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 bool __blk_mq_get_driver_tag(struct request *rq)
1108 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1109 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1110 bool shared = blk_mq_tag_busy(rq->mq_hctx);
1113 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1114 bt = &rq->mq_hctx->tags->breserved_tags;
1118 if (!hctx_may_queue(rq->mq_hctx, bt))
1120 tag = __sbitmap_queue_get(bt);
1121 if (tag == BLK_MQ_NO_TAG)
1124 rq->tag = tag + tag_offset;
1126 rq->rq_flags |= RQF_MQ_INFLIGHT;
1127 atomic_inc(&rq->mq_hctx->nr_active);
1129 rq->mq_hctx->tags->rqs[rq->tag] = rq;
1133 static bool blk_mq_get_driver_tag(struct request *rq)
1135 if (rq->tag != BLK_MQ_NO_TAG)
1137 return __blk_mq_get_driver_tag(rq);
1140 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1141 int flags, void *key)
1143 struct blk_mq_hw_ctx *hctx;
1145 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1147 spin_lock(&hctx->dispatch_wait_lock);
1148 if (!list_empty(&wait->entry)) {
1149 struct sbitmap_queue *sbq;
1151 list_del_init(&wait->entry);
1152 sbq = &hctx->tags->bitmap_tags;
1153 atomic_dec(&sbq->ws_active);
1155 spin_unlock(&hctx->dispatch_wait_lock);
1157 blk_mq_run_hw_queue(hctx, true);
1162 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1163 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1164 * restart. For both cases, take care to check the condition again after
1165 * marking us as waiting.
1167 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1170 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1171 struct wait_queue_head *wq;
1172 wait_queue_entry_t *wait;
1175 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1176 blk_mq_sched_mark_restart_hctx(hctx);
1179 * It's possible that a tag was freed in the window between the
1180 * allocation failure and adding the hardware queue to the wait
1183 * Don't clear RESTART here, someone else could have set it.
1184 * At most this will cost an extra queue run.
1186 return blk_mq_get_driver_tag(rq);
1189 wait = &hctx->dispatch_wait;
1190 if (!list_empty_careful(&wait->entry))
1193 wq = &bt_wait_ptr(sbq, hctx)->wait;
1195 spin_lock_irq(&wq->lock);
1196 spin_lock(&hctx->dispatch_wait_lock);
1197 if (!list_empty(&wait->entry)) {
1198 spin_unlock(&hctx->dispatch_wait_lock);
1199 spin_unlock_irq(&wq->lock);
1203 atomic_inc(&sbq->ws_active);
1204 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1205 __add_wait_queue(wq, wait);
1208 * It's possible that a tag was freed in the window between the
1209 * allocation failure and adding the hardware queue to the wait
1212 ret = blk_mq_get_driver_tag(rq);
1214 spin_unlock(&hctx->dispatch_wait_lock);
1215 spin_unlock_irq(&wq->lock);
1220 * We got a tag, remove ourselves from the wait queue to ensure
1221 * someone else gets the wakeup.
1223 list_del_init(&wait->entry);
1224 atomic_dec(&sbq->ws_active);
1225 spin_unlock(&hctx->dispatch_wait_lock);
1226 spin_unlock_irq(&wq->lock);
1231 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1232 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1234 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1235 * - EWMA is one simple way to compute running average value
1236 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1237 * - take 4 as factor for avoiding to get too small(0) result, and this
1238 * factor doesn't matter because EWMA decreases exponentially
1240 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1244 if (hctx->queue->elevator)
1247 ewma = hctx->dispatch_busy;
1252 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1254 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1255 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1257 hctx->dispatch_busy = ewma;
1260 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1262 static void blk_mq_handle_dev_resource(struct request *rq,
1263 struct list_head *list)
1265 struct request *next =
1266 list_first_entry_or_null(list, struct request, queuelist);
1269 * If an I/O scheduler has been configured and we got a driver tag for
1270 * the next request already, free it.
1273 blk_mq_put_driver_tag(next);
1275 list_add(&rq->queuelist, list);
1276 __blk_mq_requeue_request(rq);
1279 static void blk_mq_handle_zone_resource(struct request *rq,
1280 struct list_head *zone_list)
1283 * If we end up here it is because we cannot dispatch a request to a
1284 * specific zone due to LLD level zone-write locking or other zone
1285 * related resource not being available. In this case, set the request
1286 * aside in zone_list for retrying it later.
1288 list_add(&rq->queuelist, zone_list);
1289 __blk_mq_requeue_request(rq);
1292 enum prep_dispatch {
1294 PREP_DISPATCH_NO_TAG,
1295 PREP_DISPATCH_NO_BUDGET,
1298 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1301 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1303 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1304 blk_mq_put_driver_tag(rq);
1305 return PREP_DISPATCH_NO_BUDGET;
1308 if (!blk_mq_get_driver_tag(rq)) {
1310 * The initial allocation attempt failed, so we need to
1311 * rerun the hardware queue when a tag is freed. The
1312 * waitqueue takes care of that. If the queue is run
1313 * before we add this entry back on the dispatch list,
1314 * we'll re-run it below.
1316 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1318 * All budgets not got from this function will be put
1319 * together during handling partial dispatch
1322 blk_mq_put_dispatch_budget(rq->q);
1323 return PREP_DISPATCH_NO_TAG;
1327 return PREP_DISPATCH_OK;
1330 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1331 static void blk_mq_release_budgets(struct request_queue *q,
1332 unsigned int nr_budgets)
1336 for (i = 0; i < nr_budgets; i++)
1337 blk_mq_put_dispatch_budget(q);
1341 * Returns true if we did some work AND can potentially do more.
1343 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1344 unsigned int nr_budgets)
1346 enum prep_dispatch prep;
1347 struct request_queue *q = hctx->queue;
1348 struct request *rq, *nxt;
1350 blk_status_t ret = BLK_STS_OK;
1351 LIST_HEAD(zone_list);
1353 if (list_empty(list))
1357 * Now process all the entries, sending them to the driver.
1359 errors = queued = 0;
1361 struct blk_mq_queue_data bd;
1363 rq = list_first_entry(list, struct request, queuelist);
1365 WARN_ON_ONCE(hctx != rq->mq_hctx);
1366 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1367 if (prep != PREP_DISPATCH_OK)
1370 list_del_init(&rq->queuelist);
1375 * Flag last if we have no more requests, or if we have more
1376 * but can't assign a driver tag to it.
1378 if (list_empty(list))
1381 nxt = list_first_entry(list, struct request, queuelist);
1382 bd.last = !blk_mq_get_driver_tag(nxt);
1386 * once the request is queued to lld, no need to cover the
1391 ret = q->mq_ops->queue_rq(hctx, &bd);
1392 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1393 blk_mq_handle_dev_resource(rq, list);
1395 } else if (ret == BLK_STS_ZONE_RESOURCE) {
1397 * Move the request to zone_list and keep going through
1398 * the dispatch list to find more requests the drive can
1401 blk_mq_handle_zone_resource(rq, &zone_list);
1402 if (list_empty(list))
1407 if (unlikely(ret != BLK_STS_OK)) {
1409 blk_mq_end_request(rq, BLK_STS_IOERR);
1414 } while (!list_empty(list));
1416 if (!list_empty(&zone_list))
1417 list_splice_tail_init(&zone_list, list);
1419 hctx->dispatched[queued_to_index(queued)]++;
1422 * Any items that need requeuing? Stuff them into hctx->dispatch,
1423 * that is where we will continue on next queue run.
1425 if (!list_empty(list)) {
1427 /* For non-shared tags, the RESTART check will suffice */
1428 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1429 (hctx->flags & BLK_MQ_F_TAG_SHARED);
1430 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1432 blk_mq_release_budgets(q, nr_budgets);
1435 * If we didn't flush the entire list, we could have told
1436 * the driver there was more coming, but that turned out to
1439 if (q->mq_ops->commit_rqs && queued)
1440 q->mq_ops->commit_rqs(hctx);
1442 spin_lock(&hctx->lock);
1443 list_splice_tail_init(list, &hctx->dispatch);
1444 spin_unlock(&hctx->lock);
1447 * If SCHED_RESTART was set by the caller of this function and
1448 * it is no longer set that means that it was cleared by another
1449 * thread and hence that a queue rerun is needed.
1451 * If 'no_tag' is set, that means that we failed getting
1452 * a driver tag with an I/O scheduler attached. If our dispatch
1453 * waitqueue is no longer active, ensure that we run the queue
1454 * AFTER adding our entries back to the list.
1456 * If no I/O scheduler has been configured it is possible that
1457 * the hardware queue got stopped and restarted before requests
1458 * were pushed back onto the dispatch list. Rerun the queue to
1459 * avoid starvation. Notes:
1460 * - blk_mq_run_hw_queue() checks whether or not a queue has
1461 * been stopped before rerunning a queue.
1462 * - Some but not all block drivers stop a queue before
1463 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1466 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1467 * bit is set, run queue after a delay to avoid IO stalls
1468 * that could otherwise occur if the queue is idle. We'll do
1469 * similar if we couldn't get budget and SCHED_RESTART is set.
1471 needs_restart = blk_mq_sched_needs_restart(hctx);
1472 if (!needs_restart ||
1473 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1474 blk_mq_run_hw_queue(hctx, true);
1475 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1477 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1479 blk_mq_update_dispatch_busy(hctx, true);
1482 blk_mq_update_dispatch_busy(hctx, false);
1484 return (queued + errors) != 0;
1488 * __blk_mq_run_hw_queue - Run a hardware queue.
1489 * @hctx: Pointer to the hardware queue to run.
1491 * Send pending requests to the hardware.
1493 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1498 * We should be running this queue from one of the CPUs that
1501 * There are at least two related races now between setting
1502 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1503 * __blk_mq_run_hw_queue():
1505 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1506 * but later it becomes online, then this warning is harmless
1509 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1510 * but later it becomes offline, then the warning can't be
1511 * triggered, and we depend on blk-mq timeout handler to
1512 * handle dispatched requests to this hctx
1514 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1515 cpu_online(hctx->next_cpu)) {
1516 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1517 raw_smp_processor_id(),
1518 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1523 * We can't run the queue inline with ints disabled. Ensure that
1524 * we catch bad users of this early.
1526 WARN_ON_ONCE(in_interrupt());
1528 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1530 hctx_lock(hctx, &srcu_idx);
1531 blk_mq_sched_dispatch_requests(hctx);
1532 hctx_unlock(hctx, srcu_idx);
1535 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1537 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1539 if (cpu >= nr_cpu_ids)
1540 cpu = cpumask_first(hctx->cpumask);
1545 * It'd be great if the workqueue API had a way to pass
1546 * in a mask and had some smarts for more clever placement.
1547 * For now we just round-robin here, switching for every
1548 * BLK_MQ_CPU_WORK_BATCH queued items.
1550 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1553 int next_cpu = hctx->next_cpu;
1555 if (hctx->queue->nr_hw_queues == 1)
1556 return WORK_CPU_UNBOUND;
1558 if (--hctx->next_cpu_batch <= 0) {
1560 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1562 if (next_cpu >= nr_cpu_ids)
1563 next_cpu = blk_mq_first_mapped_cpu(hctx);
1564 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1568 * Do unbound schedule if we can't find a online CPU for this hctx,
1569 * and it should only happen in the path of handling CPU DEAD.
1571 if (!cpu_online(next_cpu)) {
1578 * Make sure to re-select CPU next time once after CPUs
1579 * in hctx->cpumask become online again.
1581 hctx->next_cpu = next_cpu;
1582 hctx->next_cpu_batch = 1;
1583 return WORK_CPU_UNBOUND;
1586 hctx->next_cpu = next_cpu;
1591 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1592 * @hctx: Pointer to the hardware queue to run.
1593 * @async: If we want to run the queue asynchronously.
1594 * @msecs: Microseconds of delay to wait before running the queue.
1596 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1597 * with a delay of @msecs.
1599 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1600 unsigned long msecs)
1602 if (unlikely(blk_mq_hctx_stopped(hctx)))
1605 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1606 int cpu = get_cpu();
1607 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1608 __blk_mq_run_hw_queue(hctx);
1616 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1617 msecs_to_jiffies(msecs));
1621 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1622 * @hctx: Pointer to the hardware queue to run.
1623 * @msecs: Microseconds of delay to wait before running the queue.
1625 * Run a hardware queue asynchronously with a delay of @msecs.
1627 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1629 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1631 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1634 * blk_mq_run_hw_queue - Start to run a hardware queue.
1635 * @hctx: Pointer to the hardware queue to run.
1636 * @async: If we want to run the queue asynchronously.
1638 * Check if the request queue is not in a quiesced state and if there are
1639 * pending requests to be sent. If this is true, run the queue to send requests
1642 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1648 * When queue is quiesced, we may be switching io scheduler, or
1649 * updating nr_hw_queues, or other things, and we can't run queue
1650 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1652 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1655 hctx_lock(hctx, &srcu_idx);
1656 need_run = !blk_queue_quiesced(hctx->queue) &&
1657 blk_mq_hctx_has_pending(hctx);
1658 hctx_unlock(hctx, srcu_idx);
1661 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1663 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1666 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1667 * @q: Pointer to the request queue to run.
1668 * @async: If we want to run the queue asynchronously.
1670 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1672 struct blk_mq_hw_ctx *hctx;
1675 queue_for_each_hw_ctx(q, hctx, i) {
1676 if (blk_mq_hctx_stopped(hctx))
1679 blk_mq_run_hw_queue(hctx, async);
1682 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1685 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1686 * @q: Pointer to the request queue to run.
1687 * @msecs: Microseconds of delay to wait before running the queues.
1689 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1691 struct blk_mq_hw_ctx *hctx;
1694 queue_for_each_hw_ctx(q, hctx, i) {
1695 if (blk_mq_hctx_stopped(hctx))
1698 blk_mq_delay_run_hw_queue(hctx, msecs);
1701 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1704 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1705 * @q: request queue.
1707 * The caller is responsible for serializing this function against
1708 * blk_mq_{start,stop}_hw_queue().
1710 bool blk_mq_queue_stopped(struct request_queue *q)
1712 struct blk_mq_hw_ctx *hctx;
1715 queue_for_each_hw_ctx(q, hctx, i)
1716 if (blk_mq_hctx_stopped(hctx))
1721 EXPORT_SYMBOL(blk_mq_queue_stopped);
1724 * This function is often used for pausing .queue_rq() by driver when
1725 * there isn't enough resource or some conditions aren't satisfied, and
1726 * BLK_STS_RESOURCE is usually returned.
1728 * We do not guarantee that dispatch can be drained or blocked
1729 * after blk_mq_stop_hw_queue() returns. Please use
1730 * blk_mq_quiesce_queue() for that requirement.
1732 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1734 cancel_delayed_work(&hctx->run_work);
1736 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1738 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1741 * This function is often used for pausing .queue_rq() by driver when
1742 * there isn't enough resource or some conditions aren't satisfied, and
1743 * BLK_STS_RESOURCE is usually returned.
1745 * We do not guarantee that dispatch can be drained or blocked
1746 * after blk_mq_stop_hw_queues() returns. Please use
1747 * blk_mq_quiesce_queue() for that requirement.
1749 void blk_mq_stop_hw_queues(struct request_queue *q)
1751 struct blk_mq_hw_ctx *hctx;
1754 queue_for_each_hw_ctx(q, hctx, i)
1755 blk_mq_stop_hw_queue(hctx);
1757 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1759 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1761 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1763 blk_mq_run_hw_queue(hctx, false);
1765 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1767 void blk_mq_start_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_start_hw_queue(hctx);
1775 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1777 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1779 if (!blk_mq_hctx_stopped(hctx))
1782 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1783 blk_mq_run_hw_queue(hctx, async);
1785 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1787 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1789 struct blk_mq_hw_ctx *hctx;
1792 queue_for_each_hw_ctx(q, hctx, i)
1793 blk_mq_start_stopped_hw_queue(hctx, async);
1795 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1797 static void blk_mq_run_work_fn(struct work_struct *work)
1799 struct blk_mq_hw_ctx *hctx;
1801 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1804 * If we are stopped, don't run the queue.
1806 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1809 __blk_mq_run_hw_queue(hctx);
1812 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1816 struct blk_mq_ctx *ctx = rq->mq_ctx;
1817 enum hctx_type type = hctx->type;
1819 lockdep_assert_held(&ctx->lock);
1821 trace_block_rq_insert(hctx->queue, rq);
1824 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1826 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1829 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1832 struct blk_mq_ctx *ctx = rq->mq_ctx;
1834 lockdep_assert_held(&ctx->lock);
1836 __blk_mq_insert_req_list(hctx, rq, at_head);
1837 blk_mq_hctx_mark_pending(hctx, ctx);
1841 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1842 * @rq: Pointer to request to be inserted.
1843 * @run_queue: If we should run the hardware queue after inserting the request.
1845 * Should only be used carefully, when the caller knows we want to
1846 * bypass a potential IO scheduler on the target device.
1848 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1851 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1853 spin_lock(&hctx->lock);
1855 list_add(&rq->queuelist, &hctx->dispatch);
1857 list_add_tail(&rq->queuelist, &hctx->dispatch);
1858 spin_unlock(&hctx->lock);
1861 blk_mq_run_hw_queue(hctx, false);
1864 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1865 struct list_head *list)
1869 enum hctx_type type = hctx->type;
1872 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1875 list_for_each_entry(rq, list, queuelist) {
1876 BUG_ON(rq->mq_ctx != ctx);
1877 trace_block_rq_insert(hctx->queue, rq);
1880 spin_lock(&ctx->lock);
1881 list_splice_tail_init(list, &ctx->rq_lists[type]);
1882 blk_mq_hctx_mark_pending(hctx, ctx);
1883 spin_unlock(&ctx->lock);
1886 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1888 struct request *rqa = container_of(a, struct request, queuelist);
1889 struct request *rqb = container_of(b, struct request, queuelist);
1891 if (rqa->mq_ctx != rqb->mq_ctx)
1892 return rqa->mq_ctx > rqb->mq_ctx;
1893 if (rqa->mq_hctx != rqb->mq_hctx)
1894 return rqa->mq_hctx > rqb->mq_hctx;
1896 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1899 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1903 if (list_empty(&plug->mq_list))
1905 list_splice_init(&plug->mq_list, &list);
1907 if (plug->rq_count > 2 && plug->multiple_queues)
1908 list_sort(NULL, &list, plug_rq_cmp);
1913 struct list_head rq_list;
1914 struct request *rq, *head_rq = list_entry_rq(list.next);
1915 struct list_head *pos = &head_rq->queuelist; /* skip first */
1916 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1917 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1918 unsigned int depth = 1;
1920 list_for_each_continue(pos, &list) {
1921 rq = list_entry_rq(pos);
1923 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1928 list_cut_before(&rq_list, &list, pos);
1929 trace_block_unplug(head_rq->q, depth, !from_schedule);
1930 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1932 } while(!list_empty(&list));
1935 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1936 unsigned int nr_segs)
1938 if (bio->bi_opf & REQ_RAHEAD)
1939 rq->cmd_flags |= REQ_FAILFAST_MASK;
1941 rq->__sector = bio->bi_iter.bi_sector;
1942 rq->write_hint = bio->bi_write_hint;
1943 blk_rq_bio_prep(rq, bio, nr_segs);
1944 blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1946 blk_account_io_start(rq);
1949 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1951 blk_qc_t *cookie, bool last)
1953 struct request_queue *q = rq->q;
1954 struct blk_mq_queue_data bd = {
1958 blk_qc_t new_cookie;
1961 new_cookie = request_to_qc_t(hctx, rq);
1964 * For OK queue, we are done. For error, caller may kill it.
1965 * Any other error (busy), just add it to our list as we
1966 * previously would have done.
1968 ret = q->mq_ops->queue_rq(hctx, &bd);
1971 blk_mq_update_dispatch_busy(hctx, false);
1972 *cookie = new_cookie;
1974 case BLK_STS_RESOURCE:
1975 case BLK_STS_DEV_RESOURCE:
1976 blk_mq_update_dispatch_busy(hctx, true);
1977 __blk_mq_requeue_request(rq);
1980 blk_mq_update_dispatch_busy(hctx, false);
1981 *cookie = BLK_QC_T_NONE;
1988 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1991 bool bypass_insert, bool last)
1993 struct request_queue *q = rq->q;
1994 bool run_queue = true;
1997 * RCU or SRCU read lock is needed before checking quiesced flag.
1999 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2000 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2001 * and avoid driver to try to dispatch again.
2003 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2005 bypass_insert = false;
2009 if (q->elevator && !bypass_insert)
2012 if (!blk_mq_get_dispatch_budget(q))
2015 if (!blk_mq_get_driver_tag(rq)) {
2016 blk_mq_put_dispatch_budget(q);
2020 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2023 return BLK_STS_RESOURCE;
2025 blk_mq_request_bypass_insert(rq, false, run_queue);
2030 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2031 * @hctx: Pointer of the associated hardware queue.
2032 * @rq: Pointer to request to be sent.
2033 * @cookie: Request queue cookie.
2035 * If the device has enough resources to accept a new request now, send the
2036 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2037 * we can try send it another time in the future. Requests inserted at this
2038 * queue have higher priority.
2040 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2041 struct request *rq, blk_qc_t *cookie)
2046 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2048 hctx_lock(hctx, &srcu_idx);
2050 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2051 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2052 blk_mq_request_bypass_insert(rq, false, true);
2053 else if (ret != BLK_STS_OK)
2054 blk_mq_end_request(rq, ret);
2056 hctx_unlock(hctx, srcu_idx);
2059 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2063 blk_qc_t unused_cookie;
2064 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2066 hctx_lock(hctx, &srcu_idx);
2067 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2068 hctx_unlock(hctx, srcu_idx);
2073 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2074 struct list_head *list)
2078 while (!list_empty(list)) {
2080 struct request *rq = list_first_entry(list, struct request,
2083 list_del_init(&rq->queuelist);
2084 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2085 if (ret != BLK_STS_OK) {
2086 if (ret == BLK_STS_RESOURCE ||
2087 ret == BLK_STS_DEV_RESOURCE) {
2088 blk_mq_request_bypass_insert(rq, false,
2092 blk_mq_end_request(rq, ret);
2098 * If we didn't flush the entire list, we could have told
2099 * the driver there was more coming, but that turned out to
2102 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs && queued)
2103 hctx->queue->mq_ops->commit_rqs(hctx);
2106 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2108 list_add_tail(&rq->queuelist, &plug->mq_list);
2110 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2111 struct request *tmp;
2113 tmp = list_first_entry(&plug->mq_list, struct request,
2115 if (tmp->q != rq->q)
2116 plug->multiple_queues = true;
2121 * blk_mq_make_request - Create and send a request to block device.
2122 * @q: Request queue pointer.
2123 * @bio: Bio pointer.
2125 * Builds up a request structure from @q and @bio and send to the device. The
2126 * request may not be queued directly to hardware if:
2127 * * This request can be merged with another one
2128 * * We want to place request at plug queue for possible future merging
2129 * * There is an IO scheduler active at this queue
2131 * It will not queue the request if there is an error with the bio, or at the
2134 * Returns: Request queue cookie.
2136 blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
2138 const int is_sync = op_is_sync(bio->bi_opf);
2139 const int is_flush_fua = op_is_flush(bio->bi_opf);
2140 struct blk_mq_alloc_data data = {
2144 struct blk_plug *plug;
2145 struct request *same_queue_rq = NULL;
2146 unsigned int nr_segs;
2150 blk_queue_bounce(q, &bio);
2151 __blk_queue_split(q, &bio, &nr_segs);
2153 if (!bio_integrity_prep(bio))
2156 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2157 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2160 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2163 rq_qos_throttle(q, bio);
2165 data.cmd_flags = bio->bi_opf;
2166 rq = __blk_mq_alloc_request(&data);
2167 if (unlikely(!rq)) {
2168 rq_qos_cleanup(q, bio);
2169 if (bio->bi_opf & REQ_NOWAIT)
2170 bio_wouldblock_error(bio);
2174 trace_block_getrq(q, bio, bio->bi_opf);
2176 rq_qos_track(q, rq, bio);
2178 cookie = request_to_qc_t(data.hctx, rq);
2180 blk_mq_bio_to_request(rq, bio, nr_segs);
2182 ret = blk_crypto_init_request(rq);
2183 if (ret != BLK_STS_OK) {
2184 bio->bi_status = ret;
2186 blk_mq_free_request(rq);
2187 return BLK_QC_T_NONE;
2190 plug = blk_mq_plug(q, bio);
2191 if (unlikely(is_flush_fua)) {
2192 /* Bypass scheduler for flush requests */
2193 blk_insert_flush(rq);
2194 blk_mq_run_hw_queue(data.hctx, true);
2195 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2196 !blk_queue_nonrot(q))) {
2198 * Use plugging if we have a ->commit_rqs() hook as well, as
2199 * we know the driver uses bd->last in a smart fashion.
2201 * Use normal plugging if this disk is slow HDD, as sequential
2202 * IO may benefit a lot from plug merging.
2204 unsigned int request_count = plug->rq_count;
2205 struct request *last = NULL;
2208 trace_block_plug(q);
2210 last = list_entry_rq(plug->mq_list.prev);
2212 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2213 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2214 blk_flush_plug_list(plug, false);
2215 trace_block_plug(q);
2218 blk_add_rq_to_plug(plug, rq);
2219 } else if (q->elevator) {
2220 /* Insert the request at the IO scheduler queue */
2221 blk_mq_sched_insert_request(rq, false, true, true);
2222 } else if (plug && !blk_queue_nomerges(q)) {
2224 * We do limited plugging. If the bio can be merged, do that.
2225 * Otherwise the existing request in the plug list will be
2226 * issued. So the plug list will have one request at most
2227 * The plug list might get flushed before this. If that happens,
2228 * the plug list is empty, and same_queue_rq is invalid.
2230 if (list_empty(&plug->mq_list))
2231 same_queue_rq = NULL;
2232 if (same_queue_rq) {
2233 list_del_init(&same_queue_rq->queuelist);
2236 blk_add_rq_to_plug(plug, rq);
2237 trace_block_plug(q);
2239 if (same_queue_rq) {
2240 data.hctx = same_queue_rq->mq_hctx;
2241 trace_block_unplug(q, 1, true);
2242 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2245 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2246 !data.hctx->dispatch_busy) {
2248 * There is no scheduler and we can try to send directly
2251 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2254 blk_mq_sched_insert_request(rq, false, true, true);
2260 return BLK_QC_T_NONE;
2262 EXPORT_SYMBOL_GPL(blk_mq_make_request); /* only for request based dm */
2264 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2265 unsigned int hctx_idx)
2269 if (tags->rqs && set->ops->exit_request) {
2272 for (i = 0; i < tags->nr_tags; i++) {
2273 struct request *rq = tags->static_rqs[i];
2277 set->ops->exit_request(set, rq, hctx_idx);
2278 tags->static_rqs[i] = NULL;
2282 while (!list_empty(&tags->page_list)) {
2283 page = list_first_entry(&tags->page_list, struct page, lru);
2284 list_del_init(&page->lru);
2286 * Remove kmemleak object previously allocated in
2287 * blk_mq_alloc_rqs().
2289 kmemleak_free(page_address(page));
2290 __free_pages(page, page->private);
2294 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2298 kfree(tags->static_rqs);
2299 tags->static_rqs = NULL;
2301 blk_mq_free_tags(tags);
2304 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2305 unsigned int hctx_idx,
2306 unsigned int nr_tags,
2307 unsigned int reserved_tags)
2309 struct blk_mq_tags *tags;
2312 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2313 if (node == NUMA_NO_NODE)
2314 node = set->numa_node;
2316 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2317 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2321 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2322 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2325 blk_mq_free_tags(tags);
2329 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2330 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2332 if (!tags->static_rqs) {
2334 blk_mq_free_tags(tags);
2341 static size_t order_to_size(unsigned int order)
2343 return (size_t)PAGE_SIZE << order;
2346 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2347 unsigned int hctx_idx, int node)
2351 if (set->ops->init_request) {
2352 ret = set->ops->init_request(set, rq, hctx_idx, node);
2357 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2361 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2362 unsigned int hctx_idx, unsigned int depth)
2364 unsigned int i, j, entries_per_page, max_order = 4;
2365 size_t rq_size, left;
2368 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2369 if (node == NUMA_NO_NODE)
2370 node = set->numa_node;
2372 INIT_LIST_HEAD(&tags->page_list);
2375 * rq_size is the size of the request plus driver payload, rounded
2376 * to the cacheline size
2378 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2380 left = rq_size * depth;
2382 for (i = 0; i < depth; ) {
2383 int this_order = max_order;
2388 while (this_order && left < order_to_size(this_order - 1))
2392 page = alloc_pages_node(node,
2393 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2399 if (order_to_size(this_order) < rq_size)
2406 page->private = this_order;
2407 list_add_tail(&page->lru, &tags->page_list);
2409 p = page_address(page);
2411 * Allow kmemleak to scan these pages as they contain pointers
2412 * to additional allocations like via ops->init_request().
2414 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2415 entries_per_page = order_to_size(this_order) / rq_size;
2416 to_do = min(entries_per_page, depth - i);
2417 left -= to_do * rq_size;
2418 for (j = 0; j < to_do; j++) {
2419 struct request *rq = p;
2421 tags->static_rqs[i] = rq;
2422 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2423 tags->static_rqs[i] = NULL;
2434 blk_mq_free_rqs(set, tags, hctx_idx);
2438 struct rq_iter_data {
2439 struct blk_mq_hw_ctx *hctx;
2443 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2445 struct rq_iter_data *iter_data = data;
2447 if (rq->mq_hctx != iter_data->hctx)
2449 iter_data->has_rq = true;
2453 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2455 struct blk_mq_tags *tags = hctx->sched_tags ?
2456 hctx->sched_tags : hctx->tags;
2457 struct rq_iter_data data = {
2461 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2465 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2466 struct blk_mq_hw_ctx *hctx)
2468 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2470 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2475 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2477 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2478 struct blk_mq_hw_ctx, cpuhp_online);
2480 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2481 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2485 * Prevent new request from being allocated on the current hctx.
2487 * The smp_mb__after_atomic() Pairs with the implied barrier in
2488 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2489 * seen once we return from the tag allocator.
2491 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2492 smp_mb__after_atomic();
2495 * Try to grab a reference to the queue and wait for any outstanding
2496 * requests. If we could not grab a reference the queue has been
2497 * frozen and there are no requests.
2499 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2500 while (blk_mq_hctx_has_requests(hctx))
2502 percpu_ref_put(&hctx->queue->q_usage_counter);
2508 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2510 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2511 struct blk_mq_hw_ctx, cpuhp_online);
2513 if (cpumask_test_cpu(cpu, hctx->cpumask))
2514 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2519 * 'cpu' is going away. splice any existing rq_list entries from this
2520 * software queue to the hw queue dispatch list, and ensure that it
2523 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2525 struct blk_mq_hw_ctx *hctx;
2526 struct blk_mq_ctx *ctx;
2528 enum hctx_type type;
2530 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2531 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2534 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2537 spin_lock(&ctx->lock);
2538 if (!list_empty(&ctx->rq_lists[type])) {
2539 list_splice_init(&ctx->rq_lists[type], &tmp);
2540 blk_mq_hctx_clear_pending(hctx, ctx);
2542 spin_unlock(&ctx->lock);
2544 if (list_empty(&tmp))
2547 spin_lock(&hctx->lock);
2548 list_splice_tail_init(&tmp, &hctx->dispatch);
2549 spin_unlock(&hctx->lock);
2551 blk_mq_run_hw_queue(hctx, true);
2555 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2557 if (!(hctx->flags & BLK_MQ_F_STACKING))
2558 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2559 &hctx->cpuhp_online);
2560 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2564 /* hctx->ctxs will be freed in queue's release handler */
2565 static void blk_mq_exit_hctx(struct request_queue *q,
2566 struct blk_mq_tag_set *set,
2567 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2569 if (blk_mq_hw_queue_mapped(hctx))
2570 blk_mq_tag_idle(hctx);
2572 if (set->ops->exit_request)
2573 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2575 if (set->ops->exit_hctx)
2576 set->ops->exit_hctx(hctx, hctx_idx);
2578 blk_mq_remove_cpuhp(hctx);
2580 spin_lock(&q->unused_hctx_lock);
2581 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2582 spin_unlock(&q->unused_hctx_lock);
2585 static void blk_mq_exit_hw_queues(struct request_queue *q,
2586 struct blk_mq_tag_set *set, int nr_queue)
2588 struct blk_mq_hw_ctx *hctx;
2591 queue_for_each_hw_ctx(q, hctx, i) {
2594 blk_mq_debugfs_unregister_hctx(hctx);
2595 blk_mq_exit_hctx(q, set, hctx, i);
2599 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2601 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2603 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2604 __alignof__(struct blk_mq_hw_ctx)) !=
2605 sizeof(struct blk_mq_hw_ctx));
2607 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2608 hw_ctx_size += sizeof(struct srcu_struct);
2613 static int blk_mq_init_hctx(struct request_queue *q,
2614 struct blk_mq_tag_set *set,
2615 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2617 hctx->queue_num = hctx_idx;
2619 if (!(hctx->flags & BLK_MQ_F_STACKING))
2620 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2621 &hctx->cpuhp_online);
2622 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2624 hctx->tags = set->tags[hctx_idx];
2626 if (set->ops->init_hctx &&
2627 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2628 goto unregister_cpu_notifier;
2630 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2636 if (set->ops->exit_hctx)
2637 set->ops->exit_hctx(hctx, hctx_idx);
2638 unregister_cpu_notifier:
2639 blk_mq_remove_cpuhp(hctx);
2643 static struct blk_mq_hw_ctx *
2644 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2647 struct blk_mq_hw_ctx *hctx;
2648 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2650 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2652 goto fail_alloc_hctx;
2654 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2657 atomic_set(&hctx->nr_active, 0);
2658 if (node == NUMA_NO_NODE)
2659 node = set->numa_node;
2660 hctx->numa_node = node;
2662 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2663 spin_lock_init(&hctx->lock);
2664 INIT_LIST_HEAD(&hctx->dispatch);
2666 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2668 INIT_LIST_HEAD(&hctx->hctx_list);
2671 * Allocate space for all possible cpus to avoid allocation at
2674 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2679 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2684 spin_lock_init(&hctx->dispatch_wait_lock);
2685 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2686 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2688 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2692 if (hctx->flags & BLK_MQ_F_BLOCKING)
2693 init_srcu_struct(hctx->srcu);
2694 blk_mq_hctx_kobj_init(hctx);
2699 sbitmap_free(&hctx->ctx_map);
2703 free_cpumask_var(hctx->cpumask);
2710 static void blk_mq_init_cpu_queues(struct request_queue *q,
2711 unsigned int nr_hw_queues)
2713 struct blk_mq_tag_set *set = q->tag_set;
2716 for_each_possible_cpu(i) {
2717 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2718 struct blk_mq_hw_ctx *hctx;
2722 spin_lock_init(&__ctx->lock);
2723 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2724 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2729 * Set local node, IFF we have more than one hw queue. If
2730 * not, we remain on the home node of the device
2732 for (j = 0; j < set->nr_maps; j++) {
2733 hctx = blk_mq_map_queue_type(q, j, i);
2734 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2735 hctx->numa_node = local_memory_node(cpu_to_node(i));
2740 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2745 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2746 set->queue_depth, set->reserved_tags);
2747 if (!set->tags[hctx_idx])
2750 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2755 blk_mq_free_rq_map(set->tags[hctx_idx]);
2756 set->tags[hctx_idx] = NULL;
2760 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2761 unsigned int hctx_idx)
2763 if (set->tags && set->tags[hctx_idx]) {
2764 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2765 blk_mq_free_rq_map(set->tags[hctx_idx]);
2766 set->tags[hctx_idx] = NULL;
2770 static void blk_mq_map_swqueue(struct request_queue *q)
2772 unsigned int i, j, hctx_idx;
2773 struct blk_mq_hw_ctx *hctx;
2774 struct blk_mq_ctx *ctx;
2775 struct blk_mq_tag_set *set = q->tag_set;
2777 queue_for_each_hw_ctx(q, hctx, i) {
2778 cpumask_clear(hctx->cpumask);
2780 hctx->dispatch_from = NULL;
2784 * Map software to hardware queues.
2786 * If the cpu isn't present, the cpu is mapped to first hctx.
2788 for_each_possible_cpu(i) {
2790 ctx = per_cpu_ptr(q->queue_ctx, i);
2791 for (j = 0; j < set->nr_maps; j++) {
2792 if (!set->map[j].nr_queues) {
2793 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2794 HCTX_TYPE_DEFAULT, i);
2797 hctx_idx = set->map[j].mq_map[i];
2798 /* unmapped hw queue can be remapped after CPU topo changed */
2799 if (!set->tags[hctx_idx] &&
2800 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2802 * If tags initialization fail for some hctx,
2803 * that hctx won't be brought online. In this
2804 * case, remap the current ctx to hctx[0] which
2805 * is guaranteed to always have tags allocated
2807 set->map[j].mq_map[i] = 0;
2810 hctx = blk_mq_map_queue_type(q, j, i);
2811 ctx->hctxs[j] = hctx;
2813 * If the CPU is already set in the mask, then we've
2814 * mapped this one already. This can happen if
2815 * devices share queues across queue maps.
2817 if (cpumask_test_cpu(i, hctx->cpumask))
2820 cpumask_set_cpu(i, hctx->cpumask);
2822 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2823 hctx->ctxs[hctx->nr_ctx++] = ctx;
2826 * If the nr_ctx type overflows, we have exceeded the
2827 * amount of sw queues we can support.
2829 BUG_ON(!hctx->nr_ctx);
2832 for (; j < HCTX_MAX_TYPES; j++)
2833 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2834 HCTX_TYPE_DEFAULT, i);
2837 queue_for_each_hw_ctx(q, hctx, i) {
2839 * If no software queues are mapped to this hardware queue,
2840 * disable it and free the request entries.
2842 if (!hctx->nr_ctx) {
2843 /* Never unmap queue 0. We need it as a
2844 * fallback in case of a new remap fails
2847 if (i && set->tags[i])
2848 blk_mq_free_map_and_requests(set, i);
2854 hctx->tags = set->tags[i];
2855 WARN_ON(!hctx->tags);
2858 * Set the map size to the number of mapped software queues.
2859 * This is more accurate and more efficient than looping
2860 * over all possibly mapped software queues.
2862 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2865 * Initialize batch roundrobin counts
2867 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2868 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2873 * Caller needs to ensure that we're either frozen/quiesced, or that
2874 * the queue isn't live yet.
2876 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2878 struct blk_mq_hw_ctx *hctx;
2881 queue_for_each_hw_ctx(q, hctx, i) {
2883 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2885 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2889 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2892 struct request_queue *q;
2894 lockdep_assert_held(&set->tag_list_lock);
2896 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2897 blk_mq_freeze_queue(q);
2898 queue_set_hctx_shared(q, shared);
2899 blk_mq_unfreeze_queue(q);
2903 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2905 struct blk_mq_tag_set *set = q->tag_set;
2907 mutex_lock(&set->tag_list_lock);
2908 list_del_rcu(&q->tag_set_list);
2909 if (list_is_singular(&set->tag_list)) {
2910 /* just transitioned to unshared */
2911 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2912 /* update existing queue */
2913 blk_mq_update_tag_set_depth(set, false);
2915 mutex_unlock(&set->tag_list_lock);
2916 INIT_LIST_HEAD(&q->tag_set_list);
2919 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2920 struct request_queue *q)
2922 mutex_lock(&set->tag_list_lock);
2925 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2927 if (!list_empty(&set->tag_list) &&
2928 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2929 set->flags |= BLK_MQ_F_TAG_SHARED;
2930 /* update existing queue */
2931 blk_mq_update_tag_set_depth(set, true);
2933 if (set->flags & BLK_MQ_F_TAG_SHARED)
2934 queue_set_hctx_shared(q, true);
2935 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2937 mutex_unlock(&set->tag_list_lock);
2940 /* All allocations will be freed in release handler of q->mq_kobj */
2941 static int blk_mq_alloc_ctxs(struct request_queue *q)
2943 struct blk_mq_ctxs *ctxs;
2946 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2950 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2951 if (!ctxs->queue_ctx)
2954 for_each_possible_cpu(cpu) {
2955 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2959 q->mq_kobj = &ctxs->kobj;
2960 q->queue_ctx = ctxs->queue_ctx;
2969 * It is the actual release handler for mq, but we do it from
2970 * request queue's release handler for avoiding use-after-free
2971 * and headache because q->mq_kobj shouldn't have been introduced,
2972 * but we can't group ctx/kctx kobj without it.
2974 void blk_mq_release(struct request_queue *q)
2976 struct blk_mq_hw_ctx *hctx, *next;
2979 queue_for_each_hw_ctx(q, hctx, i)
2980 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2982 /* all hctx are in .unused_hctx_list now */
2983 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2984 list_del_init(&hctx->hctx_list);
2985 kobject_put(&hctx->kobj);
2988 kfree(q->queue_hw_ctx);
2991 * release .mq_kobj and sw queue's kobject now because
2992 * both share lifetime with request queue.
2994 blk_mq_sysfs_deinit(q);
2997 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3000 struct request_queue *uninit_q, *q;
3002 uninit_q = __blk_alloc_queue(set->numa_node);
3004 return ERR_PTR(-ENOMEM);
3005 uninit_q->queuedata = queuedata;
3008 * Initialize the queue without an elevator. device_add_disk() will do
3009 * the initialization.
3011 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3013 blk_cleanup_queue(uninit_q);
3017 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3019 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3021 return blk_mq_init_queue_data(set, NULL);
3023 EXPORT_SYMBOL(blk_mq_init_queue);
3026 * Helper for setting up a queue with mq ops, given queue depth, and
3027 * the passed in mq ops flags.
3029 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3030 const struct blk_mq_ops *ops,
3031 unsigned int queue_depth,
3032 unsigned int set_flags)
3034 struct request_queue *q;
3037 memset(set, 0, sizeof(*set));
3039 set->nr_hw_queues = 1;
3041 set->queue_depth = queue_depth;
3042 set->numa_node = NUMA_NO_NODE;
3043 set->flags = set_flags;
3045 ret = blk_mq_alloc_tag_set(set);
3047 return ERR_PTR(ret);
3049 q = blk_mq_init_queue(set);
3051 blk_mq_free_tag_set(set);
3057 EXPORT_SYMBOL(blk_mq_init_sq_queue);
3059 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3060 struct blk_mq_tag_set *set, struct request_queue *q,
3061 int hctx_idx, int node)
3063 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3065 /* reuse dead hctx first */
3066 spin_lock(&q->unused_hctx_lock);
3067 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3068 if (tmp->numa_node == node) {
3074 list_del_init(&hctx->hctx_list);
3075 spin_unlock(&q->unused_hctx_lock);
3078 hctx = blk_mq_alloc_hctx(q, set, node);
3082 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3088 kobject_put(&hctx->kobj);
3093 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3094 struct request_queue *q)
3097 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3099 if (q->nr_hw_queues < set->nr_hw_queues) {
3100 struct blk_mq_hw_ctx **new_hctxs;
3102 new_hctxs = kcalloc_node(set->nr_hw_queues,
3103 sizeof(*new_hctxs), GFP_KERNEL,
3108 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3110 q->queue_hw_ctx = new_hctxs;
3115 /* protect against switching io scheduler */
3116 mutex_lock(&q->sysfs_lock);
3117 for (i = 0; i < set->nr_hw_queues; i++) {
3119 struct blk_mq_hw_ctx *hctx;
3121 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3123 * If the hw queue has been mapped to another numa node,
3124 * we need to realloc the hctx. If allocation fails, fallback
3125 * to use the previous one.
3127 if (hctxs[i] && (hctxs[i]->numa_node == node))
3130 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3133 blk_mq_exit_hctx(q, set, hctxs[i], i);
3137 pr_warn("Allocate new hctx on node %d fails,\
3138 fallback to previous one on node %d\n",
3139 node, hctxs[i]->numa_node);
3145 * Increasing nr_hw_queues fails. Free the newly allocated
3146 * hctxs and keep the previous q->nr_hw_queues.
3148 if (i != set->nr_hw_queues) {
3149 j = q->nr_hw_queues;
3153 end = q->nr_hw_queues;
3154 q->nr_hw_queues = set->nr_hw_queues;
3157 for (; j < end; j++) {
3158 struct blk_mq_hw_ctx *hctx = hctxs[j];
3162 blk_mq_free_map_and_requests(set, j);
3163 blk_mq_exit_hctx(q, set, hctx, j);
3167 mutex_unlock(&q->sysfs_lock);
3170 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3171 struct request_queue *q,
3174 /* mark the queue as mq asap */
3175 q->mq_ops = set->ops;
3177 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3178 blk_mq_poll_stats_bkt,
3179 BLK_MQ_POLL_STATS_BKTS, q);
3183 if (blk_mq_alloc_ctxs(q))
3186 /* init q->mq_kobj and sw queues' kobjects */
3187 blk_mq_sysfs_init(q);
3189 INIT_LIST_HEAD(&q->unused_hctx_list);
3190 spin_lock_init(&q->unused_hctx_lock);
3192 blk_mq_realloc_hw_ctxs(set, q);
3193 if (!q->nr_hw_queues)
3196 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3197 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3201 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3202 if (set->nr_maps > HCTX_TYPE_POLL &&
3203 set->map[HCTX_TYPE_POLL].nr_queues)
3204 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3206 q->sg_reserved_size = INT_MAX;
3208 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3209 INIT_LIST_HEAD(&q->requeue_list);
3210 spin_lock_init(&q->requeue_lock);
3212 q->nr_requests = set->queue_depth;
3215 * Default to classic polling
3217 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3219 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3220 blk_mq_add_queue_tag_set(set, q);
3221 blk_mq_map_swqueue(q);
3224 elevator_init_mq(q);
3229 kfree(q->queue_hw_ctx);
3230 q->nr_hw_queues = 0;
3231 blk_mq_sysfs_deinit(q);
3233 blk_stat_free_callback(q->poll_cb);
3237 return ERR_PTR(-ENOMEM);
3239 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3241 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3242 void blk_mq_exit_queue(struct request_queue *q)
3244 struct blk_mq_tag_set *set = q->tag_set;
3246 blk_mq_del_queue_tag_set(q);
3247 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3250 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3254 for (i = 0; i < set->nr_hw_queues; i++)
3255 if (!__blk_mq_alloc_map_and_request(set, i))
3262 blk_mq_free_map_and_requests(set, i);
3268 * Allocate the request maps associated with this tag_set. Note that this
3269 * may reduce the depth asked for, if memory is tight. set->queue_depth
3270 * will be updated to reflect the allocated depth.
3272 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3277 depth = set->queue_depth;
3279 err = __blk_mq_alloc_rq_maps(set);
3283 set->queue_depth >>= 1;
3284 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3288 } while (set->queue_depth);
3290 if (!set->queue_depth || err) {
3291 pr_err("blk-mq: failed to allocate request map\n");
3295 if (depth != set->queue_depth)
3296 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3297 depth, set->queue_depth);
3302 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3305 * blk_mq_map_queues() and multiple .map_queues() implementations
3306 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3307 * number of hardware queues.
3309 if (set->nr_maps == 1)
3310 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3312 if (set->ops->map_queues && !is_kdump_kernel()) {
3316 * transport .map_queues is usually done in the following
3319 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3320 * mask = get_cpu_mask(queue)
3321 * for_each_cpu(cpu, mask)
3322 * set->map[x].mq_map[cpu] = queue;
3325 * When we need to remap, the table has to be cleared for
3326 * killing stale mapping since one CPU may not be mapped
3329 for (i = 0; i < set->nr_maps; i++)
3330 blk_mq_clear_mq_map(&set->map[i]);
3332 return set->ops->map_queues(set);
3334 BUG_ON(set->nr_maps > 1);
3335 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3339 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3340 int cur_nr_hw_queues, int new_nr_hw_queues)
3342 struct blk_mq_tags **new_tags;
3344 if (cur_nr_hw_queues >= new_nr_hw_queues)
3347 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3348 GFP_KERNEL, set->numa_node);
3353 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3354 sizeof(*set->tags));
3356 set->tags = new_tags;
3357 set->nr_hw_queues = new_nr_hw_queues;
3363 * Alloc a tag set to be associated with one or more request queues.
3364 * May fail with EINVAL for various error conditions. May adjust the
3365 * requested depth down, if it's too large. In that case, the set
3366 * value will be stored in set->queue_depth.
3368 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3372 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3374 if (!set->nr_hw_queues)
3376 if (!set->queue_depth)
3378 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3381 if (!set->ops->queue_rq)
3384 if (!set->ops->get_budget ^ !set->ops->put_budget)
3387 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3388 pr_info("blk-mq: reduced tag depth to %u\n",
3390 set->queue_depth = BLK_MQ_MAX_DEPTH;
3395 else if (set->nr_maps > HCTX_MAX_TYPES)
3399 * If a crashdump is active, then we are potentially in a very
3400 * memory constrained environment. Limit us to 1 queue and
3401 * 64 tags to prevent using too much memory.
3403 if (is_kdump_kernel()) {
3404 set->nr_hw_queues = 1;
3406 set->queue_depth = min(64U, set->queue_depth);
3409 * There is no use for more h/w queues than cpus if we just have
3412 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3413 set->nr_hw_queues = nr_cpu_ids;
3415 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3419 for (i = 0; i < set->nr_maps; i++) {
3420 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3421 sizeof(set->map[i].mq_map[0]),
3422 GFP_KERNEL, set->numa_node);
3423 if (!set->map[i].mq_map)
3424 goto out_free_mq_map;
3425 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3428 ret = blk_mq_update_queue_map(set);
3430 goto out_free_mq_map;
3432 ret = blk_mq_alloc_map_and_requests(set);
3434 goto out_free_mq_map;
3436 mutex_init(&set->tag_list_lock);
3437 INIT_LIST_HEAD(&set->tag_list);
3442 for (i = 0; i < set->nr_maps; i++) {
3443 kfree(set->map[i].mq_map);
3444 set->map[i].mq_map = NULL;
3450 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3452 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3456 for (i = 0; i < set->nr_hw_queues; i++)
3457 blk_mq_free_map_and_requests(set, i);
3459 for (j = 0; j < set->nr_maps; j++) {
3460 kfree(set->map[j].mq_map);
3461 set->map[j].mq_map = NULL;
3467 EXPORT_SYMBOL(blk_mq_free_tag_set);
3469 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3471 struct blk_mq_tag_set *set = q->tag_set;
3472 struct blk_mq_hw_ctx *hctx;
3478 if (q->nr_requests == nr)
3481 blk_mq_freeze_queue(q);
3482 blk_mq_quiesce_queue(q);
3485 queue_for_each_hw_ctx(q, hctx, i) {
3489 * If we're using an MQ scheduler, just update the scheduler
3490 * queue depth. This is similar to what the old code would do.
3492 if (!hctx->sched_tags) {
3493 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3496 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3501 if (q->elevator && q->elevator->type->ops.depth_updated)
3502 q->elevator->type->ops.depth_updated(hctx);
3506 q->nr_requests = nr;
3508 blk_mq_unquiesce_queue(q);
3509 blk_mq_unfreeze_queue(q);
3515 * request_queue and elevator_type pair.
3516 * It is just used by __blk_mq_update_nr_hw_queues to cache
3517 * the elevator_type associated with a request_queue.
3519 struct blk_mq_qe_pair {
3520 struct list_head node;
3521 struct request_queue *q;
3522 struct elevator_type *type;
3526 * Cache the elevator_type in qe pair list and switch the
3527 * io scheduler to 'none'
3529 static bool blk_mq_elv_switch_none(struct list_head *head,
3530 struct request_queue *q)
3532 struct blk_mq_qe_pair *qe;
3537 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3541 INIT_LIST_HEAD(&qe->node);
3543 qe->type = q->elevator->type;
3544 list_add(&qe->node, head);
3546 mutex_lock(&q->sysfs_lock);
3548 * After elevator_switch_mq, the previous elevator_queue will be
3549 * released by elevator_release. The reference of the io scheduler
3550 * module get by elevator_get will also be put. So we need to get
3551 * a reference of the io scheduler module here to prevent it to be
3554 __module_get(qe->type->elevator_owner);
3555 elevator_switch_mq(q, NULL);
3556 mutex_unlock(&q->sysfs_lock);
3561 static void blk_mq_elv_switch_back(struct list_head *head,
3562 struct request_queue *q)
3564 struct blk_mq_qe_pair *qe;
3565 struct elevator_type *t = NULL;
3567 list_for_each_entry(qe, head, node)
3576 list_del(&qe->node);
3579 mutex_lock(&q->sysfs_lock);
3580 elevator_switch_mq(q, t);
3581 mutex_unlock(&q->sysfs_lock);
3584 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3587 struct request_queue *q;
3589 int prev_nr_hw_queues;
3591 lockdep_assert_held(&set->tag_list_lock);
3593 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3594 nr_hw_queues = nr_cpu_ids;
3595 if (nr_hw_queues < 1)
3597 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3600 list_for_each_entry(q, &set->tag_list, tag_set_list)
3601 blk_mq_freeze_queue(q);
3603 * Switch IO scheduler to 'none', cleaning up the data associated
3604 * with the previous scheduler. We will switch back once we are done
3605 * updating the new sw to hw queue mappings.
3607 list_for_each_entry(q, &set->tag_list, tag_set_list)
3608 if (!blk_mq_elv_switch_none(&head, q))
3611 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3612 blk_mq_debugfs_unregister_hctxs(q);
3613 blk_mq_sysfs_unregister(q);
3616 prev_nr_hw_queues = set->nr_hw_queues;
3617 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3621 set->nr_hw_queues = nr_hw_queues;
3623 blk_mq_update_queue_map(set);
3624 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3625 blk_mq_realloc_hw_ctxs(set, q);
3626 if (q->nr_hw_queues != set->nr_hw_queues) {
3627 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3628 nr_hw_queues, prev_nr_hw_queues);
3629 set->nr_hw_queues = prev_nr_hw_queues;
3630 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3633 blk_mq_map_swqueue(q);
3637 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3638 blk_mq_sysfs_register(q);
3639 blk_mq_debugfs_register_hctxs(q);
3643 list_for_each_entry(q, &set->tag_list, tag_set_list)
3644 blk_mq_elv_switch_back(&head, q);
3646 list_for_each_entry(q, &set->tag_list, tag_set_list)
3647 blk_mq_unfreeze_queue(q);
3650 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3652 mutex_lock(&set->tag_list_lock);
3653 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3654 mutex_unlock(&set->tag_list_lock);
3656 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3658 /* Enable polling stats and return whether they were already enabled. */
3659 static bool blk_poll_stats_enable(struct request_queue *q)
3661 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3662 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3664 blk_stat_add_callback(q, q->poll_cb);
3668 static void blk_mq_poll_stats_start(struct request_queue *q)
3671 * We don't arm the callback if polling stats are not enabled or the
3672 * callback is already active.
3674 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3675 blk_stat_is_active(q->poll_cb))
3678 blk_stat_activate_msecs(q->poll_cb, 100);
3681 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3683 struct request_queue *q = cb->data;
3686 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3687 if (cb->stat[bucket].nr_samples)
3688 q->poll_stat[bucket] = cb->stat[bucket];
3692 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3695 unsigned long ret = 0;
3699 * If stats collection isn't on, don't sleep but turn it on for
3702 if (!blk_poll_stats_enable(q))
3706 * As an optimistic guess, use half of the mean service time
3707 * for this type of request. We can (and should) make this smarter.
3708 * For instance, if the completion latencies are tight, we can
3709 * get closer than just half the mean. This is especially
3710 * important on devices where the completion latencies are longer
3711 * than ~10 usec. We do use the stats for the relevant IO size
3712 * if available which does lead to better estimates.
3714 bucket = blk_mq_poll_stats_bkt(rq);
3718 if (q->poll_stat[bucket].nr_samples)
3719 ret = (q->poll_stat[bucket].mean + 1) / 2;
3724 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3727 struct hrtimer_sleeper hs;
3728 enum hrtimer_mode mode;
3732 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3736 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3738 * 0: use half of prev avg
3739 * >0: use this specific value
3741 if (q->poll_nsec > 0)
3742 nsecs = q->poll_nsec;
3744 nsecs = blk_mq_poll_nsecs(q, rq);
3749 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3752 * This will be replaced with the stats tracking code, using
3753 * 'avg_completion_time / 2' as the pre-sleep target.
3757 mode = HRTIMER_MODE_REL;
3758 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3759 hrtimer_set_expires(&hs.timer, kt);
3762 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3764 set_current_state(TASK_UNINTERRUPTIBLE);
3765 hrtimer_sleeper_start_expires(&hs, mode);
3768 hrtimer_cancel(&hs.timer);
3769 mode = HRTIMER_MODE_ABS;
3770 } while (hs.task && !signal_pending(current));
3772 __set_current_state(TASK_RUNNING);
3773 destroy_hrtimer_on_stack(&hs.timer);
3777 static bool blk_mq_poll_hybrid(struct request_queue *q,
3778 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3782 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3785 if (!blk_qc_t_is_internal(cookie))
3786 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3788 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3790 * With scheduling, if the request has completed, we'll
3791 * get a NULL return here, as we clear the sched tag when
3792 * that happens. The request still remains valid, like always,
3793 * so we should be safe with just the NULL check.
3799 return blk_mq_poll_hybrid_sleep(q, rq);
3803 * blk_poll - poll for IO completions
3805 * @cookie: cookie passed back at IO submission time
3806 * @spin: whether to spin for completions
3809 * Poll for completions on the passed in queue. Returns number of
3810 * completed entries found. If @spin is true, then blk_poll will continue
3811 * looping until at least one completion is found, unless the task is
3812 * otherwise marked running (or we need to reschedule).
3814 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3816 struct blk_mq_hw_ctx *hctx;
3819 if (!blk_qc_t_valid(cookie) ||
3820 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3824 blk_flush_plug_list(current->plug, false);
3826 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3829 * If we sleep, have the caller restart the poll loop to reset
3830 * the state. Like for the other success return cases, the
3831 * caller is responsible for checking if the IO completed. If
3832 * the IO isn't complete, we'll get called again and will go
3833 * straight to the busy poll loop.
3835 if (blk_mq_poll_hybrid(q, hctx, cookie))
3838 hctx->poll_considered++;
3840 state = current->state;
3844 hctx->poll_invoked++;
3846 ret = q->mq_ops->poll(hctx);
3848 hctx->poll_success++;
3849 __set_current_state(TASK_RUNNING);
3853 if (signal_pending_state(state, current))
3854 __set_current_state(TASK_RUNNING);
3856 if (current->state == TASK_RUNNING)
3858 if (ret < 0 || !spin)
3861 } while (!need_resched());
3863 __set_current_state(TASK_RUNNING);
3866 EXPORT_SYMBOL_GPL(blk_poll);
3868 unsigned int blk_mq_rq_cpu(struct request *rq)
3870 return rq->mq_ctx->cpu;
3872 EXPORT_SYMBOL(blk_mq_rq_cpu);
3874 static int __init blk_mq_init(void)
3878 for_each_possible_cpu(i)
3879 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3880 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3882 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3883 "block/softirq:dead", NULL,
3884 blk_softirq_cpu_dead);
3885 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3886 blk_mq_hctx_notify_dead);
3887 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3888 blk_mq_hctx_notify_online,
3889 blk_mq_hctx_notify_offline);
3892 subsys_initcall(blk_mq_init);