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>
33 #include <linux/t10-pi.h>
36 #include "blk-mq-debugfs.h"
37 #include "blk-mq-tag.h"
40 #include "blk-mq-sched.h"
41 #include "blk-rq-qos.h"
43 static void blk_mq_poll_stats_start(struct request_queue *q);
44 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
46 static int blk_mq_poll_stats_bkt(const struct request *rq)
48 int ddir, sectors, bucket;
50 ddir = rq_data_dir(rq);
51 sectors = blk_rq_stats_sectors(rq);
53 bucket = ddir + 2 * ilog2(sectors);
57 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
58 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
64 * Check if any of the ctx, dispatch list or elevator
65 * have pending work in this hardware queue.
67 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
69 return !list_empty_careful(&hctx->dispatch) ||
70 sbitmap_any_bit_set(&hctx->ctx_map) ||
71 blk_mq_sched_has_work(hctx);
75 * Mark this ctx as having pending work in this hardware queue
77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
78 struct blk_mq_ctx *ctx)
80 const int bit = ctx->index_hw[hctx->type];
82 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
83 sbitmap_set_bit(&hctx->ctx_map, bit);
86 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
87 struct blk_mq_ctx *ctx)
89 const int bit = ctx->index_hw[hctx->type];
91 sbitmap_clear_bit(&hctx->ctx_map, bit);
95 struct hd_struct *part;
96 unsigned int *inflight;
99 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
100 struct request *rq, void *priv,
103 struct mq_inflight *mi = priv;
106 * index[0] counts the specific partition that was asked for.
108 if (rq->part == mi->part)
114 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
116 unsigned inflight[2];
117 struct mq_inflight mi = { .part = part, .inflight = inflight, };
119 inflight[0] = inflight[1] = 0;
120 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
125 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
126 struct request *rq, void *priv,
129 struct mq_inflight *mi = priv;
131 if (rq->part == mi->part)
132 mi->inflight[rq_data_dir(rq)]++;
137 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
138 unsigned int inflight[2])
140 struct mq_inflight mi = { .part = part, .inflight = inflight, };
142 inflight[0] = inflight[1] = 0;
143 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
146 void blk_freeze_queue_start(struct request_queue *q)
148 mutex_lock(&q->mq_freeze_lock);
149 if (++q->mq_freeze_depth == 1) {
150 percpu_ref_kill(&q->q_usage_counter);
151 mutex_unlock(&q->mq_freeze_lock);
153 blk_mq_run_hw_queues(q, false);
155 mutex_unlock(&q->mq_freeze_lock);
158 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
160 void blk_mq_freeze_queue_wait(struct request_queue *q)
162 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
164 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
166 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
167 unsigned long timeout)
169 return wait_event_timeout(q->mq_freeze_wq,
170 percpu_ref_is_zero(&q->q_usage_counter),
173 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
176 * Guarantee no request is in use, so we can change any data structure of
177 * the queue afterward.
179 void blk_freeze_queue(struct request_queue *q)
182 * In the !blk_mq case we are only calling this to kill the
183 * q_usage_counter, otherwise this increases the freeze depth
184 * and waits for it to return to zero. For this reason there is
185 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
186 * exported to drivers as the only user for unfreeze is blk_mq.
188 blk_freeze_queue_start(q);
189 blk_mq_freeze_queue_wait(q);
192 void blk_mq_freeze_queue(struct request_queue *q)
195 * ...just an alias to keep freeze and unfreeze actions balanced
196 * in the blk_mq_* namespace
200 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
202 void blk_mq_unfreeze_queue(struct request_queue *q)
204 mutex_lock(&q->mq_freeze_lock);
205 q->mq_freeze_depth--;
206 WARN_ON_ONCE(q->mq_freeze_depth < 0);
207 if (!q->mq_freeze_depth) {
208 percpu_ref_resurrect(&q->q_usage_counter);
209 wake_up_all(&q->mq_freeze_wq);
211 mutex_unlock(&q->mq_freeze_lock);
213 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
216 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
217 * mpt3sas driver such that this function can be removed.
219 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
221 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
223 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
226 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
229 * Note: this function does not prevent that the struct request end_io()
230 * callback function is invoked. Once this function is returned, we make
231 * sure no dispatch can happen until the queue is unquiesced via
232 * blk_mq_unquiesce_queue().
234 void blk_mq_quiesce_queue(struct request_queue *q)
236 struct blk_mq_hw_ctx *hctx;
240 blk_mq_quiesce_queue_nowait(q);
242 queue_for_each_hw_ctx(q, hctx, i) {
243 if (hctx->flags & BLK_MQ_F_BLOCKING)
244 synchronize_srcu(hctx->srcu);
251 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
254 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
257 * This function recovers queue into the state before quiescing
258 * which is done by blk_mq_quiesce_queue.
260 void blk_mq_unquiesce_queue(struct request_queue *q)
262 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
264 /* dispatch requests which are inserted during quiescing */
265 blk_mq_run_hw_queues(q, true);
267 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
269 void blk_mq_wake_waiters(struct request_queue *q)
271 struct blk_mq_hw_ctx *hctx;
274 queue_for_each_hw_ctx(q, hctx, i)
275 if (blk_mq_hw_queue_mapped(hctx))
276 blk_mq_tag_wakeup_all(hctx->tags, true);
279 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
281 return blk_mq_has_free_tags(hctx->tags);
283 EXPORT_SYMBOL(blk_mq_can_queue);
286 * Only need start/end time stamping if we have iostat or
287 * blk stats enabled, or using an IO scheduler.
289 static inline bool blk_mq_need_time_stamp(struct request *rq)
291 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
294 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
295 unsigned int tag, unsigned int op, u64 alloc_time_ns)
297 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
298 struct request *rq = tags->static_rqs[tag];
299 req_flags_t rq_flags = 0;
301 if (data->flags & BLK_MQ_REQ_INTERNAL) {
303 rq->internal_tag = tag;
305 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
306 rq_flags = RQF_MQ_INFLIGHT;
307 atomic_inc(&data->hctx->nr_active);
310 rq->internal_tag = -1;
311 data->hctx->tags->rqs[rq->tag] = rq;
314 /* csd/requeue_work/fifo_time is initialized before use */
316 rq->mq_ctx = data->ctx;
317 rq->mq_hctx = data->hctx;
318 rq->rq_flags = rq_flags;
320 if (data->flags & BLK_MQ_REQ_PREEMPT)
321 rq->rq_flags |= RQF_PREEMPT;
322 if (blk_queue_io_stat(data->q))
323 rq->rq_flags |= RQF_IO_STAT;
324 INIT_LIST_HEAD(&rq->queuelist);
325 INIT_HLIST_NODE(&rq->hash);
326 RB_CLEAR_NODE(&rq->rb_node);
329 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
330 rq->alloc_time_ns = alloc_time_ns;
332 if (blk_mq_need_time_stamp(rq))
333 rq->start_time_ns = ktime_get_ns();
335 rq->start_time_ns = 0;
336 rq->io_start_time_ns = 0;
337 rq->stats_sectors = 0;
338 rq->nr_phys_segments = 0;
339 #if defined(CONFIG_BLK_DEV_INTEGRITY)
340 rq->nr_integrity_segments = 0;
342 /* tag was already set */
344 WRITE_ONCE(rq->deadline, 0);
349 rq->end_io_data = NULL;
351 data->ctx->rq_dispatched[op_is_sync(op)]++;
352 refcount_set(&rq->ref, 1);
356 static struct request *blk_mq_get_request(struct request_queue *q,
358 struct blk_mq_alloc_data *data)
360 struct elevator_queue *e = q->elevator;
363 bool clear_ctx_on_error = false;
364 u64 alloc_time_ns = 0;
366 blk_queue_enter_live(q);
368 /* alloc_time includes depth and tag waits */
369 if (blk_queue_rq_alloc_time(q))
370 alloc_time_ns = ktime_get_ns();
373 if (likely(!data->ctx)) {
374 data->ctx = blk_mq_get_ctx(q);
375 clear_ctx_on_error = true;
377 if (likely(!data->hctx))
378 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
380 if (data->cmd_flags & REQ_NOWAIT)
381 data->flags |= BLK_MQ_REQ_NOWAIT;
384 data->flags |= BLK_MQ_REQ_INTERNAL;
387 * Flush requests are special and go directly to the
388 * dispatch list. Don't include reserved tags in the
389 * limiting, as it isn't useful.
391 if (!op_is_flush(data->cmd_flags) &&
392 e->type->ops.limit_depth &&
393 !(data->flags & BLK_MQ_REQ_RESERVED))
394 e->type->ops.limit_depth(data->cmd_flags, data);
396 blk_mq_tag_busy(data->hctx);
399 tag = blk_mq_get_tag(data);
400 if (tag == BLK_MQ_TAG_FAIL) {
401 if (clear_ctx_on_error)
407 rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
408 if (!op_is_flush(data->cmd_flags)) {
410 if (e && e->type->ops.prepare_request) {
411 if (e->type->icq_cache)
412 blk_mq_sched_assign_ioc(rq);
414 e->type->ops.prepare_request(rq, bio);
415 rq->rq_flags |= RQF_ELVPRIV;
418 data->hctx->queued++;
422 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
423 blk_mq_req_flags_t flags)
425 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
429 ret = blk_queue_enter(q, flags);
433 rq = blk_mq_get_request(q, NULL, &alloc_data);
437 return ERR_PTR(-EWOULDBLOCK);
440 rq->__sector = (sector_t) -1;
441 rq->bio = rq->biotail = NULL;
444 EXPORT_SYMBOL(blk_mq_alloc_request);
446 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
447 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
449 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
455 * If the tag allocator sleeps we could get an allocation for a
456 * different hardware context. No need to complicate the low level
457 * allocator for this for the rare use case of a command tied to
460 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
461 return ERR_PTR(-EINVAL);
463 if (hctx_idx >= q->nr_hw_queues)
464 return ERR_PTR(-EIO);
466 ret = blk_queue_enter(q, flags);
471 * Check if the hardware context is actually mapped to anything.
472 * If not tell the caller that it should skip this queue.
474 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
475 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
477 return ERR_PTR(-EXDEV);
479 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
480 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
482 rq = blk_mq_get_request(q, NULL, &alloc_data);
486 return ERR_PTR(-EWOULDBLOCK);
490 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
492 static void __blk_mq_free_request(struct request *rq)
494 struct request_queue *q = rq->q;
495 struct blk_mq_ctx *ctx = rq->mq_ctx;
496 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
497 const int sched_tag = rq->internal_tag;
499 blk_pm_mark_last_busy(rq);
502 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
504 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
505 blk_mq_sched_restart(hctx);
509 void blk_mq_free_request(struct request *rq)
511 struct request_queue *q = rq->q;
512 struct elevator_queue *e = q->elevator;
513 struct blk_mq_ctx *ctx = rq->mq_ctx;
514 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
516 if (rq->rq_flags & RQF_ELVPRIV) {
517 if (e && e->type->ops.finish_request)
518 e->type->ops.finish_request(rq);
520 put_io_context(rq->elv.icq->ioc);
525 ctx->rq_completed[rq_is_sync(rq)]++;
526 if (rq->rq_flags & RQF_MQ_INFLIGHT)
527 atomic_dec(&hctx->nr_active);
529 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
530 laptop_io_completion(q->backing_dev_info);
534 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
535 if (refcount_dec_and_test(&rq->ref))
536 __blk_mq_free_request(rq);
538 EXPORT_SYMBOL_GPL(blk_mq_free_request);
540 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
544 if (blk_mq_need_time_stamp(rq))
545 now = ktime_get_ns();
547 if (rq->rq_flags & RQF_STATS) {
548 blk_mq_poll_stats_start(rq->q);
549 blk_stat_add(rq, now);
552 if (rq->internal_tag != -1)
553 blk_mq_sched_completed_request(rq, now);
555 blk_account_io_done(rq, now);
558 rq_qos_done(rq->q, rq);
559 rq->end_io(rq, error);
561 blk_mq_free_request(rq);
564 EXPORT_SYMBOL(__blk_mq_end_request);
566 void blk_mq_end_request(struct request *rq, blk_status_t error)
568 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
570 __blk_mq_end_request(rq, error);
572 EXPORT_SYMBOL(blk_mq_end_request);
574 static void __blk_mq_complete_request_remote(void *data)
576 struct request *rq = data;
577 struct request_queue *q = rq->q;
579 q->mq_ops->complete(rq);
582 static void __blk_mq_complete_request(struct request *rq)
584 struct blk_mq_ctx *ctx = rq->mq_ctx;
585 struct request_queue *q = rq->q;
589 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
591 * Most of single queue controllers, there is only one irq vector
592 * for handling IO completion, and the only irq's affinity is set
593 * as all possible CPUs. On most of ARCHs, this affinity means the
594 * irq is handled on one specific CPU.
596 * So complete IO reqeust in softirq context in case of single queue
597 * for not degrading IO performance by irqsoff latency.
599 if (q->nr_hw_queues == 1) {
600 __blk_complete_request(rq);
605 * For a polled request, always complete locallly, it's pointless
606 * to redirect the completion.
608 if ((rq->cmd_flags & REQ_HIPRI) ||
609 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
610 q->mq_ops->complete(rq);
615 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
616 shared = cpus_share_cache(cpu, ctx->cpu);
618 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
619 rq->csd.func = __blk_mq_complete_request_remote;
622 smp_call_function_single_async(ctx->cpu, &rq->csd);
624 q->mq_ops->complete(rq);
629 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
630 __releases(hctx->srcu)
632 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
635 srcu_read_unlock(hctx->srcu, srcu_idx);
638 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
639 __acquires(hctx->srcu)
641 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
642 /* shut up gcc false positive */
646 *srcu_idx = srcu_read_lock(hctx->srcu);
650 * blk_mq_complete_request - end I/O on a request
651 * @rq: the request being processed
654 * Ends all I/O on a request. It does not handle partial completions.
655 * The actual completion happens out-of-order, through a IPI handler.
657 bool blk_mq_complete_request(struct request *rq)
659 if (unlikely(blk_should_fake_timeout(rq->q)))
661 __blk_mq_complete_request(rq);
664 EXPORT_SYMBOL(blk_mq_complete_request);
666 int blk_mq_request_started(struct request *rq)
668 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
670 EXPORT_SYMBOL_GPL(blk_mq_request_started);
672 int blk_mq_request_completed(struct request *rq)
674 return blk_mq_rq_state(rq) == MQ_RQ_COMPLETE;
676 EXPORT_SYMBOL_GPL(blk_mq_request_completed);
678 void blk_mq_start_request(struct request *rq)
680 struct request_queue *q = rq->q;
682 trace_block_rq_issue(q, rq);
684 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
685 rq->io_start_time_ns = ktime_get_ns();
686 rq->stats_sectors = blk_rq_sectors(rq);
687 rq->rq_flags |= RQF_STATS;
691 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
694 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
696 if (q->dma_drain_size && blk_rq_bytes(rq)) {
698 * Make sure space for the drain appears. We know we can do
699 * this because max_hw_segments has been adjusted to be one
700 * fewer than the device can handle.
702 rq->nr_phys_segments++;
705 #ifdef CONFIG_BLK_DEV_INTEGRITY
706 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
707 q->integrity.profile->prepare_fn(rq);
710 EXPORT_SYMBOL(blk_mq_start_request);
712 static void __blk_mq_requeue_request(struct request *rq)
714 struct request_queue *q = rq->q;
716 blk_mq_put_driver_tag(rq);
718 trace_block_rq_requeue(q, rq);
719 rq_qos_requeue(q, rq);
721 if (blk_mq_request_started(rq)) {
722 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
723 rq->rq_flags &= ~RQF_TIMED_OUT;
724 if (q->dma_drain_size && blk_rq_bytes(rq))
725 rq->nr_phys_segments--;
729 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
731 __blk_mq_requeue_request(rq);
733 /* this request will be re-inserted to io scheduler queue */
734 blk_mq_sched_requeue_request(rq);
736 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
738 EXPORT_SYMBOL(blk_mq_requeue_request);
740 static void blk_mq_requeue_work(struct work_struct *work)
742 struct request_queue *q =
743 container_of(work, struct request_queue, requeue_work.work);
745 struct request *rq, *next;
747 spin_lock_irq(&q->requeue_lock);
748 list_splice_init(&q->requeue_list, &rq_list);
749 spin_unlock_irq(&q->requeue_lock);
751 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
752 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
755 rq->rq_flags &= ~RQF_SOFTBARRIER;
756 list_del_init(&rq->queuelist);
758 * If RQF_DONTPREP, rq has contained some driver specific
759 * data, so insert it to hctx dispatch list to avoid any
762 if (rq->rq_flags & RQF_DONTPREP)
763 blk_mq_request_bypass_insert(rq, false, false);
765 blk_mq_sched_insert_request(rq, true, false, false);
768 while (!list_empty(&rq_list)) {
769 rq = list_entry(rq_list.next, struct request, queuelist);
770 list_del_init(&rq->queuelist);
771 blk_mq_sched_insert_request(rq, false, false, false);
774 blk_mq_run_hw_queues(q, false);
777 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
778 bool kick_requeue_list)
780 struct request_queue *q = rq->q;
784 * We abuse this flag that is otherwise used by the I/O scheduler to
785 * request head insertion from the workqueue.
787 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
789 spin_lock_irqsave(&q->requeue_lock, flags);
791 rq->rq_flags |= RQF_SOFTBARRIER;
792 list_add(&rq->queuelist, &q->requeue_list);
794 list_add_tail(&rq->queuelist, &q->requeue_list);
796 spin_unlock_irqrestore(&q->requeue_lock, flags);
798 if (kick_requeue_list)
799 blk_mq_kick_requeue_list(q);
802 void blk_mq_kick_requeue_list(struct request_queue *q)
804 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
806 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
808 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
811 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
812 msecs_to_jiffies(msecs));
814 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
816 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
818 if (tag < tags->nr_tags) {
819 prefetch(tags->rqs[tag]);
820 return tags->rqs[tag];
825 EXPORT_SYMBOL(blk_mq_tag_to_rq);
827 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
828 void *priv, bool reserved)
831 * If we find a request that isn't idle and the queue matches,
832 * we know the queue is busy. Return false to stop the iteration.
834 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
844 bool blk_mq_queue_inflight(struct request_queue *q)
848 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
851 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
853 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
855 req->rq_flags |= RQF_TIMED_OUT;
856 if (req->q->mq_ops->timeout) {
857 enum blk_eh_timer_return ret;
859 ret = req->q->mq_ops->timeout(req, reserved);
860 if (ret == BLK_EH_DONE)
862 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
868 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
870 unsigned long deadline;
872 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
874 if (rq->rq_flags & RQF_TIMED_OUT)
877 deadline = READ_ONCE(rq->deadline);
878 if (time_after_eq(jiffies, deadline))
883 else if (time_after(*next, deadline))
888 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
889 struct request *rq, void *priv, bool reserved)
891 unsigned long *next = priv;
894 * Just do a quick check if it is expired before locking the request in
895 * so we're not unnecessarilly synchronizing across CPUs.
897 if (!blk_mq_req_expired(rq, next))
901 * We have reason to believe the request may be expired. Take a
902 * reference on the request to lock this request lifetime into its
903 * currently allocated context to prevent it from being reallocated in
904 * the event the completion by-passes this timeout handler.
906 * If the reference was already released, then the driver beat the
907 * timeout handler to posting a natural completion.
909 if (!refcount_inc_not_zero(&rq->ref))
913 * The request is now locked and cannot be reallocated underneath the
914 * timeout handler's processing. Re-verify this exact request is truly
915 * expired; if it is not expired, then the request was completed and
916 * reallocated as a new request.
918 if (blk_mq_req_expired(rq, next))
919 blk_mq_rq_timed_out(rq, reserved);
921 if (is_flush_rq(rq, hctx))
923 else if (refcount_dec_and_test(&rq->ref))
924 __blk_mq_free_request(rq);
929 static void blk_mq_timeout_work(struct work_struct *work)
931 struct request_queue *q =
932 container_of(work, struct request_queue, timeout_work);
933 unsigned long next = 0;
934 struct blk_mq_hw_ctx *hctx;
937 /* A deadlock might occur if a request is stuck requiring a
938 * timeout at the same time a queue freeze is waiting
939 * completion, since the timeout code would not be able to
940 * acquire the queue reference here.
942 * That's why we don't use blk_queue_enter here; instead, we use
943 * percpu_ref_tryget directly, because we need to be able to
944 * obtain a reference even in the short window between the queue
945 * starting to freeze, by dropping the first reference in
946 * blk_freeze_queue_start, and the moment the last request is
947 * consumed, marked by the instant q_usage_counter reaches
950 if (!percpu_ref_tryget(&q->q_usage_counter))
953 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
956 mod_timer(&q->timeout, next);
959 * Request timeouts are handled as a forward rolling timer. If
960 * we end up here it means that no requests are pending and
961 * also that no request has been pending for a while. Mark
964 queue_for_each_hw_ctx(q, hctx, i) {
965 /* the hctx may be unmapped, so check it here */
966 if (blk_mq_hw_queue_mapped(hctx))
967 blk_mq_tag_idle(hctx);
973 struct flush_busy_ctx_data {
974 struct blk_mq_hw_ctx *hctx;
975 struct list_head *list;
978 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
980 struct flush_busy_ctx_data *flush_data = data;
981 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
982 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
983 enum hctx_type type = hctx->type;
985 spin_lock(&ctx->lock);
986 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
987 sbitmap_clear_bit(sb, bitnr);
988 spin_unlock(&ctx->lock);
993 * Process software queues that have been marked busy, splicing them
994 * to the for-dispatch
996 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
998 struct flush_busy_ctx_data data = {
1003 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1005 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1007 struct dispatch_rq_data {
1008 struct blk_mq_hw_ctx *hctx;
1012 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1015 struct dispatch_rq_data *dispatch_data = data;
1016 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1017 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1018 enum hctx_type type = hctx->type;
1020 spin_lock(&ctx->lock);
1021 if (!list_empty(&ctx->rq_lists[type])) {
1022 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1023 list_del_init(&dispatch_data->rq->queuelist);
1024 if (list_empty(&ctx->rq_lists[type]))
1025 sbitmap_clear_bit(sb, bitnr);
1027 spin_unlock(&ctx->lock);
1029 return !dispatch_data->rq;
1032 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1033 struct blk_mq_ctx *start)
1035 unsigned off = start ? start->index_hw[hctx->type] : 0;
1036 struct dispatch_rq_data data = {
1041 __sbitmap_for_each_set(&hctx->ctx_map, off,
1042 dispatch_rq_from_ctx, &data);
1047 static inline unsigned int queued_to_index(unsigned int queued)
1052 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1055 bool blk_mq_get_driver_tag(struct request *rq)
1057 struct blk_mq_alloc_data data = {
1059 .hctx = rq->mq_hctx,
1060 .flags = BLK_MQ_REQ_NOWAIT,
1061 .cmd_flags = rq->cmd_flags,
1068 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1069 data.flags |= BLK_MQ_REQ_RESERVED;
1071 shared = blk_mq_tag_busy(data.hctx);
1072 rq->tag = blk_mq_get_tag(&data);
1075 rq->rq_flags |= RQF_MQ_INFLIGHT;
1076 atomic_inc(&data.hctx->nr_active);
1078 data.hctx->tags->rqs[rq->tag] = rq;
1082 return rq->tag != -1;
1085 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1086 int flags, void *key)
1088 struct blk_mq_hw_ctx *hctx;
1090 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1092 spin_lock(&hctx->dispatch_wait_lock);
1093 if (!list_empty(&wait->entry)) {
1094 struct sbitmap_queue *sbq;
1096 list_del_init(&wait->entry);
1097 sbq = &hctx->tags->bitmap_tags;
1098 atomic_dec(&sbq->ws_active);
1100 spin_unlock(&hctx->dispatch_wait_lock);
1102 blk_mq_run_hw_queue(hctx, true);
1107 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1108 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1109 * restart. For both cases, take care to check the condition again after
1110 * marking us as waiting.
1112 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1115 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1116 struct wait_queue_head *wq;
1117 wait_queue_entry_t *wait;
1120 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1121 blk_mq_sched_mark_restart_hctx(hctx);
1124 * It's possible that a tag was freed in the window between the
1125 * allocation failure and adding the hardware queue to the wait
1128 * Don't clear RESTART here, someone else could have set it.
1129 * At most this will cost an extra queue run.
1131 return blk_mq_get_driver_tag(rq);
1134 wait = &hctx->dispatch_wait;
1135 if (!list_empty_careful(&wait->entry))
1138 wq = &bt_wait_ptr(sbq, hctx)->wait;
1140 spin_lock_irq(&wq->lock);
1141 spin_lock(&hctx->dispatch_wait_lock);
1142 if (!list_empty(&wait->entry)) {
1143 spin_unlock(&hctx->dispatch_wait_lock);
1144 spin_unlock_irq(&wq->lock);
1148 atomic_inc(&sbq->ws_active);
1149 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1150 __add_wait_queue(wq, wait);
1153 * It's possible that a tag was freed in the window between the
1154 * allocation failure and adding the hardware queue to the wait
1157 ret = blk_mq_get_driver_tag(rq);
1159 spin_unlock(&hctx->dispatch_wait_lock);
1160 spin_unlock_irq(&wq->lock);
1165 * We got a tag, remove ourselves from the wait queue to ensure
1166 * someone else gets the wakeup.
1168 list_del_init(&wait->entry);
1169 atomic_dec(&sbq->ws_active);
1170 spin_unlock(&hctx->dispatch_wait_lock);
1171 spin_unlock_irq(&wq->lock);
1176 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1177 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1179 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1180 * - EWMA is one simple way to compute running average value
1181 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1182 * - take 4 as factor for avoiding to get too small(0) result, and this
1183 * factor doesn't matter because EWMA decreases exponentially
1185 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1189 if (hctx->queue->elevator)
1192 ewma = hctx->dispatch_busy;
1197 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1199 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1200 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1202 hctx->dispatch_busy = ewma;
1205 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1207 static void blk_mq_handle_dev_resource(struct request *rq,
1208 struct list_head *list)
1210 struct request *next =
1211 list_first_entry_or_null(list, struct request, queuelist);
1214 * If an I/O scheduler has been configured and we got a driver tag for
1215 * the next request already, free it.
1218 blk_mq_put_driver_tag(next);
1220 list_add(&rq->queuelist, list);
1221 __blk_mq_requeue_request(rq);
1225 * Returns true if we did some work AND can potentially do more.
1227 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1230 struct blk_mq_hw_ctx *hctx;
1231 struct request *rq, *nxt;
1232 bool no_tag = false;
1234 blk_status_t ret = BLK_STS_OK;
1235 bool no_budget_avail = false;
1237 if (list_empty(list))
1240 WARN_ON(!list_is_singular(list) && got_budget);
1243 * Now process all the entries, sending them to the driver.
1245 errors = queued = 0;
1247 struct blk_mq_queue_data bd;
1249 rq = list_first_entry(list, struct request, queuelist);
1252 if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) {
1253 blk_mq_put_driver_tag(rq);
1254 no_budget_avail = true;
1258 if (!blk_mq_get_driver_tag(rq)) {
1260 * The initial allocation attempt failed, so we need to
1261 * rerun the hardware queue when a tag is freed. The
1262 * waitqueue takes care of that. If the queue is run
1263 * before we add this entry back on the dispatch list,
1264 * we'll re-run it below.
1266 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1267 blk_mq_put_dispatch_budget(hctx);
1269 * For non-shared tags, the RESTART check
1272 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1278 list_del_init(&rq->queuelist);
1283 * Flag last if we have no more requests, or if we have more
1284 * but can't assign a driver tag to it.
1286 if (list_empty(list))
1289 nxt = list_first_entry(list, struct request, queuelist);
1290 bd.last = !blk_mq_get_driver_tag(nxt);
1293 ret = q->mq_ops->queue_rq(hctx, &bd);
1294 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1295 blk_mq_handle_dev_resource(rq, list);
1299 if (unlikely(ret != BLK_STS_OK)) {
1301 blk_mq_end_request(rq, BLK_STS_IOERR);
1306 } while (!list_empty(list));
1308 hctx->dispatched[queued_to_index(queued)]++;
1311 * Any items that need requeuing? Stuff them into hctx->dispatch,
1312 * that is where we will continue on next queue run.
1314 if (!list_empty(list)) {
1318 * If we didn't flush the entire list, we could have told
1319 * the driver there was more coming, but that turned out to
1322 if (q->mq_ops->commit_rqs)
1323 q->mq_ops->commit_rqs(hctx);
1325 spin_lock(&hctx->lock);
1326 list_splice_tail_init(list, &hctx->dispatch);
1327 spin_unlock(&hctx->lock);
1330 * Order adding requests to hctx->dispatch and checking
1331 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1332 * in blk_mq_sched_restart(). Avoid restart code path to
1333 * miss the new added requests to hctx->dispatch, meantime
1334 * SCHED_RESTART is observed here.
1339 * If SCHED_RESTART was set by the caller of this function and
1340 * it is no longer set that means that it was cleared by another
1341 * thread and hence that a queue rerun is needed.
1343 * If 'no_tag' is set, that means that we failed getting
1344 * a driver tag with an I/O scheduler attached. If our dispatch
1345 * waitqueue is no longer active, ensure that we run the queue
1346 * AFTER adding our entries back to the list.
1348 * If no I/O scheduler has been configured it is possible that
1349 * the hardware queue got stopped and restarted before requests
1350 * were pushed back onto the dispatch list. Rerun the queue to
1351 * avoid starvation. Notes:
1352 * - blk_mq_run_hw_queue() checks whether or not a queue has
1353 * been stopped before rerunning a queue.
1354 * - Some but not all block drivers stop a queue before
1355 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1358 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1359 * bit is set, run queue after a delay to avoid IO stalls
1360 * that could otherwise occur if the queue is idle. We'll do
1361 * similar if we couldn't get budget and SCHED_RESTART is set.
1363 needs_restart = blk_mq_sched_needs_restart(hctx);
1364 if (!needs_restart ||
1365 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1366 blk_mq_run_hw_queue(hctx, true);
1367 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1369 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1371 blk_mq_update_dispatch_busy(hctx, true);
1374 blk_mq_update_dispatch_busy(hctx, false);
1377 * If the host/device is unable to accept more work, inform the
1380 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1383 return (queued + errors) != 0;
1386 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1391 * We should be running this queue from one of the CPUs that
1394 * There are at least two related races now between setting
1395 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1396 * __blk_mq_run_hw_queue():
1398 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1399 * but later it becomes online, then this warning is harmless
1402 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1403 * but later it becomes offline, then the warning can't be
1404 * triggered, and we depend on blk-mq timeout handler to
1405 * handle dispatched requests to this hctx
1407 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1408 cpu_online(hctx->next_cpu)) {
1409 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1410 raw_smp_processor_id(),
1411 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1416 * We can't run the queue inline with ints disabled. Ensure that
1417 * we catch bad users of this early.
1419 WARN_ON_ONCE(in_interrupt());
1421 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1423 hctx_lock(hctx, &srcu_idx);
1424 blk_mq_sched_dispatch_requests(hctx);
1425 hctx_unlock(hctx, srcu_idx);
1428 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1430 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1432 if (cpu >= nr_cpu_ids)
1433 cpu = cpumask_first(hctx->cpumask);
1438 * It'd be great if the workqueue API had a way to pass
1439 * in a mask and had some smarts for more clever placement.
1440 * For now we just round-robin here, switching for every
1441 * BLK_MQ_CPU_WORK_BATCH queued items.
1443 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1446 int next_cpu = hctx->next_cpu;
1448 if (hctx->queue->nr_hw_queues == 1)
1449 return WORK_CPU_UNBOUND;
1451 if (--hctx->next_cpu_batch <= 0) {
1453 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1455 if (next_cpu >= nr_cpu_ids)
1456 next_cpu = blk_mq_first_mapped_cpu(hctx);
1457 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1461 * Do unbound schedule if we can't find a online CPU for this hctx,
1462 * and it should only happen in the path of handling CPU DEAD.
1464 if (!cpu_online(next_cpu)) {
1471 * Make sure to re-select CPU next time once after CPUs
1472 * in hctx->cpumask become online again.
1474 hctx->next_cpu = next_cpu;
1475 hctx->next_cpu_batch = 1;
1476 return WORK_CPU_UNBOUND;
1479 hctx->next_cpu = next_cpu;
1483 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1484 unsigned long msecs)
1486 if (unlikely(blk_mq_hctx_stopped(hctx)))
1489 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1490 int cpu = get_cpu();
1491 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1492 __blk_mq_run_hw_queue(hctx);
1500 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1501 msecs_to_jiffies(msecs));
1504 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1506 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1508 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1510 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1516 * When queue is quiesced, we may be switching io scheduler, or
1517 * updating nr_hw_queues, or other things, and we can't run queue
1518 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1520 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1523 hctx_lock(hctx, &srcu_idx);
1524 need_run = !blk_queue_quiesced(hctx->queue) &&
1525 blk_mq_hctx_has_pending(hctx);
1526 hctx_unlock(hctx, srcu_idx);
1529 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1535 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1537 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1539 struct blk_mq_hw_ctx *hctx;
1542 queue_for_each_hw_ctx(q, hctx, i) {
1543 if (blk_mq_hctx_stopped(hctx))
1546 blk_mq_run_hw_queue(hctx, async);
1549 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1552 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1553 * @q: request queue.
1555 * The caller is responsible for serializing this function against
1556 * blk_mq_{start,stop}_hw_queue().
1558 bool blk_mq_queue_stopped(struct request_queue *q)
1560 struct blk_mq_hw_ctx *hctx;
1563 queue_for_each_hw_ctx(q, hctx, i)
1564 if (blk_mq_hctx_stopped(hctx))
1569 EXPORT_SYMBOL(blk_mq_queue_stopped);
1572 * This function is often used for pausing .queue_rq() by driver when
1573 * there isn't enough resource or some conditions aren't satisfied, and
1574 * BLK_STS_RESOURCE is usually returned.
1576 * We do not guarantee that dispatch can be drained or blocked
1577 * after blk_mq_stop_hw_queue() returns. Please use
1578 * blk_mq_quiesce_queue() for that requirement.
1580 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1582 cancel_delayed_work(&hctx->run_work);
1584 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1586 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1589 * This function is often used for pausing .queue_rq() by driver when
1590 * there isn't enough resource or some conditions aren't satisfied, and
1591 * BLK_STS_RESOURCE is usually returned.
1593 * We do not guarantee that dispatch can be drained or blocked
1594 * after blk_mq_stop_hw_queues() returns. Please use
1595 * blk_mq_quiesce_queue() for that requirement.
1597 void blk_mq_stop_hw_queues(struct request_queue *q)
1599 struct blk_mq_hw_ctx *hctx;
1602 queue_for_each_hw_ctx(q, hctx, i)
1603 blk_mq_stop_hw_queue(hctx);
1605 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1607 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1609 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1611 blk_mq_run_hw_queue(hctx, false);
1613 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1615 void blk_mq_start_hw_queues(struct request_queue *q)
1617 struct blk_mq_hw_ctx *hctx;
1620 queue_for_each_hw_ctx(q, hctx, i)
1621 blk_mq_start_hw_queue(hctx);
1623 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1625 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1627 if (!blk_mq_hctx_stopped(hctx))
1630 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1631 blk_mq_run_hw_queue(hctx, async);
1633 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1635 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1637 struct blk_mq_hw_ctx *hctx;
1640 queue_for_each_hw_ctx(q, hctx, i)
1641 blk_mq_start_stopped_hw_queue(hctx, async);
1643 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1645 static void blk_mq_run_work_fn(struct work_struct *work)
1647 struct blk_mq_hw_ctx *hctx;
1649 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1652 * If we are stopped, don't run the queue.
1654 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1657 __blk_mq_run_hw_queue(hctx);
1660 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1664 struct blk_mq_ctx *ctx = rq->mq_ctx;
1665 enum hctx_type type = hctx->type;
1667 lockdep_assert_held(&ctx->lock);
1669 trace_block_rq_insert(hctx->queue, rq);
1672 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1674 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1677 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1680 struct blk_mq_ctx *ctx = rq->mq_ctx;
1682 lockdep_assert_held(&ctx->lock);
1684 __blk_mq_insert_req_list(hctx, rq, at_head);
1685 blk_mq_hctx_mark_pending(hctx, ctx);
1689 * Should only be used carefully, when the caller knows we want to
1690 * bypass a potential IO scheduler on the target device.
1692 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1695 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1697 spin_lock(&hctx->lock);
1699 list_add(&rq->queuelist, &hctx->dispatch);
1701 list_add_tail(&rq->queuelist, &hctx->dispatch);
1702 spin_unlock(&hctx->lock);
1705 blk_mq_run_hw_queue(hctx, false);
1708 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1709 struct list_head *list)
1713 enum hctx_type type = hctx->type;
1716 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1719 list_for_each_entry(rq, list, queuelist) {
1720 BUG_ON(rq->mq_ctx != ctx);
1721 trace_block_rq_insert(hctx->queue, rq);
1724 spin_lock(&ctx->lock);
1725 list_splice_tail_init(list, &ctx->rq_lists[type]);
1726 blk_mq_hctx_mark_pending(hctx, ctx);
1727 spin_unlock(&ctx->lock);
1730 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1732 struct request *rqa = container_of(a, struct request, queuelist);
1733 struct request *rqb = container_of(b, struct request, queuelist);
1735 if (rqa->mq_ctx < rqb->mq_ctx)
1737 else if (rqa->mq_ctx > rqb->mq_ctx)
1739 else if (rqa->mq_hctx < rqb->mq_hctx)
1741 else if (rqa->mq_hctx > rqb->mq_hctx)
1744 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1747 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1749 struct blk_mq_hw_ctx *this_hctx;
1750 struct blk_mq_ctx *this_ctx;
1751 struct request_queue *this_q;
1757 list_splice_init(&plug->mq_list, &list);
1759 if (plug->rq_count > 2 && plug->multiple_queues)
1760 list_sort(NULL, &list, plug_rq_cmp);
1769 while (!list_empty(&list)) {
1770 rq = list_entry_rq(list.next);
1771 list_del_init(&rq->queuelist);
1773 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1775 trace_block_unplug(this_q, depth, !from_schedule);
1776 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1782 this_ctx = rq->mq_ctx;
1783 this_hctx = rq->mq_hctx;
1788 list_add_tail(&rq->queuelist, &rq_list);
1792 * If 'this_hctx' is set, we know we have entries to complete
1793 * on 'rq_list'. Do those.
1796 trace_block_unplug(this_q, depth, !from_schedule);
1797 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1802 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1803 unsigned int nr_segs)
1805 if (bio->bi_opf & REQ_RAHEAD)
1806 rq->cmd_flags |= REQ_FAILFAST_MASK;
1808 rq->__sector = bio->bi_iter.bi_sector;
1809 rq->write_hint = bio->bi_write_hint;
1810 blk_rq_bio_prep(rq, bio, nr_segs);
1812 blk_account_io_start(rq, true);
1815 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1817 blk_qc_t *cookie, bool last)
1819 struct request_queue *q = rq->q;
1820 struct blk_mq_queue_data bd = {
1824 blk_qc_t new_cookie;
1827 new_cookie = request_to_qc_t(hctx, rq);
1830 * For OK queue, we are done. For error, caller may kill it.
1831 * Any other error (busy), just add it to our list as we
1832 * previously would have done.
1834 ret = q->mq_ops->queue_rq(hctx, &bd);
1837 blk_mq_update_dispatch_busy(hctx, false);
1838 *cookie = new_cookie;
1840 case BLK_STS_RESOURCE:
1841 case BLK_STS_DEV_RESOURCE:
1842 blk_mq_update_dispatch_busy(hctx, true);
1843 __blk_mq_requeue_request(rq);
1846 blk_mq_update_dispatch_busy(hctx, false);
1847 *cookie = BLK_QC_T_NONE;
1854 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1857 bool bypass_insert, bool last)
1859 struct request_queue *q = rq->q;
1860 bool run_queue = true;
1863 * RCU or SRCU read lock is needed before checking quiesced flag.
1865 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1866 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1867 * and avoid driver to try to dispatch again.
1869 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1871 bypass_insert = false;
1875 if (q->elevator && !bypass_insert)
1878 if (!blk_mq_get_dispatch_budget(hctx))
1881 if (!blk_mq_get_driver_tag(rq)) {
1882 blk_mq_put_dispatch_budget(hctx);
1886 return __blk_mq_issue_directly(hctx, rq, cookie, last);
1889 return BLK_STS_RESOURCE;
1891 blk_mq_sched_insert_request(rq, false, run_queue, false);
1896 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1897 struct request *rq, blk_qc_t *cookie)
1902 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1904 hctx_lock(hctx, &srcu_idx);
1906 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1907 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1908 blk_mq_request_bypass_insert(rq, false, true);
1909 else if (ret != BLK_STS_OK)
1910 blk_mq_end_request(rq, ret);
1912 hctx_unlock(hctx, srcu_idx);
1915 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1919 blk_qc_t unused_cookie;
1920 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1922 hctx_lock(hctx, &srcu_idx);
1923 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1924 hctx_unlock(hctx, srcu_idx);
1929 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1930 struct list_head *list)
1932 while (!list_empty(list)) {
1934 struct request *rq = list_first_entry(list, struct request,
1937 list_del_init(&rq->queuelist);
1938 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1939 if (ret != BLK_STS_OK) {
1940 if (ret == BLK_STS_RESOURCE ||
1941 ret == BLK_STS_DEV_RESOURCE) {
1942 blk_mq_request_bypass_insert(rq, false,
1946 blk_mq_end_request(rq, ret);
1951 * If we didn't flush the entire list, we could have told
1952 * the driver there was more coming, but that turned out to
1955 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1956 hctx->queue->mq_ops->commit_rqs(hctx);
1959 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1961 list_add_tail(&rq->queuelist, &plug->mq_list);
1963 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1964 struct request *tmp;
1966 tmp = list_first_entry(&plug->mq_list, struct request,
1968 if (tmp->q != rq->q)
1969 plug->multiple_queues = true;
1973 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1975 const int is_sync = op_is_sync(bio->bi_opf);
1976 const int is_flush_fua = op_is_flush(bio->bi_opf);
1977 struct blk_mq_alloc_data data = { .flags = 0};
1979 struct blk_plug *plug;
1980 struct request *same_queue_rq = NULL;
1981 unsigned int nr_segs;
1984 blk_queue_bounce(q, &bio);
1985 __blk_queue_split(q, &bio, &nr_segs);
1987 if (!bio_integrity_prep(bio))
1988 return BLK_QC_T_NONE;
1990 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1991 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1992 return BLK_QC_T_NONE;
1994 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1995 return BLK_QC_T_NONE;
1997 rq_qos_throttle(q, bio);
1999 data.cmd_flags = bio->bi_opf;
2000 rq = blk_mq_get_request(q, bio, &data);
2001 if (unlikely(!rq)) {
2002 rq_qos_cleanup(q, bio);
2003 if (bio->bi_opf & REQ_NOWAIT)
2004 bio_wouldblock_error(bio);
2005 return BLK_QC_T_NONE;
2008 trace_block_getrq(q, bio, bio->bi_opf);
2010 rq_qos_track(q, rq, bio);
2012 cookie = request_to_qc_t(data.hctx, rq);
2014 blk_mq_bio_to_request(rq, bio, nr_segs);
2016 plug = blk_mq_plug(q, bio);
2017 if (unlikely(is_flush_fua)) {
2018 /* bypass scheduler for flush rq */
2019 blk_insert_flush(rq);
2020 blk_mq_run_hw_queue(data.hctx, true);
2021 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2022 !blk_queue_nonrot(q))) {
2024 * Use plugging if we have a ->commit_rqs() hook as well, as
2025 * we know the driver uses bd->last in a smart fashion.
2027 * Use normal plugging if this disk is slow HDD, as sequential
2028 * IO may benefit a lot from plug merging.
2030 unsigned int request_count = plug->rq_count;
2031 struct request *last = NULL;
2034 trace_block_plug(q);
2036 last = list_entry_rq(plug->mq_list.prev);
2038 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2039 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2040 blk_flush_plug_list(plug, false);
2041 trace_block_plug(q);
2044 blk_add_rq_to_plug(plug, rq);
2045 } else if (q->elevator) {
2046 blk_mq_sched_insert_request(rq, false, true, true);
2047 } else if (plug && !blk_queue_nomerges(q)) {
2049 * We do limited plugging. If the bio can be merged, do that.
2050 * Otherwise the existing request in the plug list will be
2051 * issued. So the plug list will have one request at most
2052 * The plug list might get flushed before this. If that happens,
2053 * the plug list is empty, and same_queue_rq is invalid.
2055 if (list_empty(&plug->mq_list))
2056 same_queue_rq = NULL;
2057 if (same_queue_rq) {
2058 list_del_init(&same_queue_rq->queuelist);
2061 blk_add_rq_to_plug(plug, rq);
2062 trace_block_plug(q);
2064 if (same_queue_rq) {
2065 data.hctx = same_queue_rq->mq_hctx;
2066 trace_block_unplug(q, 1, true);
2067 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2070 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2071 !data.hctx->dispatch_busy) {
2072 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2074 blk_mq_sched_insert_request(rq, false, true, true);
2080 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2081 unsigned int hctx_idx)
2085 if (tags->rqs && set->ops->exit_request) {
2088 for (i = 0; i < tags->nr_tags; i++) {
2089 struct request *rq = tags->static_rqs[i];
2093 set->ops->exit_request(set, rq, hctx_idx);
2094 tags->static_rqs[i] = NULL;
2098 while (!list_empty(&tags->page_list)) {
2099 page = list_first_entry(&tags->page_list, struct page, lru);
2100 list_del_init(&page->lru);
2102 * Remove kmemleak object previously allocated in
2103 * blk_mq_alloc_rqs().
2105 kmemleak_free(page_address(page));
2106 __free_pages(page, page->private);
2110 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2114 kfree(tags->static_rqs);
2115 tags->static_rqs = NULL;
2117 blk_mq_free_tags(tags);
2120 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2121 unsigned int hctx_idx,
2122 unsigned int nr_tags,
2123 unsigned int reserved_tags)
2125 struct blk_mq_tags *tags;
2128 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2129 if (node == NUMA_NO_NODE)
2130 node = set->numa_node;
2132 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2133 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2137 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2138 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2141 blk_mq_free_tags(tags);
2145 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2146 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2148 if (!tags->static_rqs) {
2150 blk_mq_free_tags(tags);
2157 static size_t order_to_size(unsigned int order)
2159 return (size_t)PAGE_SIZE << order;
2162 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2163 unsigned int hctx_idx, int node)
2167 if (set->ops->init_request) {
2168 ret = set->ops->init_request(set, rq, hctx_idx, node);
2173 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2177 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2178 unsigned int hctx_idx, unsigned int depth)
2180 unsigned int i, j, entries_per_page, max_order = 4;
2181 size_t rq_size, left;
2184 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2185 if (node == NUMA_NO_NODE)
2186 node = set->numa_node;
2188 INIT_LIST_HEAD(&tags->page_list);
2191 * rq_size is the size of the request plus driver payload, rounded
2192 * to the cacheline size
2194 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2196 left = rq_size * depth;
2198 for (i = 0; i < depth; ) {
2199 int this_order = max_order;
2204 while (this_order && left < order_to_size(this_order - 1))
2208 page = alloc_pages_node(node,
2209 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2215 if (order_to_size(this_order) < rq_size)
2222 page->private = this_order;
2223 list_add_tail(&page->lru, &tags->page_list);
2225 p = page_address(page);
2227 * Allow kmemleak to scan these pages as they contain pointers
2228 * to additional allocations like via ops->init_request().
2230 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2231 entries_per_page = order_to_size(this_order) / rq_size;
2232 to_do = min(entries_per_page, depth - i);
2233 left -= to_do * rq_size;
2234 for (j = 0; j < to_do; j++) {
2235 struct request *rq = p;
2237 tags->static_rqs[i] = rq;
2238 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2239 tags->static_rqs[i] = NULL;
2250 blk_mq_free_rqs(set, tags, hctx_idx);
2255 * 'cpu' is going away. splice any existing rq_list entries from this
2256 * software queue to the hw queue dispatch list, and ensure that it
2259 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2261 struct blk_mq_hw_ctx *hctx;
2262 struct blk_mq_ctx *ctx;
2264 enum hctx_type type;
2266 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2267 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2270 spin_lock(&ctx->lock);
2271 if (!list_empty(&ctx->rq_lists[type])) {
2272 list_splice_init(&ctx->rq_lists[type], &tmp);
2273 blk_mq_hctx_clear_pending(hctx, ctx);
2275 spin_unlock(&ctx->lock);
2277 if (list_empty(&tmp))
2280 spin_lock(&hctx->lock);
2281 list_splice_tail_init(&tmp, &hctx->dispatch);
2282 spin_unlock(&hctx->lock);
2284 blk_mq_run_hw_queue(hctx, true);
2288 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2290 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2294 /* hctx->ctxs will be freed in queue's release handler */
2295 static void blk_mq_exit_hctx(struct request_queue *q,
2296 struct blk_mq_tag_set *set,
2297 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2299 if (blk_mq_hw_queue_mapped(hctx))
2300 blk_mq_tag_idle(hctx);
2302 if (set->ops->exit_request)
2303 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2305 if (set->ops->exit_hctx)
2306 set->ops->exit_hctx(hctx, hctx_idx);
2308 blk_mq_remove_cpuhp(hctx);
2310 spin_lock(&q->unused_hctx_lock);
2311 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2312 spin_unlock(&q->unused_hctx_lock);
2315 static void blk_mq_exit_hw_queues(struct request_queue *q,
2316 struct blk_mq_tag_set *set, int nr_queue)
2318 struct blk_mq_hw_ctx *hctx;
2321 queue_for_each_hw_ctx(q, hctx, i) {
2324 blk_mq_debugfs_unregister_hctx(hctx);
2325 blk_mq_exit_hctx(q, set, hctx, i);
2329 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2331 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2333 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2334 __alignof__(struct blk_mq_hw_ctx)) !=
2335 sizeof(struct blk_mq_hw_ctx));
2337 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2338 hw_ctx_size += sizeof(struct srcu_struct);
2343 static int blk_mq_init_hctx(struct request_queue *q,
2344 struct blk_mq_tag_set *set,
2345 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2347 hctx->queue_num = hctx_idx;
2349 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2351 hctx->tags = set->tags[hctx_idx];
2353 if (set->ops->init_hctx &&
2354 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2355 goto unregister_cpu_notifier;
2357 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2363 if (set->ops->exit_hctx)
2364 set->ops->exit_hctx(hctx, hctx_idx);
2365 unregister_cpu_notifier:
2366 blk_mq_remove_cpuhp(hctx);
2370 static struct blk_mq_hw_ctx *
2371 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2374 struct blk_mq_hw_ctx *hctx;
2375 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2377 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2379 goto fail_alloc_hctx;
2381 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2384 atomic_set(&hctx->nr_active, 0);
2385 if (node == NUMA_NO_NODE)
2386 node = set->numa_node;
2387 hctx->numa_node = node;
2389 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2390 spin_lock_init(&hctx->lock);
2391 INIT_LIST_HEAD(&hctx->dispatch);
2393 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2395 INIT_LIST_HEAD(&hctx->hctx_list);
2398 * Allocate space for all possible cpus to avoid allocation at
2401 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2406 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2411 spin_lock_init(&hctx->dispatch_wait_lock);
2412 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2413 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2415 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2420 if (hctx->flags & BLK_MQ_F_BLOCKING)
2421 init_srcu_struct(hctx->srcu);
2422 blk_mq_hctx_kobj_init(hctx);
2427 sbitmap_free(&hctx->ctx_map);
2431 free_cpumask_var(hctx->cpumask);
2438 static void blk_mq_init_cpu_queues(struct request_queue *q,
2439 unsigned int nr_hw_queues)
2441 struct blk_mq_tag_set *set = q->tag_set;
2444 for_each_possible_cpu(i) {
2445 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2446 struct blk_mq_hw_ctx *hctx;
2450 spin_lock_init(&__ctx->lock);
2451 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2452 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2457 * Set local node, IFF we have more than one hw queue. If
2458 * not, we remain on the home node of the device
2460 for (j = 0; j < set->nr_maps; j++) {
2461 hctx = blk_mq_map_queue_type(q, j, i);
2462 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2463 hctx->numa_node = local_memory_node(cpu_to_node(i));
2468 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2472 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2473 set->queue_depth, set->reserved_tags);
2474 if (!set->tags[hctx_idx])
2477 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2482 blk_mq_free_rq_map(set->tags[hctx_idx]);
2483 set->tags[hctx_idx] = NULL;
2487 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2488 unsigned int hctx_idx)
2490 if (set->tags && set->tags[hctx_idx]) {
2491 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2492 blk_mq_free_rq_map(set->tags[hctx_idx]);
2493 set->tags[hctx_idx] = NULL;
2497 static void blk_mq_map_swqueue(struct request_queue *q)
2499 unsigned int i, j, hctx_idx;
2500 struct blk_mq_hw_ctx *hctx;
2501 struct blk_mq_ctx *ctx;
2502 struct blk_mq_tag_set *set = q->tag_set;
2504 queue_for_each_hw_ctx(q, hctx, i) {
2505 cpumask_clear(hctx->cpumask);
2507 hctx->dispatch_from = NULL;
2511 * Map software to hardware queues.
2513 * If the cpu isn't present, the cpu is mapped to first hctx.
2515 for_each_possible_cpu(i) {
2517 ctx = per_cpu_ptr(q->queue_ctx, i);
2518 for (j = 0; j < set->nr_maps; j++) {
2519 if (!set->map[j].nr_queues) {
2520 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2521 HCTX_TYPE_DEFAULT, i);
2524 hctx_idx = set->map[j].mq_map[i];
2525 /* unmapped hw queue can be remapped after CPU topo changed */
2526 if (!set->tags[hctx_idx] &&
2527 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2529 * If tags initialization fail for some hctx,
2530 * that hctx won't be brought online. In this
2531 * case, remap the current ctx to hctx[0] which
2532 * is guaranteed to always have tags allocated
2534 set->map[j].mq_map[i] = 0;
2537 hctx = blk_mq_map_queue_type(q, j, i);
2538 ctx->hctxs[j] = hctx;
2540 * If the CPU is already set in the mask, then we've
2541 * mapped this one already. This can happen if
2542 * devices share queues across queue maps.
2544 if (cpumask_test_cpu(i, hctx->cpumask))
2547 cpumask_set_cpu(i, hctx->cpumask);
2549 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2550 hctx->ctxs[hctx->nr_ctx++] = ctx;
2553 * If the nr_ctx type overflows, we have exceeded the
2554 * amount of sw queues we can support.
2556 BUG_ON(!hctx->nr_ctx);
2559 for (; j < HCTX_MAX_TYPES; j++)
2560 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2561 HCTX_TYPE_DEFAULT, i);
2564 queue_for_each_hw_ctx(q, hctx, i) {
2566 * If no software queues are mapped to this hardware queue,
2567 * disable it and free the request entries.
2569 if (!hctx->nr_ctx) {
2570 /* Never unmap queue 0. We need it as a
2571 * fallback in case of a new remap fails
2574 if (i && set->tags[i])
2575 blk_mq_free_map_and_requests(set, i);
2581 hctx->tags = set->tags[i];
2582 WARN_ON(!hctx->tags);
2585 * Set the map size to the number of mapped software queues.
2586 * This is more accurate and more efficient than looping
2587 * over all possibly mapped software queues.
2589 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2592 * Initialize batch roundrobin counts
2594 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2595 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2600 * Caller needs to ensure that we're either frozen/quiesced, or that
2601 * the queue isn't live yet.
2603 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2605 struct blk_mq_hw_ctx *hctx;
2608 queue_for_each_hw_ctx(q, hctx, i) {
2610 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2612 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2616 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2619 struct request_queue *q;
2621 lockdep_assert_held(&set->tag_list_lock);
2623 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2624 blk_mq_freeze_queue(q);
2625 queue_set_hctx_shared(q, shared);
2626 blk_mq_unfreeze_queue(q);
2630 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2632 struct blk_mq_tag_set *set = q->tag_set;
2634 mutex_lock(&set->tag_list_lock);
2635 list_del_rcu(&q->tag_set_list);
2636 if (list_is_singular(&set->tag_list)) {
2637 /* just transitioned to unshared */
2638 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2639 /* update existing queue */
2640 blk_mq_update_tag_set_depth(set, false);
2642 mutex_unlock(&set->tag_list_lock);
2643 INIT_LIST_HEAD(&q->tag_set_list);
2646 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2647 struct request_queue *q)
2649 mutex_lock(&set->tag_list_lock);
2652 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2654 if (!list_empty(&set->tag_list) &&
2655 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2656 set->flags |= BLK_MQ_F_TAG_SHARED;
2657 /* update existing queue */
2658 blk_mq_update_tag_set_depth(set, true);
2660 if (set->flags & BLK_MQ_F_TAG_SHARED)
2661 queue_set_hctx_shared(q, true);
2662 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2664 mutex_unlock(&set->tag_list_lock);
2667 /* All allocations will be freed in release handler of q->mq_kobj */
2668 static int blk_mq_alloc_ctxs(struct request_queue *q)
2670 struct blk_mq_ctxs *ctxs;
2673 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2677 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2678 if (!ctxs->queue_ctx)
2681 for_each_possible_cpu(cpu) {
2682 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2686 q->mq_kobj = &ctxs->kobj;
2687 q->queue_ctx = ctxs->queue_ctx;
2696 * It is the actual release handler for mq, but we do it from
2697 * request queue's release handler for avoiding use-after-free
2698 * and headache because q->mq_kobj shouldn't have been introduced,
2699 * but we can't group ctx/kctx kobj without it.
2701 void blk_mq_release(struct request_queue *q)
2703 struct blk_mq_hw_ctx *hctx, *next;
2706 queue_for_each_hw_ctx(q, hctx, i)
2707 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2709 /* all hctx are in .unused_hctx_list now */
2710 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2711 list_del_init(&hctx->hctx_list);
2712 kobject_put(&hctx->kobj);
2715 kfree(q->queue_hw_ctx);
2718 * release .mq_kobj and sw queue's kobject now because
2719 * both share lifetime with request queue.
2721 blk_mq_sysfs_deinit(q);
2724 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2726 struct request_queue *uninit_q, *q;
2728 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2730 return ERR_PTR(-ENOMEM);
2733 * Initialize the queue without an elevator. device_add_disk() will do
2734 * the initialization.
2736 q = blk_mq_init_allocated_queue(set, uninit_q, false);
2738 blk_cleanup_queue(uninit_q);
2742 EXPORT_SYMBOL(blk_mq_init_queue);
2745 * Helper for setting up a queue with mq ops, given queue depth, and
2746 * the passed in mq ops flags.
2748 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2749 const struct blk_mq_ops *ops,
2750 unsigned int queue_depth,
2751 unsigned int set_flags)
2753 struct request_queue *q;
2756 memset(set, 0, sizeof(*set));
2758 set->nr_hw_queues = 1;
2760 set->queue_depth = queue_depth;
2761 set->numa_node = NUMA_NO_NODE;
2762 set->flags = set_flags;
2764 ret = blk_mq_alloc_tag_set(set);
2766 return ERR_PTR(ret);
2768 q = blk_mq_init_queue(set);
2770 blk_mq_free_tag_set(set);
2776 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2778 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2779 struct blk_mq_tag_set *set, struct request_queue *q,
2780 int hctx_idx, int node)
2782 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2784 /* reuse dead hctx first */
2785 spin_lock(&q->unused_hctx_lock);
2786 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2787 if (tmp->numa_node == node) {
2793 list_del_init(&hctx->hctx_list);
2794 spin_unlock(&q->unused_hctx_lock);
2797 hctx = blk_mq_alloc_hctx(q, set, node);
2801 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2807 kobject_put(&hctx->kobj);
2812 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2813 struct request_queue *q)
2816 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2818 /* protect against switching io scheduler */
2819 mutex_lock(&q->sysfs_lock);
2820 for (i = 0; i < set->nr_hw_queues; i++) {
2822 struct blk_mq_hw_ctx *hctx;
2824 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2826 * If the hw queue has been mapped to another numa node,
2827 * we need to realloc the hctx. If allocation fails, fallback
2828 * to use the previous one.
2830 if (hctxs[i] && (hctxs[i]->numa_node == node))
2833 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2836 blk_mq_exit_hctx(q, set, hctxs[i], i);
2840 pr_warn("Allocate new hctx on node %d fails,\
2841 fallback to previous one on node %d\n",
2842 node, hctxs[i]->numa_node);
2848 * Increasing nr_hw_queues fails. Free the newly allocated
2849 * hctxs and keep the previous q->nr_hw_queues.
2851 if (i != set->nr_hw_queues) {
2852 j = q->nr_hw_queues;
2856 end = q->nr_hw_queues;
2857 q->nr_hw_queues = set->nr_hw_queues;
2860 for (; j < end; j++) {
2861 struct blk_mq_hw_ctx *hctx = hctxs[j];
2865 blk_mq_free_map_and_requests(set, j);
2866 blk_mq_exit_hctx(q, set, hctx, j);
2870 mutex_unlock(&q->sysfs_lock);
2874 * Maximum number of hardware queues we support. For single sets, we'll never
2875 * have more than the CPUs (software queues). For multiple sets, the tag_set
2876 * user may have set ->nr_hw_queues larger.
2878 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2880 if (set->nr_maps == 1)
2883 return max(set->nr_hw_queues, nr_cpu_ids);
2886 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2887 struct request_queue *q,
2890 /* mark the queue as mq asap */
2891 q->mq_ops = set->ops;
2893 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2894 blk_mq_poll_stats_bkt,
2895 BLK_MQ_POLL_STATS_BKTS, q);
2899 if (blk_mq_alloc_ctxs(q))
2902 /* init q->mq_kobj and sw queues' kobjects */
2903 blk_mq_sysfs_init(q);
2905 q->nr_queues = nr_hw_queues(set);
2906 q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2907 GFP_KERNEL, set->numa_node);
2908 if (!q->queue_hw_ctx)
2911 INIT_LIST_HEAD(&q->unused_hctx_list);
2912 spin_lock_init(&q->unused_hctx_lock);
2914 blk_mq_realloc_hw_ctxs(set, q);
2915 if (!q->nr_hw_queues)
2918 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2919 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2923 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2924 if (set->nr_maps > HCTX_TYPE_POLL &&
2925 set->map[HCTX_TYPE_POLL].nr_queues)
2926 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2928 q->sg_reserved_size = INT_MAX;
2930 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2931 INIT_LIST_HEAD(&q->requeue_list);
2932 spin_lock_init(&q->requeue_lock);
2934 blk_queue_make_request(q, blk_mq_make_request);
2937 * Do this after blk_queue_make_request() overrides it...
2939 q->nr_requests = set->queue_depth;
2942 * Default to classic polling
2944 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2946 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2947 blk_mq_add_queue_tag_set(set, q);
2948 blk_mq_map_swqueue(q);
2951 elevator_init_mq(q);
2956 kfree(q->queue_hw_ctx);
2957 q->nr_hw_queues = 0;
2959 blk_mq_sysfs_deinit(q);
2961 blk_stat_free_callback(q->poll_cb);
2965 return ERR_PTR(-ENOMEM);
2967 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2969 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2970 void blk_mq_exit_queue(struct request_queue *q)
2972 struct blk_mq_tag_set *set = q->tag_set;
2974 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
2975 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2976 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
2977 blk_mq_del_queue_tag_set(q);
2980 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2984 for (i = 0; i < set->nr_hw_queues; i++)
2985 if (!__blk_mq_alloc_rq_map(set, i))
2992 blk_mq_free_rq_map(set->tags[i]);
2998 * Allocate the request maps associated with this tag_set. Note that this
2999 * may reduce the depth asked for, if memory is tight. set->queue_depth
3000 * will be updated to reflect the allocated depth.
3002 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3007 depth = set->queue_depth;
3009 err = __blk_mq_alloc_rq_maps(set);
3013 set->queue_depth >>= 1;
3014 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3018 } while (set->queue_depth);
3020 if (!set->queue_depth || err) {
3021 pr_err("blk-mq: failed to allocate request map\n");
3025 if (depth != set->queue_depth)
3026 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3027 depth, set->queue_depth);
3032 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3035 * blk_mq_map_queues() and multiple .map_queues() implementations
3036 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3037 * number of hardware queues.
3039 if (set->nr_maps == 1)
3040 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3042 if (set->ops->map_queues && !is_kdump_kernel()) {
3046 * transport .map_queues is usually done in the following
3049 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3050 * mask = get_cpu_mask(queue)
3051 * for_each_cpu(cpu, mask)
3052 * set->map[x].mq_map[cpu] = queue;
3055 * When we need to remap, the table has to be cleared for
3056 * killing stale mapping since one CPU may not be mapped
3059 for (i = 0; i < set->nr_maps; i++)
3060 blk_mq_clear_mq_map(&set->map[i]);
3062 return set->ops->map_queues(set);
3064 BUG_ON(set->nr_maps > 1);
3065 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3070 * Alloc a tag set to be associated with one or more request queues.
3071 * May fail with EINVAL for various error conditions. May adjust the
3072 * requested depth down, if it's too large. In that case, the set
3073 * value will be stored in set->queue_depth.
3075 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3079 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3081 if (!set->nr_hw_queues)
3083 if (!set->queue_depth)
3085 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3088 if (!set->ops->queue_rq)
3091 if (!set->ops->get_budget ^ !set->ops->put_budget)
3094 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3095 pr_info("blk-mq: reduced tag depth to %u\n",
3097 set->queue_depth = BLK_MQ_MAX_DEPTH;
3102 else if (set->nr_maps > HCTX_MAX_TYPES)
3106 * If a crashdump is active, then we are potentially in a very
3107 * memory constrained environment. Limit us to 1 queue and
3108 * 64 tags to prevent using too much memory.
3110 if (is_kdump_kernel()) {
3111 set->nr_hw_queues = 1;
3113 set->queue_depth = min(64U, set->queue_depth);
3116 * There is no use for more h/w queues than cpus if we just have
3119 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3120 set->nr_hw_queues = nr_cpu_ids;
3122 set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3123 GFP_KERNEL, set->numa_node);
3128 for (i = 0; i < set->nr_maps; i++) {
3129 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3130 sizeof(set->map[i].mq_map[0]),
3131 GFP_KERNEL, set->numa_node);
3132 if (!set->map[i].mq_map)
3133 goto out_free_mq_map;
3134 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3137 ret = blk_mq_update_queue_map(set);
3139 goto out_free_mq_map;
3141 ret = blk_mq_alloc_rq_maps(set);
3143 goto out_free_mq_map;
3145 mutex_init(&set->tag_list_lock);
3146 INIT_LIST_HEAD(&set->tag_list);
3151 for (i = 0; i < set->nr_maps; i++) {
3152 kfree(set->map[i].mq_map);
3153 set->map[i].mq_map = NULL;
3159 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3161 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3165 for (i = 0; i < nr_hw_queues(set); i++)
3166 blk_mq_free_map_and_requests(set, i);
3168 for (j = 0; j < set->nr_maps; j++) {
3169 kfree(set->map[j].mq_map);
3170 set->map[j].mq_map = NULL;
3176 EXPORT_SYMBOL(blk_mq_free_tag_set);
3178 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3180 struct blk_mq_tag_set *set = q->tag_set;
3181 struct blk_mq_hw_ctx *hctx;
3187 if (q->nr_requests == nr)
3190 blk_mq_freeze_queue(q);
3191 blk_mq_quiesce_queue(q);
3194 queue_for_each_hw_ctx(q, hctx, i) {
3198 * If we're using an MQ scheduler, just update the scheduler
3199 * queue depth. This is similar to what the old code would do.
3201 if (!hctx->sched_tags) {
3202 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3205 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3210 if (q->elevator && q->elevator->type->ops.depth_updated)
3211 q->elevator->type->ops.depth_updated(hctx);
3215 q->nr_requests = nr;
3217 blk_mq_unquiesce_queue(q);
3218 blk_mq_unfreeze_queue(q);
3224 * request_queue and elevator_type pair.
3225 * It is just used by __blk_mq_update_nr_hw_queues to cache
3226 * the elevator_type associated with a request_queue.
3228 struct blk_mq_qe_pair {
3229 struct list_head node;
3230 struct request_queue *q;
3231 struct elevator_type *type;
3235 * Cache the elevator_type in qe pair list and switch the
3236 * io scheduler to 'none'
3238 static bool blk_mq_elv_switch_none(struct list_head *head,
3239 struct request_queue *q)
3241 struct blk_mq_qe_pair *qe;
3246 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3250 INIT_LIST_HEAD(&qe->node);
3252 qe->type = q->elevator->type;
3253 list_add(&qe->node, head);
3255 mutex_lock(&q->sysfs_lock);
3257 * After elevator_switch_mq, the previous elevator_queue will be
3258 * released by elevator_release. The reference of the io scheduler
3259 * module get by elevator_get will also be put. So we need to get
3260 * a reference of the io scheduler module here to prevent it to be
3263 __module_get(qe->type->elevator_owner);
3264 elevator_switch_mq(q, NULL);
3265 mutex_unlock(&q->sysfs_lock);
3270 static void blk_mq_elv_switch_back(struct list_head *head,
3271 struct request_queue *q)
3273 struct blk_mq_qe_pair *qe;
3274 struct elevator_type *t = NULL;
3276 list_for_each_entry(qe, head, node)
3285 list_del(&qe->node);
3288 mutex_lock(&q->sysfs_lock);
3289 elevator_switch_mq(q, t);
3290 mutex_unlock(&q->sysfs_lock);
3293 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3296 struct request_queue *q;
3298 int prev_nr_hw_queues;
3300 lockdep_assert_held(&set->tag_list_lock);
3302 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3303 nr_hw_queues = nr_cpu_ids;
3304 if (nr_hw_queues < 1)
3306 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3309 list_for_each_entry(q, &set->tag_list, tag_set_list)
3310 blk_mq_freeze_queue(q);
3312 * Sync with blk_mq_queue_tag_busy_iter.
3316 * Switch IO scheduler to 'none', cleaning up the data associated
3317 * with the previous scheduler. We will switch back once we are done
3318 * updating the new sw to hw queue mappings.
3320 list_for_each_entry(q, &set->tag_list, tag_set_list)
3321 if (!blk_mq_elv_switch_none(&head, q))
3324 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3325 blk_mq_debugfs_unregister_hctxs(q);
3326 blk_mq_sysfs_unregister(q);
3329 prev_nr_hw_queues = set->nr_hw_queues;
3330 set->nr_hw_queues = nr_hw_queues;
3332 blk_mq_update_queue_map(set);
3333 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3334 blk_mq_realloc_hw_ctxs(set, q);
3335 if (q->nr_hw_queues != set->nr_hw_queues) {
3336 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3337 nr_hw_queues, prev_nr_hw_queues);
3338 set->nr_hw_queues = prev_nr_hw_queues;
3339 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3342 blk_mq_map_swqueue(q);
3345 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3346 blk_mq_sysfs_register(q);
3347 blk_mq_debugfs_register_hctxs(q);
3351 list_for_each_entry(q, &set->tag_list, tag_set_list)
3352 blk_mq_elv_switch_back(&head, q);
3354 list_for_each_entry(q, &set->tag_list, tag_set_list)
3355 blk_mq_unfreeze_queue(q);
3358 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3360 mutex_lock(&set->tag_list_lock);
3361 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3362 mutex_unlock(&set->tag_list_lock);
3364 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3366 /* Enable polling stats and return whether they were already enabled. */
3367 static bool blk_poll_stats_enable(struct request_queue *q)
3369 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3370 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3372 blk_stat_add_callback(q, q->poll_cb);
3376 static void blk_mq_poll_stats_start(struct request_queue *q)
3379 * We don't arm the callback if polling stats are not enabled or the
3380 * callback is already active.
3382 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3383 blk_stat_is_active(q->poll_cb))
3386 blk_stat_activate_msecs(q->poll_cb, 100);
3389 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3391 struct request_queue *q = cb->data;
3394 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3395 if (cb->stat[bucket].nr_samples)
3396 q->poll_stat[bucket] = cb->stat[bucket];
3400 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3401 struct blk_mq_hw_ctx *hctx,
3404 unsigned long ret = 0;
3408 * If stats collection isn't on, don't sleep but turn it on for
3411 if (!blk_poll_stats_enable(q))
3415 * As an optimistic guess, use half of the mean service time
3416 * for this type of request. We can (and should) make this smarter.
3417 * For instance, if the completion latencies are tight, we can
3418 * get closer than just half the mean. This is especially
3419 * important on devices where the completion latencies are longer
3420 * than ~10 usec. We do use the stats for the relevant IO size
3421 * if available which does lead to better estimates.
3423 bucket = blk_mq_poll_stats_bkt(rq);
3427 if (q->poll_stat[bucket].nr_samples)
3428 ret = (q->poll_stat[bucket].mean + 1) / 2;
3433 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3434 struct blk_mq_hw_ctx *hctx,
3437 struct hrtimer_sleeper hs;
3438 enum hrtimer_mode mode;
3442 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3446 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3448 * 0: use half of prev avg
3449 * >0: use this specific value
3451 if (q->poll_nsec > 0)
3452 nsecs = q->poll_nsec;
3454 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3459 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3462 * This will be replaced with the stats tracking code, using
3463 * 'avg_completion_time / 2' as the pre-sleep target.
3467 mode = HRTIMER_MODE_REL;
3468 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3469 hrtimer_set_expires(&hs.timer, kt);
3472 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3474 set_current_state(TASK_UNINTERRUPTIBLE);
3475 hrtimer_sleeper_start_expires(&hs, mode);
3478 hrtimer_cancel(&hs.timer);
3479 mode = HRTIMER_MODE_ABS;
3480 } while (hs.task && !signal_pending(current));
3482 __set_current_state(TASK_RUNNING);
3483 destroy_hrtimer_on_stack(&hs.timer);
3487 static bool blk_mq_poll_hybrid(struct request_queue *q,
3488 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3492 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3495 if (!blk_qc_t_is_internal(cookie))
3496 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3498 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3500 * With scheduling, if the request has completed, we'll
3501 * get a NULL return here, as we clear the sched tag when
3502 * that happens. The request still remains valid, like always,
3503 * so we should be safe with just the NULL check.
3509 return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3513 * blk_poll - poll for IO completions
3515 * @cookie: cookie passed back at IO submission time
3516 * @spin: whether to spin for completions
3519 * Poll for completions on the passed in queue. Returns number of
3520 * completed entries found. If @spin is true, then blk_poll will continue
3521 * looping until at least one completion is found, unless the task is
3522 * otherwise marked running (or we need to reschedule).
3524 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3526 struct blk_mq_hw_ctx *hctx;
3529 if (!blk_qc_t_valid(cookie) ||
3530 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3534 blk_flush_plug_list(current->plug, false);
3536 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3539 * If we sleep, have the caller restart the poll loop to reset
3540 * the state. Like for the other success return cases, the
3541 * caller is responsible for checking if the IO completed. If
3542 * the IO isn't complete, we'll get called again and will go
3543 * straight to the busy poll loop.
3545 if (blk_mq_poll_hybrid(q, hctx, cookie))
3548 hctx->poll_considered++;
3550 state = current->state;
3554 hctx->poll_invoked++;
3556 ret = q->mq_ops->poll(hctx);
3558 hctx->poll_success++;
3559 __set_current_state(TASK_RUNNING);
3563 if (signal_pending_state(state, current))
3564 __set_current_state(TASK_RUNNING);
3566 if (current->state == TASK_RUNNING)
3568 if (ret < 0 || !spin)
3571 } while (!need_resched());
3573 __set_current_state(TASK_RUNNING);
3576 EXPORT_SYMBOL_GPL(blk_poll);
3578 unsigned int blk_mq_rq_cpu(struct request *rq)
3580 return rq->mq_ctx->cpu;
3582 EXPORT_SYMBOL(blk_mq_rq_cpu);
3584 static int __init blk_mq_init(void)
3586 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3587 blk_mq_hctx_notify_dead);
3590 subsys_initcall(blk_mq_init);