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)]++;
335 refcount_set(&rq->ref, 1);
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
364 * dispatch list. Don't include reserved tags in the
365 * limiting, as it isn't useful.
367 if (!op_is_flush(op) && e->type->ops.mq.limit_depth &&
368 !(data->flags & BLK_MQ_REQ_RESERVED))
369 e->type->ops.mq.limit_depth(op, data);
372 tag = blk_mq_get_tag(data);
373 if (tag == BLK_MQ_TAG_FAIL) {
374 if (put_ctx_on_error) {
375 blk_mq_put_ctx(data->ctx);
382 rq = blk_mq_rq_ctx_init(data, tag, op);
383 if (!op_is_flush(op)) {
385 if (e && e->type->ops.mq.prepare_request) {
386 if (e->type->icq_cache && rq_ioc(bio))
387 blk_mq_sched_assign_ioc(rq, bio);
389 e->type->ops.mq.prepare_request(rq, bio);
390 rq->rq_flags |= RQF_ELVPRIV;
393 data->hctx->queued++;
397 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
398 blk_mq_req_flags_t flags)
400 struct blk_mq_alloc_data alloc_data = { .flags = flags };
404 ret = blk_queue_enter(q, flags);
408 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
412 return ERR_PTR(-EWOULDBLOCK);
414 blk_mq_put_ctx(alloc_data.ctx);
417 rq->__sector = (sector_t) -1;
418 rq->bio = rq->biotail = NULL;
421 EXPORT_SYMBOL(blk_mq_alloc_request);
423 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
424 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
426 struct blk_mq_alloc_data alloc_data = { .flags = flags };
432 * If the tag allocator sleeps we could get an allocation for a
433 * different hardware context. No need to complicate the low level
434 * allocator for this for the rare use case of a command tied to
437 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
438 return ERR_PTR(-EINVAL);
440 if (hctx_idx >= q->nr_hw_queues)
441 return ERR_PTR(-EIO);
443 ret = blk_queue_enter(q, flags);
448 * Check if the hardware context is actually mapped to anything.
449 * If not tell the caller that it should skip this queue.
451 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
452 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
454 return ERR_PTR(-EXDEV);
456 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
457 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
459 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
463 return ERR_PTR(-EWOULDBLOCK);
467 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
469 static void __blk_mq_free_request(struct request *rq)
471 struct request_queue *q = rq->q;
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;
477 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
479 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
480 blk_mq_sched_restart(hctx);
484 void blk_mq_free_request(struct request *rq)
486 struct request_queue *q = rq->q;
487 struct elevator_queue *e = q->elevator;
488 struct blk_mq_ctx *ctx = rq->mq_ctx;
489 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
491 if (rq->rq_flags & RQF_ELVPRIV) {
492 if (e && e->type->ops.mq.finish_request)
493 e->type->ops.mq.finish_request(rq);
495 put_io_context(rq->elv.icq->ioc);
500 ctx->rq_completed[rq_is_sync(rq)]++;
501 if (rq->rq_flags & RQF_MQ_INFLIGHT)
502 atomic_dec(&hctx->nr_active);
504 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
505 laptop_io_completion(q->backing_dev_info);
507 wbt_done(q->rq_wb, rq);
510 blk_put_rl(blk_rq_rl(rq));
512 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
513 if (refcount_dec_and_test(&rq->ref))
514 __blk_mq_free_request(rq);
516 EXPORT_SYMBOL_GPL(blk_mq_free_request);
518 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
520 u64 now = ktime_get_ns();
522 if (rq->rq_flags & RQF_STATS) {
523 blk_mq_poll_stats_start(rq->q);
524 blk_stat_add(rq, now);
527 blk_account_io_done(rq, now);
530 wbt_done(rq->q->rq_wb, rq);
531 rq->end_io(rq, error);
533 if (unlikely(blk_bidi_rq(rq)))
534 blk_mq_free_request(rq->next_rq);
535 blk_mq_free_request(rq);
538 EXPORT_SYMBOL(__blk_mq_end_request);
540 void blk_mq_end_request(struct request *rq, blk_status_t error)
542 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
544 __blk_mq_end_request(rq, error);
546 EXPORT_SYMBOL(blk_mq_end_request);
548 static void __blk_mq_complete_request_remote(void *data)
550 struct request *rq = data;
552 rq->q->softirq_done_fn(rq);
555 static void __blk_mq_complete_request(struct request *rq)
557 struct blk_mq_ctx *ctx = rq->mq_ctx;
561 if (cmpxchg(&rq->state, MQ_RQ_IN_FLIGHT, MQ_RQ_COMPLETE) !=
565 if (rq->internal_tag != -1)
566 blk_mq_sched_completed_request(rq);
568 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
569 rq->q->softirq_done_fn(rq);
574 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
575 shared = cpus_share_cache(cpu, ctx->cpu);
577 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
578 rq->csd.func = __blk_mq_complete_request_remote;
581 smp_call_function_single_async(ctx->cpu, &rq->csd);
583 rq->q->softirq_done_fn(rq);
588 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
589 __releases(hctx->srcu)
591 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
594 srcu_read_unlock(hctx->srcu, srcu_idx);
597 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
598 __acquires(hctx->srcu)
600 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
601 /* shut up gcc false positive */
605 *srcu_idx = srcu_read_lock(hctx->srcu);
609 * blk_mq_complete_request - end I/O on a request
610 * @rq: the request being processed
613 * Ends all I/O on a request. It does not handle partial completions.
614 * The actual completion happens out-of-order, through a IPI handler.
616 void blk_mq_complete_request(struct request *rq)
618 if (unlikely(blk_should_fake_timeout(rq->q)))
620 __blk_mq_complete_request(rq);
622 EXPORT_SYMBOL(blk_mq_complete_request);
624 int blk_mq_request_started(struct request *rq)
626 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
628 EXPORT_SYMBOL_GPL(blk_mq_request_started);
630 void blk_mq_start_request(struct request *rq)
632 struct request_queue *q = rq->q;
634 blk_mq_sched_started_request(rq);
636 trace_block_rq_issue(q, rq);
638 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
639 rq->io_start_time_ns = ktime_get_ns();
640 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
641 rq->throtl_size = blk_rq_sectors(rq);
643 rq->rq_flags |= RQF_STATS;
644 wbt_issue(q->rq_wb, rq);
647 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
650 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
652 if (q->dma_drain_size && blk_rq_bytes(rq)) {
654 * Make sure space for the drain appears. We know we can do
655 * this because max_hw_segments has been adjusted to be one
656 * fewer than the device can handle.
658 rq->nr_phys_segments++;
661 EXPORT_SYMBOL(blk_mq_start_request);
663 static void __blk_mq_requeue_request(struct request *rq)
665 struct request_queue *q = rq->q;
667 blk_mq_put_driver_tag(rq);
669 trace_block_rq_requeue(q, rq);
670 wbt_requeue(q->rq_wb, rq);
672 if (blk_mq_request_started(rq)) {
673 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
674 if (q->dma_drain_size && blk_rq_bytes(rq))
675 rq->nr_phys_segments--;
679 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
681 __blk_mq_requeue_request(rq);
683 /* this request will be re-inserted to io scheduler queue */
684 blk_mq_sched_requeue_request(rq);
686 BUG_ON(blk_queued_rq(rq));
687 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
689 EXPORT_SYMBOL(blk_mq_requeue_request);
691 static void blk_mq_requeue_work(struct work_struct *work)
693 struct request_queue *q =
694 container_of(work, struct request_queue, requeue_work.work);
696 struct request *rq, *next;
698 spin_lock_irq(&q->requeue_lock);
699 list_splice_init(&q->requeue_list, &rq_list);
700 spin_unlock_irq(&q->requeue_lock);
702 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
703 if (!(rq->rq_flags & RQF_SOFTBARRIER))
706 rq->rq_flags &= ~RQF_SOFTBARRIER;
707 list_del_init(&rq->queuelist);
708 blk_mq_sched_insert_request(rq, true, false, false);
711 while (!list_empty(&rq_list)) {
712 rq = list_entry(rq_list.next, struct request, queuelist);
713 list_del_init(&rq->queuelist);
714 blk_mq_sched_insert_request(rq, false, false, false);
717 blk_mq_run_hw_queues(q, false);
720 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
721 bool kick_requeue_list)
723 struct request_queue *q = rq->q;
727 * We abuse this flag that is otherwise used by the I/O scheduler to
728 * request head insertion from the workqueue.
730 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
732 spin_lock_irqsave(&q->requeue_lock, flags);
734 rq->rq_flags |= RQF_SOFTBARRIER;
735 list_add(&rq->queuelist, &q->requeue_list);
737 list_add_tail(&rq->queuelist, &q->requeue_list);
739 spin_unlock_irqrestore(&q->requeue_lock, flags);
741 if (kick_requeue_list)
742 blk_mq_kick_requeue_list(q);
744 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
746 void blk_mq_kick_requeue_list(struct request_queue *q)
748 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
750 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
752 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
755 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
756 msecs_to_jiffies(msecs));
758 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
760 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
762 if (tag < tags->nr_tags) {
763 prefetch(tags->rqs[tag]);
764 return tags->rqs[tag];
769 EXPORT_SYMBOL(blk_mq_tag_to_rq);
771 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
773 if (req->q->mq_ops->timeout) {
774 enum blk_eh_timer_return ret;
776 ret = req->q->mq_ops->timeout(req, reserved);
777 if (ret == BLK_EH_DONE)
779 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
785 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
787 unsigned long deadline;
789 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
792 deadline = blk_rq_deadline(rq);
793 if (time_after_eq(jiffies, deadline))
798 else if (time_after(*next, deadline))
803 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
804 struct request *rq, void *priv, bool reserved)
806 unsigned long *next = priv;
809 * Just do a quick check if it is expired before locking the request in
810 * so we're not unnecessarilly synchronizing across CPUs.
812 if (!blk_mq_req_expired(rq, next))
816 * We have reason to believe the request may be expired. Take a
817 * reference on the request to lock this request lifetime into its
818 * currently allocated context to prevent it from being reallocated in
819 * the event the completion by-passes this timeout handler.
821 * If the reference was already released, then the driver beat the
822 * timeout handler to posting a natural completion.
824 if (!refcount_inc_not_zero(&rq->ref))
828 * The request is now locked and cannot be reallocated underneath the
829 * timeout handler's processing. Re-verify this exact request is truly
830 * expired; if it is not expired, then the request was completed and
831 * reallocated as a new request.
833 if (blk_mq_req_expired(rq, next))
834 blk_mq_rq_timed_out(rq, reserved);
835 if (refcount_dec_and_test(&rq->ref))
836 __blk_mq_free_request(rq);
839 static void blk_mq_timeout_work(struct work_struct *work)
841 struct request_queue *q =
842 container_of(work, struct request_queue, timeout_work);
843 unsigned long next = 0;
844 struct blk_mq_hw_ctx *hctx;
847 /* A deadlock might occur if a request is stuck requiring a
848 * timeout at the same time a queue freeze is waiting
849 * completion, since the timeout code would not be able to
850 * acquire the queue reference here.
852 * That's why we don't use blk_queue_enter here; instead, we use
853 * percpu_ref_tryget directly, because we need to be able to
854 * obtain a reference even in the short window between the queue
855 * starting to freeze, by dropping the first reference in
856 * blk_freeze_queue_start, and the moment the last request is
857 * consumed, marked by the instant q_usage_counter reaches
860 if (!percpu_ref_tryget(&q->q_usage_counter))
863 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
866 mod_timer(&q->timeout, next);
869 * Request timeouts are handled as a forward rolling timer. If
870 * we end up here it means that no requests are pending and
871 * also that no request has been pending for a while. Mark
874 queue_for_each_hw_ctx(q, hctx, i) {
875 /* the hctx may be unmapped, so check it here */
876 if (blk_mq_hw_queue_mapped(hctx))
877 blk_mq_tag_idle(hctx);
883 struct flush_busy_ctx_data {
884 struct blk_mq_hw_ctx *hctx;
885 struct list_head *list;
888 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
890 struct flush_busy_ctx_data *flush_data = data;
891 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
892 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
894 spin_lock(&ctx->lock);
895 list_splice_tail_init(&ctx->rq_list, flush_data->list);
896 sbitmap_clear_bit(sb, bitnr);
897 spin_unlock(&ctx->lock);
902 * Process software queues that have been marked busy, splicing them
903 * to the for-dispatch
905 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
907 struct flush_busy_ctx_data data = {
912 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
914 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
916 struct dispatch_rq_data {
917 struct blk_mq_hw_ctx *hctx;
921 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
924 struct dispatch_rq_data *dispatch_data = data;
925 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
926 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
928 spin_lock(&ctx->lock);
929 if (!list_empty(&ctx->rq_list)) {
930 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
931 list_del_init(&dispatch_data->rq->queuelist);
932 if (list_empty(&ctx->rq_list))
933 sbitmap_clear_bit(sb, bitnr);
935 spin_unlock(&ctx->lock);
937 return !dispatch_data->rq;
940 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
941 struct blk_mq_ctx *start)
943 unsigned off = start ? start->index_hw : 0;
944 struct dispatch_rq_data data = {
949 __sbitmap_for_each_set(&hctx->ctx_map, off,
950 dispatch_rq_from_ctx, &data);
955 static inline unsigned int queued_to_index(unsigned int queued)
960 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
963 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
966 struct blk_mq_alloc_data data = {
968 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
969 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
972 might_sleep_if(wait);
977 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
978 data.flags |= BLK_MQ_REQ_RESERVED;
980 rq->tag = blk_mq_get_tag(&data);
982 if (blk_mq_tag_busy(data.hctx)) {
983 rq->rq_flags |= RQF_MQ_INFLIGHT;
984 atomic_inc(&data.hctx->nr_active);
986 data.hctx->tags->rqs[rq->tag] = rq;
992 return rq->tag != -1;
995 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
996 int flags, void *key)
998 struct blk_mq_hw_ctx *hctx;
1000 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1002 list_del_init(&wait->entry);
1003 blk_mq_run_hw_queue(hctx, true);
1008 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1009 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1010 * restart. For both cases, take care to check the condition again after
1011 * marking us as waiting.
1013 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1016 struct blk_mq_hw_ctx *this_hctx = *hctx;
1017 struct sbq_wait_state *ws;
1018 wait_queue_entry_t *wait;
1021 if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1022 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1023 set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1026 * It's possible that a tag was freed in the window between the
1027 * allocation failure and adding the hardware queue to the wait
1030 * Don't clear RESTART here, someone else could have set it.
1031 * At most this will cost an extra queue run.
1033 return blk_mq_get_driver_tag(rq, hctx, false);
1036 wait = &this_hctx->dispatch_wait;
1037 if (!list_empty_careful(&wait->entry))
1040 spin_lock(&this_hctx->lock);
1041 if (!list_empty(&wait->entry)) {
1042 spin_unlock(&this_hctx->lock);
1046 ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1047 add_wait_queue(&ws->wait, wait);
1050 * It's possible that a tag was freed in the window between the
1051 * allocation failure and adding the hardware queue to the wait
1054 ret = blk_mq_get_driver_tag(rq, hctx, false);
1056 spin_unlock(&this_hctx->lock);
1061 * We got a tag, remove ourselves from the wait queue to ensure
1062 * someone else gets the wakeup.
1064 spin_lock_irq(&ws->wait.lock);
1065 list_del_init(&wait->entry);
1066 spin_unlock_irq(&ws->wait.lock);
1067 spin_unlock(&this_hctx->lock);
1072 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1074 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1077 struct blk_mq_hw_ctx *hctx;
1078 struct request *rq, *nxt;
1079 bool no_tag = false;
1081 blk_status_t ret = BLK_STS_OK;
1083 if (list_empty(list))
1086 WARN_ON(!list_is_singular(list) && got_budget);
1089 * Now process all the entries, sending them to the driver.
1091 errors = queued = 0;
1093 struct blk_mq_queue_data bd;
1095 rq = list_first_entry(list, struct request, queuelist);
1097 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1098 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1101 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1103 * The initial allocation attempt failed, so we need to
1104 * rerun the hardware queue when a tag is freed. The
1105 * waitqueue takes care of that. If the queue is run
1106 * before we add this entry back on the dispatch list,
1107 * we'll re-run it below.
1109 if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1110 blk_mq_put_dispatch_budget(hctx);
1112 * For non-shared tags, the RESTART check
1115 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1121 list_del_init(&rq->queuelist);
1126 * Flag last if we have no more requests, or if we have more
1127 * but can't assign a driver tag to it.
1129 if (list_empty(list))
1132 nxt = list_first_entry(list, struct request, queuelist);
1133 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1136 ret = q->mq_ops->queue_rq(hctx, &bd);
1137 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1139 * If an I/O scheduler has been configured and we got a
1140 * driver tag for the next request already, free it
1143 if (!list_empty(list)) {
1144 nxt = list_first_entry(list, struct request, queuelist);
1145 blk_mq_put_driver_tag(nxt);
1147 list_add(&rq->queuelist, list);
1148 __blk_mq_requeue_request(rq);
1152 if (unlikely(ret != BLK_STS_OK)) {
1154 blk_mq_end_request(rq, BLK_STS_IOERR);
1159 } while (!list_empty(list));
1161 hctx->dispatched[queued_to_index(queued)]++;
1164 * Any items that need requeuing? Stuff them into hctx->dispatch,
1165 * that is where we will continue on next queue run.
1167 if (!list_empty(list)) {
1170 spin_lock(&hctx->lock);
1171 list_splice_init(list, &hctx->dispatch);
1172 spin_unlock(&hctx->lock);
1175 * If SCHED_RESTART was set by the caller of this function and
1176 * it is no longer set that means that it was cleared by another
1177 * thread and hence that a queue rerun is needed.
1179 * If 'no_tag' is set, that means that we failed getting
1180 * a driver tag with an I/O scheduler attached. If our dispatch
1181 * waitqueue is no longer active, ensure that we run the queue
1182 * AFTER adding our entries back to the list.
1184 * If no I/O scheduler has been configured it is possible that
1185 * the hardware queue got stopped and restarted before requests
1186 * were pushed back onto the dispatch list. Rerun the queue to
1187 * avoid starvation. Notes:
1188 * - blk_mq_run_hw_queue() checks whether or not a queue has
1189 * been stopped before rerunning a queue.
1190 * - Some but not all block drivers stop a queue before
1191 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1194 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1195 * bit is set, run queue after a delay to avoid IO stalls
1196 * that could otherwise occur if the queue is idle.
1198 needs_restart = blk_mq_sched_needs_restart(hctx);
1199 if (!needs_restart ||
1200 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1201 blk_mq_run_hw_queue(hctx, true);
1202 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1203 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1206 return (queued + errors) != 0;
1209 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1214 * We should be running this queue from one of the CPUs that
1217 * There are at least two related races now between setting
1218 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1219 * __blk_mq_run_hw_queue():
1221 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1222 * but later it becomes online, then this warning is harmless
1225 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1226 * but later it becomes offline, then the warning can't be
1227 * triggered, and we depend on blk-mq timeout handler to
1228 * handle dispatched requests to this hctx
1230 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1231 cpu_online(hctx->next_cpu)) {
1232 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1233 raw_smp_processor_id(),
1234 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1239 * We can't run the queue inline with ints disabled. Ensure that
1240 * we catch bad users of this early.
1242 WARN_ON_ONCE(in_interrupt());
1244 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1246 hctx_lock(hctx, &srcu_idx);
1247 blk_mq_sched_dispatch_requests(hctx);
1248 hctx_unlock(hctx, srcu_idx);
1251 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1253 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1255 if (cpu >= nr_cpu_ids)
1256 cpu = cpumask_first(hctx->cpumask);
1261 * It'd be great if the workqueue API had a way to pass
1262 * in a mask and had some smarts for more clever placement.
1263 * For now we just round-robin here, switching for every
1264 * BLK_MQ_CPU_WORK_BATCH queued items.
1266 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1269 int next_cpu = hctx->next_cpu;
1271 if (hctx->queue->nr_hw_queues == 1)
1272 return WORK_CPU_UNBOUND;
1274 if (--hctx->next_cpu_batch <= 0) {
1276 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1278 if (next_cpu >= nr_cpu_ids)
1279 next_cpu = blk_mq_first_mapped_cpu(hctx);
1280 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1284 * Do unbound schedule if we can't find a online CPU for this hctx,
1285 * and it should only happen in the path of handling CPU DEAD.
1287 if (!cpu_online(next_cpu)) {
1294 * Make sure to re-select CPU next time once after CPUs
1295 * in hctx->cpumask become online again.
1297 hctx->next_cpu = next_cpu;
1298 hctx->next_cpu_batch = 1;
1299 return WORK_CPU_UNBOUND;
1302 hctx->next_cpu = next_cpu;
1306 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1307 unsigned long msecs)
1309 if (unlikely(blk_mq_hctx_stopped(hctx)))
1312 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1313 int cpu = get_cpu();
1314 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1315 __blk_mq_run_hw_queue(hctx);
1323 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1324 msecs_to_jiffies(msecs));
1327 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1329 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1331 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1333 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1339 * When queue is quiesced, we may be switching io scheduler, or
1340 * updating nr_hw_queues, or other things, and we can't run queue
1341 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1343 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1346 hctx_lock(hctx, &srcu_idx);
1347 need_run = !blk_queue_quiesced(hctx->queue) &&
1348 blk_mq_hctx_has_pending(hctx);
1349 hctx_unlock(hctx, srcu_idx);
1352 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1358 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1360 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1362 struct blk_mq_hw_ctx *hctx;
1365 queue_for_each_hw_ctx(q, hctx, i) {
1366 if (blk_mq_hctx_stopped(hctx))
1369 blk_mq_run_hw_queue(hctx, async);
1372 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1375 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1376 * @q: request queue.
1378 * The caller is responsible for serializing this function against
1379 * blk_mq_{start,stop}_hw_queue().
1381 bool blk_mq_queue_stopped(struct request_queue *q)
1383 struct blk_mq_hw_ctx *hctx;
1386 queue_for_each_hw_ctx(q, hctx, i)
1387 if (blk_mq_hctx_stopped(hctx))
1392 EXPORT_SYMBOL(blk_mq_queue_stopped);
1395 * This function is often used for pausing .queue_rq() by driver when
1396 * there isn't enough resource or some conditions aren't satisfied, and
1397 * BLK_STS_RESOURCE is usually returned.
1399 * We do not guarantee that dispatch can be drained or blocked
1400 * after blk_mq_stop_hw_queue() returns. Please use
1401 * blk_mq_quiesce_queue() for that requirement.
1403 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1405 cancel_delayed_work(&hctx->run_work);
1407 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1409 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1412 * This function is often used for pausing .queue_rq() by driver when
1413 * there isn't enough resource or some conditions aren't satisfied, and
1414 * BLK_STS_RESOURCE is usually returned.
1416 * We do not guarantee that dispatch can be drained or blocked
1417 * after blk_mq_stop_hw_queues() returns. Please use
1418 * blk_mq_quiesce_queue() for that requirement.
1420 void blk_mq_stop_hw_queues(struct request_queue *q)
1422 struct blk_mq_hw_ctx *hctx;
1425 queue_for_each_hw_ctx(q, hctx, i)
1426 blk_mq_stop_hw_queue(hctx);
1428 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1430 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1432 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1434 blk_mq_run_hw_queue(hctx, false);
1436 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1438 void blk_mq_start_hw_queues(struct request_queue *q)
1440 struct blk_mq_hw_ctx *hctx;
1443 queue_for_each_hw_ctx(q, hctx, i)
1444 blk_mq_start_hw_queue(hctx);
1446 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1448 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1450 if (!blk_mq_hctx_stopped(hctx))
1453 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1454 blk_mq_run_hw_queue(hctx, async);
1456 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1458 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1460 struct blk_mq_hw_ctx *hctx;
1463 queue_for_each_hw_ctx(q, hctx, i)
1464 blk_mq_start_stopped_hw_queue(hctx, async);
1466 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1468 static void blk_mq_run_work_fn(struct work_struct *work)
1470 struct blk_mq_hw_ctx *hctx;
1472 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1475 * If we are stopped, don't run the queue.
1477 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1480 __blk_mq_run_hw_queue(hctx);
1483 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1487 struct blk_mq_ctx *ctx = rq->mq_ctx;
1489 lockdep_assert_held(&ctx->lock);
1491 trace_block_rq_insert(hctx->queue, rq);
1494 list_add(&rq->queuelist, &ctx->rq_list);
1496 list_add_tail(&rq->queuelist, &ctx->rq_list);
1499 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1502 struct blk_mq_ctx *ctx = rq->mq_ctx;
1504 lockdep_assert_held(&ctx->lock);
1506 __blk_mq_insert_req_list(hctx, rq, at_head);
1507 blk_mq_hctx_mark_pending(hctx, ctx);
1511 * Should only be used carefully, when the caller knows we want to
1512 * bypass a potential IO scheduler on the target device.
1514 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1516 struct blk_mq_ctx *ctx = rq->mq_ctx;
1517 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1519 spin_lock(&hctx->lock);
1520 list_add_tail(&rq->queuelist, &hctx->dispatch);
1521 spin_unlock(&hctx->lock);
1524 blk_mq_run_hw_queue(hctx, false);
1527 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1528 struct list_head *list)
1532 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1535 spin_lock(&ctx->lock);
1536 while (!list_empty(list)) {
1539 rq = list_first_entry(list, struct request, queuelist);
1540 BUG_ON(rq->mq_ctx != ctx);
1541 list_del_init(&rq->queuelist);
1542 __blk_mq_insert_req_list(hctx, rq, false);
1544 blk_mq_hctx_mark_pending(hctx, ctx);
1545 spin_unlock(&ctx->lock);
1548 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1550 struct request *rqa = container_of(a, struct request, queuelist);
1551 struct request *rqb = container_of(b, struct request, queuelist);
1553 return !(rqa->mq_ctx < rqb->mq_ctx ||
1554 (rqa->mq_ctx == rqb->mq_ctx &&
1555 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1558 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1560 struct blk_mq_ctx *this_ctx;
1561 struct request_queue *this_q;
1564 LIST_HEAD(ctx_list);
1567 list_splice_init(&plug->mq_list, &list);
1569 list_sort(NULL, &list, plug_ctx_cmp);
1575 while (!list_empty(&list)) {
1576 rq = list_entry_rq(list.next);
1577 list_del_init(&rq->queuelist);
1579 if (rq->mq_ctx != this_ctx) {
1581 trace_block_unplug(this_q, depth, from_schedule);
1582 blk_mq_sched_insert_requests(this_q, this_ctx,
1587 this_ctx = rq->mq_ctx;
1593 list_add_tail(&rq->queuelist, &ctx_list);
1597 * If 'this_ctx' is set, we know we have entries to complete
1598 * on 'ctx_list'. Do those.
1601 trace_block_unplug(this_q, depth, from_schedule);
1602 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1607 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1609 blk_init_request_from_bio(rq, bio);
1611 blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1613 blk_account_io_start(rq, true);
1616 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1619 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1621 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1624 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1628 struct request_queue *q = rq->q;
1629 struct blk_mq_queue_data bd = {
1633 blk_qc_t new_cookie;
1636 new_cookie = request_to_qc_t(hctx, rq);
1639 * For OK queue, we are done. For error, caller may kill it.
1640 * Any other error (busy), just add it to our list as we
1641 * previously would have done.
1643 ret = q->mq_ops->queue_rq(hctx, &bd);
1646 *cookie = new_cookie;
1648 case BLK_STS_RESOURCE:
1649 case BLK_STS_DEV_RESOURCE:
1650 __blk_mq_requeue_request(rq);
1653 *cookie = BLK_QC_T_NONE;
1660 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1665 struct request_queue *q = rq->q;
1666 bool run_queue = true;
1669 * RCU or SRCU read lock is needed before checking quiesced flag.
1671 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1672 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1673 * and avoid driver to try to dispatch again.
1675 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1677 bypass_insert = false;
1681 if (q->elevator && !bypass_insert)
1684 if (!blk_mq_get_dispatch_budget(hctx))
1687 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1688 blk_mq_put_dispatch_budget(hctx);
1692 return __blk_mq_issue_directly(hctx, rq, cookie);
1695 return BLK_STS_RESOURCE;
1697 blk_mq_sched_insert_request(rq, false, run_queue, false);
1701 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1702 struct request *rq, blk_qc_t *cookie)
1707 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1709 hctx_lock(hctx, &srcu_idx);
1711 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1712 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1713 blk_mq_sched_insert_request(rq, false, true, false);
1714 else if (ret != BLK_STS_OK)
1715 blk_mq_end_request(rq, ret);
1717 hctx_unlock(hctx, srcu_idx);
1720 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1724 blk_qc_t unused_cookie;
1725 struct blk_mq_ctx *ctx = rq->mq_ctx;
1726 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1728 hctx_lock(hctx, &srcu_idx);
1729 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1730 hctx_unlock(hctx, srcu_idx);
1735 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1737 const int is_sync = op_is_sync(bio->bi_opf);
1738 const int is_flush_fua = op_is_flush(bio->bi_opf);
1739 struct blk_mq_alloc_data data = { .flags = 0 };
1741 unsigned int request_count = 0;
1742 struct blk_plug *plug;
1743 struct request *same_queue_rq = NULL;
1745 unsigned int wb_acct;
1747 blk_queue_bounce(q, &bio);
1749 blk_queue_split(q, &bio);
1751 if (!bio_integrity_prep(bio))
1752 return BLK_QC_T_NONE;
1754 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1755 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1756 return BLK_QC_T_NONE;
1758 if (blk_mq_sched_bio_merge(q, bio))
1759 return BLK_QC_T_NONE;
1761 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1763 trace_block_getrq(q, bio, bio->bi_opf);
1765 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1766 if (unlikely(!rq)) {
1767 __wbt_done(q->rq_wb, wb_acct);
1768 if (bio->bi_opf & REQ_NOWAIT)
1769 bio_wouldblock_error(bio);
1770 return BLK_QC_T_NONE;
1773 wbt_track(rq, wb_acct);
1775 cookie = request_to_qc_t(data.hctx, rq);
1777 plug = current->plug;
1778 if (unlikely(is_flush_fua)) {
1779 blk_mq_put_ctx(data.ctx);
1780 blk_mq_bio_to_request(rq, bio);
1782 /* bypass scheduler for flush rq */
1783 blk_insert_flush(rq);
1784 blk_mq_run_hw_queue(data.hctx, true);
1785 } else if (plug && q->nr_hw_queues == 1) {
1786 struct request *last = NULL;
1788 blk_mq_put_ctx(data.ctx);
1789 blk_mq_bio_to_request(rq, bio);
1792 * @request_count may become stale because of schedule
1793 * out, so check the list again.
1795 if (list_empty(&plug->mq_list))
1797 else if (blk_queue_nomerges(q))
1798 request_count = blk_plug_queued_count(q);
1801 trace_block_plug(q);
1803 last = list_entry_rq(plug->mq_list.prev);
1805 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1806 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1807 blk_flush_plug_list(plug, false);
1808 trace_block_plug(q);
1811 list_add_tail(&rq->queuelist, &plug->mq_list);
1812 } else if (plug && !blk_queue_nomerges(q)) {
1813 blk_mq_bio_to_request(rq, bio);
1816 * We do limited plugging. If the bio can be merged, do that.
1817 * Otherwise the existing request in the plug list will be
1818 * issued. So the plug list will have one request at most
1819 * The plug list might get flushed before this. If that happens,
1820 * the plug list is empty, and same_queue_rq is invalid.
1822 if (list_empty(&plug->mq_list))
1823 same_queue_rq = NULL;
1825 list_del_init(&same_queue_rq->queuelist);
1826 list_add_tail(&rq->queuelist, &plug->mq_list);
1828 blk_mq_put_ctx(data.ctx);
1830 if (same_queue_rq) {
1831 data.hctx = blk_mq_map_queue(q,
1832 same_queue_rq->mq_ctx->cpu);
1833 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1836 } else if (q->nr_hw_queues > 1 && is_sync) {
1837 blk_mq_put_ctx(data.ctx);
1838 blk_mq_bio_to_request(rq, bio);
1839 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1841 blk_mq_put_ctx(data.ctx);
1842 blk_mq_bio_to_request(rq, bio);
1843 blk_mq_sched_insert_request(rq, false, true, true);
1849 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1850 unsigned int hctx_idx)
1854 if (tags->rqs && set->ops->exit_request) {
1857 for (i = 0; i < tags->nr_tags; i++) {
1858 struct request *rq = tags->static_rqs[i];
1862 set->ops->exit_request(set, rq, hctx_idx);
1863 tags->static_rqs[i] = NULL;
1867 while (!list_empty(&tags->page_list)) {
1868 page = list_first_entry(&tags->page_list, struct page, lru);
1869 list_del_init(&page->lru);
1871 * Remove kmemleak object previously allocated in
1872 * blk_mq_init_rq_map().
1874 kmemleak_free(page_address(page));
1875 __free_pages(page, page->private);
1879 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1883 kfree(tags->static_rqs);
1884 tags->static_rqs = NULL;
1886 blk_mq_free_tags(tags);
1889 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1890 unsigned int hctx_idx,
1891 unsigned int nr_tags,
1892 unsigned int reserved_tags)
1894 struct blk_mq_tags *tags;
1897 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1898 if (node == NUMA_NO_NODE)
1899 node = set->numa_node;
1901 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1902 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1906 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1907 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1910 blk_mq_free_tags(tags);
1914 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1915 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1917 if (!tags->static_rqs) {
1919 blk_mq_free_tags(tags);
1926 static size_t order_to_size(unsigned int order)
1928 return (size_t)PAGE_SIZE << order;
1931 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1932 unsigned int hctx_idx, int node)
1936 if (set->ops->init_request) {
1937 ret = set->ops->init_request(set, rq, hctx_idx, node);
1942 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1946 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1947 unsigned int hctx_idx, unsigned int depth)
1949 unsigned int i, j, entries_per_page, max_order = 4;
1950 size_t rq_size, left;
1953 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1954 if (node == NUMA_NO_NODE)
1955 node = set->numa_node;
1957 INIT_LIST_HEAD(&tags->page_list);
1960 * rq_size is the size of the request plus driver payload, rounded
1961 * to the cacheline size
1963 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1965 left = rq_size * depth;
1967 for (i = 0; i < depth; ) {
1968 int this_order = max_order;
1973 while (this_order && left < order_to_size(this_order - 1))
1977 page = alloc_pages_node(node,
1978 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1984 if (order_to_size(this_order) < rq_size)
1991 page->private = this_order;
1992 list_add_tail(&page->lru, &tags->page_list);
1994 p = page_address(page);
1996 * Allow kmemleak to scan these pages as they contain pointers
1997 * to additional allocations like via ops->init_request().
1999 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2000 entries_per_page = order_to_size(this_order) / rq_size;
2001 to_do = min(entries_per_page, depth - i);
2002 left -= to_do * rq_size;
2003 for (j = 0; j < to_do; j++) {
2004 struct request *rq = p;
2006 tags->static_rqs[i] = rq;
2007 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2008 tags->static_rqs[i] = NULL;
2019 blk_mq_free_rqs(set, tags, hctx_idx);
2024 * 'cpu' is going away. splice any existing rq_list entries from this
2025 * software queue to the hw queue dispatch list, and ensure that it
2028 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2030 struct blk_mq_hw_ctx *hctx;
2031 struct blk_mq_ctx *ctx;
2034 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2035 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2037 spin_lock(&ctx->lock);
2038 if (!list_empty(&ctx->rq_list)) {
2039 list_splice_init(&ctx->rq_list, &tmp);
2040 blk_mq_hctx_clear_pending(hctx, ctx);
2042 spin_unlock(&ctx->lock);
2044 if (list_empty(&tmp))
2047 spin_lock(&hctx->lock);
2048 list_splice_tail_init(&tmp, &hctx->dispatch);
2049 spin_unlock(&hctx->lock);
2051 blk_mq_run_hw_queue(hctx, true);
2055 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2057 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2061 /* hctx->ctxs will be freed in queue's release handler */
2062 static void blk_mq_exit_hctx(struct request_queue *q,
2063 struct blk_mq_tag_set *set,
2064 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2066 blk_mq_debugfs_unregister_hctx(hctx);
2068 if (blk_mq_hw_queue_mapped(hctx))
2069 blk_mq_tag_idle(hctx);
2071 if (set->ops->exit_request)
2072 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2074 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2076 if (set->ops->exit_hctx)
2077 set->ops->exit_hctx(hctx, hctx_idx);
2079 if (hctx->flags & BLK_MQ_F_BLOCKING)
2080 cleanup_srcu_struct(hctx->srcu);
2082 blk_mq_remove_cpuhp(hctx);
2083 blk_free_flush_queue(hctx->fq);
2084 sbitmap_free(&hctx->ctx_map);
2087 static void blk_mq_exit_hw_queues(struct request_queue *q,
2088 struct blk_mq_tag_set *set, int nr_queue)
2090 struct blk_mq_hw_ctx *hctx;
2093 queue_for_each_hw_ctx(q, hctx, i) {
2096 blk_mq_exit_hctx(q, set, hctx, i);
2100 static int blk_mq_init_hctx(struct request_queue *q,
2101 struct blk_mq_tag_set *set,
2102 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2106 node = hctx->numa_node;
2107 if (node == NUMA_NO_NODE)
2108 node = hctx->numa_node = set->numa_node;
2110 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2111 spin_lock_init(&hctx->lock);
2112 INIT_LIST_HEAD(&hctx->dispatch);
2114 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2116 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2118 hctx->tags = set->tags[hctx_idx];
2121 * Allocate space for all possible cpus to avoid allocation at
2124 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2127 goto unregister_cpu_notifier;
2129 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2135 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2136 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2138 if (set->ops->init_hctx &&
2139 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2142 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2145 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2147 goto sched_exit_hctx;
2149 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2152 if (hctx->flags & BLK_MQ_F_BLOCKING)
2153 init_srcu_struct(hctx->srcu);
2155 blk_mq_debugfs_register_hctx(q, hctx);
2162 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2164 if (set->ops->exit_hctx)
2165 set->ops->exit_hctx(hctx, hctx_idx);
2167 sbitmap_free(&hctx->ctx_map);
2170 unregister_cpu_notifier:
2171 blk_mq_remove_cpuhp(hctx);
2175 static void blk_mq_init_cpu_queues(struct request_queue *q,
2176 unsigned int nr_hw_queues)
2180 for_each_possible_cpu(i) {
2181 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2182 struct blk_mq_hw_ctx *hctx;
2185 spin_lock_init(&__ctx->lock);
2186 INIT_LIST_HEAD(&__ctx->rq_list);
2190 * Set local node, IFF we have more than one hw queue. If
2191 * not, we remain on the home node of the device
2193 hctx = blk_mq_map_queue(q, i);
2194 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2195 hctx->numa_node = local_memory_node(cpu_to_node(i));
2199 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2203 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2204 set->queue_depth, set->reserved_tags);
2205 if (!set->tags[hctx_idx])
2208 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2213 blk_mq_free_rq_map(set->tags[hctx_idx]);
2214 set->tags[hctx_idx] = NULL;
2218 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2219 unsigned int hctx_idx)
2221 if (set->tags[hctx_idx]) {
2222 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2223 blk_mq_free_rq_map(set->tags[hctx_idx]);
2224 set->tags[hctx_idx] = NULL;
2228 static void blk_mq_map_swqueue(struct request_queue *q)
2230 unsigned int i, hctx_idx;
2231 struct blk_mq_hw_ctx *hctx;
2232 struct blk_mq_ctx *ctx;
2233 struct blk_mq_tag_set *set = q->tag_set;
2236 * Avoid others reading imcomplete hctx->cpumask through sysfs
2238 mutex_lock(&q->sysfs_lock);
2240 queue_for_each_hw_ctx(q, hctx, i) {
2241 cpumask_clear(hctx->cpumask);
2243 hctx->dispatch_from = NULL;
2247 * Map software to hardware queues.
2249 * If the cpu isn't present, the cpu is mapped to first hctx.
2251 for_each_possible_cpu(i) {
2252 hctx_idx = q->mq_map[i];
2253 /* unmapped hw queue can be remapped after CPU topo changed */
2254 if (!set->tags[hctx_idx] &&
2255 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2257 * If tags initialization fail for some hctx,
2258 * that hctx won't be brought online. In this
2259 * case, remap the current ctx to hctx[0] which
2260 * is guaranteed to always have tags allocated
2265 ctx = per_cpu_ptr(q->queue_ctx, i);
2266 hctx = blk_mq_map_queue(q, i);
2268 cpumask_set_cpu(i, hctx->cpumask);
2269 ctx->index_hw = hctx->nr_ctx;
2270 hctx->ctxs[hctx->nr_ctx++] = ctx;
2273 mutex_unlock(&q->sysfs_lock);
2275 queue_for_each_hw_ctx(q, hctx, i) {
2277 * If no software queues are mapped to this hardware queue,
2278 * disable it and free the request entries.
2280 if (!hctx->nr_ctx) {
2281 /* Never unmap queue 0. We need it as a
2282 * fallback in case of a new remap fails
2285 if (i && set->tags[i])
2286 blk_mq_free_map_and_requests(set, i);
2292 hctx->tags = set->tags[i];
2293 WARN_ON(!hctx->tags);
2296 * Set the map size to the number of mapped software queues.
2297 * This is more accurate and more efficient than looping
2298 * over all possibly mapped software queues.
2300 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2303 * Initialize batch roundrobin counts
2305 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2306 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2311 * Caller needs to ensure that we're either frozen/quiesced, or that
2312 * the queue isn't live yet.
2314 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2316 struct blk_mq_hw_ctx *hctx;
2319 queue_for_each_hw_ctx(q, hctx, i) {
2321 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2322 atomic_inc(&q->shared_hctx_restart);
2323 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2325 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2326 atomic_dec(&q->shared_hctx_restart);
2327 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2332 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2335 struct request_queue *q;
2337 lockdep_assert_held(&set->tag_list_lock);
2339 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2340 blk_mq_freeze_queue(q);
2341 queue_set_hctx_shared(q, shared);
2342 blk_mq_unfreeze_queue(q);
2346 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2348 struct blk_mq_tag_set *set = q->tag_set;
2350 mutex_lock(&set->tag_list_lock);
2351 list_del_rcu(&q->tag_set_list);
2352 INIT_LIST_HEAD(&q->tag_set_list);
2353 if (list_is_singular(&set->tag_list)) {
2354 /* just transitioned to unshared */
2355 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2356 /* update existing queue */
2357 blk_mq_update_tag_set_depth(set, false);
2359 mutex_unlock(&set->tag_list_lock);
2364 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2365 struct request_queue *q)
2369 mutex_lock(&set->tag_list_lock);
2372 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2374 if (!list_empty(&set->tag_list) &&
2375 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2376 set->flags |= BLK_MQ_F_TAG_SHARED;
2377 /* update existing queue */
2378 blk_mq_update_tag_set_depth(set, true);
2380 if (set->flags & BLK_MQ_F_TAG_SHARED)
2381 queue_set_hctx_shared(q, true);
2382 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2384 mutex_unlock(&set->tag_list_lock);
2388 * It is the actual release handler for mq, but we do it from
2389 * request queue's release handler for avoiding use-after-free
2390 * and headache because q->mq_kobj shouldn't have been introduced,
2391 * but we can't group ctx/kctx kobj without it.
2393 void blk_mq_release(struct request_queue *q)
2395 struct blk_mq_hw_ctx *hctx;
2398 /* hctx kobj stays in hctx */
2399 queue_for_each_hw_ctx(q, hctx, i) {
2402 kobject_put(&hctx->kobj);
2407 kfree(q->queue_hw_ctx);
2410 * release .mq_kobj and sw queue's kobject now because
2411 * both share lifetime with request queue.
2413 blk_mq_sysfs_deinit(q);
2415 free_percpu(q->queue_ctx);
2418 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2420 struct request_queue *uninit_q, *q;
2422 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2424 return ERR_PTR(-ENOMEM);
2426 q = blk_mq_init_allocated_queue(set, uninit_q);
2428 blk_cleanup_queue(uninit_q);
2432 EXPORT_SYMBOL(blk_mq_init_queue);
2434 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2436 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2438 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2439 __alignof__(struct blk_mq_hw_ctx)) !=
2440 sizeof(struct blk_mq_hw_ctx));
2442 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2443 hw_ctx_size += sizeof(struct srcu_struct);
2448 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2449 struct request_queue *q)
2452 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2454 blk_mq_sysfs_unregister(q);
2456 /* protect against switching io scheduler */
2457 mutex_lock(&q->sysfs_lock);
2458 for (i = 0; i < set->nr_hw_queues; i++) {
2464 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2465 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2470 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2477 atomic_set(&hctxs[i]->nr_active, 0);
2478 hctxs[i]->numa_node = node;
2479 hctxs[i]->queue_num = i;
2481 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2482 free_cpumask_var(hctxs[i]->cpumask);
2487 blk_mq_hctx_kobj_init(hctxs[i]);
2489 for (j = i; j < q->nr_hw_queues; j++) {
2490 struct blk_mq_hw_ctx *hctx = hctxs[j];
2494 blk_mq_free_map_and_requests(set, j);
2495 blk_mq_exit_hctx(q, set, hctx, j);
2496 kobject_put(&hctx->kobj);
2501 q->nr_hw_queues = i;
2502 mutex_unlock(&q->sysfs_lock);
2503 blk_mq_sysfs_register(q);
2506 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2507 struct request_queue *q)
2509 /* mark the queue as mq asap */
2510 q->mq_ops = set->ops;
2512 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2513 blk_mq_poll_stats_bkt,
2514 BLK_MQ_POLL_STATS_BKTS, q);
2518 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2522 /* init q->mq_kobj and sw queues' kobjects */
2523 blk_mq_sysfs_init(q);
2525 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2526 GFP_KERNEL, set->numa_node);
2527 if (!q->queue_hw_ctx)
2530 q->mq_map = set->mq_map;
2532 blk_mq_realloc_hw_ctxs(set, q);
2533 if (!q->nr_hw_queues)
2536 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2537 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2539 q->nr_queues = nr_cpu_ids;
2541 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2543 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2544 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2546 q->sg_reserved_size = INT_MAX;
2548 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2549 INIT_LIST_HEAD(&q->requeue_list);
2550 spin_lock_init(&q->requeue_lock);
2552 blk_queue_make_request(q, blk_mq_make_request);
2553 if (q->mq_ops->poll)
2554 q->poll_fn = blk_mq_poll;
2557 * Do this after blk_queue_make_request() overrides it...
2559 q->nr_requests = set->queue_depth;
2562 * Default to classic polling
2566 if (set->ops->complete)
2567 blk_queue_softirq_done(q, set->ops->complete);
2569 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2570 blk_mq_add_queue_tag_set(set, q);
2571 blk_mq_map_swqueue(q);
2573 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2576 ret = elevator_init_mq(q);
2578 return ERR_PTR(ret);
2584 kfree(q->queue_hw_ctx);
2586 free_percpu(q->queue_ctx);
2589 return ERR_PTR(-ENOMEM);
2591 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2593 void blk_mq_free_queue(struct request_queue *q)
2595 struct blk_mq_tag_set *set = q->tag_set;
2597 blk_mq_del_queue_tag_set(q);
2598 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2601 /* Basically redo blk_mq_init_queue with queue frozen */
2602 static void blk_mq_queue_reinit(struct request_queue *q)
2604 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2606 blk_mq_debugfs_unregister_hctxs(q);
2607 blk_mq_sysfs_unregister(q);
2610 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2611 * we should change hctx numa_node according to the new topology (this
2612 * involves freeing and re-allocating memory, worth doing?)
2614 blk_mq_map_swqueue(q);
2616 blk_mq_sysfs_register(q);
2617 blk_mq_debugfs_register_hctxs(q);
2620 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2624 for (i = 0; i < set->nr_hw_queues; i++)
2625 if (!__blk_mq_alloc_rq_map(set, i))
2632 blk_mq_free_rq_map(set->tags[i]);
2638 * Allocate the request maps associated with this tag_set. Note that this
2639 * may reduce the depth asked for, if memory is tight. set->queue_depth
2640 * will be updated to reflect the allocated depth.
2642 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2647 depth = set->queue_depth;
2649 err = __blk_mq_alloc_rq_maps(set);
2653 set->queue_depth >>= 1;
2654 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2658 } while (set->queue_depth);
2660 if (!set->queue_depth || err) {
2661 pr_err("blk-mq: failed to allocate request map\n");
2665 if (depth != set->queue_depth)
2666 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2667 depth, set->queue_depth);
2672 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2674 if (set->ops->map_queues) {
2677 * transport .map_queues is usually done in the following
2680 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2681 * mask = get_cpu_mask(queue)
2682 * for_each_cpu(cpu, mask)
2683 * set->mq_map[cpu] = queue;
2686 * When we need to remap, the table has to be cleared for
2687 * killing stale mapping since one CPU may not be mapped
2690 for_each_possible_cpu(cpu)
2691 set->mq_map[cpu] = 0;
2693 return set->ops->map_queues(set);
2695 return blk_mq_map_queues(set);
2699 * Alloc a tag set to be associated with one or more request queues.
2700 * May fail with EINVAL for various error conditions. May adjust the
2701 * requested depth down, if if it too large. In that case, the set
2702 * value will be stored in set->queue_depth.
2704 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2708 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2710 if (!set->nr_hw_queues)
2712 if (!set->queue_depth)
2714 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2717 if (!set->ops->queue_rq)
2720 if (!set->ops->get_budget ^ !set->ops->put_budget)
2723 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2724 pr_info("blk-mq: reduced tag depth to %u\n",
2726 set->queue_depth = BLK_MQ_MAX_DEPTH;
2730 * If a crashdump is active, then we are potentially in a very
2731 * memory constrained environment. Limit us to 1 queue and
2732 * 64 tags to prevent using too much memory.
2734 if (is_kdump_kernel()) {
2735 set->nr_hw_queues = 1;
2736 set->queue_depth = min(64U, set->queue_depth);
2739 * There is no use for more h/w queues than cpus.
2741 if (set->nr_hw_queues > nr_cpu_ids)
2742 set->nr_hw_queues = nr_cpu_ids;
2744 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2745 GFP_KERNEL, set->numa_node);
2750 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2751 GFP_KERNEL, set->numa_node);
2755 ret = blk_mq_update_queue_map(set);
2757 goto out_free_mq_map;
2759 ret = blk_mq_alloc_rq_maps(set);
2761 goto out_free_mq_map;
2763 mutex_init(&set->tag_list_lock);
2764 INIT_LIST_HEAD(&set->tag_list);
2776 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2778 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2782 for (i = 0; i < nr_cpu_ids; i++)
2783 blk_mq_free_map_and_requests(set, i);
2791 EXPORT_SYMBOL(blk_mq_free_tag_set);
2793 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2795 struct blk_mq_tag_set *set = q->tag_set;
2796 struct blk_mq_hw_ctx *hctx;
2802 blk_mq_freeze_queue(q);
2803 blk_mq_quiesce_queue(q);
2806 queue_for_each_hw_ctx(q, hctx, i) {
2810 * If we're using an MQ scheduler, just update the scheduler
2811 * queue depth. This is similar to what the old code would do.
2813 if (!hctx->sched_tags) {
2814 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2817 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2825 q->nr_requests = nr;
2827 blk_mq_unquiesce_queue(q);
2828 blk_mq_unfreeze_queue(q);
2833 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2836 struct request_queue *q;
2838 lockdep_assert_held(&set->tag_list_lock);
2840 if (nr_hw_queues > nr_cpu_ids)
2841 nr_hw_queues = nr_cpu_ids;
2842 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2845 list_for_each_entry(q, &set->tag_list, tag_set_list)
2846 blk_mq_freeze_queue(q);
2848 set->nr_hw_queues = nr_hw_queues;
2849 blk_mq_update_queue_map(set);
2850 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2851 blk_mq_realloc_hw_ctxs(set, q);
2852 blk_mq_queue_reinit(q);
2855 list_for_each_entry(q, &set->tag_list, tag_set_list)
2856 blk_mq_unfreeze_queue(q);
2859 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2861 mutex_lock(&set->tag_list_lock);
2862 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2863 mutex_unlock(&set->tag_list_lock);
2865 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2867 /* Enable polling stats and return whether they were already enabled. */
2868 static bool blk_poll_stats_enable(struct request_queue *q)
2870 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2871 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
2873 blk_stat_add_callback(q, q->poll_cb);
2877 static void blk_mq_poll_stats_start(struct request_queue *q)
2880 * We don't arm the callback if polling stats are not enabled or the
2881 * callback is already active.
2883 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2884 blk_stat_is_active(q->poll_cb))
2887 blk_stat_activate_msecs(q->poll_cb, 100);
2890 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2892 struct request_queue *q = cb->data;
2895 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2896 if (cb->stat[bucket].nr_samples)
2897 q->poll_stat[bucket] = cb->stat[bucket];
2901 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2902 struct blk_mq_hw_ctx *hctx,
2905 unsigned long ret = 0;
2909 * If stats collection isn't on, don't sleep but turn it on for
2912 if (!blk_poll_stats_enable(q))
2916 * As an optimistic guess, use half of the mean service time
2917 * for this type of request. We can (and should) make this smarter.
2918 * For instance, if the completion latencies are tight, we can
2919 * get closer than just half the mean. This is especially
2920 * important on devices where the completion latencies are longer
2921 * than ~10 usec. We do use the stats for the relevant IO size
2922 * if available which does lead to better estimates.
2924 bucket = blk_mq_poll_stats_bkt(rq);
2928 if (q->poll_stat[bucket].nr_samples)
2929 ret = (q->poll_stat[bucket].mean + 1) / 2;
2934 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2935 struct blk_mq_hw_ctx *hctx,
2938 struct hrtimer_sleeper hs;
2939 enum hrtimer_mode mode;
2943 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
2949 * -1: don't ever hybrid sleep
2950 * 0: use half of prev avg
2951 * >0: use this specific value
2953 if (q->poll_nsec == -1)
2955 else if (q->poll_nsec > 0)
2956 nsecs = q->poll_nsec;
2958 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2963 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
2966 * This will be replaced with the stats tracking code, using
2967 * 'avg_completion_time / 2' as the pre-sleep target.
2971 mode = HRTIMER_MODE_REL;
2972 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2973 hrtimer_set_expires(&hs.timer, kt);
2975 hrtimer_init_sleeper(&hs, current);
2977 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
2979 set_current_state(TASK_UNINTERRUPTIBLE);
2980 hrtimer_start_expires(&hs.timer, mode);
2983 hrtimer_cancel(&hs.timer);
2984 mode = HRTIMER_MODE_ABS;
2985 } while (hs.task && !signal_pending(current));
2987 __set_current_state(TASK_RUNNING);
2988 destroy_hrtimer_on_stack(&hs.timer);
2992 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2994 struct request_queue *q = hctx->queue;
2998 * If we sleep, have the caller restart the poll loop to reset
2999 * the state. Like for the other success return cases, the
3000 * caller is responsible for checking if the IO completed. If
3001 * the IO isn't complete, we'll get called again and will go
3002 * straight to the busy poll loop.
3004 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3007 hctx->poll_considered++;
3009 state = current->state;
3010 while (!need_resched()) {
3013 hctx->poll_invoked++;
3015 ret = q->mq_ops->poll(hctx, rq->tag);
3017 hctx->poll_success++;
3018 set_current_state(TASK_RUNNING);
3022 if (signal_pending_state(state, current))
3023 set_current_state(TASK_RUNNING);
3025 if (current->state == TASK_RUNNING)
3032 __set_current_state(TASK_RUNNING);
3036 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3038 struct blk_mq_hw_ctx *hctx;
3041 if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3044 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3045 if (!blk_qc_t_is_internal(cookie))
3046 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3048 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3050 * With scheduling, if the request has completed, we'll
3051 * get a NULL return here, as we clear the sched tag when
3052 * that happens. The request still remains valid, like always,
3053 * so we should be safe with just the NULL check.
3059 return __blk_mq_poll(hctx, rq);
3062 static int __init blk_mq_init(void)
3064 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3065 blk_mq_hctx_notify_dead);
3068 subsys_initcall(blk_mq_init);