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 list_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 hd_struct *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 (rq->part == mi->part)
109 mi->inflight[rq_data_dir(rq)]++;
114 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
116 struct mq_inflight mi = { .part = part };
118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
120 return mi.inflight[0] + mi.inflight[1];
123 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
124 unsigned int inflight[2])
126 struct mq_inflight mi = { .part = part };
128 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
129 inflight[0] = mi.inflight[0];
130 inflight[1] = mi.inflight[1];
133 void blk_freeze_queue_start(struct request_queue *q)
135 mutex_lock(&q->mq_freeze_lock);
136 if (++q->mq_freeze_depth == 1) {
137 percpu_ref_kill(&q->q_usage_counter);
138 mutex_unlock(&q->mq_freeze_lock);
140 blk_mq_run_hw_queues(q, false);
142 mutex_unlock(&q->mq_freeze_lock);
145 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
147 void blk_mq_freeze_queue_wait(struct request_queue *q)
149 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
151 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
153 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
154 unsigned long timeout)
156 return wait_event_timeout(q->mq_freeze_wq,
157 percpu_ref_is_zero(&q->q_usage_counter),
160 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
163 * Guarantee no request is in use, so we can change any data structure of
164 * the queue afterward.
166 void blk_freeze_queue(struct request_queue *q)
169 * In the !blk_mq case we are only calling this to kill the
170 * q_usage_counter, otherwise this increases the freeze depth
171 * and waits for it to return to zero. For this reason there is
172 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
173 * exported to drivers as the only user for unfreeze is blk_mq.
175 blk_freeze_queue_start(q);
176 blk_mq_freeze_queue_wait(q);
179 void blk_mq_freeze_queue(struct request_queue *q)
182 * ...just an alias to keep freeze and unfreeze actions balanced
183 * in the blk_mq_* namespace
187 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
189 void blk_mq_unfreeze_queue(struct request_queue *q)
191 mutex_lock(&q->mq_freeze_lock);
192 q->mq_freeze_depth--;
193 WARN_ON_ONCE(q->mq_freeze_depth < 0);
194 if (!q->mq_freeze_depth) {
195 percpu_ref_resurrect(&q->q_usage_counter);
196 wake_up_all(&q->mq_freeze_wq);
198 mutex_unlock(&q->mq_freeze_lock);
200 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
203 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
204 * mpt3sas driver such that this function can be removed.
206 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
208 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
210 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
213 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
216 * Note: this function does not prevent that the struct request end_io()
217 * callback function is invoked. Once this function is returned, we make
218 * sure no dispatch can happen until the queue is unquiesced via
219 * blk_mq_unquiesce_queue().
221 void blk_mq_quiesce_queue(struct request_queue *q)
223 struct blk_mq_hw_ctx *hctx;
227 blk_mq_quiesce_queue_nowait(q);
229 queue_for_each_hw_ctx(q, hctx, i) {
230 if (hctx->flags & BLK_MQ_F_BLOCKING)
231 synchronize_srcu(hctx->srcu);
238 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
241 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
244 * This function recovers queue into the state before quiescing
245 * which is done by blk_mq_quiesce_queue.
247 void blk_mq_unquiesce_queue(struct request_queue *q)
249 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
251 /* dispatch requests which are inserted during quiescing */
252 blk_mq_run_hw_queues(q, true);
254 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
256 void blk_mq_wake_waiters(struct request_queue *q)
258 struct blk_mq_hw_ctx *hctx;
261 queue_for_each_hw_ctx(q, hctx, i)
262 if (blk_mq_hw_queue_mapped(hctx))
263 blk_mq_tag_wakeup_all(hctx->tags, true);
267 * Only need start/end time stamping if we have iostat or
268 * blk stats enabled, or using an IO scheduler.
270 static inline bool blk_mq_need_time_stamp(struct request *rq)
272 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
275 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
276 unsigned int tag, u64 alloc_time_ns)
278 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
279 struct request *rq = tags->static_rqs[tag];
281 if (data->q->elevator) {
282 rq->tag = BLK_MQ_NO_TAG;
283 rq->internal_tag = tag;
286 rq->internal_tag = BLK_MQ_NO_TAG;
289 /* csd/requeue_work/fifo_time is initialized before use */
291 rq->mq_ctx = data->ctx;
292 rq->mq_hctx = data->hctx;
294 rq->cmd_flags = data->cmd_flags;
295 if (data->flags & BLK_MQ_REQ_PREEMPT)
296 rq->rq_flags |= RQF_PREEMPT;
297 if (blk_queue_io_stat(data->q))
298 rq->rq_flags |= RQF_IO_STAT;
299 INIT_LIST_HEAD(&rq->queuelist);
300 INIT_HLIST_NODE(&rq->hash);
301 RB_CLEAR_NODE(&rq->rb_node);
304 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
305 rq->alloc_time_ns = alloc_time_ns;
307 if (blk_mq_need_time_stamp(rq))
308 rq->start_time_ns = ktime_get_ns();
310 rq->start_time_ns = 0;
311 rq->io_start_time_ns = 0;
312 rq->stats_sectors = 0;
313 rq->nr_phys_segments = 0;
314 #if defined(CONFIG_BLK_DEV_INTEGRITY)
315 rq->nr_integrity_segments = 0;
317 blk_crypto_rq_set_defaults(rq);
318 /* tag was already set */
319 WRITE_ONCE(rq->deadline, 0);
324 rq->end_io_data = NULL;
326 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
327 refcount_set(&rq->ref, 1);
329 if (!op_is_flush(data->cmd_flags)) {
330 struct elevator_queue *e = data->q->elevator;
333 if (e && e->type->ops.prepare_request) {
334 if (e->type->icq_cache)
335 blk_mq_sched_assign_ioc(rq);
337 e->type->ops.prepare_request(rq);
338 rq->rq_flags |= RQF_ELVPRIV;
342 data->hctx->queued++;
346 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
348 struct request_queue *q = data->q;
349 struct elevator_queue *e = q->elevator;
350 u64 alloc_time_ns = 0;
353 /* alloc_time includes depth and tag waits */
354 if (blk_queue_rq_alloc_time(q))
355 alloc_time_ns = ktime_get_ns();
357 if (data->cmd_flags & REQ_NOWAIT)
358 data->flags |= BLK_MQ_REQ_NOWAIT;
362 * Flush requests are special and go directly to the
363 * dispatch list. Don't include reserved tags in the
364 * limiting, as it isn't useful.
366 if (!op_is_flush(data->cmd_flags) &&
367 e->type->ops.limit_depth &&
368 !(data->flags & BLK_MQ_REQ_RESERVED))
369 e->type->ops.limit_depth(data->cmd_flags, data);
373 data->ctx = blk_mq_get_ctx(q);
374 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
376 blk_mq_tag_busy(data->hctx);
379 * Waiting allocations only fail because of an inactive hctx. In that
380 * case just retry the hctx assignment and tag allocation as CPU hotplug
381 * should have migrated us to an online CPU by now.
383 tag = blk_mq_get_tag(data);
384 if (tag == BLK_MQ_NO_TAG) {
385 if (data->flags & BLK_MQ_REQ_NOWAIT)
389 * Give up the CPU and sleep for a random short time to ensure
390 * that thread using a realtime scheduling class are migrated
391 * off the the CPU, and thus off the hctx that is going away.
396 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
399 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
400 blk_mq_req_flags_t flags)
402 struct blk_mq_alloc_data data = {
410 ret = blk_queue_enter(q, flags);
414 rq = __blk_mq_alloc_request(&data);
418 rq->__sector = (sector_t) -1;
419 rq->bio = rq->biotail = NULL;
423 return ERR_PTR(-EWOULDBLOCK);
425 EXPORT_SYMBOL(blk_mq_alloc_request);
427 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
428 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
430 struct blk_mq_alloc_data data = {
435 u64 alloc_time_ns = 0;
440 /* alloc_time includes depth and tag waits */
441 if (blk_queue_rq_alloc_time(q))
442 alloc_time_ns = ktime_get_ns();
445 * If the tag allocator sleeps we could get an allocation for a
446 * different hardware context. No need to complicate the low level
447 * allocator for this for the rare use case of a command tied to
450 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
451 return ERR_PTR(-EINVAL);
453 if (hctx_idx >= q->nr_hw_queues)
454 return ERR_PTR(-EIO);
456 ret = blk_queue_enter(q, flags);
461 * Check if the hardware context is actually mapped to anything.
462 * If not tell the caller that it should skip this queue.
465 data.hctx = q->queue_hw_ctx[hctx_idx];
466 if (!blk_mq_hw_queue_mapped(data.hctx))
468 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
469 data.ctx = __blk_mq_get_ctx(q, cpu);
472 blk_mq_tag_busy(data.hctx);
475 tag = blk_mq_get_tag(&data);
476 if (tag == BLK_MQ_NO_TAG)
478 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
484 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
486 static void __blk_mq_free_request(struct request *rq)
488 struct request_queue *q = rq->q;
489 struct blk_mq_ctx *ctx = rq->mq_ctx;
490 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
491 const int sched_tag = rq->internal_tag;
493 blk_crypto_free_request(rq);
494 blk_pm_mark_last_busy(rq);
496 if (rq->tag != BLK_MQ_NO_TAG)
497 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
498 if (sched_tag != BLK_MQ_NO_TAG)
499 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
500 blk_mq_sched_restart(hctx);
504 void blk_mq_free_request(struct request *rq)
506 struct request_queue *q = rq->q;
507 struct elevator_queue *e = q->elevator;
508 struct blk_mq_ctx *ctx = rq->mq_ctx;
509 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
511 if (rq->rq_flags & RQF_ELVPRIV) {
512 if (e && e->type->ops.finish_request)
513 e->type->ops.finish_request(rq);
515 put_io_context(rq->elv.icq->ioc);
520 ctx->rq_completed[rq_is_sync(rq)]++;
521 if (rq->rq_flags & RQF_MQ_INFLIGHT)
522 atomic_dec(&hctx->nr_active);
524 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
525 laptop_io_completion(q->backing_dev_info);
529 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
530 if (refcount_dec_and_test(&rq->ref))
531 __blk_mq_free_request(rq);
533 EXPORT_SYMBOL_GPL(blk_mq_free_request);
535 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
539 if (blk_mq_need_time_stamp(rq))
540 now = ktime_get_ns();
542 if (rq->rq_flags & RQF_STATS) {
543 blk_mq_poll_stats_start(rq->q);
544 blk_stat_add(rq, now);
547 if (rq->internal_tag != BLK_MQ_NO_TAG)
548 blk_mq_sched_completed_request(rq, now);
550 blk_account_io_done(rq, now);
553 rq_qos_done(rq->q, rq);
554 rq->end_io(rq, error);
556 blk_mq_free_request(rq);
559 EXPORT_SYMBOL(__blk_mq_end_request);
561 void blk_mq_end_request(struct request *rq, blk_status_t error)
563 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
565 __blk_mq_end_request(rq, error);
567 EXPORT_SYMBOL(blk_mq_end_request);
570 * Softirq action handler - move entries to local list and loop over them
571 * while passing them to the queue registered handler.
573 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
575 struct list_head *cpu_list, local_list;
578 cpu_list = this_cpu_ptr(&blk_cpu_done);
579 list_replace_init(cpu_list, &local_list);
582 while (!list_empty(&local_list)) {
585 rq = list_entry(local_list.next, struct request, ipi_list);
586 list_del_init(&rq->ipi_list);
587 rq->q->mq_ops->complete(rq);
591 static void blk_mq_trigger_softirq(struct request *rq)
593 struct list_head *list;
596 local_irq_save(flags);
597 list = this_cpu_ptr(&blk_cpu_done);
598 list_add_tail(&rq->ipi_list, list);
601 * If the list only contains our just added request, signal a raise of
602 * the softirq. If there are already entries there, someone already
603 * raised the irq but it hasn't run yet.
605 if (list->next == &rq->ipi_list)
606 raise_softirq_irqoff(BLOCK_SOFTIRQ);
607 local_irq_restore(flags);
610 static int blk_softirq_cpu_dead(unsigned int cpu)
613 * If a CPU goes away, splice its entries to the current CPU
614 * and trigger a run of the softirq
617 list_splice_init(&per_cpu(blk_cpu_done, cpu),
618 this_cpu_ptr(&blk_cpu_done));
619 raise_softirq_irqoff(BLOCK_SOFTIRQ);
626 static void __blk_mq_complete_request_remote(void *data)
628 struct request *rq = data;
631 * For most of single queue controllers, there is only one irq vector
632 * for handling I/O completion, and the only irq's affinity is set
633 * to all possible CPUs. On most of ARCHs, this affinity means the irq
634 * is handled on one specific CPU.
636 * So complete I/O requests in softirq context in case of single queue
637 * devices to avoid degrading I/O performance due to irqsoff latency.
639 if (rq->q->nr_hw_queues == 1)
640 blk_mq_trigger_softirq(rq);
642 rq->q->mq_ops->complete(rq);
645 static inline bool blk_mq_complete_need_ipi(struct request *rq)
647 int cpu = raw_smp_processor_id();
649 if (!IS_ENABLED(CONFIG_SMP) ||
650 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
653 /* same CPU or cache domain? Complete locally */
654 if (cpu == rq->mq_ctx->cpu ||
655 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
656 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
659 /* don't try to IPI to an offline CPU */
660 return cpu_online(rq->mq_ctx->cpu);
663 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
666 blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
667 rq->tag = BLK_MQ_NO_TAG;
669 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
670 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
671 atomic_dec(&hctx->nr_active);
675 static inline void blk_mq_put_driver_tag(struct request *rq)
677 if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
680 __blk_mq_put_driver_tag(rq->mq_hctx, rq);
683 bool blk_mq_complete_request_remote(struct request *rq)
685 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
687 blk_mq_put_driver_tag(rq);
690 * For a polled request, always complete locallly, it's pointless
691 * to redirect the completion.
693 if (rq->cmd_flags & REQ_HIPRI)
696 if (blk_mq_complete_need_ipi(rq)) {
697 rq->csd.func = __blk_mq_complete_request_remote;
700 smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
702 if (rq->q->nr_hw_queues > 1)
704 blk_mq_trigger_softirq(rq);
709 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
712 * blk_mq_complete_request - end I/O on a request
713 * @rq: the request being processed
716 * Complete a request by scheduling the ->complete_rq operation.
718 void blk_mq_complete_request(struct request *rq)
720 if (!blk_mq_complete_request_remote(rq))
721 rq->q->mq_ops->complete(rq);
723 EXPORT_SYMBOL(blk_mq_complete_request);
725 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
726 __releases(hctx->srcu)
728 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
731 srcu_read_unlock(hctx->srcu, srcu_idx);
734 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
735 __acquires(hctx->srcu)
737 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
738 /* shut up gcc false positive */
742 *srcu_idx = srcu_read_lock(hctx->srcu);
746 * blk_mq_start_request - Start processing a request
747 * @rq: Pointer to request to be started
749 * Function used by device drivers to notify the block layer that a request
750 * is going to be processed now, so blk layer can do proper initializations
751 * such as starting the timeout timer.
753 void blk_mq_start_request(struct request *rq)
755 struct request_queue *q = rq->q;
757 trace_block_rq_issue(q, rq);
759 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
760 rq->io_start_time_ns = ktime_get_ns();
761 rq->stats_sectors = blk_rq_sectors(rq);
762 rq->rq_flags |= RQF_STATS;
766 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
769 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
771 #ifdef CONFIG_BLK_DEV_INTEGRITY
772 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
773 q->integrity.profile->prepare_fn(rq);
776 EXPORT_SYMBOL(blk_mq_start_request);
778 static void __blk_mq_requeue_request(struct request *rq)
780 struct request_queue *q = rq->q;
782 blk_mq_put_driver_tag(rq);
784 trace_block_rq_requeue(q, rq);
785 rq_qos_requeue(q, rq);
787 if (blk_mq_request_started(rq)) {
788 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
789 rq->rq_flags &= ~RQF_TIMED_OUT;
793 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
795 __blk_mq_requeue_request(rq);
797 /* this request will be re-inserted to io scheduler queue */
798 blk_mq_sched_requeue_request(rq);
800 BUG_ON(!list_empty(&rq->queuelist));
801 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
803 EXPORT_SYMBOL(blk_mq_requeue_request);
805 static void blk_mq_requeue_work(struct work_struct *work)
807 struct request_queue *q =
808 container_of(work, struct request_queue, requeue_work.work);
810 struct request *rq, *next;
812 spin_lock_irq(&q->requeue_lock);
813 list_splice_init(&q->requeue_list, &rq_list);
814 spin_unlock_irq(&q->requeue_lock);
816 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
817 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
820 rq->rq_flags &= ~RQF_SOFTBARRIER;
821 list_del_init(&rq->queuelist);
823 * If RQF_DONTPREP, rq has contained some driver specific
824 * data, so insert it to hctx dispatch list to avoid any
827 if (rq->rq_flags & RQF_DONTPREP)
828 blk_mq_request_bypass_insert(rq, false, false);
830 blk_mq_sched_insert_request(rq, true, false, false);
833 while (!list_empty(&rq_list)) {
834 rq = list_entry(rq_list.next, struct request, queuelist);
835 list_del_init(&rq->queuelist);
836 blk_mq_sched_insert_request(rq, false, false, false);
839 blk_mq_run_hw_queues(q, false);
842 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
843 bool kick_requeue_list)
845 struct request_queue *q = rq->q;
849 * We abuse this flag that is otherwise used by the I/O scheduler to
850 * request head insertion from the workqueue.
852 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
854 spin_lock_irqsave(&q->requeue_lock, flags);
856 rq->rq_flags |= RQF_SOFTBARRIER;
857 list_add(&rq->queuelist, &q->requeue_list);
859 list_add_tail(&rq->queuelist, &q->requeue_list);
861 spin_unlock_irqrestore(&q->requeue_lock, flags);
863 if (kick_requeue_list)
864 blk_mq_kick_requeue_list(q);
867 void blk_mq_kick_requeue_list(struct request_queue *q)
869 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
871 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
873 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
876 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
877 msecs_to_jiffies(msecs));
879 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
881 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
883 if (tag < tags->nr_tags) {
884 prefetch(tags->rqs[tag]);
885 return tags->rqs[tag];
890 EXPORT_SYMBOL(blk_mq_tag_to_rq);
892 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
893 void *priv, bool reserved)
896 * If we find a request that is inflight and the queue matches,
897 * we know the queue is busy. Return false to stop the iteration.
899 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
909 bool blk_mq_queue_inflight(struct request_queue *q)
913 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
916 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
918 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
920 req->rq_flags |= RQF_TIMED_OUT;
921 if (req->q->mq_ops->timeout) {
922 enum blk_eh_timer_return ret;
924 ret = req->q->mq_ops->timeout(req, reserved);
925 if (ret == BLK_EH_DONE)
927 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
933 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
935 unsigned long deadline;
937 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
939 if (rq->rq_flags & RQF_TIMED_OUT)
942 deadline = READ_ONCE(rq->deadline);
943 if (time_after_eq(jiffies, deadline))
948 else if (time_after(*next, deadline))
953 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
954 struct request *rq, void *priv, bool reserved)
956 unsigned long *next = priv;
959 * Just do a quick check if it is expired before locking the request in
960 * so we're not unnecessarilly synchronizing across CPUs.
962 if (!blk_mq_req_expired(rq, next))
966 * We have reason to believe the request may be expired. Take a
967 * reference on the request to lock this request lifetime into its
968 * currently allocated context to prevent it from being reallocated in
969 * the event the completion by-passes this timeout handler.
971 * If the reference was already released, then the driver beat the
972 * timeout handler to posting a natural completion.
974 if (!refcount_inc_not_zero(&rq->ref))
978 * The request is now locked and cannot be reallocated underneath the
979 * timeout handler's processing. Re-verify this exact request is truly
980 * expired; if it is not expired, then the request was completed and
981 * reallocated as a new request.
983 if (blk_mq_req_expired(rq, next))
984 blk_mq_rq_timed_out(rq, reserved);
986 if (is_flush_rq(rq, hctx))
988 else if (refcount_dec_and_test(&rq->ref))
989 __blk_mq_free_request(rq);
994 static void blk_mq_timeout_work(struct work_struct *work)
996 struct request_queue *q =
997 container_of(work, struct request_queue, timeout_work);
998 unsigned long next = 0;
999 struct blk_mq_hw_ctx *hctx;
1002 /* A deadlock might occur if a request is stuck requiring a
1003 * timeout at the same time a queue freeze is waiting
1004 * completion, since the timeout code would not be able to
1005 * acquire the queue reference here.
1007 * That's why we don't use blk_queue_enter here; instead, we use
1008 * percpu_ref_tryget directly, because we need to be able to
1009 * obtain a reference even in the short window between the queue
1010 * starting to freeze, by dropping the first reference in
1011 * blk_freeze_queue_start, and the moment the last request is
1012 * consumed, marked by the instant q_usage_counter reaches
1015 if (!percpu_ref_tryget(&q->q_usage_counter))
1018 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1021 mod_timer(&q->timeout, next);
1024 * Request timeouts are handled as a forward rolling timer. If
1025 * we end up here it means that no requests are pending and
1026 * also that no request has been pending for a while. Mark
1027 * each hctx as idle.
1029 queue_for_each_hw_ctx(q, hctx, i) {
1030 /* the hctx may be unmapped, so check it here */
1031 if (blk_mq_hw_queue_mapped(hctx))
1032 blk_mq_tag_idle(hctx);
1038 struct flush_busy_ctx_data {
1039 struct blk_mq_hw_ctx *hctx;
1040 struct list_head *list;
1043 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1045 struct flush_busy_ctx_data *flush_data = data;
1046 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1047 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1048 enum hctx_type type = hctx->type;
1050 spin_lock(&ctx->lock);
1051 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1052 sbitmap_clear_bit(sb, bitnr);
1053 spin_unlock(&ctx->lock);
1058 * Process software queues that have been marked busy, splicing them
1059 * to the for-dispatch
1061 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1063 struct flush_busy_ctx_data data = {
1068 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1070 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1072 struct dispatch_rq_data {
1073 struct blk_mq_hw_ctx *hctx;
1077 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1080 struct dispatch_rq_data *dispatch_data = data;
1081 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1082 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1083 enum hctx_type type = hctx->type;
1085 spin_lock(&ctx->lock);
1086 if (!list_empty(&ctx->rq_lists[type])) {
1087 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1088 list_del_init(&dispatch_data->rq->queuelist);
1089 if (list_empty(&ctx->rq_lists[type]))
1090 sbitmap_clear_bit(sb, bitnr);
1092 spin_unlock(&ctx->lock);
1094 return !dispatch_data->rq;
1097 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1098 struct blk_mq_ctx *start)
1100 unsigned off = start ? start->index_hw[hctx->type] : 0;
1101 struct dispatch_rq_data data = {
1106 __sbitmap_for_each_set(&hctx->ctx_map, off,
1107 dispatch_rq_from_ctx, &data);
1112 static inline unsigned int queued_to_index(unsigned int queued)
1117 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1120 static bool __blk_mq_get_driver_tag(struct request *rq)
1122 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1123 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1126 blk_mq_tag_busy(rq->mq_hctx);
1128 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1129 bt = &rq->mq_hctx->tags->breserved_tags;
1133 if (!hctx_may_queue(rq->mq_hctx, bt))
1135 tag = __sbitmap_queue_get(bt);
1136 if (tag == BLK_MQ_NO_TAG)
1139 rq->tag = tag + tag_offset;
1143 static bool blk_mq_get_driver_tag(struct request *rq)
1145 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1147 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1150 if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
1151 rq->rq_flags |= RQF_MQ_INFLIGHT;
1152 atomic_inc(&hctx->nr_active);
1154 hctx->tags->rqs[rq->tag] = rq;
1158 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1159 int flags, void *key)
1161 struct blk_mq_hw_ctx *hctx;
1163 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1165 spin_lock(&hctx->dispatch_wait_lock);
1166 if (!list_empty(&wait->entry)) {
1167 struct sbitmap_queue *sbq;
1169 list_del_init(&wait->entry);
1170 sbq = &hctx->tags->bitmap_tags;
1171 atomic_dec(&sbq->ws_active);
1173 spin_unlock(&hctx->dispatch_wait_lock);
1175 blk_mq_run_hw_queue(hctx, true);
1180 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1181 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1182 * restart. For both cases, take care to check the condition again after
1183 * marking us as waiting.
1185 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1188 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1189 struct wait_queue_head *wq;
1190 wait_queue_entry_t *wait;
1193 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1194 blk_mq_sched_mark_restart_hctx(hctx);
1197 * It's possible that a tag was freed in the window between the
1198 * allocation failure and adding the hardware queue to the wait
1201 * Don't clear RESTART here, someone else could have set it.
1202 * At most this will cost an extra queue run.
1204 return blk_mq_get_driver_tag(rq);
1207 wait = &hctx->dispatch_wait;
1208 if (!list_empty_careful(&wait->entry))
1211 wq = &bt_wait_ptr(sbq, hctx)->wait;
1213 spin_lock_irq(&wq->lock);
1214 spin_lock(&hctx->dispatch_wait_lock);
1215 if (!list_empty(&wait->entry)) {
1216 spin_unlock(&hctx->dispatch_wait_lock);
1217 spin_unlock_irq(&wq->lock);
1221 atomic_inc(&sbq->ws_active);
1222 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1223 __add_wait_queue(wq, wait);
1226 * It's possible that a tag was freed in the window between the
1227 * allocation failure and adding the hardware queue to the wait
1230 ret = blk_mq_get_driver_tag(rq);
1232 spin_unlock(&hctx->dispatch_wait_lock);
1233 spin_unlock_irq(&wq->lock);
1238 * We got a tag, remove ourselves from the wait queue to ensure
1239 * someone else gets the wakeup.
1241 list_del_init(&wait->entry);
1242 atomic_dec(&sbq->ws_active);
1243 spin_unlock(&hctx->dispatch_wait_lock);
1244 spin_unlock_irq(&wq->lock);
1249 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1250 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1252 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1253 * - EWMA is one simple way to compute running average value
1254 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1255 * - take 4 as factor for avoiding to get too small(0) result, and this
1256 * factor doesn't matter because EWMA decreases exponentially
1258 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1262 if (hctx->queue->elevator)
1265 ewma = hctx->dispatch_busy;
1270 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1272 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1273 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1275 hctx->dispatch_busy = ewma;
1278 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1280 static void blk_mq_handle_dev_resource(struct request *rq,
1281 struct list_head *list)
1283 struct request *next =
1284 list_first_entry_or_null(list, struct request, queuelist);
1287 * If an I/O scheduler has been configured and we got a driver tag for
1288 * the next request already, free it.
1291 blk_mq_put_driver_tag(next);
1293 list_add(&rq->queuelist, list);
1294 __blk_mq_requeue_request(rq);
1297 static void blk_mq_handle_zone_resource(struct request *rq,
1298 struct list_head *zone_list)
1301 * If we end up here it is because we cannot dispatch a request to a
1302 * specific zone due to LLD level zone-write locking or other zone
1303 * related resource not being available. In this case, set the request
1304 * aside in zone_list for retrying it later.
1306 list_add(&rq->queuelist, zone_list);
1307 __blk_mq_requeue_request(rq);
1310 enum prep_dispatch {
1312 PREP_DISPATCH_NO_TAG,
1313 PREP_DISPATCH_NO_BUDGET,
1316 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1319 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1321 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1322 blk_mq_put_driver_tag(rq);
1323 return PREP_DISPATCH_NO_BUDGET;
1326 if (!blk_mq_get_driver_tag(rq)) {
1328 * The initial allocation attempt failed, so we need to
1329 * rerun the hardware queue when a tag is freed. The
1330 * waitqueue takes care of that. If the queue is run
1331 * before we add this entry back on the dispatch list,
1332 * we'll re-run it below.
1334 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1336 * All budgets not got from this function will be put
1337 * together during handling partial dispatch
1340 blk_mq_put_dispatch_budget(rq->q);
1341 return PREP_DISPATCH_NO_TAG;
1345 return PREP_DISPATCH_OK;
1348 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1349 static void blk_mq_release_budgets(struct request_queue *q,
1350 unsigned int nr_budgets)
1354 for (i = 0; i < nr_budgets; i++)
1355 blk_mq_put_dispatch_budget(q);
1359 * Returns true if we did some work AND can potentially do more.
1361 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1362 unsigned int nr_budgets)
1364 enum prep_dispatch prep;
1365 struct request_queue *q = hctx->queue;
1366 struct request *rq, *nxt;
1368 blk_status_t ret = BLK_STS_OK;
1369 LIST_HEAD(zone_list);
1371 if (list_empty(list))
1375 * Now process all the entries, sending them to the driver.
1377 errors = queued = 0;
1379 struct blk_mq_queue_data bd;
1381 rq = list_first_entry(list, struct request, queuelist);
1383 WARN_ON_ONCE(hctx != rq->mq_hctx);
1384 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1385 if (prep != PREP_DISPATCH_OK)
1388 list_del_init(&rq->queuelist);
1393 * Flag last if we have no more requests, or if we have more
1394 * but can't assign a driver tag to it.
1396 if (list_empty(list))
1399 nxt = list_first_entry(list, struct request, queuelist);
1400 bd.last = !blk_mq_get_driver_tag(nxt);
1404 * once the request is queued to lld, no need to cover the
1409 ret = q->mq_ops->queue_rq(hctx, &bd);
1410 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1411 blk_mq_handle_dev_resource(rq, list);
1413 } else if (ret == BLK_STS_ZONE_RESOURCE) {
1415 * Move the request to zone_list and keep going through
1416 * the dispatch list to find more requests the drive can
1419 blk_mq_handle_zone_resource(rq, &zone_list);
1420 if (list_empty(list))
1425 if (unlikely(ret != BLK_STS_OK)) {
1427 blk_mq_end_request(rq, BLK_STS_IOERR);
1432 } while (!list_empty(list));
1434 if (!list_empty(&zone_list))
1435 list_splice_tail_init(&zone_list, list);
1437 hctx->dispatched[queued_to_index(queued)]++;
1440 * Any items that need requeuing? Stuff them into hctx->dispatch,
1441 * that is where we will continue on next queue run.
1443 if (!list_empty(list)) {
1445 /* For non-shared tags, the RESTART check will suffice */
1446 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1447 (hctx->flags & BLK_MQ_F_TAG_SHARED);
1448 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1450 blk_mq_release_budgets(q, nr_budgets);
1453 * If we didn't flush the entire list, we could have told
1454 * the driver there was more coming, but that turned out to
1457 if (q->mq_ops->commit_rqs && queued)
1458 q->mq_ops->commit_rqs(hctx);
1460 spin_lock(&hctx->lock);
1461 list_splice_tail_init(list, &hctx->dispatch);
1462 spin_unlock(&hctx->lock);
1465 * If SCHED_RESTART was set by the caller of this function and
1466 * it is no longer set that means that it was cleared by another
1467 * thread and hence that a queue rerun is needed.
1469 * If 'no_tag' is set, that means that we failed getting
1470 * a driver tag with an I/O scheduler attached. If our dispatch
1471 * waitqueue is no longer active, ensure that we run the queue
1472 * AFTER adding our entries back to the list.
1474 * If no I/O scheduler has been configured it is possible that
1475 * the hardware queue got stopped and restarted before requests
1476 * were pushed back onto the dispatch list. Rerun the queue to
1477 * avoid starvation. Notes:
1478 * - blk_mq_run_hw_queue() checks whether or not a queue has
1479 * been stopped before rerunning a queue.
1480 * - Some but not all block drivers stop a queue before
1481 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1484 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1485 * bit is set, run queue after a delay to avoid IO stalls
1486 * that could otherwise occur if the queue is idle. We'll do
1487 * similar if we couldn't get budget and SCHED_RESTART is set.
1489 needs_restart = blk_mq_sched_needs_restart(hctx);
1490 if (!needs_restart ||
1491 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1492 blk_mq_run_hw_queue(hctx, true);
1493 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1495 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1497 blk_mq_update_dispatch_busy(hctx, true);
1500 blk_mq_update_dispatch_busy(hctx, false);
1502 return (queued + errors) != 0;
1506 * __blk_mq_run_hw_queue - Run a hardware queue.
1507 * @hctx: Pointer to the hardware queue to run.
1509 * Send pending requests to the hardware.
1511 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1516 * We should be running this queue from one of the CPUs that
1519 * There are at least two related races now between setting
1520 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1521 * __blk_mq_run_hw_queue():
1523 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1524 * but later it becomes online, then this warning is harmless
1527 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1528 * but later it becomes offline, then the warning can't be
1529 * triggered, and we depend on blk-mq timeout handler to
1530 * handle dispatched requests to this hctx
1532 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1533 cpu_online(hctx->next_cpu)) {
1534 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1535 raw_smp_processor_id(),
1536 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1541 * We can't run the queue inline with ints disabled. Ensure that
1542 * we catch bad users of this early.
1544 WARN_ON_ONCE(in_interrupt());
1546 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1548 hctx_lock(hctx, &srcu_idx);
1549 blk_mq_sched_dispatch_requests(hctx);
1550 hctx_unlock(hctx, srcu_idx);
1553 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1555 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1557 if (cpu >= nr_cpu_ids)
1558 cpu = cpumask_first(hctx->cpumask);
1563 * It'd be great if the workqueue API had a way to pass
1564 * in a mask and had some smarts for more clever placement.
1565 * For now we just round-robin here, switching for every
1566 * BLK_MQ_CPU_WORK_BATCH queued items.
1568 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1571 int next_cpu = hctx->next_cpu;
1573 if (hctx->queue->nr_hw_queues == 1)
1574 return WORK_CPU_UNBOUND;
1576 if (--hctx->next_cpu_batch <= 0) {
1578 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1580 if (next_cpu >= nr_cpu_ids)
1581 next_cpu = blk_mq_first_mapped_cpu(hctx);
1582 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1586 * Do unbound schedule if we can't find a online CPU for this hctx,
1587 * and it should only happen in the path of handling CPU DEAD.
1589 if (!cpu_online(next_cpu)) {
1596 * Make sure to re-select CPU next time once after CPUs
1597 * in hctx->cpumask become online again.
1599 hctx->next_cpu = next_cpu;
1600 hctx->next_cpu_batch = 1;
1601 return WORK_CPU_UNBOUND;
1604 hctx->next_cpu = next_cpu;
1609 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1610 * @hctx: Pointer to the hardware queue to run.
1611 * @async: If we want to run the queue asynchronously.
1612 * @msecs: Microseconds of delay to wait before running the queue.
1614 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1615 * with a delay of @msecs.
1617 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1618 unsigned long msecs)
1620 if (unlikely(blk_mq_hctx_stopped(hctx)))
1623 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1624 int cpu = get_cpu();
1625 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1626 __blk_mq_run_hw_queue(hctx);
1634 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1635 msecs_to_jiffies(msecs));
1639 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1640 * @hctx: Pointer to the hardware queue to run.
1641 * @msecs: Microseconds of delay to wait before running the queue.
1643 * Run a hardware queue asynchronously with a delay of @msecs.
1645 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1647 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1649 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1652 * blk_mq_run_hw_queue - Start to run a hardware queue.
1653 * @hctx: Pointer to the hardware queue to run.
1654 * @async: If we want to run the queue asynchronously.
1656 * Check if the request queue is not in a quiesced state and if there are
1657 * pending requests to be sent. If this is true, run the queue to send requests
1660 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1666 * When queue is quiesced, we may be switching io scheduler, or
1667 * updating nr_hw_queues, or other things, and we can't run queue
1668 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1670 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1673 hctx_lock(hctx, &srcu_idx);
1674 need_run = !blk_queue_quiesced(hctx->queue) &&
1675 blk_mq_hctx_has_pending(hctx);
1676 hctx_unlock(hctx, srcu_idx);
1679 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1681 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1684 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1685 * @q: Pointer to the request queue to run.
1686 * @async: If we want to run the queue asynchronously.
1688 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1690 struct blk_mq_hw_ctx *hctx;
1693 queue_for_each_hw_ctx(q, hctx, i) {
1694 if (blk_mq_hctx_stopped(hctx))
1697 blk_mq_run_hw_queue(hctx, async);
1700 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1703 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1704 * @q: Pointer to the request queue to run.
1705 * @msecs: Microseconds of delay to wait before running the queues.
1707 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1709 struct blk_mq_hw_ctx *hctx;
1712 queue_for_each_hw_ctx(q, hctx, i) {
1713 if (blk_mq_hctx_stopped(hctx))
1716 blk_mq_delay_run_hw_queue(hctx, msecs);
1719 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1722 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1723 * @q: request queue.
1725 * The caller is responsible for serializing this function against
1726 * blk_mq_{start,stop}_hw_queue().
1728 bool blk_mq_queue_stopped(struct request_queue *q)
1730 struct blk_mq_hw_ctx *hctx;
1733 queue_for_each_hw_ctx(q, hctx, i)
1734 if (blk_mq_hctx_stopped(hctx))
1739 EXPORT_SYMBOL(blk_mq_queue_stopped);
1742 * This function is often used for pausing .queue_rq() by driver when
1743 * there isn't enough resource or some conditions aren't satisfied, and
1744 * BLK_STS_RESOURCE is usually returned.
1746 * We do not guarantee that dispatch can be drained or blocked
1747 * after blk_mq_stop_hw_queue() returns. Please use
1748 * blk_mq_quiesce_queue() for that requirement.
1750 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1752 cancel_delayed_work(&hctx->run_work);
1754 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1756 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1759 * This function is often used for pausing .queue_rq() by driver when
1760 * there isn't enough resource or some conditions aren't satisfied, and
1761 * BLK_STS_RESOURCE is usually returned.
1763 * We do not guarantee that dispatch can be drained or blocked
1764 * after blk_mq_stop_hw_queues() returns. Please use
1765 * blk_mq_quiesce_queue() for that requirement.
1767 void blk_mq_stop_hw_queues(struct request_queue *q)
1769 struct blk_mq_hw_ctx *hctx;
1772 queue_for_each_hw_ctx(q, hctx, i)
1773 blk_mq_stop_hw_queue(hctx);
1775 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1777 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1779 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1781 blk_mq_run_hw_queue(hctx, false);
1783 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1785 void blk_mq_start_hw_queues(struct request_queue *q)
1787 struct blk_mq_hw_ctx *hctx;
1790 queue_for_each_hw_ctx(q, hctx, i)
1791 blk_mq_start_hw_queue(hctx);
1793 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1795 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1797 if (!blk_mq_hctx_stopped(hctx))
1800 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1801 blk_mq_run_hw_queue(hctx, async);
1803 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1805 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1807 struct blk_mq_hw_ctx *hctx;
1810 queue_for_each_hw_ctx(q, hctx, i)
1811 blk_mq_start_stopped_hw_queue(hctx, async);
1813 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1815 static void blk_mq_run_work_fn(struct work_struct *work)
1817 struct blk_mq_hw_ctx *hctx;
1819 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1822 * If we are stopped, don't run the queue.
1824 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1827 __blk_mq_run_hw_queue(hctx);
1830 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1834 struct blk_mq_ctx *ctx = rq->mq_ctx;
1835 enum hctx_type type = hctx->type;
1837 lockdep_assert_held(&ctx->lock);
1839 trace_block_rq_insert(hctx->queue, rq);
1842 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1844 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1847 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1850 struct blk_mq_ctx *ctx = rq->mq_ctx;
1852 lockdep_assert_held(&ctx->lock);
1854 __blk_mq_insert_req_list(hctx, rq, at_head);
1855 blk_mq_hctx_mark_pending(hctx, ctx);
1859 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1860 * @rq: Pointer to request to be inserted.
1861 * @run_queue: If we should run the hardware queue after inserting the request.
1863 * Should only be used carefully, when the caller knows we want to
1864 * bypass a potential IO scheduler on the target device.
1866 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1869 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1871 spin_lock(&hctx->lock);
1873 list_add(&rq->queuelist, &hctx->dispatch);
1875 list_add_tail(&rq->queuelist, &hctx->dispatch);
1876 spin_unlock(&hctx->lock);
1879 blk_mq_run_hw_queue(hctx, false);
1882 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1883 struct list_head *list)
1887 enum hctx_type type = hctx->type;
1890 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1893 list_for_each_entry(rq, list, queuelist) {
1894 BUG_ON(rq->mq_ctx != ctx);
1895 trace_block_rq_insert(hctx->queue, rq);
1898 spin_lock(&ctx->lock);
1899 list_splice_tail_init(list, &ctx->rq_lists[type]);
1900 blk_mq_hctx_mark_pending(hctx, ctx);
1901 spin_unlock(&ctx->lock);
1904 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1906 struct request *rqa = container_of(a, struct request, queuelist);
1907 struct request *rqb = container_of(b, struct request, queuelist);
1909 if (rqa->mq_ctx != rqb->mq_ctx)
1910 return rqa->mq_ctx > rqb->mq_ctx;
1911 if (rqa->mq_hctx != rqb->mq_hctx)
1912 return rqa->mq_hctx > rqb->mq_hctx;
1914 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1917 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1921 if (list_empty(&plug->mq_list))
1923 list_splice_init(&plug->mq_list, &list);
1925 if (plug->rq_count > 2 && plug->multiple_queues)
1926 list_sort(NULL, &list, plug_rq_cmp);
1931 struct list_head rq_list;
1932 struct request *rq, *head_rq = list_entry_rq(list.next);
1933 struct list_head *pos = &head_rq->queuelist; /* skip first */
1934 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1935 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1936 unsigned int depth = 1;
1938 list_for_each_continue(pos, &list) {
1939 rq = list_entry_rq(pos);
1941 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1946 list_cut_before(&rq_list, &list, pos);
1947 trace_block_unplug(head_rq->q, depth, !from_schedule);
1948 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1950 } while(!list_empty(&list));
1953 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1954 unsigned int nr_segs)
1956 if (bio->bi_opf & REQ_RAHEAD)
1957 rq->cmd_flags |= REQ_FAILFAST_MASK;
1959 rq->__sector = bio->bi_iter.bi_sector;
1960 rq->write_hint = bio->bi_write_hint;
1961 blk_rq_bio_prep(rq, bio, nr_segs);
1962 blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1964 blk_account_io_start(rq);
1967 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1969 blk_qc_t *cookie, bool last)
1971 struct request_queue *q = rq->q;
1972 struct blk_mq_queue_data bd = {
1976 blk_qc_t new_cookie;
1979 new_cookie = request_to_qc_t(hctx, rq);
1982 * For OK queue, we are done. For error, caller may kill it.
1983 * Any other error (busy), just add it to our list as we
1984 * previously would have done.
1986 ret = q->mq_ops->queue_rq(hctx, &bd);
1989 blk_mq_update_dispatch_busy(hctx, false);
1990 *cookie = new_cookie;
1992 case BLK_STS_RESOURCE:
1993 case BLK_STS_DEV_RESOURCE:
1994 blk_mq_update_dispatch_busy(hctx, true);
1995 __blk_mq_requeue_request(rq);
1998 blk_mq_update_dispatch_busy(hctx, false);
1999 *cookie = BLK_QC_T_NONE;
2006 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2009 bool bypass_insert, bool last)
2011 struct request_queue *q = rq->q;
2012 bool run_queue = true;
2015 * RCU or SRCU read lock is needed before checking quiesced flag.
2017 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2018 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2019 * and avoid driver to try to dispatch again.
2021 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2023 bypass_insert = false;
2027 if (q->elevator && !bypass_insert)
2030 if (!blk_mq_get_dispatch_budget(q))
2033 if (!blk_mq_get_driver_tag(rq)) {
2034 blk_mq_put_dispatch_budget(q);
2038 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2041 return BLK_STS_RESOURCE;
2043 blk_mq_request_bypass_insert(rq, false, run_queue);
2048 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2049 * @hctx: Pointer of the associated hardware queue.
2050 * @rq: Pointer to request to be sent.
2051 * @cookie: Request queue cookie.
2053 * If the device has enough resources to accept a new request now, send the
2054 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2055 * we can try send it another time in the future. Requests inserted at this
2056 * queue have higher priority.
2058 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2059 struct request *rq, blk_qc_t *cookie)
2064 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2066 hctx_lock(hctx, &srcu_idx);
2068 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2069 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2070 blk_mq_request_bypass_insert(rq, false, true);
2071 else if (ret != BLK_STS_OK)
2072 blk_mq_end_request(rq, ret);
2074 hctx_unlock(hctx, srcu_idx);
2077 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2081 blk_qc_t unused_cookie;
2082 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2084 hctx_lock(hctx, &srcu_idx);
2085 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2086 hctx_unlock(hctx, srcu_idx);
2091 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2092 struct list_head *list)
2096 while (!list_empty(list)) {
2098 struct request *rq = list_first_entry(list, struct request,
2101 list_del_init(&rq->queuelist);
2102 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2103 if (ret != BLK_STS_OK) {
2104 if (ret == BLK_STS_RESOURCE ||
2105 ret == BLK_STS_DEV_RESOURCE) {
2106 blk_mq_request_bypass_insert(rq, false,
2110 blk_mq_end_request(rq, ret);
2116 * If we didn't flush the entire list, we could have told
2117 * the driver there was more coming, but that turned out to
2120 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs && queued)
2121 hctx->queue->mq_ops->commit_rqs(hctx);
2124 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2126 list_add_tail(&rq->queuelist, &plug->mq_list);
2128 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2129 struct request *tmp;
2131 tmp = list_first_entry(&plug->mq_list, struct request,
2133 if (tmp->q != rq->q)
2134 plug->multiple_queues = true;
2139 * blk_mq_make_request - Create and send a request to block device.
2140 * @q: Request queue pointer.
2141 * @bio: Bio pointer.
2143 * Builds up a request structure from @q and @bio and send to the device. The
2144 * request may not be queued directly to hardware if:
2145 * * This request can be merged with another one
2146 * * We want to place request at plug queue for possible future merging
2147 * * There is an IO scheduler active at this queue
2149 * It will not queue the request if there is an error with the bio, or at the
2152 * Returns: Request queue cookie.
2154 blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
2156 const int is_sync = op_is_sync(bio->bi_opf);
2157 const int is_flush_fua = op_is_flush(bio->bi_opf);
2158 struct blk_mq_alloc_data data = {
2162 struct blk_plug *plug;
2163 struct request *same_queue_rq = NULL;
2164 unsigned int nr_segs;
2168 blk_queue_bounce(q, &bio);
2169 __blk_queue_split(&bio, &nr_segs);
2171 if (!bio_integrity_prep(bio))
2174 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2175 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2178 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2181 rq_qos_throttle(q, bio);
2183 data.cmd_flags = bio->bi_opf;
2184 rq = __blk_mq_alloc_request(&data);
2185 if (unlikely(!rq)) {
2186 rq_qos_cleanup(q, bio);
2187 if (bio->bi_opf & REQ_NOWAIT)
2188 bio_wouldblock_error(bio);
2192 trace_block_getrq(q, bio, bio->bi_opf);
2194 rq_qos_track(q, rq, bio);
2196 cookie = request_to_qc_t(data.hctx, rq);
2198 blk_mq_bio_to_request(rq, bio, nr_segs);
2200 ret = blk_crypto_init_request(rq);
2201 if (ret != BLK_STS_OK) {
2202 bio->bi_status = ret;
2204 blk_mq_free_request(rq);
2205 return BLK_QC_T_NONE;
2208 plug = blk_mq_plug(q, bio);
2209 if (unlikely(is_flush_fua)) {
2210 /* Bypass scheduler for flush requests */
2211 blk_insert_flush(rq);
2212 blk_mq_run_hw_queue(data.hctx, true);
2213 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2214 !blk_queue_nonrot(q))) {
2216 * Use plugging if we have a ->commit_rqs() hook as well, as
2217 * we know the driver uses bd->last in a smart fashion.
2219 * Use normal plugging if this disk is slow HDD, as sequential
2220 * IO may benefit a lot from plug merging.
2222 unsigned int request_count = plug->rq_count;
2223 struct request *last = NULL;
2226 trace_block_plug(q);
2228 last = list_entry_rq(plug->mq_list.prev);
2230 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2231 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2232 blk_flush_plug_list(plug, false);
2233 trace_block_plug(q);
2236 blk_add_rq_to_plug(plug, rq);
2237 } else if (q->elevator) {
2238 /* Insert the request at the IO scheduler queue */
2239 blk_mq_sched_insert_request(rq, false, true, true);
2240 } else if (plug && !blk_queue_nomerges(q)) {
2242 * We do limited plugging. If the bio can be merged, do that.
2243 * Otherwise the existing request in the plug list will be
2244 * issued. So the plug list will have one request at most
2245 * The plug list might get flushed before this. If that happens,
2246 * the plug list is empty, and same_queue_rq is invalid.
2248 if (list_empty(&plug->mq_list))
2249 same_queue_rq = NULL;
2250 if (same_queue_rq) {
2251 list_del_init(&same_queue_rq->queuelist);
2254 blk_add_rq_to_plug(plug, rq);
2255 trace_block_plug(q);
2257 if (same_queue_rq) {
2258 data.hctx = same_queue_rq->mq_hctx;
2259 trace_block_unplug(q, 1, true);
2260 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2263 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2264 !data.hctx->dispatch_busy) {
2266 * There is no scheduler and we can try to send directly
2269 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2272 blk_mq_sched_insert_request(rq, false, true, true);
2278 return BLK_QC_T_NONE;
2280 EXPORT_SYMBOL_GPL(blk_mq_make_request); /* only for request based dm */
2282 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2283 unsigned int hctx_idx)
2287 if (tags->rqs && set->ops->exit_request) {
2290 for (i = 0; i < tags->nr_tags; i++) {
2291 struct request *rq = tags->static_rqs[i];
2295 set->ops->exit_request(set, rq, hctx_idx);
2296 tags->static_rqs[i] = NULL;
2300 while (!list_empty(&tags->page_list)) {
2301 page = list_first_entry(&tags->page_list, struct page, lru);
2302 list_del_init(&page->lru);
2304 * Remove kmemleak object previously allocated in
2305 * blk_mq_alloc_rqs().
2307 kmemleak_free(page_address(page));
2308 __free_pages(page, page->private);
2312 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2316 kfree(tags->static_rqs);
2317 tags->static_rqs = NULL;
2319 blk_mq_free_tags(tags);
2322 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2323 unsigned int hctx_idx,
2324 unsigned int nr_tags,
2325 unsigned int reserved_tags)
2327 struct blk_mq_tags *tags;
2330 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2331 if (node == NUMA_NO_NODE)
2332 node = set->numa_node;
2334 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2335 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2339 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2340 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2343 blk_mq_free_tags(tags);
2347 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2348 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2350 if (!tags->static_rqs) {
2352 blk_mq_free_tags(tags);
2359 static size_t order_to_size(unsigned int order)
2361 return (size_t)PAGE_SIZE << order;
2364 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2365 unsigned int hctx_idx, int node)
2369 if (set->ops->init_request) {
2370 ret = set->ops->init_request(set, rq, hctx_idx, node);
2375 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2379 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2380 unsigned int hctx_idx, unsigned int depth)
2382 unsigned int i, j, entries_per_page, max_order = 4;
2383 size_t rq_size, left;
2386 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2387 if (node == NUMA_NO_NODE)
2388 node = set->numa_node;
2390 INIT_LIST_HEAD(&tags->page_list);
2393 * rq_size is the size of the request plus driver payload, rounded
2394 * to the cacheline size
2396 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2398 left = rq_size * depth;
2400 for (i = 0; i < depth; ) {
2401 int this_order = max_order;
2406 while (this_order && left < order_to_size(this_order - 1))
2410 page = alloc_pages_node(node,
2411 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2417 if (order_to_size(this_order) < rq_size)
2424 page->private = this_order;
2425 list_add_tail(&page->lru, &tags->page_list);
2427 p = page_address(page);
2429 * Allow kmemleak to scan these pages as they contain pointers
2430 * to additional allocations like via ops->init_request().
2432 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2433 entries_per_page = order_to_size(this_order) / rq_size;
2434 to_do = min(entries_per_page, depth - i);
2435 left -= to_do * rq_size;
2436 for (j = 0; j < to_do; j++) {
2437 struct request *rq = p;
2439 tags->static_rqs[i] = rq;
2440 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2441 tags->static_rqs[i] = NULL;
2452 blk_mq_free_rqs(set, tags, hctx_idx);
2456 struct rq_iter_data {
2457 struct blk_mq_hw_ctx *hctx;
2461 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2463 struct rq_iter_data *iter_data = data;
2465 if (rq->mq_hctx != iter_data->hctx)
2467 iter_data->has_rq = true;
2471 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2473 struct blk_mq_tags *tags = hctx->sched_tags ?
2474 hctx->sched_tags : hctx->tags;
2475 struct rq_iter_data data = {
2479 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2483 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2484 struct blk_mq_hw_ctx *hctx)
2486 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2488 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2493 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2495 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2496 struct blk_mq_hw_ctx, cpuhp_online);
2498 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2499 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2503 * Prevent new request from being allocated on the current hctx.
2505 * The smp_mb__after_atomic() Pairs with the implied barrier in
2506 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2507 * seen once we return from the tag allocator.
2509 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2510 smp_mb__after_atomic();
2513 * Try to grab a reference to the queue and wait for any outstanding
2514 * requests. If we could not grab a reference the queue has been
2515 * frozen and there are no requests.
2517 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2518 while (blk_mq_hctx_has_requests(hctx))
2520 percpu_ref_put(&hctx->queue->q_usage_counter);
2526 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2528 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2529 struct blk_mq_hw_ctx, cpuhp_online);
2531 if (cpumask_test_cpu(cpu, hctx->cpumask))
2532 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2537 * 'cpu' is going away. splice any existing rq_list entries from this
2538 * software queue to the hw queue dispatch list, and ensure that it
2541 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2543 struct blk_mq_hw_ctx *hctx;
2544 struct blk_mq_ctx *ctx;
2546 enum hctx_type type;
2548 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2549 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2552 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2555 spin_lock(&ctx->lock);
2556 if (!list_empty(&ctx->rq_lists[type])) {
2557 list_splice_init(&ctx->rq_lists[type], &tmp);
2558 blk_mq_hctx_clear_pending(hctx, ctx);
2560 spin_unlock(&ctx->lock);
2562 if (list_empty(&tmp))
2565 spin_lock(&hctx->lock);
2566 list_splice_tail_init(&tmp, &hctx->dispatch);
2567 spin_unlock(&hctx->lock);
2569 blk_mq_run_hw_queue(hctx, true);
2573 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2575 if (!(hctx->flags & BLK_MQ_F_STACKING))
2576 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2577 &hctx->cpuhp_online);
2578 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2582 /* hctx->ctxs will be freed in queue's release handler */
2583 static void blk_mq_exit_hctx(struct request_queue *q,
2584 struct blk_mq_tag_set *set,
2585 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2587 if (blk_mq_hw_queue_mapped(hctx))
2588 blk_mq_tag_idle(hctx);
2590 if (set->ops->exit_request)
2591 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2593 if (set->ops->exit_hctx)
2594 set->ops->exit_hctx(hctx, hctx_idx);
2596 blk_mq_remove_cpuhp(hctx);
2598 spin_lock(&q->unused_hctx_lock);
2599 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2600 spin_unlock(&q->unused_hctx_lock);
2603 static void blk_mq_exit_hw_queues(struct request_queue *q,
2604 struct blk_mq_tag_set *set, int nr_queue)
2606 struct blk_mq_hw_ctx *hctx;
2609 queue_for_each_hw_ctx(q, hctx, i) {
2612 blk_mq_debugfs_unregister_hctx(hctx);
2613 blk_mq_exit_hctx(q, set, hctx, i);
2617 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2619 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2621 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2622 __alignof__(struct blk_mq_hw_ctx)) !=
2623 sizeof(struct blk_mq_hw_ctx));
2625 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2626 hw_ctx_size += sizeof(struct srcu_struct);
2631 static int blk_mq_init_hctx(struct request_queue *q,
2632 struct blk_mq_tag_set *set,
2633 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2635 hctx->queue_num = hctx_idx;
2637 if (!(hctx->flags & BLK_MQ_F_STACKING))
2638 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2639 &hctx->cpuhp_online);
2640 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2642 hctx->tags = set->tags[hctx_idx];
2644 if (set->ops->init_hctx &&
2645 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2646 goto unregister_cpu_notifier;
2648 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2654 if (set->ops->exit_hctx)
2655 set->ops->exit_hctx(hctx, hctx_idx);
2656 unregister_cpu_notifier:
2657 blk_mq_remove_cpuhp(hctx);
2661 static struct blk_mq_hw_ctx *
2662 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2665 struct blk_mq_hw_ctx *hctx;
2666 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2668 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2670 goto fail_alloc_hctx;
2672 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2675 atomic_set(&hctx->nr_active, 0);
2676 if (node == NUMA_NO_NODE)
2677 node = set->numa_node;
2678 hctx->numa_node = node;
2680 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2681 spin_lock_init(&hctx->lock);
2682 INIT_LIST_HEAD(&hctx->dispatch);
2684 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2686 INIT_LIST_HEAD(&hctx->hctx_list);
2689 * Allocate space for all possible cpus to avoid allocation at
2692 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2697 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2702 spin_lock_init(&hctx->dispatch_wait_lock);
2703 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2704 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2706 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2710 if (hctx->flags & BLK_MQ_F_BLOCKING)
2711 init_srcu_struct(hctx->srcu);
2712 blk_mq_hctx_kobj_init(hctx);
2717 sbitmap_free(&hctx->ctx_map);
2721 free_cpumask_var(hctx->cpumask);
2728 static void blk_mq_init_cpu_queues(struct request_queue *q,
2729 unsigned int nr_hw_queues)
2731 struct blk_mq_tag_set *set = q->tag_set;
2734 for_each_possible_cpu(i) {
2735 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2736 struct blk_mq_hw_ctx *hctx;
2740 spin_lock_init(&__ctx->lock);
2741 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2742 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2747 * Set local node, IFF we have more than one hw queue. If
2748 * not, we remain on the home node of the device
2750 for (j = 0; j < set->nr_maps; j++) {
2751 hctx = blk_mq_map_queue_type(q, j, i);
2752 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2753 hctx->numa_node = local_memory_node(cpu_to_node(i));
2758 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2763 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2764 set->queue_depth, set->reserved_tags);
2765 if (!set->tags[hctx_idx])
2768 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2773 blk_mq_free_rq_map(set->tags[hctx_idx]);
2774 set->tags[hctx_idx] = NULL;
2778 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2779 unsigned int hctx_idx)
2781 if (set->tags && set->tags[hctx_idx]) {
2782 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2783 blk_mq_free_rq_map(set->tags[hctx_idx]);
2784 set->tags[hctx_idx] = NULL;
2788 static void blk_mq_map_swqueue(struct request_queue *q)
2790 unsigned int i, j, hctx_idx;
2791 struct blk_mq_hw_ctx *hctx;
2792 struct blk_mq_ctx *ctx;
2793 struct blk_mq_tag_set *set = q->tag_set;
2795 queue_for_each_hw_ctx(q, hctx, i) {
2796 cpumask_clear(hctx->cpumask);
2798 hctx->dispatch_from = NULL;
2802 * Map software to hardware queues.
2804 * If the cpu isn't present, the cpu is mapped to first hctx.
2806 for_each_possible_cpu(i) {
2808 ctx = per_cpu_ptr(q->queue_ctx, i);
2809 for (j = 0; j < set->nr_maps; j++) {
2810 if (!set->map[j].nr_queues) {
2811 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2812 HCTX_TYPE_DEFAULT, i);
2815 hctx_idx = set->map[j].mq_map[i];
2816 /* unmapped hw queue can be remapped after CPU topo changed */
2817 if (!set->tags[hctx_idx] &&
2818 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2820 * If tags initialization fail for some hctx,
2821 * that hctx won't be brought online. In this
2822 * case, remap the current ctx to hctx[0] which
2823 * is guaranteed to always have tags allocated
2825 set->map[j].mq_map[i] = 0;
2828 hctx = blk_mq_map_queue_type(q, j, i);
2829 ctx->hctxs[j] = hctx;
2831 * If the CPU is already set in the mask, then we've
2832 * mapped this one already. This can happen if
2833 * devices share queues across queue maps.
2835 if (cpumask_test_cpu(i, hctx->cpumask))
2838 cpumask_set_cpu(i, hctx->cpumask);
2840 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2841 hctx->ctxs[hctx->nr_ctx++] = ctx;
2844 * If the nr_ctx type overflows, we have exceeded the
2845 * amount of sw queues we can support.
2847 BUG_ON(!hctx->nr_ctx);
2850 for (; j < HCTX_MAX_TYPES; j++)
2851 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2852 HCTX_TYPE_DEFAULT, i);
2855 queue_for_each_hw_ctx(q, hctx, i) {
2857 * If no software queues are mapped to this hardware queue,
2858 * disable it and free the request entries.
2860 if (!hctx->nr_ctx) {
2861 /* Never unmap queue 0. We need it as a
2862 * fallback in case of a new remap fails
2865 if (i && set->tags[i])
2866 blk_mq_free_map_and_requests(set, i);
2872 hctx->tags = set->tags[i];
2873 WARN_ON(!hctx->tags);
2876 * Set the map size to the number of mapped software queues.
2877 * This is more accurate and more efficient than looping
2878 * over all possibly mapped software queues.
2880 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2883 * Initialize batch roundrobin counts
2885 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2886 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2891 * Caller needs to ensure that we're either frozen/quiesced, or that
2892 * the queue isn't live yet.
2894 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2896 struct blk_mq_hw_ctx *hctx;
2899 queue_for_each_hw_ctx(q, hctx, i) {
2901 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2903 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2907 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2910 struct request_queue *q;
2912 lockdep_assert_held(&set->tag_list_lock);
2914 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2915 blk_mq_freeze_queue(q);
2916 queue_set_hctx_shared(q, shared);
2917 blk_mq_unfreeze_queue(q);
2921 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2923 struct blk_mq_tag_set *set = q->tag_set;
2925 mutex_lock(&set->tag_list_lock);
2926 list_del_rcu(&q->tag_set_list);
2927 if (list_is_singular(&set->tag_list)) {
2928 /* just transitioned to unshared */
2929 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2930 /* update existing queue */
2931 blk_mq_update_tag_set_depth(set, false);
2933 mutex_unlock(&set->tag_list_lock);
2934 INIT_LIST_HEAD(&q->tag_set_list);
2937 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2938 struct request_queue *q)
2940 mutex_lock(&set->tag_list_lock);
2943 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2945 if (!list_empty(&set->tag_list) &&
2946 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2947 set->flags |= BLK_MQ_F_TAG_SHARED;
2948 /* update existing queue */
2949 blk_mq_update_tag_set_depth(set, true);
2951 if (set->flags & BLK_MQ_F_TAG_SHARED)
2952 queue_set_hctx_shared(q, true);
2953 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2955 mutex_unlock(&set->tag_list_lock);
2958 /* All allocations will be freed in release handler of q->mq_kobj */
2959 static int blk_mq_alloc_ctxs(struct request_queue *q)
2961 struct blk_mq_ctxs *ctxs;
2964 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2968 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2969 if (!ctxs->queue_ctx)
2972 for_each_possible_cpu(cpu) {
2973 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2977 q->mq_kobj = &ctxs->kobj;
2978 q->queue_ctx = ctxs->queue_ctx;
2987 * It is the actual release handler for mq, but we do it from
2988 * request queue's release handler for avoiding use-after-free
2989 * and headache because q->mq_kobj shouldn't have been introduced,
2990 * but we can't group ctx/kctx kobj without it.
2992 void blk_mq_release(struct request_queue *q)
2994 struct blk_mq_hw_ctx *hctx, *next;
2997 queue_for_each_hw_ctx(q, hctx, i)
2998 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3000 /* all hctx are in .unused_hctx_list now */
3001 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3002 list_del_init(&hctx->hctx_list);
3003 kobject_put(&hctx->kobj);
3006 kfree(q->queue_hw_ctx);
3009 * release .mq_kobj and sw queue's kobject now because
3010 * both share lifetime with request queue.
3012 blk_mq_sysfs_deinit(q);
3015 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3018 struct request_queue *uninit_q, *q;
3020 uninit_q = __blk_alloc_queue(set->numa_node);
3022 return ERR_PTR(-ENOMEM);
3023 uninit_q->queuedata = queuedata;
3026 * Initialize the queue without an elevator. device_add_disk() will do
3027 * the initialization.
3029 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3031 blk_cleanup_queue(uninit_q);
3035 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3037 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3039 return blk_mq_init_queue_data(set, NULL);
3041 EXPORT_SYMBOL(blk_mq_init_queue);
3044 * Helper for setting up a queue with mq ops, given queue depth, and
3045 * the passed in mq ops flags.
3047 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3048 const struct blk_mq_ops *ops,
3049 unsigned int queue_depth,
3050 unsigned int set_flags)
3052 struct request_queue *q;
3055 memset(set, 0, sizeof(*set));
3057 set->nr_hw_queues = 1;
3059 set->queue_depth = queue_depth;
3060 set->numa_node = NUMA_NO_NODE;
3061 set->flags = set_flags;
3063 ret = blk_mq_alloc_tag_set(set);
3065 return ERR_PTR(ret);
3067 q = blk_mq_init_queue(set);
3069 blk_mq_free_tag_set(set);
3075 EXPORT_SYMBOL(blk_mq_init_sq_queue);
3077 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3078 struct blk_mq_tag_set *set, struct request_queue *q,
3079 int hctx_idx, int node)
3081 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3083 /* reuse dead hctx first */
3084 spin_lock(&q->unused_hctx_lock);
3085 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3086 if (tmp->numa_node == node) {
3092 list_del_init(&hctx->hctx_list);
3093 spin_unlock(&q->unused_hctx_lock);
3096 hctx = blk_mq_alloc_hctx(q, set, node);
3100 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3106 kobject_put(&hctx->kobj);
3111 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3112 struct request_queue *q)
3115 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3117 if (q->nr_hw_queues < set->nr_hw_queues) {
3118 struct blk_mq_hw_ctx **new_hctxs;
3120 new_hctxs = kcalloc_node(set->nr_hw_queues,
3121 sizeof(*new_hctxs), GFP_KERNEL,
3126 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3128 q->queue_hw_ctx = new_hctxs;
3133 /* protect against switching io scheduler */
3134 mutex_lock(&q->sysfs_lock);
3135 for (i = 0; i < set->nr_hw_queues; i++) {
3137 struct blk_mq_hw_ctx *hctx;
3139 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3141 * If the hw queue has been mapped to another numa node,
3142 * we need to realloc the hctx. If allocation fails, fallback
3143 * to use the previous one.
3145 if (hctxs[i] && (hctxs[i]->numa_node == node))
3148 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3151 blk_mq_exit_hctx(q, set, hctxs[i], i);
3155 pr_warn("Allocate new hctx on node %d fails,\
3156 fallback to previous one on node %d\n",
3157 node, hctxs[i]->numa_node);
3163 * Increasing nr_hw_queues fails. Free the newly allocated
3164 * hctxs and keep the previous q->nr_hw_queues.
3166 if (i != set->nr_hw_queues) {
3167 j = q->nr_hw_queues;
3171 end = q->nr_hw_queues;
3172 q->nr_hw_queues = set->nr_hw_queues;
3175 for (; j < end; j++) {
3176 struct blk_mq_hw_ctx *hctx = hctxs[j];
3180 blk_mq_free_map_and_requests(set, j);
3181 blk_mq_exit_hctx(q, set, hctx, j);
3185 mutex_unlock(&q->sysfs_lock);
3188 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3189 struct request_queue *q,
3192 /* mark the queue as mq asap */
3193 q->mq_ops = set->ops;
3195 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3196 blk_mq_poll_stats_bkt,
3197 BLK_MQ_POLL_STATS_BKTS, q);
3201 if (blk_mq_alloc_ctxs(q))
3204 /* init q->mq_kobj and sw queues' kobjects */
3205 blk_mq_sysfs_init(q);
3207 INIT_LIST_HEAD(&q->unused_hctx_list);
3208 spin_lock_init(&q->unused_hctx_lock);
3210 blk_mq_realloc_hw_ctxs(set, q);
3211 if (!q->nr_hw_queues)
3214 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3215 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3219 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3220 if (set->nr_maps > HCTX_TYPE_POLL &&
3221 set->map[HCTX_TYPE_POLL].nr_queues)
3222 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3224 q->sg_reserved_size = INT_MAX;
3226 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3227 INIT_LIST_HEAD(&q->requeue_list);
3228 spin_lock_init(&q->requeue_lock);
3230 q->nr_requests = set->queue_depth;
3233 * Default to classic polling
3235 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3237 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3238 blk_mq_add_queue_tag_set(set, q);
3239 blk_mq_map_swqueue(q);
3242 elevator_init_mq(q);
3247 kfree(q->queue_hw_ctx);
3248 q->nr_hw_queues = 0;
3249 blk_mq_sysfs_deinit(q);
3251 blk_stat_free_callback(q->poll_cb);
3255 return ERR_PTR(-ENOMEM);
3257 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3259 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3260 void blk_mq_exit_queue(struct request_queue *q)
3262 struct blk_mq_tag_set *set = q->tag_set;
3264 blk_mq_del_queue_tag_set(q);
3265 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3268 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3272 for (i = 0; i < set->nr_hw_queues; i++)
3273 if (!__blk_mq_alloc_map_and_request(set, i))
3280 blk_mq_free_map_and_requests(set, i);
3286 * Allocate the request maps associated with this tag_set. Note that this
3287 * may reduce the depth asked for, if memory is tight. set->queue_depth
3288 * will be updated to reflect the allocated depth.
3290 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3295 depth = set->queue_depth;
3297 err = __blk_mq_alloc_rq_maps(set);
3301 set->queue_depth >>= 1;
3302 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3306 } while (set->queue_depth);
3308 if (!set->queue_depth || err) {
3309 pr_err("blk-mq: failed to allocate request map\n");
3313 if (depth != set->queue_depth)
3314 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3315 depth, set->queue_depth);
3320 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3323 * blk_mq_map_queues() and multiple .map_queues() implementations
3324 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3325 * number of hardware queues.
3327 if (set->nr_maps == 1)
3328 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3330 if (set->ops->map_queues && !is_kdump_kernel()) {
3334 * transport .map_queues is usually done in the following
3337 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3338 * mask = get_cpu_mask(queue)
3339 * for_each_cpu(cpu, mask)
3340 * set->map[x].mq_map[cpu] = queue;
3343 * When we need to remap, the table has to be cleared for
3344 * killing stale mapping since one CPU may not be mapped
3347 for (i = 0; i < set->nr_maps; i++)
3348 blk_mq_clear_mq_map(&set->map[i]);
3350 return set->ops->map_queues(set);
3352 BUG_ON(set->nr_maps > 1);
3353 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3357 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3358 int cur_nr_hw_queues, int new_nr_hw_queues)
3360 struct blk_mq_tags **new_tags;
3362 if (cur_nr_hw_queues >= new_nr_hw_queues)
3365 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3366 GFP_KERNEL, set->numa_node);
3371 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3372 sizeof(*set->tags));
3374 set->tags = new_tags;
3375 set->nr_hw_queues = new_nr_hw_queues;
3381 * Alloc a tag set to be associated with one or more request queues.
3382 * May fail with EINVAL for various error conditions. May adjust the
3383 * requested depth down, if it's too large. In that case, the set
3384 * value will be stored in set->queue_depth.
3386 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3390 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3392 if (!set->nr_hw_queues)
3394 if (!set->queue_depth)
3396 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3399 if (!set->ops->queue_rq)
3402 if (!set->ops->get_budget ^ !set->ops->put_budget)
3405 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3406 pr_info("blk-mq: reduced tag depth to %u\n",
3408 set->queue_depth = BLK_MQ_MAX_DEPTH;
3413 else if (set->nr_maps > HCTX_MAX_TYPES)
3417 * If a crashdump is active, then we are potentially in a very
3418 * memory constrained environment. Limit us to 1 queue and
3419 * 64 tags to prevent using too much memory.
3421 if (is_kdump_kernel()) {
3422 set->nr_hw_queues = 1;
3424 set->queue_depth = min(64U, set->queue_depth);
3427 * There is no use for more h/w queues than cpus if we just have
3430 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3431 set->nr_hw_queues = nr_cpu_ids;
3433 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3437 for (i = 0; i < set->nr_maps; i++) {
3438 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3439 sizeof(set->map[i].mq_map[0]),
3440 GFP_KERNEL, set->numa_node);
3441 if (!set->map[i].mq_map)
3442 goto out_free_mq_map;
3443 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3446 ret = blk_mq_update_queue_map(set);
3448 goto out_free_mq_map;
3450 ret = blk_mq_alloc_map_and_requests(set);
3452 goto out_free_mq_map;
3454 mutex_init(&set->tag_list_lock);
3455 INIT_LIST_HEAD(&set->tag_list);
3460 for (i = 0; i < set->nr_maps; i++) {
3461 kfree(set->map[i].mq_map);
3462 set->map[i].mq_map = NULL;
3468 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3470 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3474 for (i = 0; i < set->nr_hw_queues; i++)
3475 blk_mq_free_map_and_requests(set, i);
3477 for (j = 0; j < set->nr_maps; j++) {
3478 kfree(set->map[j].mq_map);
3479 set->map[j].mq_map = NULL;
3485 EXPORT_SYMBOL(blk_mq_free_tag_set);
3487 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3489 struct blk_mq_tag_set *set = q->tag_set;
3490 struct blk_mq_hw_ctx *hctx;
3496 if (q->nr_requests == nr)
3499 blk_mq_freeze_queue(q);
3500 blk_mq_quiesce_queue(q);
3503 queue_for_each_hw_ctx(q, hctx, i) {
3507 * If we're using an MQ scheduler, just update the scheduler
3508 * queue depth. This is similar to what the old code would do.
3510 if (!hctx->sched_tags) {
3511 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3514 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3519 if (q->elevator && q->elevator->type->ops.depth_updated)
3520 q->elevator->type->ops.depth_updated(hctx);
3524 q->nr_requests = nr;
3526 blk_mq_unquiesce_queue(q);
3527 blk_mq_unfreeze_queue(q);
3533 * request_queue and elevator_type pair.
3534 * It is just used by __blk_mq_update_nr_hw_queues to cache
3535 * the elevator_type associated with a request_queue.
3537 struct blk_mq_qe_pair {
3538 struct list_head node;
3539 struct request_queue *q;
3540 struct elevator_type *type;
3544 * Cache the elevator_type in qe pair list and switch the
3545 * io scheduler to 'none'
3547 static bool blk_mq_elv_switch_none(struct list_head *head,
3548 struct request_queue *q)
3550 struct blk_mq_qe_pair *qe;
3555 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3559 INIT_LIST_HEAD(&qe->node);
3561 qe->type = q->elevator->type;
3562 list_add(&qe->node, head);
3564 mutex_lock(&q->sysfs_lock);
3566 * After elevator_switch_mq, the previous elevator_queue will be
3567 * released by elevator_release. The reference of the io scheduler
3568 * module get by elevator_get will also be put. So we need to get
3569 * a reference of the io scheduler module here to prevent it to be
3572 __module_get(qe->type->elevator_owner);
3573 elevator_switch_mq(q, NULL);
3574 mutex_unlock(&q->sysfs_lock);
3579 static void blk_mq_elv_switch_back(struct list_head *head,
3580 struct request_queue *q)
3582 struct blk_mq_qe_pair *qe;
3583 struct elevator_type *t = NULL;
3585 list_for_each_entry(qe, head, node)
3594 list_del(&qe->node);
3597 mutex_lock(&q->sysfs_lock);
3598 elevator_switch_mq(q, t);
3599 mutex_unlock(&q->sysfs_lock);
3602 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3605 struct request_queue *q;
3607 int prev_nr_hw_queues;
3609 lockdep_assert_held(&set->tag_list_lock);
3611 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3612 nr_hw_queues = nr_cpu_ids;
3613 if (nr_hw_queues < 1)
3615 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3618 list_for_each_entry(q, &set->tag_list, tag_set_list)
3619 blk_mq_freeze_queue(q);
3621 * Switch IO scheduler to 'none', cleaning up the data associated
3622 * with the previous scheduler. We will switch back once we are done
3623 * updating the new sw to hw queue mappings.
3625 list_for_each_entry(q, &set->tag_list, tag_set_list)
3626 if (!blk_mq_elv_switch_none(&head, q))
3629 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3630 blk_mq_debugfs_unregister_hctxs(q);
3631 blk_mq_sysfs_unregister(q);
3634 prev_nr_hw_queues = set->nr_hw_queues;
3635 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3639 set->nr_hw_queues = nr_hw_queues;
3641 blk_mq_update_queue_map(set);
3642 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3643 blk_mq_realloc_hw_ctxs(set, q);
3644 if (q->nr_hw_queues != set->nr_hw_queues) {
3645 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3646 nr_hw_queues, prev_nr_hw_queues);
3647 set->nr_hw_queues = prev_nr_hw_queues;
3648 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3651 blk_mq_map_swqueue(q);
3655 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3656 blk_mq_sysfs_register(q);
3657 blk_mq_debugfs_register_hctxs(q);
3661 list_for_each_entry(q, &set->tag_list, tag_set_list)
3662 blk_mq_elv_switch_back(&head, q);
3664 list_for_each_entry(q, &set->tag_list, tag_set_list)
3665 blk_mq_unfreeze_queue(q);
3668 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3670 mutex_lock(&set->tag_list_lock);
3671 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3672 mutex_unlock(&set->tag_list_lock);
3674 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3676 /* Enable polling stats and return whether they were already enabled. */
3677 static bool blk_poll_stats_enable(struct request_queue *q)
3679 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3680 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3682 blk_stat_add_callback(q, q->poll_cb);
3686 static void blk_mq_poll_stats_start(struct request_queue *q)
3689 * We don't arm the callback if polling stats are not enabled or the
3690 * callback is already active.
3692 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3693 blk_stat_is_active(q->poll_cb))
3696 blk_stat_activate_msecs(q->poll_cb, 100);
3699 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3701 struct request_queue *q = cb->data;
3704 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3705 if (cb->stat[bucket].nr_samples)
3706 q->poll_stat[bucket] = cb->stat[bucket];
3710 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3713 unsigned long ret = 0;
3717 * If stats collection isn't on, don't sleep but turn it on for
3720 if (!blk_poll_stats_enable(q))
3724 * As an optimistic guess, use half of the mean service time
3725 * for this type of request. We can (and should) make this smarter.
3726 * For instance, if the completion latencies are tight, we can
3727 * get closer than just half the mean. This is especially
3728 * important on devices where the completion latencies are longer
3729 * than ~10 usec. We do use the stats for the relevant IO size
3730 * if available which does lead to better estimates.
3732 bucket = blk_mq_poll_stats_bkt(rq);
3736 if (q->poll_stat[bucket].nr_samples)
3737 ret = (q->poll_stat[bucket].mean + 1) / 2;
3742 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3745 struct hrtimer_sleeper hs;
3746 enum hrtimer_mode mode;
3750 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3754 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3756 * 0: use half of prev avg
3757 * >0: use this specific value
3759 if (q->poll_nsec > 0)
3760 nsecs = q->poll_nsec;
3762 nsecs = blk_mq_poll_nsecs(q, rq);
3767 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3770 * This will be replaced with the stats tracking code, using
3771 * 'avg_completion_time / 2' as the pre-sleep target.
3775 mode = HRTIMER_MODE_REL;
3776 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3777 hrtimer_set_expires(&hs.timer, kt);
3780 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3782 set_current_state(TASK_UNINTERRUPTIBLE);
3783 hrtimer_sleeper_start_expires(&hs, mode);
3786 hrtimer_cancel(&hs.timer);
3787 mode = HRTIMER_MODE_ABS;
3788 } while (hs.task && !signal_pending(current));
3790 __set_current_state(TASK_RUNNING);
3791 destroy_hrtimer_on_stack(&hs.timer);
3795 static bool blk_mq_poll_hybrid(struct request_queue *q,
3796 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3800 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3803 if (!blk_qc_t_is_internal(cookie))
3804 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3806 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3808 * With scheduling, if the request has completed, we'll
3809 * get a NULL return here, as we clear the sched tag when
3810 * that happens. The request still remains valid, like always,
3811 * so we should be safe with just the NULL check.
3817 return blk_mq_poll_hybrid_sleep(q, rq);
3821 * blk_poll - poll for IO completions
3823 * @cookie: cookie passed back at IO submission time
3824 * @spin: whether to spin for completions
3827 * Poll for completions on the passed in queue. Returns number of
3828 * completed entries found. If @spin is true, then blk_poll will continue
3829 * looping until at least one completion is found, unless the task is
3830 * otherwise marked running (or we need to reschedule).
3832 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3834 struct blk_mq_hw_ctx *hctx;
3837 if (!blk_qc_t_valid(cookie) ||
3838 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3842 blk_flush_plug_list(current->plug, false);
3844 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3847 * If we sleep, have the caller restart the poll loop to reset
3848 * the state. Like for the other success return cases, the
3849 * caller is responsible for checking if the IO completed. If
3850 * the IO isn't complete, we'll get called again and will go
3851 * straight to the busy poll loop.
3853 if (blk_mq_poll_hybrid(q, hctx, cookie))
3856 hctx->poll_considered++;
3858 state = current->state;
3862 hctx->poll_invoked++;
3864 ret = q->mq_ops->poll(hctx);
3866 hctx->poll_success++;
3867 __set_current_state(TASK_RUNNING);
3871 if (signal_pending_state(state, current))
3872 __set_current_state(TASK_RUNNING);
3874 if (current->state == TASK_RUNNING)
3876 if (ret < 0 || !spin)
3879 } while (!need_resched());
3881 __set_current_state(TASK_RUNNING);
3884 EXPORT_SYMBOL_GPL(blk_poll);
3886 unsigned int blk_mq_rq_cpu(struct request *rq)
3888 return rq->mq_ctx->cpu;
3890 EXPORT_SYMBOL(blk_mq_rq_cpu);
3892 static int __init blk_mq_init(void)
3896 for_each_possible_cpu(i)
3897 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3898 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3900 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3901 "block/softirq:dead", NULL,
3902 blk_softirq_cpu_dead);
3903 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3904 blk_mq_hctx_notify_dead);
3905 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3906 blk_mq_hctx_notify_online,
3907 blk_mq_hctx_notify_offline);
3910 subsys_initcall(blk_mq_init);