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;
239 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
245 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
247 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
250 struct blk_mq_alloc_data alloc_data = { .flags = flags };
254 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
258 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
260 blk_mq_put_ctx(alloc_data.ctx);
264 return ERR_PTR(-EWOULDBLOCK);
267 rq->__sector = (sector_t) -1;
268 rq->bio = rq->biotail = NULL;
271 EXPORT_SYMBOL(blk_mq_alloc_request);
273 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
274 unsigned int flags, unsigned int hctx_idx)
276 struct blk_mq_hw_ctx *hctx;
277 struct blk_mq_ctx *ctx;
279 struct blk_mq_alloc_data alloc_data;
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 hctx = q->queue_hw_ctx[hctx_idx];
303 if (!blk_mq_hw_queue_mapped(hctx)) {
307 ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
309 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
310 rq = __blk_mq_alloc_request(&alloc_data, rw);
322 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
324 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
327 const int sched_tag = rq->internal_tag;
328 struct request_queue *q = rq->q;
330 if (rq->rq_flags & RQF_MQ_INFLIGHT)
331 atomic_dec(&hctx->nr_active);
333 wbt_done(q->rq_wb, &rq->issue_stat);
336 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
337 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
339 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
341 blk_mq_sched_completed_request(hctx, rq);
342 blk_mq_sched_restart_queues(hctx);
346 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
349 struct blk_mq_ctx *ctx = rq->mq_ctx;
351 ctx->rq_completed[rq_is_sync(rq)]++;
352 __blk_mq_finish_request(hctx, ctx, rq);
355 void blk_mq_finish_request(struct request *rq)
357 blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
360 void blk_mq_free_request(struct request *rq)
362 blk_mq_sched_put_request(rq);
364 EXPORT_SYMBOL_GPL(blk_mq_free_request);
366 inline void __blk_mq_end_request(struct request *rq, int error)
368 blk_account_io_done(rq);
371 wbt_done(rq->q->rq_wb, &rq->issue_stat);
372 rq->end_io(rq, error);
374 if (unlikely(blk_bidi_rq(rq)))
375 blk_mq_free_request(rq->next_rq);
376 blk_mq_free_request(rq);
379 EXPORT_SYMBOL(__blk_mq_end_request);
381 void blk_mq_end_request(struct request *rq, int error)
383 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
385 __blk_mq_end_request(rq, error);
387 EXPORT_SYMBOL(blk_mq_end_request);
389 static void __blk_mq_complete_request_remote(void *data)
391 struct request *rq = data;
393 rq->q->softirq_done_fn(rq);
396 static void blk_mq_ipi_complete_request(struct request *rq)
398 struct blk_mq_ctx *ctx = rq->mq_ctx;
402 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
403 rq->q->softirq_done_fn(rq);
408 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
409 shared = cpus_share_cache(cpu, ctx->cpu);
411 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
412 rq->csd.func = __blk_mq_complete_request_remote;
415 smp_call_function_single_async(ctx->cpu, &rq->csd);
417 rq->q->softirq_done_fn(rq);
422 static void blk_mq_stat_add(struct request *rq)
424 if (rq->rq_flags & RQF_STATS) {
426 * We could rq->mq_ctx here, but there's less of a risk
427 * of races if we have the completion event add the stats
428 * to the local software queue.
430 struct blk_mq_ctx *ctx;
432 ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id());
433 blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq);
437 static void __blk_mq_complete_request(struct request *rq)
439 struct request_queue *q = rq->q;
443 if (!q->softirq_done_fn)
444 blk_mq_end_request(rq, rq->errors);
446 blk_mq_ipi_complete_request(rq);
450 * blk_mq_complete_request - end I/O on a request
451 * @rq: the request being processed
454 * Ends all I/O on a request. It does not handle partial completions.
455 * The actual completion happens out-of-order, through a IPI handler.
457 void blk_mq_complete_request(struct request *rq, int error)
459 struct request_queue *q = rq->q;
461 if (unlikely(blk_should_fake_timeout(q)))
463 if (!blk_mark_rq_complete(rq)) {
465 __blk_mq_complete_request(rq);
468 EXPORT_SYMBOL(blk_mq_complete_request);
470 int blk_mq_request_started(struct request *rq)
472 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
474 EXPORT_SYMBOL_GPL(blk_mq_request_started);
476 void blk_mq_start_request(struct request *rq)
478 struct request_queue *q = rq->q;
480 blk_mq_sched_started_request(rq);
482 trace_block_rq_issue(q, rq);
484 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
485 blk_stat_set_issue_time(&rq->issue_stat);
486 rq->rq_flags |= RQF_STATS;
487 wbt_issue(q->rq_wb, &rq->issue_stat);
493 * Ensure that ->deadline is visible before set the started
494 * flag and clear the completed flag.
496 smp_mb__before_atomic();
499 * Mark us as started and clear complete. Complete might have been
500 * set if requeue raced with timeout, which then marked it as
501 * complete. So be sure to clear complete again when we start
502 * the request, otherwise we'll ignore the completion event.
504 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
505 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
506 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
507 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
509 if (q->dma_drain_size && blk_rq_bytes(rq)) {
511 * Make sure space for the drain appears. We know we can do
512 * this because max_hw_segments has been adjusted to be one
513 * fewer than the device can handle.
515 rq->nr_phys_segments++;
518 EXPORT_SYMBOL(blk_mq_start_request);
520 static void __blk_mq_requeue_request(struct request *rq)
522 struct request_queue *q = rq->q;
524 trace_block_rq_requeue(q, rq);
525 wbt_requeue(q->rq_wb, &rq->issue_stat);
526 blk_mq_sched_requeue_request(rq);
528 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
529 if (q->dma_drain_size && blk_rq_bytes(rq))
530 rq->nr_phys_segments--;
534 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
536 __blk_mq_requeue_request(rq);
538 BUG_ON(blk_queued_rq(rq));
539 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
541 EXPORT_SYMBOL(blk_mq_requeue_request);
543 static void blk_mq_requeue_work(struct work_struct *work)
545 struct request_queue *q =
546 container_of(work, struct request_queue, requeue_work.work);
548 struct request *rq, *next;
551 spin_lock_irqsave(&q->requeue_lock, flags);
552 list_splice_init(&q->requeue_list, &rq_list);
553 spin_unlock_irqrestore(&q->requeue_lock, flags);
555 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
556 if (!(rq->rq_flags & RQF_SOFTBARRIER))
559 rq->rq_flags &= ~RQF_SOFTBARRIER;
560 list_del_init(&rq->queuelist);
561 blk_mq_sched_insert_request(rq, true, false, false, true);
564 while (!list_empty(&rq_list)) {
565 rq = list_entry(rq_list.next, struct request, queuelist);
566 list_del_init(&rq->queuelist);
567 blk_mq_sched_insert_request(rq, false, false, false, true);
570 blk_mq_run_hw_queues(q, false);
573 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
574 bool kick_requeue_list)
576 struct request_queue *q = rq->q;
580 * We abuse this flag that is otherwise used by the I/O scheduler to
581 * request head insertation from the workqueue.
583 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
585 spin_lock_irqsave(&q->requeue_lock, flags);
587 rq->rq_flags |= RQF_SOFTBARRIER;
588 list_add(&rq->queuelist, &q->requeue_list);
590 list_add_tail(&rq->queuelist, &q->requeue_list);
592 spin_unlock_irqrestore(&q->requeue_lock, flags);
594 if (kick_requeue_list)
595 blk_mq_kick_requeue_list(q);
597 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
599 void blk_mq_kick_requeue_list(struct request_queue *q)
601 kblockd_schedule_delayed_work(&q->requeue_work, 0);
603 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
605 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
608 kblockd_schedule_delayed_work(&q->requeue_work,
609 msecs_to_jiffies(msecs));
611 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
613 void blk_mq_abort_requeue_list(struct request_queue *q)
618 spin_lock_irqsave(&q->requeue_lock, flags);
619 list_splice_init(&q->requeue_list, &rq_list);
620 spin_unlock_irqrestore(&q->requeue_lock, flags);
622 while (!list_empty(&rq_list)) {
625 rq = list_first_entry(&rq_list, struct request, queuelist);
626 list_del_init(&rq->queuelist);
628 blk_mq_end_request(rq, rq->errors);
631 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
633 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
635 if (tag < tags->nr_tags) {
636 prefetch(tags->rqs[tag]);
637 return tags->rqs[tag];
642 EXPORT_SYMBOL(blk_mq_tag_to_rq);
644 struct blk_mq_timeout_data {
646 unsigned int next_set;
649 void blk_mq_rq_timed_out(struct request *req, bool reserved)
651 const struct blk_mq_ops *ops = req->q->mq_ops;
652 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
655 * We know that complete is set at this point. If STARTED isn't set
656 * anymore, then the request isn't active and the "timeout" should
657 * just be ignored. This can happen due to the bitflag ordering.
658 * Timeout first checks if STARTED is set, and if it is, assumes
659 * the request is active. But if we race with completion, then
660 * we both flags will get cleared. So check here again, and ignore
661 * a timeout event with a request that isn't active.
663 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
667 ret = ops->timeout(req, reserved);
671 __blk_mq_complete_request(req);
673 case BLK_EH_RESET_TIMER:
675 blk_clear_rq_complete(req);
677 case BLK_EH_NOT_HANDLED:
680 printk(KERN_ERR "block: bad eh return: %d\n", ret);
685 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
686 struct request *rq, void *priv, bool reserved)
688 struct blk_mq_timeout_data *data = priv;
690 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
692 * If a request wasn't started before the queue was
693 * marked dying, kill it here or it'll go unnoticed.
695 if (unlikely(blk_queue_dying(rq->q))) {
697 blk_mq_end_request(rq, rq->errors);
702 if (time_after_eq(jiffies, rq->deadline)) {
703 if (!blk_mark_rq_complete(rq))
704 blk_mq_rq_timed_out(rq, reserved);
705 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
706 data->next = rq->deadline;
711 static void blk_mq_timeout_work(struct work_struct *work)
713 struct request_queue *q =
714 container_of(work, struct request_queue, timeout_work);
715 struct blk_mq_timeout_data data = {
721 /* A deadlock might occur if a request is stuck requiring a
722 * timeout at the same time a queue freeze is waiting
723 * completion, since the timeout code would not be able to
724 * acquire the queue reference here.
726 * That's why we don't use blk_queue_enter here; instead, we use
727 * percpu_ref_tryget directly, because we need to be able to
728 * obtain a reference even in the short window between the queue
729 * starting to freeze, by dropping the first reference in
730 * blk_mq_freeze_queue_start, and the moment the last request is
731 * consumed, marked by the instant q_usage_counter reaches
734 if (!percpu_ref_tryget(&q->q_usage_counter))
737 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
740 data.next = blk_rq_timeout(round_jiffies_up(data.next));
741 mod_timer(&q->timeout, data.next);
743 struct blk_mq_hw_ctx *hctx;
745 queue_for_each_hw_ctx(q, hctx, i) {
746 /* the hctx may be unmapped, so check it here */
747 if (blk_mq_hw_queue_mapped(hctx))
748 blk_mq_tag_idle(hctx);
755 * Reverse check our software queue for entries that we could potentially
756 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
757 * too much time checking for merges.
759 static bool blk_mq_attempt_merge(struct request_queue *q,
760 struct blk_mq_ctx *ctx, struct bio *bio)
765 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
771 if (!blk_rq_merge_ok(rq, bio))
774 switch (blk_try_merge(rq, bio)) {
775 case ELEVATOR_BACK_MERGE:
776 if (blk_mq_sched_allow_merge(q, rq, bio))
777 merged = bio_attempt_back_merge(q, rq, bio);
779 case ELEVATOR_FRONT_MERGE:
780 if (blk_mq_sched_allow_merge(q, rq, bio))
781 merged = bio_attempt_front_merge(q, rq, bio);
783 case ELEVATOR_DISCARD_MERGE:
784 merged = bio_attempt_discard_merge(q, rq, bio);
798 struct flush_busy_ctx_data {
799 struct blk_mq_hw_ctx *hctx;
800 struct list_head *list;
803 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
805 struct flush_busy_ctx_data *flush_data = data;
806 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
807 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
809 sbitmap_clear_bit(sb, bitnr);
810 spin_lock(&ctx->lock);
811 list_splice_tail_init(&ctx->rq_list, flush_data->list);
812 spin_unlock(&ctx->lock);
817 * Process software queues that have been marked busy, splicing them
818 * to the for-dispatch
820 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
822 struct flush_busy_ctx_data data = {
827 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
829 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
831 static inline unsigned int queued_to_index(unsigned int queued)
836 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
839 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
842 struct blk_mq_alloc_data data = {
844 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
845 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
855 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
856 data.flags |= BLK_MQ_REQ_RESERVED;
858 rq->tag = blk_mq_get_tag(&data);
860 if (blk_mq_tag_busy(data.hctx)) {
861 rq->rq_flags |= RQF_MQ_INFLIGHT;
862 atomic_inc(&data.hctx->nr_active);
864 data.hctx->tags->rqs[rq->tag] = rq;
871 static void blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
874 if (rq->tag == -1 || rq->internal_tag == -1)
877 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
880 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
881 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
882 atomic_dec(&hctx->nr_active);
887 * If we fail getting a driver tag because all the driver tags are already
888 * assigned and on the dispatch list, BUT the first entry does not have a
889 * tag, then we could deadlock. For that case, move entries with assigned
890 * driver tags to the front, leaving the set of tagged requests in the
891 * same order, and the untagged set in the same order.
893 static bool reorder_tags_to_front(struct list_head *list)
895 struct request *rq, *tmp, *first = NULL;
897 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
901 list_move(&rq->queuelist, list);
907 return first != NULL;
910 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
913 struct blk_mq_hw_ctx *hctx;
915 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
917 list_del(&wait->task_list);
918 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
919 blk_mq_run_hw_queue(hctx, true);
923 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
925 struct sbq_wait_state *ws;
928 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
929 * The thread which wins the race to grab this bit adds the hardware
930 * queue to the wait queue.
932 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
933 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
936 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
937 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
940 * As soon as this returns, it's no longer safe to fiddle with
941 * hctx->dispatch_wait, since a completion can wake up the wait queue
942 * and unlock the bit.
944 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
948 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list)
950 struct request_queue *q = hctx->queue;
952 LIST_HEAD(driver_list);
953 struct list_head *dptr;
954 int queued, ret = BLK_MQ_RQ_QUEUE_OK;
957 * Start off with dptr being NULL, so we start the first request
958 * immediately, even if we have more pending.
963 * Now process all the entries, sending them to the driver.
966 while (!list_empty(list)) {
967 struct blk_mq_queue_data bd;
969 rq = list_first_entry(list, struct request, queuelist);
970 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
971 if (!queued && reorder_tags_to_front(list))
975 * The initial allocation attempt failed, so we need to
976 * rerun the hardware queue when a tag is freed.
978 if (blk_mq_dispatch_wait_add(hctx)) {
980 * It's possible that a tag was freed in the
981 * window between the allocation failure and
982 * adding the hardware queue to the wait queue.
984 if (!blk_mq_get_driver_tag(rq, &hctx, false))
991 list_del_init(&rq->queuelist);
995 bd.last = list_empty(list);
997 ret = q->mq_ops->queue_rq(hctx, &bd);
999 case BLK_MQ_RQ_QUEUE_OK:
1002 case BLK_MQ_RQ_QUEUE_BUSY:
1003 blk_mq_put_driver_tag(hctx, rq);
1004 list_add(&rq->queuelist, list);
1005 __blk_mq_requeue_request(rq);
1008 pr_err("blk-mq: bad return on queue: %d\n", ret);
1009 case BLK_MQ_RQ_QUEUE_ERROR:
1011 blk_mq_end_request(rq, rq->errors);
1015 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
1019 * We've done the first request. If we have more than 1
1020 * left in the list, set dptr to defer issue.
1022 if (!dptr && list->next != list->prev)
1023 dptr = &driver_list;
1026 hctx->dispatched[queued_to_index(queued)]++;
1029 * Any items that need requeuing? Stuff them into hctx->dispatch,
1030 * that is where we will continue on next queue run.
1032 if (!list_empty(list)) {
1033 spin_lock(&hctx->lock);
1034 list_splice_init(list, &hctx->dispatch);
1035 spin_unlock(&hctx->lock);
1038 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1039 * it's possible the queue is stopped and restarted again
1040 * before this. Queue restart will dispatch requests. And since
1041 * requests in rq_list aren't added into hctx->dispatch yet,
1042 * the requests in rq_list might get lost.
1044 * blk_mq_run_hw_queue() already checks the STOPPED bit
1046 * If RESTART or TAG_WAITING is set, then let completion restart
1047 * the queue instead of potentially looping here.
1049 if (!blk_mq_sched_needs_restart(hctx) &&
1050 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1051 blk_mq_run_hw_queue(hctx, true);
1057 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1061 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1062 cpu_online(hctx->next_cpu));
1064 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1066 blk_mq_sched_dispatch_requests(hctx);
1069 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1070 blk_mq_sched_dispatch_requests(hctx);
1071 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1076 * It'd be great if the workqueue API had a way to pass
1077 * in a mask and had some smarts for more clever placement.
1078 * For now we just round-robin here, switching for every
1079 * BLK_MQ_CPU_WORK_BATCH queued items.
1081 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1083 if (hctx->queue->nr_hw_queues == 1)
1084 return WORK_CPU_UNBOUND;
1086 if (--hctx->next_cpu_batch <= 0) {
1089 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1090 if (next_cpu >= nr_cpu_ids)
1091 next_cpu = cpumask_first(hctx->cpumask);
1093 hctx->next_cpu = next_cpu;
1094 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1097 return hctx->next_cpu;
1100 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1102 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1103 !blk_mq_hw_queue_mapped(hctx)))
1106 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1107 int cpu = get_cpu();
1108 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1109 __blk_mq_run_hw_queue(hctx);
1117 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
1120 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1122 struct blk_mq_hw_ctx *hctx;
1125 queue_for_each_hw_ctx(q, hctx, i) {
1126 if (!blk_mq_hctx_has_pending(hctx) ||
1127 blk_mq_hctx_stopped(hctx))
1130 blk_mq_run_hw_queue(hctx, async);
1133 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1136 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1137 * @q: request queue.
1139 * The caller is responsible for serializing this function against
1140 * blk_mq_{start,stop}_hw_queue().
1142 bool blk_mq_queue_stopped(struct request_queue *q)
1144 struct blk_mq_hw_ctx *hctx;
1147 queue_for_each_hw_ctx(q, hctx, i)
1148 if (blk_mq_hctx_stopped(hctx))
1153 EXPORT_SYMBOL(blk_mq_queue_stopped);
1155 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1157 cancel_work(&hctx->run_work);
1158 cancel_delayed_work(&hctx->delay_work);
1159 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1161 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1163 void blk_mq_stop_hw_queues(struct request_queue *q)
1165 struct blk_mq_hw_ctx *hctx;
1168 queue_for_each_hw_ctx(q, hctx, i)
1169 blk_mq_stop_hw_queue(hctx);
1171 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1173 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1175 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1177 blk_mq_run_hw_queue(hctx, false);
1179 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1181 void blk_mq_start_hw_queues(struct request_queue *q)
1183 struct blk_mq_hw_ctx *hctx;
1186 queue_for_each_hw_ctx(q, hctx, i)
1187 blk_mq_start_hw_queue(hctx);
1189 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1191 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1193 if (!blk_mq_hctx_stopped(hctx))
1196 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1197 blk_mq_run_hw_queue(hctx, async);
1199 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1201 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1203 struct blk_mq_hw_ctx *hctx;
1206 queue_for_each_hw_ctx(q, hctx, i)
1207 blk_mq_start_stopped_hw_queue(hctx, async);
1209 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1211 static void blk_mq_run_work_fn(struct work_struct *work)
1213 struct blk_mq_hw_ctx *hctx;
1215 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1217 __blk_mq_run_hw_queue(hctx);
1220 static void blk_mq_delay_work_fn(struct work_struct *work)
1222 struct blk_mq_hw_ctx *hctx;
1224 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1226 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1227 __blk_mq_run_hw_queue(hctx);
1230 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1232 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1235 blk_mq_stop_hw_queue(hctx);
1236 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1237 &hctx->delay_work, msecs_to_jiffies(msecs));
1239 EXPORT_SYMBOL(blk_mq_delay_queue);
1241 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1245 struct blk_mq_ctx *ctx = rq->mq_ctx;
1247 trace_block_rq_insert(hctx->queue, rq);
1250 list_add(&rq->queuelist, &ctx->rq_list);
1252 list_add_tail(&rq->queuelist, &ctx->rq_list);
1255 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1258 struct blk_mq_ctx *ctx = rq->mq_ctx;
1260 __blk_mq_insert_req_list(hctx, rq, at_head);
1261 blk_mq_hctx_mark_pending(hctx, ctx);
1264 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1265 struct list_head *list)
1269 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1272 spin_lock(&ctx->lock);
1273 while (!list_empty(list)) {
1276 rq = list_first_entry(list, struct request, queuelist);
1277 BUG_ON(rq->mq_ctx != ctx);
1278 list_del_init(&rq->queuelist);
1279 __blk_mq_insert_req_list(hctx, rq, false);
1281 blk_mq_hctx_mark_pending(hctx, ctx);
1282 spin_unlock(&ctx->lock);
1285 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1287 struct request *rqa = container_of(a, struct request, queuelist);
1288 struct request *rqb = container_of(b, struct request, queuelist);
1290 return !(rqa->mq_ctx < rqb->mq_ctx ||
1291 (rqa->mq_ctx == rqb->mq_ctx &&
1292 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1295 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1297 struct blk_mq_ctx *this_ctx;
1298 struct request_queue *this_q;
1301 LIST_HEAD(ctx_list);
1304 list_splice_init(&plug->mq_list, &list);
1306 list_sort(NULL, &list, plug_ctx_cmp);
1312 while (!list_empty(&list)) {
1313 rq = list_entry_rq(list.next);
1314 list_del_init(&rq->queuelist);
1316 if (rq->mq_ctx != this_ctx) {
1318 trace_block_unplug(this_q, depth, from_schedule);
1319 blk_mq_sched_insert_requests(this_q, this_ctx,
1324 this_ctx = rq->mq_ctx;
1330 list_add_tail(&rq->queuelist, &ctx_list);
1334 * If 'this_ctx' is set, we know we have entries to complete
1335 * on 'ctx_list'. Do those.
1338 trace_block_unplug(this_q, depth, from_schedule);
1339 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1344 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1346 init_request_from_bio(rq, bio);
1348 blk_account_io_start(rq, true);
1351 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1353 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1354 !blk_queue_nomerges(hctx->queue);
1357 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1358 struct blk_mq_ctx *ctx,
1359 struct request *rq, struct bio *bio)
1361 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1362 blk_mq_bio_to_request(rq, bio);
1363 spin_lock(&ctx->lock);
1365 __blk_mq_insert_request(hctx, rq, false);
1366 spin_unlock(&ctx->lock);
1369 struct request_queue *q = hctx->queue;
1371 spin_lock(&ctx->lock);
1372 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1373 blk_mq_bio_to_request(rq, bio);
1377 spin_unlock(&ctx->lock);
1378 __blk_mq_finish_request(hctx, ctx, rq);
1383 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1386 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1388 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1391 static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie)
1393 struct request_queue *q = rq->q;
1394 struct blk_mq_queue_data bd = {
1399 struct blk_mq_hw_ctx *hctx;
1400 blk_qc_t new_cookie;
1406 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1409 new_cookie = request_to_qc_t(hctx, rq);
1412 * For OK queue, we are done. For error, kill it. Any other
1413 * error (busy), just add it to our list as we previously
1416 ret = q->mq_ops->queue_rq(hctx, &bd);
1417 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1418 *cookie = new_cookie;
1422 __blk_mq_requeue_request(rq);
1424 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1425 *cookie = BLK_QC_T_NONE;
1427 blk_mq_end_request(rq, rq->errors);
1432 blk_mq_sched_insert_request(rq, false, true, true, false);
1436 * Multiple hardware queue variant. This will not use per-process plugs,
1437 * but will attempt to bypass the hctx queueing if we can go straight to
1438 * hardware for SYNC IO.
1440 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1442 const int is_sync = op_is_sync(bio->bi_opf);
1443 const int is_flush_fua = op_is_flush(bio->bi_opf);
1444 struct blk_mq_alloc_data data = { .flags = 0 };
1446 unsigned int request_count = 0, srcu_idx;
1447 struct blk_plug *plug;
1448 struct request *same_queue_rq = NULL;
1450 unsigned int wb_acct;
1452 blk_queue_bounce(q, &bio);
1454 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1456 return BLK_QC_T_NONE;
1459 blk_queue_split(q, &bio, q->bio_split);
1461 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1462 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1463 return BLK_QC_T_NONE;
1465 if (blk_mq_sched_bio_merge(q, bio))
1466 return BLK_QC_T_NONE;
1468 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1470 trace_block_getrq(q, bio, bio->bi_opf);
1472 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1473 if (unlikely(!rq)) {
1474 __wbt_done(q->rq_wb, wb_acct);
1475 return BLK_QC_T_NONE;
1478 wbt_track(&rq->issue_stat, wb_acct);
1480 cookie = request_to_qc_t(data.hctx, rq);
1482 if (unlikely(is_flush_fua)) {
1485 blk_mq_bio_to_request(rq, bio);
1486 blk_insert_flush(rq);
1490 plug = current->plug;
1492 * If the driver supports defer issued based on 'last', then
1493 * queue it up like normal since we can potentially save some
1496 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1497 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1498 struct request *old_rq = NULL;
1500 blk_mq_bio_to_request(rq, bio);
1503 * We do limited plugging. If the bio can be merged, do that.
1504 * Otherwise the existing request in the plug list will be
1505 * issued. So the plug list will have one request at most
1509 * The plug list might get flushed before this. If that
1510 * happens, same_queue_rq is invalid and plug list is
1513 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1514 old_rq = same_queue_rq;
1515 list_del_init(&old_rq->queuelist);
1517 list_add_tail(&rq->queuelist, &plug->mq_list);
1518 } else /* is_sync */
1520 blk_mq_put_ctx(data.ctx);
1524 if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) {
1526 blk_mq_try_issue_directly(old_rq, &cookie);
1529 srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu);
1530 blk_mq_try_issue_directly(old_rq, &cookie);
1531 srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx);
1538 blk_mq_put_ctx(data.ctx);
1539 blk_mq_bio_to_request(rq, bio);
1540 blk_mq_sched_insert_request(rq, false, true,
1541 !is_sync || is_flush_fua, true);
1544 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1546 * For a SYNC request, send it to the hardware immediately. For
1547 * an ASYNC request, just ensure that we run it later on. The
1548 * latter allows for merging opportunities and more efficient
1552 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1554 blk_mq_put_ctx(data.ctx);
1560 * Single hardware queue variant. This will attempt to use any per-process
1561 * plug for merging and IO deferral.
1563 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1565 const int is_sync = op_is_sync(bio->bi_opf);
1566 const int is_flush_fua = op_is_flush(bio->bi_opf);
1567 struct blk_plug *plug;
1568 unsigned int request_count = 0;
1569 struct blk_mq_alloc_data data = { .flags = 0 };
1572 unsigned int wb_acct;
1574 blk_queue_bounce(q, &bio);
1576 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1578 return BLK_QC_T_NONE;
1581 blk_queue_split(q, &bio, q->bio_split);
1583 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1584 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1585 return BLK_QC_T_NONE;
1587 request_count = blk_plug_queued_count(q);
1589 if (blk_mq_sched_bio_merge(q, bio))
1590 return BLK_QC_T_NONE;
1592 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1594 trace_block_getrq(q, bio, bio->bi_opf);
1596 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1597 if (unlikely(!rq)) {
1598 __wbt_done(q->rq_wb, wb_acct);
1599 return BLK_QC_T_NONE;
1602 wbt_track(&rq->issue_stat, wb_acct);
1604 cookie = request_to_qc_t(data.hctx, rq);
1606 if (unlikely(is_flush_fua)) {
1609 blk_mq_bio_to_request(rq, bio);
1610 blk_insert_flush(rq);
1615 * A task plug currently exists. Since this is completely lockless,
1616 * utilize that to temporarily store requests until the task is
1617 * either done or scheduled away.
1619 plug = current->plug;
1621 struct request *last = NULL;
1623 blk_mq_bio_to_request(rq, bio);
1626 * @request_count may become stale because of schedule
1627 * out, so check the list again.
1629 if (list_empty(&plug->mq_list))
1632 trace_block_plug(q);
1634 last = list_entry_rq(plug->mq_list.prev);
1636 blk_mq_put_ctx(data.ctx);
1638 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1639 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1640 blk_flush_plug_list(plug, false);
1641 trace_block_plug(q);
1644 list_add_tail(&rq->queuelist, &plug->mq_list);
1650 blk_mq_put_ctx(data.ctx);
1651 blk_mq_bio_to_request(rq, bio);
1652 blk_mq_sched_insert_request(rq, false, true,
1653 !is_sync || is_flush_fua, true);
1656 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1658 * For a SYNC request, send it to the hardware immediately. For
1659 * an ASYNC request, just ensure that we run it later on. The
1660 * latter allows for merging opportunities and more efficient
1664 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1667 blk_mq_put_ctx(data.ctx);
1672 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1673 unsigned int hctx_idx)
1677 if (tags->rqs && set->ops->exit_request) {
1680 for (i = 0; i < tags->nr_tags; i++) {
1681 struct request *rq = tags->static_rqs[i];
1685 set->ops->exit_request(set->driver_data, rq,
1687 tags->static_rqs[i] = NULL;
1691 while (!list_empty(&tags->page_list)) {
1692 page = list_first_entry(&tags->page_list, struct page, lru);
1693 list_del_init(&page->lru);
1695 * Remove kmemleak object previously allocated in
1696 * blk_mq_init_rq_map().
1698 kmemleak_free(page_address(page));
1699 __free_pages(page, page->private);
1703 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1707 kfree(tags->static_rqs);
1708 tags->static_rqs = NULL;
1710 blk_mq_free_tags(tags);
1713 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1714 unsigned int hctx_idx,
1715 unsigned int nr_tags,
1716 unsigned int reserved_tags)
1718 struct blk_mq_tags *tags;
1721 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1722 if (node == NUMA_NO_NODE)
1723 node = set->numa_node;
1725 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1726 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1730 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1731 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1734 blk_mq_free_tags(tags);
1738 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1739 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1741 if (!tags->static_rqs) {
1743 blk_mq_free_tags(tags);
1750 static size_t order_to_size(unsigned int order)
1752 return (size_t)PAGE_SIZE << order;
1755 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1756 unsigned int hctx_idx, unsigned int depth)
1758 unsigned int i, j, entries_per_page, max_order = 4;
1759 size_t rq_size, left;
1762 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1763 if (node == NUMA_NO_NODE)
1764 node = set->numa_node;
1766 INIT_LIST_HEAD(&tags->page_list);
1769 * rq_size is the size of the request plus driver payload, rounded
1770 * to the cacheline size
1772 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1774 left = rq_size * depth;
1776 for (i = 0; i < depth; ) {
1777 int this_order = max_order;
1782 while (this_order && left < order_to_size(this_order - 1))
1786 page = alloc_pages_node(node,
1787 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1793 if (order_to_size(this_order) < rq_size)
1800 page->private = this_order;
1801 list_add_tail(&page->lru, &tags->page_list);
1803 p = page_address(page);
1805 * Allow kmemleak to scan these pages as they contain pointers
1806 * to additional allocations like via ops->init_request().
1808 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1809 entries_per_page = order_to_size(this_order) / rq_size;
1810 to_do = min(entries_per_page, depth - i);
1811 left -= to_do * rq_size;
1812 for (j = 0; j < to_do; j++) {
1813 struct request *rq = p;
1815 tags->static_rqs[i] = rq;
1816 if (set->ops->init_request) {
1817 if (set->ops->init_request(set->driver_data,
1820 tags->static_rqs[i] = NULL;
1832 blk_mq_free_rqs(set, tags, hctx_idx);
1837 * 'cpu' is going away. splice any existing rq_list entries from this
1838 * software queue to the hw queue dispatch list, and ensure that it
1841 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1843 struct blk_mq_hw_ctx *hctx;
1844 struct blk_mq_ctx *ctx;
1847 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1848 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1850 spin_lock(&ctx->lock);
1851 if (!list_empty(&ctx->rq_list)) {
1852 list_splice_init(&ctx->rq_list, &tmp);
1853 blk_mq_hctx_clear_pending(hctx, ctx);
1855 spin_unlock(&ctx->lock);
1857 if (list_empty(&tmp))
1860 spin_lock(&hctx->lock);
1861 list_splice_tail_init(&tmp, &hctx->dispatch);
1862 spin_unlock(&hctx->lock);
1864 blk_mq_run_hw_queue(hctx, true);
1868 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1870 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1874 /* hctx->ctxs will be freed in queue's release handler */
1875 static void blk_mq_exit_hctx(struct request_queue *q,
1876 struct blk_mq_tag_set *set,
1877 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1879 unsigned flush_start_tag = set->queue_depth;
1881 blk_mq_tag_idle(hctx);
1883 if (set->ops->exit_request)
1884 set->ops->exit_request(set->driver_data,
1885 hctx->fq->flush_rq, hctx_idx,
1886 flush_start_tag + hctx_idx);
1888 if (set->ops->exit_hctx)
1889 set->ops->exit_hctx(hctx, hctx_idx);
1891 if (hctx->flags & BLK_MQ_F_BLOCKING)
1892 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1894 blk_mq_remove_cpuhp(hctx);
1895 blk_free_flush_queue(hctx->fq);
1896 sbitmap_free(&hctx->ctx_map);
1899 static void blk_mq_exit_hw_queues(struct request_queue *q,
1900 struct blk_mq_tag_set *set, int nr_queue)
1902 struct blk_mq_hw_ctx *hctx;
1905 queue_for_each_hw_ctx(q, hctx, i) {
1908 blk_mq_exit_hctx(q, set, hctx, i);
1912 static void blk_mq_free_hw_queues(struct request_queue *q,
1913 struct blk_mq_tag_set *set)
1915 struct blk_mq_hw_ctx *hctx;
1918 queue_for_each_hw_ctx(q, hctx, i)
1919 free_cpumask_var(hctx->cpumask);
1922 static int blk_mq_init_hctx(struct request_queue *q,
1923 struct blk_mq_tag_set *set,
1924 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1927 unsigned flush_start_tag = set->queue_depth;
1929 node = hctx->numa_node;
1930 if (node == NUMA_NO_NODE)
1931 node = hctx->numa_node = set->numa_node;
1933 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1934 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1935 spin_lock_init(&hctx->lock);
1936 INIT_LIST_HEAD(&hctx->dispatch);
1938 hctx->queue_num = hctx_idx;
1939 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1941 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1943 hctx->tags = set->tags[hctx_idx];
1946 * Allocate space for all possible cpus to avoid allocation at
1949 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1952 goto unregister_cpu_notifier;
1954 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1960 if (set->ops->init_hctx &&
1961 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1964 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1968 if (set->ops->init_request &&
1969 set->ops->init_request(set->driver_data,
1970 hctx->fq->flush_rq, hctx_idx,
1971 flush_start_tag + hctx_idx, node))
1974 if (hctx->flags & BLK_MQ_F_BLOCKING)
1975 init_srcu_struct(&hctx->queue_rq_srcu);
1982 if (set->ops->exit_hctx)
1983 set->ops->exit_hctx(hctx, hctx_idx);
1985 sbitmap_free(&hctx->ctx_map);
1988 unregister_cpu_notifier:
1989 blk_mq_remove_cpuhp(hctx);
1993 static void blk_mq_init_cpu_queues(struct request_queue *q,
1994 unsigned int nr_hw_queues)
1998 for_each_possible_cpu(i) {
1999 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2000 struct blk_mq_hw_ctx *hctx;
2002 memset(__ctx, 0, sizeof(*__ctx));
2004 spin_lock_init(&__ctx->lock);
2005 INIT_LIST_HEAD(&__ctx->rq_list);
2007 blk_stat_init(&__ctx->stat[BLK_STAT_READ]);
2008 blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]);
2010 /* If the cpu isn't online, the cpu is mapped to first hctx */
2014 hctx = blk_mq_map_queue(q, i);
2017 * Set local node, IFF we have more than one hw queue. If
2018 * not, we remain on the home node of the device
2020 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2021 hctx->numa_node = local_memory_node(cpu_to_node(i));
2025 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2029 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2030 set->queue_depth, set->reserved_tags);
2031 if (!set->tags[hctx_idx])
2034 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2039 blk_mq_free_rq_map(set->tags[hctx_idx]);
2040 set->tags[hctx_idx] = NULL;
2044 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2045 unsigned int hctx_idx)
2047 if (set->tags[hctx_idx]) {
2048 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2049 blk_mq_free_rq_map(set->tags[hctx_idx]);
2050 set->tags[hctx_idx] = NULL;
2054 static void blk_mq_map_swqueue(struct request_queue *q,
2055 const struct cpumask *online_mask)
2057 unsigned int i, hctx_idx;
2058 struct blk_mq_hw_ctx *hctx;
2059 struct blk_mq_ctx *ctx;
2060 struct blk_mq_tag_set *set = q->tag_set;
2063 * Avoid others reading imcomplete hctx->cpumask through sysfs
2065 mutex_lock(&q->sysfs_lock);
2067 queue_for_each_hw_ctx(q, hctx, i) {
2068 cpumask_clear(hctx->cpumask);
2073 * Map software to hardware queues
2075 for_each_possible_cpu(i) {
2076 /* If the cpu isn't online, the cpu is mapped to first hctx */
2077 if (!cpumask_test_cpu(i, online_mask))
2080 hctx_idx = q->mq_map[i];
2081 /* unmapped hw queue can be remapped after CPU topo changed */
2082 if (!set->tags[hctx_idx] &&
2083 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2085 * If tags initialization fail for some hctx,
2086 * that hctx won't be brought online. In this
2087 * case, remap the current ctx to hctx[0] which
2088 * is guaranteed to always have tags allocated
2093 ctx = per_cpu_ptr(q->queue_ctx, i);
2094 hctx = blk_mq_map_queue(q, i);
2096 cpumask_set_cpu(i, hctx->cpumask);
2097 ctx->index_hw = hctx->nr_ctx;
2098 hctx->ctxs[hctx->nr_ctx++] = ctx;
2101 mutex_unlock(&q->sysfs_lock);
2103 queue_for_each_hw_ctx(q, hctx, i) {
2105 * If no software queues are mapped to this hardware queue,
2106 * disable it and free the request entries.
2108 if (!hctx->nr_ctx) {
2109 /* Never unmap queue 0. We need it as a
2110 * fallback in case of a new remap fails
2113 if (i && set->tags[i])
2114 blk_mq_free_map_and_requests(set, i);
2120 hctx->tags = set->tags[i];
2121 WARN_ON(!hctx->tags);
2124 * Set the map size to the number of mapped software queues.
2125 * This is more accurate and more efficient than looping
2126 * over all possibly mapped software queues.
2128 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2131 * Initialize batch roundrobin counts
2133 hctx->next_cpu = cpumask_first(hctx->cpumask);
2134 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2138 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2140 struct blk_mq_hw_ctx *hctx;
2143 queue_for_each_hw_ctx(q, hctx, i) {
2145 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2147 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2151 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2153 struct request_queue *q;
2155 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2156 blk_mq_freeze_queue(q);
2157 queue_set_hctx_shared(q, shared);
2158 blk_mq_unfreeze_queue(q);
2162 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2164 struct blk_mq_tag_set *set = q->tag_set;
2166 mutex_lock(&set->tag_list_lock);
2167 list_del_init(&q->tag_set_list);
2168 if (list_is_singular(&set->tag_list)) {
2169 /* just transitioned to unshared */
2170 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2171 /* update existing queue */
2172 blk_mq_update_tag_set_depth(set, false);
2174 mutex_unlock(&set->tag_list_lock);
2177 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2178 struct request_queue *q)
2182 mutex_lock(&set->tag_list_lock);
2184 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2185 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2186 set->flags |= BLK_MQ_F_TAG_SHARED;
2187 /* update existing queue */
2188 blk_mq_update_tag_set_depth(set, true);
2190 if (set->flags & BLK_MQ_F_TAG_SHARED)
2191 queue_set_hctx_shared(q, true);
2192 list_add_tail(&q->tag_set_list, &set->tag_list);
2194 mutex_unlock(&set->tag_list_lock);
2198 * It is the actual release handler for mq, but we do it from
2199 * request queue's release handler for avoiding use-after-free
2200 * and headache because q->mq_kobj shouldn't have been introduced,
2201 * but we can't group ctx/kctx kobj without it.
2203 void blk_mq_release(struct request_queue *q)
2205 struct blk_mq_hw_ctx *hctx;
2208 blk_mq_sched_teardown(q);
2210 /* hctx kobj stays in hctx */
2211 queue_for_each_hw_ctx(q, hctx, i) {
2220 kfree(q->queue_hw_ctx);
2222 /* ctx kobj stays in queue_ctx */
2223 free_percpu(q->queue_ctx);
2226 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2228 struct request_queue *uninit_q, *q;
2230 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2232 return ERR_PTR(-ENOMEM);
2234 q = blk_mq_init_allocated_queue(set, uninit_q);
2236 blk_cleanup_queue(uninit_q);
2240 EXPORT_SYMBOL(blk_mq_init_queue);
2242 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2243 struct request_queue *q)
2246 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2248 blk_mq_sysfs_unregister(q);
2249 for (i = 0; i < set->nr_hw_queues; i++) {
2255 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2256 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2261 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2268 atomic_set(&hctxs[i]->nr_active, 0);
2269 hctxs[i]->numa_node = node;
2270 hctxs[i]->queue_num = i;
2272 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2273 free_cpumask_var(hctxs[i]->cpumask);
2278 blk_mq_hctx_kobj_init(hctxs[i]);
2280 for (j = i; j < q->nr_hw_queues; j++) {
2281 struct blk_mq_hw_ctx *hctx = hctxs[j];
2285 blk_mq_free_map_and_requests(set, j);
2286 blk_mq_exit_hctx(q, set, hctx, j);
2287 free_cpumask_var(hctx->cpumask);
2288 kobject_put(&hctx->kobj);
2295 q->nr_hw_queues = i;
2296 blk_mq_sysfs_register(q);
2299 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2300 struct request_queue *q)
2302 /* mark the queue as mq asap */
2303 q->mq_ops = set->ops;
2305 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2309 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2310 GFP_KERNEL, set->numa_node);
2311 if (!q->queue_hw_ctx)
2314 q->mq_map = set->mq_map;
2316 blk_mq_realloc_hw_ctxs(set, q);
2317 if (!q->nr_hw_queues)
2320 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2321 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2323 q->nr_queues = nr_cpu_ids;
2325 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2327 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2328 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2330 q->sg_reserved_size = INT_MAX;
2332 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2333 INIT_LIST_HEAD(&q->requeue_list);
2334 spin_lock_init(&q->requeue_lock);
2336 if (q->nr_hw_queues > 1)
2337 blk_queue_make_request(q, blk_mq_make_request);
2339 blk_queue_make_request(q, blk_sq_make_request);
2342 * Do this after blk_queue_make_request() overrides it...
2344 q->nr_requests = set->queue_depth;
2347 * Default to classic polling
2351 if (set->ops->complete)
2352 blk_queue_softirq_done(q, set->ops->complete);
2354 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2357 mutex_lock(&all_q_mutex);
2359 list_add_tail(&q->all_q_node, &all_q_list);
2360 blk_mq_add_queue_tag_set(set, q);
2361 blk_mq_map_swqueue(q, cpu_online_mask);
2363 mutex_unlock(&all_q_mutex);
2366 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2369 ret = blk_mq_sched_init(q);
2371 return ERR_PTR(ret);
2377 kfree(q->queue_hw_ctx);
2379 free_percpu(q->queue_ctx);
2382 return ERR_PTR(-ENOMEM);
2384 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2386 void blk_mq_free_queue(struct request_queue *q)
2388 struct blk_mq_tag_set *set = q->tag_set;
2390 mutex_lock(&all_q_mutex);
2391 list_del_init(&q->all_q_node);
2392 mutex_unlock(&all_q_mutex);
2396 blk_mq_del_queue_tag_set(q);
2398 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2399 blk_mq_free_hw_queues(q, set);
2402 /* Basically redo blk_mq_init_queue with queue frozen */
2403 static void blk_mq_queue_reinit(struct request_queue *q,
2404 const struct cpumask *online_mask)
2406 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2408 blk_mq_sysfs_unregister(q);
2411 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2412 * we should change hctx numa_node according to new topology (this
2413 * involves free and re-allocate memory, worthy doing?)
2416 blk_mq_map_swqueue(q, online_mask);
2418 blk_mq_sysfs_register(q);
2422 * New online cpumask which is going to be set in this hotplug event.
2423 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2424 * one-by-one and dynamically allocating this could result in a failure.
2426 static struct cpumask cpuhp_online_new;
2428 static void blk_mq_queue_reinit_work(void)
2430 struct request_queue *q;
2432 mutex_lock(&all_q_mutex);
2434 * We need to freeze and reinit all existing queues. Freezing
2435 * involves synchronous wait for an RCU grace period and doing it
2436 * one by one may take a long time. Start freezing all queues in
2437 * one swoop and then wait for the completions so that freezing can
2438 * take place in parallel.
2440 list_for_each_entry(q, &all_q_list, all_q_node)
2441 blk_mq_freeze_queue_start(q);
2442 list_for_each_entry(q, &all_q_list, all_q_node)
2443 blk_mq_freeze_queue_wait(q);
2445 list_for_each_entry(q, &all_q_list, all_q_node)
2446 blk_mq_queue_reinit(q, &cpuhp_online_new);
2448 list_for_each_entry(q, &all_q_list, all_q_node)
2449 blk_mq_unfreeze_queue(q);
2451 mutex_unlock(&all_q_mutex);
2454 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2456 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2457 blk_mq_queue_reinit_work();
2462 * Before hotadded cpu starts handling requests, new mappings must be
2463 * established. Otherwise, these requests in hw queue might never be
2466 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2467 * for CPU0, and ctx1 for CPU1).
2469 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2470 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2472 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2473 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2474 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2477 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2479 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2480 cpumask_set_cpu(cpu, &cpuhp_online_new);
2481 blk_mq_queue_reinit_work();
2485 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2489 for (i = 0; i < set->nr_hw_queues; i++)
2490 if (!__blk_mq_alloc_rq_map(set, i))
2497 blk_mq_free_rq_map(set->tags[i]);
2503 * Allocate the request maps associated with this tag_set. Note that this
2504 * may reduce the depth asked for, if memory is tight. set->queue_depth
2505 * will be updated to reflect the allocated depth.
2507 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2512 depth = set->queue_depth;
2514 err = __blk_mq_alloc_rq_maps(set);
2518 set->queue_depth >>= 1;
2519 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2523 } while (set->queue_depth);
2525 if (!set->queue_depth || err) {
2526 pr_err("blk-mq: failed to allocate request map\n");
2530 if (depth != set->queue_depth)
2531 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2532 depth, set->queue_depth);
2538 * Alloc a tag set to be associated with one or more request queues.
2539 * May fail with EINVAL for various error conditions. May adjust the
2540 * requested depth down, if if it too large. In that case, the set
2541 * value will be stored in set->queue_depth.
2543 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2547 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2549 if (!set->nr_hw_queues)
2551 if (!set->queue_depth)
2553 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2556 if (!set->ops->queue_rq)
2559 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2560 pr_info("blk-mq: reduced tag depth to %u\n",
2562 set->queue_depth = BLK_MQ_MAX_DEPTH;
2566 * If a crashdump is active, then we are potentially in a very
2567 * memory constrained environment. Limit us to 1 queue and
2568 * 64 tags to prevent using too much memory.
2570 if (is_kdump_kernel()) {
2571 set->nr_hw_queues = 1;
2572 set->queue_depth = min(64U, set->queue_depth);
2575 * There is no use for more h/w queues than cpus.
2577 if (set->nr_hw_queues > nr_cpu_ids)
2578 set->nr_hw_queues = nr_cpu_ids;
2580 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2581 GFP_KERNEL, set->numa_node);
2586 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2587 GFP_KERNEL, set->numa_node);
2591 if (set->ops->map_queues)
2592 ret = set->ops->map_queues(set);
2594 ret = blk_mq_map_queues(set);
2596 goto out_free_mq_map;
2598 ret = blk_mq_alloc_rq_maps(set);
2600 goto out_free_mq_map;
2602 mutex_init(&set->tag_list_lock);
2603 INIT_LIST_HEAD(&set->tag_list);
2615 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2617 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2621 for (i = 0; i < nr_cpu_ids; i++)
2622 blk_mq_free_map_and_requests(set, i);
2630 EXPORT_SYMBOL(blk_mq_free_tag_set);
2632 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2634 struct blk_mq_tag_set *set = q->tag_set;
2635 struct blk_mq_hw_ctx *hctx;
2641 blk_mq_freeze_queue(q);
2642 blk_mq_quiesce_queue(q);
2645 queue_for_each_hw_ctx(q, hctx, i) {
2649 * If we're using an MQ scheduler, just update the scheduler
2650 * queue depth. This is similar to what the old code would do.
2652 if (!hctx->sched_tags) {
2653 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2654 min(nr, set->queue_depth),
2657 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2665 q->nr_requests = nr;
2667 blk_mq_unfreeze_queue(q);
2668 blk_mq_start_stopped_hw_queues(q, true);
2673 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2675 struct request_queue *q;
2677 if (nr_hw_queues > nr_cpu_ids)
2678 nr_hw_queues = nr_cpu_ids;
2679 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2682 list_for_each_entry(q, &set->tag_list, tag_set_list)
2683 blk_mq_freeze_queue(q);
2685 set->nr_hw_queues = nr_hw_queues;
2686 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2687 blk_mq_realloc_hw_ctxs(set, q);
2690 * Manually set the make_request_fn as blk_queue_make_request
2691 * resets a lot of the queue settings.
2693 if (q->nr_hw_queues > 1)
2694 q->make_request_fn = blk_mq_make_request;
2696 q->make_request_fn = blk_sq_make_request;
2698 blk_mq_queue_reinit(q, cpu_online_mask);
2701 list_for_each_entry(q, &set->tag_list, tag_set_list)
2702 blk_mq_unfreeze_queue(q);
2704 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2706 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2707 struct blk_mq_hw_ctx *hctx,
2710 struct blk_rq_stat stat[2];
2711 unsigned long ret = 0;
2714 * If stats collection isn't on, don't sleep but turn it on for
2717 if (!blk_stat_enable(q))
2721 * We don't have to do this once per IO, should optimize this
2722 * to just use the current window of stats until it changes
2724 memset(&stat, 0, sizeof(stat));
2725 blk_hctx_stat_get(hctx, stat);
2728 * As an optimistic guess, use half of the mean service time
2729 * for this type of request. We can (and should) make this smarter.
2730 * For instance, if the completion latencies are tight, we can
2731 * get closer than just half the mean. This is especially
2732 * important on devices where the completion latencies are longer
2735 if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples)
2736 ret = (stat[BLK_STAT_READ].mean + 1) / 2;
2737 else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples)
2738 ret = (stat[BLK_STAT_WRITE].mean + 1) / 2;
2743 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2744 struct blk_mq_hw_ctx *hctx,
2747 struct hrtimer_sleeper hs;
2748 enum hrtimer_mode mode;
2752 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2758 * -1: don't ever hybrid sleep
2759 * 0: use half of prev avg
2760 * >0: use this specific value
2762 if (q->poll_nsec == -1)
2764 else if (q->poll_nsec > 0)
2765 nsecs = q->poll_nsec;
2767 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2772 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2775 * This will be replaced with the stats tracking code, using
2776 * 'avg_completion_time / 2' as the pre-sleep target.
2780 mode = HRTIMER_MODE_REL;
2781 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2782 hrtimer_set_expires(&hs.timer, kt);
2784 hrtimer_init_sleeper(&hs, current);
2786 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2788 set_current_state(TASK_UNINTERRUPTIBLE);
2789 hrtimer_start_expires(&hs.timer, mode);
2792 hrtimer_cancel(&hs.timer);
2793 mode = HRTIMER_MODE_ABS;
2794 } while (hs.task && !signal_pending(current));
2796 __set_current_state(TASK_RUNNING);
2797 destroy_hrtimer_on_stack(&hs.timer);
2801 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2803 struct request_queue *q = hctx->queue;
2807 * If we sleep, have the caller restart the poll loop to reset
2808 * the state. Like for the other success return cases, the
2809 * caller is responsible for checking if the IO completed. If
2810 * the IO isn't complete, we'll get called again and will go
2811 * straight to the busy poll loop.
2813 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2816 hctx->poll_considered++;
2818 state = current->state;
2819 while (!need_resched()) {
2822 hctx->poll_invoked++;
2824 ret = q->mq_ops->poll(hctx, rq->tag);
2826 hctx->poll_success++;
2827 set_current_state(TASK_RUNNING);
2831 if (signal_pending_state(state, current))
2832 set_current_state(TASK_RUNNING);
2834 if (current->state == TASK_RUNNING)
2844 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2846 struct blk_mq_hw_ctx *hctx;
2847 struct blk_plug *plug;
2850 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2851 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2854 plug = current->plug;
2856 blk_flush_plug_list(plug, false);
2858 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2859 if (!blk_qc_t_is_internal(cookie))
2860 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2862 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2864 return __blk_mq_poll(hctx, rq);
2866 EXPORT_SYMBOL_GPL(blk_mq_poll);
2868 void blk_mq_disable_hotplug(void)
2870 mutex_lock(&all_q_mutex);
2873 void blk_mq_enable_hotplug(void)
2875 mutex_unlock(&all_q_mutex);
2878 static int __init blk_mq_init(void)
2880 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2881 blk_mq_hctx_notify_dead);
2883 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2884 blk_mq_queue_reinit_prepare,
2885 blk_mq_queue_reinit_dead);
2888 subsys_initcall(blk_mq_init);