1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/kmemleak.h>
15 #include <linux/init.h>
16 #include <linux/slab.h>
17 #include <linux/workqueue.h>
18 #include <linux/smp.h>
19 #include <linux/llist.h>
20 #include <linux/list_sort.h>
21 #include <linux/cpu.h>
22 #include <linux/cache.h>
23 #include <linux/sched/sysctl.h>
24 #include <linux/sched/topology.h>
25 #include <linux/sched/signal.h>
26 #include <linux/delay.h>
27 #include <linux/crash_dump.h>
28 #include <linux/prefetch.h>
30 #include <trace/events/block.h>
32 #include <linux/blk-mq.h>
35 #include "blk-mq-debugfs.h"
36 #include "blk-mq-tag.h"
39 #include "blk-mq-sched.h"
40 #include "blk-rq-qos.h"
42 static void blk_mq_poll_stats_start(struct request_queue *q);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
45 static int blk_mq_poll_stats_bkt(const struct request *rq)
47 int ddir, sectors, bucket;
49 ddir = rq_data_dir(rq);
50 sectors = blk_rq_stats_sectors(rq);
52 bucket = ddir + 2 * ilog2(sectors);
56 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
57 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
63 * Check if any of the ctx, dispatch list or elevator
64 * have pending work in this hardware queue.
66 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
68 return !list_empty_careful(&hctx->dispatch) ||
69 sbitmap_any_bit_set(&hctx->ctx_map) ||
70 blk_mq_sched_has_work(hctx);
74 * Mark this ctx as having pending work in this hardware queue
76 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
77 struct blk_mq_ctx *ctx)
79 const int bit = ctx->index_hw[hctx->type];
81 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
82 sbitmap_set_bit(&hctx->ctx_map, bit);
85 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
86 struct blk_mq_ctx *ctx)
88 const int bit = ctx->index_hw[hctx->type];
90 sbitmap_clear_bit(&hctx->ctx_map, bit);
94 struct hd_struct *part;
95 unsigned int *inflight;
98 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
99 struct request *rq, void *priv,
102 struct mq_inflight *mi = priv;
105 * index[0] counts the specific partition that was asked for.
107 if (rq->part == mi->part)
113 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
115 unsigned inflight[2];
116 struct mq_inflight mi = { .part = part, .inflight = inflight, };
118 inflight[0] = inflight[1] = 0;
119 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
124 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
125 struct request *rq, void *priv,
128 struct mq_inflight *mi = priv;
130 if (rq->part == mi->part)
131 mi->inflight[rq_data_dir(rq)]++;
136 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
137 unsigned int inflight[2])
139 struct mq_inflight mi = { .part = part, .inflight = inflight, };
141 inflight[0] = inflight[1] = 0;
142 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
145 void blk_freeze_queue_start(struct request_queue *q)
147 mutex_lock(&q->mq_freeze_lock);
148 if (++q->mq_freeze_depth == 1) {
149 percpu_ref_kill(&q->q_usage_counter);
150 mutex_unlock(&q->mq_freeze_lock);
152 blk_mq_run_hw_queues(q, false);
154 mutex_unlock(&q->mq_freeze_lock);
157 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
159 void blk_mq_freeze_queue_wait(struct request_queue *q)
161 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
163 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
165 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
166 unsigned long timeout)
168 return wait_event_timeout(q->mq_freeze_wq,
169 percpu_ref_is_zero(&q->q_usage_counter),
172 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
175 * Guarantee no request is in use, so we can change any data structure of
176 * the queue afterward.
178 void blk_freeze_queue(struct request_queue *q)
181 * In the !blk_mq case we are only calling this to kill the
182 * q_usage_counter, otherwise this increases the freeze depth
183 * and waits for it to return to zero. For this reason there is
184 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
185 * exported to drivers as the only user for unfreeze is blk_mq.
187 blk_freeze_queue_start(q);
188 blk_mq_freeze_queue_wait(q);
191 void blk_mq_freeze_queue(struct request_queue *q)
194 * ...just an alias to keep freeze and unfreeze actions balanced
195 * in the blk_mq_* namespace
199 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
201 void blk_mq_unfreeze_queue(struct request_queue *q)
203 mutex_lock(&q->mq_freeze_lock);
204 q->mq_freeze_depth--;
205 WARN_ON_ONCE(q->mq_freeze_depth < 0);
206 if (!q->mq_freeze_depth) {
207 percpu_ref_resurrect(&q->q_usage_counter);
208 wake_up_all(&q->mq_freeze_wq);
210 mutex_unlock(&q->mq_freeze_lock);
212 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
215 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
216 * mpt3sas driver such that this function can be removed.
218 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
220 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
222 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
225 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
228 * Note: this function does not prevent that the struct request end_io()
229 * callback function is invoked. Once this function is returned, we make
230 * sure no dispatch can happen until the queue is unquiesced via
231 * blk_mq_unquiesce_queue().
233 void blk_mq_quiesce_queue(struct request_queue *q)
235 struct blk_mq_hw_ctx *hctx;
239 blk_mq_quiesce_queue_nowait(q);
241 queue_for_each_hw_ctx(q, hctx, i) {
242 if (hctx->flags & BLK_MQ_F_BLOCKING)
243 synchronize_srcu(hctx->srcu);
250 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
253 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
256 * This function recovers queue into the state before quiescing
257 * which is done by blk_mq_quiesce_queue.
259 void blk_mq_unquiesce_queue(struct request_queue *q)
261 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
263 /* dispatch requests which are inserted during quiescing */
264 blk_mq_run_hw_queues(q, true);
266 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
268 void blk_mq_wake_waiters(struct request_queue *q)
270 struct blk_mq_hw_ctx *hctx;
273 queue_for_each_hw_ctx(q, hctx, i)
274 if (blk_mq_hw_queue_mapped(hctx))
275 blk_mq_tag_wakeup_all(hctx->tags, true);
278 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
280 return blk_mq_has_free_tags(hctx->tags);
282 EXPORT_SYMBOL(blk_mq_can_queue);
285 * Only need start/end time stamping if we have stats enabled, or using
288 static inline bool blk_mq_need_time_stamp(struct request *rq)
290 return (rq->rq_flags & RQF_IO_STAT) || rq->q->elevator;
293 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
294 unsigned int tag, unsigned int op, u64 alloc_time_ns)
296 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
297 struct request *rq = tags->static_rqs[tag];
298 req_flags_t rq_flags = 0;
300 if (data->flags & BLK_MQ_REQ_INTERNAL) {
302 rq->internal_tag = tag;
304 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
305 rq_flags = RQF_MQ_INFLIGHT;
306 atomic_inc(&data->hctx->nr_active);
309 rq->internal_tag = -1;
310 data->hctx->tags->rqs[rq->tag] = rq;
313 /* csd/requeue_work/fifo_time is initialized before use */
315 rq->mq_ctx = data->ctx;
316 rq->mq_hctx = data->hctx;
317 rq->rq_flags = rq_flags;
319 if (data->flags & BLK_MQ_REQ_PREEMPT)
320 rq->rq_flags |= RQF_PREEMPT;
321 if (blk_queue_io_stat(data->q))
322 rq->rq_flags |= RQF_IO_STAT;
323 INIT_LIST_HEAD(&rq->queuelist);
324 INIT_HLIST_NODE(&rq->hash);
325 RB_CLEAR_NODE(&rq->rb_node);
328 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
329 rq->alloc_time_ns = alloc_time_ns;
331 if (blk_mq_need_time_stamp(rq))
332 rq->start_time_ns = ktime_get_ns();
334 rq->start_time_ns = 0;
335 rq->io_start_time_ns = 0;
336 rq->stats_sectors = 0;
337 rq->nr_phys_segments = 0;
338 #if defined(CONFIG_BLK_DEV_INTEGRITY)
339 rq->nr_integrity_segments = 0;
341 /* tag was already set */
343 WRITE_ONCE(rq->deadline, 0);
348 rq->end_io_data = NULL;
350 data->ctx->rq_dispatched[op_is_sync(op)]++;
351 refcount_set(&rq->ref, 1);
355 static struct request *blk_mq_get_request(struct request_queue *q,
357 struct blk_mq_alloc_data *data)
359 struct elevator_queue *e = q->elevator;
362 bool clear_ctx_on_error = false;
363 u64 alloc_time_ns = 0;
365 blk_queue_enter_live(q);
367 /* alloc_time includes depth and tag waits */
368 if (blk_queue_rq_alloc_time(q))
369 alloc_time_ns = ktime_get_ns();
372 if (likely(!data->ctx)) {
373 data->ctx = blk_mq_get_ctx(q);
374 clear_ctx_on_error = true;
376 if (likely(!data->hctx))
377 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
379 if (data->cmd_flags & REQ_NOWAIT)
380 data->flags |= BLK_MQ_REQ_NOWAIT;
383 data->flags |= BLK_MQ_REQ_INTERNAL;
386 * Flush requests are special and go directly to the
387 * dispatch list. Don't include reserved tags in the
388 * limiting, as it isn't useful.
390 if (!op_is_flush(data->cmd_flags) &&
391 e->type->ops.limit_depth &&
392 !(data->flags & BLK_MQ_REQ_RESERVED))
393 e->type->ops.limit_depth(data->cmd_flags, data);
395 blk_mq_tag_busy(data->hctx);
398 tag = blk_mq_get_tag(data);
399 if (tag == BLK_MQ_TAG_FAIL) {
400 if (clear_ctx_on_error)
406 rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
407 if (!op_is_flush(data->cmd_flags)) {
409 if (e && e->type->ops.prepare_request) {
410 if (e->type->icq_cache)
411 blk_mq_sched_assign_ioc(rq);
413 e->type->ops.prepare_request(rq, bio);
414 rq->rq_flags |= RQF_ELVPRIV;
417 data->hctx->queued++;
421 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
422 blk_mq_req_flags_t flags)
424 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
428 ret = blk_queue_enter(q, flags);
432 rq = blk_mq_get_request(q, NULL, &alloc_data);
436 return ERR_PTR(-EWOULDBLOCK);
439 rq->__sector = (sector_t) -1;
440 rq->bio = rq->biotail = NULL;
443 EXPORT_SYMBOL(blk_mq_alloc_request);
445 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
446 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
448 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
454 * If the tag allocator sleeps we could get an allocation for a
455 * different hardware context. No need to complicate the low level
456 * allocator for this for the rare use case of a command tied to
459 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
460 return ERR_PTR(-EINVAL);
462 if (hctx_idx >= q->nr_hw_queues)
463 return ERR_PTR(-EIO);
465 ret = blk_queue_enter(q, flags);
470 * Check if the hardware context is actually mapped to anything.
471 * If not tell the caller that it should skip this queue.
473 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
474 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
476 return ERR_PTR(-EXDEV);
478 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
479 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
481 rq = blk_mq_get_request(q, NULL, &alloc_data);
485 return ERR_PTR(-EWOULDBLOCK);
489 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
491 static void __blk_mq_free_request(struct request *rq)
493 struct request_queue *q = rq->q;
494 struct blk_mq_ctx *ctx = rq->mq_ctx;
495 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
496 const int sched_tag = rq->internal_tag;
498 blk_pm_mark_last_busy(rq);
501 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
503 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
504 blk_mq_sched_restart(hctx);
508 void blk_mq_free_request(struct request *rq)
510 struct request_queue *q = rq->q;
511 struct elevator_queue *e = q->elevator;
512 struct blk_mq_ctx *ctx = rq->mq_ctx;
513 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
515 if (rq->rq_flags & RQF_ELVPRIV) {
516 if (e && e->type->ops.finish_request)
517 e->type->ops.finish_request(rq);
519 put_io_context(rq->elv.icq->ioc);
524 ctx->rq_completed[rq_is_sync(rq)]++;
525 if (rq->rq_flags & RQF_MQ_INFLIGHT)
526 atomic_dec(&hctx->nr_active);
528 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
529 laptop_io_completion(q->backing_dev_info);
533 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
534 if (refcount_dec_and_test(&rq->ref))
535 __blk_mq_free_request(rq);
537 EXPORT_SYMBOL_GPL(blk_mq_free_request);
539 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
543 if (blk_mq_need_time_stamp(rq))
544 now = ktime_get_ns();
546 if (rq->rq_flags & RQF_STATS) {
547 blk_mq_poll_stats_start(rq->q);
548 blk_stat_add(rq, now);
551 if (rq->internal_tag != -1)
552 blk_mq_sched_completed_request(rq, now);
554 blk_account_io_done(rq, now);
557 rq_qos_done(rq->q, rq);
558 rq->end_io(rq, error);
560 blk_mq_free_request(rq);
563 EXPORT_SYMBOL(__blk_mq_end_request);
565 void blk_mq_end_request(struct request *rq, blk_status_t error)
567 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
569 __blk_mq_end_request(rq, error);
571 EXPORT_SYMBOL(blk_mq_end_request);
573 static void __blk_mq_complete_request_remote(void *data)
575 struct request *rq = data;
576 struct request_queue *q = rq->q;
578 q->mq_ops->complete(rq);
581 static void __blk_mq_complete_request(struct request *rq)
583 struct blk_mq_ctx *ctx = rq->mq_ctx;
584 struct request_queue *q = rq->q;
588 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
590 * Most of single queue controllers, there is only one irq vector
591 * for handling IO completion, and the only irq's affinity is set
592 * as all possible CPUs. On most of ARCHs, this affinity means the
593 * irq is handled on one specific CPU.
595 * So complete IO reqeust in softirq context in case of single queue
596 * for not degrading IO performance by irqsoff latency.
598 if (q->nr_hw_queues == 1) {
599 __blk_complete_request(rq);
604 * For a polled request, always complete locallly, it's pointless
605 * to redirect the completion.
607 if ((rq->cmd_flags & REQ_HIPRI) ||
608 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
609 q->mq_ops->complete(rq);
614 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
615 shared = cpus_share_cache(cpu, ctx->cpu);
617 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
618 rq->csd.func = __blk_mq_complete_request_remote;
621 smp_call_function_single_async(ctx->cpu, &rq->csd);
623 q->mq_ops->complete(rq);
628 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
629 __releases(hctx->srcu)
631 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
634 srcu_read_unlock(hctx->srcu, srcu_idx);
637 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
638 __acquires(hctx->srcu)
640 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
641 /* shut up gcc false positive */
645 *srcu_idx = srcu_read_lock(hctx->srcu);
649 * blk_mq_complete_request - end I/O on a request
650 * @rq: the request being processed
653 * Ends all I/O on a request. It does not handle partial completions.
654 * The actual completion happens out-of-order, through a IPI handler.
656 bool blk_mq_complete_request(struct request *rq)
658 if (unlikely(blk_should_fake_timeout(rq->q)))
660 __blk_mq_complete_request(rq);
663 EXPORT_SYMBOL(blk_mq_complete_request);
665 int blk_mq_request_started(struct request *rq)
667 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
669 EXPORT_SYMBOL_GPL(blk_mq_request_started);
671 int blk_mq_request_completed(struct request *rq)
673 return blk_mq_rq_state(rq) == MQ_RQ_COMPLETE;
675 EXPORT_SYMBOL_GPL(blk_mq_request_completed);
677 void blk_mq_start_request(struct request *rq)
679 struct request_queue *q = rq->q;
681 trace_block_rq_issue(q, rq);
683 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
684 rq->io_start_time_ns = ktime_get_ns();
685 rq->stats_sectors = blk_rq_sectors(rq);
686 rq->rq_flags |= RQF_STATS;
690 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
693 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
695 if (q->dma_drain_size && blk_rq_bytes(rq)) {
697 * Make sure space for the drain appears. We know we can do
698 * this because max_hw_segments has been adjusted to be one
699 * fewer than the device can handle.
701 rq->nr_phys_segments++;
704 EXPORT_SYMBOL(blk_mq_start_request);
706 static void __blk_mq_requeue_request(struct request *rq)
708 struct request_queue *q = rq->q;
710 blk_mq_put_driver_tag(rq);
712 trace_block_rq_requeue(q, rq);
713 rq_qos_requeue(q, rq);
715 if (blk_mq_request_started(rq)) {
716 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
717 rq->rq_flags &= ~RQF_TIMED_OUT;
718 if (q->dma_drain_size && blk_rq_bytes(rq))
719 rq->nr_phys_segments--;
723 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
725 __blk_mq_requeue_request(rq);
727 /* this request will be re-inserted to io scheduler queue */
728 blk_mq_sched_requeue_request(rq);
730 BUG_ON(!list_empty(&rq->queuelist));
731 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
733 EXPORT_SYMBOL(blk_mq_requeue_request);
735 static void blk_mq_requeue_work(struct work_struct *work)
737 struct request_queue *q =
738 container_of(work, struct request_queue, requeue_work.work);
740 struct request *rq, *next;
742 spin_lock_irq(&q->requeue_lock);
743 list_splice_init(&q->requeue_list, &rq_list);
744 spin_unlock_irq(&q->requeue_lock);
746 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
747 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
750 rq->rq_flags &= ~RQF_SOFTBARRIER;
751 list_del_init(&rq->queuelist);
753 * If RQF_DONTPREP, rq has contained some driver specific
754 * data, so insert it to hctx dispatch list to avoid any
757 if (rq->rq_flags & RQF_DONTPREP)
758 blk_mq_request_bypass_insert(rq, false);
760 blk_mq_sched_insert_request(rq, true, false, false);
763 while (!list_empty(&rq_list)) {
764 rq = list_entry(rq_list.next, struct request, queuelist);
765 list_del_init(&rq->queuelist);
766 blk_mq_sched_insert_request(rq, false, false, false);
769 blk_mq_run_hw_queues(q, false);
772 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
773 bool kick_requeue_list)
775 struct request_queue *q = rq->q;
779 * We abuse this flag that is otherwise used by the I/O scheduler to
780 * request head insertion from the workqueue.
782 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
784 spin_lock_irqsave(&q->requeue_lock, flags);
786 rq->rq_flags |= RQF_SOFTBARRIER;
787 list_add(&rq->queuelist, &q->requeue_list);
789 list_add_tail(&rq->queuelist, &q->requeue_list);
791 spin_unlock_irqrestore(&q->requeue_lock, flags);
793 if (kick_requeue_list)
794 blk_mq_kick_requeue_list(q);
797 void blk_mq_kick_requeue_list(struct request_queue *q)
799 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
801 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
803 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
806 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
807 msecs_to_jiffies(msecs));
809 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
811 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
813 if (tag < tags->nr_tags) {
814 prefetch(tags->rqs[tag]);
815 return tags->rqs[tag];
820 EXPORT_SYMBOL(blk_mq_tag_to_rq);
822 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
823 void *priv, bool reserved)
826 * If we find a request that is inflight and the queue matches,
827 * we know the queue is busy. Return false to stop the iteration.
829 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
839 bool blk_mq_queue_inflight(struct request_queue *q)
843 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
846 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
848 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
850 req->rq_flags |= RQF_TIMED_OUT;
851 if (req->q->mq_ops->timeout) {
852 enum blk_eh_timer_return ret;
854 ret = req->q->mq_ops->timeout(req, reserved);
855 if (ret == BLK_EH_DONE)
857 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
863 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
865 unsigned long deadline;
867 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
869 if (rq->rq_flags & RQF_TIMED_OUT)
872 deadline = READ_ONCE(rq->deadline);
873 if (time_after_eq(jiffies, deadline))
878 else if (time_after(*next, deadline))
883 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
884 struct request *rq, void *priv, bool reserved)
886 unsigned long *next = priv;
889 * Just do a quick check if it is expired before locking the request in
890 * so we're not unnecessarilly synchronizing across CPUs.
892 if (!blk_mq_req_expired(rq, next))
896 * We have reason to believe the request may be expired. Take a
897 * reference on the request to lock this request lifetime into its
898 * currently allocated context to prevent it from being reallocated in
899 * the event the completion by-passes this timeout handler.
901 * If the reference was already released, then the driver beat the
902 * timeout handler to posting a natural completion.
904 if (!refcount_inc_not_zero(&rq->ref))
908 * The request is now locked and cannot be reallocated underneath the
909 * timeout handler's processing. Re-verify this exact request is truly
910 * expired; if it is not expired, then the request was completed and
911 * reallocated as a new request.
913 if (blk_mq_req_expired(rq, next))
914 blk_mq_rq_timed_out(rq, reserved);
915 if (refcount_dec_and_test(&rq->ref))
916 __blk_mq_free_request(rq);
921 static void blk_mq_timeout_work(struct work_struct *work)
923 struct request_queue *q =
924 container_of(work, struct request_queue, timeout_work);
925 unsigned long next = 0;
926 struct blk_mq_hw_ctx *hctx;
929 /* A deadlock might occur if a request is stuck requiring a
930 * timeout at the same time a queue freeze is waiting
931 * completion, since the timeout code would not be able to
932 * acquire the queue reference here.
934 * That's why we don't use blk_queue_enter here; instead, we use
935 * percpu_ref_tryget directly, because we need to be able to
936 * obtain a reference even in the short window between the queue
937 * starting to freeze, by dropping the first reference in
938 * blk_freeze_queue_start, and the moment the last request is
939 * consumed, marked by the instant q_usage_counter reaches
942 if (!percpu_ref_tryget(&q->q_usage_counter))
945 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
948 mod_timer(&q->timeout, next);
951 * Request timeouts are handled as a forward rolling timer. If
952 * we end up here it means that no requests are pending and
953 * also that no request has been pending for a while. Mark
956 queue_for_each_hw_ctx(q, hctx, i) {
957 /* the hctx may be unmapped, so check it here */
958 if (blk_mq_hw_queue_mapped(hctx))
959 blk_mq_tag_idle(hctx);
965 struct flush_busy_ctx_data {
966 struct blk_mq_hw_ctx *hctx;
967 struct list_head *list;
970 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
972 struct flush_busy_ctx_data *flush_data = data;
973 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
974 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
975 enum hctx_type type = hctx->type;
977 spin_lock(&ctx->lock);
978 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
979 sbitmap_clear_bit(sb, bitnr);
980 spin_unlock(&ctx->lock);
985 * Process software queues that have been marked busy, splicing them
986 * to the for-dispatch
988 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
990 struct flush_busy_ctx_data data = {
995 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
997 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
999 struct dispatch_rq_data {
1000 struct blk_mq_hw_ctx *hctx;
1004 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1007 struct dispatch_rq_data *dispatch_data = data;
1008 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1009 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1010 enum hctx_type type = hctx->type;
1012 spin_lock(&ctx->lock);
1013 if (!list_empty(&ctx->rq_lists[type])) {
1014 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1015 list_del_init(&dispatch_data->rq->queuelist);
1016 if (list_empty(&ctx->rq_lists[type]))
1017 sbitmap_clear_bit(sb, bitnr);
1019 spin_unlock(&ctx->lock);
1021 return !dispatch_data->rq;
1024 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1025 struct blk_mq_ctx *start)
1027 unsigned off = start ? start->index_hw[hctx->type] : 0;
1028 struct dispatch_rq_data data = {
1033 __sbitmap_for_each_set(&hctx->ctx_map, off,
1034 dispatch_rq_from_ctx, &data);
1039 static inline unsigned int queued_to_index(unsigned int queued)
1044 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1047 bool blk_mq_get_driver_tag(struct request *rq)
1049 struct blk_mq_alloc_data data = {
1051 .hctx = rq->mq_hctx,
1052 .flags = BLK_MQ_REQ_NOWAIT,
1053 .cmd_flags = rq->cmd_flags,
1060 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1061 data.flags |= BLK_MQ_REQ_RESERVED;
1063 shared = blk_mq_tag_busy(data.hctx);
1064 rq->tag = blk_mq_get_tag(&data);
1067 rq->rq_flags |= RQF_MQ_INFLIGHT;
1068 atomic_inc(&data.hctx->nr_active);
1070 data.hctx->tags->rqs[rq->tag] = rq;
1074 return rq->tag != -1;
1077 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1078 int flags, void *key)
1080 struct blk_mq_hw_ctx *hctx;
1082 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1084 spin_lock(&hctx->dispatch_wait_lock);
1085 if (!list_empty(&wait->entry)) {
1086 struct sbitmap_queue *sbq;
1088 list_del_init(&wait->entry);
1089 sbq = &hctx->tags->bitmap_tags;
1090 atomic_dec(&sbq->ws_active);
1092 spin_unlock(&hctx->dispatch_wait_lock);
1094 blk_mq_run_hw_queue(hctx, true);
1099 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1100 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1101 * restart. For both cases, take care to check the condition again after
1102 * marking us as waiting.
1104 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1107 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1108 struct wait_queue_head *wq;
1109 wait_queue_entry_t *wait;
1112 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1113 blk_mq_sched_mark_restart_hctx(hctx);
1116 * It's possible that a tag was freed in the window between the
1117 * allocation failure and adding the hardware queue to the wait
1120 * Don't clear RESTART here, someone else could have set it.
1121 * At most this will cost an extra queue run.
1123 return blk_mq_get_driver_tag(rq);
1126 wait = &hctx->dispatch_wait;
1127 if (!list_empty_careful(&wait->entry))
1130 wq = &bt_wait_ptr(sbq, hctx)->wait;
1132 spin_lock_irq(&wq->lock);
1133 spin_lock(&hctx->dispatch_wait_lock);
1134 if (!list_empty(&wait->entry)) {
1135 spin_unlock(&hctx->dispatch_wait_lock);
1136 spin_unlock_irq(&wq->lock);
1140 atomic_inc(&sbq->ws_active);
1141 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1142 __add_wait_queue(wq, wait);
1145 * It's possible that a tag was freed in the window between the
1146 * allocation failure and adding the hardware queue to the wait
1149 ret = blk_mq_get_driver_tag(rq);
1151 spin_unlock(&hctx->dispatch_wait_lock);
1152 spin_unlock_irq(&wq->lock);
1157 * We got a tag, remove ourselves from the wait queue to ensure
1158 * someone else gets the wakeup.
1160 list_del_init(&wait->entry);
1161 atomic_dec(&sbq->ws_active);
1162 spin_unlock(&hctx->dispatch_wait_lock);
1163 spin_unlock_irq(&wq->lock);
1168 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1169 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1171 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1172 * - EWMA is one simple way to compute running average value
1173 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1174 * - take 4 as factor for avoiding to get too small(0) result, and this
1175 * factor doesn't matter because EWMA decreases exponentially
1177 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1181 if (hctx->queue->elevator)
1184 ewma = hctx->dispatch_busy;
1189 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1191 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1192 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1194 hctx->dispatch_busy = ewma;
1197 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1200 * Returns true if we did some work AND can potentially do more.
1202 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1205 struct blk_mq_hw_ctx *hctx;
1206 struct request *rq, *nxt;
1207 bool no_tag = false;
1209 blk_status_t ret = BLK_STS_OK;
1211 if (list_empty(list))
1214 WARN_ON(!list_is_singular(list) && got_budget);
1217 * Now process all the entries, sending them to the driver.
1219 errors = queued = 0;
1221 struct blk_mq_queue_data bd;
1223 rq = list_first_entry(list, struct request, queuelist);
1226 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1229 if (!blk_mq_get_driver_tag(rq)) {
1231 * The initial allocation attempt failed, so we need to
1232 * rerun the hardware queue when a tag is freed. The
1233 * waitqueue takes care of that. If the queue is run
1234 * before we add this entry back on the dispatch list,
1235 * we'll re-run it below.
1237 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1238 blk_mq_put_dispatch_budget(hctx);
1240 * For non-shared tags, the RESTART check
1243 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1249 list_del_init(&rq->queuelist);
1254 * Flag last if we have no more requests, or if we have more
1255 * but can't assign a driver tag to it.
1257 if (list_empty(list))
1260 nxt = list_first_entry(list, struct request, queuelist);
1261 bd.last = !blk_mq_get_driver_tag(nxt);
1264 ret = q->mq_ops->queue_rq(hctx, &bd);
1265 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1267 * If an I/O scheduler has been configured and we got a
1268 * driver tag for the next request already, free it
1271 if (!list_empty(list)) {
1272 nxt = list_first_entry(list, struct request, queuelist);
1273 blk_mq_put_driver_tag(nxt);
1275 list_add(&rq->queuelist, list);
1276 __blk_mq_requeue_request(rq);
1280 if (unlikely(ret != BLK_STS_OK)) {
1282 blk_mq_end_request(rq, BLK_STS_IOERR);
1287 } while (!list_empty(list));
1289 hctx->dispatched[queued_to_index(queued)]++;
1292 * Any items that need requeuing? Stuff them into hctx->dispatch,
1293 * that is where we will continue on next queue run.
1295 if (!list_empty(list)) {
1299 * If we didn't flush the entire list, we could have told
1300 * the driver there was more coming, but that turned out to
1303 if (q->mq_ops->commit_rqs)
1304 q->mq_ops->commit_rqs(hctx);
1306 spin_lock(&hctx->lock);
1307 list_splice_init(list, &hctx->dispatch);
1308 spin_unlock(&hctx->lock);
1311 * If SCHED_RESTART was set by the caller of this function and
1312 * it is no longer set that means that it was cleared by another
1313 * thread and hence that a queue rerun is needed.
1315 * If 'no_tag' is set, that means that we failed getting
1316 * a driver tag with an I/O scheduler attached. If our dispatch
1317 * waitqueue is no longer active, ensure that we run the queue
1318 * AFTER adding our entries back to the list.
1320 * If no I/O scheduler has been configured it is possible that
1321 * the hardware queue got stopped and restarted before requests
1322 * were pushed back onto the dispatch list. Rerun the queue to
1323 * avoid starvation. Notes:
1324 * - blk_mq_run_hw_queue() checks whether or not a queue has
1325 * been stopped before rerunning a queue.
1326 * - Some but not all block drivers stop a queue before
1327 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1330 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1331 * bit is set, run queue after a delay to avoid IO stalls
1332 * that could otherwise occur if the queue is idle.
1334 needs_restart = blk_mq_sched_needs_restart(hctx);
1335 if (!needs_restart ||
1336 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1337 blk_mq_run_hw_queue(hctx, true);
1338 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1339 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1341 blk_mq_update_dispatch_busy(hctx, true);
1344 blk_mq_update_dispatch_busy(hctx, false);
1347 * If the host/device is unable to accept more work, inform the
1350 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1353 return (queued + errors) != 0;
1356 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1361 * We should be running this queue from one of the CPUs that
1364 * There are at least two related races now between setting
1365 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1366 * __blk_mq_run_hw_queue():
1368 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1369 * but later it becomes online, then this warning is harmless
1372 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1373 * but later it becomes offline, then the warning can't be
1374 * triggered, and we depend on blk-mq timeout handler to
1375 * handle dispatched requests to this hctx
1377 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1378 cpu_online(hctx->next_cpu)) {
1379 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1380 raw_smp_processor_id(),
1381 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1386 * We can't run the queue inline with ints disabled. Ensure that
1387 * we catch bad users of this early.
1389 WARN_ON_ONCE(in_interrupt());
1391 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1393 hctx_lock(hctx, &srcu_idx);
1394 blk_mq_sched_dispatch_requests(hctx);
1395 hctx_unlock(hctx, srcu_idx);
1398 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1400 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1402 if (cpu >= nr_cpu_ids)
1403 cpu = cpumask_first(hctx->cpumask);
1408 * It'd be great if the workqueue API had a way to pass
1409 * in a mask and had some smarts for more clever placement.
1410 * For now we just round-robin here, switching for every
1411 * BLK_MQ_CPU_WORK_BATCH queued items.
1413 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1416 int next_cpu = hctx->next_cpu;
1418 if (hctx->queue->nr_hw_queues == 1)
1419 return WORK_CPU_UNBOUND;
1421 if (--hctx->next_cpu_batch <= 0) {
1423 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1425 if (next_cpu >= nr_cpu_ids)
1426 next_cpu = blk_mq_first_mapped_cpu(hctx);
1427 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1431 * Do unbound schedule if we can't find a online CPU for this hctx,
1432 * and it should only happen in the path of handling CPU DEAD.
1434 if (!cpu_online(next_cpu)) {
1441 * Make sure to re-select CPU next time once after CPUs
1442 * in hctx->cpumask become online again.
1444 hctx->next_cpu = next_cpu;
1445 hctx->next_cpu_batch = 1;
1446 return WORK_CPU_UNBOUND;
1449 hctx->next_cpu = next_cpu;
1453 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1454 unsigned long msecs)
1456 if (unlikely(blk_mq_hctx_stopped(hctx)))
1459 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1460 int cpu = get_cpu();
1461 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1462 __blk_mq_run_hw_queue(hctx);
1470 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1471 msecs_to_jiffies(msecs));
1474 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1476 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1478 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1480 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1486 * When queue is quiesced, we may be switching io scheduler, or
1487 * updating nr_hw_queues, or other things, and we can't run queue
1488 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1490 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1493 hctx_lock(hctx, &srcu_idx);
1494 need_run = !blk_queue_quiesced(hctx->queue) &&
1495 blk_mq_hctx_has_pending(hctx);
1496 hctx_unlock(hctx, srcu_idx);
1499 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1505 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1507 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1509 struct blk_mq_hw_ctx *hctx;
1512 queue_for_each_hw_ctx(q, hctx, i) {
1513 if (blk_mq_hctx_stopped(hctx))
1516 blk_mq_run_hw_queue(hctx, async);
1519 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1522 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1523 * @q: request queue.
1525 * The caller is responsible for serializing this function against
1526 * blk_mq_{start,stop}_hw_queue().
1528 bool blk_mq_queue_stopped(struct request_queue *q)
1530 struct blk_mq_hw_ctx *hctx;
1533 queue_for_each_hw_ctx(q, hctx, i)
1534 if (blk_mq_hctx_stopped(hctx))
1539 EXPORT_SYMBOL(blk_mq_queue_stopped);
1542 * This function is often used for pausing .queue_rq() by driver when
1543 * there isn't enough resource or some conditions aren't satisfied, and
1544 * BLK_STS_RESOURCE is usually returned.
1546 * We do not guarantee that dispatch can be drained or blocked
1547 * after blk_mq_stop_hw_queue() returns. Please use
1548 * blk_mq_quiesce_queue() for that requirement.
1550 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1552 cancel_delayed_work(&hctx->run_work);
1554 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1556 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1559 * This function is often used for pausing .queue_rq() by driver when
1560 * there isn't enough resource or some conditions aren't satisfied, and
1561 * BLK_STS_RESOURCE is usually returned.
1563 * We do not guarantee that dispatch can be drained or blocked
1564 * after blk_mq_stop_hw_queues() returns. Please use
1565 * blk_mq_quiesce_queue() for that requirement.
1567 void blk_mq_stop_hw_queues(struct request_queue *q)
1569 struct blk_mq_hw_ctx *hctx;
1572 queue_for_each_hw_ctx(q, hctx, i)
1573 blk_mq_stop_hw_queue(hctx);
1575 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1577 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1579 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1581 blk_mq_run_hw_queue(hctx, false);
1583 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1585 void blk_mq_start_hw_queues(struct request_queue *q)
1587 struct blk_mq_hw_ctx *hctx;
1590 queue_for_each_hw_ctx(q, hctx, i)
1591 blk_mq_start_hw_queue(hctx);
1593 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1595 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1597 if (!blk_mq_hctx_stopped(hctx))
1600 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1601 blk_mq_run_hw_queue(hctx, async);
1603 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1605 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1607 struct blk_mq_hw_ctx *hctx;
1610 queue_for_each_hw_ctx(q, hctx, i)
1611 blk_mq_start_stopped_hw_queue(hctx, async);
1613 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1615 static void blk_mq_run_work_fn(struct work_struct *work)
1617 struct blk_mq_hw_ctx *hctx;
1619 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1622 * If we are stopped, don't run the queue.
1624 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1627 __blk_mq_run_hw_queue(hctx);
1630 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1634 struct blk_mq_ctx *ctx = rq->mq_ctx;
1635 enum hctx_type type = hctx->type;
1637 lockdep_assert_held(&ctx->lock);
1639 trace_block_rq_insert(hctx->queue, rq);
1642 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1644 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1647 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1650 struct blk_mq_ctx *ctx = rq->mq_ctx;
1652 lockdep_assert_held(&ctx->lock);
1654 __blk_mq_insert_req_list(hctx, rq, at_head);
1655 blk_mq_hctx_mark_pending(hctx, ctx);
1659 * Should only be used carefully, when the caller knows we want to
1660 * bypass a potential IO scheduler on the target device.
1662 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1664 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1666 spin_lock(&hctx->lock);
1667 list_add_tail(&rq->queuelist, &hctx->dispatch);
1668 spin_unlock(&hctx->lock);
1671 blk_mq_run_hw_queue(hctx, false);
1674 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1675 struct list_head *list)
1679 enum hctx_type type = hctx->type;
1682 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1685 list_for_each_entry(rq, list, queuelist) {
1686 BUG_ON(rq->mq_ctx != ctx);
1687 trace_block_rq_insert(hctx->queue, rq);
1690 spin_lock(&ctx->lock);
1691 list_splice_tail_init(list, &ctx->rq_lists[type]);
1692 blk_mq_hctx_mark_pending(hctx, ctx);
1693 spin_unlock(&ctx->lock);
1696 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1698 struct request *rqa = container_of(a, struct request, queuelist);
1699 struct request *rqb = container_of(b, struct request, queuelist);
1701 if (rqa->mq_ctx < rqb->mq_ctx)
1703 else if (rqa->mq_ctx > rqb->mq_ctx)
1705 else if (rqa->mq_hctx < rqb->mq_hctx)
1707 else if (rqa->mq_hctx > rqb->mq_hctx)
1710 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1713 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1715 struct blk_mq_hw_ctx *this_hctx;
1716 struct blk_mq_ctx *this_ctx;
1717 struct request_queue *this_q;
1723 list_splice_init(&plug->mq_list, &list);
1725 if (plug->rq_count > 2 && plug->multiple_queues)
1726 list_sort(NULL, &list, plug_rq_cmp);
1735 while (!list_empty(&list)) {
1736 rq = list_entry_rq(list.next);
1737 list_del_init(&rq->queuelist);
1739 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1741 trace_block_unplug(this_q, depth, !from_schedule);
1742 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1748 this_ctx = rq->mq_ctx;
1749 this_hctx = rq->mq_hctx;
1754 list_add_tail(&rq->queuelist, &rq_list);
1758 * If 'this_hctx' is set, we know we have entries to complete
1759 * on 'rq_list'. Do those.
1762 trace_block_unplug(this_q, depth, !from_schedule);
1763 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1768 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1769 unsigned int nr_segs)
1771 if (bio->bi_opf & REQ_RAHEAD)
1772 rq->cmd_flags |= REQ_FAILFAST_MASK;
1774 rq->__sector = bio->bi_iter.bi_sector;
1775 rq->write_hint = bio->bi_write_hint;
1776 blk_rq_bio_prep(rq, bio, nr_segs);
1778 blk_account_io_start(rq, true);
1781 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1783 blk_qc_t *cookie, bool last)
1785 struct request_queue *q = rq->q;
1786 struct blk_mq_queue_data bd = {
1790 blk_qc_t new_cookie;
1793 new_cookie = request_to_qc_t(hctx, rq);
1796 * For OK queue, we are done. For error, caller may kill it.
1797 * Any other error (busy), just add it to our list as we
1798 * previously would have done.
1800 ret = q->mq_ops->queue_rq(hctx, &bd);
1803 blk_mq_update_dispatch_busy(hctx, false);
1804 *cookie = new_cookie;
1806 case BLK_STS_RESOURCE:
1807 case BLK_STS_DEV_RESOURCE:
1808 blk_mq_update_dispatch_busy(hctx, true);
1809 __blk_mq_requeue_request(rq);
1812 blk_mq_update_dispatch_busy(hctx, false);
1813 *cookie = BLK_QC_T_NONE;
1820 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1823 bool bypass_insert, bool last)
1825 struct request_queue *q = rq->q;
1826 bool run_queue = true;
1829 * RCU or SRCU read lock is needed before checking quiesced flag.
1831 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1832 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1833 * and avoid driver to try to dispatch again.
1835 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1837 bypass_insert = false;
1841 if (q->elevator && !bypass_insert)
1844 if (!blk_mq_get_dispatch_budget(hctx))
1847 if (!blk_mq_get_driver_tag(rq)) {
1848 blk_mq_put_dispatch_budget(hctx);
1852 return __blk_mq_issue_directly(hctx, rq, cookie, last);
1855 return BLK_STS_RESOURCE;
1857 blk_mq_request_bypass_insert(rq, run_queue);
1861 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1862 struct request *rq, blk_qc_t *cookie)
1867 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1869 hctx_lock(hctx, &srcu_idx);
1871 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1872 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1873 blk_mq_request_bypass_insert(rq, true);
1874 else if (ret != BLK_STS_OK)
1875 blk_mq_end_request(rq, ret);
1877 hctx_unlock(hctx, srcu_idx);
1880 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1884 blk_qc_t unused_cookie;
1885 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1887 hctx_lock(hctx, &srcu_idx);
1888 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1889 hctx_unlock(hctx, srcu_idx);
1894 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1895 struct list_head *list)
1897 while (!list_empty(list)) {
1899 struct request *rq = list_first_entry(list, struct request,
1902 list_del_init(&rq->queuelist);
1903 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1904 if (ret != BLK_STS_OK) {
1905 if (ret == BLK_STS_RESOURCE ||
1906 ret == BLK_STS_DEV_RESOURCE) {
1907 blk_mq_request_bypass_insert(rq,
1911 blk_mq_end_request(rq, ret);
1916 * If we didn't flush the entire list, we could have told
1917 * the driver there was more coming, but that turned out to
1920 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1921 hctx->queue->mq_ops->commit_rqs(hctx);
1924 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1926 list_add_tail(&rq->queuelist, &plug->mq_list);
1928 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1929 struct request *tmp;
1931 tmp = list_first_entry(&plug->mq_list, struct request,
1933 if (tmp->q != rq->q)
1934 plug->multiple_queues = true;
1938 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1940 const int is_sync = op_is_sync(bio->bi_opf);
1941 const int is_flush_fua = op_is_flush(bio->bi_opf);
1942 struct blk_mq_alloc_data data = { .flags = 0};
1944 struct blk_plug *plug;
1945 struct request *same_queue_rq = NULL;
1946 unsigned int nr_segs;
1949 blk_queue_bounce(q, &bio);
1950 __blk_queue_split(q, &bio, &nr_segs);
1952 if (!bio_integrity_prep(bio))
1953 return BLK_QC_T_NONE;
1955 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1956 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1957 return BLK_QC_T_NONE;
1959 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1960 return BLK_QC_T_NONE;
1962 rq_qos_throttle(q, bio);
1964 data.cmd_flags = bio->bi_opf;
1965 rq = blk_mq_get_request(q, bio, &data);
1966 if (unlikely(!rq)) {
1967 rq_qos_cleanup(q, bio);
1969 cookie = BLK_QC_T_NONE;
1970 if (bio->bi_opf & REQ_NOWAIT_INLINE)
1971 cookie = BLK_QC_T_EAGAIN;
1972 else if (bio->bi_opf & REQ_NOWAIT)
1973 bio_wouldblock_error(bio);
1977 trace_block_getrq(q, bio, bio->bi_opf);
1979 rq_qos_track(q, rq, bio);
1981 cookie = request_to_qc_t(data.hctx, rq);
1983 blk_mq_bio_to_request(rq, bio, nr_segs);
1985 plug = blk_mq_plug(q, bio);
1986 if (unlikely(is_flush_fua)) {
1987 /* bypass scheduler for flush rq */
1988 blk_insert_flush(rq);
1989 blk_mq_run_hw_queue(data.hctx, true);
1990 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1992 * Use plugging if we have a ->commit_rqs() hook as well, as
1993 * we know the driver uses bd->last in a smart fashion.
1995 unsigned int request_count = plug->rq_count;
1996 struct request *last = NULL;
1999 trace_block_plug(q);
2001 last = list_entry_rq(plug->mq_list.prev);
2003 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2004 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2005 blk_flush_plug_list(plug, false);
2006 trace_block_plug(q);
2009 blk_add_rq_to_plug(plug, rq);
2010 } else if (plug && !blk_queue_nomerges(q)) {
2012 * We do limited plugging. If the bio can be merged, do that.
2013 * Otherwise the existing request in the plug list will be
2014 * issued. So the plug list will have one request at most
2015 * The plug list might get flushed before this. If that happens,
2016 * the plug list is empty, and same_queue_rq is invalid.
2018 if (list_empty(&plug->mq_list))
2019 same_queue_rq = NULL;
2020 if (same_queue_rq) {
2021 list_del_init(&same_queue_rq->queuelist);
2024 blk_add_rq_to_plug(plug, rq);
2025 trace_block_plug(q);
2027 if (same_queue_rq) {
2028 data.hctx = same_queue_rq->mq_hctx;
2029 trace_block_unplug(q, 1, true);
2030 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2033 } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2034 !data.hctx->dispatch_busy)) {
2035 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2037 blk_mq_sched_insert_request(rq, false, true, true);
2043 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2044 unsigned int hctx_idx)
2048 if (tags->rqs && set->ops->exit_request) {
2051 for (i = 0; i < tags->nr_tags; i++) {
2052 struct request *rq = tags->static_rqs[i];
2056 set->ops->exit_request(set, rq, hctx_idx);
2057 tags->static_rqs[i] = NULL;
2061 while (!list_empty(&tags->page_list)) {
2062 page = list_first_entry(&tags->page_list, struct page, lru);
2063 list_del_init(&page->lru);
2065 * Remove kmemleak object previously allocated in
2066 * blk_mq_alloc_rqs().
2068 kmemleak_free(page_address(page));
2069 __free_pages(page, page->private);
2073 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2077 kfree(tags->static_rqs);
2078 tags->static_rqs = NULL;
2080 blk_mq_free_tags(tags);
2083 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2084 unsigned int hctx_idx,
2085 unsigned int nr_tags,
2086 unsigned int reserved_tags)
2088 struct blk_mq_tags *tags;
2091 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2092 if (node == NUMA_NO_NODE)
2093 node = set->numa_node;
2095 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2096 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2100 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2101 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2104 blk_mq_free_tags(tags);
2108 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2109 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2111 if (!tags->static_rqs) {
2113 blk_mq_free_tags(tags);
2120 static size_t order_to_size(unsigned int order)
2122 return (size_t)PAGE_SIZE << order;
2125 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2126 unsigned int hctx_idx, int node)
2130 if (set->ops->init_request) {
2131 ret = set->ops->init_request(set, rq, hctx_idx, node);
2136 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2140 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2141 unsigned int hctx_idx, unsigned int depth)
2143 unsigned int i, j, entries_per_page, max_order = 4;
2144 size_t rq_size, left;
2147 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2148 if (node == NUMA_NO_NODE)
2149 node = set->numa_node;
2151 INIT_LIST_HEAD(&tags->page_list);
2154 * rq_size is the size of the request plus driver payload, rounded
2155 * to the cacheline size
2157 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2159 left = rq_size * depth;
2161 for (i = 0; i < depth; ) {
2162 int this_order = max_order;
2167 while (this_order && left < order_to_size(this_order - 1))
2171 page = alloc_pages_node(node,
2172 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2178 if (order_to_size(this_order) < rq_size)
2185 page->private = this_order;
2186 list_add_tail(&page->lru, &tags->page_list);
2188 p = page_address(page);
2190 * Allow kmemleak to scan these pages as they contain pointers
2191 * to additional allocations like via ops->init_request().
2193 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2194 entries_per_page = order_to_size(this_order) / rq_size;
2195 to_do = min(entries_per_page, depth - i);
2196 left -= to_do * rq_size;
2197 for (j = 0; j < to_do; j++) {
2198 struct request *rq = p;
2200 tags->static_rqs[i] = rq;
2201 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2202 tags->static_rqs[i] = NULL;
2213 blk_mq_free_rqs(set, tags, hctx_idx);
2218 * 'cpu' is going away. splice any existing rq_list entries from this
2219 * software queue to the hw queue dispatch list, and ensure that it
2222 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2224 struct blk_mq_hw_ctx *hctx;
2225 struct blk_mq_ctx *ctx;
2227 enum hctx_type type;
2229 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2230 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2233 spin_lock(&ctx->lock);
2234 if (!list_empty(&ctx->rq_lists[type])) {
2235 list_splice_init(&ctx->rq_lists[type], &tmp);
2236 blk_mq_hctx_clear_pending(hctx, ctx);
2238 spin_unlock(&ctx->lock);
2240 if (list_empty(&tmp))
2243 spin_lock(&hctx->lock);
2244 list_splice_tail_init(&tmp, &hctx->dispatch);
2245 spin_unlock(&hctx->lock);
2247 blk_mq_run_hw_queue(hctx, true);
2251 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2253 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2257 /* hctx->ctxs will be freed in queue's release handler */
2258 static void blk_mq_exit_hctx(struct request_queue *q,
2259 struct blk_mq_tag_set *set,
2260 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2262 if (blk_mq_hw_queue_mapped(hctx))
2263 blk_mq_tag_idle(hctx);
2265 if (set->ops->exit_request)
2266 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2268 if (set->ops->exit_hctx)
2269 set->ops->exit_hctx(hctx, hctx_idx);
2271 blk_mq_remove_cpuhp(hctx);
2273 spin_lock(&q->unused_hctx_lock);
2274 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2275 spin_unlock(&q->unused_hctx_lock);
2278 static void blk_mq_exit_hw_queues(struct request_queue *q,
2279 struct blk_mq_tag_set *set, int nr_queue)
2281 struct blk_mq_hw_ctx *hctx;
2284 queue_for_each_hw_ctx(q, hctx, i) {
2287 blk_mq_debugfs_unregister_hctx(hctx);
2288 blk_mq_exit_hctx(q, set, hctx, i);
2292 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2294 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2296 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2297 __alignof__(struct blk_mq_hw_ctx)) !=
2298 sizeof(struct blk_mq_hw_ctx));
2300 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2301 hw_ctx_size += sizeof(struct srcu_struct);
2306 static int blk_mq_init_hctx(struct request_queue *q,
2307 struct blk_mq_tag_set *set,
2308 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2310 hctx->queue_num = hctx_idx;
2312 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2314 hctx->tags = set->tags[hctx_idx];
2316 if (set->ops->init_hctx &&
2317 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2318 goto unregister_cpu_notifier;
2320 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2326 if (set->ops->exit_hctx)
2327 set->ops->exit_hctx(hctx, hctx_idx);
2328 unregister_cpu_notifier:
2329 blk_mq_remove_cpuhp(hctx);
2333 static struct blk_mq_hw_ctx *
2334 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2337 struct blk_mq_hw_ctx *hctx;
2338 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2340 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2342 goto fail_alloc_hctx;
2344 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2347 atomic_set(&hctx->nr_active, 0);
2348 if (node == NUMA_NO_NODE)
2349 node = set->numa_node;
2350 hctx->numa_node = node;
2352 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2353 spin_lock_init(&hctx->lock);
2354 INIT_LIST_HEAD(&hctx->dispatch);
2356 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2358 INIT_LIST_HEAD(&hctx->hctx_list);
2361 * Allocate space for all possible cpus to avoid allocation at
2364 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2369 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2374 spin_lock_init(&hctx->dispatch_wait_lock);
2375 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2376 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2378 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2383 if (hctx->flags & BLK_MQ_F_BLOCKING)
2384 init_srcu_struct(hctx->srcu);
2385 blk_mq_hctx_kobj_init(hctx);
2390 sbitmap_free(&hctx->ctx_map);
2394 free_cpumask_var(hctx->cpumask);
2401 static void blk_mq_init_cpu_queues(struct request_queue *q,
2402 unsigned int nr_hw_queues)
2404 struct blk_mq_tag_set *set = q->tag_set;
2407 for_each_possible_cpu(i) {
2408 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2409 struct blk_mq_hw_ctx *hctx;
2413 spin_lock_init(&__ctx->lock);
2414 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2415 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2420 * Set local node, IFF we have more than one hw queue. If
2421 * not, we remain on the home node of the device
2423 for (j = 0; j < set->nr_maps; j++) {
2424 hctx = blk_mq_map_queue_type(q, j, i);
2425 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2426 hctx->numa_node = local_memory_node(cpu_to_node(i));
2431 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2435 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2436 set->queue_depth, set->reserved_tags);
2437 if (!set->tags[hctx_idx])
2440 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2445 blk_mq_free_rq_map(set->tags[hctx_idx]);
2446 set->tags[hctx_idx] = NULL;
2450 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2451 unsigned int hctx_idx)
2453 if (set->tags && set->tags[hctx_idx]) {
2454 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2455 blk_mq_free_rq_map(set->tags[hctx_idx]);
2456 set->tags[hctx_idx] = NULL;
2460 static void blk_mq_map_swqueue(struct request_queue *q)
2462 unsigned int i, j, hctx_idx;
2463 struct blk_mq_hw_ctx *hctx;
2464 struct blk_mq_ctx *ctx;
2465 struct blk_mq_tag_set *set = q->tag_set;
2467 queue_for_each_hw_ctx(q, hctx, i) {
2468 cpumask_clear(hctx->cpumask);
2470 hctx->dispatch_from = NULL;
2474 * Map software to hardware queues.
2476 * If the cpu isn't present, the cpu is mapped to first hctx.
2478 for_each_possible_cpu(i) {
2479 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2480 /* unmapped hw queue can be remapped after CPU topo changed */
2481 if (!set->tags[hctx_idx] &&
2482 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2484 * If tags initialization fail for some hctx,
2485 * that hctx won't be brought online. In this
2486 * case, remap the current ctx to hctx[0] which
2487 * is guaranteed to always have tags allocated
2489 set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2492 ctx = per_cpu_ptr(q->queue_ctx, i);
2493 for (j = 0; j < set->nr_maps; j++) {
2494 if (!set->map[j].nr_queues) {
2495 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2496 HCTX_TYPE_DEFAULT, i);
2500 hctx = blk_mq_map_queue_type(q, j, i);
2501 ctx->hctxs[j] = hctx;
2503 * If the CPU is already set in the mask, then we've
2504 * mapped this one already. This can happen if
2505 * devices share queues across queue maps.
2507 if (cpumask_test_cpu(i, hctx->cpumask))
2510 cpumask_set_cpu(i, hctx->cpumask);
2512 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2513 hctx->ctxs[hctx->nr_ctx++] = ctx;
2516 * If the nr_ctx type overflows, we have exceeded the
2517 * amount of sw queues we can support.
2519 BUG_ON(!hctx->nr_ctx);
2522 for (; j < HCTX_MAX_TYPES; j++)
2523 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2524 HCTX_TYPE_DEFAULT, i);
2527 queue_for_each_hw_ctx(q, hctx, i) {
2529 * If no software queues are mapped to this hardware queue,
2530 * disable it and free the request entries.
2532 if (!hctx->nr_ctx) {
2533 /* Never unmap queue 0. We need it as a
2534 * fallback in case of a new remap fails
2537 if (i && set->tags[i])
2538 blk_mq_free_map_and_requests(set, i);
2544 hctx->tags = set->tags[i];
2545 WARN_ON(!hctx->tags);
2548 * Set the map size to the number of mapped software queues.
2549 * This is more accurate and more efficient than looping
2550 * over all possibly mapped software queues.
2552 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2555 * Initialize batch roundrobin counts
2557 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2558 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2563 * Caller needs to ensure that we're either frozen/quiesced, or that
2564 * the queue isn't live yet.
2566 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2568 struct blk_mq_hw_ctx *hctx;
2571 queue_for_each_hw_ctx(q, hctx, i) {
2573 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2575 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2579 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2582 struct request_queue *q;
2584 lockdep_assert_held(&set->tag_list_lock);
2586 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2587 blk_mq_freeze_queue(q);
2588 queue_set_hctx_shared(q, shared);
2589 blk_mq_unfreeze_queue(q);
2593 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2595 struct blk_mq_tag_set *set = q->tag_set;
2597 mutex_lock(&set->tag_list_lock);
2598 list_del_rcu(&q->tag_set_list);
2599 if (list_is_singular(&set->tag_list)) {
2600 /* just transitioned to unshared */
2601 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2602 /* update existing queue */
2603 blk_mq_update_tag_set_depth(set, false);
2605 mutex_unlock(&set->tag_list_lock);
2606 INIT_LIST_HEAD(&q->tag_set_list);
2609 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2610 struct request_queue *q)
2612 mutex_lock(&set->tag_list_lock);
2615 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2617 if (!list_empty(&set->tag_list) &&
2618 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2619 set->flags |= BLK_MQ_F_TAG_SHARED;
2620 /* update existing queue */
2621 blk_mq_update_tag_set_depth(set, true);
2623 if (set->flags & BLK_MQ_F_TAG_SHARED)
2624 queue_set_hctx_shared(q, true);
2625 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2627 mutex_unlock(&set->tag_list_lock);
2630 /* All allocations will be freed in release handler of q->mq_kobj */
2631 static int blk_mq_alloc_ctxs(struct request_queue *q)
2633 struct blk_mq_ctxs *ctxs;
2636 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2640 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2641 if (!ctxs->queue_ctx)
2644 for_each_possible_cpu(cpu) {
2645 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2649 q->mq_kobj = &ctxs->kobj;
2650 q->queue_ctx = ctxs->queue_ctx;
2659 * It is the actual release handler for mq, but we do it from
2660 * request queue's release handler for avoiding use-after-free
2661 * and headache because q->mq_kobj shouldn't have been introduced,
2662 * but we can't group ctx/kctx kobj without it.
2664 void blk_mq_release(struct request_queue *q)
2666 struct blk_mq_hw_ctx *hctx, *next;
2669 cancel_delayed_work_sync(&q->requeue_work);
2671 queue_for_each_hw_ctx(q, hctx, i)
2672 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2674 /* all hctx are in .unused_hctx_list now */
2675 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2676 list_del_init(&hctx->hctx_list);
2677 kobject_put(&hctx->kobj);
2680 kfree(q->queue_hw_ctx);
2683 * release .mq_kobj and sw queue's kobject now because
2684 * both share lifetime with request queue.
2686 blk_mq_sysfs_deinit(q);
2689 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2691 struct request_queue *uninit_q, *q;
2693 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2695 return ERR_PTR(-ENOMEM);
2698 * Initialize the queue without an elevator. device_add_disk() will do
2699 * the initialization.
2701 q = blk_mq_init_allocated_queue(set, uninit_q, false);
2703 blk_cleanup_queue(uninit_q);
2707 EXPORT_SYMBOL(blk_mq_init_queue);
2710 * Helper for setting up a queue with mq ops, given queue depth, and
2711 * the passed in mq ops flags.
2713 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2714 const struct blk_mq_ops *ops,
2715 unsigned int queue_depth,
2716 unsigned int set_flags)
2718 struct request_queue *q;
2721 memset(set, 0, sizeof(*set));
2723 set->nr_hw_queues = 1;
2725 set->queue_depth = queue_depth;
2726 set->numa_node = NUMA_NO_NODE;
2727 set->flags = set_flags;
2729 ret = blk_mq_alloc_tag_set(set);
2731 return ERR_PTR(ret);
2733 q = blk_mq_init_queue(set);
2735 blk_mq_free_tag_set(set);
2741 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2743 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2744 struct blk_mq_tag_set *set, struct request_queue *q,
2745 int hctx_idx, int node)
2747 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2749 /* reuse dead hctx first */
2750 spin_lock(&q->unused_hctx_lock);
2751 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2752 if (tmp->numa_node == node) {
2758 list_del_init(&hctx->hctx_list);
2759 spin_unlock(&q->unused_hctx_lock);
2762 hctx = blk_mq_alloc_hctx(q, set, node);
2766 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2772 kobject_put(&hctx->kobj);
2777 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2778 struct request_queue *q)
2781 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2783 /* protect against switching io scheduler */
2784 mutex_lock(&q->sysfs_lock);
2785 for (i = 0; i < set->nr_hw_queues; i++) {
2787 struct blk_mq_hw_ctx *hctx;
2789 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2791 * If the hw queue has been mapped to another numa node,
2792 * we need to realloc the hctx. If allocation fails, fallback
2793 * to use the previous one.
2795 if (hctxs[i] && (hctxs[i]->numa_node == node))
2798 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2801 blk_mq_exit_hctx(q, set, hctxs[i], i);
2805 pr_warn("Allocate new hctx on node %d fails,\
2806 fallback to previous one on node %d\n",
2807 node, hctxs[i]->numa_node);
2813 * Increasing nr_hw_queues fails. Free the newly allocated
2814 * hctxs and keep the previous q->nr_hw_queues.
2816 if (i != set->nr_hw_queues) {
2817 j = q->nr_hw_queues;
2821 end = q->nr_hw_queues;
2822 q->nr_hw_queues = set->nr_hw_queues;
2825 for (; j < end; j++) {
2826 struct blk_mq_hw_ctx *hctx = hctxs[j];
2830 blk_mq_free_map_and_requests(set, j);
2831 blk_mq_exit_hctx(q, set, hctx, j);
2835 mutex_unlock(&q->sysfs_lock);
2839 * Maximum number of hardware queues we support. For single sets, we'll never
2840 * have more than the CPUs (software queues). For multiple sets, the tag_set
2841 * user may have set ->nr_hw_queues larger.
2843 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2845 if (set->nr_maps == 1)
2848 return max(set->nr_hw_queues, nr_cpu_ids);
2851 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2852 struct request_queue *q,
2855 /* mark the queue as mq asap */
2856 q->mq_ops = set->ops;
2858 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2859 blk_mq_poll_stats_bkt,
2860 BLK_MQ_POLL_STATS_BKTS, q);
2864 if (blk_mq_alloc_ctxs(q))
2867 /* init q->mq_kobj and sw queues' kobjects */
2868 blk_mq_sysfs_init(q);
2870 q->nr_queues = nr_hw_queues(set);
2871 q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2872 GFP_KERNEL, set->numa_node);
2873 if (!q->queue_hw_ctx)
2876 INIT_LIST_HEAD(&q->unused_hctx_list);
2877 spin_lock_init(&q->unused_hctx_lock);
2879 blk_mq_realloc_hw_ctxs(set, q);
2880 if (!q->nr_hw_queues)
2883 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2884 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2888 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2889 if (set->nr_maps > HCTX_TYPE_POLL &&
2890 set->map[HCTX_TYPE_POLL].nr_queues)
2891 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2893 q->sg_reserved_size = INT_MAX;
2895 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2896 INIT_LIST_HEAD(&q->requeue_list);
2897 spin_lock_init(&q->requeue_lock);
2899 blk_queue_make_request(q, blk_mq_make_request);
2902 * Do this after blk_queue_make_request() overrides it...
2904 q->nr_requests = set->queue_depth;
2907 * Default to classic polling
2909 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2911 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2912 blk_mq_add_queue_tag_set(set, q);
2913 blk_mq_map_swqueue(q);
2916 elevator_init_mq(q);
2921 kfree(q->queue_hw_ctx);
2922 q->nr_hw_queues = 0;
2924 blk_mq_sysfs_deinit(q);
2926 blk_stat_free_callback(q->poll_cb);
2930 return ERR_PTR(-ENOMEM);
2932 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2934 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2935 void blk_mq_exit_queue(struct request_queue *q)
2937 struct blk_mq_tag_set *set = q->tag_set;
2939 blk_mq_del_queue_tag_set(q);
2940 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2943 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2947 for (i = 0; i < set->nr_hw_queues; i++)
2948 if (!__blk_mq_alloc_rq_map(set, i))
2955 blk_mq_free_rq_map(set->tags[i]);
2961 * Allocate the request maps associated with this tag_set. Note that this
2962 * may reduce the depth asked for, if memory is tight. set->queue_depth
2963 * will be updated to reflect the allocated depth.
2965 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2970 depth = set->queue_depth;
2972 err = __blk_mq_alloc_rq_maps(set);
2976 set->queue_depth >>= 1;
2977 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2981 } while (set->queue_depth);
2983 if (!set->queue_depth || err) {
2984 pr_err("blk-mq: failed to allocate request map\n");
2988 if (depth != set->queue_depth)
2989 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2990 depth, set->queue_depth);
2995 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2997 if (set->ops->map_queues && !is_kdump_kernel()) {
3001 * transport .map_queues is usually done in the following
3004 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3005 * mask = get_cpu_mask(queue)
3006 * for_each_cpu(cpu, mask)
3007 * set->map[x].mq_map[cpu] = queue;
3010 * When we need to remap, the table has to be cleared for
3011 * killing stale mapping since one CPU may not be mapped
3014 for (i = 0; i < set->nr_maps; i++)
3015 blk_mq_clear_mq_map(&set->map[i]);
3017 return set->ops->map_queues(set);
3019 BUG_ON(set->nr_maps > 1);
3020 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3025 * Alloc a tag set to be associated with one or more request queues.
3026 * May fail with EINVAL for various error conditions. May adjust the
3027 * requested depth down, if it's too large. In that case, the set
3028 * value will be stored in set->queue_depth.
3030 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3034 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3036 if (!set->nr_hw_queues)
3038 if (!set->queue_depth)
3040 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3043 if (!set->ops->queue_rq)
3046 if (!set->ops->get_budget ^ !set->ops->put_budget)
3049 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3050 pr_info("blk-mq: reduced tag depth to %u\n",
3052 set->queue_depth = BLK_MQ_MAX_DEPTH;
3057 else if (set->nr_maps > HCTX_MAX_TYPES)
3061 * If a crashdump is active, then we are potentially in a very
3062 * memory constrained environment. Limit us to 1 queue and
3063 * 64 tags to prevent using too much memory.
3065 if (is_kdump_kernel()) {
3066 set->nr_hw_queues = 1;
3068 set->queue_depth = min(64U, set->queue_depth);
3071 * There is no use for more h/w queues than cpus if we just have
3074 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3075 set->nr_hw_queues = nr_cpu_ids;
3077 set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3078 GFP_KERNEL, set->numa_node);
3083 for (i = 0; i < set->nr_maps; i++) {
3084 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3085 sizeof(set->map[i].mq_map[0]),
3086 GFP_KERNEL, set->numa_node);
3087 if (!set->map[i].mq_map)
3088 goto out_free_mq_map;
3089 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3092 ret = blk_mq_update_queue_map(set);
3094 goto out_free_mq_map;
3096 ret = blk_mq_alloc_rq_maps(set);
3098 goto out_free_mq_map;
3100 mutex_init(&set->tag_list_lock);
3101 INIT_LIST_HEAD(&set->tag_list);
3106 for (i = 0; i < set->nr_maps; i++) {
3107 kfree(set->map[i].mq_map);
3108 set->map[i].mq_map = NULL;
3114 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3116 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3120 for (i = 0; i < nr_hw_queues(set); i++)
3121 blk_mq_free_map_and_requests(set, i);
3123 for (j = 0; j < set->nr_maps; j++) {
3124 kfree(set->map[j].mq_map);
3125 set->map[j].mq_map = NULL;
3131 EXPORT_SYMBOL(blk_mq_free_tag_set);
3133 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3135 struct blk_mq_tag_set *set = q->tag_set;
3136 struct blk_mq_hw_ctx *hctx;
3142 if (q->nr_requests == nr)
3145 blk_mq_freeze_queue(q);
3146 blk_mq_quiesce_queue(q);
3149 queue_for_each_hw_ctx(q, hctx, i) {
3153 * If we're using an MQ scheduler, just update the scheduler
3154 * queue depth. This is similar to what the old code would do.
3156 if (!hctx->sched_tags) {
3157 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3160 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3165 if (q->elevator && q->elevator->type->ops.depth_updated)
3166 q->elevator->type->ops.depth_updated(hctx);
3170 q->nr_requests = nr;
3172 blk_mq_unquiesce_queue(q);
3173 blk_mq_unfreeze_queue(q);
3179 * request_queue and elevator_type pair.
3180 * It is just used by __blk_mq_update_nr_hw_queues to cache
3181 * the elevator_type associated with a request_queue.
3183 struct blk_mq_qe_pair {
3184 struct list_head node;
3185 struct request_queue *q;
3186 struct elevator_type *type;
3190 * Cache the elevator_type in qe pair list and switch the
3191 * io scheduler to 'none'
3193 static bool blk_mq_elv_switch_none(struct list_head *head,
3194 struct request_queue *q)
3196 struct blk_mq_qe_pair *qe;
3201 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3205 INIT_LIST_HEAD(&qe->node);
3207 qe->type = q->elevator->type;
3208 list_add(&qe->node, head);
3210 mutex_lock(&q->sysfs_lock);
3212 * After elevator_switch_mq, the previous elevator_queue will be
3213 * released by elevator_release. The reference of the io scheduler
3214 * module get by elevator_get will also be put. So we need to get
3215 * a reference of the io scheduler module here to prevent it to be
3218 __module_get(qe->type->elevator_owner);
3219 elevator_switch_mq(q, NULL);
3220 mutex_unlock(&q->sysfs_lock);
3225 static void blk_mq_elv_switch_back(struct list_head *head,
3226 struct request_queue *q)
3228 struct blk_mq_qe_pair *qe;
3229 struct elevator_type *t = NULL;
3231 list_for_each_entry(qe, head, node)
3240 list_del(&qe->node);
3243 mutex_lock(&q->sysfs_lock);
3244 elevator_switch_mq(q, t);
3245 mutex_unlock(&q->sysfs_lock);
3248 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3251 struct request_queue *q;
3253 int prev_nr_hw_queues;
3255 lockdep_assert_held(&set->tag_list_lock);
3257 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3258 nr_hw_queues = nr_cpu_ids;
3259 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3262 list_for_each_entry(q, &set->tag_list, tag_set_list)
3263 blk_mq_freeze_queue(q);
3265 * Sync with blk_mq_queue_tag_busy_iter.
3269 * Switch IO scheduler to 'none', cleaning up the data associated
3270 * with the previous scheduler. We will switch back once we are done
3271 * updating the new sw to hw queue mappings.
3273 list_for_each_entry(q, &set->tag_list, tag_set_list)
3274 if (!blk_mq_elv_switch_none(&head, q))
3277 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3278 blk_mq_debugfs_unregister_hctxs(q);
3279 blk_mq_sysfs_unregister(q);
3282 prev_nr_hw_queues = set->nr_hw_queues;
3283 set->nr_hw_queues = nr_hw_queues;
3284 blk_mq_update_queue_map(set);
3286 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3287 blk_mq_realloc_hw_ctxs(set, q);
3288 if (q->nr_hw_queues != set->nr_hw_queues) {
3289 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3290 nr_hw_queues, prev_nr_hw_queues);
3291 set->nr_hw_queues = prev_nr_hw_queues;
3292 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3295 blk_mq_map_swqueue(q);
3298 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3299 blk_mq_sysfs_register(q);
3300 blk_mq_debugfs_register_hctxs(q);
3304 list_for_each_entry(q, &set->tag_list, tag_set_list)
3305 blk_mq_elv_switch_back(&head, q);
3307 list_for_each_entry(q, &set->tag_list, tag_set_list)
3308 blk_mq_unfreeze_queue(q);
3311 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3313 mutex_lock(&set->tag_list_lock);
3314 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3315 mutex_unlock(&set->tag_list_lock);
3317 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3319 /* Enable polling stats and return whether they were already enabled. */
3320 static bool blk_poll_stats_enable(struct request_queue *q)
3322 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3323 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3325 blk_stat_add_callback(q, q->poll_cb);
3329 static void blk_mq_poll_stats_start(struct request_queue *q)
3332 * We don't arm the callback if polling stats are not enabled or the
3333 * callback is already active.
3335 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3336 blk_stat_is_active(q->poll_cb))
3339 blk_stat_activate_msecs(q->poll_cb, 100);
3342 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3344 struct request_queue *q = cb->data;
3347 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3348 if (cb->stat[bucket].nr_samples)
3349 q->poll_stat[bucket] = cb->stat[bucket];
3353 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3354 struct blk_mq_hw_ctx *hctx,
3357 unsigned long ret = 0;
3361 * If stats collection isn't on, don't sleep but turn it on for
3364 if (!blk_poll_stats_enable(q))
3368 * As an optimistic guess, use half of the mean service time
3369 * for this type of request. We can (and should) make this smarter.
3370 * For instance, if the completion latencies are tight, we can
3371 * get closer than just half the mean. This is especially
3372 * important on devices where the completion latencies are longer
3373 * than ~10 usec. We do use the stats for the relevant IO size
3374 * if available which does lead to better estimates.
3376 bucket = blk_mq_poll_stats_bkt(rq);
3380 if (q->poll_stat[bucket].nr_samples)
3381 ret = (q->poll_stat[bucket].mean + 1) / 2;
3386 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3387 struct blk_mq_hw_ctx *hctx,
3390 struct hrtimer_sleeper hs;
3391 enum hrtimer_mode mode;
3395 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3399 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3401 * 0: use half of prev avg
3402 * >0: use this specific value
3404 if (q->poll_nsec > 0)
3405 nsecs = q->poll_nsec;
3407 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3412 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3415 * This will be replaced with the stats tracking code, using
3416 * 'avg_completion_time / 2' as the pre-sleep target.
3420 mode = HRTIMER_MODE_REL;
3421 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3422 hrtimer_set_expires(&hs.timer, kt);
3424 hrtimer_init_sleeper(&hs, current);
3426 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3428 set_current_state(TASK_UNINTERRUPTIBLE);
3429 hrtimer_start_expires(&hs.timer, mode);
3432 hrtimer_cancel(&hs.timer);
3433 mode = HRTIMER_MODE_ABS;
3434 } while (hs.task && !signal_pending(current));
3436 __set_current_state(TASK_RUNNING);
3437 destroy_hrtimer_on_stack(&hs.timer);
3441 static bool blk_mq_poll_hybrid(struct request_queue *q,
3442 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3446 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3449 if (!blk_qc_t_is_internal(cookie))
3450 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3452 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3454 * With scheduling, if the request has completed, we'll
3455 * get a NULL return here, as we clear the sched tag when
3456 * that happens. The request still remains valid, like always,
3457 * so we should be safe with just the NULL check.
3463 return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3467 * blk_poll - poll for IO completions
3469 * @cookie: cookie passed back at IO submission time
3470 * @spin: whether to spin for completions
3473 * Poll for completions on the passed in queue. Returns number of
3474 * completed entries found. If @spin is true, then blk_poll will continue
3475 * looping until at least one completion is found, unless the task is
3476 * otherwise marked running (or we need to reschedule).
3478 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3480 struct blk_mq_hw_ctx *hctx;
3483 if (!blk_qc_t_valid(cookie) ||
3484 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3488 blk_flush_plug_list(current->plug, false);
3490 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3493 * If we sleep, have the caller restart the poll loop to reset
3494 * the state. Like for the other success return cases, the
3495 * caller is responsible for checking if the IO completed. If
3496 * the IO isn't complete, we'll get called again and will go
3497 * straight to the busy poll loop.
3499 if (blk_mq_poll_hybrid(q, hctx, cookie))
3502 hctx->poll_considered++;
3504 state = current->state;
3508 hctx->poll_invoked++;
3510 ret = q->mq_ops->poll(hctx);
3512 hctx->poll_success++;
3513 __set_current_state(TASK_RUNNING);
3517 if (signal_pending_state(state, current))
3518 __set_current_state(TASK_RUNNING);
3520 if (current->state == TASK_RUNNING)
3522 if (ret < 0 || !spin)
3525 } while (!need_resched());
3527 __set_current_state(TASK_RUNNING);
3530 EXPORT_SYMBOL_GPL(blk_poll);
3532 unsigned int blk_mq_rq_cpu(struct request *rq)
3534 return rq->mq_ctx->cpu;
3536 EXPORT_SYMBOL(blk_mq_rq_cpu);
3538 static int __init blk_mq_init(void)
3540 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3541 blk_mq_hctx_notify_dead);
3544 subsys_initcall(blk_mq_init);