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 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
158 * mpt3sas driver such that this function can be removed.
160 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
164 spin_lock_irqsave(q->queue_lock, flags);
165 queue_flag_set(QUEUE_FLAG_QUIESCED, q);
166 spin_unlock_irqrestore(q->queue_lock, flags);
168 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
171 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
174 * Note: this function does not prevent that the struct request end_io()
175 * callback function is invoked. Once this function is returned, we make
176 * sure no dispatch can happen until the queue is unquiesced via
177 * blk_mq_unquiesce_queue().
179 void blk_mq_quiesce_queue(struct request_queue *q)
181 struct blk_mq_hw_ctx *hctx;
185 blk_mq_quiesce_queue_nowait(q);
187 queue_for_each_hw_ctx(q, hctx, i) {
188 if (hctx->flags & BLK_MQ_F_BLOCKING)
189 synchronize_srcu(hctx->queue_rq_srcu);
196 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
199 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
202 * This function recovers queue into the state before quiescing
203 * which is done by blk_mq_quiesce_queue.
205 void blk_mq_unquiesce_queue(struct request_queue *q)
209 spin_lock_irqsave(q->queue_lock, flags);
210 queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
211 spin_unlock_irqrestore(q->queue_lock, flags);
213 /* dispatch requests which are inserted during quiescing */
214 blk_mq_run_hw_queues(q, true);
216 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
218 void blk_mq_wake_waiters(struct request_queue *q)
220 struct blk_mq_hw_ctx *hctx;
223 queue_for_each_hw_ctx(q, hctx, i)
224 if (blk_mq_hw_queue_mapped(hctx))
225 blk_mq_tag_wakeup_all(hctx->tags, true);
228 * If we are called because the queue has now been marked as
229 * dying, we need to ensure that processes currently waiting on
230 * the queue are notified as well.
232 wake_up_all(&q->mq_freeze_wq);
235 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
237 return blk_mq_has_free_tags(hctx->tags);
239 EXPORT_SYMBOL(blk_mq_can_queue);
241 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
242 unsigned int tag, unsigned int op)
244 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
245 struct request *rq = tags->static_rqs[tag];
249 if (data->flags & BLK_MQ_REQ_INTERNAL) {
251 rq->internal_tag = tag;
253 if (blk_mq_tag_busy(data->hctx)) {
254 rq->rq_flags = RQF_MQ_INFLIGHT;
255 atomic_inc(&data->hctx->nr_active);
258 rq->internal_tag = -1;
259 data->hctx->tags->rqs[rq->tag] = rq;
262 INIT_LIST_HEAD(&rq->queuelist);
263 /* csd/requeue_work/fifo_time is initialized before use */
265 rq->mq_ctx = data->ctx;
267 if (blk_queue_io_stat(data->q))
268 rq->rq_flags |= RQF_IO_STAT;
269 /* do not touch atomic flags, it needs atomic ops against the timer */
271 INIT_HLIST_NODE(&rq->hash);
272 RB_CLEAR_NODE(&rq->rb_node);
275 rq->start_time = jiffies;
276 #ifdef CONFIG_BLK_CGROUP
278 set_start_time_ns(rq);
279 rq->io_start_time_ns = 0;
281 rq->nr_phys_segments = 0;
282 #if defined(CONFIG_BLK_DEV_INTEGRITY)
283 rq->nr_integrity_segments = 0;
286 /* tag was already set */
289 INIT_LIST_HEAD(&rq->timeout_list);
293 rq->end_io_data = NULL;
296 data->ctx->rq_dispatched[op_is_sync(op)]++;
300 static struct request *blk_mq_get_request(struct request_queue *q,
301 struct bio *bio, unsigned int op,
302 struct blk_mq_alloc_data *data)
304 struct elevator_queue *e = q->elevator;
308 blk_queue_enter_live(q);
310 if (likely(!data->ctx))
311 data->ctx = blk_mq_get_ctx(q);
312 if (likely(!data->hctx))
313 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
315 data->flags |= BLK_MQ_REQ_NOWAIT;
318 data->flags |= BLK_MQ_REQ_INTERNAL;
321 * Flush requests are special and go directly to the
324 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
325 e->type->ops.mq.limit_depth(op, data);
328 tag = blk_mq_get_tag(data);
329 if (tag == BLK_MQ_TAG_FAIL) {
334 rq = blk_mq_rq_ctx_init(data, tag, op);
335 if (!op_is_flush(op)) {
337 if (e && e->type->ops.mq.prepare_request) {
338 if (e->type->icq_cache && rq_ioc(bio))
339 blk_mq_sched_assign_ioc(rq, bio);
341 e->type->ops.mq.prepare_request(rq, bio);
342 rq->rq_flags |= RQF_ELVPRIV;
345 data->hctx->queued++;
349 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
352 struct blk_mq_alloc_data alloc_data = { .flags = flags };
356 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
360 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
362 blk_mq_put_ctx(alloc_data.ctx);
366 return ERR_PTR(-EWOULDBLOCK);
369 rq->__sector = (sector_t) -1;
370 rq->bio = rq->biotail = NULL;
373 EXPORT_SYMBOL(blk_mq_alloc_request);
375 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
376 unsigned int op, unsigned int flags, unsigned int hctx_idx)
378 struct blk_mq_alloc_data alloc_data = { .flags = flags };
384 * If the tag allocator sleeps we could get an allocation for a
385 * different hardware context. No need to complicate the low level
386 * allocator for this for the rare use case of a command tied to
389 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
390 return ERR_PTR(-EINVAL);
392 if (hctx_idx >= q->nr_hw_queues)
393 return ERR_PTR(-EIO);
395 ret = blk_queue_enter(q, true);
400 * Check if the hardware context is actually mapped to anything.
401 * If not tell the caller that it should skip this queue.
403 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
404 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
406 return ERR_PTR(-EXDEV);
408 cpu = cpumask_first(alloc_data.hctx->cpumask);
409 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
411 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
416 return ERR_PTR(-EWOULDBLOCK);
420 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
422 void blk_mq_free_request(struct request *rq)
424 struct request_queue *q = rq->q;
425 struct elevator_queue *e = q->elevator;
426 struct blk_mq_ctx *ctx = rq->mq_ctx;
427 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
428 const int sched_tag = rq->internal_tag;
430 if (rq->rq_flags & RQF_ELVPRIV) {
431 if (e && e->type->ops.mq.finish_request)
432 e->type->ops.mq.finish_request(rq);
434 put_io_context(rq->elv.icq->ioc);
439 ctx->rq_completed[rq_is_sync(rq)]++;
440 if (rq->rq_flags & RQF_MQ_INFLIGHT)
441 atomic_dec(&hctx->nr_active);
443 wbt_done(q->rq_wb, &rq->issue_stat);
445 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
446 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
448 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
450 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
451 blk_mq_sched_restart(hctx);
454 EXPORT_SYMBOL_GPL(blk_mq_free_request);
456 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
458 blk_account_io_done(rq);
461 wbt_done(rq->q->rq_wb, &rq->issue_stat);
462 rq->end_io(rq, error);
464 if (unlikely(blk_bidi_rq(rq)))
465 blk_mq_free_request(rq->next_rq);
466 blk_mq_free_request(rq);
469 EXPORT_SYMBOL(__blk_mq_end_request);
471 void blk_mq_end_request(struct request *rq, blk_status_t error)
473 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
475 __blk_mq_end_request(rq, error);
477 EXPORT_SYMBOL(blk_mq_end_request);
479 static void __blk_mq_complete_request_remote(void *data)
481 struct request *rq = data;
483 rq->q->softirq_done_fn(rq);
486 static void __blk_mq_complete_request(struct request *rq)
488 struct blk_mq_ctx *ctx = rq->mq_ctx;
492 if (rq->internal_tag != -1)
493 blk_mq_sched_completed_request(rq);
494 if (rq->rq_flags & RQF_STATS) {
495 blk_mq_poll_stats_start(rq->q);
499 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
500 rq->q->softirq_done_fn(rq);
505 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
506 shared = cpus_share_cache(cpu, ctx->cpu);
508 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
509 rq->csd.func = __blk_mq_complete_request_remote;
512 smp_call_function_single_async(ctx->cpu, &rq->csd);
514 rq->q->softirq_done_fn(rq);
520 * blk_mq_complete_request - end I/O on a request
521 * @rq: the request being processed
524 * Ends all I/O on a request. It does not handle partial completions.
525 * The actual completion happens out-of-order, through a IPI handler.
527 void blk_mq_complete_request(struct request *rq)
529 struct request_queue *q = rq->q;
531 if (unlikely(blk_should_fake_timeout(q)))
533 if (!blk_mark_rq_complete(rq))
534 __blk_mq_complete_request(rq);
536 EXPORT_SYMBOL(blk_mq_complete_request);
538 int blk_mq_request_started(struct request *rq)
540 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
542 EXPORT_SYMBOL_GPL(blk_mq_request_started);
544 void blk_mq_start_request(struct request *rq)
546 struct request_queue *q = rq->q;
548 blk_mq_sched_started_request(rq);
550 trace_block_rq_issue(q, rq);
552 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
553 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
554 rq->rq_flags |= RQF_STATS;
555 wbt_issue(q->rq_wb, &rq->issue_stat);
561 * Ensure that ->deadline is visible before set the started
562 * flag and clear the completed flag.
564 smp_mb__before_atomic();
567 * Mark us as started and clear complete. Complete might have been
568 * set if requeue raced with timeout, which then marked it as
569 * complete. So be sure to clear complete again when we start
570 * the request, otherwise we'll ignore the completion event.
572 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
573 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
574 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
575 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
577 if (q->dma_drain_size && blk_rq_bytes(rq)) {
579 * Make sure space for the drain appears. We know we can do
580 * this because max_hw_segments has been adjusted to be one
581 * fewer than the device can handle.
583 rq->nr_phys_segments++;
586 EXPORT_SYMBOL(blk_mq_start_request);
589 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
590 * flag isn't set yet, so there may be race with timeout handler,
591 * but given rq->deadline is just set in .queue_rq() under
592 * this situation, the race won't be possible in reality because
593 * rq->timeout should be set as big enough to cover the window
594 * between blk_mq_start_request() called from .queue_rq() and
595 * clearing REQ_ATOM_STARTED here.
597 static void __blk_mq_requeue_request(struct request *rq)
599 struct request_queue *q = rq->q;
601 trace_block_rq_requeue(q, rq);
602 wbt_requeue(q->rq_wb, &rq->issue_stat);
603 blk_mq_sched_requeue_request(rq);
605 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
606 if (q->dma_drain_size && blk_rq_bytes(rq))
607 rq->nr_phys_segments--;
611 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
613 __blk_mq_requeue_request(rq);
615 BUG_ON(blk_queued_rq(rq));
616 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
618 EXPORT_SYMBOL(blk_mq_requeue_request);
620 static void blk_mq_requeue_work(struct work_struct *work)
622 struct request_queue *q =
623 container_of(work, struct request_queue, requeue_work.work);
625 struct request *rq, *next;
628 spin_lock_irqsave(&q->requeue_lock, flags);
629 list_splice_init(&q->requeue_list, &rq_list);
630 spin_unlock_irqrestore(&q->requeue_lock, flags);
632 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
633 if (!(rq->rq_flags & RQF_SOFTBARRIER))
636 rq->rq_flags &= ~RQF_SOFTBARRIER;
637 list_del_init(&rq->queuelist);
638 blk_mq_sched_insert_request(rq, true, false, false, true);
641 while (!list_empty(&rq_list)) {
642 rq = list_entry(rq_list.next, struct request, queuelist);
643 list_del_init(&rq->queuelist);
644 blk_mq_sched_insert_request(rq, false, false, false, true);
647 blk_mq_run_hw_queues(q, false);
650 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
651 bool kick_requeue_list)
653 struct request_queue *q = rq->q;
657 * We abuse this flag that is otherwise used by the I/O scheduler to
658 * request head insertation from the workqueue.
660 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
662 spin_lock_irqsave(&q->requeue_lock, flags);
664 rq->rq_flags |= RQF_SOFTBARRIER;
665 list_add(&rq->queuelist, &q->requeue_list);
667 list_add_tail(&rq->queuelist, &q->requeue_list);
669 spin_unlock_irqrestore(&q->requeue_lock, flags);
671 if (kick_requeue_list)
672 blk_mq_kick_requeue_list(q);
674 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
676 void blk_mq_kick_requeue_list(struct request_queue *q)
678 kblockd_schedule_delayed_work(&q->requeue_work, 0);
680 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
682 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
685 kblockd_schedule_delayed_work(&q->requeue_work,
686 msecs_to_jiffies(msecs));
688 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
690 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
692 if (tag < tags->nr_tags) {
693 prefetch(tags->rqs[tag]);
694 return tags->rqs[tag];
699 EXPORT_SYMBOL(blk_mq_tag_to_rq);
701 struct blk_mq_timeout_data {
703 unsigned int next_set;
706 void blk_mq_rq_timed_out(struct request *req, bool reserved)
708 const struct blk_mq_ops *ops = req->q->mq_ops;
709 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
712 * We know that complete is set at this point. If STARTED isn't set
713 * anymore, then the request isn't active and the "timeout" should
714 * just be ignored. This can happen due to the bitflag ordering.
715 * Timeout first checks if STARTED is set, and if it is, assumes
716 * the request is active. But if we race with completion, then
717 * both flags will get cleared. So check here again, and ignore
718 * a timeout event with a request that isn't active.
720 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
724 ret = ops->timeout(req, reserved);
728 __blk_mq_complete_request(req);
730 case BLK_EH_RESET_TIMER:
732 blk_clear_rq_complete(req);
734 case BLK_EH_NOT_HANDLED:
737 printk(KERN_ERR "block: bad eh return: %d\n", ret);
742 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
743 struct request *rq, void *priv, bool reserved)
745 struct blk_mq_timeout_data *data = priv;
747 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
751 * The rq being checked may have been freed and reallocated
752 * out already here, we avoid this race by checking rq->deadline
753 * and REQ_ATOM_COMPLETE flag together:
755 * - if rq->deadline is observed as new value because of
756 * reusing, the rq won't be timed out because of timing.
757 * - if rq->deadline is observed as previous value,
758 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
759 * because we put a barrier between setting rq->deadline
760 * and clearing the flag in blk_mq_start_request(), so
761 * this rq won't be timed out too.
763 if (time_after_eq(jiffies, rq->deadline)) {
764 if (!blk_mark_rq_complete(rq))
765 blk_mq_rq_timed_out(rq, reserved);
766 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
767 data->next = rq->deadline;
772 static void blk_mq_timeout_work(struct work_struct *work)
774 struct request_queue *q =
775 container_of(work, struct request_queue, timeout_work);
776 struct blk_mq_timeout_data data = {
782 /* A deadlock might occur if a request is stuck requiring a
783 * timeout at the same time a queue freeze is waiting
784 * completion, since the timeout code would not be able to
785 * acquire the queue reference here.
787 * That's why we don't use blk_queue_enter here; instead, we use
788 * percpu_ref_tryget directly, because we need to be able to
789 * obtain a reference even in the short window between the queue
790 * starting to freeze, by dropping the first reference in
791 * blk_freeze_queue_start, and the moment the last request is
792 * consumed, marked by the instant q_usage_counter reaches
795 if (!percpu_ref_tryget(&q->q_usage_counter))
798 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
801 data.next = blk_rq_timeout(round_jiffies_up(data.next));
802 mod_timer(&q->timeout, data.next);
804 struct blk_mq_hw_ctx *hctx;
806 queue_for_each_hw_ctx(q, hctx, i) {
807 /* the hctx may be unmapped, so check it here */
808 if (blk_mq_hw_queue_mapped(hctx))
809 blk_mq_tag_idle(hctx);
815 struct flush_busy_ctx_data {
816 struct blk_mq_hw_ctx *hctx;
817 struct list_head *list;
820 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
822 struct flush_busy_ctx_data *flush_data = data;
823 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
824 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
826 sbitmap_clear_bit(sb, bitnr);
827 spin_lock(&ctx->lock);
828 list_splice_tail_init(&ctx->rq_list, flush_data->list);
829 spin_unlock(&ctx->lock);
834 * Process software queues that have been marked busy, splicing them
835 * to the for-dispatch
837 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
839 struct flush_busy_ctx_data data = {
844 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
846 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
848 static inline unsigned int queued_to_index(unsigned int queued)
853 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
856 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
859 struct blk_mq_alloc_data data = {
861 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
862 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
865 might_sleep_if(wait);
870 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
871 data.flags |= BLK_MQ_REQ_RESERVED;
873 rq->tag = blk_mq_get_tag(&data);
875 if (blk_mq_tag_busy(data.hctx)) {
876 rq->rq_flags |= RQF_MQ_INFLIGHT;
877 atomic_inc(&data.hctx->nr_active);
879 data.hctx->tags->rqs[rq->tag] = rq;
885 return rq->tag != -1;
888 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
891 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
894 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
895 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
896 atomic_dec(&hctx->nr_active);
900 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
903 if (rq->tag == -1 || rq->internal_tag == -1)
906 __blk_mq_put_driver_tag(hctx, rq);
909 static void blk_mq_put_driver_tag(struct request *rq)
911 struct blk_mq_hw_ctx *hctx;
913 if (rq->tag == -1 || rq->internal_tag == -1)
916 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
917 __blk_mq_put_driver_tag(hctx, rq);
921 * If we fail getting a driver tag because all the driver tags are already
922 * assigned and on the dispatch list, BUT the first entry does not have a
923 * tag, then we could deadlock. For that case, move entries with assigned
924 * driver tags to the front, leaving the set of tagged requests in the
925 * same order, and the untagged set in the same order.
927 static bool reorder_tags_to_front(struct list_head *list)
929 struct request *rq, *tmp, *first = NULL;
931 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
935 list_move(&rq->queuelist, list);
941 return first != NULL;
944 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
947 struct blk_mq_hw_ctx *hctx;
949 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
951 list_del(&wait->entry);
952 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
953 blk_mq_run_hw_queue(hctx, true);
957 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
959 struct sbq_wait_state *ws;
962 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
963 * The thread which wins the race to grab this bit adds the hardware
964 * queue to the wait queue.
966 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
967 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
970 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
971 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
974 * As soon as this returns, it's no longer safe to fiddle with
975 * hctx->dispatch_wait, since a completion can wake up the wait queue
976 * and unlock the bit.
978 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
982 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
984 struct blk_mq_hw_ctx *hctx;
988 if (list_empty(list))
992 * Now process all the entries, sending them to the driver.
996 struct blk_mq_queue_data bd;
999 rq = list_first_entry(list, struct request, queuelist);
1000 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1001 if (!queued && reorder_tags_to_front(list))
1005 * The initial allocation attempt failed, so we need to
1006 * rerun the hardware queue when a tag is freed.
1008 if (!blk_mq_dispatch_wait_add(hctx))
1012 * It's possible that a tag was freed in the window
1013 * between the allocation failure and adding the
1014 * hardware queue to the wait queue.
1016 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1020 list_del_init(&rq->queuelist);
1025 * Flag last if we have no more requests, or if we have more
1026 * but can't assign a driver tag to it.
1028 if (list_empty(list))
1031 struct request *nxt;
1033 nxt = list_first_entry(list, struct request, queuelist);
1034 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1037 ret = q->mq_ops->queue_rq(hctx, &bd);
1038 if (ret == BLK_STS_RESOURCE) {
1039 blk_mq_put_driver_tag_hctx(hctx, rq);
1040 list_add(&rq->queuelist, list);
1041 __blk_mq_requeue_request(rq);
1045 if (unlikely(ret != BLK_STS_OK)) {
1047 blk_mq_end_request(rq, BLK_STS_IOERR);
1052 } while (!list_empty(list));
1054 hctx->dispatched[queued_to_index(queued)]++;
1057 * Any items that need requeuing? Stuff them into hctx->dispatch,
1058 * that is where we will continue on next queue run.
1060 if (!list_empty(list)) {
1062 * If an I/O scheduler has been configured and we got a driver
1063 * tag for the next request already, free it again.
1065 rq = list_first_entry(list, struct request, queuelist);
1066 blk_mq_put_driver_tag(rq);
1068 spin_lock(&hctx->lock);
1069 list_splice_init(list, &hctx->dispatch);
1070 spin_unlock(&hctx->lock);
1073 * If SCHED_RESTART was set by the caller of this function and
1074 * it is no longer set that means that it was cleared by another
1075 * thread and hence that a queue rerun is needed.
1077 * If TAG_WAITING is set that means that an I/O scheduler has
1078 * been configured and another thread is waiting for a driver
1079 * tag. To guarantee fairness, do not rerun this hardware queue
1080 * but let the other thread grab the driver tag.
1082 * If no I/O scheduler has been configured it is possible that
1083 * the hardware queue got stopped and restarted before requests
1084 * were pushed back onto the dispatch list. Rerun the queue to
1085 * avoid starvation. Notes:
1086 * - blk_mq_run_hw_queue() checks whether or not a queue has
1087 * been stopped before rerunning a queue.
1088 * - Some but not all block drivers stop a queue before
1089 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1092 if (!blk_mq_sched_needs_restart(hctx) &&
1093 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1094 blk_mq_run_hw_queue(hctx, true);
1097 return (queued + errors) != 0;
1100 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1104 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1105 cpu_online(hctx->next_cpu));
1107 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1109 blk_mq_sched_dispatch_requests(hctx);
1114 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1115 blk_mq_sched_dispatch_requests(hctx);
1116 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1121 * It'd be great if the workqueue API had a way to pass
1122 * in a mask and had some smarts for more clever placement.
1123 * For now we just round-robin here, switching for every
1124 * BLK_MQ_CPU_WORK_BATCH queued items.
1126 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1128 if (hctx->queue->nr_hw_queues == 1)
1129 return WORK_CPU_UNBOUND;
1131 if (--hctx->next_cpu_batch <= 0) {
1134 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1135 if (next_cpu >= nr_cpu_ids)
1136 next_cpu = cpumask_first(hctx->cpumask);
1138 hctx->next_cpu = next_cpu;
1139 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1142 return hctx->next_cpu;
1145 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1146 unsigned long msecs)
1148 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1151 if (unlikely(blk_mq_hctx_stopped(hctx)))
1154 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1155 int cpu = get_cpu();
1156 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1157 __blk_mq_run_hw_queue(hctx);
1165 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1167 msecs_to_jiffies(msecs));
1170 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1172 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1174 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1176 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1178 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1180 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1182 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1184 struct blk_mq_hw_ctx *hctx;
1187 queue_for_each_hw_ctx(q, hctx, i) {
1188 if (!blk_mq_hctx_has_pending(hctx) ||
1189 blk_mq_hctx_stopped(hctx))
1192 blk_mq_run_hw_queue(hctx, async);
1195 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1198 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1199 * @q: request queue.
1201 * The caller is responsible for serializing this function against
1202 * blk_mq_{start,stop}_hw_queue().
1204 bool blk_mq_queue_stopped(struct request_queue *q)
1206 struct blk_mq_hw_ctx *hctx;
1209 queue_for_each_hw_ctx(q, hctx, i)
1210 if (blk_mq_hctx_stopped(hctx))
1215 EXPORT_SYMBOL(blk_mq_queue_stopped);
1218 * This function is often used for pausing .queue_rq() by driver when
1219 * there isn't enough resource or some conditions aren't satisfied, and
1220 * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1222 * We do not guarantee that dispatch can be drained or blocked
1223 * after blk_mq_stop_hw_queue() returns. Please use
1224 * blk_mq_quiesce_queue() for that requirement.
1226 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1228 cancel_delayed_work(&hctx->run_work);
1230 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1232 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1235 * This function is often used for pausing .queue_rq() by driver when
1236 * there isn't enough resource or some conditions aren't satisfied, and
1237 * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1239 * We do not guarantee that dispatch can be drained or blocked
1240 * after blk_mq_stop_hw_queues() returns. Please use
1241 * blk_mq_quiesce_queue() for that requirement.
1243 void blk_mq_stop_hw_queues(struct request_queue *q)
1245 struct blk_mq_hw_ctx *hctx;
1248 queue_for_each_hw_ctx(q, hctx, i)
1249 blk_mq_stop_hw_queue(hctx);
1251 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1253 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1255 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1257 blk_mq_run_hw_queue(hctx, false);
1259 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1261 void blk_mq_start_hw_queues(struct request_queue *q)
1263 struct blk_mq_hw_ctx *hctx;
1266 queue_for_each_hw_ctx(q, hctx, i)
1267 blk_mq_start_hw_queue(hctx);
1269 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1271 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1273 if (!blk_mq_hctx_stopped(hctx))
1276 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1277 blk_mq_run_hw_queue(hctx, async);
1279 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1281 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1283 struct blk_mq_hw_ctx *hctx;
1286 queue_for_each_hw_ctx(q, hctx, i)
1287 blk_mq_start_stopped_hw_queue(hctx, async);
1289 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1291 static void blk_mq_run_work_fn(struct work_struct *work)
1293 struct blk_mq_hw_ctx *hctx;
1295 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1298 * If we are stopped, don't run the queue. The exception is if
1299 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1300 * the STOPPED bit and run it.
1302 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1303 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1306 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1307 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1310 __blk_mq_run_hw_queue(hctx);
1314 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1316 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1320 * Stop the hw queue, then modify currently delayed work.
1321 * This should prevent us from running the queue prematurely.
1322 * Mark the queue as auto-clearing STOPPED when it runs.
1324 blk_mq_stop_hw_queue(hctx);
1325 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1326 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1328 msecs_to_jiffies(msecs));
1330 EXPORT_SYMBOL(blk_mq_delay_queue);
1332 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1336 struct blk_mq_ctx *ctx = rq->mq_ctx;
1338 lockdep_assert_held(&ctx->lock);
1340 trace_block_rq_insert(hctx->queue, rq);
1343 list_add(&rq->queuelist, &ctx->rq_list);
1345 list_add_tail(&rq->queuelist, &ctx->rq_list);
1348 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1351 struct blk_mq_ctx *ctx = rq->mq_ctx;
1353 lockdep_assert_held(&ctx->lock);
1355 __blk_mq_insert_req_list(hctx, rq, at_head);
1356 blk_mq_hctx_mark_pending(hctx, ctx);
1359 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1360 struct list_head *list)
1364 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1367 spin_lock(&ctx->lock);
1368 while (!list_empty(list)) {
1371 rq = list_first_entry(list, struct request, queuelist);
1372 BUG_ON(rq->mq_ctx != ctx);
1373 list_del_init(&rq->queuelist);
1374 __blk_mq_insert_req_list(hctx, rq, false);
1376 blk_mq_hctx_mark_pending(hctx, ctx);
1377 spin_unlock(&ctx->lock);
1380 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1382 struct request *rqa = container_of(a, struct request, queuelist);
1383 struct request *rqb = container_of(b, struct request, queuelist);
1385 return !(rqa->mq_ctx < rqb->mq_ctx ||
1386 (rqa->mq_ctx == rqb->mq_ctx &&
1387 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1390 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1392 struct blk_mq_ctx *this_ctx;
1393 struct request_queue *this_q;
1396 LIST_HEAD(ctx_list);
1399 list_splice_init(&plug->mq_list, &list);
1401 list_sort(NULL, &list, plug_ctx_cmp);
1407 while (!list_empty(&list)) {
1408 rq = list_entry_rq(list.next);
1409 list_del_init(&rq->queuelist);
1411 if (rq->mq_ctx != this_ctx) {
1413 trace_block_unplug(this_q, depth, from_schedule);
1414 blk_mq_sched_insert_requests(this_q, this_ctx,
1419 this_ctx = rq->mq_ctx;
1425 list_add_tail(&rq->queuelist, &ctx_list);
1429 * If 'this_ctx' is set, we know we have entries to complete
1430 * on 'ctx_list'. Do those.
1433 trace_block_unplug(this_q, depth, from_schedule);
1434 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1439 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1441 blk_init_request_from_bio(rq, bio);
1443 blk_account_io_start(rq, true);
1446 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1448 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1449 !blk_queue_nomerges(hctx->queue);
1452 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1453 struct blk_mq_ctx *ctx,
1456 spin_lock(&ctx->lock);
1457 __blk_mq_insert_request(hctx, rq, false);
1458 spin_unlock(&ctx->lock);
1461 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1464 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1466 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1469 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1471 blk_qc_t *cookie, bool may_sleep)
1473 struct request_queue *q = rq->q;
1474 struct blk_mq_queue_data bd = {
1478 blk_qc_t new_cookie;
1480 bool run_queue = true;
1482 /* RCU or SRCU read lock is needed before checking quiesced flag */
1483 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1491 if (!blk_mq_get_driver_tag(rq, NULL, false))
1494 new_cookie = request_to_qc_t(hctx, rq);
1497 * For OK queue, we are done. For error, kill it. Any other
1498 * error (busy), just add it to our list as we previously
1501 ret = q->mq_ops->queue_rq(hctx, &bd);
1504 *cookie = new_cookie;
1506 case BLK_STS_RESOURCE:
1507 __blk_mq_requeue_request(rq);
1510 *cookie = BLK_QC_T_NONE;
1511 blk_mq_end_request(rq, ret);
1516 blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1519 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1520 struct request *rq, blk_qc_t *cookie)
1522 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1524 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1527 unsigned int srcu_idx;
1531 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1532 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1533 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1537 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1539 const int is_sync = op_is_sync(bio->bi_opf);
1540 const int is_flush_fua = op_is_flush(bio->bi_opf);
1541 struct blk_mq_alloc_data data = { .flags = 0 };
1543 unsigned int request_count = 0;
1544 struct blk_plug *plug;
1545 struct request *same_queue_rq = NULL;
1547 unsigned int wb_acct;
1549 blk_queue_bounce(q, &bio);
1551 blk_queue_split(q, &bio);
1553 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1555 return BLK_QC_T_NONE;
1558 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1559 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1560 return BLK_QC_T_NONE;
1562 if (blk_mq_sched_bio_merge(q, bio))
1563 return BLK_QC_T_NONE;
1565 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1567 trace_block_getrq(q, bio, bio->bi_opf);
1569 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1570 if (unlikely(!rq)) {
1571 __wbt_done(q->rq_wb, wb_acct);
1572 if (bio->bi_opf & REQ_NOWAIT)
1573 bio_wouldblock_error(bio);
1574 return BLK_QC_T_NONE;
1577 wbt_track(&rq->issue_stat, wb_acct);
1579 cookie = request_to_qc_t(data.hctx, rq);
1581 plug = current->plug;
1582 if (unlikely(is_flush_fua)) {
1583 blk_mq_put_ctx(data.ctx);
1584 blk_mq_bio_to_request(rq, bio);
1586 blk_mq_sched_insert_request(rq, false, true, true,
1589 blk_insert_flush(rq);
1590 blk_mq_run_hw_queue(data.hctx, true);
1592 } else if (plug && q->nr_hw_queues == 1) {
1593 struct request *last = NULL;
1595 blk_mq_put_ctx(data.ctx);
1596 blk_mq_bio_to_request(rq, bio);
1599 * @request_count may become stale because of schedule
1600 * out, so check the list again.
1602 if (list_empty(&plug->mq_list))
1604 else if (blk_queue_nomerges(q))
1605 request_count = blk_plug_queued_count(q);
1608 trace_block_plug(q);
1610 last = list_entry_rq(plug->mq_list.prev);
1612 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1613 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1614 blk_flush_plug_list(plug, false);
1615 trace_block_plug(q);
1618 list_add_tail(&rq->queuelist, &plug->mq_list);
1619 } else if (plug && !blk_queue_nomerges(q)) {
1620 blk_mq_bio_to_request(rq, bio);
1623 * We do limited plugging. If the bio can be merged, do that.
1624 * Otherwise the existing request in the plug list will be
1625 * issued. So the plug list will have one request at most
1626 * The plug list might get flushed before this. If that happens,
1627 * the plug list is empty, and same_queue_rq is invalid.
1629 if (list_empty(&plug->mq_list))
1630 same_queue_rq = NULL;
1632 list_del_init(&same_queue_rq->queuelist);
1633 list_add_tail(&rq->queuelist, &plug->mq_list);
1635 blk_mq_put_ctx(data.ctx);
1637 if (same_queue_rq) {
1638 data.hctx = blk_mq_map_queue(q,
1639 same_queue_rq->mq_ctx->cpu);
1640 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1643 } else if (q->nr_hw_queues > 1 && is_sync) {
1644 blk_mq_put_ctx(data.ctx);
1645 blk_mq_bio_to_request(rq, bio);
1646 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1647 } else if (q->elevator) {
1648 blk_mq_put_ctx(data.ctx);
1649 blk_mq_bio_to_request(rq, bio);
1650 blk_mq_sched_insert_request(rq, false, true, true, true);
1652 blk_mq_put_ctx(data.ctx);
1653 blk_mq_bio_to_request(rq, bio);
1654 blk_mq_queue_io(data.hctx, data.ctx, rq);
1655 blk_mq_run_hw_queue(data.hctx, true);
1661 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1662 unsigned int hctx_idx)
1666 if (tags->rqs && set->ops->exit_request) {
1669 for (i = 0; i < tags->nr_tags; i++) {
1670 struct request *rq = tags->static_rqs[i];
1674 set->ops->exit_request(set, rq, hctx_idx);
1675 tags->static_rqs[i] = NULL;
1679 while (!list_empty(&tags->page_list)) {
1680 page = list_first_entry(&tags->page_list, struct page, lru);
1681 list_del_init(&page->lru);
1683 * Remove kmemleak object previously allocated in
1684 * blk_mq_init_rq_map().
1686 kmemleak_free(page_address(page));
1687 __free_pages(page, page->private);
1691 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1695 kfree(tags->static_rqs);
1696 tags->static_rqs = NULL;
1698 blk_mq_free_tags(tags);
1701 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1702 unsigned int hctx_idx,
1703 unsigned int nr_tags,
1704 unsigned int reserved_tags)
1706 struct blk_mq_tags *tags;
1709 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1710 if (node == NUMA_NO_NODE)
1711 node = set->numa_node;
1713 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1714 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1718 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1719 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1722 blk_mq_free_tags(tags);
1726 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1727 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1729 if (!tags->static_rqs) {
1731 blk_mq_free_tags(tags);
1738 static size_t order_to_size(unsigned int order)
1740 return (size_t)PAGE_SIZE << order;
1743 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1744 unsigned int hctx_idx, unsigned int depth)
1746 unsigned int i, j, entries_per_page, max_order = 4;
1747 size_t rq_size, left;
1750 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1751 if (node == NUMA_NO_NODE)
1752 node = set->numa_node;
1754 INIT_LIST_HEAD(&tags->page_list);
1757 * rq_size is the size of the request plus driver payload, rounded
1758 * to the cacheline size
1760 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1762 left = rq_size * depth;
1764 for (i = 0; i < depth; ) {
1765 int this_order = max_order;
1770 while (this_order && left < order_to_size(this_order - 1))
1774 page = alloc_pages_node(node,
1775 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1781 if (order_to_size(this_order) < rq_size)
1788 page->private = this_order;
1789 list_add_tail(&page->lru, &tags->page_list);
1791 p = page_address(page);
1793 * Allow kmemleak to scan these pages as they contain pointers
1794 * to additional allocations like via ops->init_request().
1796 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1797 entries_per_page = order_to_size(this_order) / rq_size;
1798 to_do = min(entries_per_page, depth - i);
1799 left -= to_do * rq_size;
1800 for (j = 0; j < to_do; j++) {
1801 struct request *rq = p;
1803 tags->static_rqs[i] = rq;
1804 if (set->ops->init_request) {
1805 if (set->ops->init_request(set, rq, hctx_idx,
1807 tags->static_rqs[i] = NULL;
1819 blk_mq_free_rqs(set, tags, hctx_idx);
1824 * 'cpu' is going away. splice any existing rq_list entries from this
1825 * software queue to the hw queue dispatch list, and ensure that it
1828 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1830 struct blk_mq_hw_ctx *hctx;
1831 struct blk_mq_ctx *ctx;
1834 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1835 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1837 spin_lock(&ctx->lock);
1838 if (!list_empty(&ctx->rq_list)) {
1839 list_splice_init(&ctx->rq_list, &tmp);
1840 blk_mq_hctx_clear_pending(hctx, ctx);
1842 spin_unlock(&ctx->lock);
1844 if (list_empty(&tmp))
1847 spin_lock(&hctx->lock);
1848 list_splice_tail_init(&tmp, &hctx->dispatch);
1849 spin_unlock(&hctx->lock);
1851 blk_mq_run_hw_queue(hctx, true);
1855 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1857 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1861 /* hctx->ctxs will be freed in queue's release handler */
1862 static void blk_mq_exit_hctx(struct request_queue *q,
1863 struct blk_mq_tag_set *set,
1864 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1866 blk_mq_debugfs_unregister_hctx(hctx);
1868 blk_mq_tag_idle(hctx);
1870 if (set->ops->exit_request)
1871 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1873 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1875 if (set->ops->exit_hctx)
1876 set->ops->exit_hctx(hctx, hctx_idx);
1878 if (hctx->flags & BLK_MQ_F_BLOCKING)
1879 cleanup_srcu_struct(hctx->queue_rq_srcu);
1881 blk_mq_remove_cpuhp(hctx);
1882 blk_free_flush_queue(hctx->fq);
1883 sbitmap_free(&hctx->ctx_map);
1886 static void blk_mq_exit_hw_queues(struct request_queue *q,
1887 struct blk_mq_tag_set *set, int nr_queue)
1889 struct blk_mq_hw_ctx *hctx;
1892 queue_for_each_hw_ctx(q, hctx, i) {
1895 blk_mq_exit_hctx(q, set, hctx, i);
1899 static int blk_mq_init_hctx(struct request_queue *q,
1900 struct blk_mq_tag_set *set,
1901 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1905 node = hctx->numa_node;
1906 if (node == NUMA_NO_NODE)
1907 node = hctx->numa_node = set->numa_node;
1909 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1910 spin_lock_init(&hctx->lock);
1911 INIT_LIST_HEAD(&hctx->dispatch);
1913 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1915 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1917 hctx->tags = set->tags[hctx_idx];
1920 * Allocate space for all possible cpus to avoid allocation at
1923 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1926 goto unregister_cpu_notifier;
1928 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1934 if (set->ops->init_hctx &&
1935 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1938 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1941 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1943 goto sched_exit_hctx;
1945 if (set->ops->init_request &&
1946 set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1950 if (hctx->flags & BLK_MQ_F_BLOCKING)
1951 init_srcu_struct(hctx->queue_rq_srcu);
1953 blk_mq_debugfs_register_hctx(q, hctx);
1960 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1962 if (set->ops->exit_hctx)
1963 set->ops->exit_hctx(hctx, hctx_idx);
1965 sbitmap_free(&hctx->ctx_map);
1968 unregister_cpu_notifier:
1969 blk_mq_remove_cpuhp(hctx);
1973 static void blk_mq_init_cpu_queues(struct request_queue *q,
1974 unsigned int nr_hw_queues)
1978 for_each_possible_cpu(i) {
1979 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1980 struct blk_mq_hw_ctx *hctx;
1983 spin_lock_init(&__ctx->lock);
1984 INIT_LIST_HEAD(&__ctx->rq_list);
1987 /* If the cpu isn't online, the cpu is mapped to first hctx */
1991 hctx = blk_mq_map_queue(q, i);
1994 * Set local node, IFF we have more than one hw queue. If
1995 * not, we remain on the home node of the device
1997 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1998 hctx->numa_node = local_memory_node(cpu_to_node(i));
2002 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2006 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2007 set->queue_depth, set->reserved_tags);
2008 if (!set->tags[hctx_idx])
2011 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2016 blk_mq_free_rq_map(set->tags[hctx_idx]);
2017 set->tags[hctx_idx] = NULL;
2021 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2022 unsigned int hctx_idx)
2024 if (set->tags[hctx_idx]) {
2025 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2026 blk_mq_free_rq_map(set->tags[hctx_idx]);
2027 set->tags[hctx_idx] = NULL;
2031 static void blk_mq_map_swqueue(struct request_queue *q,
2032 const struct cpumask *online_mask)
2034 unsigned int i, hctx_idx;
2035 struct blk_mq_hw_ctx *hctx;
2036 struct blk_mq_ctx *ctx;
2037 struct blk_mq_tag_set *set = q->tag_set;
2040 * Avoid others reading imcomplete hctx->cpumask through sysfs
2042 mutex_lock(&q->sysfs_lock);
2044 queue_for_each_hw_ctx(q, hctx, i) {
2045 cpumask_clear(hctx->cpumask);
2050 * Map software to hardware queues
2052 for_each_possible_cpu(i) {
2053 /* If the cpu isn't online, the cpu is mapped to first hctx */
2054 if (!cpumask_test_cpu(i, online_mask))
2057 hctx_idx = q->mq_map[i];
2058 /* unmapped hw queue can be remapped after CPU topo changed */
2059 if (!set->tags[hctx_idx] &&
2060 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2062 * If tags initialization fail for some hctx,
2063 * that hctx won't be brought online. In this
2064 * case, remap the current ctx to hctx[0] which
2065 * is guaranteed to always have tags allocated
2070 ctx = per_cpu_ptr(q->queue_ctx, i);
2071 hctx = blk_mq_map_queue(q, i);
2073 cpumask_set_cpu(i, hctx->cpumask);
2074 ctx->index_hw = hctx->nr_ctx;
2075 hctx->ctxs[hctx->nr_ctx++] = ctx;
2078 mutex_unlock(&q->sysfs_lock);
2080 queue_for_each_hw_ctx(q, hctx, i) {
2082 * If no software queues are mapped to this hardware queue,
2083 * disable it and free the request entries.
2085 if (!hctx->nr_ctx) {
2086 /* Never unmap queue 0. We need it as a
2087 * fallback in case of a new remap fails
2090 if (i && set->tags[i])
2091 blk_mq_free_map_and_requests(set, i);
2097 hctx->tags = set->tags[i];
2098 WARN_ON(!hctx->tags);
2101 * Set the map size to the number of mapped software queues.
2102 * This is more accurate and more efficient than looping
2103 * over all possibly mapped software queues.
2105 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2108 * Initialize batch roundrobin counts
2110 hctx->next_cpu = cpumask_first(hctx->cpumask);
2111 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2116 * Caller needs to ensure that we're either frozen/quiesced, or that
2117 * the queue isn't live yet.
2119 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2121 struct blk_mq_hw_ctx *hctx;
2124 queue_for_each_hw_ctx(q, hctx, i) {
2126 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2127 atomic_inc(&q->shared_hctx_restart);
2128 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2130 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2131 atomic_dec(&q->shared_hctx_restart);
2132 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2137 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2140 struct request_queue *q;
2142 lockdep_assert_held(&set->tag_list_lock);
2144 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2145 blk_mq_freeze_queue(q);
2146 queue_set_hctx_shared(q, shared);
2147 blk_mq_unfreeze_queue(q);
2151 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2153 struct blk_mq_tag_set *set = q->tag_set;
2155 mutex_lock(&set->tag_list_lock);
2156 list_del_rcu(&q->tag_set_list);
2157 INIT_LIST_HEAD(&q->tag_set_list);
2158 if (list_is_singular(&set->tag_list)) {
2159 /* just transitioned to unshared */
2160 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2161 /* update existing queue */
2162 blk_mq_update_tag_set_depth(set, false);
2164 mutex_unlock(&set->tag_list_lock);
2169 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2170 struct request_queue *q)
2174 mutex_lock(&set->tag_list_lock);
2176 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2177 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2178 set->flags |= BLK_MQ_F_TAG_SHARED;
2179 /* update existing queue */
2180 blk_mq_update_tag_set_depth(set, true);
2182 if (set->flags & BLK_MQ_F_TAG_SHARED)
2183 queue_set_hctx_shared(q, true);
2184 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2186 mutex_unlock(&set->tag_list_lock);
2190 * It is the actual release handler for mq, but we do it from
2191 * request queue's release handler for avoiding use-after-free
2192 * and headache because q->mq_kobj shouldn't have been introduced,
2193 * but we can't group ctx/kctx kobj without it.
2195 void blk_mq_release(struct request_queue *q)
2197 struct blk_mq_hw_ctx *hctx;
2200 /* hctx kobj stays in hctx */
2201 queue_for_each_hw_ctx(q, hctx, i) {
2204 kobject_put(&hctx->kobj);
2209 kfree(q->queue_hw_ctx);
2212 * release .mq_kobj and sw queue's kobject now because
2213 * both share lifetime with request queue.
2215 blk_mq_sysfs_deinit(q);
2217 free_percpu(q->queue_ctx);
2220 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2222 struct request_queue *uninit_q, *q;
2224 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2226 return ERR_PTR(-ENOMEM);
2228 q = blk_mq_init_allocated_queue(set, uninit_q);
2230 blk_cleanup_queue(uninit_q);
2234 EXPORT_SYMBOL(blk_mq_init_queue);
2236 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2238 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2240 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2241 __alignof__(struct blk_mq_hw_ctx)) !=
2242 sizeof(struct blk_mq_hw_ctx));
2244 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2245 hw_ctx_size += sizeof(struct srcu_struct);
2250 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2251 struct request_queue *q)
2254 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2256 blk_mq_sysfs_unregister(q);
2257 for (i = 0; i < set->nr_hw_queues; i++) {
2263 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2264 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2269 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2276 atomic_set(&hctxs[i]->nr_active, 0);
2277 hctxs[i]->numa_node = node;
2278 hctxs[i]->queue_num = i;
2280 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2281 free_cpumask_var(hctxs[i]->cpumask);
2286 blk_mq_hctx_kobj_init(hctxs[i]);
2288 for (j = i; j < q->nr_hw_queues; j++) {
2289 struct blk_mq_hw_ctx *hctx = hctxs[j];
2293 blk_mq_free_map_and_requests(set, j);
2294 blk_mq_exit_hctx(q, set, hctx, j);
2295 kobject_put(&hctx->kobj);
2300 q->nr_hw_queues = i;
2301 blk_mq_sysfs_register(q);
2304 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2305 struct request_queue *q)
2307 /* mark the queue as mq asap */
2308 q->mq_ops = set->ops;
2310 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2311 blk_mq_poll_stats_bkt,
2312 BLK_MQ_POLL_STATS_BKTS, q);
2316 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2320 /* init q->mq_kobj and sw queues' kobjects */
2321 blk_mq_sysfs_init(q);
2323 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2324 GFP_KERNEL, set->numa_node);
2325 if (!q->queue_hw_ctx)
2328 q->mq_map = set->mq_map;
2330 blk_mq_realloc_hw_ctxs(set, q);
2331 if (!q->nr_hw_queues)
2334 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2335 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2337 q->nr_queues = nr_cpu_ids;
2339 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2341 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2342 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2344 q->sg_reserved_size = INT_MAX;
2346 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2347 INIT_LIST_HEAD(&q->requeue_list);
2348 spin_lock_init(&q->requeue_lock);
2350 blk_queue_make_request(q, blk_mq_make_request);
2353 * Do this after blk_queue_make_request() overrides it...
2355 q->nr_requests = set->queue_depth;
2358 * Default to classic polling
2362 if (set->ops->complete)
2363 blk_queue_softirq_done(q, set->ops->complete);
2365 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2368 mutex_lock(&all_q_mutex);
2370 list_add_tail(&q->all_q_node, &all_q_list);
2371 blk_mq_add_queue_tag_set(set, q);
2372 blk_mq_map_swqueue(q, cpu_online_mask);
2374 mutex_unlock(&all_q_mutex);
2377 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2380 ret = blk_mq_sched_init(q);
2382 return ERR_PTR(ret);
2388 kfree(q->queue_hw_ctx);
2390 free_percpu(q->queue_ctx);
2393 return ERR_PTR(-ENOMEM);
2395 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2397 void blk_mq_free_queue(struct request_queue *q)
2399 struct blk_mq_tag_set *set = q->tag_set;
2401 mutex_lock(&all_q_mutex);
2402 list_del_init(&q->all_q_node);
2403 mutex_unlock(&all_q_mutex);
2405 blk_mq_del_queue_tag_set(q);
2407 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2410 /* Basically redo blk_mq_init_queue with queue frozen */
2411 static void blk_mq_queue_reinit(struct request_queue *q,
2412 const struct cpumask *online_mask)
2414 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2416 blk_mq_debugfs_unregister_hctxs(q);
2417 blk_mq_sysfs_unregister(q);
2420 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2421 * we should change hctx numa_node according to new topology (this
2422 * involves free and re-allocate memory, worthy doing?)
2425 blk_mq_map_swqueue(q, online_mask);
2427 blk_mq_sysfs_register(q);
2428 blk_mq_debugfs_register_hctxs(q);
2432 * New online cpumask which is going to be set in this hotplug event.
2433 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2434 * one-by-one and dynamically allocating this could result in a failure.
2436 static struct cpumask cpuhp_online_new;
2438 static void blk_mq_queue_reinit_work(void)
2440 struct request_queue *q;
2442 mutex_lock(&all_q_mutex);
2444 * We need to freeze and reinit all existing queues. Freezing
2445 * involves synchronous wait for an RCU grace period and doing it
2446 * one by one may take a long time. Start freezing all queues in
2447 * one swoop and then wait for the completions so that freezing can
2448 * take place in parallel.
2450 list_for_each_entry(q, &all_q_list, all_q_node)
2451 blk_freeze_queue_start(q);
2452 list_for_each_entry(q, &all_q_list, all_q_node)
2453 blk_mq_freeze_queue_wait(q);
2455 list_for_each_entry(q, &all_q_list, all_q_node)
2456 blk_mq_queue_reinit(q, &cpuhp_online_new);
2458 list_for_each_entry(q, &all_q_list, all_q_node)
2459 blk_mq_unfreeze_queue(q);
2461 mutex_unlock(&all_q_mutex);
2464 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2466 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2467 blk_mq_queue_reinit_work();
2472 * Before hotadded cpu starts handling requests, new mappings must be
2473 * established. Otherwise, these requests in hw queue might never be
2476 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2477 * for CPU0, and ctx1 for CPU1).
2479 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2480 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2482 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2483 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2484 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2487 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2489 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2490 cpumask_set_cpu(cpu, &cpuhp_online_new);
2491 blk_mq_queue_reinit_work();
2495 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2499 for (i = 0; i < set->nr_hw_queues; i++)
2500 if (!__blk_mq_alloc_rq_map(set, i))
2507 blk_mq_free_rq_map(set->tags[i]);
2513 * Allocate the request maps associated with this tag_set. Note that this
2514 * may reduce the depth asked for, if memory is tight. set->queue_depth
2515 * will be updated to reflect the allocated depth.
2517 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2522 depth = set->queue_depth;
2524 err = __blk_mq_alloc_rq_maps(set);
2528 set->queue_depth >>= 1;
2529 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2533 } while (set->queue_depth);
2535 if (!set->queue_depth || err) {
2536 pr_err("blk-mq: failed to allocate request map\n");
2540 if (depth != set->queue_depth)
2541 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2542 depth, set->queue_depth);
2547 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2549 if (set->ops->map_queues)
2550 return set->ops->map_queues(set);
2552 return blk_mq_map_queues(set);
2556 * Alloc a tag set to be associated with one or more request queues.
2557 * May fail with EINVAL for various error conditions. May adjust the
2558 * requested depth down, if if it too large. In that case, the set
2559 * value will be stored in set->queue_depth.
2561 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2565 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2567 if (!set->nr_hw_queues)
2569 if (!set->queue_depth)
2571 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2574 if (!set->ops->queue_rq)
2577 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2578 pr_info("blk-mq: reduced tag depth to %u\n",
2580 set->queue_depth = BLK_MQ_MAX_DEPTH;
2584 * If a crashdump is active, then we are potentially in a very
2585 * memory constrained environment. Limit us to 1 queue and
2586 * 64 tags to prevent using too much memory.
2588 if (is_kdump_kernel()) {
2589 set->nr_hw_queues = 1;
2590 set->queue_depth = min(64U, set->queue_depth);
2593 * There is no use for more h/w queues than cpus.
2595 if (set->nr_hw_queues > nr_cpu_ids)
2596 set->nr_hw_queues = nr_cpu_ids;
2598 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2599 GFP_KERNEL, set->numa_node);
2604 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2605 GFP_KERNEL, set->numa_node);
2609 ret = blk_mq_update_queue_map(set);
2611 goto out_free_mq_map;
2613 ret = blk_mq_alloc_rq_maps(set);
2615 goto out_free_mq_map;
2617 mutex_init(&set->tag_list_lock);
2618 INIT_LIST_HEAD(&set->tag_list);
2630 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2632 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2636 for (i = 0; i < nr_cpu_ids; i++)
2637 blk_mq_free_map_and_requests(set, i);
2645 EXPORT_SYMBOL(blk_mq_free_tag_set);
2647 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2649 struct blk_mq_tag_set *set = q->tag_set;
2650 struct blk_mq_hw_ctx *hctx;
2656 blk_mq_freeze_queue(q);
2659 queue_for_each_hw_ctx(q, hctx, i) {
2663 * If we're using an MQ scheduler, just update the scheduler
2664 * queue depth. This is similar to what the old code would do.
2666 if (!hctx->sched_tags) {
2667 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2668 min(nr, set->queue_depth),
2671 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2679 q->nr_requests = nr;
2681 blk_mq_unfreeze_queue(q);
2686 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2689 struct request_queue *q;
2691 lockdep_assert_held(&set->tag_list_lock);
2693 if (nr_hw_queues > nr_cpu_ids)
2694 nr_hw_queues = nr_cpu_ids;
2695 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2698 list_for_each_entry(q, &set->tag_list, tag_set_list)
2699 blk_mq_freeze_queue(q);
2701 set->nr_hw_queues = nr_hw_queues;
2702 blk_mq_update_queue_map(set);
2703 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2704 blk_mq_realloc_hw_ctxs(set, q);
2705 blk_mq_queue_reinit(q, cpu_online_mask);
2708 list_for_each_entry(q, &set->tag_list, tag_set_list)
2709 blk_mq_unfreeze_queue(q);
2712 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2714 mutex_lock(&set->tag_list_lock);
2715 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2716 mutex_unlock(&set->tag_list_lock);
2718 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2720 /* Enable polling stats and return whether they were already enabled. */
2721 static bool blk_poll_stats_enable(struct request_queue *q)
2723 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2724 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2726 blk_stat_add_callback(q, q->poll_cb);
2730 static void blk_mq_poll_stats_start(struct request_queue *q)
2733 * We don't arm the callback if polling stats are not enabled or the
2734 * callback is already active.
2736 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2737 blk_stat_is_active(q->poll_cb))
2740 blk_stat_activate_msecs(q->poll_cb, 100);
2743 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2745 struct request_queue *q = cb->data;
2748 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2749 if (cb->stat[bucket].nr_samples)
2750 q->poll_stat[bucket] = cb->stat[bucket];
2754 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2755 struct blk_mq_hw_ctx *hctx,
2758 unsigned long ret = 0;
2762 * If stats collection isn't on, don't sleep but turn it on for
2765 if (!blk_poll_stats_enable(q))
2769 * As an optimistic guess, use half of the mean service time
2770 * for this type of request. We can (and should) make this smarter.
2771 * For instance, if the completion latencies are tight, we can
2772 * get closer than just half the mean. This is especially
2773 * important on devices where the completion latencies are longer
2774 * than ~10 usec. We do use the stats for the relevant IO size
2775 * if available which does lead to better estimates.
2777 bucket = blk_mq_poll_stats_bkt(rq);
2781 if (q->poll_stat[bucket].nr_samples)
2782 ret = (q->poll_stat[bucket].mean + 1) / 2;
2787 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2788 struct blk_mq_hw_ctx *hctx,
2791 struct hrtimer_sleeper hs;
2792 enum hrtimer_mode mode;
2796 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2802 * -1: don't ever hybrid sleep
2803 * 0: use half of prev avg
2804 * >0: use this specific value
2806 if (q->poll_nsec == -1)
2808 else if (q->poll_nsec > 0)
2809 nsecs = q->poll_nsec;
2811 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2816 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2819 * This will be replaced with the stats tracking code, using
2820 * 'avg_completion_time / 2' as the pre-sleep target.
2824 mode = HRTIMER_MODE_REL;
2825 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2826 hrtimer_set_expires(&hs.timer, kt);
2828 hrtimer_init_sleeper(&hs, current);
2830 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2832 set_current_state(TASK_UNINTERRUPTIBLE);
2833 hrtimer_start_expires(&hs.timer, mode);
2836 hrtimer_cancel(&hs.timer);
2837 mode = HRTIMER_MODE_ABS;
2838 } while (hs.task && !signal_pending(current));
2840 __set_current_state(TASK_RUNNING);
2841 destroy_hrtimer_on_stack(&hs.timer);
2845 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2847 struct request_queue *q = hctx->queue;
2851 * If we sleep, have the caller restart the poll loop to reset
2852 * the state. Like for the other success return cases, the
2853 * caller is responsible for checking if the IO completed. If
2854 * the IO isn't complete, we'll get called again and will go
2855 * straight to the busy poll loop.
2857 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2860 hctx->poll_considered++;
2862 state = current->state;
2863 while (!need_resched()) {
2866 hctx->poll_invoked++;
2868 ret = q->mq_ops->poll(hctx, rq->tag);
2870 hctx->poll_success++;
2871 set_current_state(TASK_RUNNING);
2875 if (signal_pending_state(state, current))
2876 set_current_state(TASK_RUNNING);
2878 if (current->state == TASK_RUNNING)
2888 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2890 struct blk_mq_hw_ctx *hctx;
2891 struct blk_plug *plug;
2894 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2895 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2898 plug = current->plug;
2900 blk_flush_plug_list(plug, false);
2902 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2903 if (!blk_qc_t_is_internal(cookie))
2904 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2906 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2908 * With scheduling, if the request has completed, we'll
2909 * get a NULL return here, as we clear the sched tag when
2910 * that happens. The request still remains valid, like always,
2911 * so we should be safe with just the NULL check.
2917 return __blk_mq_poll(hctx, rq);
2919 EXPORT_SYMBOL_GPL(blk_mq_poll);
2921 void blk_mq_disable_hotplug(void)
2923 mutex_lock(&all_q_mutex);
2926 void blk_mq_enable_hotplug(void)
2928 mutex_unlock(&all_q_mutex);
2931 static int __init blk_mq_init(void)
2933 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2934 blk_mq_hctx_notify_dead);
2936 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2937 blk_mq_queue_reinit_prepare,
2938 blk_mq_queue_reinit_dead);
2941 subsys_initcall(blk_mq_init);