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 block_device *part;
99 unsigned int inflight[2];
102 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
103 struct request *rq, void *priv,
106 struct mq_inflight *mi = priv;
108 if (rq->part == mi->part && blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
109 mi->inflight[rq_data_dir(rq)]++;
114 unsigned int blk_mq_in_flight(struct request_queue *q,
115 struct block_device *part)
117 struct mq_inflight mi = { .part = part };
119 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
121 return mi.inflight[0] + mi.inflight[1];
124 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
125 unsigned int inflight[2])
127 struct mq_inflight mi = { .part = part };
129 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
130 inflight[0] = mi.inflight[0];
131 inflight[1] = mi.inflight[1];
134 void blk_freeze_queue_start(struct request_queue *q)
136 mutex_lock(&q->mq_freeze_lock);
137 if (++q->mq_freeze_depth == 1) {
138 percpu_ref_kill(&q->q_usage_counter);
139 mutex_unlock(&q->mq_freeze_lock);
141 blk_mq_run_hw_queues(q, false);
143 mutex_unlock(&q->mq_freeze_lock);
146 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
148 void blk_mq_freeze_queue_wait(struct request_queue *q)
150 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
152 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
154 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
155 unsigned long timeout)
157 return wait_event_timeout(q->mq_freeze_wq,
158 percpu_ref_is_zero(&q->q_usage_counter),
161 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
164 * Guarantee no request is in use, so we can change any data structure of
165 * the queue afterward.
167 void blk_freeze_queue(struct request_queue *q)
170 * In the !blk_mq case we are only calling this to kill the
171 * q_usage_counter, otherwise this increases the freeze depth
172 * and waits for it to return to zero. For this reason there is
173 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
174 * exported to drivers as the only user for unfreeze is blk_mq.
176 blk_freeze_queue_start(q);
177 blk_mq_freeze_queue_wait(q);
180 void blk_mq_freeze_queue(struct request_queue *q)
183 * ...just an alias to keep freeze and unfreeze actions balanced
184 * in the blk_mq_* namespace
188 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
190 void blk_mq_unfreeze_queue(struct request_queue *q)
192 mutex_lock(&q->mq_freeze_lock);
193 q->mq_freeze_depth--;
194 WARN_ON_ONCE(q->mq_freeze_depth < 0);
195 if (!q->mq_freeze_depth) {
196 percpu_ref_resurrect(&q->q_usage_counter);
197 wake_up_all(&q->mq_freeze_wq);
199 mutex_unlock(&q->mq_freeze_lock);
201 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
204 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
205 * mpt3sas driver such that this function can be removed.
207 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
209 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
211 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
214 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
217 * Note: this function does not prevent that the struct request end_io()
218 * callback function is invoked. Once this function is returned, we make
219 * sure no dispatch can happen until the queue is unquiesced via
220 * blk_mq_unquiesce_queue().
222 void blk_mq_quiesce_queue(struct request_queue *q)
224 struct blk_mq_hw_ctx *hctx;
228 blk_mq_quiesce_queue_nowait(q);
230 queue_for_each_hw_ctx(q, hctx, i) {
231 if (hctx->flags & BLK_MQ_F_BLOCKING)
232 synchronize_srcu(hctx->srcu);
239 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
242 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
245 * This function recovers queue into the state before quiescing
246 * which is done by blk_mq_quiesce_queue.
248 void blk_mq_unquiesce_queue(struct request_queue *q)
250 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
252 /* dispatch requests which are inserted during quiescing */
253 blk_mq_run_hw_queues(q, true);
255 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
257 void blk_mq_wake_waiters(struct request_queue *q)
259 struct blk_mq_hw_ctx *hctx;
262 queue_for_each_hw_ctx(q, hctx, i)
263 if (blk_mq_hw_queue_mapped(hctx))
264 blk_mq_tag_wakeup_all(hctx->tags, true);
268 * Only need start/end time stamping if we have iostat or
269 * blk stats enabled, or using an IO scheduler.
271 static inline bool blk_mq_need_time_stamp(struct request *rq)
273 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
276 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
277 unsigned int tag, u64 alloc_time_ns)
279 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
280 struct request *rq = tags->static_rqs[tag];
282 if (data->q->elevator) {
283 rq->tag = BLK_MQ_NO_TAG;
284 rq->internal_tag = tag;
287 rq->internal_tag = BLK_MQ_NO_TAG;
290 /* csd/requeue_work/fifo_time is initialized before use */
292 rq->mq_ctx = data->ctx;
293 rq->mq_hctx = data->hctx;
295 rq->cmd_flags = data->cmd_flags;
296 if (data->flags & BLK_MQ_REQ_PREEMPT)
297 rq->rq_flags |= RQF_PREEMPT;
298 if (blk_queue_io_stat(data->q))
299 rq->rq_flags |= RQF_IO_STAT;
300 INIT_LIST_HEAD(&rq->queuelist);
301 INIT_HLIST_NODE(&rq->hash);
302 RB_CLEAR_NODE(&rq->rb_node);
305 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
306 rq->alloc_time_ns = alloc_time_ns;
308 if (blk_mq_need_time_stamp(rq))
309 rq->start_time_ns = ktime_get_ns();
311 rq->start_time_ns = 0;
312 rq->io_start_time_ns = 0;
313 rq->stats_sectors = 0;
314 rq->nr_phys_segments = 0;
315 #if defined(CONFIG_BLK_DEV_INTEGRITY)
316 rq->nr_integrity_segments = 0;
318 blk_crypto_rq_set_defaults(rq);
319 /* tag was already set */
320 WRITE_ONCE(rq->deadline, 0);
325 rq->end_io_data = NULL;
327 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
328 refcount_set(&rq->ref, 1);
330 if (!op_is_flush(data->cmd_flags)) {
331 struct elevator_queue *e = data->q->elevator;
334 if (e && e->type->ops.prepare_request) {
335 if (e->type->icq_cache)
336 blk_mq_sched_assign_ioc(rq);
338 e->type->ops.prepare_request(rq);
339 rq->rq_flags |= RQF_ELVPRIV;
343 data->hctx->queued++;
347 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
349 struct request_queue *q = data->q;
350 struct elevator_queue *e = q->elevator;
351 u64 alloc_time_ns = 0;
354 /* alloc_time includes depth and tag waits */
355 if (blk_queue_rq_alloc_time(q))
356 alloc_time_ns = ktime_get_ns();
358 if (data->cmd_flags & REQ_NOWAIT)
359 data->flags |= BLK_MQ_REQ_NOWAIT;
363 * Flush requests are special and go directly to the
364 * dispatch list. Don't include reserved tags in the
365 * limiting, as it isn't useful.
367 if (!op_is_flush(data->cmd_flags) &&
368 e->type->ops.limit_depth &&
369 !(data->flags & BLK_MQ_REQ_RESERVED))
370 e->type->ops.limit_depth(data->cmd_flags, data);
374 data->ctx = blk_mq_get_ctx(q);
375 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
377 blk_mq_tag_busy(data->hctx);
380 * Waiting allocations only fail because of an inactive hctx. In that
381 * case just retry the hctx assignment and tag allocation as CPU hotplug
382 * should have migrated us to an online CPU by now.
384 tag = blk_mq_get_tag(data);
385 if (tag == BLK_MQ_NO_TAG) {
386 if (data->flags & BLK_MQ_REQ_NOWAIT)
390 * Give up the CPU and sleep for a random short time to ensure
391 * that thread using a realtime scheduling class are migrated
392 * off the CPU, and thus off the hctx that is going away.
397 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
400 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
401 blk_mq_req_flags_t flags)
403 struct blk_mq_alloc_data data = {
411 ret = blk_queue_enter(q, flags);
415 rq = __blk_mq_alloc_request(&data);
419 rq->__sector = (sector_t) -1;
420 rq->bio = rq->biotail = NULL;
424 return ERR_PTR(-EWOULDBLOCK);
426 EXPORT_SYMBOL(blk_mq_alloc_request);
428 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
429 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
431 struct blk_mq_alloc_data data = {
436 u64 alloc_time_ns = 0;
441 /* alloc_time includes depth and tag waits */
442 if (blk_queue_rq_alloc_time(q))
443 alloc_time_ns = ktime_get_ns();
446 * If the tag allocator sleeps we could get an allocation for a
447 * different hardware context. No need to complicate the low level
448 * allocator for this for the rare use case of a command tied to
451 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
452 return ERR_PTR(-EINVAL);
454 if (hctx_idx >= q->nr_hw_queues)
455 return ERR_PTR(-EIO);
457 ret = blk_queue_enter(q, flags);
462 * Check if the hardware context is actually mapped to anything.
463 * If not tell the caller that it should skip this queue.
466 data.hctx = q->queue_hw_ctx[hctx_idx];
467 if (!blk_mq_hw_queue_mapped(data.hctx))
469 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
470 data.ctx = __blk_mq_get_ctx(q, cpu);
473 blk_mq_tag_busy(data.hctx);
476 tag = blk_mq_get_tag(&data);
477 if (tag == BLK_MQ_NO_TAG)
479 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
485 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
487 static void __blk_mq_free_request(struct request *rq)
489 struct request_queue *q = rq->q;
490 struct blk_mq_ctx *ctx = rq->mq_ctx;
491 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
492 const int sched_tag = rq->internal_tag;
494 blk_crypto_free_request(rq);
495 blk_pm_mark_last_busy(rq);
497 if (rq->tag != BLK_MQ_NO_TAG)
498 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
499 if (sched_tag != BLK_MQ_NO_TAG)
500 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
501 blk_mq_sched_restart(hctx);
505 void blk_mq_free_request(struct request *rq)
507 struct request_queue *q = rq->q;
508 struct elevator_queue *e = q->elevator;
509 struct blk_mq_ctx *ctx = rq->mq_ctx;
510 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
512 if (rq->rq_flags & RQF_ELVPRIV) {
513 if (e && e->type->ops.finish_request)
514 e->type->ops.finish_request(rq);
516 put_io_context(rq->elv.icq->ioc);
521 ctx->rq_completed[rq_is_sync(rq)]++;
522 if (rq->rq_flags & RQF_MQ_INFLIGHT)
523 __blk_mq_dec_active_requests(hctx);
525 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
526 laptop_io_completion(q->backing_dev_info);
530 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
531 if (refcount_dec_and_test(&rq->ref))
532 __blk_mq_free_request(rq);
534 EXPORT_SYMBOL_GPL(blk_mq_free_request);
536 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
540 if (blk_mq_need_time_stamp(rq))
541 now = ktime_get_ns();
543 if (rq->rq_flags & RQF_STATS) {
544 blk_mq_poll_stats_start(rq->q);
545 blk_stat_add(rq, now);
548 blk_mq_sched_completed_request(rq, now);
550 blk_account_io_done(rq, now);
553 rq_qos_done(rq->q, rq);
554 rq->end_io(rq, error);
556 blk_mq_free_request(rq);
559 EXPORT_SYMBOL(__blk_mq_end_request);
561 void blk_mq_end_request(struct request *rq, blk_status_t error)
563 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
565 __blk_mq_end_request(rq, error);
567 EXPORT_SYMBOL(blk_mq_end_request);
570 * Softirq action handler - move entries to local list and loop over them
571 * while passing them to the queue registered handler.
573 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
575 struct list_head *cpu_list, local_list;
578 cpu_list = this_cpu_ptr(&blk_cpu_done);
579 list_replace_init(cpu_list, &local_list);
582 while (!list_empty(&local_list)) {
585 rq = list_entry(local_list.next, struct request, ipi_list);
586 list_del_init(&rq->ipi_list);
587 rq->q->mq_ops->complete(rq);
591 static void blk_mq_trigger_softirq(struct request *rq)
593 struct list_head *list;
596 local_irq_save(flags);
597 list = this_cpu_ptr(&blk_cpu_done);
598 list_add_tail(&rq->ipi_list, list);
601 * If the list only contains our just added request, signal a raise of
602 * the softirq. If there are already entries there, someone already
603 * raised the irq but it hasn't run yet.
605 if (list->next == &rq->ipi_list)
606 raise_softirq_irqoff(BLOCK_SOFTIRQ);
607 local_irq_restore(flags);
610 static int blk_softirq_cpu_dead(unsigned int cpu)
613 * If a CPU goes away, splice its entries to the current CPU
614 * and trigger a run of the softirq
617 list_splice_init(&per_cpu(blk_cpu_done, cpu),
618 this_cpu_ptr(&blk_cpu_done));
619 raise_softirq_irqoff(BLOCK_SOFTIRQ);
626 static void __blk_mq_complete_request_remote(void *data)
628 struct request *rq = data;
631 * For most of single queue controllers, there is only one irq vector
632 * for handling I/O completion, and the only irq's affinity is set
633 * to all possible CPUs. On most of ARCHs, this affinity means the irq
634 * is handled on one specific CPU.
636 * So complete I/O requests in softirq context in case of single queue
637 * devices to avoid degrading I/O performance due to irqsoff latency.
639 if (rq->q->nr_hw_queues == 1)
640 blk_mq_trigger_softirq(rq);
642 rq->q->mq_ops->complete(rq);
645 static inline bool blk_mq_complete_need_ipi(struct request *rq)
647 int cpu = raw_smp_processor_id();
649 if (!IS_ENABLED(CONFIG_SMP) ||
650 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
653 /* same CPU or cache domain? Complete locally */
654 if (cpu == rq->mq_ctx->cpu ||
655 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
656 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
659 /* don't try to IPI to an offline CPU */
660 return cpu_online(rq->mq_ctx->cpu);
663 bool blk_mq_complete_request_remote(struct request *rq)
665 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
668 * For a polled request, always complete locallly, it's pointless
669 * to redirect the completion.
671 if (rq->cmd_flags & REQ_HIPRI)
674 if (blk_mq_complete_need_ipi(rq)) {
675 rq->csd.func = __blk_mq_complete_request_remote;
678 smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
680 if (rq->q->nr_hw_queues > 1)
682 blk_mq_trigger_softirq(rq);
687 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
690 * blk_mq_complete_request - end I/O on a request
691 * @rq: the request being processed
694 * Complete a request by scheduling the ->complete_rq operation.
696 void blk_mq_complete_request(struct request *rq)
698 if (!blk_mq_complete_request_remote(rq))
699 rq->q->mq_ops->complete(rq);
701 EXPORT_SYMBOL(blk_mq_complete_request);
703 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
704 __releases(hctx->srcu)
706 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
709 srcu_read_unlock(hctx->srcu, srcu_idx);
712 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
713 __acquires(hctx->srcu)
715 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
716 /* shut up gcc false positive */
720 *srcu_idx = srcu_read_lock(hctx->srcu);
724 * blk_mq_start_request - Start processing a request
725 * @rq: Pointer to request to be started
727 * Function used by device drivers to notify the block layer that a request
728 * is going to be processed now, so blk layer can do proper initializations
729 * such as starting the timeout timer.
731 void blk_mq_start_request(struct request *rq)
733 struct request_queue *q = rq->q;
735 trace_block_rq_issue(q, rq);
737 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
738 rq->io_start_time_ns = ktime_get_ns();
739 rq->stats_sectors = blk_rq_sectors(rq);
740 rq->rq_flags |= RQF_STATS;
744 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
747 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
749 #ifdef CONFIG_BLK_DEV_INTEGRITY
750 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
751 q->integrity.profile->prepare_fn(rq);
754 EXPORT_SYMBOL(blk_mq_start_request);
756 static void __blk_mq_requeue_request(struct request *rq)
758 struct request_queue *q = rq->q;
760 blk_mq_put_driver_tag(rq);
762 trace_block_rq_requeue(q, rq);
763 rq_qos_requeue(q, rq);
765 if (blk_mq_request_started(rq)) {
766 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
767 rq->rq_flags &= ~RQF_TIMED_OUT;
771 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
773 __blk_mq_requeue_request(rq);
775 /* this request will be re-inserted to io scheduler queue */
776 blk_mq_sched_requeue_request(rq);
778 BUG_ON(!list_empty(&rq->queuelist));
779 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
781 EXPORT_SYMBOL(blk_mq_requeue_request);
783 static void blk_mq_requeue_work(struct work_struct *work)
785 struct request_queue *q =
786 container_of(work, struct request_queue, requeue_work.work);
788 struct request *rq, *next;
790 spin_lock_irq(&q->requeue_lock);
791 list_splice_init(&q->requeue_list, &rq_list);
792 spin_unlock_irq(&q->requeue_lock);
794 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
795 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
798 rq->rq_flags &= ~RQF_SOFTBARRIER;
799 list_del_init(&rq->queuelist);
801 * If RQF_DONTPREP, rq has contained some driver specific
802 * data, so insert it to hctx dispatch list to avoid any
805 if (rq->rq_flags & RQF_DONTPREP)
806 blk_mq_request_bypass_insert(rq, false, false);
808 blk_mq_sched_insert_request(rq, true, false, false);
811 while (!list_empty(&rq_list)) {
812 rq = list_entry(rq_list.next, struct request, queuelist);
813 list_del_init(&rq->queuelist);
814 blk_mq_sched_insert_request(rq, false, false, false);
817 blk_mq_run_hw_queues(q, false);
820 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
821 bool kick_requeue_list)
823 struct request_queue *q = rq->q;
827 * We abuse this flag that is otherwise used by the I/O scheduler to
828 * request head insertion from the workqueue.
830 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
832 spin_lock_irqsave(&q->requeue_lock, flags);
834 rq->rq_flags |= RQF_SOFTBARRIER;
835 list_add(&rq->queuelist, &q->requeue_list);
837 list_add_tail(&rq->queuelist, &q->requeue_list);
839 spin_unlock_irqrestore(&q->requeue_lock, flags);
841 if (kick_requeue_list)
842 blk_mq_kick_requeue_list(q);
845 void blk_mq_kick_requeue_list(struct request_queue *q)
847 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
849 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
851 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
854 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
855 msecs_to_jiffies(msecs));
857 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
859 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
861 if (tag < tags->nr_tags) {
862 prefetch(tags->rqs[tag]);
863 return tags->rqs[tag];
868 EXPORT_SYMBOL(blk_mq_tag_to_rq);
870 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
871 void *priv, bool reserved)
874 * If we find a request that isn't idle and the queue matches,
875 * we know the queue is busy. Return false to stop the iteration.
877 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
887 bool blk_mq_queue_inflight(struct request_queue *q)
891 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
894 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
896 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
898 req->rq_flags |= RQF_TIMED_OUT;
899 if (req->q->mq_ops->timeout) {
900 enum blk_eh_timer_return ret;
902 ret = req->q->mq_ops->timeout(req, reserved);
903 if (ret == BLK_EH_DONE)
905 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
911 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
913 unsigned long deadline;
915 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
917 if (rq->rq_flags & RQF_TIMED_OUT)
920 deadline = READ_ONCE(rq->deadline);
921 if (time_after_eq(jiffies, deadline))
926 else if (time_after(*next, deadline))
931 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
932 struct request *rq, void *priv, bool reserved)
934 unsigned long *next = priv;
937 * Just do a quick check if it is expired before locking the request in
938 * so we're not unnecessarilly synchronizing across CPUs.
940 if (!blk_mq_req_expired(rq, next))
944 * We have reason to believe the request may be expired. Take a
945 * reference on the request to lock this request lifetime into its
946 * currently allocated context to prevent it from being reallocated in
947 * the event the completion by-passes this timeout handler.
949 * If the reference was already released, then the driver beat the
950 * timeout handler to posting a natural completion.
952 if (!refcount_inc_not_zero(&rq->ref))
956 * The request is now locked and cannot be reallocated underneath the
957 * timeout handler's processing. Re-verify this exact request is truly
958 * expired; if it is not expired, then the request was completed and
959 * reallocated as a new request.
961 if (blk_mq_req_expired(rq, next))
962 blk_mq_rq_timed_out(rq, reserved);
964 if (is_flush_rq(rq, hctx))
966 else if (refcount_dec_and_test(&rq->ref))
967 __blk_mq_free_request(rq);
972 static void blk_mq_timeout_work(struct work_struct *work)
974 struct request_queue *q =
975 container_of(work, struct request_queue, timeout_work);
976 unsigned long next = 0;
977 struct blk_mq_hw_ctx *hctx;
980 /* A deadlock might occur if a request is stuck requiring a
981 * timeout at the same time a queue freeze is waiting
982 * completion, since the timeout code would not be able to
983 * acquire the queue reference here.
985 * That's why we don't use blk_queue_enter here; instead, we use
986 * percpu_ref_tryget directly, because we need to be able to
987 * obtain a reference even in the short window between the queue
988 * starting to freeze, by dropping the first reference in
989 * blk_freeze_queue_start, and the moment the last request is
990 * consumed, marked by the instant q_usage_counter reaches
993 if (!percpu_ref_tryget(&q->q_usage_counter))
996 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
999 mod_timer(&q->timeout, next);
1002 * Request timeouts are handled as a forward rolling timer. If
1003 * we end up here it means that no requests are pending and
1004 * also that no request has been pending for a while. Mark
1005 * each hctx as idle.
1007 queue_for_each_hw_ctx(q, hctx, i) {
1008 /* the hctx may be unmapped, so check it here */
1009 if (blk_mq_hw_queue_mapped(hctx))
1010 blk_mq_tag_idle(hctx);
1016 struct flush_busy_ctx_data {
1017 struct blk_mq_hw_ctx *hctx;
1018 struct list_head *list;
1021 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1023 struct flush_busy_ctx_data *flush_data = data;
1024 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1025 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1026 enum hctx_type type = hctx->type;
1028 spin_lock(&ctx->lock);
1029 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1030 sbitmap_clear_bit(sb, bitnr);
1031 spin_unlock(&ctx->lock);
1036 * Process software queues that have been marked busy, splicing them
1037 * to the for-dispatch
1039 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1041 struct flush_busy_ctx_data data = {
1046 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1048 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1050 struct dispatch_rq_data {
1051 struct blk_mq_hw_ctx *hctx;
1055 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1058 struct dispatch_rq_data *dispatch_data = data;
1059 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1060 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1061 enum hctx_type type = hctx->type;
1063 spin_lock(&ctx->lock);
1064 if (!list_empty(&ctx->rq_lists[type])) {
1065 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1066 list_del_init(&dispatch_data->rq->queuelist);
1067 if (list_empty(&ctx->rq_lists[type]))
1068 sbitmap_clear_bit(sb, bitnr);
1070 spin_unlock(&ctx->lock);
1072 return !dispatch_data->rq;
1075 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1076 struct blk_mq_ctx *start)
1078 unsigned off = start ? start->index_hw[hctx->type] : 0;
1079 struct dispatch_rq_data data = {
1084 __sbitmap_for_each_set(&hctx->ctx_map, off,
1085 dispatch_rq_from_ctx, &data);
1090 static inline unsigned int queued_to_index(unsigned int queued)
1095 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1098 static bool __blk_mq_get_driver_tag(struct request *rq)
1100 struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags;
1101 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1104 blk_mq_tag_busy(rq->mq_hctx);
1106 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1107 bt = rq->mq_hctx->tags->breserved_tags;
1110 if (!hctx_may_queue(rq->mq_hctx, bt))
1114 tag = __sbitmap_queue_get(bt);
1115 if (tag == BLK_MQ_NO_TAG)
1118 rq->tag = tag + tag_offset;
1122 static bool blk_mq_get_driver_tag(struct request *rq)
1124 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1126 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1129 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1130 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1131 rq->rq_flags |= RQF_MQ_INFLIGHT;
1132 __blk_mq_inc_active_requests(hctx);
1134 hctx->tags->rqs[rq->tag] = rq;
1138 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1139 int flags, void *key)
1141 struct blk_mq_hw_ctx *hctx;
1143 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1145 spin_lock(&hctx->dispatch_wait_lock);
1146 if (!list_empty(&wait->entry)) {
1147 struct sbitmap_queue *sbq;
1149 list_del_init(&wait->entry);
1150 sbq = hctx->tags->bitmap_tags;
1151 atomic_dec(&sbq->ws_active);
1153 spin_unlock(&hctx->dispatch_wait_lock);
1155 blk_mq_run_hw_queue(hctx, true);
1160 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1161 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1162 * restart. For both cases, take care to check the condition again after
1163 * marking us as waiting.
1165 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1168 struct sbitmap_queue *sbq = hctx->tags->bitmap_tags;
1169 struct wait_queue_head *wq;
1170 wait_queue_entry_t *wait;
1173 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1174 blk_mq_sched_mark_restart_hctx(hctx);
1177 * It's possible that a tag was freed in the window between the
1178 * allocation failure and adding the hardware queue to the wait
1181 * Don't clear RESTART here, someone else could have set it.
1182 * At most this will cost an extra queue run.
1184 return blk_mq_get_driver_tag(rq);
1187 wait = &hctx->dispatch_wait;
1188 if (!list_empty_careful(&wait->entry))
1191 wq = &bt_wait_ptr(sbq, hctx)->wait;
1193 spin_lock_irq(&wq->lock);
1194 spin_lock(&hctx->dispatch_wait_lock);
1195 if (!list_empty(&wait->entry)) {
1196 spin_unlock(&hctx->dispatch_wait_lock);
1197 spin_unlock_irq(&wq->lock);
1201 atomic_inc(&sbq->ws_active);
1202 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1203 __add_wait_queue(wq, wait);
1206 * It's possible that a tag was freed in the window between the
1207 * allocation failure and adding the hardware queue to the wait
1210 ret = blk_mq_get_driver_tag(rq);
1212 spin_unlock(&hctx->dispatch_wait_lock);
1213 spin_unlock_irq(&wq->lock);
1218 * We got a tag, remove ourselves from the wait queue to ensure
1219 * someone else gets the wakeup.
1221 list_del_init(&wait->entry);
1222 atomic_dec(&sbq->ws_active);
1223 spin_unlock(&hctx->dispatch_wait_lock);
1224 spin_unlock_irq(&wq->lock);
1229 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1230 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1232 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1233 * - EWMA is one simple way to compute running average value
1234 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1235 * - take 4 as factor for avoiding to get too small(0) result, and this
1236 * factor doesn't matter because EWMA decreases exponentially
1238 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1242 if (hctx->queue->elevator)
1245 ewma = hctx->dispatch_busy;
1250 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1252 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1253 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1255 hctx->dispatch_busy = ewma;
1258 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1260 static void blk_mq_handle_dev_resource(struct request *rq,
1261 struct list_head *list)
1263 struct request *next =
1264 list_first_entry_or_null(list, struct request, queuelist);
1267 * If an I/O scheduler has been configured and we got a driver tag for
1268 * the next request already, free it.
1271 blk_mq_put_driver_tag(next);
1273 list_add(&rq->queuelist, list);
1274 __blk_mq_requeue_request(rq);
1277 static void blk_mq_handle_zone_resource(struct request *rq,
1278 struct list_head *zone_list)
1281 * If we end up here it is because we cannot dispatch a request to a
1282 * specific zone due to LLD level zone-write locking or other zone
1283 * related resource not being available. In this case, set the request
1284 * aside in zone_list for retrying it later.
1286 list_add(&rq->queuelist, zone_list);
1287 __blk_mq_requeue_request(rq);
1290 enum prep_dispatch {
1292 PREP_DISPATCH_NO_TAG,
1293 PREP_DISPATCH_NO_BUDGET,
1296 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1299 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1301 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1302 blk_mq_put_driver_tag(rq);
1303 return PREP_DISPATCH_NO_BUDGET;
1306 if (!blk_mq_get_driver_tag(rq)) {
1308 * The initial allocation attempt failed, so we need to
1309 * rerun the hardware queue when a tag is freed. The
1310 * waitqueue takes care of that. If the queue is run
1311 * before we add this entry back on the dispatch list,
1312 * we'll re-run it below.
1314 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1316 * All budgets not got from this function will be put
1317 * together during handling partial dispatch
1320 blk_mq_put_dispatch_budget(rq->q);
1321 return PREP_DISPATCH_NO_TAG;
1325 return PREP_DISPATCH_OK;
1328 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1329 static void blk_mq_release_budgets(struct request_queue *q,
1330 unsigned int nr_budgets)
1334 for (i = 0; i < nr_budgets; i++)
1335 blk_mq_put_dispatch_budget(q);
1339 * Returns true if we did some work AND can potentially do more.
1341 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1342 unsigned int nr_budgets)
1344 enum prep_dispatch prep;
1345 struct request_queue *q = hctx->queue;
1346 struct request *rq, *nxt;
1348 blk_status_t ret = BLK_STS_OK;
1349 LIST_HEAD(zone_list);
1351 if (list_empty(list))
1355 * Now process all the entries, sending them to the driver.
1357 errors = queued = 0;
1359 struct blk_mq_queue_data bd;
1361 rq = list_first_entry(list, struct request, queuelist);
1363 WARN_ON_ONCE(hctx != rq->mq_hctx);
1364 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1365 if (prep != PREP_DISPATCH_OK)
1368 list_del_init(&rq->queuelist);
1373 * Flag last if we have no more requests, or if we have more
1374 * but can't assign a driver tag to it.
1376 if (list_empty(list))
1379 nxt = list_first_entry(list, struct request, queuelist);
1380 bd.last = !blk_mq_get_driver_tag(nxt);
1384 * once the request is queued to lld, no need to cover the
1389 ret = q->mq_ops->queue_rq(hctx, &bd);
1394 case BLK_STS_RESOURCE:
1395 case BLK_STS_DEV_RESOURCE:
1396 blk_mq_handle_dev_resource(rq, list);
1398 case BLK_STS_ZONE_RESOURCE:
1400 * Move the request to zone_list and keep going through
1401 * the dispatch list to find more requests the drive can
1404 blk_mq_handle_zone_resource(rq, &zone_list);
1408 blk_mq_end_request(rq, BLK_STS_IOERR);
1410 } while (!list_empty(list));
1412 if (!list_empty(&zone_list))
1413 list_splice_tail_init(&zone_list, list);
1415 hctx->dispatched[queued_to_index(queued)]++;
1417 /* If we didn't flush the entire list, we could have told the driver
1418 * there was more coming, but that turned out to be a lie.
1420 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1421 q->mq_ops->commit_rqs(hctx);
1423 * Any items that need requeuing? Stuff them into hctx->dispatch,
1424 * that is where we will continue on next queue run.
1426 if (!list_empty(list)) {
1428 /* For non-shared tags, the RESTART check will suffice */
1429 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1430 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1431 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1433 blk_mq_release_budgets(q, nr_budgets);
1435 spin_lock(&hctx->lock);
1436 list_splice_tail_init(list, &hctx->dispatch);
1437 spin_unlock(&hctx->lock);
1440 * Order adding requests to hctx->dispatch and checking
1441 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1442 * in blk_mq_sched_restart(). Avoid restart code path to
1443 * miss the new added requests to hctx->dispatch, meantime
1444 * SCHED_RESTART is observed here.
1449 * If SCHED_RESTART was set by the caller of this function and
1450 * it is no longer set that means that it was cleared by another
1451 * thread and hence that a queue rerun is needed.
1453 * If 'no_tag' is set, that means that we failed getting
1454 * a driver tag with an I/O scheduler attached. If our dispatch
1455 * waitqueue is no longer active, ensure that we run the queue
1456 * AFTER adding our entries back to the list.
1458 * If no I/O scheduler has been configured it is possible that
1459 * the hardware queue got stopped and restarted before requests
1460 * were pushed back onto the dispatch list. Rerun the queue to
1461 * avoid starvation. Notes:
1462 * - blk_mq_run_hw_queue() checks whether or not a queue has
1463 * been stopped before rerunning a queue.
1464 * - Some but not all block drivers stop a queue before
1465 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1468 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1469 * bit is set, run queue after a delay to avoid IO stalls
1470 * that could otherwise occur if the queue is idle. We'll do
1471 * similar if we couldn't get budget and SCHED_RESTART is set.
1473 needs_restart = blk_mq_sched_needs_restart(hctx);
1474 if (!needs_restart ||
1475 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1476 blk_mq_run_hw_queue(hctx, true);
1477 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1479 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1481 blk_mq_update_dispatch_busy(hctx, true);
1484 blk_mq_update_dispatch_busy(hctx, false);
1486 return (queued + errors) != 0;
1490 * __blk_mq_run_hw_queue - Run a hardware queue.
1491 * @hctx: Pointer to the hardware queue to run.
1493 * Send pending requests to the hardware.
1495 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1500 * We should be running this queue from one of the CPUs that
1503 * There are at least two related races now between setting
1504 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1505 * __blk_mq_run_hw_queue():
1507 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1508 * but later it becomes online, then this warning is harmless
1511 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1512 * but later it becomes offline, then the warning can't be
1513 * triggered, and we depend on blk-mq timeout handler to
1514 * handle dispatched requests to this hctx
1516 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1517 cpu_online(hctx->next_cpu)) {
1518 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1519 raw_smp_processor_id(),
1520 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1525 * We can't run the queue inline with ints disabled. Ensure that
1526 * we catch bad users of this early.
1528 WARN_ON_ONCE(in_interrupt());
1530 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1532 hctx_lock(hctx, &srcu_idx);
1533 blk_mq_sched_dispatch_requests(hctx);
1534 hctx_unlock(hctx, srcu_idx);
1537 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1539 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1541 if (cpu >= nr_cpu_ids)
1542 cpu = cpumask_first(hctx->cpumask);
1547 * It'd be great if the workqueue API had a way to pass
1548 * in a mask and had some smarts for more clever placement.
1549 * For now we just round-robin here, switching for every
1550 * BLK_MQ_CPU_WORK_BATCH queued items.
1552 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1555 int next_cpu = hctx->next_cpu;
1557 if (hctx->queue->nr_hw_queues == 1)
1558 return WORK_CPU_UNBOUND;
1560 if (--hctx->next_cpu_batch <= 0) {
1562 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1564 if (next_cpu >= nr_cpu_ids)
1565 next_cpu = blk_mq_first_mapped_cpu(hctx);
1566 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1570 * Do unbound schedule if we can't find a online CPU for this hctx,
1571 * and it should only happen in the path of handling CPU DEAD.
1573 if (!cpu_online(next_cpu)) {
1580 * Make sure to re-select CPU next time once after CPUs
1581 * in hctx->cpumask become online again.
1583 hctx->next_cpu = next_cpu;
1584 hctx->next_cpu_batch = 1;
1585 return WORK_CPU_UNBOUND;
1588 hctx->next_cpu = next_cpu;
1593 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1594 * @hctx: Pointer to the hardware queue to run.
1595 * @async: If we want to run the queue asynchronously.
1596 * @msecs: Microseconds of delay to wait before running the queue.
1598 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1599 * with a delay of @msecs.
1601 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1602 unsigned long msecs)
1604 if (unlikely(blk_mq_hctx_stopped(hctx)))
1607 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1608 int cpu = get_cpu();
1609 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1610 __blk_mq_run_hw_queue(hctx);
1618 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1619 msecs_to_jiffies(msecs));
1623 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1624 * @hctx: Pointer to the hardware queue to run.
1625 * @msecs: Microseconds of delay to wait before running the queue.
1627 * Run a hardware queue asynchronously with a delay of @msecs.
1629 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1631 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1633 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1636 * blk_mq_run_hw_queue - Start to run a hardware queue.
1637 * @hctx: Pointer to the hardware queue to run.
1638 * @async: If we want to run the queue asynchronously.
1640 * Check if the request queue is not in a quiesced state and if there are
1641 * pending requests to be sent. If this is true, run the queue to send requests
1644 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1650 * When queue is quiesced, we may be switching io scheduler, or
1651 * updating nr_hw_queues, or other things, and we can't run queue
1652 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1654 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1657 hctx_lock(hctx, &srcu_idx);
1658 need_run = !blk_queue_quiesced(hctx->queue) &&
1659 blk_mq_hctx_has_pending(hctx);
1660 hctx_unlock(hctx, srcu_idx);
1663 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1665 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1668 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1669 * @q: Pointer to the request queue to run.
1670 * @async: If we want to run the queue asynchronously.
1672 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1674 struct blk_mq_hw_ctx *hctx;
1677 queue_for_each_hw_ctx(q, hctx, i) {
1678 if (blk_mq_hctx_stopped(hctx))
1681 blk_mq_run_hw_queue(hctx, async);
1684 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1687 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1688 * @q: Pointer to the request queue to run.
1689 * @msecs: Microseconds of delay to wait before running the queues.
1691 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1693 struct blk_mq_hw_ctx *hctx;
1696 queue_for_each_hw_ctx(q, hctx, i) {
1697 if (blk_mq_hctx_stopped(hctx))
1700 blk_mq_delay_run_hw_queue(hctx, msecs);
1703 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1706 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1707 * @q: request queue.
1709 * The caller is responsible for serializing this function against
1710 * blk_mq_{start,stop}_hw_queue().
1712 bool blk_mq_queue_stopped(struct request_queue *q)
1714 struct blk_mq_hw_ctx *hctx;
1717 queue_for_each_hw_ctx(q, hctx, i)
1718 if (blk_mq_hctx_stopped(hctx))
1723 EXPORT_SYMBOL(blk_mq_queue_stopped);
1726 * This function is often used for pausing .queue_rq() by driver when
1727 * there isn't enough resource or some conditions aren't satisfied, and
1728 * BLK_STS_RESOURCE is usually returned.
1730 * We do not guarantee that dispatch can be drained or blocked
1731 * after blk_mq_stop_hw_queue() returns. Please use
1732 * blk_mq_quiesce_queue() for that requirement.
1734 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1736 cancel_delayed_work(&hctx->run_work);
1738 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1740 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1743 * This function is often used for pausing .queue_rq() by driver when
1744 * there isn't enough resource or some conditions aren't satisfied, and
1745 * BLK_STS_RESOURCE is usually returned.
1747 * We do not guarantee that dispatch can be drained or blocked
1748 * after blk_mq_stop_hw_queues() returns. Please use
1749 * blk_mq_quiesce_queue() for that requirement.
1751 void blk_mq_stop_hw_queues(struct request_queue *q)
1753 struct blk_mq_hw_ctx *hctx;
1756 queue_for_each_hw_ctx(q, hctx, i)
1757 blk_mq_stop_hw_queue(hctx);
1759 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1761 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1763 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1765 blk_mq_run_hw_queue(hctx, false);
1767 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1769 void blk_mq_start_hw_queues(struct request_queue *q)
1771 struct blk_mq_hw_ctx *hctx;
1774 queue_for_each_hw_ctx(q, hctx, i)
1775 blk_mq_start_hw_queue(hctx);
1777 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1779 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1781 if (!blk_mq_hctx_stopped(hctx))
1784 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1785 blk_mq_run_hw_queue(hctx, async);
1787 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1789 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1791 struct blk_mq_hw_ctx *hctx;
1794 queue_for_each_hw_ctx(q, hctx, i)
1795 blk_mq_start_stopped_hw_queue(hctx, async);
1797 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1799 static void blk_mq_run_work_fn(struct work_struct *work)
1801 struct blk_mq_hw_ctx *hctx;
1803 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1806 * If we are stopped, don't run the queue.
1808 if (blk_mq_hctx_stopped(hctx))
1811 __blk_mq_run_hw_queue(hctx);
1814 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1818 struct blk_mq_ctx *ctx = rq->mq_ctx;
1819 enum hctx_type type = hctx->type;
1821 lockdep_assert_held(&ctx->lock);
1823 trace_block_rq_insert(hctx->queue, rq);
1826 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1828 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1831 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1834 struct blk_mq_ctx *ctx = rq->mq_ctx;
1836 lockdep_assert_held(&ctx->lock);
1838 __blk_mq_insert_req_list(hctx, rq, at_head);
1839 blk_mq_hctx_mark_pending(hctx, ctx);
1843 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1844 * @rq: Pointer to request to be inserted.
1845 * @at_head: true if the request should be inserted at the head of the list.
1846 * @run_queue: If we should run the hardware queue after inserting the request.
1848 * Should only be used carefully, when the caller knows we want to
1849 * bypass a potential IO scheduler on the target device.
1851 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1854 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1856 spin_lock(&hctx->lock);
1858 list_add(&rq->queuelist, &hctx->dispatch);
1860 list_add_tail(&rq->queuelist, &hctx->dispatch);
1861 spin_unlock(&hctx->lock);
1864 blk_mq_run_hw_queue(hctx, false);
1867 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1868 struct list_head *list)
1872 enum hctx_type type = hctx->type;
1875 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1878 list_for_each_entry(rq, list, queuelist) {
1879 BUG_ON(rq->mq_ctx != ctx);
1880 trace_block_rq_insert(hctx->queue, rq);
1883 spin_lock(&ctx->lock);
1884 list_splice_tail_init(list, &ctx->rq_lists[type]);
1885 blk_mq_hctx_mark_pending(hctx, ctx);
1886 spin_unlock(&ctx->lock);
1889 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1891 struct request *rqa = container_of(a, struct request, queuelist);
1892 struct request *rqb = container_of(b, struct request, queuelist);
1894 if (rqa->mq_ctx != rqb->mq_ctx)
1895 return rqa->mq_ctx > rqb->mq_ctx;
1896 if (rqa->mq_hctx != rqb->mq_hctx)
1897 return rqa->mq_hctx > rqb->mq_hctx;
1899 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1902 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1906 if (list_empty(&plug->mq_list))
1908 list_splice_init(&plug->mq_list, &list);
1910 if (plug->rq_count > 2 && plug->multiple_queues)
1911 list_sort(NULL, &list, plug_rq_cmp);
1916 struct list_head rq_list;
1917 struct request *rq, *head_rq = list_entry_rq(list.next);
1918 struct list_head *pos = &head_rq->queuelist; /* skip first */
1919 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1920 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1921 unsigned int depth = 1;
1923 list_for_each_continue(pos, &list) {
1924 rq = list_entry_rq(pos);
1926 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1931 list_cut_before(&rq_list, &list, pos);
1932 trace_block_unplug(head_rq->q, depth, !from_schedule);
1933 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1935 } while(!list_empty(&list));
1938 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1939 unsigned int nr_segs)
1943 if (bio->bi_opf & REQ_RAHEAD)
1944 rq->cmd_flags |= REQ_FAILFAST_MASK;
1946 rq->__sector = bio->bi_iter.bi_sector;
1947 rq->write_hint = bio->bi_write_hint;
1948 blk_rq_bio_prep(rq, bio, nr_segs);
1950 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1951 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1954 blk_account_io_start(rq);
1957 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1959 blk_qc_t *cookie, bool last)
1961 struct request_queue *q = rq->q;
1962 struct blk_mq_queue_data bd = {
1966 blk_qc_t new_cookie;
1969 new_cookie = request_to_qc_t(hctx, rq);
1972 * For OK queue, we are done. For error, caller may kill it.
1973 * Any other error (busy), just add it to our list as we
1974 * previously would have done.
1976 ret = q->mq_ops->queue_rq(hctx, &bd);
1979 blk_mq_update_dispatch_busy(hctx, false);
1980 *cookie = new_cookie;
1982 case BLK_STS_RESOURCE:
1983 case BLK_STS_DEV_RESOURCE:
1984 blk_mq_update_dispatch_busy(hctx, true);
1985 __blk_mq_requeue_request(rq);
1988 blk_mq_update_dispatch_busy(hctx, false);
1989 *cookie = BLK_QC_T_NONE;
1996 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1999 bool bypass_insert, bool last)
2001 struct request_queue *q = rq->q;
2002 bool run_queue = true;
2005 * RCU or SRCU read lock is needed before checking quiesced flag.
2007 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2008 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2009 * and avoid driver to try to dispatch again.
2011 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2013 bypass_insert = false;
2017 if (q->elevator && !bypass_insert)
2020 if (!blk_mq_get_dispatch_budget(q))
2023 if (!blk_mq_get_driver_tag(rq)) {
2024 blk_mq_put_dispatch_budget(q);
2028 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2031 return BLK_STS_RESOURCE;
2033 blk_mq_sched_insert_request(rq, false, run_queue, false);
2039 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2040 * @hctx: Pointer of the associated hardware queue.
2041 * @rq: Pointer to request to be sent.
2042 * @cookie: Request queue cookie.
2044 * If the device has enough resources to accept a new request now, send the
2045 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2046 * we can try send it another time in the future. Requests inserted at this
2047 * queue have higher priority.
2049 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2050 struct request *rq, blk_qc_t *cookie)
2055 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2057 hctx_lock(hctx, &srcu_idx);
2059 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2060 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2061 blk_mq_request_bypass_insert(rq, false, true);
2062 else if (ret != BLK_STS_OK)
2063 blk_mq_end_request(rq, ret);
2065 hctx_unlock(hctx, srcu_idx);
2068 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2072 blk_qc_t unused_cookie;
2073 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2075 hctx_lock(hctx, &srcu_idx);
2076 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2077 hctx_unlock(hctx, srcu_idx);
2082 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2083 struct list_head *list)
2088 while (!list_empty(list)) {
2090 struct request *rq = list_first_entry(list, struct request,
2093 list_del_init(&rq->queuelist);
2094 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2095 if (ret != BLK_STS_OK) {
2096 if (ret == BLK_STS_RESOURCE ||
2097 ret == BLK_STS_DEV_RESOURCE) {
2098 blk_mq_request_bypass_insert(rq, false,
2102 blk_mq_end_request(rq, ret);
2109 * If we didn't flush the entire list, we could have told
2110 * the driver there was more coming, but that turned out to
2113 if ((!list_empty(list) || errors) &&
2114 hctx->queue->mq_ops->commit_rqs && queued)
2115 hctx->queue->mq_ops->commit_rqs(hctx);
2118 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2120 list_add_tail(&rq->queuelist, &plug->mq_list);
2122 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2123 struct request *tmp;
2125 tmp = list_first_entry(&plug->mq_list, struct request,
2127 if (tmp->q != rq->q)
2128 plug->multiple_queues = true;
2133 * blk_mq_submit_bio - Create and send a request to block device.
2134 * @bio: Bio pointer.
2136 * Builds up a request structure from @q and @bio and send to the device. The
2137 * request may not be queued directly to hardware if:
2138 * * This request can be merged with another one
2139 * * We want to place request at plug queue for possible future merging
2140 * * There is an IO scheduler active at this queue
2142 * It will not queue the request if there is an error with the bio, or at the
2145 * Returns: Request queue cookie.
2147 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2149 struct request_queue *q = bio->bi_disk->queue;
2150 const int is_sync = op_is_sync(bio->bi_opf);
2151 const int is_flush_fua = op_is_flush(bio->bi_opf);
2152 struct blk_mq_alloc_data data = {
2156 struct blk_plug *plug;
2157 struct request *same_queue_rq = NULL;
2158 unsigned int nr_segs;
2162 blk_queue_bounce(q, &bio);
2163 __blk_queue_split(&bio, &nr_segs);
2165 if (!bio_integrity_prep(bio))
2168 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2169 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2172 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2175 rq_qos_throttle(q, bio);
2177 data.cmd_flags = bio->bi_opf;
2178 rq = __blk_mq_alloc_request(&data);
2179 if (unlikely(!rq)) {
2180 rq_qos_cleanup(q, bio);
2181 if (bio->bi_opf & REQ_NOWAIT)
2182 bio_wouldblock_error(bio);
2186 trace_block_getrq(q, bio, bio->bi_opf);
2188 rq_qos_track(q, rq, bio);
2190 cookie = request_to_qc_t(data.hctx, rq);
2192 blk_mq_bio_to_request(rq, bio, nr_segs);
2194 ret = blk_crypto_init_request(rq);
2195 if (ret != BLK_STS_OK) {
2196 bio->bi_status = ret;
2198 blk_mq_free_request(rq);
2199 return BLK_QC_T_NONE;
2202 plug = blk_mq_plug(q, bio);
2203 if (unlikely(is_flush_fua)) {
2204 /* Bypass scheduler for flush requests */
2205 blk_insert_flush(rq);
2206 blk_mq_run_hw_queue(data.hctx, true);
2207 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2208 !blk_queue_nonrot(q))) {
2210 * Use plugging if we have a ->commit_rqs() hook as well, as
2211 * we know the driver uses bd->last in a smart fashion.
2213 * Use normal plugging if this disk is slow HDD, as sequential
2214 * IO may benefit a lot from plug merging.
2216 unsigned int request_count = plug->rq_count;
2217 struct request *last = NULL;
2220 trace_block_plug(q);
2222 last = list_entry_rq(plug->mq_list.prev);
2224 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2225 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2226 blk_flush_plug_list(plug, false);
2227 trace_block_plug(q);
2230 blk_add_rq_to_plug(plug, rq);
2231 } else if (q->elevator) {
2232 /* Insert the request at the IO scheduler queue */
2233 blk_mq_sched_insert_request(rq, false, true, true);
2234 } else if (plug && !blk_queue_nomerges(q)) {
2236 * We do limited plugging. If the bio can be merged, do that.
2237 * Otherwise the existing request in the plug list will be
2238 * issued. So the plug list will have one request at most
2239 * The plug list might get flushed before this. If that happens,
2240 * the plug list is empty, and same_queue_rq is invalid.
2242 if (list_empty(&plug->mq_list))
2243 same_queue_rq = NULL;
2244 if (same_queue_rq) {
2245 list_del_init(&same_queue_rq->queuelist);
2248 blk_add_rq_to_plug(plug, rq);
2249 trace_block_plug(q);
2251 if (same_queue_rq) {
2252 data.hctx = same_queue_rq->mq_hctx;
2253 trace_block_unplug(q, 1, true);
2254 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2257 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2258 !data.hctx->dispatch_busy) {
2260 * There is no scheduler and we can try to send directly
2263 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2266 blk_mq_sched_insert_request(rq, false, true, true);
2272 return BLK_QC_T_NONE;
2275 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2276 unsigned int hctx_idx)
2280 if (tags->rqs && set->ops->exit_request) {
2283 for (i = 0; i < tags->nr_tags; i++) {
2284 struct request *rq = tags->static_rqs[i];
2288 set->ops->exit_request(set, rq, hctx_idx);
2289 tags->static_rqs[i] = NULL;
2293 while (!list_empty(&tags->page_list)) {
2294 page = list_first_entry(&tags->page_list, struct page, lru);
2295 list_del_init(&page->lru);
2297 * Remove kmemleak object previously allocated in
2298 * blk_mq_alloc_rqs().
2300 kmemleak_free(page_address(page));
2301 __free_pages(page, page->private);
2305 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags)
2309 kfree(tags->static_rqs);
2310 tags->static_rqs = NULL;
2312 blk_mq_free_tags(tags, flags);
2315 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2316 unsigned int hctx_idx,
2317 unsigned int nr_tags,
2318 unsigned int reserved_tags,
2321 struct blk_mq_tags *tags;
2324 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2325 if (node == NUMA_NO_NODE)
2326 node = set->numa_node;
2328 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags);
2332 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2333 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2336 blk_mq_free_tags(tags, flags);
2340 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2341 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2343 if (!tags->static_rqs) {
2345 blk_mq_free_tags(tags, flags);
2352 static size_t order_to_size(unsigned int order)
2354 return (size_t)PAGE_SIZE << order;
2357 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2358 unsigned int hctx_idx, int node)
2362 if (set->ops->init_request) {
2363 ret = set->ops->init_request(set, rq, hctx_idx, node);
2368 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2372 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2373 unsigned int hctx_idx, unsigned int depth)
2375 unsigned int i, j, entries_per_page, max_order = 4;
2376 size_t rq_size, left;
2379 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2380 if (node == NUMA_NO_NODE)
2381 node = set->numa_node;
2383 INIT_LIST_HEAD(&tags->page_list);
2386 * rq_size is the size of the request plus driver payload, rounded
2387 * to the cacheline size
2389 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2391 left = rq_size * depth;
2393 for (i = 0; i < depth; ) {
2394 int this_order = max_order;
2399 while (this_order && left < order_to_size(this_order - 1))
2403 page = alloc_pages_node(node,
2404 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2410 if (order_to_size(this_order) < rq_size)
2417 page->private = this_order;
2418 list_add_tail(&page->lru, &tags->page_list);
2420 p = page_address(page);
2422 * Allow kmemleak to scan these pages as they contain pointers
2423 * to additional allocations like via ops->init_request().
2425 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2426 entries_per_page = order_to_size(this_order) / rq_size;
2427 to_do = min(entries_per_page, depth - i);
2428 left -= to_do * rq_size;
2429 for (j = 0; j < to_do; j++) {
2430 struct request *rq = p;
2432 tags->static_rqs[i] = rq;
2433 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2434 tags->static_rqs[i] = NULL;
2445 blk_mq_free_rqs(set, tags, hctx_idx);
2449 struct rq_iter_data {
2450 struct blk_mq_hw_ctx *hctx;
2454 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2456 struct rq_iter_data *iter_data = data;
2458 if (rq->mq_hctx != iter_data->hctx)
2460 iter_data->has_rq = true;
2464 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2466 struct blk_mq_tags *tags = hctx->sched_tags ?
2467 hctx->sched_tags : hctx->tags;
2468 struct rq_iter_data data = {
2472 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2476 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2477 struct blk_mq_hw_ctx *hctx)
2479 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2481 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2486 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2488 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2489 struct blk_mq_hw_ctx, cpuhp_online);
2491 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2492 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2496 * Prevent new request from being allocated on the current hctx.
2498 * The smp_mb__after_atomic() Pairs with the implied barrier in
2499 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2500 * seen once we return from the tag allocator.
2502 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2503 smp_mb__after_atomic();
2506 * Try to grab a reference to the queue and wait for any outstanding
2507 * requests. If we could not grab a reference the queue has been
2508 * frozen and there are no requests.
2510 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2511 while (blk_mq_hctx_has_requests(hctx))
2513 percpu_ref_put(&hctx->queue->q_usage_counter);
2519 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2521 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2522 struct blk_mq_hw_ctx, cpuhp_online);
2524 if (cpumask_test_cpu(cpu, hctx->cpumask))
2525 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2530 * 'cpu' is going away. splice any existing rq_list entries from this
2531 * software queue to the hw queue dispatch list, and ensure that it
2534 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2536 struct blk_mq_hw_ctx *hctx;
2537 struct blk_mq_ctx *ctx;
2539 enum hctx_type type;
2541 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2542 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2545 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2548 spin_lock(&ctx->lock);
2549 if (!list_empty(&ctx->rq_lists[type])) {
2550 list_splice_init(&ctx->rq_lists[type], &tmp);
2551 blk_mq_hctx_clear_pending(hctx, ctx);
2553 spin_unlock(&ctx->lock);
2555 if (list_empty(&tmp))
2558 spin_lock(&hctx->lock);
2559 list_splice_tail_init(&tmp, &hctx->dispatch);
2560 spin_unlock(&hctx->lock);
2562 blk_mq_run_hw_queue(hctx, true);
2566 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2568 if (!(hctx->flags & BLK_MQ_F_STACKING))
2569 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2570 &hctx->cpuhp_online);
2571 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2575 /* hctx->ctxs will be freed in queue's release handler */
2576 static void blk_mq_exit_hctx(struct request_queue *q,
2577 struct blk_mq_tag_set *set,
2578 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2580 if (blk_mq_hw_queue_mapped(hctx))
2581 blk_mq_tag_idle(hctx);
2583 if (set->ops->exit_request)
2584 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2586 if (set->ops->exit_hctx)
2587 set->ops->exit_hctx(hctx, hctx_idx);
2589 blk_mq_remove_cpuhp(hctx);
2591 spin_lock(&q->unused_hctx_lock);
2592 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2593 spin_unlock(&q->unused_hctx_lock);
2596 static void blk_mq_exit_hw_queues(struct request_queue *q,
2597 struct blk_mq_tag_set *set, int nr_queue)
2599 struct blk_mq_hw_ctx *hctx;
2602 queue_for_each_hw_ctx(q, hctx, i) {
2605 blk_mq_debugfs_unregister_hctx(hctx);
2606 blk_mq_exit_hctx(q, set, hctx, i);
2610 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2612 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2614 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2615 __alignof__(struct blk_mq_hw_ctx)) !=
2616 sizeof(struct blk_mq_hw_ctx));
2618 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2619 hw_ctx_size += sizeof(struct srcu_struct);
2624 static int blk_mq_init_hctx(struct request_queue *q,
2625 struct blk_mq_tag_set *set,
2626 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2628 hctx->queue_num = hctx_idx;
2630 if (!(hctx->flags & BLK_MQ_F_STACKING))
2631 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2632 &hctx->cpuhp_online);
2633 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2635 hctx->tags = set->tags[hctx_idx];
2637 if (set->ops->init_hctx &&
2638 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2639 goto unregister_cpu_notifier;
2641 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2647 if (set->ops->exit_hctx)
2648 set->ops->exit_hctx(hctx, hctx_idx);
2649 unregister_cpu_notifier:
2650 blk_mq_remove_cpuhp(hctx);
2654 static struct blk_mq_hw_ctx *
2655 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2658 struct blk_mq_hw_ctx *hctx;
2659 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2661 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2663 goto fail_alloc_hctx;
2665 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2668 atomic_set(&hctx->nr_active, 0);
2669 atomic_set(&hctx->elevator_queued, 0);
2670 if (node == NUMA_NO_NODE)
2671 node = set->numa_node;
2672 hctx->numa_node = node;
2674 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2675 spin_lock_init(&hctx->lock);
2676 INIT_LIST_HEAD(&hctx->dispatch);
2678 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
2680 INIT_LIST_HEAD(&hctx->hctx_list);
2683 * Allocate space for all possible cpus to avoid allocation at
2686 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2691 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2696 spin_lock_init(&hctx->dispatch_wait_lock);
2697 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2698 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2700 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2704 if (hctx->flags & BLK_MQ_F_BLOCKING)
2705 init_srcu_struct(hctx->srcu);
2706 blk_mq_hctx_kobj_init(hctx);
2711 sbitmap_free(&hctx->ctx_map);
2715 free_cpumask_var(hctx->cpumask);
2722 static void blk_mq_init_cpu_queues(struct request_queue *q,
2723 unsigned int nr_hw_queues)
2725 struct blk_mq_tag_set *set = q->tag_set;
2728 for_each_possible_cpu(i) {
2729 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2730 struct blk_mq_hw_ctx *hctx;
2734 spin_lock_init(&__ctx->lock);
2735 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2736 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2741 * Set local node, IFF we have more than one hw queue. If
2742 * not, we remain on the home node of the device
2744 for (j = 0; j < set->nr_maps; j++) {
2745 hctx = blk_mq_map_queue_type(q, j, i);
2746 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2747 hctx->numa_node = cpu_to_node(i);
2752 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2755 unsigned int flags = set->flags;
2758 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2759 set->queue_depth, set->reserved_tags, flags);
2760 if (!set->tags[hctx_idx])
2763 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2768 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2769 set->tags[hctx_idx] = NULL;
2773 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2774 unsigned int hctx_idx)
2776 unsigned int flags = set->flags;
2778 if (set->tags && set->tags[hctx_idx]) {
2779 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2780 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2781 set->tags[hctx_idx] = NULL;
2785 static void blk_mq_map_swqueue(struct request_queue *q)
2787 unsigned int i, j, hctx_idx;
2788 struct blk_mq_hw_ctx *hctx;
2789 struct blk_mq_ctx *ctx;
2790 struct blk_mq_tag_set *set = q->tag_set;
2792 queue_for_each_hw_ctx(q, hctx, i) {
2793 cpumask_clear(hctx->cpumask);
2795 hctx->dispatch_from = NULL;
2799 * Map software to hardware queues.
2801 * If the cpu isn't present, the cpu is mapped to first hctx.
2803 for_each_possible_cpu(i) {
2805 ctx = per_cpu_ptr(q->queue_ctx, i);
2806 for (j = 0; j < set->nr_maps; j++) {
2807 if (!set->map[j].nr_queues) {
2808 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2809 HCTX_TYPE_DEFAULT, i);
2812 hctx_idx = set->map[j].mq_map[i];
2813 /* unmapped hw queue can be remapped after CPU topo changed */
2814 if (!set->tags[hctx_idx] &&
2815 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2817 * If tags initialization fail for some hctx,
2818 * that hctx won't be brought online. In this
2819 * case, remap the current ctx to hctx[0] which
2820 * is guaranteed to always have tags allocated
2822 set->map[j].mq_map[i] = 0;
2825 hctx = blk_mq_map_queue_type(q, j, i);
2826 ctx->hctxs[j] = hctx;
2828 * If the CPU is already set in the mask, then we've
2829 * mapped this one already. This can happen if
2830 * devices share queues across queue maps.
2832 if (cpumask_test_cpu(i, hctx->cpumask))
2835 cpumask_set_cpu(i, hctx->cpumask);
2837 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2838 hctx->ctxs[hctx->nr_ctx++] = ctx;
2841 * If the nr_ctx type overflows, we have exceeded the
2842 * amount of sw queues we can support.
2844 BUG_ON(!hctx->nr_ctx);
2847 for (; j < HCTX_MAX_TYPES; j++)
2848 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2849 HCTX_TYPE_DEFAULT, i);
2852 queue_for_each_hw_ctx(q, hctx, i) {
2854 * If no software queues are mapped to this hardware queue,
2855 * disable it and free the request entries.
2857 if (!hctx->nr_ctx) {
2858 /* Never unmap queue 0. We need it as a
2859 * fallback in case of a new remap fails
2862 if (i && set->tags[i])
2863 blk_mq_free_map_and_requests(set, i);
2869 hctx->tags = set->tags[i];
2870 WARN_ON(!hctx->tags);
2873 * Set the map size to the number of mapped software queues.
2874 * This is more accurate and more efficient than looping
2875 * over all possibly mapped software queues.
2877 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2880 * Initialize batch roundrobin counts
2882 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2883 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2888 * Caller needs to ensure that we're either frozen/quiesced, or that
2889 * the queue isn't live yet.
2891 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2893 struct blk_mq_hw_ctx *hctx;
2896 queue_for_each_hw_ctx(q, hctx, i) {
2898 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2900 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2904 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
2907 struct request_queue *q;
2909 lockdep_assert_held(&set->tag_list_lock);
2911 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2912 blk_mq_freeze_queue(q);
2913 queue_set_hctx_shared(q, shared);
2914 blk_mq_unfreeze_queue(q);
2918 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2920 struct blk_mq_tag_set *set = q->tag_set;
2922 mutex_lock(&set->tag_list_lock);
2923 list_del(&q->tag_set_list);
2924 if (list_is_singular(&set->tag_list)) {
2925 /* just transitioned to unshared */
2926 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2927 /* update existing queue */
2928 blk_mq_update_tag_set_shared(set, false);
2930 mutex_unlock(&set->tag_list_lock);
2931 INIT_LIST_HEAD(&q->tag_set_list);
2934 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2935 struct request_queue *q)
2937 mutex_lock(&set->tag_list_lock);
2940 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2942 if (!list_empty(&set->tag_list) &&
2943 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2944 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2945 /* update existing queue */
2946 blk_mq_update_tag_set_shared(set, true);
2948 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
2949 queue_set_hctx_shared(q, true);
2950 list_add_tail(&q->tag_set_list, &set->tag_list);
2952 mutex_unlock(&set->tag_list_lock);
2955 /* All allocations will be freed in release handler of q->mq_kobj */
2956 static int blk_mq_alloc_ctxs(struct request_queue *q)
2958 struct blk_mq_ctxs *ctxs;
2961 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2965 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2966 if (!ctxs->queue_ctx)
2969 for_each_possible_cpu(cpu) {
2970 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2974 q->mq_kobj = &ctxs->kobj;
2975 q->queue_ctx = ctxs->queue_ctx;
2984 * It is the actual release handler for mq, but we do it from
2985 * request queue's release handler for avoiding use-after-free
2986 * and headache because q->mq_kobj shouldn't have been introduced,
2987 * but we can't group ctx/kctx kobj without it.
2989 void blk_mq_release(struct request_queue *q)
2991 struct blk_mq_hw_ctx *hctx, *next;
2994 queue_for_each_hw_ctx(q, hctx, i)
2995 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2997 /* all hctx are in .unused_hctx_list now */
2998 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2999 list_del_init(&hctx->hctx_list);
3000 kobject_put(&hctx->kobj);
3003 kfree(q->queue_hw_ctx);
3006 * release .mq_kobj and sw queue's kobject now because
3007 * both share lifetime with request queue.
3009 blk_mq_sysfs_deinit(q);
3012 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3015 struct request_queue *uninit_q, *q;
3017 uninit_q = blk_alloc_queue(set->numa_node);
3019 return ERR_PTR(-ENOMEM);
3020 uninit_q->queuedata = queuedata;
3023 * Initialize the queue without an elevator. device_add_disk() will do
3024 * the initialization.
3026 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3028 blk_cleanup_queue(uninit_q);
3032 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3034 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3036 return blk_mq_init_queue_data(set, NULL);
3038 EXPORT_SYMBOL(blk_mq_init_queue);
3041 * Helper for setting up a queue with mq ops, given queue depth, and
3042 * the passed in mq ops flags.
3044 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3045 const struct blk_mq_ops *ops,
3046 unsigned int queue_depth,
3047 unsigned int set_flags)
3049 struct request_queue *q;
3052 memset(set, 0, sizeof(*set));
3054 set->nr_hw_queues = 1;
3056 set->queue_depth = queue_depth;
3057 set->numa_node = NUMA_NO_NODE;
3058 set->flags = set_flags;
3060 ret = blk_mq_alloc_tag_set(set);
3062 return ERR_PTR(ret);
3064 q = blk_mq_init_queue(set);
3066 blk_mq_free_tag_set(set);
3072 EXPORT_SYMBOL(blk_mq_init_sq_queue);
3074 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3075 struct blk_mq_tag_set *set, struct request_queue *q,
3076 int hctx_idx, int node)
3078 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3080 /* reuse dead hctx first */
3081 spin_lock(&q->unused_hctx_lock);
3082 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3083 if (tmp->numa_node == node) {
3089 list_del_init(&hctx->hctx_list);
3090 spin_unlock(&q->unused_hctx_lock);
3093 hctx = blk_mq_alloc_hctx(q, set, node);
3097 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3103 kobject_put(&hctx->kobj);
3108 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3109 struct request_queue *q)
3112 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3114 if (q->nr_hw_queues < set->nr_hw_queues) {
3115 struct blk_mq_hw_ctx **new_hctxs;
3117 new_hctxs = kcalloc_node(set->nr_hw_queues,
3118 sizeof(*new_hctxs), GFP_KERNEL,
3123 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3125 q->queue_hw_ctx = new_hctxs;
3130 /* protect against switching io scheduler */
3131 mutex_lock(&q->sysfs_lock);
3132 for (i = 0; i < set->nr_hw_queues; i++) {
3134 struct blk_mq_hw_ctx *hctx;
3136 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3138 * If the hw queue has been mapped to another numa node,
3139 * we need to realloc the hctx. If allocation fails, fallback
3140 * to use the previous one.
3142 if (hctxs[i] && (hctxs[i]->numa_node == node))
3145 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3148 blk_mq_exit_hctx(q, set, hctxs[i], i);
3152 pr_warn("Allocate new hctx on node %d fails,\
3153 fallback to previous one on node %d\n",
3154 node, hctxs[i]->numa_node);
3160 * Increasing nr_hw_queues fails. Free the newly allocated
3161 * hctxs and keep the previous q->nr_hw_queues.
3163 if (i != set->nr_hw_queues) {
3164 j = q->nr_hw_queues;
3168 end = q->nr_hw_queues;
3169 q->nr_hw_queues = set->nr_hw_queues;
3172 for (; j < end; j++) {
3173 struct blk_mq_hw_ctx *hctx = hctxs[j];
3177 blk_mq_free_map_and_requests(set, j);
3178 blk_mq_exit_hctx(q, set, hctx, j);
3182 mutex_unlock(&q->sysfs_lock);
3185 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3186 struct request_queue *q,
3189 /* mark the queue as mq asap */
3190 q->mq_ops = set->ops;
3192 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3193 blk_mq_poll_stats_bkt,
3194 BLK_MQ_POLL_STATS_BKTS, q);
3198 if (blk_mq_alloc_ctxs(q))
3201 /* init q->mq_kobj and sw queues' kobjects */
3202 blk_mq_sysfs_init(q);
3204 INIT_LIST_HEAD(&q->unused_hctx_list);
3205 spin_lock_init(&q->unused_hctx_lock);
3207 blk_mq_realloc_hw_ctxs(set, q);
3208 if (!q->nr_hw_queues)
3211 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3212 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3216 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3217 if (set->nr_maps > HCTX_TYPE_POLL &&
3218 set->map[HCTX_TYPE_POLL].nr_queues)
3219 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3221 q->sg_reserved_size = INT_MAX;
3223 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3224 INIT_LIST_HEAD(&q->requeue_list);
3225 spin_lock_init(&q->requeue_lock);
3227 q->nr_requests = set->queue_depth;
3230 * Default to classic polling
3232 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3234 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3235 blk_mq_add_queue_tag_set(set, q);
3236 blk_mq_map_swqueue(q);
3239 elevator_init_mq(q);
3244 kfree(q->queue_hw_ctx);
3245 q->nr_hw_queues = 0;
3246 blk_mq_sysfs_deinit(q);
3248 blk_stat_free_callback(q->poll_cb);
3252 return ERR_PTR(-ENOMEM);
3254 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3256 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3257 void blk_mq_exit_queue(struct request_queue *q)
3259 struct blk_mq_tag_set *set = q->tag_set;
3261 blk_mq_del_queue_tag_set(q);
3262 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3265 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3269 for (i = 0; i < set->nr_hw_queues; i++) {
3270 if (!__blk_mq_alloc_map_and_request(set, i))
3279 blk_mq_free_map_and_requests(set, i);
3285 * Allocate the request maps associated with this tag_set. Note that this
3286 * may reduce the depth asked for, if memory is tight. set->queue_depth
3287 * will be updated to reflect the allocated depth.
3289 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3294 depth = set->queue_depth;
3296 err = __blk_mq_alloc_rq_maps(set);
3300 set->queue_depth >>= 1;
3301 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3305 } while (set->queue_depth);
3307 if (!set->queue_depth || err) {
3308 pr_err("blk-mq: failed to allocate request map\n");
3312 if (depth != set->queue_depth)
3313 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3314 depth, set->queue_depth);
3319 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3322 * blk_mq_map_queues() and multiple .map_queues() implementations
3323 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3324 * number of hardware queues.
3326 if (set->nr_maps == 1)
3327 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3329 if (set->ops->map_queues && !is_kdump_kernel()) {
3333 * transport .map_queues is usually done in the following
3336 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3337 * mask = get_cpu_mask(queue)
3338 * for_each_cpu(cpu, mask)
3339 * set->map[x].mq_map[cpu] = queue;
3342 * When we need to remap, the table has to be cleared for
3343 * killing stale mapping since one CPU may not be mapped
3346 for (i = 0; i < set->nr_maps; i++)
3347 blk_mq_clear_mq_map(&set->map[i]);
3349 return set->ops->map_queues(set);
3351 BUG_ON(set->nr_maps > 1);
3352 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3356 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3357 int cur_nr_hw_queues, int new_nr_hw_queues)
3359 struct blk_mq_tags **new_tags;
3361 if (cur_nr_hw_queues >= new_nr_hw_queues)
3364 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3365 GFP_KERNEL, set->numa_node);
3370 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3371 sizeof(*set->tags));
3373 set->tags = new_tags;
3374 set->nr_hw_queues = new_nr_hw_queues;
3380 * Alloc a tag set to be associated with one or more request queues.
3381 * May fail with EINVAL for various error conditions. May adjust the
3382 * requested depth down, if it's too large. In that case, the set
3383 * value will be stored in set->queue_depth.
3385 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3389 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3391 if (!set->nr_hw_queues)
3393 if (!set->queue_depth)
3395 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3398 if (!set->ops->queue_rq)
3401 if (!set->ops->get_budget ^ !set->ops->put_budget)
3404 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3405 pr_info("blk-mq: reduced tag depth to %u\n",
3407 set->queue_depth = BLK_MQ_MAX_DEPTH;
3412 else if (set->nr_maps > HCTX_MAX_TYPES)
3416 * If a crashdump is active, then we are potentially in a very
3417 * memory constrained environment. Limit us to 1 queue and
3418 * 64 tags to prevent using too much memory.
3420 if (is_kdump_kernel()) {
3421 set->nr_hw_queues = 1;
3423 set->queue_depth = min(64U, set->queue_depth);
3426 * There is no use for more h/w queues than cpus if we just have
3429 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3430 set->nr_hw_queues = nr_cpu_ids;
3432 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3436 for (i = 0; i < set->nr_maps; i++) {
3437 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3438 sizeof(set->map[i].mq_map[0]),
3439 GFP_KERNEL, set->numa_node);
3440 if (!set->map[i].mq_map)
3441 goto out_free_mq_map;
3442 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3445 ret = blk_mq_update_queue_map(set);
3447 goto out_free_mq_map;
3449 ret = blk_mq_alloc_map_and_requests(set);
3451 goto out_free_mq_map;
3453 if (blk_mq_is_sbitmap_shared(set->flags)) {
3454 atomic_set(&set->active_queues_shared_sbitmap, 0);
3456 if (blk_mq_init_shared_sbitmap(set, set->flags)) {
3458 goto out_free_mq_rq_maps;
3462 mutex_init(&set->tag_list_lock);
3463 INIT_LIST_HEAD(&set->tag_list);
3467 out_free_mq_rq_maps:
3468 for (i = 0; i < set->nr_hw_queues; i++)
3469 blk_mq_free_map_and_requests(set, i);
3471 for (i = 0; i < set->nr_maps; i++) {
3472 kfree(set->map[i].mq_map);
3473 set->map[i].mq_map = NULL;
3479 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3481 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3485 for (i = 0; i < set->nr_hw_queues; i++)
3486 blk_mq_free_map_and_requests(set, i);
3488 if (blk_mq_is_sbitmap_shared(set->flags))
3489 blk_mq_exit_shared_sbitmap(set);
3491 for (j = 0; j < set->nr_maps; j++) {
3492 kfree(set->map[j].mq_map);
3493 set->map[j].mq_map = NULL;
3499 EXPORT_SYMBOL(blk_mq_free_tag_set);
3501 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3503 struct blk_mq_tag_set *set = q->tag_set;
3504 struct blk_mq_hw_ctx *hctx;
3510 if (q->nr_requests == nr)
3513 blk_mq_freeze_queue(q);
3514 blk_mq_quiesce_queue(q);
3517 queue_for_each_hw_ctx(q, hctx, i) {
3521 * If we're using an MQ scheduler, just update the scheduler
3522 * queue depth. This is similar to what the old code would do.
3524 if (!hctx->sched_tags) {
3525 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3527 if (!ret && blk_mq_is_sbitmap_shared(set->flags))
3528 blk_mq_tag_resize_shared_sbitmap(set, nr);
3530 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3535 if (q->elevator && q->elevator->type->ops.depth_updated)
3536 q->elevator->type->ops.depth_updated(hctx);
3540 q->nr_requests = nr;
3542 blk_mq_unquiesce_queue(q);
3543 blk_mq_unfreeze_queue(q);
3549 * request_queue and elevator_type pair.
3550 * It is just used by __blk_mq_update_nr_hw_queues to cache
3551 * the elevator_type associated with a request_queue.
3553 struct blk_mq_qe_pair {
3554 struct list_head node;
3555 struct request_queue *q;
3556 struct elevator_type *type;
3560 * Cache the elevator_type in qe pair list and switch the
3561 * io scheduler to 'none'
3563 static bool blk_mq_elv_switch_none(struct list_head *head,
3564 struct request_queue *q)
3566 struct blk_mq_qe_pair *qe;
3571 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3575 INIT_LIST_HEAD(&qe->node);
3577 qe->type = q->elevator->type;
3578 list_add(&qe->node, head);
3580 mutex_lock(&q->sysfs_lock);
3582 * After elevator_switch_mq, the previous elevator_queue will be
3583 * released by elevator_release. The reference of the io scheduler
3584 * module get by elevator_get will also be put. So we need to get
3585 * a reference of the io scheduler module here to prevent it to be
3588 __module_get(qe->type->elevator_owner);
3589 elevator_switch_mq(q, NULL);
3590 mutex_unlock(&q->sysfs_lock);
3595 static void blk_mq_elv_switch_back(struct list_head *head,
3596 struct request_queue *q)
3598 struct blk_mq_qe_pair *qe;
3599 struct elevator_type *t = NULL;
3601 list_for_each_entry(qe, head, node)
3610 list_del(&qe->node);
3613 mutex_lock(&q->sysfs_lock);
3614 elevator_switch_mq(q, t);
3615 mutex_unlock(&q->sysfs_lock);
3618 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3621 struct request_queue *q;
3623 int prev_nr_hw_queues;
3625 lockdep_assert_held(&set->tag_list_lock);
3627 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3628 nr_hw_queues = nr_cpu_ids;
3629 if (nr_hw_queues < 1)
3631 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3634 list_for_each_entry(q, &set->tag_list, tag_set_list)
3635 blk_mq_freeze_queue(q);
3637 * Switch IO scheduler to 'none', cleaning up the data associated
3638 * with the previous scheduler. We will switch back once we are done
3639 * updating the new sw to hw queue mappings.
3641 list_for_each_entry(q, &set->tag_list, tag_set_list)
3642 if (!blk_mq_elv_switch_none(&head, q))
3645 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3646 blk_mq_debugfs_unregister_hctxs(q);
3647 blk_mq_sysfs_unregister(q);
3650 prev_nr_hw_queues = set->nr_hw_queues;
3651 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3655 set->nr_hw_queues = nr_hw_queues;
3657 blk_mq_update_queue_map(set);
3658 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3659 blk_mq_realloc_hw_ctxs(set, q);
3660 if (q->nr_hw_queues != set->nr_hw_queues) {
3661 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3662 nr_hw_queues, prev_nr_hw_queues);
3663 set->nr_hw_queues = prev_nr_hw_queues;
3664 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3667 blk_mq_map_swqueue(q);
3671 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3672 blk_mq_sysfs_register(q);
3673 blk_mq_debugfs_register_hctxs(q);
3677 list_for_each_entry(q, &set->tag_list, tag_set_list)
3678 blk_mq_elv_switch_back(&head, q);
3680 list_for_each_entry(q, &set->tag_list, tag_set_list)
3681 blk_mq_unfreeze_queue(q);
3684 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3686 mutex_lock(&set->tag_list_lock);
3687 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3688 mutex_unlock(&set->tag_list_lock);
3690 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3692 /* Enable polling stats and return whether they were already enabled. */
3693 static bool blk_poll_stats_enable(struct request_queue *q)
3695 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3696 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3698 blk_stat_add_callback(q, q->poll_cb);
3702 static void blk_mq_poll_stats_start(struct request_queue *q)
3705 * We don't arm the callback if polling stats are not enabled or the
3706 * callback is already active.
3708 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3709 blk_stat_is_active(q->poll_cb))
3712 blk_stat_activate_msecs(q->poll_cb, 100);
3715 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3717 struct request_queue *q = cb->data;
3720 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3721 if (cb->stat[bucket].nr_samples)
3722 q->poll_stat[bucket] = cb->stat[bucket];
3726 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3729 unsigned long ret = 0;
3733 * If stats collection isn't on, don't sleep but turn it on for
3736 if (!blk_poll_stats_enable(q))
3740 * As an optimistic guess, use half of the mean service time
3741 * for this type of request. We can (and should) make this smarter.
3742 * For instance, if the completion latencies are tight, we can
3743 * get closer than just half the mean. This is especially
3744 * important on devices where the completion latencies are longer
3745 * than ~10 usec. We do use the stats for the relevant IO size
3746 * if available which does lead to better estimates.
3748 bucket = blk_mq_poll_stats_bkt(rq);
3752 if (q->poll_stat[bucket].nr_samples)
3753 ret = (q->poll_stat[bucket].mean + 1) / 2;
3758 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3761 struct hrtimer_sleeper hs;
3762 enum hrtimer_mode mode;
3766 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3770 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3772 * 0: use half of prev avg
3773 * >0: use this specific value
3775 if (q->poll_nsec > 0)
3776 nsecs = q->poll_nsec;
3778 nsecs = blk_mq_poll_nsecs(q, rq);
3783 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3786 * This will be replaced with the stats tracking code, using
3787 * 'avg_completion_time / 2' as the pre-sleep target.
3791 mode = HRTIMER_MODE_REL;
3792 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3793 hrtimer_set_expires(&hs.timer, kt);
3796 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3798 set_current_state(TASK_UNINTERRUPTIBLE);
3799 hrtimer_sleeper_start_expires(&hs, mode);
3802 hrtimer_cancel(&hs.timer);
3803 mode = HRTIMER_MODE_ABS;
3804 } while (hs.task && !signal_pending(current));
3806 __set_current_state(TASK_RUNNING);
3807 destroy_hrtimer_on_stack(&hs.timer);
3811 static bool blk_mq_poll_hybrid(struct request_queue *q,
3812 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3816 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3819 if (!blk_qc_t_is_internal(cookie))
3820 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3822 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3824 * With scheduling, if the request has completed, we'll
3825 * get a NULL return here, as we clear the sched tag when
3826 * that happens. The request still remains valid, like always,
3827 * so we should be safe with just the NULL check.
3833 return blk_mq_poll_hybrid_sleep(q, rq);
3837 * blk_poll - poll for IO completions
3839 * @cookie: cookie passed back at IO submission time
3840 * @spin: whether to spin for completions
3843 * Poll for completions on the passed in queue. Returns number of
3844 * completed entries found. If @spin is true, then blk_poll will continue
3845 * looping until at least one completion is found, unless the task is
3846 * otherwise marked running (or we need to reschedule).
3848 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3850 struct blk_mq_hw_ctx *hctx;
3853 if (!blk_qc_t_valid(cookie) ||
3854 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3858 blk_flush_plug_list(current->plug, false);
3860 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3863 * If we sleep, have the caller restart the poll loop to reset
3864 * the state. Like for the other success return cases, the
3865 * caller is responsible for checking if the IO completed. If
3866 * the IO isn't complete, we'll get called again and will go
3867 * straight to the busy poll loop.
3869 if (blk_mq_poll_hybrid(q, hctx, cookie))
3872 hctx->poll_considered++;
3874 state = current->state;
3878 hctx->poll_invoked++;
3880 ret = q->mq_ops->poll(hctx);
3882 hctx->poll_success++;
3883 __set_current_state(TASK_RUNNING);
3887 if (signal_pending_state(state, current))
3888 __set_current_state(TASK_RUNNING);
3890 if (current->state == TASK_RUNNING)
3892 if (ret < 0 || !spin)
3895 } while (!need_resched());
3897 __set_current_state(TASK_RUNNING);
3900 EXPORT_SYMBOL_GPL(blk_poll);
3902 unsigned int blk_mq_rq_cpu(struct request *rq)
3904 return rq->mq_ctx->cpu;
3906 EXPORT_SYMBOL(blk_mq_rq_cpu);
3908 static int __init blk_mq_init(void)
3912 for_each_possible_cpu(i)
3913 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3914 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3916 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3917 "block/softirq:dead", NULL,
3918 blk_softirq_cpu_dead);
3919 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3920 blk_mq_hctx_notify_dead);
3921 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3922 blk_mq_hctx_notify_online,
3923 blk_mq_hctx_notify_offline);
3926 subsys_initcall(blk_mq_init);