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];
280 req_flags_t rq_flags = 0;
282 if (data->q->elevator) {
283 rq->tag = BLK_MQ_NO_TAG;
284 rq->internal_tag = tag;
286 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
287 rq_flags = RQF_MQ_INFLIGHT;
288 atomic_inc(&data->hctx->nr_active);
291 rq->internal_tag = BLK_MQ_NO_TAG;
292 data->hctx->tags->rqs[rq->tag] = rq;
295 /* csd/requeue_work/fifo_time is initialized before use */
297 rq->mq_ctx = data->ctx;
298 rq->mq_hctx = data->hctx;
299 rq->rq_flags = rq_flags;
300 rq->cmd_flags = data->cmd_flags;
301 if (data->flags & BLK_MQ_REQ_PREEMPT)
302 rq->rq_flags |= RQF_PREEMPT;
303 if (blk_queue_io_stat(data->q))
304 rq->rq_flags |= RQF_IO_STAT;
305 INIT_LIST_HEAD(&rq->queuelist);
306 INIT_HLIST_NODE(&rq->hash);
307 RB_CLEAR_NODE(&rq->rb_node);
310 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
311 rq->alloc_time_ns = alloc_time_ns;
313 if (blk_mq_need_time_stamp(rq))
314 rq->start_time_ns = ktime_get_ns();
316 rq->start_time_ns = 0;
317 rq->io_start_time_ns = 0;
318 rq->stats_sectors = 0;
319 rq->nr_phys_segments = 0;
320 #if defined(CONFIG_BLK_DEV_INTEGRITY)
321 rq->nr_integrity_segments = 0;
323 blk_crypto_rq_set_defaults(rq);
324 /* tag was already set */
325 WRITE_ONCE(rq->deadline, 0);
330 rq->end_io_data = NULL;
332 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
333 refcount_set(&rq->ref, 1);
335 if (!op_is_flush(data->cmd_flags)) {
336 struct elevator_queue *e = data->q->elevator;
339 if (e && e->type->ops.prepare_request) {
340 if (e->type->icq_cache)
341 blk_mq_sched_assign_ioc(rq);
343 e->type->ops.prepare_request(rq);
344 rq->rq_flags |= RQF_ELVPRIV;
348 data->hctx->queued++;
352 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
354 struct request_queue *q = data->q;
355 struct elevator_queue *e = q->elevator;
356 u64 alloc_time_ns = 0;
359 /* alloc_time includes depth and tag waits */
360 if (blk_queue_rq_alloc_time(q))
361 alloc_time_ns = ktime_get_ns();
363 if (data->cmd_flags & REQ_NOWAIT)
364 data->flags |= BLK_MQ_REQ_NOWAIT;
368 * Flush requests are special and go directly to the
369 * dispatch list. Don't include reserved tags in the
370 * limiting, as it isn't useful.
372 if (!op_is_flush(data->cmd_flags) &&
373 e->type->ops.limit_depth &&
374 !(data->flags & BLK_MQ_REQ_RESERVED))
375 e->type->ops.limit_depth(data->cmd_flags, data);
379 data->ctx = blk_mq_get_ctx(q);
380 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
382 blk_mq_tag_busy(data->hctx);
385 * Waiting allocations only fail because of an inactive hctx. In that
386 * case just retry the hctx assignment and tag allocation as CPU hotplug
387 * should have migrated us to an online CPU by now.
389 tag = blk_mq_get_tag(data);
390 if (tag == BLK_MQ_NO_TAG) {
391 if (data->flags & BLK_MQ_REQ_NOWAIT)
395 * Give up the CPU and sleep for a random short time to ensure
396 * that thread using a realtime scheduling class are migrated
397 * off the the CPU, and thus off the hctx that is going away.
402 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
405 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
406 blk_mq_req_flags_t flags)
408 struct blk_mq_alloc_data data = {
416 ret = blk_queue_enter(q, flags);
420 rq = __blk_mq_alloc_request(&data);
424 rq->__sector = (sector_t) -1;
425 rq->bio = rq->biotail = NULL;
429 return ERR_PTR(-EWOULDBLOCK);
431 EXPORT_SYMBOL(blk_mq_alloc_request);
433 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
434 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
436 struct blk_mq_alloc_data data = {
441 u64 alloc_time_ns = 0;
446 /* alloc_time includes depth and tag waits */
447 if (blk_queue_rq_alloc_time(q))
448 alloc_time_ns = ktime_get_ns();
451 * If the tag allocator sleeps we could get an allocation for a
452 * different hardware context. No need to complicate the low level
453 * allocator for this for the rare use case of a command tied to
456 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
457 return ERR_PTR(-EINVAL);
459 if (hctx_idx >= q->nr_hw_queues)
460 return ERR_PTR(-EIO);
462 ret = blk_queue_enter(q, flags);
467 * Check if the hardware context is actually mapped to anything.
468 * If not tell the caller that it should skip this queue.
471 data.hctx = q->queue_hw_ctx[hctx_idx];
472 if (!blk_mq_hw_queue_mapped(data.hctx))
474 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
475 data.ctx = __blk_mq_get_ctx(q, cpu);
478 blk_mq_tag_busy(data.hctx);
481 tag = blk_mq_get_tag(&data);
482 if (tag == BLK_MQ_NO_TAG)
484 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
490 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
492 static void __blk_mq_free_request(struct request *rq)
494 struct request_queue *q = rq->q;
495 struct blk_mq_ctx *ctx = rq->mq_ctx;
496 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
497 const int sched_tag = rq->internal_tag;
499 blk_crypto_free_request(rq);
500 blk_pm_mark_last_busy(rq);
502 if (rq->tag != BLK_MQ_NO_TAG)
503 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
504 if (sched_tag != BLK_MQ_NO_TAG)
505 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
506 blk_mq_sched_restart(hctx);
510 void blk_mq_free_request(struct request *rq)
512 struct request_queue *q = rq->q;
513 struct elevator_queue *e = q->elevator;
514 struct blk_mq_ctx *ctx = rq->mq_ctx;
515 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
517 if (rq->rq_flags & RQF_ELVPRIV) {
518 if (e && e->type->ops.finish_request)
519 e->type->ops.finish_request(rq);
521 put_io_context(rq->elv.icq->ioc);
526 ctx->rq_completed[rq_is_sync(rq)]++;
527 if (rq->rq_flags & RQF_MQ_INFLIGHT)
528 atomic_dec(&hctx->nr_active);
530 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
531 laptop_io_completion(q->backing_dev_info);
535 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
536 if (refcount_dec_and_test(&rq->ref))
537 __blk_mq_free_request(rq);
539 EXPORT_SYMBOL_GPL(blk_mq_free_request);
541 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
545 if (blk_mq_need_time_stamp(rq))
546 now = ktime_get_ns();
548 if (rq->rq_flags & RQF_STATS) {
549 blk_mq_poll_stats_start(rq->q);
550 blk_stat_add(rq, now);
553 if (rq->internal_tag != BLK_MQ_NO_TAG)
554 blk_mq_sched_completed_request(rq, now);
556 blk_account_io_done(rq, now);
559 rq_qos_done(rq->q, rq);
560 rq->end_io(rq, error);
562 blk_mq_free_request(rq);
565 EXPORT_SYMBOL(__blk_mq_end_request);
567 void blk_mq_end_request(struct request *rq, blk_status_t error)
569 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
571 __blk_mq_end_request(rq, error);
573 EXPORT_SYMBOL(blk_mq_end_request);
576 * Softirq action handler - move entries to local list and loop over them
577 * while passing them to the queue registered handler.
579 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
581 struct list_head *cpu_list, local_list;
584 cpu_list = this_cpu_ptr(&blk_cpu_done);
585 list_replace_init(cpu_list, &local_list);
588 while (!list_empty(&local_list)) {
591 rq = list_entry(local_list.next, struct request, ipi_list);
592 list_del_init(&rq->ipi_list);
593 rq->q->mq_ops->complete(rq);
597 static void blk_mq_trigger_softirq(struct request *rq)
599 struct list_head *list;
602 local_irq_save(flags);
603 list = this_cpu_ptr(&blk_cpu_done);
604 list_add_tail(&rq->ipi_list, list);
607 * If the list only contains our just added request, signal a raise of
608 * the softirq. If there are already entries there, someone already
609 * raised the irq but it hasn't run yet.
611 if (list->next == &rq->ipi_list)
612 raise_softirq_irqoff(BLOCK_SOFTIRQ);
613 local_irq_restore(flags);
616 static int blk_softirq_cpu_dead(unsigned int cpu)
619 * If a CPU goes away, splice its entries to the current CPU
620 * and trigger a run of the softirq
623 list_splice_init(&per_cpu(blk_cpu_done, cpu),
624 this_cpu_ptr(&blk_cpu_done));
625 raise_softirq_irqoff(BLOCK_SOFTIRQ);
632 static void __blk_mq_complete_request_remote(void *data)
634 struct request *rq = data;
637 * For most of single queue controllers, there is only one irq vector
638 * for handling I/O completion, and the only irq's affinity is set
639 * to all possible CPUs. On most of ARCHs, this affinity means the irq
640 * is handled on one specific CPU.
642 * So complete I/O requests in softirq context in case of single queue
643 * devices to avoid degrading I/O performance due to irqsoff latency.
645 if (rq->q->nr_hw_queues == 1)
646 blk_mq_trigger_softirq(rq);
648 rq->q->mq_ops->complete(rq);
651 static inline bool blk_mq_complete_need_ipi(struct request *rq)
653 int cpu = raw_smp_processor_id();
655 if (!IS_ENABLED(CONFIG_SMP) ||
656 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
659 /* same CPU or cache domain? Complete locally */
660 if (cpu == rq->mq_ctx->cpu ||
661 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
662 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
665 /* don't try to IPI to an offline CPU */
666 return cpu_online(rq->mq_ctx->cpu);
669 bool blk_mq_complete_request_remote(struct request *rq)
671 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
673 blk_mq_put_driver_tag(rq);
676 * For a polled request, always complete locallly, it's pointless
677 * to redirect the completion.
679 if (rq->cmd_flags & REQ_HIPRI)
682 if (blk_mq_complete_need_ipi(rq)) {
683 rq->csd.func = __blk_mq_complete_request_remote;
686 smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
688 if (rq->q->nr_hw_queues > 1)
690 blk_mq_trigger_softirq(rq);
695 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
698 * blk_mq_complete_request - end I/O on a request
699 * @rq: the request being processed
702 * Complete a request by scheduling the ->complete_rq operation.
704 void blk_mq_complete_request(struct request *rq)
706 if (!blk_mq_complete_request_remote(rq))
707 rq->q->mq_ops->complete(rq);
709 EXPORT_SYMBOL(blk_mq_complete_request);
711 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
712 __releases(hctx->srcu)
714 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
717 srcu_read_unlock(hctx->srcu, srcu_idx);
720 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
721 __acquires(hctx->srcu)
723 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
724 /* shut up gcc false positive */
728 *srcu_idx = srcu_read_lock(hctx->srcu);
732 * blk_mq_start_request - Start processing a request
733 * @rq: Pointer to request to be started
735 * Function used by device drivers to notify the block layer that a request
736 * is going to be processed now, so blk layer can do proper initializations
737 * such as starting the timeout timer.
739 void blk_mq_start_request(struct request *rq)
741 struct request_queue *q = rq->q;
743 trace_block_rq_issue(q, rq);
745 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
746 rq->io_start_time_ns = ktime_get_ns();
747 rq->stats_sectors = blk_rq_sectors(rq);
748 rq->rq_flags |= RQF_STATS;
752 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
755 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
757 #ifdef CONFIG_BLK_DEV_INTEGRITY
758 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
759 q->integrity.profile->prepare_fn(rq);
762 EXPORT_SYMBOL(blk_mq_start_request);
764 static void __blk_mq_requeue_request(struct request *rq)
766 struct request_queue *q = rq->q;
768 blk_mq_put_driver_tag(rq);
770 trace_block_rq_requeue(q, rq);
771 rq_qos_requeue(q, rq);
773 if (blk_mq_request_started(rq)) {
774 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
775 rq->rq_flags &= ~RQF_TIMED_OUT;
779 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
781 __blk_mq_requeue_request(rq);
783 /* this request will be re-inserted to io scheduler queue */
784 blk_mq_sched_requeue_request(rq);
786 BUG_ON(!list_empty(&rq->queuelist));
787 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
789 EXPORT_SYMBOL(blk_mq_requeue_request);
791 static void blk_mq_requeue_work(struct work_struct *work)
793 struct request_queue *q =
794 container_of(work, struct request_queue, requeue_work.work);
796 struct request *rq, *next;
798 spin_lock_irq(&q->requeue_lock);
799 list_splice_init(&q->requeue_list, &rq_list);
800 spin_unlock_irq(&q->requeue_lock);
802 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
803 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
806 rq->rq_flags &= ~RQF_SOFTBARRIER;
807 list_del_init(&rq->queuelist);
809 * If RQF_DONTPREP, rq has contained some driver specific
810 * data, so insert it to hctx dispatch list to avoid any
813 if (rq->rq_flags & RQF_DONTPREP)
814 blk_mq_request_bypass_insert(rq, false, false);
816 blk_mq_sched_insert_request(rq, true, false, false);
819 while (!list_empty(&rq_list)) {
820 rq = list_entry(rq_list.next, struct request, queuelist);
821 list_del_init(&rq->queuelist);
822 blk_mq_sched_insert_request(rq, false, false, false);
825 blk_mq_run_hw_queues(q, false);
828 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
829 bool kick_requeue_list)
831 struct request_queue *q = rq->q;
835 * We abuse this flag that is otherwise used by the I/O scheduler to
836 * request head insertion from the workqueue.
838 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
840 spin_lock_irqsave(&q->requeue_lock, flags);
842 rq->rq_flags |= RQF_SOFTBARRIER;
843 list_add(&rq->queuelist, &q->requeue_list);
845 list_add_tail(&rq->queuelist, &q->requeue_list);
847 spin_unlock_irqrestore(&q->requeue_lock, flags);
849 if (kick_requeue_list)
850 blk_mq_kick_requeue_list(q);
853 void blk_mq_kick_requeue_list(struct request_queue *q)
855 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
857 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
859 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
862 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
863 msecs_to_jiffies(msecs));
865 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
867 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
869 if (tag < tags->nr_tags) {
870 prefetch(tags->rqs[tag]);
871 return tags->rqs[tag];
876 EXPORT_SYMBOL(blk_mq_tag_to_rq);
878 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
879 void *priv, bool reserved)
882 * If we find a request that is inflight and the queue matches,
883 * we know the queue is busy. Return false to stop the iteration.
885 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
895 bool blk_mq_queue_inflight(struct request_queue *q)
899 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
902 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
904 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
906 req->rq_flags |= RQF_TIMED_OUT;
907 if (req->q->mq_ops->timeout) {
908 enum blk_eh_timer_return ret;
910 ret = req->q->mq_ops->timeout(req, reserved);
911 if (ret == BLK_EH_DONE)
913 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
919 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
921 unsigned long deadline;
923 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
925 if (rq->rq_flags & RQF_TIMED_OUT)
928 deadline = READ_ONCE(rq->deadline);
929 if (time_after_eq(jiffies, deadline))
934 else if (time_after(*next, deadline))
939 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
940 struct request *rq, void *priv, bool reserved)
942 unsigned long *next = priv;
945 * Just do a quick check if it is expired before locking the request in
946 * so we're not unnecessarilly synchronizing across CPUs.
948 if (!blk_mq_req_expired(rq, next))
952 * We have reason to believe the request may be expired. Take a
953 * reference on the request to lock this request lifetime into its
954 * currently allocated context to prevent it from being reallocated in
955 * the event the completion by-passes this timeout handler.
957 * If the reference was already released, then the driver beat the
958 * timeout handler to posting a natural completion.
960 if (!refcount_inc_not_zero(&rq->ref))
964 * The request is now locked and cannot be reallocated underneath the
965 * timeout handler's processing. Re-verify this exact request is truly
966 * expired; if it is not expired, then the request was completed and
967 * reallocated as a new request.
969 if (blk_mq_req_expired(rq, next))
970 blk_mq_rq_timed_out(rq, reserved);
972 if (is_flush_rq(rq, hctx))
974 else if (refcount_dec_and_test(&rq->ref))
975 __blk_mq_free_request(rq);
980 static void blk_mq_timeout_work(struct work_struct *work)
982 struct request_queue *q =
983 container_of(work, struct request_queue, timeout_work);
984 unsigned long next = 0;
985 struct blk_mq_hw_ctx *hctx;
988 /* A deadlock might occur if a request is stuck requiring a
989 * timeout at the same time a queue freeze is waiting
990 * completion, since the timeout code would not be able to
991 * acquire the queue reference here.
993 * That's why we don't use blk_queue_enter here; instead, we use
994 * percpu_ref_tryget directly, because we need to be able to
995 * obtain a reference even in the short window between the queue
996 * starting to freeze, by dropping the first reference in
997 * blk_freeze_queue_start, and the moment the last request is
998 * consumed, marked by the instant q_usage_counter reaches
1001 if (!percpu_ref_tryget(&q->q_usage_counter))
1004 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1007 mod_timer(&q->timeout, next);
1010 * Request timeouts are handled as a forward rolling timer. If
1011 * we end up here it means that no requests are pending and
1012 * also that no request has been pending for a while. Mark
1013 * each hctx as idle.
1015 queue_for_each_hw_ctx(q, hctx, i) {
1016 /* the hctx may be unmapped, so check it here */
1017 if (blk_mq_hw_queue_mapped(hctx))
1018 blk_mq_tag_idle(hctx);
1024 struct flush_busy_ctx_data {
1025 struct blk_mq_hw_ctx *hctx;
1026 struct list_head *list;
1029 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1031 struct flush_busy_ctx_data *flush_data = data;
1032 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1033 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1034 enum hctx_type type = hctx->type;
1036 spin_lock(&ctx->lock);
1037 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1038 sbitmap_clear_bit(sb, bitnr);
1039 spin_unlock(&ctx->lock);
1044 * Process software queues that have been marked busy, splicing them
1045 * to the for-dispatch
1047 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1049 struct flush_busy_ctx_data data = {
1054 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1056 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1058 struct dispatch_rq_data {
1059 struct blk_mq_hw_ctx *hctx;
1063 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1066 struct dispatch_rq_data *dispatch_data = data;
1067 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1068 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1069 enum hctx_type type = hctx->type;
1071 spin_lock(&ctx->lock);
1072 if (!list_empty(&ctx->rq_lists[type])) {
1073 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1074 list_del_init(&dispatch_data->rq->queuelist);
1075 if (list_empty(&ctx->rq_lists[type]))
1076 sbitmap_clear_bit(sb, bitnr);
1078 spin_unlock(&ctx->lock);
1080 return !dispatch_data->rq;
1083 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1084 struct blk_mq_ctx *start)
1086 unsigned off = start ? start->index_hw[hctx->type] : 0;
1087 struct dispatch_rq_data data = {
1092 __sbitmap_for_each_set(&hctx->ctx_map, off,
1093 dispatch_rq_from_ctx, &data);
1098 static inline unsigned int queued_to_index(unsigned int queued)
1103 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1106 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1107 int flags, void *key)
1109 struct blk_mq_hw_ctx *hctx;
1111 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1113 spin_lock(&hctx->dispatch_wait_lock);
1114 if (!list_empty(&wait->entry)) {
1115 struct sbitmap_queue *sbq;
1117 list_del_init(&wait->entry);
1118 sbq = &hctx->tags->bitmap_tags;
1119 atomic_dec(&sbq->ws_active);
1121 spin_unlock(&hctx->dispatch_wait_lock);
1123 blk_mq_run_hw_queue(hctx, true);
1128 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1129 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1130 * restart. For both cases, take care to check the condition again after
1131 * marking us as waiting.
1133 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1136 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1137 struct wait_queue_head *wq;
1138 wait_queue_entry_t *wait;
1141 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1142 blk_mq_sched_mark_restart_hctx(hctx);
1145 * It's possible that a tag was freed in the window between the
1146 * allocation failure and adding the hardware queue to the wait
1149 * Don't clear RESTART here, someone else could have set it.
1150 * At most this will cost an extra queue run.
1152 return blk_mq_get_driver_tag(rq);
1155 wait = &hctx->dispatch_wait;
1156 if (!list_empty_careful(&wait->entry))
1159 wq = &bt_wait_ptr(sbq, hctx)->wait;
1161 spin_lock_irq(&wq->lock);
1162 spin_lock(&hctx->dispatch_wait_lock);
1163 if (!list_empty(&wait->entry)) {
1164 spin_unlock(&hctx->dispatch_wait_lock);
1165 spin_unlock_irq(&wq->lock);
1169 atomic_inc(&sbq->ws_active);
1170 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1171 __add_wait_queue(wq, wait);
1174 * It's possible that a tag was freed in the window between the
1175 * allocation failure and adding the hardware queue to the wait
1178 ret = blk_mq_get_driver_tag(rq);
1180 spin_unlock(&hctx->dispatch_wait_lock);
1181 spin_unlock_irq(&wq->lock);
1186 * We got a tag, remove ourselves from the wait queue to ensure
1187 * someone else gets the wakeup.
1189 list_del_init(&wait->entry);
1190 atomic_dec(&sbq->ws_active);
1191 spin_unlock(&hctx->dispatch_wait_lock);
1192 spin_unlock_irq(&wq->lock);
1197 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1198 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1200 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1201 * - EWMA is one simple way to compute running average value
1202 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1203 * - take 4 as factor for avoiding to get too small(0) result, and this
1204 * factor doesn't matter because EWMA decreases exponentially
1206 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1210 if (hctx->queue->elevator)
1213 ewma = hctx->dispatch_busy;
1218 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1220 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1221 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1223 hctx->dispatch_busy = ewma;
1226 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1228 static void blk_mq_handle_dev_resource(struct request *rq,
1229 struct list_head *list)
1231 struct request *next =
1232 list_first_entry_or_null(list, struct request, queuelist);
1235 * If an I/O scheduler has been configured and we got a driver tag for
1236 * the next request already, free it.
1239 blk_mq_put_driver_tag(next);
1241 list_add(&rq->queuelist, list);
1242 __blk_mq_requeue_request(rq);
1245 static void blk_mq_handle_zone_resource(struct request *rq,
1246 struct list_head *zone_list)
1249 * If we end up here it is because we cannot dispatch a request to a
1250 * specific zone due to LLD level zone-write locking or other zone
1251 * related resource not being available. In this case, set the request
1252 * aside in zone_list for retrying it later.
1254 list_add(&rq->queuelist, zone_list);
1255 __blk_mq_requeue_request(rq);
1258 enum prep_dispatch {
1260 PREP_DISPATCH_NO_TAG,
1261 PREP_DISPATCH_NO_BUDGET,
1264 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1267 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1269 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1270 blk_mq_put_driver_tag(rq);
1271 return PREP_DISPATCH_NO_BUDGET;
1274 if (!blk_mq_get_driver_tag(rq)) {
1276 * The initial allocation attempt failed, so we need to
1277 * rerun the hardware queue when a tag is freed. The
1278 * waitqueue takes care of that. If the queue is run
1279 * before we add this entry back on the dispatch list,
1280 * we'll re-run it below.
1282 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1284 * All budgets not got from this function will be put
1285 * together during handling partial dispatch
1288 blk_mq_put_dispatch_budget(rq->q);
1289 return PREP_DISPATCH_NO_TAG;
1293 return PREP_DISPATCH_OK;
1296 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1297 static void blk_mq_release_budgets(struct request_queue *q,
1298 unsigned int nr_budgets)
1302 for (i = 0; i < nr_budgets; i++)
1303 blk_mq_put_dispatch_budget(q);
1307 * Returns true if we did some work AND can potentially do more.
1309 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1310 unsigned int nr_budgets)
1312 enum prep_dispatch prep;
1313 struct request_queue *q = hctx->queue;
1314 struct request *rq, *nxt;
1316 blk_status_t ret = BLK_STS_OK;
1317 LIST_HEAD(zone_list);
1319 if (list_empty(list))
1323 * Now process all the entries, sending them to the driver.
1325 errors = queued = 0;
1327 struct blk_mq_queue_data bd;
1329 rq = list_first_entry(list, struct request, queuelist);
1331 WARN_ON_ONCE(hctx != rq->mq_hctx);
1332 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1333 if (prep != PREP_DISPATCH_OK)
1336 list_del_init(&rq->queuelist);
1341 * Flag last if we have no more requests, or if we have more
1342 * but can't assign a driver tag to it.
1344 if (list_empty(list))
1347 nxt = list_first_entry(list, struct request, queuelist);
1348 bd.last = !blk_mq_get_driver_tag(nxt);
1352 * once the request is queued to lld, no need to cover the
1357 ret = q->mq_ops->queue_rq(hctx, &bd);
1358 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1359 blk_mq_handle_dev_resource(rq, list);
1361 } else if (ret == BLK_STS_ZONE_RESOURCE) {
1363 * Move the request to zone_list and keep going through
1364 * the dispatch list to find more requests the drive can
1367 blk_mq_handle_zone_resource(rq, &zone_list);
1368 if (list_empty(list))
1373 if (unlikely(ret != BLK_STS_OK)) {
1375 blk_mq_end_request(rq, BLK_STS_IOERR);
1380 } while (!list_empty(list));
1382 if (!list_empty(&zone_list))
1383 list_splice_tail_init(&zone_list, list);
1385 hctx->dispatched[queued_to_index(queued)]++;
1388 * Any items that need requeuing? Stuff them into hctx->dispatch,
1389 * that is where we will continue on next queue run.
1391 if (!list_empty(list)) {
1393 /* For non-shared tags, the RESTART check will suffice */
1394 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1395 (hctx->flags & BLK_MQ_F_TAG_SHARED);
1396 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1398 blk_mq_release_budgets(q, nr_budgets);
1401 * If we didn't flush the entire list, we could have told
1402 * the driver there was more coming, but that turned out to
1405 if (q->mq_ops->commit_rqs && queued)
1406 q->mq_ops->commit_rqs(hctx);
1408 spin_lock(&hctx->lock);
1409 list_splice_tail_init(list, &hctx->dispatch);
1410 spin_unlock(&hctx->lock);
1413 * If SCHED_RESTART was set by the caller of this function and
1414 * it is no longer set that means that it was cleared by another
1415 * thread and hence that a queue rerun is needed.
1417 * If 'no_tag' is set, that means that we failed getting
1418 * a driver tag with an I/O scheduler attached. If our dispatch
1419 * waitqueue is no longer active, ensure that we run the queue
1420 * AFTER adding our entries back to the list.
1422 * If no I/O scheduler has been configured it is possible that
1423 * the hardware queue got stopped and restarted before requests
1424 * were pushed back onto the dispatch list. Rerun the queue to
1425 * avoid starvation. Notes:
1426 * - blk_mq_run_hw_queue() checks whether or not a queue has
1427 * been stopped before rerunning a queue.
1428 * - Some but not all block drivers stop a queue before
1429 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1432 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1433 * bit is set, run queue after a delay to avoid IO stalls
1434 * that could otherwise occur if the queue is idle. We'll do
1435 * similar if we couldn't get budget and SCHED_RESTART is set.
1437 needs_restart = blk_mq_sched_needs_restart(hctx);
1438 if (!needs_restart ||
1439 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1440 blk_mq_run_hw_queue(hctx, true);
1441 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1443 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1445 blk_mq_update_dispatch_busy(hctx, true);
1448 blk_mq_update_dispatch_busy(hctx, false);
1450 return (queued + errors) != 0;
1454 * __blk_mq_run_hw_queue - Run a hardware queue.
1455 * @hctx: Pointer to the hardware queue to run.
1457 * Send pending requests to the hardware.
1459 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1464 * We should be running this queue from one of the CPUs that
1467 * There are at least two related races now between setting
1468 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1469 * __blk_mq_run_hw_queue():
1471 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1472 * but later it becomes online, then this warning is harmless
1475 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1476 * but later it becomes offline, then the warning can't be
1477 * triggered, and we depend on blk-mq timeout handler to
1478 * handle dispatched requests to this hctx
1480 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1481 cpu_online(hctx->next_cpu)) {
1482 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1483 raw_smp_processor_id(),
1484 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1489 * We can't run the queue inline with ints disabled. Ensure that
1490 * we catch bad users of this early.
1492 WARN_ON_ONCE(in_interrupt());
1494 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1496 hctx_lock(hctx, &srcu_idx);
1497 blk_mq_sched_dispatch_requests(hctx);
1498 hctx_unlock(hctx, srcu_idx);
1501 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1503 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1505 if (cpu >= nr_cpu_ids)
1506 cpu = cpumask_first(hctx->cpumask);
1511 * It'd be great if the workqueue API had a way to pass
1512 * in a mask and had some smarts for more clever placement.
1513 * For now we just round-robin here, switching for every
1514 * BLK_MQ_CPU_WORK_BATCH queued items.
1516 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1519 int next_cpu = hctx->next_cpu;
1521 if (hctx->queue->nr_hw_queues == 1)
1522 return WORK_CPU_UNBOUND;
1524 if (--hctx->next_cpu_batch <= 0) {
1526 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1528 if (next_cpu >= nr_cpu_ids)
1529 next_cpu = blk_mq_first_mapped_cpu(hctx);
1530 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1534 * Do unbound schedule if we can't find a online CPU for this hctx,
1535 * and it should only happen in the path of handling CPU DEAD.
1537 if (!cpu_online(next_cpu)) {
1544 * Make sure to re-select CPU next time once after CPUs
1545 * in hctx->cpumask become online again.
1547 hctx->next_cpu = next_cpu;
1548 hctx->next_cpu_batch = 1;
1549 return WORK_CPU_UNBOUND;
1552 hctx->next_cpu = next_cpu;
1557 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1558 * @hctx: Pointer to the hardware queue to run.
1559 * @async: If we want to run the queue asynchronously.
1560 * @msecs: Microseconds of delay to wait before running the queue.
1562 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1563 * with a delay of @msecs.
1565 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1566 unsigned long msecs)
1568 if (unlikely(blk_mq_hctx_stopped(hctx)))
1571 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1572 int cpu = get_cpu();
1573 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1574 __blk_mq_run_hw_queue(hctx);
1582 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1583 msecs_to_jiffies(msecs));
1587 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1588 * @hctx: Pointer to the hardware queue to run.
1589 * @msecs: Microseconds of delay to wait before running the queue.
1591 * Run a hardware queue asynchronously with a delay of @msecs.
1593 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1595 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1597 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1600 * blk_mq_run_hw_queue - Start to run a hardware queue.
1601 * @hctx: Pointer to the hardware queue to run.
1602 * @async: If we want to run the queue asynchronously.
1604 * Check if the request queue is not in a quiesced state and if there are
1605 * pending requests to be sent. If this is true, run the queue to send requests
1608 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1614 * When queue is quiesced, we may be switching io scheduler, or
1615 * updating nr_hw_queues, or other things, and we can't run queue
1616 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1618 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1621 hctx_lock(hctx, &srcu_idx);
1622 need_run = !blk_queue_quiesced(hctx->queue) &&
1623 blk_mq_hctx_has_pending(hctx);
1624 hctx_unlock(hctx, srcu_idx);
1627 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1629 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1632 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1633 * @q: Pointer to the request queue to run.
1634 * @async: If we want to run the queue asynchronously.
1636 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1638 struct blk_mq_hw_ctx *hctx;
1641 queue_for_each_hw_ctx(q, hctx, i) {
1642 if (blk_mq_hctx_stopped(hctx))
1645 blk_mq_run_hw_queue(hctx, async);
1648 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1651 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1652 * @q: Pointer to the request queue to run.
1653 * @msecs: Microseconds of delay to wait before running the queues.
1655 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1657 struct blk_mq_hw_ctx *hctx;
1660 queue_for_each_hw_ctx(q, hctx, i) {
1661 if (blk_mq_hctx_stopped(hctx))
1664 blk_mq_delay_run_hw_queue(hctx, msecs);
1667 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1670 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1671 * @q: request queue.
1673 * The caller is responsible for serializing this function against
1674 * blk_mq_{start,stop}_hw_queue().
1676 bool blk_mq_queue_stopped(struct request_queue *q)
1678 struct blk_mq_hw_ctx *hctx;
1681 queue_for_each_hw_ctx(q, hctx, i)
1682 if (blk_mq_hctx_stopped(hctx))
1687 EXPORT_SYMBOL(blk_mq_queue_stopped);
1690 * This function is often used for pausing .queue_rq() by driver when
1691 * there isn't enough resource or some conditions aren't satisfied, and
1692 * BLK_STS_RESOURCE is usually returned.
1694 * We do not guarantee that dispatch can be drained or blocked
1695 * after blk_mq_stop_hw_queue() returns. Please use
1696 * blk_mq_quiesce_queue() for that requirement.
1698 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1700 cancel_delayed_work(&hctx->run_work);
1702 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1704 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1707 * This function is often used for pausing .queue_rq() by driver when
1708 * there isn't enough resource or some conditions aren't satisfied, and
1709 * BLK_STS_RESOURCE is usually returned.
1711 * We do not guarantee that dispatch can be drained or blocked
1712 * after blk_mq_stop_hw_queues() returns. Please use
1713 * blk_mq_quiesce_queue() for that requirement.
1715 void blk_mq_stop_hw_queues(struct request_queue *q)
1717 struct blk_mq_hw_ctx *hctx;
1720 queue_for_each_hw_ctx(q, hctx, i)
1721 blk_mq_stop_hw_queue(hctx);
1723 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1725 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1727 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1729 blk_mq_run_hw_queue(hctx, false);
1731 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1733 void blk_mq_start_hw_queues(struct request_queue *q)
1735 struct blk_mq_hw_ctx *hctx;
1738 queue_for_each_hw_ctx(q, hctx, i)
1739 blk_mq_start_hw_queue(hctx);
1741 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1743 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1745 if (!blk_mq_hctx_stopped(hctx))
1748 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1749 blk_mq_run_hw_queue(hctx, async);
1751 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1753 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1755 struct blk_mq_hw_ctx *hctx;
1758 queue_for_each_hw_ctx(q, hctx, i)
1759 blk_mq_start_stopped_hw_queue(hctx, async);
1761 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1763 static void blk_mq_run_work_fn(struct work_struct *work)
1765 struct blk_mq_hw_ctx *hctx;
1767 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1770 * If we are stopped, don't run the queue.
1772 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1775 __blk_mq_run_hw_queue(hctx);
1778 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1782 struct blk_mq_ctx *ctx = rq->mq_ctx;
1783 enum hctx_type type = hctx->type;
1785 lockdep_assert_held(&ctx->lock);
1787 trace_block_rq_insert(hctx->queue, rq);
1790 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1792 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1795 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1798 struct blk_mq_ctx *ctx = rq->mq_ctx;
1800 lockdep_assert_held(&ctx->lock);
1802 __blk_mq_insert_req_list(hctx, rq, at_head);
1803 blk_mq_hctx_mark_pending(hctx, ctx);
1807 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1808 * @rq: Pointer to request to be inserted.
1809 * @run_queue: If we should run the hardware queue after inserting the request.
1811 * Should only be used carefully, when the caller knows we want to
1812 * bypass a potential IO scheduler on the target device.
1814 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1817 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1819 spin_lock(&hctx->lock);
1821 list_add(&rq->queuelist, &hctx->dispatch);
1823 list_add_tail(&rq->queuelist, &hctx->dispatch);
1824 spin_unlock(&hctx->lock);
1827 blk_mq_run_hw_queue(hctx, false);
1830 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1831 struct list_head *list)
1835 enum hctx_type type = hctx->type;
1838 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1841 list_for_each_entry(rq, list, queuelist) {
1842 BUG_ON(rq->mq_ctx != ctx);
1843 trace_block_rq_insert(hctx->queue, rq);
1846 spin_lock(&ctx->lock);
1847 list_splice_tail_init(list, &ctx->rq_lists[type]);
1848 blk_mq_hctx_mark_pending(hctx, ctx);
1849 spin_unlock(&ctx->lock);
1852 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1854 struct request *rqa = container_of(a, struct request, queuelist);
1855 struct request *rqb = container_of(b, struct request, queuelist);
1857 if (rqa->mq_ctx != rqb->mq_ctx)
1858 return rqa->mq_ctx > rqb->mq_ctx;
1859 if (rqa->mq_hctx != rqb->mq_hctx)
1860 return rqa->mq_hctx > rqb->mq_hctx;
1862 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1865 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1869 if (list_empty(&plug->mq_list))
1871 list_splice_init(&plug->mq_list, &list);
1873 if (plug->rq_count > 2 && plug->multiple_queues)
1874 list_sort(NULL, &list, plug_rq_cmp);
1879 struct list_head rq_list;
1880 struct request *rq, *head_rq = list_entry_rq(list.next);
1881 struct list_head *pos = &head_rq->queuelist; /* skip first */
1882 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1883 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1884 unsigned int depth = 1;
1886 list_for_each_continue(pos, &list) {
1887 rq = list_entry_rq(pos);
1889 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1894 list_cut_before(&rq_list, &list, pos);
1895 trace_block_unplug(head_rq->q, depth, !from_schedule);
1896 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1898 } while(!list_empty(&list));
1901 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1902 unsigned int nr_segs)
1904 if (bio->bi_opf & REQ_RAHEAD)
1905 rq->cmd_flags |= REQ_FAILFAST_MASK;
1907 rq->__sector = bio->bi_iter.bi_sector;
1908 rq->write_hint = bio->bi_write_hint;
1909 blk_rq_bio_prep(rq, bio, nr_segs);
1910 blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1912 blk_account_io_start(rq);
1915 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1917 blk_qc_t *cookie, bool last)
1919 struct request_queue *q = rq->q;
1920 struct blk_mq_queue_data bd = {
1924 blk_qc_t new_cookie;
1927 new_cookie = request_to_qc_t(hctx, rq);
1930 * For OK queue, we are done. For error, caller may kill it.
1931 * Any other error (busy), just add it to our list as we
1932 * previously would have done.
1934 ret = q->mq_ops->queue_rq(hctx, &bd);
1937 blk_mq_update_dispatch_busy(hctx, false);
1938 *cookie = new_cookie;
1940 case BLK_STS_RESOURCE:
1941 case BLK_STS_DEV_RESOURCE:
1942 blk_mq_update_dispatch_busy(hctx, true);
1943 __blk_mq_requeue_request(rq);
1946 blk_mq_update_dispatch_busy(hctx, false);
1947 *cookie = BLK_QC_T_NONE;
1954 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1957 bool bypass_insert, bool last)
1959 struct request_queue *q = rq->q;
1960 bool run_queue = true;
1963 * RCU or SRCU read lock is needed before checking quiesced flag.
1965 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1966 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1967 * and avoid driver to try to dispatch again.
1969 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1971 bypass_insert = false;
1975 if (q->elevator && !bypass_insert)
1978 if (!blk_mq_get_dispatch_budget(q))
1981 if (!blk_mq_get_driver_tag(rq)) {
1982 blk_mq_put_dispatch_budget(q);
1986 return __blk_mq_issue_directly(hctx, rq, cookie, last);
1989 return BLK_STS_RESOURCE;
1991 blk_mq_request_bypass_insert(rq, false, run_queue);
1996 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
1997 * @hctx: Pointer of the associated hardware queue.
1998 * @rq: Pointer to request to be sent.
1999 * @cookie: Request queue cookie.
2001 * If the device has enough resources to accept a new request now, send the
2002 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2003 * we can try send it another time in the future. Requests inserted at this
2004 * queue have higher priority.
2006 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2007 struct request *rq, blk_qc_t *cookie)
2012 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2014 hctx_lock(hctx, &srcu_idx);
2016 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2017 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2018 blk_mq_request_bypass_insert(rq, false, true);
2019 else if (ret != BLK_STS_OK)
2020 blk_mq_end_request(rq, ret);
2022 hctx_unlock(hctx, srcu_idx);
2025 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2029 blk_qc_t unused_cookie;
2030 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2032 hctx_lock(hctx, &srcu_idx);
2033 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2034 hctx_unlock(hctx, srcu_idx);
2039 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2040 struct list_head *list)
2044 while (!list_empty(list)) {
2046 struct request *rq = list_first_entry(list, struct request,
2049 list_del_init(&rq->queuelist);
2050 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2051 if (ret != BLK_STS_OK) {
2052 if (ret == BLK_STS_RESOURCE ||
2053 ret == BLK_STS_DEV_RESOURCE) {
2054 blk_mq_request_bypass_insert(rq, false,
2058 blk_mq_end_request(rq, ret);
2064 * If we didn't flush the entire list, we could have told
2065 * the driver there was more coming, but that turned out to
2068 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs && queued)
2069 hctx->queue->mq_ops->commit_rqs(hctx);
2072 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2074 list_add_tail(&rq->queuelist, &plug->mq_list);
2076 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2077 struct request *tmp;
2079 tmp = list_first_entry(&plug->mq_list, struct request,
2081 if (tmp->q != rq->q)
2082 plug->multiple_queues = true;
2087 * blk_mq_make_request - Create and send a request to block device.
2088 * @q: Request queue pointer.
2089 * @bio: Bio pointer.
2091 * Builds up a request structure from @q and @bio and send to the device. The
2092 * request may not be queued directly to hardware if:
2093 * * This request can be merged with another one
2094 * * We want to place request at plug queue for possible future merging
2095 * * There is an IO scheduler active at this queue
2097 * It will not queue the request if there is an error with the bio, or at the
2100 * Returns: Request queue cookie.
2102 blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
2104 const int is_sync = op_is_sync(bio->bi_opf);
2105 const int is_flush_fua = op_is_flush(bio->bi_opf);
2106 struct blk_mq_alloc_data data = {
2110 struct blk_plug *plug;
2111 struct request *same_queue_rq = NULL;
2112 unsigned int nr_segs;
2116 blk_queue_bounce(q, &bio);
2117 __blk_queue_split(q, &bio, &nr_segs);
2119 if (!bio_integrity_prep(bio))
2122 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2123 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2126 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2129 rq_qos_throttle(q, bio);
2131 data.cmd_flags = bio->bi_opf;
2132 rq = __blk_mq_alloc_request(&data);
2133 if (unlikely(!rq)) {
2134 rq_qos_cleanup(q, bio);
2135 if (bio->bi_opf & REQ_NOWAIT)
2136 bio_wouldblock_error(bio);
2140 trace_block_getrq(q, bio, bio->bi_opf);
2142 rq_qos_track(q, rq, bio);
2144 cookie = request_to_qc_t(data.hctx, rq);
2146 blk_mq_bio_to_request(rq, bio, nr_segs);
2148 ret = blk_crypto_init_request(rq);
2149 if (ret != BLK_STS_OK) {
2150 bio->bi_status = ret;
2152 blk_mq_free_request(rq);
2153 return BLK_QC_T_NONE;
2156 plug = blk_mq_plug(q, bio);
2157 if (unlikely(is_flush_fua)) {
2158 /* Bypass scheduler for flush requests */
2159 blk_insert_flush(rq);
2160 blk_mq_run_hw_queue(data.hctx, true);
2161 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2162 !blk_queue_nonrot(q))) {
2164 * Use plugging if we have a ->commit_rqs() hook as well, as
2165 * we know the driver uses bd->last in a smart fashion.
2167 * Use normal plugging if this disk is slow HDD, as sequential
2168 * IO may benefit a lot from plug merging.
2170 unsigned int request_count = plug->rq_count;
2171 struct request *last = NULL;
2174 trace_block_plug(q);
2176 last = list_entry_rq(plug->mq_list.prev);
2178 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2179 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2180 blk_flush_plug_list(plug, false);
2181 trace_block_plug(q);
2184 blk_add_rq_to_plug(plug, rq);
2185 } else if (q->elevator) {
2186 /* Insert the request at the IO scheduler queue */
2187 blk_mq_sched_insert_request(rq, false, true, true);
2188 } else if (plug && !blk_queue_nomerges(q)) {
2190 * We do limited plugging. If the bio can be merged, do that.
2191 * Otherwise the existing request in the plug list will be
2192 * issued. So the plug list will have one request at most
2193 * The plug list might get flushed before this. If that happens,
2194 * the plug list is empty, and same_queue_rq is invalid.
2196 if (list_empty(&plug->mq_list))
2197 same_queue_rq = NULL;
2198 if (same_queue_rq) {
2199 list_del_init(&same_queue_rq->queuelist);
2202 blk_add_rq_to_plug(plug, rq);
2203 trace_block_plug(q);
2205 if (same_queue_rq) {
2206 data.hctx = same_queue_rq->mq_hctx;
2207 trace_block_unplug(q, 1, true);
2208 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2211 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2212 !data.hctx->dispatch_busy) {
2214 * There is no scheduler and we can try to send directly
2217 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2220 blk_mq_sched_insert_request(rq, false, true, true);
2226 return BLK_QC_T_NONE;
2228 EXPORT_SYMBOL_GPL(blk_mq_make_request); /* only for request based dm */
2230 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2231 unsigned int hctx_idx)
2235 if (tags->rqs && set->ops->exit_request) {
2238 for (i = 0; i < tags->nr_tags; i++) {
2239 struct request *rq = tags->static_rqs[i];
2243 set->ops->exit_request(set, rq, hctx_idx);
2244 tags->static_rqs[i] = NULL;
2248 while (!list_empty(&tags->page_list)) {
2249 page = list_first_entry(&tags->page_list, struct page, lru);
2250 list_del_init(&page->lru);
2252 * Remove kmemleak object previously allocated in
2253 * blk_mq_alloc_rqs().
2255 kmemleak_free(page_address(page));
2256 __free_pages(page, page->private);
2260 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2264 kfree(tags->static_rqs);
2265 tags->static_rqs = NULL;
2267 blk_mq_free_tags(tags);
2270 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2271 unsigned int hctx_idx,
2272 unsigned int nr_tags,
2273 unsigned int reserved_tags)
2275 struct blk_mq_tags *tags;
2278 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2279 if (node == NUMA_NO_NODE)
2280 node = set->numa_node;
2282 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2283 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2287 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2288 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2291 blk_mq_free_tags(tags);
2295 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2296 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2298 if (!tags->static_rqs) {
2300 blk_mq_free_tags(tags);
2307 static size_t order_to_size(unsigned int order)
2309 return (size_t)PAGE_SIZE << order;
2312 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2313 unsigned int hctx_idx, int node)
2317 if (set->ops->init_request) {
2318 ret = set->ops->init_request(set, rq, hctx_idx, node);
2323 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2327 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2328 unsigned int hctx_idx, unsigned int depth)
2330 unsigned int i, j, entries_per_page, max_order = 4;
2331 size_t rq_size, left;
2334 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2335 if (node == NUMA_NO_NODE)
2336 node = set->numa_node;
2338 INIT_LIST_HEAD(&tags->page_list);
2341 * rq_size is the size of the request plus driver payload, rounded
2342 * to the cacheline size
2344 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2346 left = rq_size * depth;
2348 for (i = 0; i < depth; ) {
2349 int this_order = max_order;
2354 while (this_order && left < order_to_size(this_order - 1))
2358 page = alloc_pages_node(node,
2359 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2365 if (order_to_size(this_order) < rq_size)
2372 page->private = this_order;
2373 list_add_tail(&page->lru, &tags->page_list);
2375 p = page_address(page);
2377 * Allow kmemleak to scan these pages as they contain pointers
2378 * to additional allocations like via ops->init_request().
2380 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2381 entries_per_page = order_to_size(this_order) / rq_size;
2382 to_do = min(entries_per_page, depth - i);
2383 left -= to_do * rq_size;
2384 for (j = 0; j < to_do; j++) {
2385 struct request *rq = p;
2387 tags->static_rqs[i] = rq;
2388 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2389 tags->static_rqs[i] = NULL;
2400 blk_mq_free_rqs(set, tags, hctx_idx);
2404 struct rq_iter_data {
2405 struct blk_mq_hw_ctx *hctx;
2409 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2411 struct rq_iter_data *iter_data = data;
2413 if (rq->mq_hctx != iter_data->hctx)
2415 iter_data->has_rq = true;
2419 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2421 struct blk_mq_tags *tags = hctx->sched_tags ?
2422 hctx->sched_tags : hctx->tags;
2423 struct rq_iter_data data = {
2427 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2431 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2432 struct blk_mq_hw_ctx *hctx)
2434 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2436 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2441 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2443 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2444 struct blk_mq_hw_ctx, cpuhp_online);
2446 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2447 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2451 * Prevent new request from being allocated on the current hctx.
2453 * The smp_mb__after_atomic() Pairs with the implied barrier in
2454 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2455 * seen once we return from the tag allocator.
2457 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2458 smp_mb__after_atomic();
2461 * Try to grab a reference to the queue and wait for any outstanding
2462 * requests. If we could not grab a reference the queue has been
2463 * frozen and there are no requests.
2465 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2466 while (blk_mq_hctx_has_requests(hctx))
2468 percpu_ref_put(&hctx->queue->q_usage_counter);
2474 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2476 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2477 struct blk_mq_hw_ctx, cpuhp_online);
2479 if (cpumask_test_cpu(cpu, hctx->cpumask))
2480 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2485 * 'cpu' is going away. splice any existing rq_list entries from this
2486 * software queue to the hw queue dispatch list, and ensure that it
2489 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2491 struct blk_mq_hw_ctx *hctx;
2492 struct blk_mq_ctx *ctx;
2494 enum hctx_type type;
2496 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2497 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2500 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2503 spin_lock(&ctx->lock);
2504 if (!list_empty(&ctx->rq_lists[type])) {
2505 list_splice_init(&ctx->rq_lists[type], &tmp);
2506 blk_mq_hctx_clear_pending(hctx, ctx);
2508 spin_unlock(&ctx->lock);
2510 if (list_empty(&tmp))
2513 spin_lock(&hctx->lock);
2514 list_splice_tail_init(&tmp, &hctx->dispatch);
2515 spin_unlock(&hctx->lock);
2517 blk_mq_run_hw_queue(hctx, true);
2521 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2523 if (!(hctx->flags & BLK_MQ_F_STACKING))
2524 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2525 &hctx->cpuhp_online);
2526 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2530 /* hctx->ctxs will be freed in queue's release handler */
2531 static void blk_mq_exit_hctx(struct request_queue *q,
2532 struct blk_mq_tag_set *set,
2533 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2535 if (blk_mq_hw_queue_mapped(hctx))
2536 blk_mq_tag_idle(hctx);
2538 if (set->ops->exit_request)
2539 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2541 if (set->ops->exit_hctx)
2542 set->ops->exit_hctx(hctx, hctx_idx);
2544 blk_mq_remove_cpuhp(hctx);
2546 spin_lock(&q->unused_hctx_lock);
2547 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2548 spin_unlock(&q->unused_hctx_lock);
2551 static void blk_mq_exit_hw_queues(struct request_queue *q,
2552 struct blk_mq_tag_set *set, int nr_queue)
2554 struct blk_mq_hw_ctx *hctx;
2557 queue_for_each_hw_ctx(q, hctx, i) {
2560 blk_mq_debugfs_unregister_hctx(hctx);
2561 blk_mq_exit_hctx(q, set, hctx, i);
2565 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2567 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2569 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2570 __alignof__(struct blk_mq_hw_ctx)) !=
2571 sizeof(struct blk_mq_hw_ctx));
2573 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2574 hw_ctx_size += sizeof(struct srcu_struct);
2579 static int blk_mq_init_hctx(struct request_queue *q,
2580 struct blk_mq_tag_set *set,
2581 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2583 hctx->queue_num = hctx_idx;
2585 if (!(hctx->flags & BLK_MQ_F_STACKING))
2586 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2587 &hctx->cpuhp_online);
2588 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2590 hctx->tags = set->tags[hctx_idx];
2592 if (set->ops->init_hctx &&
2593 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2594 goto unregister_cpu_notifier;
2596 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2602 if (set->ops->exit_hctx)
2603 set->ops->exit_hctx(hctx, hctx_idx);
2604 unregister_cpu_notifier:
2605 blk_mq_remove_cpuhp(hctx);
2609 static struct blk_mq_hw_ctx *
2610 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2613 struct blk_mq_hw_ctx *hctx;
2614 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2616 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2618 goto fail_alloc_hctx;
2620 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2623 atomic_set(&hctx->nr_active, 0);
2624 if (node == NUMA_NO_NODE)
2625 node = set->numa_node;
2626 hctx->numa_node = node;
2628 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2629 spin_lock_init(&hctx->lock);
2630 INIT_LIST_HEAD(&hctx->dispatch);
2632 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2634 INIT_LIST_HEAD(&hctx->hctx_list);
2637 * Allocate space for all possible cpus to avoid allocation at
2640 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2645 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2650 spin_lock_init(&hctx->dispatch_wait_lock);
2651 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2652 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2654 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2658 if (hctx->flags & BLK_MQ_F_BLOCKING)
2659 init_srcu_struct(hctx->srcu);
2660 blk_mq_hctx_kobj_init(hctx);
2665 sbitmap_free(&hctx->ctx_map);
2669 free_cpumask_var(hctx->cpumask);
2676 static void blk_mq_init_cpu_queues(struct request_queue *q,
2677 unsigned int nr_hw_queues)
2679 struct blk_mq_tag_set *set = q->tag_set;
2682 for_each_possible_cpu(i) {
2683 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2684 struct blk_mq_hw_ctx *hctx;
2688 spin_lock_init(&__ctx->lock);
2689 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2690 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2695 * Set local node, IFF we have more than one hw queue. If
2696 * not, we remain on the home node of the device
2698 for (j = 0; j < set->nr_maps; j++) {
2699 hctx = blk_mq_map_queue_type(q, j, i);
2700 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2701 hctx->numa_node = local_memory_node(cpu_to_node(i));
2706 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2711 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2712 set->queue_depth, set->reserved_tags);
2713 if (!set->tags[hctx_idx])
2716 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2721 blk_mq_free_rq_map(set->tags[hctx_idx]);
2722 set->tags[hctx_idx] = NULL;
2726 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2727 unsigned int hctx_idx)
2729 if (set->tags && set->tags[hctx_idx]) {
2730 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2731 blk_mq_free_rq_map(set->tags[hctx_idx]);
2732 set->tags[hctx_idx] = NULL;
2736 static void blk_mq_map_swqueue(struct request_queue *q)
2738 unsigned int i, j, hctx_idx;
2739 struct blk_mq_hw_ctx *hctx;
2740 struct blk_mq_ctx *ctx;
2741 struct blk_mq_tag_set *set = q->tag_set;
2743 queue_for_each_hw_ctx(q, hctx, i) {
2744 cpumask_clear(hctx->cpumask);
2746 hctx->dispatch_from = NULL;
2750 * Map software to hardware queues.
2752 * If the cpu isn't present, the cpu is mapped to first hctx.
2754 for_each_possible_cpu(i) {
2756 ctx = per_cpu_ptr(q->queue_ctx, i);
2757 for (j = 0; j < set->nr_maps; j++) {
2758 if (!set->map[j].nr_queues) {
2759 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2760 HCTX_TYPE_DEFAULT, i);
2763 hctx_idx = set->map[j].mq_map[i];
2764 /* unmapped hw queue can be remapped after CPU topo changed */
2765 if (!set->tags[hctx_idx] &&
2766 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2768 * If tags initialization fail for some hctx,
2769 * that hctx won't be brought online. In this
2770 * case, remap the current ctx to hctx[0] which
2771 * is guaranteed to always have tags allocated
2773 set->map[j].mq_map[i] = 0;
2776 hctx = blk_mq_map_queue_type(q, j, i);
2777 ctx->hctxs[j] = hctx;
2779 * If the CPU is already set in the mask, then we've
2780 * mapped this one already. This can happen if
2781 * devices share queues across queue maps.
2783 if (cpumask_test_cpu(i, hctx->cpumask))
2786 cpumask_set_cpu(i, hctx->cpumask);
2788 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2789 hctx->ctxs[hctx->nr_ctx++] = ctx;
2792 * If the nr_ctx type overflows, we have exceeded the
2793 * amount of sw queues we can support.
2795 BUG_ON(!hctx->nr_ctx);
2798 for (; j < HCTX_MAX_TYPES; j++)
2799 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2800 HCTX_TYPE_DEFAULT, i);
2803 queue_for_each_hw_ctx(q, hctx, i) {
2805 * If no software queues are mapped to this hardware queue,
2806 * disable it and free the request entries.
2808 if (!hctx->nr_ctx) {
2809 /* Never unmap queue 0. We need it as a
2810 * fallback in case of a new remap fails
2813 if (i && set->tags[i])
2814 blk_mq_free_map_and_requests(set, i);
2820 hctx->tags = set->tags[i];
2821 WARN_ON(!hctx->tags);
2824 * Set the map size to the number of mapped software queues.
2825 * This is more accurate and more efficient than looping
2826 * over all possibly mapped software queues.
2828 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2831 * Initialize batch roundrobin counts
2833 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2834 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2839 * Caller needs to ensure that we're either frozen/quiesced, or that
2840 * the queue isn't live yet.
2842 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2844 struct blk_mq_hw_ctx *hctx;
2847 queue_for_each_hw_ctx(q, hctx, i) {
2849 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2851 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2855 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2858 struct request_queue *q;
2860 lockdep_assert_held(&set->tag_list_lock);
2862 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2863 blk_mq_freeze_queue(q);
2864 queue_set_hctx_shared(q, shared);
2865 blk_mq_unfreeze_queue(q);
2869 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2871 struct blk_mq_tag_set *set = q->tag_set;
2873 mutex_lock(&set->tag_list_lock);
2874 list_del_rcu(&q->tag_set_list);
2875 if (list_is_singular(&set->tag_list)) {
2876 /* just transitioned to unshared */
2877 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2878 /* update existing queue */
2879 blk_mq_update_tag_set_depth(set, false);
2881 mutex_unlock(&set->tag_list_lock);
2882 INIT_LIST_HEAD(&q->tag_set_list);
2885 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2886 struct request_queue *q)
2888 mutex_lock(&set->tag_list_lock);
2891 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2893 if (!list_empty(&set->tag_list) &&
2894 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2895 set->flags |= BLK_MQ_F_TAG_SHARED;
2896 /* update existing queue */
2897 blk_mq_update_tag_set_depth(set, true);
2899 if (set->flags & BLK_MQ_F_TAG_SHARED)
2900 queue_set_hctx_shared(q, true);
2901 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2903 mutex_unlock(&set->tag_list_lock);
2906 /* All allocations will be freed in release handler of q->mq_kobj */
2907 static int blk_mq_alloc_ctxs(struct request_queue *q)
2909 struct blk_mq_ctxs *ctxs;
2912 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2916 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2917 if (!ctxs->queue_ctx)
2920 for_each_possible_cpu(cpu) {
2921 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2925 q->mq_kobj = &ctxs->kobj;
2926 q->queue_ctx = ctxs->queue_ctx;
2935 * It is the actual release handler for mq, but we do it from
2936 * request queue's release handler for avoiding use-after-free
2937 * and headache because q->mq_kobj shouldn't have been introduced,
2938 * but we can't group ctx/kctx kobj without it.
2940 void blk_mq_release(struct request_queue *q)
2942 struct blk_mq_hw_ctx *hctx, *next;
2945 queue_for_each_hw_ctx(q, hctx, i)
2946 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2948 /* all hctx are in .unused_hctx_list now */
2949 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2950 list_del_init(&hctx->hctx_list);
2951 kobject_put(&hctx->kobj);
2954 kfree(q->queue_hw_ctx);
2957 * release .mq_kobj and sw queue's kobject now because
2958 * both share lifetime with request queue.
2960 blk_mq_sysfs_deinit(q);
2963 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
2966 struct request_queue *uninit_q, *q;
2968 uninit_q = __blk_alloc_queue(set->numa_node);
2970 return ERR_PTR(-ENOMEM);
2971 uninit_q->queuedata = queuedata;
2974 * Initialize the queue without an elevator. device_add_disk() will do
2975 * the initialization.
2977 q = blk_mq_init_allocated_queue(set, uninit_q, false);
2979 blk_cleanup_queue(uninit_q);
2983 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
2985 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2987 return blk_mq_init_queue_data(set, NULL);
2989 EXPORT_SYMBOL(blk_mq_init_queue);
2992 * Helper for setting up a queue with mq ops, given queue depth, and
2993 * the passed in mq ops flags.
2995 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2996 const struct blk_mq_ops *ops,
2997 unsigned int queue_depth,
2998 unsigned int set_flags)
3000 struct request_queue *q;
3003 memset(set, 0, sizeof(*set));
3005 set->nr_hw_queues = 1;
3007 set->queue_depth = queue_depth;
3008 set->numa_node = NUMA_NO_NODE;
3009 set->flags = set_flags;
3011 ret = blk_mq_alloc_tag_set(set);
3013 return ERR_PTR(ret);
3015 q = blk_mq_init_queue(set);
3017 blk_mq_free_tag_set(set);
3023 EXPORT_SYMBOL(blk_mq_init_sq_queue);
3025 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3026 struct blk_mq_tag_set *set, struct request_queue *q,
3027 int hctx_idx, int node)
3029 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3031 /* reuse dead hctx first */
3032 spin_lock(&q->unused_hctx_lock);
3033 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3034 if (tmp->numa_node == node) {
3040 list_del_init(&hctx->hctx_list);
3041 spin_unlock(&q->unused_hctx_lock);
3044 hctx = blk_mq_alloc_hctx(q, set, node);
3048 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3054 kobject_put(&hctx->kobj);
3059 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3060 struct request_queue *q)
3063 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3065 if (q->nr_hw_queues < set->nr_hw_queues) {
3066 struct blk_mq_hw_ctx **new_hctxs;
3068 new_hctxs = kcalloc_node(set->nr_hw_queues,
3069 sizeof(*new_hctxs), GFP_KERNEL,
3074 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3076 q->queue_hw_ctx = new_hctxs;
3081 /* protect against switching io scheduler */
3082 mutex_lock(&q->sysfs_lock);
3083 for (i = 0; i < set->nr_hw_queues; i++) {
3085 struct blk_mq_hw_ctx *hctx;
3087 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3089 * If the hw queue has been mapped to another numa node,
3090 * we need to realloc the hctx. If allocation fails, fallback
3091 * to use the previous one.
3093 if (hctxs[i] && (hctxs[i]->numa_node == node))
3096 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3099 blk_mq_exit_hctx(q, set, hctxs[i], i);
3103 pr_warn("Allocate new hctx on node %d fails,\
3104 fallback to previous one on node %d\n",
3105 node, hctxs[i]->numa_node);
3111 * Increasing nr_hw_queues fails. Free the newly allocated
3112 * hctxs and keep the previous q->nr_hw_queues.
3114 if (i != set->nr_hw_queues) {
3115 j = q->nr_hw_queues;
3119 end = q->nr_hw_queues;
3120 q->nr_hw_queues = set->nr_hw_queues;
3123 for (; j < end; j++) {
3124 struct blk_mq_hw_ctx *hctx = hctxs[j];
3128 blk_mq_free_map_and_requests(set, j);
3129 blk_mq_exit_hctx(q, set, hctx, j);
3133 mutex_unlock(&q->sysfs_lock);
3136 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3137 struct request_queue *q,
3140 /* mark the queue as mq asap */
3141 q->mq_ops = set->ops;
3143 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3144 blk_mq_poll_stats_bkt,
3145 BLK_MQ_POLL_STATS_BKTS, q);
3149 if (blk_mq_alloc_ctxs(q))
3152 /* init q->mq_kobj and sw queues' kobjects */
3153 blk_mq_sysfs_init(q);
3155 INIT_LIST_HEAD(&q->unused_hctx_list);
3156 spin_lock_init(&q->unused_hctx_lock);
3158 blk_mq_realloc_hw_ctxs(set, q);
3159 if (!q->nr_hw_queues)
3162 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3163 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3167 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3168 if (set->nr_maps > HCTX_TYPE_POLL &&
3169 set->map[HCTX_TYPE_POLL].nr_queues)
3170 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3172 q->sg_reserved_size = INT_MAX;
3174 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3175 INIT_LIST_HEAD(&q->requeue_list);
3176 spin_lock_init(&q->requeue_lock);
3178 q->nr_requests = set->queue_depth;
3181 * Default to classic polling
3183 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3185 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3186 blk_mq_add_queue_tag_set(set, q);
3187 blk_mq_map_swqueue(q);
3190 elevator_init_mq(q);
3195 kfree(q->queue_hw_ctx);
3196 q->nr_hw_queues = 0;
3197 blk_mq_sysfs_deinit(q);
3199 blk_stat_free_callback(q->poll_cb);
3203 return ERR_PTR(-ENOMEM);
3205 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3207 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3208 void blk_mq_exit_queue(struct request_queue *q)
3210 struct blk_mq_tag_set *set = q->tag_set;
3212 blk_mq_del_queue_tag_set(q);
3213 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3216 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3220 for (i = 0; i < set->nr_hw_queues; i++)
3221 if (!__blk_mq_alloc_map_and_request(set, i))
3228 blk_mq_free_map_and_requests(set, i);
3234 * Allocate the request maps associated with this tag_set. Note that this
3235 * may reduce the depth asked for, if memory is tight. set->queue_depth
3236 * will be updated to reflect the allocated depth.
3238 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3243 depth = set->queue_depth;
3245 err = __blk_mq_alloc_rq_maps(set);
3249 set->queue_depth >>= 1;
3250 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3254 } while (set->queue_depth);
3256 if (!set->queue_depth || err) {
3257 pr_err("blk-mq: failed to allocate request map\n");
3261 if (depth != set->queue_depth)
3262 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3263 depth, set->queue_depth);
3268 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3271 * blk_mq_map_queues() and multiple .map_queues() implementations
3272 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3273 * number of hardware queues.
3275 if (set->nr_maps == 1)
3276 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3278 if (set->ops->map_queues && !is_kdump_kernel()) {
3282 * transport .map_queues is usually done in the following
3285 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3286 * mask = get_cpu_mask(queue)
3287 * for_each_cpu(cpu, mask)
3288 * set->map[x].mq_map[cpu] = queue;
3291 * When we need to remap, the table has to be cleared for
3292 * killing stale mapping since one CPU may not be mapped
3295 for (i = 0; i < set->nr_maps; i++)
3296 blk_mq_clear_mq_map(&set->map[i]);
3298 return set->ops->map_queues(set);
3300 BUG_ON(set->nr_maps > 1);
3301 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3305 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3306 int cur_nr_hw_queues, int new_nr_hw_queues)
3308 struct blk_mq_tags **new_tags;
3310 if (cur_nr_hw_queues >= new_nr_hw_queues)
3313 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3314 GFP_KERNEL, set->numa_node);
3319 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3320 sizeof(*set->tags));
3322 set->tags = new_tags;
3323 set->nr_hw_queues = new_nr_hw_queues;
3329 * Alloc a tag set to be associated with one or more request queues.
3330 * May fail with EINVAL for various error conditions. May adjust the
3331 * requested depth down, if it's too large. In that case, the set
3332 * value will be stored in set->queue_depth.
3334 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3338 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3340 if (!set->nr_hw_queues)
3342 if (!set->queue_depth)
3344 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3347 if (!set->ops->queue_rq)
3350 if (!set->ops->get_budget ^ !set->ops->put_budget)
3353 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3354 pr_info("blk-mq: reduced tag depth to %u\n",
3356 set->queue_depth = BLK_MQ_MAX_DEPTH;
3361 else if (set->nr_maps > HCTX_MAX_TYPES)
3365 * If a crashdump is active, then we are potentially in a very
3366 * memory constrained environment. Limit us to 1 queue and
3367 * 64 tags to prevent using too much memory.
3369 if (is_kdump_kernel()) {
3370 set->nr_hw_queues = 1;
3372 set->queue_depth = min(64U, set->queue_depth);
3375 * There is no use for more h/w queues than cpus if we just have
3378 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3379 set->nr_hw_queues = nr_cpu_ids;
3381 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3385 for (i = 0; i < set->nr_maps; i++) {
3386 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3387 sizeof(set->map[i].mq_map[0]),
3388 GFP_KERNEL, set->numa_node);
3389 if (!set->map[i].mq_map)
3390 goto out_free_mq_map;
3391 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3394 ret = blk_mq_update_queue_map(set);
3396 goto out_free_mq_map;
3398 ret = blk_mq_alloc_map_and_requests(set);
3400 goto out_free_mq_map;
3402 mutex_init(&set->tag_list_lock);
3403 INIT_LIST_HEAD(&set->tag_list);
3408 for (i = 0; i < set->nr_maps; i++) {
3409 kfree(set->map[i].mq_map);
3410 set->map[i].mq_map = NULL;
3416 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3418 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3422 for (i = 0; i < set->nr_hw_queues; i++)
3423 blk_mq_free_map_and_requests(set, i);
3425 for (j = 0; j < set->nr_maps; j++) {
3426 kfree(set->map[j].mq_map);
3427 set->map[j].mq_map = NULL;
3433 EXPORT_SYMBOL(blk_mq_free_tag_set);
3435 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3437 struct blk_mq_tag_set *set = q->tag_set;
3438 struct blk_mq_hw_ctx *hctx;
3444 if (q->nr_requests == nr)
3447 blk_mq_freeze_queue(q);
3448 blk_mq_quiesce_queue(q);
3451 queue_for_each_hw_ctx(q, hctx, i) {
3455 * If we're using an MQ scheduler, just update the scheduler
3456 * queue depth. This is similar to what the old code would do.
3458 if (!hctx->sched_tags) {
3459 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3462 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3467 if (q->elevator && q->elevator->type->ops.depth_updated)
3468 q->elevator->type->ops.depth_updated(hctx);
3472 q->nr_requests = nr;
3474 blk_mq_unquiesce_queue(q);
3475 blk_mq_unfreeze_queue(q);
3481 * request_queue and elevator_type pair.
3482 * It is just used by __blk_mq_update_nr_hw_queues to cache
3483 * the elevator_type associated with a request_queue.
3485 struct blk_mq_qe_pair {
3486 struct list_head node;
3487 struct request_queue *q;
3488 struct elevator_type *type;
3492 * Cache the elevator_type in qe pair list and switch the
3493 * io scheduler to 'none'
3495 static bool blk_mq_elv_switch_none(struct list_head *head,
3496 struct request_queue *q)
3498 struct blk_mq_qe_pair *qe;
3503 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3507 INIT_LIST_HEAD(&qe->node);
3509 qe->type = q->elevator->type;
3510 list_add(&qe->node, head);
3512 mutex_lock(&q->sysfs_lock);
3514 * After elevator_switch_mq, the previous elevator_queue will be
3515 * released by elevator_release. The reference of the io scheduler
3516 * module get by elevator_get will also be put. So we need to get
3517 * a reference of the io scheduler module here to prevent it to be
3520 __module_get(qe->type->elevator_owner);
3521 elevator_switch_mq(q, NULL);
3522 mutex_unlock(&q->sysfs_lock);
3527 static void blk_mq_elv_switch_back(struct list_head *head,
3528 struct request_queue *q)
3530 struct blk_mq_qe_pair *qe;
3531 struct elevator_type *t = NULL;
3533 list_for_each_entry(qe, head, node)
3542 list_del(&qe->node);
3545 mutex_lock(&q->sysfs_lock);
3546 elevator_switch_mq(q, t);
3547 mutex_unlock(&q->sysfs_lock);
3550 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3553 struct request_queue *q;
3555 int prev_nr_hw_queues;
3557 lockdep_assert_held(&set->tag_list_lock);
3559 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3560 nr_hw_queues = nr_cpu_ids;
3561 if (nr_hw_queues < 1)
3563 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3566 list_for_each_entry(q, &set->tag_list, tag_set_list)
3567 blk_mq_freeze_queue(q);
3569 * Switch IO scheduler to 'none', cleaning up the data associated
3570 * with the previous scheduler. We will switch back once we are done
3571 * updating the new sw to hw queue mappings.
3573 list_for_each_entry(q, &set->tag_list, tag_set_list)
3574 if (!blk_mq_elv_switch_none(&head, q))
3577 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3578 blk_mq_debugfs_unregister_hctxs(q);
3579 blk_mq_sysfs_unregister(q);
3582 prev_nr_hw_queues = set->nr_hw_queues;
3583 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3587 set->nr_hw_queues = nr_hw_queues;
3589 blk_mq_update_queue_map(set);
3590 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3591 blk_mq_realloc_hw_ctxs(set, q);
3592 if (q->nr_hw_queues != set->nr_hw_queues) {
3593 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3594 nr_hw_queues, prev_nr_hw_queues);
3595 set->nr_hw_queues = prev_nr_hw_queues;
3596 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3599 blk_mq_map_swqueue(q);
3603 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3604 blk_mq_sysfs_register(q);
3605 blk_mq_debugfs_register_hctxs(q);
3609 list_for_each_entry(q, &set->tag_list, tag_set_list)
3610 blk_mq_elv_switch_back(&head, q);
3612 list_for_each_entry(q, &set->tag_list, tag_set_list)
3613 blk_mq_unfreeze_queue(q);
3616 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3618 mutex_lock(&set->tag_list_lock);
3619 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3620 mutex_unlock(&set->tag_list_lock);
3622 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3624 /* Enable polling stats and return whether they were already enabled. */
3625 static bool blk_poll_stats_enable(struct request_queue *q)
3627 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3628 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3630 blk_stat_add_callback(q, q->poll_cb);
3634 static void blk_mq_poll_stats_start(struct request_queue *q)
3637 * We don't arm the callback if polling stats are not enabled or the
3638 * callback is already active.
3640 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3641 blk_stat_is_active(q->poll_cb))
3644 blk_stat_activate_msecs(q->poll_cb, 100);
3647 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3649 struct request_queue *q = cb->data;
3652 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3653 if (cb->stat[bucket].nr_samples)
3654 q->poll_stat[bucket] = cb->stat[bucket];
3658 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3661 unsigned long ret = 0;
3665 * If stats collection isn't on, don't sleep but turn it on for
3668 if (!blk_poll_stats_enable(q))
3672 * As an optimistic guess, use half of the mean service time
3673 * for this type of request. We can (and should) make this smarter.
3674 * For instance, if the completion latencies are tight, we can
3675 * get closer than just half the mean. This is especially
3676 * important on devices where the completion latencies are longer
3677 * than ~10 usec. We do use the stats for the relevant IO size
3678 * if available which does lead to better estimates.
3680 bucket = blk_mq_poll_stats_bkt(rq);
3684 if (q->poll_stat[bucket].nr_samples)
3685 ret = (q->poll_stat[bucket].mean + 1) / 2;
3690 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3693 struct hrtimer_sleeper hs;
3694 enum hrtimer_mode mode;
3698 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3702 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3704 * 0: use half of prev avg
3705 * >0: use this specific value
3707 if (q->poll_nsec > 0)
3708 nsecs = q->poll_nsec;
3710 nsecs = blk_mq_poll_nsecs(q, rq);
3715 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3718 * This will be replaced with the stats tracking code, using
3719 * 'avg_completion_time / 2' as the pre-sleep target.
3723 mode = HRTIMER_MODE_REL;
3724 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3725 hrtimer_set_expires(&hs.timer, kt);
3728 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3730 set_current_state(TASK_UNINTERRUPTIBLE);
3731 hrtimer_sleeper_start_expires(&hs, mode);
3734 hrtimer_cancel(&hs.timer);
3735 mode = HRTIMER_MODE_ABS;
3736 } while (hs.task && !signal_pending(current));
3738 __set_current_state(TASK_RUNNING);
3739 destroy_hrtimer_on_stack(&hs.timer);
3743 static bool blk_mq_poll_hybrid(struct request_queue *q,
3744 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3748 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3751 if (!blk_qc_t_is_internal(cookie))
3752 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3754 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3756 * With scheduling, if the request has completed, we'll
3757 * get a NULL return here, as we clear the sched tag when
3758 * that happens. The request still remains valid, like always,
3759 * so we should be safe with just the NULL check.
3765 return blk_mq_poll_hybrid_sleep(q, rq);
3769 * blk_poll - poll for IO completions
3771 * @cookie: cookie passed back at IO submission time
3772 * @spin: whether to spin for completions
3775 * Poll for completions on the passed in queue. Returns number of
3776 * completed entries found. If @spin is true, then blk_poll will continue
3777 * looping until at least one completion is found, unless the task is
3778 * otherwise marked running (or we need to reschedule).
3780 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3782 struct blk_mq_hw_ctx *hctx;
3785 if (!blk_qc_t_valid(cookie) ||
3786 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3790 blk_flush_plug_list(current->plug, false);
3792 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3795 * If we sleep, have the caller restart the poll loop to reset
3796 * the state. Like for the other success return cases, the
3797 * caller is responsible for checking if the IO completed. If
3798 * the IO isn't complete, we'll get called again and will go
3799 * straight to the busy poll loop.
3801 if (blk_mq_poll_hybrid(q, hctx, cookie))
3804 hctx->poll_considered++;
3806 state = current->state;
3810 hctx->poll_invoked++;
3812 ret = q->mq_ops->poll(hctx);
3814 hctx->poll_success++;
3815 __set_current_state(TASK_RUNNING);
3819 if (signal_pending_state(state, current))
3820 __set_current_state(TASK_RUNNING);
3822 if (current->state == TASK_RUNNING)
3824 if (ret < 0 || !spin)
3827 } while (!need_resched());
3829 __set_current_state(TASK_RUNNING);
3832 EXPORT_SYMBOL_GPL(blk_poll);
3834 unsigned int blk_mq_rq_cpu(struct request *rq)
3836 return rq->mq_ctx->cpu;
3838 EXPORT_SYMBOL(blk_mq_rq_cpu);
3840 static int __init blk_mq_init(void)
3844 for_each_possible_cpu(i)
3845 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3846 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3848 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3849 "block/softirq:dead", NULL,
3850 blk_softirq_cpu_dead);
3851 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3852 blk_mq_hctx_notify_dead);
3853 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3854 blk_mq_hctx_notify_online,
3855 blk_mq_hctx_notify_offline);
3858 subsys_initcall(blk_mq_init);