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[2];
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;
105 if (rq->part == mi->part)
106 mi->inflight[rq_data_dir(rq)]++;
111 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
113 struct mq_inflight mi = { .part = part };
115 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
117 return mi.inflight[0] + mi.inflight[1];
120 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
121 unsigned int inflight[2])
123 struct mq_inflight mi = { .part = part };
125 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
126 inflight[0] = mi.inflight[0];
127 inflight[1] = mi.inflight[1];
130 void blk_freeze_queue_start(struct request_queue *q)
132 mutex_lock(&q->mq_freeze_lock);
133 if (++q->mq_freeze_depth == 1) {
134 percpu_ref_kill(&q->q_usage_counter);
135 mutex_unlock(&q->mq_freeze_lock);
137 blk_mq_run_hw_queues(q, false);
139 mutex_unlock(&q->mq_freeze_lock);
142 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
144 void blk_mq_freeze_queue_wait(struct request_queue *q)
146 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
148 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
150 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
151 unsigned long timeout)
153 return wait_event_timeout(q->mq_freeze_wq,
154 percpu_ref_is_zero(&q->q_usage_counter),
157 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
160 * Guarantee no request is in use, so we can change any data structure of
161 * the queue afterward.
163 void blk_freeze_queue(struct request_queue *q)
166 * In the !blk_mq case we are only calling this to kill the
167 * q_usage_counter, otherwise this increases the freeze depth
168 * and waits for it to return to zero. For this reason there is
169 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
170 * exported to drivers as the only user for unfreeze is blk_mq.
172 blk_freeze_queue_start(q);
173 blk_mq_freeze_queue_wait(q);
176 void blk_mq_freeze_queue(struct request_queue *q)
179 * ...just an alias to keep freeze and unfreeze actions balanced
180 * in the blk_mq_* namespace
184 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
186 void blk_mq_unfreeze_queue(struct request_queue *q)
188 mutex_lock(&q->mq_freeze_lock);
189 q->mq_freeze_depth--;
190 WARN_ON_ONCE(q->mq_freeze_depth < 0);
191 if (!q->mq_freeze_depth) {
192 percpu_ref_resurrect(&q->q_usage_counter);
193 wake_up_all(&q->mq_freeze_wq);
195 mutex_unlock(&q->mq_freeze_lock);
197 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
200 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
201 * mpt3sas driver such that this function can be removed.
203 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
205 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
207 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
210 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
213 * Note: this function does not prevent that the struct request end_io()
214 * callback function is invoked. Once this function is returned, we make
215 * sure no dispatch can happen until the queue is unquiesced via
216 * blk_mq_unquiesce_queue().
218 void blk_mq_quiesce_queue(struct request_queue *q)
220 struct blk_mq_hw_ctx *hctx;
224 blk_mq_quiesce_queue_nowait(q);
226 queue_for_each_hw_ctx(q, hctx, i) {
227 if (hctx->flags & BLK_MQ_F_BLOCKING)
228 synchronize_srcu(hctx->srcu);
235 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
238 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
241 * This function recovers queue into the state before quiescing
242 * which is done by blk_mq_quiesce_queue.
244 void blk_mq_unquiesce_queue(struct request_queue *q)
246 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
248 /* dispatch requests which are inserted during quiescing */
249 blk_mq_run_hw_queues(q, true);
251 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
253 void blk_mq_wake_waiters(struct request_queue *q)
255 struct blk_mq_hw_ctx *hctx;
258 queue_for_each_hw_ctx(q, hctx, i)
259 if (blk_mq_hw_queue_mapped(hctx))
260 blk_mq_tag_wakeup_all(hctx->tags, true);
264 * Only need start/end time stamping if we have iostat or
265 * blk stats enabled, or using an IO scheduler.
267 static inline bool blk_mq_need_time_stamp(struct request *rq)
269 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
272 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
273 unsigned int tag, unsigned int op, u64 alloc_time_ns)
275 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
276 struct request *rq = tags->static_rqs[tag];
277 req_flags_t rq_flags = 0;
279 if (data->flags & BLK_MQ_REQ_INTERNAL) {
281 rq->internal_tag = tag;
283 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
284 rq_flags = RQF_MQ_INFLIGHT;
285 atomic_inc(&data->hctx->nr_active);
288 rq->internal_tag = -1;
289 data->hctx->tags->rqs[rq->tag] = rq;
292 /* csd/requeue_work/fifo_time is initialized before use */
294 rq->mq_ctx = data->ctx;
295 rq->mq_hctx = data->hctx;
296 rq->rq_flags = rq_flags;
298 if (data->flags & BLK_MQ_REQ_PREEMPT)
299 rq->rq_flags |= RQF_PREEMPT;
300 if (blk_queue_io_stat(data->q))
301 rq->rq_flags |= RQF_IO_STAT;
302 INIT_LIST_HEAD(&rq->queuelist);
303 INIT_HLIST_NODE(&rq->hash);
304 RB_CLEAR_NODE(&rq->rb_node);
307 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
308 rq->alloc_time_ns = alloc_time_ns;
310 if (blk_mq_need_time_stamp(rq))
311 rq->start_time_ns = ktime_get_ns();
313 rq->start_time_ns = 0;
314 rq->io_start_time_ns = 0;
315 rq->stats_sectors = 0;
316 rq->nr_phys_segments = 0;
317 #if defined(CONFIG_BLK_DEV_INTEGRITY)
318 rq->nr_integrity_segments = 0;
320 /* tag was already set */
321 WRITE_ONCE(rq->deadline, 0);
326 rq->end_io_data = NULL;
328 data->ctx->rq_dispatched[op_is_sync(op)]++;
329 refcount_set(&rq->ref, 1);
333 static struct request *blk_mq_get_request(struct request_queue *q,
335 struct blk_mq_alloc_data *data)
337 struct elevator_queue *e = q->elevator;
340 bool clear_ctx_on_error = false;
341 u64 alloc_time_ns = 0;
343 blk_queue_enter_live(q);
345 /* alloc_time includes depth and tag waits */
346 if (blk_queue_rq_alloc_time(q))
347 alloc_time_ns = ktime_get_ns();
350 if (likely(!data->ctx)) {
351 data->ctx = blk_mq_get_ctx(q);
352 clear_ctx_on_error = true;
354 if (likely(!data->hctx))
355 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
357 if (data->cmd_flags & REQ_NOWAIT)
358 data->flags |= BLK_MQ_REQ_NOWAIT;
361 data->flags |= BLK_MQ_REQ_INTERNAL;
364 * Flush requests are special and go directly to the
365 * dispatch list. Don't include reserved tags in the
366 * limiting, as it isn't useful.
368 if (!op_is_flush(data->cmd_flags) &&
369 e->type->ops.limit_depth &&
370 !(data->flags & BLK_MQ_REQ_RESERVED))
371 e->type->ops.limit_depth(data->cmd_flags, data);
373 blk_mq_tag_busy(data->hctx);
376 tag = blk_mq_get_tag(data);
377 if (tag == BLK_MQ_TAG_FAIL) {
378 if (clear_ctx_on_error)
384 rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
385 if (!op_is_flush(data->cmd_flags)) {
387 if (e && e->type->ops.prepare_request) {
388 if (e->type->icq_cache)
389 blk_mq_sched_assign_ioc(rq);
391 e->type->ops.prepare_request(rq, bio);
392 rq->rq_flags |= RQF_ELVPRIV;
395 data->hctx->queued++;
399 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
400 blk_mq_req_flags_t flags)
402 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
406 ret = blk_queue_enter(q, flags);
410 rq = blk_mq_get_request(q, NULL, &alloc_data);
414 return ERR_PTR(-EWOULDBLOCK);
417 rq->__sector = (sector_t) -1;
418 rq->bio = rq->biotail = NULL;
421 EXPORT_SYMBOL(blk_mq_alloc_request);
423 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
424 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
426 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
432 * If the tag allocator sleeps we could get an allocation for a
433 * different hardware context. No need to complicate the low level
434 * allocator for this for the rare use case of a command tied to
437 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
438 return ERR_PTR(-EINVAL);
440 if (hctx_idx >= q->nr_hw_queues)
441 return ERR_PTR(-EIO);
443 ret = blk_queue_enter(q, flags);
448 * Check if the hardware context is actually mapped to anything.
449 * If not tell the caller that it should skip this queue.
451 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
452 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
454 return ERR_PTR(-EXDEV);
456 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
457 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
459 rq = blk_mq_get_request(q, NULL, &alloc_data);
463 return ERR_PTR(-EWOULDBLOCK);
467 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
469 static void __blk_mq_free_request(struct request *rq)
471 struct request_queue *q = rq->q;
472 struct blk_mq_ctx *ctx = rq->mq_ctx;
473 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
474 const int sched_tag = rq->internal_tag;
476 blk_pm_mark_last_busy(rq);
479 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
481 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
482 blk_mq_sched_restart(hctx);
486 void blk_mq_free_request(struct request *rq)
488 struct request_queue *q = rq->q;
489 struct elevator_queue *e = q->elevator;
490 struct blk_mq_ctx *ctx = rq->mq_ctx;
491 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
493 if (rq->rq_flags & RQF_ELVPRIV) {
494 if (e && e->type->ops.finish_request)
495 e->type->ops.finish_request(rq);
497 put_io_context(rq->elv.icq->ioc);
502 ctx->rq_completed[rq_is_sync(rq)]++;
503 if (rq->rq_flags & RQF_MQ_INFLIGHT)
504 atomic_dec(&hctx->nr_active);
506 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
507 laptop_io_completion(q->backing_dev_info);
511 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
512 if (refcount_dec_and_test(&rq->ref))
513 __blk_mq_free_request(rq);
515 EXPORT_SYMBOL_GPL(blk_mq_free_request);
517 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
521 if (blk_mq_need_time_stamp(rq))
522 now = ktime_get_ns();
524 if (rq->rq_flags & RQF_STATS) {
525 blk_mq_poll_stats_start(rq->q);
526 blk_stat_add(rq, now);
529 if (rq->internal_tag != -1)
530 blk_mq_sched_completed_request(rq, now);
532 blk_account_io_done(rq, now);
535 rq_qos_done(rq->q, rq);
536 rq->end_io(rq, error);
538 blk_mq_free_request(rq);
541 EXPORT_SYMBOL(__blk_mq_end_request);
543 void blk_mq_end_request(struct request *rq, blk_status_t error)
545 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
547 __blk_mq_end_request(rq, error);
549 EXPORT_SYMBOL(blk_mq_end_request);
551 static void __blk_mq_complete_request_remote(void *data)
553 struct request *rq = data;
554 struct request_queue *q = rq->q;
556 q->mq_ops->complete(rq);
559 static void __blk_mq_complete_request(struct request *rq)
561 struct blk_mq_ctx *ctx = rq->mq_ctx;
562 struct request_queue *q = rq->q;
566 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
568 * Most of single queue controllers, there is only one irq vector
569 * for handling IO completion, and the only irq's affinity is set
570 * as all possible CPUs. On most of ARCHs, this affinity means the
571 * irq is handled on one specific CPU.
573 * So complete IO reqeust in softirq context in case of single queue
574 * for not degrading IO performance by irqsoff latency.
576 if (q->nr_hw_queues == 1) {
577 __blk_complete_request(rq);
582 * For a polled request, always complete locallly, it's pointless
583 * to redirect the completion.
585 if ((rq->cmd_flags & REQ_HIPRI) ||
586 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
587 q->mq_ops->complete(rq);
592 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
593 shared = cpus_share_cache(cpu, ctx->cpu);
595 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
596 rq->csd.func = __blk_mq_complete_request_remote;
599 smp_call_function_single_async(ctx->cpu, &rq->csd);
601 q->mq_ops->complete(rq);
606 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
607 __releases(hctx->srcu)
609 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
612 srcu_read_unlock(hctx->srcu, srcu_idx);
615 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
616 __acquires(hctx->srcu)
618 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
619 /* shut up gcc false positive */
623 *srcu_idx = srcu_read_lock(hctx->srcu);
627 * blk_mq_complete_request - end I/O on a request
628 * @rq: the request being processed
631 * Ends all I/O on a request. It does not handle partial completions.
632 * The actual completion happens out-of-order, through a IPI handler.
634 bool blk_mq_complete_request(struct request *rq)
636 if (unlikely(blk_should_fake_timeout(rq->q)))
638 __blk_mq_complete_request(rq);
641 EXPORT_SYMBOL(blk_mq_complete_request);
644 * blk_mq_start_request - Start processing a request
645 * @rq: Pointer to request to be started
647 * Function used by device drivers to notify the block layer that a request
648 * is going to be processed now, so blk layer can do proper initializations
649 * such as starting the timeout timer.
651 void blk_mq_start_request(struct request *rq)
653 struct request_queue *q = rq->q;
655 trace_block_rq_issue(q, rq);
657 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
658 rq->io_start_time_ns = ktime_get_ns();
659 rq->stats_sectors = blk_rq_sectors(rq);
660 rq->rq_flags |= RQF_STATS;
664 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
667 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
669 #ifdef CONFIG_BLK_DEV_INTEGRITY
670 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
671 q->integrity.profile->prepare_fn(rq);
674 EXPORT_SYMBOL(blk_mq_start_request);
676 static void __blk_mq_requeue_request(struct request *rq)
678 struct request_queue *q = rq->q;
680 blk_mq_put_driver_tag(rq);
682 trace_block_rq_requeue(q, rq);
683 rq_qos_requeue(q, rq);
685 if (blk_mq_request_started(rq)) {
686 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
687 rq->rq_flags &= ~RQF_TIMED_OUT;
691 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
693 __blk_mq_requeue_request(rq);
695 /* this request will be re-inserted to io scheduler queue */
696 blk_mq_sched_requeue_request(rq);
698 BUG_ON(!list_empty(&rq->queuelist));
699 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
701 EXPORT_SYMBOL(blk_mq_requeue_request);
703 static void blk_mq_requeue_work(struct work_struct *work)
705 struct request_queue *q =
706 container_of(work, struct request_queue, requeue_work.work);
708 struct request *rq, *next;
710 spin_lock_irq(&q->requeue_lock);
711 list_splice_init(&q->requeue_list, &rq_list);
712 spin_unlock_irq(&q->requeue_lock);
714 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
715 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
718 rq->rq_flags &= ~RQF_SOFTBARRIER;
719 list_del_init(&rq->queuelist);
721 * If RQF_DONTPREP, rq has contained some driver specific
722 * data, so insert it to hctx dispatch list to avoid any
725 if (rq->rq_flags & RQF_DONTPREP)
726 blk_mq_request_bypass_insert(rq, false, false);
728 blk_mq_sched_insert_request(rq, true, false, false);
731 while (!list_empty(&rq_list)) {
732 rq = list_entry(rq_list.next, struct request, queuelist);
733 list_del_init(&rq->queuelist);
734 blk_mq_sched_insert_request(rq, false, false, false);
737 blk_mq_run_hw_queues(q, false);
740 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
741 bool kick_requeue_list)
743 struct request_queue *q = rq->q;
747 * We abuse this flag that is otherwise used by the I/O scheduler to
748 * request head insertion from the workqueue.
750 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
752 spin_lock_irqsave(&q->requeue_lock, flags);
754 rq->rq_flags |= RQF_SOFTBARRIER;
755 list_add(&rq->queuelist, &q->requeue_list);
757 list_add_tail(&rq->queuelist, &q->requeue_list);
759 spin_unlock_irqrestore(&q->requeue_lock, flags);
761 if (kick_requeue_list)
762 blk_mq_kick_requeue_list(q);
765 void blk_mq_kick_requeue_list(struct request_queue *q)
767 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
769 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
771 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
774 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
775 msecs_to_jiffies(msecs));
777 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
779 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
781 if (tag < tags->nr_tags) {
782 prefetch(tags->rqs[tag]);
783 return tags->rqs[tag];
788 EXPORT_SYMBOL(blk_mq_tag_to_rq);
790 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
791 void *priv, bool reserved)
794 * If we find a request that is inflight and the queue matches,
795 * we know the queue is busy. Return false to stop the iteration.
797 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
807 bool blk_mq_queue_inflight(struct request_queue *q)
811 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
814 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
816 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
818 req->rq_flags |= RQF_TIMED_OUT;
819 if (req->q->mq_ops->timeout) {
820 enum blk_eh_timer_return ret;
822 ret = req->q->mq_ops->timeout(req, reserved);
823 if (ret == BLK_EH_DONE)
825 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
831 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
833 unsigned long deadline;
835 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
837 if (rq->rq_flags & RQF_TIMED_OUT)
840 deadline = READ_ONCE(rq->deadline);
841 if (time_after_eq(jiffies, deadline))
846 else if (time_after(*next, deadline))
851 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
852 struct request *rq, void *priv, bool reserved)
854 unsigned long *next = priv;
857 * Just do a quick check if it is expired before locking the request in
858 * so we're not unnecessarilly synchronizing across CPUs.
860 if (!blk_mq_req_expired(rq, next))
864 * We have reason to believe the request may be expired. Take a
865 * reference on the request to lock this request lifetime into its
866 * currently allocated context to prevent it from being reallocated in
867 * the event the completion by-passes this timeout handler.
869 * If the reference was already released, then the driver beat the
870 * timeout handler to posting a natural completion.
872 if (!refcount_inc_not_zero(&rq->ref))
876 * The request is now locked and cannot be reallocated underneath the
877 * timeout handler's processing. Re-verify this exact request is truly
878 * expired; if it is not expired, then the request was completed and
879 * reallocated as a new request.
881 if (blk_mq_req_expired(rq, next))
882 blk_mq_rq_timed_out(rq, reserved);
884 if (is_flush_rq(rq, hctx))
886 else if (refcount_dec_and_test(&rq->ref))
887 __blk_mq_free_request(rq);
892 static void blk_mq_timeout_work(struct work_struct *work)
894 struct request_queue *q =
895 container_of(work, struct request_queue, timeout_work);
896 unsigned long next = 0;
897 struct blk_mq_hw_ctx *hctx;
900 /* A deadlock might occur if a request is stuck requiring a
901 * timeout at the same time a queue freeze is waiting
902 * completion, since the timeout code would not be able to
903 * acquire the queue reference here.
905 * That's why we don't use blk_queue_enter here; instead, we use
906 * percpu_ref_tryget directly, because we need to be able to
907 * obtain a reference even in the short window between the queue
908 * starting to freeze, by dropping the first reference in
909 * blk_freeze_queue_start, and the moment the last request is
910 * consumed, marked by the instant q_usage_counter reaches
913 if (!percpu_ref_tryget(&q->q_usage_counter))
916 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
919 mod_timer(&q->timeout, next);
922 * Request timeouts are handled as a forward rolling timer. If
923 * we end up here it means that no requests are pending and
924 * also that no request has been pending for a while. Mark
927 queue_for_each_hw_ctx(q, hctx, i) {
928 /* the hctx may be unmapped, so check it here */
929 if (blk_mq_hw_queue_mapped(hctx))
930 blk_mq_tag_idle(hctx);
936 struct flush_busy_ctx_data {
937 struct blk_mq_hw_ctx *hctx;
938 struct list_head *list;
941 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
943 struct flush_busy_ctx_data *flush_data = data;
944 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
945 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
946 enum hctx_type type = hctx->type;
948 spin_lock(&ctx->lock);
949 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
950 sbitmap_clear_bit(sb, bitnr);
951 spin_unlock(&ctx->lock);
956 * Process software queues that have been marked busy, splicing them
957 * to the for-dispatch
959 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
961 struct flush_busy_ctx_data data = {
966 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
968 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
970 struct dispatch_rq_data {
971 struct blk_mq_hw_ctx *hctx;
975 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
978 struct dispatch_rq_data *dispatch_data = data;
979 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
980 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
981 enum hctx_type type = hctx->type;
983 spin_lock(&ctx->lock);
984 if (!list_empty(&ctx->rq_lists[type])) {
985 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
986 list_del_init(&dispatch_data->rq->queuelist);
987 if (list_empty(&ctx->rq_lists[type]))
988 sbitmap_clear_bit(sb, bitnr);
990 spin_unlock(&ctx->lock);
992 return !dispatch_data->rq;
995 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
996 struct blk_mq_ctx *start)
998 unsigned off = start ? start->index_hw[hctx->type] : 0;
999 struct dispatch_rq_data data = {
1004 __sbitmap_for_each_set(&hctx->ctx_map, off,
1005 dispatch_rq_from_ctx, &data);
1010 static inline unsigned int queued_to_index(unsigned int queued)
1015 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1018 bool blk_mq_get_driver_tag(struct request *rq)
1020 struct blk_mq_alloc_data data = {
1022 .hctx = rq->mq_hctx,
1023 .flags = BLK_MQ_REQ_NOWAIT,
1024 .cmd_flags = rq->cmd_flags,
1031 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1032 data.flags |= BLK_MQ_REQ_RESERVED;
1034 shared = blk_mq_tag_busy(data.hctx);
1035 rq->tag = blk_mq_get_tag(&data);
1038 rq->rq_flags |= RQF_MQ_INFLIGHT;
1039 atomic_inc(&data.hctx->nr_active);
1041 data.hctx->tags->rqs[rq->tag] = rq;
1044 return rq->tag != -1;
1047 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1048 int flags, void *key)
1050 struct blk_mq_hw_ctx *hctx;
1052 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1054 spin_lock(&hctx->dispatch_wait_lock);
1055 if (!list_empty(&wait->entry)) {
1056 struct sbitmap_queue *sbq;
1058 list_del_init(&wait->entry);
1059 sbq = &hctx->tags->bitmap_tags;
1060 atomic_dec(&sbq->ws_active);
1062 spin_unlock(&hctx->dispatch_wait_lock);
1064 blk_mq_run_hw_queue(hctx, true);
1069 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1070 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1071 * restart. For both cases, take care to check the condition again after
1072 * marking us as waiting.
1074 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1077 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1078 struct wait_queue_head *wq;
1079 wait_queue_entry_t *wait;
1082 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1083 blk_mq_sched_mark_restart_hctx(hctx);
1086 * It's possible that a tag was freed in the window between the
1087 * allocation failure and adding the hardware queue to the wait
1090 * Don't clear RESTART here, someone else could have set it.
1091 * At most this will cost an extra queue run.
1093 return blk_mq_get_driver_tag(rq);
1096 wait = &hctx->dispatch_wait;
1097 if (!list_empty_careful(&wait->entry))
1100 wq = &bt_wait_ptr(sbq, hctx)->wait;
1102 spin_lock_irq(&wq->lock);
1103 spin_lock(&hctx->dispatch_wait_lock);
1104 if (!list_empty(&wait->entry)) {
1105 spin_unlock(&hctx->dispatch_wait_lock);
1106 spin_unlock_irq(&wq->lock);
1110 atomic_inc(&sbq->ws_active);
1111 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1112 __add_wait_queue(wq, wait);
1115 * It's possible that a tag was freed in the window between the
1116 * allocation failure and adding the hardware queue to the wait
1119 ret = blk_mq_get_driver_tag(rq);
1121 spin_unlock(&hctx->dispatch_wait_lock);
1122 spin_unlock_irq(&wq->lock);
1127 * We got a tag, remove ourselves from the wait queue to ensure
1128 * someone else gets the wakeup.
1130 list_del_init(&wait->entry);
1131 atomic_dec(&sbq->ws_active);
1132 spin_unlock(&hctx->dispatch_wait_lock);
1133 spin_unlock_irq(&wq->lock);
1138 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1139 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1141 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1142 * - EWMA is one simple way to compute running average value
1143 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1144 * - take 4 as factor for avoiding to get too small(0) result, and this
1145 * factor doesn't matter because EWMA decreases exponentially
1147 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1151 if (hctx->queue->elevator)
1154 ewma = hctx->dispatch_busy;
1159 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1161 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1162 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1164 hctx->dispatch_busy = ewma;
1167 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1169 static void blk_mq_handle_dev_resource(struct request *rq,
1170 struct list_head *list)
1172 struct request *next =
1173 list_first_entry_or_null(list, struct request, queuelist);
1176 * If an I/O scheduler has been configured and we got a driver tag for
1177 * the next request already, free it.
1180 blk_mq_put_driver_tag(next);
1182 list_add(&rq->queuelist, list);
1183 __blk_mq_requeue_request(rq);
1187 * Returns true if we did some work AND can potentially do more.
1189 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1192 struct blk_mq_hw_ctx *hctx;
1193 struct request *rq, *nxt;
1194 bool no_tag = false;
1196 blk_status_t ret = BLK_STS_OK;
1197 bool no_budget_avail = false;
1199 if (list_empty(list))
1202 WARN_ON(!list_is_singular(list) && got_budget);
1205 * Now process all the entries, sending them to the driver.
1207 errors = queued = 0;
1209 struct blk_mq_queue_data bd;
1211 rq = list_first_entry(list, struct request, queuelist);
1214 if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) {
1215 blk_mq_put_driver_tag(rq);
1216 no_budget_avail = true;
1220 if (!blk_mq_get_driver_tag(rq)) {
1222 * The initial allocation attempt failed, so we need to
1223 * rerun the hardware queue when a tag is freed. The
1224 * waitqueue takes care of that. If the queue is run
1225 * before we add this entry back on the dispatch list,
1226 * we'll re-run it below.
1228 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1229 blk_mq_put_dispatch_budget(hctx);
1231 * For non-shared tags, the RESTART check
1234 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1240 list_del_init(&rq->queuelist);
1245 * Flag last if we have no more requests, or if we have more
1246 * but can't assign a driver tag to it.
1248 if (list_empty(list))
1251 nxt = list_first_entry(list, struct request, queuelist);
1252 bd.last = !blk_mq_get_driver_tag(nxt);
1255 ret = q->mq_ops->queue_rq(hctx, &bd);
1256 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1257 blk_mq_handle_dev_resource(rq, list);
1261 if (unlikely(ret != BLK_STS_OK)) {
1263 blk_mq_end_request(rq, BLK_STS_IOERR);
1268 } while (!list_empty(list));
1270 hctx->dispatched[queued_to_index(queued)]++;
1273 * Any items that need requeuing? Stuff them into hctx->dispatch,
1274 * that is where we will continue on next queue run.
1276 if (!list_empty(list)) {
1280 * If we didn't flush the entire list, we could have told
1281 * the driver there was more coming, but that turned out to
1284 if (q->mq_ops->commit_rqs && queued)
1285 q->mq_ops->commit_rqs(hctx);
1287 spin_lock(&hctx->lock);
1288 list_splice_tail_init(list, &hctx->dispatch);
1289 spin_unlock(&hctx->lock);
1292 * If SCHED_RESTART was set by the caller of this function and
1293 * it is no longer set that means that it was cleared by another
1294 * thread and hence that a queue rerun is needed.
1296 * If 'no_tag' is set, that means that we failed getting
1297 * a driver tag with an I/O scheduler attached. If our dispatch
1298 * waitqueue is no longer active, ensure that we run the queue
1299 * AFTER adding our entries back to the list.
1301 * If no I/O scheduler has been configured it is possible that
1302 * the hardware queue got stopped and restarted before requests
1303 * were pushed back onto the dispatch list. Rerun the queue to
1304 * avoid starvation. Notes:
1305 * - blk_mq_run_hw_queue() checks whether or not a queue has
1306 * been stopped before rerunning a queue.
1307 * - Some but not all block drivers stop a queue before
1308 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1311 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1312 * bit is set, run queue after a delay to avoid IO stalls
1313 * that could otherwise occur if the queue is idle. We'll do
1314 * similar if we couldn't get budget and SCHED_RESTART is set.
1316 needs_restart = blk_mq_sched_needs_restart(hctx);
1317 if (!needs_restart ||
1318 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1319 blk_mq_run_hw_queue(hctx, true);
1320 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1322 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1324 blk_mq_update_dispatch_busy(hctx, true);
1327 blk_mq_update_dispatch_busy(hctx, false);
1330 * If the host/device is unable to accept more work, inform the
1333 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1336 return (queued + errors) != 0;
1340 * __blk_mq_run_hw_queue - Run a hardware queue.
1341 * @hctx: Pointer to the hardware queue to run.
1343 * Send pending requests to the hardware.
1345 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1350 * We should be running this queue from one of the CPUs that
1353 * There are at least two related races now between setting
1354 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1355 * __blk_mq_run_hw_queue():
1357 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1358 * but later it becomes online, then this warning is harmless
1361 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1362 * but later it becomes offline, then the warning can't be
1363 * triggered, and we depend on blk-mq timeout handler to
1364 * handle dispatched requests to this hctx
1366 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1367 cpu_online(hctx->next_cpu)) {
1368 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1369 raw_smp_processor_id(),
1370 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1375 * We can't run the queue inline with ints disabled. Ensure that
1376 * we catch bad users of this early.
1378 WARN_ON_ONCE(in_interrupt());
1380 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1382 hctx_lock(hctx, &srcu_idx);
1383 blk_mq_sched_dispatch_requests(hctx);
1384 hctx_unlock(hctx, srcu_idx);
1387 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1389 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1391 if (cpu >= nr_cpu_ids)
1392 cpu = cpumask_first(hctx->cpumask);
1397 * It'd be great if the workqueue API had a way to pass
1398 * in a mask and had some smarts for more clever placement.
1399 * For now we just round-robin here, switching for every
1400 * BLK_MQ_CPU_WORK_BATCH queued items.
1402 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1405 int next_cpu = hctx->next_cpu;
1407 if (hctx->queue->nr_hw_queues == 1)
1408 return WORK_CPU_UNBOUND;
1410 if (--hctx->next_cpu_batch <= 0) {
1412 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1414 if (next_cpu >= nr_cpu_ids)
1415 next_cpu = blk_mq_first_mapped_cpu(hctx);
1416 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1420 * Do unbound schedule if we can't find a online CPU for this hctx,
1421 * and it should only happen in the path of handling CPU DEAD.
1423 if (!cpu_online(next_cpu)) {
1430 * Make sure to re-select CPU next time once after CPUs
1431 * in hctx->cpumask become online again.
1433 hctx->next_cpu = next_cpu;
1434 hctx->next_cpu_batch = 1;
1435 return WORK_CPU_UNBOUND;
1438 hctx->next_cpu = next_cpu;
1443 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1444 * @hctx: Pointer to the hardware queue to run.
1445 * @async: If we want to run the queue asynchronously.
1446 * @msecs: Microseconds of delay to wait before running the queue.
1448 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1449 * with a delay of @msecs.
1451 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1452 unsigned long msecs)
1454 if (unlikely(blk_mq_hctx_stopped(hctx)))
1457 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1458 int cpu = get_cpu();
1459 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1460 __blk_mq_run_hw_queue(hctx);
1468 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1469 msecs_to_jiffies(msecs));
1473 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1474 * @hctx: Pointer to the hardware queue to run.
1475 * @msecs: Microseconds of delay to wait before running the queue.
1477 * Run a hardware queue asynchronously with a delay of @msecs.
1479 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1481 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1483 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1486 * blk_mq_run_hw_queue - Start to run a hardware queue.
1487 * @hctx: Pointer to the hardware queue to run.
1488 * @async: If we want to run the queue asynchronously.
1490 * Check if the request queue is not in a quiesced state and if there are
1491 * pending requests to be sent. If this is true, run the queue to send requests
1494 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1500 * When queue is quiesced, we may be switching io scheduler, or
1501 * updating nr_hw_queues, or other things, and we can't run queue
1502 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1504 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1507 hctx_lock(hctx, &srcu_idx);
1508 need_run = !blk_queue_quiesced(hctx->queue) &&
1509 blk_mq_hctx_has_pending(hctx);
1510 hctx_unlock(hctx, srcu_idx);
1513 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1515 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1518 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1519 * @q: Pointer to the request queue to run.
1520 * @async: If we want to run the queue asynchronously.
1522 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1524 struct blk_mq_hw_ctx *hctx;
1527 queue_for_each_hw_ctx(q, hctx, i) {
1528 if (blk_mq_hctx_stopped(hctx))
1531 blk_mq_run_hw_queue(hctx, async);
1534 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1537 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1538 * @q: Pointer to the request queue to run.
1539 * @msecs: Microseconds of delay to wait before running the queues.
1541 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1543 struct blk_mq_hw_ctx *hctx;
1546 queue_for_each_hw_ctx(q, hctx, i) {
1547 if (blk_mq_hctx_stopped(hctx))
1550 blk_mq_delay_run_hw_queue(hctx, msecs);
1553 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1556 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1557 * @q: request queue.
1559 * The caller is responsible for serializing this function against
1560 * blk_mq_{start,stop}_hw_queue().
1562 bool blk_mq_queue_stopped(struct request_queue *q)
1564 struct blk_mq_hw_ctx *hctx;
1567 queue_for_each_hw_ctx(q, hctx, i)
1568 if (blk_mq_hctx_stopped(hctx))
1573 EXPORT_SYMBOL(blk_mq_queue_stopped);
1576 * This function is often used for pausing .queue_rq() by driver when
1577 * there isn't enough resource or some conditions aren't satisfied, and
1578 * BLK_STS_RESOURCE is usually returned.
1580 * We do not guarantee that dispatch can be drained or blocked
1581 * after blk_mq_stop_hw_queue() returns. Please use
1582 * blk_mq_quiesce_queue() for that requirement.
1584 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1586 cancel_delayed_work(&hctx->run_work);
1588 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1590 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1593 * This function is often used for pausing .queue_rq() by driver when
1594 * there isn't enough resource or some conditions aren't satisfied, and
1595 * BLK_STS_RESOURCE is usually returned.
1597 * We do not guarantee that dispatch can be drained or blocked
1598 * after blk_mq_stop_hw_queues() returns. Please use
1599 * blk_mq_quiesce_queue() for that requirement.
1601 void blk_mq_stop_hw_queues(struct request_queue *q)
1603 struct blk_mq_hw_ctx *hctx;
1606 queue_for_each_hw_ctx(q, hctx, i)
1607 blk_mq_stop_hw_queue(hctx);
1609 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1611 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1613 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1615 blk_mq_run_hw_queue(hctx, false);
1617 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1619 void blk_mq_start_hw_queues(struct request_queue *q)
1621 struct blk_mq_hw_ctx *hctx;
1624 queue_for_each_hw_ctx(q, hctx, i)
1625 blk_mq_start_hw_queue(hctx);
1627 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1629 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1631 if (!blk_mq_hctx_stopped(hctx))
1634 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1635 blk_mq_run_hw_queue(hctx, async);
1637 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1639 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1641 struct blk_mq_hw_ctx *hctx;
1644 queue_for_each_hw_ctx(q, hctx, i)
1645 blk_mq_start_stopped_hw_queue(hctx, async);
1647 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1649 static void blk_mq_run_work_fn(struct work_struct *work)
1651 struct blk_mq_hw_ctx *hctx;
1653 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1656 * If we are stopped, don't run the queue.
1658 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1661 __blk_mq_run_hw_queue(hctx);
1664 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1668 struct blk_mq_ctx *ctx = rq->mq_ctx;
1669 enum hctx_type type = hctx->type;
1671 lockdep_assert_held(&ctx->lock);
1673 trace_block_rq_insert(hctx->queue, rq);
1676 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1678 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1681 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1684 struct blk_mq_ctx *ctx = rq->mq_ctx;
1686 lockdep_assert_held(&ctx->lock);
1688 __blk_mq_insert_req_list(hctx, rq, at_head);
1689 blk_mq_hctx_mark_pending(hctx, ctx);
1693 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1694 * @rq: Pointer to request to be inserted.
1695 * @run_queue: If we should run the hardware queue after inserting the request.
1697 * Should only be used carefully, when the caller knows we want to
1698 * bypass a potential IO scheduler on the target device.
1700 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1703 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1705 spin_lock(&hctx->lock);
1707 list_add(&rq->queuelist, &hctx->dispatch);
1709 list_add_tail(&rq->queuelist, &hctx->dispatch);
1710 spin_unlock(&hctx->lock);
1713 blk_mq_run_hw_queue(hctx, false);
1716 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1717 struct list_head *list)
1721 enum hctx_type type = hctx->type;
1724 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1727 list_for_each_entry(rq, list, queuelist) {
1728 BUG_ON(rq->mq_ctx != ctx);
1729 trace_block_rq_insert(hctx->queue, rq);
1732 spin_lock(&ctx->lock);
1733 list_splice_tail_init(list, &ctx->rq_lists[type]);
1734 blk_mq_hctx_mark_pending(hctx, ctx);
1735 spin_unlock(&ctx->lock);
1738 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1740 struct request *rqa = container_of(a, struct request, queuelist);
1741 struct request *rqb = container_of(b, struct request, queuelist);
1743 if (rqa->mq_ctx != rqb->mq_ctx)
1744 return rqa->mq_ctx > rqb->mq_ctx;
1745 if (rqa->mq_hctx != rqb->mq_hctx)
1746 return rqa->mq_hctx > rqb->mq_hctx;
1748 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1751 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1755 if (list_empty(&plug->mq_list))
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);
1765 struct list_head rq_list;
1766 struct request *rq, *head_rq = list_entry_rq(list.next);
1767 struct list_head *pos = &head_rq->queuelist; /* skip first */
1768 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1769 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1770 unsigned int depth = 1;
1772 list_for_each_continue(pos, &list) {
1773 rq = list_entry_rq(pos);
1775 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1780 list_cut_before(&rq_list, &list, pos);
1781 trace_block_unplug(head_rq->q, depth, !from_schedule);
1782 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1784 } while(!list_empty(&list));
1787 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1788 unsigned int nr_segs)
1790 if (bio->bi_opf & REQ_RAHEAD)
1791 rq->cmd_flags |= REQ_FAILFAST_MASK;
1793 rq->__sector = bio->bi_iter.bi_sector;
1794 rq->write_hint = bio->bi_write_hint;
1795 blk_rq_bio_prep(rq, bio, nr_segs);
1797 blk_account_io_start(rq, true);
1800 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1802 blk_qc_t *cookie, bool last)
1804 struct request_queue *q = rq->q;
1805 struct blk_mq_queue_data bd = {
1809 blk_qc_t new_cookie;
1812 new_cookie = request_to_qc_t(hctx, rq);
1815 * For OK queue, we are done. For error, caller may kill it.
1816 * Any other error (busy), just add it to our list as we
1817 * previously would have done.
1819 ret = q->mq_ops->queue_rq(hctx, &bd);
1822 blk_mq_update_dispatch_busy(hctx, false);
1823 *cookie = new_cookie;
1825 case BLK_STS_RESOURCE:
1826 case BLK_STS_DEV_RESOURCE:
1827 blk_mq_update_dispatch_busy(hctx, true);
1828 __blk_mq_requeue_request(rq);
1831 blk_mq_update_dispatch_busy(hctx, false);
1832 *cookie = BLK_QC_T_NONE;
1839 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1842 bool bypass_insert, bool last)
1844 struct request_queue *q = rq->q;
1845 bool run_queue = true;
1848 * RCU or SRCU read lock is needed before checking quiesced flag.
1850 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1851 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1852 * and avoid driver to try to dispatch again.
1854 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1856 bypass_insert = false;
1860 if (q->elevator && !bypass_insert)
1863 if (!blk_mq_get_dispatch_budget(hctx))
1866 if (!blk_mq_get_driver_tag(rq)) {
1867 blk_mq_put_dispatch_budget(hctx);
1871 return __blk_mq_issue_directly(hctx, rq, cookie, last);
1874 return BLK_STS_RESOURCE;
1876 blk_mq_request_bypass_insert(rq, false, run_queue);
1881 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
1882 * @hctx: Pointer of the associated hardware queue.
1883 * @rq: Pointer to request to be sent.
1884 * @cookie: Request queue cookie.
1886 * If the device has enough resources to accept a new request now, send the
1887 * request directly to device driver. Else, insert at hctx->dispatch queue, so
1888 * we can try send it another time in the future. Requests inserted at this
1889 * queue have higher priority.
1891 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1892 struct request *rq, blk_qc_t *cookie)
1897 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1899 hctx_lock(hctx, &srcu_idx);
1901 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1902 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1903 blk_mq_request_bypass_insert(rq, false, true);
1904 else if (ret != BLK_STS_OK)
1905 blk_mq_end_request(rq, ret);
1907 hctx_unlock(hctx, srcu_idx);
1910 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1914 blk_qc_t unused_cookie;
1915 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1917 hctx_lock(hctx, &srcu_idx);
1918 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1919 hctx_unlock(hctx, srcu_idx);
1924 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1925 struct list_head *list)
1929 while (!list_empty(list)) {
1931 struct request *rq = list_first_entry(list, struct request,
1934 list_del_init(&rq->queuelist);
1935 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1936 if (ret != BLK_STS_OK) {
1937 if (ret == BLK_STS_RESOURCE ||
1938 ret == BLK_STS_DEV_RESOURCE) {
1939 blk_mq_request_bypass_insert(rq, false,
1943 blk_mq_end_request(rq, ret);
1949 * If we didn't flush the entire list, we could have told
1950 * the driver there was more coming, but that turned out to
1953 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs && queued)
1954 hctx->queue->mq_ops->commit_rqs(hctx);
1957 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1959 list_add_tail(&rq->queuelist, &plug->mq_list);
1961 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1962 struct request *tmp;
1964 tmp = list_first_entry(&plug->mq_list, struct request,
1966 if (tmp->q != rq->q)
1967 plug->multiple_queues = true;
1972 * blk_mq_make_request - Create and send a request to block device.
1973 * @q: Request queue pointer.
1974 * @bio: Bio pointer.
1976 * Builds up a request structure from @q and @bio and send to the device. The
1977 * request may not be queued directly to hardware if:
1978 * * This request can be merged with another one
1979 * * We want to place request at plug queue for possible future merging
1980 * * There is an IO scheduler active at this queue
1982 * It will not queue the request if there is an error with the bio, or at the
1985 * Returns: Request queue cookie.
1987 blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1989 const int is_sync = op_is_sync(bio->bi_opf);
1990 const int is_flush_fua = op_is_flush(bio->bi_opf);
1991 struct blk_mq_alloc_data data = { .flags = 0};
1993 struct blk_plug *plug;
1994 struct request *same_queue_rq = NULL;
1995 unsigned int nr_segs;
1998 blk_queue_bounce(q, &bio);
1999 __blk_queue_split(q, &bio, &nr_segs);
2001 if (!bio_integrity_prep(bio))
2002 return BLK_QC_T_NONE;
2004 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2005 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2006 return BLK_QC_T_NONE;
2008 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2009 return BLK_QC_T_NONE;
2011 rq_qos_throttle(q, bio);
2013 data.cmd_flags = bio->bi_opf;
2014 rq = blk_mq_get_request(q, bio, &data);
2015 if (unlikely(!rq)) {
2016 rq_qos_cleanup(q, bio);
2017 if (bio->bi_opf & REQ_NOWAIT)
2018 bio_wouldblock_error(bio);
2019 return BLK_QC_T_NONE;
2022 trace_block_getrq(q, bio, bio->bi_opf);
2024 rq_qos_track(q, rq, bio);
2026 cookie = request_to_qc_t(data.hctx, rq);
2028 blk_mq_bio_to_request(rq, bio, nr_segs);
2030 plug = blk_mq_plug(q, bio);
2031 if (unlikely(is_flush_fua)) {
2032 /* Bypass scheduler for flush requests */
2033 blk_insert_flush(rq);
2034 blk_mq_run_hw_queue(data.hctx, true);
2035 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2036 !blk_queue_nonrot(q))) {
2038 * Use plugging if we have a ->commit_rqs() hook as well, as
2039 * we know the driver uses bd->last in a smart fashion.
2041 * Use normal plugging if this disk is slow HDD, as sequential
2042 * IO may benefit a lot from plug merging.
2044 unsigned int request_count = plug->rq_count;
2045 struct request *last = NULL;
2048 trace_block_plug(q);
2050 last = list_entry_rq(plug->mq_list.prev);
2052 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2053 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2054 blk_flush_plug_list(plug, false);
2055 trace_block_plug(q);
2058 blk_add_rq_to_plug(plug, rq);
2059 } else if (q->elevator) {
2060 /* Insert the request at the IO scheduler queue */
2061 blk_mq_sched_insert_request(rq, false, true, true);
2062 } else if (plug && !blk_queue_nomerges(q)) {
2064 * We do limited plugging. If the bio can be merged, do that.
2065 * Otherwise the existing request in the plug list will be
2066 * issued. So the plug list will have one request at most
2067 * The plug list might get flushed before this. If that happens,
2068 * the plug list is empty, and same_queue_rq is invalid.
2070 if (list_empty(&plug->mq_list))
2071 same_queue_rq = NULL;
2072 if (same_queue_rq) {
2073 list_del_init(&same_queue_rq->queuelist);
2076 blk_add_rq_to_plug(plug, rq);
2077 trace_block_plug(q);
2079 if (same_queue_rq) {
2080 data.hctx = same_queue_rq->mq_hctx;
2081 trace_block_unplug(q, 1, true);
2082 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2085 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2086 !data.hctx->dispatch_busy) {
2088 * There is no scheduler and we can try to send directly
2091 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2094 blk_mq_sched_insert_request(rq, false, true, true);
2099 EXPORT_SYMBOL_GPL(blk_mq_make_request); /* only for request based dm */
2101 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2102 unsigned int hctx_idx)
2106 if (tags->rqs && set->ops->exit_request) {
2109 for (i = 0; i < tags->nr_tags; i++) {
2110 struct request *rq = tags->static_rqs[i];
2114 set->ops->exit_request(set, rq, hctx_idx);
2115 tags->static_rqs[i] = NULL;
2119 while (!list_empty(&tags->page_list)) {
2120 page = list_first_entry(&tags->page_list, struct page, lru);
2121 list_del_init(&page->lru);
2123 * Remove kmemleak object previously allocated in
2124 * blk_mq_alloc_rqs().
2126 kmemleak_free(page_address(page));
2127 __free_pages(page, page->private);
2131 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2135 kfree(tags->static_rqs);
2136 tags->static_rqs = NULL;
2138 blk_mq_free_tags(tags);
2141 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2142 unsigned int hctx_idx,
2143 unsigned int nr_tags,
2144 unsigned int reserved_tags)
2146 struct blk_mq_tags *tags;
2149 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2150 if (node == NUMA_NO_NODE)
2151 node = set->numa_node;
2153 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2154 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2158 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2159 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2162 blk_mq_free_tags(tags);
2166 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2167 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2169 if (!tags->static_rqs) {
2171 blk_mq_free_tags(tags);
2178 static size_t order_to_size(unsigned int order)
2180 return (size_t)PAGE_SIZE << order;
2183 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2184 unsigned int hctx_idx, int node)
2188 if (set->ops->init_request) {
2189 ret = set->ops->init_request(set, rq, hctx_idx, node);
2194 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2198 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2199 unsigned int hctx_idx, unsigned int depth)
2201 unsigned int i, j, entries_per_page, max_order = 4;
2202 size_t rq_size, left;
2205 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2206 if (node == NUMA_NO_NODE)
2207 node = set->numa_node;
2209 INIT_LIST_HEAD(&tags->page_list);
2212 * rq_size is the size of the request plus driver payload, rounded
2213 * to the cacheline size
2215 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2217 left = rq_size * depth;
2219 for (i = 0; i < depth; ) {
2220 int this_order = max_order;
2225 while (this_order && left < order_to_size(this_order - 1))
2229 page = alloc_pages_node(node,
2230 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2236 if (order_to_size(this_order) < rq_size)
2243 page->private = this_order;
2244 list_add_tail(&page->lru, &tags->page_list);
2246 p = page_address(page);
2248 * Allow kmemleak to scan these pages as they contain pointers
2249 * to additional allocations like via ops->init_request().
2251 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2252 entries_per_page = order_to_size(this_order) / rq_size;
2253 to_do = min(entries_per_page, depth - i);
2254 left -= to_do * rq_size;
2255 for (j = 0; j < to_do; j++) {
2256 struct request *rq = p;
2258 tags->static_rqs[i] = rq;
2259 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2260 tags->static_rqs[i] = NULL;
2271 blk_mq_free_rqs(set, tags, hctx_idx);
2276 * 'cpu' is going away. splice any existing rq_list entries from this
2277 * software queue to the hw queue dispatch list, and ensure that it
2280 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2282 struct blk_mq_hw_ctx *hctx;
2283 struct blk_mq_ctx *ctx;
2285 enum hctx_type type;
2287 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2288 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2291 spin_lock(&ctx->lock);
2292 if (!list_empty(&ctx->rq_lists[type])) {
2293 list_splice_init(&ctx->rq_lists[type], &tmp);
2294 blk_mq_hctx_clear_pending(hctx, ctx);
2296 spin_unlock(&ctx->lock);
2298 if (list_empty(&tmp))
2301 spin_lock(&hctx->lock);
2302 list_splice_tail_init(&tmp, &hctx->dispatch);
2303 spin_unlock(&hctx->lock);
2305 blk_mq_run_hw_queue(hctx, true);
2309 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2311 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2315 /* hctx->ctxs will be freed in queue's release handler */
2316 static void blk_mq_exit_hctx(struct request_queue *q,
2317 struct blk_mq_tag_set *set,
2318 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2320 if (blk_mq_hw_queue_mapped(hctx))
2321 blk_mq_tag_idle(hctx);
2323 if (set->ops->exit_request)
2324 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2326 if (set->ops->exit_hctx)
2327 set->ops->exit_hctx(hctx, hctx_idx);
2329 blk_mq_remove_cpuhp(hctx);
2331 spin_lock(&q->unused_hctx_lock);
2332 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2333 spin_unlock(&q->unused_hctx_lock);
2336 static void blk_mq_exit_hw_queues(struct request_queue *q,
2337 struct blk_mq_tag_set *set, int nr_queue)
2339 struct blk_mq_hw_ctx *hctx;
2342 queue_for_each_hw_ctx(q, hctx, i) {
2345 blk_mq_debugfs_unregister_hctx(hctx);
2346 blk_mq_exit_hctx(q, set, hctx, i);
2350 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2352 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2354 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2355 __alignof__(struct blk_mq_hw_ctx)) !=
2356 sizeof(struct blk_mq_hw_ctx));
2358 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2359 hw_ctx_size += sizeof(struct srcu_struct);
2364 static int blk_mq_init_hctx(struct request_queue *q,
2365 struct blk_mq_tag_set *set,
2366 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2368 hctx->queue_num = hctx_idx;
2370 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2372 hctx->tags = set->tags[hctx_idx];
2374 if (set->ops->init_hctx &&
2375 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2376 goto unregister_cpu_notifier;
2378 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2384 if (set->ops->exit_hctx)
2385 set->ops->exit_hctx(hctx, hctx_idx);
2386 unregister_cpu_notifier:
2387 blk_mq_remove_cpuhp(hctx);
2391 static struct blk_mq_hw_ctx *
2392 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2395 struct blk_mq_hw_ctx *hctx;
2396 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2398 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2400 goto fail_alloc_hctx;
2402 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2405 atomic_set(&hctx->nr_active, 0);
2406 if (node == NUMA_NO_NODE)
2407 node = set->numa_node;
2408 hctx->numa_node = node;
2410 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2411 spin_lock_init(&hctx->lock);
2412 INIT_LIST_HEAD(&hctx->dispatch);
2414 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2416 INIT_LIST_HEAD(&hctx->hctx_list);
2419 * Allocate space for all possible cpus to avoid allocation at
2422 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2427 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2432 spin_lock_init(&hctx->dispatch_wait_lock);
2433 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2434 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2436 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2440 if (hctx->flags & BLK_MQ_F_BLOCKING)
2441 init_srcu_struct(hctx->srcu);
2442 blk_mq_hctx_kobj_init(hctx);
2447 sbitmap_free(&hctx->ctx_map);
2451 free_cpumask_var(hctx->cpumask);
2458 static void blk_mq_init_cpu_queues(struct request_queue *q,
2459 unsigned int nr_hw_queues)
2461 struct blk_mq_tag_set *set = q->tag_set;
2464 for_each_possible_cpu(i) {
2465 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2466 struct blk_mq_hw_ctx *hctx;
2470 spin_lock_init(&__ctx->lock);
2471 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2472 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2477 * Set local node, IFF we have more than one hw queue. If
2478 * not, we remain on the home node of the device
2480 for (j = 0; j < set->nr_maps; j++) {
2481 hctx = blk_mq_map_queue_type(q, j, i);
2482 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2483 hctx->numa_node = local_memory_node(cpu_to_node(i));
2488 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2492 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2493 set->queue_depth, set->reserved_tags);
2494 if (!set->tags[hctx_idx])
2497 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2502 blk_mq_free_rq_map(set->tags[hctx_idx]);
2503 set->tags[hctx_idx] = NULL;
2507 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2508 unsigned int hctx_idx)
2510 if (set->tags && set->tags[hctx_idx]) {
2511 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2512 blk_mq_free_rq_map(set->tags[hctx_idx]);
2513 set->tags[hctx_idx] = NULL;
2517 static void blk_mq_map_swqueue(struct request_queue *q)
2519 unsigned int i, j, hctx_idx;
2520 struct blk_mq_hw_ctx *hctx;
2521 struct blk_mq_ctx *ctx;
2522 struct blk_mq_tag_set *set = q->tag_set;
2524 queue_for_each_hw_ctx(q, hctx, i) {
2525 cpumask_clear(hctx->cpumask);
2527 hctx->dispatch_from = NULL;
2531 * Map software to hardware queues.
2533 * If the cpu isn't present, the cpu is mapped to first hctx.
2535 for_each_possible_cpu(i) {
2536 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2537 /* unmapped hw queue can be remapped after CPU topo changed */
2538 if (!set->tags[hctx_idx] &&
2539 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2541 * If tags initialization fail for some hctx,
2542 * that hctx won't be brought online. In this
2543 * case, remap the current ctx to hctx[0] which
2544 * is guaranteed to always have tags allocated
2546 set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2549 ctx = per_cpu_ptr(q->queue_ctx, i);
2550 for (j = 0; j < set->nr_maps; j++) {
2551 if (!set->map[j].nr_queues) {
2552 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2553 HCTX_TYPE_DEFAULT, i);
2557 hctx = blk_mq_map_queue_type(q, j, i);
2558 ctx->hctxs[j] = hctx;
2560 * If the CPU is already set in the mask, then we've
2561 * mapped this one already. This can happen if
2562 * devices share queues across queue maps.
2564 if (cpumask_test_cpu(i, hctx->cpumask))
2567 cpumask_set_cpu(i, hctx->cpumask);
2569 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2570 hctx->ctxs[hctx->nr_ctx++] = ctx;
2573 * If the nr_ctx type overflows, we have exceeded the
2574 * amount of sw queues we can support.
2576 BUG_ON(!hctx->nr_ctx);
2579 for (; j < HCTX_MAX_TYPES; j++)
2580 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2581 HCTX_TYPE_DEFAULT, i);
2584 queue_for_each_hw_ctx(q, hctx, i) {
2586 * If no software queues are mapped to this hardware queue,
2587 * disable it and free the request entries.
2589 if (!hctx->nr_ctx) {
2590 /* Never unmap queue 0. We need it as a
2591 * fallback in case of a new remap fails
2594 if (i && set->tags[i])
2595 blk_mq_free_map_and_requests(set, i);
2601 hctx->tags = set->tags[i];
2602 WARN_ON(!hctx->tags);
2605 * Set the map size to the number of mapped software queues.
2606 * This is more accurate and more efficient than looping
2607 * over all possibly mapped software queues.
2609 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2612 * Initialize batch roundrobin counts
2614 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2615 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2620 * Caller needs to ensure that we're either frozen/quiesced, or that
2621 * the queue isn't live yet.
2623 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2625 struct blk_mq_hw_ctx *hctx;
2628 queue_for_each_hw_ctx(q, hctx, i) {
2630 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2632 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2636 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2639 struct request_queue *q;
2641 lockdep_assert_held(&set->tag_list_lock);
2643 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2644 blk_mq_freeze_queue(q);
2645 queue_set_hctx_shared(q, shared);
2646 blk_mq_unfreeze_queue(q);
2650 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2652 struct blk_mq_tag_set *set = q->tag_set;
2654 mutex_lock(&set->tag_list_lock);
2655 list_del_rcu(&q->tag_set_list);
2656 if (list_is_singular(&set->tag_list)) {
2657 /* just transitioned to unshared */
2658 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2659 /* update existing queue */
2660 blk_mq_update_tag_set_depth(set, false);
2662 mutex_unlock(&set->tag_list_lock);
2663 INIT_LIST_HEAD(&q->tag_set_list);
2666 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2667 struct request_queue *q)
2669 mutex_lock(&set->tag_list_lock);
2672 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2674 if (!list_empty(&set->tag_list) &&
2675 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2676 set->flags |= BLK_MQ_F_TAG_SHARED;
2677 /* update existing queue */
2678 blk_mq_update_tag_set_depth(set, true);
2680 if (set->flags & BLK_MQ_F_TAG_SHARED)
2681 queue_set_hctx_shared(q, true);
2682 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2684 mutex_unlock(&set->tag_list_lock);
2687 /* All allocations will be freed in release handler of q->mq_kobj */
2688 static int blk_mq_alloc_ctxs(struct request_queue *q)
2690 struct blk_mq_ctxs *ctxs;
2693 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2697 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2698 if (!ctxs->queue_ctx)
2701 for_each_possible_cpu(cpu) {
2702 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2706 q->mq_kobj = &ctxs->kobj;
2707 q->queue_ctx = ctxs->queue_ctx;
2716 * It is the actual release handler for mq, but we do it from
2717 * request queue's release handler for avoiding use-after-free
2718 * and headache because q->mq_kobj shouldn't have been introduced,
2719 * but we can't group ctx/kctx kobj without it.
2721 void blk_mq_release(struct request_queue *q)
2723 struct blk_mq_hw_ctx *hctx, *next;
2726 queue_for_each_hw_ctx(q, hctx, i)
2727 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2729 /* all hctx are in .unused_hctx_list now */
2730 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2731 list_del_init(&hctx->hctx_list);
2732 kobject_put(&hctx->kobj);
2735 kfree(q->queue_hw_ctx);
2738 * release .mq_kobj and sw queue's kobject now because
2739 * both share lifetime with request queue.
2741 blk_mq_sysfs_deinit(q);
2744 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
2747 struct request_queue *uninit_q, *q;
2749 uninit_q = __blk_alloc_queue(set->numa_node);
2751 return ERR_PTR(-ENOMEM);
2752 uninit_q->queuedata = queuedata;
2755 * Initialize the queue without an elevator. device_add_disk() will do
2756 * the initialization.
2758 q = blk_mq_init_allocated_queue(set, uninit_q, false);
2760 blk_cleanup_queue(uninit_q);
2764 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
2766 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2768 return blk_mq_init_queue_data(set, NULL);
2770 EXPORT_SYMBOL(blk_mq_init_queue);
2773 * Helper for setting up a queue with mq ops, given queue depth, and
2774 * the passed in mq ops flags.
2776 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2777 const struct blk_mq_ops *ops,
2778 unsigned int queue_depth,
2779 unsigned int set_flags)
2781 struct request_queue *q;
2784 memset(set, 0, sizeof(*set));
2786 set->nr_hw_queues = 1;
2788 set->queue_depth = queue_depth;
2789 set->numa_node = NUMA_NO_NODE;
2790 set->flags = set_flags;
2792 ret = blk_mq_alloc_tag_set(set);
2794 return ERR_PTR(ret);
2796 q = blk_mq_init_queue(set);
2798 blk_mq_free_tag_set(set);
2804 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2806 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2807 struct blk_mq_tag_set *set, struct request_queue *q,
2808 int hctx_idx, int node)
2810 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2812 /* reuse dead hctx first */
2813 spin_lock(&q->unused_hctx_lock);
2814 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2815 if (tmp->numa_node == node) {
2821 list_del_init(&hctx->hctx_list);
2822 spin_unlock(&q->unused_hctx_lock);
2825 hctx = blk_mq_alloc_hctx(q, set, node);
2829 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2835 kobject_put(&hctx->kobj);
2840 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2841 struct request_queue *q)
2844 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2846 if (q->nr_hw_queues < set->nr_hw_queues) {
2847 struct blk_mq_hw_ctx **new_hctxs;
2849 new_hctxs = kcalloc_node(set->nr_hw_queues,
2850 sizeof(*new_hctxs), GFP_KERNEL,
2855 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
2857 q->queue_hw_ctx = new_hctxs;
2862 /* protect against switching io scheduler */
2863 mutex_lock(&q->sysfs_lock);
2864 for (i = 0; i < set->nr_hw_queues; i++) {
2866 struct blk_mq_hw_ctx *hctx;
2868 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2870 * If the hw queue has been mapped to another numa node,
2871 * we need to realloc the hctx. If allocation fails, fallback
2872 * to use the previous one.
2874 if (hctxs[i] && (hctxs[i]->numa_node == node))
2877 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2880 blk_mq_exit_hctx(q, set, hctxs[i], i);
2884 pr_warn("Allocate new hctx on node %d fails,\
2885 fallback to previous one on node %d\n",
2886 node, hctxs[i]->numa_node);
2892 * Increasing nr_hw_queues fails. Free the newly allocated
2893 * hctxs and keep the previous q->nr_hw_queues.
2895 if (i != set->nr_hw_queues) {
2896 j = q->nr_hw_queues;
2900 end = q->nr_hw_queues;
2901 q->nr_hw_queues = set->nr_hw_queues;
2904 for (; j < end; j++) {
2905 struct blk_mq_hw_ctx *hctx = hctxs[j];
2909 blk_mq_free_map_and_requests(set, j);
2910 blk_mq_exit_hctx(q, set, hctx, j);
2914 mutex_unlock(&q->sysfs_lock);
2917 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2918 struct request_queue *q,
2921 /* mark the queue as mq asap */
2922 q->mq_ops = set->ops;
2924 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2925 blk_mq_poll_stats_bkt,
2926 BLK_MQ_POLL_STATS_BKTS, q);
2930 if (blk_mq_alloc_ctxs(q))
2933 /* init q->mq_kobj and sw queues' kobjects */
2934 blk_mq_sysfs_init(q);
2936 INIT_LIST_HEAD(&q->unused_hctx_list);
2937 spin_lock_init(&q->unused_hctx_lock);
2939 blk_mq_realloc_hw_ctxs(set, q);
2940 if (!q->nr_hw_queues)
2943 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2944 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2948 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2949 if (set->nr_maps > HCTX_TYPE_POLL &&
2950 set->map[HCTX_TYPE_POLL].nr_queues)
2951 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2953 q->sg_reserved_size = INT_MAX;
2955 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2956 INIT_LIST_HEAD(&q->requeue_list);
2957 spin_lock_init(&q->requeue_lock);
2959 q->nr_requests = set->queue_depth;
2962 * Default to classic polling
2964 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2966 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2967 blk_mq_add_queue_tag_set(set, q);
2968 blk_mq_map_swqueue(q);
2971 elevator_init_mq(q);
2976 kfree(q->queue_hw_ctx);
2977 q->nr_hw_queues = 0;
2978 blk_mq_sysfs_deinit(q);
2980 blk_stat_free_callback(q->poll_cb);
2984 return ERR_PTR(-ENOMEM);
2986 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2988 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2989 void blk_mq_exit_queue(struct request_queue *q)
2991 struct blk_mq_tag_set *set = q->tag_set;
2993 blk_mq_del_queue_tag_set(q);
2994 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2997 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3001 for (i = 0; i < set->nr_hw_queues; i++)
3002 if (!__blk_mq_alloc_rq_map(set, i))
3009 blk_mq_free_rq_map(set->tags[i]);
3015 * Allocate the request maps associated with this tag_set. Note that this
3016 * may reduce the depth asked for, if memory is tight. set->queue_depth
3017 * will be updated to reflect the allocated depth.
3019 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3024 depth = set->queue_depth;
3026 err = __blk_mq_alloc_rq_maps(set);
3030 set->queue_depth >>= 1;
3031 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3035 } while (set->queue_depth);
3037 if (!set->queue_depth || err) {
3038 pr_err("blk-mq: failed to allocate request map\n");
3042 if (depth != set->queue_depth)
3043 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3044 depth, set->queue_depth);
3049 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3052 * blk_mq_map_queues() and multiple .map_queues() implementations
3053 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3054 * number of hardware queues.
3056 if (set->nr_maps == 1)
3057 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3059 if (set->ops->map_queues && !is_kdump_kernel()) {
3063 * transport .map_queues is usually done in the following
3066 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3067 * mask = get_cpu_mask(queue)
3068 * for_each_cpu(cpu, mask)
3069 * set->map[x].mq_map[cpu] = queue;
3072 * When we need to remap, the table has to be cleared for
3073 * killing stale mapping since one CPU may not be mapped
3076 for (i = 0; i < set->nr_maps; i++)
3077 blk_mq_clear_mq_map(&set->map[i]);
3079 return set->ops->map_queues(set);
3081 BUG_ON(set->nr_maps > 1);
3082 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3086 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3087 int cur_nr_hw_queues, int new_nr_hw_queues)
3089 struct blk_mq_tags **new_tags;
3091 if (cur_nr_hw_queues >= new_nr_hw_queues)
3094 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3095 GFP_KERNEL, set->numa_node);
3100 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3101 sizeof(*set->tags));
3103 set->tags = new_tags;
3104 set->nr_hw_queues = new_nr_hw_queues;
3110 * Alloc a tag set to be associated with one or more request queues.
3111 * May fail with EINVAL for various error conditions. May adjust the
3112 * requested depth down, if it's too large. In that case, the set
3113 * value will be stored in set->queue_depth.
3115 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3119 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3121 if (!set->nr_hw_queues)
3123 if (!set->queue_depth)
3125 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3128 if (!set->ops->queue_rq)
3131 if (!set->ops->get_budget ^ !set->ops->put_budget)
3134 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3135 pr_info("blk-mq: reduced tag depth to %u\n",
3137 set->queue_depth = BLK_MQ_MAX_DEPTH;
3142 else if (set->nr_maps > HCTX_MAX_TYPES)
3146 * If a crashdump is active, then we are potentially in a very
3147 * memory constrained environment. Limit us to 1 queue and
3148 * 64 tags to prevent using too much memory.
3150 if (is_kdump_kernel()) {
3151 set->nr_hw_queues = 1;
3153 set->queue_depth = min(64U, set->queue_depth);
3156 * There is no use for more h/w queues than cpus if we just have
3159 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3160 set->nr_hw_queues = nr_cpu_ids;
3162 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3166 for (i = 0; i < set->nr_maps; i++) {
3167 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3168 sizeof(set->map[i].mq_map[0]),
3169 GFP_KERNEL, set->numa_node);
3170 if (!set->map[i].mq_map)
3171 goto out_free_mq_map;
3172 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3175 ret = blk_mq_update_queue_map(set);
3177 goto out_free_mq_map;
3179 ret = blk_mq_alloc_rq_maps(set);
3181 goto out_free_mq_map;
3183 mutex_init(&set->tag_list_lock);
3184 INIT_LIST_HEAD(&set->tag_list);
3189 for (i = 0; i < set->nr_maps; i++) {
3190 kfree(set->map[i].mq_map);
3191 set->map[i].mq_map = NULL;
3197 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3199 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3203 for (i = 0; i < set->nr_hw_queues; i++)
3204 blk_mq_free_map_and_requests(set, i);
3206 for (j = 0; j < set->nr_maps; j++) {
3207 kfree(set->map[j].mq_map);
3208 set->map[j].mq_map = NULL;
3214 EXPORT_SYMBOL(blk_mq_free_tag_set);
3216 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3218 struct blk_mq_tag_set *set = q->tag_set;
3219 struct blk_mq_hw_ctx *hctx;
3225 if (q->nr_requests == nr)
3228 blk_mq_freeze_queue(q);
3229 blk_mq_quiesce_queue(q);
3232 queue_for_each_hw_ctx(q, hctx, i) {
3236 * If we're using an MQ scheduler, just update the scheduler
3237 * queue depth. This is similar to what the old code would do.
3239 if (!hctx->sched_tags) {
3240 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3243 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3248 if (q->elevator && q->elevator->type->ops.depth_updated)
3249 q->elevator->type->ops.depth_updated(hctx);
3253 q->nr_requests = nr;
3255 blk_mq_unquiesce_queue(q);
3256 blk_mq_unfreeze_queue(q);
3262 * request_queue and elevator_type pair.
3263 * It is just used by __blk_mq_update_nr_hw_queues to cache
3264 * the elevator_type associated with a request_queue.
3266 struct blk_mq_qe_pair {
3267 struct list_head node;
3268 struct request_queue *q;
3269 struct elevator_type *type;
3273 * Cache the elevator_type in qe pair list and switch the
3274 * io scheduler to 'none'
3276 static bool blk_mq_elv_switch_none(struct list_head *head,
3277 struct request_queue *q)
3279 struct blk_mq_qe_pair *qe;
3284 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3288 INIT_LIST_HEAD(&qe->node);
3290 qe->type = q->elevator->type;
3291 list_add(&qe->node, head);
3293 mutex_lock(&q->sysfs_lock);
3295 * After elevator_switch_mq, the previous elevator_queue will be
3296 * released by elevator_release. The reference of the io scheduler
3297 * module get by elevator_get will also be put. So we need to get
3298 * a reference of the io scheduler module here to prevent it to be
3301 __module_get(qe->type->elevator_owner);
3302 elevator_switch_mq(q, NULL);
3303 mutex_unlock(&q->sysfs_lock);
3308 static void blk_mq_elv_switch_back(struct list_head *head,
3309 struct request_queue *q)
3311 struct blk_mq_qe_pair *qe;
3312 struct elevator_type *t = NULL;
3314 list_for_each_entry(qe, head, node)
3323 list_del(&qe->node);
3326 mutex_lock(&q->sysfs_lock);
3327 elevator_switch_mq(q, t);
3328 mutex_unlock(&q->sysfs_lock);
3331 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3334 struct request_queue *q;
3336 int prev_nr_hw_queues;
3338 lockdep_assert_held(&set->tag_list_lock);
3340 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3341 nr_hw_queues = nr_cpu_ids;
3342 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3345 list_for_each_entry(q, &set->tag_list, tag_set_list)
3346 blk_mq_freeze_queue(q);
3348 * Switch IO scheduler to 'none', cleaning up the data associated
3349 * with the previous scheduler. We will switch back once we are done
3350 * updating the new sw to hw queue mappings.
3352 list_for_each_entry(q, &set->tag_list, tag_set_list)
3353 if (!blk_mq_elv_switch_none(&head, q))
3356 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3357 blk_mq_debugfs_unregister_hctxs(q);
3358 blk_mq_sysfs_unregister(q);
3361 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3365 prev_nr_hw_queues = set->nr_hw_queues;
3366 set->nr_hw_queues = nr_hw_queues;
3367 blk_mq_update_queue_map(set);
3369 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3370 blk_mq_realloc_hw_ctxs(set, q);
3371 if (q->nr_hw_queues != set->nr_hw_queues) {
3372 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3373 nr_hw_queues, prev_nr_hw_queues);
3374 set->nr_hw_queues = prev_nr_hw_queues;
3375 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3378 blk_mq_map_swqueue(q);
3382 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3383 blk_mq_sysfs_register(q);
3384 blk_mq_debugfs_register_hctxs(q);
3388 list_for_each_entry(q, &set->tag_list, tag_set_list)
3389 blk_mq_elv_switch_back(&head, q);
3391 list_for_each_entry(q, &set->tag_list, tag_set_list)
3392 blk_mq_unfreeze_queue(q);
3395 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3397 mutex_lock(&set->tag_list_lock);
3398 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3399 mutex_unlock(&set->tag_list_lock);
3401 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3403 /* Enable polling stats and return whether they were already enabled. */
3404 static bool blk_poll_stats_enable(struct request_queue *q)
3406 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3407 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3409 blk_stat_add_callback(q, q->poll_cb);
3413 static void blk_mq_poll_stats_start(struct request_queue *q)
3416 * We don't arm the callback if polling stats are not enabled or the
3417 * callback is already active.
3419 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3420 blk_stat_is_active(q->poll_cb))
3423 blk_stat_activate_msecs(q->poll_cb, 100);
3426 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3428 struct request_queue *q = cb->data;
3431 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3432 if (cb->stat[bucket].nr_samples)
3433 q->poll_stat[bucket] = cb->stat[bucket];
3437 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3440 unsigned long ret = 0;
3444 * If stats collection isn't on, don't sleep but turn it on for
3447 if (!blk_poll_stats_enable(q))
3451 * As an optimistic guess, use half of the mean service time
3452 * for this type of request. We can (and should) make this smarter.
3453 * For instance, if the completion latencies are tight, we can
3454 * get closer than just half the mean. This is especially
3455 * important on devices where the completion latencies are longer
3456 * than ~10 usec. We do use the stats for the relevant IO size
3457 * if available which does lead to better estimates.
3459 bucket = blk_mq_poll_stats_bkt(rq);
3463 if (q->poll_stat[bucket].nr_samples)
3464 ret = (q->poll_stat[bucket].mean + 1) / 2;
3469 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3472 struct hrtimer_sleeper hs;
3473 enum hrtimer_mode mode;
3477 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3481 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3483 * 0: use half of prev avg
3484 * >0: use this specific value
3486 if (q->poll_nsec > 0)
3487 nsecs = q->poll_nsec;
3489 nsecs = blk_mq_poll_nsecs(q, rq);
3494 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3497 * This will be replaced with the stats tracking code, using
3498 * 'avg_completion_time / 2' as the pre-sleep target.
3502 mode = HRTIMER_MODE_REL;
3503 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3504 hrtimer_set_expires(&hs.timer, kt);
3507 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3509 set_current_state(TASK_UNINTERRUPTIBLE);
3510 hrtimer_sleeper_start_expires(&hs, mode);
3513 hrtimer_cancel(&hs.timer);
3514 mode = HRTIMER_MODE_ABS;
3515 } while (hs.task && !signal_pending(current));
3517 __set_current_state(TASK_RUNNING);
3518 destroy_hrtimer_on_stack(&hs.timer);
3522 static bool blk_mq_poll_hybrid(struct request_queue *q,
3523 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3527 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3530 if (!blk_qc_t_is_internal(cookie))
3531 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3533 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3535 * With scheduling, if the request has completed, we'll
3536 * get a NULL return here, as we clear the sched tag when
3537 * that happens. The request still remains valid, like always,
3538 * so we should be safe with just the NULL check.
3544 return blk_mq_poll_hybrid_sleep(q, rq);
3548 * blk_poll - poll for IO completions
3550 * @cookie: cookie passed back at IO submission time
3551 * @spin: whether to spin for completions
3554 * Poll for completions on the passed in queue. Returns number of
3555 * completed entries found. If @spin is true, then blk_poll will continue
3556 * looping until at least one completion is found, unless the task is
3557 * otherwise marked running (or we need to reschedule).
3559 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3561 struct blk_mq_hw_ctx *hctx;
3564 if (!blk_qc_t_valid(cookie) ||
3565 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3569 blk_flush_plug_list(current->plug, false);
3571 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3574 * If we sleep, have the caller restart the poll loop to reset
3575 * the state. Like for the other success return cases, the
3576 * caller is responsible for checking if the IO completed. If
3577 * the IO isn't complete, we'll get called again and will go
3578 * straight to the busy poll loop.
3580 if (blk_mq_poll_hybrid(q, hctx, cookie))
3583 hctx->poll_considered++;
3585 state = current->state;
3589 hctx->poll_invoked++;
3591 ret = q->mq_ops->poll(hctx);
3593 hctx->poll_success++;
3594 __set_current_state(TASK_RUNNING);
3598 if (signal_pending_state(state, current))
3599 __set_current_state(TASK_RUNNING);
3601 if (current->state == TASK_RUNNING)
3603 if (ret < 0 || !spin)
3606 } while (!need_resched());
3608 __set_current_state(TASK_RUNNING);
3611 EXPORT_SYMBOL_GPL(blk_poll);
3613 unsigned int blk_mq_rq_cpu(struct request *rq)
3615 return rq->mq_ctx->cpu;
3617 EXPORT_SYMBOL(blk_mq_rq_cpu);
3619 static int __init blk_mq_init(void)
3621 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3622 blk_mq_hctx_notify_dead);
3625 subsys_initcall(blk_mq_init);