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];
233 if (data->flags & BLK_MQ_REQ_INTERNAL) {
235 rq->internal_tag = tag;
237 if (blk_mq_tag_busy(data->hctx)) {
238 rq->rq_flags = RQF_MQ_INFLIGHT;
239 atomic_inc(&data->hctx->nr_active);
242 rq->internal_tag = -1;
243 data->hctx->tags->rqs[rq->tag] = rq;
246 INIT_LIST_HEAD(&rq->queuelist);
247 /* csd/requeue_work/fifo_time is initialized before use */
249 rq->mq_ctx = data->ctx;
251 if (blk_queue_io_stat(data->q))
252 rq->rq_flags |= RQF_IO_STAT;
253 /* do not touch atomic flags, it needs atomic ops against the timer */
255 INIT_HLIST_NODE(&rq->hash);
256 RB_CLEAR_NODE(&rq->rb_node);
259 rq->start_time = jiffies;
260 #ifdef CONFIG_BLK_CGROUP
262 set_start_time_ns(rq);
263 rq->io_start_time_ns = 0;
265 rq->nr_phys_segments = 0;
266 #if defined(CONFIG_BLK_DEV_INTEGRITY)
267 rq->nr_integrity_segments = 0;
270 /* tag was already set */
273 INIT_LIST_HEAD(&rq->timeout_list);
277 rq->end_io_data = NULL;
280 data->ctx->rq_dispatched[op_is_sync(op)]++;
284 static struct request *blk_mq_get_request(struct request_queue *q,
285 struct bio *bio, unsigned int op,
286 struct blk_mq_alloc_data *data)
288 struct elevator_queue *e = q->elevator;
292 blk_queue_enter_live(q);
294 if (likely(!data->ctx))
295 data->ctx = blk_mq_get_ctx(q);
296 if (likely(!data->hctx))
297 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
299 data->flags |= BLK_MQ_REQ_NOWAIT;
302 data->flags |= BLK_MQ_REQ_INTERNAL;
305 * Flush requests are special and go directly to the
308 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
309 e->type->ops.mq.limit_depth(op, data);
312 tag = blk_mq_get_tag(data);
313 if (tag == BLK_MQ_TAG_FAIL) {
318 rq = blk_mq_rq_ctx_init(data, tag, op);
319 if (!op_is_flush(op)) {
321 if (e && e->type->ops.mq.prepare_request) {
322 if (e->type->icq_cache && rq_ioc(bio))
323 blk_mq_sched_assign_ioc(rq, bio);
325 e->type->ops.mq.prepare_request(rq, bio);
326 rq->rq_flags |= RQF_ELVPRIV;
329 data->hctx->queued++;
333 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
336 struct blk_mq_alloc_data alloc_data = { .flags = flags };
340 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
344 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
346 blk_mq_put_ctx(alloc_data.ctx);
350 return ERR_PTR(-EWOULDBLOCK);
353 rq->__sector = (sector_t) -1;
354 rq->bio = rq->biotail = NULL;
357 EXPORT_SYMBOL(blk_mq_alloc_request);
359 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
360 unsigned int op, unsigned int flags, unsigned int hctx_idx)
362 struct blk_mq_alloc_data alloc_data = { .flags = flags };
368 * If the tag allocator sleeps we could get an allocation for a
369 * different hardware context. No need to complicate the low level
370 * allocator for this for the rare use case of a command tied to
373 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
374 return ERR_PTR(-EINVAL);
376 if (hctx_idx >= q->nr_hw_queues)
377 return ERR_PTR(-EIO);
379 ret = blk_queue_enter(q, true);
384 * Check if the hardware context is actually mapped to anything.
385 * If not tell the caller that it should skip this queue.
387 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
388 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
390 return ERR_PTR(-EXDEV);
392 cpu = cpumask_first(alloc_data.hctx->cpumask);
393 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
395 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
400 return ERR_PTR(-EWOULDBLOCK);
404 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
406 void blk_mq_free_request(struct request *rq)
408 struct request_queue *q = rq->q;
409 struct elevator_queue *e = q->elevator;
410 struct blk_mq_ctx *ctx = rq->mq_ctx;
411 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
412 const int sched_tag = rq->internal_tag;
414 if (rq->rq_flags & RQF_ELVPRIV) {
415 if (e && e->type->ops.mq.finish_request)
416 e->type->ops.mq.finish_request(rq);
418 put_io_context(rq->elv.icq->ioc);
423 ctx->rq_completed[rq_is_sync(rq)]++;
424 if (rq->rq_flags & RQF_MQ_INFLIGHT)
425 atomic_dec(&hctx->nr_active);
427 wbt_done(q->rq_wb, &rq->issue_stat);
429 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
430 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
432 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
434 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
435 blk_mq_sched_restart(hctx);
438 EXPORT_SYMBOL_GPL(blk_mq_free_request);
440 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
442 blk_account_io_done(rq);
445 wbt_done(rq->q->rq_wb, &rq->issue_stat);
446 rq->end_io(rq, error);
448 if (unlikely(blk_bidi_rq(rq)))
449 blk_mq_free_request(rq->next_rq);
450 blk_mq_free_request(rq);
453 EXPORT_SYMBOL(__blk_mq_end_request);
455 void blk_mq_end_request(struct request *rq, blk_status_t error)
457 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
459 __blk_mq_end_request(rq, error);
461 EXPORT_SYMBOL(blk_mq_end_request);
463 static void __blk_mq_complete_request_remote(void *data)
465 struct request *rq = data;
467 rq->q->softirq_done_fn(rq);
470 static void __blk_mq_complete_request(struct request *rq)
472 struct blk_mq_ctx *ctx = rq->mq_ctx;
476 if (rq->internal_tag != -1)
477 blk_mq_sched_completed_request(rq);
478 if (rq->rq_flags & RQF_STATS) {
479 blk_mq_poll_stats_start(rq->q);
483 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
484 rq->q->softirq_done_fn(rq);
489 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
490 shared = cpus_share_cache(cpu, ctx->cpu);
492 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
493 rq->csd.func = __blk_mq_complete_request_remote;
496 smp_call_function_single_async(ctx->cpu, &rq->csd);
498 rq->q->softirq_done_fn(rq);
504 * blk_mq_complete_request - end I/O on a request
505 * @rq: the request being processed
508 * Ends all I/O on a request. It does not handle partial completions.
509 * The actual completion happens out-of-order, through a IPI handler.
511 void blk_mq_complete_request(struct request *rq)
513 struct request_queue *q = rq->q;
515 if (unlikely(blk_should_fake_timeout(q)))
517 if (!blk_mark_rq_complete(rq))
518 __blk_mq_complete_request(rq);
520 EXPORT_SYMBOL(blk_mq_complete_request);
522 int blk_mq_request_started(struct request *rq)
524 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
526 EXPORT_SYMBOL_GPL(blk_mq_request_started);
528 void blk_mq_start_request(struct request *rq)
530 struct request_queue *q = rq->q;
532 blk_mq_sched_started_request(rq);
534 trace_block_rq_issue(q, rq);
536 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
537 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
538 rq->rq_flags |= RQF_STATS;
539 wbt_issue(q->rq_wb, &rq->issue_stat);
545 * Ensure that ->deadline is visible before set the started
546 * flag and clear the completed flag.
548 smp_mb__before_atomic();
551 * Mark us as started and clear complete. Complete might have been
552 * set if requeue raced with timeout, which then marked it as
553 * complete. So be sure to clear complete again when we start
554 * the request, otherwise we'll ignore the completion event.
556 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
557 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
558 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
559 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
561 if (q->dma_drain_size && blk_rq_bytes(rq)) {
563 * Make sure space for the drain appears. We know we can do
564 * this because max_hw_segments has been adjusted to be one
565 * fewer than the device can handle.
567 rq->nr_phys_segments++;
570 EXPORT_SYMBOL(blk_mq_start_request);
573 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
574 * flag isn't set yet, so there may be race with timeout handler,
575 * but given rq->deadline is just set in .queue_rq() under
576 * this situation, the race won't be possible in reality because
577 * rq->timeout should be set as big enough to cover the window
578 * between blk_mq_start_request() called from .queue_rq() and
579 * clearing REQ_ATOM_STARTED here.
581 static void __blk_mq_requeue_request(struct request *rq)
583 struct request_queue *q = rq->q;
585 trace_block_rq_requeue(q, rq);
586 wbt_requeue(q->rq_wb, &rq->issue_stat);
587 blk_mq_sched_requeue_request(rq);
589 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
590 if (q->dma_drain_size && blk_rq_bytes(rq))
591 rq->nr_phys_segments--;
595 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
597 __blk_mq_requeue_request(rq);
599 BUG_ON(blk_queued_rq(rq));
600 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
602 EXPORT_SYMBOL(blk_mq_requeue_request);
604 static void blk_mq_requeue_work(struct work_struct *work)
606 struct request_queue *q =
607 container_of(work, struct request_queue, requeue_work.work);
609 struct request *rq, *next;
612 spin_lock_irqsave(&q->requeue_lock, flags);
613 list_splice_init(&q->requeue_list, &rq_list);
614 spin_unlock_irqrestore(&q->requeue_lock, flags);
616 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
617 if (!(rq->rq_flags & RQF_SOFTBARRIER))
620 rq->rq_flags &= ~RQF_SOFTBARRIER;
621 list_del_init(&rq->queuelist);
622 blk_mq_sched_insert_request(rq, true, false, false, true);
625 while (!list_empty(&rq_list)) {
626 rq = list_entry(rq_list.next, struct request, queuelist);
627 list_del_init(&rq->queuelist);
628 blk_mq_sched_insert_request(rq, false, false, false, true);
631 blk_mq_run_hw_queues(q, false);
634 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
635 bool kick_requeue_list)
637 struct request_queue *q = rq->q;
641 * We abuse this flag that is otherwise used by the I/O scheduler to
642 * request head insertation from the workqueue.
644 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
646 spin_lock_irqsave(&q->requeue_lock, flags);
648 rq->rq_flags |= RQF_SOFTBARRIER;
649 list_add(&rq->queuelist, &q->requeue_list);
651 list_add_tail(&rq->queuelist, &q->requeue_list);
653 spin_unlock_irqrestore(&q->requeue_lock, flags);
655 if (kick_requeue_list)
656 blk_mq_kick_requeue_list(q);
658 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
660 void blk_mq_kick_requeue_list(struct request_queue *q)
662 kblockd_schedule_delayed_work(&q->requeue_work, 0);
664 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
666 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
669 kblockd_schedule_delayed_work(&q->requeue_work,
670 msecs_to_jiffies(msecs));
672 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
674 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
676 if (tag < tags->nr_tags) {
677 prefetch(tags->rqs[tag]);
678 return tags->rqs[tag];
683 EXPORT_SYMBOL(blk_mq_tag_to_rq);
685 struct blk_mq_timeout_data {
687 unsigned int next_set;
690 void blk_mq_rq_timed_out(struct request *req, bool reserved)
692 const struct blk_mq_ops *ops = req->q->mq_ops;
693 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
696 * We know that complete is set at this point. If STARTED isn't set
697 * anymore, then the request isn't active and the "timeout" should
698 * just be ignored. This can happen due to the bitflag ordering.
699 * Timeout first checks if STARTED is set, and if it is, assumes
700 * the request is active. But if we race with completion, then
701 * both flags will get cleared. So check here again, and ignore
702 * a timeout event with a request that isn't active.
704 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
708 ret = ops->timeout(req, reserved);
712 __blk_mq_complete_request(req);
714 case BLK_EH_RESET_TIMER:
716 blk_clear_rq_complete(req);
718 case BLK_EH_NOT_HANDLED:
721 printk(KERN_ERR "block: bad eh return: %d\n", ret);
726 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
727 struct request *rq, void *priv, bool reserved)
729 struct blk_mq_timeout_data *data = priv;
731 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
735 * The rq being checked may have been freed and reallocated
736 * out already here, we avoid this race by checking rq->deadline
737 * and REQ_ATOM_COMPLETE flag together:
739 * - if rq->deadline is observed as new value because of
740 * reusing, the rq won't be timed out because of timing.
741 * - if rq->deadline is observed as previous value,
742 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
743 * because we put a barrier between setting rq->deadline
744 * and clearing the flag in blk_mq_start_request(), so
745 * this rq won't be timed out too.
747 if (time_after_eq(jiffies, rq->deadline)) {
748 if (!blk_mark_rq_complete(rq))
749 blk_mq_rq_timed_out(rq, reserved);
750 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
751 data->next = rq->deadline;
756 static void blk_mq_timeout_work(struct work_struct *work)
758 struct request_queue *q =
759 container_of(work, struct request_queue, timeout_work);
760 struct blk_mq_timeout_data data = {
766 /* A deadlock might occur if a request is stuck requiring a
767 * timeout at the same time a queue freeze is waiting
768 * completion, since the timeout code would not be able to
769 * acquire the queue reference here.
771 * That's why we don't use blk_queue_enter here; instead, we use
772 * percpu_ref_tryget directly, because we need to be able to
773 * obtain a reference even in the short window between the queue
774 * starting to freeze, by dropping the first reference in
775 * blk_freeze_queue_start, and the moment the last request is
776 * consumed, marked by the instant q_usage_counter reaches
779 if (!percpu_ref_tryget(&q->q_usage_counter))
782 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
785 data.next = blk_rq_timeout(round_jiffies_up(data.next));
786 mod_timer(&q->timeout, data.next);
788 struct blk_mq_hw_ctx *hctx;
790 queue_for_each_hw_ctx(q, hctx, i) {
791 /* the hctx may be unmapped, so check it here */
792 if (blk_mq_hw_queue_mapped(hctx))
793 blk_mq_tag_idle(hctx);
799 struct flush_busy_ctx_data {
800 struct blk_mq_hw_ctx *hctx;
801 struct list_head *list;
804 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
806 struct flush_busy_ctx_data *flush_data = data;
807 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
808 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
810 sbitmap_clear_bit(sb, bitnr);
811 spin_lock(&ctx->lock);
812 list_splice_tail_init(&ctx->rq_list, flush_data->list);
813 spin_unlock(&ctx->lock);
818 * Process software queues that have been marked busy, splicing them
819 * to the for-dispatch
821 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
823 struct flush_busy_ctx_data data = {
828 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
830 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
832 static inline unsigned int queued_to_index(unsigned int queued)
837 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
840 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
843 struct blk_mq_alloc_data data = {
845 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
846 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
849 might_sleep_if(wait);
854 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
855 data.flags |= BLK_MQ_REQ_RESERVED;
857 rq->tag = blk_mq_get_tag(&data);
859 if (blk_mq_tag_busy(data.hctx)) {
860 rq->rq_flags |= RQF_MQ_INFLIGHT;
861 atomic_inc(&data.hctx->nr_active);
863 data.hctx->tags->rqs[rq->tag] = rq;
869 return rq->tag != -1;
872 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
875 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
878 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
879 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
880 atomic_dec(&hctx->nr_active);
884 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
887 if (rq->tag == -1 || rq->internal_tag == -1)
890 __blk_mq_put_driver_tag(hctx, rq);
893 static void blk_mq_put_driver_tag(struct request *rq)
895 struct blk_mq_hw_ctx *hctx;
897 if (rq->tag == -1 || rq->internal_tag == -1)
900 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
901 __blk_mq_put_driver_tag(hctx, rq);
905 * If we fail getting a driver tag because all the driver tags are already
906 * assigned and on the dispatch list, BUT the first entry does not have a
907 * tag, then we could deadlock. For that case, move entries with assigned
908 * driver tags to the front, leaving the set of tagged requests in the
909 * same order, and the untagged set in the same order.
911 static bool reorder_tags_to_front(struct list_head *list)
913 struct request *rq, *tmp, *first = NULL;
915 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
919 list_move(&rq->queuelist, list);
925 return first != NULL;
928 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
931 struct blk_mq_hw_ctx *hctx;
933 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
935 list_del(&wait->task_list);
936 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
937 blk_mq_run_hw_queue(hctx, true);
941 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
943 struct sbq_wait_state *ws;
946 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
947 * The thread which wins the race to grab this bit adds the hardware
948 * queue to the wait queue.
950 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
951 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
954 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
955 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
958 * As soon as this returns, it's no longer safe to fiddle with
959 * hctx->dispatch_wait, since a completion can wake up the wait queue
960 * and unlock the bit.
962 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
966 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
968 struct blk_mq_hw_ctx *hctx;
972 if (list_empty(list))
976 * Now process all the entries, sending them to the driver.
980 struct blk_mq_queue_data bd;
983 rq = list_first_entry(list, struct request, queuelist);
984 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
985 if (!queued && reorder_tags_to_front(list))
989 * The initial allocation attempt failed, so we need to
990 * rerun the hardware queue when a tag is freed.
992 if (!blk_mq_dispatch_wait_add(hctx))
996 * It's possible that a tag was freed in the window
997 * between the allocation failure and adding the
998 * hardware queue to the wait queue.
1000 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1004 list_del_init(&rq->queuelist);
1009 * Flag last if we have no more requests, or if we have more
1010 * but can't assign a driver tag to it.
1012 if (list_empty(list))
1015 struct request *nxt;
1017 nxt = list_first_entry(list, struct request, queuelist);
1018 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1021 ret = q->mq_ops->queue_rq(hctx, &bd);
1022 if (ret == BLK_STS_RESOURCE) {
1023 blk_mq_put_driver_tag_hctx(hctx, rq);
1024 list_add(&rq->queuelist, list);
1025 __blk_mq_requeue_request(rq);
1029 if (unlikely(ret != BLK_STS_OK)) {
1031 blk_mq_end_request(rq, BLK_STS_IOERR);
1036 } while (!list_empty(list));
1038 hctx->dispatched[queued_to_index(queued)]++;
1041 * Any items that need requeuing? Stuff them into hctx->dispatch,
1042 * that is where we will continue on next queue run.
1044 if (!list_empty(list)) {
1046 * If an I/O scheduler has been configured and we got a driver
1047 * tag for the next request already, free it again.
1049 rq = list_first_entry(list, struct request, queuelist);
1050 blk_mq_put_driver_tag(rq);
1052 spin_lock(&hctx->lock);
1053 list_splice_init(list, &hctx->dispatch);
1054 spin_unlock(&hctx->lock);
1057 * If SCHED_RESTART was set by the caller of this function and
1058 * it is no longer set that means that it was cleared by another
1059 * thread and hence that a queue rerun is needed.
1061 * If TAG_WAITING is set that means that an I/O scheduler has
1062 * been configured and another thread is waiting for a driver
1063 * tag. To guarantee fairness, do not rerun this hardware queue
1064 * but let the other thread grab the driver tag.
1066 * If no I/O scheduler has been configured it is possible that
1067 * the hardware queue got stopped and restarted before requests
1068 * were pushed back onto the dispatch list. Rerun the queue to
1069 * avoid starvation. Notes:
1070 * - blk_mq_run_hw_queue() checks whether or not a queue has
1071 * been stopped before rerunning a queue.
1072 * - Some but not all block drivers stop a queue before
1073 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1076 if (!blk_mq_sched_needs_restart(hctx) &&
1077 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1078 blk_mq_run_hw_queue(hctx, true);
1081 return (queued + errors) != 0;
1084 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1088 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1089 cpu_online(hctx->next_cpu));
1091 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1093 blk_mq_sched_dispatch_requests(hctx);
1098 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1099 blk_mq_sched_dispatch_requests(hctx);
1100 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1105 * It'd be great if the workqueue API had a way to pass
1106 * in a mask and had some smarts for more clever placement.
1107 * For now we just round-robin here, switching for every
1108 * BLK_MQ_CPU_WORK_BATCH queued items.
1110 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1112 if (hctx->queue->nr_hw_queues == 1)
1113 return WORK_CPU_UNBOUND;
1115 if (--hctx->next_cpu_batch <= 0) {
1118 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1119 if (next_cpu >= nr_cpu_ids)
1120 next_cpu = cpumask_first(hctx->cpumask);
1122 hctx->next_cpu = next_cpu;
1123 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1126 return hctx->next_cpu;
1129 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1130 unsigned long msecs)
1132 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1135 if (unlikely(blk_mq_hctx_stopped(hctx)))
1138 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1139 int cpu = get_cpu();
1140 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1141 __blk_mq_run_hw_queue(hctx);
1149 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1151 msecs_to_jiffies(msecs));
1154 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1156 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1158 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1160 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1162 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1164 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1166 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1168 struct blk_mq_hw_ctx *hctx;
1171 queue_for_each_hw_ctx(q, hctx, i) {
1172 if (!blk_mq_hctx_has_pending(hctx) ||
1173 blk_mq_hctx_stopped(hctx))
1176 blk_mq_run_hw_queue(hctx, async);
1179 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1182 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1183 * @q: request queue.
1185 * The caller is responsible for serializing this function against
1186 * blk_mq_{start,stop}_hw_queue().
1188 bool blk_mq_queue_stopped(struct request_queue *q)
1190 struct blk_mq_hw_ctx *hctx;
1193 queue_for_each_hw_ctx(q, hctx, i)
1194 if (blk_mq_hctx_stopped(hctx))
1199 EXPORT_SYMBOL(blk_mq_queue_stopped);
1202 * This function is often used for pausing .queue_rq() by driver when
1203 * there isn't enough resource or some conditions aren't satisfied, and
1204 * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1206 * We do not guarantee that dispatch can be drained or blocked
1207 * after blk_mq_stop_hw_queue() returns. Please use
1208 * blk_mq_quiesce_queue() for that requirement.
1210 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1212 cancel_delayed_work(&hctx->run_work);
1214 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1216 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1219 * This function is often used for pausing .queue_rq() by driver when
1220 * there isn't enough resource or some conditions aren't satisfied, and
1221 * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1223 * We do not guarantee that dispatch can be drained or blocked
1224 * after blk_mq_stop_hw_queues() returns. Please use
1225 * blk_mq_quiesce_queue() for that requirement.
1227 void blk_mq_stop_hw_queues(struct request_queue *q)
1229 struct blk_mq_hw_ctx *hctx;
1232 queue_for_each_hw_ctx(q, hctx, i)
1233 blk_mq_stop_hw_queue(hctx);
1235 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1237 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1239 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1241 blk_mq_run_hw_queue(hctx, false);
1243 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1245 void blk_mq_start_hw_queues(struct request_queue *q)
1247 struct blk_mq_hw_ctx *hctx;
1250 queue_for_each_hw_ctx(q, hctx, i)
1251 blk_mq_start_hw_queue(hctx);
1253 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1255 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1257 if (!blk_mq_hctx_stopped(hctx))
1260 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1261 blk_mq_run_hw_queue(hctx, async);
1263 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1265 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1267 struct blk_mq_hw_ctx *hctx;
1270 queue_for_each_hw_ctx(q, hctx, i)
1271 blk_mq_start_stopped_hw_queue(hctx, async);
1273 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1275 static void blk_mq_run_work_fn(struct work_struct *work)
1277 struct blk_mq_hw_ctx *hctx;
1279 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1282 * If we are stopped, don't run the queue. The exception is if
1283 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1284 * the STOPPED bit and run it.
1286 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1287 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1290 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1291 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1294 __blk_mq_run_hw_queue(hctx);
1298 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1300 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1304 * Stop the hw queue, then modify currently delayed work.
1305 * This should prevent us from running the queue prematurely.
1306 * Mark the queue as auto-clearing STOPPED when it runs.
1308 blk_mq_stop_hw_queue(hctx);
1309 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1310 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1312 msecs_to_jiffies(msecs));
1314 EXPORT_SYMBOL(blk_mq_delay_queue);
1316 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1320 struct blk_mq_ctx *ctx = rq->mq_ctx;
1322 lockdep_assert_held(&ctx->lock);
1324 trace_block_rq_insert(hctx->queue, rq);
1327 list_add(&rq->queuelist, &ctx->rq_list);
1329 list_add_tail(&rq->queuelist, &ctx->rq_list);
1332 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1335 struct blk_mq_ctx *ctx = rq->mq_ctx;
1337 lockdep_assert_held(&ctx->lock);
1339 __blk_mq_insert_req_list(hctx, rq, at_head);
1340 blk_mq_hctx_mark_pending(hctx, ctx);
1343 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1344 struct list_head *list)
1348 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1351 spin_lock(&ctx->lock);
1352 while (!list_empty(list)) {
1355 rq = list_first_entry(list, struct request, queuelist);
1356 BUG_ON(rq->mq_ctx != ctx);
1357 list_del_init(&rq->queuelist);
1358 __blk_mq_insert_req_list(hctx, rq, false);
1360 blk_mq_hctx_mark_pending(hctx, ctx);
1361 spin_unlock(&ctx->lock);
1364 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1366 struct request *rqa = container_of(a, struct request, queuelist);
1367 struct request *rqb = container_of(b, struct request, queuelist);
1369 return !(rqa->mq_ctx < rqb->mq_ctx ||
1370 (rqa->mq_ctx == rqb->mq_ctx &&
1371 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1374 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1376 struct blk_mq_ctx *this_ctx;
1377 struct request_queue *this_q;
1380 LIST_HEAD(ctx_list);
1383 list_splice_init(&plug->mq_list, &list);
1385 list_sort(NULL, &list, plug_ctx_cmp);
1391 while (!list_empty(&list)) {
1392 rq = list_entry_rq(list.next);
1393 list_del_init(&rq->queuelist);
1395 if (rq->mq_ctx != this_ctx) {
1397 trace_block_unplug(this_q, depth, from_schedule);
1398 blk_mq_sched_insert_requests(this_q, this_ctx,
1403 this_ctx = rq->mq_ctx;
1409 list_add_tail(&rq->queuelist, &ctx_list);
1413 * If 'this_ctx' is set, we know we have entries to complete
1414 * on 'ctx_list'. Do those.
1417 trace_block_unplug(this_q, depth, from_schedule);
1418 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1423 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1425 blk_init_request_from_bio(rq, bio);
1427 blk_account_io_start(rq, true);
1430 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1432 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1433 !blk_queue_nomerges(hctx->queue);
1436 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1437 struct blk_mq_ctx *ctx,
1440 spin_lock(&ctx->lock);
1441 __blk_mq_insert_request(hctx, rq, false);
1442 spin_unlock(&ctx->lock);
1445 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1448 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1450 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1453 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1455 blk_qc_t *cookie, bool may_sleep)
1457 struct request_queue *q = rq->q;
1458 struct blk_mq_queue_data bd = {
1462 blk_qc_t new_cookie;
1464 bool run_queue = true;
1466 /* RCU or SRCU read lock is needed before checking quiesced flag */
1467 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1475 if (!blk_mq_get_driver_tag(rq, NULL, false))
1478 new_cookie = request_to_qc_t(hctx, rq);
1481 * For OK queue, we are done. For error, kill it. Any other
1482 * error (busy), just add it to our list as we previously
1485 ret = q->mq_ops->queue_rq(hctx, &bd);
1488 *cookie = new_cookie;
1490 case BLK_STS_RESOURCE:
1491 __blk_mq_requeue_request(rq);
1494 *cookie = BLK_QC_T_NONE;
1495 blk_mq_end_request(rq, ret);
1500 blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1503 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1504 struct request *rq, blk_qc_t *cookie)
1506 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1508 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1511 unsigned int srcu_idx;
1515 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1516 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1517 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1521 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1523 const int is_sync = op_is_sync(bio->bi_opf);
1524 const int is_flush_fua = op_is_flush(bio->bi_opf);
1525 struct blk_mq_alloc_data data = { .flags = 0 };
1527 unsigned int request_count = 0;
1528 struct blk_plug *plug;
1529 struct request *same_queue_rq = NULL;
1531 unsigned int wb_acct;
1533 blk_queue_bounce(q, &bio);
1535 blk_queue_split(q, &bio);
1537 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1539 return BLK_QC_T_NONE;
1542 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1543 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1544 return BLK_QC_T_NONE;
1546 if (blk_mq_sched_bio_merge(q, bio))
1547 return BLK_QC_T_NONE;
1549 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1551 trace_block_getrq(q, bio, bio->bi_opf);
1553 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1554 if (unlikely(!rq)) {
1555 __wbt_done(q->rq_wb, wb_acct);
1556 if (bio->bi_opf & REQ_NOWAIT)
1557 bio_wouldblock_error(bio);
1558 return BLK_QC_T_NONE;
1561 wbt_track(&rq->issue_stat, wb_acct);
1563 cookie = request_to_qc_t(data.hctx, rq);
1565 plug = current->plug;
1566 if (unlikely(is_flush_fua)) {
1567 blk_mq_put_ctx(data.ctx);
1568 blk_mq_bio_to_request(rq, bio);
1570 blk_mq_sched_insert_request(rq, false, true, true,
1573 blk_insert_flush(rq);
1574 blk_mq_run_hw_queue(data.hctx, true);
1576 } else if (plug && q->nr_hw_queues == 1) {
1577 struct request *last = NULL;
1579 blk_mq_put_ctx(data.ctx);
1580 blk_mq_bio_to_request(rq, bio);
1583 * @request_count may become stale because of schedule
1584 * out, so check the list again.
1586 if (list_empty(&plug->mq_list))
1588 else if (blk_queue_nomerges(q))
1589 request_count = blk_plug_queued_count(q);
1592 trace_block_plug(q);
1594 last = list_entry_rq(plug->mq_list.prev);
1596 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1597 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1598 blk_flush_plug_list(plug, false);
1599 trace_block_plug(q);
1602 list_add_tail(&rq->queuelist, &plug->mq_list);
1603 } else if (plug && !blk_queue_nomerges(q)) {
1604 blk_mq_bio_to_request(rq, bio);
1607 * We do limited plugging. If the bio can be merged, do that.
1608 * Otherwise the existing request in the plug list will be
1609 * issued. So the plug list will have one request at most
1610 * The plug list might get flushed before this. If that happens,
1611 * the plug list is empty, and same_queue_rq is invalid.
1613 if (list_empty(&plug->mq_list))
1614 same_queue_rq = NULL;
1616 list_del_init(&same_queue_rq->queuelist);
1617 list_add_tail(&rq->queuelist, &plug->mq_list);
1619 blk_mq_put_ctx(data.ctx);
1621 if (same_queue_rq) {
1622 data.hctx = blk_mq_map_queue(q,
1623 same_queue_rq->mq_ctx->cpu);
1624 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1627 } else if (q->nr_hw_queues > 1 && is_sync) {
1628 blk_mq_put_ctx(data.ctx);
1629 blk_mq_bio_to_request(rq, bio);
1630 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1631 } else if (q->elevator) {
1632 blk_mq_put_ctx(data.ctx);
1633 blk_mq_bio_to_request(rq, bio);
1634 blk_mq_sched_insert_request(rq, false, true, true, true);
1636 blk_mq_put_ctx(data.ctx);
1637 blk_mq_bio_to_request(rq, bio);
1638 blk_mq_queue_io(data.hctx, data.ctx, rq);
1639 blk_mq_run_hw_queue(data.hctx, true);
1645 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1646 unsigned int hctx_idx)
1650 if (tags->rqs && set->ops->exit_request) {
1653 for (i = 0; i < tags->nr_tags; i++) {
1654 struct request *rq = tags->static_rqs[i];
1658 set->ops->exit_request(set, rq, hctx_idx);
1659 tags->static_rqs[i] = NULL;
1663 while (!list_empty(&tags->page_list)) {
1664 page = list_first_entry(&tags->page_list, struct page, lru);
1665 list_del_init(&page->lru);
1667 * Remove kmemleak object previously allocated in
1668 * blk_mq_init_rq_map().
1670 kmemleak_free(page_address(page));
1671 __free_pages(page, page->private);
1675 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1679 kfree(tags->static_rqs);
1680 tags->static_rqs = NULL;
1682 blk_mq_free_tags(tags);
1685 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1686 unsigned int hctx_idx,
1687 unsigned int nr_tags,
1688 unsigned int reserved_tags)
1690 struct blk_mq_tags *tags;
1693 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1694 if (node == NUMA_NO_NODE)
1695 node = set->numa_node;
1697 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1698 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1702 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1703 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1706 blk_mq_free_tags(tags);
1710 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1711 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1713 if (!tags->static_rqs) {
1715 blk_mq_free_tags(tags);
1722 static size_t order_to_size(unsigned int order)
1724 return (size_t)PAGE_SIZE << order;
1727 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1728 unsigned int hctx_idx, unsigned int depth)
1730 unsigned int i, j, entries_per_page, max_order = 4;
1731 size_t rq_size, left;
1734 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1735 if (node == NUMA_NO_NODE)
1736 node = set->numa_node;
1738 INIT_LIST_HEAD(&tags->page_list);
1741 * rq_size is the size of the request plus driver payload, rounded
1742 * to the cacheline size
1744 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1746 left = rq_size * depth;
1748 for (i = 0; i < depth; ) {
1749 int this_order = max_order;
1754 while (this_order && left < order_to_size(this_order - 1))
1758 page = alloc_pages_node(node,
1759 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1765 if (order_to_size(this_order) < rq_size)
1772 page->private = this_order;
1773 list_add_tail(&page->lru, &tags->page_list);
1775 p = page_address(page);
1777 * Allow kmemleak to scan these pages as they contain pointers
1778 * to additional allocations like via ops->init_request().
1780 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1781 entries_per_page = order_to_size(this_order) / rq_size;
1782 to_do = min(entries_per_page, depth - i);
1783 left -= to_do * rq_size;
1784 for (j = 0; j < to_do; j++) {
1785 struct request *rq = p;
1787 tags->static_rqs[i] = rq;
1788 if (set->ops->init_request) {
1789 if (set->ops->init_request(set, rq, hctx_idx,
1791 tags->static_rqs[i] = NULL;
1803 blk_mq_free_rqs(set, tags, hctx_idx);
1808 * 'cpu' is going away. splice any existing rq_list entries from this
1809 * software queue to the hw queue dispatch list, and ensure that it
1812 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1814 struct blk_mq_hw_ctx *hctx;
1815 struct blk_mq_ctx *ctx;
1818 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1819 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1821 spin_lock(&ctx->lock);
1822 if (!list_empty(&ctx->rq_list)) {
1823 list_splice_init(&ctx->rq_list, &tmp);
1824 blk_mq_hctx_clear_pending(hctx, ctx);
1826 spin_unlock(&ctx->lock);
1828 if (list_empty(&tmp))
1831 spin_lock(&hctx->lock);
1832 list_splice_tail_init(&tmp, &hctx->dispatch);
1833 spin_unlock(&hctx->lock);
1835 blk_mq_run_hw_queue(hctx, true);
1839 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1841 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1845 /* hctx->ctxs will be freed in queue's release handler */
1846 static void blk_mq_exit_hctx(struct request_queue *q,
1847 struct blk_mq_tag_set *set,
1848 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1850 blk_mq_debugfs_unregister_hctx(hctx);
1852 blk_mq_tag_idle(hctx);
1854 if (set->ops->exit_request)
1855 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1857 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1859 if (set->ops->exit_hctx)
1860 set->ops->exit_hctx(hctx, hctx_idx);
1862 if (hctx->flags & BLK_MQ_F_BLOCKING)
1863 cleanup_srcu_struct(hctx->queue_rq_srcu);
1865 blk_mq_remove_cpuhp(hctx);
1866 blk_free_flush_queue(hctx->fq);
1867 sbitmap_free(&hctx->ctx_map);
1870 static void blk_mq_exit_hw_queues(struct request_queue *q,
1871 struct blk_mq_tag_set *set, int nr_queue)
1873 struct blk_mq_hw_ctx *hctx;
1876 queue_for_each_hw_ctx(q, hctx, i) {
1879 blk_mq_exit_hctx(q, set, hctx, i);
1883 static int blk_mq_init_hctx(struct request_queue *q,
1884 struct blk_mq_tag_set *set,
1885 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1889 node = hctx->numa_node;
1890 if (node == NUMA_NO_NODE)
1891 node = hctx->numa_node = set->numa_node;
1893 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1894 spin_lock_init(&hctx->lock);
1895 INIT_LIST_HEAD(&hctx->dispatch);
1897 hctx->queue_num = hctx_idx;
1898 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1900 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1902 hctx->tags = set->tags[hctx_idx];
1905 * Allocate space for all possible cpus to avoid allocation at
1908 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1911 goto unregister_cpu_notifier;
1913 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1919 if (set->ops->init_hctx &&
1920 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1923 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1926 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1928 goto sched_exit_hctx;
1930 if (set->ops->init_request &&
1931 set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1935 if (hctx->flags & BLK_MQ_F_BLOCKING)
1936 init_srcu_struct(hctx->queue_rq_srcu);
1938 blk_mq_debugfs_register_hctx(q, hctx);
1945 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1947 if (set->ops->exit_hctx)
1948 set->ops->exit_hctx(hctx, hctx_idx);
1950 sbitmap_free(&hctx->ctx_map);
1953 unregister_cpu_notifier:
1954 blk_mq_remove_cpuhp(hctx);
1958 static void blk_mq_init_cpu_queues(struct request_queue *q,
1959 unsigned int nr_hw_queues)
1963 for_each_possible_cpu(i) {
1964 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1965 struct blk_mq_hw_ctx *hctx;
1968 spin_lock_init(&__ctx->lock);
1969 INIT_LIST_HEAD(&__ctx->rq_list);
1972 /* If the cpu isn't online, the cpu is mapped to first hctx */
1976 hctx = blk_mq_map_queue(q, i);
1979 * Set local node, IFF we have more than one hw queue. If
1980 * not, we remain on the home node of the device
1982 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1983 hctx->numa_node = local_memory_node(cpu_to_node(i));
1987 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1991 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1992 set->queue_depth, set->reserved_tags);
1993 if (!set->tags[hctx_idx])
1996 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2001 blk_mq_free_rq_map(set->tags[hctx_idx]);
2002 set->tags[hctx_idx] = NULL;
2006 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2007 unsigned int hctx_idx)
2009 if (set->tags[hctx_idx]) {
2010 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2011 blk_mq_free_rq_map(set->tags[hctx_idx]);
2012 set->tags[hctx_idx] = NULL;
2016 static void blk_mq_map_swqueue(struct request_queue *q,
2017 const struct cpumask *online_mask)
2019 unsigned int i, hctx_idx;
2020 struct blk_mq_hw_ctx *hctx;
2021 struct blk_mq_ctx *ctx;
2022 struct blk_mq_tag_set *set = q->tag_set;
2025 * Avoid others reading imcomplete hctx->cpumask through sysfs
2027 mutex_lock(&q->sysfs_lock);
2029 queue_for_each_hw_ctx(q, hctx, i) {
2030 cpumask_clear(hctx->cpumask);
2035 * Map software to hardware queues
2037 for_each_possible_cpu(i) {
2038 /* If the cpu isn't online, the cpu is mapped to first hctx */
2039 if (!cpumask_test_cpu(i, online_mask))
2042 hctx_idx = q->mq_map[i];
2043 /* unmapped hw queue can be remapped after CPU topo changed */
2044 if (!set->tags[hctx_idx] &&
2045 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2047 * If tags initialization fail for some hctx,
2048 * that hctx won't be brought online. In this
2049 * case, remap the current ctx to hctx[0] which
2050 * is guaranteed to always have tags allocated
2055 ctx = per_cpu_ptr(q->queue_ctx, i);
2056 hctx = blk_mq_map_queue(q, i);
2058 cpumask_set_cpu(i, hctx->cpumask);
2059 ctx->index_hw = hctx->nr_ctx;
2060 hctx->ctxs[hctx->nr_ctx++] = ctx;
2063 mutex_unlock(&q->sysfs_lock);
2065 queue_for_each_hw_ctx(q, hctx, i) {
2067 * If no software queues are mapped to this hardware queue,
2068 * disable it and free the request entries.
2070 if (!hctx->nr_ctx) {
2071 /* Never unmap queue 0. We need it as a
2072 * fallback in case of a new remap fails
2075 if (i && set->tags[i])
2076 blk_mq_free_map_and_requests(set, i);
2082 hctx->tags = set->tags[i];
2083 WARN_ON(!hctx->tags);
2086 * Set the map size to the number of mapped software queues.
2087 * This is more accurate and more efficient than looping
2088 * over all possibly mapped software queues.
2090 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2093 * Initialize batch roundrobin counts
2095 hctx->next_cpu = cpumask_first(hctx->cpumask);
2096 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2100 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2102 struct blk_mq_hw_ctx *hctx;
2105 queue_for_each_hw_ctx(q, hctx, i) {
2107 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2109 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2113 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2115 struct request_queue *q;
2117 lockdep_assert_held(&set->tag_list_lock);
2119 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2120 blk_mq_freeze_queue(q);
2121 queue_set_hctx_shared(q, shared);
2122 blk_mq_unfreeze_queue(q);
2126 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2128 struct blk_mq_tag_set *set = q->tag_set;
2130 mutex_lock(&set->tag_list_lock);
2131 list_del_rcu(&q->tag_set_list);
2132 INIT_LIST_HEAD(&q->tag_set_list);
2133 if (list_is_singular(&set->tag_list)) {
2134 /* just transitioned to unshared */
2135 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2136 /* update existing queue */
2137 blk_mq_update_tag_set_depth(set, false);
2139 mutex_unlock(&set->tag_list_lock);
2144 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2145 struct request_queue *q)
2149 mutex_lock(&set->tag_list_lock);
2151 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2152 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2153 set->flags |= BLK_MQ_F_TAG_SHARED;
2154 /* update existing queue */
2155 blk_mq_update_tag_set_depth(set, true);
2157 if (set->flags & BLK_MQ_F_TAG_SHARED)
2158 queue_set_hctx_shared(q, true);
2159 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2161 mutex_unlock(&set->tag_list_lock);
2165 * It is the actual release handler for mq, but we do it from
2166 * request queue's release handler for avoiding use-after-free
2167 * and headache because q->mq_kobj shouldn't have been introduced,
2168 * but we can't group ctx/kctx kobj without it.
2170 void blk_mq_release(struct request_queue *q)
2172 struct blk_mq_hw_ctx *hctx;
2175 /* hctx kobj stays in hctx */
2176 queue_for_each_hw_ctx(q, hctx, i) {
2179 kobject_put(&hctx->kobj);
2184 kfree(q->queue_hw_ctx);
2187 * release .mq_kobj and sw queue's kobject now because
2188 * both share lifetime with request queue.
2190 blk_mq_sysfs_deinit(q);
2192 free_percpu(q->queue_ctx);
2195 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2197 struct request_queue *uninit_q, *q;
2199 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2201 return ERR_PTR(-ENOMEM);
2203 q = blk_mq_init_allocated_queue(set, uninit_q);
2205 blk_cleanup_queue(uninit_q);
2209 EXPORT_SYMBOL(blk_mq_init_queue);
2211 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2213 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2215 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2216 __alignof__(struct blk_mq_hw_ctx)) !=
2217 sizeof(struct blk_mq_hw_ctx));
2219 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2220 hw_ctx_size += sizeof(struct srcu_struct);
2225 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2226 struct request_queue *q)
2229 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2231 blk_mq_sysfs_unregister(q);
2232 for (i = 0; i < set->nr_hw_queues; i++) {
2238 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2239 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2244 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2251 atomic_set(&hctxs[i]->nr_active, 0);
2252 hctxs[i]->numa_node = node;
2253 hctxs[i]->queue_num = i;
2255 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2256 free_cpumask_var(hctxs[i]->cpumask);
2261 blk_mq_hctx_kobj_init(hctxs[i]);
2263 for (j = i; j < q->nr_hw_queues; j++) {
2264 struct blk_mq_hw_ctx *hctx = hctxs[j];
2268 blk_mq_free_map_and_requests(set, j);
2269 blk_mq_exit_hctx(q, set, hctx, j);
2270 kobject_put(&hctx->kobj);
2275 q->nr_hw_queues = i;
2276 blk_mq_sysfs_register(q);
2279 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2280 struct request_queue *q)
2282 /* mark the queue as mq asap */
2283 q->mq_ops = set->ops;
2285 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2286 blk_mq_poll_stats_bkt,
2287 BLK_MQ_POLL_STATS_BKTS, q);
2291 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2295 /* init q->mq_kobj and sw queues' kobjects */
2296 blk_mq_sysfs_init(q);
2298 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2299 GFP_KERNEL, set->numa_node);
2300 if (!q->queue_hw_ctx)
2303 q->mq_map = set->mq_map;
2305 blk_mq_realloc_hw_ctxs(set, q);
2306 if (!q->nr_hw_queues)
2309 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2310 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2312 q->nr_queues = nr_cpu_ids;
2314 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2316 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2317 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2319 q->sg_reserved_size = INT_MAX;
2321 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2322 INIT_LIST_HEAD(&q->requeue_list);
2323 spin_lock_init(&q->requeue_lock);
2325 blk_queue_make_request(q, blk_mq_make_request);
2328 * Do this after blk_queue_make_request() overrides it...
2330 q->nr_requests = set->queue_depth;
2333 * Default to classic polling
2337 if (set->ops->complete)
2338 blk_queue_softirq_done(q, set->ops->complete);
2340 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2343 mutex_lock(&all_q_mutex);
2345 list_add_tail(&q->all_q_node, &all_q_list);
2346 blk_mq_add_queue_tag_set(set, q);
2347 blk_mq_map_swqueue(q, cpu_online_mask);
2349 mutex_unlock(&all_q_mutex);
2352 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2355 ret = blk_mq_sched_init(q);
2357 return ERR_PTR(ret);
2363 kfree(q->queue_hw_ctx);
2365 free_percpu(q->queue_ctx);
2368 return ERR_PTR(-ENOMEM);
2370 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2372 void blk_mq_free_queue(struct request_queue *q)
2374 struct blk_mq_tag_set *set = q->tag_set;
2376 mutex_lock(&all_q_mutex);
2377 list_del_init(&q->all_q_node);
2378 mutex_unlock(&all_q_mutex);
2380 blk_mq_del_queue_tag_set(q);
2382 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2385 /* Basically redo blk_mq_init_queue with queue frozen */
2386 static void blk_mq_queue_reinit(struct request_queue *q,
2387 const struct cpumask *online_mask)
2389 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2391 blk_mq_debugfs_unregister_hctxs(q);
2392 blk_mq_sysfs_unregister(q);
2395 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2396 * we should change hctx numa_node according to new topology (this
2397 * involves free and re-allocate memory, worthy doing?)
2400 blk_mq_map_swqueue(q, online_mask);
2402 blk_mq_sysfs_register(q);
2403 blk_mq_debugfs_register_hctxs(q);
2407 * New online cpumask which is going to be set in this hotplug event.
2408 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2409 * one-by-one and dynamically allocating this could result in a failure.
2411 static struct cpumask cpuhp_online_new;
2413 static void blk_mq_queue_reinit_work(void)
2415 struct request_queue *q;
2417 mutex_lock(&all_q_mutex);
2419 * We need to freeze and reinit all existing queues. Freezing
2420 * involves synchronous wait for an RCU grace period and doing it
2421 * one by one may take a long time. Start freezing all queues in
2422 * one swoop and then wait for the completions so that freezing can
2423 * take place in parallel.
2425 list_for_each_entry(q, &all_q_list, all_q_node)
2426 blk_freeze_queue_start(q);
2427 list_for_each_entry(q, &all_q_list, all_q_node)
2428 blk_mq_freeze_queue_wait(q);
2430 list_for_each_entry(q, &all_q_list, all_q_node)
2431 blk_mq_queue_reinit(q, &cpuhp_online_new);
2433 list_for_each_entry(q, &all_q_list, all_q_node)
2434 blk_mq_unfreeze_queue(q);
2436 mutex_unlock(&all_q_mutex);
2439 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2441 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2442 blk_mq_queue_reinit_work();
2447 * Before hotadded cpu starts handling requests, new mappings must be
2448 * established. Otherwise, these requests in hw queue might never be
2451 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2452 * for CPU0, and ctx1 for CPU1).
2454 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2455 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2457 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2458 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2459 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2462 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2464 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2465 cpumask_set_cpu(cpu, &cpuhp_online_new);
2466 blk_mq_queue_reinit_work();
2470 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2474 for (i = 0; i < set->nr_hw_queues; i++)
2475 if (!__blk_mq_alloc_rq_map(set, i))
2482 blk_mq_free_rq_map(set->tags[i]);
2488 * Allocate the request maps associated with this tag_set. Note that this
2489 * may reduce the depth asked for, if memory is tight. set->queue_depth
2490 * will be updated to reflect the allocated depth.
2492 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2497 depth = set->queue_depth;
2499 err = __blk_mq_alloc_rq_maps(set);
2503 set->queue_depth >>= 1;
2504 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2508 } while (set->queue_depth);
2510 if (!set->queue_depth || err) {
2511 pr_err("blk-mq: failed to allocate request map\n");
2515 if (depth != set->queue_depth)
2516 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2517 depth, set->queue_depth);
2522 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2524 if (set->ops->map_queues)
2525 return set->ops->map_queues(set);
2527 return blk_mq_map_queues(set);
2531 * Alloc a tag set to be associated with one or more request queues.
2532 * May fail with EINVAL for various error conditions. May adjust the
2533 * requested depth down, if if it too large. In that case, the set
2534 * value will be stored in set->queue_depth.
2536 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2540 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2542 if (!set->nr_hw_queues)
2544 if (!set->queue_depth)
2546 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2549 if (!set->ops->queue_rq)
2552 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2553 pr_info("blk-mq: reduced tag depth to %u\n",
2555 set->queue_depth = BLK_MQ_MAX_DEPTH;
2559 * If a crashdump is active, then we are potentially in a very
2560 * memory constrained environment. Limit us to 1 queue and
2561 * 64 tags to prevent using too much memory.
2563 if (is_kdump_kernel()) {
2564 set->nr_hw_queues = 1;
2565 set->queue_depth = min(64U, set->queue_depth);
2568 * There is no use for more h/w queues than cpus.
2570 if (set->nr_hw_queues > nr_cpu_ids)
2571 set->nr_hw_queues = nr_cpu_ids;
2573 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2574 GFP_KERNEL, set->numa_node);
2579 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2580 GFP_KERNEL, set->numa_node);
2584 ret = blk_mq_update_queue_map(set);
2586 goto out_free_mq_map;
2588 ret = blk_mq_alloc_rq_maps(set);
2590 goto out_free_mq_map;
2592 mutex_init(&set->tag_list_lock);
2593 INIT_LIST_HEAD(&set->tag_list);
2605 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2607 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2611 for (i = 0; i < nr_cpu_ids; i++)
2612 blk_mq_free_map_and_requests(set, i);
2620 EXPORT_SYMBOL(blk_mq_free_tag_set);
2622 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2624 struct blk_mq_tag_set *set = q->tag_set;
2625 struct blk_mq_hw_ctx *hctx;
2631 blk_mq_freeze_queue(q);
2634 queue_for_each_hw_ctx(q, hctx, i) {
2638 * If we're using an MQ scheduler, just update the scheduler
2639 * queue depth. This is similar to what the old code would do.
2641 if (!hctx->sched_tags) {
2642 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2643 min(nr, set->queue_depth),
2646 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2654 q->nr_requests = nr;
2656 blk_mq_unfreeze_queue(q);
2661 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2664 struct request_queue *q;
2666 lockdep_assert_held(&set->tag_list_lock);
2668 if (nr_hw_queues > nr_cpu_ids)
2669 nr_hw_queues = nr_cpu_ids;
2670 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2673 list_for_each_entry(q, &set->tag_list, tag_set_list)
2674 blk_mq_freeze_queue(q);
2676 set->nr_hw_queues = nr_hw_queues;
2677 blk_mq_update_queue_map(set);
2678 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2679 blk_mq_realloc_hw_ctxs(set, q);
2680 blk_mq_queue_reinit(q, cpu_online_mask);
2683 list_for_each_entry(q, &set->tag_list, tag_set_list)
2684 blk_mq_unfreeze_queue(q);
2687 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2689 mutex_lock(&set->tag_list_lock);
2690 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2691 mutex_unlock(&set->tag_list_lock);
2693 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2695 /* Enable polling stats and return whether they were already enabled. */
2696 static bool blk_poll_stats_enable(struct request_queue *q)
2698 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2699 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2701 blk_stat_add_callback(q, q->poll_cb);
2705 static void blk_mq_poll_stats_start(struct request_queue *q)
2708 * We don't arm the callback if polling stats are not enabled or the
2709 * callback is already active.
2711 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2712 blk_stat_is_active(q->poll_cb))
2715 blk_stat_activate_msecs(q->poll_cb, 100);
2718 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2720 struct request_queue *q = cb->data;
2723 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2724 if (cb->stat[bucket].nr_samples)
2725 q->poll_stat[bucket] = cb->stat[bucket];
2729 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2730 struct blk_mq_hw_ctx *hctx,
2733 unsigned long ret = 0;
2737 * If stats collection isn't on, don't sleep but turn it on for
2740 if (!blk_poll_stats_enable(q))
2744 * As an optimistic guess, use half of the mean service time
2745 * for this type of request. We can (and should) make this smarter.
2746 * For instance, if the completion latencies are tight, we can
2747 * get closer than just half the mean. This is especially
2748 * important on devices where the completion latencies are longer
2749 * than ~10 usec. We do use the stats for the relevant IO size
2750 * if available which does lead to better estimates.
2752 bucket = blk_mq_poll_stats_bkt(rq);
2756 if (q->poll_stat[bucket].nr_samples)
2757 ret = (q->poll_stat[bucket].mean + 1) / 2;
2762 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2763 struct blk_mq_hw_ctx *hctx,
2766 struct hrtimer_sleeper hs;
2767 enum hrtimer_mode mode;
2771 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2777 * -1: don't ever hybrid sleep
2778 * 0: use half of prev avg
2779 * >0: use this specific value
2781 if (q->poll_nsec == -1)
2783 else if (q->poll_nsec > 0)
2784 nsecs = q->poll_nsec;
2786 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2791 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2794 * This will be replaced with the stats tracking code, using
2795 * 'avg_completion_time / 2' as the pre-sleep target.
2799 mode = HRTIMER_MODE_REL;
2800 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2801 hrtimer_set_expires(&hs.timer, kt);
2803 hrtimer_init_sleeper(&hs, current);
2805 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2807 set_current_state(TASK_UNINTERRUPTIBLE);
2808 hrtimer_start_expires(&hs.timer, mode);
2811 hrtimer_cancel(&hs.timer);
2812 mode = HRTIMER_MODE_ABS;
2813 } while (hs.task && !signal_pending(current));
2815 __set_current_state(TASK_RUNNING);
2816 destroy_hrtimer_on_stack(&hs.timer);
2820 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2822 struct request_queue *q = hctx->queue;
2826 * If we sleep, have the caller restart the poll loop to reset
2827 * the state. Like for the other success return cases, the
2828 * caller is responsible for checking if the IO completed. If
2829 * the IO isn't complete, we'll get called again and will go
2830 * straight to the busy poll loop.
2832 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2835 hctx->poll_considered++;
2837 state = current->state;
2838 while (!need_resched()) {
2841 hctx->poll_invoked++;
2843 ret = q->mq_ops->poll(hctx, rq->tag);
2845 hctx->poll_success++;
2846 set_current_state(TASK_RUNNING);
2850 if (signal_pending_state(state, current))
2851 set_current_state(TASK_RUNNING);
2853 if (current->state == TASK_RUNNING)
2863 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2865 struct blk_mq_hw_ctx *hctx;
2866 struct blk_plug *plug;
2869 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2870 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2873 plug = current->plug;
2875 blk_flush_plug_list(plug, false);
2877 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2878 if (!blk_qc_t_is_internal(cookie))
2879 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2881 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2883 * With scheduling, if the request has completed, we'll
2884 * get a NULL return here, as we clear the sched tag when
2885 * that happens. The request still remains valid, like always,
2886 * so we should be safe with just the NULL check.
2892 return __blk_mq_poll(hctx, rq);
2894 EXPORT_SYMBOL_GPL(blk_mq_poll);
2896 void blk_mq_disable_hotplug(void)
2898 mutex_lock(&all_q_mutex);
2901 void blk_mq_enable_hotplug(void)
2903 mutex_unlock(&all_q_mutex);
2906 static int __init blk_mq_init(void)
2908 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2909 blk_mq_hctx_notify_dead);
2911 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2912 blk_mq_queue_reinit_prepare,
2913 blk_mq_queue_reinit_dead);
2916 subsys_initcall(blk_mq_init);