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 DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
46 static void blk_mq_poll_stats_start(struct request_queue *q);
47 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
49 static int blk_mq_poll_stats_bkt(const struct request *rq)
51 int ddir, sectors, bucket;
53 ddir = rq_data_dir(rq);
54 sectors = blk_rq_stats_sectors(rq);
56 bucket = ddir + 2 * ilog2(sectors);
60 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
61 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
67 * Check if any of the ctx, dispatch list or elevator
68 * have pending work in this hardware queue.
70 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
72 return !list_empty_careful(&hctx->dispatch) ||
73 sbitmap_any_bit_set(&hctx->ctx_map) ||
74 blk_mq_sched_has_work(hctx);
78 * Mark this ctx as having pending work in this hardware queue
80 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
81 struct blk_mq_ctx *ctx)
83 const int bit = ctx->index_hw[hctx->type];
85 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
86 sbitmap_set_bit(&hctx->ctx_map, bit);
89 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
90 struct blk_mq_ctx *ctx)
92 const int bit = ctx->index_hw[hctx->type];
94 sbitmap_clear_bit(&hctx->ctx_map, bit);
98 struct block_device *part;
99 unsigned int inflight[2];
102 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
103 struct request *rq, void *priv,
106 struct mq_inflight *mi = priv;
108 if ((!mi->part->bd_partno || rq->part == mi->part) &&
109 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
110 mi->inflight[rq_data_dir(rq)]++;
115 unsigned int blk_mq_in_flight(struct request_queue *q,
116 struct block_device *part)
118 struct mq_inflight mi = { .part = part };
120 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
122 return mi.inflight[0] + mi.inflight[1];
125 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
126 unsigned int inflight[2])
128 struct mq_inflight mi = { .part = part };
130 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
131 inflight[0] = mi.inflight[0];
132 inflight[1] = mi.inflight[1];
135 void blk_freeze_queue_start(struct request_queue *q)
137 mutex_lock(&q->mq_freeze_lock);
138 if (++q->mq_freeze_depth == 1) {
139 percpu_ref_kill(&q->q_usage_counter);
140 mutex_unlock(&q->mq_freeze_lock);
142 blk_mq_run_hw_queues(q, false);
144 mutex_unlock(&q->mq_freeze_lock);
147 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
149 void blk_mq_freeze_queue_wait(struct request_queue *q)
151 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
153 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
155 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
156 unsigned long timeout)
158 return wait_event_timeout(q->mq_freeze_wq,
159 percpu_ref_is_zero(&q->q_usage_counter),
162 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
165 * Guarantee no request is in use, so we can change any data structure of
166 * the queue afterward.
168 void blk_freeze_queue(struct request_queue *q)
171 * In the !blk_mq case we are only calling this to kill the
172 * q_usage_counter, otherwise this increases the freeze depth
173 * and waits for it to return to zero. For this reason there is
174 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
175 * exported to drivers as the only user for unfreeze is blk_mq.
177 blk_freeze_queue_start(q);
178 blk_mq_freeze_queue_wait(q);
181 void blk_mq_freeze_queue(struct request_queue *q)
184 * ...just an alias to keep freeze and unfreeze actions balanced
185 * in the blk_mq_* namespace
189 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
191 void blk_mq_unfreeze_queue(struct request_queue *q)
193 mutex_lock(&q->mq_freeze_lock);
194 q->mq_freeze_depth--;
195 WARN_ON_ONCE(q->mq_freeze_depth < 0);
196 if (!q->mq_freeze_depth) {
197 percpu_ref_resurrect(&q->q_usage_counter);
198 wake_up_all(&q->mq_freeze_wq);
200 mutex_unlock(&q->mq_freeze_lock);
202 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
205 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
206 * mpt3sas driver such that this function can be removed.
208 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
210 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
212 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
215 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
218 * Note: this function does not prevent that the struct request end_io()
219 * callback function is invoked. Once this function is returned, we make
220 * sure no dispatch can happen until the queue is unquiesced via
221 * blk_mq_unquiesce_queue().
223 void blk_mq_quiesce_queue(struct request_queue *q)
225 struct blk_mq_hw_ctx *hctx;
229 blk_mq_quiesce_queue_nowait(q);
231 queue_for_each_hw_ctx(q, hctx, i) {
232 if (hctx->flags & BLK_MQ_F_BLOCKING)
233 synchronize_srcu(hctx->srcu);
240 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
243 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
246 * This function recovers queue into the state before quiescing
247 * which is done by blk_mq_quiesce_queue.
249 void blk_mq_unquiesce_queue(struct request_queue *q)
251 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
253 /* dispatch requests which are inserted during quiescing */
254 blk_mq_run_hw_queues(q, true);
256 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
258 void blk_mq_wake_waiters(struct request_queue *q)
260 struct blk_mq_hw_ctx *hctx;
263 queue_for_each_hw_ctx(q, hctx, i)
264 if (blk_mq_hw_queue_mapped(hctx))
265 blk_mq_tag_wakeup_all(hctx->tags, true);
269 * Only need start/end time stamping if we have iostat or
270 * blk stats enabled, or using an IO scheduler.
272 static inline bool blk_mq_need_time_stamp(struct request *rq)
274 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
277 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
278 unsigned int tag, u64 alloc_time_ns)
280 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
281 struct request *rq = tags->static_rqs[tag];
283 if (data->q->elevator) {
284 rq->tag = BLK_MQ_NO_TAG;
285 rq->internal_tag = tag;
288 rq->internal_tag = BLK_MQ_NO_TAG;
291 /* csd/requeue_work/fifo_time is initialized before use */
293 rq->mq_ctx = data->ctx;
294 rq->mq_hctx = data->hctx;
296 rq->cmd_flags = data->cmd_flags;
297 if (data->flags & BLK_MQ_REQ_PM)
298 rq->rq_flags |= RQF_PM;
299 if (blk_queue_io_stat(data->q))
300 rq->rq_flags |= RQF_IO_STAT;
301 INIT_LIST_HEAD(&rq->queuelist);
302 INIT_HLIST_NODE(&rq->hash);
303 RB_CLEAR_NODE(&rq->rb_node);
306 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
307 rq->alloc_time_ns = alloc_time_ns;
309 if (blk_mq_need_time_stamp(rq))
310 rq->start_time_ns = ktime_get_ns();
312 rq->start_time_ns = 0;
313 rq->io_start_time_ns = 0;
314 rq->stats_sectors = 0;
315 rq->nr_phys_segments = 0;
316 #if defined(CONFIG_BLK_DEV_INTEGRITY)
317 rq->nr_integrity_segments = 0;
319 blk_crypto_rq_set_defaults(rq);
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(data->cmd_flags)]++;
329 refcount_set(&rq->ref, 1);
331 if (!op_is_flush(data->cmd_flags)) {
332 struct elevator_queue *e = data->q->elevator;
335 if (e && e->type->ops.prepare_request) {
336 if (e->type->icq_cache)
337 blk_mq_sched_assign_ioc(rq);
339 e->type->ops.prepare_request(rq);
340 rq->rq_flags |= RQF_ELVPRIV;
344 data->hctx->queued++;
348 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
350 struct request_queue *q = data->q;
351 struct elevator_queue *e = q->elevator;
352 u64 alloc_time_ns = 0;
355 /* alloc_time includes depth and tag waits */
356 if (blk_queue_rq_alloc_time(q))
357 alloc_time_ns = ktime_get_ns();
359 if (data->cmd_flags & REQ_NOWAIT)
360 data->flags |= BLK_MQ_REQ_NOWAIT;
364 * Flush/passthrough 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 !blk_op_is_passthrough(data->cmd_flags) &&
370 e->type->ops.limit_depth &&
371 !(data->flags & BLK_MQ_REQ_RESERVED))
372 e->type->ops.limit_depth(data->cmd_flags, data);
376 data->ctx = blk_mq_get_ctx(q);
377 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
379 blk_mq_tag_busy(data->hctx);
382 * Waiting allocations only fail because of an inactive hctx. In that
383 * case just retry the hctx assignment and tag allocation as CPU hotplug
384 * should have migrated us to an online CPU by now.
386 tag = blk_mq_get_tag(data);
387 if (tag == BLK_MQ_NO_TAG) {
388 if (data->flags & BLK_MQ_REQ_NOWAIT)
392 * Give up the CPU and sleep for a random short time to ensure
393 * that thread using a realtime scheduling class are migrated
394 * off the CPU, and thus off the hctx that is going away.
399 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
402 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
403 blk_mq_req_flags_t flags)
405 struct blk_mq_alloc_data data = {
413 ret = blk_queue_enter(q, flags);
417 rq = __blk_mq_alloc_request(&data);
421 rq->__sector = (sector_t) -1;
422 rq->bio = rq->biotail = NULL;
426 return ERR_PTR(-EWOULDBLOCK);
428 EXPORT_SYMBOL(blk_mq_alloc_request);
430 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
431 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
433 struct blk_mq_alloc_data data = {
438 u64 alloc_time_ns = 0;
443 /* alloc_time includes depth and tag waits */
444 if (blk_queue_rq_alloc_time(q))
445 alloc_time_ns = ktime_get_ns();
448 * If the tag allocator sleeps we could get an allocation for a
449 * different hardware context. No need to complicate the low level
450 * allocator for this for the rare use case of a command tied to
453 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
454 return ERR_PTR(-EINVAL);
456 if (hctx_idx >= q->nr_hw_queues)
457 return ERR_PTR(-EIO);
459 ret = blk_queue_enter(q, flags);
464 * Check if the hardware context is actually mapped to anything.
465 * If not tell the caller that it should skip this queue.
468 data.hctx = q->queue_hw_ctx[hctx_idx];
469 if (!blk_mq_hw_queue_mapped(data.hctx))
471 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
472 data.ctx = __blk_mq_get_ctx(q, cpu);
475 blk_mq_tag_busy(data.hctx);
478 tag = blk_mq_get_tag(&data);
479 if (tag == BLK_MQ_NO_TAG)
481 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
487 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
489 static void __blk_mq_free_request(struct request *rq)
491 struct request_queue *q = rq->q;
492 struct blk_mq_ctx *ctx = rq->mq_ctx;
493 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
494 const int sched_tag = rq->internal_tag;
496 blk_crypto_free_request(rq);
497 blk_pm_mark_last_busy(rq);
499 if (rq->tag != BLK_MQ_NO_TAG)
500 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
501 if (sched_tag != BLK_MQ_NO_TAG)
502 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
503 blk_mq_sched_restart(hctx);
507 void blk_mq_free_request(struct request *rq)
509 struct request_queue *q = rq->q;
510 struct elevator_queue *e = q->elevator;
511 struct blk_mq_ctx *ctx = rq->mq_ctx;
512 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
514 if (rq->rq_flags & RQF_ELVPRIV) {
515 if (e && e->type->ops.finish_request)
516 e->type->ops.finish_request(rq);
518 put_io_context(rq->elv.icq->ioc);
523 ctx->rq_completed[rq_is_sync(rq)]++;
524 if (rq->rq_flags & RQF_MQ_INFLIGHT)
525 __blk_mq_dec_active_requests(hctx);
527 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
528 laptop_io_completion(q->backing_dev_info);
532 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
533 if (refcount_dec_and_test(&rq->ref))
534 __blk_mq_free_request(rq);
536 EXPORT_SYMBOL_GPL(blk_mq_free_request);
538 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
542 if (blk_mq_need_time_stamp(rq))
543 now = ktime_get_ns();
545 if (rq->rq_flags & RQF_STATS) {
546 blk_mq_poll_stats_start(rq->q);
547 blk_stat_add(rq, now);
550 blk_mq_sched_completed_request(rq, now);
552 blk_account_io_done(rq, now);
555 rq_qos_done(rq->q, rq);
556 rq->end_io(rq, error);
558 blk_mq_free_request(rq);
561 EXPORT_SYMBOL(__blk_mq_end_request);
563 void blk_mq_end_request(struct request *rq, blk_status_t error)
565 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
567 __blk_mq_end_request(rq, error);
569 EXPORT_SYMBOL(blk_mq_end_request);
571 static void blk_complete_reqs(struct llist_head *list)
573 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
574 struct request *rq, *next;
576 llist_for_each_entry_safe(rq, next, entry, ipi_list)
577 rq->q->mq_ops->complete(rq);
580 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
582 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
585 static int blk_softirq_cpu_dead(unsigned int cpu)
587 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
591 static void __blk_mq_complete_request_remote(void *data)
593 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
596 static inline bool blk_mq_complete_need_ipi(struct request *rq)
598 int cpu = raw_smp_processor_id();
600 if (!IS_ENABLED(CONFIG_SMP) ||
601 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
604 * With force threaded interrupts enabled, raising softirq from an SMP
605 * function call will always result in waking the ksoftirqd thread.
606 * This is probably worse than completing the request on a different
609 if (force_irqthreads)
612 /* same CPU or cache domain? Complete locally */
613 if (cpu == rq->mq_ctx->cpu ||
614 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
615 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
618 /* don't try to IPI to an offline CPU */
619 return cpu_online(rq->mq_ctx->cpu);
622 static void blk_mq_complete_send_ipi(struct request *rq)
624 struct llist_head *list;
627 cpu = rq->mq_ctx->cpu;
628 list = &per_cpu(blk_cpu_done, cpu);
629 if (llist_add(&rq->ipi_list, list)) {
630 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
631 smp_call_function_single_async(cpu, &rq->csd);
635 static void blk_mq_raise_softirq(struct request *rq)
637 struct llist_head *list;
640 list = this_cpu_ptr(&blk_cpu_done);
641 if (llist_add(&rq->ipi_list, list))
642 raise_softirq(BLOCK_SOFTIRQ);
646 bool blk_mq_complete_request_remote(struct request *rq)
648 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
651 * For a polled request, always complete locallly, it's pointless
652 * to redirect the completion.
654 if (rq->cmd_flags & REQ_HIPRI)
657 if (blk_mq_complete_need_ipi(rq)) {
658 blk_mq_complete_send_ipi(rq);
662 if (rq->q->nr_hw_queues == 1) {
663 blk_mq_raise_softirq(rq);
668 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
671 * blk_mq_complete_request - end I/O on a request
672 * @rq: the request being processed
675 * Complete a request by scheduling the ->complete_rq operation.
677 void blk_mq_complete_request(struct request *rq)
679 if (!blk_mq_complete_request_remote(rq))
680 rq->q->mq_ops->complete(rq);
682 EXPORT_SYMBOL(blk_mq_complete_request);
684 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
685 __releases(hctx->srcu)
687 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
690 srcu_read_unlock(hctx->srcu, srcu_idx);
693 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
694 __acquires(hctx->srcu)
696 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
697 /* shut up gcc false positive */
701 *srcu_idx = srcu_read_lock(hctx->srcu);
705 * blk_mq_start_request - Start processing a request
706 * @rq: Pointer to request to be started
708 * Function used by device drivers to notify the block layer that a request
709 * is going to be processed now, so blk layer can do proper initializations
710 * such as starting the timeout timer.
712 void blk_mq_start_request(struct request *rq)
714 struct request_queue *q = rq->q;
716 trace_block_rq_issue(rq);
718 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
719 rq->io_start_time_ns = ktime_get_ns();
720 rq->stats_sectors = blk_rq_sectors(rq);
721 rq->rq_flags |= RQF_STATS;
725 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
728 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
730 #ifdef CONFIG_BLK_DEV_INTEGRITY
731 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
732 q->integrity.profile->prepare_fn(rq);
735 EXPORT_SYMBOL(blk_mq_start_request);
737 static void __blk_mq_requeue_request(struct request *rq)
739 struct request_queue *q = rq->q;
741 blk_mq_put_driver_tag(rq);
743 trace_block_rq_requeue(rq);
744 rq_qos_requeue(q, rq);
746 if (blk_mq_request_started(rq)) {
747 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
748 rq->rq_flags &= ~RQF_TIMED_OUT;
752 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
754 __blk_mq_requeue_request(rq);
756 /* this request will be re-inserted to io scheduler queue */
757 blk_mq_sched_requeue_request(rq);
759 BUG_ON(!list_empty(&rq->queuelist));
760 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
762 EXPORT_SYMBOL(blk_mq_requeue_request);
764 static void blk_mq_requeue_work(struct work_struct *work)
766 struct request_queue *q =
767 container_of(work, struct request_queue, requeue_work.work);
769 struct request *rq, *next;
771 spin_lock_irq(&q->requeue_lock);
772 list_splice_init(&q->requeue_list, &rq_list);
773 spin_unlock_irq(&q->requeue_lock);
775 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
776 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
779 rq->rq_flags &= ~RQF_SOFTBARRIER;
780 list_del_init(&rq->queuelist);
782 * If RQF_DONTPREP, rq has contained some driver specific
783 * data, so insert it to hctx dispatch list to avoid any
786 if (rq->rq_flags & RQF_DONTPREP)
787 blk_mq_request_bypass_insert(rq, false, false);
789 blk_mq_sched_insert_request(rq, true, false, false);
792 while (!list_empty(&rq_list)) {
793 rq = list_entry(rq_list.next, struct request, queuelist);
794 list_del_init(&rq->queuelist);
795 blk_mq_sched_insert_request(rq, false, false, false);
798 blk_mq_run_hw_queues(q, false);
801 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
802 bool kick_requeue_list)
804 struct request_queue *q = rq->q;
808 * We abuse this flag that is otherwise used by the I/O scheduler to
809 * request head insertion from the workqueue.
811 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
813 spin_lock_irqsave(&q->requeue_lock, flags);
815 rq->rq_flags |= RQF_SOFTBARRIER;
816 list_add(&rq->queuelist, &q->requeue_list);
818 list_add_tail(&rq->queuelist, &q->requeue_list);
820 spin_unlock_irqrestore(&q->requeue_lock, flags);
822 if (kick_requeue_list)
823 blk_mq_kick_requeue_list(q);
826 void blk_mq_kick_requeue_list(struct request_queue *q)
828 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
830 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
832 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
835 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
836 msecs_to_jiffies(msecs));
838 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
840 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
842 if (tag < tags->nr_tags) {
843 prefetch(tags->rqs[tag]);
844 return tags->rqs[tag];
849 EXPORT_SYMBOL(blk_mq_tag_to_rq);
851 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
852 void *priv, bool reserved)
855 * If we find a request that isn't idle and the queue matches,
856 * we know the queue is busy. Return false to stop the iteration.
858 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
868 bool blk_mq_queue_inflight(struct request_queue *q)
872 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
875 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
877 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
879 req->rq_flags |= RQF_TIMED_OUT;
880 if (req->q->mq_ops->timeout) {
881 enum blk_eh_timer_return ret;
883 ret = req->q->mq_ops->timeout(req, reserved);
884 if (ret == BLK_EH_DONE)
886 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
892 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
894 unsigned long deadline;
896 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
898 if (rq->rq_flags & RQF_TIMED_OUT)
901 deadline = READ_ONCE(rq->deadline);
902 if (time_after_eq(jiffies, deadline))
907 else if (time_after(*next, deadline))
912 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
913 struct request *rq, void *priv, bool reserved)
915 unsigned long *next = priv;
918 * Just do a quick check if it is expired before locking the request in
919 * so we're not unnecessarilly synchronizing across CPUs.
921 if (!blk_mq_req_expired(rq, next))
925 * We have reason to believe the request may be expired. Take a
926 * reference on the request to lock this request lifetime into its
927 * currently allocated context to prevent it from being reallocated in
928 * the event the completion by-passes this timeout handler.
930 * If the reference was already released, then the driver beat the
931 * timeout handler to posting a natural completion.
933 if (!refcount_inc_not_zero(&rq->ref))
937 * The request is now locked and cannot be reallocated underneath the
938 * timeout handler's processing. Re-verify this exact request is truly
939 * expired; if it is not expired, then the request was completed and
940 * reallocated as a new request.
942 if (blk_mq_req_expired(rq, next))
943 blk_mq_rq_timed_out(rq, reserved);
945 if (is_flush_rq(rq, hctx))
947 else if (refcount_dec_and_test(&rq->ref))
948 __blk_mq_free_request(rq);
953 static void blk_mq_timeout_work(struct work_struct *work)
955 struct request_queue *q =
956 container_of(work, struct request_queue, timeout_work);
957 unsigned long next = 0;
958 struct blk_mq_hw_ctx *hctx;
961 /* A deadlock might occur if a request is stuck requiring a
962 * timeout at the same time a queue freeze is waiting
963 * completion, since the timeout code would not be able to
964 * acquire the queue reference here.
966 * That's why we don't use blk_queue_enter here; instead, we use
967 * percpu_ref_tryget directly, because we need to be able to
968 * obtain a reference even in the short window between the queue
969 * starting to freeze, by dropping the first reference in
970 * blk_freeze_queue_start, and the moment the last request is
971 * consumed, marked by the instant q_usage_counter reaches
974 if (!percpu_ref_tryget(&q->q_usage_counter))
977 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
980 mod_timer(&q->timeout, next);
983 * Request timeouts are handled as a forward rolling timer. If
984 * we end up here it means that no requests are pending and
985 * also that no request has been pending for a while. Mark
988 queue_for_each_hw_ctx(q, hctx, i) {
989 /* the hctx may be unmapped, so check it here */
990 if (blk_mq_hw_queue_mapped(hctx))
991 blk_mq_tag_idle(hctx);
997 struct flush_busy_ctx_data {
998 struct blk_mq_hw_ctx *hctx;
999 struct list_head *list;
1002 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1004 struct flush_busy_ctx_data *flush_data = data;
1005 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1006 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1007 enum hctx_type type = hctx->type;
1009 spin_lock(&ctx->lock);
1010 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1011 sbitmap_clear_bit(sb, bitnr);
1012 spin_unlock(&ctx->lock);
1017 * Process software queues that have been marked busy, splicing them
1018 * to the for-dispatch
1020 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1022 struct flush_busy_ctx_data data = {
1027 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1029 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1031 struct dispatch_rq_data {
1032 struct blk_mq_hw_ctx *hctx;
1036 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1039 struct dispatch_rq_data *dispatch_data = data;
1040 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1041 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1042 enum hctx_type type = hctx->type;
1044 spin_lock(&ctx->lock);
1045 if (!list_empty(&ctx->rq_lists[type])) {
1046 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1047 list_del_init(&dispatch_data->rq->queuelist);
1048 if (list_empty(&ctx->rq_lists[type]))
1049 sbitmap_clear_bit(sb, bitnr);
1051 spin_unlock(&ctx->lock);
1053 return !dispatch_data->rq;
1056 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1057 struct blk_mq_ctx *start)
1059 unsigned off = start ? start->index_hw[hctx->type] : 0;
1060 struct dispatch_rq_data data = {
1065 __sbitmap_for_each_set(&hctx->ctx_map, off,
1066 dispatch_rq_from_ctx, &data);
1071 static inline unsigned int queued_to_index(unsigned int queued)
1076 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1079 static bool __blk_mq_get_driver_tag(struct request *rq)
1081 struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags;
1082 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1085 blk_mq_tag_busy(rq->mq_hctx);
1087 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1088 bt = rq->mq_hctx->tags->breserved_tags;
1091 if (!hctx_may_queue(rq->mq_hctx, bt))
1095 tag = __sbitmap_queue_get(bt);
1096 if (tag == BLK_MQ_NO_TAG)
1099 rq->tag = tag + tag_offset;
1103 static bool blk_mq_get_driver_tag(struct request *rq)
1105 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1107 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1110 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1111 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1112 rq->rq_flags |= RQF_MQ_INFLIGHT;
1113 __blk_mq_inc_active_requests(hctx);
1115 hctx->tags->rqs[rq->tag] = rq;
1119 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1120 int flags, void *key)
1122 struct blk_mq_hw_ctx *hctx;
1124 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1126 spin_lock(&hctx->dispatch_wait_lock);
1127 if (!list_empty(&wait->entry)) {
1128 struct sbitmap_queue *sbq;
1130 list_del_init(&wait->entry);
1131 sbq = hctx->tags->bitmap_tags;
1132 atomic_dec(&sbq->ws_active);
1134 spin_unlock(&hctx->dispatch_wait_lock);
1136 blk_mq_run_hw_queue(hctx, true);
1141 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1142 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1143 * restart. For both cases, take care to check the condition again after
1144 * marking us as waiting.
1146 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1149 struct sbitmap_queue *sbq = hctx->tags->bitmap_tags;
1150 struct wait_queue_head *wq;
1151 wait_queue_entry_t *wait;
1154 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1155 blk_mq_sched_mark_restart_hctx(hctx);
1158 * It's possible that a tag was freed in the window between the
1159 * allocation failure and adding the hardware queue to the wait
1162 * Don't clear RESTART here, someone else could have set it.
1163 * At most this will cost an extra queue run.
1165 return blk_mq_get_driver_tag(rq);
1168 wait = &hctx->dispatch_wait;
1169 if (!list_empty_careful(&wait->entry))
1172 wq = &bt_wait_ptr(sbq, hctx)->wait;
1174 spin_lock_irq(&wq->lock);
1175 spin_lock(&hctx->dispatch_wait_lock);
1176 if (!list_empty(&wait->entry)) {
1177 spin_unlock(&hctx->dispatch_wait_lock);
1178 spin_unlock_irq(&wq->lock);
1182 atomic_inc(&sbq->ws_active);
1183 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1184 __add_wait_queue(wq, wait);
1187 * It's possible that a tag was freed in the window between the
1188 * allocation failure and adding the hardware queue to the wait
1191 ret = blk_mq_get_driver_tag(rq);
1193 spin_unlock(&hctx->dispatch_wait_lock);
1194 spin_unlock_irq(&wq->lock);
1199 * We got a tag, remove ourselves from the wait queue to ensure
1200 * someone else gets the wakeup.
1202 list_del_init(&wait->entry);
1203 atomic_dec(&sbq->ws_active);
1204 spin_unlock(&hctx->dispatch_wait_lock);
1205 spin_unlock_irq(&wq->lock);
1210 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1211 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1213 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1214 * - EWMA is one simple way to compute running average value
1215 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1216 * - take 4 as factor for avoiding to get too small(0) result, and this
1217 * factor doesn't matter because EWMA decreases exponentially
1219 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1223 if (hctx->queue->elevator)
1226 ewma = hctx->dispatch_busy;
1231 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1233 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1234 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1236 hctx->dispatch_busy = ewma;
1239 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1241 static void blk_mq_handle_dev_resource(struct request *rq,
1242 struct list_head *list)
1244 struct request *next =
1245 list_first_entry_or_null(list, struct request, queuelist);
1248 * If an I/O scheduler has been configured and we got a driver tag for
1249 * the next request already, free it.
1252 blk_mq_put_driver_tag(next);
1254 list_add(&rq->queuelist, list);
1255 __blk_mq_requeue_request(rq);
1258 static void blk_mq_handle_zone_resource(struct request *rq,
1259 struct list_head *zone_list)
1262 * If we end up here it is because we cannot dispatch a request to a
1263 * specific zone due to LLD level zone-write locking or other zone
1264 * related resource not being available. In this case, set the request
1265 * aside in zone_list for retrying it later.
1267 list_add(&rq->queuelist, zone_list);
1268 __blk_mq_requeue_request(rq);
1271 enum prep_dispatch {
1273 PREP_DISPATCH_NO_TAG,
1274 PREP_DISPATCH_NO_BUDGET,
1277 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1280 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1282 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1283 blk_mq_put_driver_tag(rq);
1284 return PREP_DISPATCH_NO_BUDGET;
1287 if (!blk_mq_get_driver_tag(rq)) {
1289 * The initial allocation attempt failed, so we need to
1290 * rerun the hardware queue when a tag is freed. The
1291 * waitqueue takes care of that. If the queue is run
1292 * before we add this entry back on the dispatch list,
1293 * we'll re-run it below.
1295 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1297 * All budgets not got from this function will be put
1298 * together during handling partial dispatch
1301 blk_mq_put_dispatch_budget(rq->q);
1302 return PREP_DISPATCH_NO_TAG;
1306 return PREP_DISPATCH_OK;
1309 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1310 static void blk_mq_release_budgets(struct request_queue *q,
1311 unsigned int nr_budgets)
1315 for (i = 0; i < nr_budgets; i++)
1316 blk_mq_put_dispatch_budget(q);
1320 * Returns true if we did some work AND can potentially do more.
1322 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1323 unsigned int nr_budgets)
1325 enum prep_dispatch prep;
1326 struct request_queue *q = hctx->queue;
1327 struct request *rq, *nxt;
1329 blk_status_t ret = BLK_STS_OK;
1330 LIST_HEAD(zone_list);
1332 if (list_empty(list))
1336 * Now process all the entries, sending them to the driver.
1338 errors = queued = 0;
1340 struct blk_mq_queue_data bd;
1342 rq = list_first_entry(list, struct request, queuelist);
1344 WARN_ON_ONCE(hctx != rq->mq_hctx);
1345 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1346 if (prep != PREP_DISPATCH_OK)
1349 list_del_init(&rq->queuelist);
1354 * Flag last if we have no more requests, or if we have more
1355 * but can't assign a driver tag to it.
1357 if (list_empty(list))
1360 nxt = list_first_entry(list, struct request, queuelist);
1361 bd.last = !blk_mq_get_driver_tag(nxt);
1365 * once the request is queued to lld, no need to cover the
1370 ret = q->mq_ops->queue_rq(hctx, &bd);
1375 case BLK_STS_RESOURCE:
1376 case BLK_STS_DEV_RESOURCE:
1377 blk_mq_handle_dev_resource(rq, list);
1379 case BLK_STS_ZONE_RESOURCE:
1381 * Move the request to zone_list and keep going through
1382 * the dispatch list to find more requests the drive can
1385 blk_mq_handle_zone_resource(rq, &zone_list);
1389 blk_mq_end_request(rq, ret);
1391 } while (!list_empty(list));
1393 if (!list_empty(&zone_list))
1394 list_splice_tail_init(&zone_list, list);
1396 hctx->dispatched[queued_to_index(queued)]++;
1398 /* If we didn't flush the entire list, we could have told the driver
1399 * there was more coming, but that turned out to be a lie.
1401 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1402 q->mq_ops->commit_rqs(hctx);
1404 * Any items that need requeuing? Stuff them into hctx->dispatch,
1405 * that is where we will continue on next queue run.
1407 if (!list_empty(list)) {
1409 /* For non-shared tags, the RESTART check will suffice */
1410 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1411 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1412 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1414 blk_mq_release_budgets(q, nr_budgets);
1416 spin_lock(&hctx->lock);
1417 list_splice_tail_init(list, &hctx->dispatch);
1418 spin_unlock(&hctx->lock);
1421 * Order adding requests to hctx->dispatch and checking
1422 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1423 * in blk_mq_sched_restart(). Avoid restart code path to
1424 * miss the new added requests to hctx->dispatch, meantime
1425 * SCHED_RESTART is observed here.
1430 * If SCHED_RESTART was set by the caller of this function and
1431 * it is no longer set that means that it was cleared by another
1432 * thread and hence that a queue rerun is needed.
1434 * If 'no_tag' is set, that means that we failed getting
1435 * a driver tag with an I/O scheduler attached. If our dispatch
1436 * waitqueue is no longer active, ensure that we run the queue
1437 * AFTER adding our entries back to the list.
1439 * If no I/O scheduler has been configured it is possible that
1440 * the hardware queue got stopped and restarted before requests
1441 * were pushed back onto the dispatch list. Rerun the queue to
1442 * avoid starvation. Notes:
1443 * - blk_mq_run_hw_queue() checks whether or not a queue has
1444 * been stopped before rerunning a queue.
1445 * - Some but not all block drivers stop a queue before
1446 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1449 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1450 * bit is set, run queue after a delay to avoid IO stalls
1451 * that could otherwise occur if the queue is idle. We'll do
1452 * similar if we couldn't get budget and SCHED_RESTART is set.
1454 needs_restart = blk_mq_sched_needs_restart(hctx);
1455 if (!needs_restart ||
1456 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1457 blk_mq_run_hw_queue(hctx, true);
1458 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1460 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1462 blk_mq_update_dispatch_busy(hctx, true);
1465 blk_mq_update_dispatch_busy(hctx, false);
1467 return (queued + errors) != 0;
1471 * __blk_mq_run_hw_queue - Run a hardware queue.
1472 * @hctx: Pointer to the hardware queue to run.
1474 * Send pending requests to the hardware.
1476 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1481 * We can't run the queue inline with ints disabled. Ensure that
1482 * we catch bad users of this early.
1484 WARN_ON_ONCE(in_interrupt());
1486 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1488 hctx_lock(hctx, &srcu_idx);
1489 blk_mq_sched_dispatch_requests(hctx);
1490 hctx_unlock(hctx, srcu_idx);
1493 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1495 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1497 if (cpu >= nr_cpu_ids)
1498 cpu = cpumask_first(hctx->cpumask);
1503 * It'd be great if the workqueue API had a way to pass
1504 * in a mask and had some smarts for more clever placement.
1505 * For now we just round-robin here, switching for every
1506 * BLK_MQ_CPU_WORK_BATCH queued items.
1508 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1511 int next_cpu = hctx->next_cpu;
1513 if (hctx->queue->nr_hw_queues == 1)
1514 return WORK_CPU_UNBOUND;
1516 if (--hctx->next_cpu_batch <= 0) {
1518 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1520 if (next_cpu >= nr_cpu_ids)
1521 next_cpu = blk_mq_first_mapped_cpu(hctx);
1522 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1526 * Do unbound schedule if we can't find a online CPU for this hctx,
1527 * and it should only happen in the path of handling CPU DEAD.
1529 if (!cpu_online(next_cpu)) {
1536 * Make sure to re-select CPU next time once after CPUs
1537 * in hctx->cpumask become online again.
1539 hctx->next_cpu = next_cpu;
1540 hctx->next_cpu_batch = 1;
1541 return WORK_CPU_UNBOUND;
1544 hctx->next_cpu = next_cpu;
1549 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1550 * @hctx: Pointer to the hardware queue to run.
1551 * @async: If we want to run the queue asynchronously.
1552 * @msecs: Milliseconds of delay to wait before running the queue.
1554 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1555 * with a delay of @msecs.
1557 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1558 unsigned long msecs)
1560 if (unlikely(blk_mq_hctx_stopped(hctx)))
1563 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1564 int cpu = get_cpu();
1565 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1566 __blk_mq_run_hw_queue(hctx);
1574 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1575 msecs_to_jiffies(msecs));
1579 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1580 * @hctx: Pointer to the hardware queue to run.
1581 * @msecs: Milliseconds of delay to wait before running the queue.
1583 * Run a hardware queue asynchronously with a delay of @msecs.
1585 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1587 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1589 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1592 * blk_mq_run_hw_queue - Start to run a hardware queue.
1593 * @hctx: Pointer to the hardware queue to run.
1594 * @async: If we want to run the queue asynchronously.
1596 * Check if the request queue is not in a quiesced state and if there are
1597 * pending requests to be sent. If this is true, run the queue to send requests
1600 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1606 * When queue is quiesced, we may be switching io scheduler, or
1607 * updating nr_hw_queues, or other things, and we can't run queue
1608 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1610 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1613 hctx_lock(hctx, &srcu_idx);
1614 need_run = !blk_queue_quiesced(hctx->queue) &&
1615 blk_mq_hctx_has_pending(hctx);
1616 hctx_unlock(hctx, srcu_idx);
1619 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1621 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1624 * Is the request queue handled by an IO scheduler that does not respect
1625 * hardware queues when dispatching?
1627 static bool blk_mq_has_sqsched(struct request_queue *q)
1629 struct elevator_queue *e = q->elevator;
1631 if (e && e->type->ops.dispatch_request &&
1632 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
1638 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1641 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
1643 struct blk_mq_hw_ctx *hctx;
1646 * If the IO scheduler does not respect hardware queues when
1647 * dispatching, we just don't bother with multiple HW queues and
1648 * dispatch from hctx for the current CPU since running multiple queues
1649 * just causes lock contention inside the scheduler and pointless cache
1652 hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
1653 raw_smp_processor_id());
1654 if (!blk_mq_hctx_stopped(hctx))
1660 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1661 * @q: Pointer to the request queue to run.
1662 * @async: If we want to run the queue asynchronously.
1664 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1666 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1670 if (blk_mq_has_sqsched(q))
1671 sq_hctx = blk_mq_get_sq_hctx(q);
1672 queue_for_each_hw_ctx(q, hctx, i) {
1673 if (blk_mq_hctx_stopped(hctx))
1676 * Dispatch from this hctx either if there's no hctx preferred
1677 * by IO scheduler or if it has requests that bypass the
1680 if (!sq_hctx || sq_hctx == hctx ||
1681 !list_empty_careful(&hctx->dispatch))
1682 blk_mq_run_hw_queue(hctx, async);
1685 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1688 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1689 * @q: Pointer to the request queue to run.
1690 * @msecs: Milliseconds of delay to wait before running the queues.
1692 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1694 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1698 if (blk_mq_has_sqsched(q))
1699 sq_hctx = blk_mq_get_sq_hctx(q);
1700 queue_for_each_hw_ctx(q, hctx, i) {
1701 if (blk_mq_hctx_stopped(hctx))
1704 * Dispatch from this hctx either if there's no hctx preferred
1705 * by IO scheduler or if it has requests that bypass the
1708 if (!sq_hctx || sq_hctx == hctx ||
1709 !list_empty_careful(&hctx->dispatch))
1710 blk_mq_delay_run_hw_queue(hctx, msecs);
1713 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1716 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1717 * @q: request queue.
1719 * The caller is responsible for serializing this function against
1720 * blk_mq_{start,stop}_hw_queue().
1722 bool blk_mq_queue_stopped(struct request_queue *q)
1724 struct blk_mq_hw_ctx *hctx;
1727 queue_for_each_hw_ctx(q, hctx, i)
1728 if (blk_mq_hctx_stopped(hctx))
1733 EXPORT_SYMBOL(blk_mq_queue_stopped);
1736 * This function is often used for pausing .queue_rq() by driver when
1737 * there isn't enough resource or some conditions aren't satisfied, and
1738 * BLK_STS_RESOURCE is usually returned.
1740 * We do not guarantee that dispatch can be drained or blocked
1741 * after blk_mq_stop_hw_queue() returns. Please use
1742 * blk_mq_quiesce_queue() for that requirement.
1744 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1746 cancel_delayed_work(&hctx->run_work);
1748 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1750 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1753 * This function is often used for pausing .queue_rq() by driver when
1754 * there isn't enough resource or some conditions aren't satisfied, and
1755 * BLK_STS_RESOURCE is usually returned.
1757 * We do not guarantee that dispatch can be drained or blocked
1758 * after blk_mq_stop_hw_queues() returns. Please use
1759 * blk_mq_quiesce_queue() for that requirement.
1761 void blk_mq_stop_hw_queues(struct request_queue *q)
1763 struct blk_mq_hw_ctx *hctx;
1766 queue_for_each_hw_ctx(q, hctx, i)
1767 blk_mq_stop_hw_queue(hctx);
1769 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1771 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1773 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1775 blk_mq_run_hw_queue(hctx, false);
1777 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1779 void blk_mq_start_hw_queues(struct request_queue *q)
1781 struct blk_mq_hw_ctx *hctx;
1784 queue_for_each_hw_ctx(q, hctx, i)
1785 blk_mq_start_hw_queue(hctx);
1787 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1789 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1791 if (!blk_mq_hctx_stopped(hctx))
1794 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1795 blk_mq_run_hw_queue(hctx, async);
1797 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1799 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1801 struct blk_mq_hw_ctx *hctx;
1804 queue_for_each_hw_ctx(q, hctx, i)
1805 blk_mq_start_stopped_hw_queue(hctx, async);
1807 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1809 static void blk_mq_run_work_fn(struct work_struct *work)
1811 struct blk_mq_hw_ctx *hctx;
1813 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1816 * If we are stopped, don't run the queue.
1818 if (blk_mq_hctx_stopped(hctx))
1821 __blk_mq_run_hw_queue(hctx);
1824 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1828 struct blk_mq_ctx *ctx = rq->mq_ctx;
1829 enum hctx_type type = hctx->type;
1831 lockdep_assert_held(&ctx->lock);
1833 trace_block_rq_insert(rq);
1836 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1838 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1841 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1844 struct blk_mq_ctx *ctx = rq->mq_ctx;
1846 lockdep_assert_held(&ctx->lock);
1848 __blk_mq_insert_req_list(hctx, rq, at_head);
1849 blk_mq_hctx_mark_pending(hctx, ctx);
1853 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1854 * @rq: Pointer to request to be inserted.
1855 * @at_head: true if the request should be inserted at the head of the list.
1856 * @run_queue: If we should run the hardware queue after inserting the request.
1858 * Should only be used carefully, when the caller knows we want to
1859 * bypass a potential IO scheduler on the target device.
1861 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1864 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1866 spin_lock(&hctx->lock);
1868 list_add(&rq->queuelist, &hctx->dispatch);
1870 list_add_tail(&rq->queuelist, &hctx->dispatch);
1871 spin_unlock(&hctx->lock);
1874 blk_mq_run_hw_queue(hctx, false);
1877 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1878 struct list_head *list)
1882 enum hctx_type type = hctx->type;
1885 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1888 list_for_each_entry(rq, list, queuelist) {
1889 BUG_ON(rq->mq_ctx != ctx);
1890 trace_block_rq_insert(rq);
1893 spin_lock(&ctx->lock);
1894 list_splice_tail_init(list, &ctx->rq_lists[type]);
1895 blk_mq_hctx_mark_pending(hctx, ctx);
1896 spin_unlock(&ctx->lock);
1899 static int plug_rq_cmp(void *priv, const struct list_head *a,
1900 const struct list_head *b)
1902 struct request *rqa = container_of(a, struct request, queuelist);
1903 struct request *rqb = container_of(b, struct request, queuelist);
1905 if (rqa->mq_ctx != rqb->mq_ctx)
1906 return rqa->mq_ctx > rqb->mq_ctx;
1907 if (rqa->mq_hctx != rqb->mq_hctx)
1908 return rqa->mq_hctx > rqb->mq_hctx;
1910 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1913 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1917 if (list_empty(&plug->mq_list))
1919 list_splice_init(&plug->mq_list, &list);
1921 if (plug->rq_count > 2 && plug->multiple_queues)
1922 list_sort(NULL, &list, plug_rq_cmp);
1927 struct list_head rq_list;
1928 struct request *rq, *head_rq = list_entry_rq(list.next);
1929 struct list_head *pos = &head_rq->queuelist; /* skip first */
1930 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1931 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1932 unsigned int depth = 1;
1934 list_for_each_continue(pos, &list) {
1935 rq = list_entry_rq(pos);
1937 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1942 list_cut_before(&rq_list, &list, pos);
1943 trace_block_unplug(head_rq->q, depth, !from_schedule);
1944 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1946 } while(!list_empty(&list));
1949 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1950 unsigned int nr_segs)
1954 if (bio->bi_opf & REQ_RAHEAD)
1955 rq->cmd_flags |= REQ_FAILFAST_MASK;
1957 rq->__sector = bio->bi_iter.bi_sector;
1958 rq->write_hint = bio->bi_write_hint;
1959 blk_rq_bio_prep(rq, bio, nr_segs);
1961 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1962 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1965 blk_account_io_start(rq);
1968 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1970 blk_qc_t *cookie, bool last)
1972 struct request_queue *q = rq->q;
1973 struct blk_mq_queue_data bd = {
1977 blk_qc_t new_cookie;
1980 new_cookie = request_to_qc_t(hctx, rq);
1983 * For OK queue, we are done. For error, caller may kill it.
1984 * Any other error (busy), just add it to our list as we
1985 * previously would have done.
1987 ret = q->mq_ops->queue_rq(hctx, &bd);
1990 blk_mq_update_dispatch_busy(hctx, false);
1991 *cookie = new_cookie;
1993 case BLK_STS_RESOURCE:
1994 case BLK_STS_DEV_RESOURCE:
1995 blk_mq_update_dispatch_busy(hctx, true);
1996 __blk_mq_requeue_request(rq);
1999 blk_mq_update_dispatch_busy(hctx, false);
2000 *cookie = BLK_QC_T_NONE;
2007 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2010 bool bypass_insert, bool last)
2012 struct request_queue *q = rq->q;
2013 bool run_queue = true;
2016 * RCU or SRCU read lock is needed before checking quiesced flag.
2018 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2019 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2020 * and avoid driver to try to dispatch again.
2022 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2024 bypass_insert = false;
2028 if (q->elevator && !bypass_insert)
2031 if (!blk_mq_get_dispatch_budget(q))
2034 if (!blk_mq_get_driver_tag(rq)) {
2035 blk_mq_put_dispatch_budget(q);
2039 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2042 return BLK_STS_RESOURCE;
2044 blk_mq_sched_insert_request(rq, false, run_queue, false);
2050 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2051 * @hctx: Pointer of the associated hardware queue.
2052 * @rq: Pointer to request to be sent.
2053 * @cookie: Request queue cookie.
2055 * If the device has enough resources to accept a new request now, send the
2056 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2057 * we can try send it another time in the future. Requests inserted at this
2058 * queue have higher priority.
2060 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2061 struct request *rq, blk_qc_t *cookie)
2066 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2068 hctx_lock(hctx, &srcu_idx);
2070 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2071 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2072 blk_mq_request_bypass_insert(rq, false, true);
2073 else if (ret != BLK_STS_OK)
2074 blk_mq_end_request(rq, ret);
2076 hctx_unlock(hctx, srcu_idx);
2079 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2083 blk_qc_t unused_cookie;
2084 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2086 hctx_lock(hctx, &srcu_idx);
2087 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2088 hctx_unlock(hctx, srcu_idx);
2093 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2094 struct list_head *list)
2099 while (!list_empty(list)) {
2101 struct request *rq = list_first_entry(list, struct request,
2104 list_del_init(&rq->queuelist);
2105 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2106 if (ret != BLK_STS_OK) {
2107 if (ret == BLK_STS_RESOURCE ||
2108 ret == BLK_STS_DEV_RESOURCE) {
2109 blk_mq_request_bypass_insert(rq, false,
2113 blk_mq_end_request(rq, ret);
2120 * If we didn't flush the entire list, we could have told
2121 * the driver there was more coming, but that turned out to
2124 if ((!list_empty(list) || errors) &&
2125 hctx->queue->mq_ops->commit_rqs && queued)
2126 hctx->queue->mq_ops->commit_rqs(hctx);
2129 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2131 list_add_tail(&rq->queuelist, &plug->mq_list);
2133 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2134 struct request *tmp;
2136 tmp = list_first_entry(&plug->mq_list, struct request,
2138 if (tmp->q != rq->q)
2139 plug->multiple_queues = true;
2144 * blk_mq_submit_bio - Create and send a request to block device.
2145 * @bio: Bio pointer.
2147 * Builds up a request structure from @q and @bio and send to the device. The
2148 * request may not be queued directly to hardware if:
2149 * * This request can be merged with another one
2150 * * We want to place request at plug queue for possible future merging
2151 * * There is an IO scheduler active at this queue
2153 * It will not queue the request if there is an error with the bio, or at the
2156 * Returns: Request queue cookie.
2158 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2160 struct request_queue *q = bio->bi_bdev->bd_disk->queue;
2161 const int is_sync = op_is_sync(bio->bi_opf);
2162 const int is_flush_fua = op_is_flush(bio->bi_opf);
2163 struct blk_mq_alloc_data data = {
2167 struct blk_plug *plug;
2168 struct request *same_queue_rq = NULL;
2169 unsigned int nr_segs;
2174 blk_queue_bounce(q, &bio);
2175 __blk_queue_split(&bio, &nr_segs);
2177 if (!bio_integrity_prep(bio))
2180 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2181 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2184 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2187 rq_qos_throttle(q, bio);
2189 hipri = bio->bi_opf & REQ_HIPRI;
2191 data.cmd_flags = bio->bi_opf;
2192 rq = __blk_mq_alloc_request(&data);
2193 if (unlikely(!rq)) {
2194 rq_qos_cleanup(q, bio);
2195 if (bio->bi_opf & REQ_NOWAIT)
2196 bio_wouldblock_error(bio);
2200 trace_block_getrq(bio);
2202 rq_qos_track(q, rq, bio);
2204 cookie = request_to_qc_t(data.hctx, rq);
2206 blk_mq_bio_to_request(rq, bio, nr_segs);
2208 ret = blk_crypto_init_request(rq);
2209 if (ret != BLK_STS_OK) {
2210 bio->bi_status = ret;
2212 blk_mq_free_request(rq);
2213 return BLK_QC_T_NONE;
2216 plug = blk_mq_plug(q, bio);
2217 if (unlikely(is_flush_fua)) {
2218 /* Bypass scheduler for flush requests */
2219 blk_insert_flush(rq);
2220 blk_mq_run_hw_queue(data.hctx, true);
2221 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2222 !blk_queue_nonrot(q))) {
2224 * Use plugging if we have a ->commit_rqs() hook as well, as
2225 * we know the driver uses bd->last in a smart fashion.
2227 * Use normal plugging if this disk is slow HDD, as sequential
2228 * IO may benefit a lot from plug merging.
2230 unsigned int request_count = plug->rq_count;
2231 struct request *last = NULL;
2234 trace_block_plug(q);
2236 last = list_entry_rq(plug->mq_list.prev);
2238 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2239 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2240 blk_flush_plug_list(plug, false);
2241 trace_block_plug(q);
2244 blk_add_rq_to_plug(plug, rq);
2245 } else if (q->elevator) {
2246 /* Insert the request at the IO scheduler queue */
2247 blk_mq_sched_insert_request(rq, false, true, true);
2248 } else if (plug && !blk_queue_nomerges(q)) {
2250 * We do limited plugging. If the bio can be merged, do that.
2251 * Otherwise the existing request in the plug list will be
2252 * issued. So the plug list will have one request at most
2253 * The plug list might get flushed before this. If that happens,
2254 * the plug list is empty, and same_queue_rq is invalid.
2256 if (list_empty(&plug->mq_list))
2257 same_queue_rq = NULL;
2258 if (same_queue_rq) {
2259 list_del_init(&same_queue_rq->queuelist);
2262 blk_add_rq_to_plug(plug, rq);
2263 trace_block_plug(q);
2265 if (same_queue_rq) {
2266 data.hctx = same_queue_rq->mq_hctx;
2267 trace_block_unplug(q, 1, true);
2268 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2271 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2272 !data.hctx->dispatch_busy) {
2274 * There is no scheduler and we can try to send directly
2277 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2280 blk_mq_sched_insert_request(rq, false, true, true);
2284 return BLK_QC_T_NONE;
2288 return BLK_QC_T_NONE;
2291 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2292 unsigned int hctx_idx)
2296 if (tags->rqs && set->ops->exit_request) {
2299 for (i = 0; i < tags->nr_tags; i++) {
2300 struct request *rq = tags->static_rqs[i];
2304 set->ops->exit_request(set, rq, hctx_idx);
2305 tags->static_rqs[i] = NULL;
2309 while (!list_empty(&tags->page_list)) {
2310 page = list_first_entry(&tags->page_list, struct page, lru);
2311 list_del_init(&page->lru);
2313 * Remove kmemleak object previously allocated in
2314 * blk_mq_alloc_rqs().
2316 kmemleak_free(page_address(page));
2317 __free_pages(page, page->private);
2321 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags)
2325 kfree(tags->static_rqs);
2326 tags->static_rqs = NULL;
2328 blk_mq_free_tags(tags, flags);
2331 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2332 unsigned int hctx_idx,
2333 unsigned int nr_tags,
2334 unsigned int reserved_tags,
2337 struct blk_mq_tags *tags;
2340 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2341 if (node == NUMA_NO_NODE)
2342 node = set->numa_node;
2344 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags);
2348 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2349 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2352 blk_mq_free_tags(tags, flags);
2356 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2357 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2359 if (!tags->static_rqs) {
2361 blk_mq_free_tags(tags, flags);
2368 static size_t order_to_size(unsigned int order)
2370 return (size_t)PAGE_SIZE << order;
2373 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2374 unsigned int hctx_idx, int node)
2378 if (set->ops->init_request) {
2379 ret = set->ops->init_request(set, rq, hctx_idx, node);
2384 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2388 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2389 unsigned int hctx_idx, unsigned int depth)
2391 unsigned int i, j, entries_per_page, max_order = 4;
2392 size_t rq_size, left;
2395 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2396 if (node == NUMA_NO_NODE)
2397 node = set->numa_node;
2399 INIT_LIST_HEAD(&tags->page_list);
2402 * rq_size is the size of the request plus driver payload, rounded
2403 * to the cacheline size
2405 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2407 left = rq_size * depth;
2409 for (i = 0; i < depth; ) {
2410 int this_order = max_order;
2415 while (this_order && left < order_to_size(this_order - 1))
2419 page = alloc_pages_node(node,
2420 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2426 if (order_to_size(this_order) < rq_size)
2433 page->private = this_order;
2434 list_add_tail(&page->lru, &tags->page_list);
2436 p = page_address(page);
2438 * Allow kmemleak to scan these pages as they contain pointers
2439 * to additional allocations like via ops->init_request().
2441 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2442 entries_per_page = order_to_size(this_order) / rq_size;
2443 to_do = min(entries_per_page, depth - i);
2444 left -= to_do * rq_size;
2445 for (j = 0; j < to_do; j++) {
2446 struct request *rq = p;
2448 tags->static_rqs[i] = rq;
2449 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2450 tags->static_rqs[i] = NULL;
2461 blk_mq_free_rqs(set, tags, hctx_idx);
2465 struct rq_iter_data {
2466 struct blk_mq_hw_ctx *hctx;
2470 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2472 struct rq_iter_data *iter_data = data;
2474 if (rq->mq_hctx != iter_data->hctx)
2476 iter_data->has_rq = true;
2480 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2482 struct blk_mq_tags *tags = hctx->sched_tags ?
2483 hctx->sched_tags : hctx->tags;
2484 struct rq_iter_data data = {
2488 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2492 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2493 struct blk_mq_hw_ctx *hctx)
2495 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2497 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2502 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2504 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2505 struct blk_mq_hw_ctx, cpuhp_online);
2507 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2508 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2512 * Prevent new request from being allocated on the current hctx.
2514 * The smp_mb__after_atomic() Pairs with the implied barrier in
2515 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2516 * seen once we return from the tag allocator.
2518 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2519 smp_mb__after_atomic();
2522 * Try to grab a reference to the queue and wait for any outstanding
2523 * requests. If we could not grab a reference the queue has been
2524 * frozen and there are no requests.
2526 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2527 while (blk_mq_hctx_has_requests(hctx))
2529 percpu_ref_put(&hctx->queue->q_usage_counter);
2535 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2537 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2538 struct blk_mq_hw_ctx, cpuhp_online);
2540 if (cpumask_test_cpu(cpu, hctx->cpumask))
2541 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2546 * 'cpu' is going away. splice any existing rq_list entries from this
2547 * software queue to the hw queue dispatch list, and ensure that it
2550 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2552 struct blk_mq_hw_ctx *hctx;
2553 struct blk_mq_ctx *ctx;
2555 enum hctx_type type;
2557 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2558 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2561 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2564 spin_lock(&ctx->lock);
2565 if (!list_empty(&ctx->rq_lists[type])) {
2566 list_splice_init(&ctx->rq_lists[type], &tmp);
2567 blk_mq_hctx_clear_pending(hctx, ctx);
2569 spin_unlock(&ctx->lock);
2571 if (list_empty(&tmp))
2574 spin_lock(&hctx->lock);
2575 list_splice_tail_init(&tmp, &hctx->dispatch);
2576 spin_unlock(&hctx->lock);
2578 blk_mq_run_hw_queue(hctx, true);
2582 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2584 if (!(hctx->flags & BLK_MQ_F_STACKING))
2585 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2586 &hctx->cpuhp_online);
2587 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2591 /* hctx->ctxs will be freed in queue's release handler */
2592 static void blk_mq_exit_hctx(struct request_queue *q,
2593 struct blk_mq_tag_set *set,
2594 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2596 if (blk_mq_hw_queue_mapped(hctx))
2597 blk_mq_tag_idle(hctx);
2599 if (set->ops->exit_request)
2600 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2602 if (set->ops->exit_hctx)
2603 set->ops->exit_hctx(hctx, hctx_idx);
2605 blk_mq_remove_cpuhp(hctx);
2607 spin_lock(&q->unused_hctx_lock);
2608 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2609 spin_unlock(&q->unused_hctx_lock);
2612 static void blk_mq_exit_hw_queues(struct request_queue *q,
2613 struct blk_mq_tag_set *set, int nr_queue)
2615 struct blk_mq_hw_ctx *hctx;
2618 queue_for_each_hw_ctx(q, hctx, i) {
2621 blk_mq_debugfs_unregister_hctx(hctx);
2622 blk_mq_exit_hctx(q, set, hctx, i);
2626 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2628 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2630 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2631 __alignof__(struct blk_mq_hw_ctx)) !=
2632 sizeof(struct blk_mq_hw_ctx));
2634 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2635 hw_ctx_size += sizeof(struct srcu_struct);
2640 static int blk_mq_init_hctx(struct request_queue *q,
2641 struct blk_mq_tag_set *set,
2642 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2644 hctx->queue_num = hctx_idx;
2646 if (!(hctx->flags & BLK_MQ_F_STACKING))
2647 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2648 &hctx->cpuhp_online);
2649 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2651 hctx->tags = set->tags[hctx_idx];
2653 if (set->ops->init_hctx &&
2654 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2655 goto unregister_cpu_notifier;
2657 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2663 if (set->ops->exit_hctx)
2664 set->ops->exit_hctx(hctx, hctx_idx);
2665 unregister_cpu_notifier:
2666 blk_mq_remove_cpuhp(hctx);
2670 static struct blk_mq_hw_ctx *
2671 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2674 struct blk_mq_hw_ctx *hctx;
2675 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2677 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2679 goto fail_alloc_hctx;
2681 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2684 atomic_set(&hctx->nr_active, 0);
2685 if (node == NUMA_NO_NODE)
2686 node = set->numa_node;
2687 hctx->numa_node = node;
2689 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2690 spin_lock_init(&hctx->lock);
2691 INIT_LIST_HEAD(&hctx->dispatch);
2693 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
2695 INIT_LIST_HEAD(&hctx->hctx_list);
2698 * Allocate space for all possible cpus to avoid allocation at
2701 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2706 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2711 spin_lock_init(&hctx->dispatch_wait_lock);
2712 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2713 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2715 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2719 if (hctx->flags & BLK_MQ_F_BLOCKING)
2720 init_srcu_struct(hctx->srcu);
2721 blk_mq_hctx_kobj_init(hctx);
2726 sbitmap_free(&hctx->ctx_map);
2730 free_cpumask_var(hctx->cpumask);
2737 static void blk_mq_init_cpu_queues(struct request_queue *q,
2738 unsigned int nr_hw_queues)
2740 struct blk_mq_tag_set *set = q->tag_set;
2743 for_each_possible_cpu(i) {
2744 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2745 struct blk_mq_hw_ctx *hctx;
2749 spin_lock_init(&__ctx->lock);
2750 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2751 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2756 * Set local node, IFF we have more than one hw queue. If
2757 * not, we remain on the home node of the device
2759 for (j = 0; j < set->nr_maps; j++) {
2760 hctx = blk_mq_map_queue_type(q, j, i);
2761 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2762 hctx->numa_node = cpu_to_node(i);
2767 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2770 unsigned int flags = set->flags;
2773 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2774 set->queue_depth, set->reserved_tags, flags);
2775 if (!set->tags[hctx_idx])
2778 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2783 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2784 set->tags[hctx_idx] = NULL;
2788 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2789 unsigned int hctx_idx)
2791 unsigned int flags = set->flags;
2793 if (set->tags && set->tags[hctx_idx]) {
2794 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2795 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2796 set->tags[hctx_idx] = NULL;
2800 static void blk_mq_map_swqueue(struct request_queue *q)
2802 unsigned int i, j, hctx_idx;
2803 struct blk_mq_hw_ctx *hctx;
2804 struct blk_mq_ctx *ctx;
2805 struct blk_mq_tag_set *set = q->tag_set;
2807 queue_for_each_hw_ctx(q, hctx, i) {
2808 cpumask_clear(hctx->cpumask);
2810 hctx->dispatch_from = NULL;
2814 * Map software to hardware queues.
2816 * If the cpu isn't present, the cpu is mapped to first hctx.
2818 for_each_possible_cpu(i) {
2820 ctx = per_cpu_ptr(q->queue_ctx, i);
2821 for (j = 0; j < set->nr_maps; j++) {
2822 if (!set->map[j].nr_queues) {
2823 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2824 HCTX_TYPE_DEFAULT, i);
2827 hctx_idx = set->map[j].mq_map[i];
2828 /* unmapped hw queue can be remapped after CPU topo changed */
2829 if (!set->tags[hctx_idx] &&
2830 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2832 * If tags initialization fail for some hctx,
2833 * that hctx won't be brought online. In this
2834 * case, remap the current ctx to hctx[0] which
2835 * is guaranteed to always have tags allocated
2837 set->map[j].mq_map[i] = 0;
2840 hctx = blk_mq_map_queue_type(q, j, i);
2841 ctx->hctxs[j] = hctx;
2843 * If the CPU is already set in the mask, then we've
2844 * mapped this one already. This can happen if
2845 * devices share queues across queue maps.
2847 if (cpumask_test_cpu(i, hctx->cpumask))
2850 cpumask_set_cpu(i, hctx->cpumask);
2852 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2853 hctx->ctxs[hctx->nr_ctx++] = ctx;
2856 * If the nr_ctx type overflows, we have exceeded the
2857 * amount of sw queues we can support.
2859 BUG_ON(!hctx->nr_ctx);
2862 for (; j < HCTX_MAX_TYPES; j++)
2863 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2864 HCTX_TYPE_DEFAULT, i);
2867 queue_for_each_hw_ctx(q, hctx, i) {
2869 * If no software queues are mapped to this hardware queue,
2870 * disable it and free the request entries.
2872 if (!hctx->nr_ctx) {
2873 /* Never unmap queue 0. We need it as a
2874 * fallback in case of a new remap fails
2877 if (i && set->tags[i])
2878 blk_mq_free_map_and_requests(set, i);
2884 hctx->tags = set->tags[i];
2885 WARN_ON(!hctx->tags);
2888 * Set the map size to the number of mapped software queues.
2889 * This is more accurate and more efficient than looping
2890 * over all possibly mapped software queues.
2892 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2895 * Initialize batch roundrobin counts
2897 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2898 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2903 * Caller needs to ensure that we're either frozen/quiesced, or that
2904 * the queue isn't live yet.
2906 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2908 struct blk_mq_hw_ctx *hctx;
2911 queue_for_each_hw_ctx(q, hctx, i) {
2913 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2915 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2919 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
2922 struct request_queue *q;
2924 lockdep_assert_held(&set->tag_list_lock);
2926 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2927 blk_mq_freeze_queue(q);
2928 queue_set_hctx_shared(q, shared);
2929 blk_mq_unfreeze_queue(q);
2933 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2935 struct blk_mq_tag_set *set = q->tag_set;
2937 mutex_lock(&set->tag_list_lock);
2938 list_del(&q->tag_set_list);
2939 if (list_is_singular(&set->tag_list)) {
2940 /* just transitioned to unshared */
2941 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2942 /* update existing queue */
2943 blk_mq_update_tag_set_shared(set, false);
2945 mutex_unlock(&set->tag_list_lock);
2946 INIT_LIST_HEAD(&q->tag_set_list);
2949 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2950 struct request_queue *q)
2952 mutex_lock(&set->tag_list_lock);
2955 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2957 if (!list_empty(&set->tag_list) &&
2958 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2959 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2960 /* update existing queue */
2961 blk_mq_update_tag_set_shared(set, true);
2963 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
2964 queue_set_hctx_shared(q, true);
2965 list_add_tail(&q->tag_set_list, &set->tag_list);
2967 mutex_unlock(&set->tag_list_lock);
2970 /* All allocations will be freed in release handler of q->mq_kobj */
2971 static int blk_mq_alloc_ctxs(struct request_queue *q)
2973 struct blk_mq_ctxs *ctxs;
2976 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2980 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2981 if (!ctxs->queue_ctx)
2984 for_each_possible_cpu(cpu) {
2985 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2989 q->mq_kobj = &ctxs->kobj;
2990 q->queue_ctx = ctxs->queue_ctx;
2999 * It is the actual release handler for mq, but we do it from
3000 * request queue's release handler for avoiding use-after-free
3001 * and headache because q->mq_kobj shouldn't have been introduced,
3002 * but we can't group ctx/kctx kobj without it.
3004 void blk_mq_release(struct request_queue *q)
3006 struct blk_mq_hw_ctx *hctx, *next;
3009 queue_for_each_hw_ctx(q, hctx, i)
3010 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3012 /* all hctx are in .unused_hctx_list now */
3013 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3014 list_del_init(&hctx->hctx_list);
3015 kobject_put(&hctx->kobj);
3018 kfree(q->queue_hw_ctx);
3021 * release .mq_kobj and sw queue's kobject now because
3022 * both share lifetime with request queue.
3024 blk_mq_sysfs_deinit(q);
3027 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3030 struct request_queue *uninit_q, *q;
3032 uninit_q = blk_alloc_queue(set->numa_node);
3034 return ERR_PTR(-ENOMEM);
3035 uninit_q->queuedata = queuedata;
3038 * Initialize the queue without an elevator. device_add_disk() will do
3039 * the initialization.
3041 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3043 blk_cleanup_queue(uninit_q);
3047 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3049 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3051 return blk_mq_init_queue_data(set, NULL);
3053 EXPORT_SYMBOL(blk_mq_init_queue);
3056 * Helper for setting up a queue with mq ops, given queue depth, and
3057 * the passed in mq ops flags.
3059 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3060 const struct blk_mq_ops *ops,
3061 unsigned int queue_depth,
3062 unsigned int set_flags)
3064 struct request_queue *q;
3067 memset(set, 0, sizeof(*set));
3069 set->nr_hw_queues = 1;
3071 set->queue_depth = queue_depth;
3072 set->numa_node = NUMA_NO_NODE;
3073 set->flags = set_flags;
3075 ret = blk_mq_alloc_tag_set(set);
3077 return ERR_PTR(ret);
3079 q = blk_mq_init_queue(set);
3081 blk_mq_free_tag_set(set);
3087 EXPORT_SYMBOL(blk_mq_init_sq_queue);
3089 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3090 struct blk_mq_tag_set *set, struct request_queue *q,
3091 int hctx_idx, int node)
3093 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3095 /* reuse dead hctx first */
3096 spin_lock(&q->unused_hctx_lock);
3097 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3098 if (tmp->numa_node == node) {
3104 list_del_init(&hctx->hctx_list);
3105 spin_unlock(&q->unused_hctx_lock);
3108 hctx = blk_mq_alloc_hctx(q, set, node);
3112 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3118 kobject_put(&hctx->kobj);
3123 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3124 struct request_queue *q)
3127 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3129 if (q->nr_hw_queues < set->nr_hw_queues) {
3130 struct blk_mq_hw_ctx **new_hctxs;
3132 new_hctxs = kcalloc_node(set->nr_hw_queues,
3133 sizeof(*new_hctxs), GFP_KERNEL,
3138 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3140 q->queue_hw_ctx = new_hctxs;
3145 /* protect against switching io scheduler */
3146 mutex_lock(&q->sysfs_lock);
3147 for (i = 0; i < set->nr_hw_queues; i++) {
3149 struct blk_mq_hw_ctx *hctx;
3151 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3153 * If the hw queue has been mapped to another numa node,
3154 * we need to realloc the hctx. If allocation fails, fallback
3155 * to use the previous one.
3157 if (hctxs[i] && (hctxs[i]->numa_node == node))
3160 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3163 blk_mq_exit_hctx(q, set, hctxs[i], i);
3167 pr_warn("Allocate new hctx on node %d fails,\
3168 fallback to previous one on node %d\n",
3169 node, hctxs[i]->numa_node);
3175 * Increasing nr_hw_queues fails. Free the newly allocated
3176 * hctxs and keep the previous q->nr_hw_queues.
3178 if (i != set->nr_hw_queues) {
3179 j = q->nr_hw_queues;
3183 end = q->nr_hw_queues;
3184 q->nr_hw_queues = set->nr_hw_queues;
3187 for (; j < end; j++) {
3188 struct blk_mq_hw_ctx *hctx = hctxs[j];
3192 blk_mq_free_map_and_requests(set, j);
3193 blk_mq_exit_hctx(q, set, hctx, j);
3197 mutex_unlock(&q->sysfs_lock);
3200 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3201 struct request_queue *q,
3204 /* mark the queue as mq asap */
3205 q->mq_ops = set->ops;
3207 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3208 blk_mq_poll_stats_bkt,
3209 BLK_MQ_POLL_STATS_BKTS, q);
3213 if (blk_mq_alloc_ctxs(q))
3216 /* init q->mq_kobj and sw queues' kobjects */
3217 blk_mq_sysfs_init(q);
3219 INIT_LIST_HEAD(&q->unused_hctx_list);
3220 spin_lock_init(&q->unused_hctx_lock);
3222 blk_mq_realloc_hw_ctxs(set, q);
3223 if (!q->nr_hw_queues)
3226 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3227 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3231 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3232 if (set->nr_maps > HCTX_TYPE_POLL &&
3233 set->map[HCTX_TYPE_POLL].nr_queues)
3234 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3236 q->sg_reserved_size = INT_MAX;
3238 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3239 INIT_LIST_HEAD(&q->requeue_list);
3240 spin_lock_init(&q->requeue_lock);
3242 q->nr_requests = set->queue_depth;
3245 * Default to classic polling
3247 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3249 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3250 blk_mq_add_queue_tag_set(set, q);
3251 blk_mq_map_swqueue(q);
3254 elevator_init_mq(q);
3259 kfree(q->queue_hw_ctx);
3260 q->nr_hw_queues = 0;
3261 blk_mq_sysfs_deinit(q);
3263 blk_stat_free_callback(q->poll_cb);
3267 return ERR_PTR(-ENOMEM);
3269 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3271 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3272 void blk_mq_exit_queue(struct request_queue *q)
3274 struct blk_mq_tag_set *set = q->tag_set;
3276 blk_mq_del_queue_tag_set(q);
3277 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3280 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3284 for (i = 0; i < set->nr_hw_queues; i++) {
3285 if (!__blk_mq_alloc_map_and_request(set, i))
3294 blk_mq_free_map_and_requests(set, i);
3300 * Allocate the request maps associated with this tag_set. Note that this
3301 * may reduce the depth asked for, if memory is tight. set->queue_depth
3302 * will be updated to reflect the allocated depth.
3304 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3309 depth = set->queue_depth;
3311 err = __blk_mq_alloc_rq_maps(set);
3315 set->queue_depth >>= 1;
3316 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3320 } while (set->queue_depth);
3322 if (!set->queue_depth || err) {
3323 pr_err("blk-mq: failed to allocate request map\n");
3327 if (depth != set->queue_depth)
3328 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3329 depth, set->queue_depth);
3334 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3337 * blk_mq_map_queues() and multiple .map_queues() implementations
3338 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3339 * number of hardware queues.
3341 if (set->nr_maps == 1)
3342 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3344 if (set->ops->map_queues && !is_kdump_kernel()) {
3348 * transport .map_queues is usually done in the following
3351 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3352 * mask = get_cpu_mask(queue)
3353 * for_each_cpu(cpu, mask)
3354 * set->map[x].mq_map[cpu] = queue;
3357 * When we need to remap, the table has to be cleared for
3358 * killing stale mapping since one CPU may not be mapped
3361 for (i = 0; i < set->nr_maps; i++)
3362 blk_mq_clear_mq_map(&set->map[i]);
3364 return set->ops->map_queues(set);
3366 BUG_ON(set->nr_maps > 1);
3367 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3371 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3372 int cur_nr_hw_queues, int new_nr_hw_queues)
3374 struct blk_mq_tags **new_tags;
3376 if (cur_nr_hw_queues >= new_nr_hw_queues)
3379 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3380 GFP_KERNEL, set->numa_node);
3385 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3386 sizeof(*set->tags));
3388 set->tags = new_tags;
3389 set->nr_hw_queues = new_nr_hw_queues;
3394 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
3395 int new_nr_hw_queues)
3397 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
3401 * Alloc a tag set to be associated with one or more request queues.
3402 * May fail with EINVAL for various error conditions. May adjust the
3403 * requested depth down, if it's too large. In that case, the set
3404 * value will be stored in set->queue_depth.
3406 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3410 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3412 if (!set->nr_hw_queues)
3414 if (!set->queue_depth)
3416 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3419 if (!set->ops->queue_rq)
3422 if (!set->ops->get_budget ^ !set->ops->put_budget)
3425 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3426 pr_info("blk-mq: reduced tag depth to %u\n",
3428 set->queue_depth = BLK_MQ_MAX_DEPTH;
3433 else if (set->nr_maps > HCTX_MAX_TYPES)
3437 * If a crashdump is active, then we are potentially in a very
3438 * memory constrained environment. Limit us to 1 queue and
3439 * 64 tags to prevent using too much memory.
3441 if (is_kdump_kernel()) {
3442 set->nr_hw_queues = 1;
3444 set->queue_depth = min(64U, set->queue_depth);
3447 * There is no use for more h/w queues than cpus if we just have
3450 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3451 set->nr_hw_queues = nr_cpu_ids;
3453 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
3457 for (i = 0; i < set->nr_maps; i++) {
3458 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3459 sizeof(set->map[i].mq_map[0]),
3460 GFP_KERNEL, set->numa_node);
3461 if (!set->map[i].mq_map)
3462 goto out_free_mq_map;
3463 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3466 ret = blk_mq_update_queue_map(set);
3468 goto out_free_mq_map;
3470 ret = blk_mq_alloc_map_and_requests(set);
3472 goto out_free_mq_map;
3474 if (blk_mq_is_sbitmap_shared(set->flags)) {
3475 atomic_set(&set->active_queues_shared_sbitmap, 0);
3477 if (blk_mq_init_shared_sbitmap(set, set->flags)) {
3479 goto out_free_mq_rq_maps;
3483 mutex_init(&set->tag_list_lock);
3484 INIT_LIST_HEAD(&set->tag_list);
3488 out_free_mq_rq_maps:
3489 for (i = 0; i < set->nr_hw_queues; i++)
3490 blk_mq_free_map_and_requests(set, i);
3492 for (i = 0; i < set->nr_maps; i++) {
3493 kfree(set->map[i].mq_map);
3494 set->map[i].mq_map = NULL;
3500 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3502 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3506 for (i = 0; i < set->nr_hw_queues; i++)
3507 blk_mq_free_map_and_requests(set, i);
3509 if (blk_mq_is_sbitmap_shared(set->flags))
3510 blk_mq_exit_shared_sbitmap(set);
3512 for (j = 0; j < set->nr_maps; j++) {
3513 kfree(set->map[j].mq_map);
3514 set->map[j].mq_map = NULL;
3520 EXPORT_SYMBOL(blk_mq_free_tag_set);
3522 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3524 struct blk_mq_tag_set *set = q->tag_set;
3525 struct blk_mq_hw_ctx *hctx;
3531 if (q->nr_requests == nr)
3534 blk_mq_freeze_queue(q);
3535 blk_mq_quiesce_queue(q);
3538 queue_for_each_hw_ctx(q, hctx, i) {
3542 * If we're using an MQ scheduler, just update the scheduler
3543 * queue depth. This is similar to what the old code would do.
3545 if (!hctx->sched_tags) {
3546 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3548 if (!ret && blk_mq_is_sbitmap_shared(set->flags))
3549 blk_mq_tag_resize_shared_sbitmap(set, nr);
3551 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3556 if (q->elevator && q->elevator->type->ops.depth_updated)
3557 q->elevator->type->ops.depth_updated(hctx);
3561 q->nr_requests = nr;
3563 blk_mq_unquiesce_queue(q);
3564 blk_mq_unfreeze_queue(q);
3570 * request_queue and elevator_type pair.
3571 * It is just used by __blk_mq_update_nr_hw_queues to cache
3572 * the elevator_type associated with a request_queue.
3574 struct blk_mq_qe_pair {
3575 struct list_head node;
3576 struct request_queue *q;
3577 struct elevator_type *type;
3581 * Cache the elevator_type in qe pair list and switch the
3582 * io scheduler to 'none'
3584 static bool blk_mq_elv_switch_none(struct list_head *head,
3585 struct request_queue *q)
3587 struct blk_mq_qe_pair *qe;
3592 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3596 INIT_LIST_HEAD(&qe->node);
3598 qe->type = q->elevator->type;
3599 list_add(&qe->node, head);
3601 mutex_lock(&q->sysfs_lock);
3603 * After elevator_switch_mq, the previous elevator_queue will be
3604 * released by elevator_release. The reference of the io scheduler
3605 * module get by elevator_get will also be put. So we need to get
3606 * a reference of the io scheduler module here to prevent it to be
3609 __module_get(qe->type->elevator_owner);
3610 elevator_switch_mq(q, NULL);
3611 mutex_unlock(&q->sysfs_lock);
3616 static void blk_mq_elv_switch_back(struct list_head *head,
3617 struct request_queue *q)
3619 struct blk_mq_qe_pair *qe;
3620 struct elevator_type *t = NULL;
3622 list_for_each_entry(qe, head, node)
3631 list_del(&qe->node);
3634 mutex_lock(&q->sysfs_lock);
3635 elevator_switch_mq(q, t);
3636 mutex_unlock(&q->sysfs_lock);
3639 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3642 struct request_queue *q;
3644 int prev_nr_hw_queues;
3646 lockdep_assert_held(&set->tag_list_lock);
3648 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3649 nr_hw_queues = nr_cpu_ids;
3650 if (nr_hw_queues < 1)
3652 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3655 list_for_each_entry(q, &set->tag_list, tag_set_list)
3656 blk_mq_freeze_queue(q);
3658 * Switch IO scheduler to 'none', cleaning up the data associated
3659 * with the previous scheduler. We will switch back once we are done
3660 * updating the new sw to hw queue mappings.
3662 list_for_each_entry(q, &set->tag_list, tag_set_list)
3663 if (!blk_mq_elv_switch_none(&head, q))
3666 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3667 blk_mq_debugfs_unregister_hctxs(q);
3668 blk_mq_sysfs_unregister(q);
3671 prev_nr_hw_queues = set->nr_hw_queues;
3672 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3676 set->nr_hw_queues = nr_hw_queues;
3678 blk_mq_update_queue_map(set);
3679 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3680 blk_mq_realloc_hw_ctxs(set, q);
3681 if (q->nr_hw_queues != set->nr_hw_queues) {
3682 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3683 nr_hw_queues, prev_nr_hw_queues);
3684 set->nr_hw_queues = prev_nr_hw_queues;
3685 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3688 blk_mq_map_swqueue(q);
3692 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3693 blk_mq_sysfs_register(q);
3694 blk_mq_debugfs_register_hctxs(q);
3698 list_for_each_entry(q, &set->tag_list, tag_set_list)
3699 blk_mq_elv_switch_back(&head, q);
3701 list_for_each_entry(q, &set->tag_list, tag_set_list)
3702 blk_mq_unfreeze_queue(q);
3705 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3707 mutex_lock(&set->tag_list_lock);
3708 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3709 mutex_unlock(&set->tag_list_lock);
3711 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3713 /* Enable polling stats and return whether they were already enabled. */
3714 static bool blk_poll_stats_enable(struct request_queue *q)
3716 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3717 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3719 blk_stat_add_callback(q, q->poll_cb);
3723 static void blk_mq_poll_stats_start(struct request_queue *q)
3726 * We don't arm the callback if polling stats are not enabled or the
3727 * callback is already active.
3729 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3730 blk_stat_is_active(q->poll_cb))
3733 blk_stat_activate_msecs(q->poll_cb, 100);
3736 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3738 struct request_queue *q = cb->data;
3741 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3742 if (cb->stat[bucket].nr_samples)
3743 q->poll_stat[bucket] = cb->stat[bucket];
3747 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3750 unsigned long ret = 0;
3754 * If stats collection isn't on, don't sleep but turn it on for
3757 if (!blk_poll_stats_enable(q))
3761 * As an optimistic guess, use half of the mean service time
3762 * for this type of request. We can (and should) make this smarter.
3763 * For instance, if the completion latencies are tight, we can
3764 * get closer than just half the mean. This is especially
3765 * important on devices where the completion latencies are longer
3766 * than ~10 usec. We do use the stats for the relevant IO size
3767 * if available which does lead to better estimates.
3769 bucket = blk_mq_poll_stats_bkt(rq);
3773 if (q->poll_stat[bucket].nr_samples)
3774 ret = (q->poll_stat[bucket].mean + 1) / 2;
3779 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3782 struct hrtimer_sleeper hs;
3783 enum hrtimer_mode mode;
3787 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3791 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3793 * 0: use half of prev avg
3794 * >0: use this specific value
3796 if (q->poll_nsec > 0)
3797 nsecs = q->poll_nsec;
3799 nsecs = blk_mq_poll_nsecs(q, rq);
3804 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3807 * This will be replaced with the stats tracking code, using
3808 * 'avg_completion_time / 2' as the pre-sleep target.
3812 mode = HRTIMER_MODE_REL;
3813 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3814 hrtimer_set_expires(&hs.timer, kt);
3817 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3819 set_current_state(TASK_UNINTERRUPTIBLE);
3820 hrtimer_sleeper_start_expires(&hs, mode);
3823 hrtimer_cancel(&hs.timer);
3824 mode = HRTIMER_MODE_ABS;
3825 } while (hs.task && !signal_pending(current));
3827 __set_current_state(TASK_RUNNING);
3828 destroy_hrtimer_on_stack(&hs.timer);
3832 static bool blk_mq_poll_hybrid(struct request_queue *q,
3833 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3837 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3840 if (!blk_qc_t_is_internal(cookie))
3841 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3843 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3845 * With scheduling, if the request has completed, we'll
3846 * get a NULL return here, as we clear the sched tag when
3847 * that happens. The request still remains valid, like always,
3848 * so we should be safe with just the NULL check.
3854 return blk_mq_poll_hybrid_sleep(q, rq);
3858 * blk_poll - poll for IO completions
3860 * @cookie: cookie passed back at IO submission time
3861 * @spin: whether to spin for completions
3864 * Poll for completions on the passed in queue. Returns number of
3865 * completed entries found. If @spin is true, then blk_poll will continue
3866 * looping until at least one completion is found, unless the task is
3867 * otherwise marked running (or we need to reschedule).
3869 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3871 struct blk_mq_hw_ctx *hctx;
3874 if (!blk_qc_t_valid(cookie) ||
3875 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3879 blk_flush_plug_list(current->plug, false);
3881 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3884 * If we sleep, have the caller restart the poll loop to reset
3885 * the state. Like for the other success return cases, the
3886 * caller is responsible for checking if the IO completed. If
3887 * the IO isn't complete, we'll get called again and will go
3888 * straight to the busy poll loop. If specified not to spin,
3889 * we also should not sleep.
3891 if (spin && blk_mq_poll_hybrid(q, hctx, cookie))
3894 hctx->poll_considered++;
3896 state = current->state;
3900 hctx->poll_invoked++;
3902 ret = q->mq_ops->poll(hctx);
3904 hctx->poll_success++;
3905 __set_current_state(TASK_RUNNING);
3909 if (signal_pending_state(state, current))
3910 __set_current_state(TASK_RUNNING);
3912 if (current->state == TASK_RUNNING)
3914 if (ret < 0 || !spin)
3917 } while (!need_resched());
3919 __set_current_state(TASK_RUNNING);
3922 EXPORT_SYMBOL_GPL(blk_poll);
3924 unsigned int blk_mq_rq_cpu(struct request *rq)
3926 return rq->mq_ctx->cpu;
3928 EXPORT_SYMBOL(blk_mq_rq_cpu);
3930 static int __init blk_mq_init(void)
3934 for_each_possible_cpu(i)
3935 init_llist_head(&per_cpu(blk_cpu_done, i));
3936 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3938 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3939 "block/softirq:dead", NULL,
3940 blk_softirq_cpu_dead);
3941 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3942 blk_mq_hctx_notify_dead);
3943 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3944 blk_mq_hctx_notify_online,
3945 blk_mq_hctx_notify_offline);
3948 subsys_initcall(blk_mq_init);