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 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
66 return sbitmap_any_bit_set(&hctx->ctx_map) ||
67 !list_empty_careful(&hctx->dispatch) ||
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);
129 blk_mq_run_hw_queues(q, false);
132 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
134 void blk_mq_freeze_queue_wait(struct request_queue *q)
136 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
138 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
140 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
141 unsigned long timeout)
143 return wait_event_timeout(q->mq_freeze_wq,
144 percpu_ref_is_zero(&q->q_usage_counter),
147 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
150 * Guarantee no request is in use, so we can change any data structure of
151 * the queue afterward.
153 void blk_freeze_queue(struct request_queue *q)
156 * In the !blk_mq case we are only calling this to kill the
157 * q_usage_counter, otherwise this increases the freeze depth
158 * and waits for it to return to zero. For this reason there is
159 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
160 * exported to drivers as the only user for unfreeze is blk_mq.
162 blk_freeze_queue_start(q);
163 blk_mq_freeze_queue_wait(q);
166 void blk_mq_freeze_queue(struct request_queue *q)
169 * ...just an alias to keep freeze and unfreeze actions balanced
170 * in the blk_mq_* namespace
174 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
176 void blk_mq_unfreeze_queue(struct request_queue *q)
180 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
181 WARN_ON_ONCE(freeze_depth < 0);
183 percpu_ref_reinit(&q->q_usage_counter);
184 wake_up_all(&q->mq_freeze_wq);
187 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
190 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
191 * mpt3sas driver such that this function can be removed.
193 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
197 spin_lock_irqsave(q->queue_lock, flags);
198 queue_flag_set(QUEUE_FLAG_QUIESCED, q);
199 spin_unlock_irqrestore(q->queue_lock, flags);
201 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
204 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
207 * Note: this function does not prevent that the struct request end_io()
208 * callback function is invoked. Once this function is returned, we make
209 * sure no dispatch can happen until the queue is unquiesced via
210 * blk_mq_unquiesce_queue().
212 void blk_mq_quiesce_queue(struct request_queue *q)
214 struct blk_mq_hw_ctx *hctx;
218 blk_mq_quiesce_queue_nowait(q);
220 queue_for_each_hw_ctx(q, hctx, i) {
221 if (hctx->flags & BLK_MQ_F_BLOCKING)
222 synchronize_srcu(hctx->queue_rq_srcu);
229 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
232 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
235 * This function recovers queue into the state before quiescing
236 * which is done by blk_mq_quiesce_queue.
238 void blk_mq_unquiesce_queue(struct request_queue *q)
242 spin_lock_irqsave(q->queue_lock, flags);
243 queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
244 spin_unlock_irqrestore(q->queue_lock, flags);
246 /* dispatch requests which are inserted during quiescing */
247 blk_mq_run_hw_queues(q, true);
249 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
251 void blk_mq_wake_waiters(struct request_queue *q)
253 struct blk_mq_hw_ctx *hctx;
256 queue_for_each_hw_ctx(q, hctx, i)
257 if (blk_mq_hw_queue_mapped(hctx))
258 blk_mq_tag_wakeup_all(hctx->tags, true);
261 * If we are called because the queue has now been marked as
262 * dying, we need to ensure that processes currently waiting on
263 * the queue are notified as well.
265 wake_up_all(&q->mq_freeze_wq);
268 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
270 return blk_mq_has_free_tags(hctx->tags);
272 EXPORT_SYMBOL(blk_mq_can_queue);
274 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
275 unsigned int tag, unsigned int op)
277 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
278 struct request *rq = tags->static_rqs[tag];
282 if (data->flags & BLK_MQ_REQ_INTERNAL) {
284 rq->internal_tag = tag;
286 if (blk_mq_tag_busy(data->hctx)) {
287 rq->rq_flags = RQF_MQ_INFLIGHT;
288 atomic_inc(&data->hctx->nr_active);
291 rq->internal_tag = -1;
292 data->hctx->tags->rqs[rq->tag] = rq;
295 INIT_LIST_HEAD(&rq->queuelist);
296 /* csd/requeue_work/fifo_time is initialized before use */
298 rq->mq_ctx = data->ctx;
300 if (blk_queue_io_stat(data->q))
301 rq->rq_flags |= RQF_IO_STAT;
302 /* do not touch atomic flags, it needs atomic ops against the timer */
304 INIT_HLIST_NODE(&rq->hash);
305 RB_CLEAR_NODE(&rq->rb_node);
308 rq->start_time = jiffies;
309 #ifdef CONFIG_BLK_CGROUP
311 set_start_time_ns(rq);
312 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 */
322 INIT_LIST_HEAD(&rq->timeout_list);
326 rq->end_io_data = NULL;
329 data->ctx->rq_dispatched[op_is_sync(op)]++;
333 static struct request *blk_mq_get_request(struct request_queue *q,
334 struct bio *bio, unsigned int op,
335 struct blk_mq_alloc_data *data)
337 struct elevator_queue *e = q->elevator;
340 bool put_ctx_on_error = false;
342 blk_queue_enter_live(q);
344 if (likely(!data->ctx)) {
345 data->ctx = blk_mq_get_ctx(q);
346 put_ctx_on_error = true;
348 if (likely(!data->hctx))
349 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
351 data->flags |= BLK_MQ_REQ_NOWAIT;
354 data->flags |= BLK_MQ_REQ_INTERNAL;
357 * Flush requests are special and go directly to the
360 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
361 e->type->ops.mq.limit_depth(op, data);
364 tag = blk_mq_get_tag(data);
365 if (tag == BLK_MQ_TAG_FAIL) {
366 if (put_ctx_on_error) {
367 blk_mq_put_ctx(data->ctx);
374 rq = blk_mq_rq_ctx_init(data, tag, op);
375 if (!op_is_flush(op)) {
377 if (e && e->type->ops.mq.prepare_request) {
378 if (e->type->icq_cache && rq_ioc(bio))
379 blk_mq_sched_assign_ioc(rq, bio);
381 e->type->ops.mq.prepare_request(rq, bio);
382 rq->rq_flags |= RQF_ELVPRIV;
385 data->hctx->queued++;
389 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
392 struct blk_mq_alloc_data alloc_data = { .flags = flags };
396 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
400 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
404 return ERR_PTR(-EWOULDBLOCK);
406 blk_mq_put_ctx(alloc_data.ctx);
409 rq->__sector = (sector_t) -1;
410 rq->bio = rq->biotail = NULL;
413 EXPORT_SYMBOL(blk_mq_alloc_request);
415 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
416 unsigned int op, unsigned int flags, unsigned int hctx_idx)
418 struct blk_mq_alloc_data alloc_data = { .flags = flags };
424 * If the tag allocator sleeps we could get an allocation for a
425 * different hardware context. No need to complicate the low level
426 * allocator for this for the rare use case of a command tied to
429 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
430 return ERR_PTR(-EINVAL);
432 if (hctx_idx >= q->nr_hw_queues)
433 return ERR_PTR(-EIO);
435 ret = blk_queue_enter(q, true);
440 * Check if the hardware context is actually mapped to anything.
441 * If not tell the caller that it should skip this queue.
443 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
444 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
446 return ERR_PTR(-EXDEV);
448 cpu = cpumask_first(alloc_data.hctx->cpumask);
449 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
451 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
455 return ERR_PTR(-EWOULDBLOCK);
459 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
461 void blk_mq_free_request(struct request *rq)
463 struct request_queue *q = rq->q;
464 struct elevator_queue *e = q->elevator;
465 struct blk_mq_ctx *ctx = rq->mq_ctx;
466 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
467 const int sched_tag = rq->internal_tag;
469 if (rq->rq_flags & RQF_ELVPRIV) {
470 if (e && e->type->ops.mq.finish_request)
471 e->type->ops.mq.finish_request(rq);
473 put_io_context(rq->elv.icq->ioc);
478 ctx->rq_completed[rq_is_sync(rq)]++;
479 if (rq->rq_flags & RQF_MQ_INFLIGHT)
480 atomic_dec(&hctx->nr_active);
482 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
483 laptop_io_completion(q->backing_dev_info);
485 wbt_done(q->rq_wb, &rq->issue_stat);
488 blk_put_rl(blk_rq_rl(rq));
490 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
491 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
493 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
495 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
496 blk_mq_sched_restart(hctx);
499 EXPORT_SYMBOL_GPL(blk_mq_free_request);
501 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
503 blk_account_io_done(rq);
506 wbt_done(rq->q->rq_wb, &rq->issue_stat);
507 rq->end_io(rq, error);
509 if (unlikely(blk_bidi_rq(rq)))
510 blk_mq_free_request(rq->next_rq);
511 blk_mq_free_request(rq);
514 EXPORT_SYMBOL(__blk_mq_end_request);
516 void blk_mq_end_request(struct request *rq, blk_status_t error)
518 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
520 __blk_mq_end_request(rq, error);
522 EXPORT_SYMBOL(blk_mq_end_request);
524 static void __blk_mq_complete_request_remote(void *data)
526 struct request *rq = data;
528 rq->q->softirq_done_fn(rq);
531 static void __blk_mq_complete_request(struct request *rq)
533 struct blk_mq_ctx *ctx = rq->mq_ctx;
537 if (rq->internal_tag != -1)
538 blk_mq_sched_completed_request(rq);
539 if (rq->rq_flags & RQF_STATS) {
540 blk_mq_poll_stats_start(rq->q);
544 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
545 rq->q->softirq_done_fn(rq);
550 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
551 shared = cpus_share_cache(cpu, ctx->cpu);
553 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
554 rq->csd.func = __blk_mq_complete_request_remote;
557 smp_call_function_single_async(ctx->cpu, &rq->csd);
559 rq->q->softirq_done_fn(rq);
565 * blk_mq_complete_request - end I/O on a request
566 * @rq: the request being processed
569 * Ends all I/O on a request. It does not handle partial completions.
570 * The actual completion happens out-of-order, through a IPI handler.
572 void blk_mq_complete_request(struct request *rq)
574 struct request_queue *q = rq->q;
576 if (unlikely(blk_should_fake_timeout(q)))
578 if (!blk_mark_rq_complete(rq))
579 __blk_mq_complete_request(rq);
581 EXPORT_SYMBOL(blk_mq_complete_request);
583 int blk_mq_request_started(struct request *rq)
585 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
587 EXPORT_SYMBOL_GPL(blk_mq_request_started);
589 void blk_mq_start_request(struct request *rq)
591 struct request_queue *q = rq->q;
593 blk_mq_sched_started_request(rq);
595 trace_block_rq_issue(q, rq);
597 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
598 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
599 rq->rq_flags |= RQF_STATS;
600 wbt_issue(q->rq_wb, &rq->issue_stat);
605 WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags));
608 * Mark us as started and clear complete. Complete might have been
609 * set if requeue raced with timeout, which then marked it as
610 * complete. So be sure to clear complete again when we start
611 * the request, otherwise we'll ignore the completion event.
613 * Ensure that ->deadline is visible before we set STARTED, such that
614 * blk_mq_check_expired() is guaranteed to observe our ->deadline when
615 * it observes STARTED.
618 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
619 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) {
621 * Coherence order guarantees these consecutive stores to a
622 * single variable propagate in the specified order. Thus the
623 * clear_bit() is ordered _after_ the set bit. See
624 * blk_mq_check_expired().
626 * (the bits must be part of the same byte for this to be
629 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
632 if (q->dma_drain_size && blk_rq_bytes(rq)) {
634 * Make sure space for the drain appears. We know we can do
635 * this because max_hw_segments has been adjusted to be one
636 * fewer than the device can handle.
638 rq->nr_phys_segments++;
641 EXPORT_SYMBOL(blk_mq_start_request);
644 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
645 * flag isn't set yet, so there may be race with timeout handler,
646 * but given rq->deadline is just set in .queue_rq() under
647 * this situation, the race won't be possible in reality because
648 * rq->timeout should be set as big enough to cover the window
649 * between blk_mq_start_request() called from .queue_rq() and
650 * clearing REQ_ATOM_STARTED here.
652 static void __blk_mq_requeue_request(struct request *rq)
654 struct request_queue *q = rq->q;
656 trace_block_rq_requeue(q, rq);
657 wbt_requeue(q->rq_wb, &rq->issue_stat);
658 blk_mq_sched_requeue_request(rq);
660 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
661 if (q->dma_drain_size && blk_rq_bytes(rq))
662 rq->nr_phys_segments--;
666 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
668 __blk_mq_requeue_request(rq);
670 BUG_ON(blk_queued_rq(rq));
671 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
673 EXPORT_SYMBOL(blk_mq_requeue_request);
675 static void blk_mq_requeue_work(struct work_struct *work)
677 struct request_queue *q =
678 container_of(work, struct request_queue, requeue_work.work);
680 struct request *rq, *next;
682 spin_lock_irq(&q->requeue_lock);
683 list_splice_init(&q->requeue_list, &rq_list);
684 spin_unlock_irq(&q->requeue_lock);
686 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
687 if (!(rq->rq_flags & RQF_SOFTBARRIER))
690 rq->rq_flags &= ~RQF_SOFTBARRIER;
691 list_del_init(&rq->queuelist);
692 blk_mq_sched_insert_request(rq, true, false, false, true);
695 while (!list_empty(&rq_list)) {
696 rq = list_entry(rq_list.next, struct request, queuelist);
697 list_del_init(&rq->queuelist);
698 blk_mq_sched_insert_request(rq, false, false, false, true);
701 blk_mq_run_hw_queues(q, false);
704 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
705 bool kick_requeue_list)
707 struct request_queue *q = rq->q;
711 * We abuse this flag that is otherwise used by the I/O scheduler to
712 * request head insertation from the workqueue.
714 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
716 spin_lock_irqsave(&q->requeue_lock, flags);
718 rq->rq_flags |= RQF_SOFTBARRIER;
719 list_add(&rq->queuelist, &q->requeue_list);
721 list_add_tail(&rq->queuelist, &q->requeue_list);
723 spin_unlock_irqrestore(&q->requeue_lock, flags);
725 if (kick_requeue_list)
726 blk_mq_kick_requeue_list(q);
728 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
730 void blk_mq_kick_requeue_list(struct request_queue *q)
732 kblockd_schedule_delayed_work(&q->requeue_work, 0);
734 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
736 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
739 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
740 msecs_to_jiffies(msecs));
742 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
744 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
746 if (tag < tags->nr_tags) {
747 prefetch(tags->rqs[tag]);
748 return tags->rqs[tag];
753 EXPORT_SYMBOL(blk_mq_tag_to_rq);
755 struct blk_mq_timeout_data {
757 unsigned int next_set;
760 void blk_mq_rq_timed_out(struct request *req, bool reserved)
762 const struct blk_mq_ops *ops = req->q->mq_ops;
763 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
766 * We know that complete is set at this point. If STARTED isn't set
767 * anymore, then the request isn't active and the "timeout" should
768 * just be ignored. This can happen due to the bitflag ordering.
769 * Timeout first checks if STARTED is set, and if it is, assumes
770 * the request is active. But if we race with completion, then
771 * both flags will get cleared. So check here again, and ignore
772 * a timeout event with a request that isn't active.
774 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
778 ret = ops->timeout(req, reserved);
782 __blk_mq_complete_request(req);
784 case BLK_EH_RESET_TIMER:
786 blk_clear_rq_complete(req);
788 case BLK_EH_NOT_HANDLED:
791 printk(KERN_ERR "block: bad eh return: %d\n", ret);
796 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
797 struct request *rq, void *priv, bool reserved)
799 struct blk_mq_timeout_data *data = priv;
800 unsigned long deadline;
802 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
806 * Ensures that if we see STARTED we must also see our
807 * up-to-date deadline, see blk_mq_start_request().
811 deadline = READ_ONCE(rq->deadline);
814 * The rq being checked may have been freed and reallocated
815 * out already here, we avoid this race by checking rq->deadline
816 * and REQ_ATOM_COMPLETE flag together:
818 * - if rq->deadline is observed as new value because of
819 * reusing, the rq won't be timed out because of timing.
820 * - if rq->deadline is observed as previous value,
821 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
822 * because we put a barrier between setting rq->deadline
823 * and clearing the flag in blk_mq_start_request(), so
824 * this rq won't be timed out too.
826 if (time_after_eq(jiffies, deadline)) {
827 if (!blk_mark_rq_complete(rq)) {
829 * Again coherence order ensures that consecutive reads
830 * from the same variable must be in that order. This
831 * ensures that if we see COMPLETE clear, we must then
832 * see STARTED set and we'll ignore this timeout.
834 * (There's also the MB implied by the test_and_clear())
836 blk_mq_rq_timed_out(rq, reserved);
838 } else if (!data->next_set || time_after(data->next, deadline)) {
839 data->next = deadline;
844 static void blk_mq_timeout_work(struct work_struct *work)
846 struct request_queue *q =
847 container_of(work, struct request_queue, timeout_work);
848 struct blk_mq_timeout_data data = {
854 /* A deadlock might occur if a request is stuck requiring a
855 * timeout at the same time a queue freeze is waiting
856 * completion, since the timeout code would not be able to
857 * acquire the queue reference here.
859 * That's why we don't use blk_queue_enter here; instead, we use
860 * percpu_ref_tryget directly, because we need to be able to
861 * obtain a reference even in the short window between the queue
862 * starting to freeze, by dropping the first reference in
863 * blk_freeze_queue_start, and the moment the last request is
864 * consumed, marked by the instant q_usage_counter reaches
867 if (!percpu_ref_tryget(&q->q_usage_counter))
870 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
873 data.next = blk_rq_timeout(round_jiffies_up(data.next));
874 mod_timer(&q->timeout, data.next);
876 struct blk_mq_hw_ctx *hctx;
878 queue_for_each_hw_ctx(q, hctx, i) {
879 /* the hctx may be unmapped, so check it here */
880 if (blk_mq_hw_queue_mapped(hctx))
881 blk_mq_tag_idle(hctx);
887 struct flush_busy_ctx_data {
888 struct blk_mq_hw_ctx *hctx;
889 struct list_head *list;
892 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
894 struct flush_busy_ctx_data *flush_data = data;
895 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
896 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
898 sbitmap_clear_bit(sb, bitnr);
899 spin_lock(&ctx->lock);
900 list_splice_tail_init(&ctx->rq_list, flush_data->list);
901 spin_unlock(&ctx->lock);
906 * Process software queues that have been marked busy, splicing them
907 * to the for-dispatch
909 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
911 struct flush_busy_ctx_data data = {
916 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
918 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
920 struct dispatch_rq_data {
921 struct blk_mq_hw_ctx *hctx;
925 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
928 struct dispatch_rq_data *dispatch_data = data;
929 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
930 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
932 spin_lock(&ctx->lock);
933 if (unlikely(!list_empty(&ctx->rq_list))) {
934 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
935 list_del_init(&dispatch_data->rq->queuelist);
936 if (list_empty(&ctx->rq_list))
937 sbitmap_clear_bit(sb, bitnr);
939 spin_unlock(&ctx->lock);
941 return !dispatch_data->rq;
944 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
945 struct blk_mq_ctx *start)
947 unsigned off = start ? start->index_hw : 0;
948 struct dispatch_rq_data data = {
953 __sbitmap_for_each_set(&hctx->ctx_map, off,
954 dispatch_rq_from_ctx, &data);
959 static inline unsigned int queued_to_index(unsigned int queued)
964 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
967 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
970 struct blk_mq_alloc_data data = {
972 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
973 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
976 might_sleep_if(wait);
981 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
982 data.flags |= BLK_MQ_REQ_RESERVED;
984 rq->tag = blk_mq_get_tag(&data);
986 if (blk_mq_tag_busy(data.hctx)) {
987 rq->rq_flags |= RQF_MQ_INFLIGHT;
988 atomic_inc(&data.hctx->nr_active);
990 data.hctx->tags->rqs[rq->tag] = rq;
996 return rq->tag != -1;
999 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
1002 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
1005 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
1006 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
1007 atomic_dec(&hctx->nr_active);
1011 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
1014 if (rq->tag == -1 || rq->internal_tag == -1)
1017 __blk_mq_put_driver_tag(hctx, rq);
1020 static void blk_mq_put_driver_tag(struct request *rq)
1022 struct blk_mq_hw_ctx *hctx;
1024 if (rq->tag == -1 || rq->internal_tag == -1)
1027 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1028 __blk_mq_put_driver_tag(hctx, rq);
1032 * If we fail getting a driver tag because all the driver tags are already
1033 * assigned and on the dispatch list, BUT the first entry does not have a
1034 * tag, then we could deadlock. For that case, move entries with assigned
1035 * driver tags to the front, leaving the set of tagged requests in the
1036 * same order, and the untagged set in the same order.
1038 static bool reorder_tags_to_front(struct list_head *list)
1040 struct request *rq, *tmp, *first = NULL;
1042 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
1045 if (rq->tag != -1) {
1046 list_move(&rq->queuelist, list);
1052 return first != NULL;
1055 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
1058 struct blk_mq_hw_ctx *hctx;
1060 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1062 list_del(&wait->entry);
1063 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
1064 blk_mq_run_hw_queue(hctx, true);
1068 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
1070 struct sbq_wait_state *ws;
1073 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
1074 * The thread which wins the race to grab this bit adds the hardware
1075 * queue to the wait queue.
1077 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
1078 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
1081 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
1082 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
1085 * As soon as this returns, it's no longer safe to fiddle with
1086 * hctx->dispatch_wait, since a completion can wake up the wait queue
1087 * and unlock the bit.
1089 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
1093 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1096 struct blk_mq_hw_ctx *hctx;
1100 if (list_empty(list))
1103 WARN_ON(!list_is_singular(list) && got_budget);
1106 * Now process all the entries, sending them to the driver.
1108 errors = queued = 0;
1110 struct blk_mq_queue_data bd;
1113 rq = list_first_entry(list, struct request, queuelist);
1114 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1115 if (!queued && reorder_tags_to_front(list))
1119 * The initial allocation attempt failed, so we need to
1120 * rerun the hardware queue when a tag is freed.
1122 if (!blk_mq_dispatch_wait_add(hctx)) {
1124 blk_mq_put_dispatch_budget(hctx);
1129 * It's possible that a tag was freed in the window
1130 * between the allocation failure and adding the
1131 * hardware queue to the wait queue.
1133 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1135 blk_mq_put_dispatch_budget(hctx);
1141 ret = blk_mq_get_dispatch_budget(hctx);
1142 if (ret == BLK_STS_RESOURCE)
1144 if (ret != BLK_STS_OK)
1148 list_del_init(&rq->queuelist);
1153 * Flag last if we have no more requests, or if we have more
1154 * but can't assign a driver tag to it.
1156 if (list_empty(list))
1159 struct request *nxt;
1161 nxt = list_first_entry(list, struct request, queuelist);
1162 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1165 ret = q->mq_ops->queue_rq(hctx, &bd);
1166 if (ret == BLK_STS_RESOURCE) {
1167 blk_mq_put_driver_tag_hctx(hctx, rq);
1168 list_add(&rq->queuelist, list);
1169 __blk_mq_requeue_request(rq);
1174 if (unlikely(ret != BLK_STS_OK)) {
1176 blk_mq_end_request(rq, BLK_STS_IOERR);
1181 } while (!list_empty(list));
1183 hctx->dispatched[queued_to_index(queued)]++;
1186 * Any items that need requeuing? Stuff them into hctx->dispatch,
1187 * that is where we will continue on next queue run.
1189 if (!list_empty(list)) {
1191 * If an I/O scheduler has been configured and we got a driver
1192 * tag for the next request already, free it again.
1194 rq = list_first_entry(list, struct request, queuelist);
1195 blk_mq_put_driver_tag(rq);
1197 spin_lock(&hctx->lock);
1198 list_splice_init(list, &hctx->dispatch);
1199 spin_unlock(&hctx->lock);
1202 * If SCHED_RESTART was set by the caller of this function and
1203 * it is no longer set that means that it was cleared by another
1204 * thread and hence that a queue rerun is needed.
1206 * If TAG_WAITING is set that means that an I/O scheduler has
1207 * been configured and another thread is waiting for a driver
1208 * tag. To guarantee fairness, do not rerun this hardware queue
1209 * but let the other thread grab the driver tag.
1211 * If no I/O scheduler has been configured it is possible that
1212 * the hardware queue got stopped and restarted before requests
1213 * were pushed back onto the dispatch list. Rerun the queue to
1214 * avoid starvation. Notes:
1215 * - blk_mq_run_hw_queue() checks whether or not a queue has
1216 * been stopped before rerunning a queue.
1217 * - Some but not all block drivers stop a queue before
1218 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1221 if (!blk_mq_sched_needs_restart(hctx) &&
1222 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1223 blk_mq_run_hw_queue(hctx, true);
1226 return (queued + errors) != 0;
1229 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1234 * We should be running this queue from one of the CPUs that
1237 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1238 cpu_online(hctx->next_cpu));
1241 * We can't run the queue inline with ints disabled. Ensure that
1242 * we catch bad users of this early.
1244 WARN_ON_ONCE(in_interrupt());
1246 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1248 blk_mq_sched_dispatch_requests(hctx);
1253 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1254 blk_mq_sched_dispatch_requests(hctx);
1255 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1260 * It'd be great if the workqueue API had a way to pass
1261 * in a mask and had some smarts for more clever placement.
1262 * For now we just round-robin here, switching for every
1263 * BLK_MQ_CPU_WORK_BATCH queued items.
1265 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1267 if (hctx->queue->nr_hw_queues == 1)
1268 return WORK_CPU_UNBOUND;
1270 if (--hctx->next_cpu_batch <= 0) {
1273 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1274 if (next_cpu >= nr_cpu_ids)
1275 next_cpu = cpumask_first(hctx->cpumask);
1277 hctx->next_cpu = next_cpu;
1278 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1281 return hctx->next_cpu;
1284 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1285 unsigned long msecs)
1287 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1290 if (unlikely(blk_mq_hctx_stopped(hctx)))
1293 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1294 int cpu = get_cpu();
1295 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1296 __blk_mq_run_hw_queue(hctx);
1304 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1306 msecs_to_jiffies(msecs));
1309 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1311 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1313 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1315 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1317 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1319 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1321 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1323 struct blk_mq_hw_ctx *hctx;
1326 queue_for_each_hw_ctx(q, hctx, i) {
1327 if (!blk_mq_hctx_has_pending(hctx) ||
1328 blk_mq_hctx_stopped(hctx))
1331 blk_mq_run_hw_queue(hctx, async);
1334 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1337 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1338 * @q: request queue.
1340 * The caller is responsible for serializing this function against
1341 * blk_mq_{start,stop}_hw_queue().
1343 bool blk_mq_queue_stopped(struct request_queue *q)
1345 struct blk_mq_hw_ctx *hctx;
1348 queue_for_each_hw_ctx(q, hctx, i)
1349 if (blk_mq_hctx_stopped(hctx))
1354 EXPORT_SYMBOL(blk_mq_queue_stopped);
1357 * This function is often used for pausing .queue_rq() by driver when
1358 * there isn't enough resource or some conditions aren't satisfied, and
1359 * BLK_STS_RESOURCE is usually returned.
1361 * We do not guarantee that dispatch can be drained or blocked
1362 * after blk_mq_stop_hw_queue() returns. Please use
1363 * blk_mq_quiesce_queue() for that requirement.
1365 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1367 cancel_delayed_work(&hctx->run_work);
1369 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1371 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1374 * This function is often used for pausing .queue_rq() by driver when
1375 * there isn't enough resource or some conditions aren't satisfied, and
1376 * BLK_STS_RESOURCE is usually returned.
1378 * We do not guarantee that dispatch can be drained or blocked
1379 * after blk_mq_stop_hw_queues() returns. Please use
1380 * blk_mq_quiesce_queue() for that requirement.
1382 void blk_mq_stop_hw_queues(struct request_queue *q)
1384 struct blk_mq_hw_ctx *hctx;
1387 queue_for_each_hw_ctx(q, hctx, i)
1388 blk_mq_stop_hw_queue(hctx);
1390 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1392 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1394 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1396 blk_mq_run_hw_queue(hctx, false);
1398 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1400 void blk_mq_start_hw_queues(struct request_queue *q)
1402 struct blk_mq_hw_ctx *hctx;
1405 queue_for_each_hw_ctx(q, hctx, i)
1406 blk_mq_start_hw_queue(hctx);
1408 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1410 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1412 if (!blk_mq_hctx_stopped(hctx))
1415 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1416 blk_mq_run_hw_queue(hctx, async);
1418 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1420 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1422 struct blk_mq_hw_ctx *hctx;
1425 queue_for_each_hw_ctx(q, hctx, i)
1426 blk_mq_start_stopped_hw_queue(hctx, async);
1428 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1430 static void blk_mq_run_work_fn(struct work_struct *work)
1432 struct blk_mq_hw_ctx *hctx;
1434 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1437 * If we are stopped, don't run the queue. The exception is if
1438 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1439 * the STOPPED bit and run it.
1441 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1442 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1445 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1446 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1449 __blk_mq_run_hw_queue(hctx);
1453 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1455 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1459 * Stop the hw queue, then modify currently delayed work.
1460 * This should prevent us from running the queue prematurely.
1461 * Mark the queue as auto-clearing STOPPED when it runs.
1463 blk_mq_stop_hw_queue(hctx);
1464 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1465 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1467 msecs_to_jiffies(msecs));
1469 EXPORT_SYMBOL(blk_mq_delay_queue);
1471 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1475 struct blk_mq_ctx *ctx = rq->mq_ctx;
1477 lockdep_assert_held(&ctx->lock);
1479 trace_block_rq_insert(hctx->queue, rq);
1482 list_add(&rq->queuelist, &ctx->rq_list);
1484 list_add_tail(&rq->queuelist, &ctx->rq_list);
1487 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1490 struct blk_mq_ctx *ctx = rq->mq_ctx;
1492 lockdep_assert_held(&ctx->lock);
1494 __blk_mq_insert_req_list(hctx, rq, at_head);
1495 blk_mq_hctx_mark_pending(hctx, ctx);
1499 * Should only be used carefully, when the caller knows we want to
1500 * bypass a potential IO scheduler on the target device.
1502 void blk_mq_request_bypass_insert(struct request *rq)
1504 struct blk_mq_ctx *ctx = rq->mq_ctx;
1505 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1507 spin_lock(&hctx->lock);
1508 list_add_tail(&rq->queuelist, &hctx->dispatch);
1509 spin_unlock(&hctx->lock);
1511 blk_mq_run_hw_queue(hctx, false);
1514 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1515 struct list_head *list)
1519 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1522 spin_lock(&ctx->lock);
1523 while (!list_empty(list)) {
1526 rq = list_first_entry(list, struct request, queuelist);
1527 BUG_ON(rq->mq_ctx != ctx);
1528 list_del_init(&rq->queuelist);
1529 __blk_mq_insert_req_list(hctx, rq, false);
1531 blk_mq_hctx_mark_pending(hctx, ctx);
1532 spin_unlock(&ctx->lock);
1535 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1537 struct request *rqa = container_of(a, struct request, queuelist);
1538 struct request *rqb = container_of(b, struct request, queuelist);
1540 return !(rqa->mq_ctx < rqb->mq_ctx ||
1541 (rqa->mq_ctx == rqb->mq_ctx &&
1542 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1545 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1547 struct blk_mq_ctx *this_ctx;
1548 struct request_queue *this_q;
1551 LIST_HEAD(ctx_list);
1554 list_splice_init(&plug->mq_list, &list);
1556 list_sort(NULL, &list, plug_ctx_cmp);
1562 while (!list_empty(&list)) {
1563 rq = list_entry_rq(list.next);
1564 list_del_init(&rq->queuelist);
1566 if (rq->mq_ctx != this_ctx) {
1568 trace_block_unplug(this_q, depth, from_schedule);
1569 blk_mq_sched_insert_requests(this_q, this_ctx,
1574 this_ctx = rq->mq_ctx;
1580 list_add_tail(&rq->queuelist, &ctx_list);
1584 * If 'this_ctx' is set, we know we have entries to complete
1585 * on 'ctx_list'. Do those.
1588 trace_block_unplug(this_q, depth, from_schedule);
1589 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1594 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1596 blk_init_request_from_bio(rq, bio);
1598 blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1600 blk_account_io_start(rq, true);
1603 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1604 struct blk_mq_ctx *ctx,
1607 spin_lock(&ctx->lock);
1608 __blk_mq_insert_request(hctx, rq, false);
1609 spin_unlock(&ctx->lock);
1612 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1615 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1617 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1620 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1622 blk_qc_t *cookie, bool may_sleep)
1624 struct request_queue *q = rq->q;
1625 struct blk_mq_queue_data bd = {
1629 blk_qc_t new_cookie;
1631 bool run_queue = true;
1633 /* RCU or SRCU read lock is needed before checking quiesced flag */
1634 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1642 if (!blk_mq_get_driver_tag(rq, NULL, false))
1645 ret = blk_mq_get_dispatch_budget(hctx);
1646 if (ret == BLK_STS_RESOURCE) {
1647 blk_mq_put_driver_tag(rq);
1649 } else if (ret != BLK_STS_OK)
1652 new_cookie = request_to_qc_t(hctx, rq);
1655 * For OK queue, we are done. For error, kill it. Any other
1656 * error (busy), just add it to our list as we previously
1659 ret = q->mq_ops->queue_rq(hctx, &bd);
1662 *cookie = new_cookie;
1664 case BLK_STS_RESOURCE:
1665 __blk_mq_requeue_request(rq);
1669 *cookie = BLK_QC_T_NONE;
1670 blk_mq_end_request(rq, ret);
1675 blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1678 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1679 struct request *rq, blk_qc_t *cookie)
1681 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1683 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1686 unsigned int srcu_idx;
1690 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1691 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1692 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1696 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1698 const int is_sync = op_is_sync(bio->bi_opf);
1699 const int is_flush_fua = op_is_flush(bio->bi_opf);
1700 struct blk_mq_alloc_data data = { .flags = 0 };
1702 unsigned int request_count = 0;
1703 struct blk_plug *plug;
1704 struct request *same_queue_rq = NULL;
1706 unsigned int wb_acct;
1708 blk_queue_bounce(q, &bio);
1710 blk_queue_split(q, &bio);
1712 if (!bio_integrity_prep(bio))
1713 return BLK_QC_T_NONE;
1715 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1716 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1717 return BLK_QC_T_NONE;
1719 if (blk_mq_sched_bio_merge(q, bio))
1720 return BLK_QC_T_NONE;
1722 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1724 trace_block_getrq(q, bio, bio->bi_opf);
1726 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1727 if (unlikely(!rq)) {
1728 __wbt_done(q->rq_wb, wb_acct);
1729 if (bio->bi_opf & REQ_NOWAIT)
1730 bio_wouldblock_error(bio);
1731 return BLK_QC_T_NONE;
1734 wbt_track(&rq->issue_stat, wb_acct);
1736 cookie = request_to_qc_t(data.hctx, rq);
1738 plug = current->plug;
1739 if (unlikely(is_flush_fua)) {
1740 blk_mq_put_ctx(data.ctx);
1741 blk_mq_bio_to_request(rq, bio);
1743 blk_mq_sched_insert_request(rq, false, true, true,
1746 blk_insert_flush(rq);
1747 blk_mq_run_hw_queue(data.hctx, true);
1749 } else if (plug && q->nr_hw_queues == 1) {
1750 struct request *last = NULL;
1752 blk_mq_put_ctx(data.ctx);
1753 blk_mq_bio_to_request(rq, bio);
1756 * @request_count may become stale because of schedule
1757 * out, so check the list again.
1759 if (list_empty(&plug->mq_list))
1761 else if (blk_queue_nomerges(q))
1762 request_count = blk_plug_queued_count(q);
1765 trace_block_plug(q);
1767 last = list_entry_rq(plug->mq_list.prev);
1769 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1770 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1771 blk_flush_plug_list(plug, false);
1772 trace_block_plug(q);
1775 list_add_tail(&rq->queuelist, &plug->mq_list);
1776 } else if (plug && !blk_queue_nomerges(q)) {
1777 blk_mq_bio_to_request(rq, bio);
1780 * We do limited plugging. If the bio can be merged, do that.
1781 * Otherwise the existing request in the plug list will be
1782 * issued. So the plug list will have one request at most
1783 * The plug list might get flushed before this. If that happens,
1784 * the plug list is empty, and same_queue_rq is invalid.
1786 if (list_empty(&plug->mq_list))
1787 same_queue_rq = NULL;
1789 list_del_init(&same_queue_rq->queuelist);
1790 list_add_tail(&rq->queuelist, &plug->mq_list);
1792 blk_mq_put_ctx(data.ctx);
1794 if (same_queue_rq) {
1795 data.hctx = blk_mq_map_queue(q,
1796 same_queue_rq->mq_ctx->cpu);
1797 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1800 } else if (q->nr_hw_queues > 1 && is_sync) {
1801 blk_mq_put_ctx(data.ctx);
1802 blk_mq_bio_to_request(rq, bio);
1803 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1804 } else if (q->elevator) {
1805 blk_mq_put_ctx(data.ctx);
1806 blk_mq_bio_to_request(rq, bio);
1807 blk_mq_sched_insert_request(rq, false, true, true, true);
1809 blk_mq_put_ctx(data.ctx);
1810 blk_mq_bio_to_request(rq, bio);
1811 blk_mq_queue_io(data.hctx, data.ctx, rq);
1812 blk_mq_run_hw_queue(data.hctx, true);
1818 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1819 unsigned int hctx_idx)
1823 if (tags->rqs && set->ops->exit_request) {
1826 for (i = 0; i < tags->nr_tags; i++) {
1827 struct request *rq = tags->static_rqs[i];
1831 set->ops->exit_request(set, rq, hctx_idx);
1832 tags->static_rqs[i] = NULL;
1836 while (!list_empty(&tags->page_list)) {
1837 page = list_first_entry(&tags->page_list, struct page, lru);
1838 list_del_init(&page->lru);
1840 * Remove kmemleak object previously allocated in
1841 * blk_mq_init_rq_map().
1843 kmemleak_free(page_address(page));
1844 __free_pages(page, page->private);
1848 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1852 kfree(tags->static_rqs);
1853 tags->static_rqs = NULL;
1855 blk_mq_free_tags(tags);
1858 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1859 unsigned int hctx_idx,
1860 unsigned int nr_tags,
1861 unsigned int reserved_tags)
1863 struct blk_mq_tags *tags;
1866 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1867 if (node == NUMA_NO_NODE)
1868 node = set->numa_node;
1870 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1871 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1875 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1876 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1879 blk_mq_free_tags(tags);
1883 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1884 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1886 if (!tags->static_rqs) {
1888 blk_mq_free_tags(tags);
1895 static size_t order_to_size(unsigned int order)
1897 return (size_t)PAGE_SIZE << order;
1900 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1901 unsigned int hctx_idx, unsigned int depth)
1903 unsigned int i, j, entries_per_page, max_order = 4;
1904 size_t rq_size, left;
1907 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1908 if (node == NUMA_NO_NODE)
1909 node = set->numa_node;
1911 INIT_LIST_HEAD(&tags->page_list);
1914 * rq_size is the size of the request plus driver payload, rounded
1915 * to the cacheline size
1917 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1919 left = rq_size * depth;
1921 for (i = 0; i < depth; ) {
1922 int this_order = max_order;
1927 while (this_order && left < order_to_size(this_order - 1))
1931 page = alloc_pages_node(node,
1932 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1938 if (order_to_size(this_order) < rq_size)
1945 page->private = this_order;
1946 list_add_tail(&page->lru, &tags->page_list);
1948 p = page_address(page);
1950 * Allow kmemleak to scan these pages as they contain pointers
1951 * to additional allocations like via ops->init_request().
1953 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1954 entries_per_page = order_to_size(this_order) / rq_size;
1955 to_do = min(entries_per_page, depth - i);
1956 left -= to_do * rq_size;
1957 for (j = 0; j < to_do; j++) {
1958 struct request *rq = p;
1960 tags->static_rqs[i] = rq;
1961 if (set->ops->init_request) {
1962 if (set->ops->init_request(set, rq, hctx_idx,
1964 tags->static_rqs[i] = NULL;
1976 blk_mq_free_rqs(set, tags, hctx_idx);
1981 * 'cpu' is going away. splice any existing rq_list entries from this
1982 * software queue to the hw queue dispatch list, and ensure that it
1985 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1987 struct blk_mq_hw_ctx *hctx;
1988 struct blk_mq_ctx *ctx;
1991 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1992 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1994 spin_lock(&ctx->lock);
1995 if (!list_empty(&ctx->rq_list)) {
1996 list_splice_init(&ctx->rq_list, &tmp);
1997 blk_mq_hctx_clear_pending(hctx, ctx);
1999 spin_unlock(&ctx->lock);
2001 if (list_empty(&tmp))
2004 spin_lock(&hctx->lock);
2005 list_splice_tail_init(&tmp, &hctx->dispatch);
2006 spin_unlock(&hctx->lock);
2008 blk_mq_run_hw_queue(hctx, true);
2012 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2014 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2018 /* hctx->ctxs will be freed in queue's release handler */
2019 static void blk_mq_exit_hctx(struct request_queue *q,
2020 struct blk_mq_tag_set *set,
2021 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2023 blk_mq_debugfs_unregister_hctx(hctx);
2025 blk_mq_tag_idle(hctx);
2027 if (set->ops->exit_request)
2028 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2030 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2032 if (set->ops->exit_hctx)
2033 set->ops->exit_hctx(hctx, hctx_idx);
2035 if (hctx->flags & BLK_MQ_F_BLOCKING)
2036 cleanup_srcu_struct(hctx->queue_rq_srcu);
2038 blk_mq_remove_cpuhp(hctx);
2039 blk_free_flush_queue(hctx->fq);
2040 sbitmap_free(&hctx->ctx_map);
2043 static void blk_mq_exit_hw_queues(struct request_queue *q,
2044 struct blk_mq_tag_set *set, int nr_queue)
2046 struct blk_mq_hw_ctx *hctx;
2049 queue_for_each_hw_ctx(q, hctx, i) {
2052 blk_mq_exit_hctx(q, set, hctx, i);
2056 static int blk_mq_init_hctx(struct request_queue *q,
2057 struct blk_mq_tag_set *set,
2058 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2062 node = hctx->numa_node;
2063 if (node == NUMA_NO_NODE)
2064 node = hctx->numa_node = set->numa_node;
2066 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2067 spin_lock_init(&hctx->lock);
2068 INIT_LIST_HEAD(&hctx->dispatch);
2070 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2072 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2074 hctx->tags = set->tags[hctx_idx];
2077 * Allocate space for all possible cpus to avoid allocation at
2080 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
2083 goto unregister_cpu_notifier;
2085 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2091 if (set->ops->init_hctx &&
2092 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2095 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2098 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2100 goto sched_exit_hctx;
2102 if (set->ops->init_request &&
2103 set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
2107 if (hctx->flags & BLK_MQ_F_BLOCKING)
2108 init_srcu_struct(hctx->queue_rq_srcu);
2110 blk_mq_debugfs_register_hctx(q, hctx);
2117 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2119 if (set->ops->exit_hctx)
2120 set->ops->exit_hctx(hctx, hctx_idx);
2122 sbitmap_free(&hctx->ctx_map);
2125 unregister_cpu_notifier:
2126 blk_mq_remove_cpuhp(hctx);
2130 static void blk_mq_init_cpu_queues(struct request_queue *q,
2131 unsigned int nr_hw_queues)
2135 for_each_possible_cpu(i) {
2136 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2137 struct blk_mq_hw_ctx *hctx;
2140 spin_lock_init(&__ctx->lock);
2141 INIT_LIST_HEAD(&__ctx->rq_list);
2144 /* If the cpu isn't present, the cpu is mapped to first hctx */
2145 if (!cpu_present(i))
2148 hctx = blk_mq_map_queue(q, i);
2151 * Set local node, IFF we have more than one hw queue. If
2152 * not, we remain on the home node of the device
2154 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2155 hctx->numa_node = local_memory_node(cpu_to_node(i));
2159 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2163 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2164 set->queue_depth, set->reserved_tags);
2165 if (!set->tags[hctx_idx])
2168 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2173 blk_mq_free_rq_map(set->tags[hctx_idx]);
2174 set->tags[hctx_idx] = NULL;
2178 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2179 unsigned int hctx_idx)
2181 if (set->tags[hctx_idx]) {
2182 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2183 blk_mq_free_rq_map(set->tags[hctx_idx]);
2184 set->tags[hctx_idx] = NULL;
2188 static void blk_mq_map_swqueue(struct request_queue *q)
2190 unsigned int i, hctx_idx;
2191 struct blk_mq_hw_ctx *hctx;
2192 struct blk_mq_ctx *ctx;
2193 struct blk_mq_tag_set *set = q->tag_set;
2196 * Avoid others reading imcomplete hctx->cpumask through sysfs
2198 mutex_lock(&q->sysfs_lock);
2200 queue_for_each_hw_ctx(q, hctx, i) {
2201 cpumask_clear(hctx->cpumask);
2206 * Map software to hardware queues.
2208 * If the cpu isn't present, the cpu is mapped to first hctx.
2210 for_each_present_cpu(i) {
2211 hctx_idx = q->mq_map[i];
2212 /* unmapped hw queue can be remapped after CPU topo changed */
2213 if (!set->tags[hctx_idx] &&
2214 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2216 * If tags initialization fail for some hctx,
2217 * that hctx won't be brought online. In this
2218 * case, remap the current ctx to hctx[0] which
2219 * is guaranteed to always have tags allocated
2224 ctx = per_cpu_ptr(q->queue_ctx, i);
2225 hctx = blk_mq_map_queue(q, i);
2227 cpumask_set_cpu(i, hctx->cpumask);
2228 ctx->index_hw = hctx->nr_ctx;
2229 hctx->ctxs[hctx->nr_ctx++] = ctx;
2232 mutex_unlock(&q->sysfs_lock);
2234 queue_for_each_hw_ctx(q, hctx, i) {
2236 * If no software queues are mapped to this hardware queue,
2237 * disable it and free the request entries.
2239 if (!hctx->nr_ctx) {
2240 /* Never unmap queue 0. We need it as a
2241 * fallback in case of a new remap fails
2244 if (i && set->tags[i])
2245 blk_mq_free_map_and_requests(set, i);
2251 hctx->tags = set->tags[i];
2252 WARN_ON(!hctx->tags);
2255 * Set the map size to the number of mapped software queues.
2256 * This is more accurate and more efficient than looping
2257 * over all possibly mapped software queues.
2259 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2262 * Initialize batch roundrobin counts
2264 hctx->next_cpu = cpumask_first(hctx->cpumask);
2265 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2270 * Caller needs to ensure that we're either frozen/quiesced, or that
2271 * the queue isn't live yet.
2273 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2275 struct blk_mq_hw_ctx *hctx;
2278 queue_for_each_hw_ctx(q, hctx, i) {
2280 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2281 atomic_inc(&q->shared_hctx_restart);
2282 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2284 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2285 atomic_dec(&q->shared_hctx_restart);
2286 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2291 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2294 struct request_queue *q;
2296 lockdep_assert_held(&set->tag_list_lock);
2298 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2299 blk_mq_freeze_queue(q);
2300 queue_set_hctx_shared(q, shared);
2301 blk_mq_unfreeze_queue(q);
2305 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2307 struct blk_mq_tag_set *set = q->tag_set;
2309 mutex_lock(&set->tag_list_lock);
2310 list_del_rcu(&q->tag_set_list);
2311 INIT_LIST_HEAD(&q->tag_set_list);
2312 if (list_is_singular(&set->tag_list)) {
2313 /* just transitioned to unshared */
2314 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2315 /* update existing queue */
2316 blk_mq_update_tag_set_depth(set, false);
2318 mutex_unlock(&set->tag_list_lock);
2323 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2324 struct request_queue *q)
2328 mutex_lock(&set->tag_list_lock);
2330 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2331 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2332 set->flags |= BLK_MQ_F_TAG_SHARED;
2333 /* update existing queue */
2334 blk_mq_update_tag_set_depth(set, true);
2336 if (set->flags & BLK_MQ_F_TAG_SHARED)
2337 queue_set_hctx_shared(q, true);
2338 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2340 mutex_unlock(&set->tag_list_lock);
2344 * It is the actual release handler for mq, but we do it from
2345 * request queue's release handler for avoiding use-after-free
2346 * and headache because q->mq_kobj shouldn't have been introduced,
2347 * but we can't group ctx/kctx kobj without it.
2349 void blk_mq_release(struct request_queue *q)
2351 struct blk_mq_hw_ctx *hctx;
2354 /* hctx kobj stays in hctx */
2355 queue_for_each_hw_ctx(q, hctx, i) {
2358 kobject_put(&hctx->kobj);
2363 kfree(q->queue_hw_ctx);
2366 * release .mq_kobj and sw queue's kobject now because
2367 * both share lifetime with request queue.
2369 blk_mq_sysfs_deinit(q);
2371 free_percpu(q->queue_ctx);
2374 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2376 struct request_queue *uninit_q, *q;
2378 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2380 return ERR_PTR(-ENOMEM);
2382 q = blk_mq_init_allocated_queue(set, uninit_q);
2384 blk_cleanup_queue(uninit_q);
2388 EXPORT_SYMBOL(blk_mq_init_queue);
2390 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2392 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2394 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2395 __alignof__(struct blk_mq_hw_ctx)) !=
2396 sizeof(struct blk_mq_hw_ctx));
2398 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2399 hw_ctx_size += sizeof(struct srcu_struct);
2404 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2405 struct request_queue *q)
2408 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2410 blk_mq_sysfs_unregister(q);
2411 for (i = 0; i < set->nr_hw_queues; i++) {
2417 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2418 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2423 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2430 atomic_set(&hctxs[i]->nr_active, 0);
2431 hctxs[i]->numa_node = node;
2432 hctxs[i]->queue_num = i;
2434 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2435 free_cpumask_var(hctxs[i]->cpumask);
2440 blk_mq_hctx_kobj_init(hctxs[i]);
2442 for (j = i; j < q->nr_hw_queues; j++) {
2443 struct blk_mq_hw_ctx *hctx = hctxs[j];
2447 blk_mq_free_map_and_requests(set, j);
2448 blk_mq_exit_hctx(q, set, hctx, j);
2449 kobject_put(&hctx->kobj);
2454 q->nr_hw_queues = i;
2455 blk_mq_sysfs_register(q);
2458 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2459 struct request_queue *q)
2461 /* mark the queue as mq asap */
2462 q->mq_ops = set->ops;
2464 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2465 blk_mq_poll_stats_bkt,
2466 BLK_MQ_POLL_STATS_BKTS, q);
2470 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2474 /* init q->mq_kobj and sw queues' kobjects */
2475 blk_mq_sysfs_init(q);
2477 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2478 GFP_KERNEL, set->numa_node);
2479 if (!q->queue_hw_ctx)
2482 q->mq_map = set->mq_map;
2484 blk_mq_realloc_hw_ctxs(set, q);
2485 if (!q->nr_hw_queues)
2488 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2489 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2491 q->nr_queues = nr_cpu_ids;
2493 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2495 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2496 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2498 q->sg_reserved_size = INT_MAX;
2500 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2501 INIT_LIST_HEAD(&q->requeue_list);
2502 spin_lock_init(&q->requeue_lock);
2504 blk_queue_make_request(q, blk_mq_make_request);
2505 if (q->mq_ops->poll)
2506 q->poll_fn = blk_mq_poll;
2509 * Do this after blk_queue_make_request() overrides it...
2511 q->nr_requests = set->queue_depth;
2514 * Default to classic polling
2518 if (set->ops->complete)
2519 blk_queue_softirq_done(q, set->ops->complete);
2521 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2522 blk_mq_add_queue_tag_set(set, q);
2523 blk_mq_map_swqueue(q);
2525 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2528 ret = blk_mq_sched_init(q);
2530 return ERR_PTR(ret);
2536 kfree(q->queue_hw_ctx);
2538 free_percpu(q->queue_ctx);
2541 return ERR_PTR(-ENOMEM);
2543 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2545 void blk_mq_free_queue(struct request_queue *q)
2547 struct blk_mq_tag_set *set = q->tag_set;
2549 blk_mq_del_queue_tag_set(q);
2550 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2553 /* Basically redo blk_mq_init_queue with queue frozen */
2554 static void blk_mq_queue_reinit(struct request_queue *q)
2556 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2558 blk_mq_debugfs_unregister_hctxs(q);
2559 blk_mq_sysfs_unregister(q);
2562 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2563 * we should change hctx numa_node according to new topology (this
2564 * involves free and re-allocate memory, worthy doing?)
2567 blk_mq_map_swqueue(q);
2569 blk_mq_sysfs_register(q);
2570 blk_mq_debugfs_register_hctxs(q);
2573 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2577 for (i = 0; i < set->nr_hw_queues; i++)
2578 if (!__blk_mq_alloc_rq_map(set, i))
2585 blk_mq_free_rq_map(set->tags[i]);
2591 * Allocate the request maps associated with this tag_set. Note that this
2592 * may reduce the depth asked for, if memory is tight. set->queue_depth
2593 * will be updated to reflect the allocated depth.
2595 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2600 depth = set->queue_depth;
2602 err = __blk_mq_alloc_rq_maps(set);
2606 set->queue_depth >>= 1;
2607 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2611 } while (set->queue_depth);
2613 if (!set->queue_depth || err) {
2614 pr_err("blk-mq: failed to allocate request map\n");
2618 if (depth != set->queue_depth)
2619 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2620 depth, set->queue_depth);
2625 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2627 if (set->ops->map_queues)
2628 return set->ops->map_queues(set);
2630 return blk_mq_map_queues(set);
2634 * Alloc a tag set to be associated with one or more request queues.
2635 * May fail with EINVAL for various error conditions. May adjust the
2636 * requested depth down, if if it too large. In that case, the set
2637 * value will be stored in set->queue_depth.
2639 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2643 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2645 if (!set->nr_hw_queues)
2647 if (!set->queue_depth)
2649 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2652 if (!set->ops->queue_rq)
2655 if (!set->ops->get_budget ^ !set->ops->put_budget)
2658 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2659 pr_info("blk-mq: reduced tag depth to %u\n",
2661 set->queue_depth = BLK_MQ_MAX_DEPTH;
2665 * If a crashdump is active, then we are potentially in a very
2666 * memory constrained environment. Limit us to 1 queue and
2667 * 64 tags to prevent using too much memory.
2669 if (is_kdump_kernel()) {
2670 set->nr_hw_queues = 1;
2671 set->queue_depth = min(64U, set->queue_depth);
2674 * There is no use for more h/w queues than cpus.
2676 if (set->nr_hw_queues > nr_cpu_ids)
2677 set->nr_hw_queues = nr_cpu_ids;
2679 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2680 GFP_KERNEL, set->numa_node);
2685 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2686 GFP_KERNEL, set->numa_node);
2690 ret = blk_mq_update_queue_map(set);
2692 goto out_free_mq_map;
2694 ret = blk_mq_alloc_rq_maps(set);
2696 goto out_free_mq_map;
2698 mutex_init(&set->tag_list_lock);
2699 INIT_LIST_HEAD(&set->tag_list);
2711 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2713 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2717 for (i = 0; i < nr_cpu_ids; i++)
2718 blk_mq_free_map_and_requests(set, i);
2726 EXPORT_SYMBOL(blk_mq_free_tag_set);
2728 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2730 struct blk_mq_tag_set *set = q->tag_set;
2731 struct blk_mq_hw_ctx *hctx;
2737 blk_mq_freeze_queue(q);
2740 queue_for_each_hw_ctx(q, hctx, i) {
2744 * If we're using an MQ scheduler, just update the scheduler
2745 * queue depth. This is similar to what the old code would do.
2747 if (!hctx->sched_tags) {
2748 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2751 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2759 q->nr_requests = nr;
2761 blk_mq_unfreeze_queue(q);
2766 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2769 struct request_queue *q;
2771 lockdep_assert_held(&set->tag_list_lock);
2773 if (nr_hw_queues > nr_cpu_ids)
2774 nr_hw_queues = nr_cpu_ids;
2775 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2778 list_for_each_entry(q, &set->tag_list, tag_set_list)
2779 blk_mq_freeze_queue(q);
2781 set->nr_hw_queues = nr_hw_queues;
2782 blk_mq_update_queue_map(set);
2783 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2784 blk_mq_realloc_hw_ctxs(set, q);
2785 blk_mq_queue_reinit(q);
2788 list_for_each_entry(q, &set->tag_list, tag_set_list)
2789 blk_mq_unfreeze_queue(q);
2792 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2794 mutex_lock(&set->tag_list_lock);
2795 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2796 mutex_unlock(&set->tag_list_lock);
2798 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2800 /* Enable polling stats and return whether they were already enabled. */
2801 static bool blk_poll_stats_enable(struct request_queue *q)
2803 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2804 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2806 blk_stat_add_callback(q, q->poll_cb);
2810 static void blk_mq_poll_stats_start(struct request_queue *q)
2813 * We don't arm the callback if polling stats are not enabled or the
2814 * callback is already active.
2816 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2817 blk_stat_is_active(q->poll_cb))
2820 blk_stat_activate_msecs(q->poll_cb, 100);
2823 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2825 struct request_queue *q = cb->data;
2828 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2829 if (cb->stat[bucket].nr_samples)
2830 q->poll_stat[bucket] = cb->stat[bucket];
2834 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2835 struct blk_mq_hw_ctx *hctx,
2838 unsigned long ret = 0;
2842 * If stats collection isn't on, don't sleep but turn it on for
2845 if (!blk_poll_stats_enable(q))
2849 * As an optimistic guess, use half of the mean service time
2850 * for this type of request. We can (and should) make this smarter.
2851 * For instance, if the completion latencies are tight, we can
2852 * get closer than just half the mean. This is especially
2853 * important on devices where the completion latencies are longer
2854 * than ~10 usec. We do use the stats for the relevant IO size
2855 * if available which does lead to better estimates.
2857 bucket = blk_mq_poll_stats_bkt(rq);
2861 if (q->poll_stat[bucket].nr_samples)
2862 ret = (q->poll_stat[bucket].mean + 1) / 2;
2867 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2868 struct blk_mq_hw_ctx *hctx,
2871 struct hrtimer_sleeper hs;
2872 enum hrtimer_mode mode;
2876 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2882 * -1: don't ever hybrid sleep
2883 * 0: use half of prev avg
2884 * >0: use this specific value
2886 if (q->poll_nsec == -1)
2888 else if (q->poll_nsec > 0)
2889 nsecs = q->poll_nsec;
2891 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2896 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2899 * This will be replaced with the stats tracking code, using
2900 * 'avg_completion_time / 2' as the pre-sleep target.
2904 mode = HRTIMER_MODE_REL;
2905 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2906 hrtimer_set_expires(&hs.timer, kt);
2908 hrtimer_init_sleeper(&hs, current);
2910 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2912 set_current_state(TASK_UNINTERRUPTIBLE);
2913 hrtimer_start_expires(&hs.timer, mode);
2916 hrtimer_cancel(&hs.timer);
2917 mode = HRTIMER_MODE_ABS;
2918 } while (hs.task && !signal_pending(current));
2920 __set_current_state(TASK_RUNNING);
2921 destroy_hrtimer_on_stack(&hs.timer);
2925 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2927 struct request_queue *q = hctx->queue;
2931 * If we sleep, have the caller restart the poll loop to reset
2932 * the state. Like for the other success return cases, the
2933 * caller is responsible for checking if the IO completed. If
2934 * the IO isn't complete, we'll get called again and will go
2935 * straight to the busy poll loop.
2937 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2940 hctx->poll_considered++;
2942 state = current->state;
2943 while (!need_resched()) {
2946 hctx->poll_invoked++;
2948 ret = q->mq_ops->poll(hctx, rq->tag);
2950 hctx->poll_success++;
2951 set_current_state(TASK_RUNNING);
2955 if (signal_pending_state(state, current))
2956 set_current_state(TASK_RUNNING);
2958 if (current->state == TASK_RUNNING)
2968 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2970 struct blk_mq_hw_ctx *hctx;
2973 if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2976 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2977 if (!blk_qc_t_is_internal(cookie))
2978 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2980 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2982 * With scheduling, if the request has completed, we'll
2983 * get a NULL return here, as we clear the sched tag when
2984 * that happens. The request still remains valid, like always,
2985 * so we should be safe with just the NULL check.
2991 return __blk_mq_poll(hctx, rq);
2994 static int __init blk_mq_init(void)
2997 * See comment in block/blk.h rq_atomic_flags enum
2999 BUILD_BUG_ON((REQ_ATOM_STARTED / BITS_PER_BYTE) !=
3000 (REQ_ATOM_COMPLETE / BITS_PER_BYTE));
3002 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3003 blk_mq_hctx_notify_dead);
3006 subsys_initcall(blk_mq_init);