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-tag.h"
37 #include "blk-mq-sched.h"
39 static DEFINE_MUTEX(all_q_mutex);
40 static LIST_HEAD(all_q_list);
42 static void blk_mq_poll_stats_start(struct request_queue *q);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
45 static int blk_mq_poll_stats_bkt(const struct request *rq)
47 int ddir, bytes, bucket;
49 ddir = rq_data_dir(rq);
50 bytes = blk_rq_bytes(rq);
52 bucket = ddir + 2*(ilog2(bytes) - 9);
56 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
57 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
63 * Check if any of the ctx's have pending work in this hardware queue
65 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
67 return sbitmap_any_bit_set(&hctx->ctx_map) ||
68 !list_empty_careful(&hctx->dispatch) ||
69 blk_mq_sched_has_work(hctx);
73 * Mark this ctx as having pending work in this hardware queue
75 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
76 struct blk_mq_ctx *ctx)
78 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
79 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
82 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
83 struct blk_mq_ctx *ctx)
85 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
88 void blk_freeze_queue_start(struct request_queue *q)
92 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
93 if (freeze_depth == 1) {
94 percpu_ref_kill(&q->q_usage_counter);
95 blk_mq_run_hw_queues(q, false);
98 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
100 void blk_mq_freeze_queue_wait(struct request_queue *q)
102 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
104 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
106 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
107 unsigned long timeout)
109 return wait_event_timeout(q->mq_freeze_wq,
110 percpu_ref_is_zero(&q->q_usage_counter),
113 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
116 * Guarantee no request is in use, so we can change any data structure of
117 * the queue afterward.
119 void blk_freeze_queue(struct request_queue *q)
122 * In the !blk_mq case we are only calling this to kill the
123 * q_usage_counter, otherwise this increases the freeze depth
124 * and waits for it to return to zero. For this reason there is
125 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
126 * exported to drivers as the only user for unfreeze is blk_mq.
128 blk_freeze_queue_start(q);
129 blk_mq_freeze_queue_wait(q);
132 void blk_mq_freeze_queue(struct request_queue *q)
135 * ...just an alias to keep freeze and unfreeze actions balanced
136 * in the blk_mq_* namespace
140 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
142 void blk_mq_unfreeze_queue(struct request_queue *q)
146 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
147 WARN_ON_ONCE(freeze_depth < 0);
149 percpu_ref_reinit(&q->q_usage_counter);
150 wake_up_all(&q->mq_freeze_wq);
153 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
156 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
159 * Note: this function does not prevent that the struct request end_io()
160 * callback function is invoked. Additionally, it is not prevented that
161 * new queue_rq() calls occur unless the queue has been stopped first.
163 void blk_mq_quiesce_queue(struct request_queue *q)
165 struct blk_mq_hw_ctx *hctx;
169 blk_mq_stop_hw_queues(q);
171 queue_for_each_hw_ctx(q, hctx, i) {
172 if (hctx->flags & BLK_MQ_F_BLOCKING)
173 synchronize_srcu(&hctx->queue_rq_srcu);
180 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
182 void blk_mq_wake_waiters(struct request_queue *q)
184 struct blk_mq_hw_ctx *hctx;
187 queue_for_each_hw_ctx(q, hctx, i)
188 if (blk_mq_hw_queue_mapped(hctx))
189 blk_mq_tag_wakeup_all(hctx->tags, true);
192 * If we are called because the queue has now been marked as
193 * dying, we need to ensure that processes currently waiting on
194 * the queue are notified as well.
196 wake_up_all(&q->mq_freeze_wq);
199 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
201 return blk_mq_has_free_tags(hctx->tags);
203 EXPORT_SYMBOL(blk_mq_can_queue);
205 void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
206 struct request *rq, unsigned int op)
208 INIT_LIST_HEAD(&rq->queuelist);
209 /* csd/requeue_work/fifo_time is initialized before use */
213 if (blk_queue_io_stat(q))
214 rq->rq_flags |= RQF_IO_STAT;
215 /* do not touch atomic flags, it needs atomic ops against the timer */
217 INIT_HLIST_NODE(&rq->hash);
218 RB_CLEAR_NODE(&rq->rb_node);
221 rq->start_time = jiffies;
222 #ifdef CONFIG_BLK_CGROUP
224 set_start_time_ns(rq);
225 rq->io_start_time_ns = 0;
227 rq->nr_phys_segments = 0;
228 #if defined(CONFIG_BLK_DEV_INTEGRITY)
229 rq->nr_integrity_segments = 0;
232 /* tag was already set */
235 INIT_LIST_HEAD(&rq->timeout_list);
239 rq->end_io_data = NULL;
242 ctx->rq_dispatched[op_is_sync(op)]++;
244 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init);
246 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
252 tag = blk_mq_get_tag(data);
253 if (tag != BLK_MQ_TAG_FAIL) {
254 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
256 rq = tags->static_rqs[tag];
258 if (data->flags & BLK_MQ_REQ_INTERNAL) {
260 rq->internal_tag = tag;
262 if (blk_mq_tag_busy(data->hctx)) {
263 rq->rq_flags = RQF_MQ_INFLIGHT;
264 atomic_inc(&data->hctx->nr_active);
267 rq->internal_tag = -1;
268 data->hctx->tags->rqs[rq->tag] = rq;
271 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
277 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
279 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
282 struct blk_mq_alloc_data alloc_data = { .flags = flags };
286 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
290 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
292 blk_mq_put_ctx(alloc_data.ctx);
296 return ERR_PTR(-EWOULDBLOCK);
299 rq->__sector = (sector_t) -1;
300 rq->bio = rq->biotail = NULL;
303 EXPORT_SYMBOL(blk_mq_alloc_request);
305 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
306 unsigned int flags, unsigned int hctx_idx)
308 struct blk_mq_alloc_data alloc_data = { .flags = flags };
314 * If the tag allocator sleeps we could get an allocation for a
315 * different hardware context. No need to complicate the low level
316 * allocator for this for the rare use case of a command tied to
319 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
320 return ERR_PTR(-EINVAL);
322 if (hctx_idx >= q->nr_hw_queues)
323 return ERR_PTR(-EIO);
325 ret = blk_queue_enter(q, true);
330 * Check if the hardware context is actually mapped to anything.
331 * If not tell the caller that it should skip this queue.
333 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
334 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
336 return ERR_PTR(-EXDEV);
338 cpu = cpumask_first(alloc_data.hctx->cpumask);
339 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
341 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
346 return ERR_PTR(-EWOULDBLOCK);
350 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
352 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
355 const int sched_tag = rq->internal_tag;
356 struct request_queue *q = rq->q;
358 if (rq->rq_flags & RQF_MQ_INFLIGHT)
359 atomic_dec(&hctx->nr_active);
361 wbt_done(q->rq_wb, &rq->issue_stat);
364 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
365 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
367 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
369 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
370 blk_mq_sched_restart(hctx);
374 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
377 struct blk_mq_ctx *ctx = rq->mq_ctx;
379 ctx->rq_completed[rq_is_sync(rq)]++;
380 __blk_mq_finish_request(hctx, ctx, rq);
383 void blk_mq_finish_request(struct request *rq)
385 blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
387 EXPORT_SYMBOL_GPL(blk_mq_finish_request);
389 void blk_mq_free_request(struct request *rq)
391 blk_mq_sched_put_request(rq);
393 EXPORT_SYMBOL_GPL(blk_mq_free_request);
395 inline void __blk_mq_end_request(struct request *rq, int error)
397 blk_account_io_done(rq);
400 wbt_done(rq->q->rq_wb, &rq->issue_stat);
401 rq->end_io(rq, error);
403 if (unlikely(blk_bidi_rq(rq)))
404 blk_mq_free_request(rq->next_rq);
405 blk_mq_free_request(rq);
408 EXPORT_SYMBOL(__blk_mq_end_request);
410 void blk_mq_end_request(struct request *rq, int error)
412 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
414 __blk_mq_end_request(rq, error);
416 EXPORT_SYMBOL(blk_mq_end_request);
418 static void __blk_mq_complete_request_remote(void *data)
420 struct request *rq = data;
422 rq->q->softirq_done_fn(rq);
425 static void __blk_mq_complete_request(struct request *rq)
427 struct blk_mq_ctx *ctx = rq->mq_ctx;
431 if (rq->internal_tag != -1)
432 blk_mq_sched_completed_request(rq);
433 if (rq->rq_flags & RQF_STATS) {
434 blk_mq_poll_stats_start(rq->q);
438 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
439 rq->q->softirq_done_fn(rq);
444 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
445 shared = cpus_share_cache(cpu, ctx->cpu);
447 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
448 rq->csd.func = __blk_mq_complete_request_remote;
451 smp_call_function_single_async(ctx->cpu, &rq->csd);
453 rq->q->softirq_done_fn(rq);
459 * blk_mq_complete_request - end I/O on a request
460 * @rq: the request being processed
463 * Ends all I/O on a request. It does not handle partial completions.
464 * The actual completion happens out-of-order, through a IPI handler.
466 void blk_mq_complete_request(struct request *rq)
468 struct request_queue *q = rq->q;
470 if (unlikely(blk_should_fake_timeout(q)))
472 if (!blk_mark_rq_complete(rq))
473 __blk_mq_complete_request(rq);
475 EXPORT_SYMBOL(blk_mq_complete_request);
477 int blk_mq_request_started(struct request *rq)
479 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
481 EXPORT_SYMBOL_GPL(blk_mq_request_started);
483 void blk_mq_start_request(struct request *rq)
485 struct request_queue *q = rq->q;
487 blk_mq_sched_started_request(rq);
489 trace_block_rq_issue(q, rq);
491 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
492 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
493 rq->rq_flags |= RQF_STATS;
494 wbt_issue(q->rq_wb, &rq->issue_stat);
500 * Ensure that ->deadline is visible before set the started
501 * flag and clear the completed flag.
503 smp_mb__before_atomic();
506 * Mark us as started and clear complete. Complete might have been
507 * set if requeue raced with timeout, which then marked it as
508 * complete. So be sure to clear complete again when we start
509 * the request, otherwise we'll ignore the completion event.
511 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
512 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
513 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
514 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
516 if (q->dma_drain_size && blk_rq_bytes(rq)) {
518 * Make sure space for the drain appears. We know we can do
519 * this because max_hw_segments has been adjusted to be one
520 * fewer than the device can handle.
522 rq->nr_phys_segments++;
525 EXPORT_SYMBOL(blk_mq_start_request);
528 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
529 * flag isn't set yet, so there may be race with timeout handler,
530 * but given rq->deadline is just set in .queue_rq() under
531 * this situation, the race won't be possible in reality because
532 * rq->timeout should be set as big enough to cover the window
533 * between blk_mq_start_request() called from .queue_rq() and
534 * clearing REQ_ATOM_STARTED here.
536 static void __blk_mq_requeue_request(struct request *rq)
538 struct request_queue *q = rq->q;
540 trace_block_rq_requeue(q, rq);
541 wbt_requeue(q->rq_wb, &rq->issue_stat);
542 blk_mq_sched_requeue_request(rq);
544 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
545 if (q->dma_drain_size && blk_rq_bytes(rq))
546 rq->nr_phys_segments--;
550 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
552 __blk_mq_requeue_request(rq);
554 BUG_ON(blk_queued_rq(rq));
555 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
557 EXPORT_SYMBOL(blk_mq_requeue_request);
559 static void blk_mq_requeue_work(struct work_struct *work)
561 struct request_queue *q =
562 container_of(work, struct request_queue, requeue_work.work);
564 struct request *rq, *next;
567 spin_lock_irqsave(&q->requeue_lock, flags);
568 list_splice_init(&q->requeue_list, &rq_list);
569 spin_unlock_irqrestore(&q->requeue_lock, flags);
571 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
572 if (!(rq->rq_flags & RQF_SOFTBARRIER))
575 rq->rq_flags &= ~RQF_SOFTBARRIER;
576 list_del_init(&rq->queuelist);
577 blk_mq_sched_insert_request(rq, true, false, false, true);
580 while (!list_empty(&rq_list)) {
581 rq = list_entry(rq_list.next, struct request, queuelist);
582 list_del_init(&rq->queuelist);
583 blk_mq_sched_insert_request(rq, false, false, false, true);
586 blk_mq_run_hw_queues(q, false);
589 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
590 bool kick_requeue_list)
592 struct request_queue *q = rq->q;
596 * We abuse this flag that is otherwise used by the I/O scheduler to
597 * request head insertation from the workqueue.
599 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
601 spin_lock_irqsave(&q->requeue_lock, flags);
603 rq->rq_flags |= RQF_SOFTBARRIER;
604 list_add(&rq->queuelist, &q->requeue_list);
606 list_add_tail(&rq->queuelist, &q->requeue_list);
608 spin_unlock_irqrestore(&q->requeue_lock, flags);
610 if (kick_requeue_list)
611 blk_mq_kick_requeue_list(q);
613 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
615 void blk_mq_kick_requeue_list(struct request_queue *q)
617 kblockd_schedule_delayed_work(&q->requeue_work, 0);
619 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
621 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
624 kblockd_schedule_delayed_work(&q->requeue_work,
625 msecs_to_jiffies(msecs));
627 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
629 void blk_mq_abort_requeue_list(struct request_queue *q)
634 spin_lock_irqsave(&q->requeue_lock, flags);
635 list_splice_init(&q->requeue_list, &rq_list);
636 spin_unlock_irqrestore(&q->requeue_lock, flags);
638 while (!list_empty(&rq_list)) {
641 rq = list_first_entry(&rq_list, struct request, queuelist);
642 list_del_init(&rq->queuelist);
643 blk_mq_end_request(rq, -EIO);
646 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
648 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
650 if (tag < tags->nr_tags) {
651 prefetch(tags->rqs[tag]);
652 return tags->rqs[tag];
657 EXPORT_SYMBOL(blk_mq_tag_to_rq);
659 struct blk_mq_timeout_data {
661 unsigned int next_set;
664 void blk_mq_rq_timed_out(struct request *req, bool reserved)
666 const struct blk_mq_ops *ops = req->q->mq_ops;
667 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
670 * We know that complete is set at this point. If STARTED isn't set
671 * anymore, then the request isn't active and the "timeout" should
672 * just be ignored. This can happen due to the bitflag ordering.
673 * Timeout first checks if STARTED is set, and if it is, assumes
674 * the request is active. But if we race with completion, then
675 * both flags will get cleared. So check here again, and ignore
676 * a timeout event with a request that isn't active.
678 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
682 ret = ops->timeout(req, reserved);
686 __blk_mq_complete_request(req);
688 case BLK_EH_RESET_TIMER:
690 blk_clear_rq_complete(req);
692 case BLK_EH_NOT_HANDLED:
695 printk(KERN_ERR "block: bad eh return: %d\n", ret);
700 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
701 struct request *rq, void *priv, bool reserved)
703 struct blk_mq_timeout_data *data = priv;
705 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
709 * The rq being checked may have been freed and reallocated
710 * out already here, we avoid this race by checking rq->deadline
711 * and REQ_ATOM_COMPLETE flag together:
713 * - if rq->deadline is observed as new value because of
714 * reusing, the rq won't be timed out because of timing.
715 * - if rq->deadline is observed as previous value,
716 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
717 * because we put a barrier between setting rq->deadline
718 * and clearing the flag in blk_mq_start_request(), so
719 * this rq won't be timed out too.
721 if (time_after_eq(jiffies, rq->deadline)) {
722 if (!blk_mark_rq_complete(rq))
723 blk_mq_rq_timed_out(rq, reserved);
724 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
725 data->next = rq->deadline;
730 static void blk_mq_timeout_work(struct work_struct *work)
732 struct request_queue *q =
733 container_of(work, struct request_queue, timeout_work);
734 struct blk_mq_timeout_data data = {
740 /* A deadlock might occur if a request is stuck requiring a
741 * timeout at the same time a queue freeze is waiting
742 * completion, since the timeout code would not be able to
743 * acquire the queue reference here.
745 * That's why we don't use blk_queue_enter here; instead, we use
746 * percpu_ref_tryget directly, because we need to be able to
747 * obtain a reference even in the short window between the queue
748 * starting to freeze, by dropping the first reference in
749 * blk_freeze_queue_start, and the moment the last request is
750 * consumed, marked by the instant q_usage_counter reaches
753 if (!percpu_ref_tryget(&q->q_usage_counter))
756 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
759 data.next = blk_rq_timeout(round_jiffies_up(data.next));
760 mod_timer(&q->timeout, data.next);
762 struct blk_mq_hw_ctx *hctx;
764 queue_for_each_hw_ctx(q, hctx, i) {
765 /* the hctx may be unmapped, so check it here */
766 if (blk_mq_hw_queue_mapped(hctx))
767 blk_mq_tag_idle(hctx);
774 * Reverse check our software queue for entries that we could potentially
775 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
776 * too much time checking for merges.
778 static bool blk_mq_attempt_merge(struct request_queue *q,
779 struct blk_mq_ctx *ctx, struct bio *bio)
784 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
790 if (!blk_rq_merge_ok(rq, bio))
793 switch (blk_try_merge(rq, bio)) {
794 case ELEVATOR_BACK_MERGE:
795 if (blk_mq_sched_allow_merge(q, rq, bio))
796 merged = bio_attempt_back_merge(q, rq, bio);
798 case ELEVATOR_FRONT_MERGE:
799 if (blk_mq_sched_allow_merge(q, rq, bio))
800 merged = bio_attempt_front_merge(q, rq, bio);
802 case ELEVATOR_DISCARD_MERGE:
803 merged = bio_attempt_discard_merge(q, rq, bio);
817 struct flush_busy_ctx_data {
818 struct blk_mq_hw_ctx *hctx;
819 struct list_head *list;
822 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
824 struct flush_busy_ctx_data *flush_data = data;
825 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
826 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
828 sbitmap_clear_bit(sb, bitnr);
829 spin_lock(&ctx->lock);
830 list_splice_tail_init(&ctx->rq_list, flush_data->list);
831 spin_unlock(&ctx->lock);
836 * Process software queues that have been marked busy, splicing them
837 * to the for-dispatch
839 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
841 struct flush_busy_ctx_data data = {
846 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
848 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
850 static inline unsigned int queued_to_index(unsigned int queued)
855 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
858 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
861 struct blk_mq_alloc_data data = {
863 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
864 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
867 might_sleep_if(wait);
872 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
873 data.flags |= BLK_MQ_REQ_RESERVED;
875 rq->tag = blk_mq_get_tag(&data);
877 if (blk_mq_tag_busy(data.hctx)) {
878 rq->rq_flags |= RQF_MQ_INFLIGHT;
879 atomic_inc(&data.hctx->nr_active);
881 data.hctx->tags->rqs[rq->tag] = rq;
887 return rq->tag != -1;
890 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
893 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
896 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
897 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
898 atomic_dec(&hctx->nr_active);
902 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
905 if (rq->tag == -1 || rq->internal_tag == -1)
908 __blk_mq_put_driver_tag(hctx, rq);
911 static void blk_mq_put_driver_tag(struct request *rq)
913 struct blk_mq_hw_ctx *hctx;
915 if (rq->tag == -1 || rq->internal_tag == -1)
918 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
919 __blk_mq_put_driver_tag(hctx, rq);
923 * If we fail getting a driver tag because all the driver tags are already
924 * assigned and on the dispatch list, BUT the first entry does not have a
925 * tag, then we could deadlock. For that case, move entries with assigned
926 * driver tags to the front, leaving the set of tagged requests in the
927 * same order, and the untagged set in the same order.
929 static bool reorder_tags_to_front(struct list_head *list)
931 struct request *rq, *tmp, *first = NULL;
933 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
937 list_move(&rq->queuelist, list);
943 return first != NULL;
946 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
949 struct blk_mq_hw_ctx *hctx;
951 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
953 list_del(&wait->task_list);
954 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
955 blk_mq_run_hw_queue(hctx, true);
959 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
961 struct sbq_wait_state *ws;
964 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
965 * The thread which wins the race to grab this bit adds the hardware
966 * queue to the wait queue.
968 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
969 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
972 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
973 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
976 * As soon as this returns, it's no longer safe to fiddle with
977 * hctx->dispatch_wait, since a completion can wake up the wait queue
978 * and unlock the bit.
980 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
984 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
986 struct blk_mq_hw_ctx *hctx;
988 int errors, queued, ret = BLK_MQ_RQ_QUEUE_OK;
990 if (list_empty(list))
994 * Now process all the entries, sending them to the driver.
998 struct blk_mq_queue_data bd;
1000 rq = list_first_entry(list, struct request, queuelist);
1001 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1002 if (!queued && reorder_tags_to_front(list))
1006 * The initial allocation attempt failed, so we need to
1007 * rerun the hardware queue when a tag is freed.
1009 if (!blk_mq_dispatch_wait_add(hctx))
1013 * It's possible that a tag was freed in the window
1014 * between the allocation failure and adding the
1015 * hardware queue to the wait queue.
1017 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1021 list_del_init(&rq->queuelist);
1026 * Flag last if we have no more requests, or if we have more
1027 * but can't assign a driver tag to it.
1029 if (list_empty(list))
1032 struct request *nxt;
1034 nxt = list_first_entry(list, struct request, queuelist);
1035 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1038 ret = q->mq_ops->queue_rq(hctx, &bd);
1040 case BLK_MQ_RQ_QUEUE_OK:
1043 case BLK_MQ_RQ_QUEUE_BUSY:
1044 blk_mq_put_driver_tag_hctx(hctx, rq);
1045 list_add(&rq->queuelist, list);
1046 __blk_mq_requeue_request(rq);
1049 pr_err("blk-mq: bad return on queue: %d\n", ret);
1050 case BLK_MQ_RQ_QUEUE_ERROR:
1052 blk_mq_end_request(rq, -EIO);
1056 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
1058 } while (!list_empty(list));
1060 hctx->dispatched[queued_to_index(queued)]++;
1063 * Any items that need requeuing? Stuff them into hctx->dispatch,
1064 * that is where we will continue on next queue run.
1066 if (!list_empty(list)) {
1068 * If an I/O scheduler has been configured and we got a driver
1069 * tag for the next request already, free it again.
1071 rq = list_first_entry(list, struct request, queuelist);
1072 blk_mq_put_driver_tag(rq);
1074 spin_lock(&hctx->lock);
1075 list_splice_init(list, &hctx->dispatch);
1076 spin_unlock(&hctx->lock);
1079 * If SCHED_RESTART was set by the caller of this function and
1080 * it is no longer set that means that it was cleared by another
1081 * thread and hence that a queue rerun is needed.
1083 * If TAG_WAITING is set that means that an I/O scheduler has
1084 * been configured and another thread is waiting for a driver
1085 * tag. To guarantee fairness, do not rerun this hardware queue
1086 * but let the other thread grab the driver tag.
1088 * If no I/O scheduler has been configured it is possible that
1089 * the hardware queue got stopped and restarted before requests
1090 * were pushed back onto the dispatch list. Rerun the queue to
1091 * avoid starvation. Notes:
1092 * - blk_mq_run_hw_queue() checks whether or not a queue has
1093 * been stopped before rerunning a queue.
1094 * - Some but not all block drivers stop a queue before
1095 * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq
1098 if (!blk_mq_sched_needs_restart(hctx) &&
1099 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1100 blk_mq_run_hw_queue(hctx, true);
1103 return (queued + errors) != 0;
1106 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1110 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1111 cpu_online(hctx->next_cpu));
1113 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1115 blk_mq_sched_dispatch_requests(hctx);
1120 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1121 blk_mq_sched_dispatch_requests(hctx);
1122 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1127 * It'd be great if the workqueue API had a way to pass
1128 * in a mask and had some smarts for more clever placement.
1129 * For now we just round-robin here, switching for every
1130 * BLK_MQ_CPU_WORK_BATCH queued items.
1132 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1134 if (hctx->queue->nr_hw_queues == 1)
1135 return WORK_CPU_UNBOUND;
1137 if (--hctx->next_cpu_batch <= 0) {
1140 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1141 if (next_cpu >= nr_cpu_ids)
1142 next_cpu = cpumask_first(hctx->cpumask);
1144 hctx->next_cpu = next_cpu;
1145 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1148 return hctx->next_cpu;
1151 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1152 unsigned long msecs)
1154 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1155 !blk_mq_hw_queue_mapped(hctx)))
1158 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1159 int cpu = get_cpu();
1160 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1161 __blk_mq_run_hw_queue(hctx);
1169 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1171 msecs_to_jiffies(msecs));
1174 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1176 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1178 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1180 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1182 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1184 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1186 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1188 struct blk_mq_hw_ctx *hctx;
1191 queue_for_each_hw_ctx(q, hctx, i) {
1192 if (!blk_mq_hctx_has_pending(hctx) ||
1193 blk_mq_hctx_stopped(hctx))
1196 blk_mq_run_hw_queue(hctx, async);
1199 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1202 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1203 * @q: request queue.
1205 * The caller is responsible for serializing this function against
1206 * blk_mq_{start,stop}_hw_queue().
1208 bool blk_mq_queue_stopped(struct request_queue *q)
1210 struct blk_mq_hw_ctx *hctx;
1213 queue_for_each_hw_ctx(q, hctx, i)
1214 if (blk_mq_hctx_stopped(hctx))
1219 EXPORT_SYMBOL(blk_mq_queue_stopped);
1221 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1223 cancel_delayed_work_sync(&hctx->run_work);
1224 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1226 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1228 void blk_mq_stop_hw_queues(struct request_queue *q)
1230 struct blk_mq_hw_ctx *hctx;
1233 queue_for_each_hw_ctx(q, hctx, i)
1234 blk_mq_stop_hw_queue(hctx);
1236 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1238 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1240 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1242 blk_mq_run_hw_queue(hctx, false);
1244 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1246 void blk_mq_start_hw_queues(struct request_queue *q)
1248 struct blk_mq_hw_ctx *hctx;
1251 queue_for_each_hw_ctx(q, hctx, i)
1252 blk_mq_start_hw_queue(hctx);
1254 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1256 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1258 if (!blk_mq_hctx_stopped(hctx))
1261 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1262 blk_mq_run_hw_queue(hctx, async);
1264 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1266 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1268 struct blk_mq_hw_ctx *hctx;
1271 queue_for_each_hw_ctx(q, hctx, i)
1272 blk_mq_start_stopped_hw_queue(hctx, async);
1274 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1276 static void blk_mq_run_work_fn(struct work_struct *work)
1278 struct blk_mq_hw_ctx *hctx;
1280 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1283 * If we are stopped, don't run the queue. The exception is if
1284 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1285 * the STOPPED bit and run it.
1287 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1288 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1291 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1292 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1295 __blk_mq_run_hw_queue(hctx);
1299 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1301 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1305 * Stop the hw queue, then modify currently delayed work.
1306 * This should prevent us from running the queue prematurely.
1307 * Mark the queue as auto-clearing STOPPED when it runs.
1309 blk_mq_stop_hw_queue(hctx);
1310 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1311 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1313 msecs_to_jiffies(msecs));
1315 EXPORT_SYMBOL(blk_mq_delay_queue);
1317 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1321 struct blk_mq_ctx *ctx = rq->mq_ctx;
1323 trace_block_rq_insert(hctx->queue, rq);
1326 list_add(&rq->queuelist, &ctx->rq_list);
1328 list_add_tail(&rq->queuelist, &ctx->rq_list);
1331 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1334 struct blk_mq_ctx *ctx = rq->mq_ctx;
1336 __blk_mq_insert_req_list(hctx, rq, at_head);
1337 blk_mq_hctx_mark_pending(hctx, ctx);
1340 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1341 struct list_head *list)
1345 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1348 spin_lock(&ctx->lock);
1349 while (!list_empty(list)) {
1352 rq = list_first_entry(list, struct request, queuelist);
1353 BUG_ON(rq->mq_ctx != ctx);
1354 list_del_init(&rq->queuelist);
1355 __blk_mq_insert_req_list(hctx, rq, false);
1357 blk_mq_hctx_mark_pending(hctx, ctx);
1358 spin_unlock(&ctx->lock);
1361 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1363 struct request *rqa = container_of(a, struct request, queuelist);
1364 struct request *rqb = container_of(b, struct request, queuelist);
1366 return !(rqa->mq_ctx < rqb->mq_ctx ||
1367 (rqa->mq_ctx == rqb->mq_ctx &&
1368 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1371 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1373 struct blk_mq_ctx *this_ctx;
1374 struct request_queue *this_q;
1377 LIST_HEAD(ctx_list);
1380 list_splice_init(&plug->mq_list, &list);
1382 list_sort(NULL, &list, plug_ctx_cmp);
1388 while (!list_empty(&list)) {
1389 rq = list_entry_rq(list.next);
1390 list_del_init(&rq->queuelist);
1392 if (rq->mq_ctx != this_ctx) {
1394 trace_block_unplug(this_q, depth, from_schedule);
1395 blk_mq_sched_insert_requests(this_q, this_ctx,
1400 this_ctx = rq->mq_ctx;
1406 list_add_tail(&rq->queuelist, &ctx_list);
1410 * If 'this_ctx' is set, we know we have entries to complete
1411 * on 'ctx_list'. Do those.
1414 trace_block_unplug(this_q, depth, from_schedule);
1415 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1420 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1422 blk_init_request_from_bio(rq, bio);
1424 blk_account_io_start(rq, true);
1427 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1429 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1430 !blk_queue_nomerges(hctx->queue);
1433 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1434 struct blk_mq_ctx *ctx,
1435 struct request *rq, struct bio *bio)
1437 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1438 blk_mq_bio_to_request(rq, bio);
1439 spin_lock(&ctx->lock);
1441 __blk_mq_insert_request(hctx, rq, false);
1442 spin_unlock(&ctx->lock);
1445 struct request_queue *q = hctx->queue;
1447 spin_lock(&ctx->lock);
1448 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1449 blk_mq_bio_to_request(rq, bio);
1453 spin_unlock(&ctx->lock);
1454 __blk_mq_finish_request(hctx, ctx, rq);
1459 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1462 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1464 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1467 static void __blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie,
1470 struct request_queue *q = rq->q;
1471 struct blk_mq_queue_data bd = {
1475 struct blk_mq_hw_ctx *hctx;
1476 blk_qc_t new_cookie;
1482 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1485 new_cookie = request_to_qc_t(hctx, rq);
1488 * For OK queue, we are done. For error, kill it. Any other
1489 * error (busy), just add it to our list as we previously
1492 ret = q->mq_ops->queue_rq(hctx, &bd);
1493 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1494 *cookie = new_cookie;
1498 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1499 *cookie = BLK_QC_T_NONE;
1500 blk_mq_end_request(rq, -EIO);
1504 __blk_mq_requeue_request(rq);
1506 blk_mq_sched_insert_request(rq, false, true, false, may_sleep);
1509 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1510 struct request *rq, blk_qc_t *cookie)
1512 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1514 __blk_mq_try_issue_directly(rq, cookie, false);
1517 unsigned int srcu_idx;
1521 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1522 __blk_mq_try_issue_directly(rq, cookie, true);
1523 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1527 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1529 const int is_sync = op_is_sync(bio->bi_opf);
1530 const int is_flush_fua = op_is_flush(bio->bi_opf);
1531 struct blk_mq_alloc_data data = { .flags = 0 };
1533 unsigned int request_count = 0;
1534 struct blk_plug *plug;
1535 struct request *same_queue_rq = NULL;
1537 unsigned int wb_acct;
1539 blk_queue_bounce(q, &bio);
1541 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1543 return BLK_QC_T_NONE;
1546 blk_queue_split(q, &bio, q->bio_split);
1548 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1549 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1550 return BLK_QC_T_NONE;
1552 if (blk_mq_sched_bio_merge(q, bio))
1553 return BLK_QC_T_NONE;
1555 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1557 trace_block_getrq(q, bio, bio->bi_opf);
1559 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1560 if (unlikely(!rq)) {
1561 __wbt_done(q->rq_wb, wb_acct);
1562 return BLK_QC_T_NONE;
1565 wbt_track(&rq->issue_stat, wb_acct);
1567 cookie = request_to_qc_t(data.hctx, rq);
1569 plug = current->plug;
1570 if (unlikely(is_flush_fua)) {
1571 blk_mq_put_ctx(data.ctx);
1572 blk_mq_bio_to_request(rq, bio);
1574 blk_mq_sched_insert_request(rq, false, true, true,
1577 blk_insert_flush(rq);
1578 blk_mq_run_hw_queue(data.hctx, true);
1580 } else if (plug && q->nr_hw_queues == 1) {
1581 struct request *last = NULL;
1583 blk_mq_put_ctx(data.ctx);
1584 blk_mq_bio_to_request(rq, bio);
1587 * @request_count may become stale because of schedule
1588 * out, so check the list again.
1590 if (list_empty(&plug->mq_list))
1592 else if (blk_queue_nomerges(q))
1593 request_count = blk_plug_queued_count(q);
1596 trace_block_plug(q);
1598 last = list_entry_rq(plug->mq_list.prev);
1600 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1601 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1602 blk_flush_plug_list(plug, false);
1603 trace_block_plug(q);
1606 list_add_tail(&rq->queuelist, &plug->mq_list);
1607 } else if (plug && !blk_queue_nomerges(q)) {
1608 blk_mq_bio_to_request(rq, bio);
1611 * We do limited plugging. If the bio can be merged, do that.
1612 * Otherwise the existing request in the plug list will be
1613 * issued. So the plug list will have one request at most
1614 * The plug list might get flushed before this. If that happens,
1615 * the plug list is empty, and same_queue_rq is invalid.
1617 if (list_empty(&plug->mq_list))
1618 same_queue_rq = NULL;
1620 list_del_init(&same_queue_rq->queuelist);
1621 list_add_tail(&rq->queuelist, &plug->mq_list);
1623 blk_mq_put_ctx(data.ctx);
1626 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1628 } else if (q->nr_hw_queues > 1 && is_sync) {
1629 blk_mq_put_ctx(data.ctx);
1630 blk_mq_bio_to_request(rq, bio);
1631 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1632 } else if (q->elevator) {
1633 blk_mq_put_ctx(data.ctx);
1634 blk_mq_bio_to_request(rq, bio);
1635 blk_mq_sched_insert_request(rq, false, true, true, true);
1636 } else if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1637 blk_mq_put_ctx(data.ctx);
1638 blk_mq_run_hw_queue(data.hctx, true);
1640 blk_mq_put_ctx(data.ctx);
1645 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1646 unsigned int hctx_idx)
1650 if (tags->rqs && set->ops->exit_request) {
1653 for (i = 0; i < tags->nr_tags; i++) {
1654 struct request *rq = tags->static_rqs[i];
1658 set->ops->exit_request(set->driver_data, rq,
1660 tags->static_rqs[i] = NULL;
1664 while (!list_empty(&tags->page_list)) {
1665 page = list_first_entry(&tags->page_list, struct page, lru);
1666 list_del_init(&page->lru);
1668 * Remove kmemleak object previously allocated in
1669 * blk_mq_init_rq_map().
1671 kmemleak_free(page_address(page));
1672 __free_pages(page, page->private);
1676 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1680 kfree(tags->static_rqs);
1681 tags->static_rqs = NULL;
1683 blk_mq_free_tags(tags);
1686 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1687 unsigned int hctx_idx,
1688 unsigned int nr_tags,
1689 unsigned int reserved_tags)
1691 struct blk_mq_tags *tags;
1694 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1695 if (node == NUMA_NO_NODE)
1696 node = set->numa_node;
1698 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1699 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1703 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1704 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1707 blk_mq_free_tags(tags);
1711 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1712 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1714 if (!tags->static_rqs) {
1716 blk_mq_free_tags(tags);
1723 static size_t order_to_size(unsigned int order)
1725 return (size_t)PAGE_SIZE << order;
1728 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1729 unsigned int hctx_idx, unsigned int depth)
1731 unsigned int i, j, entries_per_page, max_order = 4;
1732 size_t rq_size, left;
1735 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1736 if (node == NUMA_NO_NODE)
1737 node = set->numa_node;
1739 INIT_LIST_HEAD(&tags->page_list);
1742 * rq_size is the size of the request plus driver payload, rounded
1743 * to the cacheline size
1745 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1747 left = rq_size * depth;
1749 for (i = 0; i < depth; ) {
1750 int this_order = max_order;
1755 while (this_order && left < order_to_size(this_order - 1))
1759 page = alloc_pages_node(node,
1760 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1766 if (order_to_size(this_order) < rq_size)
1773 page->private = this_order;
1774 list_add_tail(&page->lru, &tags->page_list);
1776 p = page_address(page);
1778 * Allow kmemleak to scan these pages as they contain pointers
1779 * to additional allocations like via ops->init_request().
1781 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1782 entries_per_page = order_to_size(this_order) / rq_size;
1783 to_do = min(entries_per_page, depth - i);
1784 left -= to_do * rq_size;
1785 for (j = 0; j < to_do; j++) {
1786 struct request *rq = p;
1788 tags->static_rqs[i] = rq;
1789 if (set->ops->init_request) {
1790 if (set->ops->init_request(set->driver_data,
1793 tags->static_rqs[i] = NULL;
1805 blk_mq_free_rqs(set, tags, hctx_idx);
1810 * 'cpu' is going away. splice any existing rq_list entries from this
1811 * software queue to the hw queue dispatch list, and ensure that it
1814 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1816 struct blk_mq_hw_ctx *hctx;
1817 struct blk_mq_ctx *ctx;
1820 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1821 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1823 spin_lock(&ctx->lock);
1824 if (!list_empty(&ctx->rq_list)) {
1825 list_splice_init(&ctx->rq_list, &tmp);
1826 blk_mq_hctx_clear_pending(hctx, ctx);
1828 spin_unlock(&ctx->lock);
1830 if (list_empty(&tmp))
1833 spin_lock(&hctx->lock);
1834 list_splice_tail_init(&tmp, &hctx->dispatch);
1835 spin_unlock(&hctx->lock);
1837 blk_mq_run_hw_queue(hctx, true);
1841 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1843 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1847 /* hctx->ctxs will be freed in queue's release handler */
1848 static void blk_mq_exit_hctx(struct request_queue *q,
1849 struct blk_mq_tag_set *set,
1850 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1852 unsigned flush_start_tag = set->queue_depth;
1854 blk_mq_tag_idle(hctx);
1856 if (set->ops->exit_request)
1857 set->ops->exit_request(set->driver_data,
1858 hctx->fq->flush_rq, hctx_idx,
1859 flush_start_tag + hctx_idx);
1861 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1863 if (set->ops->exit_hctx)
1864 set->ops->exit_hctx(hctx, hctx_idx);
1866 if (hctx->flags & BLK_MQ_F_BLOCKING)
1867 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1869 blk_mq_remove_cpuhp(hctx);
1870 blk_free_flush_queue(hctx->fq);
1871 sbitmap_free(&hctx->ctx_map);
1874 static void blk_mq_exit_hw_queues(struct request_queue *q,
1875 struct blk_mq_tag_set *set, int nr_queue)
1877 struct blk_mq_hw_ctx *hctx;
1880 queue_for_each_hw_ctx(q, hctx, i) {
1883 blk_mq_exit_hctx(q, set, hctx, i);
1887 static int blk_mq_init_hctx(struct request_queue *q,
1888 struct blk_mq_tag_set *set,
1889 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1892 unsigned flush_start_tag = set->queue_depth;
1894 node = hctx->numa_node;
1895 if (node == NUMA_NO_NODE)
1896 node = hctx->numa_node = set->numa_node;
1898 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1899 spin_lock_init(&hctx->lock);
1900 INIT_LIST_HEAD(&hctx->dispatch);
1902 hctx->queue_num = hctx_idx;
1903 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1905 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1907 hctx->tags = set->tags[hctx_idx];
1910 * Allocate space for all possible cpus to avoid allocation at
1913 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1916 goto unregister_cpu_notifier;
1918 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1924 if (set->ops->init_hctx &&
1925 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1928 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1931 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1933 goto sched_exit_hctx;
1935 if (set->ops->init_request &&
1936 set->ops->init_request(set->driver_data,
1937 hctx->fq->flush_rq, hctx_idx,
1938 flush_start_tag + hctx_idx, node))
1941 if (hctx->flags & BLK_MQ_F_BLOCKING)
1942 init_srcu_struct(&hctx->queue_rq_srcu);
1949 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1951 if (set->ops->exit_hctx)
1952 set->ops->exit_hctx(hctx, hctx_idx);
1954 sbitmap_free(&hctx->ctx_map);
1957 unregister_cpu_notifier:
1958 blk_mq_remove_cpuhp(hctx);
1962 static void blk_mq_init_cpu_queues(struct request_queue *q,
1963 unsigned int nr_hw_queues)
1967 for_each_possible_cpu(i) {
1968 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1969 struct blk_mq_hw_ctx *hctx;
1972 spin_lock_init(&__ctx->lock);
1973 INIT_LIST_HEAD(&__ctx->rq_list);
1976 /* If the cpu isn't online, the cpu is mapped to first hctx */
1980 hctx = blk_mq_map_queue(q, i);
1983 * Set local node, IFF we have more than one hw queue. If
1984 * not, we remain on the home node of the device
1986 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1987 hctx->numa_node = local_memory_node(cpu_to_node(i));
1991 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1995 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1996 set->queue_depth, set->reserved_tags);
1997 if (!set->tags[hctx_idx])
2000 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2005 blk_mq_free_rq_map(set->tags[hctx_idx]);
2006 set->tags[hctx_idx] = NULL;
2010 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2011 unsigned int hctx_idx)
2013 if (set->tags[hctx_idx]) {
2014 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2015 blk_mq_free_rq_map(set->tags[hctx_idx]);
2016 set->tags[hctx_idx] = NULL;
2020 static void blk_mq_map_swqueue(struct request_queue *q,
2021 const struct cpumask *online_mask)
2023 unsigned int i, hctx_idx;
2024 struct blk_mq_hw_ctx *hctx;
2025 struct blk_mq_ctx *ctx;
2026 struct blk_mq_tag_set *set = q->tag_set;
2029 * Avoid others reading imcomplete hctx->cpumask through sysfs
2031 mutex_lock(&q->sysfs_lock);
2033 queue_for_each_hw_ctx(q, hctx, i) {
2034 cpumask_clear(hctx->cpumask);
2039 * Map software to hardware queues
2041 for_each_possible_cpu(i) {
2042 /* If the cpu isn't online, the cpu is mapped to first hctx */
2043 if (!cpumask_test_cpu(i, online_mask))
2046 hctx_idx = q->mq_map[i];
2047 /* unmapped hw queue can be remapped after CPU topo changed */
2048 if (!set->tags[hctx_idx] &&
2049 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2051 * If tags initialization fail for some hctx,
2052 * that hctx won't be brought online. In this
2053 * case, remap the current ctx to hctx[0] which
2054 * is guaranteed to always have tags allocated
2059 ctx = per_cpu_ptr(q->queue_ctx, i);
2060 hctx = blk_mq_map_queue(q, i);
2062 cpumask_set_cpu(i, hctx->cpumask);
2063 ctx->index_hw = hctx->nr_ctx;
2064 hctx->ctxs[hctx->nr_ctx++] = ctx;
2067 mutex_unlock(&q->sysfs_lock);
2069 queue_for_each_hw_ctx(q, hctx, i) {
2071 * If no software queues are mapped to this hardware queue,
2072 * disable it and free the request entries.
2074 if (!hctx->nr_ctx) {
2075 /* Never unmap queue 0. We need it as a
2076 * fallback in case of a new remap fails
2079 if (i && set->tags[i])
2080 blk_mq_free_map_and_requests(set, i);
2086 hctx->tags = set->tags[i];
2087 WARN_ON(!hctx->tags);
2090 * Set the map size to the number of mapped software queues.
2091 * This is more accurate and more efficient than looping
2092 * over all possibly mapped software queues.
2094 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2097 * Initialize batch roundrobin counts
2099 hctx->next_cpu = cpumask_first(hctx->cpumask);
2100 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2104 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2106 struct blk_mq_hw_ctx *hctx;
2109 queue_for_each_hw_ctx(q, hctx, i) {
2111 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2113 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2117 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2119 struct request_queue *q;
2121 lockdep_assert_held(&set->tag_list_lock);
2123 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2124 blk_mq_freeze_queue(q);
2125 queue_set_hctx_shared(q, shared);
2126 blk_mq_unfreeze_queue(q);
2130 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2132 struct blk_mq_tag_set *set = q->tag_set;
2134 mutex_lock(&set->tag_list_lock);
2135 list_del_rcu(&q->tag_set_list);
2136 INIT_LIST_HEAD(&q->tag_set_list);
2137 if (list_is_singular(&set->tag_list)) {
2138 /* just transitioned to unshared */
2139 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2140 /* update existing queue */
2141 blk_mq_update_tag_set_depth(set, false);
2143 mutex_unlock(&set->tag_list_lock);
2148 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2149 struct request_queue *q)
2153 mutex_lock(&set->tag_list_lock);
2155 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2156 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2157 set->flags |= BLK_MQ_F_TAG_SHARED;
2158 /* update existing queue */
2159 blk_mq_update_tag_set_depth(set, true);
2161 if (set->flags & BLK_MQ_F_TAG_SHARED)
2162 queue_set_hctx_shared(q, true);
2163 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2165 mutex_unlock(&set->tag_list_lock);
2169 * It is the actual release handler for mq, but we do it from
2170 * request queue's release handler for avoiding use-after-free
2171 * and headache because q->mq_kobj shouldn't have been introduced,
2172 * but we can't group ctx/kctx kobj without it.
2174 void blk_mq_release(struct request_queue *q)
2176 struct blk_mq_hw_ctx *hctx;
2179 /* hctx kobj stays in hctx */
2180 queue_for_each_hw_ctx(q, hctx, i) {
2183 kobject_put(&hctx->kobj);
2188 kfree(q->queue_hw_ctx);
2191 * release .mq_kobj and sw queue's kobject now because
2192 * both share lifetime with request queue.
2194 blk_mq_sysfs_deinit(q);
2196 free_percpu(q->queue_ctx);
2199 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2201 struct request_queue *uninit_q, *q;
2203 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2205 return ERR_PTR(-ENOMEM);
2207 q = blk_mq_init_allocated_queue(set, uninit_q);
2209 blk_cleanup_queue(uninit_q);
2213 EXPORT_SYMBOL(blk_mq_init_queue);
2215 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2216 struct request_queue *q)
2219 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2221 blk_mq_sysfs_unregister(q);
2222 for (i = 0; i < set->nr_hw_queues; i++) {
2228 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2229 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2234 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2241 atomic_set(&hctxs[i]->nr_active, 0);
2242 hctxs[i]->numa_node = node;
2243 hctxs[i]->queue_num = i;
2245 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2246 free_cpumask_var(hctxs[i]->cpumask);
2251 blk_mq_hctx_kobj_init(hctxs[i]);
2253 for (j = i; j < q->nr_hw_queues; j++) {
2254 struct blk_mq_hw_ctx *hctx = hctxs[j];
2258 blk_mq_free_map_and_requests(set, j);
2259 blk_mq_exit_hctx(q, set, hctx, j);
2260 kobject_put(&hctx->kobj);
2265 q->nr_hw_queues = i;
2266 blk_mq_sysfs_register(q);
2269 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2270 struct request_queue *q)
2272 /* mark the queue as mq asap */
2273 q->mq_ops = set->ops;
2275 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2276 blk_mq_poll_stats_bkt,
2277 BLK_MQ_POLL_STATS_BKTS, q);
2281 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2285 /* init q->mq_kobj and sw queues' kobjects */
2286 blk_mq_sysfs_init(q);
2288 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2289 GFP_KERNEL, set->numa_node);
2290 if (!q->queue_hw_ctx)
2293 q->mq_map = set->mq_map;
2295 blk_mq_realloc_hw_ctxs(set, q);
2296 if (!q->nr_hw_queues)
2299 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2300 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2302 q->nr_queues = nr_cpu_ids;
2304 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2306 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2307 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2309 q->sg_reserved_size = INT_MAX;
2311 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2312 INIT_LIST_HEAD(&q->requeue_list);
2313 spin_lock_init(&q->requeue_lock);
2315 blk_queue_make_request(q, blk_mq_make_request);
2318 * Do this after blk_queue_make_request() overrides it...
2320 q->nr_requests = set->queue_depth;
2323 * Default to classic polling
2327 if (set->ops->complete)
2328 blk_queue_softirq_done(q, set->ops->complete);
2330 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2333 mutex_lock(&all_q_mutex);
2335 list_add_tail(&q->all_q_node, &all_q_list);
2336 blk_mq_add_queue_tag_set(set, q);
2337 blk_mq_map_swqueue(q, cpu_online_mask);
2339 mutex_unlock(&all_q_mutex);
2342 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2345 ret = blk_mq_sched_init(q);
2347 return ERR_PTR(ret);
2353 kfree(q->queue_hw_ctx);
2355 free_percpu(q->queue_ctx);
2358 return ERR_PTR(-ENOMEM);
2360 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2362 void blk_mq_free_queue(struct request_queue *q)
2364 struct blk_mq_tag_set *set = q->tag_set;
2366 mutex_lock(&all_q_mutex);
2367 list_del_init(&q->all_q_node);
2368 mutex_unlock(&all_q_mutex);
2370 blk_mq_del_queue_tag_set(q);
2372 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2375 /* Basically redo blk_mq_init_queue with queue frozen */
2376 static void blk_mq_queue_reinit(struct request_queue *q,
2377 const struct cpumask *online_mask)
2379 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2381 blk_mq_sysfs_unregister(q);
2384 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2385 * we should change hctx numa_node according to new topology (this
2386 * involves free and re-allocate memory, worthy doing?)
2389 blk_mq_map_swqueue(q, online_mask);
2391 blk_mq_sysfs_register(q);
2395 * New online cpumask which is going to be set in this hotplug event.
2396 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2397 * one-by-one and dynamically allocating this could result in a failure.
2399 static struct cpumask cpuhp_online_new;
2401 static void blk_mq_queue_reinit_work(void)
2403 struct request_queue *q;
2405 mutex_lock(&all_q_mutex);
2407 * We need to freeze and reinit all existing queues. Freezing
2408 * involves synchronous wait for an RCU grace period and doing it
2409 * one by one may take a long time. Start freezing all queues in
2410 * one swoop and then wait for the completions so that freezing can
2411 * take place in parallel.
2413 list_for_each_entry(q, &all_q_list, all_q_node)
2414 blk_freeze_queue_start(q);
2415 list_for_each_entry(q, &all_q_list, all_q_node)
2416 blk_mq_freeze_queue_wait(q);
2418 list_for_each_entry(q, &all_q_list, all_q_node)
2419 blk_mq_queue_reinit(q, &cpuhp_online_new);
2421 list_for_each_entry(q, &all_q_list, all_q_node)
2422 blk_mq_unfreeze_queue(q);
2424 mutex_unlock(&all_q_mutex);
2427 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2429 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2430 blk_mq_queue_reinit_work();
2435 * Before hotadded cpu starts handling requests, new mappings must be
2436 * established. Otherwise, these requests in hw queue might never be
2439 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2440 * for CPU0, and ctx1 for CPU1).
2442 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2443 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2445 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2446 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2447 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2450 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2452 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2453 cpumask_set_cpu(cpu, &cpuhp_online_new);
2454 blk_mq_queue_reinit_work();
2458 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2462 for (i = 0; i < set->nr_hw_queues; i++)
2463 if (!__blk_mq_alloc_rq_map(set, i))
2470 blk_mq_free_rq_map(set->tags[i]);
2476 * Allocate the request maps associated with this tag_set. Note that this
2477 * may reduce the depth asked for, if memory is tight. set->queue_depth
2478 * will be updated to reflect the allocated depth.
2480 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2485 depth = set->queue_depth;
2487 err = __blk_mq_alloc_rq_maps(set);
2491 set->queue_depth >>= 1;
2492 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2496 } while (set->queue_depth);
2498 if (!set->queue_depth || err) {
2499 pr_err("blk-mq: failed to allocate request map\n");
2503 if (depth != set->queue_depth)
2504 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2505 depth, set->queue_depth);
2510 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2512 if (set->ops->map_queues)
2513 return set->ops->map_queues(set);
2515 return blk_mq_map_queues(set);
2519 * Alloc a tag set to be associated with one or more request queues.
2520 * May fail with EINVAL for various error conditions. May adjust the
2521 * requested depth down, if if it too large. In that case, the set
2522 * value will be stored in set->queue_depth.
2524 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2528 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2530 if (!set->nr_hw_queues)
2532 if (!set->queue_depth)
2534 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2537 if (!set->ops->queue_rq)
2540 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2541 pr_info("blk-mq: reduced tag depth to %u\n",
2543 set->queue_depth = BLK_MQ_MAX_DEPTH;
2547 * If a crashdump is active, then we are potentially in a very
2548 * memory constrained environment. Limit us to 1 queue and
2549 * 64 tags to prevent using too much memory.
2551 if (is_kdump_kernel()) {
2552 set->nr_hw_queues = 1;
2553 set->queue_depth = min(64U, set->queue_depth);
2556 * There is no use for more h/w queues than cpus.
2558 if (set->nr_hw_queues > nr_cpu_ids)
2559 set->nr_hw_queues = nr_cpu_ids;
2561 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2562 GFP_KERNEL, set->numa_node);
2567 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2568 GFP_KERNEL, set->numa_node);
2572 ret = blk_mq_update_queue_map(set);
2574 goto out_free_mq_map;
2576 ret = blk_mq_alloc_rq_maps(set);
2578 goto out_free_mq_map;
2580 mutex_init(&set->tag_list_lock);
2581 INIT_LIST_HEAD(&set->tag_list);
2593 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2595 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2599 for (i = 0; i < nr_cpu_ids; i++)
2600 blk_mq_free_map_and_requests(set, i);
2608 EXPORT_SYMBOL(blk_mq_free_tag_set);
2610 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2612 struct blk_mq_tag_set *set = q->tag_set;
2613 struct blk_mq_hw_ctx *hctx;
2619 blk_mq_freeze_queue(q);
2620 blk_mq_quiesce_queue(q);
2623 queue_for_each_hw_ctx(q, hctx, i) {
2627 * If we're using an MQ scheduler, just update the scheduler
2628 * queue depth. This is similar to what the old code would do.
2630 if (!hctx->sched_tags) {
2631 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2632 min(nr, set->queue_depth),
2635 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2643 q->nr_requests = nr;
2645 blk_mq_unfreeze_queue(q);
2646 blk_mq_start_stopped_hw_queues(q, true);
2651 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2653 struct request_queue *q;
2655 lockdep_assert_held(&set->tag_list_lock);
2657 if (nr_hw_queues > nr_cpu_ids)
2658 nr_hw_queues = nr_cpu_ids;
2659 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2662 list_for_each_entry(q, &set->tag_list, tag_set_list)
2663 blk_mq_freeze_queue(q);
2665 set->nr_hw_queues = nr_hw_queues;
2666 blk_mq_update_queue_map(set);
2667 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2668 blk_mq_realloc_hw_ctxs(set, q);
2669 blk_mq_queue_reinit(q, cpu_online_mask);
2672 list_for_each_entry(q, &set->tag_list, tag_set_list)
2673 blk_mq_unfreeze_queue(q);
2675 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2677 /* Enable polling stats and return whether they were already enabled. */
2678 static bool blk_poll_stats_enable(struct request_queue *q)
2680 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2681 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2683 blk_stat_add_callback(q, q->poll_cb);
2687 static void blk_mq_poll_stats_start(struct request_queue *q)
2690 * We don't arm the callback if polling stats are not enabled or the
2691 * callback is already active.
2693 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2694 blk_stat_is_active(q->poll_cb))
2697 blk_stat_activate_msecs(q->poll_cb, 100);
2700 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2702 struct request_queue *q = cb->data;
2705 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2706 if (cb->stat[bucket].nr_samples)
2707 q->poll_stat[bucket] = cb->stat[bucket];
2711 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2712 struct blk_mq_hw_ctx *hctx,
2715 unsigned long ret = 0;
2719 * If stats collection isn't on, don't sleep but turn it on for
2722 if (!blk_poll_stats_enable(q))
2726 * As an optimistic guess, use half of the mean service time
2727 * for this type of request. We can (and should) make this smarter.
2728 * For instance, if the completion latencies are tight, we can
2729 * get closer than just half the mean. This is especially
2730 * important on devices where the completion latencies are longer
2731 * than ~10 usec. We do use the stats for the relevant IO size
2732 * if available which does lead to better estimates.
2734 bucket = blk_mq_poll_stats_bkt(rq);
2738 if (q->poll_stat[bucket].nr_samples)
2739 ret = (q->poll_stat[bucket].mean + 1) / 2;
2744 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2745 struct blk_mq_hw_ctx *hctx,
2748 struct hrtimer_sleeper hs;
2749 enum hrtimer_mode mode;
2753 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2759 * -1: don't ever hybrid sleep
2760 * 0: use half of prev avg
2761 * >0: use this specific value
2763 if (q->poll_nsec == -1)
2765 else if (q->poll_nsec > 0)
2766 nsecs = q->poll_nsec;
2768 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2773 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2776 * This will be replaced with the stats tracking code, using
2777 * 'avg_completion_time / 2' as the pre-sleep target.
2781 mode = HRTIMER_MODE_REL;
2782 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2783 hrtimer_set_expires(&hs.timer, kt);
2785 hrtimer_init_sleeper(&hs, current);
2787 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2789 set_current_state(TASK_UNINTERRUPTIBLE);
2790 hrtimer_start_expires(&hs.timer, mode);
2793 hrtimer_cancel(&hs.timer);
2794 mode = HRTIMER_MODE_ABS;
2795 } while (hs.task && !signal_pending(current));
2797 __set_current_state(TASK_RUNNING);
2798 destroy_hrtimer_on_stack(&hs.timer);
2802 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2804 struct request_queue *q = hctx->queue;
2808 * If we sleep, have the caller restart the poll loop to reset
2809 * the state. Like for the other success return cases, the
2810 * caller is responsible for checking if the IO completed. If
2811 * the IO isn't complete, we'll get called again and will go
2812 * straight to the busy poll loop.
2814 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2817 hctx->poll_considered++;
2819 state = current->state;
2820 while (!need_resched()) {
2823 hctx->poll_invoked++;
2825 ret = q->mq_ops->poll(hctx, rq->tag);
2827 hctx->poll_success++;
2828 set_current_state(TASK_RUNNING);
2832 if (signal_pending_state(state, current))
2833 set_current_state(TASK_RUNNING);
2835 if (current->state == TASK_RUNNING)
2845 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2847 struct blk_mq_hw_ctx *hctx;
2848 struct blk_plug *plug;
2851 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2852 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2855 plug = current->plug;
2857 blk_flush_plug_list(plug, false);
2859 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2860 if (!blk_qc_t_is_internal(cookie))
2861 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2863 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2865 * With scheduling, if the request has completed, we'll
2866 * get a NULL return here, as we clear the sched tag when
2867 * that happens. The request still remains valid, like always,
2868 * so we should be safe with just the NULL check.
2874 return __blk_mq_poll(hctx, rq);
2876 EXPORT_SYMBOL_GPL(blk_mq_poll);
2878 void blk_mq_disable_hotplug(void)
2880 mutex_lock(&all_q_mutex);
2883 void blk_mq_enable_hotplug(void)
2885 mutex_unlock(&all_q_mutex);
2888 static int __init blk_mq_init(void)
2890 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2891 blk_mq_hctx_notify_dead);
2893 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2894 blk_mq_queue_reinit_prepare,
2895 blk_mq_queue_reinit_dead);
2898 subsys_initcall(blk_mq_init);