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 void blk_mq_put_rq_ref(struct request *rq)
914 if (is_flush_rq(rq, rq->mq_hctx))
916 else if (refcount_dec_and_test(&rq->ref))
917 __blk_mq_free_request(rq);
920 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
921 struct request *rq, void *priv, bool reserved)
923 unsigned long *next = priv;
926 * Just do a quick check if it is expired before locking the request in
927 * so we're not unnecessarilly synchronizing across CPUs.
929 if (!blk_mq_req_expired(rq, next))
933 * We have reason to believe the request may be expired. Take a
934 * reference on the request to lock this request lifetime into its
935 * currently allocated context to prevent it from being reallocated in
936 * the event the completion by-passes this timeout handler.
938 * If the reference was already released, then the driver beat the
939 * timeout handler to posting a natural completion.
941 if (!refcount_inc_not_zero(&rq->ref))
945 * The request is now locked and cannot be reallocated underneath the
946 * timeout handler's processing. Re-verify this exact request is truly
947 * expired; if it is not expired, then the request was completed and
948 * reallocated as a new request.
950 if (blk_mq_req_expired(rq, next))
951 blk_mq_rq_timed_out(rq, reserved);
953 blk_mq_put_rq_ref(rq);
957 static void blk_mq_timeout_work(struct work_struct *work)
959 struct request_queue *q =
960 container_of(work, struct request_queue, timeout_work);
961 unsigned long next = 0;
962 struct blk_mq_hw_ctx *hctx;
965 /* A deadlock might occur if a request is stuck requiring a
966 * timeout at the same time a queue freeze is waiting
967 * completion, since the timeout code would not be able to
968 * acquire the queue reference here.
970 * That's why we don't use blk_queue_enter here; instead, we use
971 * percpu_ref_tryget directly, because we need to be able to
972 * obtain a reference even in the short window between the queue
973 * starting to freeze, by dropping the first reference in
974 * blk_freeze_queue_start, and the moment the last request is
975 * consumed, marked by the instant q_usage_counter reaches
978 if (!percpu_ref_tryget(&q->q_usage_counter))
981 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
984 mod_timer(&q->timeout, next);
987 * Request timeouts are handled as a forward rolling timer. If
988 * we end up here it means that no requests are pending and
989 * also that no request has been pending for a while. Mark
992 queue_for_each_hw_ctx(q, hctx, i) {
993 /* the hctx may be unmapped, so check it here */
994 if (blk_mq_hw_queue_mapped(hctx))
995 blk_mq_tag_idle(hctx);
1001 struct flush_busy_ctx_data {
1002 struct blk_mq_hw_ctx *hctx;
1003 struct list_head *list;
1006 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1008 struct flush_busy_ctx_data *flush_data = data;
1009 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1010 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1011 enum hctx_type type = hctx->type;
1013 spin_lock(&ctx->lock);
1014 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1015 sbitmap_clear_bit(sb, bitnr);
1016 spin_unlock(&ctx->lock);
1021 * Process software queues that have been marked busy, splicing them
1022 * to the for-dispatch
1024 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1026 struct flush_busy_ctx_data data = {
1031 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1033 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1035 struct dispatch_rq_data {
1036 struct blk_mq_hw_ctx *hctx;
1040 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1043 struct dispatch_rq_data *dispatch_data = data;
1044 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1045 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1046 enum hctx_type type = hctx->type;
1048 spin_lock(&ctx->lock);
1049 if (!list_empty(&ctx->rq_lists[type])) {
1050 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1051 list_del_init(&dispatch_data->rq->queuelist);
1052 if (list_empty(&ctx->rq_lists[type]))
1053 sbitmap_clear_bit(sb, bitnr);
1055 spin_unlock(&ctx->lock);
1057 return !dispatch_data->rq;
1060 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1061 struct blk_mq_ctx *start)
1063 unsigned off = start ? start->index_hw[hctx->type] : 0;
1064 struct dispatch_rq_data data = {
1069 __sbitmap_for_each_set(&hctx->ctx_map, off,
1070 dispatch_rq_from_ctx, &data);
1075 static inline unsigned int queued_to_index(unsigned int queued)
1080 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1083 static bool __blk_mq_get_driver_tag(struct request *rq)
1085 struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags;
1086 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1089 blk_mq_tag_busy(rq->mq_hctx);
1091 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1092 bt = rq->mq_hctx->tags->breserved_tags;
1095 if (!hctx_may_queue(rq->mq_hctx, bt))
1099 tag = __sbitmap_queue_get(bt);
1100 if (tag == BLK_MQ_NO_TAG)
1103 rq->tag = tag + tag_offset;
1107 bool blk_mq_get_driver_tag(struct request *rq)
1109 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1111 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1114 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1115 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1116 rq->rq_flags |= RQF_MQ_INFLIGHT;
1117 __blk_mq_inc_active_requests(hctx);
1119 hctx->tags->rqs[rq->tag] = rq;
1123 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1124 int flags, void *key)
1126 struct blk_mq_hw_ctx *hctx;
1128 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1130 spin_lock(&hctx->dispatch_wait_lock);
1131 if (!list_empty(&wait->entry)) {
1132 struct sbitmap_queue *sbq;
1134 list_del_init(&wait->entry);
1135 sbq = hctx->tags->bitmap_tags;
1136 atomic_dec(&sbq->ws_active);
1138 spin_unlock(&hctx->dispatch_wait_lock);
1140 blk_mq_run_hw_queue(hctx, true);
1145 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1146 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1147 * restart. For both cases, take care to check the condition again after
1148 * marking us as waiting.
1150 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1153 struct sbitmap_queue *sbq = hctx->tags->bitmap_tags;
1154 struct wait_queue_head *wq;
1155 wait_queue_entry_t *wait;
1158 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1159 blk_mq_sched_mark_restart_hctx(hctx);
1162 * It's possible that a tag was freed in the window between the
1163 * allocation failure and adding the hardware queue to the wait
1166 * Don't clear RESTART here, someone else could have set it.
1167 * At most this will cost an extra queue run.
1169 return blk_mq_get_driver_tag(rq);
1172 wait = &hctx->dispatch_wait;
1173 if (!list_empty_careful(&wait->entry))
1176 wq = &bt_wait_ptr(sbq, hctx)->wait;
1178 spin_lock_irq(&wq->lock);
1179 spin_lock(&hctx->dispatch_wait_lock);
1180 if (!list_empty(&wait->entry)) {
1181 spin_unlock(&hctx->dispatch_wait_lock);
1182 spin_unlock_irq(&wq->lock);
1186 atomic_inc(&sbq->ws_active);
1187 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1188 __add_wait_queue(wq, wait);
1191 * It's possible that a tag was freed in the window between the
1192 * allocation failure and adding the hardware queue to the wait
1195 ret = blk_mq_get_driver_tag(rq);
1197 spin_unlock(&hctx->dispatch_wait_lock);
1198 spin_unlock_irq(&wq->lock);
1203 * We got a tag, remove ourselves from the wait queue to ensure
1204 * someone else gets the wakeup.
1206 list_del_init(&wait->entry);
1207 atomic_dec(&sbq->ws_active);
1208 spin_unlock(&hctx->dispatch_wait_lock);
1209 spin_unlock_irq(&wq->lock);
1214 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1215 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1217 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1218 * - EWMA is one simple way to compute running average value
1219 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1220 * - take 4 as factor for avoiding to get too small(0) result, and this
1221 * factor doesn't matter because EWMA decreases exponentially
1223 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1227 ewma = hctx->dispatch_busy;
1232 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1234 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1235 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1237 hctx->dispatch_busy = ewma;
1240 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1242 static void blk_mq_handle_dev_resource(struct request *rq,
1243 struct list_head *list)
1245 struct request *next =
1246 list_first_entry_or_null(list, struct request, queuelist);
1249 * If an I/O scheduler has been configured and we got a driver tag for
1250 * the next request already, free it.
1253 blk_mq_put_driver_tag(next);
1255 list_add(&rq->queuelist, list);
1256 __blk_mq_requeue_request(rq);
1259 static void blk_mq_handle_zone_resource(struct request *rq,
1260 struct list_head *zone_list)
1263 * If we end up here it is because we cannot dispatch a request to a
1264 * specific zone due to LLD level zone-write locking or other zone
1265 * related resource not being available. In this case, set the request
1266 * aside in zone_list for retrying it later.
1268 list_add(&rq->queuelist, zone_list);
1269 __blk_mq_requeue_request(rq);
1272 enum prep_dispatch {
1274 PREP_DISPATCH_NO_TAG,
1275 PREP_DISPATCH_NO_BUDGET,
1278 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1281 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1282 int budget_token = -1;
1285 budget_token = blk_mq_get_dispatch_budget(rq->q);
1286 if (budget_token < 0) {
1287 blk_mq_put_driver_tag(rq);
1288 return PREP_DISPATCH_NO_BUDGET;
1290 blk_mq_set_rq_budget_token(rq, budget_token);
1293 if (!blk_mq_get_driver_tag(rq)) {
1295 * The initial allocation attempt failed, so we need to
1296 * rerun the hardware queue when a tag is freed. The
1297 * waitqueue takes care of that. If the queue is run
1298 * before we add this entry back on the dispatch list,
1299 * we'll re-run it below.
1301 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1303 * All budgets not got from this function will be put
1304 * together during handling partial dispatch
1307 blk_mq_put_dispatch_budget(rq->q, budget_token);
1308 return PREP_DISPATCH_NO_TAG;
1312 return PREP_DISPATCH_OK;
1315 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1316 static void blk_mq_release_budgets(struct request_queue *q,
1317 struct list_head *list)
1321 list_for_each_entry(rq, list, queuelist) {
1322 int budget_token = blk_mq_get_rq_budget_token(rq);
1324 if (budget_token >= 0)
1325 blk_mq_put_dispatch_budget(q, budget_token);
1330 * Returns true if we did some work AND can potentially do more.
1332 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1333 unsigned int nr_budgets)
1335 enum prep_dispatch prep;
1336 struct request_queue *q = hctx->queue;
1337 struct request *rq, *nxt;
1339 blk_status_t ret = BLK_STS_OK;
1340 LIST_HEAD(zone_list);
1342 if (list_empty(list))
1346 * Now process all the entries, sending them to the driver.
1348 errors = queued = 0;
1350 struct blk_mq_queue_data bd;
1352 rq = list_first_entry(list, struct request, queuelist);
1354 WARN_ON_ONCE(hctx != rq->mq_hctx);
1355 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1356 if (prep != PREP_DISPATCH_OK)
1359 list_del_init(&rq->queuelist);
1364 * Flag last if we have no more requests, or if we have more
1365 * but can't assign a driver tag to it.
1367 if (list_empty(list))
1370 nxt = list_first_entry(list, struct request, queuelist);
1371 bd.last = !blk_mq_get_driver_tag(nxt);
1375 * once the request is queued to lld, no need to cover the
1380 ret = q->mq_ops->queue_rq(hctx, &bd);
1385 case BLK_STS_RESOURCE:
1386 case BLK_STS_DEV_RESOURCE:
1387 blk_mq_handle_dev_resource(rq, list);
1389 case BLK_STS_ZONE_RESOURCE:
1391 * Move the request to zone_list and keep going through
1392 * the dispatch list to find more requests the drive can
1395 blk_mq_handle_zone_resource(rq, &zone_list);
1399 blk_mq_end_request(rq, ret);
1401 } while (!list_empty(list));
1403 if (!list_empty(&zone_list))
1404 list_splice_tail_init(&zone_list, list);
1406 hctx->dispatched[queued_to_index(queued)]++;
1408 /* If we didn't flush the entire list, we could have told the driver
1409 * there was more coming, but that turned out to be a lie.
1411 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1412 q->mq_ops->commit_rqs(hctx);
1414 * Any items that need requeuing? Stuff them into hctx->dispatch,
1415 * that is where we will continue on next queue run.
1417 if (!list_empty(list)) {
1419 /* For non-shared tags, the RESTART check will suffice */
1420 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1421 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1422 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1425 blk_mq_release_budgets(q, list);
1427 spin_lock(&hctx->lock);
1428 list_splice_tail_init(list, &hctx->dispatch);
1429 spin_unlock(&hctx->lock);
1432 * Order adding requests to hctx->dispatch and checking
1433 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1434 * in blk_mq_sched_restart(). Avoid restart code path to
1435 * miss the new added requests to hctx->dispatch, meantime
1436 * SCHED_RESTART is observed here.
1441 * If SCHED_RESTART was set by the caller of this function and
1442 * it is no longer set that means that it was cleared by another
1443 * thread and hence that a queue rerun is needed.
1445 * If 'no_tag' is set, that means that we failed getting
1446 * a driver tag with an I/O scheduler attached. If our dispatch
1447 * waitqueue is no longer active, ensure that we run the queue
1448 * AFTER adding our entries back to the list.
1450 * If no I/O scheduler has been configured it is possible that
1451 * the hardware queue got stopped and restarted before requests
1452 * were pushed back onto the dispatch list. Rerun the queue to
1453 * avoid starvation. Notes:
1454 * - blk_mq_run_hw_queue() checks whether or not a queue has
1455 * been stopped before rerunning a queue.
1456 * - Some but not all block drivers stop a queue before
1457 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1460 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1461 * bit is set, run queue after a delay to avoid IO stalls
1462 * that could otherwise occur if the queue is idle. We'll do
1463 * similar if we couldn't get budget and SCHED_RESTART is set.
1465 needs_restart = blk_mq_sched_needs_restart(hctx);
1466 if (!needs_restart ||
1467 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1468 blk_mq_run_hw_queue(hctx, true);
1469 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1471 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1473 blk_mq_update_dispatch_busy(hctx, true);
1476 blk_mq_update_dispatch_busy(hctx, false);
1478 return (queued + errors) != 0;
1482 * __blk_mq_run_hw_queue - Run a hardware queue.
1483 * @hctx: Pointer to the hardware queue to run.
1485 * Send pending requests to the hardware.
1487 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1492 * We can't run the queue inline with ints disabled. Ensure that
1493 * we catch bad users of this early.
1495 WARN_ON_ONCE(in_interrupt());
1497 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1499 hctx_lock(hctx, &srcu_idx);
1500 blk_mq_sched_dispatch_requests(hctx);
1501 hctx_unlock(hctx, srcu_idx);
1504 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1506 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1508 if (cpu >= nr_cpu_ids)
1509 cpu = cpumask_first(hctx->cpumask);
1514 * It'd be great if the workqueue API had a way to pass
1515 * in a mask and had some smarts for more clever placement.
1516 * For now we just round-robin here, switching for every
1517 * BLK_MQ_CPU_WORK_BATCH queued items.
1519 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1522 int next_cpu = hctx->next_cpu;
1524 if (hctx->queue->nr_hw_queues == 1)
1525 return WORK_CPU_UNBOUND;
1527 if (--hctx->next_cpu_batch <= 0) {
1529 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1531 if (next_cpu >= nr_cpu_ids)
1532 next_cpu = blk_mq_first_mapped_cpu(hctx);
1533 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1537 * Do unbound schedule if we can't find a online CPU for this hctx,
1538 * and it should only happen in the path of handling CPU DEAD.
1540 if (!cpu_online(next_cpu)) {
1547 * Make sure to re-select CPU next time once after CPUs
1548 * in hctx->cpumask become online again.
1550 hctx->next_cpu = next_cpu;
1551 hctx->next_cpu_batch = 1;
1552 return WORK_CPU_UNBOUND;
1555 hctx->next_cpu = next_cpu;
1560 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1561 * @hctx: Pointer to the hardware queue to run.
1562 * @async: If we want to run the queue asynchronously.
1563 * @msecs: Milliseconds of delay to wait before running the queue.
1565 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1566 * with a delay of @msecs.
1568 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1569 unsigned long msecs)
1571 if (unlikely(blk_mq_hctx_stopped(hctx)))
1574 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1575 int cpu = get_cpu();
1576 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1577 __blk_mq_run_hw_queue(hctx);
1585 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1586 msecs_to_jiffies(msecs));
1590 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1591 * @hctx: Pointer to the hardware queue to run.
1592 * @msecs: Milliseconds of delay to wait before running the queue.
1594 * Run a hardware queue asynchronously with a delay of @msecs.
1596 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1598 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1600 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1603 * blk_mq_run_hw_queue - Start to run a hardware queue.
1604 * @hctx: Pointer to the hardware queue to run.
1605 * @async: If we want to run the queue asynchronously.
1607 * Check if the request queue is not in a quiesced state and if there are
1608 * pending requests to be sent. If this is true, run the queue to send requests
1611 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1617 * When queue is quiesced, we may be switching io scheduler, or
1618 * updating nr_hw_queues, or other things, and we can't run queue
1619 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1621 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1624 hctx_lock(hctx, &srcu_idx);
1625 need_run = !blk_queue_quiesced(hctx->queue) &&
1626 blk_mq_hctx_has_pending(hctx);
1627 hctx_unlock(hctx, srcu_idx);
1630 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1632 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1635 * Is the request queue handled by an IO scheduler that does not respect
1636 * hardware queues when dispatching?
1638 static bool blk_mq_has_sqsched(struct request_queue *q)
1640 struct elevator_queue *e = q->elevator;
1642 if (e && e->type->ops.dispatch_request &&
1643 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
1649 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1652 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
1654 struct blk_mq_hw_ctx *hctx;
1657 * If the IO scheduler does not respect hardware queues when
1658 * dispatching, we just don't bother with multiple HW queues and
1659 * dispatch from hctx for the current CPU since running multiple queues
1660 * just causes lock contention inside the scheduler and pointless cache
1663 hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
1664 raw_smp_processor_id());
1665 if (!blk_mq_hctx_stopped(hctx))
1671 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1672 * @q: Pointer to the request queue to run.
1673 * @async: If we want to run the queue asynchronously.
1675 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1677 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1681 if (blk_mq_has_sqsched(q))
1682 sq_hctx = blk_mq_get_sq_hctx(q);
1683 queue_for_each_hw_ctx(q, hctx, i) {
1684 if (blk_mq_hctx_stopped(hctx))
1687 * Dispatch from this hctx either if there's no hctx preferred
1688 * by IO scheduler or if it has requests that bypass the
1691 if (!sq_hctx || sq_hctx == hctx ||
1692 !list_empty_careful(&hctx->dispatch))
1693 blk_mq_run_hw_queue(hctx, async);
1696 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1699 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1700 * @q: Pointer to the request queue to run.
1701 * @msecs: Milliseconds of delay to wait before running the queues.
1703 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1705 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1709 if (blk_mq_has_sqsched(q))
1710 sq_hctx = blk_mq_get_sq_hctx(q);
1711 queue_for_each_hw_ctx(q, hctx, i) {
1712 if (blk_mq_hctx_stopped(hctx))
1715 * Dispatch from this hctx either if there's no hctx preferred
1716 * by IO scheduler or if it has requests that bypass the
1719 if (!sq_hctx || sq_hctx == hctx ||
1720 !list_empty_careful(&hctx->dispatch))
1721 blk_mq_delay_run_hw_queue(hctx, msecs);
1724 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1727 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1728 * @q: request queue.
1730 * The caller is responsible for serializing this function against
1731 * blk_mq_{start,stop}_hw_queue().
1733 bool blk_mq_queue_stopped(struct request_queue *q)
1735 struct blk_mq_hw_ctx *hctx;
1738 queue_for_each_hw_ctx(q, hctx, i)
1739 if (blk_mq_hctx_stopped(hctx))
1744 EXPORT_SYMBOL(blk_mq_queue_stopped);
1747 * This function is often used for pausing .queue_rq() by driver when
1748 * there isn't enough resource or some conditions aren't satisfied, and
1749 * BLK_STS_RESOURCE is usually returned.
1751 * We do not guarantee that dispatch can be drained or blocked
1752 * after blk_mq_stop_hw_queue() returns. Please use
1753 * blk_mq_quiesce_queue() for that requirement.
1755 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1757 cancel_delayed_work(&hctx->run_work);
1759 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1761 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1764 * This function is often used for pausing .queue_rq() by driver when
1765 * there isn't enough resource or some conditions aren't satisfied, and
1766 * BLK_STS_RESOURCE is usually returned.
1768 * We do not guarantee that dispatch can be drained or blocked
1769 * after blk_mq_stop_hw_queues() returns. Please use
1770 * blk_mq_quiesce_queue() for that requirement.
1772 void blk_mq_stop_hw_queues(struct request_queue *q)
1774 struct blk_mq_hw_ctx *hctx;
1777 queue_for_each_hw_ctx(q, hctx, i)
1778 blk_mq_stop_hw_queue(hctx);
1780 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1782 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1784 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1786 blk_mq_run_hw_queue(hctx, false);
1788 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1790 void blk_mq_start_hw_queues(struct request_queue *q)
1792 struct blk_mq_hw_ctx *hctx;
1795 queue_for_each_hw_ctx(q, hctx, i)
1796 blk_mq_start_hw_queue(hctx);
1798 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1800 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1802 if (!blk_mq_hctx_stopped(hctx))
1805 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1806 blk_mq_run_hw_queue(hctx, async);
1808 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1810 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1812 struct blk_mq_hw_ctx *hctx;
1815 queue_for_each_hw_ctx(q, hctx, i)
1816 blk_mq_start_stopped_hw_queue(hctx, async);
1818 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1820 static void blk_mq_run_work_fn(struct work_struct *work)
1822 struct blk_mq_hw_ctx *hctx;
1824 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1827 * If we are stopped, don't run the queue.
1829 if (blk_mq_hctx_stopped(hctx))
1832 __blk_mq_run_hw_queue(hctx);
1835 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1839 struct blk_mq_ctx *ctx = rq->mq_ctx;
1840 enum hctx_type type = hctx->type;
1842 lockdep_assert_held(&ctx->lock);
1844 trace_block_rq_insert(rq);
1847 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1849 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1852 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1855 struct blk_mq_ctx *ctx = rq->mq_ctx;
1857 lockdep_assert_held(&ctx->lock);
1859 __blk_mq_insert_req_list(hctx, rq, at_head);
1860 blk_mq_hctx_mark_pending(hctx, ctx);
1864 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1865 * @rq: Pointer to request to be inserted.
1866 * @at_head: true if the request should be inserted at the head of the list.
1867 * @run_queue: If we should run the hardware queue after inserting the request.
1869 * Should only be used carefully, when the caller knows we want to
1870 * bypass a potential IO scheduler on the target device.
1872 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1875 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1877 spin_lock(&hctx->lock);
1879 list_add(&rq->queuelist, &hctx->dispatch);
1881 list_add_tail(&rq->queuelist, &hctx->dispatch);
1882 spin_unlock(&hctx->lock);
1885 blk_mq_run_hw_queue(hctx, false);
1888 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1889 struct list_head *list)
1893 enum hctx_type type = hctx->type;
1896 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1899 list_for_each_entry(rq, list, queuelist) {
1900 BUG_ON(rq->mq_ctx != ctx);
1901 trace_block_rq_insert(rq);
1904 spin_lock(&ctx->lock);
1905 list_splice_tail_init(list, &ctx->rq_lists[type]);
1906 blk_mq_hctx_mark_pending(hctx, ctx);
1907 spin_unlock(&ctx->lock);
1910 static int plug_rq_cmp(void *priv, const struct list_head *a,
1911 const struct list_head *b)
1913 struct request *rqa = container_of(a, struct request, queuelist);
1914 struct request *rqb = container_of(b, struct request, queuelist);
1916 if (rqa->mq_ctx != rqb->mq_ctx)
1917 return rqa->mq_ctx > rqb->mq_ctx;
1918 if (rqa->mq_hctx != rqb->mq_hctx)
1919 return rqa->mq_hctx > rqb->mq_hctx;
1921 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1924 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1928 if (list_empty(&plug->mq_list))
1930 list_splice_init(&plug->mq_list, &list);
1932 if (plug->rq_count > 2 && plug->multiple_queues)
1933 list_sort(NULL, &list, plug_rq_cmp);
1938 struct list_head rq_list;
1939 struct request *rq, *head_rq = list_entry_rq(list.next);
1940 struct list_head *pos = &head_rq->queuelist; /* skip first */
1941 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1942 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1943 unsigned int depth = 1;
1945 list_for_each_continue(pos, &list) {
1946 rq = list_entry_rq(pos);
1948 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1953 list_cut_before(&rq_list, &list, pos);
1954 trace_block_unplug(head_rq->q, depth, !from_schedule);
1955 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1957 } while(!list_empty(&list));
1960 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1961 unsigned int nr_segs)
1965 if (bio->bi_opf & REQ_RAHEAD)
1966 rq->cmd_flags |= REQ_FAILFAST_MASK;
1968 rq->__sector = bio->bi_iter.bi_sector;
1969 rq->write_hint = bio->bi_write_hint;
1970 blk_rq_bio_prep(rq, bio, nr_segs);
1972 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1973 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1976 blk_account_io_start(rq);
1979 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1981 blk_qc_t *cookie, bool last)
1983 struct request_queue *q = rq->q;
1984 struct blk_mq_queue_data bd = {
1988 blk_qc_t new_cookie;
1991 new_cookie = request_to_qc_t(hctx, rq);
1994 * For OK queue, we are done. For error, caller may kill it.
1995 * Any other error (busy), just add it to our list as we
1996 * previously would have done.
1998 ret = q->mq_ops->queue_rq(hctx, &bd);
2001 blk_mq_update_dispatch_busy(hctx, false);
2002 *cookie = new_cookie;
2004 case BLK_STS_RESOURCE:
2005 case BLK_STS_DEV_RESOURCE:
2006 blk_mq_update_dispatch_busy(hctx, true);
2007 __blk_mq_requeue_request(rq);
2010 blk_mq_update_dispatch_busy(hctx, false);
2011 *cookie = BLK_QC_T_NONE;
2018 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2021 bool bypass_insert, bool last)
2023 struct request_queue *q = rq->q;
2024 bool run_queue = true;
2028 * RCU or SRCU read lock is needed before checking quiesced flag.
2030 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2031 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2032 * and avoid driver to try to dispatch again.
2034 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2036 bypass_insert = false;
2040 if (q->elevator && !bypass_insert)
2043 budget_token = blk_mq_get_dispatch_budget(q);
2044 if (budget_token < 0)
2047 blk_mq_set_rq_budget_token(rq, budget_token);
2049 if (!blk_mq_get_driver_tag(rq)) {
2050 blk_mq_put_dispatch_budget(q, budget_token);
2054 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2057 return BLK_STS_RESOURCE;
2059 blk_mq_sched_insert_request(rq, false, run_queue, false);
2065 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2066 * @hctx: Pointer of the associated hardware queue.
2067 * @rq: Pointer to request to be sent.
2068 * @cookie: Request queue cookie.
2070 * If the device has enough resources to accept a new request now, send the
2071 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2072 * we can try send it another time in the future. Requests inserted at this
2073 * queue have higher priority.
2075 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2076 struct request *rq, blk_qc_t *cookie)
2081 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2083 hctx_lock(hctx, &srcu_idx);
2085 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2086 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2087 blk_mq_request_bypass_insert(rq, false, true);
2088 else if (ret != BLK_STS_OK)
2089 blk_mq_end_request(rq, ret);
2091 hctx_unlock(hctx, srcu_idx);
2094 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2098 blk_qc_t unused_cookie;
2099 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2101 hctx_lock(hctx, &srcu_idx);
2102 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2103 hctx_unlock(hctx, srcu_idx);
2108 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2109 struct list_head *list)
2114 while (!list_empty(list)) {
2116 struct request *rq = list_first_entry(list, struct request,
2119 list_del_init(&rq->queuelist);
2120 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2121 if (ret != BLK_STS_OK) {
2122 if (ret == BLK_STS_RESOURCE ||
2123 ret == BLK_STS_DEV_RESOURCE) {
2124 blk_mq_request_bypass_insert(rq, false,
2128 blk_mq_end_request(rq, ret);
2135 * If we didn't flush the entire list, we could have told
2136 * the driver there was more coming, but that turned out to
2139 if ((!list_empty(list) || errors) &&
2140 hctx->queue->mq_ops->commit_rqs && queued)
2141 hctx->queue->mq_ops->commit_rqs(hctx);
2144 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2146 list_add_tail(&rq->queuelist, &plug->mq_list);
2148 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2149 struct request *tmp;
2151 tmp = list_first_entry(&plug->mq_list, struct request,
2153 if (tmp->q != rq->q)
2154 plug->multiple_queues = true;
2159 * blk_mq_submit_bio - Create and send a request to block device.
2160 * @bio: Bio pointer.
2162 * Builds up a request structure from @q and @bio and send to the device. The
2163 * request may not be queued directly to hardware if:
2164 * * This request can be merged with another one
2165 * * We want to place request at plug queue for possible future merging
2166 * * There is an IO scheduler active at this queue
2168 * It will not queue the request if there is an error with the bio, or at the
2171 * Returns: Request queue cookie.
2173 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2175 struct request_queue *q = bio->bi_bdev->bd_disk->queue;
2176 const int is_sync = op_is_sync(bio->bi_opf);
2177 const int is_flush_fua = op_is_flush(bio->bi_opf);
2178 struct blk_mq_alloc_data data = {
2182 struct blk_plug *plug;
2183 struct request *same_queue_rq = NULL;
2184 unsigned int nr_segs;
2189 blk_queue_bounce(q, &bio);
2190 __blk_queue_split(&bio, &nr_segs);
2192 if (!bio_integrity_prep(bio))
2195 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2196 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2199 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2202 rq_qos_throttle(q, bio);
2204 hipri = bio->bi_opf & REQ_HIPRI;
2206 data.cmd_flags = bio->bi_opf;
2207 rq = __blk_mq_alloc_request(&data);
2208 if (unlikely(!rq)) {
2209 rq_qos_cleanup(q, bio);
2210 if (bio->bi_opf & REQ_NOWAIT)
2211 bio_wouldblock_error(bio);
2215 trace_block_getrq(bio);
2217 rq_qos_track(q, rq, bio);
2219 cookie = request_to_qc_t(data.hctx, rq);
2221 blk_mq_bio_to_request(rq, bio, nr_segs);
2223 ret = blk_crypto_init_request(rq);
2224 if (ret != BLK_STS_OK) {
2225 bio->bi_status = ret;
2227 blk_mq_free_request(rq);
2228 return BLK_QC_T_NONE;
2231 plug = blk_mq_plug(q, bio);
2232 if (unlikely(is_flush_fua)) {
2233 /* Bypass scheduler for flush requests */
2234 blk_insert_flush(rq);
2235 blk_mq_run_hw_queue(data.hctx, true);
2236 } else if (plug && (q->nr_hw_queues == 1 ||
2237 blk_mq_is_sbitmap_shared(rq->mq_hctx->flags) ||
2238 q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) {
2240 * Use plugging if we have a ->commit_rqs() hook as well, as
2241 * we know the driver uses bd->last in a smart fashion.
2243 * Use normal plugging if this disk is slow HDD, as sequential
2244 * IO may benefit a lot from plug merging.
2246 unsigned int request_count = plug->rq_count;
2247 struct request *last = NULL;
2250 trace_block_plug(q);
2252 last = list_entry_rq(plug->mq_list.prev);
2254 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2255 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2256 blk_flush_plug_list(plug, false);
2257 trace_block_plug(q);
2260 blk_add_rq_to_plug(plug, rq);
2261 } else if (q->elevator) {
2262 /* Insert the request at the IO scheduler queue */
2263 blk_mq_sched_insert_request(rq, false, true, true);
2264 } else if (plug && !blk_queue_nomerges(q)) {
2266 * We do limited plugging. If the bio can be merged, do that.
2267 * Otherwise the existing request in the plug list will be
2268 * issued. So the plug list will have one request at most
2269 * The plug list might get flushed before this. If that happens,
2270 * the plug list is empty, and same_queue_rq is invalid.
2272 if (list_empty(&plug->mq_list))
2273 same_queue_rq = NULL;
2274 if (same_queue_rq) {
2275 list_del_init(&same_queue_rq->queuelist);
2278 blk_add_rq_to_plug(plug, rq);
2279 trace_block_plug(q);
2281 if (same_queue_rq) {
2282 data.hctx = same_queue_rq->mq_hctx;
2283 trace_block_unplug(q, 1, true);
2284 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2287 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2288 !data.hctx->dispatch_busy) {
2290 * There is no scheduler and we can try to send directly
2293 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2296 blk_mq_sched_insert_request(rq, false, true, true);
2300 return BLK_QC_T_NONE;
2304 return BLK_QC_T_NONE;
2307 static size_t order_to_size(unsigned int order)
2309 return (size_t)PAGE_SIZE << order;
2312 /* called before freeing request pool in @tags */
2313 static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set *set,
2314 struct blk_mq_tags *tags, unsigned int hctx_idx)
2316 struct blk_mq_tags *drv_tags = set->tags[hctx_idx];
2318 unsigned long flags;
2320 list_for_each_entry(page, &tags->page_list, lru) {
2321 unsigned long start = (unsigned long)page_address(page);
2322 unsigned long end = start + order_to_size(page->private);
2325 for (i = 0; i < set->queue_depth; i++) {
2326 struct request *rq = drv_tags->rqs[i];
2327 unsigned long rq_addr = (unsigned long)rq;
2329 if (rq_addr >= start && rq_addr < end) {
2330 WARN_ON_ONCE(refcount_read(&rq->ref) != 0);
2331 cmpxchg(&drv_tags->rqs[i], rq, NULL);
2337 * Wait until all pending iteration is done.
2339 * Request reference is cleared and it is guaranteed to be observed
2340 * after the ->lock is released.
2342 spin_lock_irqsave(&drv_tags->lock, flags);
2343 spin_unlock_irqrestore(&drv_tags->lock, flags);
2346 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2347 unsigned int hctx_idx)
2351 if (tags->rqs && set->ops->exit_request) {
2354 for (i = 0; i < tags->nr_tags; i++) {
2355 struct request *rq = tags->static_rqs[i];
2359 set->ops->exit_request(set, rq, hctx_idx);
2360 tags->static_rqs[i] = NULL;
2364 blk_mq_clear_rq_mapping(set, tags, hctx_idx);
2366 while (!list_empty(&tags->page_list)) {
2367 page = list_first_entry(&tags->page_list, struct page, lru);
2368 list_del_init(&page->lru);
2370 * Remove kmemleak object previously allocated in
2371 * blk_mq_alloc_rqs().
2373 kmemleak_free(page_address(page));
2374 __free_pages(page, page->private);
2378 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags)
2382 kfree(tags->static_rqs);
2383 tags->static_rqs = NULL;
2385 blk_mq_free_tags(tags, flags);
2388 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2389 unsigned int hctx_idx,
2390 unsigned int nr_tags,
2391 unsigned int reserved_tags,
2394 struct blk_mq_tags *tags;
2397 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2398 if (node == NUMA_NO_NODE)
2399 node = set->numa_node;
2401 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags);
2405 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2406 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2409 blk_mq_free_tags(tags, flags);
2413 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2414 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2416 if (!tags->static_rqs) {
2418 blk_mq_free_tags(tags, flags);
2425 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2426 unsigned int hctx_idx, int node)
2430 if (set->ops->init_request) {
2431 ret = set->ops->init_request(set, rq, hctx_idx, node);
2436 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2440 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2441 unsigned int hctx_idx, unsigned int depth)
2443 unsigned int i, j, entries_per_page, max_order = 4;
2444 size_t rq_size, left;
2447 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2448 if (node == NUMA_NO_NODE)
2449 node = set->numa_node;
2451 INIT_LIST_HEAD(&tags->page_list);
2454 * rq_size is the size of the request plus driver payload, rounded
2455 * to the cacheline size
2457 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2459 left = rq_size * depth;
2461 for (i = 0; i < depth; ) {
2462 int this_order = max_order;
2467 while (this_order && left < order_to_size(this_order - 1))
2471 page = alloc_pages_node(node,
2472 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2478 if (order_to_size(this_order) < rq_size)
2485 page->private = this_order;
2486 list_add_tail(&page->lru, &tags->page_list);
2488 p = page_address(page);
2490 * Allow kmemleak to scan these pages as they contain pointers
2491 * to additional allocations like via ops->init_request().
2493 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2494 entries_per_page = order_to_size(this_order) / rq_size;
2495 to_do = min(entries_per_page, depth - i);
2496 left -= to_do * rq_size;
2497 for (j = 0; j < to_do; j++) {
2498 struct request *rq = p;
2500 tags->static_rqs[i] = rq;
2501 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2502 tags->static_rqs[i] = NULL;
2513 blk_mq_free_rqs(set, tags, hctx_idx);
2517 struct rq_iter_data {
2518 struct blk_mq_hw_ctx *hctx;
2522 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2524 struct rq_iter_data *iter_data = data;
2526 if (rq->mq_hctx != iter_data->hctx)
2528 iter_data->has_rq = true;
2532 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2534 struct blk_mq_tags *tags = hctx->sched_tags ?
2535 hctx->sched_tags : hctx->tags;
2536 struct rq_iter_data data = {
2540 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2544 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2545 struct blk_mq_hw_ctx *hctx)
2547 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2549 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2554 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2556 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2557 struct blk_mq_hw_ctx, cpuhp_online);
2559 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2560 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2564 * Prevent new request from being allocated on the current hctx.
2566 * The smp_mb__after_atomic() Pairs with the implied barrier in
2567 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2568 * seen once we return from the tag allocator.
2570 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2571 smp_mb__after_atomic();
2574 * Try to grab a reference to the queue and wait for any outstanding
2575 * requests. If we could not grab a reference the queue has been
2576 * frozen and there are no requests.
2578 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2579 while (blk_mq_hctx_has_requests(hctx))
2581 percpu_ref_put(&hctx->queue->q_usage_counter);
2587 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2589 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2590 struct blk_mq_hw_ctx, cpuhp_online);
2592 if (cpumask_test_cpu(cpu, hctx->cpumask))
2593 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2598 * 'cpu' is going away. splice any existing rq_list entries from this
2599 * software queue to the hw queue dispatch list, and ensure that it
2602 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2604 struct blk_mq_hw_ctx *hctx;
2605 struct blk_mq_ctx *ctx;
2607 enum hctx_type type;
2609 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2610 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2613 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2616 spin_lock(&ctx->lock);
2617 if (!list_empty(&ctx->rq_lists[type])) {
2618 list_splice_init(&ctx->rq_lists[type], &tmp);
2619 blk_mq_hctx_clear_pending(hctx, ctx);
2621 spin_unlock(&ctx->lock);
2623 if (list_empty(&tmp))
2626 spin_lock(&hctx->lock);
2627 list_splice_tail_init(&tmp, &hctx->dispatch);
2628 spin_unlock(&hctx->lock);
2630 blk_mq_run_hw_queue(hctx, true);
2634 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2636 if (!(hctx->flags & BLK_MQ_F_STACKING))
2637 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2638 &hctx->cpuhp_online);
2639 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2644 * Before freeing hw queue, clearing the flush request reference in
2645 * tags->rqs[] for avoiding potential UAF.
2647 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
2648 unsigned int queue_depth, struct request *flush_rq)
2651 unsigned long flags;
2653 /* The hw queue may not be mapped yet */
2657 WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0);
2659 for (i = 0; i < queue_depth; i++)
2660 cmpxchg(&tags->rqs[i], flush_rq, NULL);
2663 * Wait until all pending iteration is done.
2665 * Request reference is cleared and it is guaranteed to be observed
2666 * after the ->lock is released.
2668 spin_lock_irqsave(&tags->lock, flags);
2669 spin_unlock_irqrestore(&tags->lock, flags);
2672 /* hctx->ctxs will be freed in queue's release handler */
2673 static void blk_mq_exit_hctx(struct request_queue *q,
2674 struct blk_mq_tag_set *set,
2675 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2677 struct request *flush_rq = hctx->fq->flush_rq;
2679 if (blk_mq_hw_queue_mapped(hctx))
2680 blk_mq_tag_idle(hctx);
2682 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
2683 set->queue_depth, flush_rq);
2684 if (set->ops->exit_request)
2685 set->ops->exit_request(set, flush_rq, hctx_idx);
2687 if (set->ops->exit_hctx)
2688 set->ops->exit_hctx(hctx, hctx_idx);
2690 blk_mq_remove_cpuhp(hctx);
2692 spin_lock(&q->unused_hctx_lock);
2693 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2694 spin_unlock(&q->unused_hctx_lock);
2697 static void blk_mq_exit_hw_queues(struct request_queue *q,
2698 struct blk_mq_tag_set *set, int nr_queue)
2700 struct blk_mq_hw_ctx *hctx;
2703 queue_for_each_hw_ctx(q, hctx, i) {
2706 blk_mq_debugfs_unregister_hctx(hctx);
2707 blk_mq_exit_hctx(q, set, hctx, i);
2711 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2713 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2715 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2716 __alignof__(struct blk_mq_hw_ctx)) !=
2717 sizeof(struct blk_mq_hw_ctx));
2719 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2720 hw_ctx_size += sizeof(struct srcu_struct);
2725 static int blk_mq_init_hctx(struct request_queue *q,
2726 struct blk_mq_tag_set *set,
2727 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2729 hctx->queue_num = hctx_idx;
2731 if (!(hctx->flags & BLK_MQ_F_STACKING))
2732 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2733 &hctx->cpuhp_online);
2734 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2736 hctx->tags = set->tags[hctx_idx];
2738 if (set->ops->init_hctx &&
2739 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2740 goto unregister_cpu_notifier;
2742 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2748 if (set->ops->exit_hctx)
2749 set->ops->exit_hctx(hctx, hctx_idx);
2750 unregister_cpu_notifier:
2751 blk_mq_remove_cpuhp(hctx);
2755 static struct blk_mq_hw_ctx *
2756 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2759 struct blk_mq_hw_ctx *hctx;
2760 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2762 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2764 goto fail_alloc_hctx;
2766 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2769 atomic_set(&hctx->nr_active, 0);
2770 if (node == NUMA_NO_NODE)
2771 node = set->numa_node;
2772 hctx->numa_node = node;
2774 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2775 spin_lock_init(&hctx->lock);
2776 INIT_LIST_HEAD(&hctx->dispatch);
2778 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
2780 INIT_LIST_HEAD(&hctx->hctx_list);
2783 * Allocate space for all possible cpus to avoid allocation at
2786 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2791 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2792 gfp, node, false, false))
2796 spin_lock_init(&hctx->dispatch_wait_lock);
2797 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2798 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2800 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2804 if (hctx->flags & BLK_MQ_F_BLOCKING)
2805 init_srcu_struct(hctx->srcu);
2806 blk_mq_hctx_kobj_init(hctx);
2811 sbitmap_free(&hctx->ctx_map);
2815 free_cpumask_var(hctx->cpumask);
2822 static void blk_mq_init_cpu_queues(struct request_queue *q,
2823 unsigned int nr_hw_queues)
2825 struct blk_mq_tag_set *set = q->tag_set;
2828 for_each_possible_cpu(i) {
2829 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2830 struct blk_mq_hw_ctx *hctx;
2834 spin_lock_init(&__ctx->lock);
2835 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2836 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2841 * Set local node, IFF we have more than one hw queue. If
2842 * not, we remain on the home node of the device
2844 for (j = 0; j < set->nr_maps; j++) {
2845 hctx = blk_mq_map_queue_type(q, j, i);
2846 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2847 hctx->numa_node = cpu_to_node(i);
2852 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2855 unsigned int flags = set->flags;
2858 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2859 set->queue_depth, set->reserved_tags, flags);
2860 if (!set->tags[hctx_idx])
2863 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2868 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2869 set->tags[hctx_idx] = NULL;
2873 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2874 unsigned int hctx_idx)
2876 unsigned int flags = set->flags;
2878 if (set->tags && set->tags[hctx_idx]) {
2879 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2880 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2881 set->tags[hctx_idx] = NULL;
2885 static void blk_mq_map_swqueue(struct request_queue *q)
2887 unsigned int i, j, hctx_idx;
2888 struct blk_mq_hw_ctx *hctx;
2889 struct blk_mq_ctx *ctx;
2890 struct blk_mq_tag_set *set = q->tag_set;
2892 queue_for_each_hw_ctx(q, hctx, i) {
2893 cpumask_clear(hctx->cpumask);
2895 hctx->dispatch_from = NULL;
2899 * Map software to hardware queues.
2901 * If the cpu isn't present, the cpu is mapped to first hctx.
2903 for_each_possible_cpu(i) {
2905 ctx = per_cpu_ptr(q->queue_ctx, i);
2906 for (j = 0; j < set->nr_maps; j++) {
2907 if (!set->map[j].nr_queues) {
2908 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2909 HCTX_TYPE_DEFAULT, i);
2912 hctx_idx = set->map[j].mq_map[i];
2913 /* unmapped hw queue can be remapped after CPU topo changed */
2914 if (!set->tags[hctx_idx] &&
2915 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2917 * If tags initialization fail for some hctx,
2918 * that hctx won't be brought online. In this
2919 * case, remap the current ctx to hctx[0] which
2920 * is guaranteed to always have tags allocated
2922 set->map[j].mq_map[i] = 0;
2925 hctx = blk_mq_map_queue_type(q, j, i);
2926 ctx->hctxs[j] = hctx;
2928 * If the CPU is already set in the mask, then we've
2929 * mapped this one already. This can happen if
2930 * devices share queues across queue maps.
2932 if (cpumask_test_cpu(i, hctx->cpumask))
2935 cpumask_set_cpu(i, hctx->cpumask);
2937 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2938 hctx->ctxs[hctx->nr_ctx++] = ctx;
2941 * If the nr_ctx type overflows, we have exceeded the
2942 * amount of sw queues we can support.
2944 BUG_ON(!hctx->nr_ctx);
2947 for (; j < HCTX_MAX_TYPES; j++)
2948 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2949 HCTX_TYPE_DEFAULT, i);
2952 queue_for_each_hw_ctx(q, hctx, i) {
2954 * If no software queues are mapped to this hardware queue,
2955 * disable it and free the request entries.
2957 if (!hctx->nr_ctx) {
2958 /* Never unmap queue 0. We need it as a
2959 * fallback in case of a new remap fails
2962 if (i && set->tags[i])
2963 blk_mq_free_map_and_requests(set, i);
2969 hctx->tags = set->tags[i];
2970 WARN_ON(!hctx->tags);
2973 * Set the map size to the number of mapped software queues.
2974 * This is more accurate and more efficient than looping
2975 * over all possibly mapped software queues.
2977 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2980 * Initialize batch roundrobin counts
2982 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2983 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2988 * Caller needs to ensure that we're either frozen/quiesced, or that
2989 * the queue isn't live yet.
2991 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2993 struct blk_mq_hw_ctx *hctx;
2996 queue_for_each_hw_ctx(q, hctx, i) {
2998 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3000 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3004 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3007 struct request_queue *q;
3009 lockdep_assert_held(&set->tag_list_lock);
3011 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3012 blk_mq_freeze_queue(q);
3013 queue_set_hctx_shared(q, shared);
3014 blk_mq_unfreeze_queue(q);
3018 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3020 struct blk_mq_tag_set *set = q->tag_set;
3022 mutex_lock(&set->tag_list_lock);
3023 list_del(&q->tag_set_list);
3024 if (list_is_singular(&set->tag_list)) {
3025 /* just transitioned to unshared */
3026 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3027 /* update existing queue */
3028 blk_mq_update_tag_set_shared(set, false);
3030 mutex_unlock(&set->tag_list_lock);
3031 INIT_LIST_HEAD(&q->tag_set_list);
3034 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3035 struct request_queue *q)
3037 mutex_lock(&set->tag_list_lock);
3040 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3042 if (!list_empty(&set->tag_list) &&
3043 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3044 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3045 /* update existing queue */
3046 blk_mq_update_tag_set_shared(set, true);
3048 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3049 queue_set_hctx_shared(q, true);
3050 list_add_tail(&q->tag_set_list, &set->tag_list);
3052 mutex_unlock(&set->tag_list_lock);
3055 /* All allocations will be freed in release handler of q->mq_kobj */
3056 static int blk_mq_alloc_ctxs(struct request_queue *q)
3058 struct blk_mq_ctxs *ctxs;
3061 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3065 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3066 if (!ctxs->queue_ctx)
3069 for_each_possible_cpu(cpu) {
3070 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3074 q->mq_kobj = &ctxs->kobj;
3075 q->queue_ctx = ctxs->queue_ctx;
3084 * It is the actual release handler for mq, but we do it from
3085 * request queue's release handler for avoiding use-after-free
3086 * and headache because q->mq_kobj shouldn't have been introduced,
3087 * but we can't group ctx/kctx kobj without it.
3089 void blk_mq_release(struct request_queue *q)
3091 struct blk_mq_hw_ctx *hctx, *next;
3094 queue_for_each_hw_ctx(q, hctx, i)
3095 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3097 /* all hctx are in .unused_hctx_list now */
3098 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3099 list_del_init(&hctx->hctx_list);
3100 kobject_put(&hctx->kobj);
3103 kfree(q->queue_hw_ctx);
3106 * release .mq_kobj and sw queue's kobject now because
3107 * both share lifetime with request queue.
3109 blk_mq_sysfs_deinit(q);
3112 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3115 struct request_queue *q;
3118 q = blk_alloc_queue(set->numa_node);
3120 return ERR_PTR(-ENOMEM);
3121 q->queuedata = queuedata;
3122 ret = blk_mq_init_allocated_queue(set, q);
3124 blk_cleanup_queue(q);
3125 return ERR_PTR(ret);
3130 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3132 return blk_mq_init_queue_data(set, NULL);
3134 EXPORT_SYMBOL(blk_mq_init_queue);
3136 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata)
3138 struct request_queue *q;
3139 struct gendisk *disk;
3141 q = blk_mq_init_queue_data(set, queuedata);
3145 disk = __alloc_disk_node(0, set->numa_node);
3147 blk_cleanup_queue(q);
3148 return ERR_PTR(-ENOMEM);
3153 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3155 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3156 struct blk_mq_tag_set *set, struct request_queue *q,
3157 int hctx_idx, int node)
3159 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3161 /* reuse dead hctx first */
3162 spin_lock(&q->unused_hctx_lock);
3163 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3164 if (tmp->numa_node == node) {
3170 list_del_init(&hctx->hctx_list);
3171 spin_unlock(&q->unused_hctx_lock);
3174 hctx = blk_mq_alloc_hctx(q, set, node);
3178 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3184 kobject_put(&hctx->kobj);
3189 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3190 struct request_queue *q)
3193 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3195 if (q->nr_hw_queues < set->nr_hw_queues) {
3196 struct blk_mq_hw_ctx **new_hctxs;
3198 new_hctxs = kcalloc_node(set->nr_hw_queues,
3199 sizeof(*new_hctxs), GFP_KERNEL,
3204 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3206 q->queue_hw_ctx = new_hctxs;
3211 /* protect against switching io scheduler */
3212 mutex_lock(&q->sysfs_lock);
3213 for (i = 0; i < set->nr_hw_queues; i++) {
3215 struct blk_mq_hw_ctx *hctx;
3217 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3219 * If the hw queue has been mapped to another numa node,
3220 * we need to realloc the hctx. If allocation fails, fallback
3221 * to use the previous one.
3223 if (hctxs[i] && (hctxs[i]->numa_node == node))
3226 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3229 blk_mq_exit_hctx(q, set, hctxs[i], i);
3233 pr_warn("Allocate new hctx on node %d fails,\
3234 fallback to previous one on node %d\n",
3235 node, hctxs[i]->numa_node);
3241 * Increasing nr_hw_queues fails. Free the newly allocated
3242 * hctxs and keep the previous q->nr_hw_queues.
3244 if (i != set->nr_hw_queues) {
3245 j = q->nr_hw_queues;
3249 end = q->nr_hw_queues;
3250 q->nr_hw_queues = set->nr_hw_queues;
3253 for (; j < end; j++) {
3254 struct blk_mq_hw_ctx *hctx = hctxs[j];
3258 blk_mq_free_map_and_requests(set, j);
3259 blk_mq_exit_hctx(q, set, hctx, j);
3263 mutex_unlock(&q->sysfs_lock);
3266 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3267 struct request_queue *q)
3269 /* mark the queue as mq asap */
3270 q->mq_ops = set->ops;
3272 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3273 blk_mq_poll_stats_bkt,
3274 BLK_MQ_POLL_STATS_BKTS, q);
3278 if (blk_mq_alloc_ctxs(q))
3281 /* init q->mq_kobj and sw queues' kobjects */
3282 blk_mq_sysfs_init(q);
3284 INIT_LIST_HEAD(&q->unused_hctx_list);
3285 spin_lock_init(&q->unused_hctx_lock);
3287 blk_mq_realloc_hw_ctxs(set, q);
3288 if (!q->nr_hw_queues)
3291 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3292 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3296 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3297 if (set->nr_maps > HCTX_TYPE_POLL &&
3298 set->map[HCTX_TYPE_POLL].nr_queues)
3299 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3301 q->sg_reserved_size = INT_MAX;
3303 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3304 INIT_LIST_HEAD(&q->requeue_list);
3305 spin_lock_init(&q->requeue_lock);
3307 q->nr_requests = set->queue_depth;
3310 * Default to classic polling
3312 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3314 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3315 blk_mq_add_queue_tag_set(set, q);
3316 blk_mq_map_swqueue(q);
3320 kfree(q->queue_hw_ctx);
3321 q->nr_hw_queues = 0;
3322 blk_mq_sysfs_deinit(q);
3324 blk_stat_free_callback(q->poll_cb);
3330 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3332 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3333 void blk_mq_exit_queue(struct request_queue *q)
3335 struct blk_mq_tag_set *set = q->tag_set;
3337 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
3338 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3339 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
3340 blk_mq_del_queue_tag_set(q);
3343 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3347 for (i = 0; i < set->nr_hw_queues; i++) {
3348 if (!__blk_mq_alloc_map_and_request(set, i))
3357 blk_mq_free_map_and_requests(set, i);
3363 * Allocate the request maps associated with this tag_set. Note that this
3364 * may reduce the depth asked for, if memory is tight. set->queue_depth
3365 * will be updated to reflect the allocated depth.
3367 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3372 depth = set->queue_depth;
3374 err = __blk_mq_alloc_rq_maps(set);
3378 set->queue_depth >>= 1;
3379 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3383 } while (set->queue_depth);
3385 if (!set->queue_depth || err) {
3386 pr_err("blk-mq: failed to allocate request map\n");
3390 if (depth != set->queue_depth)
3391 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3392 depth, set->queue_depth);
3397 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3400 * blk_mq_map_queues() and multiple .map_queues() implementations
3401 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3402 * number of hardware queues.
3404 if (set->nr_maps == 1)
3405 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3407 if (set->ops->map_queues && !is_kdump_kernel()) {
3411 * transport .map_queues is usually done in the following
3414 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3415 * mask = get_cpu_mask(queue)
3416 * for_each_cpu(cpu, mask)
3417 * set->map[x].mq_map[cpu] = queue;
3420 * When we need to remap, the table has to be cleared for
3421 * killing stale mapping since one CPU may not be mapped
3424 for (i = 0; i < set->nr_maps; i++)
3425 blk_mq_clear_mq_map(&set->map[i]);
3427 return set->ops->map_queues(set);
3429 BUG_ON(set->nr_maps > 1);
3430 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3434 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3435 int cur_nr_hw_queues, int new_nr_hw_queues)
3437 struct blk_mq_tags **new_tags;
3439 if (cur_nr_hw_queues >= new_nr_hw_queues)
3442 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3443 GFP_KERNEL, set->numa_node);
3448 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3449 sizeof(*set->tags));
3451 set->tags = new_tags;
3452 set->nr_hw_queues = new_nr_hw_queues;
3457 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
3458 int new_nr_hw_queues)
3460 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
3464 * Alloc a tag set to be associated with one or more request queues.
3465 * May fail with EINVAL for various error conditions. May adjust the
3466 * requested depth down, if it's too large. In that case, the set
3467 * value will be stored in set->queue_depth.
3469 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3473 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3475 if (!set->nr_hw_queues)
3477 if (!set->queue_depth)
3479 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3482 if (!set->ops->queue_rq)
3485 if (!set->ops->get_budget ^ !set->ops->put_budget)
3488 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3489 pr_info("blk-mq: reduced tag depth to %u\n",
3491 set->queue_depth = BLK_MQ_MAX_DEPTH;
3496 else if (set->nr_maps > HCTX_MAX_TYPES)
3500 * If a crashdump is active, then we are potentially in a very
3501 * memory constrained environment. Limit us to 1 queue and
3502 * 64 tags to prevent using too much memory.
3504 if (is_kdump_kernel()) {
3505 set->nr_hw_queues = 1;
3507 set->queue_depth = min(64U, set->queue_depth);
3510 * There is no use for more h/w queues than cpus if we just have
3513 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3514 set->nr_hw_queues = nr_cpu_ids;
3516 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
3520 for (i = 0; i < set->nr_maps; i++) {
3521 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3522 sizeof(set->map[i].mq_map[0]),
3523 GFP_KERNEL, set->numa_node);
3524 if (!set->map[i].mq_map)
3525 goto out_free_mq_map;
3526 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3529 ret = blk_mq_update_queue_map(set);
3531 goto out_free_mq_map;
3533 ret = blk_mq_alloc_map_and_requests(set);
3535 goto out_free_mq_map;
3537 if (blk_mq_is_sbitmap_shared(set->flags)) {
3538 atomic_set(&set->active_queues_shared_sbitmap, 0);
3540 if (blk_mq_init_shared_sbitmap(set)) {
3542 goto out_free_mq_rq_maps;
3546 mutex_init(&set->tag_list_lock);
3547 INIT_LIST_HEAD(&set->tag_list);
3551 out_free_mq_rq_maps:
3552 for (i = 0; i < set->nr_hw_queues; i++)
3553 blk_mq_free_map_and_requests(set, i);
3555 for (i = 0; i < set->nr_maps; i++) {
3556 kfree(set->map[i].mq_map);
3557 set->map[i].mq_map = NULL;
3563 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3565 /* allocate and initialize a tagset for a simple single-queue device */
3566 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
3567 const struct blk_mq_ops *ops, unsigned int queue_depth,
3568 unsigned int set_flags)
3570 memset(set, 0, sizeof(*set));
3572 set->nr_hw_queues = 1;
3574 set->queue_depth = queue_depth;
3575 set->numa_node = NUMA_NO_NODE;
3576 set->flags = set_flags;
3577 return blk_mq_alloc_tag_set(set);
3579 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
3581 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3585 for (i = 0; i < set->nr_hw_queues; i++)
3586 blk_mq_free_map_and_requests(set, i);
3588 if (blk_mq_is_sbitmap_shared(set->flags))
3589 blk_mq_exit_shared_sbitmap(set);
3591 for (j = 0; j < set->nr_maps; j++) {
3592 kfree(set->map[j].mq_map);
3593 set->map[j].mq_map = NULL;
3599 EXPORT_SYMBOL(blk_mq_free_tag_set);
3601 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3603 struct blk_mq_tag_set *set = q->tag_set;
3604 struct blk_mq_hw_ctx *hctx;
3610 if (q->nr_requests == nr)
3613 blk_mq_freeze_queue(q);
3614 blk_mq_quiesce_queue(q);
3617 queue_for_each_hw_ctx(q, hctx, i) {
3621 * If we're using an MQ scheduler, just update the scheduler
3622 * queue depth. This is similar to what the old code would do.
3624 if (!hctx->sched_tags) {
3625 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3627 if (!ret && blk_mq_is_sbitmap_shared(set->flags))
3628 blk_mq_tag_resize_shared_sbitmap(set, nr);
3630 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3632 if (blk_mq_is_sbitmap_shared(set->flags)) {
3633 hctx->sched_tags->bitmap_tags =
3634 &q->sched_bitmap_tags;
3635 hctx->sched_tags->breserved_tags =
3636 &q->sched_breserved_tags;
3641 if (q->elevator && q->elevator->type->ops.depth_updated)
3642 q->elevator->type->ops.depth_updated(hctx);
3645 q->nr_requests = nr;
3646 if (q->elevator && blk_mq_is_sbitmap_shared(set->flags))
3647 sbitmap_queue_resize(&q->sched_bitmap_tags,
3648 nr - set->reserved_tags);
3651 blk_mq_unquiesce_queue(q);
3652 blk_mq_unfreeze_queue(q);
3658 * request_queue and elevator_type pair.
3659 * It is just used by __blk_mq_update_nr_hw_queues to cache
3660 * the elevator_type associated with a request_queue.
3662 struct blk_mq_qe_pair {
3663 struct list_head node;
3664 struct request_queue *q;
3665 struct elevator_type *type;
3669 * Cache the elevator_type in qe pair list and switch the
3670 * io scheduler to 'none'
3672 static bool blk_mq_elv_switch_none(struct list_head *head,
3673 struct request_queue *q)
3675 struct blk_mq_qe_pair *qe;
3680 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3684 INIT_LIST_HEAD(&qe->node);
3686 qe->type = q->elevator->type;
3687 list_add(&qe->node, head);
3689 mutex_lock(&q->sysfs_lock);
3691 * After elevator_switch_mq, the previous elevator_queue will be
3692 * released by elevator_release. The reference of the io scheduler
3693 * module get by elevator_get will also be put. So we need to get
3694 * a reference of the io scheduler module here to prevent it to be
3697 __module_get(qe->type->elevator_owner);
3698 elevator_switch_mq(q, NULL);
3699 mutex_unlock(&q->sysfs_lock);
3704 static void blk_mq_elv_switch_back(struct list_head *head,
3705 struct request_queue *q)
3707 struct blk_mq_qe_pair *qe;
3708 struct elevator_type *t = NULL;
3710 list_for_each_entry(qe, head, node)
3719 list_del(&qe->node);
3722 mutex_lock(&q->sysfs_lock);
3723 elevator_switch_mq(q, t);
3724 mutex_unlock(&q->sysfs_lock);
3727 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3730 struct request_queue *q;
3732 int prev_nr_hw_queues;
3734 lockdep_assert_held(&set->tag_list_lock);
3736 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3737 nr_hw_queues = nr_cpu_ids;
3738 if (nr_hw_queues < 1)
3740 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3743 list_for_each_entry(q, &set->tag_list, tag_set_list)
3744 blk_mq_freeze_queue(q);
3746 * Switch IO scheduler to 'none', cleaning up the data associated
3747 * with the previous scheduler. We will switch back once we are done
3748 * updating the new sw to hw queue mappings.
3750 list_for_each_entry(q, &set->tag_list, tag_set_list)
3751 if (!blk_mq_elv_switch_none(&head, q))
3754 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3755 blk_mq_debugfs_unregister_hctxs(q);
3756 blk_mq_sysfs_unregister(q);
3759 prev_nr_hw_queues = set->nr_hw_queues;
3760 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3764 set->nr_hw_queues = nr_hw_queues;
3766 blk_mq_update_queue_map(set);
3767 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3768 blk_mq_realloc_hw_ctxs(set, q);
3769 if (q->nr_hw_queues != set->nr_hw_queues) {
3770 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3771 nr_hw_queues, prev_nr_hw_queues);
3772 set->nr_hw_queues = prev_nr_hw_queues;
3773 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3776 blk_mq_map_swqueue(q);
3780 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3781 blk_mq_sysfs_register(q);
3782 blk_mq_debugfs_register_hctxs(q);
3786 list_for_each_entry(q, &set->tag_list, tag_set_list)
3787 blk_mq_elv_switch_back(&head, q);
3789 list_for_each_entry(q, &set->tag_list, tag_set_list)
3790 blk_mq_unfreeze_queue(q);
3793 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3795 mutex_lock(&set->tag_list_lock);
3796 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3797 mutex_unlock(&set->tag_list_lock);
3799 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3801 /* Enable polling stats and return whether they were already enabled. */
3802 static bool blk_poll_stats_enable(struct request_queue *q)
3804 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3805 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3807 blk_stat_add_callback(q, q->poll_cb);
3811 static void blk_mq_poll_stats_start(struct request_queue *q)
3814 * We don't arm the callback if polling stats are not enabled or the
3815 * callback is already active.
3817 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3818 blk_stat_is_active(q->poll_cb))
3821 blk_stat_activate_msecs(q->poll_cb, 100);
3824 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3826 struct request_queue *q = cb->data;
3829 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3830 if (cb->stat[bucket].nr_samples)
3831 q->poll_stat[bucket] = cb->stat[bucket];
3835 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3838 unsigned long ret = 0;
3842 * If stats collection isn't on, don't sleep but turn it on for
3845 if (!blk_poll_stats_enable(q))
3849 * As an optimistic guess, use half of the mean service time
3850 * for this type of request. We can (and should) make this smarter.
3851 * For instance, if the completion latencies are tight, we can
3852 * get closer than just half the mean. This is especially
3853 * important on devices where the completion latencies are longer
3854 * than ~10 usec. We do use the stats for the relevant IO size
3855 * if available which does lead to better estimates.
3857 bucket = blk_mq_poll_stats_bkt(rq);
3861 if (q->poll_stat[bucket].nr_samples)
3862 ret = (q->poll_stat[bucket].mean + 1) / 2;
3867 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3870 struct hrtimer_sleeper hs;
3871 enum hrtimer_mode mode;
3875 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3879 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3881 * 0: use half of prev avg
3882 * >0: use this specific value
3884 if (q->poll_nsec > 0)
3885 nsecs = q->poll_nsec;
3887 nsecs = blk_mq_poll_nsecs(q, rq);
3892 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3895 * This will be replaced with the stats tracking code, using
3896 * 'avg_completion_time / 2' as the pre-sleep target.
3900 mode = HRTIMER_MODE_REL;
3901 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3902 hrtimer_set_expires(&hs.timer, kt);
3905 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3907 set_current_state(TASK_UNINTERRUPTIBLE);
3908 hrtimer_sleeper_start_expires(&hs, mode);
3911 hrtimer_cancel(&hs.timer);
3912 mode = HRTIMER_MODE_ABS;
3913 } while (hs.task && !signal_pending(current));
3915 __set_current_state(TASK_RUNNING);
3916 destroy_hrtimer_on_stack(&hs.timer);
3920 static bool blk_mq_poll_hybrid(struct request_queue *q,
3921 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3925 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3928 if (!blk_qc_t_is_internal(cookie))
3929 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3931 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3933 * With scheduling, if the request has completed, we'll
3934 * get a NULL return here, as we clear the sched tag when
3935 * that happens. The request still remains valid, like always,
3936 * so we should be safe with just the NULL check.
3942 return blk_mq_poll_hybrid_sleep(q, rq);
3946 * blk_poll - poll for IO completions
3948 * @cookie: cookie passed back at IO submission time
3949 * @spin: whether to spin for completions
3952 * Poll for completions on the passed in queue. Returns number of
3953 * completed entries found. If @spin is true, then blk_poll will continue
3954 * looping until at least one completion is found, unless the task is
3955 * otherwise marked running (or we need to reschedule).
3957 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3959 struct blk_mq_hw_ctx *hctx;
3962 if (!blk_qc_t_valid(cookie) ||
3963 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3967 blk_flush_plug_list(current->plug, false);
3969 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3972 * If we sleep, have the caller restart the poll loop to reset
3973 * the state. Like for the other success return cases, the
3974 * caller is responsible for checking if the IO completed. If
3975 * the IO isn't complete, we'll get called again and will go
3976 * straight to the busy poll loop. If specified not to spin,
3977 * we also should not sleep.
3979 if (spin && blk_mq_poll_hybrid(q, hctx, cookie))
3982 hctx->poll_considered++;
3984 state = get_current_state();
3988 hctx->poll_invoked++;
3990 ret = q->mq_ops->poll(hctx);
3992 hctx->poll_success++;
3993 __set_current_state(TASK_RUNNING);
3997 if (signal_pending_state(state, current))
3998 __set_current_state(TASK_RUNNING);
4000 if (task_is_running(current))
4002 if (ret < 0 || !spin)
4005 } while (!need_resched());
4007 __set_current_state(TASK_RUNNING);
4010 EXPORT_SYMBOL_GPL(blk_poll);
4012 unsigned int blk_mq_rq_cpu(struct request *rq)
4014 return rq->mq_ctx->cpu;
4016 EXPORT_SYMBOL(blk_mq_rq_cpu);
4018 static int __init blk_mq_init(void)
4022 for_each_possible_cpu(i)
4023 init_llist_head(&per_cpu(blk_cpu_done, i));
4024 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4026 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4027 "block/softirq:dead", NULL,
4028 blk_softirq_cpu_dead);
4029 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4030 blk_mq_hctx_notify_dead);
4031 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4032 blk_mq_hctx_notify_online,
4033 blk_mq_hctx_notify_offline);
4036 subsys_initcall(blk_mq_init);