2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-debugfs.h"
35 #include "blk-mq-tag.h"
38 #include "blk-mq-sched.h"
40 static void blk_mq_poll_stats_start(struct request_queue *q);
41 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
43 static int blk_mq_poll_stats_bkt(const struct request *rq)
45 int ddir, bytes, bucket;
47 ddir = rq_data_dir(rq);
48 bytes = blk_rq_bytes(rq);
50 bucket = ddir + 2*(ilog2(bytes) - 9);
54 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
55 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
61 * Check if any of the ctx's have pending work in this hardware queue
63 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
65 return sbitmap_any_bit_set(&hctx->ctx_map) ||
66 !list_empty_careful(&hctx->dispatch) ||
67 blk_mq_sched_has_work(hctx);
71 * Mark this ctx as having pending work in this hardware queue
73 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
74 struct blk_mq_ctx *ctx)
76 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
77 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
80 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
81 struct blk_mq_ctx *ctx)
83 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
87 struct hd_struct *part;
88 unsigned int *inflight;
91 static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
92 struct request *rq, void *priv,
95 struct mq_inflight *mi = priv;
97 if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) &&
98 !test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) {
100 * index[0] counts the specific partition that was asked
101 * for. index[1] counts the ones that are active on the
102 * whole device, so increment that if mi->part is indeed
103 * a partition, and not a whole device.
105 if (rq->part == mi->part)
107 if (mi->part->partno)
112 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
113 unsigned int inflight[2])
115 struct mq_inflight mi = { .part = part, .inflight = inflight, };
117 inflight[0] = inflight[1] = 0;
118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
121 void blk_freeze_queue_start(struct request_queue *q)
125 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
126 if (freeze_depth == 1) {
127 percpu_ref_kill(&q->q_usage_counter);
128 blk_mq_run_hw_queues(q, false);
131 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
133 void blk_mq_freeze_queue_wait(struct request_queue *q)
135 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
137 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
139 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
140 unsigned long timeout)
142 return wait_event_timeout(q->mq_freeze_wq,
143 percpu_ref_is_zero(&q->q_usage_counter),
146 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
149 * Guarantee no request is in use, so we can change any data structure of
150 * the queue afterward.
152 void blk_freeze_queue(struct request_queue *q)
155 * In the !blk_mq case we are only calling this to kill the
156 * q_usage_counter, otherwise this increases the freeze depth
157 * and waits for it to return to zero. For this reason there is
158 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
159 * exported to drivers as the only user for unfreeze is blk_mq.
161 blk_freeze_queue_start(q);
164 blk_mq_freeze_queue_wait(q);
167 void blk_mq_freeze_queue(struct request_queue *q)
170 * ...just an alias to keep freeze and unfreeze actions balanced
171 * in the blk_mq_* namespace
175 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
177 void blk_mq_unfreeze_queue(struct request_queue *q)
181 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
182 WARN_ON_ONCE(freeze_depth < 0);
184 percpu_ref_reinit(&q->q_usage_counter);
185 wake_up_all(&q->mq_freeze_wq);
188 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
191 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
192 * mpt3sas driver such that this function can be removed.
194 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
198 spin_lock_irqsave(q->queue_lock, flags);
199 queue_flag_set(QUEUE_FLAG_QUIESCED, q);
200 spin_unlock_irqrestore(q->queue_lock, flags);
202 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
205 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
208 * Note: this function does not prevent that the struct request end_io()
209 * callback function is invoked. Once this function is returned, we make
210 * sure no dispatch can happen until the queue is unquiesced via
211 * blk_mq_unquiesce_queue().
213 void blk_mq_quiesce_queue(struct request_queue *q)
215 struct blk_mq_hw_ctx *hctx;
219 blk_mq_quiesce_queue_nowait(q);
221 queue_for_each_hw_ctx(q, hctx, i) {
222 if (hctx->flags & BLK_MQ_F_BLOCKING)
223 synchronize_srcu(hctx->queue_rq_srcu);
230 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
233 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
236 * This function recovers queue into the state before quiescing
237 * which is done by blk_mq_quiesce_queue.
239 void blk_mq_unquiesce_queue(struct request_queue *q)
243 spin_lock_irqsave(q->queue_lock, flags);
244 queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
245 spin_unlock_irqrestore(q->queue_lock, flags);
247 /* dispatch requests which are inserted during quiescing */
248 blk_mq_run_hw_queues(q, true);
250 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
252 void blk_mq_wake_waiters(struct request_queue *q)
254 struct blk_mq_hw_ctx *hctx;
257 queue_for_each_hw_ctx(q, hctx, i)
258 if (blk_mq_hw_queue_mapped(hctx))
259 blk_mq_tag_wakeup_all(hctx->tags, true);
262 * If we are called because the queue has now been marked as
263 * dying, we need to ensure that processes currently waiting on
264 * the queue are notified as well.
266 wake_up_all(&q->mq_freeze_wq);
269 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
271 return blk_mq_has_free_tags(hctx->tags);
273 EXPORT_SYMBOL(blk_mq_can_queue);
275 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
276 unsigned int tag, unsigned int op)
278 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
279 struct request *rq = tags->static_rqs[tag];
283 if (data->flags & BLK_MQ_REQ_INTERNAL) {
285 rq->internal_tag = tag;
287 if (blk_mq_tag_busy(data->hctx)) {
288 rq->rq_flags = RQF_MQ_INFLIGHT;
289 atomic_inc(&data->hctx->nr_active);
292 rq->internal_tag = -1;
293 data->hctx->tags->rqs[rq->tag] = rq;
296 INIT_LIST_HEAD(&rq->queuelist);
297 /* csd/requeue_work/fifo_time is initialized before use */
299 rq->mq_ctx = data->ctx;
301 if (blk_queue_io_stat(data->q))
302 rq->rq_flags |= RQF_IO_STAT;
303 /* do not touch atomic flags, it needs atomic ops against the timer */
305 INIT_HLIST_NODE(&rq->hash);
306 RB_CLEAR_NODE(&rq->rb_node);
309 rq->start_time = jiffies;
310 #ifdef CONFIG_BLK_CGROUP
312 set_start_time_ns(rq);
313 rq->io_start_time_ns = 0;
315 rq->nr_phys_segments = 0;
316 #if defined(CONFIG_BLK_DEV_INTEGRITY)
317 rq->nr_integrity_segments = 0;
320 /* tag was already set */
323 INIT_LIST_HEAD(&rq->timeout_list);
327 rq->end_io_data = NULL;
330 data->ctx->rq_dispatched[op_is_sync(op)]++;
334 static struct request *blk_mq_get_request(struct request_queue *q,
335 struct bio *bio, unsigned int op,
336 struct blk_mq_alloc_data *data)
338 struct elevator_queue *e = q->elevator;
341 struct blk_mq_ctx *local_ctx = NULL;
343 blk_queue_enter_live(q);
345 if (likely(!data->ctx))
346 data->ctx = local_ctx = blk_mq_get_ctx(q);
347 if (likely(!data->hctx))
348 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
350 data->flags |= BLK_MQ_REQ_NOWAIT;
353 data->flags |= BLK_MQ_REQ_INTERNAL;
356 * Flush requests are special and go directly to the
359 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
360 e->type->ops.mq.limit_depth(op, data);
363 tag = blk_mq_get_tag(data);
364 if (tag == BLK_MQ_TAG_FAIL) {
366 blk_mq_put_ctx(local_ctx);
373 rq = blk_mq_rq_ctx_init(data, tag, op);
374 if (!op_is_flush(op)) {
376 if (e && e->type->ops.mq.prepare_request) {
377 if (e->type->icq_cache && rq_ioc(bio))
378 blk_mq_sched_assign_ioc(rq, bio);
380 e->type->ops.mq.prepare_request(rq, bio);
381 rq->rq_flags |= RQF_ELVPRIV;
384 data->hctx->queued++;
388 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
391 struct blk_mq_alloc_data alloc_data = { .flags = flags };
395 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
399 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
403 return ERR_PTR(-EWOULDBLOCK);
405 blk_mq_put_ctx(alloc_data.ctx);
408 rq->__sector = (sector_t) -1;
409 rq->bio = rq->biotail = NULL;
412 EXPORT_SYMBOL(blk_mq_alloc_request);
414 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
415 unsigned int op, unsigned int flags, unsigned int hctx_idx)
417 struct blk_mq_alloc_data alloc_data = { .flags = flags };
423 * If the tag allocator sleeps we could get an allocation for a
424 * different hardware context. No need to complicate the low level
425 * allocator for this for the rare use case of a command tied to
428 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
429 return ERR_PTR(-EINVAL);
431 if (hctx_idx >= q->nr_hw_queues)
432 return ERR_PTR(-EIO);
434 ret = blk_queue_enter(q, true);
439 * Check if the hardware context is actually mapped to anything.
440 * If not tell the caller that it should skip this queue.
442 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
443 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
445 return ERR_PTR(-EXDEV);
447 cpu = cpumask_first(alloc_data.hctx->cpumask);
448 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
450 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
454 return ERR_PTR(-EWOULDBLOCK);
458 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
460 void blk_mq_free_request(struct request *rq)
462 struct request_queue *q = rq->q;
463 struct elevator_queue *e = q->elevator;
464 struct blk_mq_ctx *ctx = rq->mq_ctx;
465 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
466 const int sched_tag = rq->internal_tag;
468 if (rq->rq_flags & RQF_ELVPRIV) {
469 if (e && e->type->ops.mq.finish_request)
470 e->type->ops.mq.finish_request(rq);
472 put_io_context(rq->elv.icq->ioc);
477 ctx->rq_completed[rq_is_sync(rq)]++;
478 if (rq->rq_flags & RQF_MQ_INFLIGHT)
479 atomic_dec(&hctx->nr_active);
481 wbt_done(q->rq_wb, &rq->issue_stat);
483 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
484 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
486 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
488 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
489 blk_mq_sched_restart(hctx);
492 EXPORT_SYMBOL_GPL(blk_mq_free_request);
494 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
496 blk_account_io_done(rq);
499 wbt_done(rq->q->rq_wb, &rq->issue_stat);
500 rq->end_io(rq, error);
502 if (unlikely(blk_bidi_rq(rq)))
503 blk_mq_free_request(rq->next_rq);
504 blk_mq_free_request(rq);
507 EXPORT_SYMBOL(__blk_mq_end_request);
509 void blk_mq_end_request(struct request *rq, blk_status_t error)
511 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
513 __blk_mq_end_request(rq, error);
515 EXPORT_SYMBOL(blk_mq_end_request);
517 static void __blk_mq_complete_request_remote(void *data)
519 struct request *rq = data;
521 rq->q->softirq_done_fn(rq);
524 static void __blk_mq_complete_request(struct request *rq)
526 struct blk_mq_ctx *ctx = rq->mq_ctx;
530 if (rq->internal_tag != -1)
531 blk_mq_sched_completed_request(rq);
532 if (rq->rq_flags & RQF_STATS) {
533 blk_mq_poll_stats_start(rq->q);
537 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
538 rq->q->softirq_done_fn(rq);
543 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
544 shared = cpus_share_cache(cpu, ctx->cpu);
546 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
547 rq->csd.func = __blk_mq_complete_request_remote;
550 smp_call_function_single_async(ctx->cpu, &rq->csd);
552 rq->q->softirq_done_fn(rq);
558 * blk_mq_complete_request - end I/O on a request
559 * @rq: the request being processed
562 * Ends all I/O on a request. It does not handle partial completions.
563 * The actual completion happens out-of-order, through a IPI handler.
565 void blk_mq_complete_request(struct request *rq)
567 struct request_queue *q = rq->q;
569 if (unlikely(blk_should_fake_timeout(q)))
571 if (!blk_mark_rq_complete(rq))
572 __blk_mq_complete_request(rq);
574 EXPORT_SYMBOL(blk_mq_complete_request);
576 int blk_mq_request_started(struct request *rq)
578 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
580 EXPORT_SYMBOL_GPL(blk_mq_request_started);
582 void blk_mq_start_request(struct request *rq)
584 struct request_queue *q = rq->q;
586 blk_mq_sched_started_request(rq);
588 trace_block_rq_issue(q, rq);
590 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
591 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
592 rq->rq_flags |= RQF_STATS;
593 wbt_issue(q->rq_wb, &rq->issue_stat);
599 * Ensure that ->deadline is visible before set the started
600 * flag and clear the completed flag.
602 smp_mb__before_atomic();
605 * Mark us as started and clear complete. Complete might have been
606 * set if requeue raced with timeout, which then marked it as
607 * complete. So be sure to clear complete again when we start
608 * the request, otherwise we'll ignore the completion event.
610 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
611 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
612 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
613 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
615 if (q->dma_drain_size && blk_rq_bytes(rq)) {
617 * Make sure space for the drain appears. We know we can do
618 * this because max_hw_segments has been adjusted to be one
619 * fewer than the device can handle.
621 rq->nr_phys_segments++;
624 EXPORT_SYMBOL(blk_mq_start_request);
627 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
628 * flag isn't set yet, so there may be race with timeout handler,
629 * but given rq->deadline is just set in .queue_rq() under
630 * this situation, the race won't be possible in reality because
631 * rq->timeout should be set as big enough to cover the window
632 * between blk_mq_start_request() called from .queue_rq() and
633 * clearing REQ_ATOM_STARTED here.
635 static void __blk_mq_requeue_request(struct request *rq)
637 struct request_queue *q = rq->q;
639 trace_block_rq_requeue(q, rq);
640 wbt_requeue(q->rq_wb, &rq->issue_stat);
642 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
643 if (q->dma_drain_size && blk_rq_bytes(rq))
644 rq->nr_phys_segments--;
648 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
650 __blk_mq_requeue_request(rq);
652 /* this request will be re-inserted to io scheduler queue */
653 blk_mq_sched_requeue_request(rq);
655 BUG_ON(blk_queued_rq(rq));
656 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
658 EXPORT_SYMBOL(blk_mq_requeue_request);
660 static void blk_mq_requeue_work(struct work_struct *work)
662 struct request_queue *q =
663 container_of(work, struct request_queue, requeue_work.work);
665 struct request *rq, *next;
667 spin_lock_irq(&q->requeue_lock);
668 list_splice_init(&q->requeue_list, &rq_list);
669 spin_unlock_irq(&q->requeue_lock);
671 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
672 if (!(rq->rq_flags & RQF_SOFTBARRIER))
675 rq->rq_flags &= ~RQF_SOFTBARRIER;
676 list_del_init(&rq->queuelist);
677 blk_mq_sched_insert_request(rq, true, false, false, true);
680 while (!list_empty(&rq_list)) {
681 rq = list_entry(rq_list.next, struct request, queuelist);
682 list_del_init(&rq->queuelist);
683 blk_mq_sched_insert_request(rq, false, false, false, true);
686 blk_mq_run_hw_queues(q, false);
689 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
690 bool kick_requeue_list)
692 struct request_queue *q = rq->q;
696 * We abuse this flag that is otherwise used by the I/O scheduler to
697 * request head insertation from the workqueue.
699 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
701 spin_lock_irqsave(&q->requeue_lock, flags);
703 rq->rq_flags |= RQF_SOFTBARRIER;
704 list_add(&rq->queuelist, &q->requeue_list);
706 list_add_tail(&rq->queuelist, &q->requeue_list);
708 spin_unlock_irqrestore(&q->requeue_lock, flags);
710 if (kick_requeue_list)
711 blk_mq_kick_requeue_list(q);
713 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
715 void blk_mq_kick_requeue_list(struct request_queue *q)
717 kblockd_schedule_delayed_work(&q->requeue_work, 0);
719 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
721 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
724 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
725 msecs_to_jiffies(msecs));
727 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
729 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
731 if (tag < tags->nr_tags) {
732 prefetch(tags->rqs[tag]);
733 return tags->rqs[tag];
738 EXPORT_SYMBOL(blk_mq_tag_to_rq);
740 struct blk_mq_timeout_data {
742 unsigned int next_set;
745 void blk_mq_rq_timed_out(struct request *req, bool reserved)
747 const struct blk_mq_ops *ops = req->q->mq_ops;
748 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
751 * We know that complete is set at this point. If STARTED isn't set
752 * anymore, then the request isn't active and the "timeout" should
753 * just be ignored. This can happen due to the bitflag ordering.
754 * Timeout first checks if STARTED is set, and if it is, assumes
755 * the request is active. But if we race with completion, then
756 * both flags will get cleared. So check here again, and ignore
757 * a timeout event with a request that isn't active.
759 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
763 ret = ops->timeout(req, reserved);
767 __blk_mq_complete_request(req);
769 case BLK_EH_RESET_TIMER:
771 blk_clear_rq_complete(req);
773 case BLK_EH_NOT_HANDLED:
776 printk(KERN_ERR "block: bad eh return: %d\n", ret);
781 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
782 struct request *rq, void *priv, bool reserved)
784 struct blk_mq_timeout_data *data = priv;
786 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
790 * The rq being checked may have been freed and reallocated
791 * out already here, we avoid this race by checking rq->deadline
792 * and REQ_ATOM_COMPLETE flag together:
794 * - if rq->deadline is observed as new value because of
795 * reusing, the rq won't be timed out because of timing.
796 * - if rq->deadline is observed as previous value,
797 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
798 * because we put a barrier between setting rq->deadline
799 * and clearing the flag in blk_mq_start_request(), so
800 * this rq won't be timed out too.
802 if (time_after_eq(jiffies, rq->deadline)) {
803 if (!blk_mark_rq_complete(rq))
804 blk_mq_rq_timed_out(rq, reserved);
805 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
806 data->next = rq->deadline;
811 static void blk_mq_timeout_work(struct work_struct *work)
813 struct request_queue *q =
814 container_of(work, struct request_queue, timeout_work);
815 struct blk_mq_timeout_data data = {
821 /* A deadlock might occur if a request is stuck requiring a
822 * timeout at the same time a queue freeze is waiting
823 * completion, since the timeout code would not be able to
824 * acquire the queue reference here.
826 * That's why we don't use blk_queue_enter here; instead, we use
827 * percpu_ref_tryget directly, because we need to be able to
828 * obtain a reference even in the short window between the queue
829 * starting to freeze, by dropping the first reference in
830 * blk_freeze_queue_start, and the moment the last request is
831 * consumed, marked by the instant q_usage_counter reaches
834 if (!percpu_ref_tryget(&q->q_usage_counter))
837 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
840 data.next = blk_rq_timeout(round_jiffies_up(data.next));
841 mod_timer(&q->timeout, data.next);
843 struct blk_mq_hw_ctx *hctx;
845 queue_for_each_hw_ctx(q, hctx, i) {
846 /* the hctx may be unmapped, so check it here */
847 if (blk_mq_hw_queue_mapped(hctx))
848 blk_mq_tag_idle(hctx);
854 struct flush_busy_ctx_data {
855 struct blk_mq_hw_ctx *hctx;
856 struct list_head *list;
859 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
861 struct flush_busy_ctx_data *flush_data = data;
862 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
863 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
865 sbitmap_clear_bit(sb, bitnr);
866 spin_lock(&ctx->lock);
867 list_splice_tail_init(&ctx->rq_list, flush_data->list);
868 spin_unlock(&ctx->lock);
873 * Process software queues that have been marked busy, splicing them
874 * to the for-dispatch
876 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
878 struct flush_busy_ctx_data data = {
883 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
885 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
887 static inline unsigned int queued_to_index(unsigned int queued)
892 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
895 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
898 struct blk_mq_alloc_data data = {
900 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
901 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
904 might_sleep_if(wait);
909 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
910 data.flags |= BLK_MQ_REQ_RESERVED;
912 rq->tag = blk_mq_get_tag(&data);
914 if (blk_mq_tag_busy(data.hctx)) {
915 rq->rq_flags |= RQF_MQ_INFLIGHT;
916 atomic_inc(&data.hctx->nr_active);
918 data.hctx->tags->rqs[rq->tag] = rq;
924 return rq->tag != -1;
927 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
930 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
933 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
934 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
935 atomic_dec(&hctx->nr_active);
939 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
942 if (rq->tag == -1 || rq->internal_tag == -1)
945 __blk_mq_put_driver_tag(hctx, rq);
948 static void blk_mq_put_driver_tag(struct request *rq)
950 struct blk_mq_hw_ctx *hctx;
952 if (rq->tag == -1 || rq->internal_tag == -1)
955 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
956 __blk_mq_put_driver_tag(hctx, rq);
960 * If we fail getting a driver tag because all the driver tags are already
961 * assigned and on the dispatch list, BUT the first entry does not have a
962 * tag, then we could deadlock. For that case, move entries with assigned
963 * driver tags to the front, leaving the set of tagged requests in the
964 * same order, and the untagged set in the same order.
966 static bool reorder_tags_to_front(struct list_head *list)
968 struct request *rq, *tmp, *first = NULL;
970 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
974 list_move(&rq->queuelist, list);
980 return first != NULL;
983 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
986 struct blk_mq_hw_ctx *hctx;
988 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
990 list_del(&wait->entry);
991 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
992 blk_mq_run_hw_queue(hctx, true);
996 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
998 struct sbq_wait_state *ws;
1001 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
1002 * The thread which wins the race to grab this bit adds the hardware
1003 * queue to the wait queue.
1005 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
1006 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
1009 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
1010 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
1013 * As soon as this returns, it's no longer safe to fiddle with
1014 * hctx->dispatch_wait, since a completion can wake up the wait queue
1015 * and unlock the bit.
1017 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
1021 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
1023 struct blk_mq_hw_ctx *hctx;
1027 if (list_empty(list))
1031 * Now process all the entries, sending them to the driver.
1033 errors = queued = 0;
1035 struct blk_mq_queue_data bd;
1038 rq = list_first_entry(list, struct request, queuelist);
1039 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1040 if (!queued && reorder_tags_to_front(list))
1044 * The initial allocation attempt failed, so we need to
1045 * rerun the hardware queue when a tag is freed.
1047 if (!blk_mq_dispatch_wait_add(hctx))
1051 * It's possible that a tag was freed in the window
1052 * between the allocation failure and adding the
1053 * hardware queue to the wait queue.
1055 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1059 list_del_init(&rq->queuelist);
1064 * Flag last if we have no more requests, or if we have more
1065 * but can't assign a driver tag to it.
1067 if (list_empty(list))
1070 struct request *nxt;
1072 nxt = list_first_entry(list, struct request, queuelist);
1073 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1076 ret = q->mq_ops->queue_rq(hctx, &bd);
1077 if (ret == BLK_STS_RESOURCE) {
1078 blk_mq_put_driver_tag_hctx(hctx, rq);
1079 list_add(&rq->queuelist, list);
1080 __blk_mq_requeue_request(rq);
1084 if (unlikely(ret != BLK_STS_OK)) {
1086 blk_mq_end_request(rq, BLK_STS_IOERR);
1091 } while (!list_empty(list));
1093 hctx->dispatched[queued_to_index(queued)]++;
1096 * Any items that need requeuing? Stuff them into hctx->dispatch,
1097 * that is where we will continue on next queue run.
1099 if (!list_empty(list)) {
1101 * If an I/O scheduler has been configured and we got a driver
1102 * tag for the next request already, free it again.
1104 rq = list_first_entry(list, struct request, queuelist);
1105 blk_mq_put_driver_tag(rq);
1107 spin_lock(&hctx->lock);
1108 list_splice_init(list, &hctx->dispatch);
1109 spin_unlock(&hctx->lock);
1112 * If SCHED_RESTART was set by the caller of this function and
1113 * it is no longer set that means that it was cleared by another
1114 * thread and hence that a queue rerun is needed.
1116 * If TAG_WAITING is set that means that an I/O scheduler has
1117 * been configured and another thread is waiting for a driver
1118 * tag. To guarantee fairness, do not rerun this hardware queue
1119 * but let the other thread grab the driver tag.
1121 * If no I/O scheduler has been configured it is possible that
1122 * the hardware queue got stopped and restarted before requests
1123 * were pushed back onto the dispatch list. Rerun the queue to
1124 * avoid starvation. Notes:
1125 * - blk_mq_run_hw_queue() checks whether or not a queue has
1126 * been stopped before rerunning a queue.
1127 * - Some but not all block drivers stop a queue before
1128 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1131 if (!blk_mq_sched_needs_restart(hctx) &&
1132 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1133 blk_mq_run_hw_queue(hctx, true);
1136 return (queued + errors) != 0;
1139 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1144 * We should be running this queue from one of the CPUs that
1147 * There are at least two related races now between setting
1148 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1149 * __blk_mq_run_hw_queue():
1151 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1152 * but later it becomes online, then this warning is harmless
1155 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1156 * but later it becomes offline, then the warning can't be
1157 * triggered, and we depend on blk-mq timeout handler to
1158 * handle dispatched requests to this hctx
1160 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1161 cpu_online(hctx->next_cpu)) {
1162 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1163 raw_smp_processor_id(),
1164 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1169 * We can't run the queue inline with ints disabled. Ensure that
1170 * we catch bad users of this early.
1172 WARN_ON_ONCE(in_interrupt());
1174 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1176 blk_mq_sched_dispatch_requests(hctx);
1181 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1182 blk_mq_sched_dispatch_requests(hctx);
1183 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1188 * It'd be great if the workqueue API had a way to pass
1189 * in a mask and had some smarts for more clever placement.
1190 * For now we just round-robin here, switching for every
1191 * BLK_MQ_CPU_WORK_BATCH queued items.
1193 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1195 if (hctx->queue->nr_hw_queues == 1)
1196 return WORK_CPU_UNBOUND;
1198 if (--hctx->next_cpu_batch <= 0) {
1201 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1202 if (next_cpu >= nr_cpu_ids)
1203 next_cpu = cpumask_first(hctx->cpumask);
1205 hctx->next_cpu = next_cpu;
1206 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1209 return hctx->next_cpu;
1212 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1213 unsigned long msecs)
1215 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1218 if (unlikely(blk_mq_hctx_stopped(hctx)))
1221 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1222 int cpu = get_cpu();
1223 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1224 __blk_mq_run_hw_queue(hctx);
1232 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1234 msecs_to_jiffies(msecs));
1237 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1239 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1241 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1243 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1245 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1247 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1249 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1251 struct blk_mq_hw_ctx *hctx;
1254 queue_for_each_hw_ctx(q, hctx, i) {
1255 if (!blk_mq_hctx_has_pending(hctx) ||
1256 blk_mq_hctx_stopped(hctx))
1259 blk_mq_run_hw_queue(hctx, async);
1262 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1265 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1266 * @q: request queue.
1268 * The caller is responsible for serializing this function against
1269 * blk_mq_{start,stop}_hw_queue().
1271 bool blk_mq_queue_stopped(struct request_queue *q)
1273 struct blk_mq_hw_ctx *hctx;
1276 queue_for_each_hw_ctx(q, hctx, i)
1277 if (blk_mq_hctx_stopped(hctx))
1282 EXPORT_SYMBOL(blk_mq_queue_stopped);
1285 * This function is often used for pausing .queue_rq() by driver when
1286 * there isn't enough resource or some conditions aren't satisfied, and
1287 * BLK_STS_RESOURCE is usually returned.
1289 * We do not guarantee that dispatch can be drained or blocked
1290 * after blk_mq_stop_hw_queue() returns. Please use
1291 * blk_mq_quiesce_queue() for that requirement.
1293 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1295 cancel_delayed_work(&hctx->run_work);
1297 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1299 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1302 * This function is often used for pausing .queue_rq() by driver when
1303 * there isn't enough resource or some conditions aren't satisfied, and
1304 * BLK_STS_RESOURCE is usually returned.
1306 * We do not guarantee that dispatch can be drained or blocked
1307 * after blk_mq_stop_hw_queues() returns. Please use
1308 * blk_mq_quiesce_queue() for that requirement.
1310 void blk_mq_stop_hw_queues(struct request_queue *q)
1312 struct blk_mq_hw_ctx *hctx;
1315 queue_for_each_hw_ctx(q, hctx, i)
1316 blk_mq_stop_hw_queue(hctx);
1318 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1320 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1322 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1324 blk_mq_run_hw_queue(hctx, false);
1326 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1328 void blk_mq_start_hw_queues(struct request_queue *q)
1330 struct blk_mq_hw_ctx *hctx;
1333 queue_for_each_hw_ctx(q, hctx, i)
1334 blk_mq_start_hw_queue(hctx);
1336 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1338 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1340 if (!blk_mq_hctx_stopped(hctx))
1343 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1344 blk_mq_run_hw_queue(hctx, async);
1346 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1348 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1350 struct blk_mq_hw_ctx *hctx;
1353 queue_for_each_hw_ctx(q, hctx, i)
1354 blk_mq_start_stopped_hw_queue(hctx, async);
1356 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1358 static void blk_mq_run_work_fn(struct work_struct *work)
1360 struct blk_mq_hw_ctx *hctx;
1362 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1365 * If we are stopped, don't run the queue. The exception is if
1366 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1367 * the STOPPED bit and run it.
1369 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1370 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1373 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1374 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1377 __blk_mq_run_hw_queue(hctx);
1381 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1383 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1387 * Stop the hw queue, then modify currently delayed work.
1388 * This should prevent us from running the queue prematurely.
1389 * Mark the queue as auto-clearing STOPPED when it runs.
1391 blk_mq_stop_hw_queue(hctx);
1392 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1393 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1395 msecs_to_jiffies(msecs));
1397 EXPORT_SYMBOL(blk_mq_delay_queue);
1399 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1403 struct blk_mq_ctx *ctx = rq->mq_ctx;
1405 lockdep_assert_held(&ctx->lock);
1407 trace_block_rq_insert(hctx->queue, rq);
1410 list_add(&rq->queuelist, &ctx->rq_list);
1412 list_add_tail(&rq->queuelist, &ctx->rq_list);
1415 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1418 struct blk_mq_ctx *ctx = rq->mq_ctx;
1420 lockdep_assert_held(&ctx->lock);
1422 __blk_mq_insert_req_list(hctx, rq, at_head);
1423 blk_mq_hctx_mark_pending(hctx, ctx);
1427 * Should only be used carefully, when the caller knows we want to
1428 * bypass a potential IO scheduler on the target device.
1430 void blk_mq_request_bypass_insert(struct request *rq)
1432 struct blk_mq_ctx *ctx = rq->mq_ctx;
1433 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1435 spin_lock(&hctx->lock);
1436 list_add_tail(&rq->queuelist, &hctx->dispatch);
1437 spin_unlock(&hctx->lock);
1439 blk_mq_run_hw_queue(hctx, false);
1442 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1443 struct list_head *list)
1447 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1450 spin_lock(&ctx->lock);
1451 while (!list_empty(list)) {
1454 rq = list_first_entry(list, struct request, queuelist);
1455 BUG_ON(rq->mq_ctx != ctx);
1456 list_del_init(&rq->queuelist);
1457 __blk_mq_insert_req_list(hctx, rq, false);
1459 blk_mq_hctx_mark_pending(hctx, ctx);
1460 spin_unlock(&ctx->lock);
1463 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1465 struct request *rqa = container_of(a, struct request, queuelist);
1466 struct request *rqb = container_of(b, struct request, queuelist);
1468 return !(rqa->mq_ctx < rqb->mq_ctx ||
1469 (rqa->mq_ctx == rqb->mq_ctx &&
1470 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1473 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1475 struct blk_mq_ctx *this_ctx;
1476 struct request_queue *this_q;
1479 LIST_HEAD(ctx_list);
1482 list_splice_init(&plug->mq_list, &list);
1484 list_sort(NULL, &list, plug_ctx_cmp);
1490 while (!list_empty(&list)) {
1491 rq = list_entry_rq(list.next);
1492 list_del_init(&rq->queuelist);
1494 if (rq->mq_ctx != this_ctx) {
1496 trace_block_unplug(this_q, depth, from_schedule);
1497 blk_mq_sched_insert_requests(this_q, this_ctx,
1502 this_ctx = rq->mq_ctx;
1508 list_add_tail(&rq->queuelist, &ctx_list);
1512 * If 'this_ctx' is set, we know we have entries to complete
1513 * on 'ctx_list'. Do those.
1516 trace_block_unplug(this_q, depth, from_schedule);
1517 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1522 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1524 blk_init_request_from_bio(rq, bio);
1526 blk_account_io_start(rq, true);
1529 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1531 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1532 !blk_queue_nomerges(hctx->queue);
1535 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1536 struct blk_mq_ctx *ctx,
1539 spin_lock(&ctx->lock);
1540 __blk_mq_insert_request(hctx, rq, false);
1541 spin_unlock(&ctx->lock);
1544 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1547 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1549 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1552 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1554 blk_qc_t *cookie, bool may_sleep)
1556 struct request_queue *q = rq->q;
1557 struct blk_mq_queue_data bd = {
1561 blk_qc_t new_cookie;
1563 bool run_queue = true;
1565 /* RCU or SRCU read lock is needed before checking quiesced flag */
1566 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1574 if (!blk_mq_get_driver_tag(rq, NULL, false))
1577 new_cookie = request_to_qc_t(hctx, rq);
1580 * For OK queue, we are done. For error, kill it. Any other
1581 * error (busy), just add it to our list as we previously
1584 ret = q->mq_ops->queue_rq(hctx, &bd);
1587 *cookie = new_cookie;
1589 case BLK_STS_RESOURCE:
1590 __blk_mq_requeue_request(rq);
1593 *cookie = BLK_QC_T_NONE;
1594 blk_mq_end_request(rq, ret);
1599 blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1602 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1603 struct request *rq, blk_qc_t *cookie)
1605 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1607 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1610 unsigned int srcu_idx;
1614 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1615 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1616 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1620 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1622 const int is_sync = op_is_sync(bio->bi_opf);
1623 const int is_flush_fua = op_is_flush(bio->bi_opf);
1624 struct blk_mq_alloc_data data = { .flags = 0 };
1626 unsigned int request_count = 0;
1627 struct blk_plug *plug;
1628 struct request *same_queue_rq = NULL;
1630 unsigned int wb_acct;
1632 blk_queue_bounce(q, &bio);
1634 blk_queue_split(q, &bio);
1636 if (!bio_integrity_prep(bio))
1637 return BLK_QC_T_NONE;
1639 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1640 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1641 return BLK_QC_T_NONE;
1643 if (blk_mq_sched_bio_merge(q, bio))
1644 return BLK_QC_T_NONE;
1646 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1648 trace_block_getrq(q, bio, bio->bi_opf);
1650 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1651 if (unlikely(!rq)) {
1652 __wbt_done(q->rq_wb, wb_acct);
1653 if (bio->bi_opf & REQ_NOWAIT)
1654 bio_wouldblock_error(bio);
1655 return BLK_QC_T_NONE;
1658 wbt_track(&rq->issue_stat, wb_acct);
1660 cookie = request_to_qc_t(data.hctx, rq);
1662 plug = current->plug;
1663 if (unlikely(is_flush_fua)) {
1664 blk_mq_put_ctx(data.ctx);
1665 blk_mq_bio_to_request(rq, bio);
1667 blk_mq_sched_insert_request(rq, false, true, true,
1670 blk_insert_flush(rq);
1671 blk_mq_run_hw_queue(data.hctx, true);
1673 } else if (plug && q->nr_hw_queues == 1) {
1674 struct request *last = NULL;
1676 blk_mq_put_ctx(data.ctx);
1677 blk_mq_bio_to_request(rq, bio);
1680 * @request_count may become stale because of schedule
1681 * out, so check the list again.
1683 if (list_empty(&plug->mq_list))
1685 else if (blk_queue_nomerges(q))
1686 request_count = blk_plug_queued_count(q);
1689 trace_block_plug(q);
1691 last = list_entry_rq(plug->mq_list.prev);
1693 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1694 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1695 blk_flush_plug_list(plug, false);
1696 trace_block_plug(q);
1699 list_add_tail(&rq->queuelist, &plug->mq_list);
1700 } else if (plug && !blk_queue_nomerges(q)) {
1701 blk_mq_bio_to_request(rq, bio);
1704 * We do limited plugging. If the bio can be merged, do that.
1705 * Otherwise the existing request in the plug list will be
1706 * issued. So the plug list will have one request at most
1707 * The plug list might get flushed before this. If that happens,
1708 * the plug list is empty, and same_queue_rq is invalid.
1710 if (list_empty(&plug->mq_list))
1711 same_queue_rq = NULL;
1713 list_del_init(&same_queue_rq->queuelist);
1714 list_add_tail(&rq->queuelist, &plug->mq_list);
1716 blk_mq_put_ctx(data.ctx);
1718 if (same_queue_rq) {
1719 data.hctx = blk_mq_map_queue(q,
1720 same_queue_rq->mq_ctx->cpu);
1721 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1724 } else if (q->nr_hw_queues > 1 && is_sync) {
1725 blk_mq_put_ctx(data.ctx);
1726 blk_mq_bio_to_request(rq, bio);
1727 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1728 } else if (q->elevator) {
1729 blk_mq_put_ctx(data.ctx);
1730 blk_mq_bio_to_request(rq, bio);
1731 blk_mq_sched_insert_request(rq, false, true, true, true);
1733 blk_mq_put_ctx(data.ctx);
1734 blk_mq_bio_to_request(rq, bio);
1735 blk_mq_queue_io(data.hctx, data.ctx, rq);
1736 blk_mq_run_hw_queue(data.hctx, true);
1742 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1743 unsigned int hctx_idx)
1747 if (tags->rqs && set->ops->exit_request) {
1750 for (i = 0; i < tags->nr_tags; i++) {
1751 struct request *rq = tags->static_rqs[i];
1755 set->ops->exit_request(set, rq, hctx_idx);
1756 tags->static_rqs[i] = NULL;
1760 while (!list_empty(&tags->page_list)) {
1761 page = list_first_entry(&tags->page_list, struct page, lru);
1762 list_del_init(&page->lru);
1764 * Remove kmemleak object previously allocated in
1765 * blk_mq_init_rq_map().
1767 kmemleak_free(page_address(page));
1768 __free_pages(page, page->private);
1772 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1776 kfree(tags->static_rqs);
1777 tags->static_rqs = NULL;
1779 blk_mq_free_tags(tags);
1782 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1783 unsigned int hctx_idx,
1784 unsigned int nr_tags,
1785 unsigned int reserved_tags)
1787 struct blk_mq_tags *tags;
1790 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1791 if (node == NUMA_NO_NODE)
1792 node = set->numa_node;
1794 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1795 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1799 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1800 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1803 blk_mq_free_tags(tags);
1807 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1808 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1810 if (!tags->static_rqs) {
1812 blk_mq_free_tags(tags);
1819 static size_t order_to_size(unsigned int order)
1821 return (size_t)PAGE_SIZE << order;
1824 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1825 unsigned int hctx_idx, unsigned int depth)
1827 unsigned int i, j, entries_per_page, max_order = 4;
1828 size_t rq_size, left;
1831 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1832 if (node == NUMA_NO_NODE)
1833 node = set->numa_node;
1835 INIT_LIST_HEAD(&tags->page_list);
1838 * rq_size is the size of the request plus driver payload, rounded
1839 * to the cacheline size
1841 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1843 left = rq_size * depth;
1845 for (i = 0; i < depth; ) {
1846 int this_order = max_order;
1851 while (this_order && left < order_to_size(this_order - 1))
1855 page = alloc_pages_node(node,
1856 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1862 if (order_to_size(this_order) < rq_size)
1869 page->private = this_order;
1870 list_add_tail(&page->lru, &tags->page_list);
1872 p = page_address(page);
1874 * Allow kmemleak to scan these pages as they contain pointers
1875 * to additional allocations like via ops->init_request().
1877 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1878 entries_per_page = order_to_size(this_order) / rq_size;
1879 to_do = min(entries_per_page, depth - i);
1880 left -= to_do * rq_size;
1881 for (j = 0; j < to_do; j++) {
1882 struct request *rq = p;
1884 tags->static_rqs[i] = rq;
1885 if (set->ops->init_request) {
1886 if (set->ops->init_request(set, rq, hctx_idx,
1888 tags->static_rqs[i] = NULL;
1900 blk_mq_free_rqs(set, tags, hctx_idx);
1905 * 'cpu' is going away. splice any existing rq_list entries from this
1906 * software queue to the hw queue dispatch list, and ensure that it
1909 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1911 struct blk_mq_hw_ctx *hctx;
1912 struct blk_mq_ctx *ctx;
1915 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1916 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1918 spin_lock(&ctx->lock);
1919 if (!list_empty(&ctx->rq_list)) {
1920 list_splice_init(&ctx->rq_list, &tmp);
1921 blk_mq_hctx_clear_pending(hctx, ctx);
1923 spin_unlock(&ctx->lock);
1925 if (list_empty(&tmp))
1928 spin_lock(&hctx->lock);
1929 list_splice_tail_init(&tmp, &hctx->dispatch);
1930 spin_unlock(&hctx->lock);
1932 blk_mq_run_hw_queue(hctx, true);
1936 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1938 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1942 /* hctx->ctxs will be freed in queue's release handler */
1943 static void blk_mq_exit_hctx(struct request_queue *q,
1944 struct blk_mq_tag_set *set,
1945 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1947 blk_mq_debugfs_unregister_hctx(hctx);
1949 if (blk_mq_hw_queue_mapped(hctx))
1950 blk_mq_tag_idle(hctx);
1952 if (set->ops->exit_request)
1953 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1955 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1957 if (set->ops->exit_hctx)
1958 set->ops->exit_hctx(hctx, hctx_idx);
1960 if (hctx->flags & BLK_MQ_F_BLOCKING)
1961 cleanup_srcu_struct(hctx->queue_rq_srcu);
1963 blk_mq_remove_cpuhp(hctx);
1964 blk_free_flush_queue(hctx->fq);
1965 sbitmap_free(&hctx->ctx_map);
1968 static void blk_mq_exit_hw_queues(struct request_queue *q,
1969 struct blk_mq_tag_set *set, int nr_queue)
1971 struct blk_mq_hw_ctx *hctx;
1974 queue_for_each_hw_ctx(q, hctx, i) {
1977 blk_mq_exit_hctx(q, set, hctx, i);
1981 static int blk_mq_init_hctx(struct request_queue *q,
1982 struct blk_mq_tag_set *set,
1983 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1987 node = hctx->numa_node;
1988 if (node == NUMA_NO_NODE)
1989 node = hctx->numa_node = set->numa_node;
1991 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1992 spin_lock_init(&hctx->lock);
1993 INIT_LIST_HEAD(&hctx->dispatch);
1995 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1997 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1999 hctx->tags = set->tags[hctx_idx];
2002 * Allocate space for all possible cpus to avoid allocation at
2005 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
2008 goto unregister_cpu_notifier;
2010 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2016 if (set->ops->init_hctx &&
2017 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2020 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2023 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2025 goto sched_exit_hctx;
2027 if (set->ops->init_request &&
2028 set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
2032 if (hctx->flags & BLK_MQ_F_BLOCKING)
2033 init_srcu_struct(hctx->queue_rq_srcu);
2035 blk_mq_debugfs_register_hctx(q, hctx);
2042 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2044 if (set->ops->exit_hctx)
2045 set->ops->exit_hctx(hctx, hctx_idx);
2047 sbitmap_free(&hctx->ctx_map);
2050 unregister_cpu_notifier:
2051 blk_mq_remove_cpuhp(hctx);
2055 static void blk_mq_init_cpu_queues(struct request_queue *q,
2056 unsigned int nr_hw_queues)
2060 for_each_possible_cpu(i) {
2061 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2062 struct blk_mq_hw_ctx *hctx;
2065 spin_lock_init(&__ctx->lock);
2066 INIT_LIST_HEAD(&__ctx->rq_list);
2069 /* If the cpu isn't present, the cpu is mapped to first hctx */
2070 if (!cpu_present(i))
2073 hctx = blk_mq_map_queue(q, i);
2076 * Set local node, IFF we have more than one hw queue. If
2077 * not, we remain on the home node of the device
2079 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2080 hctx->numa_node = local_memory_node(cpu_to_node(i));
2084 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2088 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2089 set->queue_depth, set->reserved_tags);
2090 if (!set->tags[hctx_idx])
2093 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2098 blk_mq_free_rq_map(set->tags[hctx_idx]);
2099 set->tags[hctx_idx] = NULL;
2103 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2104 unsigned int hctx_idx)
2106 if (set->tags[hctx_idx]) {
2107 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2108 blk_mq_free_rq_map(set->tags[hctx_idx]);
2109 set->tags[hctx_idx] = NULL;
2113 static void blk_mq_map_swqueue(struct request_queue *q)
2115 unsigned int i, hctx_idx;
2116 struct blk_mq_hw_ctx *hctx;
2117 struct blk_mq_ctx *ctx;
2118 struct blk_mq_tag_set *set = q->tag_set;
2121 * Avoid others reading imcomplete hctx->cpumask through sysfs
2123 mutex_lock(&q->sysfs_lock);
2125 queue_for_each_hw_ctx(q, hctx, i) {
2126 cpumask_clear(hctx->cpumask);
2131 * Map software to hardware queues.
2133 * If the cpu isn't present, the cpu is mapped to first hctx.
2135 for_each_present_cpu(i) {
2136 hctx_idx = q->mq_map[i];
2137 /* unmapped hw queue can be remapped after CPU topo changed */
2138 if (!set->tags[hctx_idx] &&
2139 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2141 * If tags initialization fail for some hctx,
2142 * that hctx won't be brought online. In this
2143 * case, remap the current ctx to hctx[0] which
2144 * is guaranteed to always have tags allocated
2149 ctx = per_cpu_ptr(q->queue_ctx, i);
2150 hctx = blk_mq_map_queue(q, i);
2152 cpumask_set_cpu(i, hctx->cpumask);
2153 ctx->index_hw = hctx->nr_ctx;
2154 hctx->ctxs[hctx->nr_ctx++] = ctx;
2157 mutex_unlock(&q->sysfs_lock);
2159 queue_for_each_hw_ctx(q, hctx, i) {
2161 * If no software queues are mapped to this hardware queue,
2162 * disable it and free the request entries.
2164 if (!hctx->nr_ctx) {
2165 /* Never unmap queue 0. We need it as a
2166 * fallback in case of a new remap fails
2169 if (i && set->tags[i])
2170 blk_mq_free_map_and_requests(set, i);
2176 hctx->tags = set->tags[i];
2177 WARN_ON(!hctx->tags);
2180 * Set the map size to the number of mapped software queues.
2181 * This is more accurate and more efficient than looping
2182 * over all possibly mapped software queues.
2184 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2187 * Initialize batch roundrobin counts
2189 hctx->next_cpu = cpumask_first(hctx->cpumask);
2190 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2195 * Caller needs to ensure that we're either frozen/quiesced, or that
2196 * the queue isn't live yet.
2198 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2200 struct blk_mq_hw_ctx *hctx;
2203 queue_for_each_hw_ctx(q, hctx, i) {
2205 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2206 atomic_inc(&q->shared_hctx_restart);
2207 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2209 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2210 atomic_dec(&q->shared_hctx_restart);
2211 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2216 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2219 struct request_queue *q;
2221 lockdep_assert_held(&set->tag_list_lock);
2223 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2224 blk_mq_freeze_queue(q);
2225 queue_set_hctx_shared(q, shared);
2226 blk_mq_unfreeze_queue(q);
2230 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2232 struct blk_mq_tag_set *set = q->tag_set;
2234 mutex_lock(&set->tag_list_lock);
2235 list_del_rcu(&q->tag_set_list);
2236 INIT_LIST_HEAD(&q->tag_set_list);
2237 if (list_is_singular(&set->tag_list)) {
2238 /* just transitioned to unshared */
2239 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2240 /* update existing queue */
2241 blk_mq_update_tag_set_depth(set, false);
2243 mutex_unlock(&set->tag_list_lock);
2248 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2249 struct request_queue *q)
2253 mutex_lock(&set->tag_list_lock);
2255 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2256 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2257 set->flags |= BLK_MQ_F_TAG_SHARED;
2258 /* update existing queue */
2259 blk_mq_update_tag_set_depth(set, true);
2261 if (set->flags & BLK_MQ_F_TAG_SHARED)
2262 queue_set_hctx_shared(q, true);
2263 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2265 mutex_unlock(&set->tag_list_lock);
2269 * It is the actual release handler for mq, but we do it from
2270 * request queue's release handler for avoiding use-after-free
2271 * and headache because q->mq_kobj shouldn't have been introduced,
2272 * but we can't group ctx/kctx kobj without it.
2274 void blk_mq_release(struct request_queue *q)
2276 struct blk_mq_hw_ctx *hctx;
2279 /* hctx kobj stays in hctx */
2280 queue_for_each_hw_ctx(q, hctx, i) {
2283 kobject_put(&hctx->kobj);
2288 kfree(q->queue_hw_ctx);
2291 * release .mq_kobj and sw queue's kobject now because
2292 * both share lifetime with request queue.
2294 blk_mq_sysfs_deinit(q);
2296 free_percpu(q->queue_ctx);
2299 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2301 struct request_queue *uninit_q, *q;
2303 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2305 return ERR_PTR(-ENOMEM);
2307 q = blk_mq_init_allocated_queue(set, uninit_q);
2309 blk_cleanup_queue(uninit_q);
2313 EXPORT_SYMBOL(blk_mq_init_queue);
2315 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2317 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2319 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2320 __alignof__(struct blk_mq_hw_ctx)) !=
2321 sizeof(struct blk_mq_hw_ctx));
2323 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2324 hw_ctx_size += sizeof(struct srcu_struct);
2329 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2330 struct request_queue *q)
2333 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2335 blk_mq_sysfs_unregister(q);
2337 /* protect against switching io scheduler */
2338 mutex_lock(&q->sysfs_lock);
2339 for (i = 0; i < set->nr_hw_queues; i++) {
2345 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2346 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2351 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2358 atomic_set(&hctxs[i]->nr_active, 0);
2359 hctxs[i]->numa_node = node;
2360 hctxs[i]->queue_num = i;
2362 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2363 free_cpumask_var(hctxs[i]->cpumask);
2368 blk_mq_hctx_kobj_init(hctxs[i]);
2370 for (j = i; j < q->nr_hw_queues; j++) {
2371 struct blk_mq_hw_ctx *hctx = hctxs[j];
2375 blk_mq_free_map_and_requests(set, j);
2376 blk_mq_exit_hctx(q, set, hctx, j);
2377 kobject_put(&hctx->kobj);
2382 q->nr_hw_queues = i;
2383 mutex_unlock(&q->sysfs_lock);
2384 blk_mq_sysfs_register(q);
2387 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2388 struct request_queue *q)
2390 /* mark the queue as mq asap */
2391 q->mq_ops = set->ops;
2393 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2394 blk_mq_poll_stats_bkt,
2395 BLK_MQ_POLL_STATS_BKTS, q);
2399 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2403 /* init q->mq_kobj and sw queues' kobjects */
2404 blk_mq_sysfs_init(q);
2406 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2407 GFP_KERNEL, set->numa_node);
2408 if (!q->queue_hw_ctx)
2411 q->mq_map = set->mq_map;
2413 blk_mq_realloc_hw_ctxs(set, q);
2414 if (!q->nr_hw_queues)
2417 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2418 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2420 q->nr_queues = nr_cpu_ids;
2422 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2424 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2425 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2427 q->sg_reserved_size = INT_MAX;
2429 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2430 INIT_LIST_HEAD(&q->requeue_list);
2431 spin_lock_init(&q->requeue_lock);
2433 blk_queue_make_request(q, blk_mq_make_request);
2436 * Do this after blk_queue_make_request() overrides it...
2438 q->nr_requests = set->queue_depth;
2441 * Default to classic polling
2445 if (set->ops->complete)
2446 blk_queue_softirq_done(q, set->ops->complete);
2448 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2449 blk_mq_add_queue_tag_set(set, q);
2450 blk_mq_map_swqueue(q);
2452 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2455 ret = blk_mq_sched_init(q);
2457 return ERR_PTR(ret);
2463 kfree(q->queue_hw_ctx);
2465 free_percpu(q->queue_ctx);
2468 return ERR_PTR(-ENOMEM);
2470 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2472 void blk_mq_free_queue(struct request_queue *q)
2474 struct blk_mq_tag_set *set = q->tag_set;
2476 blk_mq_del_queue_tag_set(q);
2477 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2480 /* Basically redo blk_mq_init_queue with queue frozen */
2481 static void blk_mq_queue_reinit(struct request_queue *q)
2483 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2485 blk_mq_debugfs_unregister_hctxs(q);
2486 blk_mq_sysfs_unregister(q);
2489 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2490 * we should change hctx numa_node according to new topology (this
2491 * involves free and re-allocate memory, worthy doing?)
2494 blk_mq_map_swqueue(q);
2496 blk_mq_sysfs_register(q);
2497 blk_mq_debugfs_register_hctxs(q);
2500 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2504 for (i = 0; i < set->nr_hw_queues; i++)
2505 if (!__blk_mq_alloc_rq_map(set, i))
2512 blk_mq_free_rq_map(set->tags[i]);
2518 * Allocate the request maps associated with this tag_set. Note that this
2519 * may reduce the depth asked for, if memory is tight. set->queue_depth
2520 * will be updated to reflect the allocated depth.
2522 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2527 depth = set->queue_depth;
2529 err = __blk_mq_alloc_rq_maps(set);
2533 set->queue_depth >>= 1;
2534 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2538 } while (set->queue_depth);
2540 if (!set->queue_depth || err) {
2541 pr_err("blk-mq: failed to allocate request map\n");
2545 if (depth != set->queue_depth)
2546 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2547 depth, set->queue_depth);
2552 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2554 if (set->ops->map_queues) {
2557 * transport .map_queues is usually done in the following
2560 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2561 * mask = get_cpu_mask(queue)
2562 * for_each_cpu(cpu, mask)
2563 * set->mq_map[cpu] = queue;
2566 * When we need to remap, the table has to be cleared for
2567 * killing stale mapping since one CPU may not be mapped
2570 for_each_possible_cpu(cpu)
2571 set->mq_map[cpu] = 0;
2573 return set->ops->map_queues(set);
2575 return blk_mq_map_queues(set);
2579 * Alloc a tag set to be associated with one or more request queues.
2580 * May fail with EINVAL for various error conditions. May adjust the
2581 * requested depth down, if if it too large. In that case, the set
2582 * value will be stored in set->queue_depth.
2584 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2588 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2590 if (!set->nr_hw_queues)
2592 if (!set->queue_depth)
2594 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2597 if (!set->ops->queue_rq)
2600 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2601 pr_info("blk-mq: reduced tag depth to %u\n",
2603 set->queue_depth = BLK_MQ_MAX_DEPTH;
2607 * If a crashdump is active, then we are potentially in a very
2608 * memory constrained environment. Limit us to 1 queue and
2609 * 64 tags to prevent using too much memory.
2611 if (is_kdump_kernel()) {
2612 set->nr_hw_queues = 1;
2613 set->queue_depth = min(64U, set->queue_depth);
2616 * There is no use for more h/w queues than cpus.
2618 if (set->nr_hw_queues > nr_cpu_ids)
2619 set->nr_hw_queues = nr_cpu_ids;
2621 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2622 GFP_KERNEL, set->numa_node);
2627 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2628 GFP_KERNEL, set->numa_node);
2632 ret = blk_mq_update_queue_map(set);
2634 goto out_free_mq_map;
2636 ret = blk_mq_alloc_rq_maps(set);
2638 goto out_free_mq_map;
2640 mutex_init(&set->tag_list_lock);
2641 INIT_LIST_HEAD(&set->tag_list);
2653 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2655 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2659 for (i = 0; i < nr_cpu_ids; i++)
2660 blk_mq_free_map_and_requests(set, i);
2668 EXPORT_SYMBOL(blk_mq_free_tag_set);
2670 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2672 struct blk_mq_tag_set *set = q->tag_set;
2673 struct blk_mq_hw_ctx *hctx;
2679 blk_mq_freeze_queue(q);
2682 queue_for_each_hw_ctx(q, hctx, i) {
2686 * If we're using an MQ scheduler, just update the scheduler
2687 * queue depth. This is similar to what the old code would do.
2689 if (!hctx->sched_tags) {
2690 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2691 min(nr, set->queue_depth),
2694 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2702 q->nr_requests = nr;
2704 blk_mq_unfreeze_queue(q);
2709 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2712 struct request_queue *q;
2714 lockdep_assert_held(&set->tag_list_lock);
2716 if (nr_hw_queues > nr_cpu_ids)
2717 nr_hw_queues = nr_cpu_ids;
2718 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2721 list_for_each_entry(q, &set->tag_list, tag_set_list)
2722 blk_mq_freeze_queue(q);
2724 set->nr_hw_queues = nr_hw_queues;
2725 blk_mq_update_queue_map(set);
2726 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2727 blk_mq_realloc_hw_ctxs(set, q);
2728 blk_mq_queue_reinit(q);
2731 list_for_each_entry(q, &set->tag_list, tag_set_list)
2732 blk_mq_unfreeze_queue(q);
2735 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2737 mutex_lock(&set->tag_list_lock);
2738 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2739 mutex_unlock(&set->tag_list_lock);
2741 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2743 /* Enable polling stats and return whether they were already enabled. */
2744 static bool blk_poll_stats_enable(struct request_queue *q)
2746 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2747 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2749 blk_stat_add_callback(q, q->poll_cb);
2753 static void blk_mq_poll_stats_start(struct request_queue *q)
2756 * We don't arm the callback if polling stats are not enabled or the
2757 * callback is already active.
2759 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2760 blk_stat_is_active(q->poll_cb))
2763 blk_stat_activate_msecs(q->poll_cb, 100);
2766 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2768 struct request_queue *q = cb->data;
2771 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2772 if (cb->stat[bucket].nr_samples)
2773 q->poll_stat[bucket] = cb->stat[bucket];
2777 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2778 struct blk_mq_hw_ctx *hctx,
2781 unsigned long ret = 0;
2785 * If stats collection isn't on, don't sleep but turn it on for
2788 if (!blk_poll_stats_enable(q))
2792 * As an optimistic guess, use half of the mean service time
2793 * for this type of request. We can (and should) make this smarter.
2794 * For instance, if the completion latencies are tight, we can
2795 * get closer than just half the mean. This is especially
2796 * important on devices where the completion latencies are longer
2797 * than ~10 usec. We do use the stats for the relevant IO size
2798 * if available which does lead to better estimates.
2800 bucket = blk_mq_poll_stats_bkt(rq);
2804 if (q->poll_stat[bucket].nr_samples)
2805 ret = (q->poll_stat[bucket].mean + 1) / 2;
2810 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2811 struct blk_mq_hw_ctx *hctx,
2814 struct hrtimer_sleeper hs;
2815 enum hrtimer_mode mode;
2819 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2825 * -1: don't ever hybrid sleep
2826 * 0: use half of prev avg
2827 * >0: use this specific value
2829 if (q->poll_nsec == -1)
2831 else if (q->poll_nsec > 0)
2832 nsecs = q->poll_nsec;
2834 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2839 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2842 * This will be replaced with the stats tracking code, using
2843 * 'avg_completion_time / 2' as the pre-sleep target.
2847 mode = HRTIMER_MODE_REL;
2848 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2849 hrtimer_set_expires(&hs.timer, kt);
2851 hrtimer_init_sleeper(&hs, current);
2853 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2855 set_current_state(TASK_UNINTERRUPTIBLE);
2856 hrtimer_start_expires(&hs.timer, mode);
2859 hrtimer_cancel(&hs.timer);
2860 mode = HRTIMER_MODE_ABS;
2861 } while (hs.task && !signal_pending(current));
2863 __set_current_state(TASK_RUNNING);
2864 destroy_hrtimer_on_stack(&hs.timer);
2868 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2870 struct request_queue *q = hctx->queue;
2874 * If we sleep, have the caller restart the poll loop to reset
2875 * the state. Like for the other success return cases, the
2876 * caller is responsible for checking if the IO completed. If
2877 * the IO isn't complete, we'll get called again and will go
2878 * straight to the busy poll loop.
2880 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2883 hctx->poll_considered++;
2885 state = current->state;
2886 while (!need_resched()) {
2889 hctx->poll_invoked++;
2891 ret = q->mq_ops->poll(hctx, rq->tag);
2893 hctx->poll_success++;
2894 set_current_state(TASK_RUNNING);
2898 if (signal_pending_state(state, current))
2899 set_current_state(TASK_RUNNING);
2901 if (current->state == TASK_RUNNING)
2911 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2913 struct blk_mq_hw_ctx *hctx;
2914 struct blk_plug *plug;
2917 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2918 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2921 plug = current->plug;
2923 blk_flush_plug_list(plug, false);
2925 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2926 if (!blk_qc_t_is_internal(cookie))
2927 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2929 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2931 * With scheduling, if the request has completed, we'll
2932 * get a NULL return here, as we clear the sched tag when
2933 * that happens. The request still remains valid, like always,
2934 * so we should be safe with just the NULL check.
2940 return __blk_mq_poll(hctx, rq);
2942 EXPORT_SYMBOL_GPL(blk_mq_poll);
2944 static int __init blk_mq_init(void)
2946 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2947 blk_mq_hctx_notify_dead);
2950 subsys_initcall(blk_mq_init);