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);
46 static int blk_mq_poll_stats_bkt(const struct request *rq)
48 int ddir, bytes, bucket;
50 ddir = rq_data_dir(rq);
51 bytes = blk_rq_bytes(rq);
53 bucket = ddir + 2*(ilog2(bytes) - 9);
57 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
58 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
64 * Check if any of the ctx's have pending work in this hardware queue
66 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
68 return sbitmap_any_bit_set(&hctx->ctx_map) ||
69 !list_empty_careful(&hctx->dispatch) ||
70 blk_mq_sched_has_work(hctx);
74 * Mark this ctx as having pending work in this hardware queue
76 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
77 struct blk_mq_ctx *ctx)
79 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
80 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
83 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
84 struct blk_mq_ctx *ctx)
86 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
89 void blk_freeze_queue_start(struct request_queue *q)
93 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
94 if (freeze_depth == 1) {
95 percpu_ref_kill(&q->q_usage_counter);
96 blk_mq_run_hw_queues(q, false);
99 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
101 void blk_mq_freeze_queue_wait(struct request_queue *q)
103 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
105 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
107 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
108 unsigned long timeout)
110 return wait_event_timeout(q->mq_freeze_wq,
111 percpu_ref_is_zero(&q->q_usage_counter),
114 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
117 * Guarantee no request is in use, so we can change any data structure of
118 * the queue afterward.
120 void blk_freeze_queue(struct request_queue *q)
123 * In the !blk_mq case we are only calling this to kill the
124 * q_usage_counter, otherwise this increases the freeze depth
125 * and waits for it to return to zero. For this reason there is
126 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
127 * exported to drivers as the only user for unfreeze is blk_mq.
129 blk_freeze_queue_start(q);
130 blk_mq_freeze_queue_wait(q);
133 void blk_mq_freeze_queue(struct request_queue *q)
136 * ...just an alias to keep freeze and unfreeze actions balanced
137 * in the blk_mq_* namespace
141 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
143 void blk_mq_unfreeze_queue(struct request_queue *q)
147 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
148 WARN_ON_ONCE(freeze_depth < 0);
150 percpu_ref_reinit(&q->q_usage_counter);
151 wake_up_all(&q->mq_freeze_wq);
154 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
157 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
160 * Note: this function does not prevent that the struct request end_io()
161 * callback function is invoked. Once this function is returned, we make
162 * sure no dispatch can happen until the queue is unquiesced via
163 * blk_mq_unquiesce_queue().
165 void blk_mq_quiesce_queue(struct request_queue *q)
167 struct blk_mq_hw_ctx *hctx;
171 blk_mq_quiesce_queue_nowait(q);
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 spin_lock_irq(q->queue_lock);
194 queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
195 spin_unlock_irq(q->queue_lock);
197 /* dispatch requests which are inserted during quiescing */
198 blk_mq_run_hw_queues(q, true);
200 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
202 void blk_mq_wake_waiters(struct request_queue *q)
204 struct blk_mq_hw_ctx *hctx;
207 queue_for_each_hw_ctx(q, hctx, i)
208 if (blk_mq_hw_queue_mapped(hctx))
209 blk_mq_tag_wakeup_all(hctx->tags, true);
212 * If we are called because the queue has now been marked as
213 * dying, we need to ensure that processes currently waiting on
214 * the queue are notified as well.
216 wake_up_all(&q->mq_freeze_wq);
219 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
221 return blk_mq_has_free_tags(hctx->tags);
223 EXPORT_SYMBOL(blk_mq_can_queue);
225 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
226 unsigned int tag, unsigned int op)
228 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
229 struct request *rq = tags->static_rqs[tag];
231 if (data->flags & BLK_MQ_REQ_INTERNAL) {
233 rq->internal_tag = tag;
235 if (blk_mq_tag_busy(data->hctx)) {
236 rq->rq_flags = RQF_MQ_INFLIGHT;
237 atomic_inc(&data->hctx->nr_active);
240 rq->internal_tag = -1;
241 data->hctx->tags->rqs[rq->tag] = rq;
244 INIT_LIST_HEAD(&rq->queuelist);
245 /* csd/requeue_work/fifo_time is initialized before use */
247 rq->mq_ctx = data->ctx;
249 if (blk_queue_io_stat(data->q))
250 rq->rq_flags |= RQF_IO_STAT;
251 /* do not touch atomic flags, it needs atomic ops against the timer */
253 INIT_HLIST_NODE(&rq->hash);
254 RB_CLEAR_NODE(&rq->rb_node);
257 rq->start_time = jiffies;
258 #ifdef CONFIG_BLK_CGROUP
260 set_start_time_ns(rq);
261 rq->io_start_time_ns = 0;
263 rq->nr_phys_segments = 0;
264 #if defined(CONFIG_BLK_DEV_INTEGRITY)
265 rq->nr_integrity_segments = 0;
268 /* tag was already set */
271 INIT_LIST_HEAD(&rq->timeout_list);
275 rq->end_io_data = NULL;
278 data->ctx->rq_dispatched[op_is_sync(op)]++;
282 static struct request *blk_mq_get_request(struct request_queue *q,
283 struct bio *bio, unsigned int op,
284 struct blk_mq_alloc_data *data)
286 struct elevator_queue *e = q->elevator;
290 blk_queue_enter_live(q);
292 if (likely(!data->ctx))
293 data->ctx = blk_mq_get_ctx(q);
294 if (likely(!data->hctx))
295 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
297 data->flags |= BLK_MQ_REQ_NOWAIT;
300 data->flags |= BLK_MQ_REQ_INTERNAL;
303 * Flush requests are special and go directly to the
306 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
307 e->type->ops.mq.limit_depth(op, data);
310 tag = blk_mq_get_tag(data);
311 if (tag == BLK_MQ_TAG_FAIL) {
316 rq = blk_mq_rq_ctx_init(data, tag, op);
317 if (!op_is_flush(op)) {
319 if (e && e->type->ops.mq.prepare_request) {
320 if (e->type->icq_cache && rq_ioc(bio))
321 blk_mq_sched_assign_ioc(rq, bio);
323 e->type->ops.mq.prepare_request(rq, bio);
324 rq->rq_flags |= RQF_ELVPRIV;
327 data->hctx->queued++;
331 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
334 struct blk_mq_alloc_data alloc_data = { .flags = flags };
338 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
342 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
344 blk_mq_put_ctx(alloc_data.ctx);
348 return ERR_PTR(-EWOULDBLOCK);
351 rq->__sector = (sector_t) -1;
352 rq->bio = rq->biotail = NULL;
355 EXPORT_SYMBOL(blk_mq_alloc_request);
357 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
358 unsigned int op, unsigned int flags, unsigned int hctx_idx)
360 struct blk_mq_alloc_data alloc_data = { .flags = flags };
366 * If the tag allocator sleeps we could get an allocation for a
367 * different hardware context. No need to complicate the low level
368 * allocator for this for the rare use case of a command tied to
371 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
372 return ERR_PTR(-EINVAL);
374 if (hctx_idx >= q->nr_hw_queues)
375 return ERR_PTR(-EIO);
377 ret = blk_queue_enter(q, true);
382 * Check if the hardware context is actually mapped to anything.
383 * If not tell the caller that it should skip this queue.
385 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
386 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
388 return ERR_PTR(-EXDEV);
390 cpu = cpumask_first(alloc_data.hctx->cpumask);
391 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
393 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
398 return ERR_PTR(-EWOULDBLOCK);
402 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
404 void blk_mq_free_request(struct request *rq)
406 struct request_queue *q = rq->q;
407 struct elevator_queue *e = q->elevator;
408 struct blk_mq_ctx *ctx = rq->mq_ctx;
409 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
410 const int sched_tag = rq->internal_tag;
412 if (rq->rq_flags & RQF_ELVPRIV) {
413 if (e && e->type->ops.mq.finish_request)
414 e->type->ops.mq.finish_request(rq);
416 put_io_context(rq->elv.icq->ioc);
421 ctx->rq_completed[rq_is_sync(rq)]++;
422 if (rq->rq_flags & RQF_MQ_INFLIGHT)
423 atomic_dec(&hctx->nr_active);
425 wbt_done(q->rq_wb, &rq->issue_stat);
428 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
429 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
431 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
433 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
434 blk_mq_sched_restart(hctx);
437 EXPORT_SYMBOL_GPL(blk_mq_free_request);
439 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
441 blk_account_io_done(rq);
444 wbt_done(rq->q->rq_wb, &rq->issue_stat);
445 rq->end_io(rq, error);
447 if (unlikely(blk_bidi_rq(rq)))
448 blk_mq_free_request(rq->next_rq);
449 blk_mq_free_request(rq);
452 EXPORT_SYMBOL(__blk_mq_end_request);
454 void blk_mq_end_request(struct request *rq, blk_status_t error)
456 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
458 __blk_mq_end_request(rq, error);
460 EXPORT_SYMBOL(blk_mq_end_request);
462 static void __blk_mq_complete_request_remote(void *data)
464 struct request *rq = data;
466 rq->q->softirq_done_fn(rq);
469 static void __blk_mq_complete_request(struct request *rq)
471 struct blk_mq_ctx *ctx = rq->mq_ctx;
475 if (rq->internal_tag != -1)
476 blk_mq_sched_completed_request(rq);
477 if (rq->rq_flags & RQF_STATS) {
478 blk_mq_poll_stats_start(rq->q);
482 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
483 rq->q->softirq_done_fn(rq);
488 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
489 shared = cpus_share_cache(cpu, ctx->cpu);
491 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
492 rq->csd.func = __blk_mq_complete_request_remote;
495 smp_call_function_single_async(ctx->cpu, &rq->csd);
497 rq->q->softirq_done_fn(rq);
503 * blk_mq_complete_request - end I/O on a request
504 * @rq: the request being processed
507 * Ends all I/O on a request. It does not handle partial completions.
508 * The actual completion happens out-of-order, through a IPI handler.
510 void blk_mq_complete_request(struct request *rq)
512 struct request_queue *q = rq->q;
514 if (unlikely(blk_should_fake_timeout(q)))
516 if (!blk_mark_rq_complete(rq))
517 __blk_mq_complete_request(rq);
519 EXPORT_SYMBOL(blk_mq_complete_request);
521 int blk_mq_request_started(struct request *rq)
523 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
525 EXPORT_SYMBOL_GPL(blk_mq_request_started);
527 void blk_mq_start_request(struct request *rq)
529 struct request_queue *q = rq->q;
531 blk_mq_sched_started_request(rq);
533 trace_block_rq_issue(q, rq);
535 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
536 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
537 rq->rq_flags |= RQF_STATS;
538 wbt_issue(q->rq_wb, &rq->issue_stat);
544 * Ensure that ->deadline is visible before set the started
545 * flag and clear the completed flag.
547 smp_mb__before_atomic();
550 * Mark us as started and clear complete. Complete might have been
551 * set if requeue raced with timeout, which then marked it as
552 * complete. So be sure to clear complete again when we start
553 * the request, otherwise we'll ignore the completion event.
555 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
556 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
557 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
558 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
560 if (q->dma_drain_size && blk_rq_bytes(rq)) {
562 * Make sure space for the drain appears. We know we can do
563 * this because max_hw_segments has been adjusted to be one
564 * fewer than the device can handle.
566 rq->nr_phys_segments++;
569 EXPORT_SYMBOL(blk_mq_start_request);
572 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
573 * flag isn't set yet, so there may be race with timeout handler,
574 * but given rq->deadline is just set in .queue_rq() under
575 * this situation, the race won't be possible in reality because
576 * rq->timeout should be set as big enough to cover the window
577 * between blk_mq_start_request() called from .queue_rq() and
578 * clearing REQ_ATOM_STARTED here.
580 static void __blk_mq_requeue_request(struct request *rq)
582 struct request_queue *q = rq->q;
584 trace_block_rq_requeue(q, rq);
585 wbt_requeue(q->rq_wb, &rq->issue_stat);
586 blk_mq_sched_requeue_request(rq);
588 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
589 if (q->dma_drain_size && blk_rq_bytes(rq))
590 rq->nr_phys_segments--;
594 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
596 __blk_mq_requeue_request(rq);
598 BUG_ON(blk_queued_rq(rq));
599 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
601 EXPORT_SYMBOL(blk_mq_requeue_request);
603 static void blk_mq_requeue_work(struct work_struct *work)
605 struct request_queue *q =
606 container_of(work, struct request_queue, requeue_work.work);
608 struct request *rq, *next;
611 spin_lock_irqsave(&q->requeue_lock, flags);
612 list_splice_init(&q->requeue_list, &rq_list);
613 spin_unlock_irqrestore(&q->requeue_lock, flags);
615 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
616 if (!(rq->rq_flags & RQF_SOFTBARRIER))
619 rq->rq_flags &= ~RQF_SOFTBARRIER;
620 list_del_init(&rq->queuelist);
621 blk_mq_sched_insert_request(rq, true, false, false, true);
624 while (!list_empty(&rq_list)) {
625 rq = list_entry(rq_list.next, struct request, queuelist);
626 list_del_init(&rq->queuelist);
627 blk_mq_sched_insert_request(rq, false, false, false, true);
630 blk_mq_run_hw_queues(q, false);
633 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
634 bool kick_requeue_list)
636 struct request_queue *q = rq->q;
640 * We abuse this flag that is otherwise used by the I/O scheduler to
641 * request head insertation from the workqueue.
643 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
645 spin_lock_irqsave(&q->requeue_lock, flags);
647 rq->rq_flags |= RQF_SOFTBARRIER;
648 list_add(&rq->queuelist, &q->requeue_list);
650 list_add_tail(&rq->queuelist, &q->requeue_list);
652 spin_unlock_irqrestore(&q->requeue_lock, flags);
654 if (kick_requeue_list)
655 blk_mq_kick_requeue_list(q);
657 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
659 void blk_mq_kick_requeue_list(struct request_queue *q)
661 kblockd_schedule_delayed_work(&q->requeue_work, 0);
663 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
665 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
668 kblockd_schedule_delayed_work(&q->requeue_work,
669 msecs_to_jiffies(msecs));
671 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
673 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
675 if (tag < tags->nr_tags) {
676 prefetch(tags->rqs[tag]);
677 return tags->rqs[tag];
682 EXPORT_SYMBOL(blk_mq_tag_to_rq);
684 struct blk_mq_timeout_data {
686 unsigned int next_set;
689 void blk_mq_rq_timed_out(struct request *req, bool reserved)
691 const struct blk_mq_ops *ops = req->q->mq_ops;
692 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
695 * We know that complete is set at this point. If STARTED isn't set
696 * anymore, then the request isn't active and the "timeout" should
697 * just be ignored. This can happen due to the bitflag ordering.
698 * Timeout first checks if STARTED is set, and if it is, assumes
699 * the request is active. But if we race with completion, then
700 * both flags will get cleared. So check here again, and ignore
701 * a timeout event with a request that isn't active.
703 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
707 ret = ops->timeout(req, reserved);
711 __blk_mq_complete_request(req);
713 case BLK_EH_RESET_TIMER:
715 blk_clear_rq_complete(req);
717 case BLK_EH_NOT_HANDLED:
720 printk(KERN_ERR "block: bad eh return: %d\n", ret);
725 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
726 struct request *rq, void *priv, bool reserved)
728 struct blk_mq_timeout_data *data = priv;
730 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
734 * The rq being checked may have been freed and reallocated
735 * out already here, we avoid this race by checking rq->deadline
736 * and REQ_ATOM_COMPLETE flag together:
738 * - if rq->deadline is observed as new value because of
739 * reusing, the rq won't be timed out because of timing.
740 * - if rq->deadline is observed as previous value,
741 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
742 * because we put a barrier between setting rq->deadline
743 * and clearing the flag in blk_mq_start_request(), so
744 * this rq won't be timed out too.
746 if (time_after_eq(jiffies, rq->deadline)) {
747 if (!blk_mark_rq_complete(rq))
748 blk_mq_rq_timed_out(rq, reserved);
749 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
750 data->next = rq->deadline;
755 static void blk_mq_timeout_work(struct work_struct *work)
757 struct request_queue *q =
758 container_of(work, struct request_queue, timeout_work);
759 struct blk_mq_timeout_data data = {
765 /* A deadlock might occur if a request is stuck requiring a
766 * timeout at the same time a queue freeze is waiting
767 * completion, since the timeout code would not be able to
768 * acquire the queue reference here.
770 * That's why we don't use blk_queue_enter here; instead, we use
771 * percpu_ref_tryget directly, because we need to be able to
772 * obtain a reference even in the short window between the queue
773 * starting to freeze, by dropping the first reference in
774 * blk_freeze_queue_start, and the moment the last request is
775 * consumed, marked by the instant q_usage_counter reaches
778 if (!percpu_ref_tryget(&q->q_usage_counter))
781 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
784 data.next = blk_rq_timeout(round_jiffies_up(data.next));
785 mod_timer(&q->timeout, data.next);
787 struct blk_mq_hw_ctx *hctx;
789 queue_for_each_hw_ctx(q, hctx, i) {
790 /* the hctx may be unmapped, so check it here */
791 if (blk_mq_hw_queue_mapped(hctx))
792 blk_mq_tag_idle(hctx);
798 struct flush_busy_ctx_data {
799 struct blk_mq_hw_ctx *hctx;
800 struct list_head *list;
803 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
805 struct flush_busy_ctx_data *flush_data = data;
806 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
807 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
809 sbitmap_clear_bit(sb, bitnr);
810 spin_lock(&ctx->lock);
811 list_splice_tail_init(&ctx->rq_list, flush_data->list);
812 spin_unlock(&ctx->lock);
817 * Process software queues that have been marked busy, splicing them
818 * to the for-dispatch
820 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
822 struct flush_busy_ctx_data data = {
827 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
829 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
831 static inline unsigned int queued_to_index(unsigned int queued)
836 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
839 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
842 struct blk_mq_alloc_data data = {
844 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
845 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
848 might_sleep_if(wait);
853 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
854 data.flags |= BLK_MQ_REQ_RESERVED;
856 rq->tag = blk_mq_get_tag(&data);
858 if (blk_mq_tag_busy(data.hctx)) {
859 rq->rq_flags |= RQF_MQ_INFLIGHT;
860 atomic_inc(&data.hctx->nr_active);
862 data.hctx->tags->rqs[rq->tag] = rq;
868 return rq->tag != -1;
871 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
874 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
877 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
878 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
879 atomic_dec(&hctx->nr_active);
883 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
886 if (rq->tag == -1 || rq->internal_tag == -1)
889 __blk_mq_put_driver_tag(hctx, rq);
892 static void blk_mq_put_driver_tag(struct request *rq)
894 struct blk_mq_hw_ctx *hctx;
896 if (rq->tag == -1 || rq->internal_tag == -1)
899 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
900 __blk_mq_put_driver_tag(hctx, rq);
904 * If we fail getting a driver tag because all the driver tags are already
905 * assigned and on the dispatch list, BUT the first entry does not have a
906 * tag, then we could deadlock. For that case, move entries with assigned
907 * driver tags to the front, leaving the set of tagged requests in the
908 * same order, and the untagged set in the same order.
910 static bool reorder_tags_to_front(struct list_head *list)
912 struct request *rq, *tmp, *first = NULL;
914 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
918 list_move(&rq->queuelist, list);
924 return first != NULL;
927 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
930 struct blk_mq_hw_ctx *hctx;
932 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
934 list_del(&wait->task_list);
935 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
936 blk_mq_run_hw_queue(hctx, true);
940 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
942 struct sbq_wait_state *ws;
945 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
946 * The thread which wins the race to grab this bit adds the hardware
947 * queue to the wait queue.
949 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
950 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
953 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
954 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
957 * As soon as this returns, it's no longer safe to fiddle with
958 * hctx->dispatch_wait, since a completion can wake up the wait queue
959 * and unlock the bit.
961 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
965 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
967 struct blk_mq_hw_ctx *hctx;
971 if (list_empty(list))
975 * Now process all the entries, sending them to the driver.
979 struct blk_mq_queue_data bd;
982 rq = list_first_entry(list, struct request, queuelist);
983 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
984 if (!queued && reorder_tags_to_front(list))
988 * The initial allocation attempt failed, so we need to
989 * rerun the hardware queue when a tag is freed.
991 if (!blk_mq_dispatch_wait_add(hctx))
995 * It's possible that a tag was freed in the window
996 * between the allocation failure and adding the
997 * hardware queue to the wait queue.
999 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1003 list_del_init(&rq->queuelist);
1008 * Flag last if we have no more requests, or if we have more
1009 * but can't assign a driver tag to it.
1011 if (list_empty(list))
1014 struct request *nxt;
1016 nxt = list_first_entry(list, struct request, queuelist);
1017 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1020 ret = q->mq_ops->queue_rq(hctx, &bd);
1021 if (ret == BLK_STS_RESOURCE) {
1022 blk_mq_put_driver_tag_hctx(hctx, rq);
1023 list_add(&rq->queuelist, list);
1024 __blk_mq_requeue_request(rq);
1028 if (unlikely(ret != BLK_STS_OK)) {
1030 blk_mq_end_request(rq, BLK_STS_IOERR);
1035 } while (!list_empty(list));
1037 hctx->dispatched[queued_to_index(queued)]++;
1040 * Any items that need requeuing? Stuff them into hctx->dispatch,
1041 * that is where we will continue on next queue run.
1043 if (!list_empty(list)) {
1045 * If an I/O scheduler has been configured and we got a driver
1046 * tag for the next request already, free it again.
1048 rq = list_first_entry(list, struct request, queuelist);
1049 blk_mq_put_driver_tag(rq);
1051 spin_lock(&hctx->lock);
1052 list_splice_init(list, &hctx->dispatch);
1053 spin_unlock(&hctx->lock);
1056 * If SCHED_RESTART was set by the caller of this function and
1057 * it is no longer set that means that it was cleared by another
1058 * thread and hence that a queue rerun is needed.
1060 * If TAG_WAITING is set that means that an I/O scheduler has
1061 * been configured and another thread is waiting for a driver
1062 * tag. To guarantee fairness, do not rerun this hardware queue
1063 * but let the other thread grab the driver tag.
1065 * If no I/O scheduler has been configured it is possible that
1066 * the hardware queue got stopped and restarted before requests
1067 * were pushed back onto the dispatch list. Rerun the queue to
1068 * avoid starvation. Notes:
1069 * - blk_mq_run_hw_queue() checks whether or not a queue has
1070 * been stopped before rerunning a queue.
1071 * - Some but not all block drivers stop a queue before
1072 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1075 if (!blk_mq_sched_needs_restart(hctx) &&
1076 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1077 blk_mq_run_hw_queue(hctx, true);
1080 return (queued + errors) != 0;
1083 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1087 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1088 cpu_online(hctx->next_cpu));
1090 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1092 blk_mq_sched_dispatch_requests(hctx);
1097 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1098 blk_mq_sched_dispatch_requests(hctx);
1099 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1104 * It'd be great if the workqueue API had a way to pass
1105 * in a mask and had some smarts for more clever placement.
1106 * For now we just round-robin here, switching for every
1107 * BLK_MQ_CPU_WORK_BATCH queued items.
1109 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1111 if (hctx->queue->nr_hw_queues == 1)
1112 return WORK_CPU_UNBOUND;
1114 if (--hctx->next_cpu_batch <= 0) {
1117 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1118 if (next_cpu >= nr_cpu_ids)
1119 next_cpu = cpumask_first(hctx->cpumask);
1121 hctx->next_cpu = next_cpu;
1122 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1125 return hctx->next_cpu;
1128 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1129 unsigned long msecs)
1131 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1132 !blk_mq_hw_queue_mapped(hctx)))
1135 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1136 int cpu = get_cpu();
1137 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1138 __blk_mq_run_hw_queue(hctx);
1146 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1148 msecs_to_jiffies(msecs));
1151 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1153 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1155 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1157 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1159 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1161 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1163 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1165 struct blk_mq_hw_ctx *hctx;
1168 queue_for_each_hw_ctx(q, hctx, i) {
1169 if (!blk_mq_hctx_has_pending(hctx) ||
1170 blk_mq_hctx_stopped(hctx))
1173 blk_mq_run_hw_queue(hctx, async);
1176 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1179 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1180 * @q: request queue.
1182 * The caller is responsible for serializing this function against
1183 * blk_mq_{start,stop}_hw_queue().
1185 bool blk_mq_queue_stopped(struct request_queue *q)
1187 struct blk_mq_hw_ctx *hctx;
1190 queue_for_each_hw_ctx(q, hctx, i)
1191 if (blk_mq_hctx_stopped(hctx))
1196 EXPORT_SYMBOL(blk_mq_queue_stopped);
1199 * This function is often used for pausing .queue_rq() by driver when
1200 * there isn't enough resource or some conditions aren't satisfied, and
1201 * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1203 * We do not guarantee that dispatch can be drained or blocked
1204 * after blk_mq_stop_hw_queue() returns. Please use
1205 * blk_mq_quiesce_queue() for that requirement.
1207 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1209 cancel_delayed_work(&hctx->run_work);
1211 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1213 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1216 * This function is often used for pausing .queue_rq() by driver when
1217 * there isn't enough resource or some conditions aren't satisfied, and
1218 * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1220 * We do not guarantee that dispatch can be drained or blocked
1221 * after blk_mq_stop_hw_queues() returns. Please use
1222 * blk_mq_quiesce_queue() for that requirement.
1224 void blk_mq_stop_hw_queues(struct request_queue *q)
1226 struct blk_mq_hw_ctx *hctx;
1229 queue_for_each_hw_ctx(q, hctx, i)
1230 blk_mq_stop_hw_queue(hctx);
1232 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1234 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1236 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1238 blk_mq_run_hw_queue(hctx, false);
1240 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1242 void blk_mq_start_hw_queues(struct request_queue *q)
1244 struct blk_mq_hw_ctx *hctx;
1247 queue_for_each_hw_ctx(q, hctx, i)
1248 blk_mq_start_hw_queue(hctx);
1250 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1252 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1254 if (!blk_mq_hctx_stopped(hctx))
1257 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1258 blk_mq_run_hw_queue(hctx, async);
1260 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1262 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1264 struct blk_mq_hw_ctx *hctx;
1267 queue_for_each_hw_ctx(q, hctx, i)
1268 blk_mq_start_stopped_hw_queue(hctx, async);
1270 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1272 static void blk_mq_run_work_fn(struct work_struct *work)
1274 struct blk_mq_hw_ctx *hctx;
1276 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1279 * If we are stopped, don't run the queue. The exception is if
1280 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1281 * the STOPPED bit and run it.
1283 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1284 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1287 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1288 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1291 __blk_mq_run_hw_queue(hctx);
1295 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1297 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1301 * Stop the hw queue, then modify currently delayed work.
1302 * This should prevent us from running the queue prematurely.
1303 * Mark the queue as auto-clearing STOPPED when it runs.
1305 blk_mq_stop_hw_queue(hctx);
1306 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1307 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1309 msecs_to_jiffies(msecs));
1311 EXPORT_SYMBOL(blk_mq_delay_queue);
1313 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1317 struct blk_mq_ctx *ctx = rq->mq_ctx;
1319 trace_block_rq_insert(hctx->queue, rq);
1322 list_add(&rq->queuelist, &ctx->rq_list);
1324 list_add_tail(&rq->queuelist, &ctx->rq_list);
1327 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1330 struct blk_mq_ctx *ctx = rq->mq_ctx;
1332 __blk_mq_insert_req_list(hctx, rq, at_head);
1333 blk_mq_hctx_mark_pending(hctx, ctx);
1336 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1337 struct list_head *list)
1341 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1344 spin_lock(&ctx->lock);
1345 while (!list_empty(list)) {
1348 rq = list_first_entry(list, struct request, queuelist);
1349 BUG_ON(rq->mq_ctx != ctx);
1350 list_del_init(&rq->queuelist);
1351 __blk_mq_insert_req_list(hctx, rq, false);
1353 blk_mq_hctx_mark_pending(hctx, ctx);
1354 spin_unlock(&ctx->lock);
1357 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1359 struct request *rqa = container_of(a, struct request, queuelist);
1360 struct request *rqb = container_of(b, struct request, queuelist);
1362 return !(rqa->mq_ctx < rqb->mq_ctx ||
1363 (rqa->mq_ctx == rqb->mq_ctx &&
1364 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1367 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1369 struct blk_mq_ctx *this_ctx;
1370 struct request_queue *this_q;
1373 LIST_HEAD(ctx_list);
1376 list_splice_init(&plug->mq_list, &list);
1378 list_sort(NULL, &list, plug_ctx_cmp);
1384 while (!list_empty(&list)) {
1385 rq = list_entry_rq(list.next);
1386 list_del_init(&rq->queuelist);
1388 if (rq->mq_ctx != this_ctx) {
1390 trace_block_unplug(this_q, depth, from_schedule);
1391 blk_mq_sched_insert_requests(this_q, this_ctx,
1396 this_ctx = rq->mq_ctx;
1402 list_add_tail(&rq->queuelist, &ctx_list);
1406 * If 'this_ctx' is set, we know we have entries to complete
1407 * on 'ctx_list'. Do those.
1410 trace_block_unplug(this_q, depth, from_schedule);
1411 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1416 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1418 blk_init_request_from_bio(rq, bio);
1420 blk_account_io_start(rq, true);
1423 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1425 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1426 !blk_queue_nomerges(hctx->queue);
1429 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1430 struct blk_mq_ctx *ctx,
1433 spin_lock(&ctx->lock);
1434 __blk_mq_insert_request(hctx, rq, false);
1435 spin_unlock(&ctx->lock);
1438 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1441 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1443 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1446 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1448 blk_qc_t *cookie, bool may_sleep)
1450 struct request_queue *q = rq->q;
1451 struct blk_mq_queue_data bd = {
1455 blk_qc_t new_cookie;
1457 bool run_queue = true;
1459 /* RCU or SRCU read lock is needed before checking quiesced flag */
1460 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1468 if (!blk_mq_get_driver_tag(rq, NULL, false))
1471 new_cookie = request_to_qc_t(hctx, rq);
1474 * For OK queue, we are done. For error, kill it. Any other
1475 * error (busy), just add it to our list as we previously
1478 ret = q->mq_ops->queue_rq(hctx, &bd);
1481 *cookie = new_cookie;
1483 case BLK_STS_RESOURCE:
1484 __blk_mq_requeue_request(rq);
1487 *cookie = BLK_QC_T_NONE;
1488 blk_mq_end_request(rq, ret);
1493 blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1496 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1497 struct request *rq, blk_qc_t *cookie)
1499 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1501 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1504 unsigned int srcu_idx;
1508 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1509 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1510 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1514 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1516 const int is_sync = op_is_sync(bio->bi_opf);
1517 const int is_flush_fua = op_is_flush(bio->bi_opf);
1518 struct blk_mq_alloc_data data = { .flags = 0 };
1520 unsigned int request_count = 0;
1521 struct blk_plug *plug;
1522 struct request *same_queue_rq = NULL;
1524 unsigned int wb_acct;
1526 blk_queue_bounce(q, &bio);
1528 blk_queue_split(q, &bio);
1530 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1532 return BLK_QC_T_NONE;
1535 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1536 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1537 return BLK_QC_T_NONE;
1539 if (blk_mq_sched_bio_merge(q, bio))
1540 return BLK_QC_T_NONE;
1542 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1544 trace_block_getrq(q, bio, bio->bi_opf);
1546 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1547 if (unlikely(!rq)) {
1548 __wbt_done(q->rq_wb, wb_acct);
1549 if (bio->bi_opf & REQ_NOWAIT)
1550 bio_wouldblock_error(bio);
1551 return BLK_QC_T_NONE;
1554 wbt_track(&rq->issue_stat, wb_acct);
1556 cookie = request_to_qc_t(data.hctx, rq);
1558 plug = current->plug;
1559 if (unlikely(is_flush_fua)) {
1560 blk_mq_put_ctx(data.ctx);
1561 blk_mq_bio_to_request(rq, bio);
1563 blk_mq_sched_insert_request(rq, false, true, true,
1566 blk_insert_flush(rq);
1567 blk_mq_run_hw_queue(data.hctx, true);
1569 } else if (plug && q->nr_hw_queues == 1) {
1570 struct request *last = NULL;
1572 blk_mq_put_ctx(data.ctx);
1573 blk_mq_bio_to_request(rq, bio);
1576 * @request_count may become stale because of schedule
1577 * out, so check the list again.
1579 if (list_empty(&plug->mq_list))
1581 else if (blk_queue_nomerges(q))
1582 request_count = blk_plug_queued_count(q);
1585 trace_block_plug(q);
1587 last = list_entry_rq(plug->mq_list.prev);
1589 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1590 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1591 blk_flush_plug_list(plug, false);
1592 trace_block_plug(q);
1595 list_add_tail(&rq->queuelist, &plug->mq_list);
1596 } else if (plug && !blk_queue_nomerges(q)) {
1597 blk_mq_bio_to_request(rq, bio);
1600 * We do limited plugging. If the bio can be merged, do that.
1601 * Otherwise the existing request in the plug list will be
1602 * issued. So the plug list will have one request at most
1603 * The plug list might get flushed before this. If that happens,
1604 * the plug list is empty, and same_queue_rq is invalid.
1606 if (list_empty(&plug->mq_list))
1607 same_queue_rq = NULL;
1609 list_del_init(&same_queue_rq->queuelist);
1610 list_add_tail(&rq->queuelist, &plug->mq_list);
1612 blk_mq_put_ctx(data.ctx);
1614 if (same_queue_rq) {
1615 data.hctx = blk_mq_map_queue(q,
1616 same_queue_rq->mq_ctx->cpu);
1617 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1620 } else if (q->nr_hw_queues > 1 && is_sync) {
1621 blk_mq_put_ctx(data.ctx);
1622 blk_mq_bio_to_request(rq, bio);
1623 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1624 } else if (q->elevator) {
1625 blk_mq_put_ctx(data.ctx);
1626 blk_mq_bio_to_request(rq, bio);
1627 blk_mq_sched_insert_request(rq, false, true, true, true);
1629 blk_mq_put_ctx(data.ctx);
1630 blk_mq_bio_to_request(rq, bio);
1631 blk_mq_queue_io(data.hctx, data.ctx, rq);
1632 blk_mq_run_hw_queue(data.hctx, true);
1638 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1639 unsigned int hctx_idx)
1643 if (tags->rqs && set->ops->exit_request) {
1646 for (i = 0; i < tags->nr_tags; i++) {
1647 struct request *rq = tags->static_rqs[i];
1651 set->ops->exit_request(set, rq, hctx_idx);
1652 tags->static_rqs[i] = NULL;
1656 while (!list_empty(&tags->page_list)) {
1657 page = list_first_entry(&tags->page_list, struct page, lru);
1658 list_del_init(&page->lru);
1660 * Remove kmemleak object previously allocated in
1661 * blk_mq_init_rq_map().
1663 kmemleak_free(page_address(page));
1664 __free_pages(page, page->private);
1668 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1672 kfree(tags->static_rqs);
1673 tags->static_rqs = NULL;
1675 blk_mq_free_tags(tags);
1678 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1679 unsigned int hctx_idx,
1680 unsigned int nr_tags,
1681 unsigned int reserved_tags)
1683 struct blk_mq_tags *tags;
1686 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1687 if (node == NUMA_NO_NODE)
1688 node = set->numa_node;
1690 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1691 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1695 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1696 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1699 blk_mq_free_tags(tags);
1703 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1704 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1706 if (!tags->static_rqs) {
1708 blk_mq_free_tags(tags);
1715 static size_t order_to_size(unsigned int order)
1717 return (size_t)PAGE_SIZE << order;
1720 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1721 unsigned int hctx_idx, unsigned int depth)
1723 unsigned int i, j, entries_per_page, max_order = 4;
1724 size_t rq_size, left;
1727 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1728 if (node == NUMA_NO_NODE)
1729 node = set->numa_node;
1731 INIT_LIST_HEAD(&tags->page_list);
1734 * rq_size is the size of the request plus driver payload, rounded
1735 * to the cacheline size
1737 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1739 left = rq_size * depth;
1741 for (i = 0; i < depth; ) {
1742 int this_order = max_order;
1747 while (this_order && left < order_to_size(this_order - 1))
1751 page = alloc_pages_node(node,
1752 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1758 if (order_to_size(this_order) < rq_size)
1765 page->private = this_order;
1766 list_add_tail(&page->lru, &tags->page_list);
1768 p = page_address(page);
1770 * Allow kmemleak to scan these pages as they contain pointers
1771 * to additional allocations like via ops->init_request().
1773 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1774 entries_per_page = order_to_size(this_order) / rq_size;
1775 to_do = min(entries_per_page, depth - i);
1776 left -= to_do * rq_size;
1777 for (j = 0; j < to_do; j++) {
1778 struct request *rq = p;
1780 tags->static_rqs[i] = rq;
1781 if (set->ops->init_request) {
1782 if (set->ops->init_request(set, rq, hctx_idx,
1784 tags->static_rqs[i] = NULL;
1796 blk_mq_free_rqs(set, tags, hctx_idx);
1801 * 'cpu' is going away. splice any existing rq_list entries from this
1802 * software queue to the hw queue dispatch list, and ensure that it
1805 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1807 struct blk_mq_hw_ctx *hctx;
1808 struct blk_mq_ctx *ctx;
1811 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1812 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1814 spin_lock(&ctx->lock);
1815 if (!list_empty(&ctx->rq_list)) {
1816 list_splice_init(&ctx->rq_list, &tmp);
1817 blk_mq_hctx_clear_pending(hctx, ctx);
1819 spin_unlock(&ctx->lock);
1821 if (list_empty(&tmp))
1824 spin_lock(&hctx->lock);
1825 list_splice_tail_init(&tmp, &hctx->dispatch);
1826 spin_unlock(&hctx->lock);
1828 blk_mq_run_hw_queue(hctx, true);
1832 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1834 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1838 /* hctx->ctxs will be freed in queue's release handler */
1839 static void blk_mq_exit_hctx(struct request_queue *q,
1840 struct blk_mq_tag_set *set,
1841 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1843 blk_mq_debugfs_unregister_hctx(hctx);
1845 blk_mq_tag_idle(hctx);
1847 if (set->ops->exit_request)
1848 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1850 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1852 if (set->ops->exit_hctx)
1853 set->ops->exit_hctx(hctx, hctx_idx);
1855 if (hctx->flags & BLK_MQ_F_BLOCKING)
1856 cleanup_srcu_struct(hctx->queue_rq_srcu);
1858 blk_mq_remove_cpuhp(hctx);
1859 blk_free_flush_queue(hctx->fq);
1860 sbitmap_free(&hctx->ctx_map);
1863 static void blk_mq_exit_hw_queues(struct request_queue *q,
1864 struct blk_mq_tag_set *set, int nr_queue)
1866 struct blk_mq_hw_ctx *hctx;
1869 queue_for_each_hw_ctx(q, hctx, i) {
1872 blk_mq_exit_hctx(q, set, hctx, i);
1876 static int blk_mq_init_hctx(struct request_queue *q,
1877 struct blk_mq_tag_set *set,
1878 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1882 node = hctx->numa_node;
1883 if (node == NUMA_NO_NODE)
1884 node = hctx->numa_node = set->numa_node;
1886 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1887 spin_lock_init(&hctx->lock);
1888 INIT_LIST_HEAD(&hctx->dispatch);
1890 hctx->queue_num = hctx_idx;
1891 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1893 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1895 hctx->tags = set->tags[hctx_idx];
1898 * Allocate space for all possible cpus to avoid allocation at
1901 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1904 goto unregister_cpu_notifier;
1906 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1912 if (set->ops->init_hctx &&
1913 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1916 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1919 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1921 goto sched_exit_hctx;
1923 if (set->ops->init_request &&
1924 set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1928 if (hctx->flags & BLK_MQ_F_BLOCKING)
1929 init_srcu_struct(hctx->queue_rq_srcu);
1931 blk_mq_debugfs_register_hctx(q, hctx);
1938 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1940 if (set->ops->exit_hctx)
1941 set->ops->exit_hctx(hctx, hctx_idx);
1943 sbitmap_free(&hctx->ctx_map);
1946 unregister_cpu_notifier:
1947 blk_mq_remove_cpuhp(hctx);
1951 static void blk_mq_init_cpu_queues(struct request_queue *q,
1952 unsigned int nr_hw_queues)
1956 for_each_possible_cpu(i) {
1957 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1958 struct blk_mq_hw_ctx *hctx;
1961 spin_lock_init(&__ctx->lock);
1962 INIT_LIST_HEAD(&__ctx->rq_list);
1965 /* If the cpu isn't online, the cpu is mapped to first hctx */
1969 hctx = blk_mq_map_queue(q, i);
1972 * Set local node, IFF we have more than one hw queue. If
1973 * not, we remain on the home node of the device
1975 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1976 hctx->numa_node = local_memory_node(cpu_to_node(i));
1980 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1984 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1985 set->queue_depth, set->reserved_tags);
1986 if (!set->tags[hctx_idx])
1989 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
1994 blk_mq_free_rq_map(set->tags[hctx_idx]);
1995 set->tags[hctx_idx] = NULL;
1999 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2000 unsigned int hctx_idx)
2002 if (set->tags[hctx_idx]) {
2003 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2004 blk_mq_free_rq_map(set->tags[hctx_idx]);
2005 set->tags[hctx_idx] = NULL;
2009 static void blk_mq_map_swqueue(struct request_queue *q,
2010 const struct cpumask *online_mask)
2012 unsigned int i, hctx_idx;
2013 struct blk_mq_hw_ctx *hctx;
2014 struct blk_mq_ctx *ctx;
2015 struct blk_mq_tag_set *set = q->tag_set;
2018 * Avoid others reading imcomplete hctx->cpumask through sysfs
2020 mutex_lock(&q->sysfs_lock);
2022 queue_for_each_hw_ctx(q, hctx, i) {
2023 cpumask_clear(hctx->cpumask);
2028 * Map software to hardware queues
2030 for_each_possible_cpu(i) {
2031 /* If the cpu isn't online, the cpu is mapped to first hctx */
2032 if (!cpumask_test_cpu(i, online_mask))
2035 hctx_idx = q->mq_map[i];
2036 /* unmapped hw queue can be remapped after CPU topo changed */
2037 if (!set->tags[hctx_idx] &&
2038 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2040 * If tags initialization fail for some hctx,
2041 * that hctx won't be brought online. In this
2042 * case, remap the current ctx to hctx[0] which
2043 * is guaranteed to always have tags allocated
2048 ctx = per_cpu_ptr(q->queue_ctx, i);
2049 hctx = blk_mq_map_queue(q, i);
2051 cpumask_set_cpu(i, hctx->cpumask);
2052 ctx->index_hw = hctx->nr_ctx;
2053 hctx->ctxs[hctx->nr_ctx++] = ctx;
2056 mutex_unlock(&q->sysfs_lock);
2058 queue_for_each_hw_ctx(q, hctx, i) {
2060 * If no software queues are mapped to this hardware queue,
2061 * disable it and free the request entries.
2063 if (!hctx->nr_ctx) {
2064 /* Never unmap queue 0. We need it as a
2065 * fallback in case of a new remap fails
2068 if (i && set->tags[i])
2069 blk_mq_free_map_and_requests(set, i);
2075 hctx->tags = set->tags[i];
2076 WARN_ON(!hctx->tags);
2079 * Set the map size to the number of mapped software queues.
2080 * This is more accurate and more efficient than looping
2081 * over all possibly mapped software queues.
2083 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2086 * Initialize batch roundrobin counts
2088 hctx->next_cpu = cpumask_first(hctx->cpumask);
2089 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2093 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2095 struct blk_mq_hw_ctx *hctx;
2098 queue_for_each_hw_ctx(q, hctx, i) {
2100 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2102 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2106 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2108 struct request_queue *q;
2110 lockdep_assert_held(&set->tag_list_lock);
2112 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2113 blk_mq_freeze_queue(q);
2114 queue_set_hctx_shared(q, shared);
2115 blk_mq_unfreeze_queue(q);
2119 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2121 struct blk_mq_tag_set *set = q->tag_set;
2123 mutex_lock(&set->tag_list_lock);
2124 list_del_rcu(&q->tag_set_list);
2125 INIT_LIST_HEAD(&q->tag_set_list);
2126 if (list_is_singular(&set->tag_list)) {
2127 /* just transitioned to unshared */
2128 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2129 /* update existing queue */
2130 blk_mq_update_tag_set_depth(set, false);
2132 mutex_unlock(&set->tag_list_lock);
2137 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2138 struct request_queue *q)
2142 mutex_lock(&set->tag_list_lock);
2144 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2145 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2146 set->flags |= BLK_MQ_F_TAG_SHARED;
2147 /* update existing queue */
2148 blk_mq_update_tag_set_depth(set, true);
2150 if (set->flags & BLK_MQ_F_TAG_SHARED)
2151 queue_set_hctx_shared(q, true);
2152 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2154 mutex_unlock(&set->tag_list_lock);
2158 * It is the actual release handler for mq, but we do it from
2159 * request queue's release handler for avoiding use-after-free
2160 * and headache because q->mq_kobj shouldn't have been introduced,
2161 * but we can't group ctx/kctx kobj without it.
2163 void blk_mq_release(struct request_queue *q)
2165 struct blk_mq_hw_ctx *hctx;
2168 /* hctx kobj stays in hctx */
2169 queue_for_each_hw_ctx(q, hctx, i) {
2172 kobject_put(&hctx->kobj);
2177 kfree(q->queue_hw_ctx);
2180 * release .mq_kobj and sw queue's kobject now because
2181 * both share lifetime with request queue.
2183 blk_mq_sysfs_deinit(q);
2185 free_percpu(q->queue_ctx);
2188 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2190 struct request_queue *uninit_q, *q;
2192 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2194 return ERR_PTR(-ENOMEM);
2196 q = blk_mq_init_allocated_queue(set, uninit_q);
2198 blk_cleanup_queue(uninit_q);
2202 EXPORT_SYMBOL(blk_mq_init_queue);
2204 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2206 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2208 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2209 __alignof__(struct blk_mq_hw_ctx)) !=
2210 sizeof(struct blk_mq_hw_ctx));
2212 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2213 hw_ctx_size += sizeof(struct srcu_struct);
2218 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2219 struct request_queue *q)
2222 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2224 blk_mq_sysfs_unregister(q);
2225 for (i = 0; i < set->nr_hw_queues; i++) {
2231 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2232 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2237 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2244 atomic_set(&hctxs[i]->nr_active, 0);
2245 hctxs[i]->numa_node = node;
2246 hctxs[i]->queue_num = i;
2248 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2249 free_cpumask_var(hctxs[i]->cpumask);
2254 blk_mq_hctx_kobj_init(hctxs[i]);
2256 for (j = i; j < q->nr_hw_queues; j++) {
2257 struct blk_mq_hw_ctx *hctx = hctxs[j];
2261 blk_mq_free_map_and_requests(set, j);
2262 blk_mq_exit_hctx(q, set, hctx, j);
2263 kobject_put(&hctx->kobj);
2268 q->nr_hw_queues = i;
2269 blk_mq_sysfs_register(q);
2272 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2273 struct request_queue *q)
2275 /* mark the queue as mq asap */
2276 q->mq_ops = set->ops;
2278 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2279 blk_mq_poll_stats_bkt,
2280 BLK_MQ_POLL_STATS_BKTS, q);
2284 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2288 /* init q->mq_kobj and sw queues' kobjects */
2289 blk_mq_sysfs_init(q);
2291 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2292 GFP_KERNEL, set->numa_node);
2293 if (!q->queue_hw_ctx)
2296 q->mq_map = set->mq_map;
2298 blk_mq_realloc_hw_ctxs(set, q);
2299 if (!q->nr_hw_queues)
2302 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2303 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2305 q->nr_queues = nr_cpu_ids;
2307 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2309 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2310 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2312 q->sg_reserved_size = INT_MAX;
2314 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2315 INIT_LIST_HEAD(&q->requeue_list);
2316 spin_lock_init(&q->requeue_lock);
2318 blk_queue_make_request(q, blk_mq_make_request);
2321 * Do this after blk_queue_make_request() overrides it...
2323 q->nr_requests = set->queue_depth;
2326 * Default to classic polling
2330 if (set->ops->complete)
2331 blk_queue_softirq_done(q, set->ops->complete);
2333 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2336 mutex_lock(&all_q_mutex);
2338 list_add_tail(&q->all_q_node, &all_q_list);
2339 blk_mq_add_queue_tag_set(set, q);
2340 blk_mq_map_swqueue(q, cpu_online_mask);
2342 mutex_unlock(&all_q_mutex);
2345 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2348 ret = blk_mq_sched_init(q);
2350 return ERR_PTR(ret);
2356 kfree(q->queue_hw_ctx);
2358 free_percpu(q->queue_ctx);
2361 return ERR_PTR(-ENOMEM);
2363 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2365 void blk_mq_free_queue(struct request_queue *q)
2367 struct blk_mq_tag_set *set = q->tag_set;
2369 mutex_lock(&all_q_mutex);
2370 list_del_init(&q->all_q_node);
2371 mutex_unlock(&all_q_mutex);
2373 blk_mq_del_queue_tag_set(q);
2375 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2378 /* Basically redo blk_mq_init_queue with queue frozen */
2379 static void blk_mq_queue_reinit(struct request_queue *q,
2380 const struct cpumask *online_mask)
2382 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2384 blk_mq_debugfs_unregister_hctxs(q);
2385 blk_mq_sysfs_unregister(q);
2388 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2389 * we should change hctx numa_node according to new topology (this
2390 * involves free and re-allocate memory, worthy doing?)
2393 blk_mq_map_swqueue(q, online_mask);
2395 blk_mq_sysfs_register(q);
2396 blk_mq_debugfs_register_hctxs(q);
2400 * New online cpumask which is going to be set in this hotplug event.
2401 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2402 * one-by-one and dynamically allocating this could result in a failure.
2404 static struct cpumask cpuhp_online_new;
2406 static void blk_mq_queue_reinit_work(void)
2408 struct request_queue *q;
2410 mutex_lock(&all_q_mutex);
2412 * We need to freeze and reinit all existing queues. Freezing
2413 * involves synchronous wait for an RCU grace period and doing it
2414 * one by one may take a long time. Start freezing all queues in
2415 * one swoop and then wait for the completions so that freezing can
2416 * take place in parallel.
2418 list_for_each_entry(q, &all_q_list, all_q_node)
2419 blk_freeze_queue_start(q);
2420 list_for_each_entry(q, &all_q_list, all_q_node)
2421 blk_mq_freeze_queue_wait(q);
2423 list_for_each_entry(q, &all_q_list, all_q_node)
2424 blk_mq_queue_reinit(q, &cpuhp_online_new);
2426 list_for_each_entry(q, &all_q_list, all_q_node)
2427 blk_mq_unfreeze_queue(q);
2429 mutex_unlock(&all_q_mutex);
2432 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2434 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2435 blk_mq_queue_reinit_work();
2440 * Before hotadded cpu starts handling requests, new mappings must be
2441 * established. Otherwise, these requests in hw queue might never be
2444 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2445 * for CPU0, and ctx1 for CPU1).
2447 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2448 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2450 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2451 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2452 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2455 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2457 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2458 cpumask_set_cpu(cpu, &cpuhp_online_new);
2459 blk_mq_queue_reinit_work();
2463 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2467 for (i = 0; i < set->nr_hw_queues; i++)
2468 if (!__blk_mq_alloc_rq_map(set, i))
2475 blk_mq_free_rq_map(set->tags[i]);
2481 * Allocate the request maps associated with this tag_set. Note that this
2482 * may reduce the depth asked for, if memory is tight. set->queue_depth
2483 * will be updated to reflect the allocated depth.
2485 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2490 depth = set->queue_depth;
2492 err = __blk_mq_alloc_rq_maps(set);
2496 set->queue_depth >>= 1;
2497 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2501 } while (set->queue_depth);
2503 if (!set->queue_depth || err) {
2504 pr_err("blk-mq: failed to allocate request map\n");
2508 if (depth != set->queue_depth)
2509 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2510 depth, set->queue_depth);
2515 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2517 if (set->ops->map_queues)
2518 return set->ops->map_queues(set);
2520 return blk_mq_map_queues(set);
2524 * Alloc a tag set to be associated with one or more request queues.
2525 * May fail with EINVAL for various error conditions. May adjust the
2526 * requested depth down, if if it too large. In that case, the set
2527 * value will be stored in set->queue_depth.
2529 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2533 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2535 if (!set->nr_hw_queues)
2537 if (!set->queue_depth)
2539 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2542 if (!set->ops->queue_rq)
2545 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2546 pr_info("blk-mq: reduced tag depth to %u\n",
2548 set->queue_depth = BLK_MQ_MAX_DEPTH;
2552 * If a crashdump is active, then we are potentially in a very
2553 * memory constrained environment. Limit us to 1 queue and
2554 * 64 tags to prevent using too much memory.
2556 if (is_kdump_kernel()) {
2557 set->nr_hw_queues = 1;
2558 set->queue_depth = min(64U, set->queue_depth);
2561 * There is no use for more h/w queues than cpus.
2563 if (set->nr_hw_queues > nr_cpu_ids)
2564 set->nr_hw_queues = nr_cpu_ids;
2566 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2567 GFP_KERNEL, set->numa_node);
2572 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2573 GFP_KERNEL, set->numa_node);
2577 ret = blk_mq_update_queue_map(set);
2579 goto out_free_mq_map;
2581 ret = blk_mq_alloc_rq_maps(set);
2583 goto out_free_mq_map;
2585 mutex_init(&set->tag_list_lock);
2586 INIT_LIST_HEAD(&set->tag_list);
2598 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2600 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2604 for (i = 0; i < nr_cpu_ids; i++)
2605 blk_mq_free_map_and_requests(set, i);
2613 EXPORT_SYMBOL(blk_mq_free_tag_set);
2615 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2617 struct blk_mq_tag_set *set = q->tag_set;
2618 struct blk_mq_hw_ctx *hctx;
2624 blk_mq_freeze_queue(q);
2627 queue_for_each_hw_ctx(q, hctx, i) {
2631 * If we're using an MQ scheduler, just update the scheduler
2632 * queue depth. This is similar to what the old code would do.
2634 if (!hctx->sched_tags) {
2635 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2636 min(nr, set->queue_depth),
2639 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2647 q->nr_requests = nr;
2649 blk_mq_unfreeze_queue(q);
2654 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2657 struct request_queue *q;
2659 lockdep_assert_held(&set->tag_list_lock);
2661 if (nr_hw_queues > nr_cpu_ids)
2662 nr_hw_queues = nr_cpu_ids;
2663 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2666 list_for_each_entry(q, &set->tag_list, tag_set_list)
2667 blk_mq_freeze_queue(q);
2669 set->nr_hw_queues = nr_hw_queues;
2670 blk_mq_update_queue_map(set);
2671 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2672 blk_mq_realloc_hw_ctxs(set, q);
2673 blk_mq_queue_reinit(q, cpu_online_mask);
2676 list_for_each_entry(q, &set->tag_list, tag_set_list)
2677 blk_mq_unfreeze_queue(q);
2680 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2682 mutex_lock(&set->tag_list_lock);
2683 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2684 mutex_unlock(&set->tag_list_lock);
2686 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2688 /* Enable polling stats and return whether they were already enabled. */
2689 static bool blk_poll_stats_enable(struct request_queue *q)
2691 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2692 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2694 blk_stat_add_callback(q, q->poll_cb);
2698 static void blk_mq_poll_stats_start(struct request_queue *q)
2701 * We don't arm the callback if polling stats are not enabled or the
2702 * callback is already active.
2704 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2705 blk_stat_is_active(q->poll_cb))
2708 blk_stat_activate_msecs(q->poll_cb, 100);
2711 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2713 struct request_queue *q = cb->data;
2716 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2717 if (cb->stat[bucket].nr_samples)
2718 q->poll_stat[bucket] = cb->stat[bucket];
2722 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2723 struct blk_mq_hw_ctx *hctx,
2726 unsigned long ret = 0;
2730 * If stats collection isn't on, don't sleep but turn it on for
2733 if (!blk_poll_stats_enable(q))
2737 * As an optimistic guess, use half of the mean service time
2738 * for this type of request. We can (and should) make this smarter.
2739 * For instance, if the completion latencies are tight, we can
2740 * get closer than just half the mean. This is especially
2741 * important on devices where the completion latencies are longer
2742 * than ~10 usec. We do use the stats for the relevant IO size
2743 * if available which does lead to better estimates.
2745 bucket = blk_mq_poll_stats_bkt(rq);
2749 if (q->poll_stat[bucket].nr_samples)
2750 ret = (q->poll_stat[bucket].mean + 1) / 2;
2755 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2756 struct blk_mq_hw_ctx *hctx,
2759 struct hrtimer_sleeper hs;
2760 enum hrtimer_mode mode;
2764 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2770 * -1: don't ever hybrid sleep
2771 * 0: use half of prev avg
2772 * >0: use this specific value
2774 if (q->poll_nsec == -1)
2776 else if (q->poll_nsec > 0)
2777 nsecs = q->poll_nsec;
2779 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2784 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2787 * This will be replaced with the stats tracking code, using
2788 * 'avg_completion_time / 2' as the pre-sleep target.
2792 mode = HRTIMER_MODE_REL;
2793 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2794 hrtimer_set_expires(&hs.timer, kt);
2796 hrtimer_init_sleeper(&hs, current);
2798 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2800 set_current_state(TASK_UNINTERRUPTIBLE);
2801 hrtimer_start_expires(&hs.timer, mode);
2804 hrtimer_cancel(&hs.timer);
2805 mode = HRTIMER_MODE_ABS;
2806 } while (hs.task && !signal_pending(current));
2808 __set_current_state(TASK_RUNNING);
2809 destroy_hrtimer_on_stack(&hs.timer);
2813 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2815 struct request_queue *q = hctx->queue;
2819 * If we sleep, have the caller restart the poll loop to reset
2820 * the state. Like for the other success return cases, the
2821 * caller is responsible for checking if the IO completed. If
2822 * the IO isn't complete, we'll get called again and will go
2823 * straight to the busy poll loop.
2825 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2828 hctx->poll_considered++;
2830 state = current->state;
2831 while (!need_resched()) {
2834 hctx->poll_invoked++;
2836 ret = q->mq_ops->poll(hctx, rq->tag);
2838 hctx->poll_success++;
2839 set_current_state(TASK_RUNNING);
2843 if (signal_pending_state(state, current))
2844 set_current_state(TASK_RUNNING);
2846 if (current->state == TASK_RUNNING)
2856 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2858 struct blk_mq_hw_ctx *hctx;
2859 struct blk_plug *plug;
2862 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2863 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2866 plug = current->plug;
2868 blk_flush_plug_list(plug, false);
2870 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2871 if (!blk_qc_t_is_internal(cookie))
2872 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2874 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2876 * With scheduling, if the request has completed, we'll
2877 * get a NULL return here, as we clear the sched tag when
2878 * that happens. The request still remains valid, like always,
2879 * so we should be safe with just the NULL check.
2885 return __blk_mq_poll(hctx, rq);
2887 EXPORT_SYMBOL_GPL(blk_mq_poll);
2889 void blk_mq_disable_hotplug(void)
2891 mutex_lock(&all_q_mutex);
2894 void blk_mq_enable_hotplug(void)
2896 mutex_unlock(&all_q_mutex);
2899 static int __init blk_mq_init(void)
2901 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2902 blk_mq_hctx_notify_dead);
2904 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2905 blk_mq_queue_reinit_prepare,
2906 blk_mq_queue_reinit_dead);
2909 subsys_initcall(blk_mq_init);