2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-debugfs.h"
35 #include "blk-mq-tag.h"
38 #include "blk-mq-sched.h"
40 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie);
41 static void blk_mq_poll_stats_start(struct request_queue *q);
42 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
44 static int blk_mq_poll_stats_bkt(const struct request *rq)
46 int ddir, bytes, bucket;
48 ddir = rq_data_dir(rq);
49 bytes = blk_rq_bytes(rq);
51 bucket = ddir + 2*(ilog2(bytes) - 9);
55 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
56 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
62 * Check if any of the ctx's have pending work in this hardware queue
64 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
66 return !list_empty_careful(&hctx->dispatch) ||
67 sbitmap_any_bit_set(&hctx->ctx_map) ||
68 blk_mq_sched_has_work(hctx);
72 * Mark this ctx as having pending work in this hardware queue
74 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
75 struct blk_mq_ctx *ctx)
77 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
78 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
81 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
82 struct blk_mq_ctx *ctx)
84 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
88 struct hd_struct *part;
89 unsigned int *inflight;
92 static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
93 struct request *rq, void *priv,
96 struct mq_inflight *mi = priv;
98 if (blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) {
100 * index[0] counts the specific partition that was asked
101 * for. index[1] counts the ones that are active on the
102 * whole device, so increment that if mi->part is indeed
103 * a partition, and not a whole device.
105 if (rq->part == mi->part)
107 if (mi->part->partno)
112 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
113 unsigned int inflight[2])
115 struct mq_inflight mi = { .part = part, .inflight = inflight, };
117 inflight[0] = inflight[1] = 0;
118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
121 void blk_freeze_queue_start(struct request_queue *q)
125 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
126 if (freeze_depth == 1) {
127 percpu_ref_kill(&q->q_usage_counter);
129 blk_mq_run_hw_queues(q, false);
132 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
134 void blk_mq_freeze_queue_wait(struct request_queue *q)
136 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
138 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
140 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
141 unsigned long timeout)
143 return wait_event_timeout(q->mq_freeze_wq,
144 percpu_ref_is_zero(&q->q_usage_counter),
147 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
150 * Guarantee no request is in use, so we can change any data structure of
151 * the queue afterward.
153 void blk_freeze_queue(struct request_queue *q)
156 * In the !blk_mq case we are only calling this to kill the
157 * q_usage_counter, otherwise this increases the freeze depth
158 * and waits for it to return to zero. For this reason there is
159 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
160 * exported to drivers as the only user for unfreeze is blk_mq.
162 blk_freeze_queue_start(q);
163 blk_mq_freeze_queue_wait(q);
166 void blk_mq_freeze_queue(struct request_queue *q)
169 * ...just an alias to keep freeze and unfreeze actions balanced
170 * in the blk_mq_* namespace
174 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
176 void blk_mq_unfreeze_queue(struct request_queue *q)
180 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
181 WARN_ON_ONCE(freeze_depth < 0);
183 percpu_ref_reinit(&q->q_usage_counter);
184 wake_up_all(&q->mq_freeze_wq);
187 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
190 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
191 * mpt3sas driver such that this function can be removed.
193 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
197 spin_lock_irqsave(q->queue_lock, flags);
198 queue_flag_set(QUEUE_FLAG_QUIESCED, q);
199 spin_unlock_irqrestore(q->queue_lock, flags);
201 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
204 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
207 * Note: this function does not prevent that the struct request end_io()
208 * callback function is invoked. Once this function is returned, we make
209 * sure no dispatch can happen until the queue is unquiesced via
210 * blk_mq_unquiesce_queue().
212 void blk_mq_quiesce_queue(struct request_queue *q)
214 struct blk_mq_hw_ctx *hctx;
218 blk_mq_quiesce_queue_nowait(q);
220 queue_for_each_hw_ctx(q, hctx, i) {
221 if (hctx->flags & BLK_MQ_F_BLOCKING)
222 synchronize_srcu(hctx->queue_rq_srcu);
229 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
232 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
235 * This function recovers queue into the state before quiescing
236 * which is done by blk_mq_quiesce_queue.
238 void blk_mq_unquiesce_queue(struct request_queue *q)
242 spin_lock_irqsave(q->queue_lock, flags);
243 queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
244 spin_unlock_irqrestore(q->queue_lock, flags);
246 /* dispatch requests which are inserted during quiescing */
247 blk_mq_run_hw_queues(q, true);
249 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
251 void blk_mq_wake_waiters(struct request_queue *q)
253 struct blk_mq_hw_ctx *hctx;
256 queue_for_each_hw_ctx(q, hctx, i)
257 if (blk_mq_hw_queue_mapped(hctx))
258 blk_mq_tag_wakeup_all(hctx->tags, true);
261 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
263 return blk_mq_has_free_tags(hctx->tags);
265 EXPORT_SYMBOL(blk_mq_can_queue);
267 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
268 unsigned int tag, unsigned int op)
270 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
271 struct request *rq = tags->static_rqs[tag];
275 if (data->flags & BLK_MQ_REQ_INTERNAL) {
277 rq->internal_tag = tag;
279 if (blk_mq_tag_busy(data->hctx)) {
280 rq->rq_flags = RQF_MQ_INFLIGHT;
281 atomic_inc(&data->hctx->nr_active);
284 rq->internal_tag = -1;
285 data->hctx->tags->rqs[rq->tag] = rq;
288 INIT_LIST_HEAD(&rq->queuelist);
289 /* csd/requeue_work/fifo_time is initialized before use */
291 rq->mq_ctx = data->ctx;
293 if (data->flags & BLK_MQ_REQ_PREEMPT)
294 rq->rq_flags |= RQF_PREEMPT;
295 if (blk_queue_io_stat(data->q))
296 rq->rq_flags |= RQF_IO_STAT;
297 /* do not touch atomic flags, it needs atomic ops against the timer */
299 INIT_HLIST_NODE(&rq->hash);
300 RB_CLEAR_NODE(&rq->rb_node);
303 rq->start_time = jiffies;
304 #ifdef CONFIG_BLK_CGROUP
306 set_start_time_ns(rq);
307 rq->io_start_time_ns = 0;
309 rq->nr_phys_segments = 0;
310 #if defined(CONFIG_BLK_DEV_INTEGRITY)
311 rq->nr_integrity_segments = 0;
314 /* tag was already set */
317 INIT_LIST_HEAD(&rq->timeout_list);
321 rq->end_io_data = NULL;
324 data->ctx->rq_dispatched[op_is_sync(op)]++;
328 static struct request *blk_mq_get_request(struct request_queue *q,
329 struct bio *bio, unsigned int op,
330 struct blk_mq_alloc_data *data)
332 struct elevator_queue *e = q->elevator;
335 bool put_ctx_on_error = false;
337 blk_queue_enter_live(q);
339 if (likely(!data->ctx)) {
340 data->ctx = blk_mq_get_ctx(q);
341 put_ctx_on_error = true;
343 if (likely(!data->hctx))
344 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
346 data->flags |= BLK_MQ_REQ_NOWAIT;
349 data->flags |= BLK_MQ_REQ_INTERNAL;
352 * Flush requests are special and go directly to the
355 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
356 e->type->ops.mq.limit_depth(op, data);
359 tag = blk_mq_get_tag(data);
360 if (tag == BLK_MQ_TAG_FAIL) {
361 if (put_ctx_on_error) {
362 blk_mq_put_ctx(data->ctx);
369 rq = blk_mq_rq_ctx_init(data, tag, op);
370 if (!op_is_flush(op)) {
372 if (e && e->type->ops.mq.prepare_request) {
373 if (e->type->icq_cache && rq_ioc(bio))
374 blk_mq_sched_assign_ioc(rq, bio);
376 e->type->ops.mq.prepare_request(rq, bio);
377 rq->rq_flags |= RQF_ELVPRIV;
380 data->hctx->queued++;
384 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
385 blk_mq_req_flags_t flags)
387 struct blk_mq_alloc_data alloc_data = { .flags = flags };
391 ret = blk_queue_enter(q, flags);
395 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
399 return ERR_PTR(-EWOULDBLOCK);
401 blk_mq_put_ctx(alloc_data.ctx);
404 rq->__sector = (sector_t) -1;
405 rq->bio = rq->biotail = NULL;
408 EXPORT_SYMBOL(blk_mq_alloc_request);
410 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
411 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
413 struct blk_mq_alloc_data alloc_data = { .flags = flags };
419 * If the tag allocator sleeps we could get an allocation for a
420 * different hardware context. No need to complicate the low level
421 * allocator for this for the rare use case of a command tied to
424 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
425 return ERR_PTR(-EINVAL);
427 if (hctx_idx >= q->nr_hw_queues)
428 return ERR_PTR(-EIO);
430 ret = blk_queue_enter(q, flags);
435 * Check if the hardware context is actually mapped to anything.
436 * If not tell the caller that it should skip this queue.
438 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
439 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
441 return ERR_PTR(-EXDEV);
443 cpu = cpumask_first(alloc_data.hctx->cpumask);
444 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
446 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
450 return ERR_PTR(-EWOULDBLOCK);
454 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
456 void blk_mq_free_request(struct request *rq)
458 struct request_queue *q = rq->q;
459 struct elevator_queue *e = q->elevator;
460 struct blk_mq_ctx *ctx = rq->mq_ctx;
461 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
462 const int sched_tag = rq->internal_tag;
464 if (rq->rq_flags & RQF_ELVPRIV) {
465 if (e && e->type->ops.mq.finish_request)
466 e->type->ops.mq.finish_request(rq);
468 put_io_context(rq->elv.icq->ioc);
473 ctx->rq_completed[rq_is_sync(rq)]++;
474 if (rq->rq_flags & RQF_MQ_INFLIGHT)
475 atomic_dec(&hctx->nr_active);
477 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
478 laptop_io_completion(q->backing_dev_info);
480 wbt_done(q->rq_wb, &rq->issue_stat);
483 blk_put_rl(blk_rq_rl(rq));
485 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
486 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
487 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
489 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
491 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
492 blk_mq_sched_restart(hctx);
495 EXPORT_SYMBOL_GPL(blk_mq_free_request);
497 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
499 blk_account_io_done(rq);
502 wbt_done(rq->q->rq_wb, &rq->issue_stat);
503 rq->end_io(rq, error);
505 if (unlikely(blk_bidi_rq(rq)))
506 blk_mq_free_request(rq->next_rq);
507 blk_mq_free_request(rq);
510 EXPORT_SYMBOL(__blk_mq_end_request);
512 void blk_mq_end_request(struct request *rq, blk_status_t error)
514 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
516 __blk_mq_end_request(rq, error);
518 EXPORT_SYMBOL(blk_mq_end_request);
520 static void __blk_mq_complete_request_remote(void *data)
522 struct request *rq = data;
524 rq->q->softirq_done_fn(rq);
527 static void __blk_mq_complete_request(struct request *rq)
529 struct blk_mq_ctx *ctx = rq->mq_ctx;
533 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT);
535 if (rq->internal_tag != -1)
536 blk_mq_sched_completed_request(rq);
537 if (rq->rq_flags & RQF_STATS) {
538 blk_mq_poll_stats_start(rq->q);
542 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
543 rq->q->softirq_done_fn(rq);
548 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
549 shared = cpus_share_cache(cpu, ctx->cpu);
551 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
552 rq->csd.func = __blk_mq_complete_request_remote;
555 smp_call_function_single_async(ctx->cpu, &rq->csd);
557 rq->q->softirq_done_fn(rq);
562 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
564 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
567 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
570 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
572 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
575 *srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
578 static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate)
583 * blk_mq_rq_aborted_gstate() is used from the completion path and
584 * can thus be called from irq context. u64_stats_fetch in the
585 * middle of update on the same CPU leads to lockup. Disable irq
588 local_irq_save(flags);
589 u64_stats_update_begin(&rq->aborted_gstate_sync);
590 rq->aborted_gstate = gstate;
591 u64_stats_update_end(&rq->aborted_gstate_sync);
592 local_irq_restore(flags);
595 static u64 blk_mq_rq_aborted_gstate(struct request *rq)
601 start = u64_stats_fetch_begin(&rq->aborted_gstate_sync);
602 aborted_gstate = rq->aborted_gstate;
603 } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start));
605 return aborted_gstate;
609 * blk_mq_complete_request - end I/O on a request
610 * @rq: the request being processed
613 * Ends all I/O on a request. It does not handle partial completions.
614 * The actual completion happens out-of-order, through a IPI handler.
616 void blk_mq_complete_request(struct request *rq)
618 struct request_queue *q = rq->q;
619 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
622 if (unlikely(blk_should_fake_timeout(q)))
626 * If @rq->aborted_gstate equals the current instance, timeout is
627 * claiming @rq and we lost. This is synchronized through
628 * hctx_lock(). See blk_mq_timeout_work() for details.
630 * Completion path never blocks and we can directly use RCU here
631 * instead of hctx_lock() which can be either RCU or SRCU.
632 * However, that would complicate paths which want to synchronize
633 * against us. Let stay in sync with the issue path so that
634 * hctx_lock() covers both issue and completion paths.
636 hctx_lock(hctx, &srcu_idx);
637 if (blk_mq_rq_aborted_gstate(rq) != rq->gstate &&
638 !blk_mark_rq_complete(rq))
639 __blk_mq_complete_request(rq);
640 hctx_unlock(hctx, srcu_idx);
642 EXPORT_SYMBOL(blk_mq_complete_request);
644 int blk_mq_request_started(struct request *rq)
646 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
648 EXPORT_SYMBOL_GPL(blk_mq_request_started);
650 void blk_mq_start_request(struct request *rq)
652 struct request_queue *q = rq->q;
654 blk_mq_sched_started_request(rq);
656 trace_block_rq_issue(q, rq);
658 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
659 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
660 rq->rq_flags |= RQF_STATS;
661 wbt_issue(q->rq_wb, &rq->issue_stat);
664 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
665 WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags));
668 * Mark @rq in-flight which also advances the generation number,
669 * and register for timeout. Protect with a seqcount to allow the
670 * timeout path to read both @rq->gstate and @rq->deadline
673 * This is the only place where a request is marked in-flight. If
674 * the timeout path reads an in-flight @rq->gstate, the
675 * @rq->deadline it reads together under @rq->gstate_seq is
676 * guaranteed to be the matching one.
679 write_seqcount_begin(&rq->gstate_seq);
681 blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT);
684 write_seqcount_end(&rq->gstate_seq);
687 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
688 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
689 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
691 if (q->dma_drain_size && blk_rq_bytes(rq)) {
693 * Make sure space for the drain appears. We know we can do
694 * this because max_hw_segments has been adjusted to be one
695 * fewer than the device can handle.
697 rq->nr_phys_segments++;
700 EXPORT_SYMBOL(blk_mq_start_request);
703 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
704 * flag isn't set yet, so there may be race with timeout handler,
705 * but given rq->deadline is just set in .queue_rq() under
706 * this situation, the race won't be possible in reality because
707 * rq->timeout should be set as big enough to cover the window
708 * between blk_mq_start_request() called from .queue_rq() and
709 * clearing REQ_ATOM_STARTED here.
711 static void __blk_mq_requeue_request(struct request *rq)
713 struct request_queue *q = rq->q;
715 blk_mq_put_driver_tag(rq);
717 trace_block_rq_requeue(q, rq);
718 wbt_requeue(q->rq_wb, &rq->issue_stat);
719 blk_mq_sched_requeue_request(rq);
721 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
722 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
723 if (q->dma_drain_size && blk_rq_bytes(rq))
724 rq->nr_phys_segments--;
728 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
730 __blk_mq_requeue_request(rq);
732 BUG_ON(blk_queued_rq(rq));
733 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
735 EXPORT_SYMBOL(blk_mq_requeue_request);
737 static void blk_mq_requeue_work(struct work_struct *work)
739 struct request_queue *q =
740 container_of(work, struct request_queue, requeue_work.work);
742 struct request *rq, *next;
744 spin_lock_irq(&q->requeue_lock);
745 list_splice_init(&q->requeue_list, &rq_list);
746 spin_unlock_irq(&q->requeue_lock);
748 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
749 if (!(rq->rq_flags & RQF_SOFTBARRIER))
752 rq->rq_flags &= ~RQF_SOFTBARRIER;
753 list_del_init(&rq->queuelist);
754 blk_mq_sched_insert_request(rq, true, false, false, true);
757 while (!list_empty(&rq_list)) {
758 rq = list_entry(rq_list.next, struct request, queuelist);
759 list_del_init(&rq->queuelist);
760 blk_mq_sched_insert_request(rq, false, false, false, true);
763 blk_mq_run_hw_queues(q, false);
766 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
767 bool kick_requeue_list)
769 struct request_queue *q = rq->q;
773 * We abuse this flag that is otherwise used by the I/O scheduler to
774 * request head insertion from the workqueue.
776 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
778 spin_lock_irqsave(&q->requeue_lock, flags);
780 rq->rq_flags |= RQF_SOFTBARRIER;
781 list_add(&rq->queuelist, &q->requeue_list);
783 list_add_tail(&rq->queuelist, &q->requeue_list);
785 spin_unlock_irqrestore(&q->requeue_lock, flags);
787 if (kick_requeue_list)
788 blk_mq_kick_requeue_list(q);
790 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
792 void blk_mq_kick_requeue_list(struct request_queue *q)
794 kblockd_schedule_delayed_work(&q->requeue_work, 0);
796 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
798 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
801 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
802 msecs_to_jiffies(msecs));
804 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
806 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
808 if (tag < tags->nr_tags) {
809 prefetch(tags->rqs[tag]);
810 return tags->rqs[tag];
815 EXPORT_SYMBOL(blk_mq_tag_to_rq);
817 struct blk_mq_timeout_data {
819 unsigned int next_set;
820 unsigned int nr_expired;
823 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
825 const struct blk_mq_ops *ops = req->q->mq_ops;
826 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
829 * We know that complete is set at this point. If STARTED isn't set
830 * anymore, then the request isn't active and the "timeout" should
831 * just be ignored. This can happen due to the bitflag ordering.
832 * Timeout first checks if STARTED is set, and if it is, assumes
833 * the request is active. But if we race with completion, then
834 * both flags will get cleared. So check here again, and ignore
835 * a timeout event with a request that isn't active.
837 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
841 ret = ops->timeout(req, reserved);
845 __blk_mq_complete_request(req);
847 case BLK_EH_RESET_TIMER:
849 * As nothing prevents from completion happening while
850 * ->aborted_gstate is set, this may lead to ignored
851 * completions and further spurious timeouts.
853 blk_mq_rq_update_aborted_gstate(req, 0);
855 blk_clear_rq_complete(req);
857 case BLK_EH_NOT_HANDLED:
860 printk(KERN_ERR "block: bad eh return: %d\n", ret);
865 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
866 struct request *rq, void *priv, bool reserved)
868 struct blk_mq_timeout_data *data = priv;
869 unsigned long gstate, deadline;
874 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
877 /* read coherent snapshots of @rq->state_gen and @rq->deadline */
879 start = read_seqcount_begin(&rq->gstate_seq);
880 gstate = READ_ONCE(rq->gstate);
881 deadline = rq->deadline;
882 if (!read_seqcount_retry(&rq->gstate_seq, start))
887 /* if in-flight && overdue, mark for abortion */
888 if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT &&
889 time_after_eq(jiffies, deadline)) {
890 blk_mq_rq_update_aborted_gstate(rq, gstate);
893 } else if (!data->next_set || time_after(data->next, deadline)) {
894 data->next = deadline;
899 static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx,
900 struct request *rq, void *priv, bool reserved)
903 * We marked @rq->aborted_gstate and waited for RCU. If there were
904 * completions that we lost to, they would have finished and
905 * updated @rq->gstate by now; otherwise, the completion path is
906 * now guaranteed to see @rq->aborted_gstate and yield. If
907 * @rq->aborted_gstate still matches @rq->gstate, @rq is ours.
909 if (READ_ONCE(rq->gstate) == rq->aborted_gstate &&
910 !blk_mark_rq_complete(rq))
911 blk_mq_rq_timed_out(rq, reserved);
914 static void blk_mq_timeout_work(struct work_struct *work)
916 struct request_queue *q =
917 container_of(work, struct request_queue, timeout_work);
918 struct blk_mq_timeout_data data = {
923 struct blk_mq_hw_ctx *hctx;
926 /* A deadlock might occur if a request is stuck requiring a
927 * timeout at the same time a queue freeze is waiting
928 * completion, since the timeout code would not be able to
929 * acquire the queue reference here.
931 * That's why we don't use blk_queue_enter here; instead, we use
932 * percpu_ref_tryget directly, because we need to be able to
933 * obtain a reference even in the short window between the queue
934 * starting to freeze, by dropping the first reference in
935 * blk_freeze_queue_start, and the moment the last request is
936 * consumed, marked by the instant q_usage_counter reaches
939 if (!percpu_ref_tryget(&q->q_usage_counter))
942 /* scan for the expired ones and set their ->aborted_gstate */
943 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
945 if (data.nr_expired) {
946 bool has_rcu = false;
949 * Wait till everyone sees ->aborted_gstate. The
950 * sequential waits for SRCUs aren't ideal. If this ever
951 * becomes a problem, we can add per-hw_ctx rcu_head and
954 queue_for_each_hw_ctx(q, hctx, i) {
955 if (!hctx->nr_expired)
958 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
961 synchronize_srcu(hctx->queue_rq_srcu);
963 hctx->nr_expired = 0;
968 /* terminate the ones we won */
969 blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL);
973 data.next = blk_rq_timeout(round_jiffies_up(data.next));
974 mod_timer(&q->timeout, data.next);
976 queue_for_each_hw_ctx(q, hctx, i) {
977 /* the hctx may be unmapped, so check it here */
978 if (blk_mq_hw_queue_mapped(hctx))
979 blk_mq_tag_idle(hctx);
985 struct flush_busy_ctx_data {
986 struct blk_mq_hw_ctx *hctx;
987 struct list_head *list;
990 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
992 struct flush_busy_ctx_data *flush_data = data;
993 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
994 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
996 sbitmap_clear_bit(sb, bitnr);
997 spin_lock(&ctx->lock);
998 list_splice_tail_init(&ctx->rq_list, flush_data->list);
999 spin_unlock(&ctx->lock);
1004 * Process software queues that have been marked busy, splicing them
1005 * to the for-dispatch
1007 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1009 struct flush_busy_ctx_data data = {
1014 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1016 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1018 struct dispatch_rq_data {
1019 struct blk_mq_hw_ctx *hctx;
1023 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1026 struct dispatch_rq_data *dispatch_data = data;
1027 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1028 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1030 spin_lock(&ctx->lock);
1031 if (unlikely(!list_empty(&ctx->rq_list))) {
1032 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
1033 list_del_init(&dispatch_data->rq->queuelist);
1034 if (list_empty(&ctx->rq_list))
1035 sbitmap_clear_bit(sb, bitnr);
1037 spin_unlock(&ctx->lock);
1039 return !dispatch_data->rq;
1042 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1043 struct blk_mq_ctx *start)
1045 unsigned off = start ? start->index_hw : 0;
1046 struct dispatch_rq_data data = {
1051 __sbitmap_for_each_set(&hctx->ctx_map, off,
1052 dispatch_rq_from_ctx, &data);
1057 static inline unsigned int queued_to_index(unsigned int queued)
1062 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1065 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
1068 struct blk_mq_alloc_data data = {
1070 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
1071 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
1074 might_sleep_if(wait);
1079 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1080 data.flags |= BLK_MQ_REQ_RESERVED;
1082 rq->tag = blk_mq_get_tag(&data);
1084 if (blk_mq_tag_busy(data.hctx)) {
1085 rq->rq_flags |= RQF_MQ_INFLIGHT;
1086 atomic_inc(&data.hctx->nr_active);
1088 data.hctx->tags->rqs[rq->tag] = rq;
1094 return rq->tag != -1;
1097 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1098 int flags, void *key)
1100 struct blk_mq_hw_ctx *hctx;
1102 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1104 list_del_init(&wait->entry);
1105 blk_mq_run_hw_queue(hctx, true);
1110 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1111 * the tag wakeups. For non-shared tags, we can simply mark us nedeing a
1112 * restart. For both caes, take care to check the condition again after
1113 * marking us as waiting.
1115 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1118 struct blk_mq_hw_ctx *this_hctx = *hctx;
1119 bool shared_tags = (this_hctx->flags & BLK_MQ_F_TAG_SHARED) != 0;
1120 struct sbq_wait_state *ws;
1121 wait_queue_entry_t *wait;
1125 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1126 set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1128 wait = &this_hctx->dispatch_wait;
1129 if (!list_empty_careful(&wait->entry))
1132 spin_lock(&this_hctx->lock);
1133 if (!list_empty(&wait->entry)) {
1134 spin_unlock(&this_hctx->lock);
1138 ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1139 add_wait_queue(&ws->wait, wait);
1143 * It's possible that a tag was freed in the window between the
1144 * allocation failure and adding the hardware queue to the wait
1147 ret = blk_mq_get_driver_tag(rq, hctx, false);
1151 * Don't clear RESTART here, someone else could have set it.
1152 * At most this will cost an extra queue run.
1157 spin_unlock(&this_hctx->lock);
1162 * We got a tag, remove ourselves from the wait queue to ensure
1163 * someone else gets the wakeup.
1165 spin_lock_irq(&ws->wait.lock);
1166 list_del_init(&wait->entry);
1167 spin_unlock_irq(&ws->wait.lock);
1168 spin_unlock(&this_hctx->lock);
1173 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1176 struct blk_mq_hw_ctx *hctx;
1177 struct request *rq, *nxt;
1178 bool no_tag = false;
1181 if (list_empty(list))
1184 WARN_ON(!list_is_singular(list) && got_budget);
1187 * Now process all the entries, sending them to the driver.
1189 errors = queued = 0;
1191 struct blk_mq_queue_data bd;
1194 rq = list_first_entry(list, struct request, queuelist);
1195 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1197 * The initial allocation attempt failed, so we need to
1198 * rerun the hardware queue when a tag is freed. The
1199 * waitqueue takes care of that. If the queue is run
1200 * before we add this entry back on the dispatch list,
1201 * we'll re-run it below.
1203 if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1205 blk_mq_put_dispatch_budget(hctx);
1207 * For non-shared tags, the RESTART check
1210 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1216 if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) {
1217 blk_mq_put_driver_tag(rq);
1221 list_del_init(&rq->queuelist);
1226 * Flag last if we have no more requests, or if we have more
1227 * but can't assign a driver tag to it.
1229 if (list_empty(list))
1232 nxt = list_first_entry(list, struct request, queuelist);
1233 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1236 ret = q->mq_ops->queue_rq(hctx, &bd);
1237 if (ret == BLK_STS_RESOURCE) {
1239 * If an I/O scheduler has been configured and we got a
1240 * driver tag for the next request already, free it
1243 if (!list_empty(list)) {
1244 nxt = list_first_entry(list, struct request, queuelist);
1245 blk_mq_put_driver_tag(nxt);
1247 list_add(&rq->queuelist, list);
1248 __blk_mq_requeue_request(rq);
1252 if (unlikely(ret != BLK_STS_OK)) {
1254 blk_mq_end_request(rq, BLK_STS_IOERR);
1259 } while (!list_empty(list));
1261 hctx->dispatched[queued_to_index(queued)]++;
1264 * Any items that need requeuing? Stuff them into hctx->dispatch,
1265 * that is where we will continue on next queue run.
1267 if (!list_empty(list)) {
1268 spin_lock(&hctx->lock);
1269 list_splice_init(list, &hctx->dispatch);
1270 spin_unlock(&hctx->lock);
1273 * If SCHED_RESTART was set by the caller of this function and
1274 * it is no longer set that means that it was cleared by another
1275 * thread and hence that a queue rerun is needed.
1277 * If 'no_tag' is set, that means that we failed getting
1278 * a driver tag with an I/O scheduler attached. If our dispatch
1279 * waitqueue is no longer active, ensure that we run the queue
1280 * AFTER adding our entries back to the list.
1282 * If no I/O scheduler has been configured it is possible that
1283 * the hardware queue got stopped and restarted before requests
1284 * were pushed back onto the dispatch list. Rerun the queue to
1285 * avoid starvation. Notes:
1286 * - blk_mq_run_hw_queue() checks whether or not a queue has
1287 * been stopped before rerunning a queue.
1288 * - Some but not all block drivers stop a queue before
1289 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1292 if (!blk_mq_sched_needs_restart(hctx) ||
1293 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1294 blk_mq_run_hw_queue(hctx, true);
1297 return (queued + errors) != 0;
1300 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1305 * We should be running this queue from one of the CPUs that
1308 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1309 cpu_online(hctx->next_cpu));
1312 * We can't run the queue inline with ints disabled. Ensure that
1313 * we catch bad users of this early.
1315 WARN_ON_ONCE(in_interrupt());
1317 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1319 hctx_lock(hctx, &srcu_idx);
1320 blk_mq_sched_dispatch_requests(hctx);
1321 hctx_unlock(hctx, srcu_idx);
1325 * It'd be great if the workqueue API had a way to pass
1326 * in a mask and had some smarts for more clever placement.
1327 * For now we just round-robin here, switching for every
1328 * BLK_MQ_CPU_WORK_BATCH queued items.
1330 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1332 if (hctx->queue->nr_hw_queues == 1)
1333 return WORK_CPU_UNBOUND;
1335 if (--hctx->next_cpu_batch <= 0) {
1338 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1339 if (next_cpu >= nr_cpu_ids)
1340 next_cpu = cpumask_first(hctx->cpumask);
1342 hctx->next_cpu = next_cpu;
1343 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1346 return hctx->next_cpu;
1349 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1350 unsigned long msecs)
1352 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1355 if (unlikely(blk_mq_hctx_stopped(hctx)))
1358 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1359 int cpu = get_cpu();
1360 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1361 __blk_mq_run_hw_queue(hctx);
1369 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1371 msecs_to_jiffies(msecs));
1374 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1376 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1378 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1380 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1386 * When queue is quiesced, we may be switching io scheduler, or
1387 * updating nr_hw_queues, or other things, and we can't run queue
1388 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1390 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1393 hctx_lock(hctx, &srcu_idx);
1394 need_run = !blk_queue_quiesced(hctx->queue) &&
1395 blk_mq_hctx_has_pending(hctx);
1396 hctx_unlock(hctx, srcu_idx);
1399 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1405 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1407 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1409 struct blk_mq_hw_ctx *hctx;
1412 queue_for_each_hw_ctx(q, hctx, i) {
1413 if (blk_mq_hctx_stopped(hctx))
1416 blk_mq_run_hw_queue(hctx, async);
1419 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1422 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1423 * @q: request queue.
1425 * The caller is responsible for serializing this function against
1426 * blk_mq_{start,stop}_hw_queue().
1428 bool blk_mq_queue_stopped(struct request_queue *q)
1430 struct blk_mq_hw_ctx *hctx;
1433 queue_for_each_hw_ctx(q, hctx, i)
1434 if (blk_mq_hctx_stopped(hctx))
1439 EXPORT_SYMBOL(blk_mq_queue_stopped);
1442 * This function is often used for pausing .queue_rq() by driver when
1443 * there isn't enough resource or some conditions aren't satisfied, and
1444 * BLK_STS_RESOURCE is usually returned.
1446 * We do not guarantee that dispatch can be drained or blocked
1447 * after blk_mq_stop_hw_queue() returns. Please use
1448 * blk_mq_quiesce_queue() for that requirement.
1450 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1452 cancel_delayed_work(&hctx->run_work);
1454 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1456 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1459 * This function is often used for pausing .queue_rq() by driver when
1460 * there isn't enough resource or some conditions aren't satisfied, and
1461 * BLK_STS_RESOURCE is usually returned.
1463 * We do not guarantee that dispatch can be drained or blocked
1464 * after blk_mq_stop_hw_queues() returns. Please use
1465 * blk_mq_quiesce_queue() for that requirement.
1467 void blk_mq_stop_hw_queues(struct request_queue *q)
1469 struct blk_mq_hw_ctx *hctx;
1472 queue_for_each_hw_ctx(q, hctx, i)
1473 blk_mq_stop_hw_queue(hctx);
1475 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1477 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1479 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1481 blk_mq_run_hw_queue(hctx, false);
1483 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1485 void blk_mq_start_hw_queues(struct request_queue *q)
1487 struct blk_mq_hw_ctx *hctx;
1490 queue_for_each_hw_ctx(q, hctx, i)
1491 blk_mq_start_hw_queue(hctx);
1493 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1495 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1497 if (!blk_mq_hctx_stopped(hctx))
1500 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1501 blk_mq_run_hw_queue(hctx, async);
1503 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1505 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1507 struct blk_mq_hw_ctx *hctx;
1510 queue_for_each_hw_ctx(q, hctx, i)
1511 blk_mq_start_stopped_hw_queue(hctx, async);
1513 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1515 static void blk_mq_run_work_fn(struct work_struct *work)
1517 struct blk_mq_hw_ctx *hctx;
1519 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1522 * If we are stopped, don't run the queue. The exception is if
1523 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1524 * the STOPPED bit and run it.
1526 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1527 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1530 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1531 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1534 __blk_mq_run_hw_queue(hctx);
1538 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1540 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1544 * Stop the hw queue, then modify currently delayed work.
1545 * This should prevent us from running the queue prematurely.
1546 * Mark the queue as auto-clearing STOPPED when it runs.
1548 blk_mq_stop_hw_queue(hctx);
1549 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1550 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1552 msecs_to_jiffies(msecs));
1554 EXPORT_SYMBOL(blk_mq_delay_queue);
1556 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1560 struct blk_mq_ctx *ctx = rq->mq_ctx;
1562 lockdep_assert_held(&ctx->lock);
1564 trace_block_rq_insert(hctx->queue, rq);
1567 list_add(&rq->queuelist, &ctx->rq_list);
1569 list_add_tail(&rq->queuelist, &ctx->rq_list);
1572 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1575 struct blk_mq_ctx *ctx = rq->mq_ctx;
1577 lockdep_assert_held(&ctx->lock);
1579 __blk_mq_insert_req_list(hctx, rq, at_head);
1580 blk_mq_hctx_mark_pending(hctx, ctx);
1584 * Should only be used carefully, when the caller knows we want to
1585 * bypass a potential IO scheduler on the target device.
1587 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1589 struct blk_mq_ctx *ctx = rq->mq_ctx;
1590 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1592 spin_lock(&hctx->lock);
1593 list_add_tail(&rq->queuelist, &hctx->dispatch);
1594 spin_unlock(&hctx->lock);
1597 blk_mq_run_hw_queue(hctx, false);
1600 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1601 struct list_head *list)
1605 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1608 spin_lock(&ctx->lock);
1609 while (!list_empty(list)) {
1612 rq = list_first_entry(list, struct request, queuelist);
1613 BUG_ON(rq->mq_ctx != ctx);
1614 list_del_init(&rq->queuelist);
1615 __blk_mq_insert_req_list(hctx, rq, false);
1617 blk_mq_hctx_mark_pending(hctx, ctx);
1618 spin_unlock(&ctx->lock);
1621 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1623 struct request *rqa = container_of(a, struct request, queuelist);
1624 struct request *rqb = container_of(b, struct request, queuelist);
1626 return !(rqa->mq_ctx < rqb->mq_ctx ||
1627 (rqa->mq_ctx == rqb->mq_ctx &&
1628 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1631 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1633 struct blk_mq_ctx *this_ctx;
1634 struct request_queue *this_q;
1637 LIST_HEAD(ctx_list);
1640 list_splice_init(&plug->mq_list, &list);
1642 list_sort(NULL, &list, plug_ctx_cmp);
1648 while (!list_empty(&list)) {
1649 rq = list_entry_rq(list.next);
1650 list_del_init(&rq->queuelist);
1652 if (rq->mq_ctx != this_ctx) {
1654 trace_block_unplug(this_q, depth, from_schedule);
1655 blk_mq_sched_insert_requests(this_q, this_ctx,
1660 this_ctx = rq->mq_ctx;
1666 list_add_tail(&rq->queuelist, &ctx_list);
1670 * If 'this_ctx' is set, we know we have entries to complete
1671 * on 'ctx_list'. Do those.
1674 trace_block_unplug(this_q, depth, from_schedule);
1675 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1680 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1682 blk_init_request_from_bio(rq, bio);
1684 blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1686 blk_account_io_start(rq, true);
1689 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1690 struct blk_mq_ctx *ctx,
1693 spin_lock(&ctx->lock);
1694 __blk_mq_insert_request(hctx, rq, false);
1695 spin_unlock(&ctx->lock);
1698 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1701 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1703 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1706 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1710 struct request_queue *q = rq->q;
1711 struct blk_mq_queue_data bd = {
1715 blk_qc_t new_cookie;
1717 bool run_queue = true;
1719 /* RCU or SRCU read lock is needed before checking quiesced flag */
1720 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1728 if (!blk_mq_get_driver_tag(rq, NULL, false))
1731 if (!blk_mq_get_dispatch_budget(hctx)) {
1732 blk_mq_put_driver_tag(rq);
1736 new_cookie = request_to_qc_t(hctx, rq);
1739 * For OK queue, we are done. For error, kill it. Any other
1740 * error (busy), just add it to our list as we previously
1743 ret = q->mq_ops->queue_rq(hctx, &bd);
1746 *cookie = new_cookie;
1748 case BLK_STS_RESOURCE:
1749 __blk_mq_requeue_request(rq);
1752 *cookie = BLK_QC_T_NONE;
1753 blk_mq_end_request(rq, ret);
1758 blk_mq_sched_insert_request(rq, false, run_queue, false,
1759 hctx->flags & BLK_MQ_F_BLOCKING);
1762 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1763 struct request *rq, blk_qc_t *cookie)
1767 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1769 hctx_lock(hctx, &srcu_idx);
1770 __blk_mq_try_issue_directly(hctx, rq, cookie);
1771 hctx_unlock(hctx, srcu_idx);
1774 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1776 const int is_sync = op_is_sync(bio->bi_opf);
1777 const int is_flush_fua = op_is_flush(bio->bi_opf);
1778 struct blk_mq_alloc_data data = { .flags = 0 };
1780 unsigned int request_count = 0;
1781 struct blk_plug *plug;
1782 struct request *same_queue_rq = NULL;
1784 unsigned int wb_acct;
1786 blk_queue_bounce(q, &bio);
1788 blk_queue_split(q, &bio);
1790 if (!bio_integrity_prep(bio))
1791 return BLK_QC_T_NONE;
1793 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1794 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1795 return BLK_QC_T_NONE;
1797 if (blk_mq_sched_bio_merge(q, bio))
1798 return BLK_QC_T_NONE;
1800 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1802 trace_block_getrq(q, bio, bio->bi_opf);
1804 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1805 if (unlikely(!rq)) {
1806 __wbt_done(q->rq_wb, wb_acct);
1807 if (bio->bi_opf & REQ_NOWAIT)
1808 bio_wouldblock_error(bio);
1809 return BLK_QC_T_NONE;
1812 wbt_track(&rq->issue_stat, wb_acct);
1814 cookie = request_to_qc_t(data.hctx, rq);
1816 plug = current->plug;
1817 if (unlikely(is_flush_fua)) {
1818 blk_mq_put_ctx(data.ctx);
1819 blk_mq_bio_to_request(rq, bio);
1821 /* bypass scheduler for flush rq */
1822 blk_insert_flush(rq);
1823 blk_mq_run_hw_queue(data.hctx, true);
1824 } else if (plug && q->nr_hw_queues == 1) {
1825 struct request *last = NULL;
1827 blk_mq_put_ctx(data.ctx);
1828 blk_mq_bio_to_request(rq, bio);
1831 * @request_count may become stale because of schedule
1832 * out, so check the list again.
1834 if (list_empty(&plug->mq_list))
1836 else if (blk_queue_nomerges(q))
1837 request_count = blk_plug_queued_count(q);
1840 trace_block_plug(q);
1842 last = list_entry_rq(plug->mq_list.prev);
1844 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1845 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1846 blk_flush_plug_list(plug, false);
1847 trace_block_plug(q);
1850 list_add_tail(&rq->queuelist, &plug->mq_list);
1851 } else if (plug && !blk_queue_nomerges(q)) {
1852 blk_mq_bio_to_request(rq, bio);
1855 * We do limited plugging. If the bio can be merged, do that.
1856 * Otherwise the existing request in the plug list will be
1857 * issued. So the plug list will have one request at most
1858 * The plug list might get flushed before this. If that happens,
1859 * the plug list is empty, and same_queue_rq is invalid.
1861 if (list_empty(&plug->mq_list))
1862 same_queue_rq = NULL;
1864 list_del_init(&same_queue_rq->queuelist);
1865 list_add_tail(&rq->queuelist, &plug->mq_list);
1867 blk_mq_put_ctx(data.ctx);
1869 if (same_queue_rq) {
1870 data.hctx = blk_mq_map_queue(q,
1871 same_queue_rq->mq_ctx->cpu);
1872 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1875 } else if (q->nr_hw_queues > 1 && is_sync) {
1876 blk_mq_put_ctx(data.ctx);
1877 blk_mq_bio_to_request(rq, bio);
1878 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1879 } else if (q->elevator) {
1880 blk_mq_put_ctx(data.ctx);
1881 blk_mq_bio_to_request(rq, bio);
1882 blk_mq_sched_insert_request(rq, false, true, true, true);
1884 blk_mq_put_ctx(data.ctx);
1885 blk_mq_bio_to_request(rq, bio);
1886 blk_mq_queue_io(data.hctx, data.ctx, rq);
1887 blk_mq_run_hw_queue(data.hctx, true);
1893 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1894 unsigned int hctx_idx)
1898 if (tags->rqs && set->ops->exit_request) {
1901 for (i = 0; i < tags->nr_tags; i++) {
1902 struct request *rq = tags->static_rqs[i];
1906 set->ops->exit_request(set, rq, hctx_idx);
1907 tags->static_rqs[i] = NULL;
1911 while (!list_empty(&tags->page_list)) {
1912 page = list_first_entry(&tags->page_list, struct page, lru);
1913 list_del_init(&page->lru);
1915 * Remove kmemleak object previously allocated in
1916 * blk_mq_init_rq_map().
1918 kmemleak_free(page_address(page));
1919 __free_pages(page, page->private);
1923 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1927 kfree(tags->static_rqs);
1928 tags->static_rqs = NULL;
1930 blk_mq_free_tags(tags);
1933 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1934 unsigned int hctx_idx,
1935 unsigned int nr_tags,
1936 unsigned int reserved_tags)
1938 struct blk_mq_tags *tags;
1941 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1942 if (node == NUMA_NO_NODE)
1943 node = set->numa_node;
1945 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1946 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1950 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1951 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1954 blk_mq_free_tags(tags);
1958 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1959 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1961 if (!tags->static_rqs) {
1963 blk_mq_free_tags(tags);
1970 static size_t order_to_size(unsigned int order)
1972 return (size_t)PAGE_SIZE << order;
1975 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1976 unsigned int hctx_idx, int node)
1980 if (set->ops->init_request) {
1981 ret = set->ops->init_request(set, rq, hctx_idx, node);
1986 seqcount_init(&rq->gstate_seq);
1987 u64_stats_init(&rq->aborted_gstate_sync);
1991 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1992 unsigned int hctx_idx, unsigned int depth)
1994 unsigned int i, j, entries_per_page, max_order = 4;
1995 size_t rq_size, left;
1998 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1999 if (node == NUMA_NO_NODE)
2000 node = set->numa_node;
2002 INIT_LIST_HEAD(&tags->page_list);
2005 * rq_size is the size of the request plus driver payload, rounded
2006 * to the cacheline size
2008 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2010 left = rq_size * depth;
2012 for (i = 0; i < depth; ) {
2013 int this_order = max_order;
2018 while (this_order && left < order_to_size(this_order - 1))
2022 page = alloc_pages_node(node,
2023 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2029 if (order_to_size(this_order) < rq_size)
2036 page->private = this_order;
2037 list_add_tail(&page->lru, &tags->page_list);
2039 p = page_address(page);
2041 * Allow kmemleak to scan these pages as they contain pointers
2042 * to additional allocations like via ops->init_request().
2044 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2045 entries_per_page = order_to_size(this_order) / rq_size;
2046 to_do = min(entries_per_page, depth - i);
2047 left -= to_do * rq_size;
2048 for (j = 0; j < to_do; j++) {
2049 struct request *rq = p;
2051 tags->static_rqs[i] = rq;
2052 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2053 tags->static_rqs[i] = NULL;
2064 blk_mq_free_rqs(set, tags, hctx_idx);
2069 * 'cpu' is going away. splice any existing rq_list entries from this
2070 * software queue to the hw queue dispatch list, and ensure that it
2073 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2075 struct blk_mq_hw_ctx *hctx;
2076 struct blk_mq_ctx *ctx;
2079 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2080 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2082 spin_lock(&ctx->lock);
2083 if (!list_empty(&ctx->rq_list)) {
2084 list_splice_init(&ctx->rq_list, &tmp);
2085 blk_mq_hctx_clear_pending(hctx, ctx);
2087 spin_unlock(&ctx->lock);
2089 if (list_empty(&tmp))
2092 spin_lock(&hctx->lock);
2093 list_splice_tail_init(&tmp, &hctx->dispatch);
2094 spin_unlock(&hctx->lock);
2096 blk_mq_run_hw_queue(hctx, true);
2100 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2102 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2106 /* hctx->ctxs will be freed in queue's release handler */
2107 static void blk_mq_exit_hctx(struct request_queue *q,
2108 struct blk_mq_tag_set *set,
2109 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2111 blk_mq_debugfs_unregister_hctx(hctx);
2113 if (blk_mq_hw_queue_mapped(hctx))
2114 blk_mq_tag_idle(hctx);
2116 if (set->ops->exit_request)
2117 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2119 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2121 if (set->ops->exit_hctx)
2122 set->ops->exit_hctx(hctx, hctx_idx);
2124 if (hctx->flags & BLK_MQ_F_BLOCKING)
2125 cleanup_srcu_struct(hctx->queue_rq_srcu);
2127 blk_mq_remove_cpuhp(hctx);
2128 blk_free_flush_queue(hctx->fq);
2129 sbitmap_free(&hctx->ctx_map);
2132 static void blk_mq_exit_hw_queues(struct request_queue *q,
2133 struct blk_mq_tag_set *set, int nr_queue)
2135 struct blk_mq_hw_ctx *hctx;
2138 queue_for_each_hw_ctx(q, hctx, i) {
2141 blk_mq_exit_hctx(q, set, hctx, i);
2145 static int blk_mq_init_hctx(struct request_queue *q,
2146 struct blk_mq_tag_set *set,
2147 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2151 node = hctx->numa_node;
2152 if (node == NUMA_NO_NODE)
2153 node = hctx->numa_node = set->numa_node;
2155 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2156 spin_lock_init(&hctx->lock);
2157 INIT_LIST_HEAD(&hctx->dispatch);
2159 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2161 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2163 hctx->tags = set->tags[hctx_idx];
2166 * Allocate space for all possible cpus to avoid allocation at
2169 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2172 goto unregister_cpu_notifier;
2174 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2180 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2181 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2183 if (set->ops->init_hctx &&
2184 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2187 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2190 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2192 goto sched_exit_hctx;
2194 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2197 if (hctx->flags & BLK_MQ_F_BLOCKING)
2198 init_srcu_struct(hctx->queue_rq_srcu);
2200 blk_mq_debugfs_register_hctx(q, hctx);
2207 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2209 if (set->ops->exit_hctx)
2210 set->ops->exit_hctx(hctx, hctx_idx);
2212 sbitmap_free(&hctx->ctx_map);
2215 unregister_cpu_notifier:
2216 blk_mq_remove_cpuhp(hctx);
2220 static void blk_mq_init_cpu_queues(struct request_queue *q,
2221 unsigned int nr_hw_queues)
2225 for_each_possible_cpu(i) {
2226 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2227 struct blk_mq_hw_ctx *hctx;
2230 spin_lock_init(&__ctx->lock);
2231 INIT_LIST_HEAD(&__ctx->rq_list);
2234 /* If the cpu isn't present, the cpu is mapped to first hctx */
2235 if (!cpu_present(i))
2238 hctx = blk_mq_map_queue(q, i);
2241 * Set local node, IFF we have more than one hw queue. If
2242 * not, we remain on the home node of the device
2244 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2245 hctx->numa_node = local_memory_node(cpu_to_node(i));
2249 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2253 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2254 set->queue_depth, set->reserved_tags);
2255 if (!set->tags[hctx_idx])
2258 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2263 blk_mq_free_rq_map(set->tags[hctx_idx]);
2264 set->tags[hctx_idx] = NULL;
2268 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2269 unsigned int hctx_idx)
2271 if (set->tags[hctx_idx]) {
2272 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2273 blk_mq_free_rq_map(set->tags[hctx_idx]);
2274 set->tags[hctx_idx] = NULL;
2278 static void blk_mq_map_swqueue(struct request_queue *q)
2280 unsigned int i, hctx_idx;
2281 struct blk_mq_hw_ctx *hctx;
2282 struct blk_mq_ctx *ctx;
2283 struct blk_mq_tag_set *set = q->tag_set;
2286 * Avoid others reading imcomplete hctx->cpumask through sysfs
2288 mutex_lock(&q->sysfs_lock);
2290 queue_for_each_hw_ctx(q, hctx, i) {
2291 cpumask_clear(hctx->cpumask);
2296 * Map software to hardware queues.
2298 * If the cpu isn't present, the cpu is mapped to first hctx.
2300 for_each_present_cpu(i) {
2301 hctx_idx = q->mq_map[i];
2302 /* unmapped hw queue can be remapped after CPU topo changed */
2303 if (!set->tags[hctx_idx] &&
2304 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2306 * If tags initialization fail for some hctx,
2307 * that hctx won't be brought online. In this
2308 * case, remap the current ctx to hctx[0] which
2309 * is guaranteed to always have tags allocated
2314 ctx = per_cpu_ptr(q->queue_ctx, i);
2315 hctx = blk_mq_map_queue(q, i);
2317 cpumask_set_cpu(i, hctx->cpumask);
2318 ctx->index_hw = hctx->nr_ctx;
2319 hctx->ctxs[hctx->nr_ctx++] = ctx;
2322 mutex_unlock(&q->sysfs_lock);
2324 queue_for_each_hw_ctx(q, hctx, i) {
2326 * If no software queues are mapped to this hardware queue,
2327 * disable it and free the request entries.
2329 if (!hctx->nr_ctx) {
2330 /* Never unmap queue 0. We need it as a
2331 * fallback in case of a new remap fails
2334 if (i && set->tags[i])
2335 blk_mq_free_map_and_requests(set, i);
2341 hctx->tags = set->tags[i];
2342 WARN_ON(!hctx->tags);
2345 * Set the map size to the number of mapped software queues.
2346 * This is more accurate and more efficient than looping
2347 * over all possibly mapped software queues.
2349 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2352 * Initialize batch roundrobin counts
2354 hctx->next_cpu = cpumask_first(hctx->cpumask);
2355 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2360 * Caller needs to ensure that we're either frozen/quiesced, or that
2361 * the queue isn't live yet.
2363 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2365 struct blk_mq_hw_ctx *hctx;
2368 queue_for_each_hw_ctx(q, hctx, i) {
2370 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2371 atomic_inc(&q->shared_hctx_restart);
2372 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2374 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2375 atomic_dec(&q->shared_hctx_restart);
2376 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2381 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2384 struct request_queue *q;
2386 lockdep_assert_held(&set->tag_list_lock);
2388 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2389 blk_mq_freeze_queue(q);
2390 queue_set_hctx_shared(q, shared);
2391 blk_mq_unfreeze_queue(q);
2395 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2397 struct blk_mq_tag_set *set = q->tag_set;
2399 mutex_lock(&set->tag_list_lock);
2400 list_del_rcu(&q->tag_set_list);
2401 INIT_LIST_HEAD(&q->tag_set_list);
2402 if (list_is_singular(&set->tag_list)) {
2403 /* just transitioned to unshared */
2404 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2405 /* update existing queue */
2406 blk_mq_update_tag_set_depth(set, false);
2408 mutex_unlock(&set->tag_list_lock);
2413 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2414 struct request_queue *q)
2418 mutex_lock(&set->tag_list_lock);
2421 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2423 if (!list_empty(&set->tag_list) &&
2424 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2425 set->flags |= BLK_MQ_F_TAG_SHARED;
2426 /* update existing queue */
2427 blk_mq_update_tag_set_depth(set, true);
2429 if (set->flags & BLK_MQ_F_TAG_SHARED)
2430 queue_set_hctx_shared(q, true);
2431 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2433 mutex_unlock(&set->tag_list_lock);
2437 * It is the actual release handler for mq, but we do it from
2438 * request queue's release handler for avoiding use-after-free
2439 * and headache because q->mq_kobj shouldn't have been introduced,
2440 * but we can't group ctx/kctx kobj without it.
2442 void blk_mq_release(struct request_queue *q)
2444 struct blk_mq_hw_ctx *hctx;
2447 /* hctx kobj stays in hctx */
2448 queue_for_each_hw_ctx(q, hctx, i) {
2451 kobject_put(&hctx->kobj);
2456 kfree(q->queue_hw_ctx);
2459 * release .mq_kobj and sw queue's kobject now because
2460 * both share lifetime with request queue.
2462 blk_mq_sysfs_deinit(q);
2464 free_percpu(q->queue_ctx);
2467 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2469 struct request_queue *uninit_q, *q;
2471 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2473 return ERR_PTR(-ENOMEM);
2475 q = blk_mq_init_allocated_queue(set, uninit_q);
2477 blk_cleanup_queue(uninit_q);
2481 EXPORT_SYMBOL(blk_mq_init_queue);
2483 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2485 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2487 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2488 __alignof__(struct blk_mq_hw_ctx)) !=
2489 sizeof(struct blk_mq_hw_ctx));
2491 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2492 hw_ctx_size += sizeof(struct srcu_struct);
2497 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2498 struct request_queue *q)
2501 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2503 blk_mq_sysfs_unregister(q);
2505 /* protect against switching io scheduler */
2506 mutex_lock(&q->sysfs_lock);
2507 for (i = 0; i < set->nr_hw_queues; i++) {
2513 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2514 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2519 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2526 atomic_set(&hctxs[i]->nr_active, 0);
2527 hctxs[i]->numa_node = node;
2528 hctxs[i]->queue_num = i;
2530 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2531 free_cpumask_var(hctxs[i]->cpumask);
2536 blk_mq_hctx_kobj_init(hctxs[i]);
2538 for (j = i; j < q->nr_hw_queues; j++) {
2539 struct blk_mq_hw_ctx *hctx = hctxs[j];
2543 blk_mq_free_map_and_requests(set, j);
2544 blk_mq_exit_hctx(q, set, hctx, j);
2545 kobject_put(&hctx->kobj);
2550 q->nr_hw_queues = i;
2551 mutex_unlock(&q->sysfs_lock);
2552 blk_mq_sysfs_register(q);
2555 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2556 struct request_queue *q)
2558 /* mark the queue as mq asap */
2559 q->mq_ops = set->ops;
2561 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2562 blk_mq_poll_stats_bkt,
2563 BLK_MQ_POLL_STATS_BKTS, q);
2567 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2571 /* init q->mq_kobj and sw queues' kobjects */
2572 blk_mq_sysfs_init(q);
2574 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2575 GFP_KERNEL, set->numa_node);
2576 if (!q->queue_hw_ctx)
2579 q->mq_map = set->mq_map;
2581 blk_mq_realloc_hw_ctxs(set, q);
2582 if (!q->nr_hw_queues)
2585 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2586 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2588 q->nr_queues = nr_cpu_ids;
2590 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2592 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2593 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2595 q->sg_reserved_size = INT_MAX;
2597 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2598 INIT_LIST_HEAD(&q->requeue_list);
2599 spin_lock_init(&q->requeue_lock);
2601 blk_queue_make_request(q, blk_mq_make_request);
2602 if (q->mq_ops->poll)
2603 q->poll_fn = blk_mq_poll;
2606 * Do this after blk_queue_make_request() overrides it...
2608 q->nr_requests = set->queue_depth;
2611 * Default to classic polling
2615 if (set->ops->complete)
2616 blk_queue_softirq_done(q, set->ops->complete);
2618 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2619 blk_mq_add_queue_tag_set(set, q);
2620 blk_mq_map_swqueue(q);
2622 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2625 ret = blk_mq_sched_init(q);
2627 return ERR_PTR(ret);
2633 kfree(q->queue_hw_ctx);
2635 free_percpu(q->queue_ctx);
2638 return ERR_PTR(-ENOMEM);
2640 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2642 void blk_mq_free_queue(struct request_queue *q)
2644 struct blk_mq_tag_set *set = q->tag_set;
2646 blk_mq_del_queue_tag_set(q);
2647 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2650 /* Basically redo blk_mq_init_queue with queue frozen */
2651 static void blk_mq_queue_reinit(struct request_queue *q)
2653 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2655 blk_mq_debugfs_unregister_hctxs(q);
2656 blk_mq_sysfs_unregister(q);
2659 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2660 * we should change hctx numa_node according to the new topology (this
2661 * involves freeing and re-allocating memory, worth doing?)
2663 blk_mq_map_swqueue(q);
2665 blk_mq_sysfs_register(q);
2666 blk_mq_debugfs_register_hctxs(q);
2669 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2673 for (i = 0; i < set->nr_hw_queues; i++)
2674 if (!__blk_mq_alloc_rq_map(set, i))
2681 blk_mq_free_rq_map(set->tags[i]);
2687 * Allocate the request maps associated with this tag_set. Note that this
2688 * may reduce the depth asked for, if memory is tight. set->queue_depth
2689 * will be updated to reflect the allocated depth.
2691 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2696 depth = set->queue_depth;
2698 err = __blk_mq_alloc_rq_maps(set);
2702 set->queue_depth >>= 1;
2703 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2707 } while (set->queue_depth);
2709 if (!set->queue_depth || err) {
2710 pr_err("blk-mq: failed to allocate request map\n");
2714 if (depth != set->queue_depth)
2715 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2716 depth, set->queue_depth);
2721 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2723 if (set->ops->map_queues) {
2726 * transport .map_queues is usually done in the following
2729 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2730 * mask = get_cpu_mask(queue)
2731 * for_each_cpu(cpu, mask)
2732 * set->mq_map[cpu] = queue;
2735 * When we need to remap, the table has to be cleared for
2736 * killing stale mapping since one CPU may not be mapped
2739 for_each_possible_cpu(cpu)
2740 set->mq_map[cpu] = 0;
2742 return set->ops->map_queues(set);
2744 return blk_mq_map_queues(set);
2748 * Alloc a tag set to be associated with one or more request queues.
2749 * May fail with EINVAL for various error conditions. May adjust the
2750 * requested depth down, if if it too large. In that case, the set
2751 * value will be stored in set->queue_depth.
2753 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2757 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2759 if (!set->nr_hw_queues)
2761 if (!set->queue_depth)
2763 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2766 if (!set->ops->queue_rq)
2769 if (!set->ops->get_budget ^ !set->ops->put_budget)
2772 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2773 pr_info("blk-mq: reduced tag depth to %u\n",
2775 set->queue_depth = BLK_MQ_MAX_DEPTH;
2779 * If a crashdump is active, then we are potentially in a very
2780 * memory constrained environment. Limit us to 1 queue and
2781 * 64 tags to prevent using too much memory.
2783 if (is_kdump_kernel()) {
2784 set->nr_hw_queues = 1;
2785 set->queue_depth = min(64U, set->queue_depth);
2788 * There is no use for more h/w queues than cpus.
2790 if (set->nr_hw_queues > nr_cpu_ids)
2791 set->nr_hw_queues = nr_cpu_ids;
2793 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2794 GFP_KERNEL, set->numa_node);
2799 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2800 GFP_KERNEL, set->numa_node);
2804 ret = blk_mq_update_queue_map(set);
2806 goto out_free_mq_map;
2808 ret = blk_mq_alloc_rq_maps(set);
2810 goto out_free_mq_map;
2812 mutex_init(&set->tag_list_lock);
2813 INIT_LIST_HEAD(&set->tag_list);
2825 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2827 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2831 for (i = 0; i < nr_cpu_ids; i++)
2832 blk_mq_free_map_and_requests(set, i);
2840 EXPORT_SYMBOL(blk_mq_free_tag_set);
2842 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2844 struct blk_mq_tag_set *set = q->tag_set;
2845 struct blk_mq_hw_ctx *hctx;
2851 blk_mq_freeze_queue(q);
2852 blk_mq_quiesce_queue(q);
2855 queue_for_each_hw_ctx(q, hctx, i) {
2859 * If we're using an MQ scheduler, just update the scheduler
2860 * queue depth. This is similar to what the old code would do.
2862 if (!hctx->sched_tags) {
2863 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2866 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2874 q->nr_requests = nr;
2876 blk_mq_unquiesce_queue(q);
2877 blk_mq_unfreeze_queue(q);
2882 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2885 struct request_queue *q;
2887 lockdep_assert_held(&set->tag_list_lock);
2889 if (nr_hw_queues > nr_cpu_ids)
2890 nr_hw_queues = nr_cpu_ids;
2891 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2894 list_for_each_entry(q, &set->tag_list, tag_set_list)
2895 blk_mq_freeze_queue(q);
2897 set->nr_hw_queues = nr_hw_queues;
2898 blk_mq_update_queue_map(set);
2899 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2900 blk_mq_realloc_hw_ctxs(set, q);
2901 blk_mq_queue_reinit(q);
2904 list_for_each_entry(q, &set->tag_list, tag_set_list)
2905 blk_mq_unfreeze_queue(q);
2908 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2910 mutex_lock(&set->tag_list_lock);
2911 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2912 mutex_unlock(&set->tag_list_lock);
2914 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2916 /* Enable polling stats and return whether they were already enabled. */
2917 static bool blk_poll_stats_enable(struct request_queue *q)
2919 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2920 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2922 blk_stat_add_callback(q, q->poll_cb);
2926 static void blk_mq_poll_stats_start(struct request_queue *q)
2929 * We don't arm the callback if polling stats are not enabled or the
2930 * callback is already active.
2932 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2933 blk_stat_is_active(q->poll_cb))
2936 blk_stat_activate_msecs(q->poll_cb, 100);
2939 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2941 struct request_queue *q = cb->data;
2944 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2945 if (cb->stat[bucket].nr_samples)
2946 q->poll_stat[bucket] = cb->stat[bucket];
2950 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2951 struct blk_mq_hw_ctx *hctx,
2954 unsigned long ret = 0;
2958 * If stats collection isn't on, don't sleep but turn it on for
2961 if (!blk_poll_stats_enable(q))
2965 * As an optimistic guess, use half of the mean service time
2966 * for this type of request. We can (and should) make this smarter.
2967 * For instance, if the completion latencies are tight, we can
2968 * get closer than just half the mean. This is especially
2969 * important on devices where the completion latencies are longer
2970 * than ~10 usec. We do use the stats for the relevant IO size
2971 * if available which does lead to better estimates.
2973 bucket = blk_mq_poll_stats_bkt(rq);
2977 if (q->poll_stat[bucket].nr_samples)
2978 ret = (q->poll_stat[bucket].mean + 1) / 2;
2983 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2984 struct blk_mq_hw_ctx *hctx,
2987 struct hrtimer_sleeper hs;
2988 enum hrtimer_mode mode;
2992 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2998 * -1: don't ever hybrid sleep
2999 * 0: use half of prev avg
3000 * >0: use this specific value
3002 if (q->poll_nsec == -1)
3004 else if (q->poll_nsec > 0)
3005 nsecs = q->poll_nsec;
3007 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3012 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
3015 * This will be replaced with the stats tracking code, using
3016 * 'avg_completion_time / 2' as the pre-sleep target.
3020 mode = HRTIMER_MODE_REL;
3021 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3022 hrtimer_set_expires(&hs.timer, kt);
3024 hrtimer_init_sleeper(&hs, current);
3026 if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) &&
3027 blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
3029 set_current_state(TASK_UNINTERRUPTIBLE);
3030 hrtimer_start_expires(&hs.timer, mode);
3033 hrtimer_cancel(&hs.timer);
3034 mode = HRTIMER_MODE_ABS;
3035 } while (hs.task && !signal_pending(current));
3037 __set_current_state(TASK_RUNNING);
3038 destroy_hrtimer_on_stack(&hs.timer);
3042 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3044 struct request_queue *q = hctx->queue;
3048 * If we sleep, have the caller restart the poll loop to reset
3049 * the state. Like for the other success return cases, the
3050 * caller is responsible for checking if the IO completed. If
3051 * the IO isn't complete, we'll get called again and will go
3052 * straight to the busy poll loop.
3054 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3057 hctx->poll_considered++;
3059 state = current->state;
3060 while (!need_resched()) {
3063 hctx->poll_invoked++;
3065 ret = q->mq_ops->poll(hctx, rq->tag);
3067 hctx->poll_success++;
3068 set_current_state(TASK_RUNNING);
3072 if (signal_pending_state(state, current))
3073 set_current_state(TASK_RUNNING);
3075 if (current->state == TASK_RUNNING)
3085 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3087 struct blk_mq_hw_ctx *hctx;
3090 if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3093 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3094 if (!blk_qc_t_is_internal(cookie))
3095 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3097 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3099 * With scheduling, if the request has completed, we'll
3100 * get a NULL return here, as we clear the sched tag when
3101 * that happens. The request still remains valid, like always,
3102 * so we should be safe with just the NULL check.
3108 return __blk_mq_poll(hctx, rq);
3111 static int __init blk_mq_init(void)
3113 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3114 blk_mq_hctx_notify_dead);
3117 subsys_initcall(blk_mq_init);