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 DEFINE_MUTEX(all_q_mutex);
41 static LIST_HEAD(all_q_list);
43 static void blk_mq_poll_stats_start(struct request_queue *q);
44 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
45 static void __blk_mq_stop_hw_queues(struct request_queue *q, bool sync);
47 static int blk_mq_poll_stats_bkt(const struct request *rq)
49 int ddir, bytes, bucket;
51 ddir = rq_data_dir(rq);
52 bytes = blk_rq_bytes(rq);
54 bucket = ddir + 2*(ilog2(bytes) - 9);
58 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
59 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
65 * Check if any of the ctx's have pending work in this hardware queue
67 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
69 return sbitmap_any_bit_set(&hctx->ctx_map) ||
70 !list_empty_careful(&hctx->dispatch) ||
71 blk_mq_sched_has_work(hctx);
75 * Mark this ctx as having pending work in this hardware queue
77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
78 struct blk_mq_ctx *ctx)
80 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
81 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
84 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
85 struct blk_mq_ctx *ctx)
87 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
90 void blk_freeze_queue_start(struct request_queue *q)
94 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
95 if (freeze_depth == 1) {
96 percpu_ref_kill(&q->q_usage_counter);
97 blk_mq_run_hw_queues(q, false);
100 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
102 void blk_mq_freeze_queue_wait(struct request_queue *q)
104 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
106 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
108 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
109 unsigned long timeout)
111 return wait_event_timeout(q->mq_freeze_wq,
112 percpu_ref_is_zero(&q->q_usage_counter),
115 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
118 * Guarantee no request is in use, so we can change any data structure of
119 * the queue afterward.
121 void blk_freeze_queue(struct request_queue *q)
124 * In the !blk_mq case we are only calling this to kill the
125 * q_usage_counter, otherwise this increases the freeze depth
126 * and waits for it to return to zero. For this reason there is
127 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
128 * exported to drivers as the only user for unfreeze is blk_mq.
130 blk_freeze_queue_start(q);
131 blk_mq_freeze_queue_wait(q);
134 void blk_mq_freeze_queue(struct request_queue *q)
137 * ...just an alias to keep freeze and unfreeze actions balanced
138 * in the blk_mq_* namespace
142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
144 void blk_mq_unfreeze_queue(struct request_queue *q)
148 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
149 WARN_ON_ONCE(freeze_depth < 0);
151 percpu_ref_reinit(&q->q_usage_counter);
152 wake_up_all(&q->mq_freeze_wq);
155 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
158 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
161 * Note: this function does not prevent that the struct request end_io()
162 * callback function is invoked. Additionally, it is not prevented that
163 * new queue_rq() calls occur unless the queue has been stopped first.
165 void blk_mq_quiesce_queue(struct request_queue *q)
167 struct blk_mq_hw_ctx *hctx;
171 __blk_mq_stop_hw_queues(q, true);
173 queue_for_each_hw_ctx(q, hctx, i) {
174 if (hctx->flags & BLK_MQ_F_BLOCKING)
175 synchronize_srcu(&hctx->queue_rq_srcu);
182 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
185 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
188 * This function recovers queue into the state before quiescing
189 * which is done by blk_mq_quiesce_queue.
191 void blk_mq_unquiesce_queue(struct request_queue *q)
193 blk_mq_start_stopped_hw_queues(q, true);
195 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
197 void blk_mq_wake_waiters(struct request_queue *q)
199 struct blk_mq_hw_ctx *hctx;
202 queue_for_each_hw_ctx(q, hctx, i)
203 if (blk_mq_hw_queue_mapped(hctx))
204 blk_mq_tag_wakeup_all(hctx->tags, true);
207 * If we are called because the queue has now been marked as
208 * dying, we need to ensure that processes currently waiting on
209 * the queue are notified as well.
211 wake_up_all(&q->mq_freeze_wq);
214 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
216 return blk_mq_has_free_tags(hctx->tags);
218 EXPORT_SYMBOL(blk_mq_can_queue);
220 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
221 unsigned int tag, unsigned int op)
223 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
224 struct request *rq = tags->static_rqs[tag];
226 if (data->flags & BLK_MQ_REQ_INTERNAL) {
228 rq->internal_tag = tag;
230 if (blk_mq_tag_busy(data->hctx)) {
231 rq->rq_flags = RQF_MQ_INFLIGHT;
232 atomic_inc(&data->hctx->nr_active);
235 rq->internal_tag = -1;
236 data->hctx->tags->rqs[rq->tag] = rq;
239 INIT_LIST_HEAD(&rq->queuelist);
240 /* csd/requeue_work/fifo_time is initialized before use */
242 rq->mq_ctx = data->ctx;
244 if (blk_queue_io_stat(data->q))
245 rq->rq_flags |= RQF_IO_STAT;
246 /* do not touch atomic flags, it needs atomic ops against the timer */
248 INIT_HLIST_NODE(&rq->hash);
249 RB_CLEAR_NODE(&rq->rb_node);
252 rq->start_time = jiffies;
253 #ifdef CONFIG_BLK_CGROUP
255 set_start_time_ns(rq);
256 rq->io_start_time_ns = 0;
258 rq->nr_phys_segments = 0;
259 #if defined(CONFIG_BLK_DEV_INTEGRITY)
260 rq->nr_integrity_segments = 0;
263 /* tag was already set */
266 INIT_LIST_HEAD(&rq->timeout_list);
270 rq->end_io_data = NULL;
273 data->ctx->rq_dispatched[op_is_sync(op)]++;
277 static struct request *blk_mq_get_request(struct request_queue *q,
278 struct bio *bio, unsigned int op,
279 struct blk_mq_alloc_data *data)
281 struct elevator_queue *e = q->elevator;
285 blk_queue_enter_live(q);
287 if (likely(!data->ctx))
288 data->ctx = blk_mq_get_ctx(q);
289 if (likely(!data->hctx))
290 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
293 data->flags |= BLK_MQ_REQ_INTERNAL;
296 * Flush requests are special and go directly to the
299 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
300 e->type->ops.mq.limit_depth(op, data);
303 tag = blk_mq_get_tag(data);
304 if (tag == BLK_MQ_TAG_FAIL) {
309 rq = blk_mq_rq_ctx_init(data, tag, op);
310 if (!op_is_flush(op)) {
312 if (e && e->type->ops.mq.prepare_request) {
313 if (e->type->icq_cache && rq_ioc(bio))
314 blk_mq_sched_assign_ioc(rq, bio);
316 e->type->ops.mq.prepare_request(rq, bio);
317 rq->rq_flags |= RQF_ELVPRIV;
320 data->hctx->queued++;
324 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
327 struct blk_mq_alloc_data alloc_data = { .flags = flags };
331 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
335 rq = blk_mq_get_request(q, NULL, rw, &alloc_data);
337 blk_mq_put_ctx(alloc_data.ctx);
341 return ERR_PTR(-EWOULDBLOCK);
344 rq->__sector = (sector_t) -1;
345 rq->bio = rq->biotail = NULL;
348 EXPORT_SYMBOL(blk_mq_alloc_request);
350 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
351 unsigned int flags, unsigned int hctx_idx)
353 struct blk_mq_alloc_data alloc_data = { .flags = flags };
359 * If the tag allocator sleeps we could get an allocation for a
360 * different hardware context. No need to complicate the low level
361 * allocator for this for the rare use case of a command tied to
364 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
365 return ERR_PTR(-EINVAL);
367 if (hctx_idx >= q->nr_hw_queues)
368 return ERR_PTR(-EIO);
370 ret = blk_queue_enter(q, true);
375 * Check if the hardware context is actually mapped to anything.
376 * If not tell the caller that it should skip this queue.
378 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
379 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
381 return ERR_PTR(-EXDEV);
383 cpu = cpumask_first(alloc_data.hctx->cpumask);
384 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
386 rq = blk_mq_get_request(q, NULL, rw, &alloc_data);
391 return ERR_PTR(-EWOULDBLOCK);
395 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
397 void blk_mq_free_request(struct request *rq)
399 struct request_queue *q = rq->q;
400 struct elevator_queue *e = q->elevator;
401 struct blk_mq_ctx *ctx = rq->mq_ctx;
402 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
403 const int sched_tag = rq->internal_tag;
405 if (rq->rq_flags & RQF_ELVPRIV) {
406 if (e && e->type->ops.mq.finish_request)
407 e->type->ops.mq.finish_request(rq);
409 put_io_context(rq->elv.icq->ioc);
414 ctx->rq_completed[rq_is_sync(rq)]++;
415 if (rq->rq_flags & RQF_MQ_INFLIGHT)
416 atomic_dec(&hctx->nr_active);
418 wbt_done(q->rq_wb, &rq->issue_stat);
421 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
422 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
424 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
426 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
427 blk_mq_sched_restart(hctx);
430 EXPORT_SYMBOL_GPL(blk_mq_free_request);
432 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
434 blk_account_io_done(rq);
437 wbt_done(rq->q->rq_wb, &rq->issue_stat);
438 rq->end_io(rq, error);
440 if (unlikely(blk_bidi_rq(rq)))
441 blk_mq_free_request(rq->next_rq);
442 blk_mq_free_request(rq);
445 EXPORT_SYMBOL(__blk_mq_end_request);
447 void blk_mq_end_request(struct request *rq, blk_status_t error)
449 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
451 __blk_mq_end_request(rq, error);
453 EXPORT_SYMBOL(blk_mq_end_request);
455 static void __blk_mq_complete_request_remote(void *data)
457 struct request *rq = data;
459 rq->q->softirq_done_fn(rq);
462 static void __blk_mq_complete_request(struct request *rq)
464 struct blk_mq_ctx *ctx = rq->mq_ctx;
468 if (rq->internal_tag != -1)
469 blk_mq_sched_completed_request(rq);
470 if (rq->rq_flags & RQF_STATS) {
471 blk_mq_poll_stats_start(rq->q);
475 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
476 rq->q->softirq_done_fn(rq);
481 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
482 shared = cpus_share_cache(cpu, ctx->cpu);
484 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
485 rq->csd.func = __blk_mq_complete_request_remote;
488 smp_call_function_single_async(ctx->cpu, &rq->csd);
490 rq->q->softirq_done_fn(rq);
496 * blk_mq_complete_request - end I/O on a request
497 * @rq: the request being processed
500 * Ends all I/O on a request. It does not handle partial completions.
501 * The actual completion happens out-of-order, through a IPI handler.
503 void blk_mq_complete_request(struct request *rq)
505 struct request_queue *q = rq->q;
507 if (unlikely(blk_should_fake_timeout(q)))
509 if (!blk_mark_rq_complete(rq))
510 __blk_mq_complete_request(rq);
512 EXPORT_SYMBOL(blk_mq_complete_request);
514 int blk_mq_request_started(struct request *rq)
516 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
518 EXPORT_SYMBOL_GPL(blk_mq_request_started);
520 void blk_mq_start_request(struct request *rq)
522 struct request_queue *q = rq->q;
524 blk_mq_sched_started_request(rq);
526 trace_block_rq_issue(q, rq);
528 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
529 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
530 rq->rq_flags |= RQF_STATS;
531 wbt_issue(q->rq_wb, &rq->issue_stat);
537 * Ensure that ->deadline is visible before set the started
538 * flag and clear the completed flag.
540 smp_mb__before_atomic();
543 * Mark us as started and clear complete. Complete might have been
544 * set if requeue raced with timeout, which then marked it as
545 * complete. So be sure to clear complete again when we start
546 * the request, otherwise we'll ignore the completion event.
548 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
549 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
550 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
551 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
553 if (q->dma_drain_size && blk_rq_bytes(rq)) {
555 * Make sure space for the drain appears. We know we can do
556 * this because max_hw_segments has been adjusted to be one
557 * fewer than the device can handle.
559 rq->nr_phys_segments++;
562 EXPORT_SYMBOL(blk_mq_start_request);
565 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
566 * flag isn't set yet, so there may be race with timeout handler,
567 * but given rq->deadline is just set in .queue_rq() under
568 * this situation, the race won't be possible in reality because
569 * rq->timeout should be set as big enough to cover the window
570 * between blk_mq_start_request() called from .queue_rq() and
571 * clearing REQ_ATOM_STARTED here.
573 static void __blk_mq_requeue_request(struct request *rq)
575 struct request_queue *q = rq->q;
577 trace_block_rq_requeue(q, rq);
578 wbt_requeue(q->rq_wb, &rq->issue_stat);
579 blk_mq_sched_requeue_request(rq);
581 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
582 if (q->dma_drain_size && blk_rq_bytes(rq))
583 rq->nr_phys_segments--;
587 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
589 __blk_mq_requeue_request(rq);
591 BUG_ON(blk_queued_rq(rq));
592 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
594 EXPORT_SYMBOL(blk_mq_requeue_request);
596 static void blk_mq_requeue_work(struct work_struct *work)
598 struct request_queue *q =
599 container_of(work, struct request_queue, requeue_work.work);
601 struct request *rq, *next;
604 spin_lock_irqsave(&q->requeue_lock, flags);
605 list_splice_init(&q->requeue_list, &rq_list);
606 spin_unlock_irqrestore(&q->requeue_lock, flags);
608 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
609 if (!(rq->rq_flags & RQF_SOFTBARRIER))
612 rq->rq_flags &= ~RQF_SOFTBARRIER;
613 list_del_init(&rq->queuelist);
614 blk_mq_sched_insert_request(rq, true, false, false, true);
617 while (!list_empty(&rq_list)) {
618 rq = list_entry(rq_list.next, struct request, queuelist);
619 list_del_init(&rq->queuelist);
620 blk_mq_sched_insert_request(rq, false, false, false, true);
623 blk_mq_run_hw_queues(q, false);
626 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
627 bool kick_requeue_list)
629 struct request_queue *q = rq->q;
633 * We abuse this flag that is otherwise used by the I/O scheduler to
634 * request head insertation from the workqueue.
636 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
638 spin_lock_irqsave(&q->requeue_lock, flags);
640 rq->rq_flags |= RQF_SOFTBARRIER;
641 list_add(&rq->queuelist, &q->requeue_list);
643 list_add_tail(&rq->queuelist, &q->requeue_list);
645 spin_unlock_irqrestore(&q->requeue_lock, flags);
647 if (kick_requeue_list)
648 blk_mq_kick_requeue_list(q);
650 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
652 void blk_mq_kick_requeue_list(struct request_queue *q)
654 kblockd_schedule_delayed_work(&q->requeue_work, 0);
656 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
658 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
661 kblockd_schedule_delayed_work(&q->requeue_work,
662 msecs_to_jiffies(msecs));
664 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
666 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
668 if (tag < tags->nr_tags) {
669 prefetch(tags->rqs[tag]);
670 return tags->rqs[tag];
675 EXPORT_SYMBOL(blk_mq_tag_to_rq);
677 struct blk_mq_timeout_data {
679 unsigned int next_set;
682 void blk_mq_rq_timed_out(struct request *req, bool reserved)
684 const struct blk_mq_ops *ops = req->q->mq_ops;
685 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
688 * We know that complete is set at this point. If STARTED isn't set
689 * anymore, then the request isn't active and the "timeout" should
690 * just be ignored. This can happen due to the bitflag ordering.
691 * Timeout first checks if STARTED is set, and if it is, assumes
692 * the request is active. But if we race with completion, then
693 * both flags will get cleared. So check here again, and ignore
694 * a timeout event with a request that isn't active.
696 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
700 ret = ops->timeout(req, reserved);
704 __blk_mq_complete_request(req);
706 case BLK_EH_RESET_TIMER:
708 blk_clear_rq_complete(req);
710 case BLK_EH_NOT_HANDLED:
713 printk(KERN_ERR "block: bad eh return: %d\n", ret);
718 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
719 struct request *rq, void *priv, bool reserved)
721 struct blk_mq_timeout_data *data = priv;
723 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
727 * The rq being checked may have been freed and reallocated
728 * out already here, we avoid this race by checking rq->deadline
729 * and REQ_ATOM_COMPLETE flag together:
731 * - if rq->deadline is observed as new value because of
732 * reusing, the rq won't be timed out because of timing.
733 * - if rq->deadline is observed as previous value,
734 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
735 * because we put a barrier between setting rq->deadline
736 * and clearing the flag in blk_mq_start_request(), so
737 * this rq won't be timed out too.
739 if (time_after_eq(jiffies, rq->deadline)) {
740 if (!blk_mark_rq_complete(rq))
741 blk_mq_rq_timed_out(rq, reserved);
742 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
743 data->next = rq->deadline;
748 static void blk_mq_timeout_work(struct work_struct *work)
750 struct request_queue *q =
751 container_of(work, struct request_queue, timeout_work);
752 struct blk_mq_timeout_data data = {
758 /* A deadlock might occur if a request is stuck requiring a
759 * timeout at the same time a queue freeze is waiting
760 * completion, since the timeout code would not be able to
761 * acquire the queue reference here.
763 * That's why we don't use blk_queue_enter here; instead, we use
764 * percpu_ref_tryget directly, because we need to be able to
765 * obtain a reference even in the short window between the queue
766 * starting to freeze, by dropping the first reference in
767 * blk_freeze_queue_start, and the moment the last request is
768 * consumed, marked by the instant q_usage_counter reaches
771 if (!percpu_ref_tryget(&q->q_usage_counter))
774 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
777 data.next = blk_rq_timeout(round_jiffies_up(data.next));
778 mod_timer(&q->timeout, data.next);
780 struct blk_mq_hw_ctx *hctx;
782 queue_for_each_hw_ctx(q, hctx, i) {
783 /* the hctx may be unmapped, so check it here */
784 if (blk_mq_hw_queue_mapped(hctx))
785 blk_mq_tag_idle(hctx);
791 struct flush_busy_ctx_data {
792 struct blk_mq_hw_ctx *hctx;
793 struct list_head *list;
796 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
798 struct flush_busy_ctx_data *flush_data = data;
799 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
800 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
802 sbitmap_clear_bit(sb, bitnr);
803 spin_lock(&ctx->lock);
804 list_splice_tail_init(&ctx->rq_list, flush_data->list);
805 spin_unlock(&ctx->lock);
810 * Process software queues that have been marked busy, splicing them
811 * to the for-dispatch
813 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
815 struct flush_busy_ctx_data data = {
820 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
822 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
824 static inline unsigned int queued_to_index(unsigned int queued)
829 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
832 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
835 struct blk_mq_alloc_data data = {
837 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
838 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
841 might_sleep_if(wait);
846 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
847 data.flags |= BLK_MQ_REQ_RESERVED;
849 rq->tag = blk_mq_get_tag(&data);
851 if (blk_mq_tag_busy(data.hctx)) {
852 rq->rq_flags |= RQF_MQ_INFLIGHT;
853 atomic_inc(&data.hctx->nr_active);
855 data.hctx->tags->rqs[rq->tag] = rq;
861 return rq->tag != -1;
864 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
867 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
870 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
871 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
872 atomic_dec(&hctx->nr_active);
876 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
879 if (rq->tag == -1 || rq->internal_tag == -1)
882 __blk_mq_put_driver_tag(hctx, rq);
885 static void blk_mq_put_driver_tag(struct request *rq)
887 struct blk_mq_hw_ctx *hctx;
889 if (rq->tag == -1 || rq->internal_tag == -1)
892 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
893 __blk_mq_put_driver_tag(hctx, rq);
897 * If we fail getting a driver tag because all the driver tags are already
898 * assigned and on the dispatch list, BUT the first entry does not have a
899 * tag, then we could deadlock. For that case, move entries with assigned
900 * driver tags to the front, leaving the set of tagged requests in the
901 * same order, and the untagged set in the same order.
903 static bool reorder_tags_to_front(struct list_head *list)
905 struct request *rq, *tmp, *first = NULL;
907 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
911 list_move(&rq->queuelist, list);
917 return first != NULL;
920 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
923 struct blk_mq_hw_ctx *hctx;
925 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
927 list_del(&wait->task_list);
928 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
929 blk_mq_run_hw_queue(hctx, true);
933 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
935 struct sbq_wait_state *ws;
938 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
939 * The thread which wins the race to grab this bit adds the hardware
940 * queue to the wait queue.
942 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
943 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
946 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
947 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
950 * As soon as this returns, it's no longer safe to fiddle with
951 * hctx->dispatch_wait, since a completion can wake up the wait queue
952 * and unlock the bit.
954 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
958 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
960 struct blk_mq_hw_ctx *hctx;
964 if (list_empty(list))
968 * Now process all the entries, sending them to the driver.
972 struct blk_mq_queue_data bd;
975 rq = list_first_entry(list, struct request, queuelist);
976 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
977 if (!queued && reorder_tags_to_front(list))
981 * The initial allocation attempt failed, so we need to
982 * rerun the hardware queue when a tag is freed.
984 if (!blk_mq_dispatch_wait_add(hctx))
988 * It's possible that a tag was freed in the window
989 * between the allocation failure and adding the
990 * hardware queue to the wait queue.
992 if (!blk_mq_get_driver_tag(rq, &hctx, false))
996 list_del_init(&rq->queuelist);
1001 * Flag last if we have no more requests, or if we have more
1002 * but can't assign a driver tag to it.
1004 if (list_empty(list))
1007 struct request *nxt;
1009 nxt = list_first_entry(list, struct request, queuelist);
1010 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1013 ret = q->mq_ops->queue_rq(hctx, &bd);
1014 if (ret == BLK_STS_RESOURCE) {
1015 blk_mq_put_driver_tag_hctx(hctx, rq);
1016 list_add(&rq->queuelist, list);
1017 __blk_mq_requeue_request(rq);
1021 if (unlikely(ret != BLK_STS_OK)) {
1023 blk_mq_end_request(rq, BLK_STS_IOERR);
1028 } while (!list_empty(list));
1030 hctx->dispatched[queued_to_index(queued)]++;
1033 * Any items that need requeuing? Stuff them into hctx->dispatch,
1034 * that is where we will continue on next queue run.
1036 if (!list_empty(list)) {
1038 * If an I/O scheduler has been configured and we got a driver
1039 * tag for the next request already, free it again.
1041 rq = list_first_entry(list, struct request, queuelist);
1042 blk_mq_put_driver_tag(rq);
1044 spin_lock(&hctx->lock);
1045 list_splice_init(list, &hctx->dispatch);
1046 spin_unlock(&hctx->lock);
1049 * If SCHED_RESTART was set by the caller of this function and
1050 * it is no longer set that means that it was cleared by another
1051 * thread and hence that a queue rerun is needed.
1053 * If TAG_WAITING is set that means that an I/O scheduler has
1054 * been configured and another thread is waiting for a driver
1055 * tag. To guarantee fairness, do not rerun this hardware queue
1056 * but let the other thread grab the driver tag.
1058 * If no I/O scheduler has been configured it is possible that
1059 * the hardware queue got stopped and restarted before requests
1060 * were pushed back onto the dispatch list. Rerun the queue to
1061 * avoid starvation. Notes:
1062 * - blk_mq_run_hw_queue() checks whether or not a queue has
1063 * been stopped before rerunning a queue.
1064 * - Some but not all block drivers stop a queue before
1065 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1068 if (!blk_mq_sched_needs_restart(hctx) &&
1069 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1070 blk_mq_run_hw_queue(hctx, true);
1073 return (queued + errors) != 0;
1076 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1080 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1081 cpu_online(hctx->next_cpu));
1083 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1085 blk_mq_sched_dispatch_requests(hctx);
1090 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1091 blk_mq_sched_dispatch_requests(hctx);
1092 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1097 * It'd be great if the workqueue API had a way to pass
1098 * in a mask and had some smarts for more clever placement.
1099 * For now we just round-robin here, switching for every
1100 * BLK_MQ_CPU_WORK_BATCH queued items.
1102 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1104 if (hctx->queue->nr_hw_queues == 1)
1105 return WORK_CPU_UNBOUND;
1107 if (--hctx->next_cpu_batch <= 0) {
1110 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1111 if (next_cpu >= nr_cpu_ids)
1112 next_cpu = cpumask_first(hctx->cpumask);
1114 hctx->next_cpu = next_cpu;
1115 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1118 return hctx->next_cpu;
1121 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1122 unsigned long msecs)
1124 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1125 !blk_mq_hw_queue_mapped(hctx)))
1128 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1129 int cpu = get_cpu();
1130 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1131 __blk_mq_run_hw_queue(hctx);
1139 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1141 msecs_to_jiffies(msecs));
1144 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1146 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1148 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1150 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1152 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1154 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1156 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1158 struct blk_mq_hw_ctx *hctx;
1161 queue_for_each_hw_ctx(q, hctx, i) {
1162 if (!blk_mq_hctx_has_pending(hctx) ||
1163 blk_mq_hctx_stopped(hctx))
1166 blk_mq_run_hw_queue(hctx, async);
1169 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1172 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1173 * @q: request queue.
1175 * The caller is responsible for serializing this function against
1176 * blk_mq_{start,stop}_hw_queue().
1178 bool blk_mq_queue_stopped(struct request_queue *q)
1180 struct blk_mq_hw_ctx *hctx;
1183 queue_for_each_hw_ctx(q, hctx, i)
1184 if (blk_mq_hctx_stopped(hctx))
1189 EXPORT_SYMBOL(blk_mq_queue_stopped);
1191 static void __blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx, bool sync)
1194 cancel_delayed_work_sync(&hctx->run_work);
1196 cancel_delayed_work(&hctx->run_work);
1198 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1201 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1203 __blk_mq_stop_hw_queue(hctx, false);
1205 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1207 static void __blk_mq_stop_hw_queues(struct request_queue *q, bool sync)
1209 struct blk_mq_hw_ctx *hctx;
1212 queue_for_each_hw_ctx(q, hctx, i)
1213 __blk_mq_stop_hw_queue(hctx, sync);
1216 void blk_mq_stop_hw_queues(struct request_queue *q)
1218 __blk_mq_stop_hw_queues(q, false);
1220 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1222 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1224 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1226 blk_mq_run_hw_queue(hctx, false);
1228 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1230 void blk_mq_start_hw_queues(struct request_queue *q)
1232 struct blk_mq_hw_ctx *hctx;
1235 queue_for_each_hw_ctx(q, hctx, i)
1236 blk_mq_start_hw_queue(hctx);
1238 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1240 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1242 if (!blk_mq_hctx_stopped(hctx))
1245 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1246 blk_mq_run_hw_queue(hctx, async);
1248 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1250 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1252 struct blk_mq_hw_ctx *hctx;
1255 queue_for_each_hw_ctx(q, hctx, i)
1256 blk_mq_start_stopped_hw_queue(hctx, async);
1258 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1260 static void blk_mq_run_work_fn(struct work_struct *work)
1262 struct blk_mq_hw_ctx *hctx;
1264 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1267 * If we are stopped, don't run the queue. The exception is if
1268 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1269 * the STOPPED bit and run it.
1271 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1272 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1275 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1276 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1279 __blk_mq_run_hw_queue(hctx);
1283 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1285 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1289 * Stop the hw queue, then modify currently delayed work.
1290 * This should prevent us from running the queue prematurely.
1291 * Mark the queue as auto-clearing STOPPED when it runs.
1293 blk_mq_stop_hw_queue(hctx);
1294 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1295 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1297 msecs_to_jiffies(msecs));
1299 EXPORT_SYMBOL(blk_mq_delay_queue);
1301 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1305 struct blk_mq_ctx *ctx = rq->mq_ctx;
1307 trace_block_rq_insert(hctx->queue, rq);
1310 list_add(&rq->queuelist, &ctx->rq_list);
1312 list_add_tail(&rq->queuelist, &ctx->rq_list);
1315 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1318 struct blk_mq_ctx *ctx = rq->mq_ctx;
1320 __blk_mq_insert_req_list(hctx, rq, at_head);
1321 blk_mq_hctx_mark_pending(hctx, ctx);
1324 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1325 struct list_head *list)
1329 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1332 spin_lock(&ctx->lock);
1333 while (!list_empty(list)) {
1336 rq = list_first_entry(list, struct request, queuelist);
1337 BUG_ON(rq->mq_ctx != ctx);
1338 list_del_init(&rq->queuelist);
1339 __blk_mq_insert_req_list(hctx, rq, false);
1341 blk_mq_hctx_mark_pending(hctx, ctx);
1342 spin_unlock(&ctx->lock);
1345 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1347 struct request *rqa = container_of(a, struct request, queuelist);
1348 struct request *rqb = container_of(b, struct request, queuelist);
1350 return !(rqa->mq_ctx < rqb->mq_ctx ||
1351 (rqa->mq_ctx == rqb->mq_ctx &&
1352 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1355 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1357 struct blk_mq_ctx *this_ctx;
1358 struct request_queue *this_q;
1361 LIST_HEAD(ctx_list);
1364 list_splice_init(&plug->mq_list, &list);
1366 list_sort(NULL, &list, plug_ctx_cmp);
1372 while (!list_empty(&list)) {
1373 rq = list_entry_rq(list.next);
1374 list_del_init(&rq->queuelist);
1376 if (rq->mq_ctx != this_ctx) {
1378 trace_block_unplug(this_q, depth, from_schedule);
1379 blk_mq_sched_insert_requests(this_q, this_ctx,
1384 this_ctx = rq->mq_ctx;
1390 list_add_tail(&rq->queuelist, &ctx_list);
1394 * If 'this_ctx' is set, we know we have entries to complete
1395 * on 'ctx_list'. Do those.
1398 trace_block_unplug(this_q, depth, from_schedule);
1399 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1404 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1406 blk_init_request_from_bio(rq, bio);
1408 blk_account_io_start(rq, true);
1411 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1413 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1414 !blk_queue_nomerges(hctx->queue);
1417 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1418 struct blk_mq_ctx *ctx,
1421 spin_lock(&ctx->lock);
1422 __blk_mq_insert_request(hctx, rq, false);
1423 spin_unlock(&ctx->lock);
1426 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1429 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1431 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1434 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1436 blk_qc_t *cookie, bool may_sleep)
1438 struct request_queue *q = rq->q;
1439 struct blk_mq_queue_data bd = {
1443 blk_qc_t new_cookie;
1445 bool run_queue = true;
1447 if (blk_mq_hctx_stopped(hctx)) {
1455 if (!blk_mq_get_driver_tag(rq, NULL, false))
1458 new_cookie = request_to_qc_t(hctx, rq);
1461 * For OK queue, we are done. For error, kill it. Any other
1462 * error (busy), just add it to our list as we previously
1465 ret = q->mq_ops->queue_rq(hctx, &bd);
1468 *cookie = new_cookie;
1470 case BLK_STS_RESOURCE:
1471 __blk_mq_requeue_request(rq);
1474 *cookie = BLK_QC_T_NONE;
1475 blk_mq_end_request(rq, ret);
1480 blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1483 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1484 struct request *rq, blk_qc_t *cookie)
1486 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1488 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1491 unsigned int srcu_idx;
1495 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1496 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1497 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1501 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1503 const int is_sync = op_is_sync(bio->bi_opf);
1504 const int is_flush_fua = op_is_flush(bio->bi_opf);
1505 struct blk_mq_alloc_data data = { .flags = 0 };
1507 unsigned int request_count = 0;
1508 struct blk_plug *plug;
1509 struct request *same_queue_rq = NULL;
1511 unsigned int wb_acct;
1513 blk_queue_bounce(q, &bio);
1515 blk_queue_split(q, &bio);
1517 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1519 return BLK_QC_T_NONE;
1522 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1523 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1524 return BLK_QC_T_NONE;
1526 if (blk_mq_sched_bio_merge(q, bio))
1527 return BLK_QC_T_NONE;
1529 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1531 trace_block_getrq(q, bio, bio->bi_opf);
1533 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1534 if (unlikely(!rq)) {
1535 __wbt_done(q->rq_wb, wb_acct);
1536 return BLK_QC_T_NONE;
1539 wbt_track(&rq->issue_stat, wb_acct);
1541 cookie = request_to_qc_t(data.hctx, rq);
1543 plug = current->plug;
1544 if (unlikely(is_flush_fua)) {
1545 blk_mq_put_ctx(data.ctx);
1546 blk_mq_bio_to_request(rq, bio);
1548 blk_mq_sched_insert_request(rq, false, true, true,
1551 blk_insert_flush(rq);
1552 blk_mq_run_hw_queue(data.hctx, true);
1554 } else if (plug && q->nr_hw_queues == 1) {
1555 struct request *last = NULL;
1557 blk_mq_put_ctx(data.ctx);
1558 blk_mq_bio_to_request(rq, bio);
1561 * @request_count may become stale because of schedule
1562 * out, so check the list again.
1564 if (list_empty(&plug->mq_list))
1566 else if (blk_queue_nomerges(q))
1567 request_count = blk_plug_queued_count(q);
1570 trace_block_plug(q);
1572 last = list_entry_rq(plug->mq_list.prev);
1574 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1575 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1576 blk_flush_plug_list(plug, false);
1577 trace_block_plug(q);
1580 list_add_tail(&rq->queuelist, &plug->mq_list);
1581 } else if (plug && !blk_queue_nomerges(q)) {
1582 blk_mq_bio_to_request(rq, bio);
1585 * We do limited plugging. If the bio can be merged, do that.
1586 * Otherwise the existing request in the plug list will be
1587 * issued. So the plug list will have one request at most
1588 * The plug list might get flushed before this. If that happens,
1589 * the plug list is empty, and same_queue_rq is invalid.
1591 if (list_empty(&plug->mq_list))
1592 same_queue_rq = NULL;
1594 list_del_init(&same_queue_rq->queuelist);
1595 list_add_tail(&rq->queuelist, &plug->mq_list);
1597 blk_mq_put_ctx(data.ctx);
1599 if (same_queue_rq) {
1600 data.hctx = blk_mq_map_queue(q,
1601 same_queue_rq->mq_ctx->cpu);
1602 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1605 } else if (q->nr_hw_queues > 1 && is_sync) {
1606 blk_mq_put_ctx(data.ctx);
1607 blk_mq_bio_to_request(rq, bio);
1608 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1609 } else if (q->elevator) {
1610 blk_mq_put_ctx(data.ctx);
1611 blk_mq_bio_to_request(rq, bio);
1612 blk_mq_sched_insert_request(rq, false, true, true, true);
1614 blk_mq_put_ctx(data.ctx);
1615 blk_mq_bio_to_request(rq, bio);
1616 blk_mq_queue_io(data.hctx, data.ctx, rq);
1617 blk_mq_run_hw_queue(data.hctx, true);
1623 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1624 unsigned int hctx_idx)
1628 if (tags->rqs && set->ops->exit_request) {
1631 for (i = 0; i < tags->nr_tags; i++) {
1632 struct request *rq = tags->static_rqs[i];
1636 set->ops->exit_request(set, rq, hctx_idx);
1637 tags->static_rqs[i] = NULL;
1641 while (!list_empty(&tags->page_list)) {
1642 page = list_first_entry(&tags->page_list, struct page, lru);
1643 list_del_init(&page->lru);
1645 * Remove kmemleak object previously allocated in
1646 * blk_mq_init_rq_map().
1648 kmemleak_free(page_address(page));
1649 __free_pages(page, page->private);
1653 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1657 kfree(tags->static_rqs);
1658 tags->static_rqs = NULL;
1660 blk_mq_free_tags(tags);
1663 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1664 unsigned int hctx_idx,
1665 unsigned int nr_tags,
1666 unsigned int reserved_tags)
1668 struct blk_mq_tags *tags;
1671 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1672 if (node == NUMA_NO_NODE)
1673 node = set->numa_node;
1675 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1676 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1680 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1681 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1684 blk_mq_free_tags(tags);
1688 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1689 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1691 if (!tags->static_rqs) {
1693 blk_mq_free_tags(tags);
1700 static size_t order_to_size(unsigned int order)
1702 return (size_t)PAGE_SIZE << order;
1705 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1706 unsigned int hctx_idx, unsigned int depth)
1708 unsigned int i, j, entries_per_page, max_order = 4;
1709 size_t rq_size, left;
1712 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1713 if (node == NUMA_NO_NODE)
1714 node = set->numa_node;
1716 INIT_LIST_HEAD(&tags->page_list);
1719 * rq_size is the size of the request plus driver payload, rounded
1720 * to the cacheline size
1722 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1724 left = rq_size * depth;
1726 for (i = 0; i < depth; ) {
1727 int this_order = max_order;
1732 while (this_order && left < order_to_size(this_order - 1))
1736 page = alloc_pages_node(node,
1737 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1743 if (order_to_size(this_order) < rq_size)
1750 page->private = this_order;
1751 list_add_tail(&page->lru, &tags->page_list);
1753 p = page_address(page);
1755 * Allow kmemleak to scan these pages as they contain pointers
1756 * to additional allocations like via ops->init_request().
1758 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1759 entries_per_page = order_to_size(this_order) / rq_size;
1760 to_do = min(entries_per_page, depth - i);
1761 left -= to_do * rq_size;
1762 for (j = 0; j < to_do; j++) {
1763 struct request *rq = p;
1765 tags->static_rqs[i] = rq;
1766 if (set->ops->init_request) {
1767 if (set->ops->init_request(set, rq, hctx_idx,
1769 tags->static_rqs[i] = NULL;
1781 blk_mq_free_rqs(set, tags, hctx_idx);
1786 * 'cpu' is going away. splice any existing rq_list entries from this
1787 * software queue to the hw queue dispatch list, and ensure that it
1790 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1792 struct blk_mq_hw_ctx *hctx;
1793 struct blk_mq_ctx *ctx;
1796 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1797 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1799 spin_lock(&ctx->lock);
1800 if (!list_empty(&ctx->rq_list)) {
1801 list_splice_init(&ctx->rq_list, &tmp);
1802 blk_mq_hctx_clear_pending(hctx, ctx);
1804 spin_unlock(&ctx->lock);
1806 if (list_empty(&tmp))
1809 spin_lock(&hctx->lock);
1810 list_splice_tail_init(&tmp, &hctx->dispatch);
1811 spin_unlock(&hctx->lock);
1813 blk_mq_run_hw_queue(hctx, true);
1817 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1819 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1823 /* hctx->ctxs will be freed in queue's release handler */
1824 static void blk_mq_exit_hctx(struct request_queue *q,
1825 struct blk_mq_tag_set *set,
1826 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1828 blk_mq_debugfs_unregister_hctx(hctx);
1830 blk_mq_tag_idle(hctx);
1832 if (set->ops->exit_request)
1833 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1835 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1837 if (set->ops->exit_hctx)
1838 set->ops->exit_hctx(hctx, hctx_idx);
1840 if (hctx->flags & BLK_MQ_F_BLOCKING)
1841 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1843 blk_mq_remove_cpuhp(hctx);
1844 blk_free_flush_queue(hctx->fq);
1845 sbitmap_free(&hctx->ctx_map);
1848 static void blk_mq_exit_hw_queues(struct request_queue *q,
1849 struct blk_mq_tag_set *set, int nr_queue)
1851 struct blk_mq_hw_ctx *hctx;
1854 queue_for_each_hw_ctx(q, hctx, i) {
1857 blk_mq_exit_hctx(q, set, hctx, i);
1861 static int blk_mq_init_hctx(struct request_queue *q,
1862 struct blk_mq_tag_set *set,
1863 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1867 node = hctx->numa_node;
1868 if (node == NUMA_NO_NODE)
1869 node = hctx->numa_node = set->numa_node;
1871 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1872 spin_lock_init(&hctx->lock);
1873 INIT_LIST_HEAD(&hctx->dispatch);
1875 hctx->queue_num = hctx_idx;
1876 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1878 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1880 hctx->tags = set->tags[hctx_idx];
1883 * Allocate space for all possible cpus to avoid allocation at
1886 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1889 goto unregister_cpu_notifier;
1891 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1897 if (set->ops->init_hctx &&
1898 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1901 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1904 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1906 goto sched_exit_hctx;
1908 if (set->ops->init_request &&
1909 set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1913 if (hctx->flags & BLK_MQ_F_BLOCKING)
1914 init_srcu_struct(&hctx->queue_rq_srcu);
1916 blk_mq_debugfs_register_hctx(q, hctx);
1923 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1925 if (set->ops->exit_hctx)
1926 set->ops->exit_hctx(hctx, hctx_idx);
1928 sbitmap_free(&hctx->ctx_map);
1931 unregister_cpu_notifier:
1932 blk_mq_remove_cpuhp(hctx);
1936 static void blk_mq_init_cpu_queues(struct request_queue *q,
1937 unsigned int nr_hw_queues)
1941 for_each_possible_cpu(i) {
1942 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1943 struct blk_mq_hw_ctx *hctx;
1946 spin_lock_init(&__ctx->lock);
1947 INIT_LIST_HEAD(&__ctx->rq_list);
1950 /* If the cpu isn't online, the cpu is mapped to first hctx */
1954 hctx = blk_mq_map_queue(q, i);
1957 * Set local node, IFF we have more than one hw queue. If
1958 * not, we remain on the home node of the device
1960 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1961 hctx->numa_node = local_memory_node(cpu_to_node(i));
1965 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1969 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1970 set->queue_depth, set->reserved_tags);
1971 if (!set->tags[hctx_idx])
1974 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
1979 blk_mq_free_rq_map(set->tags[hctx_idx]);
1980 set->tags[hctx_idx] = NULL;
1984 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
1985 unsigned int hctx_idx)
1987 if (set->tags[hctx_idx]) {
1988 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
1989 blk_mq_free_rq_map(set->tags[hctx_idx]);
1990 set->tags[hctx_idx] = NULL;
1994 static void blk_mq_map_swqueue(struct request_queue *q,
1995 const struct cpumask *online_mask)
1997 unsigned int i, hctx_idx;
1998 struct blk_mq_hw_ctx *hctx;
1999 struct blk_mq_ctx *ctx;
2000 struct blk_mq_tag_set *set = q->tag_set;
2003 * Avoid others reading imcomplete hctx->cpumask through sysfs
2005 mutex_lock(&q->sysfs_lock);
2007 queue_for_each_hw_ctx(q, hctx, i) {
2008 cpumask_clear(hctx->cpumask);
2013 * Map software to hardware queues
2015 for_each_possible_cpu(i) {
2016 /* If the cpu isn't online, the cpu is mapped to first hctx */
2017 if (!cpumask_test_cpu(i, online_mask))
2020 hctx_idx = q->mq_map[i];
2021 /* unmapped hw queue can be remapped after CPU topo changed */
2022 if (!set->tags[hctx_idx] &&
2023 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2025 * If tags initialization fail for some hctx,
2026 * that hctx won't be brought online. In this
2027 * case, remap the current ctx to hctx[0] which
2028 * is guaranteed to always have tags allocated
2033 ctx = per_cpu_ptr(q->queue_ctx, i);
2034 hctx = blk_mq_map_queue(q, i);
2036 cpumask_set_cpu(i, hctx->cpumask);
2037 ctx->index_hw = hctx->nr_ctx;
2038 hctx->ctxs[hctx->nr_ctx++] = ctx;
2041 mutex_unlock(&q->sysfs_lock);
2043 queue_for_each_hw_ctx(q, hctx, i) {
2045 * If no software queues are mapped to this hardware queue,
2046 * disable it and free the request entries.
2048 if (!hctx->nr_ctx) {
2049 /* Never unmap queue 0. We need it as a
2050 * fallback in case of a new remap fails
2053 if (i && set->tags[i])
2054 blk_mq_free_map_and_requests(set, i);
2060 hctx->tags = set->tags[i];
2061 WARN_ON(!hctx->tags);
2064 * Set the map size to the number of mapped software queues.
2065 * This is more accurate and more efficient than looping
2066 * over all possibly mapped software queues.
2068 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2071 * Initialize batch roundrobin counts
2073 hctx->next_cpu = cpumask_first(hctx->cpumask);
2074 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2078 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2080 struct blk_mq_hw_ctx *hctx;
2083 queue_for_each_hw_ctx(q, hctx, i) {
2085 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2087 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2091 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2093 struct request_queue *q;
2095 lockdep_assert_held(&set->tag_list_lock);
2097 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2098 blk_mq_freeze_queue(q);
2099 queue_set_hctx_shared(q, shared);
2100 blk_mq_unfreeze_queue(q);
2104 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2106 struct blk_mq_tag_set *set = q->tag_set;
2108 mutex_lock(&set->tag_list_lock);
2109 list_del_rcu(&q->tag_set_list);
2110 INIT_LIST_HEAD(&q->tag_set_list);
2111 if (list_is_singular(&set->tag_list)) {
2112 /* just transitioned to unshared */
2113 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2114 /* update existing queue */
2115 blk_mq_update_tag_set_depth(set, false);
2117 mutex_unlock(&set->tag_list_lock);
2122 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2123 struct request_queue *q)
2127 mutex_lock(&set->tag_list_lock);
2129 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2130 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2131 set->flags |= BLK_MQ_F_TAG_SHARED;
2132 /* update existing queue */
2133 blk_mq_update_tag_set_depth(set, true);
2135 if (set->flags & BLK_MQ_F_TAG_SHARED)
2136 queue_set_hctx_shared(q, true);
2137 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2139 mutex_unlock(&set->tag_list_lock);
2143 * It is the actual release handler for mq, but we do it from
2144 * request queue's release handler for avoiding use-after-free
2145 * and headache because q->mq_kobj shouldn't have been introduced,
2146 * but we can't group ctx/kctx kobj without it.
2148 void blk_mq_release(struct request_queue *q)
2150 struct blk_mq_hw_ctx *hctx;
2153 /* hctx kobj stays in hctx */
2154 queue_for_each_hw_ctx(q, hctx, i) {
2157 kobject_put(&hctx->kobj);
2162 kfree(q->queue_hw_ctx);
2165 * release .mq_kobj and sw queue's kobject now because
2166 * both share lifetime with request queue.
2168 blk_mq_sysfs_deinit(q);
2170 free_percpu(q->queue_ctx);
2173 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2175 struct request_queue *uninit_q, *q;
2177 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2179 return ERR_PTR(-ENOMEM);
2181 q = blk_mq_init_allocated_queue(set, uninit_q);
2183 blk_cleanup_queue(uninit_q);
2187 EXPORT_SYMBOL(blk_mq_init_queue);
2189 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2190 struct request_queue *q)
2193 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2195 blk_mq_sysfs_unregister(q);
2196 for (i = 0; i < set->nr_hw_queues; i++) {
2202 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2203 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2208 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2215 atomic_set(&hctxs[i]->nr_active, 0);
2216 hctxs[i]->numa_node = node;
2217 hctxs[i]->queue_num = i;
2219 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2220 free_cpumask_var(hctxs[i]->cpumask);
2225 blk_mq_hctx_kobj_init(hctxs[i]);
2227 for (j = i; j < q->nr_hw_queues; j++) {
2228 struct blk_mq_hw_ctx *hctx = hctxs[j];
2232 blk_mq_free_map_and_requests(set, j);
2233 blk_mq_exit_hctx(q, set, hctx, j);
2234 kobject_put(&hctx->kobj);
2239 q->nr_hw_queues = i;
2240 blk_mq_sysfs_register(q);
2243 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2244 struct request_queue *q)
2246 /* mark the queue as mq asap */
2247 q->mq_ops = set->ops;
2249 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2250 blk_mq_poll_stats_bkt,
2251 BLK_MQ_POLL_STATS_BKTS, q);
2255 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2259 /* init q->mq_kobj and sw queues' kobjects */
2260 blk_mq_sysfs_init(q);
2262 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2263 GFP_KERNEL, set->numa_node);
2264 if (!q->queue_hw_ctx)
2267 q->mq_map = set->mq_map;
2269 blk_mq_realloc_hw_ctxs(set, q);
2270 if (!q->nr_hw_queues)
2273 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2274 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2276 q->nr_queues = nr_cpu_ids;
2278 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2280 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2281 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2283 q->sg_reserved_size = INT_MAX;
2285 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2286 INIT_LIST_HEAD(&q->requeue_list);
2287 spin_lock_init(&q->requeue_lock);
2289 blk_queue_make_request(q, blk_mq_make_request);
2292 * Do this after blk_queue_make_request() overrides it...
2294 q->nr_requests = set->queue_depth;
2297 * Default to classic polling
2301 if (set->ops->complete)
2302 blk_queue_softirq_done(q, set->ops->complete);
2304 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2307 mutex_lock(&all_q_mutex);
2309 list_add_tail(&q->all_q_node, &all_q_list);
2310 blk_mq_add_queue_tag_set(set, q);
2311 blk_mq_map_swqueue(q, cpu_online_mask);
2313 mutex_unlock(&all_q_mutex);
2316 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2319 ret = blk_mq_sched_init(q);
2321 return ERR_PTR(ret);
2327 kfree(q->queue_hw_ctx);
2329 free_percpu(q->queue_ctx);
2332 return ERR_PTR(-ENOMEM);
2334 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2336 void blk_mq_free_queue(struct request_queue *q)
2338 struct blk_mq_tag_set *set = q->tag_set;
2340 mutex_lock(&all_q_mutex);
2341 list_del_init(&q->all_q_node);
2342 mutex_unlock(&all_q_mutex);
2344 blk_mq_del_queue_tag_set(q);
2346 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2349 /* Basically redo blk_mq_init_queue with queue frozen */
2350 static void blk_mq_queue_reinit(struct request_queue *q,
2351 const struct cpumask *online_mask)
2353 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2355 blk_mq_debugfs_unregister_hctxs(q);
2356 blk_mq_sysfs_unregister(q);
2359 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2360 * we should change hctx numa_node according to new topology (this
2361 * involves free and re-allocate memory, worthy doing?)
2364 blk_mq_map_swqueue(q, online_mask);
2366 blk_mq_sysfs_register(q);
2367 blk_mq_debugfs_register_hctxs(q);
2371 * New online cpumask which is going to be set in this hotplug event.
2372 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2373 * one-by-one and dynamically allocating this could result in a failure.
2375 static struct cpumask cpuhp_online_new;
2377 static void blk_mq_queue_reinit_work(void)
2379 struct request_queue *q;
2381 mutex_lock(&all_q_mutex);
2383 * We need to freeze and reinit all existing queues. Freezing
2384 * involves synchronous wait for an RCU grace period and doing it
2385 * one by one may take a long time. Start freezing all queues in
2386 * one swoop and then wait for the completions so that freezing can
2387 * take place in parallel.
2389 list_for_each_entry(q, &all_q_list, all_q_node)
2390 blk_freeze_queue_start(q);
2391 list_for_each_entry(q, &all_q_list, all_q_node)
2392 blk_mq_freeze_queue_wait(q);
2394 list_for_each_entry(q, &all_q_list, all_q_node)
2395 blk_mq_queue_reinit(q, &cpuhp_online_new);
2397 list_for_each_entry(q, &all_q_list, all_q_node)
2398 blk_mq_unfreeze_queue(q);
2400 mutex_unlock(&all_q_mutex);
2403 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2405 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2406 blk_mq_queue_reinit_work();
2411 * Before hotadded cpu starts handling requests, new mappings must be
2412 * established. Otherwise, these requests in hw queue might never be
2415 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2416 * for CPU0, and ctx1 for CPU1).
2418 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2419 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2421 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2422 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2423 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2426 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2428 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2429 cpumask_set_cpu(cpu, &cpuhp_online_new);
2430 blk_mq_queue_reinit_work();
2434 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2438 for (i = 0; i < set->nr_hw_queues; i++)
2439 if (!__blk_mq_alloc_rq_map(set, i))
2446 blk_mq_free_rq_map(set->tags[i]);
2452 * Allocate the request maps associated with this tag_set. Note that this
2453 * may reduce the depth asked for, if memory is tight. set->queue_depth
2454 * will be updated to reflect the allocated depth.
2456 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2461 depth = set->queue_depth;
2463 err = __blk_mq_alloc_rq_maps(set);
2467 set->queue_depth >>= 1;
2468 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2472 } while (set->queue_depth);
2474 if (!set->queue_depth || err) {
2475 pr_err("blk-mq: failed to allocate request map\n");
2479 if (depth != set->queue_depth)
2480 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2481 depth, set->queue_depth);
2486 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2488 if (set->ops->map_queues)
2489 return set->ops->map_queues(set);
2491 return blk_mq_map_queues(set);
2495 * Alloc a tag set to be associated with one or more request queues.
2496 * May fail with EINVAL for various error conditions. May adjust the
2497 * requested depth down, if if it too large. In that case, the set
2498 * value will be stored in set->queue_depth.
2500 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2504 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2506 if (!set->nr_hw_queues)
2508 if (!set->queue_depth)
2510 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2513 if (!set->ops->queue_rq)
2516 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2517 pr_info("blk-mq: reduced tag depth to %u\n",
2519 set->queue_depth = BLK_MQ_MAX_DEPTH;
2523 * If a crashdump is active, then we are potentially in a very
2524 * memory constrained environment. Limit us to 1 queue and
2525 * 64 tags to prevent using too much memory.
2527 if (is_kdump_kernel()) {
2528 set->nr_hw_queues = 1;
2529 set->queue_depth = min(64U, set->queue_depth);
2532 * There is no use for more h/w queues than cpus.
2534 if (set->nr_hw_queues > nr_cpu_ids)
2535 set->nr_hw_queues = nr_cpu_ids;
2537 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2538 GFP_KERNEL, set->numa_node);
2543 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2544 GFP_KERNEL, set->numa_node);
2548 ret = blk_mq_update_queue_map(set);
2550 goto out_free_mq_map;
2552 ret = blk_mq_alloc_rq_maps(set);
2554 goto out_free_mq_map;
2556 mutex_init(&set->tag_list_lock);
2557 INIT_LIST_HEAD(&set->tag_list);
2569 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2571 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2575 for (i = 0; i < nr_cpu_ids; i++)
2576 blk_mq_free_map_and_requests(set, i);
2584 EXPORT_SYMBOL(blk_mq_free_tag_set);
2586 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2588 struct blk_mq_tag_set *set = q->tag_set;
2589 struct blk_mq_hw_ctx *hctx;
2595 blk_mq_freeze_queue(q);
2598 queue_for_each_hw_ctx(q, hctx, i) {
2602 * If we're using an MQ scheduler, just update the scheduler
2603 * queue depth. This is similar to what the old code would do.
2605 if (!hctx->sched_tags) {
2606 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2607 min(nr, set->queue_depth),
2610 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2618 q->nr_requests = nr;
2620 blk_mq_unfreeze_queue(q);
2625 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2628 struct request_queue *q;
2630 lockdep_assert_held(&set->tag_list_lock);
2632 if (nr_hw_queues > nr_cpu_ids)
2633 nr_hw_queues = nr_cpu_ids;
2634 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2637 list_for_each_entry(q, &set->tag_list, tag_set_list)
2638 blk_mq_freeze_queue(q);
2640 set->nr_hw_queues = nr_hw_queues;
2641 blk_mq_update_queue_map(set);
2642 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2643 blk_mq_realloc_hw_ctxs(set, q);
2644 blk_mq_queue_reinit(q, cpu_online_mask);
2647 list_for_each_entry(q, &set->tag_list, tag_set_list)
2648 blk_mq_unfreeze_queue(q);
2651 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2653 mutex_lock(&set->tag_list_lock);
2654 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2655 mutex_unlock(&set->tag_list_lock);
2657 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2659 /* Enable polling stats and return whether they were already enabled. */
2660 static bool blk_poll_stats_enable(struct request_queue *q)
2662 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2663 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2665 blk_stat_add_callback(q, q->poll_cb);
2669 static void blk_mq_poll_stats_start(struct request_queue *q)
2672 * We don't arm the callback if polling stats are not enabled or the
2673 * callback is already active.
2675 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2676 blk_stat_is_active(q->poll_cb))
2679 blk_stat_activate_msecs(q->poll_cb, 100);
2682 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2684 struct request_queue *q = cb->data;
2687 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2688 if (cb->stat[bucket].nr_samples)
2689 q->poll_stat[bucket] = cb->stat[bucket];
2693 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2694 struct blk_mq_hw_ctx *hctx,
2697 unsigned long ret = 0;
2701 * If stats collection isn't on, don't sleep but turn it on for
2704 if (!blk_poll_stats_enable(q))
2708 * As an optimistic guess, use half of the mean service time
2709 * for this type of request. We can (and should) make this smarter.
2710 * For instance, if the completion latencies are tight, we can
2711 * get closer than just half the mean. This is especially
2712 * important on devices where the completion latencies are longer
2713 * than ~10 usec. We do use the stats for the relevant IO size
2714 * if available which does lead to better estimates.
2716 bucket = blk_mq_poll_stats_bkt(rq);
2720 if (q->poll_stat[bucket].nr_samples)
2721 ret = (q->poll_stat[bucket].mean + 1) / 2;
2726 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2727 struct blk_mq_hw_ctx *hctx,
2730 struct hrtimer_sleeper hs;
2731 enum hrtimer_mode mode;
2735 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2741 * -1: don't ever hybrid sleep
2742 * 0: use half of prev avg
2743 * >0: use this specific value
2745 if (q->poll_nsec == -1)
2747 else if (q->poll_nsec > 0)
2748 nsecs = q->poll_nsec;
2750 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2755 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2758 * This will be replaced with the stats tracking code, using
2759 * 'avg_completion_time / 2' as the pre-sleep target.
2763 mode = HRTIMER_MODE_REL;
2764 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2765 hrtimer_set_expires(&hs.timer, kt);
2767 hrtimer_init_sleeper(&hs, current);
2769 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2771 set_current_state(TASK_UNINTERRUPTIBLE);
2772 hrtimer_start_expires(&hs.timer, mode);
2775 hrtimer_cancel(&hs.timer);
2776 mode = HRTIMER_MODE_ABS;
2777 } while (hs.task && !signal_pending(current));
2779 __set_current_state(TASK_RUNNING);
2780 destroy_hrtimer_on_stack(&hs.timer);
2784 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2786 struct request_queue *q = hctx->queue;
2790 * If we sleep, have the caller restart the poll loop to reset
2791 * the state. Like for the other success return cases, the
2792 * caller is responsible for checking if the IO completed. If
2793 * the IO isn't complete, we'll get called again and will go
2794 * straight to the busy poll loop.
2796 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2799 hctx->poll_considered++;
2801 state = current->state;
2802 while (!need_resched()) {
2805 hctx->poll_invoked++;
2807 ret = q->mq_ops->poll(hctx, rq->tag);
2809 hctx->poll_success++;
2810 set_current_state(TASK_RUNNING);
2814 if (signal_pending_state(state, current))
2815 set_current_state(TASK_RUNNING);
2817 if (current->state == TASK_RUNNING)
2827 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2829 struct blk_mq_hw_ctx *hctx;
2830 struct blk_plug *plug;
2833 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2834 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2837 plug = current->plug;
2839 blk_flush_plug_list(plug, false);
2841 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2842 if (!blk_qc_t_is_internal(cookie))
2843 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2845 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2847 * With scheduling, if the request has completed, we'll
2848 * get a NULL return here, as we clear the sched tag when
2849 * that happens. The request still remains valid, like always,
2850 * so we should be safe with just the NULL check.
2856 return __blk_mq_poll(hctx, rq);
2858 EXPORT_SYMBOL_GPL(blk_mq_poll);
2860 void blk_mq_disable_hotplug(void)
2862 mutex_lock(&all_q_mutex);
2865 void blk_mq_enable_hotplug(void)
2867 mutex_unlock(&all_q_mutex);
2870 static int __init blk_mq_init(void)
2872 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2873 blk_mq_hctx_notify_dead);
2875 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2876 blk_mq_queue_reinit_prepare,
2877 blk_mq_queue_reinit_dead);
2880 subsys_initcall(blk_mq_init);