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;
99 * index[0] counts the specific partition that was asked for. index[1]
100 * counts the ones that are active on the whole device, so increment
101 * that if mi->part is indeed a partition, and not a whole device.
103 if (rq->part == mi->part)
105 if (mi->part->partno)
109 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
110 unsigned int inflight[2])
112 struct mq_inflight mi = { .part = part, .inflight = inflight, };
114 inflight[0] = inflight[1] = 0;
115 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
118 static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
119 struct request *rq, void *priv,
122 struct mq_inflight *mi = priv;
124 if (rq->part == mi->part)
125 mi->inflight[rq_data_dir(rq)]++;
128 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
129 unsigned int inflight[2])
131 struct mq_inflight mi = { .part = part, .inflight = inflight, };
133 inflight[0] = inflight[1] = 0;
134 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
137 void blk_freeze_queue_start(struct request_queue *q)
141 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
142 if (freeze_depth == 1) {
143 percpu_ref_kill(&q->q_usage_counter);
145 blk_mq_run_hw_queues(q, false);
148 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
150 void blk_mq_freeze_queue_wait(struct request_queue *q)
152 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
154 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
156 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
157 unsigned long timeout)
159 return wait_event_timeout(q->mq_freeze_wq,
160 percpu_ref_is_zero(&q->q_usage_counter),
163 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
166 * Guarantee no request is in use, so we can change any data structure of
167 * the queue afterward.
169 void blk_freeze_queue(struct request_queue *q)
172 * In the !blk_mq case we are only calling this to kill the
173 * q_usage_counter, otherwise this increases the freeze depth
174 * and waits for it to return to zero. For this reason there is
175 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
176 * exported to drivers as the only user for unfreeze is blk_mq.
178 blk_freeze_queue_start(q);
181 blk_mq_freeze_queue_wait(q);
184 void blk_mq_freeze_queue(struct request_queue *q)
187 * ...just an alias to keep freeze and unfreeze actions balanced
188 * in the blk_mq_* namespace
192 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
194 void blk_mq_unfreeze_queue(struct request_queue *q)
198 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
199 WARN_ON_ONCE(freeze_depth < 0);
201 percpu_ref_reinit(&q->q_usage_counter);
202 wake_up_all(&q->mq_freeze_wq);
205 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
208 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
209 * mpt3sas driver such that this function can be removed.
211 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
213 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
215 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
218 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
221 * Note: this function does not prevent that the struct request end_io()
222 * callback function is invoked. Once this function is returned, we make
223 * sure no dispatch can happen until the queue is unquiesced via
224 * blk_mq_unquiesce_queue().
226 void blk_mq_quiesce_queue(struct request_queue *q)
228 struct blk_mq_hw_ctx *hctx;
232 blk_mq_quiesce_queue_nowait(q);
234 queue_for_each_hw_ctx(q, hctx, i) {
235 if (hctx->flags & BLK_MQ_F_BLOCKING)
236 synchronize_srcu(hctx->srcu);
243 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
246 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
249 * This function recovers queue into the state before quiescing
250 * which is done by blk_mq_quiesce_queue.
252 void blk_mq_unquiesce_queue(struct request_queue *q)
254 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
256 /* dispatch requests which are inserted during quiescing */
257 blk_mq_run_hw_queues(q, true);
259 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
261 void blk_mq_wake_waiters(struct request_queue *q)
263 struct blk_mq_hw_ctx *hctx;
266 queue_for_each_hw_ctx(q, hctx, i)
267 if (blk_mq_hw_queue_mapped(hctx))
268 blk_mq_tag_wakeup_all(hctx->tags, true);
271 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
273 return blk_mq_has_free_tags(hctx->tags);
275 EXPORT_SYMBOL(blk_mq_can_queue);
277 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
278 unsigned int tag, unsigned int op)
280 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
281 struct request *rq = tags->static_rqs[tag];
282 req_flags_t rq_flags = 0;
284 if (data->flags & BLK_MQ_REQ_INTERNAL) {
286 rq->internal_tag = tag;
288 if (blk_mq_tag_busy(data->hctx)) {
289 rq_flags = RQF_MQ_INFLIGHT;
290 atomic_inc(&data->hctx->nr_active);
293 rq->internal_tag = -1;
294 data->hctx->tags->rqs[rq->tag] = rq;
297 /* csd/requeue_work/fifo_time is initialized before use */
299 rq->mq_ctx = data->ctx;
300 rq->rq_flags = rq_flags;
303 if (data->flags & BLK_MQ_REQ_PREEMPT)
304 rq->rq_flags |= RQF_PREEMPT;
305 if (blk_queue_io_stat(data->q))
306 rq->rq_flags |= RQF_IO_STAT;
307 INIT_LIST_HEAD(&rq->queuelist);
308 INIT_HLIST_NODE(&rq->hash);
309 RB_CLEAR_NODE(&rq->rb_node);
312 rq->start_time_ns = ktime_get_ns();
313 rq->io_start_time_ns = 0;
314 rq->nr_phys_segments = 0;
315 #if defined(CONFIG_BLK_DEV_INTEGRITY)
316 rq->nr_integrity_segments = 0;
319 /* tag was already set */
323 INIT_LIST_HEAD(&rq->timeout_list);
327 rq->end_io_data = NULL;
330 #ifdef CONFIG_BLK_CGROUP
334 data->ctx->rq_dispatched[op_is_sync(op)]++;
338 static struct request *blk_mq_get_request(struct request_queue *q,
339 struct bio *bio, unsigned int op,
340 struct blk_mq_alloc_data *data)
342 struct elevator_queue *e = q->elevator;
345 bool put_ctx_on_error = false;
347 blk_queue_enter_live(q);
349 if (likely(!data->ctx)) {
350 data->ctx = blk_mq_get_ctx(q);
351 put_ctx_on_error = true;
353 if (likely(!data->hctx))
354 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
356 data->flags |= BLK_MQ_REQ_NOWAIT;
359 data->flags |= BLK_MQ_REQ_INTERNAL;
362 * Flush requests are special and go directly to the
365 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
366 e->type->ops.mq.limit_depth(op, data);
369 tag = blk_mq_get_tag(data);
370 if (tag == BLK_MQ_TAG_FAIL) {
371 if (put_ctx_on_error) {
372 blk_mq_put_ctx(data->ctx);
379 rq = blk_mq_rq_ctx_init(data, tag, op);
380 if (!op_is_flush(op)) {
382 if (e && e->type->ops.mq.prepare_request) {
383 if (e->type->icq_cache && rq_ioc(bio))
384 blk_mq_sched_assign_ioc(rq, bio);
386 e->type->ops.mq.prepare_request(rq, bio);
387 rq->rq_flags |= RQF_ELVPRIV;
390 data->hctx->queued++;
394 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
395 blk_mq_req_flags_t flags)
397 struct blk_mq_alloc_data alloc_data = { .flags = flags };
401 ret = blk_queue_enter(q, flags);
405 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
409 return ERR_PTR(-EWOULDBLOCK);
411 blk_mq_put_ctx(alloc_data.ctx);
414 rq->__sector = (sector_t) -1;
415 rq->bio = rq->biotail = NULL;
418 EXPORT_SYMBOL(blk_mq_alloc_request);
420 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
421 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
423 struct blk_mq_alloc_data alloc_data = { .flags = flags };
429 * If the tag allocator sleeps we could get an allocation for a
430 * different hardware context. No need to complicate the low level
431 * allocator for this for the rare use case of a command tied to
434 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
435 return ERR_PTR(-EINVAL);
437 if (hctx_idx >= q->nr_hw_queues)
438 return ERR_PTR(-EIO);
440 ret = blk_queue_enter(q, flags);
445 * Check if the hardware context is actually mapped to anything.
446 * If not tell the caller that it should skip this queue.
448 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
449 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
451 return ERR_PTR(-EXDEV);
453 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
454 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
456 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
460 return ERR_PTR(-EWOULDBLOCK);
464 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
466 void blk_mq_free_request(struct request *rq)
468 struct request_queue *q = rq->q;
469 struct elevator_queue *e = q->elevator;
470 struct blk_mq_ctx *ctx = rq->mq_ctx;
471 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
472 const int sched_tag = rq->internal_tag;
474 if (rq->rq_flags & RQF_ELVPRIV) {
475 if (e && e->type->ops.mq.finish_request)
476 e->type->ops.mq.finish_request(rq);
478 put_io_context(rq->elv.icq->ioc);
483 ctx->rq_completed[rq_is_sync(rq)]++;
484 if (rq->rq_flags & RQF_MQ_INFLIGHT)
485 atomic_dec(&hctx->nr_active);
487 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
488 laptop_io_completion(q->backing_dev_info);
490 wbt_done(q->rq_wb, rq);
493 blk_put_rl(blk_rq_rl(rq));
495 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
497 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
499 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
500 blk_mq_sched_restart(hctx);
503 EXPORT_SYMBOL_GPL(blk_mq_free_request);
505 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
507 u64 now = ktime_get_ns();
509 if (rq->rq_flags & RQF_STATS) {
510 blk_mq_poll_stats_start(rq->q);
511 blk_stat_add(rq, now);
514 blk_account_io_done(rq, now);
517 wbt_done(rq->q->rq_wb, rq);
518 rq->end_io(rq, error);
520 if (unlikely(blk_bidi_rq(rq)))
521 blk_mq_free_request(rq->next_rq);
522 blk_mq_free_request(rq);
525 EXPORT_SYMBOL(__blk_mq_end_request);
527 void blk_mq_end_request(struct request *rq, blk_status_t error)
529 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
531 __blk_mq_end_request(rq, error);
533 EXPORT_SYMBOL(blk_mq_end_request);
535 static void __blk_mq_complete_request_remote(void *data)
537 struct request *rq = data;
539 rq->q->softirq_done_fn(rq);
542 static void __blk_mq_complete_request(struct request *rq)
544 struct blk_mq_ctx *ctx = rq->mq_ctx;
548 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT);
549 blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE);
551 if (rq->internal_tag != -1)
552 blk_mq_sched_completed_request(rq);
554 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
555 rq->q->softirq_done_fn(rq);
560 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
561 shared = cpus_share_cache(cpu, ctx->cpu);
563 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
564 rq->csd.func = __blk_mq_complete_request_remote;
567 smp_call_function_single_async(ctx->cpu, &rq->csd);
569 rq->q->softirq_done_fn(rq);
574 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
575 __releases(hctx->srcu)
577 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
580 srcu_read_unlock(hctx->srcu, srcu_idx);
583 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
584 __acquires(hctx->srcu)
586 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
587 /* shut up gcc false positive */
591 *srcu_idx = srcu_read_lock(hctx->srcu);
594 static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate)
599 * blk_mq_rq_aborted_gstate() is used from the completion path and
600 * can thus be called from irq context. u64_stats_fetch in the
601 * middle of update on the same CPU leads to lockup. Disable irq
604 local_irq_save(flags);
605 u64_stats_update_begin(&rq->aborted_gstate_sync);
606 rq->aborted_gstate = gstate;
607 u64_stats_update_end(&rq->aborted_gstate_sync);
608 local_irq_restore(flags);
611 static u64 blk_mq_rq_aborted_gstate(struct request *rq)
617 start = u64_stats_fetch_begin(&rq->aborted_gstate_sync);
618 aborted_gstate = rq->aborted_gstate;
619 } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start));
621 return aborted_gstate;
625 * blk_mq_complete_request - end I/O on a request
626 * @rq: the request being processed
629 * Ends all I/O on a request. It does not handle partial completions.
630 * The actual completion happens out-of-order, through a IPI handler.
632 void blk_mq_complete_request(struct request *rq)
634 struct request_queue *q = rq->q;
635 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
638 if (unlikely(blk_should_fake_timeout(q)))
642 * If @rq->aborted_gstate equals the current instance, timeout is
643 * claiming @rq and we lost. This is synchronized through
644 * hctx_lock(). See blk_mq_timeout_work() for details.
646 * Completion path never blocks and we can directly use RCU here
647 * instead of hctx_lock() which can be either RCU or SRCU.
648 * However, that would complicate paths which want to synchronize
649 * against us. Let stay in sync with the issue path so that
650 * hctx_lock() covers both issue and completion paths.
652 hctx_lock(hctx, &srcu_idx);
653 if (blk_mq_rq_aborted_gstate(rq) != rq->gstate)
654 __blk_mq_complete_request(rq);
655 hctx_unlock(hctx, srcu_idx);
657 EXPORT_SYMBOL(blk_mq_complete_request);
659 int blk_mq_request_started(struct request *rq)
661 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
663 EXPORT_SYMBOL_GPL(blk_mq_request_started);
665 void blk_mq_start_request(struct request *rq)
667 struct request_queue *q = rq->q;
669 blk_mq_sched_started_request(rq);
671 trace_block_rq_issue(q, rq);
673 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
674 rq->io_start_time_ns = ktime_get_ns();
675 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
676 rq->throtl_size = blk_rq_sectors(rq);
678 rq->rq_flags |= RQF_STATS;
679 wbt_issue(q->rq_wb, rq);
682 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
685 * Mark @rq in-flight which also advances the generation number,
686 * and register for timeout. Protect with a seqcount to allow the
687 * timeout path to read both @rq->gstate and @rq->deadline
690 * This is the only place where a request is marked in-flight. If
691 * the timeout path reads an in-flight @rq->gstate, the
692 * @rq->deadline it reads together under @rq->gstate_seq is
693 * guaranteed to be the matching one.
696 write_seqcount_begin(&rq->gstate_seq);
698 blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT);
701 write_seqcount_end(&rq->gstate_seq);
704 if (q->dma_drain_size && blk_rq_bytes(rq)) {
706 * Make sure space for the drain appears. We know we can do
707 * this because max_hw_segments has been adjusted to be one
708 * fewer than the device can handle.
710 rq->nr_phys_segments++;
713 EXPORT_SYMBOL(blk_mq_start_request);
716 * When we reach here because queue is busy, it's safe to change the state
717 * to IDLE without checking @rq->aborted_gstate because we should still be
718 * holding the RCU read lock and thus protected against timeout.
720 static void __blk_mq_requeue_request(struct request *rq)
722 struct request_queue *q = rq->q;
724 blk_mq_put_driver_tag(rq);
726 trace_block_rq_requeue(q, rq);
727 wbt_requeue(q->rq_wb, rq);
729 if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) {
730 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
731 if (q->dma_drain_size && blk_rq_bytes(rq))
732 rq->nr_phys_segments--;
736 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
738 __blk_mq_requeue_request(rq);
740 /* this request will be re-inserted to io scheduler queue */
741 blk_mq_sched_requeue_request(rq);
743 BUG_ON(blk_queued_rq(rq));
744 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
746 EXPORT_SYMBOL(blk_mq_requeue_request);
748 static void blk_mq_requeue_work(struct work_struct *work)
750 struct request_queue *q =
751 container_of(work, struct request_queue, requeue_work.work);
753 struct request *rq, *next;
755 spin_lock_irq(&q->requeue_lock);
756 list_splice_init(&q->requeue_list, &rq_list);
757 spin_unlock_irq(&q->requeue_lock);
759 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
760 if (!(rq->rq_flags & RQF_SOFTBARRIER))
763 rq->rq_flags &= ~RQF_SOFTBARRIER;
764 list_del_init(&rq->queuelist);
765 blk_mq_sched_insert_request(rq, true, false, false);
768 while (!list_empty(&rq_list)) {
769 rq = list_entry(rq_list.next, struct request, queuelist);
770 list_del_init(&rq->queuelist);
771 blk_mq_sched_insert_request(rq, false, false, false);
774 blk_mq_run_hw_queues(q, false);
777 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
778 bool kick_requeue_list)
780 struct request_queue *q = rq->q;
784 * We abuse this flag that is otherwise used by the I/O scheduler to
785 * request head insertion from the workqueue.
787 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
789 spin_lock_irqsave(&q->requeue_lock, flags);
791 rq->rq_flags |= RQF_SOFTBARRIER;
792 list_add(&rq->queuelist, &q->requeue_list);
794 list_add_tail(&rq->queuelist, &q->requeue_list);
796 spin_unlock_irqrestore(&q->requeue_lock, flags);
798 if (kick_requeue_list)
799 blk_mq_kick_requeue_list(q);
801 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
803 void blk_mq_kick_requeue_list(struct request_queue *q)
805 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
807 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
809 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
812 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
813 msecs_to_jiffies(msecs));
815 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
817 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
819 if (tag < tags->nr_tags) {
820 prefetch(tags->rqs[tag]);
821 return tags->rqs[tag];
826 EXPORT_SYMBOL(blk_mq_tag_to_rq);
828 struct blk_mq_timeout_data {
830 unsigned int next_set;
831 unsigned int nr_expired;
834 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
836 const struct blk_mq_ops *ops = req->q->mq_ops;
837 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
839 req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED;
842 ret = ops->timeout(req, reserved);
846 __blk_mq_complete_request(req);
848 case BLK_EH_RESET_TIMER:
850 * As nothing prevents from completion happening while
851 * ->aborted_gstate is set, this may lead to ignored
852 * completions and further spurious timeouts.
854 blk_mq_rq_update_aborted_gstate(req, 0);
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 (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED)
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 = blk_rq_deadline(rq);
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 (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) &&
910 READ_ONCE(rq->gstate) == rq->aborted_gstate)
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->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);
977 * Request timeouts are handled as a forward rolling timer. If
978 * we end up here it means that no requests are pending and
979 * also that no request has been pending for a while. Mark
982 queue_for_each_hw_ctx(q, hctx, i) {
983 /* the hctx may be unmapped, so check it here */
984 if (blk_mq_hw_queue_mapped(hctx))
985 blk_mq_tag_idle(hctx);
991 struct flush_busy_ctx_data {
992 struct blk_mq_hw_ctx *hctx;
993 struct list_head *list;
996 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
998 struct flush_busy_ctx_data *flush_data = data;
999 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1000 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1002 spin_lock(&ctx->lock);
1003 list_splice_tail_init(&ctx->rq_list, flush_data->list);
1004 sbitmap_clear_bit(sb, bitnr);
1005 spin_unlock(&ctx->lock);
1010 * Process software queues that have been marked busy, splicing them
1011 * to the for-dispatch
1013 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1015 struct flush_busy_ctx_data data = {
1020 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1022 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1024 struct dispatch_rq_data {
1025 struct blk_mq_hw_ctx *hctx;
1029 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1032 struct dispatch_rq_data *dispatch_data = data;
1033 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1034 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1036 spin_lock(&ctx->lock);
1037 if (unlikely(!list_empty(&ctx->rq_list))) {
1038 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
1039 list_del_init(&dispatch_data->rq->queuelist);
1040 if (list_empty(&ctx->rq_list))
1041 sbitmap_clear_bit(sb, bitnr);
1043 spin_unlock(&ctx->lock);
1045 return !dispatch_data->rq;
1048 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1049 struct blk_mq_ctx *start)
1051 unsigned off = start ? start->index_hw : 0;
1052 struct dispatch_rq_data data = {
1057 __sbitmap_for_each_set(&hctx->ctx_map, off,
1058 dispatch_rq_from_ctx, &data);
1063 static inline unsigned int queued_to_index(unsigned int queued)
1068 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1071 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
1074 struct blk_mq_alloc_data data = {
1076 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
1077 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
1080 might_sleep_if(wait);
1085 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1086 data.flags |= BLK_MQ_REQ_RESERVED;
1088 rq->tag = blk_mq_get_tag(&data);
1090 if (blk_mq_tag_busy(data.hctx)) {
1091 rq->rq_flags |= RQF_MQ_INFLIGHT;
1092 atomic_inc(&data.hctx->nr_active);
1094 data.hctx->tags->rqs[rq->tag] = rq;
1100 return rq->tag != -1;
1103 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1104 int flags, void *key)
1106 struct blk_mq_hw_ctx *hctx;
1108 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1110 list_del_init(&wait->entry);
1111 blk_mq_run_hw_queue(hctx, true);
1116 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1117 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1118 * restart. For both cases, take care to check the condition again after
1119 * marking us as waiting.
1121 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1124 struct blk_mq_hw_ctx *this_hctx = *hctx;
1125 struct sbq_wait_state *ws;
1126 wait_queue_entry_t *wait;
1129 if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1130 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1131 set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1134 * It's possible that a tag was freed in the window between the
1135 * allocation failure and adding the hardware queue to the wait
1138 * Don't clear RESTART here, someone else could have set it.
1139 * At most this will cost an extra queue run.
1141 return blk_mq_get_driver_tag(rq, hctx, false);
1144 wait = &this_hctx->dispatch_wait;
1145 if (!list_empty_careful(&wait->entry))
1148 spin_lock(&this_hctx->lock);
1149 if (!list_empty(&wait->entry)) {
1150 spin_unlock(&this_hctx->lock);
1154 ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1155 add_wait_queue(&ws->wait, wait);
1158 * It's possible that a tag was freed in the window between the
1159 * allocation failure and adding the hardware queue to the wait
1162 ret = blk_mq_get_driver_tag(rq, hctx, false);
1164 spin_unlock(&this_hctx->lock);
1169 * We got a tag, remove ourselves from the wait queue to ensure
1170 * someone else gets the wakeup.
1172 spin_lock_irq(&ws->wait.lock);
1173 list_del_init(&wait->entry);
1174 spin_unlock_irq(&ws->wait.lock);
1175 spin_unlock(&this_hctx->lock);
1180 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1182 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1185 struct blk_mq_hw_ctx *hctx;
1186 struct request *rq, *nxt;
1187 bool no_tag = false;
1189 blk_status_t ret = BLK_STS_OK;
1191 if (list_empty(list))
1194 WARN_ON(!list_is_singular(list) && got_budget);
1197 * Now process all the entries, sending them to the driver.
1199 errors = queued = 0;
1201 struct blk_mq_queue_data bd;
1203 rq = list_first_entry(list, struct request, queuelist);
1205 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1206 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1209 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1211 * The initial allocation attempt failed, so we need to
1212 * rerun the hardware queue when a tag is freed. The
1213 * waitqueue takes care of that. If the queue is run
1214 * before we add this entry back on the dispatch list,
1215 * we'll re-run it below.
1217 if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1218 blk_mq_put_dispatch_budget(hctx);
1220 * For non-shared tags, the RESTART check
1223 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1229 list_del_init(&rq->queuelist);
1234 * Flag last if we have no more requests, or if we have more
1235 * but can't assign a driver tag to it.
1237 if (list_empty(list))
1240 nxt = list_first_entry(list, struct request, queuelist);
1241 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1244 ret = q->mq_ops->queue_rq(hctx, &bd);
1245 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1247 * If an I/O scheduler has been configured and we got a
1248 * driver tag for the next request already, free it
1251 if (!list_empty(list)) {
1252 nxt = list_first_entry(list, struct request, queuelist);
1253 blk_mq_put_driver_tag(nxt);
1255 list_add(&rq->queuelist, list);
1256 __blk_mq_requeue_request(rq);
1260 if (unlikely(ret != BLK_STS_OK)) {
1262 blk_mq_end_request(rq, BLK_STS_IOERR);
1267 } while (!list_empty(list));
1269 hctx->dispatched[queued_to_index(queued)]++;
1272 * Any items that need requeuing? Stuff them into hctx->dispatch,
1273 * that is where we will continue on next queue run.
1275 if (!list_empty(list)) {
1278 spin_lock(&hctx->lock);
1279 list_splice_init(list, &hctx->dispatch);
1280 spin_unlock(&hctx->lock);
1283 * If SCHED_RESTART was set by the caller of this function and
1284 * it is no longer set that means that it was cleared by another
1285 * thread and hence that a queue rerun is needed.
1287 * If 'no_tag' is set, that means that we failed getting
1288 * a driver tag with an I/O scheduler attached. If our dispatch
1289 * waitqueue is no longer active, ensure that we run the queue
1290 * AFTER adding our entries back to the list.
1292 * If no I/O scheduler has been configured it is possible that
1293 * the hardware queue got stopped and restarted before requests
1294 * were pushed back onto the dispatch list. Rerun the queue to
1295 * avoid starvation. Notes:
1296 * - blk_mq_run_hw_queue() checks whether or not a queue has
1297 * been stopped before rerunning a queue.
1298 * - Some but not all block drivers stop a queue before
1299 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1302 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1303 * bit is set, run queue after a delay to avoid IO stalls
1304 * that could otherwise occur if the queue is idle.
1306 needs_restart = blk_mq_sched_needs_restart(hctx);
1307 if (!needs_restart ||
1308 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1309 blk_mq_run_hw_queue(hctx, true);
1310 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1311 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1314 return (queued + errors) != 0;
1317 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1322 * We should be running this queue from one of the CPUs that
1325 * There are at least two related races now between setting
1326 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1327 * __blk_mq_run_hw_queue():
1329 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1330 * but later it becomes online, then this warning is harmless
1333 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1334 * but later it becomes offline, then the warning can't be
1335 * triggered, and we depend on blk-mq timeout handler to
1336 * handle dispatched requests to this hctx
1338 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1339 cpu_online(hctx->next_cpu)) {
1340 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1341 raw_smp_processor_id(),
1342 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1347 * We can't run the queue inline with ints disabled. Ensure that
1348 * we catch bad users of this early.
1350 WARN_ON_ONCE(in_interrupt());
1352 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1354 hctx_lock(hctx, &srcu_idx);
1355 blk_mq_sched_dispatch_requests(hctx);
1356 hctx_unlock(hctx, srcu_idx);
1359 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1361 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1363 if (cpu >= nr_cpu_ids)
1364 cpu = cpumask_first(hctx->cpumask);
1369 * It'd be great if the workqueue API had a way to pass
1370 * in a mask and had some smarts for more clever placement.
1371 * For now we just round-robin here, switching for every
1372 * BLK_MQ_CPU_WORK_BATCH queued items.
1374 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1377 int next_cpu = hctx->next_cpu;
1379 if (hctx->queue->nr_hw_queues == 1)
1380 return WORK_CPU_UNBOUND;
1382 if (--hctx->next_cpu_batch <= 0) {
1384 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1386 if (next_cpu >= nr_cpu_ids)
1387 next_cpu = blk_mq_first_mapped_cpu(hctx);
1388 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1392 * Do unbound schedule if we can't find a online CPU for this hctx,
1393 * and it should only happen in the path of handling CPU DEAD.
1395 if (!cpu_online(next_cpu)) {
1402 * Make sure to re-select CPU next time once after CPUs
1403 * in hctx->cpumask become online again.
1405 hctx->next_cpu = next_cpu;
1406 hctx->next_cpu_batch = 1;
1407 return WORK_CPU_UNBOUND;
1410 hctx->next_cpu = next_cpu;
1414 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1415 unsigned long msecs)
1417 if (unlikely(blk_mq_hctx_stopped(hctx)))
1420 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1421 int cpu = get_cpu();
1422 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1423 __blk_mq_run_hw_queue(hctx);
1431 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1432 msecs_to_jiffies(msecs));
1435 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1437 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1439 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1441 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1447 * When queue is quiesced, we may be switching io scheduler, or
1448 * updating nr_hw_queues, or other things, and we can't run queue
1449 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1451 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1454 hctx_lock(hctx, &srcu_idx);
1455 need_run = !blk_queue_quiesced(hctx->queue) &&
1456 blk_mq_hctx_has_pending(hctx);
1457 hctx_unlock(hctx, srcu_idx);
1460 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1466 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1468 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1470 struct blk_mq_hw_ctx *hctx;
1473 queue_for_each_hw_ctx(q, hctx, i) {
1474 if (blk_mq_hctx_stopped(hctx))
1477 blk_mq_run_hw_queue(hctx, async);
1480 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1483 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1484 * @q: request queue.
1486 * The caller is responsible for serializing this function against
1487 * blk_mq_{start,stop}_hw_queue().
1489 bool blk_mq_queue_stopped(struct request_queue *q)
1491 struct blk_mq_hw_ctx *hctx;
1494 queue_for_each_hw_ctx(q, hctx, i)
1495 if (blk_mq_hctx_stopped(hctx))
1500 EXPORT_SYMBOL(blk_mq_queue_stopped);
1503 * This function is often used for pausing .queue_rq() by driver when
1504 * there isn't enough resource or some conditions aren't satisfied, and
1505 * BLK_STS_RESOURCE is usually returned.
1507 * We do not guarantee that dispatch can be drained or blocked
1508 * after blk_mq_stop_hw_queue() returns. Please use
1509 * blk_mq_quiesce_queue() for that requirement.
1511 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1513 cancel_delayed_work(&hctx->run_work);
1515 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1517 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1520 * This function is often used for pausing .queue_rq() by driver when
1521 * there isn't enough resource or some conditions aren't satisfied, and
1522 * BLK_STS_RESOURCE is usually returned.
1524 * We do not guarantee that dispatch can be drained or blocked
1525 * after blk_mq_stop_hw_queues() returns. Please use
1526 * blk_mq_quiesce_queue() for that requirement.
1528 void blk_mq_stop_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_stop_hw_queue(hctx);
1536 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1538 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1540 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1542 blk_mq_run_hw_queue(hctx, false);
1544 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1546 void blk_mq_start_hw_queues(struct request_queue *q)
1548 struct blk_mq_hw_ctx *hctx;
1551 queue_for_each_hw_ctx(q, hctx, i)
1552 blk_mq_start_hw_queue(hctx);
1554 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1556 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1558 if (!blk_mq_hctx_stopped(hctx))
1561 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1562 blk_mq_run_hw_queue(hctx, async);
1564 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1566 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1568 struct blk_mq_hw_ctx *hctx;
1571 queue_for_each_hw_ctx(q, hctx, i)
1572 blk_mq_start_stopped_hw_queue(hctx, async);
1574 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1576 static void blk_mq_run_work_fn(struct work_struct *work)
1578 struct blk_mq_hw_ctx *hctx;
1580 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1583 * If we are stopped, don't run the queue.
1585 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1586 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1588 __blk_mq_run_hw_queue(hctx);
1591 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1595 struct blk_mq_ctx *ctx = rq->mq_ctx;
1597 lockdep_assert_held(&ctx->lock);
1599 trace_block_rq_insert(hctx->queue, rq);
1602 list_add(&rq->queuelist, &ctx->rq_list);
1604 list_add_tail(&rq->queuelist, &ctx->rq_list);
1607 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1610 struct blk_mq_ctx *ctx = rq->mq_ctx;
1612 lockdep_assert_held(&ctx->lock);
1614 __blk_mq_insert_req_list(hctx, rq, at_head);
1615 blk_mq_hctx_mark_pending(hctx, ctx);
1619 * Should only be used carefully, when the caller knows we want to
1620 * bypass a potential IO scheduler on the target device.
1622 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1624 struct blk_mq_ctx *ctx = rq->mq_ctx;
1625 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1627 spin_lock(&hctx->lock);
1628 list_add_tail(&rq->queuelist, &hctx->dispatch);
1629 spin_unlock(&hctx->lock);
1632 blk_mq_run_hw_queue(hctx, false);
1635 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1636 struct list_head *list)
1640 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1643 spin_lock(&ctx->lock);
1644 while (!list_empty(list)) {
1647 rq = list_first_entry(list, struct request, queuelist);
1648 BUG_ON(rq->mq_ctx != ctx);
1649 list_del_init(&rq->queuelist);
1650 __blk_mq_insert_req_list(hctx, rq, false);
1652 blk_mq_hctx_mark_pending(hctx, ctx);
1653 spin_unlock(&ctx->lock);
1656 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1658 struct request *rqa = container_of(a, struct request, queuelist);
1659 struct request *rqb = container_of(b, struct request, queuelist);
1661 return !(rqa->mq_ctx < rqb->mq_ctx ||
1662 (rqa->mq_ctx == rqb->mq_ctx &&
1663 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1666 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1668 struct blk_mq_ctx *this_ctx;
1669 struct request_queue *this_q;
1672 LIST_HEAD(ctx_list);
1675 list_splice_init(&plug->mq_list, &list);
1677 list_sort(NULL, &list, plug_ctx_cmp);
1683 while (!list_empty(&list)) {
1684 rq = list_entry_rq(list.next);
1685 list_del_init(&rq->queuelist);
1687 if (rq->mq_ctx != this_ctx) {
1689 trace_block_unplug(this_q, depth, from_schedule);
1690 blk_mq_sched_insert_requests(this_q, this_ctx,
1695 this_ctx = rq->mq_ctx;
1701 list_add_tail(&rq->queuelist, &ctx_list);
1705 * If 'this_ctx' is set, we know we have entries to complete
1706 * on 'ctx_list'. Do those.
1709 trace_block_unplug(this_q, depth, from_schedule);
1710 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1715 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1717 blk_init_request_from_bio(rq, bio);
1719 blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1721 blk_account_io_start(rq, true);
1724 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1725 struct blk_mq_ctx *ctx,
1728 spin_lock(&ctx->lock);
1729 __blk_mq_insert_request(hctx, rq, false);
1730 spin_unlock(&ctx->lock);
1733 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1736 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1738 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1741 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1745 struct request_queue *q = rq->q;
1746 struct blk_mq_queue_data bd = {
1750 blk_qc_t new_cookie;
1753 new_cookie = request_to_qc_t(hctx, rq);
1756 * For OK queue, we are done. For error, caller may kill it.
1757 * Any other error (busy), just add it to our list as we
1758 * previously would have done.
1760 ret = q->mq_ops->queue_rq(hctx, &bd);
1763 *cookie = new_cookie;
1765 case BLK_STS_RESOURCE:
1766 case BLK_STS_DEV_RESOURCE:
1767 __blk_mq_requeue_request(rq);
1770 *cookie = BLK_QC_T_NONE;
1777 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1782 struct request_queue *q = rq->q;
1783 bool run_queue = true;
1786 * RCU or SRCU read lock is needed before checking quiesced flag.
1788 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1789 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1790 * and avoid driver to try to dispatch again.
1792 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1794 bypass_insert = false;
1798 if (q->elevator && !bypass_insert)
1801 if (!blk_mq_get_dispatch_budget(hctx))
1804 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1805 blk_mq_put_dispatch_budget(hctx);
1809 return __blk_mq_issue_directly(hctx, rq, cookie);
1812 return BLK_STS_RESOURCE;
1814 blk_mq_sched_insert_request(rq, false, run_queue, false);
1818 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1819 struct request *rq, blk_qc_t *cookie)
1824 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1826 hctx_lock(hctx, &srcu_idx);
1828 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1829 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1830 blk_mq_sched_insert_request(rq, false, true, false);
1831 else if (ret != BLK_STS_OK)
1832 blk_mq_end_request(rq, ret);
1834 hctx_unlock(hctx, srcu_idx);
1837 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1841 blk_qc_t unused_cookie;
1842 struct blk_mq_ctx *ctx = rq->mq_ctx;
1843 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1845 hctx_lock(hctx, &srcu_idx);
1846 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1847 hctx_unlock(hctx, srcu_idx);
1852 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1854 const int is_sync = op_is_sync(bio->bi_opf);
1855 const int is_flush_fua = op_is_flush(bio->bi_opf);
1856 struct blk_mq_alloc_data data = { .flags = 0 };
1858 unsigned int request_count = 0;
1859 struct blk_plug *plug;
1860 struct request *same_queue_rq = NULL;
1862 unsigned int wb_acct;
1864 blk_queue_bounce(q, &bio);
1866 blk_queue_split(q, &bio);
1868 if (!bio_integrity_prep(bio))
1869 return BLK_QC_T_NONE;
1871 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1872 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1873 return BLK_QC_T_NONE;
1875 if (blk_mq_sched_bio_merge(q, bio))
1876 return BLK_QC_T_NONE;
1878 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1880 trace_block_getrq(q, bio, bio->bi_opf);
1882 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1883 if (unlikely(!rq)) {
1884 __wbt_done(q->rq_wb, wb_acct);
1885 if (bio->bi_opf & REQ_NOWAIT)
1886 bio_wouldblock_error(bio);
1887 return BLK_QC_T_NONE;
1890 wbt_track(rq, wb_acct);
1892 cookie = request_to_qc_t(data.hctx, rq);
1894 plug = current->plug;
1895 if (unlikely(is_flush_fua)) {
1896 blk_mq_put_ctx(data.ctx);
1897 blk_mq_bio_to_request(rq, bio);
1899 /* bypass scheduler for flush rq */
1900 blk_insert_flush(rq);
1901 blk_mq_run_hw_queue(data.hctx, true);
1902 } else if (plug && q->nr_hw_queues == 1) {
1903 struct request *last = NULL;
1905 blk_mq_put_ctx(data.ctx);
1906 blk_mq_bio_to_request(rq, bio);
1909 * @request_count may become stale because of schedule
1910 * out, so check the list again.
1912 if (list_empty(&plug->mq_list))
1914 else if (blk_queue_nomerges(q))
1915 request_count = blk_plug_queued_count(q);
1918 trace_block_plug(q);
1920 last = list_entry_rq(plug->mq_list.prev);
1922 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1923 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1924 blk_flush_plug_list(plug, false);
1925 trace_block_plug(q);
1928 list_add_tail(&rq->queuelist, &plug->mq_list);
1929 } else if (plug && !blk_queue_nomerges(q)) {
1930 blk_mq_bio_to_request(rq, bio);
1933 * We do limited plugging. If the bio can be merged, do that.
1934 * Otherwise the existing request in the plug list will be
1935 * issued. So the plug list will have one request at most
1936 * The plug list might get flushed before this. If that happens,
1937 * the plug list is empty, and same_queue_rq is invalid.
1939 if (list_empty(&plug->mq_list))
1940 same_queue_rq = NULL;
1942 list_del_init(&same_queue_rq->queuelist);
1943 list_add_tail(&rq->queuelist, &plug->mq_list);
1945 blk_mq_put_ctx(data.ctx);
1947 if (same_queue_rq) {
1948 data.hctx = blk_mq_map_queue(q,
1949 same_queue_rq->mq_ctx->cpu);
1950 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1953 } else if (q->nr_hw_queues > 1 && is_sync) {
1954 blk_mq_put_ctx(data.ctx);
1955 blk_mq_bio_to_request(rq, bio);
1956 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1957 } else if (q->elevator) {
1958 blk_mq_put_ctx(data.ctx);
1959 blk_mq_bio_to_request(rq, bio);
1960 blk_mq_sched_insert_request(rq, false, true, true);
1962 blk_mq_put_ctx(data.ctx);
1963 blk_mq_bio_to_request(rq, bio);
1964 blk_mq_queue_io(data.hctx, data.ctx, rq);
1965 blk_mq_run_hw_queue(data.hctx, true);
1971 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1972 unsigned int hctx_idx)
1976 if (tags->rqs && set->ops->exit_request) {
1979 for (i = 0; i < tags->nr_tags; i++) {
1980 struct request *rq = tags->static_rqs[i];
1984 set->ops->exit_request(set, rq, hctx_idx);
1985 tags->static_rqs[i] = NULL;
1989 while (!list_empty(&tags->page_list)) {
1990 page = list_first_entry(&tags->page_list, struct page, lru);
1991 list_del_init(&page->lru);
1993 * Remove kmemleak object previously allocated in
1994 * blk_mq_init_rq_map().
1996 kmemleak_free(page_address(page));
1997 __free_pages(page, page->private);
2001 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2005 kfree(tags->static_rqs);
2006 tags->static_rqs = NULL;
2008 blk_mq_free_tags(tags);
2011 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2012 unsigned int hctx_idx,
2013 unsigned int nr_tags,
2014 unsigned int reserved_tags)
2016 struct blk_mq_tags *tags;
2019 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2020 if (node == NUMA_NO_NODE)
2021 node = set->numa_node;
2023 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2024 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2028 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2029 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2032 blk_mq_free_tags(tags);
2036 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2037 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2039 if (!tags->static_rqs) {
2041 blk_mq_free_tags(tags);
2048 static size_t order_to_size(unsigned int order)
2050 return (size_t)PAGE_SIZE << order;
2053 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2054 unsigned int hctx_idx, int node)
2058 if (set->ops->init_request) {
2059 ret = set->ops->init_request(set, rq, hctx_idx, node);
2064 seqcount_init(&rq->gstate_seq);
2065 u64_stats_init(&rq->aborted_gstate_sync);
2067 * start gstate with gen 1 instead of 0, otherwise it will be equal
2068 * to aborted_gstate, and be identified timed out by
2069 * blk_mq_terminate_expired.
2071 WRITE_ONCE(rq->gstate, MQ_RQ_GEN_INC);
2076 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2077 unsigned int hctx_idx, unsigned int depth)
2079 unsigned int i, j, entries_per_page, max_order = 4;
2080 size_t rq_size, left;
2083 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2084 if (node == NUMA_NO_NODE)
2085 node = set->numa_node;
2087 INIT_LIST_HEAD(&tags->page_list);
2090 * rq_size is the size of the request plus driver payload, rounded
2091 * to the cacheline size
2093 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2095 left = rq_size * depth;
2097 for (i = 0; i < depth; ) {
2098 int this_order = max_order;
2103 while (this_order && left < order_to_size(this_order - 1))
2107 page = alloc_pages_node(node,
2108 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2114 if (order_to_size(this_order) < rq_size)
2121 page->private = this_order;
2122 list_add_tail(&page->lru, &tags->page_list);
2124 p = page_address(page);
2126 * Allow kmemleak to scan these pages as they contain pointers
2127 * to additional allocations like via ops->init_request().
2129 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2130 entries_per_page = order_to_size(this_order) / rq_size;
2131 to_do = min(entries_per_page, depth - i);
2132 left -= to_do * rq_size;
2133 for (j = 0; j < to_do; j++) {
2134 struct request *rq = p;
2136 tags->static_rqs[i] = rq;
2137 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2138 tags->static_rqs[i] = NULL;
2149 blk_mq_free_rqs(set, tags, hctx_idx);
2154 * 'cpu' is going away. splice any existing rq_list entries from this
2155 * software queue to the hw queue dispatch list, and ensure that it
2158 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2160 struct blk_mq_hw_ctx *hctx;
2161 struct blk_mq_ctx *ctx;
2164 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2165 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2167 spin_lock(&ctx->lock);
2168 if (!list_empty(&ctx->rq_list)) {
2169 list_splice_init(&ctx->rq_list, &tmp);
2170 blk_mq_hctx_clear_pending(hctx, ctx);
2172 spin_unlock(&ctx->lock);
2174 if (list_empty(&tmp))
2177 spin_lock(&hctx->lock);
2178 list_splice_tail_init(&tmp, &hctx->dispatch);
2179 spin_unlock(&hctx->lock);
2181 blk_mq_run_hw_queue(hctx, true);
2185 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2187 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2191 /* hctx->ctxs will be freed in queue's release handler */
2192 static void blk_mq_exit_hctx(struct request_queue *q,
2193 struct blk_mq_tag_set *set,
2194 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2196 blk_mq_debugfs_unregister_hctx(hctx);
2198 if (blk_mq_hw_queue_mapped(hctx))
2199 blk_mq_tag_idle(hctx);
2201 if (set->ops->exit_request)
2202 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2204 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2206 if (set->ops->exit_hctx)
2207 set->ops->exit_hctx(hctx, hctx_idx);
2209 if (hctx->flags & BLK_MQ_F_BLOCKING)
2210 cleanup_srcu_struct(hctx->srcu);
2212 blk_mq_remove_cpuhp(hctx);
2213 blk_free_flush_queue(hctx->fq);
2214 sbitmap_free(&hctx->ctx_map);
2217 static void blk_mq_exit_hw_queues(struct request_queue *q,
2218 struct blk_mq_tag_set *set, int nr_queue)
2220 struct blk_mq_hw_ctx *hctx;
2223 queue_for_each_hw_ctx(q, hctx, i) {
2226 blk_mq_exit_hctx(q, set, hctx, i);
2230 static int blk_mq_init_hctx(struct request_queue *q,
2231 struct blk_mq_tag_set *set,
2232 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2236 node = hctx->numa_node;
2237 if (node == NUMA_NO_NODE)
2238 node = hctx->numa_node = set->numa_node;
2240 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2241 spin_lock_init(&hctx->lock);
2242 INIT_LIST_HEAD(&hctx->dispatch);
2244 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2246 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2248 hctx->tags = set->tags[hctx_idx];
2251 * Allocate space for all possible cpus to avoid allocation at
2254 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2257 goto unregister_cpu_notifier;
2259 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2265 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2266 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2268 if (set->ops->init_hctx &&
2269 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2272 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2275 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2277 goto sched_exit_hctx;
2279 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2282 if (hctx->flags & BLK_MQ_F_BLOCKING)
2283 init_srcu_struct(hctx->srcu);
2285 blk_mq_debugfs_register_hctx(q, hctx);
2292 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2294 if (set->ops->exit_hctx)
2295 set->ops->exit_hctx(hctx, hctx_idx);
2297 sbitmap_free(&hctx->ctx_map);
2300 unregister_cpu_notifier:
2301 blk_mq_remove_cpuhp(hctx);
2305 static void blk_mq_init_cpu_queues(struct request_queue *q,
2306 unsigned int nr_hw_queues)
2310 for_each_possible_cpu(i) {
2311 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2312 struct blk_mq_hw_ctx *hctx;
2315 spin_lock_init(&__ctx->lock);
2316 INIT_LIST_HEAD(&__ctx->rq_list);
2320 * Set local node, IFF we have more than one hw queue. If
2321 * not, we remain on the home node of the device
2323 hctx = blk_mq_map_queue(q, i);
2324 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2325 hctx->numa_node = local_memory_node(cpu_to_node(i));
2329 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2333 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2334 set->queue_depth, set->reserved_tags);
2335 if (!set->tags[hctx_idx])
2338 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2343 blk_mq_free_rq_map(set->tags[hctx_idx]);
2344 set->tags[hctx_idx] = NULL;
2348 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2349 unsigned int hctx_idx)
2351 if (set->tags[hctx_idx]) {
2352 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2353 blk_mq_free_rq_map(set->tags[hctx_idx]);
2354 set->tags[hctx_idx] = NULL;
2358 static void blk_mq_map_swqueue(struct request_queue *q)
2360 unsigned int i, hctx_idx;
2361 struct blk_mq_hw_ctx *hctx;
2362 struct blk_mq_ctx *ctx;
2363 struct blk_mq_tag_set *set = q->tag_set;
2366 * Avoid others reading imcomplete hctx->cpumask through sysfs
2368 mutex_lock(&q->sysfs_lock);
2370 queue_for_each_hw_ctx(q, hctx, i) {
2371 cpumask_clear(hctx->cpumask);
2376 * Map software to hardware queues.
2378 * If the cpu isn't present, the cpu is mapped to first hctx.
2380 for_each_possible_cpu(i) {
2381 hctx_idx = q->mq_map[i];
2382 /* unmapped hw queue can be remapped after CPU topo changed */
2383 if (!set->tags[hctx_idx] &&
2384 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2386 * If tags initialization fail for some hctx,
2387 * that hctx won't be brought online. In this
2388 * case, remap the current ctx to hctx[0] which
2389 * is guaranteed to always have tags allocated
2394 ctx = per_cpu_ptr(q->queue_ctx, i);
2395 hctx = blk_mq_map_queue(q, i);
2397 cpumask_set_cpu(i, hctx->cpumask);
2398 ctx->index_hw = hctx->nr_ctx;
2399 hctx->ctxs[hctx->nr_ctx++] = ctx;
2402 mutex_unlock(&q->sysfs_lock);
2404 queue_for_each_hw_ctx(q, hctx, i) {
2406 * If no software queues are mapped to this hardware queue,
2407 * disable it and free the request entries.
2409 if (!hctx->nr_ctx) {
2410 /* Never unmap queue 0. We need it as a
2411 * fallback in case of a new remap fails
2414 if (i && set->tags[i])
2415 blk_mq_free_map_and_requests(set, i);
2421 hctx->tags = set->tags[i];
2422 WARN_ON(!hctx->tags);
2425 * Set the map size to the number of mapped software queues.
2426 * This is more accurate and more efficient than looping
2427 * over all possibly mapped software queues.
2429 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2432 * Initialize batch roundrobin counts
2434 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2435 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2440 * Caller needs to ensure that we're either frozen/quiesced, or that
2441 * the queue isn't live yet.
2443 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2445 struct blk_mq_hw_ctx *hctx;
2448 queue_for_each_hw_ctx(q, hctx, i) {
2450 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2451 atomic_inc(&q->shared_hctx_restart);
2452 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2454 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2455 atomic_dec(&q->shared_hctx_restart);
2456 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2461 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2464 struct request_queue *q;
2466 lockdep_assert_held(&set->tag_list_lock);
2468 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2469 blk_mq_freeze_queue(q);
2470 queue_set_hctx_shared(q, shared);
2471 blk_mq_unfreeze_queue(q);
2475 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2477 struct blk_mq_tag_set *set = q->tag_set;
2479 mutex_lock(&set->tag_list_lock);
2480 list_del_rcu(&q->tag_set_list);
2481 INIT_LIST_HEAD(&q->tag_set_list);
2482 if (list_is_singular(&set->tag_list)) {
2483 /* just transitioned to unshared */
2484 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2485 /* update existing queue */
2486 blk_mq_update_tag_set_depth(set, false);
2488 mutex_unlock(&set->tag_list_lock);
2493 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2494 struct request_queue *q)
2498 mutex_lock(&set->tag_list_lock);
2501 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2503 if (!list_empty(&set->tag_list) &&
2504 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2505 set->flags |= BLK_MQ_F_TAG_SHARED;
2506 /* update existing queue */
2507 blk_mq_update_tag_set_depth(set, true);
2509 if (set->flags & BLK_MQ_F_TAG_SHARED)
2510 queue_set_hctx_shared(q, true);
2511 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2513 mutex_unlock(&set->tag_list_lock);
2517 * It is the actual release handler for mq, but we do it from
2518 * request queue's release handler for avoiding use-after-free
2519 * and headache because q->mq_kobj shouldn't have been introduced,
2520 * but we can't group ctx/kctx kobj without it.
2522 void blk_mq_release(struct request_queue *q)
2524 struct blk_mq_hw_ctx *hctx;
2527 /* hctx kobj stays in hctx */
2528 queue_for_each_hw_ctx(q, hctx, i) {
2531 kobject_put(&hctx->kobj);
2536 kfree(q->queue_hw_ctx);
2539 * release .mq_kobj and sw queue's kobject now because
2540 * both share lifetime with request queue.
2542 blk_mq_sysfs_deinit(q);
2544 free_percpu(q->queue_ctx);
2547 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2549 struct request_queue *uninit_q, *q;
2551 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2553 return ERR_PTR(-ENOMEM);
2555 q = blk_mq_init_allocated_queue(set, uninit_q);
2557 blk_cleanup_queue(uninit_q);
2561 EXPORT_SYMBOL(blk_mq_init_queue);
2563 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2565 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2567 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2568 __alignof__(struct blk_mq_hw_ctx)) !=
2569 sizeof(struct blk_mq_hw_ctx));
2571 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2572 hw_ctx_size += sizeof(struct srcu_struct);
2577 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2578 struct request_queue *q)
2581 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2583 blk_mq_sysfs_unregister(q);
2585 /* protect against switching io scheduler */
2586 mutex_lock(&q->sysfs_lock);
2587 for (i = 0; i < set->nr_hw_queues; i++) {
2593 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2594 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2599 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2606 atomic_set(&hctxs[i]->nr_active, 0);
2607 hctxs[i]->numa_node = node;
2608 hctxs[i]->queue_num = i;
2610 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2611 free_cpumask_var(hctxs[i]->cpumask);
2616 blk_mq_hctx_kobj_init(hctxs[i]);
2618 for (j = i; j < q->nr_hw_queues; j++) {
2619 struct blk_mq_hw_ctx *hctx = hctxs[j];
2623 blk_mq_free_map_and_requests(set, j);
2624 blk_mq_exit_hctx(q, set, hctx, j);
2625 kobject_put(&hctx->kobj);
2630 q->nr_hw_queues = i;
2631 mutex_unlock(&q->sysfs_lock);
2632 blk_mq_sysfs_register(q);
2635 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2636 struct request_queue *q)
2638 /* mark the queue as mq asap */
2639 q->mq_ops = set->ops;
2641 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2642 blk_mq_poll_stats_bkt,
2643 BLK_MQ_POLL_STATS_BKTS, q);
2647 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2651 /* init q->mq_kobj and sw queues' kobjects */
2652 blk_mq_sysfs_init(q);
2654 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2655 GFP_KERNEL, set->numa_node);
2656 if (!q->queue_hw_ctx)
2659 q->mq_map = set->mq_map;
2661 blk_mq_realloc_hw_ctxs(set, q);
2662 if (!q->nr_hw_queues)
2665 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2666 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2668 q->nr_queues = nr_cpu_ids;
2670 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2672 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2673 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2675 q->sg_reserved_size = INT_MAX;
2677 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2678 INIT_LIST_HEAD(&q->requeue_list);
2679 spin_lock_init(&q->requeue_lock);
2681 blk_queue_make_request(q, blk_mq_make_request);
2682 if (q->mq_ops->poll)
2683 q->poll_fn = blk_mq_poll;
2686 * Do this after blk_queue_make_request() overrides it...
2688 q->nr_requests = set->queue_depth;
2691 * Default to classic polling
2695 if (set->ops->complete)
2696 blk_queue_softirq_done(q, set->ops->complete);
2698 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2699 blk_mq_add_queue_tag_set(set, q);
2700 blk_mq_map_swqueue(q);
2702 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2705 ret = blk_mq_sched_init(q);
2707 return ERR_PTR(ret);
2713 kfree(q->queue_hw_ctx);
2715 free_percpu(q->queue_ctx);
2718 return ERR_PTR(-ENOMEM);
2720 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2722 void blk_mq_free_queue(struct request_queue *q)
2724 struct blk_mq_tag_set *set = q->tag_set;
2726 blk_mq_del_queue_tag_set(q);
2727 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2730 /* Basically redo blk_mq_init_queue with queue frozen */
2731 static void blk_mq_queue_reinit(struct request_queue *q)
2733 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2735 blk_mq_debugfs_unregister_hctxs(q);
2736 blk_mq_sysfs_unregister(q);
2739 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2740 * we should change hctx numa_node according to the new topology (this
2741 * involves freeing and re-allocating memory, worth doing?)
2743 blk_mq_map_swqueue(q);
2745 blk_mq_sysfs_register(q);
2746 blk_mq_debugfs_register_hctxs(q);
2749 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2753 for (i = 0; i < set->nr_hw_queues; i++)
2754 if (!__blk_mq_alloc_rq_map(set, i))
2761 blk_mq_free_rq_map(set->tags[i]);
2767 * Allocate the request maps associated with this tag_set. Note that this
2768 * may reduce the depth asked for, if memory is tight. set->queue_depth
2769 * will be updated to reflect the allocated depth.
2771 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2776 depth = set->queue_depth;
2778 err = __blk_mq_alloc_rq_maps(set);
2782 set->queue_depth >>= 1;
2783 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2787 } while (set->queue_depth);
2789 if (!set->queue_depth || err) {
2790 pr_err("blk-mq: failed to allocate request map\n");
2794 if (depth != set->queue_depth)
2795 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2796 depth, set->queue_depth);
2801 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2803 if (set->ops->map_queues) {
2806 * transport .map_queues is usually done in the following
2809 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2810 * mask = get_cpu_mask(queue)
2811 * for_each_cpu(cpu, mask)
2812 * set->mq_map[cpu] = queue;
2815 * When we need to remap, the table has to be cleared for
2816 * killing stale mapping since one CPU may not be mapped
2819 for_each_possible_cpu(cpu)
2820 set->mq_map[cpu] = 0;
2822 return set->ops->map_queues(set);
2824 return blk_mq_map_queues(set);
2828 * Alloc a tag set to be associated with one or more request queues.
2829 * May fail with EINVAL for various error conditions. May adjust the
2830 * requested depth down, if if it too large. In that case, the set
2831 * value will be stored in set->queue_depth.
2833 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2837 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2839 if (!set->nr_hw_queues)
2841 if (!set->queue_depth)
2843 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2846 if (!set->ops->queue_rq)
2849 if (!set->ops->get_budget ^ !set->ops->put_budget)
2852 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2853 pr_info("blk-mq: reduced tag depth to %u\n",
2855 set->queue_depth = BLK_MQ_MAX_DEPTH;
2859 * If a crashdump is active, then we are potentially in a very
2860 * memory constrained environment. Limit us to 1 queue and
2861 * 64 tags to prevent using too much memory.
2863 if (is_kdump_kernel()) {
2864 set->nr_hw_queues = 1;
2865 set->queue_depth = min(64U, set->queue_depth);
2868 * There is no use for more h/w queues than cpus.
2870 if (set->nr_hw_queues > nr_cpu_ids)
2871 set->nr_hw_queues = nr_cpu_ids;
2873 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2874 GFP_KERNEL, set->numa_node);
2879 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2880 GFP_KERNEL, set->numa_node);
2884 ret = blk_mq_update_queue_map(set);
2886 goto out_free_mq_map;
2888 ret = blk_mq_alloc_rq_maps(set);
2890 goto out_free_mq_map;
2892 mutex_init(&set->tag_list_lock);
2893 INIT_LIST_HEAD(&set->tag_list);
2905 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2907 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2911 for (i = 0; i < nr_cpu_ids; i++)
2912 blk_mq_free_map_and_requests(set, i);
2920 EXPORT_SYMBOL(blk_mq_free_tag_set);
2922 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2924 struct blk_mq_tag_set *set = q->tag_set;
2925 struct blk_mq_hw_ctx *hctx;
2931 blk_mq_freeze_queue(q);
2932 blk_mq_quiesce_queue(q);
2935 queue_for_each_hw_ctx(q, hctx, i) {
2939 * If we're using an MQ scheduler, just update the scheduler
2940 * queue depth. This is similar to what the old code would do.
2942 if (!hctx->sched_tags) {
2943 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2946 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2954 q->nr_requests = nr;
2956 blk_mq_unquiesce_queue(q);
2957 blk_mq_unfreeze_queue(q);
2962 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2965 struct request_queue *q;
2967 lockdep_assert_held(&set->tag_list_lock);
2969 if (nr_hw_queues > nr_cpu_ids)
2970 nr_hw_queues = nr_cpu_ids;
2971 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2974 list_for_each_entry(q, &set->tag_list, tag_set_list)
2975 blk_mq_freeze_queue(q);
2977 set->nr_hw_queues = nr_hw_queues;
2978 blk_mq_update_queue_map(set);
2979 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2980 blk_mq_realloc_hw_ctxs(set, q);
2981 blk_mq_queue_reinit(q);
2984 list_for_each_entry(q, &set->tag_list, tag_set_list)
2985 blk_mq_unfreeze_queue(q);
2988 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2990 mutex_lock(&set->tag_list_lock);
2991 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2992 mutex_unlock(&set->tag_list_lock);
2994 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2996 /* Enable polling stats and return whether they were already enabled. */
2997 static bool blk_poll_stats_enable(struct request_queue *q)
2999 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3000 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3002 blk_stat_add_callback(q, q->poll_cb);
3006 static void blk_mq_poll_stats_start(struct request_queue *q)
3009 * We don't arm the callback if polling stats are not enabled or the
3010 * callback is already active.
3012 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3013 blk_stat_is_active(q->poll_cb))
3016 blk_stat_activate_msecs(q->poll_cb, 100);
3019 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3021 struct request_queue *q = cb->data;
3024 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3025 if (cb->stat[bucket].nr_samples)
3026 q->poll_stat[bucket] = cb->stat[bucket];
3030 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3031 struct blk_mq_hw_ctx *hctx,
3034 unsigned long ret = 0;
3038 * If stats collection isn't on, don't sleep but turn it on for
3041 if (!blk_poll_stats_enable(q))
3045 * As an optimistic guess, use half of the mean service time
3046 * for this type of request. We can (and should) make this smarter.
3047 * For instance, if the completion latencies are tight, we can
3048 * get closer than just half the mean. This is especially
3049 * important on devices where the completion latencies are longer
3050 * than ~10 usec. We do use the stats for the relevant IO size
3051 * if available which does lead to better estimates.
3053 bucket = blk_mq_poll_stats_bkt(rq);
3057 if (q->poll_stat[bucket].nr_samples)
3058 ret = (q->poll_stat[bucket].mean + 1) / 2;
3063 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3064 struct blk_mq_hw_ctx *hctx,
3067 struct hrtimer_sleeper hs;
3068 enum hrtimer_mode mode;
3072 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3078 * -1: don't ever hybrid sleep
3079 * 0: use half of prev avg
3080 * >0: use this specific value
3082 if (q->poll_nsec == -1)
3084 else if (q->poll_nsec > 0)
3085 nsecs = q->poll_nsec;
3087 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3092 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3095 * This will be replaced with the stats tracking code, using
3096 * 'avg_completion_time / 2' as the pre-sleep target.
3100 mode = HRTIMER_MODE_REL;
3101 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3102 hrtimer_set_expires(&hs.timer, kt);
3104 hrtimer_init_sleeper(&hs, current);
3106 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3108 set_current_state(TASK_UNINTERRUPTIBLE);
3109 hrtimer_start_expires(&hs.timer, mode);
3112 hrtimer_cancel(&hs.timer);
3113 mode = HRTIMER_MODE_ABS;
3114 } while (hs.task && !signal_pending(current));
3116 __set_current_state(TASK_RUNNING);
3117 destroy_hrtimer_on_stack(&hs.timer);
3121 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3123 struct request_queue *q = hctx->queue;
3127 * If we sleep, have the caller restart the poll loop to reset
3128 * the state. Like for the other success return cases, the
3129 * caller is responsible for checking if the IO completed. If
3130 * the IO isn't complete, we'll get called again and will go
3131 * straight to the busy poll loop.
3133 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3136 hctx->poll_considered++;
3138 state = current->state;
3139 while (!need_resched()) {
3142 hctx->poll_invoked++;
3144 ret = q->mq_ops->poll(hctx, rq->tag);
3146 hctx->poll_success++;
3147 set_current_state(TASK_RUNNING);
3151 if (signal_pending_state(state, current))
3152 set_current_state(TASK_RUNNING);
3154 if (current->state == TASK_RUNNING)
3161 __set_current_state(TASK_RUNNING);
3165 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3167 struct blk_mq_hw_ctx *hctx;
3170 if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3173 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3174 if (!blk_qc_t_is_internal(cookie))
3175 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3177 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3179 * With scheduling, if the request has completed, we'll
3180 * get a NULL return here, as we clear the sched tag when
3181 * that happens. The request still remains valid, like always,
3182 * so we should be safe with just the NULL check.
3188 return __blk_mq_poll(hctx, rq);
3191 static int __init blk_mq_init(void)
3193 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3194 blk_mq_hctx_notify_dead);
3197 subsys_initcall(blk_mq_init);