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, bytes, bucket;
49 ddir = rq_data_dir(rq);
50 bytes = blk_rq_bytes(rq);
52 bucket = ddir + 2*(ilog2(bytes) - 9);
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)
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 if (blk_mq_need_time_stamp(rq))
329 rq->start_time_ns = ktime_get_ns();
331 rq->start_time_ns = 0;
332 rq->io_start_time_ns = 0;
333 rq->nr_phys_segments = 0;
334 #if defined(CONFIG_BLK_DEV_INTEGRITY)
335 rq->nr_integrity_segments = 0;
337 /* tag was already set */
339 WRITE_ONCE(rq->deadline, 0);
344 rq->end_io_data = NULL;
346 data->ctx->rq_dispatched[op_is_sync(op)]++;
347 refcount_set(&rq->ref, 1);
351 static struct request *blk_mq_get_request(struct request_queue *q,
353 struct blk_mq_alloc_data *data)
355 struct elevator_queue *e = q->elevator;
358 bool clear_ctx_on_error = false;
360 blk_queue_enter_live(q);
362 if (likely(!data->ctx)) {
363 data->ctx = blk_mq_get_ctx(q);
364 clear_ctx_on_error = true;
366 if (likely(!data->hctx))
367 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
369 if (data->cmd_flags & REQ_NOWAIT)
370 data->flags |= BLK_MQ_REQ_NOWAIT;
373 data->flags |= BLK_MQ_REQ_INTERNAL;
376 * Flush requests are special and go directly to the
377 * dispatch list. Don't include reserved tags in the
378 * limiting, as it isn't useful.
380 if (!op_is_flush(data->cmd_flags) &&
381 e->type->ops.limit_depth &&
382 !(data->flags & BLK_MQ_REQ_RESERVED))
383 e->type->ops.limit_depth(data->cmd_flags, data);
385 blk_mq_tag_busy(data->hctx);
388 tag = blk_mq_get_tag(data);
389 if (tag == BLK_MQ_TAG_FAIL) {
390 if (clear_ctx_on_error)
396 rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags);
397 if (!op_is_flush(data->cmd_flags)) {
399 if (e && e->type->ops.prepare_request) {
400 if (e->type->icq_cache)
401 blk_mq_sched_assign_ioc(rq);
403 e->type->ops.prepare_request(rq, bio);
404 rq->rq_flags |= RQF_ELVPRIV;
407 data->hctx->queued++;
411 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
412 blk_mq_req_flags_t flags)
414 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
418 ret = blk_queue_enter(q, flags);
422 rq = blk_mq_get_request(q, NULL, &alloc_data);
426 return ERR_PTR(-EWOULDBLOCK);
429 rq->__sector = (sector_t) -1;
430 rq->bio = rq->biotail = NULL;
433 EXPORT_SYMBOL(blk_mq_alloc_request);
435 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
436 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
438 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
444 * If the tag allocator sleeps we could get an allocation for a
445 * different hardware context. No need to complicate the low level
446 * allocator for this for the rare use case of a command tied to
449 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
450 return ERR_PTR(-EINVAL);
452 if (hctx_idx >= q->nr_hw_queues)
453 return ERR_PTR(-EIO);
455 ret = blk_queue_enter(q, flags);
460 * Check if the hardware context is actually mapped to anything.
461 * If not tell the caller that it should skip this queue.
463 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
464 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
466 return ERR_PTR(-EXDEV);
468 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
469 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
471 rq = blk_mq_get_request(q, NULL, &alloc_data);
475 return ERR_PTR(-EWOULDBLOCK);
479 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
481 static void __blk_mq_free_request(struct request *rq)
483 struct request_queue *q = rq->q;
484 struct blk_mq_ctx *ctx = rq->mq_ctx;
485 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
486 const int sched_tag = rq->internal_tag;
488 blk_pm_mark_last_busy(rq);
491 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
493 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
494 blk_mq_sched_restart(hctx);
498 void blk_mq_free_request(struct request *rq)
500 struct request_queue *q = rq->q;
501 struct elevator_queue *e = q->elevator;
502 struct blk_mq_ctx *ctx = rq->mq_ctx;
503 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
505 if (rq->rq_flags & RQF_ELVPRIV) {
506 if (e && e->type->ops.finish_request)
507 e->type->ops.finish_request(rq);
509 put_io_context(rq->elv.icq->ioc);
514 ctx->rq_completed[rq_is_sync(rq)]++;
515 if (rq->rq_flags & RQF_MQ_INFLIGHT)
516 atomic_dec(&hctx->nr_active);
518 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
519 laptop_io_completion(q->backing_dev_info);
523 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
524 if (refcount_dec_and_test(&rq->ref))
525 __blk_mq_free_request(rq);
527 EXPORT_SYMBOL_GPL(blk_mq_free_request);
529 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
533 if (blk_mq_need_time_stamp(rq))
534 now = ktime_get_ns();
536 if (rq->rq_flags & RQF_STATS) {
537 blk_mq_poll_stats_start(rq->q);
538 blk_stat_add(rq, now);
541 if (rq->internal_tag != -1)
542 blk_mq_sched_completed_request(rq, now);
544 blk_account_io_done(rq, now);
547 rq_qos_done(rq->q, rq);
548 rq->end_io(rq, error);
550 blk_mq_free_request(rq);
553 EXPORT_SYMBOL(__blk_mq_end_request);
555 void blk_mq_end_request(struct request *rq, blk_status_t error)
557 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
559 __blk_mq_end_request(rq, error);
561 EXPORT_SYMBOL(blk_mq_end_request);
563 static void __blk_mq_complete_request_remote(void *data)
565 struct request *rq = data;
566 struct request_queue *q = rq->q;
568 q->mq_ops->complete(rq);
571 static void __blk_mq_complete_request(struct request *rq)
573 struct blk_mq_ctx *ctx = rq->mq_ctx;
574 struct request_queue *q = rq->q;
578 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
580 * Most of single queue controllers, there is only one irq vector
581 * for handling IO completion, and the only irq's affinity is set
582 * as all possible CPUs. On most of ARCHs, this affinity means the
583 * irq is handled on one specific CPU.
585 * So complete IO reqeust in softirq context in case of single queue
586 * for not degrading IO performance by irqsoff latency.
588 if (q->nr_hw_queues == 1) {
589 __blk_complete_request(rq);
594 * For a polled request, always complete locallly, it's pointless
595 * to redirect the completion.
597 if ((rq->cmd_flags & REQ_HIPRI) ||
598 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
599 q->mq_ops->complete(rq);
604 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
605 shared = cpus_share_cache(cpu, ctx->cpu);
607 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
608 rq->csd.func = __blk_mq_complete_request_remote;
611 smp_call_function_single_async(ctx->cpu, &rq->csd);
613 q->mq_ops->complete(rq);
618 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
619 __releases(hctx->srcu)
621 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
624 srcu_read_unlock(hctx->srcu, srcu_idx);
627 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
628 __acquires(hctx->srcu)
630 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
631 /* shut up gcc false positive */
635 *srcu_idx = srcu_read_lock(hctx->srcu);
639 * blk_mq_complete_request - end I/O on a request
640 * @rq: the request being processed
643 * Ends all I/O on a request. It does not handle partial completions.
644 * The actual completion happens out-of-order, through a IPI handler.
646 bool blk_mq_complete_request(struct request *rq)
648 if (unlikely(blk_should_fake_timeout(rq->q)))
650 __blk_mq_complete_request(rq);
653 EXPORT_SYMBOL(blk_mq_complete_request);
655 void blk_mq_complete_request_sync(struct request *rq)
657 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
658 rq->q->mq_ops->complete(rq);
660 EXPORT_SYMBOL_GPL(blk_mq_complete_request_sync);
662 int blk_mq_request_started(struct request *rq)
664 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
666 EXPORT_SYMBOL_GPL(blk_mq_request_started);
668 void blk_mq_start_request(struct request *rq)
670 struct request_queue *q = rq->q;
672 blk_mq_sched_started_request(rq);
674 trace_block_rq_issue(q, rq);
676 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
677 rq->io_start_time_ns = ktime_get_ns();
678 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
679 rq->throtl_size = blk_rq_sectors(rq);
681 rq->rq_flags |= RQF_STATS;
685 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
688 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
690 if (q->dma_drain_size && blk_rq_bytes(rq)) {
692 * Make sure space for the drain appears. We know we can do
693 * this because max_hw_segments has been adjusted to be one
694 * fewer than the device can handle.
696 rq->nr_phys_segments++;
699 EXPORT_SYMBOL(blk_mq_start_request);
701 static void __blk_mq_requeue_request(struct request *rq)
703 struct request_queue *q = rq->q;
705 blk_mq_put_driver_tag(rq);
707 trace_block_rq_requeue(q, rq);
708 rq_qos_requeue(q, rq);
710 if (blk_mq_request_started(rq)) {
711 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
712 rq->rq_flags &= ~RQF_TIMED_OUT;
713 if (q->dma_drain_size && blk_rq_bytes(rq))
714 rq->nr_phys_segments--;
718 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
720 __blk_mq_requeue_request(rq);
722 /* this request will be re-inserted to io scheduler queue */
723 blk_mq_sched_requeue_request(rq);
725 BUG_ON(!list_empty(&rq->queuelist));
726 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
728 EXPORT_SYMBOL(blk_mq_requeue_request);
730 static void blk_mq_requeue_work(struct work_struct *work)
732 struct request_queue *q =
733 container_of(work, struct request_queue, requeue_work.work);
735 struct request *rq, *next;
737 spin_lock_irq(&q->requeue_lock);
738 list_splice_init(&q->requeue_list, &rq_list);
739 spin_unlock_irq(&q->requeue_lock);
741 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
742 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
745 rq->rq_flags &= ~RQF_SOFTBARRIER;
746 list_del_init(&rq->queuelist);
748 * If RQF_DONTPREP, rq has contained some driver specific
749 * data, so insert it to hctx dispatch list to avoid any
752 if (rq->rq_flags & RQF_DONTPREP)
753 blk_mq_request_bypass_insert(rq, false);
755 blk_mq_sched_insert_request(rq, true, false, false);
758 while (!list_empty(&rq_list)) {
759 rq = list_entry(rq_list.next, struct request, queuelist);
760 list_del_init(&rq->queuelist);
761 blk_mq_sched_insert_request(rq, false, false, false);
764 blk_mq_run_hw_queues(q, false);
767 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
768 bool kick_requeue_list)
770 struct request_queue *q = rq->q;
774 * We abuse this flag that is otherwise used by the I/O scheduler to
775 * request head insertion from the workqueue.
777 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
779 spin_lock_irqsave(&q->requeue_lock, flags);
781 rq->rq_flags |= RQF_SOFTBARRIER;
782 list_add(&rq->queuelist, &q->requeue_list);
784 list_add_tail(&rq->queuelist, &q->requeue_list);
786 spin_unlock_irqrestore(&q->requeue_lock, flags);
788 if (kick_requeue_list)
789 blk_mq_kick_requeue_list(q);
792 void blk_mq_kick_requeue_list(struct request_queue *q)
794 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
796 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
798 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
801 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
802 msecs_to_jiffies(msecs));
804 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
806 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
808 if (tag < tags->nr_tags) {
809 prefetch(tags->rqs[tag]);
810 return tags->rqs[tag];
815 EXPORT_SYMBOL(blk_mq_tag_to_rq);
817 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
818 void *priv, bool reserved)
821 * If we find a request that is inflight and the queue matches,
822 * we know the queue is busy. Return false to stop the iteration.
824 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
834 bool blk_mq_queue_inflight(struct request_queue *q)
838 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
841 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
843 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
845 req->rq_flags |= RQF_TIMED_OUT;
846 if (req->q->mq_ops->timeout) {
847 enum blk_eh_timer_return ret;
849 ret = req->q->mq_ops->timeout(req, reserved);
850 if (ret == BLK_EH_DONE)
852 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
858 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
860 unsigned long deadline;
862 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
864 if (rq->rq_flags & RQF_TIMED_OUT)
867 deadline = READ_ONCE(rq->deadline);
868 if (time_after_eq(jiffies, deadline))
873 else if (time_after(*next, deadline))
878 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
879 struct request *rq, void *priv, bool reserved)
881 unsigned long *next = priv;
884 * Just do a quick check if it is expired before locking the request in
885 * so we're not unnecessarilly synchronizing across CPUs.
887 if (!blk_mq_req_expired(rq, next))
891 * We have reason to believe the request may be expired. Take a
892 * reference on the request to lock this request lifetime into its
893 * currently allocated context to prevent it from being reallocated in
894 * the event the completion by-passes this timeout handler.
896 * If the reference was already released, then the driver beat the
897 * timeout handler to posting a natural completion.
899 if (!refcount_inc_not_zero(&rq->ref))
903 * The request is now locked and cannot be reallocated underneath the
904 * timeout handler's processing. Re-verify this exact request is truly
905 * expired; if it is not expired, then the request was completed and
906 * reallocated as a new request.
908 if (blk_mq_req_expired(rq, next))
909 blk_mq_rq_timed_out(rq, reserved);
910 if (refcount_dec_and_test(&rq->ref))
911 __blk_mq_free_request(rq);
916 static void blk_mq_timeout_work(struct work_struct *work)
918 struct request_queue *q =
919 container_of(work, struct request_queue, timeout_work);
920 unsigned long next = 0;
921 struct blk_mq_hw_ctx *hctx;
924 /* A deadlock might occur if a request is stuck requiring a
925 * timeout at the same time a queue freeze is waiting
926 * completion, since the timeout code would not be able to
927 * acquire the queue reference here.
929 * That's why we don't use blk_queue_enter here; instead, we use
930 * percpu_ref_tryget directly, because we need to be able to
931 * obtain a reference even in the short window between the queue
932 * starting to freeze, by dropping the first reference in
933 * blk_freeze_queue_start, and the moment the last request is
934 * consumed, marked by the instant q_usage_counter reaches
937 if (!percpu_ref_tryget(&q->q_usage_counter))
940 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
943 mod_timer(&q->timeout, next);
946 * Request timeouts are handled as a forward rolling timer. If
947 * we end up here it means that no requests are pending and
948 * also that no request has been pending for a while. Mark
951 queue_for_each_hw_ctx(q, hctx, i) {
952 /* the hctx may be unmapped, so check it here */
953 if (blk_mq_hw_queue_mapped(hctx))
954 blk_mq_tag_idle(hctx);
960 struct flush_busy_ctx_data {
961 struct blk_mq_hw_ctx *hctx;
962 struct list_head *list;
965 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
967 struct flush_busy_ctx_data *flush_data = data;
968 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
969 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
970 enum hctx_type type = hctx->type;
972 spin_lock(&ctx->lock);
973 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
974 sbitmap_clear_bit(sb, bitnr);
975 spin_unlock(&ctx->lock);
980 * Process software queues that have been marked busy, splicing them
981 * to the for-dispatch
983 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
985 struct flush_busy_ctx_data data = {
990 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
992 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
994 struct dispatch_rq_data {
995 struct blk_mq_hw_ctx *hctx;
999 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1002 struct dispatch_rq_data *dispatch_data = data;
1003 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1004 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1005 enum hctx_type type = hctx->type;
1007 spin_lock(&ctx->lock);
1008 if (!list_empty(&ctx->rq_lists[type])) {
1009 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1010 list_del_init(&dispatch_data->rq->queuelist);
1011 if (list_empty(&ctx->rq_lists[type]))
1012 sbitmap_clear_bit(sb, bitnr);
1014 spin_unlock(&ctx->lock);
1016 return !dispatch_data->rq;
1019 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1020 struct blk_mq_ctx *start)
1022 unsigned off = start ? start->index_hw[hctx->type] : 0;
1023 struct dispatch_rq_data data = {
1028 __sbitmap_for_each_set(&hctx->ctx_map, off,
1029 dispatch_rq_from_ctx, &data);
1034 static inline unsigned int queued_to_index(unsigned int queued)
1039 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1042 bool blk_mq_get_driver_tag(struct request *rq)
1044 struct blk_mq_alloc_data data = {
1046 .hctx = rq->mq_hctx,
1047 .flags = BLK_MQ_REQ_NOWAIT,
1048 .cmd_flags = rq->cmd_flags,
1055 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1056 data.flags |= BLK_MQ_REQ_RESERVED;
1058 shared = blk_mq_tag_busy(data.hctx);
1059 rq->tag = blk_mq_get_tag(&data);
1062 rq->rq_flags |= RQF_MQ_INFLIGHT;
1063 atomic_inc(&data.hctx->nr_active);
1065 data.hctx->tags->rqs[rq->tag] = rq;
1069 return rq->tag != -1;
1072 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1073 int flags, void *key)
1075 struct blk_mq_hw_ctx *hctx;
1077 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1079 spin_lock(&hctx->dispatch_wait_lock);
1080 if (!list_empty(&wait->entry)) {
1081 struct sbitmap_queue *sbq;
1083 list_del_init(&wait->entry);
1084 sbq = &hctx->tags->bitmap_tags;
1085 atomic_dec(&sbq->ws_active);
1087 spin_unlock(&hctx->dispatch_wait_lock);
1089 blk_mq_run_hw_queue(hctx, true);
1094 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1095 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1096 * restart. For both cases, take care to check the condition again after
1097 * marking us as waiting.
1099 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1102 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1103 struct wait_queue_head *wq;
1104 wait_queue_entry_t *wait;
1107 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1108 blk_mq_sched_mark_restart_hctx(hctx);
1111 * It's possible that a tag was freed in the window between the
1112 * allocation failure and adding the hardware queue to the wait
1115 * Don't clear RESTART here, someone else could have set it.
1116 * At most this will cost an extra queue run.
1118 return blk_mq_get_driver_tag(rq);
1121 wait = &hctx->dispatch_wait;
1122 if (!list_empty_careful(&wait->entry))
1125 wq = &bt_wait_ptr(sbq, hctx)->wait;
1127 spin_lock_irq(&wq->lock);
1128 spin_lock(&hctx->dispatch_wait_lock);
1129 if (!list_empty(&wait->entry)) {
1130 spin_unlock(&hctx->dispatch_wait_lock);
1131 spin_unlock_irq(&wq->lock);
1135 atomic_inc(&sbq->ws_active);
1136 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1137 __add_wait_queue(wq, wait);
1140 * It's possible that a tag was freed in the window between the
1141 * allocation failure and adding the hardware queue to the wait
1144 ret = blk_mq_get_driver_tag(rq);
1146 spin_unlock(&hctx->dispatch_wait_lock);
1147 spin_unlock_irq(&wq->lock);
1152 * We got a tag, remove ourselves from the wait queue to ensure
1153 * someone else gets the wakeup.
1155 list_del_init(&wait->entry);
1156 atomic_dec(&sbq->ws_active);
1157 spin_unlock(&hctx->dispatch_wait_lock);
1158 spin_unlock_irq(&wq->lock);
1163 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1164 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1166 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1167 * - EWMA is one simple way to compute running average value
1168 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1169 * - take 4 as factor for avoiding to get too small(0) result, and this
1170 * factor doesn't matter because EWMA decreases exponentially
1172 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1176 if (hctx->queue->elevator)
1179 ewma = hctx->dispatch_busy;
1184 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1186 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1187 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1189 hctx->dispatch_busy = ewma;
1192 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1195 * Returns true if we did some work AND can potentially do more.
1197 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1200 struct blk_mq_hw_ctx *hctx;
1201 struct request *rq, *nxt;
1202 bool no_tag = false;
1204 blk_status_t ret = BLK_STS_OK;
1206 if (list_empty(list))
1209 WARN_ON(!list_is_singular(list) && got_budget);
1212 * Now process all the entries, sending them to the driver.
1214 errors = queued = 0;
1216 struct blk_mq_queue_data bd;
1218 rq = list_first_entry(list, struct request, queuelist);
1221 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1224 if (!blk_mq_get_driver_tag(rq)) {
1226 * The initial allocation attempt failed, so we need to
1227 * rerun the hardware queue when a tag is freed. The
1228 * waitqueue takes care of that. If the queue is run
1229 * before we add this entry back on the dispatch list,
1230 * we'll re-run it below.
1232 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1233 blk_mq_put_dispatch_budget(hctx);
1235 * For non-shared tags, the RESTART check
1238 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1244 list_del_init(&rq->queuelist);
1249 * Flag last if we have no more requests, or if we have more
1250 * but can't assign a driver tag to it.
1252 if (list_empty(list))
1255 nxt = list_first_entry(list, struct request, queuelist);
1256 bd.last = !blk_mq_get_driver_tag(nxt);
1259 ret = q->mq_ops->queue_rq(hctx, &bd);
1260 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1262 * If an I/O scheduler has been configured and we got a
1263 * driver tag for the next request already, free it
1266 if (!list_empty(list)) {
1267 nxt = list_first_entry(list, struct request, queuelist);
1268 blk_mq_put_driver_tag(nxt);
1270 list_add(&rq->queuelist, list);
1271 __blk_mq_requeue_request(rq);
1275 if (unlikely(ret != BLK_STS_OK)) {
1277 blk_mq_end_request(rq, BLK_STS_IOERR);
1282 } while (!list_empty(list));
1284 hctx->dispatched[queued_to_index(queued)]++;
1287 * Any items that need requeuing? Stuff them into hctx->dispatch,
1288 * that is where we will continue on next queue run.
1290 if (!list_empty(list)) {
1294 * If we didn't flush the entire list, we could have told
1295 * the driver there was more coming, but that turned out to
1298 if (q->mq_ops->commit_rqs)
1299 q->mq_ops->commit_rqs(hctx);
1301 spin_lock(&hctx->lock);
1302 list_splice_init(list, &hctx->dispatch);
1303 spin_unlock(&hctx->lock);
1306 * If SCHED_RESTART was set by the caller of this function and
1307 * it is no longer set that means that it was cleared by another
1308 * thread and hence that a queue rerun is needed.
1310 * If 'no_tag' is set, that means that we failed getting
1311 * a driver tag with an I/O scheduler attached. If our dispatch
1312 * waitqueue is no longer active, ensure that we run the queue
1313 * AFTER adding our entries back to the list.
1315 * If no I/O scheduler has been configured it is possible that
1316 * the hardware queue got stopped and restarted before requests
1317 * were pushed back onto the dispatch list. Rerun the queue to
1318 * avoid starvation. Notes:
1319 * - blk_mq_run_hw_queue() checks whether or not a queue has
1320 * been stopped before rerunning a queue.
1321 * - Some but not all block drivers stop a queue before
1322 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1325 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1326 * bit is set, run queue after a delay to avoid IO stalls
1327 * that could otherwise occur if the queue is idle.
1329 needs_restart = blk_mq_sched_needs_restart(hctx);
1330 if (!needs_restart ||
1331 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1332 blk_mq_run_hw_queue(hctx, true);
1333 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1334 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1336 blk_mq_update_dispatch_busy(hctx, true);
1339 blk_mq_update_dispatch_busy(hctx, false);
1342 * If the host/device is unable to accept more work, inform the
1345 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1348 return (queued + errors) != 0;
1351 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1356 * We should be running this queue from one of the CPUs that
1359 * There are at least two related races now between setting
1360 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1361 * __blk_mq_run_hw_queue():
1363 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1364 * but later it becomes online, then this warning is harmless
1367 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1368 * but later it becomes offline, then the warning can't be
1369 * triggered, and we depend on blk-mq timeout handler to
1370 * handle dispatched requests to this hctx
1372 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1373 cpu_online(hctx->next_cpu)) {
1374 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1375 raw_smp_processor_id(),
1376 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1381 * We can't run the queue inline with ints disabled. Ensure that
1382 * we catch bad users of this early.
1384 WARN_ON_ONCE(in_interrupt());
1386 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1388 hctx_lock(hctx, &srcu_idx);
1389 blk_mq_sched_dispatch_requests(hctx);
1390 hctx_unlock(hctx, srcu_idx);
1393 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1395 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1397 if (cpu >= nr_cpu_ids)
1398 cpu = cpumask_first(hctx->cpumask);
1403 * It'd be great if the workqueue API had a way to pass
1404 * in a mask and had some smarts for more clever placement.
1405 * For now we just round-robin here, switching for every
1406 * BLK_MQ_CPU_WORK_BATCH queued items.
1408 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1411 int next_cpu = hctx->next_cpu;
1413 if (hctx->queue->nr_hw_queues == 1)
1414 return WORK_CPU_UNBOUND;
1416 if (--hctx->next_cpu_batch <= 0) {
1418 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1420 if (next_cpu >= nr_cpu_ids)
1421 next_cpu = blk_mq_first_mapped_cpu(hctx);
1422 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1426 * Do unbound schedule if we can't find a online CPU for this hctx,
1427 * and it should only happen in the path of handling CPU DEAD.
1429 if (!cpu_online(next_cpu)) {
1436 * Make sure to re-select CPU next time once after CPUs
1437 * in hctx->cpumask become online again.
1439 hctx->next_cpu = next_cpu;
1440 hctx->next_cpu_batch = 1;
1441 return WORK_CPU_UNBOUND;
1444 hctx->next_cpu = next_cpu;
1448 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1449 unsigned long msecs)
1451 if (unlikely(blk_mq_hctx_stopped(hctx)))
1454 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1455 int cpu = get_cpu();
1456 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1457 __blk_mq_run_hw_queue(hctx);
1465 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1466 msecs_to_jiffies(msecs));
1469 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1471 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1473 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1475 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1481 * When queue is quiesced, we may be switching io scheduler, or
1482 * updating nr_hw_queues, or other things, and we can't run queue
1483 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1485 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1488 hctx_lock(hctx, &srcu_idx);
1489 need_run = !blk_queue_quiesced(hctx->queue) &&
1490 blk_mq_hctx_has_pending(hctx);
1491 hctx_unlock(hctx, srcu_idx);
1494 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1500 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1502 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1504 struct blk_mq_hw_ctx *hctx;
1507 queue_for_each_hw_ctx(q, hctx, i) {
1508 if (blk_mq_hctx_stopped(hctx))
1511 blk_mq_run_hw_queue(hctx, async);
1514 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1517 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1518 * @q: request queue.
1520 * The caller is responsible for serializing this function against
1521 * blk_mq_{start,stop}_hw_queue().
1523 bool blk_mq_queue_stopped(struct request_queue *q)
1525 struct blk_mq_hw_ctx *hctx;
1528 queue_for_each_hw_ctx(q, hctx, i)
1529 if (blk_mq_hctx_stopped(hctx))
1534 EXPORT_SYMBOL(blk_mq_queue_stopped);
1537 * This function is often used for pausing .queue_rq() by driver when
1538 * there isn't enough resource or some conditions aren't satisfied, and
1539 * BLK_STS_RESOURCE is usually returned.
1541 * We do not guarantee that dispatch can be drained or blocked
1542 * after blk_mq_stop_hw_queue() returns. Please use
1543 * blk_mq_quiesce_queue() for that requirement.
1545 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1547 cancel_delayed_work(&hctx->run_work);
1549 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1551 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1554 * This function is often used for pausing .queue_rq() by driver when
1555 * there isn't enough resource or some conditions aren't satisfied, and
1556 * BLK_STS_RESOURCE is usually returned.
1558 * We do not guarantee that dispatch can be drained or blocked
1559 * after blk_mq_stop_hw_queues() returns. Please use
1560 * blk_mq_quiesce_queue() for that requirement.
1562 void blk_mq_stop_hw_queues(struct request_queue *q)
1564 struct blk_mq_hw_ctx *hctx;
1567 queue_for_each_hw_ctx(q, hctx, i)
1568 blk_mq_stop_hw_queue(hctx);
1570 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1572 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1574 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1576 blk_mq_run_hw_queue(hctx, false);
1578 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1580 void blk_mq_start_hw_queues(struct request_queue *q)
1582 struct blk_mq_hw_ctx *hctx;
1585 queue_for_each_hw_ctx(q, hctx, i)
1586 blk_mq_start_hw_queue(hctx);
1588 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1590 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1592 if (!blk_mq_hctx_stopped(hctx))
1595 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1596 blk_mq_run_hw_queue(hctx, async);
1598 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1600 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1602 struct blk_mq_hw_ctx *hctx;
1605 queue_for_each_hw_ctx(q, hctx, i)
1606 blk_mq_start_stopped_hw_queue(hctx, async);
1608 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1610 static void blk_mq_run_work_fn(struct work_struct *work)
1612 struct blk_mq_hw_ctx *hctx;
1614 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1617 * If we are stopped, don't run the queue.
1619 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1622 __blk_mq_run_hw_queue(hctx);
1625 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1629 struct blk_mq_ctx *ctx = rq->mq_ctx;
1630 enum hctx_type type = hctx->type;
1632 lockdep_assert_held(&ctx->lock);
1634 trace_block_rq_insert(hctx->queue, rq);
1637 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1639 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1642 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1645 struct blk_mq_ctx *ctx = rq->mq_ctx;
1647 lockdep_assert_held(&ctx->lock);
1649 __blk_mq_insert_req_list(hctx, rq, at_head);
1650 blk_mq_hctx_mark_pending(hctx, ctx);
1654 * Should only be used carefully, when the caller knows we want to
1655 * bypass a potential IO scheduler on the target device.
1657 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1659 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1661 spin_lock(&hctx->lock);
1662 list_add_tail(&rq->queuelist, &hctx->dispatch);
1663 spin_unlock(&hctx->lock);
1666 blk_mq_run_hw_queue(hctx, false);
1669 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1670 struct list_head *list)
1674 enum hctx_type type = hctx->type;
1677 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1680 list_for_each_entry(rq, list, queuelist) {
1681 BUG_ON(rq->mq_ctx != ctx);
1682 trace_block_rq_insert(hctx->queue, rq);
1685 spin_lock(&ctx->lock);
1686 list_splice_tail_init(list, &ctx->rq_lists[type]);
1687 blk_mq_hctx_mark_pending(hctx, ctx);
1688 spin_unlock(&ctx->lock);
1691 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1693 struct request *rqa = container_of(a, struct request, queuelist);
1694 struct request *rqb = container_of(b, struct request, queuelist);
1696 if (rqa->mq_ctx < rqb->mq_ctx)
1698 else if (rqa->mq_ctx > rqb->mq_ctx)
1700 else if (rqa->mq_hctx < rqb->mq_hctx)
1702 else if (rqa->mq_hctx > rqb->mq_hctx)
1705 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1708 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1710 struct blk_mq_hw_ctx *this_hctx;
1711 struct blk_mq_ctx *this_ctx;
1712 struct request_queue *this_q;
1718 list_splice_init(&plug->mq_list, &list);
1720 if (plug->rq_count > 2 && plug->multiple_queues)
1721 list_sort(NULL, &list, plug_rq_cmp);
1730 while (!list_empty(&list)) {
1731 rq = list_entry_rq(list.next);
1732 list_del_init(&rq->queuelist);
1734 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1736 trace_block_unplug(this_q, depth, !from_schedule);
1737 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1743 this_ctx = rq->mq_ctx;
1744 this_hctx = rq->mq_hctx;
1749 list_add_tail(&rq->queuelist, &rq_list);
1753 * If 'this_hctx' is set, we know we have entries to complete
1754 * on 'rq_list'. Do those.
1757 trace_block_unplug(this_q, depth, !from_schedule);
1758 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1763 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1764 unsigned int nr_segs)
1766 if (bio->bi_opf & REQ_RAHEAD)
1767 rq->cmd_flags |= REQ_FAILFAST_MASK;
1769 rq->__sector = bio->bi_iter.bi_sector;
1770 rq->write_hint = bio->bi_write_hint;
1771 blk_rq_bio_prep(rq, bio, nr_segs);
1773 blk_account_io_start(rq, true);
1776 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1778 blk_qc_t *cookie, bool last)
1780 struct request_queue *q = rq->q;
1781 struct blk_mq_queue_data bd = {
1785 blk_qc_t new_cookie;
1788 new_cookie = request_to_qc_t(hctx, rq);
1791 * For OK queue, we are done. For error, caller may kill it.
1792 * Any other error (busy), just add it to our list as we
1793 * previously would have done.
1795 ret = q->mq_ops->queue_rq(hctx, &bd);
1798 blk_mq_update_dispatch_busy(hctx, false);
1799 *cookie = new_cookie;
1801 case BLK_STS_RESOURCE:
1802 case BLK_STS_DEV_RESOURCE:
1803 blk_mq_update_dispatch_busy(hctx, true);
1804 __blk_mq_requeue_request(rq);
1807 blk_mq_update_dispatch_busy(hctx, false);
1808 *cookie = BLK_QC_T_NONE;
1815 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1818 bool bypass_insert, bool last)
1820 struct request_queue *q = rq->q;
1821 bool run_queue = true;
1824 * RCU or SRCU read lock is needed before checking quiesced flag.
1826 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1827 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1828 * and avoid driver to try to dispatch again.
1830 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1832 bypass_insert = false;
1836 if (q->elevator && !bypass_insert)
1839 if (!blk_mq_get_dispatch_budget(hctx))
1842 if (!blk_mq_get_driver_tag(rq)) {
1843 blk_mq_put_dispatch_budget(hctx);
1847 return __blk_mq_issue_directly(hctx, rq, cookie, last);
1850 return BLK_STS_RESOURCE;
1852 blk_mq_request_bypass_insert(rq, run_queue);
1856 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1857 struct request *rq, blk_qc_t *cookie)
1862 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1864 hctx_lock(hctx, &srcu_idx);
1866 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1867 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1868 blk_mq_request_bypass_insert(rq, true);
1869 else if (ret != BLK_STS_OK)
1870 blk_mq_end_request(rq, ret);
1872 hctx_unlock(hctx, srcu_idx);
1875 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1879 blk_qc_t unused_cookie;
1880 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1882 hctx_lock(hctx, &srcu_idx);
1883 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1884 hctx_unlock(hctx, srcu_idx);
1889 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1890 struct list_head *list)
1892 while (!list_empty(list)) {
1894 struct request *rq = list_first_entry(list, struct request,
1897 list_del_init(&rq->queuelist);
1898 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1899 if (ret != BLK_STS_OK) {
1900 if (ret == BLK_STS_RESOURCE ||
1901 ret == BLK_STS_DEV_RESOURCE) {
1902 blk_mq_request_bypass_insert(rq,
1906 blk_mq_end_request(rq, ret);
1911 * If we didn't flush the entire list, we could have told
1912 * the driver there was more coming, but that turned out to
1915 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1916 hctx->queue->mq_ops->commit_rqs(hctx);
1919 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1921 list_add_tail(&rq->queuelist, &plug->mq_list);
1923 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1924 struct request *tmp;
1926 tmp = list_first_entry(&plug->mq_list, struct request,
1928 if (tmp->q != rq->q)
1929 plug->multiple_queues = true;
1933 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1935 const int is_sync = op_is_sync(bio->bi_opf);
1936 const int is_flush_fua = op_is_flush(bio->bi_opf);
1937 struct blk_mq_alloc_data data = { .flags = 0};
1939 struct blk_plug *plug;
1940 struct request *same_queue_rq = NULL;
1941 unsigned int nr_segs;
1944 blk_queue_bounce(q, &bio);
1945 __blk_queue_split(q, &bio, &nr_segs);
1947 if (!bio_integrity_prep(bio))
1948 return BLK_QC_T_NONE;
1950 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1951 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1952 return BLK_QC_T_NONE;
1954 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1955 return BLK_QC_T_NONE;
1957 rq_qos_throttle(q, bio);
1959 data.cmd_flags = bio->bi_opf;
1960 rq = blk_mq_get_request(q, bio, &data);
1961 if (unlikely(!rq)) {
1962 rq_qos_cleanup(q, bio);
1964 cookie = BLK_QC_T_NONE;
1965 if (bio->bi_opf & REQ_NOWAIT_INLINE)
1966 cookie = BLK_QC_T_EAGAIN;
1967 else if (bio->bi_opf & REQ_NOWAIT)
1968 bio_wouldblock_error(bio);
1972 trace_block_getrq(q, bio, bio->bi_opf);
1974 rq_qos_track(q, rq, bio);
1976 cookie = request_to_qc_t(data.hctx, rq);
1978 blk_mq_bio_to_request(rq, bio, nr_segs);
1980 plug = blk_mq_plug(q, bio);
1981 if (unlikely(is_flush_fua)) {
1982 /* bypass scheduler for flush rq */
1983 blk_insert_flush(rq);
1984 blk_mq_run_hw_queue(data.hctx, true);
1985 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1987 * Use plugging if we have a ->commit_rqs() hook as well, as
1988 * we know the driver uses bd->last in a smart fashion.
1990 unsigned int request_count = plug->rq_count;
1991 struct request *last = NULL;
1994 trace_block_plug(q);
1996 last = list_entry_rq(plug->mq_list.prev);
1998 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1999 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2000 blk_flush_plug_list(plug, false);
2001 trace_block_plug(q);
2004 blk_add_rq_to_plug(plug, rq);
2005 } else if (plug && !blk_queue_nomerges(q)) {
2007 * We do limited plugging. If the bio can be merged, do that.
2008 * Otherwise the existing request in the plug list will be
2009 * issued. So the plug list will have one request at most
2010 * The plug list might get flushed before this. If that happens,
2011 * the plug list is empty, and same_queue_rq is invalid.
2013 if (list_empty(&plug->mq_list))
2014 same_queue_rq = NULL;
2015 if (same_queue_rq) {
2016 list_del_init(&same_queue_rq->queuelist);
2019 blk_add_rq_to_plug(plug, rq);
2020 trace_block_plug(q);
2022 if (same_queue_rq) {
2023 data.hctx = same_queue_rq->mq_hctx;
2024 trace_block_unplug(q, 1, true);
2025 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2028 } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2029 !data.hctx->dispatch_busy)) {
2030 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2032 blk_mq_sched_insert_request(rq, false, true, true);
2038 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2039 unsigned int hctx_idx)
2043 if (tags->rqs && set->ops->exit_request) {
2046 for (i = 0; i < tags->nr_tags; i++) {
2047 struct request *rq = tags->static_rqs[i];
2051 set->ops->exit_request(set, rq, hctx_idx);
2052 tags->static_rqs[i] = NULL;
2056 while (!list_empty(&tags->page_list)) {
2057 page = list_first_entry(&tags->page_list, struct page, lru);
2058 list_del_init(&page->lru);
2060 * Remove kmemleak object previously allocated in
2061 * blk_mq_alloc_rqs().
2063 kmemleak_free(page_address(page));
2064 __free_pages(page, page->private);
2068 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2072 kfree(tags->static_rqs);
2073 tags->static_rqs = NULL;
2075 blk_mq_free_tags(tags);
2078 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2079 unsigned int hctx_idx,
2080 unsigned int nr_tags,
2081 unsigned int reserved_tags)
2083 struct blk_mq_tags *tags;
2086 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2087 if (node == NUMA_NO_NODE)
2088 node = set->numa_node;
2090 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2091 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2095 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2096 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2099 blk_mq_free_tags(tags);
2103 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2104 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2106 if (!tags->static_rqs) {
2108 blk_mq_free_tags(tags);
2115 static size_t order_to_size(unsigned int order)
2117 return (size_t)PAGE_SIZE << order;
2120 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2121 unsigned int hctx_idx, int node)
2125 if (set->ops->init_request) {
2126 ret = set->ops->init_request(set, rq, hctx_idx, node);
2131 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2135 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2136 unsigned int hctx_idx, unsigned int depth)
2138 unsigned int i, j, entries_per_page, max_order = 4;
2139 size_t rq_size, left;
2142 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2143 if (node == NUMA_NO_NODE)
2144 node = set->numa_node;
2146 INIT_LIST_HEAD(&tags->page_list);
2149 * rq_size is the size of the request plus driver payload, rounded
2150 * to the cacheline size
2152 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2154 left = rq_size * depth;
2156 for (i = 0; i < depth; ) {
2157 int this_order = max_order;
2162 while (this_order && left < order_to_size(this_order - 1))
2166 page = alloc_pages_node(node,
2167 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2173 if (order_to_size(this_order) < rq_size)
2180 page->private = this_order;
2181 list_add_tail(&page->lru, &tags->page_list);
2183 p = page_address(page);
2185 * Allow kmemleak to scan these pages as they contain pointers
2186 * to additional allocations like via ops->init_request().
2188 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2189 entries_per_page = order_to_size(this_order) / rq_size;
2190 to_do = min(entries_per_page, depth - i);
2191 left -= to_do * rq_size;
2192 for (j = 0; j < to_do; j++) {
2193 struct request *rq = p;
2195 tags->static_rqs[i] = rq;
2196 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2197 tags->static_rqs[i] = NULL;
2208 blk_mq_free_rqs(set, tags, hctx_idx);
2213 * 'cpu' is going away. splice any existing rq_list entries from this
2214 * software queue to the hw queue dispatch list, and ensure that it
2217 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2219 struct blk_mq_hw_ctx *hctx;
2220 struct blk_mq_ctx *ctx;
2222 enum hctx_type type;
2224 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2225 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2228 spin_lock(&ctx->lock);
2229 if (!list_empty(&ctx->rq_lists[type])) {
2230 list_splice_init(&ctx->rq_lists[type], &tmp);
2231 blk_mq_hctx_clear_pending(hctx, ctx);
2233 spin_unlock(&ctx->lock);
2235 if (list_empty(&tmp))
2238 spin_lock(&hctx->lock);
2239 list_splice_tail_init(&tmp, &hctx->dispatch);
2240 spin_unlock(&hctx->lock);
2242 blk_mq_run_hw_queue(hctx, true);
2246 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2248 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2252 /* hctx->ctxs will be freed in queue's release handler */
2253 static void blk_mq_exit_hctx(struct request_queue *q,
2254 struct blk_mq_tag_set *set,
2255 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2257 if (blk_mq_hw_queue_mapped(hctx))
2258 blk_mq_tag_idle(hctx);
2260 if (set->ops->exit_request)
2261 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2263 if (set->ops->exit_hctx)
2264 set->ops->exit_hctx(hctx, hctx_idx);
2266 blk_mq_remove_cpuhp(hctx);
2268 spin_lock(&q->unused_hctx_lock);
2269 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2270 spin_unlock(&q->unused_hctx_lock);
2273 static void blk_mq_exit_hw_queues(struct request_queue *q,
2274 struct blk_mq_tag_set *set, int nr_queue)
2276 struct blk_mq_hw_ctx *hctx;
2279 queue_for_each_hw_ctx(q, hctx, i) {
2282 blk_mq_debugfs_unregister_hctx(hctx);
2283 blk_mq_exit_hctx(q, set, hctx, i);
2287 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2289 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2291 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2292 __alignof__(struct blk_mq_hw_ctx)) !=
2293 sizeof(struct blk_mq_hw_ctx));
2295 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2296 hw_ctx_size += sizeof(struct srcu_struct);
2301 static int blk_mq_init_hctx(struct request_queue *q,
2302 struct blk_mq_tag_set *set,
2303 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2305 hctx->queue_num = hctx_idx;
2307 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2309 hctx->tags = set->tags[hctx_idx];
2311 if (set->ops->init_hctx &&
2312 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2313 goto unregister_cpu_notifier;
2315 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2321 if (set->ops->exit_hctx)
2322 set->ops->exit_hctx(hctx, hctx_idx);
2323 unregister_cpu_notifier:
2324 blk_mq_remove_cpuhp(hctx);
2328 static struct blk_mq_hw_ctx *
2329 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2332 struct blk_mq_hw_ctx *hctx;
2333 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2335 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2337 goto fail_alloc_hctx;
2339 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2342 atomic_set(&hctx->nr_active, 0);
2343 if (node == NUMA_NO_NODE)
2344 node = set->numa_node;
2345 hctx->numa_node = node;
2347 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2348 spin_lock_init(&hctx->lock);
2349 INIT_LIST_HEAD(&hctx->dispatch);
2351 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2353 INIT_LIST_HEAD(&hctx->hctx_list);
2356 * Allocate space for all possible cpus to avoid allocation at
2359 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2364 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2369 spin_lock_init(&hctx->dispatch_wait_lock);
2370 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2371 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2373 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2378 if (hctx->flags & BLK_MQ_F_BLOCKING)
2379 init_srcu_struct(hctx->srcu);
2380 blk_mq_hctx_kobj_init(hctx);
2385 sbitmap_free(&hctx->ctx_map);
2389 free_cpumask_var(hctx->cpumask);
2396 static void blk_mq_init_cpu_queues(struct request_queue *q,
2397 unsigned int nr_hw_queues)
2399 struct blk_mq_tag_set *set = q->tag_set;
2402 for_each_possible_cpu(i) {
2403 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2404 struct blk_mq_hw_ctx *hctx;
2408 spin_lock_init(&__ctx->lock);
2409 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2410 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2415 * Set local node, IFF we have more than one hw queue. If
2416 * not, we remain on the home node of the device
2418 for (j = 0; j < set->nr_maps; j++) {
2419 hctx = blk_mq_map_queue_type(q, j, i);
2420 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2421 hctx->numa_node = local_memory_node(cpu_to_node(i));
2426 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2430 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2431 set->queue_depth, set->reserved_tags);
2432 if (!set->tags[hctx_idx])
2435 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2440 blk_mq_free_rq_map(set->tags[hctx_idx]);
2441 set->tags[hctx_idx] = NULL;
2445 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2446 unsigned int hctx_idx)
2448 if (set->tags && set->tags[hctx_idx]) {
2449 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2450 blk_mq_free_rq_map(set->tags[hctx_idx]);
2451 set->tags[hctx_idx] = NULL;
2455 static void blk_mq_map_swqueue(struct request_queue *q)
2457 unsigned int i, j, hctx_idx;
2458 struct blk_mq_hw_ctx *hctx;
2459 struct blk_mq_ctx *ctx;
2460 struct blk_mq_tag_set *set = q->tag_set;
2463 * Avoid others reading imcomplete hctx->cpumask through sysfs
2465 mutex_lock(&q->sysfs_lock);
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 mutex_unlock(&q->sysfs_lock);
2529 queue_for_each_hw_ctx(q, hctx, i) {
2531 * If no software queues are mapped to this hardware queue,
2532 * disable it and free the request entries.
2534 if (!hctx->nr_ctx) {
2535 /* Never unmap queue 0. We need it as a
2536 * fallback in case of a new remap fails
2539 if (i && set->tags[i])
2540 blk_mq_free_map_and_requests(set, i);
2546 hctx->tags = set->tags[i];
2547 WARN_ON(!hctx->tags);
2550 * Set the map size to the number of mapped software queues.
2551 * This is more accurate and more efficient than looping
2552 * over all possibly mapped software queues.
2554 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2557 * Initialize batch roundrobin counts
2559 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2560 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2565 * Caller needs to ensure that we're either frozen/quiesced, or that
2566 * the queue isn't live yet.
2568 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2570 struct blk_mq_hw_ctx *hctx;
2573 queue_for_each_hw_ctx(q, hctx, i) {
2575 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2577 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2581 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2584 struct request_queue *q;
2586 lockdep_assert_held(&set->tag_list_lock);
2588 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2589 blk_mq_freeze_queue(q);
2590 queue_set_hctx_shared(q, shared);
2591 blk_mq_unfreeze_queue(q);
2595 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2597 struct blk_mq_tag_set *set = q->tag_set;
2599 mutex_lock(&set->tag_list_lock);
2600 list_del_rcu(&q->tag_set_list);
2601 if (list_is_singular(&set->tag_list)) {
2602 /* just transitioned to unshared */
2603 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2604 /* update existing queue */
2605 blk_mq_update_tag_set_depth(set, false);
2607 mutex_unlock(&set->tag_list_lock);
2608 INIT_LIST_HEAD(&q->tag_set_list);
2611 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2612 struct request_queue *q)
2614 mutex_lock(&set->tag_list_lock);
2617 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2619 if (!list_empty(&set->tag_list) &&
2620 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2621 set->flags |= BLK_MQ_F_TAG_SHARED;
2622 /* update existing queue */
2623 blk_mq_update_tag_set_depth(set, true);
2625 if (set->flags & BLK_MQ_F_TAG_SHARED)
2626 queue_set_hctx_shared(q, true);
2627 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2629 mutex_unlock(&set->tag_list_lock);
2632 /* All allocations will be freed in release handler of q->mq_kobj */
2633 static int blk_mq_alloc_ctxs(struct request_queue *q)
2635 struct blk_mq_ctxs *ctxs;
2638 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2642 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2643 if (!ctxs->queue_ctx)
2646 for_each_possible_cpu(cpu) {
2647 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2651 q->mq_kobj = &ctxs->kobj;
2652 q->queue_ctx = ctxs->queue_ctx;
2661 * It is the actual release handler for mq, but we do it from
2662 * request queue's release handler for avoiding use-after-free
2663 * and headache because q->mq_kobj shouldn't have been introduced,
2664 * but we can't group ctx/kctx kobj without it.
2666 void blk_mq_release(struct request_queue *q)
2668 struct blk_mq_hw_ctx *hctx, *next;
2671 cancel_delayed_work_sync(&q->requeue_work);
2673 queue_for_each_hw_ctx(q, hctx, i)
2674 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2676 /* all hctx are in .unused_hctx_list now */
2677 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2678 list_del_init(&hctx->hctx_list);
2679 kobject_put(&hctx->kobj);
2682 kfree(q->queue_hw_ctx);
2685 * release .mq_kobj and sw queue's kobject now because
2686 * both share lifetime with request queue.
2688 blk_mq_sysfs_deinit(q);
2691 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2693 struct request_queue *uninit_q, *q;
2695 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2697 return ERR_PTR(-ENOMEM);
2699 q = blk_mq_init_allocated_queue(set, uninit_q);
2701 blk_cleanup_queue(uninit_q);
2705 EXPORT_SYMBOL(blk_mq_init_queue);
2708 * Helper for setting up a queue with mq ops, given queue depth, and
2709 * the passed in mq ops flags.
2711 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2712 const struct blk_mq_ops *ops,
2713 unsigned int queue_depth,
2714 unsigned int set_flags)
2716 struct request_queue *q;
2719 memset(set, 0, sizeof(*set));
2721 set->nr_hw_queues = 1;
2723 set->queue_depth = queue_depth;
2724 set->numa_node = NUMA_NO_NODE;
2725 set->flags = set_flags;
2727 ret = blk_mq_alloc_tag_set(set);
2729 return ERR_PTR(ret);
2731 q = blk_mq_init_queue(set);
2733 blk_mq_free_tag_set(set);
2739 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2741 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2742 struct blk_mq_tag_set *set, struct request_queue *q,
2743 int hctx_idx, int node)
2745 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2747 /* reuse dead hctx first */
2748 spin_lock(&q->unused_hctx_lock);
2749 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2750 if (tmp->numa_node == node) {
2756 list_del_init(&hctx->hctx_list);
2757 spin_unlock(&q->unused_hctx_lock);
2760 hctx = blk_mq_alloc_hctx(q, set, node);
2764 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2770 kobject_put(&hctx->kobj);
2775 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2776 struct request_queue *q)
2779 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2781 /* protect against switching io scheduler */
2782 mutex_lock(&q->sysfs_lock);
2783 for (i = 0; i < set->nr_hw_queues; i++) {
2785 struct blk_mq_hw_ctx *hctx;
2787 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2789 * If the hw queue has been mapped to another numa node,
2790 * we need to realloc the hctx. If allocation fails, fallback
2791 * to use the previous one.
2793 if (hctxs[i] && (hctxs[i]->numa_node == node))
2796 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2799 blk_mq_exit_hctx(q, set, hctxs[i], i);
2803 pr_warn("Allocate new hctx on node %d fails,\
2804 fallback to previous one on node %d\n",
2805 node, hctxs[i]->numa_node);
2811 * Increasing nr_hw_queues fails. Free the newly allocated
2812 * hctxs and keep the previous q->nr_hw_queues.
2814 if (i != set->nr_hw_queues) {
2815 j = q->nr_hw_queues;
2819 end = q->nr_hw_queues;
2820 q->nr_hw_queues = set->nr_hw_queues;
2823 for (; j < end; j++) {
2824 struct blk_mq_hw_ctx *hctx = hctxs[j];
2828 blk_mq_free_map_and_requests(set, j);
2829 blk_mq_exit_hctx(q, set, hctx, j);
2833 mutex_unlock(&q->sysfs_lock);
2837 * Maximum number of hardware queues we support. For single sets, we'll never
2838 * have more than the CPUs (software queues). For multiple sets, the tag_set
2839 * user may have set ->nr_hw_queues larger.
2841 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2843 if (set->nr_maps == 1)
2846 return max(set->nr_hw_queues, nr_cpu_ids);
2849 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2850 struct request_queue *q)
2852 /* mark the queue as mq asap */
2853 q->mq_ops = set->ops;
2855 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2856 blk_mq_poll_stats_bkt,
2857 BLK_MQ_POLL_STATS_BKTS, q);
2861 if (blk_mq_alloc_ctxs(q))
2864 /* init q->mq_kobj and sw queues' kobjects */
2865 blk_mq_sysfs_init(q);
2867 q->nr_queues = nr_hw_queues(set);
2868 q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2869 GFP_KERNEL, set->numa_node);
2870 if (!q->queue_hw_ctx)
2873 INIT_LIST_HEAD(&q->unused_hctx_list);
2874 spin_lock_init(&q->unused_hctx_lock);
2876 blk_mq_realloc_hw_ctxs(set, q);
2877 if (!q->nr_hw_queues)
2880 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2881 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2885 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2886 if (set->nr_maps > HCTX_TYPE_POLL &&
2887 set->map[HCTX_TYPE_POLL].nr_queues)
2888 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2890 q->sg_reserved_size = INT_MAX;
2892 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2893 INIT_LIST_HEAD(&q->requeue_list);
2894 spin_lock_init(&q->requeue_lock);
2896 blk_queue_make_request(q, blk_mq_make_request);
2899 * Do this after blk_queue_make_request() overrides it...
2901 q->nr_requests = set->queue_depth;
2904 * Default to classic polling
2906 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2908 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2909 blk_mq_add_queue_tag_set(set, q);
2910 blk_mq_map_swqueue(q);
2912 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2915 ret = elevator_init_mq(q);
2917 return ERR_PTR(ret);
2923 kfree(q->queue_hw_ctx);
2925 blk_mq_sysfs_deinit(q);
2927 blk_stat_free_callback(q->poll_cb);
2931 return ERR_PTR(-ENOMEM);
2933 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2935 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2936 void blk_mq_exit_queue(struct request_queue *q)
2938 struct blk_mq_tag_set *set = q->tag_set;
2940 blk_mq_del_queue_tag_set(q);
2941 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2944 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2948 for (i = 0; i < set->nr_hw_queues; i++)
2949 if (!__blk_mq_alloc_rq_map(set, i))
2956 blk_mq_free_rq_map(set->tags[i]);
2962 * Allocate the request maps associated with this tag_set. Note that this
2963 * may reduce the depth asked for, if memory is tight. set->queue_depth
2964 * will be updated to reflect the allocated depth.
2966 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2971 depth = set->queue_depth;
2973 err = __blk_mq_alloc_rq_maps(set);
2977 set->queue_depth >>= 1;
2978 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2982 } while (set->queue_depth);
2984 if (!set->queue_depth || err) {
2985 pr_err("blk-mq: failed to allocate request map\n");
2989 if (depth != set->queue_depth)
2990 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2991 depth, set->queue_depth);
2996 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2998 if (set->ops->map_queues && !is_kdump_kernel()) {
3002 * transport .map_queues is usually done in the following
3005 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3006 * mask = get_cpu_mask(queue)
3007 * for_each_cpu(cpu, mask)
3008 * set->map[x].mq_map[cpu] = queue;
3011 * When we need to remap, the table has to be cleared for
3012 * killing stale mapping since one CPU may not be mapped
3015 for (i = 0; i < set->nr_maps; i++)
3016 blk_mq_clear_mq_map(&set->map[i]);
3018 return set->ops->map_queues(set);
3020 BUG_ON(set->nr_maps > 1);
3021 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3026 * Alloc a tag set to be associated with one or more request queues.
3027 * May fail with EINVAL for various error conditions. May adjust the
3028 * requested depth down, if it's too large. In that case, the set
3029 * value will be stored in set->queue_depth.
3031 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3035 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3037 if (!set->nr_hw_queues)
3039 if (!set->queue_depth)
3041 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3044 if (!set->ops->queue_rq)
3047 if (!set->ops->get_budget ^ !set->ops->put_budget)
3050 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3051 pr_info("blk-mq: reduced tag depth to %u\n",
3053 set->queue_depth = BLK_MQ_MAX_DEPTH;
3058 else if (set->nr_maps > HCTX_MAX_TYPES)
3062 * If a crashdump is active, then we are potentially in a very
3063 * memory constrained environment. Limit us to 1 queue and
3064 * 64 tags to prevent using too much memory.
3066 if (is_kdump_kernel()) {
3067 set->nr_hw_queues = 1;
3069 set->queue_depth = min(64U, set->queue_depth);
3072 * There is no use for more h/w queues than cpus if we just have
3075 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3076 set->nr_hw_queues = nr_cpu_ids;
3078 set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3079 GFP_KERNEL, set->numa_node);
3084 for (i = 0; i < set->nr_maps; i++) {
3085 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3086 sizeof(set->map[i].mq_map[0]),
3087 GFP_KERNEL, set->numa_node);
3088 if (!set->map[i].mq_map)
3089 goto out_free_mq_map;
3090 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3093 ret = blk_mq_update_queue_map(set);
3095 goto out_free_mq_map;
3097 ret = blk_mq_alloc_rq_maps(set);
3099 goto out_free_mq_map;
3101 mutex_init(&set->tag_list_lock);
3102 INIT_LIST_HEAD(&set->tag_list);
3107 for (i = 0; i < set->nr_maps; i++) {
3108 kfree(set->map[i].mq_map);
3109 set->map[i].mq_map = NULL;
3115 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3117 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3121 for (i = 0; i < nr_hw_queues(set); i++)
3122 blk_mq_free_map_and_requests(set, i);
3124 for (j = 0; j < set->nr_maps; j++) {
3125 kfree(set->map[j].mq_map);
3126 set->map[j].mq_map = NULL;
3132 EXPORT_SYMBOL(blk_mq_free_tag_set);
3134 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3136 struct blk_mq_tag_set *set = q->tag_set;
3137 struct blk_mq_hw_ctx *hctx;
3143 if (q->nr_requests == nr)
3146 blk_mq_freeze_queue(q);
3147 blk_mq_quiesce_queue(q);
3150 queue_for_each_hw_ctx(q, hctx, i) {
3154 * If we're using an MQ scheduler, just update the scheduler
3155 * queue depth. This is similar to what the old code would do.
3157 if (!hctx->sched_tags) {
3158 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3161 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3166 if (q->elevator && q->elevator->type->ops.depth_updated)
3167 q->elevator->type->ops.depth_updated(hctx);
3171 q->nr_requests = nr;
3173 blk_mq_unquiesce_queue(q);
3174 blk_mq_unfreeze_queue(q);
3180 * request_queue and elevator_type pair.
3181 * It is just used by __blk_mq_update_nr_hw_queues to cache
3182 * the elevator_type associated with a request_queue.
3184 struct blk_mq_qe_pair {
3185 struct list_head node;
3186 struct request_queue *q;
3187 struct elevator_type *type;
3191 * Cache the elevator_type in qe pair list and switch the
3192 * io scheduler to 'none'
3194 static bool blk_mq_elv_switch_none(struct list_head *head,
3195 struct request_queue *q)
3197 struct blk_mq_qe_pair *qe;
3202 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3206 INIT_LIST_HEAD(&qe->node);
3208 qe->type = q->elevator->type;
3209 list_add(&qe->node, head);
3211 mutex_lock(&q->sysfs_lock);
3213 * After elevator_switch_mq, the previous elevator_queue will be
3214 * released by elevator_release. The reference of the io scheduler
3215 * module get by elevator_get will also be put. So we need to get
3216 * a reference of the io scheduler module here to prevent it to be
3219 __module_get(qe->type->elevator_owner);
3220 elevator_switch_mq(q, NULL);
3221 mutex_unlock(&q->sysfs_lock);
3226 static void blk_mq_elv_switch_back(struct list_head *head,
3227 struct request_queue *q)
3229 struct blk_mq_qe_pair *qe;
3230 struct elevator_type *t = NULL;
3232 list_for_each_entry(qe, head, node)
3241 list_del(&qe->node);
3244 mutex_lock(&q->sysfs_lock);
3245 elevator_switch_mq(q, t);
3246 mutex_unlock(&q->sysfs_lock);
3249 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3252 struct request_queue *q;
3254 int prev_nr_hw_queues;
3256 lockdep_assert_held(&set->tag_list_lock);
3258 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3259 nr_hw_queues = nr_cpu_ids;
3260 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3263 list_for_each_entry(q, &set->tag_list, tag_set_list)
3264 blk_mq_freeze_queue(q);
3266 * Sync with blk_mq_queue_tag_busy_iter.
3270 * Switch IO scheduler to 'none', cleaning up the data associated
3271 * with the previous scheduler. We will switch back once we are done
3272 * updating the new sw to hw queue mappings.
3274 list_for_each_entry(q, &set->tag_list, tag_set_list)
3275 if (!blk_mq_elv_switch_none(&head, q))
3278 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3279 blk_mq_debugfs_unregister_hctxs(q);
3280 blk_mq_sysfs_unregister(q);
3283 prev_nr_hw_queues = set->nr_hw_queues;
3284 set->nr_hw_queues = nr_hw_queues;
3285 blk_mq_update_queue_map(set);
3287 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3288 blk_mq_realloc_hw_ctxs(set, q);
3289 if (q->nr_hw_queues != set->nr_hw_queues) {
3290 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3291 nr_hw_queues, prev_nr_hw_queues);
3292 set->nr_hw_queues = prev_nr_hw_queues;
3293 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3296 blk_mq_map_swqueue(q);
3299 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3300 blk_mq_sysfs_register(q);
3301 blk_mq_debugfs_register_hctxs(q);
3305 list_for_each_entry(q, &set->tag_list, tag_set_list)
3306 blk_mq_elv_switch_back(&head, q);
3308 list_for_each_entry(q, &set->tag_list, tag_set_list)
3309 blk_mq_unfreeze_queue(q);
3312 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3314 mutex_lock(&set->tag_list_lock);
3315 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3316 mutex_unlock(&set->tag_list_lock);
3318 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3320 /* Enable polling stats and return whether they were already enabled. */
3321 static bool blk_poll_stats_enable(struct request_queue *q)
3323 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3324 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3326 blk_stat_add_callback(q, q->poll_cb);
3330 static void blk_mq_poll_stats_start(struct request_queue *q)
3333 * We don't arm the callback if polling stats are not enabled or the
3334 * callback is already active.
3336 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3337 blk_stat_is_active(q->poll_cb))
3340 blk_stat_activate_msecs(q->poll_cb, 100);
3343 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3345 struct request_queue *q = cb->data;
3348 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3349 if (cb->stat[bucket].nr_samples)
3350 q->poll_stat[bucket] = cb->stat[bucket];
3354 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3355 struct blk_mq_hw_ctx *hctx,
3358 unsigned long ret = 0;
3362 * If stats collection isn't on, don't sleep but turn it on for
3365 if (!blk_poll_stats_enable(q))
3369 * As an optimistic guess, use half of the mean service time
3370 * for this type of request. We can (and should) make this smarter.
3371 * For instance, if the completion latencies are tight, we can
3372 * get closer than just half the mean. This is especially
3373 * important on devices where the completion latencies are longer
3374 * than ~10 usec. We do use the stats for the relevant IO size
3375 * if available which does lead to better estimates.
3377 bucket = blk_mq_poll_stats_bkt(rq);
3381 if (q->poll_stat[bucket].nr_samples)
3382 ret = (q->poll_stat[bucket].mean + 1) / 2;
3387 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3388 struct blk_mq_hw_ctx *hctx,
3391 struct hrtimer_sleeper hs;
3392 enum hrtimer_mode mode;
3396 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3400 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3402 * 0: use half of prev avg
3403 * >0: use this specific value
3405 if (q->poll_nsec > 0)
3406 nsecs = q->poll_nsec;
3408 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3413 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3416 * This will be replaced with the stats tracking code, using
3417 * 'avg_completion_time / 2' as the pre-sleep target.
3421 mode = HRTIMER_MODE_REL;
3422 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3423 hrtimer_set_expires(&hs.timer, kt);
3425 hrtimer_init_sleeper(&hs, current);
3427 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3429 set_current_state(TASK_UNINTERRUPTIBLE);
3430 hrtimer_start_expires(&hs.timer, mode);
3433 hrtimer_cancel(&hs.timer);
3434 mode = HRTIMER_MODE_ABS;
3435 } while (hs.task && !signal_pending(current));
3437 __set_current_state(TASK_RUNNING);
3438 destroy_hrtimer_on_stack(&hs.timer);
3442 static bool blk_mq_poll_hybrid(struct request_queue *q,
3443 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3447 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3450 if (!blk_qc_t_is_internal(cookie))
3451 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3453 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3455 * With scheduling, if the request has completed, we'll
3456 * get a NULL return here, as we clear the sched tag when
3457 * that happens. The request still remains valid, like always,
3458 * so we should be safe with just the NULL check.
3464 return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3468 * blk_poll - poll for IO completions
3470 * @cookie: cookie passed back at IO submission time
3471 * @spin: whether to spin for completions
3474 * Poll for completions on the passed in queue. Returns number of
3475 * completed entries found. If @spin is true, then blk_poll will continue
3476 * looping until at least one completion is found, unless the task is
3477 * otherwise marked running (or we need to reschedule).
3479 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3481 struct blk_mq_hw_ctx *hctx;
3484 if (!blk_qc_t_valid(cookie) ||
3485 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3489 blk_flush_plug_list(current->plug, false);
3491 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3494 * If we sleep, have the caller restart the poll loop to reset
3495 * the state. Like for the other success return cases, the
3496 * caller is responsible for checking if the IO completed. If
3497 * the IO isn't complete, we'll get called again and will go
3498 * straight to the busy poll loop.
3500 if (blk_mq_poll_hybrid(q, hctx, cookie))
3503 hctx->poll_considered++;
3505 state = current->state;
3509 hctx->poll_invoked++;
3511 ret = q->mq_ops->poll(hctx);
3513 hctx->poll_success++;
3514 __set_current_state(TASK_RUNNING);
3518 if (signal_pending_state(state, current))
3519 __set_current_state(TASK_RUNNING);
3521 if (current->state == TASK_RUNNING)
3523 if (ret < 0 || !spin)
3526 } while (!need_resched());
3528 __set_current_state(TASK_RUNNING);
3531 EXPORT_SYMBOL_GPL(blk_poll);
3533 unsigned int blk_mq_rq_cpu(struct request *rq)
3535 return rq->mq_ctx->cpu;
3537 EXPORT_SYMBOL(blk_mq_rq_cpu);
3539 static int __init blk_mq_init(void)
3541 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3542 blk_mq_hctx_notify_dead);
3545 subsys_initcall(blk_mq_init);