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 (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) &&
99 !test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) {
101 * index[0] counts the specific partition that was asked
102 * for. index[1] counts the ones that are active on the
103 * whole device, so increment that if mi->part is indeed
104 * a partition, and not a whole device.
106 if (rq->part == mi->part)
108 if (mi->part->partno)
113 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
114 unsigned int inflight[2])
116 struct mq_inflight mi = { .part = part, .inflight = inflight, };
118 inflight[0] = inflight[1] = 0;
119 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
122 void blk_freeze_queue_start(struct request_queue *q)
126 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
127 if (freeze_depth == 1) {
128 percpu_ref_kill(&q->q_usage_counter);
130 blk_mq_run_hw_queues(q, false);
133 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
135 void blk_mq_freeze_queue_wait(struct request_queue *q)
137 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
139 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
141 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
142 unsigned long timeout)
144 return wait_event_timeout(q->mq_freeze_wq,
145 percpu_ref_is_zero(&q->q_usage_counter),
148 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
151 * Guarantee no request is in use, so we can change any data structure of
152 * the queue afterward.
154 void blk_freeze_queue(struct request_queue *q)
157 * In the !blk_mq case we are only calling this to kill the
158 * q_usage_counter, otherwise this increases the freeze depth
159 * and waits for it to return to zero. For this reason there is
160 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
161 * exported to drivers as the only user for unfreeze is blk_mq.
163 blk_freeze_queue_start(q);
164 blk_mq_freeze_queue_wait(q);
167 void blk_mq_freeze_queue(struct request_queue *q)
170 * ...just an alias to keep freeze and unfreeze actions balanced
171 * in the blk_mq_* namespace
175 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
177 void blk_mq_unfreeze_queue(struct request_queue *q)
181 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
182 WARN_ON_ONCE(freeze_depth < 0);
184 percpu_ref_reinit(&q->q_usage_counter);
185 wake_up_all(&q->mq_freeze_wq);
188 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
191 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
192 * mpt3sas driver such that this function can be removed.
194 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
198 spin_lock_irqsave(q->queue_lock, flags);
199 queue_flag_set(QUEUE_FLAG_QUIESCED, q);
200 spin_unlock_irqrestore(q->queue_lock, flags);
202 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
205 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
208 * Note: this function does not prevent that the struct request end_io()
209 * callback function is invoked. Once this function is returned, we make
210 * sure no dispatch can happen until the queue is unquiesced via
211 * blk_mq_unquiesce_queue().
213 void blk_mq_quiesce_queue(struct request_queue *q)
215 struct blk_mq_hw_ctx *hctx;
219 blk_mq_quiesce_queue_nowait(q);
221 queue_for_each_hw_ctx(q, hctx, i) {
222 if (hctx->flags & BLK_MQ_F_BLOCKING)
223 synchronize_srcu(hctx->queue_rq_srcu);
230 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
233 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
236 * This function recovers queue into the state before quiescing
237 * which is done by blk_mq_quiesce_queue.
239 void blk_mq_unquiesce_queue(struct request_queue *q)
243 spin_lock_irqsave(q->queue_lock, flags);
244 queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
245 spin_unlock_irqrestore(q->queue_lock, flags);
247 /* dispatch requests which are inserted during quiescing */
248 blk_mq_run_hw_queues(q, true);
250 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
252 void blk_mq_wake_waiters(struct request_queue *q)
254 struct blk_mq_hw_ctx *hctx;
257 queue_for_each_hw_ctx(q, hctx, i)
258 if (blk_mq_hw_queue_mapped(hctx))
259 blk_mq_tag_wakeup_all(hctx->tags, true);
262 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
264 return blk_mq_has_free_tags(hctx->tags);
266 EXPORT_SYMBOL(blk_mq_can_queue);
268 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
269 unsigned int tag, unsigned int op)
271 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
272 struct request *rq = tags->static_rqs[tag];
276 if (data->flags & BLK_MQ_REQ_INTERNAL) {
278 rq->internal_tag = tag;
280 if (blk_mq_tag_busy(data->hctx)) {
281 rq->rq_flags = RQF_MQ_INFLIGHT;
282 atomic_inc(&data->hctx->nr_active);
285 rq->internal_tag = -1;
286 data->hctx->tags->rqs[rq->tag] = rq;
289 INIT_LIST_HEAD(&rq->queuelist);
290 /* csd/requeue_work/fifo_time is initialized before use */
292 rq->mq_ctx = data->ctx;
294 if (data->flags & BLK_MQ_REQ_PREEMPT)
295 rq->rq_flags |= RQF_PREEMPT;
296 if (blk_queue_io_stat(data->q))
297 rq->rq_flags |= RQF_IO_STAT;
298 /* do not touch atomic flags, it needs atomic ops against the timer */
300 INIT_HLIST_NODE(&rq->hash);
301 RB_CLEAR_NODE(&rq->rb_node);
304 rq->start_time = jiffies;
305 #ifdef CONFIG_BLK_CGROUP
307 set_start_time_ns(rq);
308 rq->io_start_time_ns = 0;
310 rq->nr_phys_segments = 0;
311 #if defined(CONFIG_BLK_DEV_INTEGRITY)
312 rq->nr_integrity_segments = 0;
315 /* tag was already set */
318 INIT_LIST_HEAD(&rq->timeout_list);
322 rq->end_io_data = NULL;
325 data->ctx->rq_dispatched[op_is_sync(op)]++;
329 static struct request *blk_mq_get_request(struct request_queue *q,
330 struct bio *bio, unsigned int op,
331 struct blk_mq_alloc_data *data)
333 struct elevator_queue *e = q->elevator;
336 bool put_ctx_on_error = false;
338 blk_queue_enter_live(q);
340 if (likely(!data->ctx)) {
341 data->ctx = blk_mq_get_ctx(q);
342 put_ctx_on_error = true;
344 if (likely(!data->hctx))
345 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
347 data->flags |= BLK_MQ_REQ_NOWAIT;
350 data->flags |= BLK_MQ_REQ_INTERNAL;
353 * Flush requests are special and go directly to the
356 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
357 e->type->ops.mq.limit_depth(op, data);
360 tag = blk_mq_get_tag(data);
361 if (tag == BLK_MQ_TAG_FAIL) {
362 if (put_ctx_on_error) {
363 blk_mq_put_ctx(data->ctx);
370 rq = blk_mq_rq_ctx_init(data, tag, op);
371 if (!op_is_flush(op)) {
373 if (e && e->type->ops.mq.prepare_request) {
374 if (e->type->icq_cache && rq_ioc(bio))
375 blk_mq_sched_assign_ioc(rq, bio);
377 e->type->ops.mq.prepare_request(rq, bio);
378 rq->rq_flags |= RQF_ELVPRIV;
381 data->hctx->queued++;
385 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
386 blk_mq_req_flags_t flags)
388 struct blk_mq_alloc_data alloc_data = { .flags = flags };
392 ret = blk_queue_enter(q, flags);
396 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
400 return ERR_PTR(-EWOULDBLOCK);
402 blk_mq_put_ctx(alloc_data.ctx);
405 rq->__sector = (sector_t) -1;
406 rq->bio = rq->biotail = NULL;
409 EXPORT_SYMBOL(blk_mq_alloc_request);
411 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
412 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
414 struct blk_mq_alloc_data alloc_data = { .flags = flags };
420 * If the tag allocator sleeps we could get an allocation for a
421 * different hardware context. No need to complicate the low level
422 * allocator for this for the rare use case of a command tied to
425 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
426 return ERR_PTR(-EINVAL);
428 if (hctx_idx >= q->nr_hw_queues)
429 return ERR_PTR(-EIO);
431 ret = blk_queue_enter(q, flags);
436 * Check if the hardware context is actually mapped to anything.
437 * If not tell the caller that it should skip this queue.
439 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
440 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
442 return ERR_PTR(-EXDEV);
444 cpu = cpumask_first(alloc_data.hctx->cpumask);
445 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
447 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
451 return ERR_PTR(-EWOULDBLOCK);
455 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
457 void blk_mq_free_request(struct request *rq)
459 struct request_queue *q = rq->q;
460 struct elevator_queue *e = q->elevator;
461 struct blk_mq_ctx *ctx = rq->mq_ctx;
462 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
463 const int sched_tag = rq->internal_tag;
465 if (rq->rq_flags & RQF_ELVPRIV) {
466 if (e && e->type->ops.mq.finish_request)
467 e->type->ops.mq.finish_request(rq);
469 put_io_context(rq->elv.icq->ioc);
474 ctx->rq_completed[rq_is_sync(rq)]++;
475 if (rq->rq_flags & RQF_MQ_INFLIGHT)
476 atomic_dec(&hctx->nr_active);
478 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
479 laptop_io_completion(q->backing_dev_info);
481 wbt_done(q->rq_wb, &rq->issue_stat);
484 blk_put_rl(blk_rq_rl(rq));
486 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
487 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
488 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
490 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
492 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
493 blk_mq_sched_restart(hctx);
496 EXPORT_SYMBOL_GPL(blk_mq_free_request);
498 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
500 blk_account_io_done(rq);
503 wbt_done(rq->q->rq_wb, &rq->issue_stat);
504 rq->end_io(rq, error);
506 if (unlikely(blk_bidi_rq(rq)))
507 blk_mq_free_request(rq->next_rq);
508 blk_mq_free_request(rq);
511 EXPORT_SYMBOL(__blk_mq_end_request);
513 void blk_mq_end_request(struct request *rq, blk_status_t error)
515 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
517 __blk_mq_end_request(rq, error);
519 EXPORT_SYMBOL(blk_mq_end_request);
521 static void __blk_mq_complete_request_remote(void *data)
523 struct request *rq = data;
525 rq->q->softirq_done_fn(rq);
528 static void __blk_mq_complete_request(struct request *rq)
530 struct blk_mq_ctx *ctx = rq->mq_ctx;
534 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT);
536 if (rq->internal_tag != -1)
537 blk_mq_sched_completed_request(rq);
538 if (rq->rq_flags & RQF_STATS) {
539 blk_mq_poll_stats_start(rq->q);
543 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
544 rq->q->softirq_done_fn(rq);
549 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
550 shared = cpus_share_cache(cpu, ctx->cpu);
552 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
553 rq->csd.func = __blk_mq_complete_request_remote;
556 smp_call_function_single_async(ctx->cpu, &rq->csd);
558 rq->q->softirq_done_fn(rq);
563 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
565 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
568 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
571 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
573 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
576 *srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
579 static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate)
584 * blk_mq_rq_aborted_gstate() is used from the completion path and
585 * can thus be called from irq context. u64_stats_fetch in the
586 * middle of update on the same CPU leads to lockup. Disable irq
589 local_irq_save(flags);
590 u64_stats_update_begin(&rq->aborted_gstate_sync);
591 rq->aborted_gstate = gstate;
592 u64_stats_update_end(&rq->aborted_gstate_sync);
593 local_irq_restore(flags);
596 static u64 blk_mq_rq_aborted_gstate(struct request *rq)
602 start = u64_stats_fetch_begin(&rq->aborted_gstate_sync);
603 aborted_gstate = rq->aborted_gstate;
604 } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start));
606 return aborted_gstate;
610 * blk_mq_complete_request - end I/O on a request
611 * @rq: the request being processed
614 * Ends all I/O on a request. It does not handle partial completions.
615 * The actual completion happens out-of-order, through a IPI handler.
617 void blk_mq_complete_request(struct request *rq)
619 struct request_queue *q = rq->q;
620 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
623 if (unlikely(blk_should_fake_timeout(q)))
627 * If @rq->aborted_gstate equals the current instance, timeout is
628 * claiming @rq and we lost. This is synchronized through
629 * hctx_lock(). See blk_mq_timeout_work() for details.
631 * Completion path never blocks and we can directly use RCU here
632 * instead of hctx_lock() which can be either RCU or SRCU.
633 * However, that would complicate paths which want to synchronize
634 * against us. Let stay in sync with the issue path so that
635 * hctx_lock() covers both issue and completion paths.
637 hctx_lock(hctx, &srcu_idx);
638 if (blk_mq_rq_aborted_gstate(rq) != rq->gstate &&
639 !blk_mark_rq_complete(rq))
640 __blk_mq_complete_request(rq);
641 hctx_unlock(hctx, srcu_idx);
643 EXPORT_SYMBOL(blk_mq_complete_request);
645 int blk_mq_request_started(struct request *rq)
647 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
649 EXPORT_SYMBOL_GPL(blk_mq_request_started);
651 void blk_mq_start_request(struct request *rq)
653 struct request_queue *q = rq->q;
655 blk_mq_sched_started_request(rq);
657 trace_block_rq_issue(q, rq);
659 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
660 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
661 rq->rq_flags |= RQF_STATS;
662 wbt_issue(q->rq_wb, &rq->issue_stat);
665 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
666 WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags));
669 * Mark @rq in-flight which also advances the generation number,
670 * and register for timeout. Protect with a seqcount to allow the
671 * timeout path to read both @rq->gstate and @rq->deadline
674 * This is the only place where a request is marked in-flight. If
675 * the timeout path reads an in-flight @rq->gstate, the
676 * @rq->deadline it reads together under @rq->gstate_seq is
677 * guaranteed to be the matching one.
680 write_seqcount_begin(&rq->gstate_seq);
682 blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT);
685 write_seqcount_end(&rq->gstate_seq);
688 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
689 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
690 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
692 if (q->dma_drain_size && blk_rq_bytes(rq)) {
694 * Make sure space for the drain appears. We know we can do
695 * this because max_hw_segments has been adjusted to be one
696 * fewer than the device can handle.
698 rq->nr_phys_segments++;
701 EXPORT_SYMBOL(blk_mq_start_request);
704 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
705 * flag isn't set yet, so there may be race with timeout handler,
706 * but given rq->deadline is just set in .queue_rq() under
707 * this situation, the race won't be possible in reality because
708 * rq->timeout should be set as big enough to cover the window
709 * between blk_mq_start_request() called from .queue_rq() and
710 * clearing REQ_ATOM_STARTED here.
712 static void __blk_mq_requeue_request(struct request *rq)
714 struct request_queue *q = rq->q;
716 blk_mq_put_driver_tag(rq);
718 trace_block_rq_requeue(q, rq);
719 wbt_requeue(q->rq_wb, &rq->issue_stat);
720 blk_mq_sched_requeue_request(rq);
722 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
723 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
724 if (q->dma_drain_size && blk_rq_bytes(rq))
725 rq->nr_phys_segments--;
729 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
731 __blk_mq_requeue_request(rq);
733 BUG_ON(blk_queued_rq(rq));
734 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
736 EXPORT_SYMBOL(blk_mq_requeue_request);
738 static void blk_mq_requeue_work(struct work_struct *work)
740 struct request_queue *q =
741 container_of(work, struct request_queue, requeue_work.work);
743 struct request *rq, *next;
745 spin_lock_irq(&q->requeue_lock);
746 list_splice_init(&q->requeue_list, &rq_list);
747 spin_unlock_irq(&q->requeue_lock);
749 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
750 if (!(rq->rq_flags & RQF_SOFTBARRIER))
753 rq->rq_flags &= ~RQF_SOFTBARRIER;
754 list_del_init(&rq->queuelist);
755 blk_mq_sched_insert_request(rq, true, false, false, true);
758 while (!list_empty(&rq_list)) {
759 rq = list_entry(rq_list.next, struct request, queuelist);
760 list_del_init(&rq->queuelist);
761 blk_mq_sched_insert_request(rq, false, false, false, true);
764 blk_mq_run_hw_queues(q, false);
767 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
768 bool kick_requeue_list)
770 struct request_queue *q = rq->q;
774 * We abuse this flag that is otherwise used by the I/O scheduler to
775 * request head insertion from the workqueue.
777 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
779 spin_lock_irqsave(&q->requeue_lock, flags);
781 rq->rq_flags |= RQF_SOFTBARRIER;
782 list_add(&rq->queuelist, &q->requeue_list);
784 list_add_tail(&rq->queuelist, &q->requeue_list);
786 spin_unlock_irqrestore(&q->requeue_lock, flags);
788 if (kick_requeue_list)
789 blk_mq_kick_requeue_list(q);
791 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
793 void blk_mq_kick_requeue_list(struct request_queue *q)
795 kblockd_schedule_delayed_work(&q->requeue_work, 0);
797 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
799 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
802 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
803 msecs_to_jiffies(msecs));
805 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
807 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
809 if (tag < tags->nr_tags) {
810 prefetch(tags->rqs[tag]);
811 return tags->rqs[tag];
816 EXPORT_SYMBOL(blk_mq_tag_to_rq);
818 struct blk_mq_timeout_data {
820 unsigned int next_set;
821 unsigned int nr_expired;
824 void blk_mq_rq_timed_out(struct request *req, bool reserved)
826 const struct blk_mq_ops *ops = req->q->mq_ops;
827 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
830 * We know that complete is set at this point. If STARTED isn't set
831 * anymore, then the request isn't active and the "timeout" should
832 * just be ignored. This can happen due to the bitflag ordering.
833 * Timeout first checks if STARTED is set, and if it is, assumes
834 * the request is active. But if we race with completion, then
835 * both flags will get cleared. So check here again, and ignore
836 * a timeout event with a request that isn't active.
838 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
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);
856 blk_clear_rq_complete(req);
858 case BLK_EH_NOT_HANDLED:
861 printk(KERN_ERR "block: bad eh return: %d\n", ret);
866 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
867 struct request *rq, void *priv, bool reserved)
869 struct blk_mq_timeout_data *data = priv;
870 unsigned long gstate, deadline;
875 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
878 /* read coherent snapshots of @rq->state_gen and @rq->deadline */
880 start = read_seqcount_begin(&rq->gstate_seq);
881 gstate = READ_ONCE(rq->gstate);
882 deadline = rq->deadline;
883 if (!read_seqcount_retry(&rq->gstate_seq, start))
888 /* if in-flight && overdue, mark for abortion */
889 if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT &&
890 time_after_eq(jiffies, deadline)) {
891 blk_mq_rq_update_aborted_gstate(rq, gstate);
894 } else if (!data->next_set || time_after(data->next, deadline)) {
895 data->next = deadline;
900 static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx,
901 struct request *rq, void *priv, bool reserved)
904 * We marked @rq->aborted_gstate and waited for RCU. If there were
905 * completions that we lost to, they would have finished and
906 * updated @rq->gstate by now; otherwise, the completion path is
907 * now guaranteed to see @rq->aborted_gstate and yield. If
908 * @rq->aborted_gstate still matches @rq->gstate, @rq is ours.
910 if (READ_ONCE(rq->gstate) == rq->aborted_gstate &&
911 !blk_mark_rq_complete(rq))
912 blk_mq_rq_timed_out(rq, reserved);
915 static void blk_mq_timeout_work(struct work_struct *work)
917 struct request_queue *q =
918 container_of(work, struct request_queue, timeout_work);
919 struct blk_mq_timeout_data data = {
924 struct blk_mq_hw_ctx *hctx;
927 /* A deadlock might occur if a request is stuck requiring a
928 * timeout at the same time a queue freeze is waiting
929 * completion, since the timeout code would not be able to
930 * acquire the queue reference here.
932 * That's why we don't use blk_queue_enter here; instead, we use
933 * percpu_ref_tryget directly, because we need to be able to
934 * obtain a reference even in the short window between the queue
935 * starting to freeze, by dropping the first reference in
936 * blk_freeze_queue_start, and the moment the last request is
937 * consumed, marked by the instant q_usage_counter reaches
940 if (!percpu_ref_tryget(&q->q_usage_counter))
943 /* scan for the expired ones and set their ->aborted_gstate */
944 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
946 if (data.nr_expired) {
947 bool has_rcu = false;
950 * Wait till everyone sees ->aborted_gstate. The
951 * sequential waits for SRCUs aren't ideal. If this ever
952 * becomes a problem, we can add per-hw_ctx rcu_head and
955 queue_for_each_hw_ctx(q, hctx, i) {
956 if (!hctx->nr_expired)
959 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
962 synchronize_srcu(hctx->queue_rq_srcu);
964 hctx->nr_expired = 0;
969 /* terminate the ones we won */
970 blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL);
974 data.next = blk_rq_timeout(round_jiffies_up(data.next));
975 mod_timer(&q->timeout, data.next);
977 queue_for_each_hw_ctx(q, hctx, i) {
978 /* the hctx may be unmapped, so check it here */
979 if (blk_mq_hw_queue_mapped(hctx))
980 blk_mq_tag_idle(hctx);
986 struct flush_busy_ctx_data {
987 struct blk_mq_hw_ctx *hctx;
988 struct list_head *list;
991 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
993 struct flush_busy_ctx_data *flush_data = data;
994 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
995 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
997 sbitmap_clear_bit(sb, bitnr);
998 spin_lock(&ctx->lock);
999 list_splice_tail_init(&ctx->rq_list, flush_data->list);
1000 spin_unlock(&ctx->lock);
1005 * Process software queues that have been marked busy, splicing them
1006 * to the for-dispatch
1008 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1010 struct flush_busy_ctx_data data = {
1015 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1017 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1019 struct dispatch_rq_data {
1020 struct blk_mq_hw_ctx *hctx;
1024 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1027 struct dispatch_rq_data *dispatch_data = data;
1028 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1029 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1031 spin_lock(&ctx->lock);
1032 if (unlikely(!list_empty(&ctx->rq_list))) {
1033 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
1034 list_del_init(&dispatch_data->rq->queuelist);
1035 if (list_empty(&ctx->rq_list))
1036 sbitmap_clear_bit(sb, bitnr);
1038 spin_unlock(&ctx->lock);
1040 return !dispatch_data->rq;
1043 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1044 struct blk_mq_ctx *start)
1046 unsigned off = start ? start->index_hw : 0;
1047 struct dispatch_rq_data data = {
1052 __sbitmap_for_each_set(&hctx->ctx_map, off,
1053 dispatch_rq_from_ctx, &data);
1058 static inline unsigned int queued_to_index(unsigned int queued)
1063 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1066 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
1069 struct blk_mq_alloc_data data = {
1071 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
1072 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
1075 might_sleep_if(wait);
1080 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1081 data.flags |= BLK_MQ_REQ_RESERVED;
1083 rq->tag = blk_mq_get_tag(&data);
1085 if (blk_mq_tag_busy(data.hctx)) {
1086 rq->rq_flags |= RQF_MQ_INFLIGHT;
1087 atomic_inc(&data.hctx->nr_active);
1089 data.hctx->tags->rqs[rq->tag] = rq;
1095 return rq->tag != -1;
1098 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1099 int flags, void *key)
1101 struct blk_mq_hw_ctx *hctx;
1103 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1105 list_del_init(&wait->entry);
1106 blk_mq_run_hw_queue(hctx, true);
1111 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1112 * the tag wakeups. For non-shared tags, we can simply mark us nedeing a
1113 * restart. For both caes, take care to check the condition again after
1114 * marking us as waiting.
1116 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1119 struct blk_mq_hw_ctx *this_hctx = *hctx;
1120 bool shared_tags = (this_hctx->flags & BLK_MQ_F_TAG_SHARED) != 0;
1121 struct sbq_wait_state *ws;
1122 wait_queue_entry_t *wait;
1126 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1127 set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1129 wait = &this_hctx->dispatch_wait;
1130 if (!list_empty_careful(&wait->entry))
1133 spin_lock(&this_hctx->lock);
1134 if (!list_empty(&wait->entry)) {
1135 spin_unlock(&this_hctx->lock);
1139 ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1140 add_wait_queue(&ws->wait, wait);
1144 * It's possible that a tag was freed in the window between the
1145 * allocation failure and adding the hardware queue to the wait
1148 ret = blk_mq_get_driver_tag(rq, hctx, false);
1152 * Don't clear RESTART here, someone else could have set it.
1153 * At most this will cost an extra queue run.
1158 spin_unlock(&this_hctx->lock);
1163 * We got a tag, remove ourselves from the wait queue to ensure
1164 * someone else gets the wakeup.
1166 spin_lock_irq(&ws->wait.lock);
1167 list_del_init(&wait->entry);
1168 spin_unlock_irq(&ws->wait.lock);
1169 spin_unlock(&this_hctx->lock);
1174 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1177 struct blk_mq_hw_ctx *hctx;
1178 struct request *rq, *nxt;
1179 bool no_tag = false;
1182 if (list_empty(list))
1185 WARN_ON(!list_is_singular(list) && got_budget);
1188 * Now process all the entries, sending them to the driver.
1190 errors = queued = 0;
1192 struct blk_mq_queue_data bd;
1195 rq = list_first_entry(list, struct request, queuelist);
1196 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1198 * The initial allocation attempt failed, so we need to
1199 * rerun the hardware queue when a tag is freed. The
1200 * waitqueue takes care of that. If the queue is run
1201 * before we add this entry back on the dispatch list,
1202 * we'll re-run it below.
1204 if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1206 blk_mq_put_dispatch_budget(hctx);
1208 * For non-shared tags, the RESTART check
1211 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1217 if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) {
1218 blk_mq_put_driver_tag(rq);
1222 list_del_init(&rq->queuelist);
1227 * Flag last if we have no more requests, or if we have more
1228 * but can't assign a driver tag to it.
1230 if (list_empty(list))
1233 nxt = list_first_entry(list, struct request, queuelist);
1234 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1237 ret = q->mq_ops->queue_rq(hctx, &bd);
1238 if (ret == BLK_STS_RESOURCE) {
1240 * If an I/O scheduler has been configured and we got a
1241 * driver tag for the next request already, free it
1244 if (!list_empty(list)) {
1245 nxt = list_first_entry(list, struct request, queuelist);
1246 blk_mq_put_driver_tag(nxt);
1248 list_add(&rq->queuelist, list);
1249 __blk_mq_requeue_request(rq);
1253 if (unlikely(ret != BLK_STS_OK)) {
1255 blk_mq_end_request(rq, BLK_STS_IOERR);
1260 } while (!list_empty(list));
1262 hctx->dispatched[queued_to_index(queued)]++;
1265 * Any items that need requeuing? Stuff them into hctx->dispatch,
1266 * that is where we will continue on next queue run.
1268 if (!list_empty(list)) {
1269 spin_lock(&hctx->lock);
1270 list_splice_init(list, &hctx->dispatch);
1271 spin_unlock(&hctx->lock);
1274 * If SCHED_RESTART was set by the caller of this function and
1275 * it is no longer set that means that it was cleared by another
1276 * thread and hence that a queue rerun is needed.
1278 * If 'no_tag' is set, that means that we failed getting
1279 * a driver tag with an I/O scheduler attached. If our dispatch
1280 * waitqueue is no longer active, ensure that we run the queue
1281 * AFTER adding our entries back to the list.
1283 * If no I/O scheduler has been configured it is possible that
1284 * the hardware queue got stopped and restarted before requests
1285 * were pushed back onto the dispatch list. Rerun the queue to
1286 * avoid starvation. Notes:
1287 * - blk_mq_run_hw_queue() checks whether or not a queue has
1288 * been stopped before rerunning a queue.
1289 * - Some but not all block drivers stop a queue before
1290 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1293 if (!blk_mq_sched_needs_restart(hctx) ||
1294 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1295 blk_mq_run_hw_queue(hctx, true);
1298 return (queued + errors) != 0;
1301 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1306 * We should be running this queue from one of the CPUs that
1309 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1310 cpu_online(hctx->next_cpu));
1313 * We can't run the queue inline with ints disabled. Ensure that
1314 * we catch bad users of this early.
1316 WARN_ON_ONCE(in_interrupt());
1318 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1320 hctx_lock(hctx, &srcu_idx);
1321 blk_mq_sched_dispatch_requests(hctx);
1322 hctx_unlock(hctx, srcu_idx);
1326 * It'd be great if the workqueue API had a way to pass
1327 * in a mask and had some smarts for more clever placement.
1328 * For now we just round-robin here, switching for every
1329 * BLK_MQ_CPU_WORK_BATCH queued items.
1331 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1333 if (hctx->queue->nr_hw_queues == 1)
1334 return WORK_CPU_UNBOUND;
1336 if (--hctx->next_cpu_batch <= 0) {
1339 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1340 if (next_cpu >= nr_cpu_ids)
1341 next_cpu = cpumask_first(hctx->cpumask);
1343 hctx->next_cpu = next_cpu;
1344 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1347 return hctx->next_cpu;
1350 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1351 unsigned long msecs)
1353 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1356 if (unlikely(blk_mq_hctx_stopped(hctx)))
1359 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1360 int cpu = get_cpu();
1361 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1362 __blk_mq_run_hw_queue(hctx);
1370 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1372 msecs_to_jiffies(msecs));
1375 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1377 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1379 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1381 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1387 * When queue is quiesced, we may be switching io scheduler, or
1388 * updating nr_hw_queues, or other things, and we can't run queue
1389 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1391 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1394 hctx_lock(hctx, &srcu_idx);
1395 need_run = !blk_queue_quiesced(hctx->queue) &&
1396 blk_mq_hctx_has_pending(hctx);
1397 hctx_unlock(hctx, srcu_idx);
1400 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1406 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1408 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1410 struct blk_mq_hw_ctx *hctx;
1413 queue_for_each_hw_ctx(q, hctx, i) {
1414 if (blk_mq_hctx_stopped(hctx))
1417 blk_mq_run_hw_queue(hctx, async);
1420 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1423 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1424 * @q: request queue.
1426 * The caller is responsible for serializing this function against
1427 * blk_mq_{start,stop}_hw_queue().
1429 bool blk_mq_queue_stopped(struct request_queue *q)
1431 struct blk_mq_hw_ctx *hctx;
1434 queue_for_each_hw_ctx(q, hctx, i)
1435 if (blk_mq_hctx_stopped(hctx))
1440 EXPORT_SYMBOL(blk_mq_queue_stopped);
1443 * This function is often used for pausing .queue_rq() by driver when
1444 * there isn't enough resource or some conditions aren't satisfied, and
1445 * BLK_STS_RESOURCE is usually returned.
1447 * We do not guarantee that dispatch can be drained or blocked
1448 * after blk_mq_stop_hw_queue() returns. Please use
1449 * blk_mq_quiesce_queue() for that requirement.
1451 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1453 cancel_delayed_work(&hctx->run_work);
1455 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1457 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1460 * This function is often used for pausing .queue_rq() by driver when
1461 * there isn't enough resource or some conditions aren't satisfied, and
1462 * BLK_STS_RESOURCE is usually returned.
1464 * We do not guarantee that dispatch can be drained or blocked
1465 * after blk_mq_stop_hw_queues() returns. Please use
1466 * blk_mq_quiesce_queue() for that requirement.
1468 void blk_mq_stop_hw_queues(struct request_queue *q)
1470 struct blk_mq_hw_ctx *hctx;
1473 queue_for_each_hw_ctx(q, hctx, i)
1474 blk_mq_stop_hw_queue(hctx);
1476 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1478 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1480 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1482 blk_mq_run_hw_queue(hctx, false);
1484 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1486 void blk_mq_start_hw_queues(struct request_queue *q)
1488 struct blk_mq_hw_ctx *hctx;
1491 queue_for_each_hw_ctx(q, hctx, i)
1492 blk_mq_start_hw_queue(hctx);
1494 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1496 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1498 if (!blk_mq_hctx_stopped(hctx))
1501 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1502 blk_mq_run_hw_queue(hctx, async);
1504 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1506 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1508 struct blk_mq_hw_ctx *hctx;
1511 queue_for_each_hw_ctx(q, hctx, i)
1512 blk_mq_start_stopped_hw_queue(hctx, async);
1514 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1516 static void blk_mq_run_work_fn(struct work_struct *work)
1518 struct blk_mq_hw_ctx *hctx;
1520 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1523 * If we are stopped, don't run the queue. The exception is if
1524 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1525 * the STOPPED bit and run it.
1527 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1528 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1531 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1532 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1535 __blk_mq_run_hw_queue(hctx);
1539 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1541 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1545 * Stop the hw queue, then modify currently delayed work.
1546 * This should prevent us from running the queue prematurely.
1547 * Mark the queue as auto-clearing STOPPED when it runs.
1549 blk_mq_stop_hw_queue(hctx);
1550 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1551 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1553 msecs_to_jiffies(msecs));
1555 EXPORT_SYMBOL(blk_mq_delay_queue);
1557 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1561 struct blk_mq_ctx *ctx = rq->mq_ctx;
1563 lockdep_assert_held(&ctx->lock);
1565 trace_block_rq_insert(hctx->queue, rq);
1568 list_add(&rq->queuelist, &ctx->rq_list);
1570 list_add_tail(&rq->queuelist, &ctx->rq_list);
1573 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1576 struct blk_mq_ctx *ctx = rq->mq_ctx;
1578 lockdep_assert_held(&ctx->lock);
1580 __blk_mq_insert_req_list(hctx, rq, at_head);
1581 blk_mq_hctx_mark_pending(hctx, ctx);
1585 * Should only be used carefully, when the caller knows we want to
1586 * bypass a potential IO scheduler on the target device.
1588 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1590 struct blk_mq_ctx *ctx = rq->mq_ctx;
1591 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1593 spin_lock(&hctx->lock);
1594 list_add_tail(&rq->queuelist, &hctx->dispatch);
1595 spin_unlock(&hctx->lock);
1598 blk_mq_run_hw_queue(hctx, false);
1601 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1602 struct list_head *list)
1606 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1609 spin_lock(&ctx->lock);
1610 while (!list_empty(list)) {
1613 rq = list_first_entry(list, struct request, queuelist);
1614 BUG_ON(rq->mq_ctx != ctx);
1615 list_del_init(&rq->queuelist);
1616 __blk_mq_insert_req_list(hctx, rq, false);
1618 blk_mq_hctx_mark_pending(hctx, ctx);
1619 spin_unlock(&ctx->lock);
1622 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1624 struct request *rqa = container_of(a, struct request, queuelist);
1625 struct request *rqb = container_of(b, struct request, queuelist);
1627 return !(rqa->mq_ctx < rqb->mq_ctx ||
1628 (rqa->mq_ctx == rqb->mq_ctx &&
1629 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1632 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1634 struct blk_mq_ctx *this_ctx;
1635 struct request_queue *this_q;
1638 LIST_HEAD(ctx_list);
1641 list_splice_init(&plug->mq_list, &list);
1643 list_sort(NULL, &list, plug_ctx_cmp);
1649 while (!list_empty(&list)) {
1650 rq = list_entry_rq(list.next);
1651 list_del_init(&rq->queuelist);
1653 if (rq->mq_ctx != this_ctx) {
1655 trace_block_unplug(this_q, depth, from_schedule);
1656 blk_mq_sched_insert_requests(this_q, this_ctx,
1661 this_ctx = rq->mq_ctx;
1667 list_add_tail(&rq->queuelist, &ctx_list);
1671 * If 'this_ctx' is set, we know we have entries to complete
1672 * on 'ctx_list'. Do those.
1675 trace_block_unplug(this_q, depth, from_schedule);
1676 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1681 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1683 blk_init_request_from_bio(rq, bio);
1685 blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1687 blk_account_io_start(rq, true);
1690 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1691 struct blk_mq_ctx *ctx,
1694 spin_lock(&ctx->lock);
1695 __blk_mq_insert_request(hctx, rq, false);
1696 spin_unlock(&ctx->lock);
1699 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1702 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1704 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1707 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1711 struct request_queue *q = rq->q;
1712 struct blk_mq_queue_data bd = {
1716 blk_qc_t new_cookie;
1718 bool run_queue = true;
1720 /* RCU or SRCU read lock is needed before checking quiesced flag */
1721 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1729 if (!blk_mq_get_driver_tag(rq, NULL, false))
1732 if (!blk_mq_get_dispatch_budget(hctx)) {
1733 blk_mq_put_driver_tag(rq);
1737 new_cookie = request_to_qc_t(hctx, rq);
1740 * For OK queue, we are done. For error, kill it. Any other
1741 * error (busy), just add it to our list as we previously
1744 ret = q->mq_ops->queue_rq(hctx, &bd);
1747 *cookie = new_cookie;
1749 case BLK_STS_RESOURCE:
1750 __blk_mq_requeue_request(rq);
1753 *cookie = BLK_QC_T_NONE;
1754 blk_mq_end_request(rq, ret);
1759 blk_mq_sched_insert_request(rq, false, run_queue, false,
1760 hctx->flags & BLK_MQ_F_BLOCKING);
1763 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1764 struct request *rq, blk_qc_t *cookie)
1768 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1770 hctx_lock(hctx, &srcu_idx);
1771 __blk_mq_try_issue_directly(hctx, rq, cookie);
1772 hctx_unlock(hctx, srcu_idx);
1775 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1777 const int is_sync = op_is_sync(bio->bi_opf);
1778 const int is_flush_fua = op_is_flush(bio->bi_opf);
1779 struct blk_mq_alloc_data data = { .flags = 0 };
1781 unsigned int request_count = 0;
1782 struct blk_plug *plug;
1783 struct request *same_queue_rq = NULL;
1785 unsigned int wb_acct;
1787 blk_queue_bounce(q, &bio);
1789 blk_queue_split(q, &bio);
1791 if (!bio_integrity_prep(bio))
1792 return BLK_QC_T_NONE;
1794 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1795 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1796 return BLK_QC_T_NONE;
1798 if (blk_mq_sched_bio_merge(q, bio))
1799 return BLK_QC_T_NONE;
1801 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1803 trace_block_getrq(q, bio, bio->bi_opf);
1805 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1806 if (unlikely(!rq)) {
1807 __wbt_done(q->rq_wb, wb_acct);
1808 if (bio->bi_opf & REQ_NOWAIT)
1809 bio_wouldblock_error(bio);
1810 return BLK_QC_T_NONE;
1813 wbt_track(&rq->issue_stat, wb_acct);
1815 cookie = request_to_qc_t(data.hctx, rq);
1817 plug = current->plug;
1818 if (unlikely(is_flush_fua)) {
1819 blk_mq_put_ctx(data.ctx);
1820 blk_mq_bio_to_request(rq, bio);
1822 /* bypass scheduler for flush rq */
1823 blk_insert_flush(rq);
1824 blk_mq_run_hw_queue(data.hctx, true);
1825 } else if (plug && q->nr_hw_queues == 1) {
1826 struct request *last = NULL;
1828 blk_mq_put_ctx(data.ctx);
1829 blk_mq_bio_to_request(rq, bio);
1832 * @request_count may become stale because of schedule
1833 * out, so check the list again.
1835 if (list_empty(&plug->mq_list))
1837 else if (blk_queue_nomerges(q))
1838 request_count = blk_plug_queued_count(q);
1841 trace_block_plug(q);
1843 last = list_entry_rq(plug->mq_list.prev);
1845 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1846 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1847 blk_flush_plug_list(plug, false);
1848 trace_block_plug(q);
1851 list_add_tail(&rq->queuelist, &plug->mq_list);
1852 } else if (plug && !blk_queue_nomerges(q)) {
1853 blk_mq_bio_to_request(rq, bio);
1856 * We do limited plugging. If the bio can be merged, do that.
1857 * Otherwise the existing request in the plug list will be
1858 * issued. So the plug list will have one request at most
1859 * The plug list might get flushed before this. If that happens,
1860 * the plug list is empty, and same_queue_rq is invalid.
1862 if (list_empty(&plug->mq_list))
1863 same_queue_rq = NULL;
1865 list_del_init(&same_queue_rq->queuelist);
1866 list_add_tail(&rq->queuelist, &plug->mq_list);
1868 blk_mq_put_ctx(data.ctx);
1870 if (same_queue_rq) {
1871 data.hctx = blk_mq_map_queue(q,
1872 same_queue_rq->mq_ctx->cpu);
1873 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1876 } else if (q->nr_hw_queues > 1 && is_sync) {
1877 blk_mq_put_ctx(data.ctx);
1878 blk_mq_bio_to_request(rq, bio);
1879 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1880 } else if (q->elevator) {
1881 blk_mq_put_ctx(data.ctx);
1882 blk_mq_bio_to_request(rq, bio);
1883 blk_mq_sched_insert_request(rq, false, true, true, true);
1885 blk_mq_put_ctx(data.ctx);
1886 blk_mq_bio_to_request(rq, bio);
1887 blk_mq_queue_io(data.hctx, data.ctx, rq);
1888 blk_mq_run_hw_queue(data.hctx, true);
1894 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1895 unsigned int hctx_idx)
1899 if (tags->rqs && set->ops->exit_request) {
1902 for (i = 0; i < tags->nr_tags; i++) {
1903 struct request *rq = tags->static_rqs[i];
1907 set->ops->exit_request(set, rq, hctx_idx);
1908 tags->static_rqs[i] = NULL;
1912 while (!list_empty(&tags->page_list)) {
1913 page = list_first_entry(&tags->page_list, struct page, lru);
1914 list_del_init(&page->lru);
1916 * Remove kmemleak object previously allocated in
1917 * blk_mq_init_rq_map().
1919 kmemleak_free(page_address(page));
1920 __free_pages(page, page->private);
1924 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1928 kfree(tags->static_rqs);
1929 tags->static_rqs = NULL;
1931 blk_mq_free_tags(tags);
1934 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1935 unsigned int hctx_idx,
1936 unsigned int nr_tags,
1937 unsigned int reserved_tags)
1939 struct blk_mq_tags *tags;
1942 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1943 if (node == NUMA_NO_NODE)
1944 node = set->numa_node;
1946 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1947 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1951 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1952 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1955 blk_mq_free_tags(tags);
1959 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1960 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1962 if (!tags->static_rqs) {
1964 blk_mq_free_tags(tags);
1971 static size_t order_to_size(unsigned int order)
1973 return (size_t)PAGE_SIZE << order;
1976 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1977 unsigned int hctx_idx, int node)
1981 if (set->ops->init_request) {
1982 ret = set->ops->init_request(set, rq, hctx_idx, node);
1987 seqcount_init(&rq->gstate_seq);
1988 u64_stats_init(&rq->aborted_gstate_sync);
1992 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1993 unsigned int hctx_idx, unsigned int depth)
1995 unsigned int i, j, entries_per_page, max_order = 4;
1996 size_t rq_size, left;
1999 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2000 if (node == NUMA_NO_NODE)
2001 node = set->numa_node;
2003 INIT_LIST_HEAD(&tags->page_list);
2006 * rq_size is the size of the request plus driver payload, rounded
2007 * to the cacheline size
2009 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2011 left = rq_size * depth;
2013 for (i = 0; i < depth; ) {
2014 int this_order = max_order;
2019 while (this_order && left < order_to_size(this_order - 1))
2023 page = alloc_pages_node(node,
2024 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2030 if (order_to_size(this_order) < rq_size)
2037 page->private = this_order;
2038 list_add_tail(&page->lru, &tags->page_list);
2040 p = page_address(page);
2042 * Allow kmemleak to scan these pages as they contain pointers
2043 * to additional allocations like via ops->init_request().
2045 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2046 entries_per_page = order_to_size(this_order) / rq_size;
2047 to_do = min(entries_per_page, depth - i);
2048 left -= to_do * rq_size;
2049 for (j = 0; j < to_do; j++) {
2050 struct request *rq = p;
2052 tags->static_rqs[i] = rq;
2053 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2054 tags->static_rqs[i] = NULL;
2065 blk_mq_free_rqs(set, tags, hctx_idx);
2070 * 'cpu' is going away. splice any existing rq_list entries from this
2071 * software queue to the hw queue dispatch list, and ensure that it
2074 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2076 struct blk_mq_hw_ctx *hctx;
2077 struct blk_mq_ctx *ctx;
2080 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2081 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2083 spin_lock(&ctx->lock);
2084 if (!list_empty(&ctx->rq_list)) {
2085 list_splice_init(&ctx->rq_list, &tmp);
2086 blk_mq_hctx_clear_pending(hctx, ctx);
2088 spin_unlock(&ctx->lock);
2090 if (list_empty(&tmp))
2093 spin_lock(&hctx->lock);
2094 list_splice_tail_init(&tmp, &hctx->dispatch);
2095 spin_unlock(&hctx->lock);
2097 blk_mq_run_hw_queue(hctx, true);
2101 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2103 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2107 /* hctx->ctxs will be freed in queue's release handler */
2108 static void blk_mq_exit_hctx(struct request_queue *q,
2109 struct blk_mq_tag_set *set,
2110 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2112 blk_mq_debugfs_unregister_hctx(hctx);
2114 if (blk_mq_hw_queue_mapped(hctx))
2115 blk_mq_tag_idle(hctx);
2117 if (set->ops->exit_request)
2118 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2120 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2122 if (set->ops->exit_hctx)
2123 set->ops->exit_hctx(hctx, hctx_idx);
2125 if (hctx->flags & BLK_MQ_F_BLOCKING)
2126 cleanup_srcu_struct(hctx->queue_rq_srcu);
2128 blk_mq_remove_cpuhp(hctx);
2129 blk_free_flush_queue(hctx->fq);
2130 sbitmap_free(&hctx->ctx_map);
2133 static void blk_mq_exit_hw_queues(struct request_queue *q,
2134 struct blk_mq_tag_set *set, int nr_queue)
2136 struct blk_mq_hw_ctx *hctx;
2139 queue_for_each_hw_ctx(q, hctx, i) {
2142 blk_mq_exit_hctx(q, set, hctx, i);
2146 static int blk_mq_init_hctx(struct request_queue *q,
2147 struct blk_mq_tag_set *set,
2148 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2152 node = hctx->numa_node;
2153 if (node == NUMA_NO_NODE)
2154 node = hctx->numa_node = set->numa_node;
2156 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2157 spin_lock_init(&hctx->lock);
2158 INIT_LIST_HEAD(&hctx->dispatch);
2160 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2162 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2164 hctx->tags = set->tags[hctx_idx];
2167 * Allocate space for all possible cpus to avoid allocation at
2170 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2173 goto unregister_cpu_notifier;
2175 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2181 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2182 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2184 if (set->ops->init_hctx &&
2185 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2188 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2191 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2193 goto sched_exit_hctx;
2195 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2198 if (hctx->flags & BLK_MQ_F_BLOCKING)
2199 init_srcu_struct(hctx->queue_rq_srcu);
2201 blk_mq_debugfs_register_hctx(q, hctx);
2208 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2210 if (set->ops->exit_hctx)
2211 set->ops->exit_hctx(hctx, hctx_idx);
2213 sbitmap_free(&hctx->ctx_map);
2216 unregister_cpu_notifier:
2217 blk_mq_remove_cpuhp(hctx);
2221 static void blk_mq_init_cpu_queues(struct request_queue *q,
2222 unsigned int nr_hw_queues)
2226 for_each_possible_cpu(i) {
2227 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2228 struct blk_mq_hw_ctx *hctx;
2231 spin_lock_init(&__ctx->lock);
2232 INIT_LIST_HEAD(&__ctx->rq_list);
2235 /* If the cpu isn't present, the cpu is mapped to first hctx */
2236 if (!cpu_present(i))
2239 hctx = blk_mq_map_queue(q, i);
2242 * Set local node, IFF we have more than one hw queue. If
2243 * not, we remain on the home node of the device
2245 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2246 hctx->numa_node = local_memory_node(cpu_to_node(i));
2250 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2254 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2255 set->queue_depth, set->reserved_tags);
2256 if (!set->tags[hctx_idx])
2259 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2264 blk_mq_free_rq_map(set->tags[hctx_idx]);
2265 set->tags[hctx_idx] = NULL;
2269 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2270 unsigned int hctx_idx)
2272 if (set->tags[hctx_idx]) {
2273 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2274 blk_mq_free_rq_map(set->tags[hctx_idx]);
2275 set->tags[hctx_idx] = NULL;
2279 static void blk_mq_map_swqueue(struct request_queue *q)
2281 unsigned int i, hctx_idx;
2282 struct blk_mq_hw_ctx *hctx;
2283 struct blk_mq_ctx *ctx;
2284 struct blk_mq_tag_set *set = q->tag_set;
2287 * Avoid others reading imcomplete hctx->cpumask through sysfs
2289 mutex_lock(&q->sysfs_lock);
2291 queue_for_each_hw_ctx(q, hctx, i) {
2292 cpumask_clear(hctx->cpumask);
2297 * Map software to hardware queues.
2299 * If the cpu isn't present, the cpu is mapped to first hctx.
2301 for_each_present_cpu(i) {
2302 hctx_idx = q->mq_map[i];
2303 /* unmapped hw queue can be remapped after CPU topo changed */
2304 if (!set->tags[hctx_idx] &&
2305 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2307 * If tags initialization fail for some hctx,
2308 * that hctx won't be brought online. In this
2309 * case, remap the current ctx to hctx[0] which
2310 * is guaranteed to always have tags allocated
2315 ctx = per_cpu_ptr(q->queue_ctx, i);
2316 hctx = blk_mq_map_queue(q, i);
2318 cpumask_set_cpu(i, hctx->cpumask);
2319 ctx->index_hw = hctx->nr_ctx;
2320 hctx->ctxs[hctx->nr_ctx++] = ctx;
2323 mutex_unlock(&q->sysfs_lock);
2325 queue_for_each_hw_ctx(q, hctx, i) {
2327 * If no software queues are mapped to this hardware queue,
2328 * disable it and free the request entries.
2330 if (!hctx->nr_ctx) {
2331 /* Never unmap queue 0. We need it as a
2332 * fallback in case of a new remap fails
2335 if (i && set->tags[i])
2336 blk_mq_free_map_and_requests(set, i);
2342 hctx->tags = set->tags[i];
2343 WARN_ON(!hctx->tags);
2346 * Set the map size to the number of mapped software queues.
2347 * This is more accurate and more efficient than looping
2348 * over all possibly mapped software queues.
2350 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2353 * Initialize batch roundrobin counts
2355 hctx->next_cpu = cpumask_first(hctx->cpumask);
2356 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2361 * Caller needs to ensure that we're either frozen/quiesced, or that
2362 * the queue isn't live yet.
2364 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2366 struct blk_mq_hw_ctx *hctx;
2369 queue_for_each_hw_ctx(q, hctx, i) {
2371 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2372 atomic_inc(&q->shared_hctx_restart);
2373 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2375 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2376 atomic_dec(&q->shared_hctx_restart);
2377 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2382 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2385 struct request_queue *q;
2387 lockdep_assert_held(&set->tag_list_lock);
2389 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2390 blk_mq_freeze_queue(q);
2391 queue_set_hctx_shared(q, shared);
2392 blk_mq_unfreeze_queue(q);
2396 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2398 struct blk_mq_tag_set *set = q->tag_set;
2400 mutex_lock(&set->tag_list_lock);
2401 list_del_rcu(&q->tag_set_list);
2402 INIT_LIST_HEAD(&q->tag_set_list);
2403 if (list_is_singular(&set->tag_list)) {
2404 /* just transitioned to unshared */
2405 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2406 /* update existing queue */
2407 blk_mq_update_tag_set_depth(set, false);
2409 mutex_unlock(&set->tag_list_lock);
2414 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2415 struct request_queue *q)
2419 mutex_lock(&set->tag_list_lock);
2422 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2424 if (!list_empty(&set->tag_list) &&
2425 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2426 set->flags |= BLK_MQ_F_TAG_SHARED;
2427 /* update existing queue */
2428 blk_mq_update_tag_set_depth(set, true);
2430 if (set->flags & BLK_MQ_F_TAG_SHARED)
2431 queue_set_hctx_shared(q, true);
2432 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2434 mutex_unlock(&set->tag_list_lock);
2438 * It is the actual release handler for mq, but we do it from
2439 * request queue's release handler for avoiding use-after-free
2440 * and headache because q->mq_kobj shouldn't have been introduced,
2441 * but we can't group ctx/kctx kobj without it.
2443 void blk_mq_release(struct request_queue *q)
2445 struct blk_mq_hw_ctx *hctx;
2448 /* hctx kobj stays in hctx */
2449 queue_for_each_hw_ctx(q, hctx, i) {
2452 kobject_put(&hctx->kobj);
2457 kfree(q->queue_hw_ctx);
2460 * release .mq_kobj and sw queue's kobject now because
2461 * both share lifetime with request queue.
2463 blk_mq_sysfs_deinit(q);
2465 free_percpu(q->queue_ctx);
2468 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2470 struct request_queue *uninit_q, *q;
2472 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2474 return ERR_PTR(-ENOMEM);
2476 q = blk_mq_init_allocated_queue(set, uninit_q);
2478 blk_cleanup_queue(uninit_q);
2482 EXPORT_SYMBOL(blk_mq_init_queue);
2484 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2486 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2488 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2489 __alignof__(struct blk_mq_hw_ctx)) !=
2490 sizeof(struct blk_mq_hw_ctx));
2492 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2493 hw_ctx_size += sizeof(struct srcu_struct);
2498 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2499 struct request_queue *q)
2502 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2504 blk_mq_sysfs_unregister(q);
2506 /* protect against switching io scheduler */
2507 mutex_lock(&q->sysfs_lock);
2508 for (i = 0; i < set->nr_hw_queues; i++) {
2514 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2515 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2520 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2527 atomic_set(&hctxs[i]->nr_active, 0);
2528 hctxs[i]->numa_node = node;
2529 hctxs[i]->queue_num = i;
2531 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2532 free_cpumask_var(hctxs[i]->cpumask);
2537 blk_mq_hctx_kobj_init(hctxs[i]);
2539 for (j = i; j < q->nr_hw_queues; j++) {
2540 struct blk_mq_hw_ctx *hctx = hctxs[j];
2544 blk_mq_free_map_and_requests(set, j);
2545 blk_mq_exit_hctx(q, set, hctx, j);
2546 kobject_put(&hctx->kobj);
2551 q->nr_hw_queues = i;
2552 mutex_unlock(&q->sysfs_lock);
2553 blk_mq_sysfs_register(q);
2556 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2557 struct request_queue *q)
2559 /* mark the queue as mq asap */
2560 q->mq_ops = set->ops;
2562 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2563 blk_mq_poll_stats_bkt,
2564 BLK_MQ_POLL_STATS_BKTS, q);
2568 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2572 /* init q->mq_kobj and sw queues' kobjects */
2573 blk_mq_sysfs_init(q);
2575 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2576 GFP_KERNEL, set->numa_node);
2577 if (!q->queue_hw_ctx)
2580 q->mq_map = set->mq_map;
2582 blk_mq_realloc_hw_ctxs(set, q);
2583 if (!q->nr_hw_queues)
2586 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2587 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2589 q->nr_queues = nr_cpu_ids;
2591 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2593 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2594 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2596 q->sg_reserved_size = INT_MAX;
2598 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2599 INIT_LIST_HEAD(&q->requeue_list);
2600 spin_lock_init(&q->requeue_lock);
2602 blk_queue_make_request(q, blk_mq_make_request);
2603 if (q->mq_ops->poll)
2604 q->poll_fn = blk_mq_poll;
2607 * Do this after blk_queue_make_request() overrides it...
2609 q->nr_requests = set->queue_depth;
2612 * Default to classic polling
2616 if (set->ops->complete)
2617 blk_queue_softirq_done(q, set->ops->complete);
2619 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2620 blk_mq_add_queue_tag_set(set, q);
2621 blk_mq_map_swqueue(q);
2623 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2626 ret = blk_mq_sched_init(q);
2628 return ERR_PTR(ret);
2634 kfree(q->queue_hw_ctx);
2636 free_percpu(q->queue_ctx);
2639 return ERR_PTR(-ENOMEM);
2641 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2643 void blk_mq_free_queue(struct request_queue *q)
2645 struct blk_mq_tag_set *set = q->tag_set;
2647 blk_mq_del_queue_tag_set(q);
2648 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2651 /* Basically redo blk_mq_init_queue with queue frozen */
2652 static void blk_mq_queue_reinit(struct request_queue *q)
2654 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2656 blk_mq_debugfs_unregister_hctxs(q);
2657 blk_mq_sysfs_unregister(q);
2660 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2661 * we should change hctx numa_node according to the new topology (this
2662 * involves freeing and re-allocating memory, worth doing?)
2664 blk_mq_map_swqueue(q);
2666 blk_mq_sysfs_register(q);
2667 blk_mq_debugfs_register_hctxs(q);
2670 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2674 for (i = 0; i < set->nr_hw_queues; i++)
2675 if (!__blk_mq_alloc_rq_map(set, i))
2682 blk_mq_free_rq_map(set->tags[i]);
2688 * Allocate the request maps associated with this tag_set. Note that this
2689 * may reduce the depth asked for, if memory is tight. set->queue_depth
2690 * will be updated to reflect the allocated depth.
2692 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2697 depth = set->queue_depth;
2699 err = __blk_mq_alloc_rq_maps(set);
2703 set->queue_depth >>= 1;
2704 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2708 } while (set->queue_depth);
2710 if (!set->queue_depth || err) {
2711 pr_err("blk-mq: failed to allocate request map\n");
2715 if (depth != set->queue_depth)
2716 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2717 depth, set->queue_depth);
2722 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2724 if (set->ops->map_queues) {
2727 * transport .map_queues is usually done in the following
2730 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2731 * mask = get_cpu_mask(queue)
2732 * for_each_cpu(cpu, mask)
2733 * set->mq_map[cpu] = queue;
2736 * When we need to remap, the table has to be cleared for
2737 * killing stale mapping since one CPU may not be mapped
2740 for_each_possible_cpu(cpu)
2741 set->mq_map[cpu] = 0;
2743 return set->ops->map_queues(set);
2745 return blk_mq_map_queues(set);
2749 * Alloc a tag set to be associated with one or more request queues.
2750 * May fail with EINVAL for various error conditions. May adjust the
2751 * requested depth down, if if it too large. In that case, the set
2752 * value will be stored in set->queue_depth.
2754 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2758 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2760 if (!set->nr_hw_queues)
2762 if (!set->queue_depth)
2764 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2767 if (!set->ops->queue_rq)
2770 if (!set->ops->get_budget ^ !set->ops->put_budget)
2773 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2774 pr_info("blk-mq: reduced tag depth to %u\n",
2776 set->queue_depth = BLK_MQ_MAX_DEPTH;
2780 * If a crashdump is active, then we are potentially in a very
2781 * memory constrained environment. Limit us to 1 queue and
2782 * 64 tags to prevent using too much memory.
2784 if (is_kdump_kernel()) {
2785 set->nr_hw_queues = 1;
2786 set->queue_depth = min(64U, set->queue_depth);
2789 * There is no use for more h/w queues than cpus.
2791 if (set->nr_hw_queues > nr_cpu_ids)
2792 set->nr_hw_queues = nr_cpu_ids;
2794 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2795 GFP_KERNEL, set->numa_node);
2800 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2801 GFP_KERNEL, set->numa_node);
2805 ret = blk_mq_update_queue_map(set);
2807 goto out_free_mq_map;
2809 ret = blk_mq_alloc_rq_maps(set);
2811 goto out_free_mq_map;
2813 mutex_init(&set->tag_list_lock);
2814 INIT_LIST_HEAD(&set->tag_list);
2826 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2828 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2832 for (i = 0; i < nr_cpu_ids; i++)
2833 blk_mq_free_map_and_requests(set, i);
2841 EXPORT_SYMBOL(blk_mq_free_tag_set);
2843 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2845 struct blk_mq_tag_set *set = q->tag_set;
2846 struct blk_mq_hw_ctx *hctx;
2852 blk_mq_freeze_queue(q);
2853 blk_mq_quiesce_queue(q);
2856 queue_for_each_hw_ctx(q, hctx, i) {
2860 * If we're using an MQ scheduler, just update the scheduler
2861 * queue depth. This is similar to what the old code would do.
2863 if (!hctx->sched_tags) {
2864 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2867 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2875 q->nr_requests = nr;
2877 blk_mq_unquiesce_queue(q);
2878 blk_mq_unfreeze_queue(q);
2883 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2886 struct request_queue *q;
2888 lockdep_assert_held(&set->tag_list_lock);
2890 if (nr_hw_queues > nr_cpu_ids)
2891 nr_hw_queues = nr_cpu_ids;
2892 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2895 list_for_each_entry(q, &set->tag_list, tag_set_list)
2896 blk_mq_freeze_queue(q);
2898 set->nr_hw_queues = nr_hw_queues;
2899 blk_mq_update_queue_map(set);
2900 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2901 blk_mq_realloc_hw_ctxs(set, q);
2902 blk_mq_queue_reinit(q);
2905 list_for_each_entry(q, &set->tag_list, tag_set_list)
2906 blk_mq_unfreeze_queue(q);
2909 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2911 mutex_lock(&set->tag_list_lock);
2912 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2913 mutex_unlock(&set->tag_list_lock);
2915 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2917 /* Enable polling stats and return whether they were already enabled. */
2918 static bool blk_poll_stats_enable(struct request_queue *q)
2920 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2921 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2923 blk_stat_add_callback(q, q->poll_cb);
2927 static void blk_mq_poll_stats_start(struct request_queue *q)
2930 * We don't arm the callback if polling stats are not enabled or the
2931 * callback is already active.
2933 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2934 blk_stat_is_active(q->poll_cb))
2937 blk_stat_activate_msecs(q->poll_cb, 100);
2940 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2942 struct request_queue *q = cb->data;
2945 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2946 if (cb->stat[bucket].nr_samples)
2947 q->poll_stat[bucket] = cb->stat[bucket];
2951 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2952 struct blk_mq_hw_ctx *hctx,
2955 unsigned long ret = 0;
2959 * If stats collection isn't on, don't sleep but turn it on for
2962 if (!blk_poll_stats_enable(q))
2966 * As an optimistic guess, use half of the mean service time
2967 * for this type of request. We can (and should) make this smarter.
2968 * For instance, if the completion latencies are tight, we can
2969 * get closer than just half the mean. This is especially
2970 * important on devices where the completion latencies are longer
2971 * than ~10 usec. We do use the stats for the relevant IO size
2972 * if available which does lead to better estimates.
2974 bucket = blk_mq_poll_stats_bkt(rq);
2978 if (q->poll_stat[bucket].nr_samples)
2979 ret = (q->poll_stat[bucket].mean + 1) / 2;
2984 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2985 struct blk_mq_hw_ctx *hctx,
2988 struct hrtimer_sleeper hs;
2989 enum hrtimer_mode mode;
2993 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2999 * -1: don't ever hybrid sleep
3000 * 0: use half of prev avg
3001 * >0: use this specific value
3003 if (q->poll_nsec == -1)
3005 else if (q->poll_nsec > 0)
3006 nsecs = q->poll_nsec;
3008 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3013 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
3016 * This will be replaced with the stats tracking code, using
3017 * 'avg_completion_time / 2' as the pre-sleep target.
3021 mode = HRTIMER_MODE_REL;
3022 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3023 hrtimer_set_expires(&hs.timer, kt);
3025 hrtimer_init_sleeper(&hs, current);
3027 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
3029 set_current_state(TASK_UNINTERRUPTIBLE);
3030 hrtimer_start_expires(&hs.timer, mode);
3033 hrtimer_cancel(&hs.timer);
3034 mode = HRTIMER_MODE_ABS;
3035 } while (hs.task && !signal_pending(current));
3037 __set_current_state(TASK_RUNNING);
3038 destroy_hrtimer_on_stack(&hs.timer);
3042 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3044 struct request_queue *q = hctx->queue;
3048 * If we sleep, have the caller restart the poll loop to reset
3049 * the state. Like for the other success return cases, the
3050 * caller is responsible for checking if the IO completed. If
3051 * the IO isn't complete, we'll get called again and will go
3052 * straight to the busy poll loop.
3054 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3057 hctx->poll_considered++;
3059 state = current->state;
3060 while (!need_resched()) {
3063 hctx->poll_invoked++;
3065 ret = q->mq_ops->poll(hctx, rq->tag);
3067 hctx->poll_success++;
3068 set_current_state(TASK_RUNNING);
3072 if (signal_pending_state(state, current))
3073 set_current_state(TASK_RUNNING);
3075 if (current->state == TASK_RUNNING)
3085 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3087 struct blk_mq_hw_ctx *hctx;
3090 if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3093 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3094 if (!blk_qc_t_is_internal(cookie))
3095 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3097 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3099 * With scheduling, if the request has completed, we'll
3100 * get a NULL return here, as we clear the sched tag when
3101 * that happens. The request still remains valid, like always,
3102 * so we should be safe with just the NULL check.
3108 return __blk_mq_poll(hctx, rq);
3111 static int __init blk_mq_init(void)
3113 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3114 blk_mq_hctx_notify_dead);
3117 subsys_initcall(blk_mq_init);