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>
29 #include <linux/blk-crypto.h>
31 #include <trace/events/block.h>
33 #include <linux/blk-mq.h>
34 #include <linux/t10-pi.h>
37 #include "blk-mq-debugfs.h"
38 #include "blk-mq-tag.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
44 static void blk_mq_poll_stats_start(struct request_queue *q);
45 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
47 static int blk_mq_poll_stats_bkt(const struct request *rq)
49 int ddir, sectors, bucket;
51 ddir = rq_data_dir(rq);
52 sectors = blk_rq_stats_sectors(rq);
54 bucket = ddir + 2 * ilog2(sectors);
58 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
59 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
65 * Check if any of the ctx, dispatch list or elevator
66 * have pending work in this hardware queue.
68 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
70 return !list_empty_careful(&hctx->dispatch) ||
71 sbitmap_any_bit_set(&hctx->ctx_map) ||
72 blk_mq_sched_has_work(hctx);
76 * Mark this ctx as having pending work in this hardware queue
78 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
79 struct blk_mq_ctx *ctx)
81 const int bit = ctx->index_hw[hctx->type];
83 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
84 sbitmap_set_bit(&hctx->ctx_map, bit);
87 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
88 struct blk_mq_ctx *ctx)
90 const int bit = ctx->index_hw[hctx->type];
92 sbitmap_clear_bit(&hctx->ctx_map, bit);
96 struct hd_struct *part;
97 unsigned int inflight[2];
100 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
101 struct request *rq, void *priv,
104 struct mq_inflight *mi = priv;
106 if (rq->part == mi->part)
107 mi->inflight[rq_data_dir(rq)]++;
112 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
114 struct mq_inflight mi = { .part = part };
116 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
118 return mi.inflight[0] + mi.inflight[1];
121 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
122 unsigned int inflight[2])
124 struct mq_inflight mi = { .part = part };
126 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
127 inflight[0] = mi.inflight[0];
128 inflight[1] = mi.inflight[1];
131 void blk_freeze_queue_start(struct request_queue *q)
133 mutex_lock(&q->mq_freeze_lock);
134 if (++q->mq_freeze_depth == 1) {
135 percpu_ref_kill(&q->q_usage_counter);
136 mutex_unlock(&q->mq_freeze_lock);
138 blk_mq_run_hw_queues(q, false);
140 mutex_unlock(&q->mq_freeze_lock);
143 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
145 void blk_mq_freeze_queue_wait(struct request_queue *q)
147 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
149 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
151 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
152 unsigned long timeout)
154 return wait_event_timeout(q->mq_freeze_wq,
155 percpu_ref_is_zero(&q->q_usage_counter),
158 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
161 * Guarantee no request is in use, so we can change any data structure of
162 * the queue afterward.
164 void blk_freeze_queue(struct request_queue *q)
167 * In the !blk_mq case we are only calling this to kill the
168 * q_usage_counter, otherwise this increases the freeze depth
169 * and waits for it to return to zero. For this reason there is
170 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
171 * exported to drivers as the only user for unfreeze is blk_mq.
173 blk_freeze_queue_start(q);
174 blk_mq_freeze_queue_wait(q);
177 void blk_mq_freeze_queue(struct request_queue *q)
180 * ...just an alias to keep freeze and unfreeze actions balanced
181 * in the blk_mq_* namespace
185 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
187 void blk_mq_unfreeze_queue(struct request_queue *q)
189 mutex_lock(&q->mq_freeze_lock);
190 q->mq_freeze_depth--;
191 WARN_ON_ONCE(q->mq_freeze_depth < 0);
192 if (!q->mq_freeze_depth) {
193 percpu_ref_resurrect(&q->q_usage_counter);
194 wake_up_all(&q->mq_freeze_wq);
196 mutex_unlock(&q->mq_freeze_lock);
198 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
201 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
202 * mpt3sas driver such that this function can be removed.
204 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
206 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
208 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
211 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
214 * Note: this function does not prevent that the struct request end_io()
215 * callback function is invoked. Once this function is returned, we make
216 * sure no dispatch can happen until the queue is unquiesced via
217 * blk_mq_unquiesce_queue().
219 void blk_mq_quiesce_queue(struct request_queue *q)
221 struct blk_mq_hw_ctx *hctx;
225 blk_mq_quiesce_queue_nowait(q);
227 queue_for_each_hw_ctx(q, hctx, i) {
228 if (hctx->flags & BLK_MQ_F_BLOCKING)
229 synchronize_srcu(hctx->srcu);
236 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
239 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
242 * This function recovers queue into the state before quiescing
243 * which is done by blk_mq_quiesce_queue.
245 void blk_mq_unquiesce_queue(struct request_queue *q)
247 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
249 /* dispatch requests which are inserted during quiescing */
250 blk_mq_run_hw_queues(q, true);
252 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
254 void blk_mq_wake_waiters(struct request_queue *q)
256 struct blk_mq_hw_ctx *hctx;
259 queue_for_each_hw_ctx(q, hctx, i)
260 if (blk_mq_hw_queue_mapped(hctx))
261 blk_mq_tag_wakeup_all(hctx->tags, true);
265 * Only need start/end time stamping if we have iostat or
266 * blk stats enabled, or using an IO scheduler.
268 static inline bool blk_mq_need_time_stamp(struct request *rq)
270 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
273 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
274 unsigned int tag, u64 alloc_time_ns)
276 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
277 struct request *rq = tags->static_rqs[tag];
278 req_flags_t rq_flags = 0;
280 if (data->flags & BLK_MQ_REQ_INTERNAL) {
281 rq->tag = BLK_MQ_NO_TAG;
282 rq->internal_tag = tag;
284 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
285 rq_flags = RQF_MQ_INFLIGHT;
286 atomic_inc(&data->hctx->nr_active);
289 rq->internal_tag = BLK_MQ_NO_TAG;
290 data->hctx->tags->rqs[rq->tag] = rq;
293 /* csd/requeue_work/fifo_time is initialized before use */
295 rq->mq_ctx = data->ctx;
296 rq->mq_hctx = data->hctx;
297 rq->rq_flags = rq_flags;
298 rq->cmd_flags = data->cmd_flags;
299 if (data->flags & BLK_MQ_REQ_PREEMPT)
300 rq->rq_flags |= RQF_PREEMPT;
301 if (blk_queue_io_stat(data->q))
302 rq->rq_flags |= RQF_IO_STAT;
303 INIT_LIST_HEAD(&rq->queuelist);
304 INIT_HLIST_NODE(&rq->hash);
305 RB_CLEAR_NODE(&rq->rb_node);
308 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
309 rq->alloc_time_ns = alloc_time_ns;
311 if (blk_mq_need_time_stamp(rq))
312 rq->start_time_ns = ktime_get_ns();
314 rq->start_time_ns = 0;
315 rq->io_start_time_ns = 0;
316 rq->stats_sectors = 0;
317 rq->nr_phys_segments = 0;
318 #if defined(CONFIG_BLK_DEV_INTEGRITY)
319 rq->nr_integrity_segments = 0;
321 blk_crypto_rq_set_defaults(rq);
322 /* tag was already set */
323 WRITE_ONCE(rq->deadline, 0);
328 rq->end_io_data = NULL;
330 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
331 refcount_set(&rq->ref, 1);
333 if (!op_is_flush(data->cmd_flags)) {
334 struct elevator_queue *e = data->q->elevator;
337 if (e && e->type->ops.prepare_request) {
338 if (e->type->icq_cache)
339 blk_mq_sched_assign_ioc(rq);
341 e->type->ops.prepare_request(rq);
342 rq->rq_flags |= RQF_ELVPRIV;
346 data->hctx->queued++;
350 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
352 struct request_queue *q = data->q;
353 struct elevator_queue *e = q->elevator;
354 u64 alloc_time_ns = 0;
357 /* alloc_time includes depth and tag waits */
358 if (blk_queue_rq_alloc_time(q))
359 alloc_time_ns = ktime_get_ns();
361 if (data->cmd_flags & REQ_NOWAIT)
362 data->flags |= BLK_MQ_REQ_NOWAIT;
365 data->flags |= BLK_MQ_REQ_INTERNAL;
368 * Flush requests are special and go directly to the
369 * dispatch list. Don't include reserved tags in the
370 * limiting, as it isn't useful.
372 if (!op_is_flush(data->cmd_flags) &&
373 e->type->ops.limit_depth &&
374 !(data->flags & BLK_MQ_REQ_RESERVED))
375 e->type->ops.limit_depth(data->cmd_flags, data);
379 data->ctx = blk_mq_get_ctx(q);
380 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
381 if (!(data->flags & BLK_MQ_REQ_INTERNAL))
382 blk_mq_tag_busy(data->hctx);
385 * Waiting allocations only fail because of an inactive hctx. In that
386 * case just retry the hctx assignment and tag allocation as CPU hotplug
387 * should have migrated us to an online CPU by now.
389 tag = blk_mq_get_tag(data);
390 if (tag == BLK_MQ_NO_TAG) {
391 if (data->flags & BLK_MQ_REQ_NOWAIT)
395 * Give up the CPU and sleep for a random short time to ensure
396 * that thread using a realtime scheduling class are migrated
397 * off the the CPU, and thus off the hctx that is going away.
402 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
405 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
406 blk_mq_req_flags_t flags)
408 struct blk_mq_alloc_data data = {
416 ret = blk_queue_enter(q, flags);
420 rq = __blk_mq_alloc_request(&data);
424 rq->__sector = (sector_t) -1;
425 rq->bio = rq->biotail = NULL;
429 return ERR_PTR(-EWOULDBLOCK);
431 EXPORT_SYMBOL(blk_mq_alloc_request);
433 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
434 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
436 struct blk_mq_alloc_data data = {
441 u64 alloc_time_ns = 0;
446 /* alloc_time includes depth and tag waits */
447 if (blk_queue_rq_alloc_time(q))
448 alloc_time_ns = ktime_get_ns();
451 * If the tag allocator sleeps we could get an allocation for a
452 * different hardware context. No need to complicate the low level
453 * allocator for this for the rare use case of a command tied to
456 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
457 return ERR_PTR(-EINVAL);
459 if (hctx_idx >= q->nr_hw_queues)
460 return ERR_PTR(-EIO);
462 ret = blk_queue_enter(q, flags);
467 * Check if the hardware context is actually mapped to anything.
468 * If not tell the caller that it should skip this queue.
471 data.hctx = q->queue_hw_ctx[hctx_idx];
472 if (!blk_mq_hw_queue_mapped(data.hctx))
474 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
475 data.ctx = __blk_mq_get_ctx(q, cpu);
478 data.flags |= BLK_MQ_REQ_INTERNAL;
480 blk_mq_tag_busy(data.hctx);
483 tag = blk_mq_get_tag(&data);
484 if (tag == BLK_MQ_NO_TAG)
486 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
492 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
494 static void __blk_mq_free_request(struct request *rq)
496 struct request_queue *q = rq->q;
497 struct blk_mq_ctx *ctx = rq->mq_ctx;
498 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
499 const int sched_tag = rq->internal_tag;
501 blk_crypto_free_request(rq);
502 blk_pm_mark_last_busy(rq);
504 if (rq->tag != BLK_MQ_NO_TAG)
505 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
506 if (sched_tag != BLK_MQ_NO_TAG)
507 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
508 blk_mq_sched_restart(hctx);
512 void blk_mq_free_request(struct request *rq)
514 struct request_queue *q = rq->q;
515 struct elevator_queue *e = q->elevator;
516 struct blk_mq_ctx *ctx = rq->mq_ctx;
517 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
519 if (rq->rq_flags & RQF_ELVPRIV) {
520 if (e && e->type->ops.finish_request)
521 e->type->ops.finish_request(rq);
523 put_io_context(rq->elv.icq->ioc);
528 ctx->rq_completed[rq_is_sync(rq)]++;
529 if (rq->rq_flags & RQF_MQ_INFLIGHT)
530 atomic_dec(&hctx->nr_active);
532 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
533 laptop_io_completion(q->backing_dev_info);
537 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
538 if (refcount_dec_and_test(&rq->ref))
539 __blk_mq_free_request(rq);
541 EXPORT_SYMBOL_GPL(blk_mq_free_request);
543 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
547 if (blk_mq_need_time_stamp(rq))
548 now = ktime_get_ns();
550 if (rq->rq_flags & RQF_STATS) {
551 blk_mq_poll_stats_start(rq->q);
552 blk_stat_add(rq, now);
555 if (rq->internal_tag != BLK_MQ_NO_TAG)
556 blk_mq_sched_completed_request(rq, now);
558 blk_account_io_done(rq, now);
561 rq_qos_done(rq->q, rq);
562 rq->end_io(rq, error);
564 blk_mq_free_request(rq);
567 EXPORT_SYMBOL(__blk_mq_end_request);
569 void blk_mq_end_request(struct request *rq, blk_status_t error)
571 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
573 __blk_mq_end_request(rq, error);
575 EXPORT_SYMBOL(blk_mq_end_request);
577 static void __blk_mq_complete_request_remote(void *data)
579 struct request *rq = data;
580 struct request_queue *q = rq->q;
582 q->mq_ops->complete(rq);
586 * blk_mq_force_complete_rq() - Force complete the request, bypassing any error
587 * injection that could drop the completion.
588 * @rq: Request to be force completed
590 * Drivers should use blk_mq_complete_request() to complete requests in their
591 * normal IO path. For timeout error recovery, drivers may call this forced
592 * completion routine after they've reclaimed timed out requests to bypass
593 * potentially subsequent fake timeouts.
595 void blk_mq_force_complete_rq(struct request *rq)
597 struct blk_mq_ctx *ctx = rq->mq_ctx;
598 struct request_queue *q = rq->q;
602 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
604 * Most of single queue controllers, there is only one irq vector
605 * for handling IO completion, and the only irq's affinity is set
606 * as all possible CPUs. On most of ARCHs, this affinity means the
607 * irq is handled on one specific CPU.
609 * So complete IO reqeust in softirq context in case of single queue
610 * for not degrading IO performance by irqsoff latency.
612 if (q->nr_hw_queues == 1) {
613 __blk_complete_request(rq);
618 * For a polled request, always complete locallly, it's pointless
619 * to redirect the completion.
621 if ((rq->cmd_flags & REQ_HIPRI) ||
622 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
623 q->mq_ops->complete(rq);
628 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
629 shared = cpus_share_cache(cpu, ctx->cpu);
631 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
632 rq->csd.func = __blk_mq_complete_request_remote;
635 smp_call_function_single_async(ctx->cpu, &rq->csd);
637 q->mq_ops->complete(rq);
641 EXPORT_SYMBOL_GPL(blk_mq_force_complete_rq);
643 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
644 __releases(hctx->srcu)
646 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
649 srcu_read_unlock(hctx->srcu, srcu_idx);
652 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
653 __acquires(hctx->srcu)
655 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
656 /* shut up gcc false positive */
660 *srcu_idx = srcu_read_lock(hctx->srcu);
664 * blk_mq_complete_request - end I/O on a request
665 * @rq: the request being processed
668 * Ends all I/O on a request. It does not handle partial completions.
669 * The actual completion happens out-of-order, through a IPI handler.
671 bool blk_mq_complete_request(struct request *rq)
673 if (unlikely(blk_should_fake_timeout(rq->q)))
675 blk_mq_force_complete_rq(rq);
678 EXPORT_SYMBOL(blk_mq_complete_request);
681 * blk_mq_start_request - Start processing a request
682 * @rq: Pointer to request to be started
684 * Function used by device drivers to notify the block layer that a request
685 * is going to be processed now, so blk layer can do proper initializations
686 * such as starting the timeout timer.
688 void blk_mq_start_request(struct request *rq)
690 struct request_queue *q = rq->q;
692 trace_block_rq_issue(q, rq);
694 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
695 rq->io_start_time_ns = ktime_get_ns();
696 rq->stats_sectors = blk_rq_sectors(rq);
697 rq->rq_flags |= RQF_STATS;
701 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
704 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
706 #ifdef CONFIG_BLK_DEV_INTEGRITY
707 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
708 q->integrity.profile->prepare_fn(rq);
711 EXPORT_SYMBOL(blk_mq_start_request);
713 static void __blk_mq_requeue_request(struct request *rq)
715 struct request_queue *q = rq->q;
717 blk_mq_put_driver_tag(rq);
719 trace_block_rq_requeue(q, rq);
720 rq_qos_requeue(q, rq);
722 if (blk_mq_request_started(rq)) {
723 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
724 rq->rq_flags &= ~RQF_TIMED_OUT;
728 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
730 __blk_mq_requeue_request(rq);
732 /* this request will be re-inserted to io scheduler queue */
733 blk_mq_sched_requeue_request(rq);
735 BUG_ON(!list_empty(&rq->queuelist));
736 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
738 EXPORT_SYMBOL(blk_mq_requeue_request);
740 static void blk_mq_requeue_work(struct work_struct *work)
742 struct request_queue *q =
743 container_of(work, struct request_queue, requeue_work.work);
745 struct request *rq, *next;
747 spin_lock_irq(&q->requeue_lock);
748 list_splice_init(&q->requeue_list, &rq_list);
749 spin_unlock_irq(&q->requeue_lock);
751 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
752 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
755 rq->rq_flags &= ~RQF_SOFTBARRIER;
756 list_del_init(&rq->queuelist);
758 * If RQF_DONTPREP, rq has contained some driver specific
759 * data, so insert it to hctx dispatch list to avoid any
762 if (rq->rq_flags & RQF_DONTPREP)
763 blk_mq_request_bypass_insert(rq, false, false);
765 blk_mq_sched_insert_request(rq, true, false, false);
768 while (!list_empty(&rq_list)) {
769 rq = list_entry(rq_list.next, struct request, queuelist);
770 list_del_init(&rq->queuelist);
771 blk_mq_sched_insert_request(rq, false, false, false);
774 blk_mq_run_hw_queues(q, false);
777 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
778 bool kick_requeue_list)
780 struct request_queue *q = rq->q;
784 * We abuse this flag that is otherwise used by the I/O scheduler to
785 * request head insertion from the workqueue.
787 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
789 spin_lock_irqsave(&q->requeue_lock, flags);
791 rq->rq_flags |= RQF_SOFTBARRIER;
792 list_add(&rq->queuelist, &q->requeue_list);
794 list_add_tail(&rq->queuelist, &q->requeue_list);
796 spin_unlock_irqrestore(&q->requeue_lock, flags);
798 if (kick_requeue_list)
799 blk_mq_kick_requeue_list(q);
802 void blk_mq_kick_requeue_list(struct request_queue *q)
804 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
806 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
808 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
811 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
812 msecs_to_jiffies(msecs));
814 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
816 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
818 if (tag < tags->nr_tags) {
819 prefetch(tags->rqs[tag]);
820 return tags->rqs[tag];
825 EXPORT_SYMBOL(blk_mq_tag_to_rq);
827 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
828 void *priv, bool reserved)
831 * If we find a request that is inflight and the queue matches,
832 * we know the queue is busy. Return false to stop the iteration.
834 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
844 bool blk_mq_queue_inflight(struct request_queue *q)
848 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
851 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
853 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
855 req->rq_flags |= RQF_TIMED_OUT;
856 if (req->q->mq_ops->timeout) {
857 enum blk_eh_timer_return ret;
859 ret = req->q->mq_ops->timeout(req, reserved);
860 if (ret == BLK_EH_DONE)
862 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
868 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
870 unsigned long deadline;
872 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
874 if (rq->rq_flags & RQF_TIMED_OUT)
877 deadline = READ_ONCE(rq->deadline);
878 if (time_after_eq(jiffies, deadline))
883 else if (time_after(*next, deadline))
888 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
889 struct request *rq, void *priv, bool reserved)
891 unsigned long *next = priv;
894 * Just do a quick check if it is expired before locking the request in
895 * so we're not unnecessarilly synchronizing across CPUs.
897 if (!blk_mq_req_expired(rq, next))
901 * We have reason to believe the request may be expired. Take a
902 * reference on the request to lock this request lifetime into its
903 * currently allocated context to prevent it from being reallocated in
904 * the event the completion by-passes this timeout handler.
906 * If the reference was already released, then the driver beat the
907 * timeout handler to posting a natural completion.
909 if (!refcount_inc_not_zero(&rq->ref))
913 * The request is now locked and cannot be reallocated underneath the
914 * timeout handler's processing. Re-verify this exact request is truly
915 * expired; if it is not expired, then the request was completed and
916 * reallocated as a new request.
918 if (blk_mq_req_expired(rq, next))
919 blk_mq_rq_timed_out(rq, reserved);
921 if (is_flush_rq(rq, hctx))
923 else if (refcount_dec_and_test(&rq->ref))
924 __blk_mq_free_request(rq);
929 static void blk_mq_timeout_work(struct work_struct *work)
931 struct request_queue *q =
932 container_of(work, struct request_queue, timeout_work);
933 unsigned long next = 0;
934 struct blk_mq_hw_ctx *hctx;
937 /* A deadlock might occur if a request is stuck requiring a
938 * timeout at the same time a queue freeze is waiting
939 * completion, since the timeout code would not be able to
940 * acquire the queue reference here.
942 * That's why we don't use blk_queue_enter here; instead, we use
943 * percpu_ref_tryget directly, because we need to be able to
944 * obtain a reference even in the short window between the queue
945 * starting to freeze, by dropping the first reference in
946 * blk_freeze_queue_start, and the moment the last request is
947 * consumed, marked by the instant q_usage_counter reaches
950 if (!percpu_ref_tryget(&q->q_usage_counter))
953 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
956 mod_timer(&q->timeout, next);
959 * Request timeouts are handled as a forward rolling timer. If
960 * we end up here it means that no requests are pending and
961 * also that no request has been pending for a while. Mark
964 queue_for_each_hw_ctx(q, hctx, i) {
965 /* the hctx may be unmapped, so check it here */
966 if (blk_mq_hw_queue_mapped(hctx))
967 blk_mq_tag_idle(hctx);
973 struct flush_busy_ctx_data {
974 struct blk_mq_hw_ctx *hctx;
975 struct list_head *list;
978 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
980 struct flush_busy_ctx_data *flush_data = data;
981 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
982 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
983 enum hctx_type type = hctx->type;
985 spin_lock(&ctx->lock);
986 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
987 sbitmap_clear_bit(sb, bitnr);
988 spin_unlock(&ctx->lock);
993 * Process software queues that have been marked busy, splicing them
994 * to the for-dispatch
996 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
998 struct flush_busy_ctx_data data = {
1003 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1005 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1007 struct dispatch_rq_data {
1008 struct blk_mq_hw_ctx *hctx;
1012 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1015 struct dispatch_rq_data *dispatch_data = data;
1016 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1017 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1018 enum hctx_type type = hctx->type;
1020 spin_lock(&ctx->lock);
1021 if (!list_empty(&ctx->rq_lists[type])) {
1022 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1023 list_del_init(&dispatch_data->rq->queuelist);
1024 if (list_empty(&ctx->rq_lists[type]))
1025 sbitmap_clear_bit(sb, bitnr);
1027 spin_unlock(&ctx->lock);
1029 return !dispatch_data->rq;
1032 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1033 struct blk_mq_ctx *start)
1035 unsigned off = start ? start->index_hw[hctx->type] : 0;
1036 struct dispatch_rq_data data = {
1041 __sbitmap_for_each_set(&hctx->ctx_map, off,
1042 dispatch_rq_from_ctx, &data);
1047 static inline unsigned int queued_to_index(unsigned int queued)
1052 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1055 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1056 int flags, void *key)
1058 struct blk_mq_hw_ctx *hctx;
1060 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1062 spin_lock(&hctx->dispatch_wait_lock);
1063 if (!list_empty(&wait->entry)) {
1064 struct sbitmap_queue *sbq;
1066 list_del_init(&wait->entry);
1067 sbq = &hctx->tags->bitmap_tags;
1068 atomic_dec(&sbq->ws_active);
1070 spin_unlock(&hctx->dispatch_wait_lock);
1072 blk_mq_run_hw_queue(hctx, true);
1077 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1078 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1079 * restart. For both cases, take care to check the condition again after
1080 * marking us as waiting.
1082 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1085 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1086 struct wait_queue_head *wq;
1087 wait_queue_entry_t *wait;
1090 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1091 blk_mq_sched_mark_restart_hctx(hctx);
1094 * It's possible that a tag was freed in the window between the
1095 * allocation failure and adding the hardware queue to the wait
1098 * Don't clear RESTART here, someone else could have set it.
1099 * At most this will cost an extra queue run.
1101 return blk_mq_get_driver_tag(rq);
1104 wait = &hctx->dispatch_wait;
1105 if (!list_empty_careful(&wait->entry))
1108 wq = &bt_wait_ptr(sbq, hctx)->wait;
1110 spin_lock_irq(&wq->lock);
1111 spin_lock(&hctx->dispatch_wait_lock);
1112 if (!list_empty(&wait->entry)) {
1113 spin_unlock(&hctx->dispatch_wait_lock);
1114 spin_unlock_irq(&wq->lock);
1118 atomic_inc(&sbq->ws_active);
1119 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1120 __add_wait_queue(wq, wait);
1123 * It's possible that a tag was freed in the window between the
1124 * allocation failure and adding the hardware queue to the wait
1127 ret = blk_mq_get_driver_tag(rq);
1129 spin_unlock(&hctx->dispatch_wait_lock);
1130 spin_unlock_irq(&wq->lock);
1135 * We got a tag, remove ourselves from the wait queue to ensure
1136 * someone else gets the wakeup.
1138 list_del_init(&wait->entry);
1139 atomic_dec(&sbq->ws_active);
1140 spin_unlock(&hctx->dispatch_wait_lock);
1141 spin_unlock_irq(&wq->lock);
1146 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1147 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1149 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1150 * - EWMA is one simple way to compute running average value
1151 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1152 * - take 4 as factor for avoiding to get too small(0) result, and this
1153 * factor doesn't matter because EWMA decreases exponentially
1155 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1159 if (hctx->queue->elevator)
1162 ewma = hctx->dispatch_busy;
1167 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1169 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1170 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1172 hctx->dispatch_busy = ewma;
1175 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1177 static void blk_mq_handle_dev_resource(struct request *rq,
1178 struct list_head *list)
1180 struct request *next =
1181 list_first_entry_or_null(list, struct request, queuelist);
1184 * If an I/O scheduler has been configured and we got a driver tag for
1185 * the next request already, free it.
1188 blk_mq_put_driver_tag(next);
1190 list_add(&rq->queuelist, list);
1191 __blk_mq_requeue_request(rq);
1194 static void blk_mq_handle_zone_resource(struct request *rq,
1195 struct list_head *zone_list)
1198 * If we end up here it is because we cannot dispatch a request to a
1199 * specific zone due to LLD level zone-write locking or other zone
1200 * related resource not being available. In this case, set the request
1201 * aside in zone_list for retrying it later.
1203 list_add(&rq->queuelist, zone_list);
1204 __blk_mq_requeue_request(rq);
1208 * Returns true if we did some work AND can potentially do more.
1210 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1213 struct blk_mq_hw_ctx *hctx;
1214 struct request *rq, *nxt;
1215 bool no_tag = false;
1217 blk_status_t ret = BLK_STS_OK;
1218 bool no_budget_avail = false;
1219 LIST_HEAD(zone_list);
1221 if (list_empty(list))
1224 WARN_ON(!list_is_singular(list) && got_budget);
1227 * Now process all the entries, sending them to the driver.
1229 errors = queued = 0;
1231 struct blk_mq_queue_data bd;
1233 rq = list_first_entry(list, struct request, queuelist);
1236 if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) {
1237 blk_mq_put_driver_tag(rq);
1238 no_budget_avail = true;
1242 if (!blk_mq_get_driver_tag(rq)) {
1244 * The initial allocation attempt failed, so we need to
1245 * rerun the hardware queue when a tag is freed. The
1246 * waitqueue takes care of that. If the queue is run
1247 * before we add this entry back on the dispatch list,
1248 * we'll re-run it below.
1250 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1251 blk_mq_put_dispatch_budget(hctx);
1253 * For non-shared tags, the RESTART check
1256 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1262 list_del_init(&rq->queuelist);
1267 * Flag last if we have no more requests, or if we have more
1268 * but can't assign a driver tag to it.
1270 if (list_empty(list))
1273 nxt = list_first_entry(list, struct request, queuelist);
1274 bd.last = !blk_mq_get_driver_tag(nxt);
1277 ret = q->mq_ops->queue_rq(hctx, &bd);
1278 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1279 blk_mq_handle_dev_resource(rq, list);
1281 } else if (ret == BLK_STS_ZONE_RESOURCE) {
1283 * Move the request to zone_list and keep going through
1284 * the dispatch list to find more requests the drive can
1287 blk_mq_handle_zone_resource(rq, &zone_list);
1288 if (list_empty(list))
1293 if (unlikely(ret != BLK_STS_OK)) {
1295 blk_mq_end_request(rq, BLK_STS_IOERR);
1300 } while (!list_empty(list));
1302 if (!list_empty(&zone_list))
1303 list_splice_tail_init(&zone_list, list);
1305 hctx->dispatched[queued_to_index(queued)]++;
1308 * Any items that need requeuing? Stuff them into hctx->dispatch,
1309 * that is where we will continue on next queue run.
1311 if (!list_empty(list)) {
1315 * If we didn't flush the entire list, we could have told
1316 * the driver there was more coming, but that turned out to
1319 if (q->mq_ops->commit_rqs && queued)
1320 q->mq_ops->commit_rqs(hctx);
1322 spin_lock(&hctx->lock);
1323 list_splice_tail_init(list, &hctx->dispatch);
1324 spin_unlock(&hctx->lock);
1327 * If SCHED_RESTART was set by the caller of this function and
1328 * it is no longer set that means that it was cleared by another
1329 * thread and hence that a queue rerun is needed.
1331 * If 'no_tag' is set, that means that we failed getting
1332 * a driver tag with an I/O scheduler attached. If our dispatch
1333 * waitqueue is no longer active, ensure that we run the queue
1334 * AFTER adding our entries back to the list.
1336 * If no I/O scheduler has been configured it is possible that
1337 * the hardware queue got stopped and restarted before requests
1338 * were pushed back onto the dispatch list. Rerun the queue to
1339 * avoid starvation. Notes:
1340 * - blk_mq_run_hw_queue() checks whether or not a queue has
1341 * been stopped before rerunning a queue.
1342 * - Some but not all block drivers stop a queue before
1343 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1346 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1347 * bit is set, run queue after a delay to avoid IO stalls
1348 * that could otherwise occur if the queue is idle. We'll do
1349 * similar if we couldn't get budget and SCHED_RESTART is set.
1351 needs_restart = blk_mq_sched_needs_restart(hctx);
1352 if (!needs_restart ||
1353 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1354 blk_mq_run_hw_queue(hctx, true);
1355 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1357 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1359 blk_mq_update_dispatch_busy(hctx, true);
1362 blk_mq_update_dispatch_busy(hctx, false);
1365 * If the host/device is unable to accept more work, inform the
1368 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1371 return (queued + errors) != 0;
1375 * __blk_mq_run_hw_queue - Run a hardware queue.
1376 * @hctx: Pointer to the hardware queue to run.
1378 * Send pending requests to the hardware.
1380 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1385 * We should be running this queue from one of the CPUs that
1388 * There are at least two related races now between setting
1389 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1390 * __blk_mq_run_hw_queue():
1392 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1393 * but later it becomes online, then this warning is harmless
1396 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1397 * but later it becomes offline, then the warning can't be
1398 * triggered, and we depend on blk-mq timeout handler to
1399 * handle dispatched requests to this hctx
1401 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1402 cpu_online(hctx->next_cpu)) {
1403 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1404 raw_smp_processor_id(),
1405 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1410 * We can't run the queue inline with ints disabled. Ensure that
1411 * we catch bad users of this early.
1413 WARN_ON_ONCE(in_interrupt());
1415 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1417 hctx_lock(hctx, &srcu_idx);
1418 blk_mq_sched_dispatch_requests(hctx);
1419 hctx_unlock(hctx, srcu_idx);
1422 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1424 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1426 if (cpu >= nr_cpu_ids)
1427 cpu = cpumask_first(hctx->cpumask);
1432 * It'd be great if the workqueue API had a way to pass
1433 * in a mask and had some smarts for more clever placement.
1434 * For now we just round-robin here, switching for every
1435 * BLK_MQ_CPU_WORK_BATCH queued items.
1437 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1440 int next_cpu = hctx->next_cpu;
1442 if (hctx->queue->nr_hw_queues == 1)
1443 return WORK_CPU_UNBOUND;
1445 if (--hctx->next_cpu_batch <= 0) {
1447 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1449 if (next_cpu >= nr_cpu_ids)
1450 next_cpu = blk_mq_first_mapped_cpu(hctx);
1451 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1455 * Do unbound schedule if we can't find a online CPU for this hctx,
1456 * and it should only happen in the path of handling CPU DEAD.
1458 if (!cpu_online(next_cpu)) {
1465 * Make sure to re-select CPU next time once after CPUs
1466 * in hctx->cpumask become online again.
1468 hctx->next_cpu = next_cpu;
1469 hctx->next_cpu_batch = 1;
1470 return WORK_CPU_UNBOUND;
1473 hctx->next_cpu = next_cpu;
1478 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1479 * @hctx: Pointer to the hardware queue to run.
1480 * @async: If we want to run the queue asynchronously.
1481 * @msecs: Microseconds of delay to wait before running the queue.
1483 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1484 * with a delay of @msecs.
1486 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1487 unsigned long msecs)
1489 if (unlikely(blk_mq_hctx_stopped(hctx)))
1492 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1493 int cpu = get_cpu();
1494 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1495 __blk_mq_run_hw_queue(hctx);
1503 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1504 msecs_to_jiffies(msecs));
1508 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1509 * @hctx: Pointer to the hardware queue to run.
1510 * @msecs: Microseconds of delay to wait before running the queue.
1512 * Run a hardware queue asynchronously with a delay of @msecs.
1514 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1516 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1518 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1521 * blk_mq_run_hw_queue - Start to run a hardware queue.
1522 * @hctx: Pointer to the hardware queue to run.
1523 * @async: If we want to run the queue asynchronously.
1525 * Check if the request queue is not in a quiesced state and if there are
1526 * pending requests to be sent. If this is true, run the queue to send requests
1529 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1535 * When queue is quiesced, we may be switching io scheduler, or
1536 * updating nr_hw_queues, or other things, and we can't run queue
1537 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1539 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1542 hctx_lock(hctx, &srcu_idx);
1543 need_run = !blk_queue_quiesced(hctx->queue) &&
1544 blk_mq_hctx_has_pending(hctx);
1545 hctx_unlock(hctx, srcu_idx);
1548 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1550 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1553 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1554 * @q: Pointer to the request queue to run.
1555 * @async: If we want to run the queue asynchronously.
1557 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1559 struct blk_mq_hw_ctx *hctx;
1562 queue_for_each_hw_ctx(q, hctx, i) {
1563 if (blk_mq_hctx_stopped(hctx))
1566 blk_mq_run_hw_queue(hctx, async);
1569 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1572 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1573 * @q: Pointer to the request queue to run.
1574 * @msecs: Microseconds of delay to wait before running the queues.
1576 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1578 struct blk_mq_hw_ctx *hctx;
1581 queue_for_each_hw_ctx(q, hctx, i) {
1582 if (blk_mq_hctx_stopped(hctx))
1585 blk_mq_delay_run_hw_queue(hctx, msecs);
1588 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1591 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1592 * @q: request queue.
1594 * The caller is responsible for serializing this function against
1595 * blk_mq_{start,stop}_hw_queue().
1597 bool blk_mq_queue_stopped(struct request_queue *q)
1599 struct blk_mq_hw_ctx *hctx;
1602 queue_for_each_hw_ctx(q, hctx, i)
1603 if (blk_mq_hctx_stopped(hctx))
1608 EXPORT_SYMBOL(blk_mq_queue_stopped);
1611 * This function is often used for pausing .queue_rq() by driver when
1612 * there isn't enough resource or some conditions aren't satisfied, and
1613 * BLK_STS_RESOURCE is usually returned.
1615 * We do not guarantee that dispatch can be drained or blocked
1616 * after blk_mq_stop_hw_queue() returns. Please use
1617 * blk_mq_quiesce_queue() for that requirement.
1619 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1621 cancel_delayed_work(&hctx->run_work);
1623 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1625 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1628 * This function is often used for pausing .queue_rq() by driver when
1629 * there isn't enough resource or some conditions aren't satisfied, and
1630 * BLK_STS_RESOURCE is usually returned.
1632 * We do not guarantee that dispatch can be drained or blocked
1633 * after blk_mq_stop_hw_queues() returns. Please use
1634 * blk_mq_quiesce_queue() for that requirement.
1636 void blk_mq_stop_hw_queues(struct request_queue *q)
1638 struct blk_mq_hw_ctx *hctx;
1641 queue_for_each_hw_ctx(q, hctx, i)
1642 blk_mq_stop_hw_queue(hctx);
1644 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1646 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1648 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1650 blk_mq_run_hw_queue(hctx, false);
1652 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1654 void blk_mq_start_hw_queues(struct request_queue *q)
1656 struct blk_mq_hw_ctx *hctx;
1659 queue_for_each_hw_ctx(q, hctx, i)
1660 blk_mq_start_hw_queue(hctx);
1662 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1664 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1666 if (!blk_mq_hctx_stopped(hctx))
1669 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1670 blk_mq_run_hw_queue(hctx, async);
1672 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1674 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1676 struct blk_mq_hw_ctx *hctx;
1679 queue_for_each_hw_ctx(q, hctx, i)
1680 blk_mq_start_stopped_hw_queue(hctx, async);
1682 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1684 static void blk_mq_run_work_fn(struct work_struct *work)
1686 struct blk_mq_hw_ctx *hctx;
1688 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1691 * If we are stopped, don't run the queue.
1693 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1696 __blk_mq_run_hw_queue(hctx);
1699 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1703 struct blk_mq_ctx *ctx = rq->mq_ctx;
1704 enum hctx_type type = hctx->type;
1706 lockdep_assert_held(&ctx->lock);
1708 trace_block_rq_insert(hctx->queue, rq);
1711 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1713 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1716 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1719 struct blk_mq_ctx *ctx = rq->mq_ctx;
1721 lockdep_assert_held(&ctx->lock);
1723 __blk_mq_insert_req_list(hctx, rq, at_head);
1724 blk_mq_hctx_mark_pending(hctx, ctx);
1728 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1729 * @rq: Pointer to request to be inserted.
1730 * @run_queue: If we should run the hardware queue after inserting the request.
1732 * Should only be used carefully, when the caller knows we want to
1733 * bypass a potential IO scheduler on the target device.
1735 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1738 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1740 spin_lock(&hctx->lock);
1742 list_add(&rq->queuelist, &hctx->dispatch);
1744 list_add_tail(&rq->queuelist, &hctx->dispatch);
1745 spin_unlock(&hctx->lock);
1748 blk_mq_run_hw_queue(hctx, false);
1751 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1752 struct list_head *list)
1756 enum hctx_type type = hctx->type;
1759 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1762 list_for_each_entry(rq, list, queuelist) {
1763 BUG_ON(rq->mq_ctx != ctx);
1764 trace_block_rq_insert(hctx->queue, rq);
1767 spin_lock(&ctx->lock);
1768 list_splice_tail_init(list, &ctx->rq_lists[type]);
1769 blk_mq_hctx_mark_pending(hctx, ctx);
1770 spin_unlock(&ctx->lock);
1773 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1775 struct request *rqa = container_of(a, struct request, queuelist);
1776 struct request *rqb = container_of(b, struct request, queuelist);
1778 if (rqa->mq_ctx != rqb->mq_ctx)
1779 return rqa->mq_ctx > rqb->mq_ctx;
1780 if (rqa->mq_hctx != rqb->mq_hctx)
1781 return rqa->mq_hctx > rqb->mq_hctx;
1783 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1786 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1790 if (list_empty(&plug->mq_list))
1792 list_splice_init(&plug->mq_list, &list);
1794 if (plug->rq_count > 2 && plug->multiple_queues)
1795 list_sort(NULL, &list, plug_rq_cmp);
1800 struct list_head rq_list;
1801 struct request *rq, *head_rq = list_entry_rq(list.next);
1802 struct list_head *pos = &head_rq->queuelist; /* skip first */
1803 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1804 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1805 unsigned int depth = 1;
1807 list_for_each_continue(pos, &list) {
1808 rq = list_entry_rq(pos);
1810 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1815 list_cut_before(&rq_list, &list, pos);
1816 trace_block_unplug(head_rq->q, depth, !from_schedule);
1817 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1819 } while(!list_empty(&list));
1822 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1823 unsigned int nr_segs)
1825 if (bio->bi_opf & REQ_RAHEAD)
1826 rq->cmd_flags |= REQ_FAILFAST_MASK;
1828 rq->__sector = bio->bi_iter.bi_sector;
1829 rq->write_hint = bio->bi_write_hint;
1830 blk_rq_bio_prep(rq, bio, nr_segs);
1831 blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1833 blk_account_io_start(rq);
1836 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1838 blk_qc_t *cookie, bool last)
1840 struct request_queue *q = rq->q;
1841 struct blk_mq_queue_data bd = {
1845 blk_qc_t new_cookie;
1848 new_cookie = request_to_qc_t(hctx, rq);
1851 * For OK queue, we are done. For error, caller may kill it.
1852 * Any other error (busy), just add it to our list as we
1853 * previously would have done.
1855 ret = q->mq_ops->queue_rq(hctx, &bd);
1858 blk_mq_update_dispatch_busy(hctx, false);
1859 *cookie = new_cookie;
1861 case BLK_STS_RESOURCE:
1862 case BLK_STS_DEV_RESOURCE:
1863 blk_mq_update_dispatch_busy(hctx, true);
1864 __blk_mq_requeue_request(rq);
1867 blk_mq_update_dispatch_busy(hctx, false);
1868 *cookie = BLK_QC_T_NONE;
1875 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1878 bool bypass_insert, bool last)
1880 struct request_queue *q = rq->q;
1881 bool run_queue = true;
1884 * RCU or SRCU read lock is needed before checking quiesced flag.
1886 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1887 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1888 * and avoid driver to try to dispatch again.
1890 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1892 bypass_insert = false;
1896 if (q->elevator && !bypass_insert)
1899 if (!blk_mq_get_dispatch_budget(hctx))
1902 if (!blk_mq_get_driver_tag(rq)) {
1903 blk_mq_put_dispatch_budget(hctx);
1907 return __blk_mq_issue_directly(hctx, rq, cookie, last);
1910 return BLK_STS_RESOURCE;
1912 blk_mq_request_bypass_insert(rq, false, run_queue);
1917 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
1918 * @hctx: Pointer of the associated hardware queue.
1919 * @rq: Pointer to request to be sent.
1920 * @cookie: Request queue cookie.
1922 * If the device has enough resources to accept a new request now, send the
1923 * request directly to device driver. Else, insert at hctx->dispatch queue, so
1924 * we can try send it another time in the future. Requests inserted at this
1925 * queue have higher priority.
1927 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1928 struct request *rq, blk_qc_t *cookie)
1933 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1935 hctx_lock(hctx, &srcu_idx);
1937 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1938 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1939 blk_mq_request_bypass_insert(rq, false, true);
1940 else if (ret != BLK_STS_OK)
1941 blk_mq_end_request(rq, ret);
1943 hctx_unlock(hctx, srcu_idx);
1946 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1950 blk_qc_t unused_cookie;
1951 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1953 hctx_lock(hctx, &srcu_idx);
1954 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1955 hctx_unlock(hctx, srcu_idx);
1960 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1961 struct list_head *list)
1965 while (!list_empty(list)) {
1967 struct request *rq = list_first_entry(list, struct request,
1970 list_del_init(&rq->queuelist);
1971 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1972 if (ret != BLK_STS_OK) {
1973 if (ret == BLK_STS_RESOURCE ||
1974 ret == BLK_STS_DEV_RESOURCE) {
1975 blk_mq_request_bypass_insert(rq, false,
1979 blk_mq_end_request(rq, ret);
1985 * If we didn't flush the entire list, we could have told
1986 * the driver there was more coming, but that turned out to
1989 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs && queued)
1990 hctx->queue->mq_ops->commit_rqs(hctx);
1993 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1995 list_add_tail(&rq->queuelist, &plug->mq_list);
1997 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1998 struct request *tmp;
2000 tmp = list_first_entry(&plug->mq_list, struct request,
2002 if (tmp->q != rq->q)
2003 plug->multiple_queues = true;
2008 * blk_mq_make_request - Create and send a request to block device.
2009 * @q: Request queue pointer.
2010 * @bio: Bio pointer.
2012 * Builds up a request structure from @q and @bio and send to the device. The
2013 * request may not be queued directly to hardware if:
2014 * * This request can be merged with another one
2015 * * We want to place request at plug queue for possible future merging
2016 * * There is an IO scheduler active at this queue
2018 * It will not queue the request if there is an error with the bio, or at the
2021 * Returns: Request queue cookie.
2023 blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
2025 const int is_sync = op_is_sync(bio->bi_opf);
2026 const int is_flush_fua = op_is_flush(bio->bi_opf);
2027 struct blk_mq_alloc_data data = {
2031 struct blk_plug *plug;
2032 struct request *same_queue_rq = NULL;
2033 unsigned int nr_segs;
2037 blk_queue_bounce(q, &bio);
2038 __blk_queue_split(q, &bio, &nr_segs);
2040 if (!bio_integrity_prep(bio))
2043 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2044 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2047 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2050 rq_qos_throttle(q, bio);
2052 data.cmd_flags = bio->bi_opf;
2053 rq = __blk_mq_alloc_request(&data);
2054 if (unlikely(!rq)) {
2055 rq_qos_cleanup(q, bio);
2056 if (bio->bi_opf & REQ_NOWAIT)
2057 bio_wouldblock_error(bio);
2061 trace_block_getrq(q, bio, bio->bi_opf);
2063 rq_qos_track(q, rq, bio);
2065 cookie = request_to_qc_t(data.hctx, rq);
2067 blk_mq_bio_to_request(rq, bio, nr_segs);
2069 ret = blk_crypto_init_request(rq);
2070 if (ret != BLK_STS_OK) {
2071 bio->bi_status = ret;
2073 blk_mq_free_request(rq);
2074 return BLK_QC_T_NONE;
2077 plug = blk_mq_plug(q, bio);
2078 if (unlikely(is_flush_fua)) {
2079 /* Bypass scheduler for flush requests */
2080 blk_insert_flush(rq);
2081 blk_mq_run_hw_queue(data.hctx, true);
2082 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2083 !blk_queue_nonrot(q))) {
2085 * Use plugging if we have a ->commit_rqs() hook as well, as
2086 * we know the driver uses bd->last in a smart fashion.
2088 * Use normal plugging if this disk is slow HDD, as sequential
2089 * IO may benefit a lot from plug merging.
2091 unsigned int request_count = plug->rq_count;
2092 struct request *last = NULL;
2095 trace_block_plug(q);
2097 last = list_entry_rq(plug->mq_list.prev);
2099 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2100 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2101 blk_flush_plug_list(plug, false);
2102 trace_block_plug(q);
2105 blk_add_rq_to_plug(plug, rq);
2106 } else if (q->elevator) {
2107 /* Insert the request at the IO scheduler queue */
2108 blk_mq_sched_insert_request(rq, false, true, true);
2109 } else if (plug && !blk_queue_nomerges(q)) {
2111 * We do limited plugging. If the bio can be merged, do that.
2112 * Otherwise the existing request in the plug list will be
2113 * issued. So the plug list will have one request at most
2114 * The plug list might get flushed before this. If that happens,
2115 * the plug list is empty, and same_queue_rq is invalid.
2117 if (list_empty(&plug->mq_list))
2118 same_queue_rq = NULL;
2119 if (same_queue_rq) {
2120 list_del_init(&same_queue_rq->queuelist);
2123 blk_add_rq_to_plug(plug, rq);
2124 trace_block_plug(q);
2126 if (same_queue_rq) {
2127 data.hctx = same_queue_rq->mq_hctx;
2128 trace_block_unplug(q, 1, true);
2129 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2132 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2133 !data.hctx->dispatch_busy) {
2135 * There is no scheduler and we can try to send directly
2138 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2141 blk_mq_sched_insert_request(rq, false, true, true);
2147 return BLK_QC_T_NONE;
2149 EXPORT_SYMBOL_GPL(blk_mq_make_request); /* only for request based dm */
2151 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2152 unsigned int hctx_idx)
2156 if (tags->rqs && set->ops->exit_request) {
2159 for (i = 0; i < tags->nr_tags; i++) {
2160 struct request *rq = tags->static_rqs[i];
2164 set->ops->exit_request(set, rq, hctx_idx);
2165 tags->static_rqs[i] = NULL;
2169 while (!list_empty(&tags->page_list)) {
2170 page = list_first_entry(&tags->page_list, struct page, lru);
2171 list_del_init(&page->lru);
2173 * Remove kmemleak object previously allocated in
2174 * blk_mq_alloc_rqs().
2176 kmemleak_free(page_address(page));
2177 __free_pages(page, page->private);
2181 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2185 kfree(tags->static_rqs);
2186 tags->static_rqs = NULL;
2188 blk_mq_free_tags(tags);
2191 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2192 unsigned int hctx_idx,
2193 unsigned int nr_tags,
2194 unsigned int reserved_tags)
2196 struct blk_mq_tags *tags;
2199 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2200 if (node == NUMA_NO_NODE)
2201 node = set->numa_node;
2203 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2204 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2208 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2209 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2212 blk_mq_free_tags(tags);
2216 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2217 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2219 if (!tags->static_rqs) {
2221 blk_mq_free_tags(tags);
2228 static size_t order_to_size(unsigned int order)
2230 return (size_t)PAGE_SIZE << order;
2233 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2234 unsigned int hctx_idx, int node)
2238 if (set->ops->init_request) {
2239 ret = set->ops->init_request(set, rq, hctx_idx, node);
2244 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2248 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2249 unsigned int hctx_idx, unsigned int depth)
2251 unsigned int i, j, entries_per_page, max_order = 4;
2252 size_t rq_size, left;
2255 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2256 if (node == NUMA_NO_NODE)
2257 node = set->numa_node;
2259 INIT_LIST_HEAD(&tags->page_list);
2262 * rq_size is the size of the request plus driver payload, rounded
2263 * to the cacheline size
2265 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2267 left = rq_size * depth;
2269 for (i = 0; i < depth; ) {
2270 int this_order = max_order;
2275 while (this_order && left < order_to_size(this_order - 1))
2279 page = alloc_pages_node(node,
2280 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2286 if (order_to_size(this_order) < rq_size)
2293 page->private = this_order;
2294 list_add_tail(&page->lru, &tags->page_list);
2296 p = page_address(page);
2298 * Allow kmemleak to scan these pages as they contain pointers
2299 * to additional allocations like via ops->init_request().
2301 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2302 entries_per_page = order_to_size(this_order) / rq_size;
2303 to_do = min(entries_per_page, depth - i);
2304 left -= to_do * rq_size;
2305 for (j = 0; j < to_do; j++) {
2306 struct request *rq = p;
2308 tags->static_rqs[i] = rq;
2309 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2310 tags->static_rqs[i] = NULL;
2321 blk_mq_free_rqs(set, tags, hctx_idx);
2325 struct rq_iter_data {
2326 struct blk_mq_hw_ctx *hctx;
2330 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2332 struct rq_iter_data *iter_data = data;
2334 if (rq->mq_hctx != iter_data->hctx)
2336 iter_data->has_rq = true;
2340 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2342 struct blk_mq_tags *tags = hctx->sched_tags ?
2343 hctx->sched_tags : hctx->tags;
2344 struct rq_iter_data data = {
2348 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2352 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2353 struct blk_mq_hw_ctx *hctx)
2355 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2357 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2362 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2364 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2365 struct blk_mq_hw_ctx, cpuhp_online);
2367 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2368 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2372 * Prevent new request from being allocated on the current hctx.
2374 * The smp_mb__after_atomic() Pairs with the implied barrier in
2375 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2376 * seen once we return from the tag allocator.
2378 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2379 smp_mb__after_atomic();
2382 * Try to grab a reference to the queue and wait for any outstanding
2383 * requests. If we could not grab a reference the queue has been
2384 * frozen and there are no requests.
2386 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2387 while (blk_mq_hctx_has_requests(hctx))
2389 percpu_ref_put(&hctx->queue->q_usage_counter);
2395 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2397 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2398 struct blk_mq_hw_ctx, cpuhp_online);
2400 if (cpumask_test_cpu(cpu, hctx->cpumask))
2401 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2406 * 'cpu' is going away. splice any existing rq_list entries from this
2407 * software queue to the hw queue dispatch list, and ensure that it
2410 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2412 struct blk_mq_hw_ctx *hctx;
2413 struct blk_mq_ctx *ctx;
2415 enum hctx_type type;
2417 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2418 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2421 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2424 spin_lock(&ctx->lock);
2425 if (!list_empty(&ctx->rq_lists[type])) {
2426 list_splice_init(&ctx->rq_lists[type], &tmp);
2427 blk_mq_hctx_clear_pending(hctx, ctx);
2429 spin_unlock(&ctx->lock);
2431 if (list_empty(&tmp))
2434 spin_lock(&hctx->lock);
2435 list_splice_tail_init(&tmp, &hctx->dispatch);
2436 spin_unlock(&hctx->lock);
2438 blk_mq_run_hw_queue(hctx, true);
2442 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2444 if (!(hctx->flags & BLK_MQ_F_STACKING))
2445 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2446 &hctx->cpuhp_online);
2447 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2451 /* hctx->ctxs will be freed in queue's release handler */
2452 static void blk_mq_exit_hctx(struct request_queue *q,
2453 struct blk_mq_tag_set *set,
2454 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2456 if (blk_mq_hw_queue_mapped(hctx))
2457 blk_mq_tag_idle(hctx);
2459 if (set->ops->exit_request)
2460 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2462 if (set->ops->exit_hctx)
2463 set->ops->exit_hctx(hctx, hctx_idx);
2465 blk_mq_remove_cpuhp(hctx);
2467 spin_lock(&q->unused_hctx_lock);
2468 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2469 spin_unlock(&q->unused_hctx_lock);
2472 static void blk_mq_exit_hw_queues(struct request_queue *q,
2473 struct blk_mq_tag_set *set, int nr_queue)
2475 struct blk_mq_hw_ctx *hctx;
2478 queue_for_each_hw_ctx(q, hctx, i) {
2481 blk_mq_debugfs_unregister_hctx(hctx);
2482 blk_mq_exit_hctx(q, set, hctx, i);
2486 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2488 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2490 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2491 __alignof__(struct blk_mq_hw_ctx)) !=
2492 sizeof(struct blk_mq_hw_ctx));
2494 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2495 hw_ctx_size += sizeof(struct srcu_struct);
2500 static int blk_mq_init_hctx(struct request_queue *q,
2501 struct blk_mq_tag_set *set,
2502 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2504 hctx->queue_num = hctx_idx;
2506 if (!(hctx->flags & BLK_MQ_F_STACKING))
2507 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2508 &hctx->cpuhp_online);
2509 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2511 hctx->tags = set->tags[hctx_idx];
2513 if (set->ops->init_hctx &&
2514 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2515 goto unregister_cpu_notifier;
2517 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2523 if (set->ops->exit_hctx)
2524 set->ops->exit_hctx(hctx, hctx_idx);
2525 unregister_cpu_notifier:
2526 blk_mq_remove_cpuhp(hctx);
2530 static struct blk_mq_hw_ctx *
2531 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2534 struct blk_mq_hw_ctx *hctx;
2535 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2537 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2539 goto fail_alloc_hctx;
2541 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2544 atomic_set(&hctx->nr_active, 0);
2545 if (node == NUMA_NO_NODE)
2546 node = set->numa_node;
2547 hctx->numa_node = node;
2549 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2550 spin_lock_init(&hctx->lock);
2551 INIT_LIST_HEAD(&hctx->dispatch);
2553 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2555 INIT_LIST_HEAD(&hctx->hctx_list);
2558 * Allocate space for all possible cpus to avoid allocation at
2561 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2566 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2571 spin_lock_init(&hctx->dispatch_wait_lock);
2572 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2573 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2575 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2579 if (hctx->flags & BLK_MQ_F_BLOCKING)
2580 init_srcu_struct(hctx->srcu);
2581 blk_mq_hctx_kobj_init(hctx);
2586 sbitmap_free(&hctx->ctx_map);
2590 free_cpumask_var(hctx->cpumask);
2597 static void blk_mq_init_cpu_queues(struct request_queue *q,
2598 unsigned int nr_hw_queues)
2600 struct blk_mq_tag_set *set = q->tag_set;
2603 for_each_possible_cpu(i) {
2604 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2605 struct blk_mq_hw_ctx *hctx;
2609 spin_lock_init(&__ctx->lock);
2610 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2611 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2616 * Set local node, IFF we have more than one hw queue. If
2617 * not, we remain on the home node of the device
2619 for (j = 0; j < set->nr_maps; j++) {
2620 hctx = blk_mq_map_queue_type(q, j, i);
2621 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2622 hctx->numa_node = local_memory_node(cpu_to_node(i));
2627 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2632 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2633 set->queue_depth, set->reserved_tags);
2634 if (!set->tags[hctx_idx])
2637 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2642 blk_mq_free_rq_map(set->tags[hctx_idx]);
2643 set->tags[hctx_idx] = NULL;
2647 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2648 unsigned int hctx_idx)
2650 if (set->tags && set->tags[hctx_idx]) {
2651 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2652 blk_mq_free_rq_map(set->tags[hctx_idx]);
2653 set->tags[hctx_idx] = NULL;
2657 static void blk_mq_map_swqueue(struct request_queue *q)
2659 unsigned int i, j, hctx_idx;
2660 struct blk_mq_hw_ctx *hctx;
2661 struct blk_mq_ctx *ctx;
2662 struct blk_mq_tag_set *set = q->tag_set;
2664 queue_for_each_hw_ctx(q, hctx, i) {
2665 cpumask_clear(hctx->cpumask);
2667 hctx->dispatch_from = NULL;
2671 * Map software to hardware queues.
2673 * If the cpu isn't present, the cpu is mapped to first hctx.
2675 for_each_possible_cpu(i) {
2677 ctx = per_cpu_ptr(q->queue_ctx, i);
2678 for (j = 0; j < set->nr_maps; j++) {
2679 if (!set->map[j].nr_queues) {
2680 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2681 HCTX_TYPE_DEFAULT, i);
2684 hctx_idx = set->map[j].mq_map[i];
2685 /* unmapped hw queue can be remapped after CPU topo changed */
2686 if (!set->tags[hctx_idx] &&
2687 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2689 * If tags initialization fail for some hctx,
2690 * that hctx won't be brought online. In this
2691 * case, remap the current ctx to hctx[0] which
2692 * is guaranteed to always have tags allocated
2694 set->map[j].mq_map[i] = 0;
2697 hctx = blk_mq_map_queue_type(q, j, i);
2698 ctx->hctxs[j] = hctx;
2700 * If the CPU is already set in the mask, then we've
2701 * mapped this one already. This can happen if
2702 * devices share queues across queue maps.
2704 if (cpumask_test_cpu(i, hctx->cpumask))
2707 cpumask_set_cpu(i, hctx->cpumask);
2709 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2710 hctx->ctxs[hctx->nr_ctx++] = ctx;
2713 * If the nr_ctx type overflows, we have exceeded the
2714 * amount of sw queues we can support.
2716 BUG_ON(!hctx->nr_ctx);
2719 for (; j < HCTX_MAX_TYPES; j++)
2720 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2721 HCTX_TYPE_DEFAULT, i);
2724 queue_for_each_hw_ctx(q, hctx, i) {
2726 * If no software queues are mapped to this hardware queue,
2727 * disable it and free the request entries.
2729 if (!hctx->nr_ctx) {
2730 /* Never unmap queue 0. We need it as a
2731 * fallback in case of a new remap fails
2734 if (i && set->tags[i])
2735 blk_mq_free_map_and_requests(set, i);
2741 hctx->tags = set->tags[i];
2742 WARN_ON(!hctx->tags);
2745 * Set the map size to the number of mapped software queues.
2746 * This is more accurate and more efficient than looping
2747 * over all possibly mapped software queues.
2749 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2752 * Initialize batch roundrobin counts
2754 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2755 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2760 * Caller needs to ensure that we're either frozen/quiesced, or that
2761 * the queue isn't live yet.
2763 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2765 struct blk_mq_hw_ctx *hctx;
2768 queue_for_each_hw_ctx(q, hctx, i) {
2770 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2772 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2776 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2779 struct request_queue *q;
2781 lockdep_assert_held(&set->tag_list_lock);
2783 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2784 blk_mq_freeze_queue(q);
2785 queue_set_hctx_shared(q, shared);
2786 blk_mq_unfreeze_queue(q);
2790 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2792 struct blk_mq_tag_set *set = q->tag_set;
2794 mutex_lock(&set->tag_list_lock);
2795 list_del_rcu(&q->tag_set_list);
2796 if (list_is_singular(&set->tag_list)) {
2797 /* just transitioned to unshared */
2798 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2799 /* update existing queue */
2800 blk_mq_update_tag_set_depth(set, false);
2802 mutex_unlock(&set->tag_list_lock);
2803 INIT_LIST_HEAD(&q->tag_set_list);
2806 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2807 struct request_queue *q)
2809 mutex_lock(&set->tag_list_lock);
2812 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2814 if (!list_empty(&set->tag_list) &&
2815 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2816 set->flags |= BLK_MQ_F_TAG_SHARED;
2817 /* update existing queue */
2818 blk_mq_update_tag_set_depth(set, true);
2820 if (set->flags & BLK_MQ_F_TAG_SHARED)
2821 queue_set_hctx_shared(q, true);
2822 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2824 mutex_unlock(&set->tag_list_lock);
2827 /* All allocations will be freed in release handler of q->mq_kobj */
2828 static int blk_mq_alloc_ctxs(struct request_queue *q)
2830 struct blk_mq_ctxs *ctxs;
2833 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2837 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2838 if (!ctxs->queue_ctx)
2841 for_each_possible_cpu(cpu) {
2842 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2846 q->mq_kobj = &ctxs->kobj;
2847 q->queue_ctx = ctxs->queue_ctx;
2856 * It is the actual release handler for mq, but we do it from
2857 * request queue's release handler for avoiding use-after-free
2858 * and headache because q->mq_kobj shouldn't have been introduced,
2859 * but we can't group ctx/kctx kobj without it.
2861 void blk_mq_release(struct request_queue *q)
2863 struct blk_mq_hw_ctx *hctx, *next;
2866 queue_for_each_hw_ctx(q, hctx, i)
2867 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2869 /* all hctx are in .unused_hctx_list now */
2870 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2871 list_del_init(&hctx->hctx_list);
2872 kobject_put(&hctx->kobj);
2875 kfree(q->queue_hw_ctx);
2878 * release .mq_kobj and sw queue's kobject now because
2879 * both share lifetime with request queue.
2881 blk_mq_sysfs_deinit(q);
2884 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
2887 struct request_queue *uninit_q, *q;
2889 uninit_q = __blk_alloc_queue(set->numa_node);
2891 return ERR_PTR(-ENOMEM);
2892 uninit_q->queuedata = queuedata;
2895 * Initialize the queue without an elevator. device_add_disk() will do
2896 * the initialization.
2898 q = blk_mq_init_allocated_queue(set, uninit_q, false);
2900 blk_cleanup_queue(uninit_q);
2904 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
2906 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2908 return blk_mq_init_queue_data(set, NULL);
2910 EXPORT_SYMBOL(blk_mq_init_queue);
2913 * Helper for setting up a queue with mq ops, given queue depth, and
2914 * the passed in mq ops flags.
2916 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2917 const struct blk_mq_ops *ops,
2918 unsigned int queue_depth,
2919 unsigned int set_flags)
2921 struct request_queue *q;
2924 memset(set, 0, sizeof(*set));
2926 set->nr_hw_queues = 1;
2928 set->queue_depth = queue_depth;
2929 set->numa_node = NUMA_NO_NODE;
2930 set->flags = set_flags;
2932 ret = blk_mq_alloc_tag_set(set);
2934 return ERR_PTR(ret);
2936 q = blk_mq_init_queue(set);
2938 blk_mq_free_tag_set(set);
2944 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2946 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2947 struct blk_mq_tag_set *set, struct request_queue *q,
2948 int hctx_idx, int node)
2950 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2952 /* reuse dead hctx first */
2953 spin_lock(&q->unused_hctx_lock);
2954 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2955 if (tmp->numa_node == node) {
2961 list_del_init(&hctx->hctx_list);
2962 spin_unlock(&q->unused_hctx_lock);
2965 hctx = blk_mq_alloc_hctx(q, set, node);
2969 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2975 kobject_put(&hctx->kobj);
2980 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2981 struct request_queue *q)
2984 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2986 if (q->nr_hw_queues < set->nr_hw_queues) {
2987 struct blk_mq_hw_ctx **new_hctxs;
2989 new_hctxs = kcalloc_node(set->nr_hw_queues,
2990 sizeof(*new_hctxs), GFP_KERNEL,
2995 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
2997 q->queue_hw_ctx = new_hctxs;
3002 /* protect against switching io scheduler */
3003 mutex_lock(&q->sysfs_lock);
3004 for (i = 0; i < set->nr_hw_queues; i++) {
3006 struct blk_mq_hw_ctx *hctx;
3008 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3010 * If the hw queue has been mapped to another numa node,
3011 * we need to realloc the hctx. If allocation fails, fallback
3012 * to use the previous one.
3014 if (hctxs[i] && (hctxs[i]->numa_node == node))
3017 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3020 blk_mq_exit_hctx(q, set, hctxs[i], i);
3024 pr_warn("Allocate new hctx on node %d fails,\
3025 fallback to previous one on node %d\n",
3026 node, hctxs[i]->numa_node);
3032 * Increasing nr_hw_queues fails. Free the newly allocated
3033 * hctxs and keep the previous q->nr_hw_queues.
3035 if (i != set->nr_hw_queues) {
3036 j = q->nr_hw_queues;
3040 end = q->nr_hw_queues;
3041 q->nr_hw_queues = set->nr_hw_queues;
3044 for (; j < end; j++) {
3045 struct blk_mq_hw_ctx *hctx = hctxs[j];
3049 blk_mq_free_map_and_requests(set, j);
3050 blk_mq_exit_hctx(q, set, hctx, j);
3054 mutex_unlock(&q->sysfs_lock);
3057 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3058 struct request_queue *q,
3061 /* mark the queue as mq asap */
3062 q->mq_ops = set->ops;
3064 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3065 blk_mq_poll_stats_bkt,
3066 BLK_MQ_POLL_STATS_BKTS, q);
3070 if (blk_mq_alloc_ctxs(q))
3073 /* init q->mq_kobj and sw queues' kobjects */
3074 blk_mq_sysfs_init(q);
3076 INIT_LIST_HEAD(&q->unused_hctx_list);
3077 spin_lock_init(&q->unused_hctx_lock);
3079 blk_mq_realloc_hw_ctxs(set, q);
3080 if (!q->nr_hw_queues)
3083 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3084 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3088 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3089 if (set->nr_maps > HCTX_TYPE_POLL &&
3090 set->map[HCTX_TYPE_POLL].nr_queues)
3091 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3093 q->sg_reserved_size = INT_MAX;
3095 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3096 INIT_LIST_HEAD(&q->requeue_list);
3097 spin_lock_init(&q->requeue_lock);
3099 q->nr_requests = set->queue_depth;
3102 * Default to classic polling
3104 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3106 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3107 blk_mq_add_queue_tag_set(set, q);
3108 blk_mq_map_swqueue(q);
3111 elevator_init_mq(q);
3116 kfree(q->queue_hw_ctx);
3117 q->nr_hw_queues = 0;
3118 blk_mq_sysfs_deinit(q);
3120 blk_stat_free_callback(q->poll_cb);
3124 return ERR_PTR(-ENOMEM);
3126 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3128 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3129 void blk_mq_exit_queue(struct request_queue *q)
3131 struct blk_mq_tag_set *set = q->tag_set;
3133 blk_mq_del_queue_tag_set(q);
3134 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3137 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3141 for (i = 0; i < set->nr_hw_queues; i++)
3142 if (!__blk_mq_alloc_map_and_request(set, i))
3149 blk_mq_free_map_and_requests(set, i);
3155 * Allocate the request maps associated with this tag_set. Note that this
3156 * may reduce the depth asked for, if memory is tight. set->queue_depth
3157 * will be updated to reflect the allocated depth.
3159 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3164 depth = set->queue_depth;
3166 err = __blk_mq_alloc_rq_maps(set);
3170 set->queue_depth >>= 1;
3171 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3175 } while (set->queue_depth);
3177 if (!set->queue_depth || err) {
3178 pr_err("blk-mq: failed to allocate request map\n");
3182 if (depth != set->queue_depth)
3183 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3184 depth, set->queue_depth);
3189 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3192 * blk_mq_map_queues() and multiple .map_queues() implementations
3193 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3194 * number of hardware queues.
3196 if (set->nr_maps == 1)
3197 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3199 if (set->ops->map_queues && !is_kdump_kernel()) {
3203 * transport .map_queues is usually done in the following
3206 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3207 * mask = get_cpu_mask(queue)
3208 * for_each_cpu(cpu, mask)
3209 * set->map[x].mq_map[cpu] = queue;
3212 * When we need to remap, the table has to be cleared for
3213 * killing stale mapping since one CPU may not be mapped
3216 for (i = 0; i < set->nr_maps; i++)
3217 blk_mq_clear_mq_map(&set->map[i]);
3219 return set->ops->map_queues(set);
3221 BUG_ON(set->nr_maps > 1);
3222 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3226 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3227 int cur_nr_hw_queues, int new_nr_hw_queues)
3229 struct blk_mq_tags **new_tags;
3231 if (cur_nr_hw_queues >= new_nr_hw_queues)
3234 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3235 GFP_KERNEL, set->numa_node);
3240 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3241 sizeof(*set->tags));
3243 set->tags = new_tags;
3244 set->nr_hw_queues = new_nr_hw_queues;
3250 * Alloc a tag set to be associated with one or more request queues.
3251 * May fail with EINVAL for various error conditions. May adjust the
3252 * requested depth down, if it's too large. In that case, the set
3253 * value will be stored in set->queue_depth.
3255 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3259 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3261 if (!set->nr_hw_queues)
3263 if (!set->queue_depth)
3265 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3268 if (!set->ops->queue_rq)
3271 if (!set->ops->get_budget ^ !set->ops->put_budget)
3274 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3275 pr_info("blk-mq: reduced tag depth to %u\n",
3277 set->queue_depth = BLK_MQ_MAX_DEPTH;
3282 else if (set->nr_maps > HCTX_MAX_TYPES)
3286 * If a crashdump is active, then we are potentially in a very
3287 * memory constrained environment. Limit us to 1 queue and
3288 * 64 tags to prevent using too much memory.
3290 if (is_kdump_kernel()) {
3291 set->nr_hw_queues = 1;
3293 set->queue_depth = min(64U, set->queue_depth);
3296 * There is no use for more h/w queues than cpus if we just have
3299 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3300 set->nr_hw_queues = nr_cpu_ids;
3302 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3306 for (i = 0; i < set->nr_maps; i++) {
3307 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3308 sizeof(set->map[i].mq_map[0]),
3309 GFP_KERNEL, set->numa_node);
3310 if (!set->map[i].mq_map)
3311 goto out_free_mq_map;
3312 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3315 ret = blk_mq_update_queue_map(set);
3317 goto out_free_mq_map;
3319 ret = blk_mq_alloc_map_and_requests(set);
3321 goto out_free_mq_map;
3323 mutex_init(&set->tag_list_lock);
3324 INIT_LIST_HEAD(&set->tag_list);
3329 for (i = 0; i < set->nr_maps; i++) {
3330 kfree(set->map[i].mq_map);
3331 set->map[i].mq_map = NULL;
3337 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3339 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3343 for (i = 0; i < set->nr_hw_queues; i++)
3344 blk_mq_free_map_and_requests(set, i);
3346 for (j = 0; j < set->nr_maps; j++) {
3347 kfree(set->map[j].mq_map);
3348 set->map[j].mq_map = NULL;
3354 EXPORT_SYMBOL(blk_mq_free_tag_set);
3356 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3358 struct blk_mq_tag_set *set = q->tag_set;
3359 struct blk_mq_hw_ctx *hctx;
3365 if (q->nr_requests == nr)
3368 blk_mq_freeze_queue(q);
3369 blk_mq_quiesce_queue(q);
3372 queue_for_each_hw_ctx(q, hctx, i) {
3376 * If we're using an MQ scheduler, just update the scheduler
3377 * queue depth. This is similar to what the old code would do.
3379 if (!hctx->sched_tags) {
3380 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3383 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3388 if (q->elevator && q->elevator->type->ops.depth_updated)
3389 q->elevator->type->ops.depth_updated(hctx);
3393 q->nr_requests = nr;
3395 blk_mq_unquiesce_queue(q);
3396 blk_mq_unfreeze_queue(q);
3402 * request_queue and elevator_type pair.
3403 * It is just used by __blk_mq_update_nr_hw_queues to cache
3404 * the elevator_type associated with a request_queue.
3406 struct blk_mq_qe_pair {
3407 struct list_head node;
3408 struct request_queue *q;
3409 struct elevator_type *type;
3413 * Cache the elevator_type in qe pair list and switch the
3414 * io scheduler to 'none'
3416 static bool blk_mq_elv_switch_none(struct list_head *head,
3417 struct request_queue *q)
3419 struct blk_mq_qe_pair *qe;
3424 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3428 INIT_LIST_HEAD(&qe->node);
3430 qe->type = q->elevator->type;
3431 list_add(&qe->node, head);
3433 mutex_lock(&q->sysfs_lock);
3435 * After elevator_switch_mq, the previous elevator_queue will be
3436 * released by elevator_release. The reference of the io scheduler
3437 * module get by elevator_get will also be put. So we need to get
3438 * a reference of the io scheduler module here to prevent it to be
3441 __module_get(qe->type->elevator_owner);
3442 elevator_switch_mq(q, NULL);
3443 mutex_unlock(&q->sysfs_lock);
3448 static void blk_mq_elv_switch_back(struct list_head *head,
3449 struct request_queue *q)
3451 struct blk_mq_qe_pair *qe;
3452 struct elevator_type *t = NULL;
3454 list_for_each_entry(qe, head, node)
3463 list_del(&qe->node);
3466 mutex_lock(&q->sysfs_lock);
3467 elevator_switch_mq(q, t);
3468 mutex_unlock(&q->sysfs_lock);
3471 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3474 struct request_queue *q;
3476 int prev_nr_hw_queues;
3478 lockdep_assert_held(&set->tag_list_lock);
3480 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3481 nr_hw_queues = nr_cpu_ids;
3482 if (nr_hw_queues < 1)
3484 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3487 list_for_each_entry(q, &set->tag_list, tag_set_list)
3488 blk_mq_freeze_queue(q);
3490 * Switch IO scheduler to 'none', cleaning up the data associated
3491 * with the previous scheduler. We will switch back once we are done
3492 * updating the new sw to hw queue mappings.
3494 list_for_each_entry(q, &set->tag_list, tag_set_list)
3495 if (!blk_mq_elv_switch_none(&head, q))
3498 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3499 blk_mq_debugfs_unregister_hctxs(q);
3500 blk_mq_sysfs_unregister(q);
3503 prev_nr_hw_queues = set->nr_hw_queues;
3504 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3508 set->nr_hw_queues = nr_hw_queues;
3510 blk_mq_update_queue_map(set);
3511 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3512 blk_mq_realloc_hw_ctxs(set, q);
3513 if (q->nr_hw_queues != set->nr_hw_queues) {
3514 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3515 nr_hw_queues, prev_nr_hw_queues);
3516 set->nr_hw_queues = prev_nr_hw_queues;
3517 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3520 blk_mq_map_swqueue(q);
3524 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3525 blk_mq_sysfs_register(q);
3526 blk_mq_debugfs_register_hctxs(q);
3530 list_for_each_entry(q, &set->tag_list, tag_set_list)
3531 blk_mq_elv_switch_back(&head, q);
3533 list_for_each_entry(q, &set->tag_list, tag_set_list)
3534 blk_mq_unfreeze_queue(q);
3537 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3539 mutex_lock(&set->tag_list_lock);
3540 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3541 mutex_unlock(&set->tag_list_lock);
3543 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3545 /* Enable polling stats and return whether they were already enabled. */
3546 static bool blk_poll_stats_enable(struct request_queue *q)
3548 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3549 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3551 blk_stat_add_callback(q, q->poll_cb);
3555 static void blk_mq_poll_stats_start(struct request_queue *q)
3558 * We don't arm the callback if polling stats are not enabled or the
3559 * callback is already active.
3561 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3562 blk_stat_is_active(q->poll_cb))
3565 blk_stat_activate_msecs(q->poll_cb, 100);
3568 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3570 struct request_queue *q = cb->data;
3573 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3574 if (cb->stat[bucket].nr_samples)
3575 q->poll_stat[bucket] = cb->stat[bucket];
3579 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3582 unsigned long ret = 0;
3586 * If stats collection isn't on, don't sleep but turn it on for
3589 if (!blk_poll_stats_enable(q))
3593 * As an optimistic guess, use half of the mean service time
3594 * for this type of request. We can (and should) make this smarter.
3595 * For instance, if the completion latencies are tight, we can
3596 * get closer than just half the mean. This is especially
3597 * important on devices where the completion latencies are longer
3598 * than ~10 usec. We do use the stats for the relevant IO size
3599 * if available which does lead to better estimates.
3601 bucket = blk_mq_poll_stats_bkt(rq);
3605 if (q->poll_stat[bucket].nr_samples)
3606 ret = (q->poll_stat[bucket].mean + 1) / 2;
3611 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3614 struct hrtimer_sleeper hs;
3615 enum hrtimer_mode mode;
3619 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3623 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3625 * 0: use half of prev avg
3626 * >0: use this specific value
3628 if (q->poll_nsec > 0)
3629 nsecs = q->poll_nsec;
3631 nsecs = blk_mq_poll_nsecs(q, rq);
3636 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3639 * This will be replaced with the stats tracking code, using
3640 * 'avg_completion_time / 2' as the pre-sleep target.
3644 mode = HRTIMER_MODE_REL;
3645 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3646 hrtimer_set_expires(&hs.timer, kt);
3649 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3651 set_current_state(TASK_UNINTERRUPTIBLE);
3652 hrtimer_sleeper_start_expires(&hs, mode);
3655 hrtimer_cancel(&hs.timer);
3656 mode = HRTIMER_MODE_ABS;
3657 } while (hs.task && !signal_pending(current));
3659 __set_current_state(TASK_RUNNING);
3660 destroy_hrtimer_on_stack(&hs.timer);
3664 static bool blk_mq_poll_hybrid(struct request_queue *q,
3665 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3669 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3672 if (!blk_qc_t_is_internal(cookie))
3673 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3675 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3677 * With scheduling, if the request has completed, we'll
3678 * get a NULL return here, as we clear the sched tag when
3679 * that happens. The request still remains valid, like always,
3680 * so we should be safe with just the NULL check.
3686 return blk_mq_poll_hybrid_sleep(q, rq);
3690 * blk_poll - poll for IO completions
3692 * @cookie: cookie passed back at IO submission time
3693 * @spin: whether to spin for completions
3696 * Poll for completions on the passed in queue. Returns number of
3697 * completed entries found. If @spin is true, then blk_poll will continue
3698 * looping until at least one completion is found, unless the task is
3699 * otherwise marked running (or we need to reschedule).
3701 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3703 struct blk_mq_hw_ctx *hctx;
3706 if (!blk_qc_t_valid(cookie) ||
3707 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3711 blk_flush_plug_list(current->plug, false);
3713 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3716 * If we sleep, have the caller restart the poll loop to reset
3717 * the state. Like for the other success return cases, the
3718 * caller is responsible for checking if the IO completed. If
3719 * the IO isn't complete, we'll get called again and will go
3720 * straight to the busy poll loop.
3722 if (blk_mq_poll_hybrid(q, hctx, cookie))
3725 hctx->poll_considered++;
3727 state = current->state;
3731 hctx->poll_invoked++;
3733 ret = q->mq_ops->poll(hctx);
3735 hctx->poll_success++;
3736 __set_current_state(TASK_RUNNING);
3740 if (signal_pending_state(state, current))
3741 __set_current_state(TASK_RUNNING);
3743 if (current->state == TASK_RUNNING)
3745 if (ret < 0 || !spin)
3748 } while (!need_resched());
3750 __set_current_state(TASK_RUNNING);
3753 EXPORT_SYMBOL_GPL(blk_poll);
3755 unsigned int blk_mq_rq_cpu(struct request *rq)
3757 return rq->mq_ctx->cpu;
3759 EXPORT_SYMBOL(blk_mq_rq_cpu);
3761 static int __init blk_mq_init(void)
3763 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3764 blk_mq_hctx_notify_dead);
3765 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3766 blk_mq_hctx_notify_online,
3767 blk_mq_hctx_notify_offline);
3770 subsys_initcall(blk_mq_init);