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);
165 blk_mq_freeze_queue_wait(q);
168 void blk_mq_freeze_queue(struct request_queue *q)
171 * ...just an alias to keep freeze and unfreeze actions balanced
172 * in the blk_mq_* namespace
176 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
178 void blk_mq_unfreeze_queue(struct request_queue *q)
182 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
183 WARN_ON_ONCE(freeze_depth < 0);
185 percpu_ref_reinit(&q->q_usage_counter);
186 wake_up_all(&q->mq_freeze_wq);
189 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
192 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
193 * mpt3sas driver such that this function can be removed.
195 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
197 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
199 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
202 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
205 * Note: this function does not prevent that the struct request end_io()
206 * callback function is invoked. Once this function is returned, we make
207 * sure no dispatch can happen until the queue is unquiesced via
208 * blk_mq_unquiesce_queue().
210 void blk_mq_quiesce_queue(struct request_queue *q)
212 struct blk_mq_hw_ctx *hctx;
216 blk_mq_quiesce_queue_nowait(q);
218 queue_for_each_hw_ctx(q, hctx, i) {
219 if (hctx->flags & BLK_MQ_F_BLOCKING)
220 synchronize_srcu(hctx->srcu);
227 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
230 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
233 * This function recovers queue into the state before quiescing
234 * which is done by blk_mq_quiesce_queue.
236 void blk_mq_unquiesce_queue(struct request_queue *q)
238 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
240 /* dispatch requests which are inserted during quiescing */
241 blk_mq_run_hw_queues(q, true);
243 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
245 void blk_mq_wake_waiters(struct request_queue *q)
247 struct blk_mq_hw_ctx *hctx;
250 queue_for_each_hw_ctx(q, hctx, i)
251 if (blk_mq_hw_queue_mapped(hctx))
252 blk_mq_tag_wakeup_all(hctx->tags, true);
255 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
257 return blk_mq_has_free_tags(hctx->tags);
259 EXPORT_SYMBOL(blk_mq_can_queue);
261 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
262 unsigned int tag, unsigned int op)
264 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
265 struct request *rq = tags->static_rqs[tag];
266 req_flags_t rq_flags = 0;
268 if (data->flags & BLK_MQ_REQ_INTERNAL) {
270 rq->internal_tag = tag;
272 if (blk_mq_tag_busy(data->hctx)) {
273 rq_flags = RQF_MQ_INFLIGHT;
274 atomic_inc(&data->hctx->nr_active);
277 rq->internal_tag = -1;
278 data->hctx->tags->rqs[rq->tag] = rq;
281 /* csd/requeue_work/fifo_time is initialized before use */
283 rq->mq_ctx = data->ctx;
284 rq->rq_flags = rq_flags;
287 if (data->flags & BLK_MQ_REQ_PREEMPT)
288 rq->rq_flags |= RQF_PREEMPT;
289 if (blk_queue_io_stat(data->q))
290 rq->rq_flags |= RQF_IO_STAT;
291 INIT_LIST_HEAD(&rq->queuelist);
292 INIT_HLIST_NODE(&rq->hash);
293 RB_CLEAR_NODE(&rq->rb_node);
296 rq->start_time = jiffies;
297 rq->nr_phys_segments = 0;
298 #if defined(CONFIG_BLK_DEV_INTEGRITY)
299 rq->nr_integrity_segments = 0;
302 /* tag was already set */
306 INIT_LIST_HEAD(&rq->timeout_list);
310 rq->end_io_data = NULL;
313 #ifdef CONFIG_BLK_CGROUP
315 set_start_time_ns(rq);
316 rq->io_start_time_ns = 0;
319 data->ctx->rq_dispatched[op_is_sync(op)]++;
323 static struct request *blk_mq_get_request(struct request_queue *q,
324 struct bio *bio, unsigned int op,
325 struct blk_mq_alloc_data *data)
327 struct elevator_queue *e = q->elevator;
330 bool put_ctx_on_error = false;
332 blk_queue_enter_live(q);
334 if (likely(!data->ctx)) {
335 data->ctx = blk_mq_get_ctx(q);
336 put_ctx_on_error = true;
338 if (likely(!data->hctx))
339 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
341 data->flags |= BLK_MQ_REQ_NOWAIT;
344 data->flags |= BLK_MQ_REQ_INTERNAL;
347 * Flush requests are special and go directly to the
350 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
351 e->type->ops.mq.limit_depth(op, data);
354 tag = blk_mq_get_tag(data);
355 if (tag == BLK_MQ_TAG_FAIL) {
356 if (put_ctx_on_error) {
357 blk_mq_put_ctx(data->ctx);
364 rq = blk_mq_rq_ctx_init(data, tag, op);
365 if (!op_is_flush(op)) {
367 if (e && e->type->ops.mq.prepare_request) {
368 if (e->type->icq_cache && rq_ioc(bio))
369 blk_mq_sched_assign_ioc(rq, bio);
371 e->type->ops.mq.prepare_request(rq, bio);
372 rq->rq_flags |= RQF_ELVPRIV;
375 data->hctx->queued++;
379 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
380 blk_mq_req_flags_t flags)
382 struct blk_mq_alloc_data alloc_data = { .flags = flags };
386 ret = blk_queue_enter(q, flags);
390 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
394 return ERR_PTR(-EWOULDBLOCK);
396 blk_mq_put_ctx(alloc_data.ctx);
399 rq->__sector = (sector_t) -1;
400 rq->bio = rq->biotail = NULL;
403 EXPORT_SYMBOL(blk_mq_alloc_request);
405 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
406 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
408 struct blk_mq_alloc_data alloc_data = { .flags = flags };
414 * If the tag allocator sleeps we could get an allocation for a
415 * different hardware context. No need to complicate the low level
416 * allocator for this for the rare use case of a command tied to
419 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
420 return ERR_PTR(-EINVAL);
422 if (hctx_idx >= q->nr_hw_queues)
423 return ERR_PTR(-EIO);
425 ret = blk_queue_enter(q, flags);
430 * Check if the hardware context is actually mapped to anything.
431 * If not tell the caller that it should skip this queue.
433 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
434 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
436 return ERR_PTR(-EXDEV);
438 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
439 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
441 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
445 return ERR_PTR(-EWOULDBLOCK);
449 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
451 void blk_mq_free_request(struct request *rq)
453 struct request_queue *q = rq->q;
454 struct elevator_queue *e = q->elevator;
455 struct blk_mq_ctx *ctx = rq->mq_ctx;
456 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
457 const int sched_tag = rq->internal_tag;
459 if (rq->rq_flags & RQF_ELVPRIV) {
460 if (e && e->type->ops.mq.finish_request)
461 e->type->ops.mq.finish_request(rq);
463 put_io_context(rq->elv.icq->ioc);
468 ctx->rq_completed[rq_is_sync(rq)]++;
469 if (rq->rq_flags & RQF_MQ_INFLIGHT)
470 atomic_dec(&hctx->nr_active);
472 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
473 laptop_io_completion(q->backing_dev_info);
475 wbt_done(q->rq_wb, &rq->issue_stat);
478 blk_put_rl(blk_rq_rl(rq));
480 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
482 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
484 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
485 blk_mq_sched_restart(hctx);
488 EXPORT_SYMBOL_GPL(blk_mq_free_request);
490 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
492 blk_account_io_done(rq);
495 wbt_done(rq->q->rq_wb, &rq->issue_stat);
496 rq->end_io(rq, error);
498 if (unlikely(blk_bidi_rq(rq)))
499 blk_mq_free_request(rq->next_rq);
500 blk_mq_free_request(rq);
503 EXPORT_SYMBOL(__blk_mq_end_request);
505 void blk_mq_end_request(struct request *rq, blk_status_t error)
507 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
509 __blk_mq_end_request(rq, error);
511 EXPORT_SYMBOL(blk_mq_end_request);
513 static void __blk_mq_complete_request_remote(void *data)
515 struct request *rq = data;
517 rq->q->softirq_done_fn(rq);
520 static void __blk_mq_complete_request(struct request *rq)
522 struct blk_mq_ctx *ctx = rq->mq_ctx;
526 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT);
527 blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE);
529 if (rq->internal_tag != -1)
530 blk_mq_sched_completed_request(rq);
531 if (rq->rq_flags & RQF_STATS) {
532 blk_mq_poll_stats_start(rq->q);
536 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
537 rq->q->softirq_done_fn(rq);
542 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
543 shared = cpus_share_cache(cpu, ctx->cpu);
545 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
546 rq->csd.func = __blk_mq_complete_request_remote;
549 smp_call_function_single_async(ctx->cpu, &rq->csd);
551 rq->q->softirq_done_fn(rq);
556 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
557 __releases(hctx->srcu)
559 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
562 srcu_read_unlock(hctx->srcu, srcu_idx);
565 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
566 __acquires(hctx->srcu)
568 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
569 /* shut up gcc false positive */
573 *srcu_idx = srcu_read_lock(hctx->srcu);
576 static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate)
581 * blk_mq_rq_aborted_gstate() is used from the completion path and
582 * can thus be called from irq context. u64_stats_fetch in the
583 * middle of update on the same CPU leads to lockup. Disable irq
586 local_irq_save(flags);
587 u64_stats_update_begin(&rq->aborted_gstate_sync);
588 rq->aborted_gstate = gstate;
589 u64_stats_update_end(&rq->aborted_gstate_sync);
590 local_irq_restore(flags);
593 static u64 blk_mq_rq_aborted_gstate(struct request *rq)
599 start = u64_stats_fetch_begin(&rq->aborted_gstate_sync);
600 aborted_gstate = rq->aborted_gstate;
601 } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start));
603 return aborted_gstate;
607 * blk_mq_complete_request - end I/O on a request
608 * @rq: the request being processed
611 * Ends all I/O on a request. It does not handle partial completions.
612 * The actual completion happens out-of-order, through a IPI handler.
614 void blk_mq_complete_request(struct request *rq)
616 struct request_queue *q = rq->q;
617 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
620 if (unlikely(blk_should_fake_timeout(q)))
624 * If @rq->aborted_gstate equals the current instance, timeout is
625 * claiming @rq and we lost. This is synchronized through
626 * hctx_lock(). See blk_mq_timeout_work() for details.
628 * Completion path never blocks and we can directly use RCU here
629 * instead of hctx_lock() which can be either RCU or SRCU.
630 * However, that would complicate paths which want to synchronize
631 * against us. Let stay in sync with the issue path so that
632 * hctx_lock() covers both issue and completion paths.
634 hctx_lock(hctx, &srcu_idx);
635 if (blk_mq_rq_aborted_gstate(rq) != rq->gstate)
636 __blk_mq_complete_request(rq);
637 hctx_unlock(hctx, srcu_idx);
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 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
657 rq->rq_flags |= RQF_STATS;
658 wbt_issue(q->rq_wb, &rq->issue_stat);
661 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
664 * Mark @rq in-flight which also advances the generation number,
665 * and register for timeout. Protect with a seqcount to allow the
666 * timeout path to read both @rq->gstate and @rq->deadline
669 * This is the only place where a request is marked in-flight. If
670 * the timeout path reads an in-flight @rq->gstate, the
671 * @rq->deadline it reads together under @rq->gstate_seq is
672 * guaranteed to be the matching one.
675 write_seqcount_begin(&rq->gstate_seq);
677 blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT);
680 write_seqcount_end(&rq->gstate_seq);
683 if (q->dma_drain_size && blk_rq_bytes(rq)) {
685 * Make sure space for the drain appears. We know we can do
686 * this because max_hw_segments has been adjusted to be one
687 * fewer than the device can handle.
689 rq->nr_phys_segments++;
692 EXPORT_SYMBOL(blk_mq_start_request);
695 * When we reach here because queue is busy, it's safe to change the state
696 * to IDLE without checking @rq->aborted_gstate because we should still be
697 * holding the RCU read lock and thus protected against timeout.
699 static void __blk_mq_requeue_request(struct request *rq)
701 struct request_queue *q = rq->q;
703 blk_mq_put_driver_tag(rq);
705 trace_block_rq_requeue(q, rq);
706 wbt_requeue(q->rq_wb, &rq->issue_stat);
708 if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) {
709 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
710 if (q->dma_drain_size && blk_rq_bytes(rq))
711 rq->nr_phys_segments--;
715 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
717 __blk_mq_requeue_request(rq);
719 /* this request will be re-inserted to io scheduler queue */
720 blk_mq_sched_requeue_request(rq);
722 BUG_ON(blk_queued_rq(rq));
723 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
725 EXPORT_SYMBOL(blk_mq_requeue_request);
727 static void blk_mq_requeue_work(struct work_struct *work)
729 struct request_queue *q =
730 container_of(work, struct request_queue, requeue_work.work);
732 struct request *rq, *next;
734 spin_lock_irq(&q->requeue_lock);
735 list_splice_init(&q->requeue_list, &rq_list);
736 spin_unlock_irq(&q->requeue_lock);
738 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
739 if (!(rq->rq_flags & RQF_SOFTBARRIER))
742 rq->rq_flags &= ~RQF_SOFTBARRIER;
743 list_del_init(&rq->queuelist);
744 blk_mq_sched_insert_request(rq, true, false, false);
747 while (!list_empty(&rq_list)) {
748 rq = list_entry(rq_list.next, struct request, queuelist);
749 list_del_init(&rq->queuelist);
750 blk_mq_sched_insert_request(rq, false, false, false);
753 blk_mq_run_hw_queues(q, false);
756 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
757 bool kick_requeue_list)
759 struct request_queue *q = rq->q;
763 * We abuse this flag that is otherwise used by the I/O scheduler to
764 * request head insertion from the workqueue.
766 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
768 spin_lock_irqsave(&q->requeue_lock, flags);
770 rq->rq_flags |= RQF_SOFTBARRIER;
771 list_add(&rq->queuelist, &q->requeue_list);
773 list_add_tail(&rq->queuelist, &q->requeue_list);
775 spin_unlock_irqrestore(&q->requeue_lock, flags);
777 if (kick_requeue_list)
778 blk_mq_kick_requeue_list(q);
780 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
782 void blk_mq_kick_requeue_list(struct request_queue *q)
784 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
786 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
788 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
791 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
792 msecs_to_jiffies(msecs));
794 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
796 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
798 if (tag < tags->nr_tags) {
799 prefetch(tags->rqs[tag]);
800 return tags->rqs[tag];
805 EXPORT_SYMBOL(blk_mq_tag_to_rq);
807 struct blk_mq_timeout_data {
809 unsigned int next_set;
810 unsigned int nr_expired;
813 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
815 const struct blk_mq_ops *ops = req->q->mq_ops;
816 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
818 req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED;
821 ret = ops->timeout(req, reserved);
825 __blk_mq_complete_request(req);
827 case BLK_EH_RESET_TIMER:
829 * As nothing prevents from completion happening while
830 * ->aborted_gstate is set, this may lead to ignored
831 * completions and further spurious timeouts.
833 blk_mq_rq_update_aborted_gstate(req, 0);
836 case BLK_EH_NOT_HANDLED:
839 printk(KERN_ERR "block: bad eh return: %d\n", ret);
844 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
845 struct request *rq, void *priv, bool reserved)
847 struct blk_mq_timeout_data *data = priv;
848 unsigned long gstate, deadline;
853 if (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED)
856 /* read coherent snapshots of @rq->state_gen and @rq->deadline */
858 start = read_seqcount_begin(&rq->gstate_seq);
859 gstate = READ_ONCE(rq->gstate);
860 deadline = blk_rq_deadline(rq);
861 if (!read_seqcount_retry(&rq->gstate_seq, start))
866 /* if in-flight && overdue, mark for abortion */
867 if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT &&
868 time_after_eq(jiffies, deadline)) {
869 blk_mq_rq_update_aborted_gstate(rq, gstate);
872 } else if (!data->next_set || time_after(data->next, deadline)) {
873 data->next = deadline;
878 static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx,
879 struct request *rq, void *priv, bool reserved)
882 * We marked @rq->aborted_gstate and waited for RCU. If there were
883 * completions that we lost to, they would have finished and
884 * updated @rq->gstate by now; otherwise, the completion path is
885 * now guaranteed to see @rq->aborted_gstate and yield. If
886 * @rq->aborted_gstate still matches @rq->gstate, @rq is ours.
888 if (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) &&
889 READ_ONCE(rq->gstate) == rq->aborted_gstate)
890 blk_mq_rq_timed_out(rq, reserved);
893 static void blk_mq_timeout_work(struct work_struct *work)
895 struct request_queue *q =
896 container_of(work, struct request_queue, timeout_work);
897 struct blk_mq_timeout_data data = {
902 struct blk_mq_hw_ctx *hctx;
905 /* A deadlock might occur if a request is stuck requiring a
906 * timeout at the same time a queue freeze is waiting
907 * completion, since the timeout code would not be able to
908 * acquire the queue reference here.
910 * That's why we don't use blk_queue_enter here; instead, we use
911 * percpu_ref_tryget directly, because we need to be able to
912 * obtain a reference even in the short window between the queue
913 * starting to freeze, by dropping the first reference in
914 * blk_freeze_queue_start, and the moment the last request is
915 * consumed, marked by the instant q_usage_counter reaches
918 if (!percpu_ref_tryget(&q->q_usage_counter))
921 /* scan for the expired ones and set their ->aborted_gstate */
922 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
924 if (data.nr_expired) {
925 bool has_rcu = false;
928 * Wait till everyone sees ->aborted_gstate. The
929 * sequential waits for SRCUs aren't ideal. If this ever
930 * becomes a problem, we can add per-hw_ctx rcu_head and
933 queue_for_each_hw_ctx(q, hctx, i) {
934 if (!hctx->nr_expired)
937 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
940 synchronize_srcu(hctx->srcu);
942 hctx->nr_expired = 0;
947 /* terminate the ones we won */
948 blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL);
952 data.next = blk_rq_timeout(round_jiffies_up(data.next));
953 mod_timer(&q->timeout, data.next);
956 * Request timeouts are handled as a forward rolling timer. If
957 * we end up here it means that no requests are pending and
958 * also that no request has been pending for a while. Mark
961 queue_for_each_hw_ctx(q, hctx, i) {
962 /* the hctx may be unmapped, so check it here */
963 if (blk_mq_hw_queue_mapped(hctx))
964 blk_mq_tag_idle(hctx);
970 struct flush_busy_ctx_data {
971 struct blk_mq_hw_ctx *hctx;
972 struct list_head *list;
975 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
977 struct flush_busy_ctx_data *flush_data = data;
978 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
979 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
981 spin_lock(&ctx->lock);
982 list_splice_tail_init(&ctx->rq_list, flush_data->list);
983 sbitmap_clear_bit(sb, bitnr);
984 spin_unlock(&ctx->lock);
989 * Process software queues that have been marked busy, splicing them
990 * to the for-dispatch
992 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
994 struct flush_busy_ctx_data data = {
999 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1001 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1003 struct dispatch_rq_data {
1004 struct blk_mq_hw_ctx *hctx;
1008 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1011 struct dispatch_rq_data *dispatch_data = data;
1012 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1013 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1015 spin_lock(&ctx->lock);
1016 if (unlikely(!list_empty(&ctx->rq_list))) {
1017 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
1018 list_del_init(&dispatch_data->rq->queuelist);
1019 if (list_empty(&ctx->rq_list))
1020 sbitmap_clear_bit(sb, bitnr);
1022 spin_unlock(&ctx->lock);
1024 return !dispatch_data->rq;
1027 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1028 struct blk_mq_ctx *start)
1030 unsigned off = start ? start->index_hw : 0;
1031 struct dispatch_rq_data data = {
1036 __sbitmap_for_each_set(&hctx->ctx_map, off,
1037 dispatch_rq_from_ctx, &data);
1042 static inline unsigned int queued_to_index(unsigned int queued)
1047 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1050 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
1053 struct blk_mq_alloc_data data = {
1055 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
1056 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
1059 might_sleep_if(wait);
1064 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1065 data.flags |= BLK_MQ_REQ_RESERVED;
1067 rq->tag = blk_mq_get_tag(&data);
1069 if (blk_mq_tag_busy(data.hctx)) {
1070 rq->rq_flags |= RQF_MQ_INFLIGHT;
1071 atomic_inc(&data.hctx->nr_active);
1073 data.hctx->tags->rqs[rq->tag] = rq;
1079 return rq->tag != -1;
1082 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1083 int flags, void *key)
1085 struct blk_mq_hw_ctx *hctx;
1087 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1089 list_del_init(&wait->entry);
1090 blk_mq_run_hw_queue(hctx, true);
1095 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1096 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1097 * restart. For both cases, take care to check the condition again after
1098 * marking us as waiting.
1100 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1103 struct blk_mq_hw_ctx *this_hctx = *hctx;
1104 struct sbq_wait_state *ws;
1105 wait_queue_entry_t *wait;
1108 if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1109 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1110 set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1113 * It's possible that a tag was freed in the window between the
1114 * allocation failure and adding the hardware queue to the wait
1117 * Don't clear RESTART here, someone else could have set it.
1118 * At most this will cost an extra queue run.
1120 return blk_mq_get_driver_tag(rq, hctx, false);
1123 wait = &this_hctx->dispatch_wait;
1124 if (!list_empty_careful(&wait->entry))
1127 spin_lock(&this_hctx->lock);
1128 if (!list_empty(&wait->entry)) {
1129 spin_unlock(&this_hctx->lock);
1133 ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1134 add_wait_queue(&ws->wait, wait);
1137 * It's possible that a tag was freed in the window between the
1138 * allocation failure and adding the hardware queue to the wait
1141 ret = blk_mq_get_driver_tag(rq, hctx, false);
1143 spin_unlock(&this_hctx->lock);
1148 * We got a tag, remove ourselves from the wait queue to ensure
1149 * someone else gets the wakeup.
1151 spin_lock_irq(&ws->wait.lock);
1152 list_del_init(&wait->entry);
1153 spin_unlock_irq(&ws->wait.lock);
1154 spin_unlock(&this_hctx->lock);
1159 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1161 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1164 struct blk_mq_hw_ctx *hctx;
1165 struct request *rq, *nxt;
1166 bool no_tag = false;
1168 blk_status_t ret = BLK_STS_OK;
1170 if (list_empty(list))
1173 WARN_ON(!list_is_singular(list) && got_budget);
1176 * Now process all the entries, sending them to the driver.
1178 errors = queued = 0;
1180 struct blk_mq_queue_data bd;
1182 rq = list_first_entry(list, struct request, queuelist);
1184 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1185 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1188 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1190 * The initial allocation attempt failed, so we need to
1191 * rerun the hardware queue when a tag is freed. The
1192 * waitqueue takes care of that. If the queue is run
1193 * before we add this entry back on the dispatch list,
1194 * we'll re-run it below.
1196 if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1197 blk_mq_put_dispatch_budget(hctx);
1199 * For non-shared tags, the RESTART check
1202 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1208 list_del_init(&rq->queuelist);
1213 * Flag last if we have no more requests, or if we have more
1214 * but can't assign a driver tag to it.
1216 if (list_empty(list))
1219 nxt = list_first_entry(list, struct request, queuelist);
1220 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1223 ret = q->mq_ops->queue_rq(hctx, &bd);
1224 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1226 * If an I/O scheduler has been configured and we got a
1227 * driver tag for the next request already, free it
1230 if (!list_empty(list)) {
1231 nxt = list_first_entry(list, struct request, queuelist);
1232 blk_mq_put_driver_tag(nxt);
1234 list_add(&rq->queuelist, list);
1235 __blk_mq_requeue_request(rq);
1239 if (unlikely(ret != BLK_STS_OK)) {
1241 blk_mq_end_request(rq, BLK_STS_IOERR);
1246 } while (!list_empty(list));
1248 hctx->dispatched[queued_to_index(queued)]++;
1251 * Any items that need requeuing? Stuff them into hctx->dispatch,
1252 * that is where we will continue on next queue run.
1254 if (!list_empty(list)) {
1257 spin_lock(&hctx->lock);
1258 list_splice_init(list, &hctx->dispatch);
1259 spin_unlock(&hctx->lock);
1262 * If SCHED_RESTART was set by the caller of this function and
1263 * it is no longer set that means that it was cleared by another
1264 * thread and hence that a queue rerun is needed.
1266 * If 'no_tag' is set, that means that we failed getting
1267 * a driver tag with an I/O scheduler attached. If our dispatch
1268 * waitqueue is no longer active, ensure that we run the queue
1269 * AFTER adding our entries back to the list.
1271 * If no I/O scheduler has been configured it is possible that
1272 * the hardware queue got stopped and restarted before requests
1273 * were pushed back onto the dispatch list. Rerun the queue to
1274 * avoid starvation. Notes:
1275 * - blk_mq_run_hw_queue() checks whether or not a queue has
1276 * been stopped before rerunning a queue.
1277 * - Some but not all block drivers stop a queue before
1278 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1281 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1282 * bit is set, run queue after a delay to avoid IO stalls
1283 * that could otherwise occur if the queue is idle.
1285 needs_restart = blk_mq_sched_needs_restart(hctx);
1286 if (!needs_restart ||
1287 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1288 blk_mq_run_hw_queue(hctx, true);
1289 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1290 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1293 return (queued + errors) != 0;
1296 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1301 * We should be running this queue from one of the CPUs that
1304 * There are at least two related races now between setting
1305 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1306 * __blk_mq_run_hw_queue():
1308 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1309 * but later it becomes online, then this warning is harmless
1312 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1313 * but later it becomes offline, then the warning can't be
1314 * triggered, and we depend on blk-mq timeout handler to
1315 * handle dispatched requests to this hctx
1317 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1318 cpu_online(hctx->next_cpu)) {
1319 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1320 raw_smp_processor_id(),
1321 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1326 * We can't run the queue inline with ints disabled. Ensure that
1327 * we catch bad users of this early.
1329 WARN_ON_ONCE(in_interrupt());
1331 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1333 hctx_lock(hctx, &srcu_idx);
1334 blk_mq_sched_dispatch_requests(hctx);
1335 hctx_unlock(hctx, srcu_idx);
1338 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1340 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1342 if (cpu >= nr_cpu_ids)
1343 cpu = cpumask_first(hctx->cpumask);
1348 * It'd be great if the workqueue API had a way to pass
1349 * in a mask and had some smarts for more clever placement.
1350 * For now we just round-robin here, switching for every
1351 * BLK_MQ_CPU_WORK_BATCH queued items.
1353 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1356 int next_cpu = hctx->next_cpu;
1358 if (hctx->queue->nr_hw_queues == 1)
1359 return WORK_CPU_UNBOUND;
1361 if (--hctx->next_cpu_batch <= 0) {
1363 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1365 if (next_cpu >= nr_cpu_ids)
1366 next_cpu = blk_mq_first_mapped_cpu(hctx);
1367 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1371 * Do unbound schedule if we can't find a online CPU for this hctx,
1372 * and it should only happen in the path of handling CPU DEAD.
1374 if (!cpu_online(next_cpu)) {
1381 * Make sure to re-select CPU next time once after CPUs
1382 * in hctx->cpumask become online again.
1384 hctx->next_cpu = next_cpu;
1385 hctx->next_cpu_batch = 1;
1386 return WORK_CPU_UNBOUND;
1389 hctx->next_cpu = next_cpu;
1393 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1394 unsigned long msecs)
1396 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1399 if (unlikely(blk_mq_hctx_stopped(hctx)))
1402 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1403 int cpu = get_cpu();
1404 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1405 __blk_mq_run_hw_queue(hctx);
1413 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1414 msecs_to_jiffies(msecs));
1417 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1419 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1421 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1423 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1429 * When queue is quiesced, we may be switching io scheduler, or
1430 * updating nr_hw_queues, or other things, and we can't run queue
1431 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1433 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1436 hctx_lock(hctx, &srcu_idx);
1437 need_run = !blk_queue_quiesced(hctx->queue) &&
1438 blk_mq_hctx_has_pending(hctx);
1439 hctx_unlock(hctx, srcu_idx);
1442 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1448 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1450 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1452 struct blk_mq_hw_ctx *hctx;
1455 queue_for_each_hw_ctx(q, hctx, i) {
1456 if (blk_mq_hctx_stopped(hctx))
1459 blk_mq_run_hw_queue(hctx, async);
1462 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1465 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1466 * @q: request queue.
1468 * The caller is responsible for serializing this function against
1469 * blk_mq_{start,stop}_hw_queue().
1471 bool blk_mq_queue_stopped(struct request_queue *q)
1473 struct blk_mq_hw_ctx *hctx;
1476 queue_for_each_hw_ctx(q, hctx, i)
1477 if (blk_mq_hctx_stopped(hctx))
1482 EXPORT_SYMBOL(blk_mq_queue_stopped);
1485 * This function is often used for pausing .queue_rq() by driver when
1486 * there isn't enough resource or some conditions aren't satisfied, and
1487 * BLK_STS_RESOURCE is usually returned.
1489 * We do not guarantee that dispatch can be drained or blocked
1490 * after blk_mq_stop_hw_queue() returns. Please use
1491 * blk_mq_quiesce_queue() for that requirement.
1493 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1495 cancel_delayed_work(&hctx->run_work);
1497 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1499 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1502 * This function is often used for pausing .queue_rq() by driver when
1503 * there isn't enough resource or some conditions aren't satisfied, and
1504 * BLK_STS_RESOURCE is usually returned.
1506 * We do not guarantee that dispatch can be drained or blocked
1507 * after blk_mq_stop_hw_queues() returns. Please use
1508 * blk_mq_quiesce_queue() for that requirement.
1510 void blk_mq_stop_hw_queues(struct request_queue *q)
1512 struct blk_mq_hw_ctx *hctx;
1515 queue_for_each_hw_ctx(q, hctx, i)
1516 blk_mq_stop_hw_queue(hctx);
1518 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1520 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1522 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1524 blk_mq_run_hw_queue(hctx, false);
1526 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1528 void blk_mq_start_hw_queues(struct request_queue *q)
1530 struct blk_mq_hw_ctx *hctx;
1533 queue_for_each_hw_ctx(q, hctx, i)
1534 blk_mq_start_hw_queue(hctx);
1536 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1538 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1540 if (!blk_mq_hctx_stopped(hctx))
1543 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1544 blk_mq_run_hw_queue(hctx, async);
1546 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1548 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1550 struct blk_mq_hw_ctx *hctx;
1553 queue_for_each_hw_ctx(q, hctx, i)
1554 blk_mq_start_stopped_hw_queue(hctx, async);
1556 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1558 static void blk_mq_run_work_fn(struct work_struct *work)
1560 struct blk_mq_hw_ctx *hctx;
1562 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1565 * If we are stopped, don't run the queue. The exception is if
1566 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1567 * the STOPPED bit and run it.
1569 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1570 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1573 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1574 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1577 __blk_mq_run_hw_queue(hctx);
1581 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1583 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1587 * Stop the hw queue, then modify currently delayed work.
1588 * This should prevent us from running the queue prematurely.
1589 * Mark the queue as auto-clearing STOPPED when it runs.
1591 blk_mq_stop_hw_queue(hctx);
1592 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1593 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1595 msecs_to_jiffies(msecs));
1597 EXPORT_SYMBOL(blk_mq_delay_queue);
1599 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1603 struct blk_mq_ctx *ctx = rq->mq_ctx;
1605 lockdep_assert_held(&ctx->lock);
1607 trace_block_rq_insert(hctx->queue, rq);
1610 list_add(&rq->queuelist, &ctx->rq_list);
1612 list_add_tail(&rq->queuelist, &ctx->rq_list);
1615 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1618 struct blk_mq_ctx *ctx = rq->mq_ctx;
1620 lockdep_assert_held(&ctx->lock);
1622 __blk_mq_insert_req_list(hctx, rq, at_head);
1623 blk_mq_hctx_mark_pending(hctx, ctx);
1627 * Should only be used carefully, when the caller knows we want to
1628 * bypass a potential IO scheduler on the target device.
1630 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1632 struct blk_mq_ctx *ctx = rq->mq_ctx;
1633 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1635 spin_lock(&hctx->lock);
1636 list_add_tail(&rq->queuelist, &hctx->dispatch);
1637 spin_unlock(&hctx->lock);
1640 blk_mq_run_hw_queue(hctx, false);
1643 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1644 struct list_head *list)
1648 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1651 spin_lock(&ctx->lock);
1652 while (!list_empty(list)) {
1655 rq = list_first_entry(list, struct request, queuelist);
1656 BUG_ON(rq->mq_ctx != ctx);
1657 list_del_init(&rq->queuelist);
1658 __blk_mq_insert_req_list(hctx, rq, false);
1660 blk_mq_hctx_mark_pending(hctx, ctx);
1661 spin_unlock(&ctx->lock);
1664 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1666 struct request *rqa = container_of(a, struct request, queuelist);
1667 struct request *rqb = container_of(b, struct request, queuelist);
1669 return !(rqa->mq_ctx < rqb->mq_ctx ||
1670 (rqa->mq_ctx == rqb->mq_ctx &&
1671 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1674 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1676 struct blk_mq_ctx *this_ctx;
1677 struct request_queue *this_q;
1680 LIST_HEAD(ctx_list);
1683 list_splice_init(&plug->mq_list, &list);
1685 list_sort(NULL, &list, plug_ctx_cmp);
1691 while (!list_empty(&list)) {
1692 rq = list_entry_rq(list.next);
1693 list_del_init(&rq->queuelist);
1695 if (rq->mq_ctx != this_ctx) {
1697 trace_block_unplug(this_q, depth, from_schedule);
1698 blk_mq_sched_insert_requests(this_q, this_ctx,
1703 this_ctx = rq->mq_ctx;
1709 list_add_tail(&rq->queuelist, &ctx_list);
1713 * If 'this_ctx' is set, we know we have entries to complete
1714 * on 'ctx_list'. Do those.
1717 trace_block_unplug(this_q, depth, from_schedule);
1718 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1723 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1725 blk_init_request_from_bio(rq, bio);
1727 blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1729 blk_account_io_start(rq, true);
1732 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1733 struct blk_mq_ctx *ctx,
1736 spin_lock(&ctx->lock);
1737 __blk_mq_insert_request(hctx, rq, false);
1738 spin_unlock(&ctx->lock);
1741 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1744 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1746 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1749 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1753 struct request_queue *q = rq->q;
1754 struct blk_mq_queue_data bd = {
1758 blk_qc_t new_cookie;
1761 new_cookie = request_to_qc_t(hctx, rq);
1764 * For OK queue, we are done. For error, caller may kill it.
1765 * Any other error (busy), just add it to our list as we
1766 * previously would have done.
1768 ret = q->mq_ops->queue_rq(hctx, &bd);
1771 *cookie = new_cookie;
1773 case BLK_STS_RESOURCE:
1774 case BLK_STS_DEV_RESOURCE:
1775 __blk_mq_requeue_request(rq);
1778 *cookie = BLK_QC_T_NONE;
1785 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1790 struct request_queue *q = rq->q;
1791 bool run_queue = true;
1794 * RCU or SRCU read lock is needed before checking quiesced flag.
1796 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1797 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1798 * and avoid driver to try to dispatch again.
1800 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1802 bypass_insert = false;
1806 if (q->elevator && !bypass_insert)
1809 if (!blk_mq_get_dispatch_budget(hctx))
1812 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1813 blk_mq_put_dispatch_budget(hctx);
1817 return __blk_mq_issue_directly(hctx, rq, cookie);
1820 return BLK_STS_RESOURCE;
1822 blk_mq_sched_insert_request(rq, false, run_queue, false);
1826 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1827 struct request *rq, blk_qc_t *cookie)
1832 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1834 hctx_lock(hctx, &srcu_idx);
1836 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1837 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1838 blk_mq_sched_insert_request(rq, false, true, false);
1839 else if (ret != BLK_STS_OK)
1840 blk_mq_end_request(rq, ret);
1842 hctx_unlock(hctx, srcu_idx);
1845 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1849 blk_qc_t unused_cookie;
1850 struct blk_mq_ctx *ctx = rq->mq_ctx;
1851 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1853 hctx_lock(hctx, &srcu_idx);
1854 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1855 hctx_unlock(hctx, srcu_idx);
1860 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1862 const int is_sync = op_is_sync(bio->bi_opf);
1863 const int is_flush_fua = op_is_flush(bio->bi_opf);
1864 struct blk_mq_alloc_data data = { .flags = 0 };
1866 unsigned int request_count = 0;
1867 struct blk_plug *plug;
1868 struct request *same_queue_rq = NULL;
1870 unsigned int wb_acct;
1872 blk_queue_bounce(q, &bio);
1874 blk_queue_split(q, &bio);
1876 if (!bio_integrity_prep(bio))
1877 return BLK_QC_T_NONE;
1879 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1880 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1881 return BLK_QC_T_NONE;
1883 if (blk_mq_sched_bio_merge(q, bio))
1884 return BLK_QC_T_NONE;
1886 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1888 trace_block_getrq(q, bio, bio->bi_opf);
1890 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1891 if (unlikely(!rq)) {
1892 __wbt_done(q->rq_wb, wb_acct);
1893 if (bio->bi_opf & REQ_NOWAIT)
1894 bio_wouldblock_error(bio);
1895 return BLK_QC_T_NONE;
1898 wbt_track(&rq->issue_stat, wb_acct);
1900 cookie = request_to_qc_t(data.hctx, rq);
1902 plug = current->plug;
1903 if (unlikely(is_flush_fua)) {
1904 blk_mq_put_ctx(data.ctx);
1905 blk_mq_bio_to_request(rq, bio);
1907 /* bypass scheduler for flush rq */
1908 blk_insert_flush(rq);
1909 blk_mq_run_hw_queue(data.hctx, true);
1910 } else if (plug && q->nr_hw_queues == 1) {
1911 struct request *last = NULL;
1913 blk_mq_put_ctx(data.ctx);
1914 blk_mq_bio_to_request(rq, bio);
1917 * @request_count may become stale because of schedule
1918 * out, so check the list again.
1920 if (list_empty(&plug->mq_list))
1922 else if (blk_queue_nomerges(q))
1923 request_count = blk_plug_queued_count(q);
1926 trace_block_plug(q);
1928 last = list_entry_rq(plug->mq_list.prev);
1930 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1931 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1932 blk_flush_plug_list(plug, false);
1933 trace_block_plug(q);
1936 list_add_tail(&rq->queuelist, &plug->mq_list);
1937 } else if (plug && !blk_queue_nomerges(q)) {
1938 blk_mq_bio_to_request(rq, bio);
1941 * We do limited plugging. If the bio can be merged, do that.
1942 * Otherwise the existing request in the plug list will be
1943 * issued. So the plug list will have one request at most
1944 * The plug list might get flushed before this. If that happens,
1945 * the plug list is empty, and same_queue_rq is invalid.
1947 if (list_empty(&plug->mq_list))
1948 same_queue_rq = NULL;
1950 list_del_init(&same_queue_rq->queuelist);
1951 list_add_tail(&rq->queuelist, &plug->mq_list);
1953 blk_mq_put_ctx(data.ctx);
1955 if (same_queue_rq) {
1956 data.hctx = blk_mq_map_queue(q,
1957 same_queue_rq->mq_ctx->cpu);
1958 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1961 } else if (q->nr_hw_queues > 1 && is_sync) {
1962 blk_mq_put_ctx(data.ctx);
1963 blk_mq_bio_to_request(rq, bio);
1964 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1965 } else if (q->elevator) {
1966 blk_mq_put_ctx(data.ctx);
1967 blk_mq_bio_to_request(rq, bio);
1968 blk_mq_sched_insert_request(rq, false, true, true);
1970 blk_mq_put_ctx(data.ctx);
1971 blk_mq_bio_to_request(rq, bio);
1972 blk_mq_queue_io(data.hctx, data.ctx, rq);
1973 blk_mq_run_hw_queue(data.hctx, true);
1979 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1980 unsigned int hctx_idx)
1984 if (tags->rqs && set->ops->exit_request) {
1987 for (i = 0; i < tags->nr_tags; i++) {
1988 struct request *rq = tags->static_rqs[i];
1992 set->ops->exit_request(set, rq, hctx_idx);
1993 tags->static_rqs[i] = NULL;
1997 while (!list_empty(&tags->page_list)) {
1998 page = list_first_entry(&tags->page_list, struct page, lru);
1999 list_del_init(&page->lru);
2001 * Remove kmemleak object previously allocated in
2002 * blk_mq_init_rq_map().
2004 kmemleak_free(page_address(page));
2005 __free_pages(page, page->private);
2009 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2013 kfree(tags->static_rqs);
2014 tags->static_rqs = NULL;
2016 blk_mq_free_tags(tags);
2019 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2020 unsigned int hctx_idx,
2021 unsigned int nr_tags,
2022 unsigned int reserved_tags)
2024 struct blk_mq_tags *tags;
2027 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2028 if (node == NUMA_NO_NODE)
2029 node = set->numa_node;
2031 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2032 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2036 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2037 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2040 blk_mq_free_tags(tags);
2044 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2045 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2047 if (!tags->static_rqs) {
2049 blk_mq_free_tags(tags);
2056 static size_t order_to_size(unsigned int order)
2058 return (size_t)PAGE_SIZE << order;
2061 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2062 unsigned int hctx_idx, int node)
2066 if (set->ops->init_request) {
2067 ret = set->ops->init_request(set, rq, hctx_idx, node);
2072 seqcount_init(&rq->gstate_seq);
2073 u64_stats_init(&rq->aborted_gstate_sync);
2077 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2078 unsigned int hctx_idx, unsigned int depth)
2080 unsigned int i, j, entries_per_page, max_order = 4;
2081 size_t rq_size, left;
2084 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2085 if (node == NUMA_NO_NODE)
2086 node = set->numa_node;
2088 INIT_LIST_HEAD(&tags->page_list);
2091 * rq_size is the size of the request plus driver payload, rounded
2092 * to the cacheline size
2094 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2096 left = rq_size * depth;
2098 for (i = 0; i < depth; ) {
2099 int this_order = max_order;
2104 while (this_order && left < order_to_size(this_order - 1))
2108 page = alloc_pages_node(node,
2109 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2115 if (order_to_size(this_order) < rq_size)
2122 page->private = this_order;
2123 list_add_tail(&page->lru, &tags->page_list);
2125 p = page_address(page);
2127 * Allow kmemleak to scan these pages as they contain pointers
2128 * to additional allocations like via ops->init_request().
2130 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2131 entries_per_page = order_to_size(this_order) / rq_size;
2132 to_do = min(entries_per_page, depth - i);
2133 left -= to_do * rq_size;
2134 for (j = 0; j < to_do; j++) {
2135 struct request *rq = p;
2137 tags->static_rqs[i] = rq;
2138 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2139 tags->static_rqs[i] = NULL;
2150 blk_mq_free_rqs(set, tags, hctx_idx);
2155 * 'cpu' is going away. splice any existing rq_list entries from this
2156 * software queue to the hw queue dispatch list, and ensure that it
2159 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2161 struct blk_mq_hw_ctx *hctx;
2162 struct blk_mq_ctx *ctx;
2165 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2166 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2168 spin_lock(&ctx->lock);
2169 if (!list_empty(&ctx->rq_list)) {
2170 list_splice_init(&ctx->rq_list, &tmp);
2171 blk_mq_hctx_clear_pending(hctx, ctx);
2173 spin_unlock(&ctx->lock);
2175 if (list_empty(&tmp))
2178 spin_lock(&hctx->lock);
2179 list_splice_tail_init(&tmp, &hctx->dispatch);
2180 spin_unlock(&hctx->lock);
2182 blk_mq_run_hw_queue(hctx, true);
2186 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2188 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2192 /* hctx->ctxs will be freed in queue's release handler */
2193 static void blk_mq_exit_hctx(struct request_queue *q,
2194 struct blk_mq_tag_set *set,
2195 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2197 blk_mq_debugfs_unregister_hctx(hctx);
2199 if (blk_mq_hw_queue_mapped(hctx))
2200 blk_mq_tag_idle(hctx);
2202 if (set->ops->exit_request)
2203 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2205 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2207 if (set->ops->exit_hctx)
2208 set->ops->exit_hctx(hctx, hctx_idx);
2210 if (hctx->flags & BLK_MQ_F_BLOCKING)
2211 cleanup_srcu_struct(hctx->srcu);
2213 blk_mq_remove_cpuhp(hctx);
2214 blk_free_flush_queue(hctx->fq);
2215 sbitmap_free(&hctx->ctx_map);
2218 static void blk_mq_exit_hw_queues(struct request_queue *q,
2219 struct blk_mq_tag_set *set, int nr_queue)
2221 struct blk_mq_hw_ctx *hctx;
2224 queue_for_each_hw_ctx(q, hctx, i) {
2227 blk_mq_exit_hctx(q, set, hctx, i);
2231 static int blk_mq_init_hctx(struct request_queue *q,
2232 struct blk_mq_tag_set *set,
2233 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2237 node = hctx->numa_node;
2238 if (node == NUMA_NO_NODE)
2239 node = hctx->numa_node = set->numa_node;
2241 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2242 spin_lock_init(&hctx->lock);
2243 INIT_LIST_HEAD(&hctx->dispatch);
2245 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2247 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2249 hctx->tags = set->tags[hctx_idx];
2252 * Allocate space for all possible cpus to avoid allocation at
2255 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2258 goto unregister_cpu_notifier;
2260 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2266 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2267 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2269 if (set->ops->init_hctx &&
2270 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2273 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2276 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2278 goto sched_exit_hctx;
2280 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2283 if (hctx->flags & BLK_MQ_F_BLOCKING)
2284 init_srcu_struct(hctx->srcu);
2286 blk_mq_debugfs_register_hctx(q, hctx);
2293 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2295 if (set->ops->exit_hctx)
2296 set->ops->exit_hctx(hctx, hctx_idx);
2298 sbitmap_free(&hctx->ctx_map);
2301 unregister_cpu_notifier:
2302 blk_mq_remove_cpuhp(hctx);
2306 static void blk_mq_init_cpu_queues(struct request_queue *q,
2307 unsigned int nr_hw_queues)
2311 for_each_possible_cpu(i) {
2312 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2313 struct blk_mq_hw_ctx *hctx;
2316 spin_lock_init(&__ctx->lock);
2317 INIT_LIST_HEAD(&__ctx->rq_list);
2321 * Set local node, IFF we have more than one hw queue. If
2322 * not, we remain on the home node of the device
2324 hctx = blk_mq_map_queue(q, i);
2325 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2326 hctx->numa_node = local_memory_node(cpu_to_node(i));
2330 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2334 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2335 set->queue_depth, set->reserved_tags);
2336 if (!set->tags[hctx_idx])
2339 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2344 blk_mq_free_rq_map(set->tags[hctx_idx]);
2345 set->tags[hctx_idx] = NULL;
2349 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2350 unsigned int hctx_idx)
2352 if (set->tags[hctx_idx]) {
2353 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2354 blk_mq_free_rq_map(set->tags[hctx_idx]);
2355 set->tags[hctx_idx] = NULL;
2359 static void blk_mq_map_swqueue(struct request_queue *q)
2361 unsigned int i, hctx_idx;
2362 struct blk_mq_hw_ctx *hctx;
2363 struct blk_mq_ctx *ctx;
2364 struct blk_mq_tag_set *set = q->tag_set;
2367 * Avoid others reading imcomplete hctx->cpumask through sysfs
2369 mutex_lock(&q->sysfs_lock);
2371 queue_for_each_hw_ctx(q, hctx, i) {
2372 cpumask_clear(hctx->cpumask);
2377 * Map software to hardware queues.
2379 * If the cpu isn't present, the cpu is mapped to first hctx.
2381 for_each_possible_cpu(i) {
2382 hctx_idx = q->mq_map[i];
2383 /* unmapped hw queue can be remapped after CPU topo changed */
2384 if (!set->tags[hctx_idx] &&
2385 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2387 * If tags initialization fail for some hctx,
2388 * that hctx won't be brought online. In this
2389 * case, remap the current ctx to hctx[0] which
2390 * is guaranteed to always have tags allocated
2395 ctx = per_cpu_ptr(q->queue_ctx, i);
2396 hctx = blk_mq_map_queue(q, i);
2398 cpumask_set_cpu(i, hctx->cpumask);
2399 ctx->index_hw = hctx->nr_ctx;
2400 hctx->ctxs[hctx->nr_ctx++] = ctx;
2403 mutex_unlock(&q->sysfs_lock);
2405 queue_for_each_hw_ctx(q, hctx, i) {
2407 * If no software queues are mapped to this hardware queue,
2408 * disable it and free the request entries.
2410 if (!hctx->nr_ctx) {
2411 /* Never unmap queue 0. We need it as a
2412 * fallback in case of a new remap fails
2415 if (i && set->tags[i])
2416 blk_mq_free_map_and_requests(set, i);
2422 hctx->tags = set->tags[i];
2423 WARN_ON(!hctx->tags);
2426 * Set the map size to the number of mapped software queues.
2427 * This is more accurate and more efficient than looping
2428 * over all possibly mapped software queues.
2430 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2433 * Initialize batch roundrobin counts
2435 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2436 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2441 * Caller needs to ensure that we're either frozen/quiesced, or that
2442 * the queue isn't live yet.
2444 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2446 struct blk_mq_hw_ctx *hctx;
2449 queue_for_each_hw_ctx(q, hctx, i) {
2451 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2452 atomic_inc(&q->shared_hctx_restart);
2453 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2455 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2456 atomic_dec(&q->shared_hctx_restart);
2457 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2462 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2465 struct request_queue *q;
2467 lockdep_assert_held(&set->tag_list_lock);
2469 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2470 blk_mq_freeze_queue(q);
2471 queue_set_hctx_shared(q, shared);
2472 blk_mq_unfreeze_queue(q);
2476 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2478 struct blk_mq_tag_set *set = q->tag_set;
2480 mutex_lock(&set->tag_list_lock);
2481 list_del_rcu(&q->tag_set_list);
2482 INIT_LIST_HEAD(&q->tag_set_list);
2483 if (list_is_singular(&set->tag_list)) {
2484 /* just transitioned to unshared */
2485 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2486 /* update existing queue */
2487 blk_mq_update_tag_set_depth(set, false);
2489 mutex_unlock(&set->tag_list_lock);
2494 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2495 struct request_queue *q)
2499 mutex_lock(&set->tag_list_lock);
2502 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2504 if (!list_empty(&set->tag_list) &&
2505 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2506 set->flags |= BLK_MQ_F_TAG_SHARED;
2507 /* update existing queue */
2508 blk_mq_update_tag_set_depth(set, true);
2510 if (set->flags & BLK_MQ_F_TAG_SHARED)
2511 queue_set_hctx_shared(q, true);
2512 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2514 mutex_unlock(&set->tag_list_lock);
2518 * It is the actual release handler for mq, but we do it from
2519 * request queue's release handler for avoiding use-after-free
2520 * and headache because q->mq_kobj shouldn't have been introduced,
2521 * but we can't group ctx/kctx kobj without it.
2523 void blk_mq_release(struct request_queue *q)
2525 struct blk_mq_hw_ctx *hctx;
2528 /* hctx kobj stays in hctx */
2529 queue_for_each_hw_ctx(q, hctx, i) {
2532 kobject_put(&hctx->kobj);
2537 kfree(q->queue_hw_ctx);
2540 * release .mq_kobj and sw queue's kobject now because
2541 * both share lifetime with request queue.
2543 blk_mq_sysfs_deinit(q);
2545 free_percpu(q->queue_ctx);
2548 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2550 struct request_queue *uninit_q, *q;
2552 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2554 return ERR_PTR(-ENOMEM);
2556 q = blk_mq_init_allocated_queue(set, uninit_q);
2558 blk_cleanup_queue(uninit_q);
2562 EXPORT_SYMBOL(blk_mq_init_queue);
2564 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2566 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2568 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2569 __alignof__(struct blk_mq_hw_ctx)) !=
2570 sizeof(struct blk_mq_hw_ctx));
2572 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2573 hw_ctx_size += sizeof(struct srcu_struct);
2578 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2579 struct request_queue *q)
2582 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2584 blk_mq_sysfs_unregister(q);
2586 /* protect against switching io scheduler */
2587 mutex_lock(&q->sysfs_lock);
2588 for (i = 0; i < set->nr_hw_queues; i++) {
2594 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2595 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2600 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2607 atomic_set(&hctxs[i]->nr_active, 0);
2608 hctxs[i]->numa_node = node;
2609 hctxs[i]->queue_num = i;
2611 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2612 free_cpumask_var(hctxs[i]->cpumask);
2617 blk_mq_hctx_kobj_init(hctxs[i]);
2619 for (j = i; j < q->nr_hw_queues; j++) {
2620 struct blk_mq_hw_ctx *hctx = hctxs[j];
2624 blk_mq_free_map_and_requests(set, j);
2625 blk_mq_exit_hctx(q, set, hctx, j);
2626 kobject_put(&hctx->kobj);
2631 q->nr_hw_queues = i;
2632 mutex_unlock(&q->sysfs_lock);
2633 blk_mq_sysfs_register(q);
2636 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2637 struct request_queue *q)
2639 /* mark the queue as mq asap */
2640 q->mq_ops = set->ops;
2642 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2643 blk_mq_poll_stats_bkt,
2644 BLK_MQ_POLL_STATS_BKTS, q);
2648 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2652 /* init q->mq_kobj and sw queues' kobjects */
2653 blk_mq_sysfs_init(q);
2655 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2656 GFP_KERNEL, set->numa_node);
2657 if (!q->queue_hw_ctx)
2660 q->mq_map = set->mq_map;
2662 blk_mq_realloc_hw_ctxs(set, q);
2663 if (!q->nr_hw_queues)
2666 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2667 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2669 q->nr_queues = nr_cpu_ids;
2671 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2673 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2674 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2676 q->sg_reserved_size = INT_MAX;
2678 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2679 INIT_LIST_HEAD(&q->requeue_list);
2680 spin_lock_init(&q->requeue_lock);
2682 blk_queue_make_request(q, blk_mq_make_request);
2683 if (q->mq_ops->poll)
2684 q->poll_fn = blk_mq_poll;
2687 * Do this after blk_queue_make_request() overrides it...
2689 q->nr_requests = set->queue_depth;
2692 * Default to classic polling
2696 if (set->ops->complete)
2697 blk_queue_softirq_done(q, set->ops->complete);
2699 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2700 blk_mq_add_queue_tag_set(set, q);
2701 blk_mq_map_swqueue(q);
2703 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2706 ret = blk_mq_sched_init(q);
2708 return ERR_PTR(ret);
2714 kfree(q->queue_hw_ctx);
2716 free_percpu(q->queue_ctx);
2719 return ERR_PTR(-ENOMEM);
2721 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2723 void blk_mq_free_queue(struct request_queue *q)
2725 struct blk_mq_tag_set *set = q->tag_set;
2727 blk_mq_del_queue_tag_set(q);
2728 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2731 /* Basically redo blk_mq_init_queue with queue frozen */
2732 static void blk_mq_queue_reinit(struct request_queue *q)
2734 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2736 blk_mq_debugfs_unregister_hctxs(q);
2737 blk_mq_sysfs_unregister(q);
2740 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2741 * we should change hctx numa_node according to the new topology (this
2742 * involves freeing and re-allocating memory, worth doing?)
2744 blk_mq_map_swqueue(q);
2746 blk_mq_sysfs_register(q);
2747 blk_mq_debugfs_register_hctxs(q);
2750 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2754 for (i = 0; i < set->nr_hw_queues; i++)
2755 if (!__blk_mq_alloc_rq_map(set, i))
2762 blk_mq_free_rq_map(set->tags[i]);
2768 * Allocate the request maps associated with this tag_set. Note that this
2769 * may reduce the depth asked for, if memory is tight. set->queue_depth
2770 * will be updated to reflect the allocated depth.
2772 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2777 depth = set->queue_depth;
2779 err = __blk_mq_alloc_rq_maps(set);
2783 set->queue_depth >>= 1;
2784 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2788 } while (set->queue_depth);
2790 if (!set->queue_depth || err) {
2791 pr_err("blk-mq: failed to allocate request map\n");
2795 if (depth != set->queue_depth)
2796 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2797 depth, set->queue_depth);
2802 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2804 if (set->ops->map_queues) {
2807 * transport .map_queues is usually done in the following
2810 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2811 * mask = get_cpu_mask(queue)
2812 * for_each_cpu(cpu, mask)
2813 * set->mq_map[cpu] = queue;
2816 * When we need to remap, the table has to be cleared for
2817 * killing stale mapping since one CPU may not be mapped
2820 for_each_possible_cpu(cpu)
2821 set->mq_map[cpu] = 0;
2823 return set->ops->map_queues(set);
2825 return blk_mq_map_queues(set);
2829 * Alloc a tag set to be associated with one or more request queues.
2830 * May fail with EINVAL for various error conditions. May adjust the
2831 * requested depth down, if if it too large. In that case, the set
2832 * value will be stored in set->queue_depth.
2834 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2838 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2840 if (!set->nr_hw_queues)
2842 if (!set->queue_depth)
2844 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2847 if (!set->ops->queue_rq)
2850 if (!set->ops->get_budget ^ !set->ops->put_budget)
2853 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2854 pr_info("blk-mq: reduced tag depth to %u\n",
2856 set->queue_depth = BLK_MQ_MAX_DEPTH;
2860 * If a crashdump is active, then we are potentially in a very
2861 * memory constrained environment. Limit us to 1 queue and
2862 * 64 tags to prevent using too much memory.
2864 if (is_kdump_kernel()) {
2865 set->nr_hw_queues = 1;
2866 set->queue_depth = min(64U, set->queue_depth);
2869 * There is no use for more h/w queues than cpus.
2871 if (set->nr_hw_queues > nr_cpu_ids)
2872 set->nr_hw_queues = nr_cpu_ids;
2874 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2875 GFP_KERNEL, set->numa_node);
2880 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2881 GFP_KERNEL, set->numa_node);
2885 ret = blk_mq_update_queue_map(set);
2887 goto out_free_mq_map;
2889 ret = blk_mq_alloc_rq_maps(set);
2891 goto out_free_mq_map;
2893 mutex_init(&set->tag_list_lock);
2894 INIT_LIST_HEAD(&set->tag_list);
2906 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2908 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2912 for (i = 0; i < nr_cpu_ids; i++)
2913 blk_mq_free_map_and_requests(set, i);
2921 EXPORT_SYMBOL(blk_mq_free_tag_set);
2923 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2925 struct blk_mq_tag_set *set = q->tag_set;
2926 struct blk_mq_hw_ctx *hctx;
2932 blk_mq_freeze_queue(q);
2933 blk_mq_quiesce_queue(q);
2936 queue_for_each_hw_ctx(q, hctx, i) {
2940 * If we're using an MQ scheduler, just update the scheduler
2941 * queue depth. This is similar to what the old code would do.
2943 if (!hctx->sched_tags) {
2944 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2947 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2955 q->nr_requests = nr;
2957 blk_mq_unquiesce_queue(q);
2958 blk_mq_unfreeze_queue(q);
2963 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2966 struct request_queue *q;
2968 lockdep_assert_held(&set->tag_list_lock);
2970 if (nr_hw_queues > nr_cpu_ids)
2971 nr_hw_queues = nr_cpu_ids;
2972 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2975 list_for_each_entry(q, &set->tag_list, tag_set_list)
2976 blk_mq_freeze_queue(q);
2978 set->nr_hw_queues = nr_hw_queues;
2979 blk_mq_update_queue_map(set);
2980 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2981 blk_mq_realloc_hw_ctxs(set, q);
2982 blk_mq_queue_reinit(q);
2985 list_for_each_entry(q, &set->tag_list, tag_set_list)
2986 blk_mq_unfreeze_queue(q);
2989 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2991 mutex_lock(&set->tag_list_lock);
2992 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2993 mutex_unlock(&set->tag_list_lock);
2995 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2997 /* Enable polling stats and return whether they were already enabled. */
2998 static bool blk_poll_stats_enable(struct request_queue *q)
3000 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3001 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3003 blk_stat_add_callback(q, q->poll_cb);
3007 static void blk_mq_poll_stats_start(struct request_queue *q)
3010 * We don't arm the callback if polling stats are not enabled or the
3011 * callback is already active.
3013 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3014 blk_stat_is_active(q->poll_cb))
3017 blk_stat_activate_msecs(q->poll_cb, 100);
3020 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3022 struct request_queue *q = cb->data;
3025 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3026 if (cb->stat[bucket].nr_samples)
3027 q->poll_stat[bucket] = cb->stat[bucket];
3031 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3032 struct blk_mq_hw_ctx *hctx,
3035 unsigned long ret = 0;
3039 * If stats collection isn't on, don't sleep but turn it on for
3042 if (!blk_poll_stats_enable(q))
3046 * As an optimistic guess, use half of the mean service time
3047 * for this type of request. We can (and should) make this smarter.
3048 * For instance, if the completion latencies are tight, we can
3049 * get closer than just half the mean. This is especially
3050 * important on devices where the completion latencies are longer
3051 * than ~10 usec. We do use the stats for the relevant IO size
3052 * if available which does lead to better estimates.
3054 bucket = blk_mq_poll_stats_bkt(rq);
3058 if (q->poll_stat[bucket].nr_samples)
3059 ret = (q->poll_stat[bucket].mean + 1) / 2;
3064 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3065 struct blk_mq_hw_ctx *hctx,
3068 struct hrtimer_sleeper hs;
3069 enum hrtimer_mode mode;
3073 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3079 * -1: don't ever hybrid sleep
3080 * 0: use half of prev avg
3081 * >0: use this specific value
3083 if (q->poll_nsec == -1)
3085 else if (q->poll_nsec > 0)
3086 nsecs = q->poll_nsec;
3088 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3093 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3096 * This will be replaced with the stats tracking code, using
3097 * 'avg_completion_time / 2' as the pre-sleep target.
3101 mode = HRTIMER_MODE_REL;
3102 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3103 hrtimer_set_expires(&hs.timer, kt);
3105 hrtimer_init_sleeper(&hs, current);
3107 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3109 set_current_state(TASK_UNINTERRUPTIBLE);
3110 hrtimer_start_expires(&hs.timer, mode);
3113 hrtimer_cancel(&hs.timer);
3114 mode = HRTIMER_MODE_ABS;
3115 } while (hs.task && !signal_pending(current));
3117 __set_current_state(TASK_RUNNING);
3118 destroy_hrtimer_on_stack(&hs.timer);
3122 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3124 struct request_queue *q = hctx->queue;
3128 * If we sleep, have the caller restart the poll loop to reset
3129 * the state. Like for the other success return cases, the
3130 * caller is responsible for checking if the IO completed. If
3131 * the IO isn't complete, we'll get called again and will go
3132 * straight to the busy poll loop.
3134 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3137 hctx->poll_considered++;
3139 state = current->state;
3140 while (!need_resched()) {
3143 hctx->poll_invoked++;
3145 ret = q->mq_ops->poll(hctx, rq->tag);
3147 hctx->poll_success++;
3148 set_current_state(TASK_RUNNING);
3152 if (signal_pending_state(state, current))
3153 set_current_state(TASK_RUNNING);
3155 if (current->state == TASK_RUNNING)
3162 __set_current_state(TASK_RUNNING);
3166 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3168 struct blk_mq_hw_ctx *hctx;
3171 if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3174 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3175 if (!blk_qc_t_is_internal(cookie))
3176 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3178 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3180 * With scheduling, if the request has completed, we'll
3181 * get a NULL return here, as we clear the sched tag when
3182 * that happens. The request still remains valid, like always,
3183 * so we should be safe with just the NULL check.
3189 return __blk_mq_poll(hctx, rq);
3192 static int __init blk_mq_init(void)
3194 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3195 blk_mq_hctx_notify_dead);
3198 subsys_initcall(blk_mq_init);