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;
334 /* tag was already set */
336 WRITE_ONCE(rq->deadline, 0);
341 rq->end_io_data = NULL;
343 data->ctx->rq_dispatched[op_is_sync(op)]++;
344 refcount_set(&rq->ref, 1);
348 static struct request *blk_mq_get_request(struct request_queue *q,
350 struct blk_mq_alloc_data *data)
352 struct elevator_queue *e = q->elevator;
355 bool put_ctx_on_error = false;
357 blk_queue_enter_live(q);
359 if (likely(!data->ctx)) {
360 data->ctx = blk_mq_get_ctx(q);
361 put_ctx_on_error = true;
363 if (likely(!data->hctx))
364 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
366 if (data->cmd_flags & REQ_NOWAIT)
367 data->flags |= BLK_MQ_REQ_NOWAIT;
370 data->flags |= BLK_MQ_REQ_INTERNAL;
373 * Flush requests are special and go directly to the
374 * dispatch list. Don't include reserved tags in the
375 * limiting, as it isn't useful.
377 if (!op_is_flush(data->cmd_flags) &&
378 e->type->ops.limit_depth &&
379 !(data->flags & BLK_MQ_REQ_RESERVED))
380 e->type->ops.limit_depth(data->cmd_flags, data);
382 blk_mq_tag_busy(data->hctx);
385 tag = blk_mq_get_tag(data);
386 if (tag == BLK_MQ_TAG_FAIL) {
387 if (put_ctx_on_error) {
388 blk_mq_put_ctx(data->ctx);
395 rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags);
396 if (!op_is_flush(data->cmd_flags)) {
398 if (e && e->type->ops.prepare_request) {
399 if (e->type->icq_cache)
400 blk_mq_sched_assign_ioc(rq);
402 e->type->ops.prepare_request(rq, bio);
403 rq->rq_flags |= RQF_ELVPRIV;
406 data->hctx->queued++;
410 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
411 blk_mq_req_flags_t flags)
413 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
417 ret = blk_queue_enter(q, flags);
421 rq = blk_mq_get_request(q, NULL, &alloc_data);
425 return ERR_PTR(-EWOULDBLOCK);
427 blk_mq_put_ctx(alloc_data.ctx);
430 rq->__sector = (sector_t) -1;
431 rq->bio = rq->biotail = NULL;
434 EXPORT_SYMBOL(blk_mq_alloc_request);
436 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
437 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
439 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
445 * If the tag allocator sleeps we could get an allocation for a
446 * different hardware context. No need to complicate the low level
447 * allocator for this for the rare use case of a command tied to
450 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
451 return ERR_PTR(-EINVAL);
453 if (hctx_idx >= q->nr_hw_queues)
454 return ERR_PTR(-EIO);
456 ret = blk_queue_enter(q, flags);
461 * Check if the hardware context is actually mapped to anything.
462 * If not tell the caller that it should skip this queue.
464 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
465 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
467 return ERR_PTR(-EXDEV);
469 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
470 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
472 rq = blk_mq_get_request(q, NULL, &alloc_data);
476 return ERR_PTR(-EWOULDBLOCK);
480 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
482 static void __blk_mq_free_request(struct request *rq)
484 struct request_queue *q = rq->q;
485 struct blk_mq_ctx *ctx = rq->mq_ctx;
486 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
487 const int sched_tag = rq->internal_tag;
489 blk_pm_mark_last_busy(rq);
492 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
494 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
495 blk_mq_sched_restart(hctx);
499 void blk_mq_free_request(struct request *rq)
501 struct request_queue *q = rq->q;
502 struct elevator_queue *e = q->elevator;
503 struct blk_mq_ctx *ctx = rq->mq_ctx;
504 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
506 if (rq->rq_flags & RQF_ELVPRIV) {
507 if (e && e->type->ops.finish_request)
508 e->type->ops.finish_request(rq);
510 put_io_context(rq->elv.icq->ioc);
515 ctx->rq_completed[rq_is_sync(rq)]++;
516 if (rq->rq_flags & RQF_MQ_INFLIGHT)
517 atomic_dec(&hctx->nr_active);
519 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
520 laptop_io_completion(q->backing_dev_info);
524 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
525 if (refcount_dec_and_test(&rq->ref))
526 __blk_mq_free_request(rq);
528 EXPORT_SYMBOL_GPL(blk_mq_free_request);
530 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
534 if (blk_mq_need_time_stamp(rq))
535 now = ktime_get_ns();
537 if (rq->rq_flags & RQF_STATS) {
538 blk_mq_poll_stats_start(rq->q);
539 blk_stat_add(rq, now);
542 if (rq->internal_tag != -1)
543 blk_mq_sched_completed_request(rq, now);
545 blk_account_io_done(rq, now);
548 rq_qos_done(rq->q, rq);
549 rq->end_io(rq, error);
551 blk_mq_free_request(rq);
554 EXPORT_SYMBOL(__blk_mq_end_request);
556 void blk_mq_end_request(struct request *rq, blk_status_t error)
558 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
560 __blk_mq_end_request(rq, error);
562 EXPORT_SYMBOL(blk_mq_end_request);
564 static void __blk_mq_complete_request_remote(void *data)
566 struct request *rq = data;
567 struct request_queue *q = rq->q;
569 q->mq_ops->complete(rq);
572 static void __blk_mq_complete_request(struct request *rq)
574 struct blk_mq_ctx *ctx = rq->mq_ctx;
575 struct request_queue *q = rq->q;
579 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
581 * Most of single queue controllers, there is only one irq vector
582 * for handling IO completion, and the only irq's affinity is set
583 * as all possible CPUs. On most of ARCHs, this affinity means the
584 * irq is handled on one specific CPU.
586 * So complete IO reqeust in softirq context in case of single queue
587 * for not degrading IO performance by irqsoff latency.
589 if (q->nr_hw_queues == 1) {
590 __blk_complete_request(rq);
595 * For a polled request, always complete locallly, it's pointless
596 * to redirect the completion.
598 if ((rq->cmd_flags & REQ_HIPRI) ||
599 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
600 q->mq_ops->complete(rq);
605 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
606 shared = cpus_share_cache(cpu, ctx->cpu);
608 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
609 rq->csd.func = __blk_mq_complete_request_remote;
612 smp_call_function_single_async(ctx->cpu, &rq->csd);
614 q->mq_ops->complete(rq);
619 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
620 __releases(hctx->srcu)
622 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
625 srcu_read_unlock(hctx->srcu, srcu_idx);
628 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
629 __acquires(hctx->srcu)
631 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
632 /* shut up gcc false positive */
636 *srcu_idx = srcu_read_lock(hctx->srcu);
640 * blk_mq_complete_request - end I/O on a request
641 * @rq: the request being processed
644 * Ends all I/O on a request. It does not handle partial completions.
645 * The actual completion happens out-of-order, through a IPI handler.
647 bool blk_mq_complete_request(struct request *rq)
649 if (unlikely(blk_should_fake_timeout(rq->q)))
651 __blk_mq_complete_request(rq);
654 EXPORT_SYMBOL(blk_mq_complete_request);
656 int blk_mq_request_started(struct request *rq)
658 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
660 EXPORT_SYMBOL_GPL(blk_mq_request_started);
662 void blk_mq_start_request(struct request *rq)
664 struct request_queue *q = rq->q;
666 blk_mq_sched_started_request(rq);
668 trace_block_rq_issue(q, rq);
670 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
671 rq->io_start_time_ns = ktime_get_ns();
672 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
673 rq->throtl_size = blk_rq_sectors(rq);
675 rq->rq_flags |= RQF_STATS;
679 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
682 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
684 if (q->dma_drain_size && blk_rq_bytes(rq)) {
686 * Make sure space for the drain appears. We know we can do
687 * this because max_hw_segments has been adjusted to be one
688 * fewer than the device can handle.
690 rq->nr_phys_segments++;
693 EXPORT_SYMBOL(blk_mq_start_request);
695 static void __blk_mq_requeue_request(struct request *rq)
697 struct request_queue *q = rq->q;
699 blk_mq_put_driver_tag(rq);
701 trace_block_rq_requeue(q, rq);
702 rq_qos_requeue(q, rq);
704 if (blk_mq_request_started(rq)) {
705 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
706 rq->rq_flags &= ~RQF_TIMED_OUT;
707 if (q->dma_drain_size && blk_rq_bytes(rq))
708 rq->nr_phys_segments--;
712 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
714 __blk_mq_requeue_request(rq);
716 /* this request will be re-inserted to io scheduler queue */
717 blk_mq_sched_requeue_request(rq);
719 BUG_ON(!list_empty(&rq->queuelist));
720 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
722 EXPORT_SYMBOL(blk_mq_requeue_request);
724 static void blk_mq_requeue_work(struct work_struct *work)
726 struct request_queue *q =
727 container_of(work, struct request_queue, requeue_work.work);
729 struct request *rq, *next;
731 spin_lock_irq(&q->requeue_lock);
732 list_splice_init(&q->requeue_list, &rq_list);
733 spin_unlock_irq(&q->requeue_lock);
735 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
736 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
739 rq->rq_flags &= ~RQF_SOFTBARRIER;
740 list_del_init(&rq->queuelist);
742 * If RQF_DONTPREP, rq has contained some driver specific
743 * data, so insert it to hctx dispatch list to avoid any
746 if (rq->rq_flags & RQF_DONTPREP)
747 blk_mq_request_bypass_insert(rq, false);
749 blk_mq_sched_insert_request(rq, true, false, false);
752 while (!list_empty(&rq_list)) {
753 rq = list_entry(rq_list.next, struct request, queuelist);
754 list_del_init(&rq->queuelist);
755 blk_mq_sched_insert_request(rq, false, false, false);
758 blk_mq_run_hw_queues(q, false);
761 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
762 bool kick_requeue_list)
764 struct request_queue *q = rq->q;
768 * We abuse this flag that is otherwise used by the I/O scheduler to
769 * request head insertion from the workqueue.
771 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
773 spin_lock_irqsave(&q->requeue_lock, flags);
775 rq->rq_flags |= RQF_SOFTBARRIER;
776 list_add(&rq->queuelist, &q->requeue_list);
778 list_add_tail(&rq->queuelist, &q->requeue_list);
780 spin_unlock_irqrestore(&q->requeue_lock, flags);
782 if (kick_requeue_list)
783 blk_mq_kick_requeue_list(q);
785 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
787 void blk_mq_kick_requeue_list(struct request_queue *q)
789 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
791 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
793 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
796 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
797 msecs_to_jiffies(msecs));
799 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
801 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
803 if (tag < tags->nr_tags) {
804 prefetch(tags->rqs[tag]);
805 return tags->rqs[tag];
810 EXPORT_SYMBOL(blk_mq_tag_to_rq);
812 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
813 void *priv, bool reserved)
816 * If we find a request that is inflight and the queue matches,
817 * we know the queue is busy. Return false to stop the iteration.
819 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
829 bool blk_mq_queue_inflight(struct request_queue *q)
833 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
836 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
838 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
840 req->rq_flags |= RQF_TIMED_OUT;
841 if (req->q->mq_ops->timeout) {
842 enum blk_eh_timer_return ret;
844 ret = req->q->mq_ops->timeout(req, reserved);
845 if (ret == BLK_EH_DONE)
847 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
853 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
855 unsigned long deadline;
857 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
859 if (rq->rq_flags & RQF_TIMED_OUT)
862 deadline = READ_ONCE(rq->deadline);
863 if (time_after_eq(jiffies, deadline))
868 else if (time_after(*next, deadline))
873 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
874 struct request *rq, void *priv, bool reserved)
876 unsigned long *next = priv;
879 * Just do a quick check if it is expired before locking the request in
880 * so we're not unnecessarilly synchronizing across CPUs.
882 if (!blk_mq_req_expired(rq, next))
886 * We have reason to believe the request may be expired. Take a
887 * reference on the request to lock this request lifetime into its
888 * currently allocated context to prevent it from being reallocated in
889 * the event the completion by-passes this timeout handler.
891 * If the reference was already released, then the driver beat the
892 * timeout handler to posting a natural completion.
894 if (!refcount_inc_not_zero(&rq->ref))
898 * The request is now locked and cannot be reallocated underneath the
899 * timeout handler's processing. Re-verify this exact request is truly
900 * expired; if it is not expired, then the request was completed and
901 * reallocated as a new request.
903 if (blk_mq_req_expired(rq, next))
904 blk_mq_rq_timed_out(rq, reserved);
905 if (refcount_dec_and_test(&rq->ref))
906 __blk_mq_free_request(rq);
911 static void blk_mq_timeout_work(struct work_struct *work)
913 struct request_queue *q =
914 container_of(work, struct request_queue, timeout_work);
915 unsigned long next = 0;
916 struct blk_mq_hw_ctx *hctx;
919 /* A deadlock might occur if a request is stuck requiring a
920 * timeout at the same time a queue freeze is waiting
921 * completion, since the timeout code would not be able to
922 * acquire the queue reference here.
924 * That's why we don't use blk_queue_enter here; instead, we use
925 * percpu_ref_tryget directly, because we need to be able to
926 * obtain a reference even in the short window between the queue
927 * starting to freeze, by dropping the first reference in
928 * blk_freeze_queue_start, and the moment the last request is
929 * consumed, marked by the instant q_usage_counter reaches
932 if (!percpu_ref_tryget(&q->q_usage_counter))
935 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
938 mod_timer(&q->timeout, next);
941 * Request timeouts are handled as a forward rolling timer. If
942 * we end up here it means that no requests are pending and
943 * also that no request has been pending for a while. Mark
946 queue_for_each_hw_ctx(q, hctx, i) {
947 /* the hctx may be unmapped, so check it here */
948 if (blk_mq_hw_queue_mapped(hctx))
949 blk_mq_tag_idle(hctx);
955 struct flush_busy_ctx_data {
956 struct blk_mq_hw_ctx *hctx;
957 struct list_head *list;
960 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
962 struct flush_busy_ctx_data *flush_data = data;
963 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
964 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
965 enum hctx_type type = hctx->type;
967 spin_lock(&ctx->lock);
968 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
969 sbitmap_clear_bit(sb, bitnr);
970 spin_unlock(&ctx->lock);
975 * Process software queues that have been marked busy, splicing them
976 * to the for-dispatch
978 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
980 struct flush_busy_ctx_data data = {
985 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
987 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
989 struct dispatch_rq_data {
990 struct blk_mq_hw_ctx *hctx;
994 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
997 struct dispatch_rq_data *dispatch_data = data;
998 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
999 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1000 enum hctx_type type = hctx->type;
1002 spin_lock(&ctx->lock);
1003 if (!list_empty(&ctx->rq_lists[type])) {
1004 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1005 list_del_init(&dispatch_data->rq->queuelist);
1006 if (list_empty(&ctx->rq_lists[type]))
1007 sbitmap_clear_bit(sb, bitnr);
1009 spin_unlock(&ctx->lock);
1011 return !dispatch_data->rq;
1014 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1015 struct blk_mq_ctx *start)
1017 unsigned off = start ? start->index_hw[hctx->type] : 0;
1018 struct dispatch_rq_data data = {
1023 __sbitmap_for_each_set(&hctx->ctx_map, off,
1024 dispatch_rq_from_ctx, &data);
1029 static inline unsigned int queued_to_index(unsigned int queued)
1034 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1037 bool blk_mq_get_driver_tag(struct request *rq)
1039 struct blk_mq_alloc_data data = {
1041 .hctx = rq->mq_hctx,
1042 .flags = BLK_MQ_REQ_NOWAIT,
1043 .cmd_flags = rq->cmd_flags,
1050 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1051 data.flags |= BLK_MQ_REQ_RESERVED;
1053 shared = blk_mq_tag_busy(data.hctx);
1054 rq->tag = blk_mq_get_tag(&data);
1057 rq->rq_flags |= RQF_MQ_INFLIGHT;
1058 atomic_inc(&data.hctx->nr_active);
1060 data.hctx->tags->rqs[rq->tag] = rq;
1064 return rq->tag != -1;
1067 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1068 int flags, void *key)
1070 struct blk_mq_hw_ctx *hctx;
1072 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1074 spin_lock(&hctx->dispatch_wait_lock);
1075 list_del_init(&wait->entry);
1076 spin_unlock(&hctx->dispatch_wait_lock);
1078 blk_mq_run_hw_queue(hctx, true);
1083 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1084 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1085 * restart. For both cases, take care to check the condition again after
1086 * marking us as waiting.
1088 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1091 struct wait_queue_head *wq;
1092 wait_queue_entry_t *wait;
1095 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1096 blk_mq_sched_mark_restart_hctx(hctx);
1099 * It's possible that a tag was freed in the window between the
1100 * allocation failure and adding the hardware queue to the wait
1103 * Don't clear RESTART here, someone else could have set it.
1104 * At most this will cost an extra queue run.
1106 return blk_mq_get_driver_tag(rq);
1109 wait = &hctx->dispatch_wait;
1110 if (!list_empty_careful(&wait->entry))
1113 wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait;
1115 spin_lock_irq(&wq->lock);
1116 spin_lock(&hctx->dispatch_wait_lock);
1117 if (!list_empty(&wait->entry)) {
1118 spin_unlock(&hctx->dispatch_wait_lock);
1119 spin_unlock_irq(&wq->lock);
1123 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1124 __add_wait_queue(wq, wait);
1127 * It's possible that a tag was freed in the window between the
1128 * allocation failure and adding the hardware queue to the wait
1131 ret = blk_mq_get_driver_tag(rq);
1133 spin_unlock(&hctx->dispatch_wait_lock);
1134 spin_unlock_irq(&wq->lock);
1139 * We got a tag, remove ourselves from the wait queue to ensure
1140 * someone else gets the wakeup.
1142 list_del_init(&wait->entry);
1143 spin_unlock(&hctx->dispatch_wait_lock);
1144 spin_unlock_irq(&wq->lock);
1149 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1150 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1152 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1153 * - EWMA is one simple way to compute running average value
1154 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1155 * - take 4 as factor for avoiding to get too small(0) result, and this
1156 * factor doesn't matter because EWMA decreases exponentially
1158 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1162 if (hctx->queue->elevator)
1165 ewma = hctx->dispatch_busy;
1170 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1172 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1173 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1175 hctx->dispatch_busy = ewma;
1178 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1181 * Returns true if we did some work AND can potentially do more.
1183 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1186 struct blk_mq_hw_ctx *hctx;
1187 struct request *rq, *nxt;
1188 bool no_tag = false;
1190 blk_status_t ret = BLK_STS_OK;
1192 if (list_empty(list))
1195 WARN_ON(!list_is_singular(list) && got_budget);
1198 * Now process all the entries, sending them to the driver.
1200 errors = queued = 0;
1202 struct blk_mq_queue_data bd;
1204 rq = list_first_entry(list, struct request, queuelist);
1207 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1210 if (!blk_mq_get_driver_tag(rq)) {
1212 * The initial allocation attempt failed, so we need to
1213 * rerun the hardware queue when a tag is freed. The
1214 * waitqueue takes care of that. If the queue is run
1215 * before we add this entry back on the dispatch list,
1216 * we'll re-run it below.
1218 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1219 blk_mq_put_dispatch_budget(hctx);
1221 * For non-shared tags, the RESTART check
1224 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1230 list_del_init(&rq->queuelist);
1235 * Flag last if we have no more requests, or if we have more
1236 * but can't assign a driver tag to it.
1238 if (list_empty(list))
1241 nxt = list_first_entry(list, struct request, queuelist);
1242 bd.last = !blk_mq_get_driver_tag(nxt);
1245 ret = q->mq_ops->queue_rq(hctx, &bd);
1246 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1248 * If an I/O scheduler has been configured and we got a
1249 * driver tag for the next request already, free it
1252 if (!list_empty(list)) {
1253 nxt = list_first_entry(list, struct request, queuelist);
1254 blk_mq_put_driver_tag(nxt);
1256 list_add(&rq->queuelist, list);
1257 __blk_mq_requeue_request(rq);
1261 if (unlikely(ret != BLK_STS_OK)) {
1263 blk_mq_end_request(rq, BLK_STS_IOERR);
1268 } while (!list_empty(list));
1270 hctx->dispatched[queued_to_index(queued)]++;
1273 * Any items that need requeuing? Stuff them into hctx->dispatch,
1274 * that is where we will continue on next queue run.
1276 if (!list_empty(list)) {
1280 * If we didn't flush the entire list, we could have told
1281 * the driver there was more coming, but that turned out to
1284 if (q->mq_ops->commit_rqs)
1285 q->mq_ops->commit_rqs(hctx);
1287 spin_lock(&hctx->lock);
1288 list_splice_init(list, &hctx->dispatch);
1289 spin_unlock(&hctx->lock);
1292 * If SCHED_RESTART was set by the caller of this function and
1293 * it is no longer set that means that it was cleared by another
1294 * thread and hence that a queue rerun is needed.
1296 * If 'no_tag' is set, that means that we failed getting
1297 * a driver tag with an I/O scheduler attached. If our dispatch
1298 * waitqueue is no longer active, ensure that we run the queue
1299 * AFTER adding our entries back to the list.
1301 * If no I/O scheduler has been configured it is possible that
1302 * the hardware queue got stopped and restarted before requests
1303 * were pushed back onto the dispatch list. Rerun the queue to
1304 * avoid starvation. Notes:
1305 * - blk_mq_run_hw_queue() checks whether or not a queue has
1306 * been stopped before rerunning a queue.
1307 * - Some but not all block drivers stop a queue before
1308 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1311 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1312 * bit is set, run queue after a delay to avoid IO stalls
1313 * that could otherwise occur if the queue is idle.
1315 needs_restart = blk_mq_sched_needs_restart(hctx);
1316 if (!needs_restart ||
1317 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1318 blk_mq_run_hw_queue(hctx, true);
1319 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1320 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1322 blk_mq_update_dispatch_busy(hctx, true);
1325 blk_mq_update_dispatch_busy(hctx, false);
1328 * If the host/device is unable to accept more work, inform the
1331 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1334 return (queued + errors) != 0;
1337 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1342 * We should be running this queue from one of the CPUs that
1345 * There are at least two related races now between setting
1346 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1347 * __blk_mq_run_hw_queue():
1349 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1350 * but later it becomes online, then this warning is harmless
1353 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1354 * but later it becomes offline, then the warning can't be
1355 * triggered, and we depend on blk-mq timeout handler to
1356 * handle dispatched requests to this hctx
1358 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1359 cpu_online(hctx->next_cpu)) {
1360 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1361 raw_smp_processor_id(),
1362 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1367 * We can't run the queue inline with ints disabled. Ensure that
1368 * we catch bad users of this early.
1370 WARN_ON_ONCE(in_interrupt());
1372 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1374 hctx_lock(hctx, &srcu_idx);
1375 blk_mq_sched_dispatch_requests(hctx);
1376 hctx_unlock(hctx, srcu_idx);
1379 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1381 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1383 if (cpu >= nr_cpu_ids)
1384 cpu = cpumask_first(hctx->cpumask);
1389 * It'd be great if the workqueue API had a way to pass
1390 * in a mask and had some smarts for more clever placement.
1391 * For now we just round-robin here, switching for every
1392 * BLK_MQ_CPU_WORK_BATCH queued items.
1394 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1397 int next_cpu = hctx->next_cpu;
1399 if (hctx->queue->nr_hw_queues == 1)
1400 return WORK_CPU_UNBOUND;
1402 if (--hctx->next_cpu_batch <= 0) {
1404 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1406 if (next_cpu >= nr_cpu_ids)
1407 next_cpu = blk_mq_first_mapped_cpu(hctx);
1408 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1412 * Do unbound schedule if we can't find a online CPU for this hctx,
1413 * and it should only happen in the path of handling CPU DEAD.
1415 if (!cpu_online(next_cpu)) {
1422 * Make sure to re-select CPU next time once after CPUs
1423 * in hctx->cpumask become online again.
1425 hctx->next_cpu = next_cpu;
1426 hctx->next_cpu_batch = 1;
1427 return WORK_CPU_UNBOUND;
1430 hctx->next_cpu = next_cpu;
1434 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1435 unsigned long msecs)
1437 if (unlikely(blk_mq_hctx_stopped(hctx)))
1440 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1441 int cpu = get_cpu();
1442 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1443 __blk_mq_run_hw_queue(hctx);
1451 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1452 msecs_to_jiffies(msecs));
1455 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1457 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1459 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1461 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1467 * When queue is quiesced, we may be switching io scheduler, or
1468 * updating nr_hw_queues, or other things, and we can't run queue
1469 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1471 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1474 hctx_lock(hctx, &srcu_idx);
1475 need_run = !blk_queue_quiesced(hctx->queue) &&
1476 blk_mq_hctx_has_pending(hctx);
1477 hctx_unlock(hctx, srcu_idx);
1480 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1486 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1488 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1490 struct blk_mq_hw_ctx *hctx;
1493 queue_for_each_hw_ctx(q, hctx, i) {
1494 if (blk_mq_hctx_stopped(hctx))
1497 blk_mq_run_hw_queue(hctx, async);
1500 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1503 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1504 * @q: request queue.
1506 * The caller is responsible for serializing this function against
1507 * blk_mq_{start,stop}_hw_queue().
1509 bool blk_mq_queue_stopped(struct request_queue *q)
1511 struct blk_mq_hw_ctx *hctx;
1514 queue_for_each_hw_ctx(q, hctx, i)
1515 if (blk_mq_hctx_stopped(hctx))
1520 EXPORT_SYMBOL(blk_mq_queue_stopped);
1523 * This function is often used for pausing .queue_rq() by driver when
1524 * there isn't enough resource or some conditions aren't satisfied, and
1525 * BLK_STS_RESOURCE is usually returned.
1527 * We do not guarantee that dispatch can be drained or blocked
1528 * after blk_mq_stop_hw_queue() returns. Please use
1529 * blk_mq_quiesce_queue() for that requirement.
1531 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1533 cancel_delayed_work(&hctx->run_work);
1535 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1537 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1540 * This function is often used for pausing .queue_rq() by driver when
1541 * there isn't enough resource or some conditions aren't satisfied, and
1542 * BLK_STS_RESOURCE is usually returned.
1544 * We do not guarantee that dispatch can be drained or blocked
1545 * after blk_mq_stop_hw_queues() returns. Please use
1546 * blk_mq_quiesce_queue() for that requirement.
1548 void blk_mq_stop_hw_queues(struct request_queue *q)
1550 struct blk_mq_hw_ctx *hctx;
1553 queue_for_each_hw_ctx(q, hctx, i)
1554 blk_mq_stop_hw_queue(hctx);
1556 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1558 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1560 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1562 blk_mq_run_hw_queue(hctx, false);
1564 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1566 void blk_mq_start_hw_queues(struct request_queue *q)
1568 struct blk_mq_hw_ctx *hctx;
1571 queue_for_each_hw_ctx(q, hctx, i)
1572 blk_mq_start_hw_queue(hctx);
1574 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1576 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1578 if (!blk_mq_hctx_stopped(hctx))
1581 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1582 blk_mq_run_hw_queue(hctx, async);
1584 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1586 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1588 struct blk_mq_hw_ctx *hctx;
1591 queue_for_each_hw_ctx(q, hctx, i)
1592 blk_mq_start_stopped_hw_queue(hctx, async);
1594 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1596 static void blk_mq_run_work_fn(struct work_struct *work)
1598 struct blk_mq_hw_ctx *hctx;
1600 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1603 * If we are stopped, don't run the queue.
1605 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1608 __blk_mq_run_hw_queue(hctx);
1611 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1615 struct blk_mq_ctx *ctx = rq->mq_ctx;
1616 enum hctx_type type = hctx->type;
1618 lockdep_assert_held(&ctx->lock);
1620 trace_block_rq_insert(hctx->queue, rq);
1623 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1625 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1628 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1631 struct blk_mq_ctx *ctx = rq->mq_ctx;
1633 lockdep_assert_held(&ctx->lock);
1635 __blk_mq_insert_req_list(hctx, rq, at_head);
1636 blk_mq_hctx_mark_pending(hctx, ctx);
1640 * Should only be used carefully, when the caller knows we want to
1641 * bypass a potential IO scheduler on the target device.
1643 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1645 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1647 spin_lock(&hctx->lock);
1648 list_add_tail(&rq->queuelist, &hctx->dispatch);
1649 spin_unlock(&hctx->lock);
1652 blk_mq_run_hw_queue(hctx, false);
1655 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1656 struct list_head *list)
1660 enum hctx_type type = hctx->type;
1663 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1666 list_for_each_entry(rq, list, queuelist) {
1667 BUG_ON(rq->mq_ctx != ctx);
1668 trace_block_rq_insert(hctx->queue, rq);
1671 spin_lock(&ctx->lock);
1672 list_splice_tail_init(list, &ctx->rq_lists[type]);
1673 blk_mq_hctx_mark_pending(hctx, ctx);
1674 spin_unlock(&ctx->lock);
1677 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1679 struct request *rqa = container_of(a, struct request, queuelist);
1680 struct request *rqb = container_of(b, struct request, queuelist);
1682 if (rqa->mq_ctx < rqb->mq_ctx)
1684 else if (rqa->mq_ctx > rqb->mq_ctx)
1686 else if (rqa->mq_hctx < rqb->mq_hctx)
1688 else if (rqa->mq_hctx > rqb->mq_hctx)
1691 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1694 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1696 struct blk_mq_hw_ctx *this_hctx;
1697 struct blk_mq_ctx *this_ctx;
1698 struct request_queue *this_q;
1704 list_splice_init(&plug->mq_list, &list);
1707 if (plug->rq_count > 2 && plug->multiple_queues)
1708 list_sort(NULL, &list, plug_rq_cmp);
1715 while (!list_empty(&list)) {
1716 rq = list_entry_rq(list.next);
1717 list_del_init(&rq->queuelist);
1719 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1721 trace_block_unplug(this_q, depth, !from_schedule);
1722 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1728 this_ctx = rq->mq_ctx;
1729 this_hctx = rq->mq_hctx;
1734 list_add_tail(&rq->queuelist, &rq_list);
1738 * If 'this_hctx' is set, we know we have entries to complete
1739 * on 'rq_list'. Do those.
1742 trace_block_unplug(this_q, depth, !from_schedule);
1743 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1748 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1750 blk_init_request_from_bio(rq, bio);
1752 blk_account_io_start(rq, true);
1755 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1757 blk_qc_t *cookie, bool last)
1759 struct request_queue *q = rq->q;
1760 struct blk_mq_queue_data bd = {
1764 blk_qc_t new_cookie;
1767 new_cookie = request_to_qc_t(hctx, rq);
1770 * For OK queue, we are done. For error, caller may kill it.
1771 * Any other error (busy), just add it to our list as we
1772 * previously would have done.
1774 ret = q->mq_ops->queue_rq(hctx, &bd);
1777 blk_mq_update_dispatch_busy(hctx, false);
1778 *cookie = new_cookie;
1780 case BLK_STS_RESOURCE:
1781 case BLK_STS_DEV_RESOURCE:
1782 blk_mq_update_dispatch_busy(hctx, true);
1783 __blk_mq_requeue_request(rq);
1786 blk_mq_update_dispatch_busy(hctx, false);
1787 *cookie = BLK_QC_T_NONE;
1794 blk_status_t blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1797 bool bypass, bool last)
1799 struct request_queue *q = rq->q;
1800 bool run_queue = true;
1801 blk_status_t ret = BLK_STS_RESOURCE;
1805 hctx_lock(hctx, &srcu_idx);
1807 * hctx_lock is needed before checking quiesced flag.
1809 * When queue is stopped or quiesced, ignore 'bypass', insert
1810 * and return BLK_STS_OK to caller, and avoid driver to try to
1813 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) {
1819 if (unlikely(q->elevator && !bypass))
1822 if (!blk_mq_get_dispatch_budget(hctx))
1825 if (!blk_mq_get_driver_tag(rq)) {
1826 blk_mq_put_dispatch_budget(hctx);
1831 * Always add a request that has been through
1832 *.queue_rq() to the hardware dispatch list.
1835 ret = __blk_mq_issue_directly(hctx, rq, cookie, last);
1837 hctx_unlock(hctx, srcu_idx);
1841 case BLK_STS_DEV_RESOURCE:
1842 case BLK_STS_RESOURCE:
1844 blk_mq_request_bypass_insert(rq, run_queue);
1846 * We have to return BLK_STS_OK for the DM
1847 * to avoid livelock. Otherwise, we return
1848 * the real result to indicate whether the
1849 * request is direct-issued successfully.
1851 ret = bypass ? BLK_STS_OK : ret;
1852 } else if (!bypass) {
1853 blk_mq_sched_insert_request(rq, false,
1859 blk_mq_end_request(rq, ret);
1866 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1867 struct list_head *list)
1870 blk_status_t ret = BLK_STS_OK;
1872 while (!list_empty(list)) {
1873 struct request *rq = list_first_entry(list, struct request,
1876 list_del_init(&rq->queuelist);
1877 if (ret == BLK_STS_OK)
1878 ret = blk_mq_try_issue_directly(hctx, rq, &unused,
1882 blk_mq_sched_insert_request(rq, false, true, false);
1886 * If we didn't flush the entire list, we could have told
1887 * the driver there was more coming, but that turned out to
1890 if (ret != BLK_STS_OK && hctx->queue->mq_ops->commit_rqs)
1891 hctx->queue->mq_ops->commit_rqs(hctx);
1894 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1896 list_add_tail(&rq->queuelist, &plug->mq_list);
1898 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1899 struct request *tmp;
1901 tmp = list_first_entry(&plug->mq_list, struct request,
1903 if (tmp->q != rq->q)
1904 plug->multiple_queues = true;
1908 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1910 const int is_sync = op_is_sync(bio->bi_opf);
1911 const int is_flush_fua = op_is_flush(bio->bi_opf);
1912 struct blk_mq_alloc_data data = { .flags = 0};
1914 struct blk_plug *plug;
1915 struct request *same_queue_rq = NULL;
1918 blk_queue_bounce(q, &bio);
1920 blk_queue_split(q, &bio);
1922 if (!bio_integrity_prep(bio))
1923 return BLK_QC_T_NONE;
1925 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1926 blk_attempt_plug_merge(q, bio, &same_queue_rq))
1927 return BLK_QC_T_NONE;
1929 if (blk_mq_sched_bio_merge(q, bio))
1930 return BLK_QC_T_NONE;
1932 rq_qos_throttle(q, bio);
1934 data.cmd_flags = bio->bi_opf;
1935 rq = blk_mq_get_request(q, bio, &data);
1936 if (unlikely(!rq)) {
1937 rq_qos_cleanup(q, bio);
1938 if (bio->bi_opf & REQ_NOWAIT)
1939 bio_wouldblock_error(bio);
1940 return BLK_QC_T_NONE;
1943 trace_block_getrq(q, bio, bio->bi_opf);
1945 rq_qos_track(q, rq, bio);
1947 cookie = request_to_qc_t(data.hctx, rq);
1949 plug = current->plug;
1950 if (unlikely(is_flush_fua)) {
1951 blk_mq_put_ctx(data.ctx);
1952 blk_mq_bio_to_request(rq, bio);
1954 /* bypass scheduler for flush rq */
1955 blk_insert_flush(rq);
1956 blk_mq_run_hw_queue(data.hctx, true);
1957 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1959 * Use plugging if we have a ->commit_rqs() hook as well, as
1960 * we know the driver uses bd->last in a smart fashion.
1962 unsigned int request_count = plug->rq_count;
1963 struct request *last = NULL;
1965 blk_mq_put_ctx(data.ctx);
1966 blk_mq_bio_to_request(rq, bio);
1969 trace_block_plug(q);
1971 last = list_entry_rq(plug->mq_list.prev);
1973 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1974 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1975 blk_flush_plug_list(plug, false);
1976 trace_block_plug(q);
1979 blk_add_rq_to_plug(plug, rq);
1980 } else if (plug && !blk_queue_nomerges(q)) {
1981 blk_mq_bio_to_request(rq, bio);
1984 * We do limited plugging. If the bio can be merged, do that.
1985 * Otherwise the existing request in the plug list will be
1986 * issued. So the plug list will have one request at most
1987 * The plug list might get flushed before this. If that happens,
1988 * the plug list is empty, and same_queue_rq is invalid.
1990 if (list_empty(&plug->mq_list))
1991 same_queue_rq = NULL;
1992 if (same_queue_rq) {
1993 list_del_init(&same_queue_rq->queuelist);
1996 blk_add_rq_to_plug(plug, rq);
1998 blk_mq_put_ctx(data.ctx);
2000 if (same_queue_rq) {
2001 data.hctx = same_queue_rq->mq_hctx;
2002 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2003 &cookie, false, true);
2005 } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2006 !data.hctx->dispatch_busy)) {
2007 blk_mq_put_ctx(data.ctx);
2008 blk_mq_bio_to_request(rq, bio);
2009 blk_mq_try_issue_directly(data.hctx, rq, &cookie, false, true);
2011 blk_mq_put_ctx(data.ctx);
2012 blk_mq_bio_to_request(rq, bio);
2013 blk_mq_sched_insert_request(rq, false, true, true);
2019 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2020 unsigned int hctx_idx)
2024 if (tags->rqs && set->ops->exit_request) {
2027 for (i = 0; i < tags->nr_tags; i++) {
2028 struct request *rq = tags->static_rqs[i];
2032 set->ops->exit_request(set, rq, hctx_idx);
2033 tags->static_rqs[i] = NULL;
2037 while (!list_empty(&tags->page_list)) {
2038 page = list_first_entry(&tags->page_list, struct page, lru);
2039 list_del_init(&page->lru);
2041 * Remove kmemleak object previously allocated in
2042 * blk_mq_init_rq_map().
2044 kmemleak_free(page_address(page));
2045 __free_pages(page, page->private);
2049 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2053 kfree(tags->static_rqs);
2054 tags->static_rqs = NULL;
2056 blk_mq_free_tags(tags);
2059 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2060 unsigned int hctx_idx,
2061 unsigned int nr_tags,
2062 unsigned int reserved_tags)
2064 struct blk_mq_tags *tags;
2067 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2068 if (node == NUMA_NO_NODE)
2069 node = set->numa_node;
2071 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2072 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2076 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2077 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2080 blk_mq_free_tags(tags);
2084 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2085 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2087 if (!tags->static_rqs) {
2089 blk_mq_free_tags(tags);
2096 static size_t order_to_size(unsigned int order)
2098 return (size_t)PAGE_SIZE << order;
2101 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2102 unsigned int hctx_idx, int node)
2106 if (set->ops->init_request) {
2107 ret = set->ops->init_request(set, rq, hctx_idx, node);
2112 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2116 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2117 unsigned int hctx_idx, unsigned int depth)
2119 unsigned int i, j, entries_per_page, max_order = 4;
2120 size_t rq_size, left;
2123 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2124 if (node == NUMA_NO_NODE)
2125 node = set->numa_node;
2127 INIT_LIST_HEAD(&tags->page_list);
2130 * rq_size is the size of the request plus driver payload, rounded
2131 * to the cacheline size
2133 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2135 left = rq_size * depth;
2137 for (i = 0; i < depth; ) {
2138 int this_order = max_order;
2143 while (this_order && left < order_to_size(this_order - 1))
2147 page = alloc_pages_node(node,
2148 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2154 if (order_to_size(this_order) < rq_size)
2161 page->private = this_order;
2162 list_add_tail(&page->lru, &tags->page_list);
2164 p = page_address(page);
2166 * Allow kmemleak to scan these pages as they contain pointers
2167 * to additional allocations like via ops->init_request().
2169 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2170 entries_per_page = order_to_size(this_order) / rq_size;
2171 to_do = min(entries_per_page, depth - i);
2172 left -= to_do * rq_size;
2173 for (j = 0; j < to_do; j++) {
2174 struct request *rq = p;
2176 tags->static_rqs[i] = rq;
2177 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2178 tags->static_rqs[i] = NULL;
2189 blk_mq_free_rqs(set, tags, hctx_idx);
2194 * 'cpu' is going away. splice any existing rq_list entries from this
2195 * software queue to the hw queue dispatch list, and ensure that it
2198 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2200 struct blk_mq_hw_ctx *hctx;
2201 struct blk_mq_ctx *ctx;
2203 enum hctx_type type;
2205 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2206 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2209 spin_lock(&ctx->lock);
2210 if (!list_empty(&ctx->rq_lists[type])) {
2211 list_splice_init(&ctx->rq_lists[type], &tmp);
2212 blk_mq_hctx_clear_pending(hctx, ctx);
2214 spin_unlock(&ctx->lock);
2216 if (list_empty(&tmp))
2219 spin_lock(&hctx->lock);
2220 list_splice_tail_init(&tmp, &hctx->dispatch);
2221 spin_unlock(&hctx->lock);
2223 blk_mq_run_hw_queue(hctx, true);
2227 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2229 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2233 /* hctx->ctxs will be freed in queue's release handler */
2234 static void blk_mq_exit_hctx(struct request_queue *q,
2235 struct blk_mq_tag_set *set,
2236 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2238 if (blk_mq_hw_queue_mapped(hctx))
2239 blk_mq_tag_idle(hctx);
2241 if (set->ops->exit_request)
2242 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2244 if (set->ops->exit_hctx)
2245 set->ops->exit_hctx(hctx, hctx_idx);
2247 if (hctx->flags & BLK_MQ_F_BLOCKING)
2248 cleanup_srcu_struct(hctx->srcu);
2250 blk_mq_remove_cpuhp(hctx);
2251 blk_free_flush_queue(hctx->fq);
2252 sbitmap_free(&hctx->ctx_map);
2255 static void blk_mq_exit_hw_queues(struct request_queue *q,
2256 struct blk_mq_tag_set *set, int nr_queue)
2258 struct blk_mq_hw_ctx *hctx;
2261 queue_for_each_hw_ctx(q, hctx, i) {
2264 blk_mq_debugfs_unregister_hctx(hctx);
2265 blk_mq_exit_hctx(q, set, hctx, i);
2269 static int blk_mq_init_hctx(struct request_queue *q,
2270 struct blk_mq_tag_set *set,
2271 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2275 node = hctx->numa_node;
2276 if (node == NUMA_NO_NODE)
2277 node = hctx->numa_node = set->numa_node;
2279 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2280 spin_lock_init(&hctx->lock);
2281 INIT_LIST_HEAD(&hctx->dispatch);
2283 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2285 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2287 hctx->tags = set->tags[hctx_idx];
2290 * Allocate space for all possible cpus to avoid allocation at
2293 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2294 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node);
2296 goto unregister_cpu_notifier;
2298 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2299 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node))
2304 spin_lock_init(&hctx->dispatch_wait_lock);
2305 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2306 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2308 if (set->ops->init_hctx &&
2309 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2312 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2313 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
2317 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2320 if (hctx->flags & BLK_MQ_F_BLOCKING)
2321 init_srcu_struct(hctx->srcu);
2328 if (set->ops->exit_hctx)
2329 set->ops->exit_hctx(hctx, hctx_idx);
2331 sbitmap_free(&hctx->ctx_map);
2334 unregister_cpu_notifier:
2335 blk_mq_remove_cpuhp(hctx);
2339 static void blk_mq_init_cpu_queues(struct request_queue *q,
2340 unsigned int nr_hw_queues)
2342 struct blk_mq_tag_set *set = q->tag_set;
2345 for_each_possible_cpu(i) {
2346 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2347 struct blk_mq_hw_ctx *hctx;
2351 spin_lock_init(&__ctx->lock);
2352 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2353 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2358 * Set local node, IFF we have more than one hw queue. If
2359 * not, we remain on the home node of the device
2361 for (j = 0; j < set->nr_maps; j++) {
2362 hctx = blk_mq_map_queue_type(q, j, i);
2363 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2364 hctx->numa_node = local_memory_node(cpu_to_node(i));
2369 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2373 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2374 set->queue_depth, set->reserved_tags);
2375 if (!set->tags[hctx_idx])
2378 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2383 blk_mq_free_rq_map(set->tags[hctx_idx]);
2384 set->tags[hctx_idx] = NULL;
2388 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2389 unsigned int hctx_idx)
2391 if (set->tags && set->tags[hctx_idx]) {
2392 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2393 blk_mq_free_rq_map(set->tags[hctx_idx]);
2394 set->tags[hctx_idx] = NULL;
2398 static void blk_mq_map_swqueue(struct request_queue *q)
2400 unsigned int i, j, hctx_idx;
2401 struct blk_mq_hw_ctx *hctx;
2402 struct blk_mq_ctx *ctx;
2403 struct blk_mq_tag_set *set = q->tag_set;
2406 * Avoid others reading imcomplete hctx->cpumask through sysfs
2408 mutex_lock(&q->sysfs_lock);
2410 queue_for_each_hw_ctx(q, hctx, i) {
2411 cpumask_clear(hctx->cpumask);
2413 hctx->dispatch_from = NULL;
2417 * Map software to hardware queues.
2419 * If the cpu isn't present, the cpu is mapped to first hctx.
2421 for_each_possible_cpu(i) {
2422 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2423 /* unmapped hw queue can be remapped after CPU topo changed */
2424 if (!set->tags[hctx_idx] &&
2425 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2427 * If tags initialization fail for some hctx,
2428 * that hctx won't be brought online. In this
2429 * case, remap the current ctx to hctx[0] which
2430 * is guaranteed to always have tags allocated
2432 set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2435 ctx = per_cpu_ptr(q->queue_ctx, i);
2436 for (j = 0; j < set->nr_maps; j++) {
2437 if (!set->map[j].nr_queues) {
2438 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2439 HCTX_TYPE_DEFAULT, i);
2443 hctx = blk_mq_map_queue_type(q, j, i);
2444 ctx->hctxs[j] = hctx;
2446 * If the CPU is already set in the mask, then we've
2447 * mapped this one already. This can happen if
2448 * devices share queues across queue maps.
2450 if (cpumask_test_cpu(i, hctx->cpumask))
2453 cpumask_set_cpu(i, hctx->cpumask);
2455 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2456 hctx->ctxs[hctx->nr_ctx++] = ctx;
2459 * If the nr_ctx type overflows, we have exceeded the
2460 * amount of sw queues we can support.
2462 BUG_ON(!hctx->nr_ctx);
2465 for (; j < HCTX_MAX_TYPES; j++)
2466 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2467 HCTX_TYPE_DEFAULT, i);
2470 mutex_unlock(&q->sysfs_lock);
2472 queue_for_each_hw_ctx(q, hctx, i) {
2474 * If no software queues are mapped to this hardware queue,
2475 * disable it and free the request entries.
2477 if (!hctx->nr_ctx) {
2478 /* Never unmap queue 0. We need it as a
2479 * fallback in case of a new remap fails
2482 if (i && set->tags[i])
2483 blk_mq_free_map_and_requests(set, i);
2489 hctx->tags = set->tags[i];
2490 WARN_ON(!hctx->tags);
2493 * Set the map size to the number of mapped software queues.
2494 * This is more accurate and more efficient than looping
2495 * over all possibly mapped software queues.
2497 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2500 * Initialize batch roundrobin counts
2502 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2503 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2508 * Caller needs to ensure that we're either frozen/quiesced, or that
2509 * the queue isn't live yet.
2511 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2513 struct blk_mq_hw_ctx *hctx;
2516 queue_for_each_hw_ctx(q, hctx, i) {
2518 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2520 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2524 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2527 struct request_queue *q;
2529 lockdep_assert_held(&set->tag_list_lock);
2531 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2532 blk_mq_freeze_queue(q);
2533 queue_set_hctx_shared(q, shared);
2534 blk_mq_unfreeze_queue(q);
2538 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2540 struct blk_mq_tag_set *set = q->tag_set;
2542 mutex_lock(&set->tag_list_lock);
2543 list_del_rcu(&q->tag_set_list);
2544 if (list_is_singular(&set->tag_list)) {
2545 /* just transitioned to unshared */
2546 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2547 /* update existing queue */
2548 blk_mq_update_tag_set_depth(set, false);
2550 mutex_unlock(&set->tag_list_lock);
2551 INIT_LIST_HEAD(&q->tag_set_list);
2554 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2555 struct request_queue *q)
2557 mutex_lock(&set->tag_list_lock);
2560 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2562 if (!list_empty(&set->tag_list) &&
2563 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2564 set->flags |= BLK_MQ_F_TAG_SHARED;
2565 /* update existing queue */
2566 blk_mq_update_tag_set_depth(set, true);
2568 if (set->flags & BLK_MQ_F_TAG_SHARED)
2569 queue_set_hctx_shared(q, true);
2570 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2572 mutex_unlock(&set->tag_list_lock);
2575 /* All allocations will be freed in release handler of q->mq_kobj */
2576 static int blk_mq_alloc_ctxs(struct request_queue *q)
2578 struct blk_mq_ctxs *ctxs;
2581 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2585 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2586 if (!ctxs->queue_ctx)
2589 for_each_possible_cpu(cpu) {
2590 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2594 q->mq_kobj = &ctxs->kobj;
2595 q->queue_ctx = ctxs->queue_ctx;
2604 * It is the actual release handler for mq, but we do it from
2605 * request queue's release handler for avoiding use-after-free
2606 * and headache because q->mq_kobj shouldn't have been introduced,
2607 * but we can't group ctx/kctx kobj without it.
2609 void blk_mq_release(struct request_queue *q)
2611 struct blk_mq_hw_ctx *hctx;
2614 /* hctx kobj stays in hctx */
2615 queue_for_each_hw_ctx(q, hctx, i) {
2618 kobject_put(&hctx->kobj);
2621 kfree(q->queue_hw_ctx);
2624 * release .mq_kobj and sw queue's kobject now because
2625 * both share lifetime with request queue.
2627 blk_mq_sysfs_deinit(q);
2630 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2632 struct request_queue *uninit_q, *q;
2634 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2636 return ERR_PTR(-ENOMEM);
2638 q = blk_mq_init_allocated_queue(set, uninit_q);
2640 blk_cleanup_queue(uninit_q);
2644 EXPORT_SYMBOL(blk_mq_init_queue);
2647 * Helper for setting up a queue with mq ops, given queue depth, and
2648 * the passed in mq ops flags.
2650 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2651 const struct blk_mq_ops *ops,
2652 unsigned int queue_depth,
2653 unsigned int set_flags)
2655 struct request_queue *q;
2658 memset(set, 0, sizeof(*set));
2660 set->nr_hw_queues = 1;
2662 set->queue_depth = queue_depth;
2663 set->numa_node = NUMA_NO_NODE;
2664 set->flags = set_flags;
2666 ret = blk_mq_alloc_tag_set(set);
2668 return ERR_PTR(ret);
2670 q = blk_mq_init_queue(set);
2672 blk_mq_free_tag_set(set);
2678 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2680 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2682 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2684 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2685 __alignof__(struct blk_mq_hw_ctx)) !=
2686 sizeof(struct blk_mq_hw_ctx));
2688 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2689 hw_ctx_size += sizeof(struct srcu_struct);
2694 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2695 struct blk_mq_tag_set *set, struct request_queue *q,
2696 int hctx_idx, int node)
2698 struct blk_mq_hw_ctx *hctx;
2700 hctx = kzalloc_node(blk_mq_hw_ctx_size(set),
2701 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2706 if (!zalloc_cpumask_var_node(&hctx->cpumask,
2707 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2713 atomic_set(&hctx->nr_active, 0);
2714 hctx->numa_node = node;
2715 hctx->queue_num = hctx_idx;
2717 if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) {
2718 free_cpumask_var(hctx->cpumask);
2722 blk_mq_hctx_kobj_init(hctx);
2727 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2728 struct request_queue *q)
2731 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2733 /* protect against switching io scheduler */
2734 mutex_lock(&q->sysfs_lock);
2735 for (i = 0; i < set->nr_hw_queues; i++) {
2737 struct blk_mq_hw_ctx *hctx;
2739 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2741 * If the hw queue has been mapped to another numa node,
2742 * we need to realloc the hctx. If allocation fails, fallback
2743 * to use the previous one.
2745 if (hctxs[i] && (hctxs[i]->numa_node == node))
2748 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2751 blk_mq_exit_hctx(q, set, hctxs[i], i);
2752 kobject_put(&hctxs[i]->kobj);
2757 pr_warn("Allocate new hctx on node %d fails,\
2758 fallback to previous one on node %d\n",
2759 node, hctxs[i]->numa_node);
2765 * Increasing nr_hw_queues fails. Free the newly allocated
2766 * hctxs and keep the previous q->nr_hw_queues.
2768 if (i != set->nr_hw_queues) {
2769 j = q->nr_hw_queues;
2773 end = q->nr_hw_queues;
2774 q->nr_hw_queues = set->nr_hw_queues;
2777 for (; j < end; j++) {
2778 struct blk_mq_hw_ctx *hctx = hctxs[j];
2782 blk_mq_free_map_and_requests(set, j);
2783 blk_mq_exit_hctx(q, set, hctx, j);
2784 kobject_put(&hctx->kobj);
2789 mutex_unlock(&q->sysfs_lock);
2793 * Maximum number of hardware queues we support. For single sets, we'll never
2794 * have more than the CPUs (software queues). For multiple sets, the tag_set
2795 * user may have set ->nr_hw_queues larger.
2797 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2799 if (set->nr_maps == 1)
2802 return max(set->nr_hw_queues, nr_cpu_ids);
2805 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2806 struct request_queue *q)
2808 /* mark the queue as mq asap */
2809 q->mq_ops = set->ops;
2811 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2812 blk_mq_poll_stats_bkt,
2813 BLK_MQ_POLL_STATS_BKTS, q);
2817 if (blk_mq_alloc_ctxs(q))
2820 /* init q->mq_kobj and sw queues' kobjects */
2821 blk_mq_sysfs_init(q);
2823 q->nr_queues = nr_hw_queues(set);
2824 q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2825 GFP_KERNEL, set->numa_node);
2826 if (!q->queue_hw_ctx)
2829 blk_mq_realloc_hw_ctxs(set, q);
2830 if (!q->nr_hw_queues)
2833 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2834 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2838 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2839 if (set->nr_maps > HCTX_TYPE_POLL &&
2840 set->map[HCTX_TYPE_POLL].nr_queues)
2841 blk_queue_flag_set(QUEUE_FLAG_POLL, 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[HCTX_TYPE_DEFAULT]);
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 if (q->nr_requests == nr)
3095 blk_mq_freeze_queue(q);
3096 blk_mq_quiesce_queue(q);
3099 queue_for_each_hw_ctx(q, hctx, i) {
3103 * If we're using an MQ scheduler, just update the scheduler
3104 * queue depth. This is similar to what the old code would do.
3106 if (!hctx->sched_tags) {
3107 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3110 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3118 q->nr_requests = nr;
3120 blk_mq_unquiesce_queue(q);
3121 blk_mq_unfreeze_queue(q);
3127 * request_queue and elevator_type pair.
3128 * It is just used by __blk_mq_update_nr_hw_queues to cache
3129 * the elevator_type associated with a request_queue.
3131 struct blk_mq_qe_pair {
3132 struct list_head node;
3133 struct request_queue *q;
3134 struct elevator_type *type;
3138 * Cache the elevator_type in qe pair list and switch the
3139 * io scheduler to 'none'
3141 static bool blk_mq_elv_switch_none(struct list_head *head,
3142 struct request_queue *q)
3144 struct blk_mq_qe_pair *qe;
3149 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3153 INIT_LIST_HEAD(&qe->node);
3155 qe->type = q->elevator->type;
3156 list_add(&qe->node, head);
3158 mutex_lock(&q->sysfs_lock);
3160 * After elevator_switch_mq, the previous elevator_queue will be
3161 * released by elevator_release. The reference of the io scheduler
3162 * module get by elevator_get will also be put. So we need to get
3163 * a reference of the io scheduler module here to prevent it to be
3166 __module_get(qe->type->elevator_owner);
3167 elevator_switch_mq(q, NULL);
3168 mutex_unlock(&q->sysfs_lock);
3173 static void blk_mq_elv_switch_back(struct list_head *head,
3174 struct request_queue *q)
3176 struct blk_mq_qe_pair *qe;
3177 struct elevator_type *t = NULL;
3179 list_for_each_entry(qe, head, node)
3188 list_del(&qe->node);
3191 mutex_lock(&q->sysfs_lock);
3192 elevator_switch_mq(q, t);
3193 mutex_unlock(&q->sysfs_lock);
3196 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3199 struct request_queue *q;
3201 int prev_nr_hw_queues;
3203 lockdep_assert_held(&set->tag_list_lock);
3205 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3206 nr_hw_queues = nr_cpu_ids;
3207 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3210 list_for_each_entry(q, &set->tag_list, tag_set_list)
3211 blk_mq_freeze_queue(q);
3213 * Sync with blk_mq_queue_tag_busy_iter.
3217 * Switch IO scheduler to 'none', cleaning up the data associated
3218 * with the previous scheduler. We will switch back once we are done
3219 * updating the new sw to hw queue mappings.
3221 list_for_each_entry(q, &set->tag_list, tag_set_list)
3222 if (!blk_mq_elv_switch_none(&head, q))
3225 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3226 blk_mq_debugfs_unregister_hctxs(q);
3227 blk_mq_sysfs_unregister(q);
3230 prev_nr_hw_queues = set->nr_hw_queues;
3231 set->nr_hw_queues = nr_hw_queues;
3232 blk_mq_update_queue_map(set);
3234 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3235 blk_mq_realloc_hw_ctxs(set, q);
3236 if (q->nr_hw_queues != set->nr_hw_queues) {
3237 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3238 nr_hw_queues, prev_nr_hw_queues);
3239 set->nr_hw_queues = prev_nr_hw_queues;
3240 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3243 blk_mq_map_swqueue(q);
3246 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3247 blk_mq_sysfs_register(q);
3248 blk_mq_debugfs_register_hctxs(q);
3252 list_for_each_entry(q, &set->tag_list, tag_set_list)
3253 blk_mq_elv_switch_back(&head, q);
3255 list_for_each_entry(q, &set->tag_list, tag_set_list)
3256 blk_mq_unfreeze_queue(q);
3259 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3261 mutex_lock(&set->tag_list_lock);
3262 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3263 mutex_unlock(&set->tag_list_lock);
3265 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3267 /* Enable polling stats and return whether they were already enabled. */
3268 static bool blk_poll_stats_enable(struct request_queue *q)
3270 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3271 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3273 blk_stat_add_callback(q, q->poll_cb);
3277 static void blk_mq_poll_stats_start(struct request_queue *q)
3280 * We don't arm the callback if polling stats are not enabled or the
3281 * callback is already active.
3283 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3284 blk_stat_is_active(q->poll_cb))
3287 blk_stat_activate_msecs(q->poll_cb, 100);
3290 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3292 struct request_queue *q = cb->data;
3295 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3296 if (cb->stat[bucket].nr_samples)
3297 q->poll_stat[bucket] = cb->stat[bucket];
3301 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3302 struct blk_mq_hw_ctx *hctx,
3305 unsigned long ret = 0;
3309 * If stats collection isn't on, don't sleep but turn it on for
3312 if (!blk_poll_stats_enable(q))
3316 * As an optimistic guess, use half of the mean service time
3317 * for this type of request. We can (and should) make this smarter.
3318 * For instance, if the completion latencies are tight, we can
3319 * get closer than just half the mean. This is especially
3320 * important on devices where the completion latencies are longer
3321 * than ~10 usec. We do use the stats for the relevant IO size
3322 * if available which does lead to better estimates.
3324 bucket = blk_mq_poll_stats_bkt(rq);
3328 if (q->poll_stat[bucket].nr_samples)
3329 ret = (q->poll_stat[bucket].mean + 1) / 2;
3334 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3335 struct blk_mq_hw_ctx *hctx,
3338 struct hrtimer_sleeper hs;
3339 enum hrtimer_mode mode;
3343 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3347 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3349 * 0: use half of prev avg
3350 * >0: use this specific value
3352 if (q->poll_nsec > 0)
3353 nsecs = q->poll_nsec;
3355 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3360 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3363 * This will be replaced with the stats tracking code, using
3364 * 'avg_completion_time / 2' as the pre-sleep target.
3368 mode = HRTIMER_MODE_REL;
3369 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3370 hrtimer_set_expires(&hs.timer, kt);
3372 hrtimer_init_sleeper(&hs, current);
3374 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3376 set_current_state(TASK_UNINTERRUPTIBLE);
3377 hrtimer_start_expires(&hs.timer, mode);
3380 hrtimer_cancel(&hs.timer);
3381 mode = HRTIMER_MODE_ABS;
3382 } while (hs.task && !signal_pending(current));
3384 __set_current_state(TASK_RUNNING);
3385 destroy_hrtimer_on_stack(&hs.timer);
3389 static bool blk_mq_poll_hybrid(struct request_queue *q,
3390 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3394 if (q->poll_nsec == -1)
3397 if (!blk_qc_t_is_internal(cookie))
3398 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3400 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3402 * With scheduling, if the request has completed, we'll
3403 * get a NULL return here, as we clear the sched tag when
3404 * that happens. The request still remains valid, like always,
3405 * so we should be safe with just the NULL check.
3411 return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3415 * blk_poll - poll for IO completions
3417 * @cookie: cookie passed back at IO submission time
3418 * @spin: whether to spin for completions
3421 * Poll for completions on the passed in queue. Returns number of
3422 * completed entries found. If @spin is true, then blk_poll will continue
3423 * looping until at least one completion is found, unless the task is
3424 * otherwise marked running (or we need to reschedule).
3426 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3428 struct blk_mq_hw_ctx *hctx;
3431 if (!blk_qc_t_valid(cookie) ||
3432 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3436 blk_flush_plug_list(current->plug, false);
3438 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3441 * If we sleep, have the caller restart the poll loop to reset
3442 * the state. Like for the other success return cases, the
3443 * caller is responsible for checking if the IO completed. If
3444 * the IO isn't complete, we'll get called again and will go
3445 * straight to the busy poll loop.
3447 if (blk_mq_poll_hybrid(q, hctx, cookie))
3450 hctx->poll_considered++;
3452 state = current->state;
3456 hctx->poll_invoked++;
3458 ret = q->mq_ops->poll(hctx);
3460 hctx->poll_success++;
3461 __set_current_state(TASK_RUNNING);
3465 if (signal_pending_state(state, current))
3466 __set_current_state(TASK_RUNNING);
3468 if (current->state == TASK_RUNNING)
3470 if (ret < 0 || !spin)
3473 } while (!need_resched());
3475 __set_current_state(TASK_RUNNING);
3478 EXPORT_SYMBOL_GPL(blk_poll);
3480 unsigned int blk_mq_rq_cpu(struct request *rq)
3482 return rq->mq_ctx->cpu;
3484 EXPORT_SYMBOL(blk_mq_rq_cpu);
3486 static int __init blk_mq_init(void)
3488 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3489 blk_mq_hctx_notify_dead);
3492 subsys_initcall(blk_mq_init);