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 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie);
42 static void blk_mq_poll_stats_start(struct request_queue *q);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
45 static int blk_mq_poll_stats_bkt(const struct request *rq)
47 int ddir, bytes, bucket;
49 ddir = rq_data_dir(rq);
50 bytes = blk_rq_bytes(rq);
52 bucket = ddir + 2*(ilog2(bytes) - 9);
56 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
57 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
63 * Check if any of the ctx's have pending work in this hardware queue
65 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
67 return !list_empty_careful(&hctx->dispatch) ||
68 sbitmap_any_bit_set(&hctx->ctx_map) ||
69 blk_mq_sched_has_work(hctx);
73 * Mark this ctx as having pending work in this hardware queue
75 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
76 struct blk_mq_ctx *ctx)
78 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
79 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
82 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
83 struct blk_mq_ctx *ctx)
85 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
89 struct hd_struct *part;
90 unsigned int *inflight;
93 static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
94 struct request *rq, void *priv,
97 struct mq_inflight *mi = priv;
100 * index[0] counts the specific partition that was asked for. index[1]
101 * counts the ones that are active on the whole device, so increment
102 * that if mi->part is indeed a partition, and not a whole device.
104 if (rq->part == mi->part)
106 if (mi->part->partno)
110 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
111 unsigned int inflight[2])
113 struct mq_inflight mi = { .part = part, .inflight = inflight, };
115 inflight[0] = inflight[1] = 0;
116 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
119 static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
120 struct request *rq, void *priv,
123 struct mq_inflight *mi = priv;
125 if (rq->part == mi->part)
126 mi->inflight[rq_data_dir(rq)]++;
129 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
130 unsigned int inflight[2])
132 struct mq_inflight mi = { .part = part, .inflight = inflight, };
134 inflight[0] = inflight[1] = 0;
135 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
138 void blk_freeze_queue_start(struct request_queue *q)
142 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
143 if (freeze_depth == 1) {
144 percpu_ref_kill(&q->q_usage_counter);
146 blk_mq_run_hw_queues(q, false);
149 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
151 void blk_mq_freeze_queue_wait(struct request_queue *q)
153 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
155 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
157 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
158 unsigned long timeout)
160 return wait_event_timeout(q->mq_freeze_wq,
161 percpu_ref_is_zero(&q->q_usage_counter),
164 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
167 * Guarantee no request is in use, so we can change any data structure of
168 * the queue afterward.
170 void blk_freeze_queue(struct request_queue *q)
173 * In the !blk_mq case we are only calling this to kill the
174 * q_usage_counter, otherwise this increases the freeze depth
175 * and waits for it to return to zero. For this reason there is
176 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
177 * exported to drivers as the only user for unfreeze is blk_mq.
179 blk_freeze_queue_start(q);
182 blk_mq_freeze_queue_wait(q);
185 void blk_mq_freeze_queue(struct request_queue *q)
188 * ...just an alias to keep freeze and unfreeze actions balanced
189 * in the blk_mq_* namespace
193 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
195 void blk_mq_unfreeze_queue(struct request_queue *q)
199 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
200 WARN_ON_ONCE(freeze_depth < 0);
202 percpu_ref_resurrect(&q->q_usage_counter);
203 wake_up_all(&q->mq_freeze_wq);
206 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
209 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
210 * mpt3sas driver such that this function can be removed.
212 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
214 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
216 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
219 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
222 * Note: this function does not prevent that the struct request end_io()
223 * callback function is invoked. Once this function is returned, we make
224 * sure no dispatch can happen until the queue is unquiesced via
225 * blk_mq_unquiesce_queue().
227 void blk_mq_quiesce_queue(struct request_queue *q)
229 struct blk_mq_hw_ctx *hctx;
233 blk_mq_quiesce_queue_nowait(q);
235 queue_for_each_hw_ctx(q, hctx, i) {
236 if (hctx->flags & BLK_MQ_F_BLOCKING)
237 synchronize_srcu(hctx->srcu);
244 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
247 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
250 * This function recovers queue into the state before quiescing
251 * which is done by blk_mq_quiesce_queue.
253 void blk_mq_unquiesce_queue(struct request_queue *q)
255 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
257 /* dispatch requests which are inserted during quiescing */
258 blk_mq_run_hw_queues(q, true);
260 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
262 void blk_mq_wake_waiters(struct request_queue *q)
264 struct blk_mq_hw_ctx *hctx;
267 queue_for_each_hw_ctx(q, hctx, i)
268 if (blk_mq_hw_queue_mapped(hctx))
269 blk_mq_tag_wakeup_all(hctx->tags, true);
272 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
274 return blk_mq_has_free_tags(hctx->tags);
276 EXPORT_SYMBOL(blk_mq_can_queue);
278 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
279 unsigned int tag, unsigned int op)
281 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
282 struct request *rq = tags->static_rqs[tag];
283 req_flags_t rq_flags = 0;
285 if (data->flags & BLK_MQ_REQ_INTERNAL) {
287 rq->internal_tag = tag;
289 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
290 rq_flags = RQF_MQ_INFLIGHT;
291 atomic_inc(&data->hctx->nr_active);
294 rq->internal_tag = -1;
295 data->hctx->tags->rqs[rq->tag] = rq;
298 /* csd/requeue_work/fifo_time is initialized before use */
300 rq->mq_ctx = data->ctx;
301 rq->rq_flags = rq_flags;
304 if (data->flags & BLK_MQ_REQ_PREEMPT)
305 rq->rq_flags |= RQF_PREEMPT;
306 if (blk_queue_io_stat(data->q))
307 rq->rq_flags |= RQF_IO_STAT;
308 INIT_LIST_HEAD(&rq->queuelist);
309 INIT_HLIST_NODE(&rq->hash);
310 RB_CLEAR_NODE(&rq->rb_node);
313 rq->start_time_ns = ktime_get_ns();
314 rq->io_start_time_ns = 0;
315 rq->nr_phys_segments = 0;
316 #if defined(CONFIG_BLK_DEV_INTEGRITY)
317 rq->nr_integrity_segments = 0;
320 /* tag was already set */
324 INIT_LIST_HEAD(&rq->timeout_list);
328 rq->end_io_data = NULL;
331 #ifdef CONFIG_BLK_CGROUP
335 data->ctx->rq_dispatched[op_is_sync(op)]++;
336 refcount_set(&rq->ref, 1);
340 static struct request *blk_mq_get_request(struct request_queue *q,
341 struct bio *bio, unsigned int op,
342 struct blk_mq_alloc_data *data)
344 struct elevator_queue *e = q->elevator;
347 bool put_ctx_on_error = false;
349 blk_queue_enter_live(q);
351 if (likely(!data->ctx)) {
352 data->ctx = blk_mq_get_ctx(q);
353 put_ctx_on_error = true;
355 if (likely(!data->hctx))
356 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
358 data->flags |= BLK_MQ_REQ_NOWAIT;
361 data->flags |= BLK_MQ_REQ_INTERNAL;
364 * Flush requests are special and go directly to the
365 * dispatch list. Don't include reserved tags in the
366 * limiting, as it isn't useful.
368 if (!op_is_flush(op) && e->type->ops.mq.limit_depth &&
369 !(data->flags & BLK_MQ_REQ_RESERVED))
370 e->type->ops.mq.limit_depth(op, data);
372 blk_mq_tag_busy(data->hctx);
375 tag = blk_mq_get_tag(data);
376 if (tag == BLK_MQ_TAG_FAIL) {
377 if (put_ctx_on_error) {
378 blk_mq_put_ctx(data->ctx);
385 rq = blk_mq_rq_ctx_init(data, tag, op);
386 if (!op_is_flush(op)) {
388 if (e && e->type->ops.mq.prepare_request) {
389 if (e->type->icq_cache && rq_ioc(bio))
390 blk_mq_sched_assign_ioc(rq, bio);
392 e->type->ops.mq.prepare_request(rq, bio);
393 rq->rq_flags |= RQF_ELVPRIV;
396 data->hctx->queued++;
400 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
401 blk_mq_req_flags_t flags)
403 struct blk_mq_alloc_data alloc_data = { .flags = flags };
407 ret = blk_queue_enter(q, flags);
411 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
415 return ERR_PTR(-EWOULDBLOCK);
417 blk_mq_put_ctx(alloc_data.ctx);
420 rq->__sector = (sector_t) -1;
421 rq->bio = rq->biotail = NULL;
424 EXPORT_SYMBOL(blk_mq_alloc_request);
426 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
427 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
429 struct blk_mq_alloc_data alloc_data = { .flags = flags };
435 * If the tag allocator sleeps we could get an allocation for a
436 * different hardware context. No need to complicate the low level
437 * allocator for this for the rare use case of a command tied to
440 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
441 return ERR_PTR(-EINVAL);
443 if (hctx_idx >= q->nr_hw_queues)
444 return ERR_PTR(-EIO);
446 ret = blk_queue_enter(q, flags);
451 * Check if the hardware context is actually mapped to anything.
452 * If not tell the caller that it should skip this queue.
454 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
455 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
457 return ERR_PTR(-EXDEV);
459 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
460 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
462 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
466 return ERR_PTR(-EWOULDBLOCK);
470 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
472 static void __blk_mq_free_request(struct request *rq)
474 struct request_queue *q = rq->q;
475 struct blk_mq_ctx *ctx = rq->mq_ctx;
476 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
477 const int sched_tag = rq->internal_tag;
479 blk_pm_mark_last_busy(rq);
481 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
483 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
484 blk_mq_sched_restart(hctx);
488 void blk_mq_free_request(struct request *rq)
490 struct request_queue *q = rq->q;
491 struct elevator_queue *e = q->elevator;
492 struct blk_mq_ctx *ctx = rq->mq_ctx;
493 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
495 if (rq->rq_flags & RQF_ELVPRIV) {
496 if (e && e->type->ops.mq.finish_request)
497 e->type->ops.mq.finish_request(rq);
499 put_io_context(rq->elv.icq->ioc);
504 ctx->rq_completed[rq_is_sync(rq)]++;
505 if (rq->rq_flags & RQF_MQ_INFLIGHT)
506 atomic_dec(&hctx->nr_active);
508 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
509 laptop_io_completion(q->backing_dev_info);
514 blk_put_rl(blk_rq_rl(rq));
516 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
517 if (refcount_dec_and_test(&rq->ref))
518 __blk_mq_free_request(rq);
520 EXPORT_SYMBOL_GPL(blk_mq_free_request);
522 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
524 u64 now = ktime_get_ns();
526 if (rq->rq_flags & RQF_STATS) {
527 blk_mq_poll_stats_start(rq->q);
528 blk_stat_add(rq, now);
531 if (rq->internal_tag != -1)
532 blk_mq_sched_completed_request(rq, now);
534 blk_account_io_done(rq, now);
537 rq_qos_done(rq->q, rq);
538 rq->end_io(rq, error);
540 if (unlikely(blk_bidi_rq(rq)))
541 blk_mq_free_request(rq->next_rq);
542 blk_mq_free_request(rq);
545 EXPORT_SYMBOL(__blk_mq_end_request);
547 void blk_mq_end_request(struct request *rq, blk_status_t error)
549 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
551 __blk_mq_end_request(rq, error);
553 EXPORT_SYMBOL(blk_mq_end_request);
555 static void __blk_mq_complete_request_remote(void *data)
557 struct request *rq = data;
559 rq->q->softirq_done_fn(rq);
562 static void __blk_mq_complete_request(struct request *rq)
564 struct blk_mq_ctx *ctx = rq->mq_ctx;
568 if (!blk_mq_mark_complete(rq))
572 * Most of single queue controllers, there is only one irq vector
573 * for handling IO completion, and the only irq's affinity is set
574 * as all possible CPUs. On most of ARCHs, this affinity means the
575 * irq is handled on one specific CPU.
577 * So complete IO reqeust in softirq context in case of single queue
578 * for not degrading IO performance by irqsoff latency.
580 if (rq->q->nr_hw_queues == 1) {
581 __blk_complete_request(rq);
585 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
586 rq->q->softirq_done_fn(rq);
591 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
592 shared = cpus_share_cache(cpu, ctx->cpu);
594 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
595 rq->csd.func = __blk_mq_complete_request_remote;
598 smp_call_function_single_async(ctx->cpu, &rq->csd);
600 rq->q->softirq_done_fn(rq);
605 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
606 __releases(hctx->srcu)
608 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
611 srcu_read_unlock(hctx->srcu, srcu_idx);
614 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
615 __acquires(hctx->srcu)
617 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
618 /* shut up gcc false positive */
622 *srcu_idx = srcu_read_lock(hctx->srcu);
626 * blk_mq_complete_request - end I/O on a request
627 * @rq: the request being processed
630 * Ends all I/O on a request. It does not handle partial completions.
631 * The actual completion happens out-of-order, through a IPI handler.
633 void blk_mq_complete_request(struct request *rq)
635 if (unlikely(blk_should_fake_timeout(rq->q)))
637 __blk_mq_complete_request(rq);
639 EXPORT_SYMBOL(blk_mq_complete_request);
641 int blk_mq_request_started(struct request *rq)
643 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
645 EXPORT_SYMBOL_GPL(blk_mq_request_started);
647 void blk_mq_start_request(struct request *rq)
649 struct request_queue *q = rq->q;
651 blk_mq_sched_started_request(rq);
653 trace_block_rq_issue(q, rq);
655 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
656 rq->io_start_time_ns = ktime_get_ns();
657 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
658 rq->throtl_size = blk_rq_sectors(rq);
660 rq->rq_flags |= RQF_STATS;
664 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
667 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
669 if (q->dma_drain_size && blk_rq_bytes(rq)) {
671 * Make sure space for the drain appears. We know we can do
672 * this because max_hw_segments has been adjusted to be one
673 * fewer than the device can handle.
675 rq->nr_phys_segments++;
678 EXPORT_SYMBOL(blk_mq_start_request);
680 static void __blk_mq_requeue_request(struct request *rq)
682 struct request_queue *q = rq->q;
684 blk_mq_put_driver_tag(rq);
686 trace_block_rq_requeue(q, rq);
687 rq_qos_requeue(q, rq);
689 if (blk_mq_request_started(rq)) {
690 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
691 rq->rq_flags &= ~RQF_TIMED_OUT;
692 if (q->dma_drain_size && blk_rq_bytes(rq))
693 rq->nr_phys_segments--;
697 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
699 __blk_mq_requeue_request(rq);
701 /* this request will be re-inserted to io scheduler queue */
702 blk_mq_sched_requeue_request(rq);
704 BUG_ON(blk_queued_rq(rq));
705 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
707 EXPORT_SYMBOL(blk_mq_requeue_request);
709 static void blk_mq_requeue_work(struct work_struct *work)
711 struct request_queue *q =
712 container_of(work, struct request_queue, requeue_work.work);
714 struct request *rq, *next;
716 spin_lock_irq(&q->requeue_lock);
717 list_splice_init(&q->requeue_list, &rq_list);
718 spin_unlock_irq(&q->requeue_lock);
720 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
721 if (!(rq->rq_flags & RQF_SOFTBARRIER))
724 rq->rq_flags &= ~RQF_SOFTBARRIER;
725 list_del_init(&rq->queuelist);
726 blk_mq_sched_insert_request(rq, true, false, false);
729 while (!list_empty(&rq_list)) {
730 rq = list_entry(rq_list.next, struct request, queuelist);
731 list_del_init(&rq->queuelist);
732 blk_mq_sched_insert_request(rq, false, false, false);
735 blk_mq_run_hw_queues(q, false);
738 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
739 bool kick_requeue_list)
741 struct request_queue *q = rq->q;
745 * We abuse this flag that is otherwise used by the I/O scheduler to
746 * request head insertion from the workqueue.
748 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
750 spin_lock_irqsave(&q->requeue_lock, flags);
752 rq->rq_flags |= RQF_SOFTBARRIER;
753 list_add(&rq->queuelist, &q->requeue_list);
755 list_add_tail(&rq->queuelist, &q->requeue_list);
757 spin_unlock_irqrestore(&q->requeue_lock, flags);
759 if (kick_requeue_list)
760 blk_mq_kick_requeue_list(q);
762 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
764 void blk_mq_kick_requeue_list(struct request_queue *q)
766 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
768 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
770 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
773 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
774 msecs_to_jiffies(msecs));
776 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
778 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
780 if (tag < tags->nr_tags) {
781 prefetch(tags->rqs[tag]);
782 return tags->rqs[tag];
787 EXPORT_SYMBOL(blk_mq_tag_to_rq);
789 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
791 req->rq_flags |= RQF_TIMED_OUT;
792 if (req->q->mq_ops->timeout) {
793 enum blk_eh_timer_return ret;
795 ret = req->q->mq_ops->timeout(req, reserved);
796 if (ret == BLK_EH_DONE)
798 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
804 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
806 unsigned long deadline;
808 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
810 if (rq->rq_flags & RQF_TIMED_OUT)
813 deadline = blk_rq_deadline(rq);
814 if (time_after_eq(jiffies, deadline))
819 else if (time_after(*next, deadline))
824 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
825 struct request *rq, void *priv, bool reserved)
827 unsigned long *next = priv;
830 * Just do a quick check if it is expired before locking the request in
831 * so we're not unnecessarilly synchronizing across CPUs.
833 if (!blk_mq_req_expired(rq, next))
837 * We have reason to believe the request may be expired. Take a
838 * reference on the request to lock this request lifetime into its
839 * currently allocated context to prevent it from being reallocated in
840 * the event the completion by-passes this timeout handler.
842 * If the reference was already released, then the driver beat the
843 * timeout handler to posting a natural completion.
845 if (!refcount_inc_not_zero(&rq->ref))
849 * The request is now locked and cannot be reallocated underneath the
850 * timeout handler's processing. Re-verify this exact request is truly
851 * expired; if it is not expired, then the request was completed and
852 * reallocated as a new request.
854 if (blk_mq_req_expired(rq, next))
855 blk_mq_rq_timed_out(rq, reserved);
856 if (refcount_dec_and_test(&rq->ref))
857 __blk_mq_free_request(rq);
860 static void blk_mq_timeout_work(struct work_struct *work)
862 struct request_queue *q =
863 container_of(work, struct request_queue, timeout_work);
864 unsigned long next = 0;
865 struct blk_mq_hw_ctx *hctx;
868 /* A deadlock might occur if a request is stuck requiring a
869 * timeout at the same time a queue freeze is waiting
870 * completion, since the timeout code would not be able to
871 * acquire the queue reference here.
873 * That's why we don't use blk_queue_enter here; instead, we use
874 * percpu_ref_tryget directly, because we need to be able to
875 * obtain a reference even in the short window between the queue
876 * starting to freeze, by dropping the first reference in
877 * blk_freeze_queue_start, and the moment the last request is
878 * consumed, marked by the instant q_usage_counter reaches
881 if (!percpu_ref_tryget(&q->q_usage_counter))
884 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
887 mod_timer(&q->timeout, next);
890 * Request timeouts are handled as a forward rolling timer. If
891 * we end up here it means that no requests are pending and
892 * also that no request has been pending for a while. Mark
895 queue_for_each_hw_ctx(q, hctx, i) {
896 /* the hctx may be unmapped, so check it here */
897 if (blk_mq_hw_queue_mapped(hctx))
898 blk_mq_tag_idle(hctx);
904 struct flush_busy_ctx_data {
905 struct blk_mq_hw_ctx *hctx;
906 struct list_head *list;
909 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
911 struct flush_busy_ctx_data *flush_data = data;
912 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
913 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
915 spin_lock(&ctx->lock);
916 list_splice_tail_init(&ctx->rq_list, flush_data->list);
917 sbitmap_clear_bit(sb, bitnr);
918 spin_unlock(&ctx->lock);
923 * Process software queues that have been marked busy, splicing them
924 * to the for-dispatch
926 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
928 struct flush_busy_ctx_data data = {
933 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
935 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
937 struct dispatch_rq_data {
938 struct blk_mq_hw_ctx *hctx;
942 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
945 struct dispatch_rq_data *dispatch_data = data;
946 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
947 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
949 spin_lock(&ctx->lock);
950 if (!list_empty(&ctx->rq_list)) {
951 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
952 list_del_init(&dispatch_data->rq->queuelist);
953 if (list_empty(&ctx->rq_list))
954 sbitmap_clear_bit(sb, bitnr);
956 spin_unlock(&ctx->lock);
958 return !dispatch_data->rq;
961 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
962 struct blk_mq_ctx *start)
964 unsigned off = start ? start->index_hw : 0;
965 struct dispatch_rq_data data = {
970 __sbitmap_for_each_set(&hctx->ctx_map, off,
971 dispatch_rq_from_ctx, &data);
976 static inline unsigned int queued_to_index(unsigned int queued)
981 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
984 bool blk_mq_get_driver_tag(struct request *rq)
986 struct blk_mq_alloc_data data = {
988 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
989 .flags = BLK_MQ_REQ_NOWAIT,
996 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
997 data.flags |= BLK_MQ_REQ_RESERVED;
999 shared = blk_mq_tag_busy(data.hctx);
1000 rq->tag = blk_mq_get_tag(&data);
1003 rq->rq_flags |= RQF_MQ_INFLIGHT;
1004 atomic_inc(&data.hctx->nr_active);
1006 data.hctx->tags->rqs[rq->tag] = rq;
1010 return rq->tag != -1;
1013 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1014 int flags, void *key)
1016 struct blk_mq_hw_ctx *hctx;
1018 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1020 spin_lock(&hctx->dispatch_wait_lock);
1021 list_del_init(&wait->entry);
1022 spin_unlock(&hctx->dispatch_wait_lock);
1024 blk_mq_run_hw_queue(hctx, true);
1029 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1030 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1031 * restart. For both cases, take care to check the condition again after
1032 * marking us as waiting.
1034 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1037 struct wait_queue_head *wq;
1038 wait_queue_entry_t *wait;
1041 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1042 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
1043 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
1046 * It's possible that a tag was freed in the window between the
1047 * allocation failure and adding the hardware queue to the wait
1050 * Don't clear RESTART here, someone else could have set it.
1051 * At most this will cost an extra queue run.
1053 return blk_mq_get_driver_tag(rq);
1056 wait = &hctx->dispatch_wait;
1057 if (!list_empty_careful(&wait->entry))
1060 wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait;
1062 spin_lock_irq(&wq->lock);
1063 spin_lock(&hctx->dispatch_wait_lock);
1064 if (!list_empty(&wait->entry)) {
1065 spin_unlock(&hctx->dispatch_wait_lock);
1066 spin_unlock_irq(&wq->lock);
1070 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1071 __add_wait_queue(wq, wait);
1074 * It's possible that a tag was freed in the window between the
1075 * allocation failure and adding the hardware queue to the wait
1078 ret = blk_mq_get_driver_tag(rq);
1080 spin_unlock(&hctx->dispatch_wait_lock);
1081 spin_unlock_irq(&wq->lock);
1086 * We got a tag, remove ourselves from the wait queue to ensure
1087 * someone else gets the wakeup.
1089 list_del_init(&wait->entry);
1090 spin_unlock(&hctx->dispatch_wait_lock);
1091 spin_unlock_irq(&wq->lock);
1096 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1097 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1099 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1100 * - EWMA is one simple way to compute running average value
1101 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1102 * - take 4 as factor for avoiding to get too small(0) result, and this
1103 * factor doesn't matter because EWMA decreases exponentially
1105 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1109 if (hctx->queue->elevator)
1112 ewma = hctx->dispatch_busy;
1117 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1119 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1120 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1122 hctx->dispatch_busy = ewma;
1125 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1128 * Returns true if we did some work AND can potentially do more.
1130 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1133 struct blk_mq_hw_ctx *hctx;
1134 struct request *rq, *nxt;
1135 bool no_tag = false;
1137 blk_status_t ret = BLK_STS_OK;
1139 if (list_empty(list))
1142 WARN_ON(!list_is_singular(list) && got_budget);
1145 * Now process all the entries, sending them to the driver.
1147 errors = queued = 0;
1149 struct blk_mq_queue_data bd;
1151 rq = list_first_entry(list, struct request, queuelist);
1153 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1154 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1157 if (!blk_mq_get_driver_tag(rq)) {
1159 * The initial allocation attempt failed, so we need to
1160 * rerun the hardware queue when a tag is freed. The
1161 * waitqueue takes care of that. If the queue is run
1162 * before we add this entry back on the dispatch list,
1163 * we'll re-run it below.
1165 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1166 blk_mq_put_dispatch_budget(hctx);
1168 * For non-shared tags, the RESTART check
1171 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1177 list_del_init(&rq->queuelist);
1182 * Flag last if we have no more requests, or if we have more
1183 * but can't assign a driver tag to it.
1185 if (list_empty(list))
1188 nxt = list_first_entry(list, struct request, queuelist);
1189 bd.last = !blk_mq_get_driver_tag(nxt);
1192 ret = q->mq_ops->queue_rq(hctx, &bd);
1193 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1195 * If an I/O scheduler has been configured and we got a
1196 * driver tag for the next request already, free it
1199 if (!list_empty(list)) {
1200 nxt = list_first_entry(list, struct request, queuelist);
1201 blk_mq_put_driver_tag(nxt);
1203 list_add(&rq->queuelist, list);
1204 __blk_mq_requeue_request(rq);
1208 if (unlikely(ret != BLK_STS_OK)) {
1210 blk_mq_end_request(rq, BLK_STS_IOERR);
1215 } while (!list_empty(list));
1217 hctx->dispatched[queued_to_index(queued)]++;
1220 * Any items that need requeuing? Stuff them into hctx->dispatch,
1221 * that is where we will continue on next queue run.
1223 if (!list_empty(list)) {
1226 spin_lock(&hctx->lock);
1227 list_splice_init(list, &hctx->dispatch);
1228 spin_unlock(&hctx->lock);
1231 * If SCHED_RESTART was set by the caller of this function and
1232 * it is no longer set that means that it was cleared by another
1233 * thread and hence that a queue rerun is needed.
1235 * If 'no_tag' is set, that means that we failed getting
1236 * a driver tag with an I/O scheduler attached. If our dispatch
1237 * waitqueue is no longer active, ensure that we run the queue
1238 * AFTER adding our entries back to the list.
1240 * If no I/O scheduler has been configured it is possible that
1241 * the hardware queue got stopped and restarted before requests
1242 * were pushed back onto the dispatch list. Rerun the queue to
1243 * avoid starvation. Notes:
1244 * - blk_mq_run_hw_queue() checks whether or not a queue has
1245 * been stopped before rerunning a queue.
1246 * - Some but not all block drivers stop a queue before
1247 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1250 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1251 * bit is set, run queue after a delay to avoid IO stalls
1252 * that could otherwise occur if the queue is idle.
1254 needs_restart = blk_mq_sched_needs_restart(hctx);
1255 if (!needs_restart ||
1256 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1257 blk_mq_run_hw_queue(hctx, true);
1258 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1259 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1261 blk_mq_update_dispatch_busy(hctx, true);
1264 blk_mq_update_dispatch_busy(hctx, false);
1267 * If the host/device is unable to accept more work, inform the
1270 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1273 return (queued + errors) != 0;
1276 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1281 * We should be running this queue from one of the CPUs that
1284 * There are at least two related races now between setting
1285 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1286 * __blk_mq_run_hw_queue():
1288 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1289 * but later it becomes online, then this warning is harmless
1292 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1293 * but later it becomes offline, then the warning can't be
1294 * triggered, and we depend on blk-mq timeout handler to
1295 * handle dispatched requests to this hctx
1297 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1298 cpu_online(hctx->next_cpu)) {
1299 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1300 raw_smp_processor_id(),
1301 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1306 * We can't run the queue inline with ints disabled. Ensure that
1307 * we catch bad users of this early.
1309 WARN_ON_ONCE(in_interrupt());
1311 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1313 hctx_lock(hctx, &srcu_idx);
1314 blk_mq_sched_dispatch_requests(hctx);
1315 hctx_unlock(hctx, srcu_idx);
1318 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1320 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1322 if (cpu >= nr_cpu_ids)
1323 cpu = cpumask_first(hctx->cpumask);
1328 * It'd be great if the workqueue API had a way to pass
1329 * in a mask and had some smarts for more clever placement.
1330 * For now we just round-robin here, switching for every
1331 * BLK_MQ_CPU_WORK_BATCH queued items.
1333 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1336 int next_cpu = hctx->next_cpu;
1338 if (hctx->queue->nr_hw_queues == 1)
1339 return WORK_CPU_UNBOUND;
1341 if (--hctx->next_cpu_batch <= 0) {
1343 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1345 if (next_cpu >= nr_cpu_ids)
1346 next_cpu = blk_mq_first_mapped_cpu(hctx);
1347 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1351 * Do unbound schedule if we can't find a online CPU for this hctx,
1352 * and it should only happen in the path of handling CPU DEAD.
1354 if (!cpu_online(next_cpu)) {
1361 * Make sure to re-select CPU next time once after CPUs
1362 * in hctx->cpumask become online again.
1364 hctx->next_cpu = next_cpu;
1365 hctx->next_cpu_batch = 1;
1366 return WORK_CPU_UNBOUND;
1369 hctx->next_cpu = next_cpu;
1373 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1374 unsigned long msecs)
1376 if (unlikely(blk_mq_hctx_stopped(hctx)))
1379 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1380 int cpu = get_cpu();
1381 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1382 __blk_mq_run_hw_queue(hctx);
1390 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1391 msecs_to_jiffies(msecs));
1394 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1396 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1398 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1400 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1406 * When queue is quiesced, we may be switching io scheduler, or
1407 * updating nr_hw_queues, or other things, and we can't run queue
1408 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1410 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1413 hctx_lock(hctx, &srcu_idx);
1414 need_run = !blk_queue_quiesced(hctx->queue) &&
1415 blk_mq_hctx_has_pending(hctx);
1416 hctx_unlock(hctx, srcu_idx);
1419 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1425 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1427 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1429 struct blk_mq_hw_ctx *hctx;
1432 queue_for_each_hw_ctx(q, hctx, i) {
1433 if (blk_mq_hctx_stopped(hctx))
1436 blk_mq_run_hw_queue(hctx, async);
1439 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1442 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1443 * @q: request queue.
1445 * The caller is responsible for serializing this function against
1446 * blk_mq_{start,stop}_hw_queue().
1448 bool blk_mq_queue_stopped(struct request_queue *q)
1450 struct blk_mq_hw_ctx *hctx;
1453 queue_for_each_hw_ctx(q, hctx, i)
1454 if (blk_mq_hctx_stopped(hctx))
1459 EXPORT_SYMBOL(blk_mq_queue_stopped);
1462 * This function is often used for pausing .queue_rq() by driver when
1463 * there isn't enough resource or some conditions aren't satisfied, and
1464 * BLK_STS_RESOURCE is usually returned.
1466 * We do not guarantee that dispatch can be drained or blocked
1467 * after blk_mq_stop_hw_queue() returns. Please use
1468 * blk_mq_quiesce_queue() for that requirement.
1470 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1472 cancel_delayed_work(&hctx->run_work);
1474 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1476 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1479 * This function is often used for pausing .queue_rq() by driver when
1480 * there isn't enough resource or some conditions aren't satisfied, and
1481 * BLK_STS_RESOURCE is usually returned.
1483 * We do not guarantee that dispatch can be drained or blocked
1484 * after blk_mq_stop_hw_queues() returns. Please use
1485 * blk_mq_quiesce_queue() for that requirement.
1487 void blk_mq_stop_hw_queues(struct request_queue *q)
1489 struct blk_mq_hw_ctx *hctx;
1492 queue_for_each_hw_ctx(q, hctx, i)
1493 blk_mq_stop_hw_queue(hctx);
1495 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1497 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1499 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1501 blk_mq_run_hw_queue(hctx, false);
1503 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1505 void blk_mq_start_hw_queues(struct request_queue *q)
1507 struct blk_mq_hw_ctx *hctx;
1510 queue_for_each_hw_ctx(q, hctx, i)
1511 blk_mq_start_hw_queue(hctx);
1513 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1515 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1517 if (!blk_mq_hctx_stopped(hctx))
1520 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1521 blk_mq_run_hw_queue(hctx, async);
1523 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1525 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1527 struct blk_mq_hw_ctx *hctx;
1530 queue_for_each_hw_ctx(q, hctx, i)
1531 blk_mq_start_stopped_hw_queue(hctx, async);
1533 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1535 static void blk_mq_run_work_fn(struct work_struct *work)
1537 struct blk_mq_hw_ctx *hctx;
1539 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1542 * If we are stopped, don't run the queue.
1544 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1547 __blk_mq_run_hw_queue(hctx);
1550 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1554 struct blk_mq_ctx *ctx = rq->mq_ctx;
1556 lockdep_assert_held(&ctx->lock);
1558 trace_block_rq_insert(hctx->queue, rq);
1561 list_add(&rq->queuelist, &ctx->rq_list);
1563 list_add_tail(&rq->queuelist, &ctx->rq_list);
1566 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1569 struct blk_mq_ctx *ctx = rq->mq_ctx;
1571 lockdep_assert_held(&ctx->lock);
1573 __blk_mq_insert_req_list(hctx, rq, at_head);
1574 blk_mq_hctx_mark_pending(hctx, ctx);
1578 * Should only be used carefully, when the caller knows we want to
1579 * bypass a potential IO scheduler on the target device.
1581 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1583 struct blk_mq_ctx *ctx = rq->mq_ctx;
1584 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1586 spin_lock(&hctx->lock);
1587 list_add_tail(&rq->queuelist, &hctx->dispatch);
1588 spin_unlock(&hctx->lock);
1591 blk_mq_run_hw_queue(hctx, false);
1594 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1595 struct list_head *list)
1601 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1604 list_for_each_entry(rq, list, queuelist) {
1605 BUG_ON(rq->mq_ctx != ctx);
1606 trace_block_rq_insert(hctx->queue, rq);
1609 spin_lock(&ctx->lock);
1610 list_splice_tail_init(list, &ctx->rq_list);
1611 blk_mq_hctx_mark_pending(hctx, ctx);
1612 spin_unlock(&ctx->lock);
1615 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1617 struct request *rqa = container_of(a, struct request, queuelist);
1618 struct request *rqb = container_of(b, struct request, queuelist);
1620 return !(rqa->mq_ctx < rqb->mq_ctx ||
1621 (rqa->mq_ctx == rqb->mq_ctx &&
1622 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1625 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1627 struct blk_mq_ctx *this_ctx;
1628 struct request_queue *this_q;
1631 LIST_HEAD(ctx_list);
1634 list_splice_init(&plug->mq_list, &list);
1636 list_sort(NULL, &list, plug_ctx_cmp);
1642 while (!list_empty(&list)) {
1643 rq = list_entry_rq(list.next);
1644 list_del_init(&rq->queuelist);
1646 if (rq->mq_ctx != this_ctx) {
1648 trace_block_unplug(this_q, depth, !from_schedule);
1649 blk_mq_sched_insert_requests(this_q, this_ctx,
1654 this_ctx = rq->mq_ctx;
1660 list_add_tail(&rq->queuelist, &ctx_list);
1664 * If 'this_ctx' is set, we know we have entries to complete
1665 * on 'ctx_list'. Do those.
1668 trace_block_unplug(this_q, depth, !from_schedule);
1669 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1674 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1676 blk_init_request_from_bio(rq, bio);
1678 blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1680 blk_account_io_start(rq, true);
1683 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1686 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1688 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1691 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1695 struct request_queue *q = rq->q;
1696 struct blk_mq_queue_data bd = {
1700 blk_qc_t new_cookie;
1703 new_cookie = request_to_qc_t(hctx, rq);
1706 * For OK queue, we are done. For error, caller may kill it.
1707 * Any other error (busy), just add it to our list as we
1708 * previously would have done.
1710 ret = q->mq_ops->queue_rq(hctx, &bd);
1713 blk_mq_update_dispatch_busy(hctx, false);
1714 *cookie = new_cookie;
1716 case BLK_STS_RESOURCE:
1717 case BLK_STS_DEV_RESOURCE:
1719 * If direct dispatch fails, we cannot allow any merging on
1720 * this IO. Drivers (like SCSI) may have set up permanent state
1721 * for this request, like SG tables and mappings, and if we
1722 * merge to it later on then we'll still only do IO to the
1725 rq->cmd_flags |= REQ_NOMERGE;
1727 blk_mq_update_dispatch_busy(hctx, true);
1728 __blk_mq_requeue_request(rq);
1731 blk_mq_update_dispatch_busy(hctx, false);
1732 *cookie = BLK_QC_T_NONE;
1740 * Don't allow direct dispatch of anything but regular reads/writes,
1741 * as some of the other commands can potentially share request space
1742 * with data we need for the IO scheduler. If we attempt a direct dispatch
1743 * on those and fail, we can't safely add it to the scheduler afterwards
1744 * without potentially overwriting data that the driver has already written.
1746 static bool blk_rq_can_direct_dispatch(struct request *rq)
1748 return req_op(rq) == REQ_OP_READ || req_op(rq) == REQ_OP_WRITE;
1751 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1756 struct request_queue *q = rq->q;
1757 bool run_queue = true;
1760 * RCU or SRCU read lock is needed before checking quiesced flag.
1762 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1763 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1764 * and avoid driver to try to dispatch again.
1766 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1768 bypass_insert = false;
1772 if (!blk_rq_can_direct_dispatch(rq) || (q->elevator && !bypass_insert))
1775 if (!blk_mq_get_dispatch_budget(hctx))
1778 if (!blk_mq_get_driver_tag(rq)) {
1779 blk_mq_put_dispatch_budget(hctx);
1783 return __blk_mq_issue_directly(hctx, rq, cookie);
1786 return BLK_STS_RESOURCE;
1788 blk_mq_sched_insert_request(rq, false, run_queue, false);
1792 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1793 struct request *rq, blk_qc_t *cookie)
1798 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1800 hctx_lock(hctx, &srcu_idx);
1802 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1803 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1804 blk_mq_sched_insert_request(rq, false, true, false);
1805 else if (ret != BLK_STS_OK)
1806 blk_mq_end_request(rq, ret);
1808 hctx_unlock(hctx, srcu_idx);
1811 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1815 blk_qc_t unused_cookie;
1816 struct blk_mq_ctx *ctx = rq->mq_ctx;
1817 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1819 hctx_lock(hctx, &srcu_idx);
1820 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1821 hctx_unlock(hctx, srcu_idx);
1826 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1827 struct list_head *list)
1829 while (!list_empty(list)) {
1831 struct request *rq = list_first_entry(list, struct request,
1834 if (!blk_rq_can_direct_dispatch(rq))
1837 list_del_init(&rq->queuelist);
1838 ret = blk_mq_request_issue_directly(rq);
1839 if (ret != BLK_STS_OK) {
1840 if (ret == BLK_STS_RESOURCE ||
1841 ret == BLK_STS_DEV_RESOURCE) {
1842 list_add(&rq->queuelist, list);
1845 blk_mq_end_request(rq, ret);
1850 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1852 const int is_sync = op_is_sync(bio->bi_opf);
1853 const int is_flush_fua = op_is_flush(bio->bi_opf);
1854 struct blk_mq_alloc_data data = { .flags = 0 };
1856 unsigned int request_count = 0;
1857 struct blk_plug *plug;
1858 struct request *same_queue_rq = NULL;
1861 blk_queue_bounce(q, &bio);
1863 blk_queue_split(q, &bio);
1865 if (!bio_integrity_prep(bio))
1866 return BLK_QC_T_NONE;
1868 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1869 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1870 return BLK_QC_T_NONE;
1872 if (blk_mq_sched_bio_merge(q, bio))
1873 return BLK_QC_T_NONE;
1875 rq_qos_throttle(q, bio, NULL);
1877 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1878 if (unlikely(!rq)) {
1879 rq_qos_cleanup(q, bio);
1880 if (bio->bi_opf & REQ_NOWAIT)
1881 bio_wouldblock_error(bio);
1882 return BLK_QC_T_NONE;
1885 trace_block_getrq(q, bio, bio->bi_opf);
1887 rq_qos_track(q, rq, bio);
1889 cookie = request_to_qc_t(data.hctx, rq);
1891 plug = current->plug;
1892 if (unlikely(is_flush_fua)) {
1893 blk_mq_put_ctx(data.ctx);
1894 blk_mq_bio_to_request(rq, bio);
1896 /* bypass scheduler for flush rq */
1897 blk_insert_flush(rq);
1898 blk_mq_run_hw_queue(data.hctx, true);
1899 } else if (plug && q->nr_hw_queues == 1) {
1900 struct request *last = NULL;
1902 blk_mq_put_ctx(data.ctx);
1903 blk_mq_bio_to_request(rq, bio);
1906 * @request_count may become stale because of schedule
1907 * out, so check the list again.
1909 if (list_empty(&plug->mq_list))
1911 else if (blk_queue_nomerges(q))
1912 request_count = blk_plug_queued_count(q);
1915 trace_block_plug(q);
1917 last = list_entry_rq(plug->mq_list.prev);
1919 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1920 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1921 blk_flush_plug_list(plug, false);
1922 trace_block_plug(q);
1925 list_add_tail(&rq->queuelist, &plug->mq_list);
1926 } else if (plug && !blk_queue_nomerges(q)) {
1927 blk_mq_bio_to_request(rq, bio);
1930 * We do limited plugging. If the bio can be merged, do that.
1931 * Otherwise the existing request in the plug list will be
1932 * issued. So the plug list will have one request at most
1933 * The plug list might get flushed before this. If that happens,
1934 * the plug list is empty, and same_queue_rq is invalid.
1936 if (list_empty(&plug->mq_list))
1937 same_queue_rq = NULL;
1939 list_del_init(&same_queue_rq->queuelist);
1940 list_add_tail(&rq->queuelist, &plug->mq_list);
1942 blk_mq_put_ctx(data.ctx);
1944 if (same_queue_rq) {
1945 data.hctx = blk_mq_map_queue(q,
1946 same_queue_rq->mq_ctx->cpu);
1947 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1950 } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
1951 !data.hctx->dispatch_busy)) {
1952 blk_mq_put_ctx(data.ctx);
1953 blk_mq_bio_to_request(rq, bio);
1954 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1956 blk_mq_put_ctx(data.ctx);
1957 blk_mq_bio_to_request(rq, bio);
1958 blk_mq_sched_insert_request(rq, false, true, true);
1964 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1965 unsigned int hctx_idx)
1969 if (tags->rqs && set->ops->exit_request) {
1972 for (i = 0; i < tags->nr_tags; i++) {
1973 struct request *rq = tags->static_rqs[i];
1977 set->ops->exit_request(set, rq, hctx_idx);
1978 tags->static_rqs[i] = NULL;
1982 while (!list_empty(&tags->page_list)) {
1983 page = list_first_entry(&tags->page_list, struct page, lru);
1984 list_del_init(&page->lru);
1986 * Remove kmemleak object previously allocated in
1987 * blk_mq_init_rq_map().
1989 kmemleak_free(page_address(page));
1990 __free_pages(page, page->private);
1994 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1998 kfree(tags->static_rqs);
1999 tags->static_rqs = NULL;
2001 blk_mq_free_tags(tags);
2004 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2005 unsigned int hctx_idx,
2006 unsigned int nr_tags,
2007 unsigned int reserved_tags)
2009 struct blk_mq_tags *tags;
2012 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2013 if (node == NUMA_NO_NODE)
2014 node = set->numa_node;
2016 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2017 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2021 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2022 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2025 blk_mq_free_tags(tags);
2029 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2030 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2032 if (!tags->static_rqs) {
2034 blk_mq_free_tags(tags);
2041 static size_t order_to_size(unsigned int order)
2043 return (size_t)PAGE_SIZE << order;
2046 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2047 unsigned int hctx_idx, int node)
2051 if (set->ops->init_request) {
2052 ret = set->ops->init_request(set, rq, hctx_idx, node);
2057 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2061 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2062 unsigned int hctx_idx, unsigned int depth)
2064 unsigned int i, j, entries_per_page, max_order = 4;
2065 size_t rq_size, left;
2068 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2069 if (node == NUMA_NO_NODE)
2070 node = set->numa_node;
2072 INIT_LIST_HEAD(&tags->page_list);
2075 * rq_size is the size of the request plus driver payload, rounded
2076 * to the cacheline size
2078 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2080 left = rq_size * depth;
2082 for (i = 0; i < depth; ) {
2083 int this_order = max_order;
2088 while (this_order && left < order_to_size(this_order - 1))
2092 page = alloc_pages_node(node,
2093 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2099 if (order_to_size(this_order) < rq_size)
2106 page->private = this_order;
2107 list_add_tail(&page->lru, &tags->page_list);
2109 p = page_address(page);
2111 * Allow kmemleak to scan these pages as they contain pointers
2112 * to additional allocations like via ops->init_request().
2114 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2115 entries_per_page = order_to_size(this_order) / rq_size;
2116 to_do = min(entries_per_page, depth - i);
2117 left -= to_do * rq_size;
2118 for (j = 0; j < to_do; j++) {
2119 struct request *rq = p;
2121 tags->static_rqs[i] = rq;
2122 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2123 tags->static_rqs[i] = NULL;
2134 blk_mq_free_rqs(set, tags, hctx_idx);
2139 * 'cpu' is going away. splice any existing rq_list entries from this
2140 * software queue to the hw queue dispatch list, and ensure that it
2143 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2145 struct blk_mq_hw_ctx *hctx;
2146 struct blk_mq_ctx *ctx;
2149 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2150 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2152 spin_lock(&ctx->lock);
2153 if (!list_empty(&ctx->rq_list)) {
2154 list_splice_init(&ctx->rq_list, &tmp);
2155 blk_mq_hctx_clear_pending(hctx, ctx);
2157 spin_unlock(&ctx->lock);
2159 if (list_empty(&tmp))
2162 spin_lock(&hctx->lock);
2163 list_splice_tail_init(&tmp, &hctx->dispatch);
2164 spin_unlock(&hctx->lock);
2166 blk_mq_run_hw_queue(hctx, true);
2170 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2172 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2176 /* hctx->ctxs will be freed in queue's release handler */
2177 static void blk_mq_exit_hctx(struct request_queue *q,
2178 struct blk_mq_tag_set *set,
2179 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2181 if (blk_mq_hw_queue_mapped(hctx))
2182 blk_mq_tag_idle(hctx);
2184 if (set->ops->exit_request)
2185 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2187 if (set->ops->exit_hctx)
2188 set->ops->exit_hctx(hctx, hctx_idx);
2190 if (hctx->flags & BLK_MQ_F_BLOCKING)
2191 cleanup_srcu_struct(hctx->srcu);
2193 blk_mq_remove_cpuhp(hctx);
2194 blk_free_flush_queue(hctx->fq);
2195 sbitmap_free(&hctx->ctx_map);
2198 static void blk_mq_exit_hw_queues(struct request_queue *q,
2199 struct blk_mq_tag_set *set, int nr_queue)
2201 struct blk_mq_hw_ctx *hctx;
2204 queue_for_each_hw_ctx(q, hctx, i) {
2207 blk_mq_debugfs_unregister_hctx(hctx);
2208 blk_mq_exit_hctx(q, set, hctx, i);
2212 static int blk_mq_init_hctx(struct request_queue *q,
2213 struct blk_mq_tag_set *set,
2214 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2218 node = hctx->numa_node;
2219 if (node == NUMA_NO_NODE)
2220 node = hctx->numa_node = set->numa_node;
2222 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2223 spin_lock_init(&hctx->lock);
2224 INIT_LIST_HEAD(&hctx->dispatch);
2226 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2228 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2230 hctx->tags = set->tags[hctx_idx];
2233 * Allocate space for all possible cpus to avoid allocation at
2236 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2237 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node);
2239 goto unregister_cpu_notifier;
2241 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2242 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node))
2247 spin_lock_init(&hctx->dispatch_wait_lock);
2248 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2249 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2251 if (set->ops->init_hctx &&
2252 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2255 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2256 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
2260 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2263 if (hctx->flags & BLK_MQ_F_BLOCKING)
2264 init_srcu_struct(hctx->srcu);
2271 if (set->ops->exit_hctx)
2272 set->ops->exit_hctx(hctx, hctx_idx);
2274 sbitmap_free(&hctx->ctx_map);
2277 unregister_cpu_notifier:
2278 blk_mq_remove_cpuhp(hctx);
2282 static void blk_mq_init_cpu_queues(struct request_queue *q,
2283 unsigned int nr_hw_queues)
2287 for_each_possible_cpu(i) {
2288 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2289 struct blk_mq_hw_ctx *hctx;
2292 spin_lock_init(&__ctx->lock);
2293 INIT_LIST_HEAD(&__ctx->rq_list);
2297 * Set local node, IFF we have more than one hw queue. If
2298 * not, we remain on the home node of the device
2300 hctx = blk_mq_map_queue(q, i);
2301 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2302 hctx->numa_node = local_memory_node(cpu_to_node(i));
2306 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2310 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2311 set->queue_depth, set->reserved_tags);
2312 if (!set->tags[hctx_idx])
2315 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2320 blk_mq_free_rq_map(set->tags[hctx_idx]);
2321 set->tags[hctx_idx] = NULL;
2325 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2326 unsigned int hctx_idx)
2328 if (set->tags[hctx_idx]) {
2329 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2330 blk_mq_free_rq_map(set->tags[hctx_idx]);
2331 set->tags[hctx_idx] = NULL;
2335 static void blk_mq_map_swqueue(struct request_queue *q)
2337 unsigned int i, hctx_idx;
2338 struct blk_mq_hw_ctx *hctx;
2339 struct blk_mq_ctx *ctx;
2340 struct blk_mq_tag_set *set = q->tag_set;
2343 * Avoid others reading imcomplete hctx->cpumask through sysfs
2345 mutex_lock(&q->sysfs_lock);
2347 queue_for_each_hw_ctx(q, hctx, i) {
2348 cpumask_clear(hctx->cpumask);
2350 hctx->dispatch_from = NULL;
2354 * Map software to hardware queues.
2356 * If the cpu isn't present, the cpu is mapped to first hctx.
2358 for_each_possible_cpu(i) {
2359 hctx_idx = q->mq_map[i];
2360 /* unmapped hw queue can be remapped after CPU topo changed */
2361 if (!set->tags[hctx_idx] &&
2362 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2364 * If tags initialization fail for some hctx,
2365 * that hctx won't be brought online. In this
2366 * case, remap the current ctx to hctx[0] which
2367 * is guaranteed to always have tags allocated
2372 ctx = per_cpu_ptr(q->queue_ctx, i);
2373 hctx = blk_mq_map_queue(q, i);
2375 cpumask_set_cpu(i, hctx->cpumask);
2376 ctx->index_hw = hctx->nr_ctx;
2377 hctx->ctxs[hctx->nr_ctx++] = ctx;
2380 mutex_unlock(&q->sysfs_lock);
2382 queue_for_each_hw_ctx(q, hctx, i) {
2384 * If no software queues are mapped to this hardware queue,
2385 * disable it and free the request entries.
2387 if (!hctx->nr_ctx) {
2388 /* Never unmap queue 0. We need it as a
2389 * fallback in case of a new remap fails
2392 if (i && set->tags[i])
2393 blk_mq_free_map_and_requests(set, i);
2399 hctx->tags = set->tags[i];
2400 WARN_ON(!hctx->tags);
2403 * Set the map size to the number of mapped software queues.
2404 * This is more accurate and more efficient than looping
2405 * over all possibly mapped software queues.
2407 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2410 * Initialize batch roundrobin counts
2412 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2413 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2418 * Caller needs to ensure that we're either frozen/quiesced, or that
2419 * the queue isn't live yet.
2421 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2423 struct blk_mq_hw_ctx *hctx;
2426 queue_for_each_hw_ctx(q, hctx, i) {
2428 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2430 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2434 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2437 struct request_queue *q;
2439 lockdep_assert_held(&set->tag_list_lock);
2441 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2442 blk_mq_freeze_queue(q);
2443 queue_set_hctx_shared(q, shared);
2444 blk_mq_unfreeze_queue(q);
2448 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2450 struct blk_mq_tag_set *set = q->tag_set;
2452 mutex_lock(&set->tag_list_lock);
2453 list_del_rcu(&q->tag_set_list);
2454 if (list_is_singular(&set->tag_list)) {
2455 /* just transitioned to unshared */
2456 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2457 /* update existing queue */
2458 blk_mq_update_tag_set_depth(set, false);
2460 mutex_unlock(&set->tag_list_lock);
2461 INIT_LIST_HEAD(&q->tag_set_list);
2464 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2465 struct request_queue *q)
2469 mutex_lock(&set->tag_list_lock);
2472 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2474 if (!list_empty(&set->tag_list) &&
2475 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2476 set->flags |= BLK_MQ_F_TAG_SHARED;
2477 /* update existing queue */
2478 blk_mq_update_tag_set_depth(set, true);
2480 if (set->flags & BLK_MQ_F_TAG_SHARED)
2481 queue_set_hctx_shared(q, true);
2482 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2484 mutex_unlock(&set->tag_list_lock);
2488 * It is the actual release handler for mq, but we do it from
2489 * request queue's release handler for avoiding use-after-free
2490 * and headache because q->mq_kobj shouldn't have been introduced,
2491 * but we can't group ctx/kctx kobj without it.
2493 void blk_mq_release(struct request_queue *q)
2495 struct blk_mq_hw_ctx *hctx;
2498 /* hctx kobj stays in hctx */
2499 queue_for_each_hw_ctx(q, hctx, i) {
2502 kobject_put(&hctx->kobj);
2507 kfree(q->queue_hw_ctx);
2510 * release .mq_kobj and sw queue's kobject now because
2511 * both share lifetime with request queue.
2513 blk_mq_sysfs_deinit(q);
2515 free_percpu(q->queue_ctx);
2518 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2520 struct request_queue *uninit_q, *q;
2522 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2524 return ERR_PTR(-ENOMEM);
2526 q = blk_mq_init_allocated_queue(set, uninit_q);
2528 blk_cleanup_queue(uninit_q);
2532 EXPORT_SYMBOL(blk_mq_init_queue);
2535 * Helper for setting up a queue with mq ops, given queue depth, and
2536 * the passed in mq ops flags.
2538 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2539 const struct blk_mq_ops *ops,
2540 unsigned int queue_depth,
2541 unsigned int set_flags)
2543 struct request_queue *q;
2546 memset(set, 0, sizeof(*set));
2548 set->nr_hw_queues = 1;
2549 set->queue_depth = queue_depth;
2550 set->numa_node = NUMA_NO_NODE;
2551 set->flags = set_flags;
2553 ret = blk_mq_alloc_tag_set(set);
2555 return ERR_PTR(ret);
2557 q = blk_mq_init_queue(set);
2559 blk_mq_free_tag_set(set);
2565 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2567 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2569 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2571 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2572 __alignof__(struct blk_mq_hw_ctx)) !=
2573 sizeof(struct blk_mq_hw_ctx));
2575 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2576 hw_ctx_size += sizeof(struct srcu_struct);
2581 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2582 struct blk_mq_tag_set *set, struct request_queue *q,
2583 int hctx_idx, int node)
2585 struct blk_mq_hw_ctx *hctx;
2587 hctx = kzalloc_node(blk_mq_hw_ctx_size(set),
2588 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2593 if (!zalloc_cpumask_var_node(&hctx->cpumask,
2594 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2600 atomic_set(&hctx->nr_active, 0);
2601 hctx->numa_node = node;
2602 hctx->queue_num = hctx_idx;
2604 if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) {
2605 free_cpumask_var(hctx->cpumask);
2609 blk_mq_hctx_kobj_init(hctx);
2614 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2615 struct request_queue *q)
2618 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2620 /* protect against switching io scheduler */
2621 mutex_lock(&q->sysfs_lock);
2622 for (i = 0; i < set->nr_hw_queues; i++) {
2624 struct blk_mq_hw_ctx *hctx;
2626 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2628 * If the hw queue has been mapped to another numa node,
2629 * we need to realloc the hctx. If allocation fails, fallback
2630 * to use the previous one.
2632 if (hctxs[i] && (hctxs[i]->numa_node == node))
2635 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2638 blk_mq_exit_hctx(q, set, hctxs[i], i);
2639 kobject_put(&hctxs[i]->kobj);
2644 pr_warn("Allocate new hctx on node %d fails,\
2645 fallback to previous one on node %d\n",
2646 node, hctxs[i]->numa_node);
2652 * Increasing nr_hw_queues fails. Free the newly allocated
2653 * hctxs and keep the previous q->nr_hw_queues.
2655 if (i != set->nr_hw_queues) {
2656 j = q->nr_hw_queues;
2660 end = q->nr_hw_queues;
2661 q->nr_hw_queues = set->nr_hw_queues;
2664 for (; j < end; j++) {
2665 struct blk_mq_hw_ctx *hctx = hctxs[j];
2669 blk_mq_free_map_and_requests(set, j);
2670 blk_mq_exit_hctx(q, set, hctx, j);
2671 kobject_put(&hctx->kobj);
2676 mutex_unlock(&q->sysfs_lock);
2679 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2680 struct request_queue *q)
2682 /* mark the queue as mq asap */
2683 q->mq_ops = set->ops;
2685 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2686 blk_mq_poll_stats_bkt,
2687 BLK_MQ_POLL_STATS_BKTS, q);
2691 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2695 /* init q->mq_kobj and sw queues' kobjects */
2696 blk_mq_sysfs_init(q);
2698 q->queue_hw_ctx = kcalloc_node(nr_cpu_ids, sizeof(*(q->queue_hw_ctx)),
2699 GFP_KERNEL, set->numa_node);
2700 if (!q->queue_hw_ctx)
2703 q->mq_map = set->mq_map;
2705 blk_mq_realloc_hw_ctxs(set, q);
2706 if (!q->nr_hw_queues)
2709 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2710 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2712 q->nr_queues = nr_cpu_ids;
2714 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2716 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2717 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2719 q->sg_reserved_size = INT_MAX;
2721 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2722 INIT_LIST_HEAD(&q->requeue_list);
2723 spin_lock_init(&q->requeue_lock);
2725 blk_queue_make_request(q, blk_mq_make_request);
2726 if (q->mq_ops->poll)
2727 q->poll_fn = blk_mq_poll;
2730 * Do this after blk_queue_make_request() overrides it...
2732 q->nr_requests = set->queue_depth;
2735 * Default to classic polling
2739 if (set->ops->complete)
2740 blk_queue_softirq_done(q, set->ops->complete);
2742 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2743 blk_mq_add_queue_tag_set(set, q);
2744 blk_mq_map_swqueue(q);
2746 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2749 ret = elevator_init_mq(q);
2751 return ERR_PTR(ret);
2757 kfree(q->queue_hw_ctx);
2759 free_percpu(q->queue_ctx);
2762 return ERR_PTR(-ENOMEM);
2764 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2766 void blk_mq_free_queue(struct request_queue *q)
2768 struct blk_mq_tag_set *set = q->tag_set;
2770 blk_mq_del_queue_tag_set(q);
2771 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2774 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2778 for (i = 0; i < set->nr_hw_queues; i++)
2779 if (!__blk_mq_alloc_rq_map(set, i))
2786 blk_mq_free_rq_map(set->tags[i]);
2792 * Allocate the request maps associated with this tag_set. Note that this
2793 * may reduce the depth asked for, if memory is tight. set->queue_depth
2794 * will be updated to reflect the allocated depth.
2796 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2801 depth = set->queue_depth;
2803 err = __blk_mq_alloc_rq_maps(set);
2807 set->queue_depth >>= 1;
2808 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2812 } while (set->queue_depth);
2814 if (!set->queue_depth || err) {
2815 pr_err("blk-mq: failed to allocate request map\n");
2819 if (depth != set->queue_depth)
2820 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2821 depth, set->queue_depth);
2826 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2828 if (set->ops->map_queues) {
2830 * transport .map_queues is usually done in the following
2833 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2834 * mask = get_cpu_mask(queue)
2835 * for_each_cpu(cpu, mask)
2836 * set->mq_map[cpu] = queue;
2839 * When we need to remap, the table has to be cleared for
2840 * killing stale mapping since one CPU may not be mapped
2843 blk_mq_clear_mq_map(set);
2845 return set->ops->map_queues(set);
2847 return blk_mq_map_queues(set);
2851 * Alloc a tag set to be associated with one or more request queues.
2852 * May fail with EINVAL for various error conditions. May adjust the
2853 * requested depth down, if it's too large. In that case, the set
2854 * value will be stored in set->queue_depth.
2856 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2860 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2862 if (!set->nr_hw_queues)
2864 if (!set->queue_depth)
2866 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2869 if (!set->ops->queue_rq)
2872 if (!set->ops->get_budget ^ !set->ops->put_budget)
2875 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2876 pr_info("blk-mq: reduced tag depth to %u\n",
2878 set->queue_depth = BLK_MQ_MAX_DEPTH;
2882 * If a crashdump is active, then we are potentially in a very
2883 * memory constrained environment. Limit us to 1 queue and
2884 * 64 tags to prevent using too much memory.
2886 if (is_kdump_kernel()) {
2887 set->nr_hw_queues = 1;
2888 set->queue_depth = min(64U, set->queue_depth);
2891 * There is no use for more h/w queues than cpus.
2893 if (set->nr_hw_queues > nr_cpu_ids)
2894 set->nr_hw_queues = nr_cpu_ids;
2896 set->tags = kcalloc_node(nr_cpu_ids, sizeof(struct blk_mq_tags *),
2897 GFP_KERNEL, set->numa_node);
2902 set->mq_map = kcalloc_node(nr_cpu_ids, sizeof(*set->mq_map),
2903 GFP_KERNEL, set->numa_node);
2907 ret = blk_mq_update_queue_map(set);
2909 goto out_free_mq_map;
2911 ret = blk_mq_alloc_rq_maps(set);
2913 goto out_free_mq_map;
2915 mutex_init(&set->tag_list_lock);
2916 INIT_LIST_HEAD(&set->tag_list);
2928 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2930 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2934 for (i = 0; i < nr_cpu_ids; i++)
2935 blk_mq_free_map_and_requests(set, i);
2943 EXPORT_SYMBOL(blk_mq_free_tag_set);
2945 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2947 struct blk_mq_tag_set *set = q->tag_set;
2948 struct blk_mq_hw_ctx *hctx;
2954 blk_mq_freeze_queue(q);
2955 blk_mq_quiesce_queue(q);
2958 queue_for_each_hw_ctx(q, hctx, i) {
2962 * If we're using an MQ scheduler, just update the scheduler
2963 * queue depth. This is similar to what the old code would do.
2965 if (!hctx->sched_tags) {
2966 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2969 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2977 q->nr_requests = nr;
2979 blk_mq_unquiesce_queue(q);
2980 blk_mq_unfreeze_queue(q);
2986 * request_queue and elevator_type pair.
2987 * It is just used by __blk_mq_update_nr_hw_queues to cache
2988 * the elevator_type associated with a request_queue.
2990 struct blk_mq_qe_pair {
2991 struct list_head node;
2992 struct request_queue *q;
2993 struct elevator_type *type;
2997 * Cache the elevator_type in qe pair list and switch the
2998 * io scheduler to 'none'
3000 static bool blk_mq_elv_switch_none(struct list_head *head,
3001 struct request_queue *q)
3003 struct blk_mq_qe_pair *qe;
3008 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3012 INIT_LIST_HEAD(&qe->node);
3014 qe->type = q->elevator->type;
3015 list_add(&qe->node, head);
3017 mutex_lock(&q->sysfs_lock);
3019 * After elevator_switch_mq, the previous elevator_queue will be
3020 * released by elevator_release. The reference of the io scheduler
3021 * module get by elevator_get will also be put. So we need to get
3022 * a reference of the io scheduler module here to prevent it to be
3025 __module_get(qe->type->elevator_owner);
3026 elevator_switch_mq(q, NULL);
3027 mutex_unlock(&q->sysfs_lock);
3032 static void blk_mq_elv_switch_back(struct list_head *head,
3033 struct request_queue *q)
3035 struct blk_mq_qe_pair *qe;
3036 struct elevator_type *t = NULL;
3038 list_for_each_entry(qe, head, node)
3047 list_del(&qe->node);
3050 mutex_lock(&q->sysfs_lock);
3051 elevator_switch_mq(q, t);
3052 mutex_unlock(&q->sysfs_lock);
3055 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3058 struct request_queue *q;
3060 int prev_nr_hw_queues;
3062 lockdep_assert_held(&set->tag_list_lock);
3064 if (nr_hw_queues > nr_cpu_ids)
3065 nr_hw_queues = nr_cpu_ids;
3066 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3069 list_for_each_entry(q, &set->tag_list, tag_set_list)
3070 blk_mq_freeze_queue(q);
3072 * Sync with blk_mq_queue_tag_busy_iter.
3076 * Switch IO scheduler to 'none', cleaning up the data associated
3077 * with the previous scheduler. We will switch back once we are done
3078 * updating the new sw to hw queue mappings.
3080 list_for_each_entry(q, &set->tag_list, tag_set_list)
3081 if (!blk_mq_elv_switch_none(&head, q))
3084 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3085 blk_mq_debugfs_unregister_hctxs(q);
3086 blk_mq_sysfs_unregister(q);
3089 prev_nr_hw_queues = set->nr_hw_queues;
3090 set->nr_hw_queues = nr_hw_queues;
3091 blk_mq_update_queue_map(set);
3093 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3094 blk_mq_realloc_hw_ctxs(set, q);
3095 if (q->nr_hw_queues != set->nr_hw_queues) {
3096 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3097 nr_hw_queues, prev_nr_hw_queues);
3098 set->nr_hw_queues = prev_nr_hw_queues;
3099 blk_mq_map_queues(set);
3102 blk_mq_map_swqueue(q);
3105 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3106 blk_mq_sysfs_register(q);
3107 blk_mq_debugfs_register_hctxs(q);
3111 list_for_each_entry(q, &set->tag_list, tag_set_list)
3112 blk_mq_elv_switch_back(&head, q);
3114 list_for_each_entry(q, &set->tag_list, tag_set_list)
3115 blk_mq_unfreeze_queue(q);
3118 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3120 mutex_lock(&set->tag_list_lock);
3121 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3122 mutex_unlock(&set->tag_list_lock);
3124 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3126 /* Enable polling stats and return whether they were already enabled. */
3127 static bool blk_poll_stats_enable(struct request_queue *q)
3129 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3130 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3132 blk_stat_add_callback(q, q->poll_cb);
3136 static void blk_mq_poll_stats_start(struct request_queue *q)
3139 * We don't arm the callback if polling stats are not enabled or the
3140 * callback is already active.
3142 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3143 blk_stat_is_active(q->poll_cb))
3146 blk_stat_activate_msecs(q->poll_cb, 100);
3149 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3151 struct request_queue *q = cb->data;
3154 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3155 if (cb->stat[bucket].nr_samples)
3156 q->poll_stat[bucket] = cb->stat[bucket];
3160 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3161 struct blk_mq_hw_ctx *hctx,
3164 unsigned long ret = 0;
3168 * If stats collection isn't on, don't sleep but turn it on for
3171 if (!blk_poll_stats_enable(q))
3175 * As an optimistic guess, use half of the mean service time
3176 * for this type of request. We can (and should) make this smarter.
3177 * For instance, if the completion latencies are tight, we can
3178 * get closer than just half the mean. This is especially
3179 * important on devices where the completion latencies are longer
3180 * than ~10 usec. We do use the stats for the relevant IO size
3181 * if available which does lead to better estimates.
3183 bucket = blk_mq_poll_stats_bkt(rq);
3187 if (q->poll_stat[bucket].nr_samples)
3188 ret = (q->poll_stat[bucket].mean + 1) / 2;
3193 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3194 struct blk_mq_hw_ctx *hctx,
3197 struct hrtimer_sleeper hs;
3198 enum hrtimer_mode mode;
3202 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3208 * -1: don't ever hybrid sleep
3209 * 0: use half of prev avg
3210 * >0: use this specific value
3212 if (q->poll_nsec == -1)
3214 else if (q->poll_nsec > 0)
3215 nsecs = q->poll_nsec;
3217 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3222 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3225 * This will be replaced with the stats tracking code, using
3226 * 'avg_completion_time / 2' as the pre-sleep target.
3230 mode = HRTIMER_MODE_REL;
3231 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3232 hrtimer_set_expires(&hs.timer, kt);
3234 hrtimer_init_sleeper(&hs, current);
3236 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3238 set_current_state(TASK_UNINTERRUPTIBLE);
3239 hrtimer_start_expires(&hs.timer, mode);
3242 hrtimer_cancel(&hs.timer);
3243 mode = HRTIMER_MODE_ABS;
3244 } while (hs.task && !signal_pending(current));
3246 __set_current_state(TASK_RUNNING);
3247 destroy_hrtimer_on_stack(&hs.timer);
3251 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3253 struct request_queue *q = hctx->queue;
3257 * If we sleep, have the caller restart the poll loop to reset
3258 * the state. Like for the other success return cases, the
3259 * caller is responsible for checking if the IO completed. If
3260 * the IO isn't complete, we'll get called again and will go
3261 * straight to the busy poll loop.
3263 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3266 hctx->poll_considered++;
3268 state = current->state;
3269 while (!need_resched()) {
3272 hctx->poll_invoked++;
3274 ret = q->mq_ops->poll(hctx, rq->tag);
3276 hctx->poll_success++;
3277 set_current_state(TASK_RUNNING);
3281 if (signal_pending_state(state, current))
3282 set_current_state(TASK_RUNNING);
3284 if (current->state == TASK_RUNNING)
3291 __set_current_state(TASK_RUNNING);
3295 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3297 struct blk_mq_hw_ctx *hctx;
3300 if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3303 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3304 if (!blk_qc_t_is_internal(cookie))
3305 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3307 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3309 * With scheduling, if the request has completed, we'll
3310 * get a NULL return here, as we clear the sched tag when
3311 * that happens. The request still remains valid, like always,
3312 * so we should be safe with just the NULL check.
3318 return __blk_mq_poll(hctx, rq);
3321 static int __init blk_mq_init(void)
3323 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3324 blk_mq_hctx_notify_dead);
3327 subsys_initcall(blk_mq_init);