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_ns = ktime_get_ns();
313 rq->io_start_time_ns = 0;
314 rq->nr_phys_segments = 0;
315 #if defined(CONFIG_BLK_DEV_INTEGRITY)
316 rq->nr_integrity_segments = 0;
319 /* tag was already set */
323 INIT_LIST_HEAD(&rq->timeout_list);
327 rq->end_io_data = NULL;
330 #ifdef CONFIG_BLK_CGROUP
334 data->ctx->rq_dispatched[op_is_sync(op)]++;
338 static struct request *blk_mq_get_request(struct request_queue *q,
339 struct bio *bio, unsigned int op,
340 struct blk_mq_alloc_data *data)
342 struct elevator_queue *e = q->elevator;
345 bool put_ctx_on_error = false;
347 blk_queue_enter_live(q);
349 if (likely(!data->ctx)) {
350 data->ctx = blk_mq_get_ctx(q);
351 put_ctx_on_error = true;
353 if (likely(!data->hctx))
354 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
356 data->flags |= BLK_MQ_REQ_NOWAIT;
359 data->flags |= BLK_MQ_REQ_INTERNAL;
362 * Flush requests are special and go directly to the
363 * dispatch list. Don't include reserved tags in the
364 * limiting, as it isn't useful.
366 if (!op_is_flush(op) && e->type->ops.mq.limit_depth &&
367 !(data->flags & BLK_MQ_REQ_RESERVED))
368 e->type->ops.mq.limit_depth(op, data);
371 tag = blk_mq_get_tag(data);
372 if (tag == BLK_MQ_TAG_FAIL) {
373 if (put_ctx_on_error) {
374 blk_mq_put_ctx(data->ctx);
381 rq = blk_mq_rq_ctx_init(data, tag, op);
382 if (!op_is_flush(op)) {
384 if (e && e->type->ops.mq.prepare_request) {
385 if (e->type->icq_cache && rq_ioc(bio))
386 blk_mq_sched_assign_ioc(rq, bio);
388 e->type->ops.mq.prepare_request(rq, bio);
389 rq->rq_flags |= RQF_ELVPRIV;
392 data->hctx->queued++;
396 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
397 blk_mq_req_flags_t flags)
399 struct blk_mq_alloc_data alloc_data = { .flags = flags };
403 ret = blk_queue_enter(q, flags);
407 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
411 return ERR_PTR(-EWOULDBLOCK);
413 blk_mq_put_ctx(alloc_data.ctx);
416 rq->__sector = (sector_t) -1;
417 rq->bio = rq->biotail = NULL;
420 EXPORT_SYMBOL(blk_mq_alloc_request);
422 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
423 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
425 struct blk_mq_alloc_data alloc_data = { .flags = flags };
431 * If the tag allocator sleeps we could get an allocation for a
432 * different hardware context. No need to complicate the low level
433 * allocator for this for the rare use case of a command tied to
436 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
437 return ERR_PTR(-EINVAL);
439 if (hctx_idx >= q->nr_hw_queues)
440 return ERR_PTR(-EIO);
442 ret = blk_queue_enter(q, flags);
447 * Check if the hardware context is actually mapped to anything.
448 * If not tell the caller that it should skip this queue.
450 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
451 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
453 return ERR_PTR(-EXDEV);
455 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
456 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
458 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
462 return ERR_PTR(-EWOULDBLOCK);
466 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
468 void blk_mq_free_request(struct request *rq)
470 struct request_queue *q = rq->q;
471 struct elevator_queue *e = q->elevator;
472 struct blk_mq_ctx *ctx = rq->mq_ctx;
473 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
474 const int sched_tag = rq->internal_tag;
476 if (rq->rq_flags & RQF_ELVPRIV) {
477 if (e && e->type->ops.mq.finish_request)
478 e->type->ops.mq.finish_request(rq);
480 put_io_context(rq->elv.icq->ioc);
485 ctx->rq_completed[rq_is_sync(rq)]++;
486 if (rq->rq_flags & RQF_MQ_INFLIGHT)
487 atomic_dec(&hctx->nr_active);
489 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
490 laptop_io_completion(q->backing_dev_info);
492 wbt_done(q->rq_wb, rq);
495 blk_put_rl(blk_rq_rl(rq));
497 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
499 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
501 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
502 blk_mq_sched_restart(hctx);
505 EXPORT_SYMBOL_GPL(blk_mq_free_request);
507 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
509 u64 now = ktime_get_ns();
511 if (rq->rq_flags & RQF_STATS) {
512 blk_mq_poll_stats_start(rq->q);
513 blk_stat_add(rq, now);
516 blk_account_io_done(rq, now);
519 wbt_done(rq->q->rq_wb, rq);
520 rq->end_io(rq, error);
522 if (unlikely(blk_bidi_rq(rq)))
523 blk_mq_free_request(rq->next_rq);
524 blk_mq_free_request(rq);
527 EXPORT_SYMBOL(__blk_mq_end_request);
529 void blk_mq_end_request(struct request *rq, blk_status_t error)
531 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
533 __blk_mq_end_request(rq, error);
535 EXPORT_SYMBOL(blk_mq_end_request);
537 static void __blk_mq_complete_request_remote(void *data)
539 struct request *rq = data;
541 rq->q->softirq_done_fn(rq);
544 static void __blk_mq_complete_request(struct request *rq)
546 struct blk_mq_ctx *ctx = rq->mq_ctx;
550 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT);
551 blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE);
553 if (rq->internal_tag != -1)
554 blk_mq_sched_completed_request(rq);
556 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
557 rq->q->softirq_done_fn(rq);
562 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
563 shared = cpus_share_cache(cpu, ctx->cpu);
565 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
566 rq->csd.func = __blk_mq_complete_request_remote;
569 smp_call_function_single_async(ctx->cpu, &rq->csd);
571 rq->q->softirq_done_fn(rq);
576 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
577 __releases(hctx->srcu)
579 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
582 srcu_read_unlock(hctx->srcu, srcu_idx);
585 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
586 __acquires(hctx->srcu)
588 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
589 /* shut up gcc false positive */
593 *srcu_idx = srcu_read_lock(hctx->srcu);
596 static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate)
601 * blk_mq_rq_aborted_gstate() is used from the completion path and
602 * can thus be called from irq context. u64_stats_fetch in the
603 * middle of update on the same CPU leads to lockup. Disable irq
606 local_irq_save(flags);
607 u64_stats_update_begin(&rq->aborted_gstate_sync);
608 rq->aborted_gstate = gstate;
609 u64_stats_update_end(&rq->aborted_gstate_sync);
610 local_irq_restore(flags);
613 static u64 blk_mq_rq_aborted_gstate(struct request *rq)
619 start = u64_stats_fetch_begin(&rq->aborted_gstate_sync);
620 aborted_gstate = rq->aborted_gstate;
621 } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start));
623 return aborted_gstate;
627 * blk_mq_complete_request - end I/O on a request
628 * @rq: the request being processed
631 * Ends all I/O on a request. It does not handle partial completions.
632 * The actual completion happens out-of-order, through a IPI handler.
634 void blk_mq_complete_request(struct request *rq)
636 struct request_queue *q = rq->q;
637 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
640 if (unlikely(blk_should_fake_timeout(q)))
644 * If @rq->aborted_gstate equals the current instance, timeout is
645 * claiming @rq and we lost. This is synchronized through
646 * hctx_lock(). See blk_mq_timeout_work() for details.
648 * Completion path never blocks and we can directly use RCU here
649 * instead of hctx_lock() which can be either RCU or SRCU.
650 * However, that would complicate paths which want to synchronize
651 * against us. Let stay in sync with the issue path so that
652 * hctx_lock() covers both issue and completion paths.
654 hctx_lock(hctx, &srcu_idx);
655 if (blk_mq_rq_aborted_gstate(rq) != rq->gstate)
656 __blk_mq_complete_request(rq);
657 hctx_unlock(hctx, srcu_idx);
659 EXPORT_SYMBOL(blk_mq_complete_request);
661 int blk_mq_request_started(struct request *rq)
663 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
665 EXPORT_SYMBOL_GPL(blk_mq_request_started);
667 void blk_mq_start_request(struct request *rq)
669 struct request_queue *q = rq->q;
671 blk_mq_sched_started_request(rq);
673 trace_block_rq_issue(q, rq);
675 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
676 rq->io_start_time_ns = ktime_get_ns();
677 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
678 rq->throtl_size = blk_rq_sectors(rq);
680 rq->rq_flags |= RQF_STATS;
681 wbt_issue(q->rq_wb, rq);
684 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
687 * Mark @rq in-flight which also advances the generation number,
688 * and register for timeout. Protect with a seqcount to allow the
689 * timeout path to read both @rq->gstate and @rq->deadline
692 * This is the only place where a request is marked in-flight. If
693 * the timeout path reads an in-flight @rq->gstate, the
694 * @rq->deadline it reads together under @rq->gstate_seq is
695 * guaranteed to be the matching one.
698 write_seqcount_begin(&rq->gstate_seq);
700 blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT);
703 write_seqcount_end(&rq->gstate_seq);
706 if (q->dma_drain_size && blk_rq_bytes(rq)) {
708 * Make sure space for the drain appears. We know we can do
709 * this because max_hw_segments has been adjusted to be one
710 * fewer than the device can handle.
712 rq->nr_phys_segments++;
715 EXPORT_SYMBOL(blk_mq_start_request);
718 * When we reach here because queue is busy, it's safe to change the state
719 * to IDLE without checking @rq->aborted_gstate because we should still be
720 * holding the RCU read lock and thus protected against timeout.
722 static void __blk_mq_requeue_request(struct request *rq)
724 struct request_queue *q = rq->q;
726 blk_mq_put_driver_tag(rq);
728 trace_block_rq_requeue(q, rq);
729 wbt_requeue(q->rq_wb, rq);
731 if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) {
732 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
733 if (q->dma_drain_size && blk_rq_bytes(rq))
734 rq->nr_phys_segments--;
738 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
740 __blk_mq_requeue_request(rq);
742 /* this request will be re-inserted to io scheduler queue */
743 blk_mq_sched_requeue_request(rq);
745 BUG_ON(blk_queued_rq(rq));
746 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
748 EXPORT_SYMBOL(blk_mq_requeue_request);
750 static void blk_mq_requeue_work(struct work_struct *work)
752 struct request_queue *q =
753 container_of(work, struct request_queue, requeue_work.work);
755 struct request *rq, *next;
757 spin_lock_irq(&q->requeue_lock);
758 list_splice_init(&q->requeue_list, &rq_list);
759 spin_unlock_irq(&q->requeue_lock);
761 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
762 if (!(rq->rq_flags & RQF_SOFTBARRIER))
765 rq->rq_flags &= ~RQF_SOFTBARRIER;
766 list_del_init(&rq->queuelist);
767 blk_mq_sched_insert_request(rq, true, false, false);
770 while (!list_empty(&rq_list)) {
771 rq = list_entry(rq_list.next, struct request, queuelist);
772 list_del_init(&rq->queuelist);
773 blk_mq_sched_insert_request(rq, false, false, false);
776 blk_mq_run_hw_queues(q, false);
779 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
780 bool kick_requeue_list)
782 struct request_queue *q = rq->q;
786 * We abuse this flag that is otherwise used by the I/O scheduler to
787 * request head insertion from the workqueue.
789 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
791 spin_lock_irqsave(&q->requeue_lock, flags);
793 rq->rq_flags |= RQF_SOFTBARRIER;
794 list_add(&rq->queuelist, &q->requeue_list);
796 list_add_tail(&rq->queuelist, &q->requeue_list);
798 spin_unlock_irqrestore(&q->requeue_lock, flags);
800 if (kick_requeue_list)
801 blk_mq_kick_requeue_list(q);
803 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
805 void blk_mq_kick_requeue_list(struct request_queue *q)
807 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
809 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
811 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
814 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
815 msecs_to_jiffies(msecs));
817 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
819 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
821 if (tag < tags->nr_tags) {
822 prefetch(tags->rqs[tag]);
823 return tags->rqs[tag];
828 EXPORT_SYMBOL(blk_mq_tag_to_rq);
830 struct blk_mq_timeout_data {
832 unsigned int next_set;
833 unsigned int nr_expired;
836 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
838 const struct blk_mq_ops *ops = req->q->mq_ops;
839 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
841 req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED;
844 ret = ops->timeout(req, reserved);
848 __blk_mq_complete_request(req);
850 case BLK_EH_RESET_TIMER:
852 * As nothing prevents from completion happening while
853 * ->aborted_gstate is set, this may lead to ignored
854 * completions and further spurious timeouts.
856 blk_mq_rq_update_aborted_gstate(req, 0);
859 case BLK_EH_NOT_HANDLED:
862 printk(KERN_ERR "block: bad eh return: %d\n", ret);
867 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
868 struct request *rq, void *priv, bool reserved)
870 struct blk_mq_timeout_data *data = priv;
871 unsigned long gstate, deadline;
876 if (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED)
879 /* read coherent snapshots of @rq->state_gen and @rq->deadline */
881 start = read_seqcount_begin(&rq->gstate_seq);
882 gstate = READ_ONCE(rq->gstate);
883 deadline = blk_rq_deadline(rq);
884 if (!read_seqcount_retry(&rq->gstate_seq, start))
889 /* if in-flight && overdue, mark for abortion */
890 if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT &&
891 time_after_eq(jiffies, deadline)) {
892 blk_mq_rq_update_aborted_gstate(rq, gstate);
895 } else if (!data->next_set || time_after(data->next, deadline)) {
896 data->next = deadline;
901 static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx,
902 struct request *rq, void *priv, bool reserved)
905 * We marked @rq->aborted_gstate and waited for RCU. If there were
906 * completions that we lost to, they would have finished and
907 * updated @rq->gstate by now; otherwise, the completion path is
908 * now guaranteed to see @rq->aborted_gstate and yield. If
909 * @rq->aborted_gstate still matches @rq->gstate, @rq is ours.
911 if (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) &&
912 READ_ONCE(rq->gstate) == rq->aborted_gstate)
913 blk_mq_rq_timed_out(rq, reserved);
916 static void blk_mq_timeout_work(struct work_struct *work)
918 struct request_queue *q =
919 container_of(work, struct request_queue, timeout_work);
920 struct blk_mq_timeout_data data = {
925 struct blk_mq_hw_ctx *hctx;
928 /* A deadlock might occur if a request is stuck requiring a
929 * timeout at the same time a queue freeze is waiting
930 * completion, since the timeout code would not be able to
931 * acquire the queue reference here.
933 * That's why we don't use blk_queue_enter here; instead, we use
934 * percpu_ref_tryget directly, because we need to be able to
935 * obtain a reference even in the short window between the queue
936 * starting to freeze, by dropping the first reference in
937 * blk_freeze_queue_start, and the moment the last request is
938 * consumed, marked by the instant q_usage_counter reaches
941 if (!percpu_ref_tryget(&q->q_usage_counter))
944 /* scan for the expired ones and set their ->aborted_gstate */
945 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
947 if (data.nr_expired) {
948 bool has_rcu = false;
951 * Wait till everyone sees ->aborted_gstate. The
952 * sequential waits for SRCUs aren't ideal. If this ever
953 * becomes a problem, we can add per-hw_ctx rcu_head and
956 queue_for_each_hw_ctx(q, hctx, i) {
957 if (!hctx->nr_expired)
960 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
963 synchronize_srcu(hctx->srcu);
965 hctx->nr_expired = 0;
970 /* terminate the ones we won */
971 blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL);
975 data.next = blk_rq_timeout(round_jiffies_up(data.next));
976 mod_timer(&q->timeout, data.next);
979 * Request timeouts are handled as a forward rolling timer. If
980 * we end up here it means that no requests are pending and
981 * also that no request has been pending for a while. Mark
984 queue_for_each_hw_ctx(q, hctx, i) {
985 /* the hctx may be unmapped, so check it here */
986 if (blk_mq_hw_queue_mapped(hctx))
987 blk_mq_tag_idle(hctx);
993 struct flush_busy_ctx_data {
994 struct blk_mq_hw_ctx *hctx;
995 struct list_head *list;
998 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1000 struct flush_busy_ctx_data *flush_data = data;
1001 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1002 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1004 spin_lock(&ctx->lock);
1005 list_splice_tail_init(&ctx->rq_list, flush_data->list);
1006 sbitmap_clear_bit(sb, bitnr);
1007 spin_unlock(&ctx->lock);
1012 * Process software queues that have been marked busy, splicing them
1013 * to the for-dispatch
1015 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1017 struct flush_busy_ctx_data data = {
1022 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1024 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1026 struct dispatch_rq_data {
1027 struct blk_mq_hw_ctx *hctx;
1031 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1034 struct dispatch_rq_data *dispatch_data = data;
1035 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1036 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1038 spin_lock(&ctx->lock);
1039 if (unlikely(!list_empty(&ctx->rq_list))) {
1040 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
1041 list_del_init(&dispatch_data->rq->queuelist);
1042 if (list_empty(&ctx->rq_list))
1043 sbitmap_clear_bit(sb, bitnr);
1045 spin_unlock(&ctx->lock);
1047 return !dispatch_data->rq;
1050 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1051 struct blk_mq_ctx *start)
1053 unsigned off = start ? start->index_hw : 0;
1054 struct dispatch_rq_data data = {
1059 __sbitmap_for_each_set(&hctx->ctx_map, off,
1060 dispatch_rq_from_ctx, &data);
1065 static inline unsigned int queued_to_index(unsigned int queued)
1070 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1073 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
1076 struct blk_mq_alloc_data data = {
1078 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
1079 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
1082 might_sleep_if(wait);
1087 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1088 data.flags |= BLK_MQ_REQ_RESERVED;
1090 rq->tag = blk_mq_get_tag(&data);
1092 if (blk_mq_tag_busy(data.hctx)) {
1093 rq->rq_flags |= RQF_MQ_INFLIGHT;
1094 atomic_inc(&data.hctx->nr_active);
1096 data.hctx->tags->rqs[rq->tag] = rq;
1102 return rq->tag != -1;
1105 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1106 int flags, void *key)
1108 struct blk_mq_hw_ctx *hctx;
1110 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1112 list_del_init(&wait->entry);
1113 blk_mq_run_hw_queue(hctx, true);
1118 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1119 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1120 * restart. For both cases, take care to check the condition again after
1121 * marking us as waiting.
1123 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1126 struct blk_mq_hw_ctx *this_hctx = *hctx;
1127 struct sbq_wait_state *ws;
1128 wait_queue_entry_t *wait;
1131 if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1132 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1133 set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1136 * It's possible that a tag was freed in the window between the
1137 * allocation failure and adding the hardware queue to the wait
1140 * Don't clear RESTART here, someone else could have set it.
1141 * At most this will cost an extra queue run.
1143 return blk_mq_get_driver_tag(rq, hctx, false);
1146 wait = &this_hctx->dispatch_wait;
1147 if (!list_empty_careful(&wait->entry))
1150 spin_lock(&this_hctx->lock);
1151 if (!list_empty(&wait->entry)) {
1152 spin_unlock(&this_hctx->lock);
1156 ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1157 add_wait_queue(&ws->wait, wait);
1160 * It's possible that a tag was freed in the window between the
1161 * allocation failure and adding the hardware queue to the wait
1164 ret = blk_mq_get_driver_tag(rq, hctx, false);
1166 spin_unlock(&this_hctx->lock);
1171 * We got a tag, remove ourselves from the wait queue to ensure
1172 * someone else gets the wakeup.
1174 spin_lock_irq(&ws->wait.lock);
1175 list_del_init(&wait->entry);
1176 spin_unlock_irq(&ws->wait.lock);
1177 spin_unlock(&this_hctx->lock);
1182 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1184 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1187 struct blk_mq_hw_ctx *hctx;
1188 struct request *rq, *nxt;
1189 bool no_tag = false;
1191 blk_status_t ret = BLK_STS_OK;
1193 if (list_empty(list))
1196 WARN_ON(!list_is_singular(list) && got_budget);
1199 * Now process all the entries, sending them to the driver.
1201 errors = queued = 0;
1203 struct blk_mq_queue_data bd;
1205 rq = list_first_entry(list, struct request, queuelist);
1207 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1208 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1211 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1213 * The initial allocation attempt failed, so we need to
1214 * rerun the hardware queue when a tag is freed. The
1215 * waitqueue takes care of that. If the queue is run
1216 * before we add this entry back on the dispatch list,
1217 * we'll re-run it below.
1219 if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1220 blk_mq_put_dispatch_budget(hctx);
1222 * For non-shared tags, the RESTART check
1225 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1231 list_del_init(&rq->queuelist);
1236 * Flag last if we have no more requests, or if we have more
1237 * but can't assign a driver tag to it.
1239 if (list_empty(list))
1242 nxt = list_first_entry(list, struct request, queuelist);
1243 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1246 ret = q->mq_ops->queue_rq(hctx, &bd);
1247 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1249 * If an I/O scheduler has been configured and we got a
1250 * driver tag for the next request already, free it
1253 if (!list_empty(list)) {
1254 nxt = list_first_entry(list, struct request, queuelist);
1255 blk_mq_put_driver_tag(nxt);
1257 list_add(&rq->queuelist, list);
1258 __blk_mq_requeue_request(rq);
1262 if (unlikely(ret != BLK_STS_OK)) {
1264 blk_mq_end_request(rq, BLK_STS_IOERR);
1269 } while (!list_empty(list));
1271 hctx->dispatched[queued_to_index(queued)]++;
1274 * Any items that need requeuing? Stuff them into hctx->dispatch,
1275 * that is where we will continue on next queue run.
1277 if (!list_empty(list)) {
1280 spin_lock(&hctx->lock);
1281 list_splice_init(list, &hctx->dispatch);
1282 spin_unlock(&hctx->lock);
1285 * If SCHED_RESTART was set by the caller of this function and
1286 * it is no longer set that means that it was cleared by another
1287 * thread and hence that a queue rerun is needed.
1289 * If 'no_tag' is set, that means that we failed getting
1290 * a driver tag with an I/O scheduler attached. If our dispatch
1291 * waitqueue is no longer active, ensure that we run the queue
1292 * AFTER adding our entries back to the list.
1294 * If no I/O scheduler has been configured it is possible that
1295 * the hardware queue got stopped and restarted before requests
1296 * were pushed back onto the dispatch list. Rerun the queue to
1297 * avoid starvation. Notes:
1298 * - blk_mq_run_hw_queue() checks whether or not a queue has
1299 * been stopped before rerunning a queue.
1300 * - Some but not all block drivers stop a queue before
1301 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1304 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1305 * bit is set, run queue after a delay to avoid IO stalls
1306 * that could otherwise occur if the queue is idle.
1308 needs_restart = blk_mq_sched_needs_restart(hctx);
1309 if (!needs_restart ||
1310 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1311 blk_mq_run_hw_queue(hctx, true);
1312 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1313 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1316 return (queued + errors) != 0;
1319 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1324 * We should be running this queue from one of the CPUs that
1327 * There are at least two related races now between setting
1328 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1329 * __blk_mq_run_hw_queue():
1331 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1332 * but later it becomes online, then this warning is harmless
1335 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1336 * but later it becomes offline, then the warning can't be
1337 * triggered, and we depend on blk-mq timeout handler to
1338 * handle dispatched requests to this hctx
1340 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1341 cpu_online(hctx->next_cpu)) {
1342 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1343 raw_smp_processor_id(),
1344 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1349 * We can't run the queue inline with ints disabled. Ensure that
1350 * we catch bad users of this early.
1352 WARN_ON_ONCE(in_interrupt());
1354 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1356 hctx_lock(hctx, &srcu_idx);
1357 blk_mq_sched_dispatch_requests(hctx);
1358 hctx_unlock(hctx, srcu_idx);
1361 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1363 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1365 if (cpu >= nr_cpu_ids)
1366 cpu = cpumask_first(hctx->cpumask);
1371 * It'd be great if the workqueue API had a way to pass
1372 * in a mask and had some smarts for more clever placement.
1373 * For now we just round-robin here, switching for every
1374 * BLK_MQ_CPU_WORK_BATCH queued items.
1376 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1379 int next_cpu = hctx->next_cpu;
1381 if (hctx->queue->nr_hw_queues == 1)
1382 return WORK_CPU_UNBOUND;
1384 if (--hctx->next_cpu_batch <= 0) {
1386 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1388 if (next_cpu >= nr_cpu_ids)
1389 next_cpu = blk_mq_first_mapped_cpu(hctx);
1390 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1394 * Do unbound schedule if we can't find a online CPU for this hctx,
1395 * and it should only happen in the path of handling CPU DEAD.
1397 if (!cpu_online(next_cpu)) {
1404 * Make sure to re-select CPU next time once after CPUs
1405 * in hctx->cpumask become online again.
1407 hctx->next_cpu = next_cpu;
1408 hctx->next_cpu_batch = 1;
1409 return WORK_CPU_UNBOUND;
1412 hctx->next_cpu = next_cpu;
1416 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1417 unsigned long msecs)
1419 if (unlikely(blk_mq_hctx_stopped(hctx)))
1422 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1423 int cpu = get_cpu();
1424 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1425 __blk_mq_run_hw_queue(hctx);
1433 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1434 msecs_to_jiffies(msecs));
1437 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1439 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1441 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1443 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1449 * When queue is quiesced, we may be switching io scheduler, or
1450 * updating nr_hw_queues, or other things, and we can't run queue
1451 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1453 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1456 hctx_lock(hctx, &srcu_idx);
1457 need_run = !blk_queue_quiesced(hctx->queue) &&
1458 blk_mq_hctx_has_pending(hctx);
1459 hctx_unlock(hctx, srcu_idx);
1462 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1468 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1470 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1472 struct blk_mq_hw_ctx *hctx;
1475 queue_for_each_hw_ctx(q, hctx, i) {
1476 if (blk_mq_hctx_stopped(hctx))
1479 blk_mq_run_hw_queue(hctx, async);
1482 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1485 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1486 * @q: request queue.
1488 * The caller is responsible for serializing this function against
1489 * blk_mq_{start,stop}_hw_queue().
1491 bool blk_mq_queue_stopped(struct request_queue *q)
1493 struct blk_mq_hw_ctx *hctx;
1496 queue_for_each_hw_ctx(q, hctx, i)
1497 if (blk_mq_hctx_stopped(hctx))
1502 EXPORT_SYMBOL(blk_mq_queue_stopped);
1505 * This function is often used for pausing .queue_rq() by driver when
1506 * there isn't enough resource or some conditions aren't satisfied, and
1507 * BLK_STS_RESOURCE is usually returned.
1509 * We do not guarantee that dispatch can be drained or blocked
1510 * after blk_mq_stop_hw_queue() returns. Please use
1511 * blk_mq_quiesce_queue() for that requirement.
1513 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1515 cancel_delayed_work(&hctx->run_work);
1517 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1519 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1522 * This function is often used for pausing .queue_rq() by driver when
1523 * there isn't enough resource or some conditions aren't satisfied, and
1524 * BLK_STS_RESOURCE is usually returned.
1526 * We do not guarantee that dispatch can be drained or blocked
1527 * after blk_mq_stop_hw_queues() returns. Please use
1528 * blk_mq_quiesce_queue() for that requirement.
1530 void blk_mq_stop_hw_queues(struct request_queue *q)
1532 struct blk_mq_hw_ctx *hctx;
1535 queue_for_each_hw_ctx(q, hctx, i)
1536 blk_mq_stop_hw_queue(hctx);
1538 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1540 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1542 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1544 blk_mq_run_hw_queue(hctx, false);
1546 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1548 void blk_mq_start_hw_queues(struct request_queue *q)
1550 struct blk_mq_hw_ctx *hctx;
1553 queue_for_each_hw_ctx(q, hctx, i)
1554 blk_mq_start_hw_queue(hctx);
1556 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1558 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1560 if (!blk_mq_hctx_stopped(hctx))
1563 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1564 blk_mq_run_hw_queue(hctx, async);
1566 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1568 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1570 struct blk_mq_hw_ctx *hctx;
1573 queue_for_each_hw_ctx(q, hctx, i)
1574 blk_mq_start_stopped_hw_queue(hctx, async);
1576 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1578 static void blk_mq_run_work_fn(struct work_struct *work)
1580 struct blk_mq_hw_ctx *hctx;
1582 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1585 * If we are stopped, don't run the queue.
1587 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1588 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1590 __blk_mq_run_hw_queue(hctx);
1593 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1597 struct blk_mq_ctx *ctx = rq->mq_ctx;
1599 lockdep_assert_held(&ctx->lock);
1601 trace_block_rq_insert(hctx->queue, rq);
1604 list_add(&rq->queuelist, &ctx->rq_list);
1606 list_add_tail(&rq->queuelist, &ctx->rq_list);
1609 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1612 struct blk_mq_ctx *ctx = rq->mq_ctx;
1614 lockdep_assert_held(&ctx->lock);
1616 __blk_mq_insert_req_list(hctx, rq, at_head);
1617 blk_mq_hctx_mark_pending(hctx, ctx);
1621 * Should only be used carefully, when the caller knows we want to
1622 * bypass a potential IO scheduler on the target device.
1624 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1626 struct blk_mq_ctx *ctx = rq->mq_ctx;
1627 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1629 spin_lock(&hctx->lock);
1630 list_add_tail(&rq->queuelist, &hctx->dispatch);
1631 spin_unlock(&hctx->lock);
1634 blk_mq_run_hw_queue(hctx, false);
1637 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1638 struct list_head *list)
1642 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1645 spin_lock(&ctx->lock);
1646 while (!list_empty(list)) {
1649 rq = list_first_entry(list, struct request, queuelist);
1650 BUG_ON(rq->mq_ctx != ctx);
1651 list_del_init(&rq->queuelist);
1652 __blk_mq_insert_req_list(hctx, rq, false);
1654 blk_mq_hctx_mark_pending(hctx, ctx);
1655 spin_unlock(&ctx->lock);
1658 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1660 struct request *rqa = container_of(a, struct request, queuelist);
1661 struct request *rqb = container_of(b, struct request, queuelist);
1663 return !(rqa->mq_ctx < rqb->mq_ctx ||
1664 (rqa->mq_ctx == rqb->mq_ctx &&
1665 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1668 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1670 struct blk_mq_ctx *this_ctx;
1671 struct request_queue *this_q;
1674 LIST_HEAD(ctx_list);
1677 list_splice_init(&plug->mq_list, &list);
1679 list_sort(NULL, &list, plug_ctx_cmp);
1685 while (!list_empty(&list)) {
1686 rq = list_entry_rq(list.next);
1687 list_del_init(&rq->queuelist);
1689 if (rq->mq_ctx != this_ctx) {
1691 trace_block_unplug(this_q, depth, from_schedule);
1692 blk_mq_sched_insert_requests(this_q, this_ctx,
1697 this_ctx = rq->mq_ctx;
1703 list_add_tail(&rq->queuelist, &ctx_list);
1707 * If 'this_ctx' is set, we know we have entries to complete
1708 * on 'ctx_list'. Do those.
1711 trace_block_unplug(this_q, depth, from_schedule);
1712 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1717 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1719 blk_init_request_from_bio(rq, bio);
1721 blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1723 blk_account_io_start(rq, true);
1726 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1727 struct blk_mq_ctx *ctx,
1730 spin_lock(&ctx->lock);
1731 __blk_mq_insert_request(hctx, rq, false);
1732 spin_unlock(&ctx->lock);
1735 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1738 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1740 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1743 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1747 struct request_queue *q = rq->q;
1748 struct blk_mq_queue_data bd = {
1752 blk_qc_t new_cookie;
1755 new_cookie = request_to_qc_t(hctx, rq);
1758 * For OK queue, we are done. For error, caller may kill it.
1759 * Any other error (busy), just add it to our list as we
1760 * previously would have done.
1762 ret = q->mq_ops->queue_rq(hctx, &bd);
1765 *cookie = new_cookie;
1767 case BLK_STS_RESOURCE:
1768 case BLK_STS_DEV_RESOURCE:
1769 __blk_mq_requeue_request(rq);
1772 *cookie = BLK_QC_T_NONE;
1779 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1784 struct request_queue *q = rq->q;
1785 bool run_queue = true;
1788 * RCU or SRCU read lock is needed before checking quiesced flag.
1790 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1791 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1792 * and avoid driver to try to dispatch again.
1794 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1796 bypass_insert = false;
1800 if (q->elevator && !bypass_insert)
1803 if (!blk_mq_get_dispatch_budget(hctx))
1806 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1807 blk_mq_put_dispatch_budget(hctx);
1811 return __blk_mq_issue_directly(hctx, rq, cookie);
1814 return BLK_STS_RESOURCE;
1816 blk_mq_sched_insert_request(rq, false, run_queue, false);
1820 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1821 struct request *rq, blk_qc_t *cookie)
1826 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1828 hctx_lock(hctx, &srcu_idx);
1830 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1831 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1832 blk_mq_sched_insert_request(rq, false, true, false);
1833 else if (ret != BLK_STS_OK)
1834 blk_mq_end_request(rq, ret);
1836 hctx_unlock(hctx, srcu_idx);
1839 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1843 blk_qc_t unused_cookie;
1844 struct blk_mq_ctx *ctx = rq->mq_ctx;
1845 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1847 hctx_lock(hctx, &srcu_idx);
1848 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1849 hctx_unlock(hctx, srcu_idx);
1854 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1856 const int is_sync = op_is_sync(bio->bi_opf);
1857 const int is_flush_fua = op_is_flush(bio->bi_opf);
1858 struct blk_mq_alloc_data data = { .flags = 0 };
1860 unsigned int request_count = 0;
1861 struct blk_plug *plug;
1862 struct request *same_queue_rq = NULL;
1864 unsigned int wb_acct;
1866 blk_queue_bounce(q, &bio);
1868 blk_queue_split(q, &bio);
1870 if (!bio_integrity_prep(bio))
1871 return BLK_QC_T_NONE;
1873 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1874 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1875 return BLK_QC_T_NONE;
1877 if (blk_mq_sched_bio_merge(q, bio))
1878 return BLK_QC_T_NONE;
1880 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1882 trace_block_getrq(q, bio, bio->bi_opf);
1884 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1885 if (unlikely(!rq)) {
1886 __wbt_done(q->rq_wb, wb_acct);
1887 if (bio->bi_opf & REQ_NOWAIT)
1888 bio_wouldblock_error(bio);
1889 return BLK_QC_T_NONE;
1892 wbt_track(rq, wb_acct);
1894 cookie = request_to_qc_t(data.hctx, rq);
1896 plug = current->plug;
1897 if (unlikely(is_flush_fua)) {
1898 blk_mq_put_ctx(data.ctx);
1899 blk_mq_bio_to_request(rq, bio);
1901 /* bypass scheduler for flush rq */
1902 blk_insert_flush(rq);
1903 blk_mq_run_hw_queue(data.hctx, true);
1904 } else if (plug && q->nr_hw_queues == 1) {
1905 struct request *last = NULL;
1907 blk_mq_put_ctx(data.ctx);
1908 blk_mq_bio_to_request(rq, bio);
1911 * @request_count may become stale because of schedule
1912 * out, so check the list again.
1914 if (list_empty(&plug->mq_list))
1916 else if (blk_queue_nomerges(q))
1917 request_count = blk_plug_queued_count(q);
1920 trace_block_plug(q);
1922 last = list_entry_rq(plug->mq_list.prev);
1924 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1925 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1926 blk_flush_plug_list(plug, false);
1927 trace_block_plug(q);
1930 list_add_tail(&rq->queuelist, &plug->mq_list);
1931 } else if (plug && !blk_queue_nomerges(q)) {
1932 blk_mq_bio_to_request(rq, bio);
1935 * We do limited plugging. If the bio can be merged, do that.
1936 * Otherwise the existing request in the plug list will be
1937 * issued. So the plug list will have one request at most
1938 * The plug list might get flushed before this. If that happens,
1939 * the plug list is empty, and same_queue_rq is invalid.
1941 if (list_empty(&plug->mq_list))
1942 same_queue_rq = NULL;
1944 list_del_init(&same_queue_rq->queuelist);
1945 list_add_tail(&rq->queuelist, &plug->mq_list);
1947 blk_mq_put_ctx(data.ctx);
1949 if (same_queue_rq) {
1950 data.hctx = blk_mq_map_queue(q,
1951 same_queue_rq->mq_ctx->cpu);
1952 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1955 } else if (q->nr_hw_queues > 1 && is_sync) {
1956 blk_mq_put_ctx(data.ctx);
1957 blk_mq_bio_to_request(rq, bio);
1958 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1959 } else if (q->elevator) {
1960 blk_mq_put_ctx(data.ctx);
1961 blk_mq_bio_to_request(rq, bio);
1962 blk_mq_sched_insert_request(rq, false, true, true);
1964 blk_mq_put_ctx(data.ctx);
1965 blk_mq_bio_to_request(rq, bio);
1966 blk_mq_queue_io(data.hctx, data.ctx, rq);
1967 blk_mq_run_hw_queue(data.hctx, true);
1973 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1974 unsigned int hctx_idx)
1978 if (tags->rqs && set->ops->exit_request) {
1981 for (i = 0; i < tags->nr_tags; i++) {
1982 struct request *rq = tags->static_rqs[i];
1986 set->ops->exit_request(set, rq, hctx_idx);
1987 tags->static_rqs[i] = NULL;
1991 while (!list_empty(&tags->page_list)) {
1992 page = list_first_entry(&tags->page_list, struct page, lru);
1993 list_del_init(&page->lru);
1995 * Remove kmemleak object previously allocated in
1996 * blk_mq_init_rq_map().
1998 kmemleak_free(page_address(page));
1999 __free_pages(page, page->private);
2003 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2007 kfree(tags->static_rqs);
2008 tags->static_rqs = NULL;
2010 blk_mq_free_tags(tags);
2013 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2014 unsigned int hctx_idx,
2015 unsigned int nr_tags,
2016 unsigned int reserved_tags)
2018 struct blk_mq_tags *tags;
2021 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2022 if (node == NUMA_NO_NODE)
2023 node = set->numa_node;
2025 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2026 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2030 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2031 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2034 blk_mq_free_tags(tags);
2038 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2039 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2041 if (!tags->static_rqs) {
2043 blk_mq_free_tags(tags);
2050 static size_t order_to_size(unsigned int order)
2052 return (size_t)PAGE_SIZE << order;
2055 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2056 unsigned int hctx_idx, int node)
2060 if (set->ops->init_request) {
2061 ret = set->ops->init_request(set, rq, hctx_idx, node);
2066 seqcount_init(&rq->gstate_seq);
2067 u64_stats_init(&rq->aborted_gstate_sync);
2069 * start gstate with gen 1 instead of 0, otherwise it will be equal
2070 * to aborted_gstate, and be identified timed out by
2071 * blk_mq_terminate_expired.
2073 WRITE_ONCE(rq->gstate, MQ_RQ_GEN_INC);
2078 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2079 unsigned int hctx_idx, unsigned int depth)
2081 unsigned int i, j, entries_per_page, max_order = 4;
2082 size_t rq_size, left;
2085 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2086 if (node == NUMA_NO_NODE)
2087 node = set->numa_node;
2089 INIT_LIST_HEAD(&tags->page_list);
2092 * rq_size is the size of the request plus driver payload, rounded
2093 * to the cacheline size
2095 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2097 left = rq_size * depth;
2099 for (i = 0; i < depth; ) {
2100 int this_order = max_order;
2105 while (this_order && left < order_to_size(this_order - 1))
2109 page = alloc_pages_node(node,
2110 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2116 if (order_to_size(this_order) < rq_size)
2123 page->private = this_order;
2124 list_add_tail(&page->lru, &tags->page_list);
2126 p = page_address(page);
2128 * Allow kmemleak to scan these pages as they contain pointers
2129 * to additional allocations like via ops->init_request().
2131 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2132 entries_per_page = order_to_size(this_order) / rq_size;
2133 to_do = min(entries_per_page, depth - i);
2134 left -= to_do * rq_size;
2135 for (j = 0; j < to_do; j++) {
2136 struct request *rq = p;
2138 tags->static_rqs[i] = rq;
2139 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2140 tags->static_rqs[i] = NULL;
2151 blk_mq_free_rqs(set, tags, hctx_idx);
2156 * 'cpu' is going away. splice any existing rq_list entries from this
2157 * software queue to the hw queue dispatch list, and ensure that it
2160 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2162 struct blk_mq_hw_ctx *hctx;
2163 struct blk_mq_ctx *ctx;
2166 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2167 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2169 spin_lock(&ctx->lock);
2170 if (!list_empty(&ctx->rq_list)) {
2171 list_splice_init(&ctx->rq_list, &tmp);
2172 blk_mq_hctx_clear_pending(hctx, ctx);
2174 spin_unlock(&ctx->lock);
2176 if (list_empty(&tmp))
2179 spin_lock(&hctx->lock);
2180 list_splice_tail_init(&tmp, &hctx->dispatch);
2181 spin_unlock(&hctx->lock);
2183 blk_mq_run_hw_queue(hctx, true);
2187 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2189 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2193 /* hctx->ctxs will be freed in queue's release handler */
2194 static void blk_mq_exit_hctx(struct request_queue *q,
2195 struct blk_mq_tag_set *set,
2196 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2198 blk_mq_debugfs_unregister_hctx(hctx);
2200 if (blk_mq_hw_queue_mapped(hctx))
2201 blk_mq_tag_idle(hctx);
2203 if (set->ops->exit_request)
2204 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2206 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2208 if (set->ops->exit_hctx)
2209 set->ops->exit_hctx(hctx, hctx_idx);
2211 if (hctx->flags & BLK_MQ_F_BLOCKING)
2212 cleanup_srcu_struct(hctx->srcu);
2214 blk_mq_remove_cpuhp(hctx);
2215 blk_free_flush_queue(hctx->fq);
2216 sbitmap_free(&hctx->ctx_map);
2219 static void blk_mq_exit_hw_queues(struct request_queue *q,
2220 struct blk_mq_tag_set *set, int nr_queue)
2222 struct blk_mq_hw_ctx *hctx;
2225 queue_for_each_hw_ctx(q, hctx, i) {
2228 blk_mq_exit_hctx(q, set, hctx, i);
2232 static int blk_mq_init_hctx(struct request_queue *q,
2233 struct blk_mq_tag_set *set,
2234 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2238 node = hctx->numa_node;
2239 if (node == NUMA_NO_NODE)
2240 node = hctx->numa_node = set->numa_node;
2242 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2243 spin_lock_init(&hctx->lock);
2244 INIT_LIST_HEAD(&hctx->dispatch);
2246 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2248 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2250 hctx->tags = set->tags[hctx_idx];
2253 * Allocate space for all possible cpus to avoid allocation at
2256 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2259 goto unregister_cpu_notifier;
2261 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2267 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2268 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2270 if (set->ops->init_hctx &&
2271 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2274 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2277 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2279 goto sched_exit_hctx;
2281 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2284 if (hctx->flags & BLK_MQ_F_BLOCKING)
2285 init_srcu_struct(hctx->srcu);
2287 blk_mq_debugfs_register_hctx(q, hctx);
2294 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2296 if (set->ops->exit_hctx)
2297 set->ops->exit_hctx(hctx, hctx_idx);
2299 sbitmap_free(&hctx->ctx_map);
2302 unregister_cpu_notifier:
2303 blk_mq_remove_cpuhp(hctx);
2307 static void blk_mq_init_cpu_queues(struct request_queue *q,
2308 unsigned int nr_hw_queues)
2312 for_each_possible_cpu(i) {
2313 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2314 struct blk_mq_hw_ctx *hctx;
2317 spin_lock_init(&__ctx->lock);
2318 INIT_LIST_HEAD(&__ctx->rq_list);
2322 * Set local node, IFF we have more than one hw queue. If
2323 * not, we remain on the home node of the device
2325 hctx = blk_mq_map_queue(q, i);
2326 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2327 hctx->numa_node = local_memory_node(cpu_to_node(i));
2331 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2335 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2336 set->queue_depth, set->reserved_tags);
2337 if (!set->tags[hctx_idx])
2340 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2345 blk_mq_free_rq_map(set->tags[hctx_idx]);
2346 set->tags[hctx_idx] = NULL;
2350 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2351 unsigned int hctx_idx)
2353 if (set->tags[hctx_idx]) {
2354 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2355 blk_mq_free_rq_map(set->tags[hctx_idx]);
2356 set->tags[hctx_idx] = NULL;
2360 static void blk_mq_map_swqueue(struct request_queue *q)
2362 unsigned int i, hctx_idx;
2363 struct blk_mq_hw_ctx *hctx;
2364 struct blk_mq_ctx *ctx;
2365 struct blk_mq_tag_set *set = q->tag_set;
2368 * Avoid others reading imcomplete hctx->cpumask through sysfs
2370 mutex_lock(&q->sysfs_lock);
2372 queue_for_each_hw_ctx(q, hctx, i) {
2373 cpumask_clear(hctx->cpumask);
2378 * Map software to hardware queues.
2380 * If the cpu isn't present, the cpu is mapped to first hctx.
2382 for_each_possible_cpu(i) {
2383 hctx_idx = q->mq_map[i];
2384 /* unmapped hw queue can be remapped after CPU topo changed */
2385 if (!set->tags[hctx_idx] &&
2386 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2388 * If tags initialization fail for some hctx,
2389 * that hctx won't be brought online. In this
2390 * case, remap the current ctx to hctx[0] which
2391 * is guaranteed to always have tags allocated
2396 ctx = per_cpu_ptr(q->queue_ctx, i);
2397 hctx = blk_mq_map_queue(q, i);
2399 cpumask_set_cpu(i, hctx->cpumask);
2400 ctx->index_hw = hctx->nr_ctx;
2401 hctx->ctxs[hctx->nr_ctx++] = ctx;
2404 mutex_unlock(&q->sysfs_lock);
2406 queue_for_each_hw_ctx(q, hctx, i) {
2408 * If no software queues are mapped to this hardware queue,
2409 * disable it and free the request entries.
2411 if (!hctx->nr_ctx) {
2412 /* Never unmap queue 0. We need it as a
2413 * fallback in case of a new remap fails
2416 if (i && set->tags[i])
2417 blk_mq_free_map_and_requests(set, i);
2423 hctx->tags = set->tags[i];
2424 WARN_ON(!hctx->tags);
2427 * Set the map size to the number of mapped software queues.
2428 * This is more accurate and more efficient than looping
2429 * over all possibly mapped software queues.
2431 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2434 * Initialize batch roundrobin counts
2436 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2437 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2442 * Caller needs to ensure that we're either frozen/quiesced, or that
2443 * the queue isn't live yet.
2445 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2447 struct blk_mq_hw_ctx *hctx;
2450 queue_for_each_hw_ctx(q, hctx, i) {
2452 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2453 atomic_inc(&q->shared_hctx_restart);
2454 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2456 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2457 atomic_dec(&q->shared_hctx_restart);
2458 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2463 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2466 struct request_queue *q;
2468 lockdep_assert_held(&set->tag_list_lock);
2470 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2471 blk_mq_freeze_queue(q);
2472 queue_set_hctx_shared(q, shared);
2473 blk_mq_unfreeze_queue(q);
2477 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2479 struct blk_mq_tag_set *set = q->tag_set;
2481 mutex_lock(&set->tag_list_lock);
2482 list_del_rcu(&q->tag_set_list);
2483 INIT_LIST_HEAD(&q->tag_set_list);
2484 if (list_is_singular(&set->tag_list)) {
2485 /* just transitioned to unshared */
2486 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2487 /* update existing queue */
2488 blk_mq_update_tag_set_depth(set, false);
2490 mutex_unlock(&set->tag_list_lock);
2495 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2496 struct request_queue *q)
2500 mutex_lock(&set->tag_list_lock);
2503 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2505 if (!list_empty(&set->tag_list) &&
2506 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2507 set->flags |= BLK_MQ_F_TAG_SHARED;
2508 /* update existing queue */
2509 blk_mq_update_tag_set_depth(set, true);
2511 if (set->flags & BLK_MQ_F_TAG_SHARED)
2512 queue_set_hctx_shared(q, true);
2513 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2515 mutex_unlock(&set->tag_list_lock);
2519 * It is the actual release handler for mq, but we do it from
2520 * request queue's release handler for avoiding use-after-free
2521 * and headache because q->mq_kobj shouldn't have been introduced,
2522 * but we can't group ctx/kctx kobj without it.
2524 void blk_mq_release(struct request_queue *q)
2526 struct blk_mq_hw_ctx *hctx;
2529 /* hctx kobj stays in hctx */
2530 queue_for_each_hw_ctx(q, hctx, i) {
2533 kobject_put(&hctx->kobj);
2538 kfree(q->queue_hw_ctx);
2541 * release .mq_kobj and sw queue's kobject now because
2542 * both share lifetime with request queue.
2544 blk_mq_sysfs_deinit(q);
2546 free_percpu(q->queue_ctx);
2549 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2551 struct request_queue *uninit_q, *q;
2553 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2555 return ERR_PTR(-ENOMEM);
2557 q = blk_mq_init_allocated_queue(set, uninit_q);
2559 blk_cleanup_queue(uninit_q);
2563 EXPORT_SYMBOL(blk_mq_init_queue);
2565 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2567 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2569 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2570 __alignof__(struct blk_mq_hw_ctx)) !=
2571 sizeof(struct blk_mq_hw_ctx));
2573 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2574 hw_ctx_size += sizeof(struct srcu_struct);
2579 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2580 struct request_queue *q)
2583 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2585 blk_mq_sysfs_unregister(q);
2587 /* protect against switching io scheduler */
2588 mutex_lock(&q->sysfs_lock);
2589 for (i = 0; i < set->nr_hw_queues; i++) {
2595 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2596 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2601 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2608 atomic_set(&hctxs[i]->nr_active, 0);
2609 hctxs[i]->numa_node = node;
2610 hctxs[i]->queue_num = i;
2612 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2613 free_cpumask_var(hctxs[i]->cpumask);
2618 blk_mq_hctx_kobj_init(hctxs[i]);
2620 for (j = i; j < q->nr_hw_queues; j++) {
2621 struct blk_mq_hw_ctx *hctx = hctxs[j];
2625 blk_mq_free_map_and_requests(set, j);
2626 blk_mq_exit_hctx(q, set, hctx, j);
2627 kobject_put(&hctx->kobj);
2632 q->nr_hw_queues = i;
2633 mutex_unlock(&q->sysfs_lock);
2634 blk_mq_sysfs_register(q);
2637 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2638 struct request_queue *q)
2640 /* mark the queue as mq asap */
2641 q->mq_ops = set->ops;
2643 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2644 blk_mq_poll_stats_bkt,
2645 BLK_MQ_POLL_STATS_BKTS, q);
2649 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2653 /* init q->mq_kobj and sw queues' kobjects */
2654 blk_mq_sysfs_init(q);
2656 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2657 GFP_KERNEL, set->numa_node);
2658 if (!q->queue_hw_ctx)
2661 q->mq_map = set->mq_map;
2663 blk_mq_realloc_hw_ctxs(set, q);
2664 if (!q->nr_hw_queues)
2667 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2668 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2670 q->nr_queues = nr_cpu_ids;
2672 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2674 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2675 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2677 q->sg_reserved_size = INT_MAX;
2679 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2680 INIT_LIST_HEAD(&q->requeue_list);
2681 spin_lock_init(&q->requeue_lock);
2683 blk_queue_make_request(q, blk_mq_make_request);
2684 if (q->mq_ops->poll)
2685 q->poll_fn = blk_mq_poll;
2688 * Do this after blk_queue_make_request() overrides it...
2690 q->nr_requests = set->queue_depth;
2693 * Default to classic polling
2697 if (set->ops->complete)
2698 blk_queue_softirq_done(q, set->ops->complete);
2700 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2701 blk_mq_add_queue_tag_set(set, q);
2702 blk_mq_map_swqueue(q);
2704 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2707 ret = blk_mq_sched_init(q);
2709 return ERR_PTR(ret);
2715 kfree(q->queue_hw_ctx);
2717 free_percpu(q->queue_ctx);
2720 return ERR_PTR(-ENOMEM);
2722 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2724 void blk_mq_free_queue(struct request_queue *q)
2726 struct blk_mq_tag_set *set = q->tag_set;
2728 blk_mq_del_queue_tag_set(q);
2729 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2732 /* Basically redo blk_mq_init_queue with queue frozen */
2733 static void blk_mq_queue_reinit(struct request_queue *q)
2735 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2737 blk_mq_debugfs_unregister_hctxs(q);
2738 blk_mq_sysfs_unregister(q);
2741 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2742 * we should change hctx numa_node according to the new topology (this
2743 * involves freeing and re-allocating memory, worth doing?)
2745 blk_mq_map_swqueue(q);
2747 blk_mq_sysfs_register(q);
2748 blk_mq_debugfs_register_hctxs(q);
2751 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2755 for (i = 0; i < set->nr_hw_queues; i++)
2756 if (!__blk_mq_alloc_rq_map(set, i))
2763 blk_mq_free_rq_map(set->tags[i]);
2769 * Allocate the request maps associated with this tag_set. Note that this
2770 * may reduce the depth asked for, if memory is tight. set->queue_depth
2771 * will be updated to reflect the allocated depth.
2773 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2778 depth = set->queue_depth;
2780 err = __blk_mq_alloc_rq_maps(set);
2784 set->queue_depth >>= 1;
2785 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2789 } while (set->queue_depth);
2791 if (!set->queue_depth || err) {
2792 pr_err("blk-mq: failed to allocate request map\n");
2796 if (depth != set->queue_depth)
2797 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2798 depth, set->queue_depth);
2803 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2805 if (set->ops->map_queues) {
2808 * transport .map_queues is usually done in the following
2811 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2812 * mask = get_cpu_mask(queue)
2813 * for_each_cpu(cpu, mask)
2814 * set->mq_map[cpu] = queue;
2817 * When we need to remap, the table has to be cleared for
2818 * killing stale mapping since one CPU may not be mapped
2821 for_each_possible_cpu(cpu)
2822 set->mq_map[cpu] = 0;
2824 return set->ops->map_queues(set);
2826 return blk_mq_map_queues(set);
2830 * Alloc a tag set to be associated with one or more request queues.
2831 * May fail with EINVAL for various error conditions. May adjust the
2832 * requested depth down, if if it too large. In that case, the set
2833 * value will be stored in set->queue_depth.
2835 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2839 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2841 if (!set->nr_hw_queues)
2843 if (!set->queue_depth)
2845 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2848 if (!set->ops->queue_rq)
2851 if (!set->ops->get_budget ^ !set->ops->put_budget)
2854 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2855 pr_info("blk-mq: reduced tag depth to %u\n",
2857 set->queue_depth = BLK_MQ_MAX_DEPTH;
2861 * If a crashdump is active, then we are potentially in a very
2862 * memory constrained environment. Limit us to 1 queue and
2863 * 64 tags to prevent using too much memory.
2865 if (is_kdump_kernel()) {
2866 set->nr_hw_queues = 1;
2867 set->queue_depth = min(64U, set->queue_depth);
2870 * There is no use for more h/w queues than cpus.
2872 if (set->nr_hw_queues > nr_cpu_ids)
2873 set->nr_hw_queues = nr_cpu_ids;
2875 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2876 GFP_KERNEL, set->numa_node);
2881 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2882 GFP_KERNEL, set->numa_node);
2886 ret = blk_mq_update_queue_map(set);
2888 goto out_free_mq_map;
2890 ret = blk_mq_alloc_rq_maps(set);
2892 goto out_free_mq_map;
2894 mutex_init(&set->tag_list_lock);
2895 INIT_LIST_HEAD(&set->tag_list);
2907 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2909 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2913 for (i = 0; i < nr_cpu_ids; i++)
2914 blk_mq_free_map_and_requests(set, i);
2922 EXPORT_SYMBOL(blk_mq_free_tag_set);
2924 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2926 struct blk_mq_tag_set *set = q->tag_set;
2927 struct blk_mq_hw_ctx *hctx;
2933 blk_mq_freeze_queue(q);
2934 blk_mq_quiesce_queue(q);
2937 queue_for_each_hw_ctx(q, hctx, i) {
2941 * If we're using an MQ scheduler, just update the scheduler
2942 * queue depth. This is similar to what the old code would do.
2944 if (!hctx->sched_tags) {
2945 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2948 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2956 q->nr_requests = nr;
2958 blk_mq_unquiesce_queue(q);
2959 blk_mq_unfreeze_queue(q);
2964 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2967 struct request_queue *q;
2969 lockdep_assert_held(&set->tag_list_lock);
2971 if (nr_hw_queues > nr_cpu_ids)
2972 nr_hw_queues = nr_cpu_ids;
2973 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2976 list_for_each_entry(q, &set->tag_list, tag_set_list)
2977 blk_mq_freeze_queue(q);
2979 set->nr_hw_queues = nr_hw_queues;
2980 blk_mq_update_queue_map(set);
2981 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2982 blk_mq_realloc_hw_ctxs(set, q);
2983 blk_mq_queue_reinit(q);
2986 list_for_each_entry(q, &set->tag_list, tag_set_list)
2987 blk_mq_unfreeze_queue(q);
2990 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2992 mutex_lock(&set->tag_list_lock);
2993 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2994 mutex_unlock(&set->tag_list_lock);
2996 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2998 /* Enable polling stats and return whether they were already enabled. */
2999 static bool blk_poll_stats_enable(struct request_queue *q)
3001 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3002 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3004 blk_stat_add_callback(q, q->poll_cb);
3008 static void blk_mq_poll_stats_start(struct request_queue *q)
3011 * We don't arm the callback if polling stats are not enabled or the
3012 * callback is already active.
3014 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3015 blk_stat_is_active(q->poll_cb))
3018 blk_stat_activate_msecs(q->poll_cb, 100);
3021 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3023 struct request_queue *q = cb->data;
3026 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3027 if (cb->stat[bucket].nr_samples)
3028 q->poll_stat[bucket] = cb->stat[bucket];
3032 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3033 struct blk_mq_hw_ctx *hctx,
3036 unsigned long ret = 0;
3040 * If stats collection isn't on, don't sleep but turn it on for
3043 if (!blk_poll_stats_enable(q))
3047 * As an optimistic guess, use half of the mean service time
3048 * for this type of request. We can (and should) make this smarter.
3049 * For instance, if the completion latencies are tight, we can
3050 * get closer than just half the mean. This is especially
3051 * important on devices where the completion latencies are longer
3052 * than ~10 usec. We do use the stats for the relevant IO size
3053 * if available which does lead to better estimates.
3055 bucket = blk_mq_poll_stats_bkt(rq);
3059 if (q->poll_stat[bucket].nr_samples)
3060 ret = (q->poll_stat[bucket].mean + 1) / 2;
3065 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3066 struct blk_mq_hw_ctx *hctx,
3069 struct hrtimer_sleeper hs;
3070 enum hrtimer_mode mode;
3074 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3080 * -1: don't ever hybrid sleep
3081 * 0: use half of prev avg
3082 * >0: use this specific value
3084 if (q->poll_nsec == -1)
3086 else if (q->poll_nsec > 0)
3087 nsecs = q->poll_nsec;
3089 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3094 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3097 * This will be replaced with the stats tracking code, using
3098 * 'avg_completion_time / 2' as the pre-sleep target.
3102 mode = HRTIMER_MODE_REL;
3103 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3104 hrtimer_set_expires(&hs.timer, kt);
3106 hrtimer_init_sleeper(&hs, current);
3108 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3110 set_current_state(TASK_UNINTERRUPTIBLE);
3111 hrtimer_start_expires(&hs.timer, mode);
3114 hrtimer_cancel(&hs.timer);
3115 mode = HRTIMER_MODE_ABS;
3116 } while (hs.task && !signal_pending(current));
3118 __set_current_state(TASK_RUNNING);
3119 destroy_hrtimer_on_stack(&hs.timer);
3123 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3125 struct request_queue *q = hctx->queue;
3129 * If we sleep, have the caller restart the poll loop to reset
3130 * the state. Like for the other success return cases, the
3131 * caller is responsible for checking if the IO completed. If
3132 * the IO isn't complete, we'll get called again and will go
3133 * straight to the busy poll loop.
3135 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3138 hctx->poll_considered++;
3140 state = current->state;
3141 while (!need_resched()) {
3144 hctx->poll_invoked++;
3146 ret = q->mq_ops->poll(hctx, rq->tag);
3148 hctx->poll_success++;
3149 set_current_state(TASK_RUNNING);
3153 if (signal_pending_state(state, current))
3154 set_current_state(TASK_RUNNING);
3156 if (current->state == TASK_RUNNING)
3163 __set_current_state(TASK_RUNNING);
3167 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3169 struct blk_mq_hw_ctx *hctx;
3172 if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3175 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3176 if (!blk_qc_t_is_internal(cookie))
3177 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3179 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3181 * With scheduling, if the request has completed, we'll
3182 * get a NULL return here, as we clear the sched tag when
3183 * that happens. The request still remains valid, like always,
3184 * so we should be safe with just the NULL check.
3190 return __blk_mq_poll(hctx, rq);
3193 static int __init blk_mq_init(void)
3195 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3196 blk_mq_hctx_notify_dead);
3199 subsys_initcall(blk_mq_init);