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 const struct blk_mq_ops *ops = req->q->mq_ops;
774 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
777 ret = ops->timeout(req, reserved);
781 if (blk_mq_rq_state(req) == MQ_RQ_IN_FLIGHT)
782 __blk_mq_complete_request(req);
784 case BLK_EH_RESET_TIMER:
787 case BLK_EH_NOT_HANDLED:
790 printk(KERN_ERR "block: bad eh return: %d\n", ret);
795 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
797 unsigned long deadline;
799 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
802 deadline = blk_rq_deadline(rq);
803 if (time_after_eq(jiffies, deadline))
808 else if (time_after(*next, deadline))
813 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
814 struct request *rq, void *priv, bool reserved)
816 unsigned long *next = priv;
819 * Just do a quick check if it is expired before locking the request in
820 * so we're not unnecessarilly synchronizing across CPUs.
822 if (!blk_mq_req_expired(rq, next))
826 * We have reason to believe the request may be expired. Take a
827 * reference on the request to lock this request lifetime into its
828 * currently allocated context to prevent it from being reallocated in
829 * the event the completion by-passes this timeout handler.
831 * If the reference was already released, then the driver beat the
832 * timeout handler to posting a natural completion.
834 if (!refcount_inc_not_zero(&rq->ref))
838 * The request is now locked and cannot be reallocated underneath the
839 * timeout handler's processing. Re-verify this exact request is truly
840 * expired; if it is not expired, then the request was completed and
841 * reallocated as a new request.
843 if (blk_mq_req_expired(rq, next))
844 blk_mq_rq_timed_out(rq, reserved);
845 if (refcount_dec_and_test(&rq->ref))
846 __blk_mq_free_request(rq);
849 static void blk_mq_timeout_work(struct work_struct *work)
851 struct request_queue *q =
852 container_of(work, struct request_queue, timeout_work);
853 unsigned long next = 0;
854 struct blk_mq_hw_ctx *hctx;
857 /* A deadlock might occur if a request is stuck requiring a
858 * timeout at the same time a queue freeze is waiting
859 * completion, since the timeout code would not be able to
860 * acquire the queue reference here.
862 * That's why we don't use blk_queue_enter here; instead, we use
863 * percpu_ref_tryget directly, because we need to be able to
864 * obtain a reference even in the short window between the queue
865 * starting to freeze, by dropping the first reference in
866 * blk_freeze_queue_start, and the moment the last request is
867 * consumed, marked by the instant q_usage_counter reaches
870 if (!percpu_ref_tryget(&q->q_usage_counter))
873 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
876 mod_timer(&q->timeout, next);
879 * Request timeouts are handled as a forward rolling timer. If
880 * we end up here it means that no requests are pending and
881 * also that no request has been pending for a while. Mark
884 queue_for_each_hw_ctx(q, hctx, i) {
885 /* the hctx may be unmapped, so check it here */
886 if (blk_mq_hw_queue_mapped(hctx))
887 blk_mq_tag_idle(hctx);
893 struct flush_busy_ctx_data {
894 struct blk_mq_hw_ctx *hctx;
895 struct list_head *list;
898 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
900 struct flush_busy_ctx_data *flush_data = data;
901 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
902 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
904 spin_lock(&ctx->lock);
905 list_splice_tail_init(&ctx->rq_list, flush_data->list);
906 sbitmap_clear_bit(sb, bitnr);
907 spin_unlock(&ctx->lock);
912 * Process software queues that have been marked busy, splicing them
913 * to the for-dispatch
915 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
917 struct flush_busy_ctx_data data = {
922 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
924 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
926 struct dispatch_rq_data {
927 struct blk_mq_hw_ctx *hctx;
931 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
934 struct dispatch_rq_data *dispatch_data = data;
935 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
936 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
938 spin_lock(&ctx->lock);
939 if (!list_empty(&ctx->rq_list)) {
940 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
941 list_del_init(&dispatch_data->rq->queuelist);
942 if (list_empty(&ctx->rq_list))
943 sbitmap_clear_bit(sb, bitnr);
945 spin_unlock(&ctx->lock);
947 return !dispatch_data->rq;
950 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
951 struct blk_mq_ctx *start)
953 unsigned off = start ? start->index_hw : 0;
954 struct dispatch_rq_data data = {
959 __sbitmap_for_each_set(&hctx->ctx_map, off,
960 dispatch_rq_from_ctx, &data);
965 static inline unsigned int queued_to_index(unsigned int queued)
970 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
973 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
976 struct blk_mq_alloc_data data = {
978 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
979 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
982 might_sleep_if(wait);
987 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
988 data.flags |= BLK_MQ_REQ_RESERVED;
990 rq->tag = blk_mq_get_tag(&data);
992 if (blk_mq_tag_busy(data.hctx)) {
993 rq->rq_flags |= RQF_MQ_INFLIGHT;
994 atomic_inc(&data.hctx->nr_active);
996 data.hctx->tags->rqs[rq->tag] = rq;
1002 return rq->tag != -1;
1005 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1006 int flags, void *key)
1008 struct blk_mq_hw_ctx *hctx;
1010 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1012 list_del_init(&wait->entry);
1013 blk_mq_run_hw_queue(hctx, true);
1018 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1019 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1020 * restart. For both cases, take care to check the condition again after
1021 * marking us as waiting.
1023 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1026 struct blk_mq_hw_ctx *this_hctx = *hctx;
1027 struct sbq_wait_state *ws;
1028 wait_queue_entry_t *wait;
1031 if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1032 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1033 set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1036 * It's possible that a tag was freed in the window between the
1037 * allocation failure and adding the hardware queue to the wait
1040 * Don't clear RESTART here, someone else could have set it.
1041 * At most this will cost an extra queue run.
1043 return blk_mq_get_driver_tag(rq, hctx, false);
1046 wait = &this_hctx->dispatch_wait;
1047 if (!list_empty_careful(&wait->entry))
1050 spin_lock(&this_hctx->lock);
1051 if (!list_empty(&wait->entry)) {
1052 spin_unlock(&this_hctx->lock);
1056 ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1057 add_wait_queue(&ws->wait, wait);
1060 * It's possible that a tag was freed in the window between the
1061 * allocation failure and adding the hardware queue to the wait
1064 ret = blk_mq_get_driver_tag(rq, hctx, false);
1066 spin_unlock(&this_hctx->lock);
1071 * We got a tag, remove ourselves from the wait queue to ensure
1072 * someone else gets the wakeup.
1074 spin_lock_irq(&ws->wait.lock);
1075 list_del_init(&wait->entry);
1076 spin_unlock_irq(&ws->wait.lock);
1077 spin_unlock(&this_hctx->lock);
1082 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1084 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1087 struct blk_mq_hw_ctx *hctx;
1088 struct request *rq, *nxt;
1089 bool no_tag = false;
1091 blk_status_t ret = BLK_STS_OK;
1093 if (list_empty(list))
1096 WARN_ON(!list_is_singular(list) && got_budget);
1099 * Now process all the entries, sending them to the driver.
1101 errors = queued = 0;
1103 struct blk_mq_queue_data bd;
1105 rq = list_first_entry(list, struct request, queuelist);
1107 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1108 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1111 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1113 * The initial allocation attempt failed, so we need to
1114 * rerun the hardware queue when a tag is freed. The
1115 * waitqueue takes care of that. If the queue is run
1116 * before we add this entry back on the dispatch list,
1117 * we'll re-run it below.
1119 if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1120 blk_mq_put_dispatch_budget(hctx);
1122 * For non-shared tags, the RESTART check
1125 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1131 list_del_init(&rq->queuelist);
1136 * Flag last if we have no more requests, or if we have more
1137 * but can't assign a driver tag to it.
1139 if (list_empty(list))
1142 nxt = list_first_entry(list, struct request, queuelist);
1143 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1146 ret = q->mq_ops->queue_rq(hctx, &bd);
1147 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1149 * If an I/O scheduler has been configured and we got a
1150 * driver tag for the next request already, free it
1153 if (!list_empty(list)) {
1154 nxt = list_first_entry(list, struct request, queuelist);
1155 blk_mq_put_driver_tag(nxt);
1157 list_add(&rq->queuelist, list);
1158 __blk_mq_requeue_request(rq);
1162 if (unlikely(ret != BLK_STS_OK)) {
1164 blk_mq_end_request(rq, BLK_STS_IOERR);
1169 } while (!list_empty(list));
1171 hctx->dispatched[queued_to_index(queued)]++;
1174 * Any items that need requeuing? Stuff them into hctx->dispatch,
1175 * that is where we will continue on next queue run.
1177 if (!list_empty(list)) {
1180 spin_lock(&hctx->lock);
1181 list_splice_init(list, &hctx->dispatch);
1182 spin_unlock(&hctx->lock);
1185 * If SCHED_RESTART was set by the caller of this function and
1186 * it is no longer set that means that it was cleared by another
1187 * thread and hence that a queue rerun is needed.
1189 * If 'no_tag' is set, that means that we failed getting
1190 * a driver tag with an I/O scheduler attached. If our dispatch
1191 * waitqueue is no longer active, ensure that we run the queue
1192 * AFTER adding our entries back to the list.
1194 * If no I/O scheduler has been configured it is possible that
1195 * the hardware queue got stopped and restarted before requests
1196 * were pushed back onto the dispatch list. Rerun the queue to
1197 * avoid starvation. Notes:
1198 * - blk_mq_run_hw_queue() checks whether or not a queue has
1199 * been stopped before rerunning a queue.
1200 * - Some but not all block drivers stop a queue before
1201 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1204 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1205 * bit is set, run queue after a delay to avoid IO stalls
1206 * that could otherwise occur if the queue is idle.
1208 needs_restart = blk_mq_sched_needs_restart(hctx);
1209 if (!needs_restart ||
1210 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1211 blk_mq_run_hw_queue(hctx, true);
1212 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1213 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1216 return (queued + errors) != 0;
1219 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1224 * We should be running this queue from one of the CPUs that
1227 * There are at least two related races now between setting
1228 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1229 * __blk_mq_run_hw_queue():
1231 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1232 * but later it becomes online, then this warning is harmless
1235 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1236 * but later it becomes offline, then the warning can't be
1237 * triggered, and we depend on blk-mq timeout handler to
1238 * handle dispatched requests to this hctx
1240 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1241 cpu_online(hctx->next_cpu)) {
1242 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1243 raw_smp_processor_id(),
1244 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1249 * We can't run the queue inline with ints disabled. Ensure that
1250 * we catch bad users of this early.
1252 WARN_ON_ONCE(in_interrupt());
1254 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1256 hctx_lock(hctx, &srcu_idx);
1257 blk_mq_sched_dispatch_requests(hctx);
1258 hctx_unlock(hctx, srcu_idx);
1261 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1263 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1265 if (cpu >= nr_cpu_ids)
1266 cpu = cpumask_first(hctx->cpumask);
1271 * It'd be great if the workqueue API had a way to pass
1272 * in a mask and had some smarts for more clever placement.
1273 * For now we just round-robin here, switching for every
1274 * BLK_MQ_CPU_WORK_BATCH queued items.
1276 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1279 int next_cpu = hctx->next_cpu;
1281 if (hctx->queue->nr_hw_queues == 1)
1282 return WORK_CPU_UNBOUND;
1284 if (--hctx->next_cpu_batch <= 0) {
1286 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1288 if (next_cpu >= nr_cpu_ids)
1289 next_cpu = blk_mq_first_mapped_cpu(hctx);
1290 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1294 * Do unbound schedule if we can't find a online CPU for this hctx,
1295 * and it should only happen in the path of handling CPU DEAD.
1297 if (!cpu_online(next_cpu)) {
1304 * Make sure to re-select CPU next time once after CPUs
1305 * in hctx->cpumask become online again.
1307 hctx->next_cpu = next_cpu;
1308 hctx->next_cpu_batch = 1;
1309 return WORK_CPU_UNBOUND;
1312 hctx->next_cpu = next_cpu;
1316 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1317 unsigned long msecs)
1319 if (unlikely(blk_mq_hctx_stopped(hctx)))
1322 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1323 int cpu = get_cpu();
1324 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1325 __blk_mq_run_hw_queue(hctx);
1333 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1334 msecs_to_jiffies(msecs));
1337 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1339 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1341 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1343 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1349 * When queue is quiesced, we may be switching io scheduler, or
1350 * updating nr_hw_queues, or other things, and we can't run queue
1351 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1353 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1356 hctx_lock(hctx, &srcu_idx);
1357 need_run = !blk_queue_quiesced(hctx->queue) &&
1358 blk_mq_hctx_has_pending(hctx);
1359 hctx_unlock(hctx, srcu_idx);
1362 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1368 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1370 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1372 struct blk_mq_hw_ctx *hctx;
1375 queue_for_each_hw_ctx(q, hctx, i) {
1376 if (blk_mq_hctx_stopped(hctx))
1379 blk_mq_run_hw_queue(hctx, async);
1382 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1385 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1386 * @q: request queue.
1388 * The caller is responsible for serializing this function against
1389 * blk_mq_{start,stop}_hw_queue().
1391 bool blk_mq_queue_stopped(struct request_queue *q)
1393 struct blk_mq_hw_ctx *hctx;
1396 queue_for_each_hw_ctx(q, hctx, i)
1397 if (blk_mq_hctx_stopped(hctx))
1402 EXPORT_SYMBOL(blk_mq_queue_stopped);
1405 * This function is often used for pausing .queue_rq() by driver when
1406 * there isn't enough resource or some conditions aren't satisfied, and
1407 * BLK_STS_RESOURCE is usually returned.
1409 * We do not guarantee that dispatch can be drained or blocked
1410 * after blk_mq_stop_hw_queue() returns. Please use
1411 * blk_mq_quiesce_queue() for that requirement.
1413 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1415 cancel_delayed_work(&hctx->run_work);
1417 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1419 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1422 * This function is often used for pausing .queue_rq() by driver when
1423 * there isn't enough resource or some conditions aren't satisfied, and
1424 * BLK_STS_RESOURCE is usually returned.
1426 * We do not guarantee that dispatch can be drained or blocked
1427 * after blk_mq_stop_hw_queues() returns. Please use
1428 * blk_mq_quiesce_queue() for that requirement.
1430 void blk_mq_stop_hw_queues(struct request_queue *q)
1432 struct blk_mq_hw_ctx *hctx;
1435 queue_for_each_hw_ctx(q, hctx, i)
1436 blk_mq_stop_hw_queue(hctx);
1438 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1440 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1442 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1444 blk_mq_run_hw_queue(hctx, false);
1446 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1448 void blk_mq_start_hw_queues(struct request_queue *q)
1450 struct blk_mq_hw_ctx *hctx;
1453 queue_for_each_hw_ctx(q, hctx, i)
1454 blk_mq_start_hw_queue(hctx);
1456 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1458 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1460 if (!blk_mq_hctx_stopped(hctx))
1463 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1464 blk_mq_run_hw_queue(hctx, async);
1466 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1468 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1470 struct blk_mq_hw_ctx *hctx;
1473 queue_for_each_hw_ctx(q, hctx, i)
1474 blk_mq_start_stopped_hw_queue(hctx, async);
1476 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1478 static void blk_mq_run_work_fn(struct work_struct *work)
1480 struct blk_mq_hw_ctx *hctx;
1482 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1485 * If we are stopped, don't run the queue.
1487 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1488 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1490 __blk_mq_run_hw_queue(hctx);
1493 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1497 struct blk_mq_ctx *ctx = rq->mq_ctx;
1499 lockdep_assert_held(&ctx->lock);
1501 trace_block_rq_insert(hctx->queue, rq);
1504 list_add(&rq->queuelist, &ctx->rq_list);
1506 list_add_tail(&rq->queuelist, &ctx->rq_list);
1509 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1512 struct blk_mq_ctx *ctx = rq->mq_ctx;
1514 lockdep_assert_held(&ctx->lock);
1516 __blk_mq_insert_req_list(hctx, rq, at_head);
1517 blk_mq_hctx_mark_pending(hctx, ctx);
1521 * Should only be used carefully, when the caller knows we want to
1522 * bypass a potential IO scheduler on the target device.
1524 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1526 struct blk_mq_ctx *ctx = rq->mq_ctx;
1527 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1529 spin_lock(&hctx->lock);
1530 list_add_tail(&rq->queuelist, &hctx->dispatch);
1531 spin_unlock(&hctx->lock);
1534 blk_mq_run_hw_queue(hctx, false);
1537 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1538 struct list_head *list)
1542 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1545 spin_lock(&ctx->lock);
1546 while (!list_empty(list)) {
1549 rq = list_first_entry(list, struct request, queuelist);
1550 BUG_ON(rq->mq_ctx != ctx);
1551 list_del_init(&rq->queuelist);
1552 __blk_mq_insert_req_list(hctx, rq, false);
1554 blk_mq_hctx_mark_pending(hctx, ctx);
1555 spin_unlock(&ctx->lock);
1558 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1560 struct request *rqa = container_of(a, struct request, queuelist);
1561 struct request *rqb = container_of(b, struct request, queuelist);
1563 return !(rqa->mq_ctx < rqb->mq_ctx ||
1564 (rqa->mq_ctx == rqb->mq_ctx &&
1565 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1568 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1570 struct blk_mq_ctx *this_ctx;
1571 struct request_queue *this_q;
1574 LIST_HEAD(ctx_list);
1577 list_splice_init(&plug->mq_list, &list);
1579 list_sort(NULL, &list, plug_ctx_cmp);
1585 while (!list_empty(&list)) {
1586 rq = list_entry_rq(list.next);
1587 list_del_init(&rq->queuelist);
1589 if (rq->mq_ctx != this_ctx) {
1591 trace_block_unplug(this_q, depth, from_schedule);
1592 blk_mq_sched_insert_requests(this_q, this_ctx,
1597 this_ctx = rq->mq_ctx;
1603 list_add_tail(&rq->queuelist, &ctx_list);
1607 * If 'this_ctx' is set, we know we have entries to complete
1608 * on 'ctx_list'. Do those.
1611 trace_block_unplug(this_q, depth, from_schedule);
1612 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1617 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1619 blk_init_request_from_bio(rq, bio);
1621 blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1623 blk_account_io_start(rq, true);
1626 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1629 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1631 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1634 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1638 struct request_queue *q = rq->q;
1639 struct blk_mq_queue_data bd = {
1643 blk_qc_t new_cookie;
1646 new_cookie = request_to_qc_t(hctx, rq);
1649 * For OK queue, we are done. For error, caller may kill it.
1650 * Any other error (busy), just add it to our list as we
1651 * previously would have done.
1653 ret = q->mq_ops->queue_rq(hctx, &bd);
1656 *cookie = new_cookie;
1658 case BLK_STS_RESOURCE:
1659 case BLK_STS_DEV_RESOURCE:
1660 __blk_mq_requeue_request(rq);
1663 *cookie = BLK_QC_T_NONE;
1670 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1675 struct request_queue *q = rq->q;
1676 bool run_queue = true;
1679 * RCU or SRCU read lock is needed before checking quiesced flag.
1681 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1682 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1683 * and avoid driver to try to dispatch again.
1685 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1687 bypass_insert = false;
1691 if (q->elevator && !bypass_insert)
1694 if (!blk_mq_get_dispatch_budget(hctx))
1697 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1698 blk_mq_put_dispatch_budget(hctx);
1702 return __blk_mq_issue_directly(hctx, rq, cookie);
1705 return BLK_STS_RESOURCE;
1707 blk_mq_sched_insert_request(rq, false, run_queue, false);
1711 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1712 struct request *rq, blk_qc_t *cookie)
1717 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1719 hctx_lock(hctx, &srcu_idx);
1721 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1722 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1723 blk_mq_sched_insert_request(rq, false, true, false);
1724 else if (ret != BLK_STS_OK)
1725 blk_mq_end_request(rq, ret);
1727 hctx_unlock(hctx, srcu_idx);
1730 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1734 blk_qc_t unused_cookie;
1735 struct blk_mq_ctx *ctx = rq->mq_ctx;
1736 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1738 hctx_lock(hctx, &srcu_idx);
1739 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1740 hctx_unlock(hctx, srcu_idx);
1745 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1747 const int is_sync = op_is_sync(bio->bi_opf);
1748 const int is_flush_fua = op_is_flush(bio->bi_opf);
1749 struct blk_mq_alloc_data data = { .flags = 0 };
1751 unsigned int request_count = 0;
1752 struct blk_plug *plug;
1753 struct request *same_queue_rq = NULL;
1755 unsigned int wb_acct;
1757 blk_queue_bounce(q, &bio);
1759 blk_queue_split(q, &bio);
1761 if (!bio_integrity_prep(bio))
1762 return BLK_QC_T_NONE;
1764 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1765 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1766 return BLK_QC_T_NONE;
1768 if (blk_mq_sched_bio_merge(q, bio))
1769 return BLK_QC_T_NONE;
1771 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1773 trace_block_getrq(q, bio, bio->bi_opf);
1775 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1776 if (unlikely(!rq)) {
1777 __wbt_done(q->rq_wb, wb_acct);
1778 if (bio->bi_opf & REQ_NOWAIT)
1779 bio_wouldblock_error(bio);
1780 return BLK_QC_T_NONE;
1783 wbt_track(rq, wb_acct);
1785 cookie = request_to_qc_t(data.hctx, rq);
1787 plug = current->plug;
1788 if (unlikely(is_flush_fua)) {
1789 blk_mq_put_ctx(data.ctx);
1790 blk_mq_bio_to_request(rq, bio);
1792 /* bypass scheduler for flush rq */
1793 blk_insert_flush(rq);
1794 blk_mq_run_hw_queue(data.hctx, true);
1795 } else if (plug && q->nr_hw_queues == 1) {
1796 struct request *last = NULL;
1798 blk_mq_put_ctx(data.ctx);
1799 blk_mq_bio_to_request(rq, bio);
1802 * @request_count may become stale because of schedule
1803 * out, so check the list again.
1805 if (list_empty(&plug->mq_list))
1807 else if (blk_queue_nomerges(q))
1808 request_count = blk_plug_queued_count(q);
1811 trace_block_plug(q);
1813 last = list_entry_rq(plug->mq_list.prev);
1815 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1816 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1817 blk_flush_plug_list(plug, false);
1818 trace_block_plug(q);
1821 list_add_tail(&rq->queuelist, &plug->mq_list);
1822 } else if (plug && !blk_queue_nomerges(q)) {
1823 blk_mq_bio_to_request(rq, bio);
1826 * We do limited plugging. If the bio can be merged, do that.
1827 * Otherwise the existing request in the plug list will be
1828 * issued. So the plug list will have one request at most
1829 * The plug list might get flushed before this. If that happens,
1830 * the plug list is empty, and same_queue_rq is invalid.
1832 if (list_empty(&plug->mq_list))
1833 same_queue_rq = NULL;
1835 list_del_init(&same_queue_rq->queuelist);
1836 list_add_tail(&rq->queuelist, &plug->mq_list);
1838 blk_mq_put_ctx(data.ctx);
1840 if (same_queue_rq) {
1841 data.hctx = blk_mq_map_queue(q,
1842 same_queue_rq->mq_ctx->cpu);
1843 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1846 } else if (q->nr_hw_queues > 1 && is_sync) {
1847 blk_mq_put_ctx(data.ctx);
1848 blk_mq_bio_to_request(rq, bio);
1849 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1851 blk_mq_put_ctx(data.ctx);
1852 blk_mq_bio_to_request(rq, bio);
1853 blk_mq_sched_insert_request(rq, false, true, true);
1859 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1860 unsigned int hctx_idx)
1864 if (tags->rqs && set->ops->exit_request) {
1867 for (i = 0; i < tags->nr_tags; i++) {
1868 struct request *rq = tags->static_rqs[i];
1872 set->ops->exit_request(set, rq, hctx_idx);
1873 tags->static_rqs[i] = NULL;
1877 while (!list_empty(&tags->page_list)) {
1878 page = list_first_entry(&tags->page_list, struct page, lru);
1879 list_del_init(&page->lru);
1881 * Remove kmemleak object previously allocated in
1882 * blk_mq_init_rq_map().
1884 kmemleak_free(page_address(page));
1885 __free_pages(page, page->private);
1889 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1893 kfree(tags->static_rqs);
1894 tags->static_rqs = NULL;
1896 blk_mq_free_tags(tags);
1899 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1900 unsigned int hctx_idx,
1901 unsigned int nr_tags,
1902 unsigned int reserved_tags)
1904 struct blk_mq_tags *tags;
1907 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1908 if (node == NUMA_NO_NODE)
1909 node = set->numa_node;
1911 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1912 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1916 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1917 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1920 blk_mq_free_tags(tags);
1924 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1925 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1927 if (!tags->static_rqs) {
1929 blk_mq_free_tags(tags);
1936 static size_t order_to_size(unsigned int order)
1938 return (size_t)PAGE_SIZE << order;
1941 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1942 unsigned int hctx_idx, int node)
1946 if (set->ops->init_request) {
1947 ret = set->ops->init_request(set, rq, hctx_idx, node);
1952 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1956 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1957 unsigned int hctx_idx, unsigned int depth)
1959 unsigned int i, j, entries_per_page, max_order = 4;
1960 size_t rq_size, left;
1963 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1964 if (node == NUMA_NO_NODE)
1965 node = set->numa_node;
1967 INIT_LIST_HEAD(&tags->page_list);
1970 * rq_size is the size of the request plus driver payload, rounded
1971 * to the cacheline size
1973 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1975 left = rq_size * depth;
1977 for (i = 0; i < depth; ) {
1978 int this_order = max_order;
1983 while (this_order && left < order_to_size(this_order - 1))
1987 page = alloc_pages_node(node,
1988 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1994 if (order_to_size(this_order) < rq_size)
2001 page->private = this_order;
2002 list_add_tail(&page->lru, &tags->page_list);
2004 p = page_address(page);
2006 * Allow kmemleak to scan these pages as they contain pointers
2007 * to additional allocations like via ops->init_request().
2009 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2010 entries_per_page = order_to_size(this_order) / rq_size;
2011 to_do = min(entries_per_page, depth - i);
2012 left -= to_do * rq_size;
2013 for (j = 0; j < to_do; j++) {
2014 struct request *rq = p;
2016 tags->static_rqs[i] = rq;
2017 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2018 tags->static_rqs[i] = NULL;
2029 blk_mq_free_rqs(set, tags, hctx_idx);
2034 * 'cpu' is going away. splice any existing rq_list entries from this
2035 * software queue to the hw queue dispatch list, and ensure that it
2038 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2040 struct blk_mq_hw_ctx *hctx;
2041 struct blk_mq_ctx *ctx;
2044 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2045 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2047 spin_lock(&ctx->lock);
2048 if (!list_empty(&ctx->rq_list)) {
2049 list_splice_init(&ctx->rq_list, &tmp);
2050 blk_mq_hctx_clear_pending(hctx, ctx);
2052 spin_unlock(&ctx->lock);
2054 if (list_empty(&tmp))
2057 spin_lock(&hctx->lock);
2058 list_splice_tail_init(&tmp, &hctx->dispatch);
2059 spin_unlock(&hctx->lock);
2061 blk_mq_run_hw_queue(hctx, true);
2065 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2067 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2071 /* hctx->ctxs will be freed in queue's release handler */
2072 static void blk_mq_exit_hctx(struct request_queue *q,
2073 struct blk_mq_tag_set *set,
2074 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2076 blk_mq_debugfs_unregister_hctx(hctx);
2078 if (blk_mq_hw_queue_mapped(hctx))
2079 blk_mq_tag_idle(hctx);
2081 if (set->ops->exit_request)
2082 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2084 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2086 if (set->ops->exit_hctx)
2087 set->ops->exit_hctx(hctx, hctx_idx);
2089 if (hctx->flags & BLK_MQ_F_BLOCKING)
2090 cleanup_srcu_struct(hctx->srcu);
2092 blk_mq_remove_cpuhp(hctx);
2093 blk_free_flush_queue(hctx->fq);
2094 sbitmap_free(&hctx->ctx_map);
2097 static void blk_mq_exit_hw_queues(struct request_queue *q,
2098 struct blk_mq_tag_set *set, int nr_queue)
2100 struct blk_mq_hw_ctx *hctx;
2103 queue_for_each_hw_ctx(q, hctx, i) {
2106 blk_mq_exit_hctx(q, set, hctx, i);
2110 static int blk_mq_init_hctx(struct request_queue *q,
2111 struct blk_mq_tag_set *set,
2112 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2116 node = hctx->numa_node;
2117 if (node == NUMA_NO_NODE)
2118 node = hctx->numa_node = set->numa_node;
2120 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2121 spin_lock_init(&hctx->lock);
2122 INIT_LIST_HEAD(&hctx->dispatch);
2124 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2126 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2128 hctx->tags = set->tags[hctx_idx];
2131 * Allocate space for all possible cpus to avoid allocation at
2134 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2137 goto unregister_cpu_notifier;
2139 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2145 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2146 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2148 if (set->ops->init_hctx &&
2149 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2152 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2155 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2157 goto sched_exit_hctx;
2159 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2162 if (hctx->flags & BLK_MQ_F_BLOCKING)
2163 init_srcu_struct(hctx->srcu);
2165 blk_mq_debugfs_register_hctx(q, hctx);
2172 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2174 if (set->ops->exit_hctx)
2175 set->ops->exit_hctx(hctx, hctx_idx);
2177 sbitmap_free(&hctx->ctx_map);
2180 unregister_cpu_notifier:
2181 blk_mq_remove_cpuhp(hctx);
2185 static void blk_mq_init_cpu_queues(struct request_queue *q,
2186 unsigned int nr_hw_queues)
2190 for_each_possible_cpu(i) {
2191 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2192 struct blk_mq_hw_ctx *hctx;
2195 spin_lock_init(&__ctx->lock);
2196 INIT_LIST_HEAD(&__ctx->rq_list);
2200 * Set local node, IFF we have more than one hw queue. If
2201 * not, we remain on the home node of the device
2203 hctx = blk_mq_map_queue(q, i);
2204 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2205 hctx->numa_node = local_memory_node(cpu_to_node(i));
2209 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2213 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2214 set->queue_depth, set->reserved_tags);
2215 if (!set->tags[hctx_idx])
2218 ret = blk_mq_alloc_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_free_map_and_requests(struct blk_mq_tag_set *set,
2229 unsigned int hctx_idx)
2231 if (set->tags[hctx_idx]) {
2232 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2233 blk_mq_free_rq_map(set->tags[hctx_idx]);
2234 set->tags[hctx_idx] = NULL;
2238 static void blk_mq_map_swqueue(struct request_queue *q)
2240 unsigned int i, hctx_idx;
2241 struct blk_mq_hw_ctx *hctx;
2242 struct blk_mq_ctx *ctx;
2243 struct blk_mq_tag_set *set = q->tag_set;
2246 * Avoid others reading imcomplete hctx->cpumask through sysfs
2248 mutex_lock(&q->sysfs_lock);
2250 queue_for_each_hw_ctx(q, hctx, i) {
2251 cpumask_clear(hctx->cpumask);
2253 hctx->dispatch_from = NULL;
2257 * Map software to hardware queues.
2259 * If the cpu isn't present, the cpu is mapped to first hctx.
2261 for_each_possible_cpu(i) {
2262 hctx_idx = q->mq_map[i];
2263 /* unmapped hw queue can be remapped after CPU topo changed */
2264 if (!set->tags[hctx_idx] &&
2265 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2267 * If tags initialization fail for some hctx,
2268 * that hctx won't be brought online. In this
2269 * case, remap the current ctx to hctx[0] which
2270 * is guaranteed to always have tags allocated
2275 ctx = per_cpu_ptr(q->queue_ctx, i);
2276 hctx = blk_mq_map_queue(q, i);
2278 cpumask_set_cpu(i, hctx->cpumask);
2279 ctx->index_hw = hctx->nr_ctx;
2280 hctx->ctxs[hctx->nr_ctx++] = ctx;
2283 mutex_unlock(&q->sysfs_lock);
2285 queue_for_each_hw_ctx(q, hctx, i) {
2287 * If no software queues are mapped to this hardware queue,
2288 * disable it and free the request entries.
2290 if (!hctx->nr_ctx) {
2291 /* Never unmap queue 0. We need it as a
2292 * fallback in case of a new remap fails
2295 if (i && set->tags[i])
2296 blk_mq_free_map_and_requests(set, i);
2302 hctx->tags = set->tags[i];
2303 WARN_ON(!hctx->tags);
2306 * Set the map size to the number of mapped software queues.
2307 * This is more accurate and more efficient than looping
2308 * over all possibly mapped software queues.
2310 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2313 * Initialize batch roundrobin counts
2315 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2316 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2321 * Caller needs to ensure that we're either frozen/quiesced, or that
2322 * the queue isn't live yet.
2324 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2326 struct blk_mq_hw_ctx *hctx;
2329 queue_for_each_hw_ctx(q, hctx, i) {
2331 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2332 atomic_inc(&q->shared_hctx_restart);
2333 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2335 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2336 atomic_dec(&q->shared_hctx_restart);
2337 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2342 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2345 struct request_queue *q;
2347 lockdep_assert_held(&set->tag_list_lock);
2349 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2350 blk_mq_freeze_queue(q);
2351 queue_set_hctx_shared(q, shared);
2352 blk_mq_unfreeze_queue(q);
2356 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2358 struct blk_mq_tag_set *set = q->tag_set;
2360 mutex_lock(&set->tag_list_lock);
2361 list_del_rcu(&q->tag_set_list);
2362 INIT_LIST_HEAD(&q->tag_set_list);
2363 if (list_is_singular(&set->tag_list)) {
2364 /* just transitioned to unshared */
2365 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2366 /* update existing queue */
2367 blk_mq_update_tag_set_depth(set, false);
2369 mutex_unlock(&set->tag_list_lock);
2374 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2375 struct request_queue *q)
2379 mutex_lock(&set->tag_list_lock);
2382 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2384 if (!list_empty(&set->tag_list) &&
2385 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2386 set->flags |= BLK_MQ_F_TAG_SHARED;
2387 /* update existing queue */
2388 blk_mq_update_tag_set_depth(set, true);
2390 if (set->flags & BLK_MQ_F_TAG_SHARED)
2391 queue_set_hctx_shared(q, true);
2392 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2394 mutex_unlock(&set->tag_list_lock);
2398 * It is the actual release handler for mq, but we do it from
2399 * request queue's release handler for avoiding use-after-free
2400 * and headache because q->mq_kobj shouldn't have been introduced,
2401 * but we can't group ctx/kctx kobj without it.
2403 void blk_mq_release(struct request_queue *q)
2405 struct blk_mq_hw_ctx *hctx;
2408 /* hctx kobj stays in hctx */
2409 queue_for_each_hw_ctx(q, hctx, i) {
2412 kobject_put(&hctx->kobj);
2417 kfree(q->queue_hw_ctx);
2420 * release .mq_kobj and sw queue's kobject now because
2421 * both share lifetime with request queue.
2423 blk_mq_sysfs_deinit(q);
2425 free_percpu(q->queue_ctx);
2428 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2430 struct request_queue *uninit_q, *q;
2432 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2434 return ERR_PTR(-ENOMEM);
2436 q = blk_mq_init_allocated_queue(set, uninit_q);
2438 blk_cleanup_queue(uninit_q);
2442 EXPORT_SYMBOL(blk_mq_init_queue);
2444 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2446 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2448 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2449 __alignof__(struct blk_mq_hw_ctx)) !=
2450 sizeof(struct blk_mq_hw_ctx));
2452 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2453 hw_ctx_size += sizeof(struct srcu_struct);
2458 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2459 struct request_queue *q)
2462 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2464 blk_mq_sysfs_unregister(q);
2466 /* protect against switching io scheduler */
2467 mutex_lock(&q->sysfs_lock);
2468 for (i = 0; i < set->nr_hw_queues; i++) {
2474 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2475 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2480 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2487 atomic_set(&hctxs[i]->nr_active, 0);
2488 hctxs[i]->numa_node = node;
2489 hctxs[i]->queue_num = i;
2491 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2492 free_cpumask_var(hctxs[i]->cpumask);
2497 blk_mq_hctx_kobj_init(hctxs[i]);
2499 for (j = i; j < q->nr_hw_queues; j++) {
2500 struct blk_mq_hw_ctx *hctx = hctxs[j];
2504 blk_mq_free_map_and_requests(set, j);
2505 blk_mq_exit_hctx(q, set, hctx, j);
2506 kobject_put(&hctx->kobj);
2511 q->nr_hw_queues = i;
2512 mutex_unlock(&q->sysfs_lock);
2513 blk_mq_sysfs_register(q);
2516 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2517 struct request_queue *q)
2519 /* mark the queue as mq asap */
2520 q->mq_ops = set->ops;
2522 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2523 blk_mq_poll_stats_bkt,
2524 BLK_MQ_POLL_STATS_BKTS, q);
2528 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2532 /* init q->mq_kobj and sw queues' kobjects */
2533 blk_mq_sysfs_init(q);
2535 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2536 GFP_KERNEL, set->numa_node);
2537 if (!q->queue_hw_ctx)
2540 q->mq_map = set->mq_map;
2542 blk_mq_realloc_hw_ctxs(set, q);
2543 if (!q->nr_hw_queues)
2546 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2547 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2549 q->nr_queues = nr_cpu_ids;
2551 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2553 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2554 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2556 q->sg_reserved_size = INT_MAX;
2558 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2559 INIT_LIST_HEAD(&q->requeue_list);
2560 spin_lock_init(&q->requeue_lock);
2562 blk_queue_make_request(q, blk_mq_make_request);
2563 if (q->mq_ops->poll)
2564 q->poll_fn = blk_mq_poll;
2567 * Do this after blk_queue_make_request() overrides it...
2569 q->nr_requests = set->queue_depth;
2572 * Default to classic polling
2576 if (set->ops->complete)
2577 blk_queue_softirq_done(q, set->ops->complete);
2579 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2580 blk_mq_add_queue_tag_set(set, q);
2581 blk_mq_map_swqueue(q);
2583 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2586 ret = blk_mq_sched_init(q);
2588 return ERR_PTR(ret);
2594 kfree(q->queue_hw_ctx);
2596 free_percpu(q->queue_ctx);
2599 return ERR_PTR(-ENOMEM);
2601 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2603 void blk_mq_free_queue(struct request_queue *q)
2605 struct blk_mq_tag_set *set = q->tag_set;
2607 blk_mq_del_queue_tag_set(q);
2608 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2611 /* Basically redo blk_mq_init_queue with queue frozen */
2612 static void blk_mq_queue_reinit(struct request_queue *q)
2614 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2616 blk_mq_debugfs_unregister_hctxs(q);
2617 blk_mq_sysfs_unregister(q);
2620 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2621 * we should change hctx numa_node according to the new topology (this
2622 * involves freeing and re-allocating memory, worth doing?)
2624 blk_mq_map_swqueue(q);
2626 blk_mq_sysfs_register(q);
2627 blk_mq_debugfs_register_hctxs(q);
2630 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2634 for (i = 0; i < set->nr_hw_queues; i++)
2635 if (!__blk_mq_alloc_rq_map(set, i))
2642 blk_mq_free_rq_map(set->tags[i]);
2648 * Allocate the request maps associated with this tag_set. Note that this
2649 * may reduce the depth asked for, if memory is tight. set->queue_depth
2650 * will be updated to reflect the allocated depth.
2652 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2657 depth = set->queue_depth;
2659 err = __blk_mq_alloc_rq_maps(set);
2663 set->queue_depth >>= 1;
2664 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2668 } while (set->queue_depth);
2670 if (!set->queue_depth || err) {
2671 pr_err("blk-mq: failed to allocate request map\n");
2675 if (depth != set->queue_depth)
2676 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2677 depth, set->queue_depth);
2682 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2684 if (set->ops->map_queues) {
2687 * transport .map_queues is usually done in the following
2690 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2691 * mask = get_cpu_mask(queue)
2692 * for_each_cpu(cpu, mask)
2693 * set->mq_map[cpu] = queue;
2696 * When we need to remap, the table has to be cleared for
2697 * killing stale mapping since one CPU may not be mapped
2700 for_each_possible_cpu(cpu)
2701 set->mq_map[cpu] = 0;
2703 return set->ops->map_queues(set);
2705 return blk_mq_map_queues(set);
2709 * Alloc a tag set to be associated with one or more request queues.
2710 * May fail with EINVAL for various error conditions. May adjust the
2711 * requested depth down, if if it too large. In that case, the set
2712 * value will be stored in set->queue_depth.
2714 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2718 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2720 if (!set->nr_hw_queues)
2722 if (!set->queue_depth)
2724 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2727 if (!set->ops->queue_rq)
2730 if (!set->ops->get_budget ^ !set->ops->put_budget)
2733 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2734 pr_info("blk-mq: reduced tag depth to %u\n",
2736 set->queue_depth = BLK_MQ_MAX_DEPTH;
2740 * If a crashdump is active, then we are potentially in a very
2741 * memory constrained environment. Limit us to 1 queue and
2742 * 64 tags to prevent using too much memory.
2744 if (is_kdump_kernel()) {
2745 set->nr_hw_queues = 1;
2746 set->queue_depth = min(64U, set->queue_depth);
2749 * There is no use for more h/w queues than cpus.
2751 if (set->nr_hw_queues > nr_cpu_ids)
2752 set->nr_hw_queues = nr_cpu_ids;
2754 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2755 GFP_KERNEL, set->numa_node);
2760 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2761 GFP_KERNEL, set->numa_node);
2765 ret = blk_mq_update_queue_map(set);
2767 goto out_free_mq_map;
2769 ret = blk_mq_alloc_rq_maps(set);
2771 goto out_free_mq_map;
2773 mutex_init(&set->tag_list_lock);
2774 INIT_LIST_HEAD(&set->tag_list);
2786 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2788 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2792 for (i = 0; i < nr_cpu_ids; i++)
2793 blk_mq_free_map_and_requests(set, i);
2801 EXPORT_SYMBOL(blk_mq_free_tag_set);
2803 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2805 struct blk_mq_tag_set *set = q->tag_set;
2806 struct blk_mq_hw_ctx *hctx;
2812 blk_mq_freeze_queue(q);
2813 blk_mq_quiesce_queue(q);
2816 queue_for_each_hw_ctx(q, hctx, i) {
2820 * If we're using an MQ scheduler, just update the scheduler
2821 * queue depth. This is similar to what the old code would do.
2823 if (!hctx->sched_tags) {
2824 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2827 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2835 q->nr_requests = nr;
2837 blk_mq_unquiesce_queue(q);
2838 blk_mq_unfreeze_queue(q);
2843 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2846 struct request_queue *q;
2848 lockdep_assert_held(&set->tag_list_lock);
2850 if (nr_hw_queues > nr_cpu_ids)
2851 nr_hw_queues = nr_cpu_ids;
2852 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2855 list_for_each_entry(q, &set->tag_list, tag_set_list)
2856 blk_mq_freeze_queue(q);
2858 set->nr_hw_queues = nr_hw_queues;
2859 blk_mq_update_queue_map(set);
2860 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2861 blk_mq_realloc_hw_ctxs(set, q);
2862 blk_mq_queue_reinit(q);
2865 list_for_each_entry(q, &set->tag_list, tag_set_list)
2866 blk_mq_unfreeze_queue(q);
2869 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2871 mutex_lock(&set->tag_list_lock);
2872 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2873 mutex_unlock(&set->tag_list_lock);
2875 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2877 /* Enable polling stats and return whether they were already enabled. */
2878 static bool blk_poll_stats_enable(struct request_queue *q)
2880 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2881 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
2883 blk_stat_add_callback(q, q->poll_cb);
2887 static void blk_mq_poll_stats_start(struct request_queue *q)
2890 * We don't arm the callback if polling stats are not enabled or the
2891 * callback is already active.
2893 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2894 blk_stat_is_active(q->poll_cb))
2897 blk_stat_activate_msecs(q->poll_cb, 100);
2900 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2902 struct request_queue *q = cb->data;
2905 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2906 if (cb->stat[bucket].nr_samples)
2907 q->poll_stat[bucket] = cb->stat[bucket];
2911 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2912 struct blk_mq_hw_ctx *hctx,
2915 unsigned long ret = 0;
2919 * If stats collection isn't on, don't sleep but turn it on for
2922 if (!blk_poll_stats_enable(q))
2926 * As an optimistic guess, use half of the mean service time
2927 * for this type of request. We can (and should) make this smarter.
2928 * For instance, if the completion latencies are tight, we can
2929 * get closer than just half the mean. This is especially
2930 * important on devices where the completion latencies are longer
2931 * than ~10 usec. We do use the stats for the relevant IO size
2932 * if available which does lead to better estimates.
2934 bucket = blk_mq_poll_stats_bkt(rq);
2938 if (q->poll_stat[bucket].nr_samples)
2939 ret = (q->poll_stat[bucket].mean + 1) / 2;
2944 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2945 struct blk_mq_hw_ctx *hctx,
2948 struct hrtimer_sleeper hs;
2949 enum hrtimer_mode mode;
2953 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
2959 * -1: don't ever hybrid sleep
2960 * 0: use half of prev avg
2961 * >0: use this specific value
2963 if (q->poll_nsec == -1)
2965 else if (q->poll_nsec > 0)
2966 nsecs = q->poll_nsec;
2968 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2973 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
2976 * This will be replaced with the stats tracking code, using
2977 * 'avg_completion_time / 2' as the pre-sleep target.
2981 mode = HRTIMER_MODE_REL;
2982 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2983 hrtimer_set_expires(&hs.timer, kt);
2985 hrtimer_init_sleeper(&hs, current);
2987 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
2989 set_current_state(TASK_UNINTERRUPTIBLE);
2990 hrtimer_start_expires(&hs.timer, mode);
2993 hrtimer_cancel(&hs.timer);
2994 mode = HRTIMER_MODE_ABS;
2995 } while (hs.task && !signal_pending(current));
2997 __set_current_state(TASK_RUNNING);
2998 destroy_hrtimer_on_stack(&hs.timer);
3002 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3004 struct request_queue *q = hctx->queue;
3008 * If we sleep, have the caller restart the poll loop to reset
3009 * the state. Like for the other success return cases, the
3010 * caller is responsible for checking if the IO completed. If
3011 * the IO isn't complete, we'll get called again and will go
3012 * straight to the busy poll loop.
3014 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3017 hctx->poll_considered++;
3019 state = current->state;
3020 while (!need_resched()) {
3023 hctx->poll_invoked++;
3025 ret = q->mq_ops->poll(hctx, rq->tag);
3027 hctx->poll_success++;
3028 set_current_state(TASK_RUNNING);
3032 if (signal_pending_state(state, current))
3033 set_current_state(TASK_RUNNING);
3035 if (current->state == TASK_RUNNING)
3042 __set_current_state(TASK_RUNNING);
3046 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3048 struct blk_mq_hw_ctx *hctx;
3051 if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3054 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3055 if (!blk_qc_t_is_internal(cookie))
3056 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3058 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3060 * With scheduling, if the request has completed, we'll
3061 * get a NULL return here, as we clear the sched tag when
3062 * that happens. The request still remains valid, like always,
3063 * so we should be safe with just the NULL check.
3069 return __blk_mq_poll(hctx, rq);
3072 static int __init blk_mq_init(void)
3074 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3075 blk_mq_hctx_notify_dead);
3078 subsys_initcall(blk_mq_init);