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))
739 rq->rq_flags &= ~RQF_SOFTBARRIER;
740 list_del_init(&rq->queuelist);
741 blk_mq_sched_insert_request(rq, true, false, false);
744 while (!list_empty(&rq_list)) {
745 rq = list_entry(rq_list.next, struct request, queuelist);
746 list_del_init(&rq->queuelist);
747 blk_mq_sched_insert_request(rq, false, false, false);
750 blk_mq_run_hw_queues(q, false);
753 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
754 bool kick_requeue_list)
756 struct request_queue *q = rq->q;
760 * We abuse this flag that is otherwise used by the I/O scheduler to
761 * request head insertion from the workqueue.
763 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
765 spin_lock_irqsave(&q->requeue_lock, flags);
767 rq->rq_flags |= RQF_SOFTBARRIER;
768 list_add(&rq->queuelist, &q->requeue_list);
770 list_add_tail(&rq->queuelist, &q->requeue_list);
772 spin_unlock_irqrestore(&q->requeue_lock, flags);
774 if (kick_requeue_list)
775 blk_mq_kick_requeue_list(q);
777 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
779 void blk_mq_kick_requeue_list(struct request_queue *q)
781 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
783 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
785 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
788 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
789 msecs_to_jiffies(msecs));
791 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
793 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
795 if (tag < tags->nr_tags) {
796 prefetch(tags->rqs[tag]);
797 return tags->rqs[tag];
802 EXPORT_SYMBOL(blk_mq_tag_to_rq);
804 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
805 void *priv, bool reserved)
808 * If we find a request that is inflight and the queue matches,
809 * we know the queue is busy. Return false to stop the iteration.
811 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
821 bool blk_mq_queue_inflight(struct request_queue *q)
825 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
828 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
830 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
832 req->rq_flags |= RQF_TIMED_OUT;
833 if (req->q->mq_ops->timeout) {
834 enum blk_eh_timer_return ret;
836 ret = req->q->mq_ops->timeout(req, reserved);
837 if (ret == BLK_EH_DONE)
839 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
845 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
847 unsigned long deadline;
849 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
851 if (rq->rq_flags & RQF_TIMED_OUT)
854 deadline = READ_ONCE(rq->deadline);
855 if (time_after_eq(jiffies, deadline))
860 else if (time_after(*next, deadline))
865 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
866 struct request *rq, void *priv, bool reserved)
868 unsigned long *next = priv;
871 * Just do a quick check if it is expired before locking the request in
872 * so we're not unnecessarilly synchronizing across CPUs.
874 if (!blk_mq_req_expired(rq, next))
878 * We have reason to believe the request may be expired. Take a
879 * reference on the request to lock this request lifetime into its
880 * currently allocated context to prevent it from being reallocated in
881 * the event the completion by-passes this timeout handler.
883 * If the reference was already released, then the driver beat the
884 * timeout handler to posting a natural completion.
886 if (!refcount_inc_not_zero(&rq->ref))
890 * The request is now locked and cannot be reallocated underneath the
891 * timeout handler's processing. Re-verify this exact request is truly
892 * expired; if it is not expired, then the request was completed and
893 * reallocated as a new request.
895 if (blk_mq_req_expired(rq, next))
896 blk_mq_rq_timed_out(rq, reserved);
897 if (refcount_dec_and_test(&rq->ref))
898 __blk_mq_free_request(rq);
903 static void blk_mq_timeout_work(struct work_struct *work)
905 struct request_queue *q =
906 container_of(work, struct request_queue, timeout_work);
907 unsigned long next = 0;
908 struct blk_mq_hw_ctx *hctx;
911 /* A deadlock might occur if a request is stuck requiring a
912 * timeout at the same time a queue freeze is waiting
913 * completion, since the timeout code would not be able to
914 * acquire the queue reference here.
916 * That's why we don't use blk_queue_enter here; instead, we use
917 * percpu_ref_tryget directly, because we need to be able to
918 * obtain a reference even in the short window between the queue
919 * starting to freeze, by dropping the first reference in
920 * blk_freeze_queue_start, and the moment the last request is
921 * consumed, marked by the instant q_usage_counter reaches
924 if (!percpu_ref_tryget(&q->q_usage_counter))
927 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
930 mod_timer(&q->timeout, next);
933 * Request timeouts are handled as a forward rolling timer. If
934 * we end up here it means that no requests are pending and
935 * also that no request has been pending for a while. Mark
938 queue_for_each_hw_ctx(q, hctx, i) {
939 /* the hctx may be unmapped, so check it here */
940 if (blk_mq_hw_queue_mapped(hctx))
941 blk_mq_tag_idle(hctx);
947 struct flush_busy_ctx_data {
948 struct blk_mq_hw_ctx *hctx;
949 struct list_head *list;
952 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
954 struct flush_busy_ctx_data *flush_data = data;
955 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
956 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
957 enum hctx_type type = hctx->type;
959 spin_lock(&ctx->lock);
960 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
961 sbitmap_clear_bit(sb, bitnr);
962 spin_unlock(&ctx->lock);
967 * Process software queues that have been marked busy, splicing them
968 * to the for-dispatch
970 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
972 struct flush_busy_ctx_data data = {
977 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
979 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
981 struct dispatch_rq_data {
982 struct blk_mq_hw_ctx *hctx;
986 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
989 struct dispatch_rq_data *dispatch_data = data;
990 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
991 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
992 enum hctx_type type = hctx->type;
994 spin_lock(&ctx->lock);
995 if (!list_empty(&ctx->rq_lists[type])) {
996 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
997 list_del_init(&dispatch_data->rq->queuelist);
998 if (list_empty(&ctx->rq_lists[type]))
999 sbitmap_clear_bit(sb, bitnr);
1001 spin_unlock(&ctx->lock);
1003 return !dispatch_data->rq;
1006 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1007 struct blk_mq_ctx *start)
1009 unsigned off = start ? start->index_hw[hctx->type] : 0;
1010 struct dispatch_rq_data data = {
1015 __sbitmap_for_each_set(&hctx->ctx_map, off,
1016 dispatch_rq_from_ctx, &data);
1021 static inline unsigned int queued_to_index(unsigned int queued)
1026 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1029 bool blk_mq_get_driver_tag(struct request *rq)
1031 struct blk_mq_alloc_data data = {
1033 .hctx = rq->mq_hctx,
1034 .flags = BLK_MQ_REQ_NOWAIT,
1035 .cmd_flags = rq->cmd_flags,
1042 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1043 data.flags |= BLK_MQ_REQ_RESERVED;
1045 shared = blk_mq_tag_busy(data.hctx);
1046 rq->tag = blk_mq_get_tag(&data);
1049 rq->rq_flags |= RQF_MQ_INFLIGHT;
1050 atomic_inc(&data.hctx->nr_active);
1052 data.hctx->tags->rqs[rq->tag] = rq;
1056 return rq->tag != -1;
1059 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1060 int flags, void *key)
1062 struct blk_mq_hw_ctx *hctx;
1064 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1066 spin_lock(&hctx->dispatch_wait_lock);
1067 list_del_init(&wait->entry);
1068 spin_unlock(&hctx->dispatch_wait_lock);
1070 blk_mq_run_hw_queue(hctx, true);
1075 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1076 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1077 * restart. For both cases, take care to check the condition again after
1078 * marking us as waiting.
1080 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1083 struct wait_queue_head *wq;
1084 wait_queue_entry_t *wait;
1087 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1088 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
1089 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
1092 * It's possible that a tag was freed in the window between the
1093 * allocation failure and adding the hardware queue to the wait
1096 * Don't clear RESTART here, someone else could have set it.
1097 * At most this will cost an extra queue run.
1099 return blk_mq_get_driver_tag(rq);
1102 wait = &hctx->dispatch_wait;
1103 if (!list_empty_careful(&wait->entry))
1106 wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait;
1108 spin_lock_irq(&wq->lock);
1109 spin_lock(&hctx->dispatch_wait_lock);
1110 if (!list_empty(&wait->entry)) {
1111 spin_unlock(&hctx->dispatch_wait_lock);
1112 spin_unlock_irq(&wq->lock);
1116 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1117 __add_wait_queue(wq, wait);
1120 * It's possible that a tag was freed in the window between the
1121 * allocation failure and adding the hardware queue to the wait
1124 ret = blk_mq_get_driver_tag(rq);
1126 spin_unlock(&hctx->dispatch_wait_lock);
1127 spin_unlock_irq(&wq->lock);
1132 * We got a tag, remove ourselves from the wait queue to ensure
1133 * someone else gets the wakeup.
1135 list_del_init(&wait->entry);
1136 spin_unlock(&hctx->dispatch_wait_lock);
1137 spin_unlock_irq(&wq->lock);
1142 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1143 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1145 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1146 * - EWMA is one simple way to compute running average value
1147 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1148 * - take 4 as factor for avoiding to get too small(0) result, and this
1149 * factor doesn't matter because EWMA decreases exponentially
1151 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1155 if (hctx->queue->elevator)
1158 ewma = hctx->dispatch_busy;
1163 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1165 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1166 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1168 hctx->dispatch_busy = ewma;
1171 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1174 * Returns true if we did some work AND can potentially do more.
1176 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1179 struct blk_mq_hw_ctx *hctx;
1180 struct request *rq, *nxt;
1181 bool no_tag = false;
1183 blk_status_t ret = BLK_STS_OK;
1185 if (list_empty(list))
1188 WARN_ON(!list_is_singular(list) && got_budget);
1191 * Now process all the entries, sending them to the driver.
1193 errors = queued = 0;
1195 struct blk_mq_queue_data bd;
1197 rq = list_first_entry(list, struct request, queuelist);
1200 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1203 if (!blk_mq_get_driver_tag(rq)) {
1205 * The initial allocation attempt failed, so we need to
1206 * rerun the hardware queue when a tag is freed. The
1207 * waitqueue takes care of that. If the queue is run
1208 * before we add this entry back on the dispatch list,
1209 * we'll re-run it below.
1211 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1212 blk_mq_put_dispatch_budget(hctx);
1214 * For non-shared tags, the RESTART check
1217 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1223 list_del_init(&rq->queuelist);
1228 * Flag last if we have no more requests, or if we have more
1229 * but can't assign a driver tag to it.
1231 if (list_empty(list))
1234 nxt = list_first_entry(list, struct request, queuelist);
1235 bd.last = !blk_mq_get_driver_tag(nxt);
1238 ret = q->mq_ops->queue_rq(hctx, &bd);
1239 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1241 * If an I/O scheduler has been configured and we got a
1242 * driver tag for the next request already, free it
1245 if (!list_empty(list)) {
1246 nxt = list_first_entry(list, struct request, queuelist);
1247 blk_mq_put_driver_tag(nxt);
1249 list_add(&rq->queuelist, list);
1250 __blk_mq_requeue_request(rq);
1254 if (unlikely(ret != BLK_STS_OK)) {
1256 blk_mq_end_request(rq, BLK_STS_IOERR);
1261 } while (!list_empty(list));
1263 hctx->dispatched[queued_to_index(queued)]++;
1266 * Any items that need requeuing? Stuff them into hctx->dispatch,
1267 * that is where we will continue on next queue run.
1269 if (!list_empty(list)) {
1273 * If we didn't flush the entire list, we could have told
1274 * the driver there was more coming, but that turned out to
1277 if (q->mq_ops->commit_rqs)
1278 q->mq_ops->commit_rqs(hctx);
1280 spin_lock(&hctx->lock);
1281 list_splice_init(list, &hctx->dispatch);
1282 spin_unlock(&hctx->lock);
1285 * If SCHED_RESTART was set by the caller of this function and
1286 * it is no longer set that means that it was cleared by another
1287 * thread and hence that a queue rerun is needed.
1289 * If 'no_tag' is set, that means that we failed getting
1290 * a driver tag with an I/O scheduler attached. If our dispatch
1291 * waitqueue is no longer active, ensure that we run the queue
1292 * AFTER adding our entries back to the list.
1294 * If no I/O scheduler has been configured it is possible that
1295 * the hardware queue got stopped and restarted before requests
1296 * were pushed back onto the dispatch list. Rerun the queue to
1297 * avoid starvation. Notes:
1298 * - blk_mq_run_hw_queue() checks whether or not a queue has
1299 * been stopped before rerunning a queue.
1300 * - Some but not all block drivers stop a queue before
1301 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1304 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1305 * bit is set, run queue after a delay to avoid IO stalls
1306 * that could otherwise occur if the queue is idle.
1308 needs_restart = blk_mq_sched_needs_restart(hctx);
1309 if (!needs_restart ||
1310 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1311 blk_mq_run_hw_queue(hctx, true);
1312 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1313 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1315 blk_mq_update_dispatch_busy(hctx, true);
1318 blk_mq_update_dispatch_busy(hctx, false);
1321 * If the host/device is unable to accept more work, inform the
1324 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1327 return (queued + errors) != 0;
1330 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1335 * We should be running this queue from one of the CPUs that
1338 * There are at least two related races now between setting
1339 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1340 * __blk_mq_run_hw_queue():
1342 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1343 * but later it becomes online, then this warning is harmless
1346 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1347 * but later it becomes offline, then the warning can't be
1348 * triggered, and we depend on blk-mq timeout handler to
1349 * handle dispatched requests to this hctx
1351 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1352 cpu_online(hctx->next_cpu)) {
1353 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1354 raw_smp_processor_id(),
1355 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1360 * We can't run the queue inline with ints disabled. Ensure that
1361 * we catch bad users of this early.
1363 WARN_ON_ONCE(in_interrupt());
1365 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1367 hctx_lock(hctx, &srcu_idx);
1368 blk_mq_sched_dispatch_requests(hctx);
1369 hctx_unlock(hctx, srcu_idx);
1372 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1374 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1376 if (cpu >= nr_cpu_ids)
1377 cpu = cpumask_first(hctx->cpumask);
1382 * It'd be great if the workqueue API had a way to pass
1383 * in a mask and had some smarts for more clever placement.
1384 * For now we just round-robin here, switching for every
1385 * BLK_MQ_CPU_WORK_BATCH queued items.
1387 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1390 int next_cpu = hctx->next_cpu;
1392 if (hctx->queue->nr_hw_queues == 1)
1393 return WORK_CPU_UNBOUND;
1395 if (--hctx->next_cpu_batch <= 0) {
1397 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1399 if (next_cpu >= nr_cpu_ids)
1400 next_cpu = blk_mq_first_mapped_cpu(hctx);
1401 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1405 * Do unbound schedule if we can't find a online CPU for this hctx,
1406 * and it should only happen in the path of handling CPU DEAD.
1408 if (!cpu_online(next_cpu)) {
1415 * Make sure to re-select CPU next time once after CPUs
1416 * in hctx->cpumask become online again.
1418 hctx->next_cpu = next_cpu;
1419 hctx->next_cpu_batch = 1;
1420 return WORK_CPU_UNBOUND;
1423 hctx->next_cpu = next_cpu;
1427 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1428 unsigned long msecs)
1430 if (unlikely(blk_mq_hctx_stopped(hctx)))
1433 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1434 int cpu = get_cpu();
1435 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1436 __blk_mq_run_hw_queue(hctx);
1444 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1445 msecs_to_jiffies(msecs));
1448 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1450 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1452 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1454 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1460 * When queue is quiesced, we may be switching io scheduler, or
1461 * updating nr_hw_queues, or other things, and we can't run queue
1462 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1464 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1467 hctx_lock(hctx, &srcu_idx);
1468 need_run = !blk_queue_quiesced(hctx->queue) &&
1469 blk_mq_hctx_has_pending(hctx);
1470 hctx_unlock(hctx, srcu_idx);
1473 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1479 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1481 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1483 struct blk_mq_hw_ctx *hctx;
1486 queue_for_each_hw_ctx(q, hctx, i) {
1487 if (blk_mq_hctx_stopped(hctx))
1490 blk_mq_run_hw_queue(hctx, async);
1493 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1496 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1497 * @q: request queue.
1499 * The caller is responsible for serializing this function against
1500 * blk_mq_{start,stop}_hw_queue().
1502 bool blk_mq_queue_stopped(struct request_queue *q)
1504 struct blk_mq_hw_ctx *hctx;
1507 queue_for_each_hw_ctx(q, hctx, i)
1508 if (blk_mq_hctx_stopped(hctx))
1513 EXPORT_SYMBOL(blk_mq_queue_stopped);
1516 * This function is often used for pausing .queue_rq() by driver when
1517 * there isn't enough resource or some conditions aren't satisfied, and
1518 * BLK_STS_RESOURCE is usually returned.
1520 * We do not guarantee that dispatch can be drained or blocked
1521 * after blk_mq_stop_hw_queue() returns. Please use
1522 * blk_mq_quiesce_queue() for that requirement.
1524 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1526 cancel_delayed_work(&hctx->run_work);
1528 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1530 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1533 * This function is often used for pausing .queue_rq() by driver when
1534 * there isn't enough resource or some conditions aren't satisfied, and
1535 * BLK_STS_RESOURCE is usually returned.
1537 * We do not guarantee that dispatch can be drained or blocked
1538 * after blk_mq_stop_hw_queues() returns. Please use
1539 * blk_mq_quiesce_queue() for that requirement.
1541 void blk_mq_stop_hw_queues(struct request_queue *q)
1543 struct blk_mq_hw_ctx *hctx;
1546 queue_for_each_hw_ctx(q, hctx, i)
1547 blk_mq_stop_hw_queue(hctx);
1549 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1551 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1553 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1555 blk_mq_run_hw_queue(hctx, false);
1557 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1559 void blk_mq_start_hw_queues(struct request_queue *q)
1561 struct blk_mq_hw_ctx *hctx;
1564 queue_for_each_hw_ctx(q, hctx, i)
1565 blk_mq_start_hw_queue(hctx);
1567 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1569 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1571 if (!blk_mq_hctx_stopped(hctx))
1574 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1575 blk_mq_run_hw_queue(hctx, async);
1577 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1579 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1581 struct blk_mq_hw_ctx *hctx;
1584 queue_for_each_hw_ctx(q, hctx, i)
1585 blk_mq_start_stopped_hw_queue(hctx, async);
1587 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1589 static void blk_mq_run_work_fn(struct work_struct *work)
1591 struct blk_mq_hw_ctx *hctx;
1593 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1596 * If we are stopped, don't run the queue.
1598 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1601 __blk_mq_run_hw_queue(hctx);
1604 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1608 struct blk_mq_ctx *ctx = rq->mq_ctx;
1609 enum hctx_type type = hctx->type;
1611 lockdep_assert_held(&ctx->lock);
1613 trace_block_rq_insert(hctx->queue, rq);
1616 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1618 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1621 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1624 struct blk_mq_ctx *ctx = rq->mq_ctx;
1626 lockdep_assert_held(&ctx->lock);
1628 __blk_mq_insert_req_list(hctx, rq, at_head);
1629 blk_mq_hctx_mark_pending(hctx, ctx);
1633 * Should only be used carefully, when the caller knows we want to
1634 * bypass a potential IO scheduler on the target device.
1636 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1638 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1640 spin_lock(&hctx->lock);
1641 list_add_tail(&rq->queuelist, &hctx->dispatch);
1642 spin_unlock(&hctx->lock);
1645 blk_mq_run_hw_queue(hctx, false);
1648 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1649 struct list_head *list)
1653 enum hctx_type type = hctx->type;
1656 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1659 list_for_each_entry(rq, list, queuelist) {
1660 BUG_ON(rq->mq_ctx != ctx);
1661 trace_block_rq_insert(hctx->queue, rq);
1664 spin_lock(&ctx->lock);
1665 list_splice_tail_init(list, &ctx->rq_lists[type]);
1666 blk_mq_hctx_mark_pending(hctx, ctx);
1667 spin_unlock(&ctx->lock);
1670 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1672 struct request *rqa = container_of(a, struct request, queuelist);
1673 struct request *rqb = container_of(b, struct request, queuelist);
1675 if (rqa->mq_ctx < rqb->mq_ctx)
1677 else if (rqa->mq_ctx > rqb->mq_ctx)
1679 else if (rqa->mq_hctx < rqb->mq_hctx)
1681 else if (rqa->mq_hctx > rqb->mq_hctx)
1684 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1687 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1689 struct blk_mq_hw_ctx *this_hctx;
1690 struct blk_mq_ctx *this_ctx;
1691 struct request_queue *this_q;
1697 list_splice_init(&plug->mq_list, &list);
1700 if (plug->rq_count > 2 && plug->multiple_queues)
1701 list_sort(NULL, &list, plug_rq_cmp);
1708 while (!list_empty(&list)) {
1709 rq = list_entry_rq(list.next);
1710 list_del_init(&rq->queuelist);
1712 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1714 trace_block_unplug(this_q, depth, !from_schedule);
1715 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1721 this_ctx = rq->mq_ctx;
1722 this_hctx = rq->mq_hctx;
1727 list_add_tail(&rq->queuelist, &rq_list);
1731 * If 'this_hctx' is set, we know we have entries to complete
1732 * on 'rq_list'. Do those.
1735 trace_block_unplug(this_q, depth, !from_schedule);
1736 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1741 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1743 blk_init_request_from_bio(rq, bio);
1745 blk_account_io_start(rq, true);
1748 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1750 blk_qc_t *cookie, bool last)
1752 struct request_queue *q = rq->q;
1753 struct blk_mq_queue_data bd = {
1757 blk_qc_t new_cookie;
1760 new_cookie = request_to_qc_t(hctx, rq);
1763 * For OK queue, we are done. For error, caller may kill it.
1764 * Any other error (busy), just add it to our list as we
1765 * previously would have done.
1767 ret = q->mq_ops->queue_rq(hctx, &bd);
1770 blk_mq_update_dispatch_busy(hctx, false);
1771 *cookie = new_cookie;
1773 case BLK_STS_RESOURCE:
1774 case BLK_STS_DEV_RESOURCE:
1775 blk_mq_update_dispatch_busy(hctx, true);
1776 __blk_mq_requeue_request(rq);
1779 blk_mq_update_dispatch_busy(hctx, false);
1780 *cookie = BLK_QC_T_NONE;
1787 blk_status_t blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1790 bool bypass, bool last)
1792 struct request_queue *q = rq->q;
1793 bool run_queue = true;
1794 blk_status_t ret = BLK_STS_RESOURCE;
1798 hctx_lock(hctx, &srcu_idx);
1800 * hctx_lock is needed before checking quiesced flag.
1802 * When queue is stopped or quiesced, ignore 'bypass', insert
1803 * and return BLK_STS_OK to caller, and avoid driver to try to
1806 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) {
1812 if (unlikely(q->elevator && !bypass))
1815 if (!blk_mq_get_dispatch_budget(hctx))
1818 if (!blk_mq_get_driver_tag(rq)) {
1819 blk_mq_put_dispatch_budget(hctx);
1824 * Always add a request that has been through
1825 *.queue_rq() to the hardware dispatch list.
1828 ret = __blk_mq_issue_directly(hctx, rq, cookie, last);
1830 hctx_unlock(hctx, srcu_idx);
1834 case BLK_STS_DEV_RESOURCE:
1835 case BLK_STS_RESOURCE:
1837 blk_mq_request_bypass_insert(rq, run_queue);
1839 * We have to return BLK_STS_OK for the DM
1840 * to avoid livelock. Otherwise, we return
1841 * the real result to indicate whether the
1842 * request is direct-issued successfully.
1844 ret = bypass ? BLK_STS_OK : ret;
1845 } else if (!bypass) {
1846 blk_mq_sched_insert_request(rq, false,
1852 blk_mq_end_request(rq, ret);
1859 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1860 struct list_head *list)
1863 blk_status_t ret = BLK_STS_OK;
1865 while (!list_empty(list)) {
1866 struct request *rq = list_first_entry(list, struct request,
1869 list_del_init(&rq->queuelist);
1870 if (ret == BLK_STS_OK)
1871 ret = blk_mq_try_issue_directly(hctx, rq, &unused,
1875 blk_mq_sched_insert_request(rq, false, true, false);
1879 * If we didn't flush the entire list, we could have told
1880 * the driver there was more coming, but that turned out to
1883 if (ret != BLK_STS_OK && hctx->queue->mq_ops->commit_rqs)
1884 hctx->queue->mq_ops->commit_rqs(hctx);
1887 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1889 list_add_tail(&rq->queuelist, &plug->mq_list);
1891 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1892 struct request *tmp;
1894 tmp = list_first_entry(&plug->mq_list, struct request,
1896 if (tmp->q != rq->q)
1897 plug->multiple_queues = true;
1901 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1903 const int is_sync = op_is_sync(bio->bi_opf);
1904 const int is_flush_fua = op_is_flush(bio->bi_opf);
1905 struct blk_mq_alloc_data data = { .flags = 0, .cmd_flags = bio->bi_opf };
1907 struct blk_plug *plug;
1908 struct request *same_queue_rq = NULL;
1911 blk_queue_bounce(q, &bio);
1913 blk_queue_split(q, &bio);
1915 if (!bio_integrity_prep(bio))
1916 return BLK_QC_T_NONE;
1918 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1919 blk_attempt_plug_merge(q, bio, &same_queue_rq))
1920 return BLK_QC_T_NONE;
1922 if (blk_mq_sched_bio_merge(q, bio))
1923 return BLK_QC_T_NONE;
1925 rq_qos_throttle(q, bio);
1927 rq = blk_mq_get_request(q, bio, &data);
1928 if (unlikely(!rq)) {
1929 rq_qos_cleanup(q, bio);
1930 if (bio->bi_opf & REQ_NOWAIT)
1931 bio_wouldblock_error(bio);
1932 return BLK_QC_T_NONE;
1935 trace_block_getrq(q, bio, bio->bi_opf);
1937 rq_qos_track(q, rq, bio);
1939 cookie = request_to_qc_t(data.hctx, rq);
1941 plug = current->plug;
1942 if (unlikely(is_flush_fua)) {
1943 blk_mq_put_ctx(data.ctx);
1944 blk_mq_bio_to_request(rq, bio);
1946 /* bypass scheduler for flush rq */
1947 blk_insert_flush(rq);
1948 blk_mq_run_hw_queue(data.hctx, true);
1949 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1951 * Use plugging if we have a ->commit_rqs() hook as well, as
1952 * we know the driver uses bd->last in a smart fashion.
1954 unsigned int request_count = plug->rq_count;
1955 struct request *last = NULL;
1957 blk_mq_put_ctx(data.ctx);
1958 blk_mq_bio_to_request(rq, bio);
1961 trace_block_plug(q);
1963 last = list_entry_rq(plug->mq_list.prev);
1965 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1966 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1967 blk_flush_plug_list(plug, false);
1968 trace_block_plug(q);
1971 blk_add_rq_to_plug(plug, rq);
1972 } else if (plug && !blk_queue_nomerges(q)) {
1973 blk_mq_bio_to_request(rq, bio);
1976 * We do limited plugging. If the bio can be merged, do that.
1977 * Otherwise the existing request in the plug list will be
1978 * issued. So the plug list will have one request at most
1979 * The plug list might get flushed before this. If that happens,
1980 * the plug list is empty, and same_queue_rq is invalid.
1982 if (list_empty(&plug->mq_list))
1983 same_queue_rq = NULL;
1984 if (same_queue_rq) {
1985 list_del_init(&same_queue_rq->queuelist);
1988 blk_add_rq_to_plug(plug, rq);
1990 blk_mq_put_ctx(data.ctx);
1992 if (same_queue_rq) {
1993 data.hctx = same_queue_rq->mq_hctx;
1994 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1995 &cookie, false, true);
1997 } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
1998 !data.hctx->dispatch_busy)) {
1999 blk_mq_put_ctx(data.ctx);
2000 blk_mq_bio_to_request(rq, bio);
2001 blk_mq_try_issue_directly(data.hctx, rq, &cookie, false, true);
2003 blk_mq_put_ctx(data.ctx);
2004 blk_mq_bio_to_request(rq, bio);
2005 blk_mq_sched_insert_request(rq, false, true, true);
2011 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2012 unsigned int hctx_idx)
2016 if (tags->rqs && set->ops->exit_request) {
2019 for (i = 0; i < tags->nr_tags; i++) {
2020 struct request *rq = tags->static_rqs[i];
2024 set->ops->exit_request(set, rq, hctx_idx);
2025 tags->static_rqs[i] = NULL;
2029 while (!list_empty(&tags->page_list)) {
2030 page = list_first_entry(&tags->page_list, struct page, lru);
2031 list_del_init(&page->lru);
2033 * Remove kmemleak object previously allocated in
2034 * blk_mq_init_rq_map().
2036 kmemleak_free(page_address(page));
2037 __free_pages(page, page->private);
2041 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2045 kfree(tags->static_rqs);
2046 tags->static_rqs = NULL;
2048 blk_mq_free_tags(tags);
2051 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2052 unsigned int hctx_idx,
2053 unsigned int nr_tags,
2054 unsigned int reserved_tags)
2056 struct blk_mq_tags *tags;
2059 node = blk_mq_hw_queue_to_node(&set->map[0], hctx_idx);
2060 if (node == NUMA_NO_NODE)
2061 node = set->numa_node;
2063 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2064 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2068 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2069 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2072 blk_mq_free_tags(tags);
2076 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2077 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2079 if (!tags->static_rqs) {
2081 blk_mq_free_tags(tags);
2088 static size_t order_to_size(unsigned int order)
2090 return (size_t)PAGE_SIZE << order;
2093 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2094 unsigned int hctx_idx, int node)
2098 if (set->ops->init_request) {
2099 ret = set->ops->init_request(set, rq, hctx_idx, node);
2104 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2108 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2109 unsigned int hctx_idx, unsigned int depth)
2111 unsigned int i, j, entries_per_page, max_order = 4;
2112 size_t rq_size, left;
2115 node = blk_mq_hw_queue_to_node(&set->map[0], hctx_idx);
2116 if (node == NUMA_NO_NODE)
2117 node = set->numa_node;
2119 INIT_LIST_HEAD(&tags->page_list);
2122 * rq_size is the size of the request plus driver payload, rounded
2123 * to the cacheline size
2125 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2127 left = rq_size * depth;
2129 for (i = 0; i < depth; ) {
2130 int this_order = max_order;
2135 while (this_order && left < order_to_size(this_order - 1))
2139 page = alloc_pages_node(node,
2140 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2146 if (order_to_size(this_order) < rq_size)
2153 page->private = this_order;
2154 list_add_tail(&page->lru, &tags->page_list);
2156 p = page_address(page);
2158 * Allow kmemleak to scan these pages as they contain pointers
2159 * to additional allocations like via ops->init_request().
2161 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2162 entries_per_page = order_to_size(this_order) / rq_size;
2163 to_do = min(entries_per_page, depth - i);
2164 left -= to_do * rq_size;
2165 for (j = 0; j < to_do; j++) {
2166 struct request *rq = p;
2168 tags->static_rqs[i] = rq;
2169 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2170 tags->static_rqs[i] = NULL;
2181 blk_mq_free_rqs(set, tags, hctx_idx);
2186 * 'cpu' is going away. splice any existing rq_list entries from this
2187 * software queue to the hw queue dispatch list, and ensure that it
2190 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2192 struct blk_mq_hw_ctx *hctx;
2193 struct blk_mq_ctx *ctx;
2195 enum hctx_type type;
2197 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2198 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2201 spin_lock(&ctx->lock);
2202 if (!list_empty(&ctx->rq_lists[type])) {
2203 list_splice_init(&ctx->rq_lists[type], &tmp);
2204 blk_mq_hctx_clear_pending(hctx, ctx);
2206 spin_unlock(&ctx->lock);
2208 if (list_empty(&tmp))
2211 spin_lock(&hctx->lock);
2212 list_splice_tail_init(&tmp, &hctx->dispatch);
2213 spin_unlock(&hctx->lock);
2215 blk_mq_run_hw_queue(hctx, true);
2219 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2221 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2225 /* hctx->ctxs will be freed in queue's release handler */
2226 static void blk_mq_exit_hctx(struct request_queue *q,
2227 struct blk_mq_tag_set *set,
2228 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2230 if (blk_mq_hw_queue_mapped(hctx))
2231 blk_mq_tag_idle(hctx);
2233 if (set->ops->exit_request)
2234 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2236 if (set->ops->exit_hctx)
2237 set->ops->exit_hctx(hctx, hctx_idx);
2239 if (hctx->flags & BLK_MQ_F_BLOCKING)
2240 cleanup_srcu_struct(hctx->srcu);
2242 blk_mq_remove_cpuhp(hctx);
2243 blk_free_flush_queue(hctx->fq);
2244 sbitmap_free(&hctx->ctx_map);
2247 static void blk_mq_exit_hw_queues(struct request_queue *q,
2248 struct blk_mq_tag_set *set, int nr_queue)
2250 struct blk_mq_hw_ctx *hctx;
2253 queue_for_each_hw_ctx(q, hctx, i) {
2256 blk_mq_debugfs_unregister_hctx(hctx);
2257 blk_mq_exit_hctx(q, set, hctx, i);
2261 static int blk_mq_init_hctx(struct request_queue *q,
2262 struct blk_mq_tag_set *set,
2263 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2267 node = hctx->numa_node;
2268 if (node == NUMA_NO_NODE)
2269 node = hctx->numa_node = set->numa_node;
2271 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2272 spin_lock_init(&hctx->lock);
2273 INIT_LIST_HEAD(&hctx->dispatch);
2275 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2277 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2279 hctx->tags = set->tags[hctx_idx];
2282 * Allocate space for all possible cpus to avoid allocation at
2285 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2286 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node);
2288 goto unregister_cpu_notifier;
2290 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2291 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node))
2296 spin_lock_init(&hctx->dispatch_wait_lock);
2297 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2298 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2300 if (set->ops->init_hctx &&
2301 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2304 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2305 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
2309 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2312 if (hctx->flags & BLK_MQ_F_BLOCKING)
2313 init_srcu_struct(hctx->srcu);
2320 if (set->ops->exit_hctx)
2321 set->ops->exit_hctx(hctx, hctx_idx);
2323 sbitmap_free(&hctx->ctx_map);
2326 unregister_cpu_notifier:
2327 blk_mq_remove_cpuhp(hctx);
2331 static void blk_mq_init_cpu_queues(struct request_queue *q,
2332 unsigned int nr_hw_queues)
2334 struct blk_mq_tag_set *set = q->tag_set;
2337 for_each_possible_cpu(i) {
2338 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2339 struct blk_mq_hw_ctx *hctx;
2343 spin_lock_init(&__ctx->lock);
2344 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2345 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2350 * Set local node, IFF we have more than one hw queue. If
2351 * not, we remain on the home node of the device
2353 for (j = 0; j < set->nr_maps; j++) {
2354 hctx = blk_mq_map_queue_type(q, j, i);
2355 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2356 hctx->numa_node = local_memory_node(cpu_to_node(i));
2361 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2365 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2366 set->queue_depth, set->reserved_tags);
2367 if (!set->tags[hctx_idx])
2370 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2375 blk_mq_free_rq_map(set->tags[hctx_idx]);
2376 set->tags[hctx_idx] = NULL;
2380 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2381 unsigned int hctx_idx)
2383 if (set->tags && set->tags[hctx_idx]) {
2384 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2385 blk_mq_free_rq_map(set->tags[hctx_idx]);
2386 set->tags[hctx_idx] = NULL;
2390 static void blk_mq_map_swqueue(struct request_queue *q)
2392 unsigned int i, j, hctx_idx;
2393 struct blk_mq_hw_ctx *hctx;
2394 struct blk_mq_ctx *ctx;
2395 struct blk_mq_tag_set *set = q->tag_set;
2398 * Avoid others reading imcomplete hctx->cpumask through sysfs
2400 mutex_lock(&q->sysfs_lock);
2402 queue_for_each_hw_ctx(q, hctx, i) {
2403 cpumask_clear(hctx->cpumask);
2405 hctx->dispatch_from = NULL;
2409 * Map software to hardware queues.
2411 * If the cpu isn't present, the cpu is mapped to first hctx.
2413 for_each_possible_cpu(i) {
2414 hctx_idx = set->map[0].mq_map[i];
2415 /* unmapped hw queue can be remapped after CPU topo changed */
2416 if (!set->tags[hctx_idx] &&
2417 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2419 * If tags initialization fail for some hctx,
2420 * that hctx won't be brought online. In this
2421 * case, remap the current ctx to hctx[0] which
2422 * is guaranteed to always have tags allocated
2424 set->map[0].mq_map[i] = 0;
2427 ctx = per_cpu_ptr(q->queue_ctx, i);
2428 for (j = 0; j < set->nr_maps; j++) {
2429 if (!set->map[j].nr_queues)
2432 hctx = blk_mq_map_queue_type(q, j, i);
2435 * If the CPU is already set in the mask, then we've
2436 * mapped this one already. This can happen if
2437 * devices share queues across queue maps.
2439 if (cpumask_test_cpu(i, hctx->cpumask))
2442 cpumask_set_cpu(i, hctx->cpumask);
2444 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2445 hctx->ctxs[hctx->nr_ctx++] = ctx;
2448 * If the nr_ctx type overflows, we have exceeded the
2449 * amount of sw queues we can support.
2451 BUG_ON(!hctx->nr_ctx);
2455 mutex_unlock(&q->sysfs_lock);
2457 queue_for_each_hw_ctx(q, hctx, i) {
2459 * If no software queues are mapped to this hardware queue,
2460 * disable it and free the request entries.
2462 if (!hctx->nr_ctx) {
2463 /* Never unmap queue 0. We need it as a
2464 * fallback in case of a new remap fails
2467 if (i && set->tags[i])
2468 blk_mq_free_map_and_requests(set, i);
2474 hctx->tags = set->tags[i];
2475 WARN_ON(!hctx->tags);
2478 * Set the map size to the number of mapped software queues.
2479 * This is more accurate and more efficient than looping
2480 * over all possibly mapped software queues.
2482 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2485 * Initialize batch roundrobin counts
2487 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2488 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2493 * Caller needs to ensure that we're either frozen/quiesced, or that
2494 * the queue isn't live yet.
2496 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2498 struct blk_mq_hw_ctx *hctx;
2501 queue_for_each_hw_ctx(q, hctx, i) {
2503 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2505 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2509 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2512 struct request_queue *q;
2514 lockdep_assert_held(&set->tag_list_lock);
2516 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2517 blk_mq_freeze_queue(q);
2518 queue_set_hctx_shared(q, shared);
2519 blk_mq_unfreeze_queue(q);
2523 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2525 struct blk_mq_tag_set *set = q->tag_set;
2527 mutex_lock(&set->tag_list_lock);
2528 list_del_rcu(&q->tag_set_list);
2529 if (list_is_singular(&set->tag_list)) {
2530 /* just transitioned to unshared */
2531 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2532 /* update existing queue */
2533 blk_mq_update_tag_set_depth(set, false);
2535 mutex_unlock(&set->tag_list_lock);
2536 INIT_LIST_HEAD(&q->tag_set_list);
2539 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2540 struct request_queue *q)
2542 mutex_lock(&set->tag_list_lock);
2545 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2547 if (!list_empty(&set->tag_list) &&
2548 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2549 set->flags |= BLK_MQ_F_TAG_SHARED;
2550 /* update existing queue */
2551 blk_mq_update_tag_set_depth(set, true);
2553 if (set->flags & BLK_MQ_F_TAG_SHARED)
2554 queue_set_hctx_shared(q, true);
2555 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2557 mutex_unlock(&set->tag_list_lock);
2560 /* All allocations will be freed in release handler of q->mq_kobj */
2561 static int blk_mq_alloc_ctxs(struct request_queue *q)
2563 struct blk_mq_ctxs *ctxs;
2566 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2570 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2571 if (!ctxs->queue_ctx)
2574 for_each_possible_cpu(cpu) {
2575 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2579 q->mq_kobj = &ctxs->kobj;
2580 q->queue_ctx = ctxs->queue_ctx;
2589 * It is the actual release handler for mq, but we do it from
2590 * request queue's release handler for avoiding use-after-free
2591 * and headache because q->mq_kobj shouldn't have been introduced,
2592 * but we can't group ctx/kctx kobj without it.
2594 void blk_mq_release(struct request_queue *q)
2596 struct blk_mq_hw_ctx *hctx;
2599 /* hctx kobj stays in hctx */
2600 queue_for_each_hw_ctx(q, hctx, i) {
2603 kobject_put(&hctx->kobj);
2606 kfree(q->queue_hw_ctx);
2609 * release .mq_kobj and sw queue's kobject now because
2610 * both share lifetime with request queue.
2612 blk_mq_sysfs_deinit(q);
2615 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2617 struct request_queue *uninit_q, *q;
2619 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2621 return ERR_PTR(-ENOMEM);
2623 q = blk_mq_init_allocated_queue(set, uninit_q);
2625 blk_cleanup_queue(uninit_q);
2629 EXPORT_SYMBOL(blk_mq_init_queue);
2632 * Helper for setting up a queue with mq ops, given queue depth, and
2633 * the passed in mq ops flags.
2635 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2636 const struct blk_mq_ops *ops,
2637 unsigned int queue_depth,
2638 unsigned int set_flags)
2640 struct request_queue *q;
2643 memset(set, 0, sizeof(*set));
2645 set->nr_hw_queues = 1;
2647 set->queue_depth = queue_depth;
2648 set->numa_node = NUMA_NO_NODE;
2649 set->flags = set_flags;
2651 ret = blk_mq_alloc_tag_set(set);
2653 return ERR_PTR(ret);
2655 q = blk_mq_init_queue(set);
2657 blk_mq_free_tag_set(set);
2663 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2665 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2667 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2669 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2670 __alignof__(struct blk_mq_hw_ctx)) !=
2671 sizeof(struct blk_mq_hw_ctx));
2673 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2674 hw_ctx_size += sizeof(struct srcu_struct);
2679 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2680 struct blk_mq_tag_set *set, struct request_queue *q,
2681 int hctx_idx, int node)
2683 struct blk_mq_hw_ctx *hctx;
2685 hctx = kzalloc_node(blk_mq_hw_ctx_size(set),
2686 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2691 if (!zalloc_cpumask_var_node(&hctx->cpumask,
2692 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2698 atomic_set(&hctx->nr_active, 0);
2699 hctx->numa_node = node;
2700 hctx->queue_num = hctx_idx;
2702 if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) {
2703 free_cpumask_var(hctx->cpumask);
2707 blk_mq_hctx_kobj_init(hctx);
2712 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2713 struct request_queue *q)
2716 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2718 /* protect against switching io scheduler */
2719 mutex_lock(&q->sysfs_lock);
2720 for (i = 0; i < set->nr_hw_queues; i++) {
2722 struct blk_mq_hw_ctx *hctx;
2724 node = blk_mq_hw_queue_to_node(&set->map[0], i);
2726 * If the hw queue has been mapped to another numa node,
2727 * we need to realloc the hctx. If allocation fails, fallback
2728 * to use the previous one.
2730 if (hctxs[i] && (hctxs[i]->numa_node == node))
2733 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2736 blk_mq_exit_hctx(q, set, hctxs[i], i);
2737 kobject_put(&hctxs[i]->kobj);
2742 pr_warn("Allocate new hctx on node %d fails,\
2743 fallback to previous one on node %d\n",
2744 node, hctxs[i]->numa_node);
2750 * Increasing nr_hw_queues fails. Free the newly allocated
2751 * hctxs and keep the previous q->nr_hw_queues.
2753 if (i != set->nr_hw_queues) {
2754 j = q->nr_hw_queues;
2758 end = q->nr_hw_queues;
2759 q->nr_hw_queues = set->nr_hw_queues;
2762 for (; j < end; j++) {
2763 struct blk_mq_hw_ctx *hctx = hctxs[j];
2767 blk_mq_free_map_and_requests(set, j);
2768 blk_mq_exit_hctx(q, set, hctx, j);
2769 kobject_put(&hctx->kobj);
2774 mutex_unlock(&q->sysfs_lock);
2778 * Maximum number of hardware queues we support. For single sets, we'll never
2779 * have more than the CPUs (software queues). For multiple sets, the tag_set
2780 * user may have set ->nr_hw_queues larger.
2782 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2784 if (set->nr_maps == 1)
2787 return max(set->nr_hw_queues, nr_cpu_ids);
2790 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2791 struct request_queue *q)
2793 /* mark the queue as mq asap */
2794 q->mq_ops = set->ops;
2796 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2797 blk_mq_poll_stats_bkt,
2798 BLK_MQ_POLL_STATS_BKTS, q);
2802 if (blk_mq_alloc_ctxs(q))
2805 /* init q->mq_kobj and sw queues' kobjects */
2806 blk_mq_sysfs_init(q);
2808 q->nr_queues = nr_hw_queues(set);
2809 q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2810 GFP_KERNEL, set->numa_node);
2811 if (!q->queue_hw_ctx)
2814 blk_mq_realloc_hw_ctxs(set, q);
2815 if (!q->nr_hw_queues)
2818 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2819 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2823 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2824 if (set->nr_maps > HCTX_TYPE_POLL &&
2825 set->map[HCTX_TYPE_POLL].nr_queues)
2826 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2828 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2829 blk_queue_flag_set(QUEUE_FLAG_NO_SG_MERGE, q);
2831 q->sg_reserved_size = INT_MAX;
2833 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2834 INIT_LIST_HEAD(&q->requeue_list);
2835 spin_lock_init(&q->requeue_lock);
2837 blk_queue_make_request(q, blk_mq_make_request);
2840 * Do this after blk_queue_make_request() overrides it...
2842 q->nr_requests = set->queue_depth;
2845 * Default to classic polling
2849 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2850 blk_mq_add_queue_tag_set(set, q);
2851 blk_mq_map_swqueue(q);
2853 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2856 ret = elevator_init_mq(q);
2858 return ERR_PTR(ret);
2864 kfree(q->queue_hw_ctx);
2866 blk_mq_sysfs_deinit(q);
2869 return ERR_PTR(-ENOMEM);
2871 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2873 void blk_mq_free_queue(struct request_queue *q)
2875 struct blk_mq_tag_set *set = q->tag_set;
2877 blk_mq_del_queue_tag_set(q);
2878 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2881 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2885 for (i = 0; i < set->nr_hw_queues; i++)
2886 if (!__blk_mq_alloc_rq_map(set, i))
2893 blk_mq_free_rq_map(set->tags[i]);
2899 * Allocate the request maps associated with this tag_set. Note that this
2900 * may reduce the depth asked for, if memory is tight. set->queue_depth
2901 * will be updated to reflect the allocated depth.
2903 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2908 depth = set->queue_depth;
2910 err = __blk_mq_alloc_rq_maps(set);
2914 set->queue_depth >>= 1;
2915 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2919 } while (set->queue_depth);
2921 if (!set->queue_depth || err) {
2922 pr_err("blk-mq: failed to allocate request map\n");
2926 if (depth != set->queue_depth)
2927 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2928 depth, set->queue_depth);
2933 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2935 if (set->ops->map_queues && !is_kdump_kernel()) {
2939 * transport .map_queues is usually done in the following
2942 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2943 * mask = get_cpu_mask(queue)
2944 * for_each_cpu(cpu, mask)
2945 * set->map[x].mq_map[cpu] = queue;
2948 * When we need to remap, the table has to be cleared for
2949 * killing stale mapping since one CPU may not be mapped
2952 for (i = 0; i < set->nr_maps; i++)
2953 blk_mq_clear_mq_map(&set->map[i]);
2955 return set->ops->map_queues(set);
2957 BUG_ON(set->nr_maps > 1);
2958 return blk_mq_map_queues(&set->map[0]);
2963 * Alloc a tag set to be associated with one or more request queues.
2964 * May fail with EINVAL for various error conditions. May adjust the
2965 * requested depth down, if it's too large. In that case, the set
2966 * value will be stored in set->queue_depth.
2968 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2972 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2974 if (!set->nr_hw_queues)
2976 if (!set->queue_depth)
2978 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2981 if (!set->ops->queue_rq)
2984 if (!set->ops->get_budget ^ !set->ops->put_budget)
2987 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2988 pr_info("blk-mq: reduced tag depth to %u\n",
2990 set->queue_depth = BLK_MQ_MAX_DEPTH;
2995 else if (set->nr_maps > HCTX_MAX_TYPES)
2999 * If a crashdump is active, then we are potentially in a very
3000 * memory constrained environment. Limit us to 1 queue and
3001 * 64 tags to prevent using too much memory.
3003 if (is_kdump_kernel()) {
3004 set->nr_hw_queues = 1;
3006 set->queue_depth = min(64U, set->queue_depth);
3009 * There is no use for more h/w queues than cpus if we just have
3012 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3013 set->nr_hw_queues = nr_cpu_ids;
3015 set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3016 GFP_KERNEL, set->numa_node);
3021 for (i = 0; i < set->nr_maps; i++) {
3022 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3023 sizeof(set->map[i].mq_map[0]),
3024 GFP_KERNEL, set->numa_node);
3025 if (!set->map[i].mq_map)
3026 goto out_free_mq_map;
3027 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3030 ret = blk_mq_update_queue_map(set);
3032 goto out_free_mq_map;
3034 ret = blk_mq_alloc_rq_maps(set);
3036 goto out_free_mq_map;
3038 mutex_init(&set->tag_list_lock);
3039 INIT_LIST_HEAD(&set->tag_list);
3044 for (i = 0; i < set->nr_maps; i++) {
3045 kfree(set->map[i].mq_map);
3046 set->map[i].mq_map = NULL;
3052 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3054 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3058 for (i = 0; i < nr_hw_queues(set); i++)
3059 blk_mq_free_map_and_requests(set, i);
3061 for (j = 0; j < set->nr_maps; j++) {
3062 kfree(set->map[j].mq_map);
3063 set->map[j].mq_map = NULL;
3069 EXPORT_SYMBOL(blk_mq_free_tag_set);
3071 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3073 struct blk_mq_tag_set *set = q->tag_set;
3074 struct blk_mq_hw_ctx *hctx;
3080 blk_mq_freeze_queue(q);
3081 blk_mq_quiesce_queue(q);
3084 queue_for_each_hw_ctx(q, hctx, i) {
3088 * If we're using an MQ scheduler, just update the scheduler
3089 * queue depth. This is similar to what the old code would do.
3091 if (!hctx->sched_tags) {
3092 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3095 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3103 q->nr_requests = nr;
3105 blk_mq_unquiesce_queue(q);
3106 blk_mq_unfreeze_queue(q);
3112 * request_queue and elevator_type pair.
3113 * It is just used by __blk_mq_update_nr_hw_queues to cache
3114 * the elevator_type associated with a request_queue.
3116 struct blk_mq_qe_pair {
3117 struct list_head node;
3118 struct request_queue *q;
3119 struct elevator_type *type;
3123 * Cache the elevator_type in qe pair list and switch the
3124 * io scheduler to 'none'
3126 static bool blk_mq_elv_switch_none(struct list_head *head,
3127 struct request_queue *q)
3129 struct blk_mq_qe_pair *qe;
3134 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3138 INIT_LIST_HEAD(&qe->node);
3140 qe->type = q->elevator->type;
3141 list_add(&qe->node, head);
3143 mutex_lock(&q->sysfs_lock);
3145 * After elevator_switch_mq, the previous elevator_queue will be
3146 * released by elevator_release. The reference of the io scheduler
3147 * module get by elevator_get will also be put. So we need to get
3148 * a reference of the io scheduler module here to prevent it to be
3151 __module_get(qe->type->elevator_owner);
3152 elevator_switch_mq(q, NULL);
3153 mutex_unlock(&q->sysfs_lock);
3158 static void blk_mq_elv_switch_back(struct list_head *head,
3159 struct request_queue *q)
3161 struct blk_mq_qe_pair *qe;
3162 struct elevator_type *t = NULL;
3164 list_for_each_entry(qe, head, node)
3173 list_del(&qe->node);
3176 mutex_lock(&q->sysfs_lock);
3177 elevator_switch_mq(q, t);
3178 mutex_unlock(&q->sysfs_lock);
3181 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3184 struct request_queue *q;
3186 int prev_nr_hw_queues;
3188 lockdep_assert_held(&set->tag_list_lock);
3190 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3191 nr_hw_queues = nr_cpu_ids;
3192 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3195 list_for_each_entry(q, &set->tag_list, tag_set_list)
3196 blk_mq_freeze_queue(q);
3198 * Sync with blk_mq_queue_tag_busy_iter.
3202 * Switch IO scheduler to 'none', cleaning up the data associated
3203 * with the previous scheduler. We will switch back once we are done
3204 * updating the new sw to hw queue mappings.
3206 list_for_each_entry(q, &set->tag_list, tag_set_list)
3207 if (!blk_mq_elv_switch_none(&head, q))
3210 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3211 blk_mq_debugfs_unregister_hctxs(q);
3212 blk_mq_sysfs_unregister(q);
3215 prev_nr_hw_queues = set->nr_hw_queues;
3216 set->nr_hw_queues = nr_hw_queues;
3217 blk_mq_update_queue_map(set);
3219 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3220 blk_mq_realloc_hw_ctxs(set, q);
3221 if (q->nr_hw_queues != set->nr_hw_queues) {
3222 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3223 nr_hw_queues, prev_nr_hw_queues);
3224 set->nr_hw_queues = prev_nr_hw_queues;
3225 blk_mq_map_queues(&set->map[0]);
3228 blk_mq_map_swqueue(q);
3231 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3232 blk_mq_sysfs_register(q);
3233 blk_mq_debugfs_register_hctxs(q);
3237 list_for_each_entry(q, &set->tag_list, tag_set_list)
3238 blk_mq_elv_switch_back(&head, q);
3240 list_for_each_entry(q, &set->tag_list, tag_set_list)
3241 blk_mq_unfreeze_queue(q);
3244 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3246 mutex_lock(&set->tag_list_lock);
3247 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3248 mutex_unlock(&set->tag_list_lock);
3250 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3252 /* Enable polling stats and return whether they were already enabled. */
3253 static bool blk_poll_stats_enable(struct request_queue *q)
3255 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3256 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3258 blk_stat_add_callback(q, q->poll_cb);
3262 static void blk_mq_poll_stats_start(struct request_queue *q)
3265 * We don't arm the callback if polling stats are not enabled or the
3266 * callback is already active.
3268 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3269 blk_stat_is_active(q->poll_cb))
3272 blk_stat_activate_msecs(q->poll_cb, 100);
3275 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3277 struct request_queue *q = cb->data;
3280 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3281 if (cb->stat[bucket].nr_samples)
3282 q->poll_stat[bucket] = cb->stat[bucket];
3286 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3287 struct blk_mq_hw_ctx *hctx,
3290 unsigned long ret = 0;
3294 * If stats collection isn't on, don't sleep but turn it on for
3297 if (!blk_poll_stats_enable(q))
3301 * As an optimistic guess, use half of the mean service time
3302 * for this type of request. We can (and should) make this smarter.
3303 * For instance, if the completion latencies are tight, we can
3304 * get closer than just half the mean. This is especially
3305 * important on devices where the completion latencies are longer
3306 * than ~10 usec. We do use the stats for the relevant IO size
3307 * if available which does lead to better estimates.
3309 bucket = blk_mq_poll_stats_bkt(rq);
3313 if (q->poll_stat[bucket].nr_samples)
3314 ret = (q->poll_stat[bucket].mean + 1) / 2;
3319 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3320 struct blk_mq_hw_ctx *hctx,
3323 struct hrtimer_sleeper hs;
3324 enum hrtimer_mode mode;
3328 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3332 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3334 * 0: use half of prev avg
3335 * >0: use this specific value
3337 if (q->poll_nsec > 0)
3338 nsecs = q->poll_nsec;
3340 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3345 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3348 * This will be replaced with the stats tracking code, using
3349 * 'avg_completion_time / 2' as the pre-sleep target.
3353 mode = HRTIMER_MODE_REL;
3354 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3355 hrtimer_set_expires(&hs.timer, kt);
3357 hrtimer_init_sleeper(&hs, current);
3359 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3361 set_current_state(TASK_UNINTERRUPTIBLE);
3362 hrtimer_start_expires(&hs.timer, mode);
3365 hrtimer_cancel(&hs.timer);
3366 mode = HRTIMER_MODE_ABS;
3367 } while (hs.task && !signal_pending(current));
3369 __set_current_state(TASK_RUNNING);
3370 destroy_hrtimer_on_stack(&hs.timer);
3374 static bool blk_mq_poll_hybrid(struct request_queue *q,
3375 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3379 if (q->poll_nsec == -1)
3382 if (!blk_qc_t_is_internal(cookie))
3383 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3385 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3387 * With scheduling, if the request has completed, we'll
3388 * get a NULL return here, as we clear the sched tag when
3389 * that happens. The request still remains valid, like always,
3390 * so we should be safe with just the NULL check.
3396 return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3400 * blk_poll - poll for IO completions
3402 * @cookie: cookie passed back at IO submission time
3403 * @spin: whether to spin for completions
3406 * Poll for completions on the passed in queue. Returns number of
3407 * completed entries found. If @spin is true, then blk_poll will continue
3408 * looping until at least one completion is found, unless the task is
3409 * otherwise marked running (or we need to reschedule).
3411 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3413 struct blk_mq_hw_ctx *hctx;
3416 if (!blk_qc_t_valid(cookie) ||
3417 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3421 blk_flush_plug_list(current->plug, false);
3423 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3426 * If we sleep, have the caller restart the poll loop to reset
3427 * the state. Like for the other success return cases, the
3428 * caller is responsible for checking if the IO completed. If
3429 * the IO isn't complete, we'll get called again and will go
3430 * straight to the busy poll loop.
3432 if (blk_mq_poll_hybrid(q, hctx, cookie))
3435 hctx->poll_considered++;
3437 state = current->state;
3441 hctx->poll_invoked++;
3443 ret = q->mq_ops->poll(hctx);
3445 hctx->poll_success++;
3446 __set_current_state(TASK_RUNNING);
3450 if (signal_pending_state(state, current))
3451 __set_current_state(TASK_RUNNING);
3453 if (current->state == TASK_RUNNING)
3455 if (ret < 0 || !spin)
3458 } while (!need_resched());
3460 __set_current_state(TASK_RUNNING);
3463 EXPORT_SYMBOL_GPL(blk_poll);
3465 unsigned int blk_mq_rq_cpu(struct request *rq)
3467 return rq->mq_ctx->cpu;
3469 EXPORT_SYMBOL(blk_mq_rq_cpu);
3471 static int __init blk_mq_init(void)
3473 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3474 blk_mq_hctx_notify_dead);
3477 subsys_initcall(blk_mq_init);