * @rq: the request being processed
452 * Ends all I/O on a request. It does not handle partial completions.
453 * The actual completion happens out-of-order, through a IPI handler.
455 void blk_mq_complete_request(struct request *rq, int error)
457 struct request_queue *q = rq->q;
459 if (unlikely(blk_should_fake_timeout(q)))
461 if (!blk_mark_rq_complete(rq)) {
463 __blk_mq_complete_request(rq);
466 EXPORT_SYMBOL(blk_mq_complete_request);
468 int blk_mq_request_started(struct request *rq)
470 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
472 EXPORT_SYMBOL_GPL(blk_mq_request_started);
474 void blk_mq_start_request(struct request *rq)
476 struct request_queue *q = rq->q;
478 blk_mq_sched_started_request(rq);
480 trace_block_rq_issue(q, rq);
482 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
483 blk_stat_set_issue_time(&rq->issue_stat);
484 rq->rq_flags |= RQF_STATS;
485 wbt_issue(q->rq_wb, &rq->issue_stat);
491 * Ensure that ->deadline is visible before set the started
492 * flag and clear the completed flag.
494 smp_mb__before_atomic();
497 * Mark us as started and clear complete. Complete might have been
498 * set if requeue raced with timeout, which then marked it as
499 * complete. So be sure to clear complete again when we start
500 * the request, otherwise we'll ignore the completion event.
502 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
503 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
504 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
505 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
507 if (q->dma_drain_size && blk_rq_bytes(rq)) {
509 * Make sure space for the drain appears. We know we can do
510 * this because max_hw_segments has been adjusted to be one
511 * fewer than the device can handle.
513 rq->nr_phys_segments++;
516 EXPORT_SYMBOL(blk_mq_start_request);
518 static void __blk_mq_requeue_request(struct request *rq)
520 struct request_queue *q = rq->q;
522 trace_block_rq_requeue(q, rq);
523 wbt_requeue(q->rq_wb, &rq->issue_stat);
524 blk_mq_sched_requeue_request(rq);
526 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
527 if (q->dma_drain_size && blk_rq_bytes(rq))
528 rq->nr_phys_segments--;
532 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
534 __blk_mq_requeue_request(rq);
536 BUG_ON(blk_queued_rq(rq));
537 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
539 EXPORT_SYMBOL(blk_mq_requeue_request);
541 static void blk_mq_requeue_work(struct work_struct *work)
543 struct request_queue *q =
544 container_of(work, struct request_queue, requeue_work.work);
546 struct request *rq, *next;
549 spin_lock_irqsave(&q->requeue_lock, flags);
550 list_splice_init(&q->requeue_list, &rq_list);
551 spin_unlock_irqrestore(&q->requeue_lock, flags);
553 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
554 if (!(rq->rq_flags & RQF_SOFTBARRIER))
557 rq->rq_flags &= ~RQF_SOFTBARRIER;
558 list_del_init(&rq->queuelist);
559 blk_mq_sched_insert_request(rq, true, false, false, true);
562 while (!list_empty(&rq_list)) {
563 rq = list_entry(rq_list.next, struct request, queuelist);
564 list_del_init(&rq->queuelist);
565 blk_mq_sched_insert_request(rq, false, false, false, true);
568 blk_mq_run_hw_queues(q, false);
571 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
572 bool kick_requeue_list)
574 struct request_queue *q = rq->q;
578 * We abuse this flag that is otherwise used by the I/O scheduler to
579 * request head insertation from the workqueue.
581 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
583 spin_lock_irqsave(&q->requeue_lock, flags);
585 rq->rq_flags |= RQF_SOFTBARRIER;
586 list_add(&rq->queuelist, &q->requeue_list);
588 list_add_tail(&rq->queuelist, &q->requeue_list);
590 spin_unlock_irqrestore(&q->requeue_lock, flags);
592 if (kick_requeue_list)
593 blk_mq_kick_requeue_list(q);
595 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
597 void blk_mq_kick_requeue_list(struct request_queue *q)
599 kblockd_schedule_delayed_work(&q->requeue_work, 0);
601 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
603 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
606 kblockd_schedule_delayed_work(&q->requeue_work,
607 msecs_to_jiffies(msecs));
609 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
611 void blk_mq_abort_requeue_list(struct request_queue *q)
616 spin_lock_irqsave(&q->requeue_lock, flags);
617 list_splice_init(&q->requeue_list, &rq_list);
618 spin_unlock_irqrestore(&q->requeue_lock, flags);
620 while (!list_empty(&rq_list)) {
623 rq = list_first_entry(&rq_list, struct request, queuelist);
624 list_del_init(&rq->queuelist);
626 blk_mq_end_request(rq, rq->errors);
629 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
631 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
633 if (tag < tags->nr_tags) {
634 prefetch(tags->rqs[tag]);
635 return tags->rqs[tag];
640 EXPORT_SYMBOL(blk_mq_tag_to_rq);
642 struct blk_mq_timeout_data {
644 unsigned int next_set;
647 void blk_mq_rq_timed_out(struct request *req, bool reserved)
649 const struct blk_mq_ops *ops = req->q->mq_ops;
650 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
653 * We know that complete is set at this point. If STARTED isn't set
654 * anymore, then the request isn't active and the "timeout" should
655 * just be ignored. This can happen due to the bitflag ordering.
656 * Timeout first checks if STARTED is set, and if it is, assumes
657 * the request is active. But if we race with completion, then
658 * we both flags will get cleared. So check here again, and ignore
659 * a timeout event with a request that isn't active.
661 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
665 ret = ops->timeout(req, reserved);
669 __blk_mq_complete_request(req);
671 case BLK_EH_RESET_TIMER:
673 blk_clear_rq_complete(req);
675 case BLK_EH_NOT_HANDLED:
678 printk(KERN_ERR "block: bad eh return: %d\n", ret);
683 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
684 struct request *rq, void *priv, bool reserved)
686 struct blk_mq_timeout_data *data = priv;
688 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
690 * If a request wasn't started before the queue was
691 * marked dying, kill it here or it'll go unnoticed.
693 if (unlikely(blk_queue_dying(rq->q))) {
695 blk_mq_end_request(rq, rq->errors);
700 if (time_after_eq(jiffies, rq->deadline)) {
701 if (!blk_mark_rq_complete(rq))
702 blk_mq_rq_timed_out(rq, reserved);
703 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
704 data->next = rq->deadline;
709 static void blk_mq_timeout_work(struct work_struct *work)
711 struct request_queue *q =
712 container_of(work, struct request_queue, timeout_work);
713 struct blk_mq_timeout_data data = {
719 /* A deadlock might occur if a request is stuck requiring a
720 * timeout at the same time a queue freeze is waiting
721 * completion, since the timeout code would not be able to
722 * acquire the queue reference here.
724 * That's why we don't use blk_queue_enter here; instead, we use
725 * percpu_ref_tryget directly, because we need to be able to
726 * obtain a reference even in the short window between the queue
727 * starting to freeze, by dropping the first reference in
728 * blk_mq_freeze_queue_start, and the moment the last request is
729 * consumed, marked by the instant q_usage_counter reaches
732 if (!percpu_ref_tryget(&q->q_usage_counter))
735 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
738 data.next = blk_rq_timeout(round_jiffies_up(data.next));
739 mod_timer(&q->timeout, data.next);
741 struct blk_mq_hw_ctx *hctx;
743 queue_for_each_hw_ctx(q, hctx, i) {
744 /* the hctx may be unmapped, so check it here */
745 if (blk_mq_hw_queue_mapped(hctx))
746 blk_mq_tag_idle(hctx);
753 * Reverse check our software queue for entries that we could potentially
754 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
755 * too much time checking for merges.
757 static bool blk_mq_attempt_merge(struct request_queue *q,
758 struct blk_mq_ctx *ctx, struct bio *bio)
763 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
769 if (!blk_rq_merge_ok(rq, bio))
772 switch (blk_try_merge(rq, bio)) {
773 case ELEVATOR_BACK_MERGE:
774 if (blk_mq_sched_allow_merge(q, rq, bio))
775 merged = bio_attempt_back_merge(q, rq, bio);
777 case ELEVATOR_FRONT_MERGE:
778 if (blk_mq_sched_allow_merge(q, rq, bio))
779 merged = bio_attempt_front_merge(q, rq, bio);
781 case ELEVATOR_DISCARD_MERGE:
782 merged = bio_attempt_discard_merge(q, rq, bio);
796 struct flush_busy_ctx_data {
797 struct blk_mq_hw_ctx *hctx;
798 struct list_head *list;
801 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
803 struct flush_busy_ctx_data *flush_data = data;
804 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
805 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
807 sbitmap_clear_bit(sb, bitnr);
808 spin_lock(&ctx->lock);
809 list_splice_tail_init(&ctx->rq_list, flush_data->list);
810 spin_unlock(&ctx->lock);
815 * Process software queues that have been marked busy, splicing them
816 * to the for-dispatch
818 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
820 struct flush_busy_ctx_data data = {
825 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
827 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
829 static inline unsigned int queued_to_index(unsigned int queued)
834 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
837 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
840 struct blk_mq_alloc_data data = {
842 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
843 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
853 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
854 data.flags |= BLK_MQ_REQ_RESERVED;
856 rq->tag = blk_mq_get_tag(&data);
858 if (blk_mq_tag_busy(data.hctx)) {
859 rq->rq_flags |= RQF_MQ_INFLIGHT;
860 atomic_inc(&data.hctx->nr_active);
862 data.hctx->tags->rqs[rq->tag] = rq;
869 static void blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
872 if (rq->tag == -1 || rq->internal_tag == -1)
875 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
878 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
879 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
880 atomic_dec(&hctx->nr_active);
885 * If we fail getting a driver tag because all the driver tags are already
886 * assigned and on the dispatch list, BUT the first entry does not have a
887 * tag, then we could deadlock. For that case, move entries with assigned
888 * driver tags to the front, leaving the set of tagged requests in the
889 * same order, and the untagged set in the same order.
891 static bool reorder_tags_to_front(struct list_head *list)
893 struct request *rq, *tmp, *first = NULL;
895 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
899 list_move(&rq->queuelist, list);
905 return first != NULL;
908 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
911 struct blk_mq_hw_ctx *hctx;
913 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
915 list_del(&wait->task_list);
916 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
917 blk_mq_run_hw_queue(hctx, true);
921 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
923 struct sbq_wait_state *ws;
926 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
927 * The thread which wins the race to grab this bit adds the hardware
928 * queue to the wait queue.
930 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
931 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
934 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
935 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
938 * As soon as this returns, it's no longer safe to fiddle with
939 * hctx->dispatch_wait, since a completion can wake up the wait queue
940 * and unlock the bit.
942 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
946 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list)
948 struct request_queue *q = hctx->queue;
950 LIST_HEAD(driver_list);
951 struct list_head *dptr;
952 int queued, ret = BLK_MQ_RQ_QUEUE_OK;
955 * Start off with dptr being NULL, so we start the first request
956 * immediately, even if we have more pending.
961 * Now process all the entries, sending them to the driver.
964 while (!list_empty(list)) {
965 struct blk_mq_queue_data bd;
967 rq = list_first_entry(list, struct request, queuelist);
968 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
969 if (!queued && reorder_tags_to_front(list))
973 * The initial allocation attempt failed, so we need to
974 * rerun the hardware queue when a tag is freed.
976 if (blk_mq_dispatch_wait_add(hctx)) {
978 * It's possible that a tag was freed in the
979 * window between the allocation failure and
980 * adding the hardware queue to the wait queue.
982 if (!blk_mq_get_driver_tag(rq, &hctx, false))
989 list_del_init(&rq->queuelist);
993 bd.last = list_empty(list);
995 ret = q->mq_ops->queue_rq(hctx, &bd);
997 case BLK_MQ_RQ_QUEUE_OK:
1000 case BLK_MQ_RQ_QUEUE_BUSY:
1001 blk_mq_put_driver_tag(hctx, rq);
1002 list_add(&rq->queuelist, list);
1003 __blk_mq_requeue_request(rq);
1006 pr_err("blk-mq: bad return on queue: %d\n", ret);
1007 case BLK_MQ_RQ_QUEUE_ERROR:
1009 blk_mq_end_request(rq, rq->errors);
1013 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
1017 * We've done the first request. If we have more than 1
1018 * left in the list, set dptr to defer issue.
1020 if (!dptr && list->next != list->prev)
1021 dptr = &driver_list;
1024 hctx->dispatched[queued_to_index(queued)]++;
1027 * Any items that need requeuing? Stuff them into hctx->dispatch,
1028 * that is where we will continue on next queue run.
1030 if (!list_empty(list)) {
1031 spin_lock(&hctx->lock);
1032 list_splice_init(list, &hctx->dispatch);
1033 spin_unlock(&hctx->lock);
1036 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1037 * it's possible the queue is stopped and restarted again
1038 * before this. Queue restart will dispatch requests. And since
1039 * requests in rq_list aren't added into hctx->dispatch yet,
1040 * the requests in rq_list might get lost.
1042 * blk_mq_run_hw_queue() already checks the STOPPED bit
1044 * If RESTART or TAG_WAITING is set, then let completion restart
1045 * the queue instead of potentially looping here.
1047 if (!blk_mq_sched_needs_restart(hctx) &&
1048 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1049 blk_mq_run_hw_queue(hctx, true);
1055 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1059 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1060 cpu_online(hctx->next_cpu));
1062 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1064 blk_mq_sched_dispatch_requests(hctx);
1067 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1068 blk_mq_sched_dispatch_requests(hctx);
1069 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1074 * It'd be great if the workqueue API had a way to pass
1075 * in a mask and had some smarts for more clever placement.
1076 * For now we just round-robin here, switching for every
1077 * BLK_MQ_CPU_WORK_BATCH queued items.
1079 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1081 if (hctx->queue->nr_hw_queues == 1)
1082 return WORK_CPU_UNBOUND;
1084 if (--hctx->next_cpu_batch <= 0) {
1087 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1088 if (next_cpu >= nr_cpu_ids)
1089 next_cpu = cpumask_first(hctx->cpumask);
1091 hctx->next_cpu = next_cpu;
1092 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1095 return hctx->next_cpu;
1098 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1100 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1101 !blk_mq_hw_queue_mapped(hctx)))
1104 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1105 int cpu = get_cpu();
1106 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1107 __blk_mq_run_hw_queue(hctx);
1115 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
1118 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1120 struct blk_mq_hw_ctx *hctx;
1123 queue_for_each_hw_ctx(q, hctx, i) {
1124 if (!blk_mq_hctx_has_pending(hctx) ||
1125 blk_mq_hctx_stopped(hctx))
1128 blk_mq_run_hw_queue(hctx, async);
1131 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1134 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1135 * @q: request queue.
1137 * The caller is responsible for serializing this function against
1138 * blk_mq_{start,stop}_hw_queue().
1140 bool blk_mq_queue_stopped(struct request_queue *q)
1142 struct blk_mq_hw_ctx *hctx;
1145 queue_for_each_hw_ctx(q, hctx, i)
1146 if (blk_mq_hctx_stopped(hctx))
1151 EXPORT_SYMBOL(blk_mq_queue_stopped);
1153 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1155 cancel_work(&hctx->run_work);
1156 cancel_delayed_work(&hctx->delay_work);
1157 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1159 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1161 void blk_mq_stop_hw_queues(struct request_queue *q)
1163 struct blk_mq_hw_ctx *hctx;
1166 queue_for_each_hw_ctx(q, hctx, i)
1167 blk_mq_stop_hw_queue(hctx);
1169 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1171 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1173 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1175 blk_mq_run_hw_queue(hctx, false);
1177 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1179 void blk_mq_start_hw_queues(struct request_queue *q)
1181 struct blk_mq_hw_ctx *hctx;
1184 queue_for_each_hw_ctx(q, hctx, i)
1185 blk_mq_start_hw_queue(hctx);
1187 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1189 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1191 if (!blk_mq_hctx_stopped(hctx))
1194 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1195 blk_mq_run_hw_queue(hctx, async);
1197 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1199 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1201 struct blk_mq_hw_ctx *hctx;
1204 queue_for_each_hw_ctx(q, hctx, i)
1205 blk_mq_start_stopped_hw_queue(hctx, async);
1207 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1209 static void blk_mq_run_work_fn(struct work_struct *work)
1211 struct blk_mq_hw_ctx *hctx;
1213 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1215 __blk_mq_run_hw_queue(hctx);
1218 static void blk_mq_delay_work_fn(struct work_struct *work)
1220 struct blk_mq_hw_ctx *hctx;
1222 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1224 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1225 __blk_mq_run_hw_queue(hctx);
1228 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1230 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1233 blk_mq_stop_hw_queue(hctx);
1234 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1235 &hctx->delay_work, msecs_to_jiffies(msecs));
1237 EXPORT_SYMBOL(blk_mq_delay_queue);
1239 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1243 struct blk_mq_ctx *ctx = rq->mq_ctx;
1245 trace_block_rq_insert(hctx->queue, rq);
1248 list_add(&rq->queuelist, &ctx->rq_list);
1250 list_add_tail(&rq->queuelist, &ctx->rq_list);
1253 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1256 struct blk_mq_ctx *ctx = rq->mq_ctx;
1258 __blk_mq_insert_req_list(hctx, rq, at_head);
1259 blk_mq_hctx_mark_pending(hctx, ctx);
1262 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1263 struct list_head *list)
1267 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1270 spin_lock(&ctx->lock);
1271 while (!list_empty(list)) {
1274 rq = list_first_entry(list, struct request, queuelist);
1275 BUG_ON(rq->mq_ctx != ctx);
1276 list_del_init(&rq->queuelist);
1277 __blk_mq_insert_req_list(hctx, rq, false);
1279 blk_mq_hctx_mark_pending(hctx, ctx);
1280 spin_unlock(&ctx->lock);
1283 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1285 struct request *rqa = container_of(a, struct request, queuelist);
1286 struct request *rqb = container_of(b, struct request, queuelist);
1288 return !(rqa->mq_ctx < rqb->mq_ctx ||
1289 (rqa->mq_ctx == rqb->mq_ctx &&
1290 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1293 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1295 struct blk_mq_ctx *this_ctx;
1296 struct request_queue *this_q;
1299 LIST_HEAD(ctx_list);
1302 list_splice_init(&plug->mq_list, &list);
1304 list_sort(NULL, &list, plug_ctx_cmp);
1310 while (!list_empty(&list)) {
1311 rq = list_entry_rq(list.next);
1312 list_del_init(&rq->queuelist);
1314 if (rq->mq_ctx != this_ctx) {
1316 trace_block_unplug(this_q, depth, from_schedule);
1317 blk_mq_sched_insert_requests(this_q, this_ctx,
1322 this_ctx = rq->mq_ctx;
1328 list_add_tail(&rq->queuelist, &ctx_list);
1332 * If 'this_ctx' is set, we know we have entries to complete
1333 * on 'ctx_list'. Do those.
1336 trace_block_unplug(this_q, depth, from_schedule);
1337 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1342 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1344 init_request_from_bio(rq, bio);
1346 blk_account_io_start(rq, true);
1349 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1351 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1352 !blk_queue_nomerges(hctx->queue);
1355 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1356 struct blk_mq_ctx *ctx,
1357 struct request *rq, struct bio *bio)
1359 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1360 blk_mq_bio_to_request(rq, bio);
1361 spin_lock(&ctx->lock);
1363 __blk_mq_insert_request(hctx, rq, false);
1364 spin_unlock(&ctx->lock);
1367 struct request_queue *q = hctx->queue;
1369 spin_lock(&ctx->lock);
1370 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1371 blk_mq_bio_to_request(rq, bio);
1375 spin_unlock(&ctx->lock);
1376 __blk_mq_finish_request(hctx, ctx, rq);
1381 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1384 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1386 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1389 static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie)
1391 struct request_queue *q = rq->q;
1392 struct blk_mq_queue_data bd = {
1397 struct blk_mq_hw_ctx *hctx;
1398 blk_qc_t new_cookie;
1404 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1407 new_cookie = request_to_qc_t(hctx, rq);
1410 * For OK queue, we are done. For error, kill it. Any other
1411 * error (busy), just add it to our list as we previously
1414 ret = q->mq_ops->queue_rq(hctx, &bd);
1415 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1416 *cookie = new_cookie;
1420 __blk_mq_requeue_request(rq);
1422 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1423 *cookie = BLK_QC_T_NONE;
1425 blk_mq_end_request(rq, rq->errors);
1430 blk_mq_sched_insert_request(rq, false, true, true, false);
1434 * Multiple hardware queue variant. This will not use per-process plugs,
1435 * but will attempt to bypass the hctx queueing if we can go straight to
1436 * hardware for SYNC IO.
1438 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1440 const int is_sync = op_is_sync(bio->bi_opf);
1441 const int is_flush_fua = op_is_flush(bio->bi_opf);
1442 struct blk_mq_alloc_data data = { .flags = 0 };
1444 unsigned int request_count = 0, srcu_idx;
1445 struct blk_plug *plug;
1446 struct request *same_queue_rq = NULL;
1448 unsigned int wb_acct;
1450 blk_queue_bounce(q, &bio);
1452 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1454 return BLK_QC_T_NONE;
1457 blk_queue_split(q, &bio, q->bio_split);
1459 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1460 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1461 return BLK_QC_T_NONE;
1463 if (blk_mq_sched_bio_merge(q, bio))
1464 return BLK_QC_T_NONE;
1466 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1468 trace_block_getrq(q, bio, bio->bi_opf);
1470 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1471 if (unlikely(!rq)) {
1472 __wbt_done(q->rq_wb, wb_acct);
1473 return BLK_QC_T_NONE;
1476 wbt_track(&rq->issue_stat, wb_acct);
1478 cookie = request_to_qc_t(data.hctx, rq);
1480 if (unlikely(is_flush_fua)) {
1483 blk_mq_bio_to_request(rq, bio);
1484 blk_insert_flush(rq);
1488 plug = current->plug;
1490 * If the driver supports defer issued based on 'last', then
1491 * queue it up like normal since we can potentially save some
1494 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1495 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1496 struct request *old_rq = NULL;
1498 blk_mq_bio_to_request(rq, bio);
1501 * We do limited plugging. If the bio can be merged, do that.
1502 * Otherwise the existing request in the plug list will be
1503 * issued. So the plug list will have one request at most
1507 * The plug list might get flushed before this. If that
1508 * happens, same_queue_rq is invalid and plug list is
1511 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1512 old_rq = same_queue_rq;
1513 list_del_init(&old_rq->queuelist);
1515 list_add_tail(&rq->queuelist, &plug->mq_list);
1516 } else /* is_sync */
1518 blk_mq_put_ctx(data.ctx);
1522 if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) {
1524 blk_mq_try_issue_directly(old_rq, &cookie);
1527 srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu);
1528 blk_mq_try_issue_directly(old_rq, &cookie);
1529 srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx);
1536 blk_mq_put_ctx(data.ctx);
1537 blk_mq_bio_to_request(rq, bio);
1538 blk_mq_sched_insert_request(rq, false, true,
1539 !is_sync || is_flush_fua, true);
1542 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1544 * For a SYNC request, send it to the hardware immediately. For
1545 * an ASYNC request, just ensure that we run it later on. The
1546 * latter allows for merging opportunities and more efficient
1550 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1552 blk_mq_put_ctx(data.ctx);
1558 * Single hardware queue variant. This will attempt to use any per-process
1559 * plug for merging and IO deferral.
1561 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1563 const int is_sync = op_is_sync(bio->bi_opf);
1564 const int is_flush_fua = op_is_flush(bio->bi_opf);
1565 struct blk_plug *plug;
1566 unsigned int request_count = 0;
1567 struct blk_mq_alloc_data data = { .flags = 0 };
1570 unsigned int wb_acct;
1572 blk_queue_bounce(q, &bio);
1574 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1576 return BLK_QC_T_NONE;
1579 blk_queue_split(q, &bio, q->bio_split);
1581 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1582 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1583 return BLK_QC_T_NONE;
1585 request_count = blk_plug_queued_count(q);
1587 if (blk_mq_sched_bio_merge(q, bio))
1588 return BLK_QC_T_NONE;
1590 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1592 trace_block_getrq(q, bio, bio->bi_opf);
1594 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1595 if (unlikely(!rq)) {
1596 __wbt_done(q->rq_wb, wb_acct);
1597 return BLK_QC_T_NONE;
1600 wbt_track(&rq->issue_stat, wb_acct);
1602 cookie = request_to_qc_t(data.hctx, rq);
1604 if (unlikely(is_flush_fua)) {
1607 blk_mq_bio_to_request(rq, bio);
1608 blk_insert_flush(rq);
1613 * A task plug currently exists. Since this is completely lockless,
1614 * utilize that to temporarily store requests until the task is
1615 * either done or scheduled away.
1617 plug = current->plug;
1619 struct request *last = NULL;
1621 blk_mq_bio_to_request(rq, bio);
1624 * @request_count may become stale because of schedule
1625 * out, so check the list again.
1627 if (list_empty(&plug->mq_list))
1630 trace_block_plug(q);
1632 last = list_entry_rq(plug->mq_list.prev);
1634 blk_mq_put_ctx(data.ctx);
1636 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1637 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1638 blk_flush_plug_list(plug, false);
1639 trace_block_plug(q);
1642 list_add_tail(&rq->queuelist, &plug->mq_list);
1648 blk_mq_put_ctx(data.ctx);
1649 blk_mq_bio_to_request(rq, bio);
1650 blk_mq_sched_insert_request(rq, false, true,
1651 !is_sync || is_flush_fua, true);
1654 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1656 * For a SYNC request, send it to the hardware immediately. For
1657 * an ASYNC request, just ensure that we run it later on. The
1658 * latter allows for merging opportunities and more efficient
1662 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1665 blk_mq_put_ctx(data.ctx);
1670 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1671 unsigned int hctx_idx)
1675 if (tags->rqs && set->ops->exit_request) {
1678 for (i = 0; i < tags->nr_tags; i++) {
1679 struct request *rq = tags->static_rqs[i];
1683 set->ops->exit_request(set->driver_data, rq,
1685 tags->static_rqs[i] = NULL;
1689 while (!list_empty(&tags->page_list)) {
1690 page = list_first_entry(&tags->page_list, struct page, lru);
1691 list_del_init(&page->lru);
1693 * Remove kmemleak object previously allocated in
1694 * blk_mq_init_rq_map().
1696 kmemleak_free(page_address(page));
1697 __free_pages(page, page->private);
1701 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1705 kfree(tags->static_rqs);
1706 tags->static_rqs = NULL;
1708 blk_mq_free_tags(tags);
1711 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1712 unsigned int hctx_idx,
1713 unsigned int nr_tags,
1714 unsigned int reserved_tags)
1716 struct blk_mq_tags *tags;
1719 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1720 if (node == NUMA_NO_NODE)
1721 node = set->numa_node;
1723 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1724 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1728 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1729 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1732 blk_mq_free_tags(tags);
1736 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1737 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1739 if (!tags->static_rqs) {
1741 blk_mq_free_tags(tags);
1748 static size_t order_to_size(unsigned int order)
1750 return (size_t)PAGE_SIZE << order;
1753 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1754 unsigned int hctx_idx, unsigned int depth)
1756 unsigned int i, j, entries_per_page, max_order = 4;
1757 size_t rq_size, left;
1760 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1761 if (node == NUMA_NO_NODE)
1762 node = set->numa_node;
1764 INIT_LIST_HEAD(&tags->page_list);
1767 * rq_size is the size of the request plus driver payload, rounded
1768 * to the cacheline size
1770 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1772 left = rq_size * depth;
1774 for (i = 0; i < depth; ) {
1775 int this_order = max_order;
1780 while (this_order && left < order_to_size(this_order - 1))
1784 page = alloc_pages_node(node,
1785 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1791 if (order_to_size(this_order) < rq_size)
1798 page->private = this_order;
1799 list_add_tail(&page->lru, &tags->page_list);
1801 p = page_address(page);
1803 * Allow kmemleak to scan these pages as they contain pointers
1804 * to additional allocations like via ops->init_request().
1806 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1807 entries_per_page = order_to_size(this_order) / rq_size;
1808 to_do = min(entries_per_page, depth - i);
1809 left -= to_do * rq_size;
1810 for (j = 0; j < to_do; j++) {
1811 struct request *rq = p;
1813 tags->static_rqs[i] = rq;
1814 if (set->ops->init_request) {
1815 if (set->ops->init_request(set->driver_data,
1818 tags->static_rqs[i] = NULL;
1830 blk_mq_free_rqs(set, tags, hctx_idx);
1835 * 'cpu' is going away. splice any existing rq_list entries from this
1836 * software queue to the hw queue dispatch list, and ensure that it
1839 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1841 struct blk_mq_hw_ctx *hctx;
1842 struct blk_mq_ctx *ctx;
1845 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1846 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1848 spin_lock(&ctx->lock);
1849 if (!list_empty(&ctx->rq_list)) {
1850 list_splice_init(&ctx->rq_list, &tmp);
1851 blk_mq_hctx_clear_pending(hctx, ctx);
1853 spin_unlock(&ctx->lock);
1855 if (list_empty(&tmp))
1858 spin_lock(&hctx->lock);
1859 list_splice_tail_init(&tmp, &hctx->dispatch);
1860 spin_unlock(&hctx->lock);
1862 blk_mq_run_hw_queue(hctx, true);
1866 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1868 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1872 /* hctx->ctxs will be freed in queue's release handler */
1873 static void blk_mq_exit_hctx(struct request_queue *q,
1874 struct blk_mq_tag_set *set,
1875 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1877 unsigned flush_start_tag = set->queue_depth;
1879 blk_mq_tag_idle(hctx);
1881 if (set->ops->exit_request)
1882 set->ops->exit_request(set->driver_data,
1883 hctx->fq->flush_rq, hctx_idx,
1884 flush_start_tag + hctx_idx);
1886 if (set->ops->exit_hctx)
1887 set->ops->exit_hctx(hctx, hctx_idx);
1889 if (hctx->flags & BLK_MQ_F_BLOCKING)
1890 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1892 blk_mq_remove_cpuhp(hctx);
1893 blk_free_flush_queue(hctx->fq);
1894 sbitmap_free(&hctx->ctx_map);
1897 static void blk_mq_exit_hw_queues(struct request_queue *q,
1898 struct blk_mq_tag_set *set, int nr_queue)
1900 struct blk_mq_hw_ctx *hctx;
1903 queue_for_each_hw_ctx(q, hctx, i) {
1906 blk_mq_exit_hctx(q, set, hctx, i);
1910 static void blk_mq_free_hw_queues(struct request_queue *q,
1911 struct blk_mq_tag_set *set)
1913 struct blk_mq_hw_ctx *hctx;
1916 queue_for_each_hw_ctx(q, hctx, i)
1917 free_cpumask_var(hctx->cpumask);
1920 static int blk_mq_init_hctx(struct request_queue *q,
1921 struct blk_mq_tag_set *set,
1922 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1925 unsigned flush_start_tag = set->queue_depth;
1927 node = hctx->numa_node;
1928 if (node == NUMA_NO_NODE)
1929 node = hctx->numa_node = set->numa_node;
1931 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1932 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1933 spin_lock_init(&hctx->lock);
1934 INIT_LIST_HEAD(&hctx->dispatch);
1936 hctx->queue_num = hctx_idx;
1937 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1939 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1941 hctx->tags = set->tags[hctx_idx];
1944 * Allocate space for all possible cpus to avoid allocation at
1947 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1950 goto unregister_cpu_notifier;
1952 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1958 if (set->ops->init_hctx &&
1959 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1962 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1966 if (set->ops->init_request &&
1967 set->ops->init_request(set->driver_data,
1968 hctx->fq->flush_rq, hctx_idx,
1969 flush_start_tag + hctx_idx, node))
1972 if (hctx->flags & BLK_MQ_F_BLOCKING)
1973 init_srcu_struct(&hctx->queue_rq_srcu);
1980 if (set->ops->exit_hctx)
1981 set->ops->exit_hctx(hctx, hctx_idx);
1983 sbitmap_free(&hctx->ctx_map);
1986 unregister_cpu_notifier:
1987 blk_mq_remove_cpuhp(hctx);
1991 static void blk_mq_init_cpu_queues(struct request_queue *q,
1992 unsigned int nr_hw_queues)
1996 for_each_possible_cpu(i) {
1997 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1998 struct blk_mq_hw_ctx *hctx;
2000 memset(__ctx, 0, sizeof(*__ctx));
2002 spin_lock_init(&__ctx->lock);
2003 INIT_LIST_HEAD(&__ctx->rq_list);
2005 blk_stat_init(&__ctx->stat[BLK_STAT_READ]);
2006 blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]);
2008 /* If the cpu isn't online, the cpu is mapped to first hctx */
2012 hctx = blk_mq_map_queue(q, i);
2015 * Set local node, IFF we have more than one hw queue. If
2016 * not, we remain on the home node of the device
2018 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2019 hctx->numa_node = local_memory_node(cpu_to_node(i));
2023 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2027 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2028 set->queue_depth, set->reserved_tags);
2029 if (!set->tags[hctx_idx])
2032 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2037 blk_mq_free_rq_map(set->tags[hctx_idx]);
2038 set->tags[hctx_idx] = NULL;
2042 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2043 unsigned int hctx_idx)
2045 if (set->tags[hctx_idx]) {
2046 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2047 blk_mq_free_rq_map(set->tags[hctx_idx]);
2048 set->tags[hctx_idx] = NULL;
2052 static void blk_mq_map_swqueue(struct request_queue *q,
2053 const struct cpumask *online_mask)
2055 unsigned int i, hctx_idx;
2056 struct blk_mq_hw_ctx *hctx;
2057 struct blk_mq_ctx *ctx;
2058 struct blk_mq_tag_set *set = q->tag_set;
2061 * Avoid others reading imcomplete hctx->cpumask through sysfs
2063 mutex_lock(&q->sysfs_lock);
2065 queue_for_each_hw_ctx(q, hctx, i) {
2066 cpumask_clear(hctx->cpumask);
2071 * Map software to hardware queues
2073 for_each_possible_cpu(i) {
2074 /* If the cpu isn't online, the cpu is mapped to first hctx */
2075 if (!cpumask_test_cpu(i, online_mask))
2078 hctx_idx = q->mq_map[i];
2079 /* unmapped hw queue can be remapped after CPU topo changed */
2080 if (!set->tags[hctx_idx] &&
2081 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2083 * If tags initialization fail for some hctx,
2084 * that hctx won't be brought online. In this
2085 * case, remap the current ctx to hctx[0] which
2086 * is guaranteed to always have tags allocated
2091 ctx = per_cpu_ptr(q->queue_ctx, i);
2092 hctx = blk_mq_map_queue(q, i);
2094 cpumask_set_cpu(i, hctx->cpumask);
2095 ctx->index_hw = hctx->nr_ctx;
2096 hctx->ctxs[hctx->nr_ctx++] = ctx;
2099 mutex_unlock(&q->sysfs_lock);
2101 queue_for_each_hw_ctx(q, hctx, i) {
2103 * If no software queues are mapped to this hardware queue,
2104 * disable it and free the request entries.
2106 if (!hctx->nr_ctx) {
2107 /* Never unmap queue 0. We need it as a
2108 * fallback in case of a new remap fails
2111 if (i && set->tags[i])
2112 blk_mq_free_map_and_requests(set, i);
2118 hctx->tags = set->tags[i];
2119 WARN_ON(!hctx->tags);
2122 * Set the map size to the number of mapped software queues.
2123 * This is more accurate and more efficient than looping
2124 * over all possibly mapped software queues.
2126 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2129 * Initialize batch roundrobin counts
2131 hctx->next_cpu = cpumask_first(hctx->cpumask);
2132 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2136 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2138 struct blk_mq_hw_ctx *hctx;
2141 queue_for_each_hw_ctx(q, hctx, i) {
2143 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2145 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2149 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2151 struct request_queue *q;
2153 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2154 blk_mq_freeze_queue(q);
2155 queue_set_hctx_shared(q, shared);
2156 blk_mq_unfreeze_queue(q);
2160 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2162 struct blk_mq_tag_set *set = q->tag_set;
2164 mutex_lock(&set->tag_list_lock);
2165 list_del_init(&q->tag_set_list);
2166 if (list_is_singular(&set->tag_list)) {
2167 /* just transitioned to unshared */
2168 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2169 /* update existing queue */
2170 blk_mq_update_tag_set_depth(set, false);
2172 mutex_unlock(&set->tag_list_lock);
2175 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2176 struct request_queue *q)
2180 mutex_lock(&set->tag_list_lock);
2182 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2183 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2184 set->flags |= BLK_MQ_F_TAG_SHARED;
2185 /* update existing queue */
2186 blk_mq_update_tag_set_depth(set, true);
2188 if (set->flags & BLK_MQ_F_TAG_SHARED)
2189 queue_set_hctx_shared(q, true);
2190 list_add_tail(&q->tag_set_list, &set->tag_list);
2192 mutex_unlock(&set->tag_list_lock);
2196 * It is the actual release handler for mq, but we do it from
2197 * request queue's release handler for avoiding use-after-free
2198 * and headache because q->mq_kobj shouldn't have been introduced,
2199 * but we can't group ctx/kctx kobj without it.
2201 void blk_mq_release(struct request_queue *q)
2203 struct blk_mq_hw_ctx *hctx;
2206 blk_mq_sched_teardown(q);
2208 /* hctx kobj stays in hctx */
2209 queue_for_each_hw_ctx(q, hctx, i) {
2218 kfree(q->queue_hw_ctx);
2220 /* ctx kobj stays in queue_ctx */
2221 free_percpu(q->queue_ctx);
2224 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2226 struct request_queue *uninit_q, *q;
2228 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2230 return ERR_PTR(-ENOMEM);
2232 q = blk_mq_init_allocated_queue(set, uninit_q);
2234 blk_cleanup_queue(uninit_q);
2238 EXPORT_SYMBOL(blk_mq_init_queue);
2240 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2241 struct request_queue *q)
2244 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2246 blk_mq_sysfs_unregister(q);
2247 for (i = 0; i < set->nr_hw_queues; i++) {
2253 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2254 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2259 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2266 atomic_set(&hctxs[i]->nr_active, 0);
2267 hctxs[i]->numa_node = node;
2268 hctxs[i]->queue_num = i;
2270 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2271 free_cpumask_var(hctxs[i]->cpumask);
2276 blk_mq_hctx_kobj_init(hctxs[i]);
2278 for (j = i; j < q->nr_hw_queues; j++) {
2279 struct blk_mq_hw_ctx *hctx = hctxs[j];
2283 blk_mq_free_map_and_requests(set, j);
2284 blk_mq_exit_hctx(q, set, hctx, j);
2285 free_cpumask_var(hctx->cpumask);
2286 kobject_put(&hctx->kobj);
2293 q->nr_hw_queues = i;
2294 blk_mq_sysfs_register(q);
2297 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2298 struct request_queue *q)
2300 /* mark the queue as mq asap */
2301 q->mq_ops = set->ops;
2303 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2307 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2308 GFP_KERNEL, set->numa_node);
2309 if (!q->queue_hw_ctx)
2312 q->mq_map = set->mq_map;
2314 blk_mq_realloc_hw_ctxs(set, q);
2315 if (!q->nr_hw_queues)
2318 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2319 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2321 q->nr_queues = nr_cpu_ids;
2323 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2325 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2326 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2328 q->sg_reserved_size = INT_MAX;
2330 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2331 INIT_LIST_HEAD(&q->requeue_list);
2332 spin_lock_init(&q->requeue_lock);
2334 if (q->nr_hw_queues > 1)
2335 blk_queue_make_request(q, blk_mq_make_request);
2337 blk_queue_make_request(q, blk_sq_make_request);
2340 * Do this after blk_queue_make_request() overrides it...
2342 q->nr_requests = set->queue_depth;
2345 * Default to classic polling
2349 if (set->ops->complete)
2350 blk_queue_softirq_done(q, set->ops->complete);
2352 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2355 mutex_lock(&all_q_mutex);
2357 list_add_tail(&q->all_q_node, &all_q_list);
2358 blk_mq_add_queue_tag_set(set, q);
2359 blk_mq_map_swqueue(q, cpu_online_mask);
2361 mutex_unlock(&all_q_mutex);
2364 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2367 ret = blk_mq_sched_init(q);
2369 return ERR_PTR(ret);
2375 kfree(q->queue_hw_ctx);
2377 free_percpu(q->queue_ctx);
2380 return ERR_PTR(-ENOMEM);
2382 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2384 void blk_mq_free_queue(struct request_queue *q)
2386 struct blk_mq_tag_set *set = q->tag_set;
2388 mutex_lock(&all_q_mutex);
2389 list_del_init(&q->all_q_node);
2390 mutex_unlock(&all_q_mutex);
2394 blk_mq_del_queue_tag_set(q);
2396 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2397 blk_mq_free_hw_queues(q, set);
2400 /* Basically redo blk_mq_init_queue with queue frozen */
2401 static void blk_mq_queue_reinit(struct request_queue *q,
2402 const struct cpumask *online_mask)
2404 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2406 blk_mq_sysfs_unregister(q);
2409 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2410 * we should change hctx numa_node according to new topology (this
2411 * involves free and re-allocate memory, worthy doing?)
2414 blk_mq_map_swqueue(q, online_mask);
2416 blk_mq_sysfs_register(q);
2420 * New online cpumask which is going to be set in this hotplug event.
2421 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2422 * one-by-one and dynamically allocating this could result in a failure.
2424 static struct cpumask cpuhp_online_new;
2426 static void blk_mq_queue_reinit_work(void)
2428 struct request_queue *q;
2430 mutex_lock(&all_q_mutex);
2432 * We need to freeze and reinit all existing queues. Freezing
2433 * involves synchronous wait for an RCU grace period and doing it
2434 * one by one may take a long time. Start freezing all queues in
2435 * one swoop and then wait for the completions so that freezing can
2436 * take place in parallel.
2438 list_for_each_entry(q, &all_q_list, all_q_node)
2439 blk_mq_freeze_queue_start(q);
2440 list_for_each_entry(q, &all_q_list, all_q_node)
2441 blk_mq_freeze_queue_wait(q);
2443 list_for_each_entry(q, &all_q_list, all_q_node)
2444 blk_mq_queue_reinit(q, &cpuhp_online_new);
2446 list_for_each_entry(q, &all_q_list, all_q_node)
2447 blk_mq_unfreeze_queue(q);
2449 mutex_unlock(&all_q_mutex);
2452 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2454 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2455 blk_mq_queue_reinit_work();
2460 * Before hotadded cpu starts handling requests, new mappings must be
2461 * established. Otherwise, these requests in hw queue might never be
2464 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2465 * for CPU0, and ctx1 for CPU1).
2467 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2468 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2470 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2471 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2472 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2475 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2477 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2478 cpumask_set_cpu(cpu, &cpuhp_online_new);
2479 blk_mq_queue_reinit_work();
2483 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2487 for (i = 0; i < set->nr_hw_queues; i++)
2488 if (!__blk_mq_alloc_rq_map(set, i))
2495 blk_mq_free_rq_map(set->tags[i]);
2501 * Allocate the request maps associated with this tag_set. Note that this
2502 * may reduce the depth asked for, if memory is tight. set->queue_depth
2503 * will be updated to reflect the allocated depth.
2505 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2510 depth = set->queue_depth;
2512 err = __blk_mq_alloc_rq_maps(set);
2516 set->queue_depth >>= 1;
2517 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2521 } while (set->queue_depth);
2523 if (!set->queue_depth || err) {
2524 pr_err("blk-mq: failed to allocate request map\n");
2528 if (depth != set->queue_depth)
2529 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2530 depth, set->queue_depth);
2536 * Alloc a tag set to be associated with one or more request queues.
2537 * May fail with EINVAL for various error conditions. May adjust the
2538 * requested depth down, if if it too large. In that case, the set
2539 * value will be stored in set->queue_depth.
2541 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2545 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2547 if (!set->nr_hw_queues)
2549 if (!set->queue_depth)
2551 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2554 if (!set->ops->queue_rq)
2557 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2558 pr_info("blk-mq: reduced tag depth to %u\n",
2560 set->queue_depth = BLK_MQ_MAX_DEPTH;
2564 * If a crashdump is active, then we are potentially in a very
2565 * memory constrained environment. Limit us to 1 queue and
2566 * 64 tags to prevent using too much memory.
2568 if (is_kdump_kernel()) {
2569 set->nr_hw_queues = 1;
2570 set->queue_depth = min(64U, set->queue_depth);
2573 * There is no use for more h/w queues than cpus.
2575 if (set->nr_hw_queues > nr_cpu_ids)
2576 set->nr_hw_queues = nr_cpu_ids;
2578 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2579 GFP_KERNEL, set->numa_node);
2584 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2585 GFP_KERNEL, set->numa_node);
2589 if (set->ops->map_queues)
2590 ret = set->ops->map_queues(set);
2592 ret = blk_mq_map_queues(set);
2594 goto out_free_mq_map;
2596 ret = blk_mq_alloc_rq_maps(set);
2598 goto out_free_mq_map;
2600 mutex_init(&set->tag_list_lock);
2601 INIT_LIST_HEAD(&set->tag_list);
2613 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2615 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2619 for (i = 0; i < nr_cpu_ids; i++)
2620 blk_mq_free_map_and_requests(set, i);
2628 EXPORT_SYMBOL(blk_mq_free_tag_set);
2630 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2632 struct blk_mq_tag_set *set = q->tag_set;
2633 struct blk_mq_hw_ctx *hctx;
2639 blk_mq_freeze_queue(q);
2640 blk_mq_quiesce_queue(q);
2643 queue_for_each_hw_ctx(q, hctx, i) {
2647 * If we're using an MQ scheduler, just update the scheduler
2648 * queue depth. This is similar to what the old code would do.
2650 if (!hctx->sched_tags) {
2651 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2652 min(nr, set->queue_depth),
2655 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2663 q->nr_requests = nr;
2665 blk_mq_unfreeze_queue(q);
2666 blk_mq_start_stopped_hw_queues(q, true);
2671 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2673 struct request_queue *q;
2675 if (nr_hw_queues > nr_cpu_ids)
2676 nr_hw_queues = nr_cpu_ids;
2677 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2680 list_for_each_entry(q, &set->tag_list, tag_set_list)
2681 blk_mq_freeze_queue(q);
2683 set->nr_hw_queues = nr_hw_queues;
2684 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2685 blk_mq_realloc_hw_ctxs(set, q);
2688 * Manually set the make_request_fn as blk_queue_make_request
2689 * resets a lot of the queue settings.
2691 if (q->nr_hw_queues > 1)
2692 q->make_request_fn = blk_mq_make_request;
2694 q->make_request_fn = blk_sq_make_request;
2696 blk_mq_queue_reinit(q, cpu_online_mask);
2699 list_for_each_entry(q, &set->tag_list, tag_set_list)
2700 blk_mq_unfreeze_queue(q);
2702 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2704 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2705 struct blk_mq_hw_ctx *hctx,
2708 struct blk_rq_stat stat[2];
2709 unsigned long ret = 0;
2712 * If stats collection isn't on, don't sleep but turn it on for
2715 if (!blk_stat_enable(q))
2719 * We don't have to do this once per IO, should optimize this
2720 * to just use the current window of stats until it changes
2722 memset(&stat, 0, sizeof(stat));
2723 blk_hctx_stat_get(hctx, stat);
2726 * As an optimistic guess, use half of the mean service time
2727 * for this type of request. We can (and should) make this smarter.
2728 * For instance, if the completion latencies are tight, we can
2729 * get closer than just half the mean. This is especially
2730 * important on devices where the completion latencies are longer
2733 if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples)
2734 ret = (stat[BLK_STAT_READ].mean + 1) / 2;
2735 else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples)
2736 ret = (stat[BLK_STAT_WRITE].mean + 1) / 2;
2741 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2742 struct blk_mq_hw_ctx *hctx,
2745 struct hrtimer_sleeper hs;
2746 enum hrtimer_mode mode;
2750 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2756 * -1: don't ever hybrid sleep
2757 * 0: use half of prev avg
2758 * >0: use this specific value
2760 if (q->poll_nsec == -1)
2762 else if (q->poll_nsec > 0)
2763 nsecs = q->poll_nsec;
2765 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2770 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2773 * This will be replaced with the stats tracking code, using
2774 * 'avg_completion_time / 2' as the pre-sleep target.
2778 mode = HRTIMER_MODE_REL;
2779 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2780 hrtimer_set_expires(&hs.timer, kt);
2782 hrtimer_init_sleeper(&hs, current);
2784 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2786 set_current_state(TASK_UNINTERRUPTIBLE);
2787 hrtimer_start_expires(&hs.timer, mode);
2790 hrtimer_cancel(&hs.timer);
2791 mode = HRTIMER_MODE_ABS;
2792 } while (hs.task && !signal_pending(current));
2794 __set_current_state(TASK_RUNNING);
2795 destroy_hrtimer_on_stack(&hs.timer);
2799 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2801 struct request_queue *q = hctx->queue;
2805 * If we sleep, have the caller restart the poll loop to reset
2806 * the state. Like for the other success return cases, the
2807 * caller is responsible for checking if the IO completed. If
2808 * the IO isn't complete, we'll get called again and will go
2809 * straight to the busy poll loop.
2811 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2814 hctx->poll_considered++;
2816 state = current->state;
2817 while (!need_resched()) {
2820 hctx->poll_invoked++;
2822 ret = q->mq_ops->poll(hctx, rq->tag);
2824 hctx->poll_success++;
2825 set_current_state(TASK_RUNNING);
2829 if (signal_pending_state(state, current))
2830 set_current_state(TASK_RUNNING);
2832 if (current->state == TASK_RUNNING)
2842 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2844 struct blk_mq_hw_ctx *hctx;
2845 struct blk_plug *plug;
2848 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2849 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2852 plug = current->plug;
2854 blk_flush_plug_list(plug, false);
2856 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2857 if (!blk_qc_t_is_internal(cookie))
2858 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2860 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2862 return __blk_mq_poll(hctx, rq);
2864 EXPORT_SYMBOL_GPL(blk_mq_poll);
2866 void blk_mq_disable_hotplug(void)
2868 mutex_lock(&all_q_mutex);
2871 void blk_mq_enable_hotplug(void)
2873 mutex_unlock(&all_q_mutex);
2876 static int __init blk_mq_init(void)
2878 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2879 blk_mq_hctx_notify_dead);
2881 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2882 blk_mq_queue_reinit_prepare,
2883 blk_mq_queue_reinit_dead);
2886 subsys_initcall(blk_mq_init);