2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-debugfs.h"
35 #include "blk-mq-tag.h"
38 #include "blk-mq-sched.h"
40 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie);
41 static void blk_mq_poll_stats_start(struct request_queue *q);
42 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
44 static int blk_mq_poll_stats_bkt(const struct request *rq)
46 int ddir, bytes, bucket;
48 ddir = rq_data_dir(rq);
49 bytes = blk_rq_bytes(rq);
51 bucket = ddir + 2*(ilog2(bytes) - 9);
55 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
56 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
62 * Check if any of the ctx's have pending work in this hardware queue
64 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
66 return !list_empty_careful(&hctx->dispatch) ||
67 sbitmap_any_bit_set(&hctx->ctx_map) ||
68 blk_mq_sched_has_work(hctx);
72 * Mark this ctx as having pending work in this hardware queue
74 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
75 struct blk_mq_ctx *ctx)
77 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
78 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
81 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
82 struct blk_mq_ctx *ctx)
84 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
88 struct hd_struct *part;
89 unsigned int *inflight;
92 static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
93 struct request *rq, void *priv,
96 struct mq_inflight *mi = priv;
99 * index[0] counts the specific partition that was asked for. index[1]
100 * counts the ones that are active on the whole device, so increment
101 * that if mi->part is indeed a partition, and not a whole device.
103 if (rq->part == mi->part)
105 if (mi->part->partno)
109 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
110 unsigned int inflight[2])
112 struct mq_inflight mi = { .part = part, .inflight = inflight, };
114 inflight[0] = inflight[1] = 0;
115 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
118 static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
119 struct request *rq, void *priv,
122 struct mq_inflight *mi = priv;
124 if (rq->part == mi->part)
125 mi->inflight[rq_data_dir(rq)]++;
128 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
129 unsigned int inflight[2])
131 struct mq_inflight mi = { .part = part, .inflight = inflight, };
133 inflight[0] = inflight[1] = 0;
134 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
137 void blk_freeze_queue_start(struct request_queue *q)
141 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
142 if (freeze_depth == 1) {
143 percpu_ref_kill(&q->q_usage_counter);
145 blk_mq_run_hw_queues(q, false);
148 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
150 void blk_mq_freeze_queue_wait(struct request_queue *q)
152 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
154 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
156 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
157 unsigned long timeout)
159 return wait_event_timeout(q->mq_freeze_wq,
160 percpu_ref_is_zero(&q->q_usage_counter),
163 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
166 * Guarantee no request is in use, so we can change any data structure of
167 * the queue afterward.
169 void blk_freeze_queue(struct request_queue *q)
172 * In the !blk_mq case we are only calling this to kill the
173 * q_usage_counter, otherwise this increases the freeze depth
174 * and waits for it to return to zero. For this reason there is
175 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
176 * exported to drivers as the only user for unfreeze is blk_mq.
178 blk_freeze_queue_start(q);
181 blk_mq_freeze_queue_wait(q);
184 void blk_mq_freeze_queue(struct request_queue *q)
187 * ...just an alias to keep freeze and unfreeze actions balanced
188 * in the blk_mq_* namespace
192 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
194 void blk_mq_unfreeze_queue(struct request_queue *q)
198 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
199 WARN_ON_ONCE(freeze_depth < 0);
201 percpu_ref_reinit(&q->q_usage_counter);
202 wake_up_all(&q->mq_freeze_wq);
205 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
208 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
209 * mpt3sas driver such that this function can be removed.
211 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
213 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
215 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
218 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
221 * Note: this function does not prevent that the struct request end_io()
222 * callback function is invoked. Once this function is returned, we make
223 * sure no dispatch can happen until the queue is unquiesced via
224 * blk_mq_unquiesce_queue().
226 void blk_mq_quiesce_queue(struct request_queue *q)
228 struct blk_mq_hw_ctx *hctx;
232 blk_mq_quiesce_queue_nowait(q);
234 queue_for_each_hw_ctx(q, hctx, i) {
235 if (hctx->flags & BLK_MQ_F_BLOCKING)
236 synchronize_srcu(hctx->srcu);
243 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
246 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
249 * This function recovers queue into the state before quiescing
250 * which is done by blk_mq_quiesce_queue.
252 void blk_mq_unquiesce_queue(struct request_queue *q)
254 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
256 /* dispatch requests which are inserted during quiescing */
257 blk_mq_run_hw_queues(q, true);
259 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
261 void blk_mq_wake_waiters(struct request_queue *q)
263 struct blk_mq_hw_ctx *hctx;
266 queue_for_each_hw_ctx(q, hctx, i)
267 if (blk_mq_hw_queue_mapped(hctx))
268 blk_mq_tag_wakeup_all(hctx->tags, true);
271 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
273 return blk_mq_has_free_tags(hctx->tags);
275 EXPORT_SYMBOL(blk_mq_can_queue);
277 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
278 unsigned int tag, unsigned int op)
280 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
281 struct request *rq = tags->static_rqs[tag];
282 req_flags_t rq_flags = 0;
284 if (data->flags & BLK_MQ_REQ_INTERNAL) {
286 rq->internal_tag = tag;
288 if (blk_mq_tag_busy(data->hctx)) {
289 rq_flags = RQF_MQ_INFLIGHT;
290 atomic_inc(&data->hctx->nr_active);
293 rq->internal_tag = -1;
294 data->hctx->tags->rqs[rq->tag] = rq;
297 /* csd/requeue_work/fifo_time is initialized before use */
299 rq->mq_ctx = data->ctx;
300 rq->rq_flags = rq_flags;
303 if (data->flags & BLK_MQ_REQ_PREEMPT)
304 rq->rq_flags |= RQF_PREEMPT;
305 if (blk_queue_io_stat(data->q))
306 rq->rq_flags |= RQF_IO_STAT;
307 INIT_LIST_HEAD(&rq->queuelist);
308 INIT_HLIST_NODE(&rq->hash);
309 RB_CLEAR_NODE(&rq->rb_node);
312 rq->start_time = jiffies;
313 rq->nr_phys_segments = 0;
314 #if defined(CONFIG_BLK_DEV_INTEGRITY)
315 rq->nr_integrity_segments = 0;
318 /* tag was already set */
322 INIT_LIST_HEAD(&rq->timeout_list);
326 rq->end_io_data = NULL;
329 #ifdef CONFIG_BLK_CGROUP
331 set_start_time_ns(rq);
332 rq->io_start_time_ns = 0;
335 data->ctx->rq_dispatched[op_is_sync(op)]++;
339 static struct request *blk_mq_get_request(struct request_queue *q,
340 struct bio *bio, unsigned int op,
341 struct blk_mq_alloc_data *data)
343 struct elevator_queue *e = q->elevator;
346 bool put_ctx_on_error = false;
348 blk_queue_enter_live(q);
350 if (likely(!data->ctx)) {
351 data->ctx = blk_mq_get_ctx(q);
352 put_ctx_on_error = true;
354 if (likely(!data->hctx))
355 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
357 data->flags |= BLK_MQ_REQ_NOWAIT;
360 data->flags |= BLK_MQ_REQ_INTERNAL;
363 * Flush requests are special and go directly to the
366 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
367 e->type->ops.mq.limit_depth(op, data);
370 tag = blk_mq_get_tag(data);
371 if (tag == BLK_MQ_TAG_FAIL) {
372 if (put_ctx_on_error) {
373 blk_mq_put_ctx(data->ctx);
380 rq = blk_mq_rq_ctx_init(data, tag, op);
381 if (!op_is_flush(op)) {
383 if (e && e->type->ops.mq.prepare_request) {
384 if (e->type->icq_cache && rq_ioc(bio))
385 blk_mq_sched_assign_ioc(rq, bio);
387 e->type->ops.mq.prepare_request(rq, bio);
388 rq->rq_flags |= RQF_ELVPRIV;
391 data->hctx->queued++;
395 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
396 blk_mq_req_flags_t flags)
398 struct blk_mq_alloc_data alloc_data = { .flags = flags };
402 ret = blk_queue_enter(q, flags);
406 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
410 return ERR_PTR(-EWOULDBLOCK);
412 blk_mq_put_ctx(alloc_data.ctx);
415 rq->__sector = (sector_t) -1;
416 rq->bio = rq->biotail = NULL;
419 EXPORT_SYMBOL(blk_mq_alloc_request);
421 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
422 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
424 struct blk_mq_alloc_data alloc_data = { .flags = flags };
430 * If the tag allocator sleeps we could get an allocation for a
431 * different hardware context. No need to complicate the low level
432 * allocator for this for the rare use case of a command tied to
435 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
436 return ERR_PTR(-EINVAL);
438 if (hctx_idx >= q->nr_hw_queues)
439 return ERR_PTR(-EIO);
441 ret = blk_queue_enter(q, flags);
446 * Check if the hardware context is actually mapped to anything.
447 * If not tell the caller that it should skip this queue.
449 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
450 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
452 return ERR_PTR(-EXDEV);
454 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
455 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
457 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
461 return ERR_PTR(-EWOULDBLOCK);
465 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
467 void blk_mq_free_request(struct request *rq)
469 struct request_queue *q = rq->q;
470 struct elevator_queue *e = q->elevator;
471 struct blk_mq_ctx *ctx = rq->mq_ctx;
472 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
473 const int sched_tag = rq->internal_tag;
475 if (rq->rq_flags & RQF_ELVPRIV) {
476 if (e && e->type->ops.mq.finish_request)
477 e->type->ops.mq.finish_request(rq);
479 put_io_context(rq->elv.icq->ioc);
484 ctx->rq_completed[rq_is_sync(rq)]++;
485 if (rq->rq_flags & RQF_MQ_INFLIGHT)
486 atomic_dec(&hctx->nr_active);
488 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
489 laptop_io_completion(q->backing_dev_info);
491 wbt_done(q->rq_wb, &rq->issue_stat);
494 blk_put_rl(blk_rq_rl(rq));
496 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
498 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
500 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
501 blk_mq_sched_restart(hctx);
504 EXPORT_SYMBOL_GPL(blk_mq_free_request);
506 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
508 blk_account_io_done(rq);
511 wbt_done(rq->q->rq_wb, &rq->issue_stat);
512 rq->end_io(rq, error);
514 if (unlikely(blk_bidi_rq(rq)))
515 blk_mq_free_request(rq->next_rq);
516 blk_mq_free_request(rq);
519 EXPORT_SYMBOL(__blk_mq_end_request);
521 void blk_mq_end_request(struct request *rq, blk_status_t error)
523 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
525 __blk_mq_end_request(rq, error);
527 EXPORT_SYMBOL(blk_mq_end_request);
529 static void __blk_mq_complete_request_remote(void *data)
531 struct request *rq = data;
533 rq->q->softirq_done_fn(rq);
536 static void __blk_mq_complete_request(struct request *rq)
538 struct blk_mq_ctx *ctx = rq->mq_ctx;
542 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT);
543 blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE);
545 if (rq->internal_tag != -1)
546 blk_mq_sched_completed_request(rq);
547 if (rq->rq_flags & RQF_STATS) {
548 blk_mq_poll_stats_start(rq->q);
552 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
553 rq->q->softirq_done_fn(rq);
558 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
559 shared = cpus_share_cache(cpu, ctx->cpu);
561 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
562 rq->csd.func = __blk_mq_complete_request_remote;
565 smp_call_function_single_async(ctx->cpu, &rq->csd);
567 rq->q->softirq_done_fn(rq);
572 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
573 __releases(hctx->srcu)
575 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
578 srcu_read_unlock(hctx->srcu, srcu_idx);
581 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
582 __acquires(hctx->srcu)
584 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
585 /* shut up gcc false positive */
589 *srcu_idx = srcu_read_lock(hctx->srcu);
592 static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate)
597 * blk_mq_rq_aborted_gstate() is used from the completion path and
598 * can thus be called from irq context. u64_stats_fetch in the
599 * middle of update on the same CPU leads to lockup. Disable irq
602 local_irq_save(flags);
603 u64_stats_update_begin(&rq->aborted_gstate_sync);
604 rq->aborted_gstate = gstate;
605 u64_stats_update_end(&rq->aborted_gstate_sync);
606 local_irq_restore(flags);
609 static u64 blk_mq_rq_aborted_gstate(struct request *rq)
615 start = u64_stats_fetch_begin(&rq->aborted_gstate_sync);
616 aborted_gstate = rq->aborted_gstate;
617 } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start));
619 return aborted_gstate;
623 * blk_mq_complete_request - end I/O on a request
624 * @rq: the request being processed
627 * Ends all I/O on a request. It does not handle partial completions.
628 * The actual completion happens out-of-order, through a IPI handler.
630 void blk_mq_complete_request(struct request *rq)
632 struct request_queue *q = rq->q;
633 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
636 if (unlikely(blk_should_fake_timeout(q)))
640 * If @rq->aborted_gstate equals the current instance, timeout is
641 * claiming @rq and we lost. This is synchronized through
642 * hctx_lock(). See blk_mq_timeout_work() for details.
644 * Completion path never blocks and we can directly use RCU here
645 * instead of hctx_lock() which can be either RCU or SRCU.
646 * However, that would complicate paths which want to synchronize
647 * against us. Let stay in sync with the issue path so that
648 * hctx_lock() covers both issue and completion paths.
650 hctx_lock(hctx, &srcu_idx);
651 if (blk_mq_rq_aborted_gstate(rq) != rq->gstate)
652 __blk_mq_complete_request(rq);
653 hctx_unlock(hctx, srcu_idx);
655 EXPORT_SYMBOL(blk_mq_complete_request);
657 int blk_mq_request_started(struct request *rq)
659 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
661 EXPORT_SYMBOL_GPL(blk_mq_request_started);
663 void blk_mq_start_request(struct request *rq)
665 struct request_queue *q = rq->q;
667 blk_mq_sched_started_request(rq);
669 trace_block_rq_issue(q, rq);
671 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
672 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
673 rq->rq_flags |= RQF_STATS;
674 wbt_issue(q->rq_wb, &rq->issue_stat);
677 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
680 * Mark @rq in-flight which also advances the generation number,
681 * and register for timeout. Protect with a seqcount to allow the
682 * timeout path to read both @rq->gstate and @rq->deadline
685 * This is the only place where a request is marked in-flight. If
686 * the timeout path reads an in-flight @rq->gstate, the
687 * @rq->deadline it reads together under @rq->gstate_seq is
688 * guaranteed to be the matching one.
691 write_seqcount_begin(&rq->gstate_seq);
693 blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT);
696 write_seqcount_end(&rq->gstate_seq);
699 if (q->dma_drain_size && blk_rq_bytes(rq)) {
701 * Make sure space for the drain appears. We know we can do
702 * this because max_hw_segments has been adjusted to be one
703 * fewer than the device can handle.
705 rq->nr_phys_segments++;
708 EXPORT_SYMBOL(blk_mq_start_request);
711 * When we reach here because queue is busy, it's safe to change the state
712 * to IDLE without checking @rq->aborted_gstate because we should still be
713 * holding the RCU read lock and thus protected against timeout.
715 static void __blk_mq_requeue_request(struct request *rq)
717 struct request_queue *q = rq->q;
719 blk_mq_put_driver_tag(rq);
721 trace_block_rq_requeue(q, rq);
722 wbt_requeue(q->rq_wb, &rq->issue_stat);
724 if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) {
725 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
726 if (q->dma_drain_size && blk_rq_bytes(rq))
727 rq->nr_phys_segments--;
731 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
733 __blk_mq_requeue_request(rq);
735 /* this request will be re-inserted to io scheduler queue */
736 blk_mq_sched_requeue_request(rq);
738 BUG_ON(blk_queued_rq(rq));
739 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
741 EXPORT_SYMBOL(blk_mq_requeue_request);
743 static void blk_mq_requeue_work(struct work_struct *work)
745 struct request_queue *q =
746 container_of(work, struct request_queue, requeue_work.work);
748 struct request *rq, *next;
750 spin_lock_irq(&q->requeue_lock);
751 list_splice_init(&q->requeue_list, &rq_list);
752 spin_unlock_irq(&q->requeue_lock);
754 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
755 if (!(rq->rq_flags & RQF_SOFTBARRIER))
758 rq->rq_flags &= ~RQF_SOFTBARRIER;
759 list_del_init(&rq->queuelist);
760 blk_mq_sched_insert_request(rq, true, false, false);
763 while (!list_empty(&rq_list)) {
764 rq = list_entry(rq_list.next, struct request, queuelist);
765 list_del_init(&rq->queuelist);
766 blk_mq_sched_insert_request(rq, false, false, false);
769 blk_mq_run_hw_queues(q, false);
772 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
773 bool kick_requeue_list)
775 struct request_queue *q = rq->q;
779 * We abuse this flag that is otherwise used by the I/O scheduler to
780 * request head insertion from the workqueue.
782 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
784 spin_lock_irqsave(&q->requeue_lock, flags);
786 rq->rq_flags |= RQF_SOFTBARRIER;
787 list_add(&rq->queuelist, &q->requeue_list);
789 list_add_tail(&rq->queuelist, &q->requeue_list);
791 spin_unlock_irqrestore(&q->requeue_lock, flags);
793 if (kick_requeue_list)
794 blk_mq_kick_requeue_list(q);
796 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
798 void blk_mq_kick_requeue_list(struct request_queue *q)
800 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
802 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
804 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
807 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
808 msecs_to_jiffies(msecs));
810 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
812 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
814 if (tag < tags->nr_tags) {
815 prefetch(tags->rqs[tag]);
816 return tags->rqs[tag];
821 EXPORT_SYMBOL(blk_mq_tag_to_rq);
823 struct blk_mq_timeout_data {
825 unsigned int next_set;
826 unsigned int nr_expired;
829 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
831 const struct blk_mq_ops *ops = req->q->mq_ops;
832 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
834 req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED;
837 ret = ops->timeout(req, reserved);
841 __blk_mq_complete_request(req);
843 case BLK_EH_RESET_TIMER:
845 * As nothing prevents from completion happening while
846 * ->aborted_gstate is set, this may lead to ignored
847 * completions and further spurious timeouts.
849 blk_mq_rq_update_aborted_gstate(req, 0);
852 case BLK_EH_NOT_HANDLED:
855 printk(KERN_ERR "block: bad eh return: %d\n", ret);
860 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
861 struct request *rq, void *priv, bool reserved)
863 struct blk_mq_timeout_data *data = priv;
864 unsigned long gstate, deadline;
869 if (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED)
872 /* read coherent snapshots of @rq->state_gen and @rq->deadline */
874 start = read_seqcount_begin(&rq->gstate_seq);
875 gstate = READ_ONCE(rq->gstate);
876 deadline = blk_rq_deadline(rq);
877 if (!read_seqcount_retry(&rq->gstate_seq, start))
882 /* if in-flight && overdue, mark for abortion */
883 if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT &&
884 time_after_eq(jiffies, deadline)) {
885 blk_mq_rq_update_aborted_gstate(rq, gstate);
888 } else if (!data->next_set || time_after(data->next, deadline)) {
889 data->next = deadline;
894 static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx,
895 struct request *rq, void *priv, bool reserved)
898 * We marked @rq->aborted_gstate and waited for RCU. If there were
899 * completions that we lost to, they would have finished and
900 * updated @rq->gstate by now; otherwise, the completion path is
901 * now guaranteed to see @rq->aborted_gstate and yield. If
902 * @rq->aborted_gstate still matches @rq->gstate, @rq is ours.
904 if (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) &&
905 READ_ONCE(rq->gstate) == rq->aborted_gstate)
906 blk_mq_rq_timed_out(rq, reserved);
909 static void blk_mq_timeout_work(struct work_struct *work)
911 struct request_queue *q =
912 container_of(work, struct request_queue, timeout_work);
913 struct blk_mq_timeout_data data = {
918 struct blk_mq_hw_ctx *hctx;
921 /* A deadlock might occur if a request is stuck requiring a
922 * timeout at the same time a queue freeze is waiting
923 * completion, since the timeout code would not be able to
924 * acquire the queue reference here.
926 * That's why we don't use blk_queue_enter here; instead, we use
927 * percpu_ref_tryget directly, because we need to be able to
928 * obtain a reference even in the short window between the queue
929 * starting to freeze, by dropping the first reference in
930 * blk_freeze_queue_start, and the moment the last request is
931 * consumed, marked by the instant q_usage_counter reaches
934 if (!percpu_ref_tryget(&q->q_usage_counter))
937 /* scan for the expired ones and set their ->aborted_gstate */
938 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
940 if (data.nr_expired) {
941 bool has_rcu = false;
944 * Wait till everyone sees ->aborted_gstate. The
945 * sequential waits for SRCUs aren't ideal. If this ever
946 * becomes a problem, we can add per-hw_ctx rcu_head and
949 queue_for_each_hw_ctx(q, hctx, i) {
950 if (!hctx->nr_expired)
953 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
956 synchronize_srcu(hctx->srcu);
958 hctx->nr_expired = 0;
963 /* terminate the ones we won */
964 blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL);
968 data.next = blk_rq_timeout(round_jiffies_up(data.next));
969 mod_timer(&q->timeout, data.next);
972 * Request timeouts are handled as a forward rolling timer. If
973 * we end up here it means that no requests are pending and
974 * also that no request has been pending for a while. Mark
977 queue_for_each_hw_ctx(q, hctx, i) {
978 /* the hctx may be unmapped, so check it here */
979 if (blk_mq_hw_queue_mapped(hctx))
980 blk_mq_tag_idle(hctx);
986 struct flush_busy_ctx_data {
987 struct blk_mq_hw_ctx *hctx;
988 struct list_head *list;
991 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
993 struct flush_busy_ctx_data *flush_data = data;
994 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
995 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
997 spin_lock(&ctx->lock);
998 list_splice_tail_init(&ctx->rq_list, flush_data->list);
999 sbitmap_clear_bit(sb, bitnr);
1000 spin_unlock(&ctx->lock);
1005 * Process software queues that have been marked busy, splicing them
1006 * to the for-dispatch
1008 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1010 struct flush_busy_ctx_data data = {
1015 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1017 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1019 struct dispatch_rq_data {
1020 struct blk_mq_hw_ctx *hctx;
1024 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1027 struct dispatch_rq_data *dispatch_data = data;
1028 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1029 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1031 spin_lock(&ctx->lock);
1032 if (unlikely(!list_empty(&ctx->rq_list))) {
1033 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
1034 list_del_init(&dispatch_data->rq->queuelist);
1035 if (list_empty(&ctx->rq_list))
1036 sbitmap_clear_bit(sb, bitnr);
1038 spin_unlock(&ctx->lock);
1040 return !dispatch_data->rq;
1043 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1044 struct blk_mq_ctx *start)
1046 unsigned off = start ? start->index_hw : 0;
1047 struct dispatch_rq_data data = {
1052 __sbitmap_for_each_set(&hctx->ctx_map, off,
1053 dispatch_rq_from_ctx, &data);
1058 static inline unsigned int queued_to_index(unsigned int queued)
1063 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1066 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
1069 struct blk_mq_alloc_data data = {
1071 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
1072 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
1075 might_sleep_if(wait);
1080 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1081 data.flags |= BLK_MQ_REQ_RESERVED;
1083 rq->tag = blk_mq_get_tag(&data);
1085 if (blk_mq_tag_busy(data.hctx)) {
1086 rq->rq_flags |= RQF_MQ_INFLIGHT;
1087 atomic_inc(&data.hctx->nr_active);
1089 data.hctx->tags->rqs[rq->tag] = rq;
1095 return rq->tag != -1;
1098 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1099 int flags, void *key)
1101 struct blk_mq_hw_ctx *hctx;
1103 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1105 list_del_init(&wait->entry);
1106 blk_mq_run_hw_queue(hctx, true);
1111 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1112 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1113 * restart. For both cases, take care to check the condition again after
1114 * marking us as waiting.
1116 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1119 struct blk_mq_hw_ctx *this_hctx = *hctx;
1120 struct sbq_wait_state *ws;
1121 wait_queue_entry_t *wait;
1124 if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1125 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1126 set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1129 * It's possible that a tag was freed in the window between the
1130 * allocation failure and adding the hardware queue to the wait
1133 * Don't clear RESTART here, someone else could have set it.
1134 * At most this will cost an extra queue run.
1136 return blk_mq_get_driver_tag(rq, hctx, false);
1139 wait = &this_hctx->dispatch_wait;
1140 if (!list_empty_careful(&wait->entry))
1143 spin_lock(&this_hctx->lock);
1144 if (!list_empty(&wait->entry)) {
1145 spin_unlock(&this_hctx->lock);
1149 ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1150 add_wait_queue(&ws->wait, wait);
1153 * It's possible that a tag was freed in the window between the
1154 * allocation failure and adding the hardware queue to the wait
1157 ret = blk_mq_get_driver_tag(rq, hctx, false);
1159 spin_unlock(&this_hctx->lock);
1164 * We got a tag, remove ourselves from the wait queue to ensure
1165 * someone else gets the wakeup.
1167 spin_lock_irq(&ws->wait.lock);
1168 list_del_init(&wait->entry);
1169 spin_unlock_irq(&ws->wait.lock);
1170 spin_unlock(&this_hctx->lock);
1175 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1177 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1180 struct blk_mq_hw_ctx *hctx;
1181 struct request *rq, *nxt;
1182 bool no_tag = false;
1184 blk_status_t ret = BLK_STS_OK;
1186 if (list_empty(list))
1189 WARN_ON(!list_is_singular(list) && got_budget);
1192 * Now process all the entries, sending them to the driver.
1194 errors = queued = 0;
1196 struct blk_mq_queue_data bd;
1198 rq = list_first_entry(list, struct request, queuelist);
1200 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1201 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1204 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1206 * The initial allocation attempt failed, so we need to
1207 * rerun the hardware queue when a tag is freed. The
1208 * waitqueue takes care of that. If the queue is run
1209 * before we add this entry back on the dispatch list,
1210 * we'll re-run it below.
1212 if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1213 blk_mq_put_dispatch_budget(hctx);
1215 * For non-shared tags, the RESTART check
1218 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1224 list_del_init(&rq->queuelist);
1229 * Flag last if we have no more requests, or if we have more
1230 * but can't assign a driver tag to it.
1232 if (list_empty(list))
1235 nxt = list_first_entry(list, struct request, queuelist);
1236 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1239 ret = q->mq_ops->queue_rq(hctx, &bd);
1240 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1242 * If an I/O scheduler has been configured and we got a
1243 * driver tag for the next request already, free it
1246 if (!list_empty(list)) {
1247 nxt = list_first_entry(list, struct request, queuelist);
1248 blk_mq_put_driver_tag(nxt);
1250 list_add(&rq->queuelist, list);
1251 __blk_mq_requeue_request(rq);
1255 if (unlikely(ret != BLK_STS_OK)) {
1257 blk_mq_end_request(rq, BLK_STS_IOERR);
1262 } while (!list_empty(list));
1264 hctx->dispatched[queued_to_index(queued)]++;
1267 * Any items that need requeuing? Stuff them into hctx->dispatch,
1268 * that is where we will continue on next queue run.
1270 if (!list_empty(list)) {
1273 spin_lock(&hctx->lock);
1274 list_splice_init(list, &hctx->dispatch);
1275 spin_unlock(&hctx->lock);
1278 * If SCHED_RESTART was set by the caller of this function and
1279 * it is no longer set that means that it was cleared by another
1280 * thread and hence that a queue rerun is needed.
1282 * If 'no_tag' is set, that means that we failed getting
1283 * a driver tag with an I/O scheduler attached. If our dispatch
1284 * waitqueue is no longer active, ensure that we run the queue
1285 * AFTER adding our entries back to the list.
1287 * If no I/O scheduler has been configured it is possible that
1288 * the hardware queue got stopped and restarted before requests
1289 * were pushed back onto the dispatch list. Rerun the queue to
1290 * avoid starvation. Notes:
1291 * - blk_mq_run_hw_queue() checks whether or not a queue has
1292 * been stopped before rerunning a queue.
1293 * - Some but not all block drivers stop a queue before
1294 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1297 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1298 * bit is set, run queue after a delay to avoid IO stalls
1299 * that could otherwise occur if the queue is idle.
1301 needs_restart = blk_mq_sched_needs_restart(hctx);
1302 if (!needs_restart ||
1303 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1304 blk_mq_run_hw_queue(hctx, true);
1305 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1306 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1309 return (queued + errors) != 0;
1312 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1317 * We should be running this queue from one of the CPUs that
1320 * There are at least two related races now between setting
1321 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1322 * __blk_mq_run_hw_queue():
1324 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1325 * but later it becomes online, then this warning is harmless
1328 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1329 * but later it becomes offline, then the warning can't be
1330 * triggered, and we depend on blk-mq timeout handler to
1331 * handle dispatched requests to this hctx
1333 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1334 cpu_online(hctx->next_cpu)) {
1335 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1336 raw_smp_processor_id(),
1337 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1342 * We can't run the queue inline with ints disabled. Ensure that
1343 * we catch bad users of this early.
1345 WARN_ON_ONCE(in_interrupt());
1347 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1349 hctx_lock(hctx, &srcu_idx);
1350 blk_mq_sched_dispatch_requests(hctx);
1351 hctx_unlock(hctx, srcu_idx);
1354 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1356 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1358 if (cpu >= nr_cpu_ids)
1359 cpu = cpumask_first(hctx->cpumask);
1364 * It'd be great if the workqueue API had a way to pass
1365 * in a mask and had some smarts for more clever placement.
1366 * For now we just round-robin here, switching for every
1367 * BLK_MQ_CPU_WORK_BATCH queued items.
1369 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1372 int next_cpu = hctx->next_cpu;
1374 if (hctx->queue->nr_hw_queues == 1)
1375 return WORK_CPU_UNBOUND;
1377 if (--hctx->next_cpu_batch <= 0) {
1379 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1381 if (next_cpu >= nr_cpu_ids)
1382 next_cpu = blk_mq_first_mapped_cpu(hctx);
1383 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1387 * Do unbound schedule if we can't find a online CPU for this hctx,
1388 * and it should only happen in the path of handling CPU DEAD.
1390 if (!cpu_online(next_cpu)) {
1397 * Make sure to re-select CPU next time once after CPUs
1398 * in hctx->cpumask become online again.
1400 hctx->next_cpu = next_cpu;
1401 hctx->next_cpu_batch = 1;
1402 return WORK_CPU_UNBOUND;
1405 hctx->next_cpu = next_cpu;
1409 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1410 unsigned long msecs)
1412 if (unlikely(blk_mq_hctx_stopped(hctx)))
1415 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1416 int cpu = get_cpu();
1417 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1418 __blk_mq_run_hw_queue(hctx);
1426 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1427 msecs_to_jiffies(msecs));
1430 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1432 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1434 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1436 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1442 * When queue is quiesced, we may be switching io scheduler, or
1443 * updating nr_hw_queues, or other things, and we can't run queue
1444 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1446 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1449 hctx_lock(hctx, &srcu_idx);
1450 need_run = !blk_queue_quiesced(hctx->queue) &&
1451 blk_mq_hctx_has_pending(hctx);
1452 hctx_unlock(hctx, srcu_idx);
1455 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1461 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1463 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1465 struct blk_mq_hw_ctx *hctx;
1468 queue_for_each_hw_ctx(q, hctx, i) {
1469 if (blk_mq_hctx_stopped(hctx))
1472 blk_mq_run_hw_queue(hctx, async);
1475 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1478 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1479 * @q: request queue.
1481 * The caller is responsible for serializing this function against
1482 * blk_mq_{start,stop}_hw_queue().
1484 bool blk_mq_queue_stopped(struct request_queue *q)
1486 struct blk_mq_hw_ctx *hctx;
1489 queue_for_each_hw_ctx(q, hctx, i)
1490 if (blk_mq_hctx_stopped(hctx))
1495 EXPORT_SYMBOL(blk_mq_queue_stopped);
1498 * This function is often used for pausing .queue_rq() by driver when
1499 * there isn't enough resource or some conditions aren't satisfied, and
1500 * BLK_STS_RESOURCE is usually returned.
1502 * We do not guarantee that dispatch can be drained or blocked
1503 * after blk_mq_stop_hw_queue() returns. Please use
1504 * blk_mq_quiesce_queue() for that requirement.
1506 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1508 cancel_delayed_work(&hctx->run_work);
1510 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1512 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1515 * This function is often used for pausing .queue_rq() by driver when
1516 * there isn't enough resource or some conditions aren't satisfied, and
1517 * BLK_STS_RESOURCE is usually returned.
1519 * We do not guarantee that dispatch can be drained or blocked
1520 * after blk_mq_stop_hw_queues() returns. Please use
1521 * blk_mq_quiesce_queue() for that requirement.
1523 void blk_mq_stop_hw_queues(struct request_queue *q)
1525 struct blk_mq_hw_ctx *hctx;
1528 queue_for_each_hw_ctx(q, hctx, i)
1529 blk_mq_stop_hw_queue(hctx);
1531 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1533 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1535 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1537 blk_mq_run_hw_queue(hctx, false);
1539 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1541 void blk_mq_start_hw_queues(struct request_queue *q)
1543 struct blk_mq_hw_ctx *hctx;
1546 queue_for_each_hw_ctx(q, hctx, i)
1547 blk_mq_start_hw_queue(hctx);
1549 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1551 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1553 if (!blk_mq_hctx_stopped(hctx))
1556 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1557 blk_mq_run_hw_queue(hctx, async);
1559 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1561 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1563 struct blk_mq_hw_ctx *hctx;
1566 queue_for_each_hw_ctx(q, hctx, i)
1567 blk_mq_start_stopped_hw_queue(hctx, async);
1569 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1571 static void blk_mq_run_work_fn(struct work_struct *work)
1573 struct blk_mq_hw_ctx *hctx;
1575 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1578 * If we are stopped, don't run the queue.
1580 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1581 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1583 __blk_mq_run_hw_queue(hctx);
1586 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1590 struct blk_mq_ctx *ctx = rq->mq_ctx;
1592 lockdep_assert_held(&ctx->lock);
1594 trace_block_rq_insert(hctx->queue, rq);
1597 list_add(&rq->queuelist, &ctx->rq_list);
1599 list_add_tail(&rq->queuelist, &ctx->rq_list);
1602 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1605 struct blk_mq_ctx *ctx = rq->mq_ctx;
1607 lockdep_assert_held(&ctx->lock);
1609 __blk_mq_insert_req_list(hctx, rq, at_head);
1610 blk_mq_hctx_mark_pending(hctx, ctx);
1614 * Should only be used carefully, when the caller knows we want to
1615 * bypass a potential IO scheduler on the target device.
1617 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1619 struct blk_mq_ctx *ctx = rq->mq_ctx;
1620 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1622 spin_lock(&hctx->lock);
1623 list_add_tail(&rq->queuelist, &hctx->dispatch);
1624 spin_unlock(&hctx->lock);
1627 blk_mq_run_hw_queue(hctx, false);
1630 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1631 struct list_head *list)
1635 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1638 spin_lock(&ctx->lock);
1639 while (!list_empty(list)) {
1642 rq = list_first_entry(list, struct request, queuelist);
1643 BUG_ON(rq->mq_ctx != ctx);
1644 list_del_init(&rq->queuelist);
1645 __blk_mq_insert_req_list(hctx, rq, false);
1647 blk_mq_hctx_mark_pending(hctx, ctx);
1648 spin_unlock(&ctx->lock);
1651 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1653 struct request *rqa = container_of(a, struct request, queuelist);
1654 struct request *rqb = container_of(b, struct request, queuelist);
1656 return !(rqa->mq_ctx < rqb->mq_ctx ||
1657 (rqa->mq_ctx == rqb->mq_ctx &&
1658 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1661 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1663 struct blk_mq_ctx *this_ctx;
1664 struct request_queue *this_q;
1667 LIST_HEAD(ctx_list);
1670 list_splice_init(&plug->mq_list, &list);
1672 list_sort(NULL, &list, plug_ctx_cmp);
1678 while (!list_empty(&list)) {
1679 rq = list_entry_rq(list.next);
1680 list_del_init(&rq->queuelist);
1682 if (rq->mq_ctx != this_ctx) {
1684 trace_block_unplug(this_q, depth, from_schedule);
1685 blk_mq_sched_insert_requests(this_q, this_ctx,
1690 this_ctx = rq->mq_ctx;
1696 list_add_tail(&rq->queuelist, &ctx_list);
1700 * If 'this_ctx' is set, we know we have entries to complete
1701 * on 'ctx_list'. Do those.
1704 trace_block_unplug(this_q, depth, from_schedule);
1705 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1710 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1712 blk_init_request_from_bio(rq, bio);
1714 blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1716 blk_account_io_start(rq, true);
1719 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1720 struct blk_mq_ctx *ctx,
1723 spin_lock(&ctx->lock);
1724 __blk_mq_insert_request(hctx, rq, false);
1725 spin_unlock(&ctx->lock);
1728 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1731 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1733 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1736 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1740 struct request_queue *q = rq->q;
1741 struct blk_mq_queue_data bd = {
1745 blk_qc_t new_cookie;
1748 new_cookie = request_to_qc_t(hctx, rq);
1751 * For OK queue, we are done. For error, caller may kill it.
1752 * Any other error (busy), just add it to our list as we
1753 * previously would have done.
1755 ret = q->mq_ops->queue_rq(hctx, &bd);
1758 *cookie = new_cookie;
1760 case BLK_STS_RESOURCE:
1761 case BLK_STS_DEV_RESOURCE:
1762 __blk_mq_requeue_request(rq);
1765 *cookie = BLK_QC_T_NONE;
1772 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1777 struct request_queue *q = rq->q;
1778 bool run_queue = true;
1781 * RCU or SRCU read lock is needed before checking quiesced flag.
1783 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1784 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1785 * and avoid driver to try to dispatch again.
1787 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1789 bypass_insert = false;
1793 if (q->elevator && !bypass_insert)
1796 if (!blk_mq_get_dispatch_budget(hctx))
1799 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1800 blk_mq_put_dispatch_budget(hctx);
1804 return __blk_mq_issue_directly(hctx, rq, cookie);
1807 return BLK_STS_RESOURCE;
1809 blk_mq_sched_insert_request(rq, false, run_queue, false);
1813 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1814 struct request *rq, blk_qc_t *cookie)
1819 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1821 hctx_lock(hctx, &srcu_idx);
1823 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1824 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1825 blk_mq_sched_insert_request(rq, false, true, false);
1826 else if (ret != BLK_STS_OK)
1827 blk_mq_end_request(rq, ret);
1829 hctx_unlock(hctx, srcu_idx);
1832 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1836 blk_qc_t unused_cookie;
1837 struct blk_mq_ctx *ctx = rq->mq_ctx;
1838 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1840 hctx_lock(hctx, &srcu_idx);
1841 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1842 hctx_unlock(hctx, srcu_idx);
1847 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1849 const int is_sync = op_is_sync(bio->bi_opf);
1850 const int is_flush_fua = op_is_flush(bio->bi_opf);
1851 struct blk_mq_alloc_data data = { .flags = 0 };
1853 unsigned int request_count = 0;
1854 struct blk_plug *plug;
1855 struct request *same_queue_rq = NULL;
1857 unsigned int wb_acct;
1859 blk_queue_bounce(q, &bio);
1861 blk_queue_split(q, &bio);
1863 if (!bio_integrity_prep(bio))
1864 return BLK_QC_T_NONE;
1866 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1867 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1868 return BLK_QC_T_NONE;
1870 if (blk_mq_sched_bio_merge(q, bio))
1871 return BLK_QC_T_NONE;
1873 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1875 trace_block_getrq(q, bio, bio->bi_opf);
1877 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1878 if (unlikely(!rq)) {
1879 __wbt_done(q->rq_wb, wb_acct);
1880 if (bio->bi_opf & REQ_NOWAIT)
1881 bio_wouldblock_error(bio);
1882 return BLK_QC_T_NONE;
1885 wbt_track(&rq->issue_stat, wb_acct);
1887 cookie = request_to_qc_t(data.hctx, rq);
1889 plug = current->plug;
1890 if (unlikely(is_flush_fua)) {
1891 blk_mq_put_ctx(data.ctx);
1892 blk_mq_bio_to_request(rq, bio);
1894 /* bypass scheduler for flush rq */
1895 blk_insert_flush(rq);
1896 blk_mq_run_hw_queue(data.hctx, true);
1897 } else if (plug && q->nr_hw_queues == 1) {
1898 struct request *last = NULL;
1900 blk_mq_put_ctx(data.ctx);
1901 blk_mq_bio_to_request(rq, bio);
1904 * @request_count may become stale because of schedule
1905 * out, so check the list again.
1907 if (list_empty(&plug->mq_list))
1909 else if (blk_queue_nomerges(q))
1910 request_count = blk_plug_queued_count(q);
1913 trace_block_plug(q);
1915 last = list_entry_rq(plug->mq_list.prev);
1917 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1918 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1919 blk_flush_plug_list(plug, false);
1920 trace_block_plug(q);
1923 list_add_tail(&rq->queuelist, &plug->mq_list);
1924 } else if (plug && !blk_queue_nomerges(q)) {
1925 blk_mq_bio_to_request(rq, bio);
1928 * We do limited plugging. If the bio can be merged, do that.
1929 * Otherwise the existing request in the plug list will be
1930 * issued. So the plug list will have one request at most
1931 * The plug list might get flushed before this. If that happens,
1932 * the plug list is empty, and same_queue_rq is invalid.
1934 if (list_empty(&plug->mq_list))
1935 same_queue_rq = NULL;
1937 list_del_init(&same_queue_rq->queuelist);
1938 list_add_tail(&rq->queuelist, &plug->mq_list);
1940 blk_mq_put_ctx(data.ctx);
1942 if (same_queue_rq) {
1943 data.hctx = blk_mq_map_queue(q,
1944 same_queue_rq->mq_ctx->cpu);
1945 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1948 } else if (q->nr_hw_queues > 1 && is_sync) {
1949 blk_mq_put_ctx(data.ctx);
1950 blk_mq_bio_to_request(rq, bio);
1951 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1952 } else if (q->elevator) {
1953 blk_mq_put_ctx(data.ctx);
1954 blk_mq_bio_to_request(rq, bio);
1955 blk_mq_sched_insert_request(rq, false, true, true);
1957 blk_mq_put_ctx(data.ctx);
1958 blk_mq_bio_to_request(rq, bio);
1959 blk_mq_queue_io(data.hctx, data.ctx, rq);
1960 blk_mq_run_hw_queue(data.hctx, true);
1966 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1967 unsigned int hctx_idx)
1971 if (tags->rqs && set->ops->exit_request) {
1974 for (i = 0; i < tags->nr_tags; i++) {
1975 struct request *rq = tags->static_rqs[i];
1979 set->ops->exit_request(set, rq, hctx_idx);
1980 tags->static_rqs[i] = NULL;
1984 while (!list_empty(&tags->page_list)) {
1985 page = list_first_entry(&tags->page_list, struct page, lru);
1986 list_del_init(&page->lru);
1988 * Remove kmemleak object previously allocated in
1989 * blk_mq_init_rq_map().
1991 kmemleak_free(page_address(page));
1992 __free_pages(page, page->private);
1996 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2000 kfree(tags->static_rqs);
2001 tags->static_rqs = NULL;
2003 blk_mq_free_tags(tags);
2006 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2007 unsigned int hctx_idx,
2008 unsigned int nr_tags,
2009 unsigned int reserved_tags)
2011 struct blk_mq_tags *tags;
2014 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2015 if (node == NUMA_NO_NODE)
2016 node = set->numa_node;
2018 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2019 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2023 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2024 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2027 blk_mq_free_tags(tags);
2031 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2032 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2034 if (!tags->static_rqs) {
2036 blk_mq_free_tags(tags);
2043 static size_t order_to_size(unsigned int order)
2045 return (size_t)PAGE_SIZE << order;
2048 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2049 unsigned int hctx_idx, int node)
2053 if (set->ops->init_request) {
2054 ret = set->ops->init_request(set, rq, hctx_idx, node);
2059 seqcount_init(&rq->gstate_seq);
2060 u64_stats_init(&rq->aborted_gstate_sync);
2062 * start gstate with gen 1 instead of 0, otherwise it will be equal
2063 * to aborted_gstate, and be identified timed out by
2064 * blk_mq_terminate_expired.
2066 WRITE_ONCE(rq->gstate, MQ_RQ_GEN_INC);
2071 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2072 unsigned int hctx_idx, unsigned int depth)
2074 unsigned int i, j, entries_per_page, max_order = 4;
2075 size_t rq_size, left;
2078 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2079 if (node == NUMA_NO_NODE)
2080 node = set->numa_node;
2082 INIT_LIST_HEAD(&tags->page_list);
2085 * rq_size is the size of the request plus driver payload, rounded
2086 * to the cacheline size
2088 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2090 left = rq_size * depth;
2092 for (i = 0; i < depth; ) {
2093 int this_order = max_order;
2098 while (this_order && left < order_to_size(this_order - 1))
2102 page = alloc_pages_node(node,
2103 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2109 if (order_to_size(this_order) < rq_size)
2116 page->private = this_order;
2117 list_add_tail(&page->lru, &tags->page_list);
2119 p = page_address(page);
2121 * Allow kmemleak to scan these pages as they contain pointers
2122 * to additional allocations like via ops->init_request().
2124 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2125 entries_per_page = order_to_size(this_order) / rq_size;
2126 to_do = min(entries_per_page, depth - i);
2127 left -= to_do * rq_size;
2128 for (j = 0; j < to_do; j++) {
2129 struct request *rq = p;
2131 tags->static_rqs[i] = rq;
2132 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2133 tags->static_rqs[i] = NULL;
2144 blk_mq_free_rqs(set, tags, hctx_idx);
2149 * 'cpu' is going away. splice any existing rq_list entries from this
2150 * software queue to the hw queue dispatch list, and ensure that it
2153 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2155 struct blk_mq_hw_ctx *hctx;
2156 struct blk_mq_ctx *ctx;
2159 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2160 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2162 spin_lock(&ctx->lock);
2163 if (!list_empty(&ctx->rq_list)) {
2164 list_splice_init(&ctx->rq_list, &tmp);
2165 blk_mq_hctx_clear_pending(hctx, ctx);
2167 spin_unlock(&ctx->lock);
2169 if (list_empty(&tmp))
2172 spin_lock(&hctx->lock);
2173 list_splice_tail_init(&tmp, &hctx->dispatch);
2174 spin_unlock(&hctx->lock);
2176 blk_mq_run_hw_queue(hctx, true);
2180 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2182 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2186 /* hctx->ctxs will be freed in queue's release handler */
2187 static void blk_mq_exit_hctx(struct request_queue *q,
2188 struct blk_mq_tag_set *set,
2189 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2191 blk_mq_debugfs_unregister_hctx(hctx);
2193 if (blk_mq_hw_queue_mapped(hctx))
2194 blk_mq_tag_idle(hctx);
2196 if (set->ops->exit_request)
2197 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2199 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2201 if (set->ops->exit_hctx)
2202 set->ops->exit_hctx(hctx, hctx_idx);
2204 if (hctx->flags & BLK_MQ_F_BLOCKING)
2205 cleanup_srcu_struct(hctx->srcu);
2207 blk_mq_remove_cpuhp(hctx);
2208 blk_free_flush_queue(hctx->fq);
2209 sbitmap_free(&hctx->ctx_map);
2212 static void blk_mq_exit_hw_queues(struct request_queue *q,
2213 struct blk_mq_tag_set *set, int nr_queue)
2215 struct blk_mq_hw_ctx *hctx;
2218 queue_for_each_hw_ctx(q, hctx, i) {
2221 blk_mq_exit_hctx(q, set, hctx, i);
2225 static int blk_mq_init_hctx(struct request_queue *q,
2226 struct blk_mq_tag_set *set,
2227 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2231 node = hctx->numa_node;
2232 if (node == NUMA_NO_NODE)
2233 node = hctx->numa_node = set->numa_node;
2235 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2236 spin_lock_init(&hctx->lock);
2237 INIT_LIST_HEAD(&hctx->dispatch);
2239 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2241 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2243 hctx->tags = set->tags[hctx_idx];
2246 * Allocate space for all possible cpus to avoid allocation at
2249 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2252 goto unregister_cpu_notifier;
2254 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2260 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2261 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2263 if (set->ops->init_hctx &&
2264 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2267 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2270 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2272 goto sched_exit_hctx;
2274 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2277 if (hctx->flags & BLK_MQ_F_BLOCKING)
2278 init_srcu_struct(hctx->srcu);
2280 blk_mq_debugfs_register_hctx(q, hctx);
2287 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2289 if (set->ops->exit_hctx)
2290 set->ops->exit_hctx(hctx, hctx_idx);
2292 sbitmap_free(&hctx->ctx_map);
2295 unregister_cpu_notifier:
2296 blk_mq_remove_cpuhp(hctx);
2300 static void blk_mq_init_cpu_queues(struct request_queue *q,
2301 unsigned int nr_hw_queues)
2305 for_each_possible_cpu(i) {
2306 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2307 struct blk_mq_hw_ctx *hctx;
2310 spin_lock_init(&__ctx->lock);
2311 INIT_LIST_HEAD(&__ctx->rq_list);
2315 * Set local node, IFF we have more than one hw queue. If
2316 * not, we remain on the home node of the device
2318 hctx = blk_mq_map_queue(q, i);
2319 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2320 hctx->numa_node = local_memory_node(cpu_to_node(i));
2324 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2328 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2329 set->queue_depth, set->reserved_tags);
2330 if (!set->tags[hctx_idx])
2333 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2338 blk_mq_free_rq_map(set->tags[hctx_idx]);
2339 set->tags[hctx_idx] = NULL;
2343 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2344 unsigned int hctx_idx)
2346 if (set->tags[hctx_idx]) {
2347 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2348 blk_mq_free_rq_map(set->tags[hctx_idx]);
2349 set->tags[hctx_idx] = NULL;
2353 static void blk_mq_map_swqueue(struct request_queue *q)
2355 unsigned int i, hctx_idx;
2356 struct blk_mq_hw_ctx *hctx;
2357 struct blk_mq_ctx *ctx;
2358 struct blk_mq_tag_set *set = q->tag_set;
2361 * Avoid others reading imcomplete hctx->cpumask through sysfs
2363 mutex_lock(&q->sysfs_lock);
2365 queue_for_each_hw_ctx(q, hctx, i) {
2366 cpumask_clear(hctx->cpumask);
2371 * Map software to hardware queues.
2373 * If the cpu isn't present, the cpu is mapped to first hctx.
2375 for_each_possible_cpu(i) {
2376 hctx_idx = q->mq_map[i];
2377 /* unmapped hw queue can be remapped after CPU topo changed */
2378 if (!set->tags[hctx_idx] &&
2379 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2381 * If tags initialization fail for some hctx,
2382 * that hctx won't be brought online. In this
2383 * case, remap the current ctx to hctx[0] which
2384 * is guaranteed to always have tags allocated
2389 ctx = per_cpu_ptr(q->queue_ctx, i);
2390 hctx = blk_mq_map_queue(q, i);
2392 cpumask_set_cpu(i, hctx->cpumask);
2393 ctx->index_hw = hctx->nr_ctx;
2394 hctx->ctxs[hctx->nr_ctx++] = ctx;
2397 mutex_unlock(&q->sysfs_lock);
2399 queue_for_each_hw_ctx(q, hctx, i) {
2401 * If no software queues are mapped to this hardware queue,
2402 * disable it and free the request entries.
2404 if (!hctx->nr_ctx) {
2405 /* Never unmap queue 0. We need it as a
2406 * fallback in case of a new remap fails
2409 if (i && set->tags[i])
2410 blk_mq_free_map_and_requests(set, i);
2416 hctx->tags = set->tags[i];
2417 WARN_ON(!hctx->tags);
2420 * Set the map size to the number of mapped software queues.
2421 * This is more accurate and more efficient than looping
2422 * over all possibly mapped software queues.
2424 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2427 * Initialize batch roundrobin counts
2429 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2430 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2435 * Caller needs to ensure that we're either frozen/quiesced, or that
2436 * the queue isn't live yet.
2438 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2440 struct blk_mq_hw_ctx *hctx;
2443 queue_for_each_hw_ctx(q, hctx, i) {
2445 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2446 atomic_inc(&q->shared_hctx_restart);
2447 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2449 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2450 atomic_dec(&q->shared_hctx_restart);
2451 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2456 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2459 struct request_queue *q;
2461 lockdep_assert_held(&set->tag_list_lock);
2463 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2464 blk_mq_freeze_queue(q);
2465 queue_set_hctx_shared(q, shared);
2466 blk_mq_unfreeze_queue(q);
2470 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2472 struct blk_mq_tag_set *set = q->tag_set;
2474 mutex_lock(&set->tag_list_lock);
2475 list_del_rcu(&q->tag_set_list);
2476 INIT_LIST_HEAD(&q->tag_set_list);
2477 if (list_is_singular(&set->tag_list)) {
2478 /* just transitioned to unshared */
2479 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2480 /* update existing queue */
2481 blk_mq_update_tag_set_depth(set, false);
2483 mutex_unlock(&set->tag_list_lock);
2488 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2489 struct request_queue *q)
2493 mutex_lock(&set->tag_list_lock);
2496 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2498 if (!list_empty(&set->tag_list) &&
2499 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2500 set->flags |= BLK_MQ_F_TAG_SHARED;
2501 /* update existing queue */
2502 blk_mq_update_tag_set_depth(set, true);
2504 if (set->flags & BLK_MQ_F_TAG_SHARED)
2505 queue_set_hctx_shared(q, true);
2506 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2508 mutex_unlock(&set->tag_list_lock);
2512 * It is the actual release handler for mq, but we do it from
2513 * request queue's release handler for avoiding use-after-free
2514 * and headache because q->mq_kobj shouldn't have been introduced,
2515 * but we can't group ctx/kctx kobj without it.
2517 void blk_mq_release(struct request_queue *q)
2519 struct blk_mq_hw_ctx *hctx;
2522 /* hctx kobj stays in hctx */
2523 queue_for_each_hw_ctx(q, hctx, i) {
2526 kobject_put(&hctx->kobj);
2531 kfree(q->queue_hw_ctx);
2534 * release .mq_kobj and sw queue's kobject now because
2535 * both share lifetime with request queue.
2537 blk_mq_sysfs_deinit(q);
2539 free_percpu(q->queue_ctx);
2542 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2544 struct request_queue *uninit_q, *q;
2546 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2548 return ERR_PTR(-ENOMEM);
2550 q = blk_mq_init_allocated_queue(set, uninit_q);
2552 blk_cleanup_queue(uninit_q);
2556 EXPORT_SYMBOL(blk_mq_init_queue);
2558 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2560 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2562 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2563 __alignof__(struct blk_mq_hw_ctx)) !=
2564 sizeof(struct blk_mq_hw_ctx));
2566 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2567 hw_ctx_size += sizeof(struct srcu_struct);
2572 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2573 struct request_queue *q)
2576 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2578 blk_mq_sysfs_unregister(q);
2580 /* protect against switching io scheduler */
2581 mutex_lock(&q->sysfs_lock);
2582 for (i = 0; i < set->nr_hw_queues; i++) {
2588 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2589 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2594 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2601 atomic_set(&hctxs[i]->nr_active, 0);
2602 hctxs[i]->numa_node = node;
2603 hctxs[i]->queue_num = i;
2605 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2606 free_cpumask_var(hctxs[i]->cpumask);
2611 blk_mq_hctx_kobj_init(hctxs[i]);
2613 for (j = i; j < q->nr_hw_queues; j++) {
2614 struct blk_mq_hw_ctx *hctx = hctxs[j];
2618 blk_mq_free_map_and_requests(set, j);
2619 blk_mq_exit_hctx(q, set, hctx, j);
2620 kobject_put(&hctx->kobj);
2625 q->nr_hw_queues = i;
2626 mutex_unlock(&q->sysfs_lock);
2627 blk_mq_sysfs_register(q);
2630 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2631 struct request_queue *q)
2633 /* mark the queue as mq asap */
2634 q->mq_ops = set->ops;
2636 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2637 blk_mq_poll_stats_bkt,
2638 BLK_MQ_POLL_STATS_BKTS, q);
2642 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2646 /* init q->mq_kobj and sw queues' kobjects */
2647 blk_mq_sysfs_init(q);
2649 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2650 GFP_KERNEL, set->numa_node);
2651 if (!q->queue_hw_ctx)
2654 q->mq_map = set->mq_map;
2656 blk_mq_realloc_hw_ctxs(set, q);
2657 if (!q->nr_hw_queues)
2660 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2661 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2663 q->nr_queues = nr_cpu_ids;
2665 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2667 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2668 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2670 q->sg_reserved_size = INT_MAX;
2672 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2673 INIT_LIST_HEAD(&q->requeue_list);
2674 spin_lock_init(&q->requeue_lock);
2676 blk_queue_make_request(q, blk_mq_make_request);
2677 if (q->mq_ops->poll)
2678 q->poll_fn = blk_mq_poll;
2681 * Do this after blk_queue_make_request() overrides it...
2683 q->nr_requests = set->queue_depth;
2686 * Default to classic polling
2690 if (set->ops->complete)
2691 blk_queue_softirq_done(q, set->ops->complete);
2693 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2694 blk_mq_add_queue_tag_set(set, q);
2695 blk_mq_map_swqueue(q);
2697 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2700 ret = blk_mq_sched_init(q);
2702 return ERR_PTR(ret);
2708 kfree(q->queue_hw_ctx);
2710 free_percpu(q->queue_ctx);
2713 return ERR_PTR(-ENOMEM);
2715 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2717 void blk_mq_free_queue(struct request_queue *q)
2719 struct blk_mq_tag_set *set = q->tag_set;
2721 blk_mq_del_queue_tag_set(q);
2722 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2725 /* Basically redo blk_mq_init_queue with queue frozen */
2726 static void blk_mq_queue_reinit(struct request_queue *q)
2728 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2730 blk_mq_debugfs_unregister_hctxs(q);
2731 blk_mq_sysfs_unregister(q);
2734 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2735 * we should change hctx numa_node according to the new topology (this
2736 * involves freeing and re-allocating memory, worth doing?)
2738 blk_mq_map_swqueue(q);
2740 blk_mq_sysfs_register(q);
2741 blk_mq_debugfs_register_hctxs(q);
2744 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2748 for (i = 0; i < set->nr_hw_queues; i++)
2749 if (!__blk_mq_alloc_rq_map(set, i))
2756 blk_mq_free_rq_map(set->tags[i]);
2762 * Allocate the request maps associated with this tag_set. Note that this
2763 * may reduce the depth asked for, if memory is tight. set->queue_depth
2764 * will be updated to reflect the allocated depth.
2766 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2771 depth = set->queue_depth;
2773 err = __blk_mq_alloc_rq_maps(set);
2777 set->queue_depth >>= 1;
2778 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2782 } while (set->queue_depth);
2784 if (!set->queue_depth || err) {
2785 pr_err("blk-mq: failed to allocate request map\n");
2789 if (depth != set->queue_depth)
2790 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2791 depth, set->queue_depth);
2796 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2798 if (set->ops->map_queues) {
2801 * transport .map_queues is usually done in the following
2804 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2805 * mask = get_cpu_mask(queue)
2806 * for_each_cpu(cpu, mask)
2807 * set->mq_map[cpu] = queue;
2810 * When we need to remap, the table has to be cleared for
2811 * killing stale mapping since one CPU may not be mapped
2814 for_each_possible_cpu(cpu)
2815 set->mq_map[cpu] = 0;
2817 return set->ops->map_queues(set);
2819 return blk_mq_map_queues(set);
2823 * Alloc a tag set to be associated with one or more request queues.
2824 * May fail with EINVAL for various error conditions. May adjust the
2825 * requested depth down, if if it too large. In that case, the set
2826 * value will be stored in set->queue_depth.
2828 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2832 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2834 if (!set->nr_hw_queues)
2836 if (!set->queue_depth)
2838 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2841 if (!set->ops->queue_rq)
2844 if (!set->ops->get_budget ^ !set->ops->put_budget)
2847 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2848 pr_info("blk-mq: reduced tag depth to %u\n",
2850 set->queue_depth = BLK_MQ_MAX_DEPTH;
2854 * If a crashdump is active, then we are potentially in a very
2855 * memory constrained environment. Limit us to 1 queue and
2856 * 64 tags to prevent using too much memory.
2858 if (is_kdump_kernel()) {
2859 set->nr_hw_queues = 1;
2860 set->queue_depth = min(64U, set->queue_depth);
2863 * There is no use for more h/w queues than cpus.
2865 if (set->nr_hw_queues > nr_cpu_ids)
2866 set->nr_hw_queues = nr_cpu_ids;
2868 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2869 GFP_KERNEL, set->numa_node);
2874 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2875 GFP_KERNEL, set->numa_node);
2879 ret = blk_mq_update_queue_map(set);
2881 goto out_free_mq_map;
2883 ret = blk_mq_alloc_rq_maps(set);
2885 goto out_free_mq_map;
2887 mutex_init(&set->tag_list_lock);
2888 INIT_LIST_HEAD(&set->tag_list);
2900 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2902 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2906 for (i = 0; i < nr_cpu_ids; i++)
2907 blk_mq_free_map_and_requests(set, i);
2915 EXPORT_SYMBOL(blk_mq_free_tag_set);
2917 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2919 struct blk_mq_tag_set *set = q->tag_set;
2920 struct blk_mq_hw_ctx *hctx;
2926 blk_mq_freeze_queue(q);
2927 blk_mq_quiesce_queue(q);
2930 queue_for_each_hw_ctx(q, hctx, i) {
2934 * If we're using an MQ scheduler, just update the scheduler
2935 * queue depth. This is similar to what the old code would do.
2937 if (!hctx->sched_tags) {
2938 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2941 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2949 q->nr_requests = nr;
2951 blk_mq_unquiesce_queue(q);
2952 blk_mq_unfreeze_queue(q);
2957 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2960 struct request_queue *q;
2962 lockdep_assert_held(&set->tag_list_lock);
2964 if (nr_hw_queues > nr_cpu_ids)
2965 nr_hw_queues = nr_cpu_ids;
2966 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2969 list_for_each_entry(q, &set->tag_list, tag_set_list)
2970 blk_mq_freeze_queue(q);
2972 set->nr_hw_queues = nr_hw_queues;
2973 blk_mq_update_queue_map(set);
2974 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2975 blk_mq_realloc_hw_ctxs(set, q);
2976 blk_mq_queue_reinit(q);
2979 list_for_each_entry(q, &set->tag_list, tag_set_list)
2980 blk_mq_unfreeze_queue(q);
2983 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2985 mutex_lock(&set->tag_list_lock);
2986 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2987 mutex_unlock(&set->tag_list_lock);
2989 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2991 /* Enable polling stats and return whether they were already enabled. */
2992 static bool blk_poll_stats_enable(struct request_queue *q)
2994 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2995 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
2997 blk_stat_add_callback(q, q->poll_cb);
3001 static void blk_mq_poll_stats_start(struct request_queue *q)
3004 * We don't arm the callback if polling stats are not enabled or the
3005 * callback is already active.
3007 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3008 blk_stat_is_active(q->poll_cb))
3011 blk_stat_activate_msecs(q->poll_cb, 100);
3014 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3016 struct request_queue *q = cb->data;
3019 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3020 if (cb->stat[bucket].nr_samples)
3021 q->poll_stat[bucket] = cb->stat[bucket];
3025 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3026 struct blk_mq_hw_ctx *hctx,
3029 unsigned long ret = 0;
3033 * If stats collection isn't on, don't sleep but turn it on for
3036 if (!blk_poll_stats_enable(q))
3040 * As an optimistic guess, use half of the mean service time
3041 * for this type of request. We can (and should) make this smarter.
3042 * For instance, if the completion latencies are tight, we can
3043 * get closer than just half the mean. This is especially
3044 * important on devices where the completion latencies are longer
3045 * than ~10 usec. We do use the stats for the relevant IO size
3046 * if available which does lead to better estimates.
3048 bucket = blk_mq_poll_stats_bkt(rq);
3052 if (q->poll_stat[bucket].nr_samples)
3053 ret = (q->poll_stat[bucket].mean + 1) / 2;
3058 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3059 struct blk_mq_hw_ctx *hctx,
3062 struct hrtimer_sleeper hs;
3063 enum hrtimer_mode mode;
3067 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3073 * -1: don't ever hybrid sleep
3074 * 0: use half of prev avg
3075 * >0: use this specific value
3077 if (q->poll_nsec == -1)
3079 else if (q->poll_nsec > 0)
3080 nsecs = q->poll_nsec;
3082 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3087 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3090 * This will be replaced with the stats tracking code, using
3091 * 'avg_completion_time / 2' as the pre-sleep target.
3095 mode = HRTIMER_MODE_REL;
3096 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3097 hrtimer_set_expires(&hs.timer, kt);
3099 hrtimer_init_sleeper(&hs, current);
3101 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3103 set_current_state(TASK_UNINTERRUPTIBLE);
3104 hrtimer_start_expires(&hs.timer, mode);
3107 hrtimer_cancel(&hs.timer);
3108 mode = HRTIMER_MODE_ABS;
3109 } while (hs.task && !signal_pending(current));
3111 __set_current_state(TASK_RUNNING);
3112 destroy_hrtimer_on_stack(&hs.timer);
3116 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3118 struct request_queue *q = hctx->queue;
3122 * If we sleep, have the caller restart the poll loop to reset
3123 * the state. Like for the other success return cases, the
3124 * caller is responsible for checking if the IO completed. If
3125 * the IO isn't complete, we'll get called again and will go
3126 * straight to the busy poll loop.
3128 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3131 hctx->poll_considered++;
3133 state = current->state;
3134 while (!need_resched()) {
3137 hctx->poll_invoked++;
3139 ret = q->mq_ops->poll(hctx, rq->tag);
3141 hctx->poll_success++;
3142 set_current_state(TASK_RUNNING);
3146 if (signal_pending_state(state, current))
3147 set_current_state(TASK_RUNNING);
3149 if (current->state == TASK_RUNNING)
3156 __set_current_state(TASK_RUNNING);
3160 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3162 struct blk_mq_hw_ctx *hctx;
3165 if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3168 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3169 if (!blk_qc_t_is_internal(cookie))
3170 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3172 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3174 * With scheduling, if the request has completed, we'll
3175 * get a NULL return here, as we clear the sched tag when
3176 * that happens. The request still remains valid, like always,
3177 * so we should be safe with just the NULL check.
3183 return __blk_mq_poll(hctx, rq);
3186 static int __init blk_mq_init(void)
3188 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3189 blk_mq_hctx_notify_dead);
3192 subsys_initcall(blk_mq_init);