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"
39 #include "blk-rq-qos.h"
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 const int bit = ctx->index_hw[hctx->type];
79 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
80 sbitmap_set_bit(&hctx->ctx_map, bit);
83 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
84 struct blk_mq_ctx *ctx)
86 const int bit = ctx->index_hw[hctx->type];
88 sbitmap_clear_bit(&hctx->ctx_map, bit);
92 struct hd_struct *part;
93 unsigned int *inflight;
96 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
97 struct request *rq, void *priv,
100 struct mq_inflight *mi = priv;
103 * index[0] counts the specific partition that was asked for.
105 if (rq->part == mi->part)
111 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
113 unsigned inflight[2];
114 struct mq_inflight mi = { .part = part, .inflight = inflight, };
116 inflight[0] = inflight[1] = 0;
117 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
122 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
123 struct request *rq, void *priv,
126 struct mq_inflight *mi = priv;
128 if (rq->part == mi->part)
129 mi->inflight[rq_data_dir(rq)]++;
134 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
135 unsigned int inflight[2])
137 struct mq_inflight mi = { .part = part, .inflight = inflight, };
139 inflight[0] = inflight[1] = 0;
140 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
143 void blk_freeze_queue_start(struct request_queue *q)
147 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
148 if (freeze_depth == 1) {
149 percpu_ref_kill(&q->q_usage_counter);
151 blk_mq_run_hw_queues(q, false);
154 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
156 void blk_mq_freeze_queue_wait(struct request_queue *q)
158 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
160 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
162 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
163 unsigned long timeout)
165 return wait_event_timeout(q->mq_freeze_wq,
166 percpu_ref_is_zero(&q->q_usage_counter),
169 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
172 * Guarantee no request is in use, so we can change any data structure of
173 * the queue afterward.
175 void blk_freeze_queue(struct request_queue *q)
178 * In the !blk_mq case we are only calling this to kill the
179 * q_usage_counter, otherwise this increases the freeze depth
180 * and waits for it to return to zero. For this reason there is
181 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
182 * exported to drivers as the only user for unfreeze is blk_mq.
184 blk_freeze_queue_start(q);
185 blk_mq_freeze_queue_wait(q);
188 void blk_mq_freeze_queue(struct request_queue *q)
191 * ...just an alias to keep freeze and unfreeze actions balanced
192 * in the blk_mq_* namespace
196 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
198 void blk_mq_unfreeze_queue(struct request_queue *q)
202 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
203 WARN_ON_ONCE(freeze_depth < 0);
205 percpu_ref_resurrect(&q->q_usage_counter);
206 wake_up_all(&q->mq_freeze_wq);
209 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
212 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
213 * mpt3sas driver such that this function can be removed.
215 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
217 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
219 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
222 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
225 * Note: this function does not prevent that the struct request end_io()
226 * callback function is invoked. Once this function is returned, we make
227 * sure no dispatch can happen until the queue is unquiesced via
228 * blk_mq_unquiesce_queue().
230 void blk_mq_quiesce_queue(struct request_queue *q)
232 struct blk_mq_hw_ctx *hctx;
236 blk_mq_quiesce_queue_nowait(q);
238 queue_for_each_hw_ctx(q, hctx, i) {
239 if (hctx->flags & BLK_MQ_F_BLOCKING)
240 synchronize_srcu(hctx->srcu);
247 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
250 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
253 * This function recovers queue into the state before quiescing
254 * which is done by blk_mq_quiesce_queue.
256 void blk_mq_unquiesce_queue(struct request_queue *q)
258 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
260 /* dispatch requests which are inserted during quiescing */
261 blk_mq_run_hw_queues(q, true);
263 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
265 void blk_mq_wake_waiters(struct request_queue *q)
267 struct blk_mq_hw_ctx *hctx;
270 queue_for_each_hw_ctx(q, hctx, i)
271 if (blk_mq_hw_queue_mapped(hctx))
272 blk_mq_tag_wakeup_all(hctx->tags, true);
275 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
277 return blk_mq_has_free_tags(hctx->tags);
279 EXPORT_SYMBOL(blk_mq_can_queue);
282 * Only need start/end time stamping if we have stats enabled, or using
285 static inline bool blk_mq_need_time_stamp(struct request *rq)
287 return (rq->rq_flags & RQF_IO_STAT) || rq->q->elevator;
290 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
291 unsigned int tag, unsigned int op)
293 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
294 struct request *rq = tags->static_rqs[tag];
295 req_flags_t rq_flags = 0;
297 if (data->flags & BLK_MQ_REQ_INTERNAL) {
299 rq->internal_tag = tag;
301 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
302 rq_flags = RQF_MQ_INFLIGHT;
303 atomic_inc(&data->hctx->nr_active);
306 rq->internal_tag = -1;
307 data->hctx->tags->rqs[rq->tag] = rq;
310 /* csd/requeue_work/fifo_time is initialized before use */
312 rq->mq_ctx = data->ctx;
313 rq->mq_hctx = data->hctx;
314 rq->rq_flags = rq_flags;
316 if (data->flags & BLK_MQ_REQ_PREEMPT)
317 rq->rq_flags |= RQF_PREEMPT;
318 if (blk_queue_io_stat(data->q))
319 rq->rq_flags |= RQF_IO_STAT;
320 INIT_LIST_HEAD(&rq->queuelist);
321 INIT_HLIST_NODE(&rq->hash);
322 RB_CLEAR_NODE(&rq->rb_node);
325 if (blk_mq_need_time_stamp(rq))
326 rq->start_time_ns = ktime_get_ns();
328 rq->start_time_ns = 0;
329 rq->io_start_time_ns = 0;
330 rq->nr_phys_segments = 0;
331 #if defined(CONFIG_BLK_DEV_INTEGRITY)
332 rq->nr_integrity_segments = 0;
335 /* tag was already set */
337 WRITE_ONCE(rq->deadline, 0);
342 rq->end_io_data = NULL;
345 data->ctx->rq_dispatched[op_is_sync(op)]++;
346 refcount_set(&rq->ref, 1);
350 static struct request *blk_mq_get_request(struct request_queue *q,
352 struct blk_mq_alloc_data *data)
354 struct elevator_queue *e = q->elevator;
357 bool put_ctx_on_error = false;
359 blk_queue_enter_live(q);
361 if (likely(!data->ctx)) {
362 data->ctx = blk_mq_get_ctx(q);
363 put_ctx_on_error = true;
365 if (likely(!data->hctx))
366 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
368 if (data->cmd_flags & REQ_NOWAIT)
369 data->flags |= BLK_MQ_REQ_NOWAIT;
372 data->flags |= BLK_MQ_REQ_INTERNAL;
375 * Flush requests are special and go directly to the
376 * dispatch list. Don't include reserved tags in the
377 * limiting, as it isn't useful.
379 if (!op_is_flush(data->cmd_flags) &&
380 e->type->ops.limit_depth &&
381 !(data->flags & BLK_MQ_REQ_RESERVED))
382 e->type->ops.limit_depth(data->cmd_flags, data);
384 blk_mq_tag_busy(data->hctx);
387 tag = blk_mq_get_tag(data);
388 if (tag == BLK_MQ_TAG_FAIL) {
389 if (put_ctx_on_error) {
390 blk_mq_put_ctx(data->ctx);
397 rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags);
398 if (!op_is_flush(data->cmd_flags)) {
400 if (e && e->type->ops.prepare_request) {
401 if (e->type->icq_cache)
402 blk_mq_sched_assign_ioc(rq);
404 e->type->ops.prepare_request(rq, bio);
405 rq->rq_flags |= RQF_ELVPRIV;
408 data->hctx->queued++;
412 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
413 blk_mq_req_flags_t flags)
415 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
419 ret = blk_queue_enter(q, flags);
423 rq = blk_mq_get_request(q, NULL, &alloc_data);
427 return ERR_PTR(-EWOULDBLOCK);
429 blk_mq_put_ctx(alloc_data.ctx);
432 rq->__sector = (sector_t) -1;
433 rq->bio = rq->biotail = NULL;
436 EXPORT_SYMBOL(blk_mq_alloc_request);
438 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
439 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
441 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
447 * If the tag allocator sleeps we could get an allocation for a
448 * different hardware context. No need to complicate the low level
449 * allocator for this for the rare use case of a command tied to
452 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
453 return ERR_PTR(-EINVAL);
455 if (hctx_idx >= q->nr_hw_queues)
456 return ERR_PTR(-EIO);
458 ret = blk_queue_enter(q, flags);
463 * Check if the hardware context is actually mapped to anything.
464 * If not tell the caller that it should skip this queue.
466 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
467 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
469 return ERR_PTR(-EXDEV);
471 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
472 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
474 rq = blk_mq_get_request(q, NULL, &alloc_data);
478 return ERR_PTR(-EWOULDBLOCK);
482 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
484 static void __blk_mq_free_request(struct request *rq)
486 struct request_queue *q = rq->q;
487 struct blk_mq_ctx *ctx = rq->mq_ctx;
488 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
489 const int sched_tag = rq->internal_tag;
491 blk_pm_mark_last_busy(rq);
494 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
496 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
497 blk_mq_sched_restart(hctx);
501 void blk_mq_free_request(struct request *rq)
503 struct request_queue *q = rq->q;
504 struct elevator_queue *e = q->elevator;
505 struct blk_mq_ctx *ctx = rq->mq_ctx;
506 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
508 if (rq->rq_flags & RQF_ELVPRIV) {
509 if (e && e->type->ops.finish_request)
510 e->type->ops.finish_request(rq);
512 put_io_context(rq->elv.icq->ioc);
517 ctx->rq_completed[rq_is_sync(rq)]++;
518 if (rq->rq_flags & RQF_MQ_INFLIGHT)
519 atomic_dec(&hctx->nr_active);
521 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
522 laptop_io_completion(q->backing_dev_info);
526 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
527 if (refcount_dec_and_test(&rq->ref))
528 __blk_mq_free_request(rq);
530 EXPORT_SYMBOL_GPL(blk_mq_free_request);
532 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
536 if (blk_mq_need_time_stamp(rq))
537 now = ktime_get_ns();
539 if (rq->rq_flags & RQF_STATS) {
540 blk_mq_poll_stats_start(rq->q);
541 blk_stat_add(rq, now);
544 if (rq->internal_tag != -1)
545 blk_mq_sched_completed_request(rq, now);
547 blk_account_io_done(rq, now);
550 rq_qos_done(rq->q, rq);
551 rq->end_io(rq, error);
553 if (unlikely(blk_bidi_rq(rq)))
554 blk_mq_free_request(rq->next_rq);
555 blk_mq_free_request(rq);
558 EXPORT_SYMBOL(__blk_mq_end_request);
560 void blk_mq_end_request(struct request *rq, blk_status_t error)
562 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
564 __blk_mq_end_request(rq, error);
566 EXPORT_SYMBOL(blk_mq_end_request);
568 static void __blk_mq_complete_request_remote(void *data)
570 struct request *rq = data;
571 struct request_queue *q = rq->q;
573 q->mq_ops->complete(rq);
576 static void __blk_mq_complete_request(struct request *rq)
578 struct blk_mq_ctx *ctx = rq->mq_ctx;
579 struct request_queue *q = rq->q;
583 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
585 * Most of single queue controllers, there is only one irq vector
586 * for handling IO completion, and the only irq's affinity is set
587 * as all possible CPUs. On most of ARCHs, this affinity means the
588 * irq is handled on one specific CPU.
590 * So complete IO reqeust in softirq context in case of single queue
591 * for not degrading IO performance by irqsoff latency.
593 if (q->nr_hw_queues == 1) {
594 __blk_complete_request(rq);
599 * For a polled request, always complete locallly, it's pointless
600 * to redirect the completion.
602 if ((rq->cmd_flags & REQ_HIPRI) ||
603 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
604 q->mq_ops->complete(rq);
609 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
610 shared = cpus_share_cache(cpu, ctx->cpu);
612 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
613 rq->csd.func = __blk_mq_complete_request_remote;
616 smp_call_function_single_async(ctx->cpu, &rq->csd);
618 q->mq_ops->complete(rq);
623 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
624 __releases(hctx->srcu)
626 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
629 srcu_read_unlock(hctx->srcu, srcu_idx);
632 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
633 __acquires(hctx->srcu)
635 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
636 /* shut up gcc false positive */
640 *srcu_idx = srcu_read_lock(hctx->srcu);
644 * blk_mq_complete_request - end I/O on a request
645 * @rq: the request being processed
648 * Ends all I/O on a request. It does not handle partial completions.
649 * The actual completion happens out-of-order, through a IPI handler.
651 bool blk_mq_complete_request(struct request *rq)
653 if (unlikely(blk_should_fake_timeout(rq->q)))
655 __blk_mq_complete_request(rq);
658 EXPORT_SYMBOL(blk_mq_complete_request);
660 int blk_mq_request_started(struct request *rq)
662 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
664 EXPORT_SYMBOL_GPL(blk_mq_request_started);
666 void blk_mq_start_request(struct request *rq)
668 struct request_queue *q = rq->q;
670 blk_mq_sched_started_request(rq);
672 trace_block_rq_issue(q, rq);
674 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
675 rq->io_start_time_ns = ktime_get_ns();
676 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
677 rq->throtl_size = blk_rq_sectors(rq);
679 rq->rq_flags |= RQF_STATS;
683 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
686 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
688 if (q->dma_drain_size && blk_rq_bytes(rq)) {
690 * Make sure space for the drain appears. We know we can do
691 * this because max_hw_segments has been adjusted to be one
692 * fewer than the device can handle.
694 rq->nr_phys_segments++;
697 EXPORT_SYMBOL(blk_mq_start_request);
699 static void __blk_mq_requeue_request(struct request *rq)
701 struct request_queue *q = rq->q;
703 blk_mq_put_driver_tag(rq);
705 trace_block_rq_requeue(q, rq);
706 rq_qos_requeue(q, rq);
708 if (blk_mq_request_started(rq)) {
709 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
710 rq->rq_flags &= ~RQF_TIMED_OUT;
711 if (q->dma_drain_size && blk_rq_bytes(rq))
712 rq->nr_phys_segments--;
716 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
718 __blk_mq_requeue_request(rq);
720 /* this request will be re-inserted to io scheduler queue */
721 blk_mq_sched_requeue_request(rq);
723 BUG_ON(!list_empty(&rq->queuelist));
724 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
726 EXPORT_SYMBOL(blk_mq_requeue_request);
728 static void blk_mq_requeue_work(struct work_struct *work)
730 struct request_queue *q =
731 container_of(work, struct request_queue, requeue_work.work);
733 struct request *rq, *next;
735 spin_lock_irq(&q->requeue_lock);
736 list_splice_init(&q->requeue_list, &rq_list);
737 spin_unlock_irq(&q->requeue_lock);
739 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
740 if (!(rq->rq_flags & RQF_SOFTBARRIER))
743 rq->rq_flags &= ~RQF_SOFTBARRIER;
744 list_del_init(&rq->queuelist);
745 blk_mq_sched_insert_request(rq, true, false, false);
748 while (!list_empty(&rq_list)) {
749 rq = list_entry(rq_list.next, struct request, queuelist);
750 list_del_init(&rq->queuelist);
751 blk_mq_sched_insert_request(rq, false, false, false);
754 blk_mq_run_hw_queues(q, false);
757 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
758 bool kick_requeue_list)
760 struct request_queue *q = rq->q;
764 * We abuse this flag that is otherwise used by the I/O scheduler to
765 * request head insertion from the workqueue.
767 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
769 spin_lock_irqsave(&q->requeue_lock, flags);
771 rq->rq_flags |= RQF_SOFTBARRIER;
772 list_add(&rq->queuelist, &q->requeue_list);
774 list_add_tail(&rq->queuelist, &q->requeue_list);
776 spin_unlock_irqrestore(&q->requeue_lock, flags);
778 if (kick_requeue_list)
779 blk_mq_kick_requeue_list(q);
781 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
783 void blk_mq_kick_requeue_list(struct request_queue *q)
785 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
787 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
789 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
792 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
793 msecs_to_jiffies(msecs));
795 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
797 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
799 if (tag < tags->nr_tags) {
800 prefetch(tags->rqs[tag]);
801 return tags->rqs[tag];
806 EXPORT_SYMBOL(blk_mq_tag_to_rq);
808 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
809 void *priv, bool reserved)
812 * If we find a request that is inflight and the queue matches,
813 * we know the queue is busy. Return false to stop the iteration.
815 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
825 bool blk_mq_queue_inflight(struct request_queue *q)
829 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
832 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
834 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
836 req->rq_flags |= RQF_TIMED_OUT;
837 if (req->q->mq_ops->timeout) {
838 enum blk_eh_timer_return ret;
840 ret = req->q->mq_ops->timeout(req, reserved);
841 if (ret == BLK_EH_DONE)
843 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
849 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
851 unsigned long deadline;
853 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
855 if (rq->rq_flags & RQF_TIMED_OUT)
858 deadline = READ_ONCE(rq->deadline);
859 if (time_after_eq(jiffies, deadline))
864 else if (time_after(*next, deadline))
869 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
870 struct request *rq, void *priv, bool reserved)
872 unsigned long *next = priv;
875 * Just do a quick check if it is expired before locking the request in
876 * so we're not unnecessarilly synchronizing across CPUs.
878 if (!blk_mq_req_expired(rq, next))
882 * We have reason to believe the request may be expired. Take a
883 * reference on the request to lock this request lifetime into its
884 * currently allocated context to prevent it from being reallocated in
885 * the event the completion by-passes this timeout handler.
887 * If the reference was already released, then the driver beat the
888 * timeout handler to posting a natural completion.
890 if (!refcount_inc_not_zero(&rq->ref))
894 * The request is now locked and cannot be reallocated underneath the
895 * timeout handler's processing. Re-verify this exact request is truly
896 * expired; if it is not expired, then the request was completed and
897 * reallocated as a new request.
899 if (blk_mq_req_expired(rq, next))
900 blk_mq_rq_timed_out(rq, reserved);
901 if (refcount_dec_and_test(&rq->ref))
902 __blk_mq_free_request(rq);
907 static void blk_mq_timeout_work(struct work_struct *work)
909 struct request_queue *q =
910 container_of(work, struct request_queue, timeout_work);
911 unsigned long next = 0;
912 struct blk_mq_hw_ctx *hctx;
915 /* A deadlock might occur if a request is stuck requiring a
916 * timeout at the same time a queue freeze is waiting
917 * completion, since the timeout code would not be able to
918 * acquire the queue reference here.
920 * That's why we don't use blk_queue_enter here; instead, we use
921 * percpu_ref_tryget directly, because we need to be able to
922 * obtain a reference even in the short window between the queue
923 * starting to freeze, by dropping the first reference in
924 * blk_freeze_queue_start, and the moment the last request is
925 * consumed, marked by the instant q_usage_counter reaches
928 if (!percpu_ref_tryget(&q->q_usage_counter))
931 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
934 mod_timer(&q->timeout, next);
937 * Request timeouts are handled as a forward rolling timer. If
938 * we end up here it means that no requests are pending and
939 * also that no request has been pending for a while. Mark
942 queue_for_each_hw_ctx(q, hctx, i) {
943 /* the hctx may be unmapped, so check it here */
944 if (blk_mq_hw_queue_mapped(hctx))
945 blk_mq_tag_idle(hctx);
951 struct flush_busy_ctx_data {
952 struct blk_mq_hw_ctx *hctx;
953 struct list_head *list;
956 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
958 struct flush_busy_ctx_data *flush_data = data;
959 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
960 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
961 enum hctx_type type = hctx->type;
963 spin_lock(&ctx->lock);
964 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
965 sbitmap_clear_bit(sb, bitnr);
966 spin_unlock(&ctx->lock);
971 * Process software queues that have been marked busy, splicing them
972 * to the for-dispatch
974 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
976 struct flush_busy_ctx_data data = {
981 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
983 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
985 struct dispatch_rq_data {
986 struct blk_mq_hw_ctx *hctx;
990 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
993 struct dispatch_rq_data *dispatch_data = data;
994 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
995 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
996 enum hctx_type type = hctx->type;
998 spin_lock(&ctx->lock);
999 if (!list_empty(&ctx->rq_lists[type])) {
1000 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1001 list_del_init(&dispatch_data->rq->queuelist);
1002 if (list_empty(&ctx->rq_lists[type]))
1003 sbitmap_clear_bit(sb, bitnr);
1005 spin_unlock(&ctx->lock);
1007 return !dispatch_data->rq;
1010 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1011 struct blk_mq_ctx *start)
1013 unsigned off = start ? start->index_hw[hctx->type] : 0;
1014 struct dispatch_rq_data data = {
1019 __sbitmap_for_each_set(&hctx->ctx_map, off,
1020 dispatch_rq_from_ctx, &data);
1025 static inline unsigned int queued_to_index(unsigned int queued)
1030 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1033 bool blk_mq_get_driver_tag(struct request *rq)
1035 struct blk_mq_alloc_data data = {
1037 .hctx = rq->mq_hctx,
1038 .flags = BLK_MQ_REQ_NOWAIT,
1039 .cmd_flags = rq->cmd_flags,
1046 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1047 data.flags |= BLK_MQ_REQ_RESERVED;
1049 shared = blk_mq_tag_busy(data.hctx);
1050 rq->tag = blk_mq_get_tag(&data);
1053 rq->rq_flags |= RQF_MQ_INFLIGHT;
1054 atomic_inc(&data.hctx->nr_active);
1056 data.hctx->tags->rqs[rq->tag] = rq;
1060 return rq->tag != -1;
1063 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1064 int flags, void *key)
1066 struct blk_mq_hw_ctx *hctx;
1068 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1070 spin_lock(&hctx->dispatch_wait_lock);
1071 list_del_init(&wait->entry);
1072 spin_unlock(&hctx->dispatch_wait_lock);
1074 blk_mq_run_hw_queue(hctx, true);
1079 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1080 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1081 * restart. For both cases, take care to check the condition again after
1082 * marking us as waiting.
1084 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1087 struct wait_queue_head *wq;
1088 wait_queue_entry_t *wait;
1091 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1092 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
1093 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
1096 * It's possible that a tag was freed in the window between the
1097 * allocation failure and adding the hardware queue to the wait
1100 * Don't clear RESTART here, someone else could have set it.
1101 * At most this will cost an extra queue run.
1103 return blk_mq_get_driver_tag(rq);
1106 wait = &hctx->dispatch_wait;
1107 if (!list_empty_careful(&wait->entry))
1110 wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait;
1112 spin_lock_irq(&wq->lock);
1113 spin_lock(&hctx->dispatch_wait_lock);
1114 if (!list_empty(&wait->entry)) {
1115 spin_unlock(&hctx->dispatch_wait_lock);
1116 spin_unlock_irq(&wq->lock);
1120 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1121 __add_wait_queue(wq, wait);
1124 * It's possible that a tag was freed in the window between the
1125 * allocation failure and adding the hardware queue to the wait
1128 ret = blk_mq_get_driver_tag(rq);
1130 spin_unlock(&hctx->dispatch_wait_lock);
1131 spin_unlock_irq(&wq->lock);
1136 * We got a tag, remove ourselves from the wait queue to ensure
1137 * someone else gets the wakeup.
1139 list_del_init(&wait->entry);
1140 spin_unlock(&hctx->dispatch_wait_lock);
1141 spin_unlock_irq(&wq->lock);
1146 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1147 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1149 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1150 * - EWMA is one simple way to compute running average value
1151 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1152 * - take 4 as factor for avoiding to get too small(0) result, and this
1153 * factor doesn't matter because EWMA decreases exponentially
1155 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1159 if (hctx->queue->elevator)
1162 ewma = hctx->dispatch_busy;
1167 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1169 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1170 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1172 hctx->dispatch_busy = ewma;
1175 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1178 * Returns true if we did some work AND can potentially do more.
1180 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1183 struct blk_mq_hw_ctx *hctx;
1184 struct request *rq, *nxt;
1185 bool no_tag = false;
1187 blk_status_t ret = BLK_STS_OK;
1189 if (list_empty(list))
1192 WARN_ON(!list_is_singular(list) && got_budget);
1195 * Now process all the entries, sending them to the driver.
1197 errors = queued = 0;
1199 struct blk_mq_queue_data bd;
1201 rq = list_first_entry(list, struct request, queuelist);
1204 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1207 if (!blk_mq_get_driver_tag(rq)) {
1209 * The initial allocation attempt failed, so we need to
1210 * rerun the hardware queue when a tag is freed. The
1211 * waitqueue takes care of that. If the queue is run
1212 * before we add this entry back on the dispatch list,
1213 * we'll re-run it below.
1215 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1216 blk_mq_put_dispatch_budget(hctx);
1218 * For non-shared tags, the RESTART check
1221 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1227 list_del_init(&rq->queuelist);
1232 * Flag last if we have no more requests, or if we have more
1233 * but can't assign a driver tag to it.
1235 if (list_empty(list))
1238 nxt = list_first_entry(list, struct request, queuelist);
1239 bd.last = !blk_mq_get_driver_tag(nxt);
1242 ret = q->mq_ops->queue_rq(hctx, &bd);
1243 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1245 * If an I/O scheduler has been configured and we got a
1246 * driver tag for the next request already, free it
1249 if (!list_empty(list)) {
1250 nxt = list_first_entry(list, struct request, queuelist);
1251 blk_mq_put_driver_tag(nxt);
1253 list_add(&rq->queuelist, list);
1254 __blk_mq_requeue_request(rq);
1258 if (unlikely(ret != BLK_STS_OK)) {
1260 blk_mq_end_request(rq, BLK_STS_IOERR);
1265 } while (!list_empty(list));
1267 hctx->dispatched[queued_to_index(queued)]++;
1270 * Any items that need requeuing? Stuff them into hctx->dispatch,
1271 * that is where we will continue on next queue run.
1273 if (!list_empty(list)) {
1277 * If we didn't flush the entire list, we could have told
1278 * the driver there was more coming, but that turned out to
1281 if (q->mq_ops->commit_rqs)
1282 q->mq_ops->commit_rqs(hctx);
1284 spin_lock(&hctx->lock);
1285 list_splice_init(list, &hctx->dispatch);
1286 spin_unlock(&hctx->lock);
1289 * If SCHED_RESTART was set by the caller of this function and
1290 * it is no longer set that means that it was cleared by another
1291 * thread and hence that a queue rerun is needed.
1293 * If 'no_tag' is set, that means that we failed getting
1294 * a driver tag with an I/O scheduler attached. If our dispatch
1295 * waitqueue is no longer active, ensure that we run the queue
1296 * AFTER adding our entries back to the list.
1298 * If no I/O scheduler has been configured it is possible that
1299 * the hardware queue got stopped and restarted before requests
1300 * were pushed back onto the dispatch list. Rerun the queue to
1301 * avoid starvation. Notes:
1302 * - blk_mq_run_hw_queue() checks whether or not a queue has
1303 * been stopped before rerunning a queue.
1304 * - Some but not all block drivers stop a queue before
1305 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1308 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1309 * bit is set, run queue after a delay to avoid IO stalls
1310 * that could otherwise occur if the queue is idle.
1312 needs_restart = blk_mq_sched_needs_restart(hctx);
1313 if (!needs_restart ||
1314 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1315 blk_mq_run_hw_queue(hctx, true);
1316 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1317 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1319 blk_mq_update_dispatch_busy(hctx, true);
1322 blk_mq_update_dispatch_busy(hctx, false);
1325 * If the host/device is unable to accept more work, inform the
1328 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1331 return (queued + errors) != 0;
1334 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1339 * We should be running this queue from one of the CPUs that
1342 * There are at least two related races now between setting
1343 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1344 * __blk_mq_run_hw_queue():
1346 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1347 * but later it becomes online, then this warning is harmless
1350 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1351 * but later it becomes offline, then the warning can't be
1352 * triggered, and we depend on blk-mq timeout handler to
1353 * handle dispatched requests to this hctx
1355 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1356 cpu_online(hctx->next_cpu)) {
1357 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1358 raw_smp_processor_id(),
1359 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1364 * We can't run the queue inline with ints disabled. Ensure that
1365 * we catch bad users of this early.
1367 WARN_ON_ONCE(in_interrupt());
1369 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1371 hctx_lock(hctx, &srcu_idx);
1372 blk_mq_sched_dispatch_requests(hctx);
1373 hctx_unlock(hctx, srcu_idx);
1376 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1378 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1380 if (cpu >= nr_cpu_ids)
1381 cpu = cpumask_first(hctx->cpumask);
1386 * It'd be great if the workqueue API had a way to pass
1387 * in a mask and had some smarts for more clever placement.
1388 * For now we just round-robin here, switching for every
1389 * BLK_MQ_CPU_WORK_BATCH queued items.
1391 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1394 int next_cpu = hctx->next_cpu;
1396 if (hctx->queue->nr_hw_queues == 1)
1397 return WORK_CPU_UNBOUND;
1399 if (--hctx->next_cpu_batch <= 0) {
1401 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1403 if (next_cpu >= nr_cpu_ids)
1404 next_cpu = blk_mq_first_mapped_cpu(hctx);
1405 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1409 * Do unbound schedule if we can't find a online CPU for this hctx,
1410 * and it should only happen in the path of handling CPU DEAD.
1412 if (!cpu_online(next_cpu)) {
1419 * Make sure to re-select CPU next time once after CPUs
1420 * in hctx->cpumask become online again.
1422 hctx->next_cpu = next_cpu;
1423 hctx->next_cpu_batch = 1;
1424 return WORK_CPU_UNBOUND;
1427 hctx->next_cpu = next_cpu;
1431 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1432 unsigned long msecs)
1434 if (unlikely(blk_mq_hctx_stopped(hctx)))
1437 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1438 int cpu = get_cpu();
1439 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1440 __blk_mq_run_hw_queue(hctx);
1448 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1449 msecs_to_jiffies(msecs));
1452 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1454 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1456 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1458 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1464 * When queue is quiesced, we may be switching io scheduler, or
1465 * updating nr_hw_queues, or other things, and we can't run queue
1466 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1468 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1471 hctx_lock(hctx, &srcu_idx);
1472 need_run = !blk_queue_quiesced(hctx->queue) &&
1473 blk_mq_hctx_has_pending(hctx);
1474 hctx_unlock(hctx, srcu_idx);
1477 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1483 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1485 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1487 struct blk_mq_hw_ctx *hctx;
1490 queue_for_each_hw_ctx(q, hctx, i) {
1491 if (blk_mq_hctx_stopped(hctx))
1494 blk_mq_run_hw_queue(hctx, async);
1497 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1500 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1501 * @q: request queue.
1503 * The caller is responsible for serializing this function against
1504 * blk_mq_{start,stop}_hw_queue().
1506 bool blk_mq_queue_stopped(struct request_queue *q)
1508 struct blk_mq_hw_ctx *hctx;
1511 queue_for_each_hw_ctx(q, hctx, i)
1512 if (blk_mq_hctx_stopped(hctx))
1517 EXPORT_SYMBOL(blk_mq_queue_stopped);
1520 * This function is often used for pausing .queue_rq() by driver when
1521 * there isn't enough resource or some conditions aren't satisfied, and
1522 * BLK_STS_RESOURCE is usually returned.
1524 * We do not guarantee that dispatch can be drained or blocked
1525 * after blk_mq_stop_hw_queue() returns. Please use
1526 * blk_mq_quiesce_queue() for that requirement.
1528 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1530 cancel_delayed_work(&hctx->run_work);
1532 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1534 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1537 * This function is often used for pausing .queue_rq() by driver when
1538 * there isn't enough resource or some conditions aren't satisfied, and
1539 * BLK_STS_RESOURCE is usually returned.
1541 * We do not guarantee that dispatch can be drained or blocked
1542 * after blk_mq_stop_hw_queues() returns. Please use
1543 * blk_mq_quiesce_queue() for that requirement.
1545 void blk_mq_stop_hw_queues(struct request_queue *q)
1547 struct blk_mq_hw_ctx *hctx;
1550 queue_for_each_hw_ctx(q, hctx, i)
1551 blk_mq_stop_hw_queue(hctx);
1553 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1555 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1557 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1559 blk_mq_run_hw_queue(hctx, false);
1561 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1563 void blk_mq_start_hw_queues(struct request_queue *q)
1565 struct blk_mq_hw_ctx *hctx;
1568 queue_for_each_hw_ctx(q, hctx, i)
1569 blk_mq_start_hw_queue(hctx);
1571 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1573 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1575 if (!blk_mq_hctx_stopped(hctx))
1578 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1579 blk_mq_run_hw_queue(hctx, async);
1581 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1583 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1585 struct blk_mq_hw_ctx *hctx;
1588 queue_for_each_hw_ctx(q, hctx, i)
1589 blk_mq_start_stopped_hw_queue(hctx, async);
1591 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1593 static void blk_mq_run_work_fn(struct work_struct *work)
1595 struct blk_mq_hw_ctx *hctx;
1597 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1600 * If we are stopped, don't run the queue.
1602 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1605 __blk_mq_run_hw_queue(hctx);
1608 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1612 struct blk_mq_ctx *ctx = rq->mq_ctx;
1613 enum hctx_type type = hctx->type;
1615 lockdep_assert_held(&ctx->lock);
1617 trace_block_rq_insert(hctx->queue, rq);
1620 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1622 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1625 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1628 struct blk_mq_ctx *ctx = rq->mq_ctx;
1630 lockdep_assert_held(&ctx->lock);
1632 __blk_mq_insert_req_list(hctx, rq, at_head);
1633 blk_mq_hctx_mark_pending(hctx, ctx);
1637 * Should only be used carefully, when the caller knows we want to
1638 * bypass a potential IO scheduler on the target device.
1640 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1642 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1644 spin_lock(&hctx->lock);
1645 list_add_tail(&rq->queuelist, &hctx->dispatch);
1646 spin_unlock(&hctx->lock);
1649 blk_mq_run_hw_queue(hctx, false);
1652 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1653 struct list_head *list)
1657 enum hctx_type type = hctx->type;
1660 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1663 list_for_each_entry(rq, list, queuelist) {
1664 BUG_ON(rq->mq_ctx != ctx);
1665 trace_block_rq_insert(hctx->queue, rq);
1668 spin_lock(&ctx->lock);
1669 list_splice_tail_init(list, &ctx->rq_lists[type]);
1670 blk_mq_hctx_mark_pending(hctx, ctx);
1671 spin_unlock(&ctx->lock);
1674 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1676 struct request *rqa = container_of(a, struct request, queuelist);
1677 struct request *rqb = container_of(b, struct request, queuelist);
1679 if (rqa->mq_ctx < rqb->mq_ctx)
1681 else if (rqa->mq_ctx > rqb->mq_ctx)
1683 else if (rqa->mq_hctx < rqb->mq_hctx)
1685 else if (rqa->mq_hctx > rqb->mq_hctx)
1688 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1691 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1693 struct blk_mq_hw_ctx *this_hctx;
1694 struct blk_mq_ctx *this_ctx;
1695 struct request_queue *this_q;
1701 list_splice_init(&plug->mq_list, &list);
1704 if (plug->rq_count > 2 && plug->multiple_queues)
1705 list_sort(NULL, &list, plug_rq_cmp);
1712 while (!list_empty(&list)) {
1713 rq = list_entry_rq(list.next);
1714 list_del_init(&rq->queuelist);
1716 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1718 trace_block_unplug(this_q, depth, !from_schedule);
1719 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1725 this_ctx = rq->mq_ctx;
1726 this_hctx = rq->mq_hctx;
1731 list_add_tail(&rq->queuelist, &rq_list);
1735 * If 'this_hctx' is set, we know we have entries to complete
1736 * on 'rq_list'. Do those.
1739 trace_block_unplug(this_q, depth, !from_schedule);
1740 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1745 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1747 blk_init_request_from_bio(rq, bio);
1749 blk_account_io_start(rq, true);
1752 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1755 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1757 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1760 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1762 blk_qc_t *cookie, bool last)
1764 struct request_queue *q = rq->q;
1765 struct blk_mq_queue_data bd = {
1769 blk_qc_t new_cookie;
1772 new_cookie = request_to_qc_t(hctx, rq);
1775 * For OK queue, we are done. For error, caller may kill it.
1776 * Any other error (busy), just add it to our list as we
1777 * previously would have done.
1779 ret = q->mq_ops->queue_rq(hctx, &bd);
1782 blk_mq_update_dispatch_busy(hctx, false);
1783 *cookie = new_cookie;
1785 case BLK_STS_RESOURCE:
1786 case BLK_STS_DEV_RESOURCE:
1787 blk_mq_update_dispatch_busy(hctx, true);
1788 __blk_mq_requeue_request(rq);
1791 blk_mq_update_dispatch_busy(hctx, false);
1792 *cookie = BLK_QC_T_NONE;
1799 blk_status_t blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1802 bool bypass, bool last)
1804 struct request_queue *q = rq->q;
1805 bool run_queue = true;
1806 blk_status_t ret = BLK_STS_RESOURCE;
1810 hctx_lock(hctx, &srcu_idx);
1812 * hctx_lock is needed before checking quiesced flag.
1814 * When queue is stopped or quiesced, ignore 'bypass', insert
1815 * and return BLK_STS_OK to caller, and avoid driver to try to
1818 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) {
1824 if (unlikely(q->elevator && !bypass))
1827 if (!blk_mq_get_dispatch_budget(hctx))
1830 if (!blk_mq_get_driver_tag(rq)) {
1831 blk_mq_put_dispatch_budget(hctx);
1836 * Always add a request that has been through
1837 *.queue_rq() to the hardware dispatch list.
1840 ret = __blk_mq_issue_directly(hctx, rq, cookie, last);
1842 hctx_unlock(hctx, srcu_idx);
1846 case BLK_STS_DEV_RESOURCE:
1847 case BLK_STS_RESOURCE:
1849 blk_mq_request_bypass_insert(rq, run_queue);
1851 * We have to return BLK_STS_OK for the DM
1852 * to avoid livelock. Otherwise, we return
1853 * the real result to indicate whether the
1854 * request is direct-issued successfully.
1856 ret = bypass ? BLK_STS_OK : ret;
1857 } else if (!bypass) {
1858 blk_mq_sched_insert_request(rq, false,
1864 blk_mq_end_request(rq, ret);
1871 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1872 struct list_head *list)
1875 blk_status_t ret = BLK_STS_OK;
1877 while (!list_empty(list)) {
1878 struct request *rq = list_first_entry(list, struct request,
1881 list_del_init(&rq->queuelist);
1882 if (ret == BLK_STS_OK)
1883 ret = blk_mq_try_issue_directly(hctx, rq, &unused,
1887 blk_mq_sched_insert_request(rq, false, true, false);
1891 * If we didn't flush the entire list, we could have told
1892 * the driver there was more coming, but that turned out to
1895 if (ret != BLK_STS_OK && hctx->queue->mq_ops->commit_rqs)
1896 hctx->queue->mq_ops->commit_rqs(hctx);
1899 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1901 list_add_tail(&rq->queuelist, &plug->mq_list);
1903 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1904 struct request *tmp;
1906 tmp = list_first_entry(&plug->mq_list, struct request,
1908 if (tmp->q != rq->q)
1909 plug->multiple_queues = true;
1913 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1915 const int is_sync = op_is_sync(bio->bi_opf);
1916 const int is_flush_fua = op_is_flush(bio->bi_opf);
1917 struct blk_mq_alloc_data data = { .flags = 0, .cmd_flags = bio->bi_opf };
1919 struct blk_plug *plug;
1920 struct request *same_queue_rq = NULL;
1923 blk_queue_bounce(q, &bio);
1925 blk_queue_split(q, &bio);
1927 if (!bio_integrity_prep(bio))
1928 return BLK_QC_T_NONE;
1930 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1931 blk_attempt_plug_merge(q, bio, &same_queue_rq))
1932 return BLK_QC_T_NONE;
1934 if (blk_mq_sched_bio_merge(q, bio))
1935 return BLK_QC_T_NONE;
1937 rq_qos_throttle(q, bio);
1939 rq = blk_mq_get_request(q, bio, &data);
1940 if (unlikely(!rq)) {
1941 rq_qos_cleanup(q, bio);
1942 if (bio->bi_opf & REQ_NOWAIT)
1943 bio_wouldblock_error(bio);
1944 return BLK_QC_T_NONE;
1947 trace_block_getrq(q, bio, bio->bi_opf);
1949 rq_qos_track(q, rq, bio);
1951 cookie = request_to_qc_t(data.hctx, rq);
1953 plug = current->plug;
1954 if (unlikely(is_flush_fua)) {
1955 blk_mq_put_ctx(data.ctx);
1956 blk_mq_bio_to_request(rq, bio);
1958 /* bypass scheduler for flush rq */
1959 blk_insert_flush(rq);
1960 blk_mq_run_hw_queue(data.hctx, true);
1961 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1963 * Use plugging if we have a ->commit_rqs() hook as well, as
1964 * we know the driver uses bd->last in a smart fashion.
1966 unsigned int request_count = plug->rq_count;
1967 struct request *last = NULL;
1969 blk_mq_put_ctx(data.ctx);
1970 blk_mq_bio_to_request(rq, bio);
1973 trace_block_plug(q);
1975 last = list_entry_rq(plug->mq_list.prev);
1977 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1978 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1979 blk_flush_plug_list(plug, false);
1980 trace_block_plug(q);
1983 blk_add_rq_to_plug(plug, rq);
1984 } else if (plug && !blk_queue_nomerges(q)) {
1985 blk_mq_bio_to_request(rq, bio);
1988 * We do limited plugging. If the bio can be merged, do that.
1989 * Otherwise the existing request in the plug list will be
1990 * issued. So the plug list will have one request at most
1991 * The plug list might get flushed before this. If that happens,
1992 * the plug list is empty, and same_queue_rq is invalid.
1994 if (list_empty(&plug->mq_list))
1995 same_queue_rq = NULL;
1996 if (same_queue_rq) {
1997 list_del_init(&same_queue_rq->queuelist);
2000 blk_add_rq_to_plug(plug, rq);
2002 blk_mq_put_ctx(data.ctx);
2004 if (same_queue_rq) {
2005 data.hctx = same_queue_rq->mq_hctx;
2006 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2007 &cookie, false, true);
2009 } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2010 !data.hctx->dispatch_busy)) {
2011 blk_mq_put_ctx(data.ctx);
2012 blk_mq_bio_to_request(rq, bio);
2013 blk_mq_try_issue_directly(data.hctx, rq, &cookie, false, true);
2015 blk_mq_put_ctx(data.ctx);
2016 blk_mq_bio_to_request(rq, bio);
2017 blk_mq_sched_insert_request(rq, false, true, true);
2023 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2024 unsigned int hctx_idx)
2028 if (tags->rqs && set->ops->exit_request) {
2031 for (i = 0; i < tags->nr_tags; i++) {
2032 struct request *rq = tags->static_rqs[i];
2036 set->ops->exit_request(set, rq, hctx_idx);
2037 tags->static_rqs[i] = NULL;
2041 while (!list_empty(&tags->page_list)) {
2042 page = list_first_entry(&tags->page_list, struct page, lru);
2043 list_del_init(&page->lru);
2045 * Remove kmemleak object previously allocated in
2046 * blk_mq_init_rq_map().
2048 kmemleak_free(page_address(page));
2049 __free_pages(page, page->private);
2053 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2057 kfree(tags->static_rqs);
2058 tags->static_rqs = NULL;
2060 blk_mq_free_tags(tags);
2063 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2064 unsigned int hctx_idx,
2065 unsigned int nr_tags,
2066 unsigned int reserved_tags)
2068 struct blk_mq_tags *tags;
2071 node = blk_mq_hw_queue_to_node(&set->map[0], hctx_idx);
2072 if (node == NUMA_NO_NODE)
2073 node = set->numa_node;
2075 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2076 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2080 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2081 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2084 blk_mq_free_tags(tags);
2088 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2089 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2091 if (!tags->static_rqs) {
2093 blk_mq_free_tags(tags);
2100 static size_t order_to_size(unsigned int order)
2102 return (size_t)PAGE_SIZE << order;
2105 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2106 unsigned int hctx_idx, int node)
2110 if (set->ops->init_request) {
2111 ret = set->ops->init_request(set, rq, hctx_idx, node);
2116 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2120 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2121 unsigned int hctx_idx, unsigned int depth)
2123 unsigned int i, j, entries_per_page, max_order = 4;
2124 size_t rq_size, left;
2127 node = blk_mq_hw_queue_to_node(&set->map[0], hctx_idx);
2128 if (node == NUMA_NO_NODE)
2129 node = set->numa_node;
2131 INIT_LIST_HEAD(&tags->page_list);
2134 * rq_size is the size of the request plus driver payload, rounded
2135 * to the cacheline size
2137 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2139 left = rq_size * depth;
2141 for (i = 0; i < depth; ) {
2142 int this_order = max_order;
2147 while (this_order && left < order_to_size(this_order - 1))
2151 page = alloc_pages_node(node,
2152 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2158 if (order_to_size(this_order) < rq_size)
2165 page->private = this_order;
2166 list_add_tail(&page->lru, &tags->page_list);
2168 p = page_address(page);
2170 * Allow kmemleak to scan these pages as they contain pointers
2171 * to additional allocations like via ops->init_request().
2173 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2174 entries_per_page = order_to_size(this_order) / rq_size;
2175 to_do = min(entries_per_page, depth - i);
2176 left -= to_do * rq_size;
2177 for (j = 0; j < to_do; j++) {
2178 struct request *rq = p;
2180 tags->static_rqs[i] = rq;
2181 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2182 tags->static_rqs[i] = NULL;
2193 blk_mq_free_rqs(set, tags, hctx_idx);
2198 * 'cpu' is going away. splice any existing rq_list entries from this
2199 * software queue to the hw queue dispatch list, and ensure that it
2202 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2204 struct blk_mq_hw_ctx *hctx;
2205 struct blk_mq_ctx *ctx;
2207 enum hctx_type type;
2209 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2210 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2213 spin_lock(&ctx->lock);
2214 if (!list_empty(&ctx->rq_lists[type])) {
2215 list_splice_init(&ctx->rq_lists[type], &tmp);
2216 blk_mq_hctx_clear_pending(hctx, ctx);
2218 spin_unlock(&ctx->lock);
2220 if (list_empty(&tmp))
2223 spin_lock(&hctx->lock);
2224 list_splice_tail_init(&tmp, &hctx->dispatch);
2225 spin_unlock(&hctx->lock);
2227 blk_mq_run_hw_queue(hctx, true);
2231 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2233 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2237 /* hctx->ctxs will be freed in queue's release handler */
2238 static void blk_mq_exit_hctx(struct request_queue *q,
2239 struct blk_mq_tag_set *set,
2240 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2242 if (blk_mq_hw_queue_mapped(hctx))
2243 blk_mq_tag_idle(hctx);
2245 if (set->ops->exit_request)
2246 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2248 if (set->ops->exit_hctx)
2249 set->ops->exit_hctx(hctx, hctx_idx);
2251 if (hctx->flags & BLK_MQ_F_BLOCKING)
2252 cleanup_srcu_struct(hctx->srcu);
2254 blk_mq_remove_cpuhp(hctx);
2255 blk_free_flush_queue(hctx->fq);
2256 sbitmap_free(&hctx->ctx_map);
2259 static void blk_mq_exit_hw_queues(struct request_queue *q,
2260 struct blk_mq_tag_set *set, int nr_queue)
2262 struct blk_mq_hw_ctx *hctx;
2265 queue_for_each_hw_ctx(q, hctx, i) {
2268 blk_mq_debugfs_unregister_hctx(hctx);
2269 blk_mq_exit_hctx(q, set, hctx, i);
2273 static int blk_mq_init_hctx(struct request_queue *q,
2274 struct blk_mq_tag_set *set,
2275 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2279 node = hctx->numa_node;
2280 if (node == NUMA_NO_NODE)
2281 node = hctx->numa_node = set->numa_node;
2283 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2284 spin_lock_init(&hctx->lock);
2285 INIT_LIST_HEAD(&hctx->dispatch);
2287 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2289 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2291 hctx->tags = set->tags[hctx_idx];
2294 * Allocate space for all possible cpus to avoid allocation at
2297 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2298 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node);
2300 goto unregister_cpu_notifier;
2302 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2303 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node))
2308 spin_lock_init(&hctx->dispatch_wait_lock);
2309 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2310 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2312 if (set->ops->init_hctx &&
2313 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2316 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2317 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
2321 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2324 if (hctx->flags & BLK_MQ_F_BLOCKING)
2325 init_srcu_struct(hctx->srcu);
2332 if (set->ops->exit_hctx)
2333 set->ops->exit_hctx(hctx, hctx_idx);
2335 sbitmap_free(&hctx->ctx_map);
2338 unregister_cpu_notifier:
2339 blk_mq_remove_cpuhp(hctx);
2343 static void blk_mq_init_cpu_queues(struct request_queue *q,
2344 unsigned int nr_hw_queues)
2346 struct blk_mq_tag_set *set = q->tag_set;
2349 for_each_possible_cpu(i) {
2350 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2351 struct blk_mq_hw_ctx *hctx;
2355 spin_lock_init(&__ctx->lock);
2356 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2357 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2362 * Set local node, IFF we have more than one hw queue. If
2363 * not, we remain on the home node of the device
2365 for (j = 0; j < set->nr_maps; j++) {
2366 hctx = blk_mq_map_queue_type(q, j, i);
2367 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2368 hctx->numa_node = local_memory_node(cpu_to_node(i));
2373 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2377 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2378 set->queue_depth, set->reserved_tags);
2379 if (!set->tags[hctx_idx])
2382 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2387 blk_mq_free_rq_map(set->tags[hctx_idx]);
2388 set->tags[hctx_idx] = NULL;
2392 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2393 unsigned int hctx_idx)
2395 if (set->tags && set->tags[hctx_idx]) {
2396 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2397 blk_mq_free_rq_map(set->tags[hctx_idx]);
2398 set->tags[hctx_idx] = NULL;
2402 static void blk_mq_map_swqueue(struct request_queue *q)
2404 unsigned int i, j, hctx_idx;
2405 struct blk_mq_hw_ctx *hctx;
2406 struct blk_mq_ctx *ctx;
2407 struct blk_mq_tag_set *set = q->tag_set;
2410 * Avoid others reading imcomplete hctx->cpumask through sysfs
2412 mutex_lock(&q->sysfs_lock);
2414 queue_for_each_hw_ctx(q, hctx, i) {
2415 cpumask_clear(hctx->cpumask);
2417 hctx->dispatch_from = NULL;
2421 * Map software to hardware queues.
2423 * If the cpu isn't present, the cpu is mapped to first hctx.
2425 for_each_possible_cpu(i) {
2426 hctx_idx = set->map[0].mq_map[i];
2427 /* unmapped hw queue can be remapped after CPU topo changed */
2428 if (!set->tags[hctx_idx] &&
2429 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2431 * If tags initialization fail for some hctx,
2432 * that hctx won't be brought online. In this
2433 * case, remap the current ctx to hctx[0] which
2434 * is guaranteed to always have tags allocated
2436 set->map[0].mq_map[i] = 0;
2439 ctx = per_cpu_ptr(q->queue_ctx, i);
2440 for (j = 0; j < set->nr_maps; j++) {
2441 if (!set->map[j].nr_queues)
2444 hctx = blk_mq_map_queue_type(q, j, i);
2447 * If the CPU is already set in the mask, then we've
2448 * mapped this one already. This can happen if
2449 * devices share queues across queue maps.
2451 if (cpumask_test_cpu(i, hctx->cpumask))
2454 cpumask_set_cpu(i, hctx->cpumask);
2456 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2457 hctx->ctxs[hctx->nr_ctx++] = ctx;
2460 * If the nr_ctx type overflows, we have exceeded the
2461 * amount of sw queues we can support.
2463 BUG_ON(!hctx->nr_ctx);
2467 mutex_unlock(&q->sysfs_lock);
2469 queue_for_each_hw_ctx(q, hctx, i) {
2471 * If no software queues are mapped to this hardware queue,
2472 * disable it and free the request entries.
2474 if (!hctx->nr_ctx) {
2475 /* Never unmap queue 0. We need it as a
2476 * fallback in case of a new remap fails
2479 if (i && set->tags[i])
2480 blk_mq_free_map_and_requests(set, i);
2486 hctx->tags = set->tags[i];
2487 WARN_ON(!hctx->tags);
2490 * Set the map size to the number of mapped software queues.
2491 * This is more accurate and more efficient than looping
2492 * over all possibly mapped software queues.
2494 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2497 * Initialize batch roundrobin counts
2499 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2500 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2505 * Caller needs to ensure that we're either frozen/quiesced, or that
2506 * the queue isn't live yet.
2508 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2510 struct blk_mq_hw_ctx *hctx;
2513 queue_for_each_hw_ctx(q, hctx, i) {
2515 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2517 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2521 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2524 struct request_queue *q;
2526 lockdep_assert_held(&set->tag_list_lock);
2528 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2529 blk_mq_freeze_queue(q);
2530 queue_set_hctx_shared(q, shared);
2531 blk_mq_unfreeze_queue(q);
2535 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2537 struct blk_mq_tag_set *set = q->tag_set;
2539 mutex_lock(&set->tag_list_lock);
2540 list_del_rcu(&q->tag_set_list);
2541 if (list_is_singular(&set->tag_list)) {
2542 /* just transitioned to unshared */
2543 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2544 /* update existing queue */
2545 blk_mq_update_tag_set_depth(set, false);
2547 mutex_unlock(&set->tag_list_lock);
2548 INIT_LIST_HEAD(&q->tag_set_list);
2551 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2552 struct request_queue *q)
2554 mutex_lock(&set->tag_list_lock);
2557 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2559 if (!list_empty(&set->tag_list) &&
2560 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2561 set->flags |= BLK_MQ_F_TAG_SHARED;
2562 /* update existing queue */
2563 blk_mq_update_tag_set_depth(set, true);
2565 if (set->flags & BLK_MQ_F_TAG_SHARED)
2566 queue_set_hctx_shared(q, true);
2567 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2569 mutex_unlock(&set->tag_list_lock);
2572 /* All allocations will be freed in release handler of q->mq_kobj */
2573 static int blk_mq_alloc_ctxs(struct request_queue *q)
2575 struct blk_mq_ctxs *ctxs;
2578 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2582 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2583 if (!ctxs->queue_ctx)
2586 for_each_possible_cpu(cpu) {
2587 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2591 q->mq_kobj = &ctxs->kobj;
2592 q->queue_ctx = ctxs->queue_ctx;
2601 * It is the actual release handler for mq, but we do it from
2602 * request queue's release handler for avoiding use-after-free
2603 * and headache because q->mq_kobj shouldn't have been introduced,
2604 * but we can't group ctx/kctx kobj without it.
2606 void blk_mq_release(struct request_queue *q)
2608 struct blk_mq_hw_ctx *hctx;
2611 /* hctx kobj stays in hctx */
2612 queue_for_each_hw_ctx(q, hctx, i) {
2615 kobject_put(&hctx->kobj);
2618 kfree(q->queue_hw_ctx);
2621 * release .mq_kobj and sw queue's kobject now because
2622 * both share lifetime with request queue.
2624 blk_mq_sysfs_deinit(q);
2627 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2629 struct request_queue *uninit_q, *q;
2631 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2633 return ERR_PTR(-ENOMEM);
2635 q = blk_mq_init_allocated_queue(set, uninit_q);
2637 blk_cleanup_queue(uninit_q);
2641 EXPORT_SYMBOL(blk_mq_init_queue);
2644 * Helper for setting up a queue with mq ops, given queue depth, and
2645 * the passed in mq ops flags.
2647 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2648 const struct blk_mq_ops *ops,
2649 unsigned int queue_depth,
2650 unsigned int set_flags)
2652 struct request_queue *q;
2655 memset(set, 0, sizeof(*set));
2657 set->nr_hw_queues = 1;
2659 set->queue_depth = queue_depth;
2660 set->numa_node = NUMA_NO_NODE;
2661 set->flags = set_flags;
2663 ret = blk_mq_alloc_tag_set(set);
2665 return ERR_PTR(ret);
2667 q = blk_mq_init_queue(set);
2669 blk_mq_free_tag_set(set);
2675 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2677 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2679 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2681 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2682 __alignof__(struct blk_mq_hw_ctx)) !=
2683 sizeof(struct blk_mq_hw_ctx));
2685 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2686 hw_ctx_size += sizeof(struct srcu_struct);
2691 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2692 struct blk_mq_tag_set *set, struct request_queue *q,
2693 int hctx_idx, int node)
2695 struct blk_mq_hw_ctx *hctx;
2697 hctx = kzalloc_node(blk_mq_hw_ctx_size(set),
2698 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2703 if (!zalloc_cpumask_var_node(&hctx->cpumask,
2704 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2710 atomic_set(&hctx->nr_active, 0);
2711 hctx->numa_node = node;
2712 hctx->queue_num = hctx_idx;
2714 if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) {
2715 free_cpumask_var(hctx->cpumask);
2719 blk_mq_hctx_kobj_init(hctx);
2724 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2725 struct request_queue *q)
2728 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2730 /* protect against switching io scheduler */
2731 mutex_lock(&q->sysfs_lock);
2732 for (i = 0; i < set->nr_hw_queues; i++) {
2734 struct blk_mq_hw_ctx *hctx;
2736 node = blk_mq_hw_queue_to_node(&set->map[0], i);
2738 * If the hw queue has been mapped to another numa node,
2739 * we need to realloc the hctx. If allocation fails, fallback
2740 * to use the previous one.
2742 if (hctxs[i] && (hctxs[i]->numa_node == node))
2745 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2748 blk_mq_exit_hctx(q, set, hctxs[i], i);
2749 kobject_put(&hctxs[i]->kobj);
2754 pr_warn("Allocate new hctx on node %d fails,\
2755 fallback to previous one on node %d\n",
2756 node, hctxs[i]->numa_node);
2762 * Increasing nr_hw_queues fails. Free the newly allocated
2763 * hctxs and keep the previous q->nr_hw_queues.
2765 if (i != set->nr_hw_queues) {
2766 j = q->nr_hw_queues;
2770 end = q->nr_hw_queues;
2771 q->nr_hw_queues = set->nr_hw_queues;
2774 for (; j < end; j++) {
2775 struct blk_mq_hw_ctx *hctx = hctxs[j];
2779 blk_mq_free_map_and_requests(set, j);
2780 blk_mq_exit_hctx(q, set, hctx, j);
2781 kobject_put(&hctx->kobj);
2786 mutex_unlock(&q->sysfs_lock);
2790 * Maximum number of hardware queues we support. For single sets, we'll never
2791 * have more than the CPUs (software queues). For multiple sets, the tag_set
2792 * user may have set ->nr_hw_queues larger.
2794 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2796 if (set->nr_maps == 1)
2799 return max(set->nr_hw_queues, nr_cpu_ids);
2802 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2803 struct request_queue *q)
2805 /* mark the queue as mq asap */
2806 q->mq_ops = set->ops;
2808 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2809 blk_mq_poll_stats_bkt,
2810 BLK_MQ_POLL_STATS_BKTS, q);
2814 if (blk_mq_alloc_ctxs(q))
2817 /* init q->mq_kobj and sw queues' kobjects */
2818 blk_mq_sysfs_init(q);
2820 q->nr_queues = nr_hw_queues(set);
2821 q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2822 GFP_KERNEL, set->numa_node);
2823 if (!q->queue_hw_ctx)
2826 blk_mq_realloc_hw_ctxs(set, q);
2827 if (!q->nr_hw_queues)
2830 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2831 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2835 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2836 if (set->nr_maps > HCTX_TYPE_POLL &&
2837 set->map[HCTX_TYPE_POLL].nr_queues)
2838 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2840 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2841 blk_queue_flag_set(QUEUE_FLAG_NO_SG_MERGE, q);
2843 q->sg_reserved_size = INT_MAX;
2845 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2846 INIT_LIST_HEAD(&q->requeue_list);
2847 spin_lock_init(&q->requeue_lock);
2849 blk_queue_make_request(q, blk_mq_make_request);
2852 * Do this after blk_queue_make_request() overrides it...
2854 q->nr_requests = set->queue_depth;
2857 * Default to classic polling
2861 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2862 blk_mq_add_queue_tag_set(set, q);
2863 blk_mq_map_swqueue(q);
2865 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2868 ret = elevator_init_mq(q);
2870 return ERR_PTR(ret);
2876 kfree(q->queue_hw_ctx);
2878 blk_mq_sysfs_deinit(q);
2881 return ERR_PTR(-ENOMEM);
2883 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2885 void blk_mq_free_queue(struct request_queue *q)
2887 struct blk_mq_tag_set *set = q->tag_set;
2889 blk_mq_del_queue_tag_set(q);
2890 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2893 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2897 for (i = 0; i < set->nr_hw_queues; i++)
2898 if (!__blk_mq_alloc_rq_map(set, i))
2905 blk_mq_free_rq_map(set->tags[i]);
2911 * Allocate the request maps associated with this tag_set. Note that this
2912 * may reduce the depth asked for, if memory is tight. set->queue_depth
2913 * will be updated to reflect the allocated depth.
2915 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2920 depth = set->queue_depth;
2922 err = __blk_mq_alloc_rq_maps(set);
2926 set->queue_depth >>= 1;
2927 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2931 } while (set->queue_depth);
2933 if (!set->queue_depth || err) {
2934 pr_err("blk-mq: failed to allocate request map\n");
2938 if (depth != set->queue_depth)
2939 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2940 depth, set->queue_depth);
2945 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2947 if (set->ops->map_queues && !is_kdump_kernel()) {
2951 * transport .map_queues is usually done in the following
2954 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2955 * mask = get_cpu_mask(queue)
2956 * for_each_cpu(cpu, mask)
2957 * set->map[x].mq_map[cpu] = queue;
2960 * When we need to remap, the table has to be cleared for
2961 * killing stale mapping since one CPU may not be mapped
2964 for (i = 0; i < set->nr_maps; i++)
2965 blk_mq_clear_mq_map(&set->map[i]);
2967 return set->ops->map_queues(set);
2969 BUG_ON(set->nr_maps > 1);
2970 return blk_mq_map_queues(&set->map[0]);
2975 * Alloc a tag set to be associated with one or more request queues.
2976 * May fail with EINVAL for various error conditions. May adjust the
2977 * requested depth down, if it's too large. In that case, the set
2978 * value will be stored in set->queue_depth.
2980 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2984 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2986 if (!set->nr_hw_queues)
2988 if (!set->queue_depth)
2990 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2993 if (!set->ops->queue_rq)
2996 if (!set->ops->get_budget ^ !set->ops->put_budget)
2999 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3000 pr_info("blk-mq: reduced tag depth to %u\n",
3002 set->queue_depth = BLK_MQ_MAX_DEPTH;
3007 else if (set->nr_maps > HCTX_MAX_TYPES)
3011 * If a crashdump is active, then we are potentially in a very
3012 * memory constrained environment. Limit us to 1 queue and
3013 * 64 tags to prevent using too much memory.
3015 if (is_kdump_kernel()) {
3016 set->nr_hw_queues = 1;
3018 set->queue_depth = min(64U, set->queue_depth);
3021 * There is no use for more h/w queues than cpus if we just have
3024 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3025 set->nr_hw_queues = nr_cpu_ids;
3027 set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3028 GFP_KERNEL, set->numa_node);
3033 for (i = 0; i < set->nr_maps; i++) {
3034 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3035 sizeof(set->map[i].mq_map[0]),
3036 GFP_KERNEL, set->numa_node);
3037 if (!set->map[i].mq_map)
3038 goto out_free_mq_map;
3039 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3042 ret = blk_mq_update_queue_map(set);
3044 goto out_free_mq_map;
3046 ret = blk_mq_alloc_rq_maps(set);
3048 goto out_free_mq_map;
3050 mutex_init(&set->tag_list_lock);
3051 INIT_LIST_HEAD(&set->tag_list);
3056 for (i = 0; i < set->nr_maps; i++) {
3057 kfree(set->map[i].mq_map);
3058 set->map[i].mq_map = NULL;
3064 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3066 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3070 for (i = 0; i < nr_hw_queues(set); i++)
3071 blk_mq_free_map_and_requests(set, i);
3073 for (j = 0; j < set->nr_maps; j++) {
3074 kfree(set->map[j].mq_map);
3075 set->map[j].mq_map = NULL;
3081 EXPORT_SYMBOL(blk_mq_free_tag_set);
3083 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3085 struct blk_mq_tag_set *set = q->tag_set;
3086 struct blk_mq_hw_ctx *hctx;
3092 blk_mq_freeze_queue(q);
3093 blk_mq_quiesce_queue(q);
3096 queue_for_each_hw_ctx(q, hctx, i) {
3100 * If we're using an MQ scheduler, just update the scheduler
3101 * queue depth. This is similar to what the old code would do.
3103 if (!hctx->sched_tags) {
3104 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3107 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3115 q->nr_requests = nr;
3117 blk_mq_unquiesce_queue(q);
3118 blk_mq_unfreeze_queue(q);
3124 * request_queue and elevator_type pair.
3125 * It is just used by __blk_mq_update_nr_hw_queues to cache
3126 * the elevator_type associated with a request_queue.
3128 struct blk_mq_qe_pair {
3129 struct list_head node;
3130 struct request_queue *q;
3131 struct elevator_type *type;
3135 * Cache the elevator_type in qe pair list and switch the
3136 * io scheduler to 'none'
3138 static bool blk_mq_elv_switch_none(struct list_head *head,
3139 struct request_queue *q)
3141 struct blk_mq_qe_pair *qe;
3146 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3150 INIT_LIST_HEAD(&qe->node);
3152 qe->type = q->elevator->type;
3153 list_add(&qe->node, head);
3155 mutex_lock(&q->sysfs_lock);
3157 * After elevator_switch_mq, the previous elevator_queue will be
3158 * released by elevator_release. The reference of the io scheduler
3159 * module get by elevator_get will also be put. So we need to get
3160 * a reference of the io scheduler module here to prevent it to be
3163 __module_get(qe->type->elevator_owner);
3164 elevator_switch_mq(q, NULL);
3165 mutex_unlock(&q->sysfs_lock);
3170 static void blk_mq_elv_switch_back(struct list_head *head,
3171 struct request_queue *q)
3173 struct blk_mq_qe_pair *qe;
3174 struct elevator_type *t = NULL;
3176 list_for_each_entry(qe, head, node)
3185 list_del(&qe->node);
3188 mutex_lock(&q->sysfs_lock);
3189 elevator_switch_mq(q, t);
3190 mutex_unlock(&q->sysfs_lock);
3193 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3196 struct request_queue *q;
3198 int prev_nr_hw_queues;
3200 lockdep_assert_held(&set->tag_list_lock);
3202 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3203 nr_hw_queues = nr_cpu_ids;
3204 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3207 list_for_each_entry(q, &set->tag_list, tag_set_list)
3208 blk_mq_freeze_queue(q);
3210 * Sync with blk_mq_queue_tag_busy_iter.
3214 * Switch IO scheduler to 'none', cleaning up the data associated
3215 * with the previous scheduler. We will switch back once we are done
3216 * updating the new sw to hw queue mappings.
3218 list_for_each_entry(q, &set->tag_list, tag_set_list)
3219 if (!blk_mq_elv_switch_none(&head, q))
3222 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3223 blk_mq_debugfs_unregister_hctxs(q);
3224 blk_mq_sysfs_unregister(q);
3227 prev_nr_hw_queues = set->nr_hw_queues;
3228 set->nr_hw_queues = nr_hw_queues;
3229 blk_mq_update_queue_map(set);
3231 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3232 blk_mq_realloc_hw_ctxs(set, q);
3233 if (q->nr_hw_queues != set->nr_hw_queues) {
3234 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3235 nr_hw_queues, prev_nr_hw_queues);
3236 set->nr_hw_queues = prev_nr_hw_queues;
3237 blk_mq_map_queues(&set->map[0]);
3240 blk_mq_map_swqueue(q);
3243 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3244 blk_mq_sysfs_register(q);
3245 blk_mq_debugfs_register_hctxs(q);
3249 list_for_each_entry(q, &set->tag_list, tag_set_list)
3250 blk_mq_elv_switch_back(&head, q);
3252 list_for_each_entry(q, &set->tag_list, tag_set_list)
3253 blk_mq_unfreeze_queue(q);
3256 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3258 mutex_lock(&set->tag_list_lock);
3259 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3260 mutex_unlock(&set->tag_list_lock);
3262 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3264 /* Enable polling stats and return whether they were already enabled. */
3265 static bool blk_poll_stats_enable(struct request_queue *q)
3267 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3268 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3270 blk_stat_add_callback(q, q->poll_cb);
3274 static void blk_mq_poll_stats_start(struct request_queue *q)
3277 * We don't arm the callback if polling stats are not enabled or the
3278 * callback is already active.
3280 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3281 blk_stat_is_active(q->poll_cb))
3284 blk_stat_activate_msecs(q->poll_cb, 100);
3287 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3289 struct request_queue *q = cb->data;
3292 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3293 if (cb->stat[bucket].nr_samples)
3294 q->poll_stat[bucket] = cb->stat[bucket];
3298 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3299 struct blk_mq_hw_ctx *hctx,
3302 unsigned long ret = 0;
3306 * If stats collection isn't on, don't sleep but turn it on for
3309 if (!blk_poll_stats_enable(q))
3313 * As an optimistic guess, use half of the mean service time
3314 * for this type of request. We can (and should) make this smarter.
3315 * For instance, if the completion latencies are tight, we can
3316 * get closer than just half the mean. This is especially
3317 * important on devices where the completion latencies are longer
3318 * than ~10 usec. We do use the stats for the relevant IO size
3319 * if available which does lead to better estimates.
3321 bucket = blk_mq_poll_stats_bkt(rq);
3325 if (q->poll_stat[bucket].nr_samples)
3326 ret = (q->poll_stat[bucket].mean + 1) / 2;
3331 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3332 struct blk_mq_hw_ctx *hctx,
3335 struct hrtimer_sleeper hs;
3336 enum hrtimer_mode mode;
3340 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3344 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3346 * 0: use half of prev avg
3347 * >0: use this specific value
3349 if (q->poll_nsec > 0)
3350 nsecs = q->poll_nsec;
3352 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3357 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3360 * This will be replaced with the stats tracking code, using
3361 * 'avg_completion_time / 2' as the pre-sleep target.
3365 mode = HRTIMER_MODE_REL;
3366 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3367 hrtimer_set_expires(&hs.timer, kt);
3369 hrtimer_init_sleeper(&hs, current);
3371 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3373 set_current_state(TASK_UNINTERRUPTIBLE);
3374 hrtimer_start_expires(&hs.timer, mode);
3377 hrtimer_cancel(&hs.timer);
3378 mode = HRTIMER_MODE_ABS;
3379 } while (hs.task && !signal_pending(current));
3381 __set_current_state(TASK_RUNNING);
3382 destroy_hrtimer_on_stack(&hs.timer);
3386 static bool blk_mq_poll_hybrid(struct request_queue *q,
3387 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3391 if (q->poll_nsec == -1)
3394 if (!blk_qc_t_is_internal(cookie))
3395 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3397 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3399 * With scheduling, if the request has completed, we'll
3400 * get a NULL return here, as we clear the sched tag when
3401 * that happens. The request still remains valid, like always,
3402 * so we should be safe with just the NULL check.
3408 return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3412 * blk_poll - poll for IO completions
3414 * @cookie: cookie passed back at IO submission time
3415 * @spin: whether to spin for completions
3418 * Poll for completions on the passed in queue. Returns number of
3419 * completed entries found. If @spin is true, then blk_poll will continue
3420 * looping until at least one completion is found, unless the task is
3421 * otherwise marked running (or we need to reschedule).
3423 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3425 struct blk_mq_hw_ctx *hctx;
3428 if (!blk_qc_t_valid(cookie) ||
3429 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3433 blk_flush_plug_list(current->plug, false);
3435 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3438 * If we sleep, have the caller restart the poll loop to reset
3439 * the state. Like for the other success return cases, the
3440 * caller is responsible for checking if the IO completed. If
3441 * the IO isn't complete, we'll get called again and will go
3442 * straight to the busy poll loop.
3444 if (blk_mq_poll_hybrid(q, hctx, cookie))
3447 hctx->poll_considered++;
3449 state = current->state;
3453 hctx->poll_invoked++;
3455 ret = q->mq_ops->poll(hctx);
3457 hctx->poll_success++;
3458 __set_current_state(TASK_RUNNING);
3462 if (signal_pending_state(state, current))
3463 __set_current_state(TASK_RUNNING);
3465 if (current->state == TASK_RUNNING)
3467 if (ret < 0 || !spin)
3470 } while (!need_resched());
3472 __set_current_state(TASK_RUNNING);
3475 EXPORT_SYMBOL_GPL(blk_poll);
3477 unsigned int blk_mq_rq_cpu(struct request *rq)
3479 return rq->mq_ctx->cpu;
3481 EXPORT_SYMBOL(blk_mq_rq_cpu);
3483 static int __init blk_mq_init(void)
3485 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3486 blk_mq_hctx_notify_dead);
3489 subsys_initcall(blk_mq_init);