2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-debugfs.h"
35 #include "blk-mq-tag.h"
38 #include "blk-mq-sched.h"
40 static DEFINE_MUTEX(all_q_mutex);
41 static LIST_HEAD(all_q_list);
43 static void blk_mq_poll_stats_start(struct request_queue *q);
44 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
46 static int blk_mq_poll_stats_bkt(const struct request *rq)
48 int ddir, bytes, bucket;
50 ddir = rq_data_dir(rq);
51 bytes = blk_rq_bytes(rq);
53 bucket = ddir + 2*(ilog2(bytes) - 9);
57 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
58 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
64 * Check if any of the ctx's have pending work in this hardware queue
66 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
68 return sbitmap_any_bit_set(&hctx->ctx_map) ||
69 !list_empty_careful(&hctx->dispatch) ||
70 blk_mq_sched_has_work(hctx);
74 * Mark this ctx as having pending work in this hardware queue
76 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
77 struct blk_mq_ctx *ctx)
79 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
80 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
83 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
84 struct blk_mq_ctx *ctx)
86 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
89 void blk_freeze_queue_start(struct request_queue *q)
93 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
94 if (freeze_depth == 1) {
95 percpu_ref_kill(&q->q_usage_counter);
96 blk_mq_run_hw_queues(q, false);
99 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
101 void blk_mq_freeze_queue_wait(struct request_queue *q)
103 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
105 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
107 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
108 unsigned long timeout)
110 return wait_event_timeout(q->mq_freeze_wq,
111 percpu_ref_is_zero(&q->q_usage_counter),
114 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
117 * Guarantee no request is in use, so we can change any data structure of
118 * the queue afterward.
120 void blk_freeze_queue(struct request_queue *q)
123 * In the !blk_mq case we are only calling this to kill the
124 * q_usage_counter, otherwise this increases the freeze depth
125 * and waits for it to return to zero. For this reason there is
126 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
127 * exported to drivers as the only user for unfreeze is blk_mq.
129 blk_freeze_queue_start(q);
130 blk_mq_freeze_queue_wait(q);
133 void blk_mq_freeze_queue(struct request_queue *q)
136 * ...just an alias to keep freeze and unfreeze actions balanced
137 * in the blk_mq_* namespace
141 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
143 void blk_mq_unfreeze_queue(struct request_queue *q)
147 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
148 WARN_ON_ONCE(freeze_depth < 0);
150 percpu_ref_reinit(&q->q_usage_counter);
151 wake_up_all(&q->mq_freeze_wq);
154 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
157 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
160 * Note: this function does not prevent that the struct request end_io()
161 * callback function is invoked. Once this function is returned, we make
162 * sure no dispatch can happen until the queue is unquiesced via
163 * blk_mq_unquiesce_queue().
165 void blk_mq_quiesce_queue(struct request_queue *q)
167 struct blk_mq_hw_ctx *hctx;
171 blk_mq_quiesce_queue_nowait(q);
173 queue_for_each_hw_ctx(q, hctx, i) {
174 if (hctx->flags & BLK_MQ_F_BLOCKING)
175 synchronize_srcu(&hctx->queue_rq_srcu);
182 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
185 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
188 * This function recovers queue into the state before quiescing
189 * which is done by blk_mq_quiesce_queue.
191 void blk_mq_unquiesce_queue(struct request_queue *q)
193 spin_lock_irq(q->queue_lock);
194 queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
195 spin_unlock_irq(q->queue_lock);
197 /* dispatch requests which are inserted during quiescing */
198 blk_mq_run_hw_queues(q, true);
200 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
202 void blk_mq_wake_waiters(struct request_queue *q)
204 struct blk_mq_hw_ctx *hctx;
207 queue_for_each_hw_ctx(q, hctx, i)
208 if (blk_mq_hw_queue_mapped(hctx))
209 blk_mq_tag_wakeup_all(hctx->tags, true);
212 * If we are called because the queue has now been marked as
213 * dying, we need to ensure that processes currently waiting on
214 * the queue are notified as well.
216 wake_up_all(&q->mq_freeze_wq);
219 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
221 return blk_mq_has_free_tags(hctx->tags);
223 EXPORT_SYMBOL(blk_mq_can_queue);
225 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
226 unsigned int tag, unsigned int op)
228 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
229 struct request *rq = tags->static_rqs[tag];
231 if (data->flags & BLK_MQ_REQ_INTERNAL) {
233 rq->internal_tag = tag;
235 if (blk_mq_tag_busy(data->hctx)) {
236 rq->rq_flags = RQF_MQ_INFLIGHT;
237 atomic_inc(&data->hctx->nr_active);
240 rq->internal_tag = -1;
241 data->hctx->tags->rqs[rq->tag] = rq;
244 INIT_LIST_HEAD(&rq->queuelist);
245 /* csd/requeue_work/fifo_time is initialized before use */
247 rq->mq_ctx = data->ctx;
249 if (blk_queue_io_stat(data->q))
250 rq->rq_flags |= RQF_IO_STAT;
251 /* do not touch atomic flags, it needs atomic ops against the timer */
253 INIT_HLIST_NODE(&rq->hash);
254 RB_CLEAR_NODE(&rq->rb_node);
257 rq->start_time = jiffies;
258 #ifdef CONFIG_BLK_CGROUP
260 set_start_time_ns(rq);
261 rq->io_start_time_ns = 0;
263 rq->nr_phys_segments = 0;
264 #if defined(CONFIG_BLK_DEV_INTEGRITY)
265 rq->nr_integrity_segments = 0;
268 /* tag was already set */
271 INIT_LIST_HEAD(&rq->timeout_list);
275 rq->end_io_data = NULL;
278 data->ctx->rq_dispatched[op_is_sync(op)]++;
282 static struct request *blk_mq_get_request(struct request_queue *q,
283 struct bio *bio, unsigned int op,
284 struct blk_mq_alloc_data *data)
286 struct elevator_queue *e = q->elevator;
290 blk_queue_enter_live(q);
292 if (likely(!data->ctx))
293 data->ctx = blk_mq_get_ctx(q);
294 if (likely(!data->hctx))
295 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
298 data->flags |= BLK_MQ_REQ_INTERNAL;
301 * Flush requests are special and go directly to the
304 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
305 e->type->ops.mq.limit_depth(op, data);
308 tag = blk_mq_get_tag(data);
309 if (tag == BLK_MQ_TAG_FAIL) {
314 rq = blk_mq_rq_ctx_init(data, tag, op);
315 if (!op_is_flush(op)) {
317 if (e && e->type->ops.mq.prepare_request) {
318 if (e->type->icq_cache && rq_ioc(bio))
319 blk_mq_sched_assign_ioc(rq, bio);
321 e->type->ops.mq.prepare_request(rq, bio);
322 rq->rq_flags |= RQF_ELVPRIV;
325 data->hctx->queued++;
329 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
332 struct blk_mq_alloc_data alloc_data = { .flags = flags };
336 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
340 rq = blk_mq_get_request(q, NULL, rw, &alloc_data);
342 blk_mq_put_ctx(alloc_data.ctx);
346 return ERR_PTR(-EWOULDBLOCK);
349 rq->__sector = (sector_t) -1;
350 rq->bio = rq->biotail = NULL;
353 EXPORT_SYMBOL(blk_mq_alloc_request);
355 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
356 unsigned int flags, unsigned int hctx_idx)
358 struct blk_mq_alloc_data alloc_data = { .flags = flags };
364 * If the tag allocator sleeps we could get an allocation for a
365 * different hardware context. No need to complicate the low level
366 * allocator for this for the rare use case of a command tied to
369 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
370 return ERR_PTR(-EINVAL);
372 if (hctx_idx >= q->nr_hw_queues)
373 return ERR_PTR(-EIO);
375 ret = blk_queue_enter(q, true);
380 * Check if the hardware context is actually mapped to anything.
381 * If not tell the caller that it should skip this queue.
383 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
384 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
386 return ERR_PTR(-EXDEV);
388 cpu = cpumask_first(alloc_data.hctx->cpumask);
389 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
391 rq = blk_mq_get_request(q, NULL, rw, &alloc_data);
396 return ERR_PTR(-EWOULDBLOCK);
400 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
402 void blk_mq_free_request(struct request *rq)
404 struct request_queue *q = rq->q;
405 struct elevator_queue *e = q->elevator;
406 struct blk_mq_ctx *ctx = rq->mq_ctx;
407 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
408 const int sched_tag = rq->internal_tag;
410 if (rq->rq_flags & RQF_ELVPRIV) {
411 if (e && e->type->ops.mq.finish_request)
412 e->type->ops.mq.finish_request(rq);
414 put_io_context(rq->elv.icq->ioc);
419 ctx->rq_completed[rq_is_sync(rq)]++;
420 if (rq->rq_flags & RQF_MQ_INFLIGHT)
421 atomic_dec(&hctx->nr_active);
423 wbt_done(q->rq_wb, &rq->issue_stat);
426 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
427 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
429 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
431 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
432 blk_mq_sched_restart(hctx);
435 EXPORT_SYMBOL_GPL(blk_mq_free_request);
437 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
439 blk_account_io_done(rq);
442 wbt_done(rq->q->rq_wb, &rq->issue_stat);
443 rq->end_io(rq, error);
445 if (unlikely(blk_bidi_rq(rq)))
446 blk_mq_free_request(rq->next_rq);
447 blk_mq_free_request(rq);
450 EXPORT_SYMBOL(__blk_mq_end_request);
452 void blk_mq_end_request(struct request *rq, blk_status_t error)
454 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
456 __blk_mq_end_request(rq, error);
458 EXPORT_SYMBOL(blk_mq_end_request);
460 static void __blk_mq_complete_request_remote(void *data)
462 struct request *rq = data;
464 rq->q->softirq_done_fn(rq);
467 static void __blk_mq_complete_request(struct request *rq)
469 struct blk_mq_ctx *ctx = rq->mq_ctx;
473 if (rq->internal_tag != -1)
474 blk_mq_sched_completed_request(rq);
475 if (rq->rq_flags & RQF_STATS) {
476 blk_mq_poll_stats_start(rq->q);
480 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
481 rq->q->softirq_done_fn(rq);
486 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
487 shared = cpus_share_cache(cpu, ctx->cpu);
489 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
490 rq->csd.func = __blk_mq_complete_request_remote;
493 smp_call_function_single_async(ctx->cpu, &rq->csd);
495 rq->q->softirq_done_fn(rq);
501 * blk_mq_complete_request - end I/O on a request
502 * @rq: the request being processed
505 * Ends all I/O on a request. It does not handle partial completions.
506 * The actual completion happens out-of-order, through a IPI handler.
508 void blk_mq_complete_request(struct request *rq)
510 struct request_queue *q = rq->q;
512 if (unlikely(blk_should_fake_timeout(q)))
514 if (!blk_mark_rq_complete(rq))
515 __blk_mq_complete_request(rq);
517 EXPORT_SYMBOL(blk_mq_complete_request);
519 int blk_mq_request_started(struct request *rq)
521 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
523 EXPORT_SYMBOL_GPL(blk_mq_request_started);
525 void blk_mq_start_request(struct request *rq)
527 struct request_queue *q = rq->q;
529 blk_mq_sched_started_request(rq);
531 trace_block_rq_issue(q, rq);
533 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
534 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
535 rq->rq_flags |= RQF_STATS;
536 wbt_issue(q->rq_wb, &rq->issue_stat);
542 * Ensure that ->deadline is visible before set the started
543 * flag and clear the completed flag.
545 smp_mb__before_atomic();
548 * Mark us as started and clear complete. Complete might have been
549 * set if requeue raced with timeout, which then marked it as
550 * complete. So be sure to clear complete again when we start
551 * the request, otherwise we'll ignore the completion event.
553 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
554 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
555 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
556 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
558 if (q->dma_drain_size && blk_rq_bytes(rq)) {
560 * Make sure space for the drain appears. We know we can do
561 * this because max_hw_segments has been adjusted to be one
562 * fewer than the device can handle.
564 rq->nr_phys_segments++;
567 EXPORT_SYMBOL(blk_mq_start_request);
570 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
571 * flag isn't set yet, so there may be race with timeout handler,
572 * but given rq->deadline is just set in .queue_rq() under
573 * this situation, the race won't be possible in reality because
574 * rq->timeout should be set as big enough to cover the window
575 * between blk_mq_start_request() called from .queue_rq() and
576 * clearing REQ_ATOM_STARTED here.
578 static void __blk_mq_requeue_request(struct request *rq)
580 struct request_queue *q = rq->q;
582 trace_block_rq_requeue(q, rq);
583 wbt_requeue(q->rq_wb, &rq->issue_stat);
584 blk_mq_sched_requeue_request(rq);
586 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
587 if (q->dma_drain_size && blk_rq_bytes(rq))
588 rq->nr_phys_segments--;
592 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
594 __blk_mq_requeue_request(rq);
596 BUG_ON(blk_queued_rq(rq));
597 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
599 EXPORT_SYMBOL(blk_mq_requeue_request);
601 static void blk_mq_requeue_work(struct work_struct *work)
603 struct request_queue *q =
604 container_of(work, struct request_queue, requeue_work.work);
606 struct request *rq, *next;
609 spin_lock_irqsave(&q->requeue_lock, flags);
610 list_splice_init(&q->requeue_list, &rq_list);
611 spin_unlock_irqrestore(&q->requeue_lock, flags);
613 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
614 if (!(rq->rq_flags & RQF_SOFTBARRIER))
617 rq->rq_flags &= ~RQF_SOFTBARRIER;
618 list_del_init(&rq->queuelist);
619 blk_mq_sched_insert_request(rq, true, false, false, true);
622 while (!list_empty(&rq_list)) {
623 rq = list_entry(rq_list.next, struct request, queuelist);
624 list_del_init(&rq->queuelist);
625 blk_mq_sched_insert_request(rq, false, false, false, true);
628 blk_mq_run_hw_queues(q, false);
631 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
632 bool kick_requeue_list)
634 struct request_queue *q = rq->q;
638 * We abuse this flag that is otherwise used by the I/O scheduler to
639 * request head insertation from the workqueue.
641 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
643 spin_lock_irqsave(&q->requeue_lock, flags);
645 rq->rq_flags |= RQF_SOFTBARRIER;
646 list_add(&rq->queuelist, &q->requeue_list);
648 list_add_tail(&rq->queuelist, &q->requeue_list);
650 spin_unlock_irqrestore(&q->requeue_lock, flags);
652 if (kick_requeue_list)
653 blk_mq_kick_requeue_list(q);
655 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
657 void blk_mq_kick_requeue_list(struct request_queue *q)
659 kblockd_schedule_delayed_work(&q->requeue_work, 0);
661 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
663 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
666 kblockd_schedule_delayed_work(&q->requeue_work,
667 msecs_to_jiffies(msecs));
669 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
671 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
673 if (tag < tags->nr_tags) {
674 prefetch(tags->rqs[tag]);
675 return tags->rqs[tag];
680 EXPORT_SYMBOL(blk_mq_tag_to_rq);
682 struct blk_mq_timeout_data {
684 unsigned int next_set;
687 void blk_mq_rq_timed_out(struct request *req, bool reserved)
689 const struct blk_mq_ops *ops = req->q->mq_ops;
690 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
693 * We know that complete is set at this point. If STARTED isn't set
694 * anymore, then the request isn't active and the "timeout" should
695 * just be ignored. This can happen due to the bitflag ordering.
696 * Timeout first checks if STARTED is set, and if it is, assumes
697 * the request is active. But if we race with completion, then
698 * both flags will get cleared. So check here again, and ignore
699 * a timeout event with a request that isn't active.
701 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
705 ret = ops->timeout(req, reserved);
709 __blk_mq_complete_request(req);
711 case BLK_EH_RESET_TIMER:
713 blk_clear_rq_complete(req);
715 case BLK_EH_NOT_HANDLED:
718 printk(KERN_ERR "block: bad eh return: %d\n", ret);
723 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
724 struct request *rq, void *priv, bool reserved)
726 struct blk_mq_timeout_data *data = priv;
728 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
732 * The rq being checked may have been freed and reallocated
733 * out already here, we avoid this race by checking rq->deadline
734 * and REQ_ATOM_COMPLETE flag together:
736 * - if rq->deadline is observed as new value because of
737 * reusing, the rq won't be timed out because of timing.
738 * - if rq->deadline is observed as previous value,
739 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
740 * because we put a barrier between setting rq->deadline
741 * and clearing the flag in blk_mq_start_request(), so
742 * this rq won't be timed out too.
744 if (time_after_eq(jiffies, rq->deadline)) {
745 if (!blk_mark_rq_complete(rq))
746 blk_mq_rq_timed_out(rq, reserved);
747 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
748 data->next = rq->deadline;
753 static void blk_mq_timeout_work(struct work_struct *work)
755 struct request_queue *q =
756 container_of(work, struct request_queue, timeout_work);
757 struct blk_mq_timeout_data data = {
763 /* A deadlock might occur if a request is stuck requiring a
764 * timeout at the same time a queue freeze is waiting
765 * completion, since the timeout code would not be able to
766 * acquire the queue reference here.
768 * That's why we don't use blk_queue_enter here; instead, we use
769 * percpu_ref_tryget directly, because we need to be able to
770 * obtain a reference even in the short window between the queue
771 * starting to freeze, by dropping the first reference in
772 * blk_freeze_queue_start, and the moment the last request is
773 * consumed, marked by the instant q_usage_counter reaches
776 if (!percpu_ref_tryget(&q->q_usage_counter))
779 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
782 data.next = blk_rq_timeout(round_jiffies_up(data.next));
783 mod_timer(&q->timeout, data.next);
785 struct blk_mq_hw_ctx *hctx;
787 queue_for_each_hw_ctx(q, hctx, i) {
788 /* the hctx may be unmapped, so check it here */
789 if (blk_mq_hw_queue_mapped(hctx))
790 blk_mq_tag_idle(hctx);
796 struct flush_busy_ctx_data {
797 struct blk_mq_hw_ctx *hctx;
798 struct list_head *list;
801 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
803 struct flush_busy_ctx_data *flush_data = data;
804 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
805 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
807 sbitmap_clear_bit(sb, bitnr);
808 spin_lock(&ctx->lock);
809 list_splice_tail_init(&ctx->rq_list, flush_data->list);
810 spin_unlock(&ctx->lock);
815 * Process software queues that have been marked busy, splicing them
816 * to the for-dispatch
818 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
820 struct flush_busy_ctx_data data = {
825 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
827 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
829 static inline unsigned int queued_to_index(unsigned int queued)
834 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
837 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
840 struct blk_mq_alloc_data data = {
842 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
843 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
846 might_sleep_if(wait);
851 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
852 data.flags |= BLK_MQ_REQ_RESERVED;
854 rq->tag = blk_mq_get_tag(&data);
856 if (blk_mq_tag_busy(data.hctx)) {
857 rq->rq_flags |= RQF_MQ_INFLIGHT;
858 atomic_inc(&data.hctx->nr_active);
860 data.hctx->tags->rqs[rq->tag] = rq;
866 return rq->tag != -1;
869 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
872 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
875 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
876 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
877 atomic_dec(&hctx->nr_active);
881 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
884 if (rq->tag == -1 || rq->internal_tag == -1)
887 __blk_mq_put_driver_tag(hctx, rq);
890 static void blk_mq_put_driver_tag(struct request *rq)
892 struct blk_mq_hw_ctx *hctx;
894 if (rq->tag == -1 || rq->internal_tag == -1)
897 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
898 __blk_mq_put_driver_tag(hctx, rq);
902 * If we fail getting a driver tag because all the driver tags are already
903 * assigned and on the dispatch list, BUT the first entry does not have a
904 * tag, then we could deadlock. For that case, move entries with assigned
905 * driver tags to the front, leaving the set of tagged requests in the
906 * same order, and the untagged set in the same order.
908 static bool reorder_tags_to_front(struct list_head *list)
910 struct request *rq, *tmp, *first = NULL;
912 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
916 list_move(&rq->queuelist, list);
922 return first != NULL;
925 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
928 struct blk_mq_hw_ctx *hctx;
930 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
932 list_del(&wait->task_list);
933 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
934 blk_mq_run_hw_queue(hctx, true);
938 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
940 struct sbq_wait_state *ws;
943 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
944 * The thread which wins the race to grab this bit adds the hardware
945 * queue to the wait queue.
947 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
948 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
951 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
952 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
955 * As soon as this returns, it's no longer safe to fiddle with
956 * hctx->dispatch_wait, since a completion can wake up the wait queue
957 * and unlock the bit.
959 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
963 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
965 struct blk_mq_hw_ctx *hctx;
969 if (list_empty(list))
973 * Now process all the entries, sending them to the driver.
977 struct blk_mq_queue_data bd;
980 rq = list_first_entry(list, struct request, queuelist);
981 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
982 if (!queued && reorder_tags_to_front(list))
986 * The initial allocation attempt failed, so we need to
987 * rerun the hardware queue when a tag is freed.
989 if (!blk_mq_dispatch_wait_add(hctx))
993 * It's possible that a tag was freed in the window
994 * between the allocation failure and adding the
995 * hardware queue to the wait queue.
997 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1001 list_del_init(&rq->queuelist);
1006 * Flag last if we have no more requests, or if we have more
1007 * but can't assign a driver tag to it.
1009 if (list_empty(list))
1012 struct request *nxt;
1014 nxt = list_first_entry(list, struct request, queuelist);
1015 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1018 ret = q->mq_ops->queue_rq(hctx, &bd);
1019 if (ret == BLK_STS_RESOURCE) {
1020 blk_mq_put_driver_tag_hctx(hctx, rq);
1021 list_add(&rq->queuelist, list);
1022 __blk_mq_requeue_request(rq);
1026 if (unlikely(ret != BLK_STS_OK)) {
1028 blk_mq_end_request(rq, BLK_STS_IOERR);
1033 } while (!list_empty(list));
1035 hctx->dispatched[queued_to_index(queued)]++;
1038 * Any items that need requeuing? Stuff them into hctx->dispatch,
1039 * that is where we will continue on next queue run.
1041 if (!list_empty(list)) {
1043 * If an I/O scheduler has been configured and we got a driver
1044 * tag for the next request already, free it again.
1046 rq = list_first_entry(list, struct request, queuelist);
1047 blk_mq_put_driver_tag(rq);
1049 spin_lock(&hctx->lock);
1050 list_splice_init(list, &hctx->dispatch);
1051 spin_unlock(&hctx->lock);
1054 * If SCHED_RESTART was set by the caller of this function and
1055 * it is no longer set that means that it was cleared by another
1056 * thread and hence that a queue rerun is needed.
1058 * If TAG_WAITING is set that means that an I/O scheduler has
1059 * been configured and another thread is waiting for a driver
1060 * tag. To guarantee fairness, do not rerun this hardware queue
1061 * but let the other thread grab the driver tag.
1063 * If no I/O scheduler has been configured it is possible that
1064 * the hardware queue got stopped and restarted before requests
1065 * were pushed back onto the dispatch list. Rerun the queue to
1066 * avoid starvation. Notes:
1067 * - blk_mq_run_hw_queue() checks whether or not a queue has
1068 * been stopped before rerunning a queue.
1069 * - Some but not all block drivers stop a queue before
1070 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1073 if (!blk_mq_sched_needs_restart(hctx) &&
1074 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1075 blk_mq_run_hw_queue(hctx, true);
1078 return (queued + errors) != 0;
1081 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1085 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1086 cpu_online(hctx->next_cpu));
1088 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1090 blk_mq_sched_dispatch_requests(hctx);
1095 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1096 blk_mq_sched_dispatch_requests(hctx);
1097 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1102 * It'd be great if the workqueue API had a way to pass
1103 * in a mask and had some smarts for more clever placement.
1104 * For now we just round-robin here, switching for every
1105 * BLK_MQ_CPU_WORK_BATCH queued items.
1107 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1109 if (hctx->queue->nr_hw_queues == 1)
1110 return WORK_CPU_UNBOUND;
1112 if (--hctx->next_cpu_batch <= 0) {
1115 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1116 if (next_cpu >= nr_cpu_ids)
1117 next_cpu = cpumask_first(hctx->cpumask);
1119 hctx->next_cpu = next_cpu;
1120 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1123 return hctx->next_cpu;
1126 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1127 unsigned long msecs)
1129 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1130 !blk_mq_hw_queue_mapped(hctx)))
1133 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1134 int cpu = get_cpu();
1135 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1136 __blk_mq_run_hw_queue(hctx);
1144 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1146 msecs_to_jiffies(msecs));
1149 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1151 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1153 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1155 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1157 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1159 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1161 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1163 struct blk_mq_hw_ctx *hctx;
1166 queue_for_each_hw_ctx(q, hctx, i) {
1167 if (!blk_mq_hctx_has_pending(hctx) ||
1168 blk_mq_hctx_stopped(hctx))
1171 blk_mq_run_hw_queue(hctx, async);
1174 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1177 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1178 * @q: request queue.
1180 * The caller is responsible for serializing this function against
1181 * blk_mq_{start,stop}_hw_queue().
1183 bool blk_mq_queue_stopped(struct request_queue *q)
1185 struct blk_mq_hw_ctx *hctx;
1188 queue_for_each_hw_ctx(q, hctx, i)
1189 if (blk_mq_hctx_stopped(hctx))
1194 EXPORT_SYMBOL(blk_mq_queue_stopped);
1197 * This function is often used for pausing .queue_rq() by driver when
1198 * there isn't enough resource or some conditions aren't satisfied, and
1199 * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1201 * We do not guarantee that dispatch can be drained or blocked
1202 * after blk_mq_stop_hw_queue() returns. Please use
1203 * blk_mq_quiesce_queue() for that requirement.
1205 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1207 cancel_delayed_work(&hctx->run_work);
1209 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1211 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1214 * This function is often used for pausing .queue_rq() by driver when
1215 * there isn't enough resource or some conditions aren't satisfied, and
1216 * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1218 * We do not guarantee that dispatch can be drained or blocked
1219 * after blk_mq_stop_hw_queues() returns. Please use
1220 * blk_mq_quiesce_queue() for that requirement.
1222 void blk_mq_stop_hw_queues(struct request_queue *q)
1224 struct blk_mq_hw_ctx *hctx;
1227 queue_for_each_hw_ctx(q, hctx, i)
1228 blk_mq_stop_hw_queue(hctx);
1230 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1232 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1234 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1236 blk_mq_run_hw_queue(hctx, false);
1238 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1240 void blk_mq_start_hw_queues(struct request_queue *q)
1242 struct blk_mq_hw_ctx *hctx;
1245 queue_for_each_hw_ctx(q, hctx, i)
1246 blk_mq_start_hw_queue(hctx);
1248 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1250 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1252 if (!blk_mq_hctx_stopped(hctx))
1255 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1256 blk_mq_run_hw_queue(hctx, async);
1258 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1260 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1262 struct blk_mq_hw_ctx *hctx;
1265 queue_for_each_hw_ctx(q, hctx, i)
1266 blk_mq_start_stopped_hw_queue(hctx, async);
1268 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1270 static void blk_mq_run_work_fn(struct work_struct *work)
1272 struct blk_mq_hw_ctx *hctx;
1274 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1277 * If we are stopped, don't run the queue. The exception is if
1278 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1279 * the STOPPED bit and run it.
1281 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1282 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1285 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1286 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1289 __blk_mq_run_hw_queue(hctx);
1293 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1295 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1299 * Stop the hw queue, then modify currently delayed work.
1300 * This should prevent us from running the queue prematurely.
1301 * Mark the queue as auto-clearing STOPPED when it runs.
1303 blk_mq_stop_hw_queue(hctx);
1304 set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1305 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1307 msecs_to_jiffies(msecs));
1309 EXPORT_SYMBOL(blk_mq_delay_queue);
1311 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1315 struct blk_mq_ctx *ctx = rq->mq_ctx;
1317 trace_block_rq_insert(hctx->queue, rq);
1320 list_add(&rq->queuelist, &ctx->rq_list);
1322 list_add_tail(&rq->queuelist, &ctx->rq_list);
1325 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1328 struct blk_mq_ctx *ctx = rq->mq_ctx;
1330 __blk_mq_insert_req_list(hctx, rq, at_head);
1331 blk_mq_hctx_mark_pending(hctx, ctx);
1334 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1335 struct list_head *list)
1339 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1342 spin_lock(&ctx->lock);
1343 while (!list_empty(list)) {
1346 rq = list_first_entry(list, struct request, queuelist);
1347 BUG_ON(rq->mq_ctx != ctx);
1348 list_del_init(&rq->queuelist);
1349 __blk_mq_insert_req_list(hctx, rq, false);
1351 blk_mq_hctx_mark_pending(hctx, ctx);
1352 spin_unlock(&ctx->lock);
1355 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1357 struct request *rqa = container_of(a, struct request, queuelist);
1358 struct request *rqb = container_of(b, struct request, queuelist);
1360 return !(rqa->mq_ctx < rqb->mq_ctx ||
1361 (rqa->mq_ctx == rqb->mq_ctx &&
1362 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1365 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1367 struct blk_mq_ctx *this_ctx;
1368 struct request_queue *this_q;
1371 LIST_HEAD(ctx_list);
1374 list_splice_init(&plug->mq_list, &list);
1376 list_sort(NULL, &list, plug_ctx_cmp);
1382 while (!list_empty(&list)) {
1383 rq = list_entry_rq(list.next);
1384 list_del_init(&rq->queuelist);
1386 if (rq->mq_ctx != this_ctx) {
1388 trace_block_unplug(this_q, depth, from_schedule);
1389 blk_mq_sched_insert_requests(this_q, this_ctx,
1394 this_ctx = rq->mq_ctx;
1400 list_add_tail(&rq->queuelist, &ctx_list);
1404 * If 'this_ctx' is set, we know we have entries to complete
1405 * on 'ctx_list'. Do those.
1408 trace_block_unplug(this_q, depth, from_schedule);
1409 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1414 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1416 blk_init_request_from_bio(rq, bio);
1418 blk_account_io_start(rq, true);
1421 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1423 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1424 !blk_queue_nomerges(hctx->queue);
1427 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1428 struct blk_mq_ctx *ctx,
1431 spin_lock(&ctx->lock);
1432 __blk_mq_insert_request(hctx, rq, false);
1433 spin_unlock(&ctx->lock);
1436 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1439 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1441 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1444 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1446 blk_qc_t *cookie, bool may_sleep)
1448 struct request_queue *q = rq->q;
1449 struct blk_mq_queue_data bd = {
1453 blk_qc_t new_cookie;
1455 bool run_queue = true;
1457 /* RCU or SRCU read lock is needed before checking quiesced flag */
1458 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1466 if (!blk_mq_get_driver_tag(rq, NULL, false))
1469 new_cookie = request_to_qc_t(hctx, rq);
1472 * For OK queue, we are done. For error, kill it. Any other
1473 * error (busy), just add it to our list as we previously
1476 ret = q->mq_ops->queue_rq(hctx, &bd);
1479 *cookie = new_cookie;
1481 case BLK_STS_RESOURCE:
1482 __blk_mq_requeue_request(rq);
1485 *cookie = BLK_QC_T_NONE;
1486 blk_mq_end_request(rq, ret);
1491 blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1494 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1495 struct request *rq, blk_qc_t *cookie)
1497 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1499 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1502 unsigned int srcu_idx;
1506 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1507 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1508 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1512 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1514 const int is_sync = op_is_sync(bio->bi_opf);
1515 const int is_flush_fua = op_is_flush(bio->bi_opf);
1516 struct blk_mq_alloc_data data = { .flags = 0 };
1518 unsigned int request_count = 0;
1519 struct blk_plug *plug;
1520 struct request *same_queue_rq = NULL;
1522 unsigned int wb_acct;
1524 blk_queue_bounce(q, &bio);
1526 blk_queue_split(q, &bio);
1528 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1530 return BLK_QC_T_NONE;
1533 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1534 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1535 return BLK_QC_T_NONE;
1537 if (blk_mq_sched_bio_merge(q, bio))
1538 return BLK_QC_T_NONE;
1540 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1542 trace_block_getrq(q, bio, bio->bi_opf);
1544 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1545 if (unlikely(!rq)) {
1546 __wbt_done(q->rq_wb, wb_acct);
1547 return BLK_QC_T_NONE;
1550 wbt_track(&rq->issue_stat, wb_acct);
1552 cookie = request_to_qc_t(data.hctx, rq);
1554 plug = current->plug;
1555 if (unlikely(is_flush_fua)) {
1556 blk_mq_put_ctx(data.ctx);
1557 blk_mq_bio_to_request(rq, bio);
1559 blk_mq_sched_insert_request(rq, false, true, true,
1562 blk_insert_flush(rq);
1563 blk_mq_run_hw_queue(data.hctx, true);
1565 } else if (plug && q->nr_hw_queues == 1) {
1566 struct request *last = NULL;
1568 blk_mq_put_ctx(data.ctx);
1569 blk_mq_bio_to_request(rq, bio);
1572 * @request_count may become stale because of schedule
1573 * out, so check the list again.
1575 if (list_empty(&plug->mq_list))
1577 else if (blk_queue_nomerges(q))
1578 request_count = blk_plug_queued_count(q);
1581 trace_block_plug(q);
1583 last = list_entry_rq(plug->mq_list.prev);
1585 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1586 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1587 blk_flush_plug_list(plug, false);
1588 trace_block_plug(q);
1591 list_add_tail(&rq->queuelist, &plug->mq_list);
1592 } else if (plug && !blk_queue_nomerges(q)) {
1593 blk_mq_bio_to_request(rq, bio);
1596 * We do limited plugging. If the bio can be merged, do that.
1597 * Otherwise the existing request in the plug list will be
1598 * issued. So the plug list will have one request at most
1599 * The plug list might get flushed before this. If that happens,
1600 * the plug list is empty, and same_queue_rq is invalid.
1602 if (list_empty(&plug->mq_list))
1603 same_queue_rq = NULL;
1605 list_del_init(&same_queue_rq->queuelist);
1606 list_add_tail(&rq->queuelist, &plug->mq_list);
1608 blk_mq_put_ctx(data.ctx);
1610 if (same_queue_rq) {
1611 data.hctx = blk_mq_map_queue(q,
1612 same_queue_rq->mq_ctx->cpu);
1613 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1616 } else if (q->nr_hw_queues > 1 && is_sync) {
1617 blk_mq_put_ctx(data.ctx);
1618 blk_mq_bio_to_request(rq, bio);
1619 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1620 } else if (q->elevator) {
1621 blk_mq_put_ctx(data.ctx);
1622 blk_mq_bio_to_request(rq, bio);
1623 blk_mq_sched_insert_request(rq, false, true, true, true);
1625 blk_mq_put_ctx(data.ctx);
1626 blk_mq_bio_to_request(rq, bio);
1627 blk_mq_queue_io(data.hctx, data.ctx, rq);
1628 blk_mq_run_hw_queue(data.hctx, true);
1634 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1635 unsigned int hctx_idx)
1639 if (tags->rqs && set->ops->exit_request) {
1642 for (i = 0; i < tags->nr_tags; i++) {
1643 struct request *rq = tags->static_rqs[i];
1647 set->ops->exit_request(set, rq, hctx_idx);
1648 tags->static_rqs[i] = NULL;
1652 while (!list_empty(&tags->page_list)) {
1653 page = list_first_entry(&tags->page_list, struct page, lru);
1654 list_del_init(&page->lru);
1656 * Remove kmemleak object previously allocated in
1657 * blk_mq_init_rq_map().
1659 kmemleak_free(page_address(page));
1660 __free_pages(page, page->private);
1664 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1668 kfree(tags->static_rqs);
1669 tags->static_rqs = NULL;
1671 blk_mq_free_tags(tags);
1674 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1675 unsigned int hctx_idx,
1676 unsigned int nr_tags,
1677 unsigned int reserved_tags)
1679 struct blk_mq_tags *tags;
1682 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1683 if (node == NUMA_NO_NODE)
1684 node = set->numa_node;
1686 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1687 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1691 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1692 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1695 blk_mq_free_tags(tags);
1699 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1700 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1702 if (!tags->static_rqs) {
1704 blk_mq_free_tags(tags);
1711 static size_t order_to_size(unsigned int order)
1713 return (size_t)PAGE_SIZE << order;
1716 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1717 unsigned int hctx_idx, unsigned int depth)
1719 unsigned int i, j, entries_per_page, max_order = 4;
1720 size_t rq_size, left;
1723 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1724 if (node == NUMA_NO_NODE)
1725 node = set->numa_node;
1727 INIT_LIST_HEAD(&tags->page_list);
1730 * rq_size is the size of the request plus driver payload, rounded
1731 * to the cacheline size
1733 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1735 left = rq_size * depth;
1737 for (i = 0; i < depth; ) {
1738 int this_order = max_order;
1743 while (this_order && left < order_to_size(this_order - 1))
1747 page = alloc_pages_node(node,
1748 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1754 if (order_to_size(this_order) < rq_size)
1761 page->private = this_order;
1762 list_add_tail(&page->lru, &tags->page_list);
1764 p = page_address(page);
1766 * Allow kmemleak to scan these pages as they contain pointers
1767 * to additional allocations like via ops->init_request().
1769 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1770 entries_per_page = order_to_size(this_order) / rq_size;
1771 to_do = min(entries_per_page, depth - i);
1772 left -= to_do * rq_size;
1773 for (j = 0; j < to_do; j++) {
1774 struct request *rq = p;
1776 tags->static_rqs[i] = rq;
1777 if (set->ops->init_request) {
1778 if (set->ops->init_request(set, rq, hctx_idx,
1780 tags->static_rqs[i] = NULL;
1792 blk_mq_free_rqs(set, tags, hctx_idx);
1797 * 'cpu' is going away. splice any existing rq_list entries from this
1798 * software queue to the hw queue dispatch list, and ensure that it
1801 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1803 struct blk_mq_hw_ctx *hctx;
1804 struct blk_mq_ctx *ctx;
1807 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1808 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1810 spin_lock(&ctx->lock);
1811 if (!list_empty(&ctx->rq_list)) {
1812 list_splice_init(&ctx->rq_list, &tmp);
1813 blk_mq_hctx_clear_pending(hctx, ctx);
1815 spin_unlock(&ctx->lock);
1817 if (list_empty(&tmp))
1820 spin_lock(&hctx->lock);
1821 list_splice_tail_init(&tmp, &hctx->dispatch);
1822 spin_unlock(&hctx->lock);
1824 blk_mq_run_hw_queue(hctx, true);
1828 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1830 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1834 /* hctx->ctxs will be freed in queue's release handler */
1835 static void blk_mq_exit_hctx(struct request_queue *q,
1836 struct blk_mq_tag_set *set,
1837 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1839 blk_mq_debugfs_unregister_hctx(hctx);
1841 blk_mq_tag_idle(hctx);
1843 if (set->ops->exit_request)
1844 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1846 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1848 if (set->ops->exit_hctx)
1849 set->ops->exit_hctx(hctx, hctx_idx);
1851 if (hctx->flags & BLK_MQ_F_BLOCKING)
1852 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1854 blk_mq_remove_cpuhp(hctx);
1855 blk_free_flush_queue(hctx->fq);
1856 sbitmap_free(&hctx->ctx_map);
1859 static void blk_mq_exit_hw_queues(struct request_queue *q,
1860 struct blk_mq_tag_set *set, int nr_queue)
1862 struct blk_mq_hw_ctx *hctx;
1865 queue_for_each_hw_ctx(q, hctx, i) {
1868 blk_mq_exit_hctx(q, set, hctx, i);
1872 static int blk_mq_init_hctx(struct request_queue *q,
1873 struct blk_mq_tag_set *set,
1874 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1878 node = hctx->numa_node;
1879 if (node == NUMA_NO_NODE)
1880 node = hctx->numa_node = set->numa_node;
1882 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1883 spin_lock_init(&hctx->lock);
1884 INIT_LIST_HEAD(&hctx->dispatch);
1886 hctx->queue_num = hctx_idx;
1887 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1889 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1891 hctx->tags = set->tags[hctx_idx];
1894 * Allocate space for all possible cpus to avoid allocation at
1897 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1900 goto unregister_cpu_notifier;
1902 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1908 if (set->ops->init_hctx &&
1909 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1912 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1915 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1917 goto sched_exit_hctx;
1919 if (set->ops->init_request &&
1920 set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1924 if (hctx->flags & BLK_MQ_F_BLOCKING)
1925 init_srcu_struct(&hctx->queue_rq_srcu);
1927 blk_mq_debugfs_register_hctx(q, hctx);
1934 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1936 if (set->ops->exit_hctx)
1937 set->ops->exit_hctx(hctx, hctx_idx);
1939 sbitmap_free(&hctx->ctx_map);
1942 unregister_cpu_notifier:
1943 blk_mq_remove_cpuhp(hctx);
1947 static void blk_mq_init_cpu_queues(struct request_queue *q,
1948 unsigned int nr_hw_queues)
1952 for_each_possible_cpu(i) {
1953 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1954 struct blk_mq_hw_ctx *hctx;
1957 spin_lock_init(&__ctx->lock);
1958 INIT_LIST_HEAD(&__ctx->rq_list);
1961 /* If the cpu isn't online, the cpu is mapped to first hctx */
1965 hctx = blk_mq_map_queue(q, i);
1968 * Set local node, IFF we have more than one hw queue. If
1969 * not, we remain on the home node of the device
1971 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1972 hctx->numa_node = local_memory_node(cpu_to_node(i));
1976 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1980 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1981 set->queue_depth, set->reserved_tags);
1982 if (!set->tags[hctx_idx])
1985 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
1990 blk_mq_free_rq_map(set->tags[hctx_idx]);
1991 set->tags[hctx_idx] = NULL;
1995 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
1996 unsigned int hctx_idx)
1998 if (set->tags[hctx_idx]) {
1999 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2000 blk_mq_free_rq_map(set->tags[hctx_idx]);
2001 set->tags[hctx_idx] = NULL;
2005 static void blk_mq_map_swqueue(struct request_queue *q,
2006 const struct cpumask *online_mask)
2008 unsigned int i, hctx_idx;
2009 struct blk_mq_hw_ctx *hctx;
2010 struct blk_mq_ctx *ctx;
2011 struct blk_mq_tag_set *set = q->tag_set;
2014 * Avoid others reading imcomplete hctx->cpumask through sysfs
2016 mutex_lock(&q->sysfs_lock);
2018 queue_for_each_hw_ctx(q, hctx, i) {
2019 cpumask_clear(hctx->cpumask);
2024 * Map software to hardware queues
2026 for_each_possible_cpu(i) {
2027 /* If the cpu isn't online, the cpu is mapped to first hctx */
2028 if (!cpumask_test_cpu(i, online_mask))
2031 hctx_idx = q->mq_map[i];
2032 /* unmapped hw queue can be remapped after CPU topo changed */
2033 if (!set->tags[hctx_idx] &&
2034 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2036 * If tags initialization fail for some hctx,
2037 * that hctx won't be brought online. In this
2038 * case, remap the current ctx to hctx[0] which
2039 * is guaranteed to always have tags allocated
2044 ctx = per_cpu_ptr(q->queue_ctx, i);
2045 hctx = blk_mq_map_queue(q, i);
2047 cpumask_set_cpu(i, hctx->cpumask);
2048 ctx->index_hw = hctx->nr_ctx;
2049 hctx->ctxs[hctx->nr_ctx++] = ctx;
2052 mutex_unlock(&q->sysfs_lock);
2054 queue_for_each_hw_ctx(q, hctx, i) {
2056 * If no software queues are mapped to this hardware queue,
2057 * disable it and free the request entries.
2059 if (!hctx->nr_ctx) {
2060 /* Never unmap queue 0. We need it as a
2061 * fallback in case of a new remap fails
2064 if (i && set->tags[i])
2065 blk_mq_free_map_and_requests(set, i);
2071 hctx->tags = set->tags[i];
2072 WARN_ON(!hctx->tags);
2075 * Set the map size to the number of mapped software queues.
2076 * This is more accurate and more efficient than looping
2077 * over all possibly mapped software queues.
2079 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2082 * Initialize batch roundrobin counts
2084 hctx->next_cpu = cpumask_first(hctx->cpumask);
2085 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2089 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2091 struct blk_mq_hw_ctx *hctx;
2094 queue_for_each_hw_ctx(q, hctx, i) {
2096 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2098 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2102 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2104 struct request_queue *q;
2106 lockdep_assert_held(&set->tag_list_lock);
2108 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2109 blk_mq_freeze_queue(q);
2110 queue_set_hctx_shared(q, shared);
2111 blk_mq_unfreeze_queue(q);
2115 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2117 struct blk_mq_tag_set *set = q->tag_set;
2119 mutex_lock(&set->tag_list_lock);
2120 list_del_rcu(&q->tag_set_list);
2121 INIT_LIST_HEAD(&q->tag_set_list);
2122 if (list_is_singular(&set->tag_list)) {
2123 /* just transitioned to unshared */
2124 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2125 /* update existing queue */
2126 blk_mq_update_tag_set_depth(set, false);
2128 mutex_unlock(&set->tag_list_lock);
2133 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2134 struct request_queue *q)
2138 mutex_lock(&set->tag_list_lock);
2140 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2141 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2142 set->flags |= BLK_MQ_F_TAG_SHARED;
2143 /* update existing queue */
2144 blk_mq_update_tag_set_depth(set, true);
2146 if (set->flags & BLK_MQ_F_TAG_SHARED)
2147 queue_set_hctx_shared(q, true);
2148 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2150 mutex_unlock(&set->tag_list_lock);
2154 * It is the actual release handler for mq, but we do it from
2155 * request queue's release handler for avoiding use-after-free
2156 * and headache because q->mq_kobj shouldn't have been introduced,
2157 * but we can't group ctx/kctx kobj without it.
2159 void blk_mq_release(struct request_queue *q)
2161 struct blk_mq_hw_ctx *hctx;
2164 /* hctx kobj stays in hctx */
2165 queue_for_each_hw_ctx(q, hctx, i) {
2168 kobject_put(&hctx->kobj);
2173 kfree(q->queue_hw_ctx);
2176 * release .mq_kobj and sw queue's kobject now because
2177 * both share lifetime with request queue.
2179 blk_mq_sysfs_deinit(q);
2181 free_percpu(q->queue_ctx);
2184 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2186 struct request_queue *uninit_q, *q;
2188 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2190 return ERR_PTR(-ENOMEM);
2192 q = blk_mq_init_allocated_queue(set, uninit_q);
2194 blk_cleanup_queue(uninit_q);
2198 EXPORT_SYMBOL(blk_mq_init_queue);
2200 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2201 struct request_queue *q)
2204 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2206 blk_mq_sysfs_unregister(q);
2207 for (i = 0; i < set->nr_hw_queues; i++) {
2213 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2214 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2219 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2226 atomic_set(&hctxs[i]->nr_active, 0);
2227 hctxs[i]->numa_node = node;
2228 hctxs[i]->queue_num = i;
2230 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2231 free_cpumask_var(hctxs[i]->cpumask);
2236 blk_mq_hctx_kobj_init(hctxs[i]);
2238 for (j = i; j < q->nr_hw_queues; j++) {
2239 struct blk_mq_hw_ctx *hctx = hctxs[j];
2243 blk_mq_free_map_and_requests(set, j);
2244 blk_mq_exit_hctx(q, set, hctx, j);
2245 kobject_put(&hctx->kobj);
2250 q->nr_hw_queues = i;
2251 blk_mq_sysfs_register(q);
2254 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2255 struct request_queue *q)
2257 /* mark the queue as mq asap */
2258 q->mq_ops = set->ops;
2260 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2261 blk_mq_poll_stats_bkt,
2262 BLK_MQ_POLL_STATS_BKTS, q);
2266 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2270 /* init q->mq_kobj and sw queues' kobjects */
2271 blk_mq_sysfs_init(q);
2273 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2274 GFP_KERNEL, set->numa_node);
2275 if (!q->queue_hw_ctx)
2278 q->mq_map = set->mq_map;
2280 blk_mq_realloc_hw_ctxs(set, q);
2281 if (!q->nr_hw_queues)
2284 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2285 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2287 q->nr_queues = nr_cpu_ids;
2289 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2291 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2292 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2294 q->sg_reserved_size = INT_MAX;
2296 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2297 INIT_LIST_HEAD(&q->requeue_list);
2298 spin_lock_init(&q->requeue_lock);
2300 blk_queue_make_request(q, blk_mq_make_request);
2303 * Do this after blk_queue_make_request() overrides it...
2305 q->nr_requests = set->queue_depth;
2308 * Default to classic polling
2312 if (set->ops->complete)
2313 blk_queue_softirq_done(q, set->ops->complete);
2315 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2318 mutex_lock(&all_q_mutex);
2320 list_add_tail(&q->all_q_node, &all_q_list);
2321 blk_mq_add_queue_tag_set(set, q);
2322 blk_mq_map_swqueue(q, cpu_online_mask);
2324 mutex_unlock(&all_q_mutex);
2327 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2330 ret = blk_mq_sched_init(q);
2332 return ERR_PTR(ret);
2338 kfree(q->queue_hw_ctx);
2340 free_percpu(q->queue_ctx);
2343 return ERR_PTR(-ENOMEM);
2345 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2347 void blk_mq_free_queue(struct request_queue *q)
2349 struct blk_mq_tag_set *set = q->tag_set;
2351 mutex_lock(&all_q_mutex);
2352 list_del_init(&q->all_q_node);
2353 mutex_unlock(&all_q_mutex);
2355 blk_mq_del_queue_tag_set(q);
2357 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2360 /* Basically redo blk_mq_init_queue with queue frozen */
2361 static void blk_mq_queue_reinit(struct request_queue *q,
2362 const struct cpumask *online_mask)
2364 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2366 blk_mq_debugfs_unregister_hctxs(q);
2367 blk_mq_sysfs_unregister(q);
2370 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2371 * we should change hctx numa_node according to new topology (this
2372 * involves free and re-allocate memory, worthy doing?)
2375 blk_mq_map_swqueue(q, online_mask);
2377 blk_mq_sysfs_register(q);
2378 blk_mq_debugfs_register_hctxs(q);
2382 * New online cpumask which is going to be set in this hotplug event.
2383 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2384 * one-by-one and dynamically allocating this could result in a failure.
2386 static struct cpumask cpuhp_online_new;
2388 static void blk_mq_queue_reinit_work(void)
2390 struct request_queue *q;
2392 mutex_lock(&all_q_mutex);
2394 * We need to freeze and reinit all existing queues. Freezing
2395 * involves synchronous wait for an RCU grace period and doing it
2396 * one by one may take a long time. Start freezing all queues in
2397 * one swoop and then wait for the completions so that freezing can
2398 * take place in parallel.
2400 list_for_each_entry(q, &all_q_list, all_q_node)
2401 blk_freeze_queue_start(q);
2402 list_for_each_entry(q, &all_q_list, all_q_node)
2403 blk_mq_freeze_queue_wait(q);
2405 list_for_each_entry(q, &all_q_list, all_q_node)
2406 blk_mq_queue_reinit(q, &cpuhp_online_new);
2408 list_for_each_entry(q, &all_q_list, all_q_node)
2409 blk_mq_unfreeze_queue(q);
2411 mutex_unlock(&all_q_mutex);
2414 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2416 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2417 blk_mq_queue_reinit_work();
2422 * Before hotadded cpu starts handling requests, new mappings must be
2423 * established. Otherwise, these requests in hw queue might never be
2426 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2427 * for CPU0, and ctx1 for CPU1).
2429 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2430 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2432 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2433 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2434 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2437 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2439 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2440 cpumask_set_cpu(cpu, &cpuhp_online_new);
2441 blk_mq_queue_reinit_work();
2445 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2449 for (i = 0; i < set->nr_hw_queues; i++)
2450 if (!__blk_mq_alloc_rq_map(set, i))
2457 blk_mq_free_rq_map(set->tags[i]);
2463 * Allocate the request maps associated with this tag_set. Note that this
2464 * may reduce the depth asked for, if memory is tight. set->queue_depth
2465 * will be updated to reflect the allocated depth.
2467 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2472 depth = set->queue_depth;
2474 err = __blk_mq_alloc_rq_maps(set);
2478 set->queue_depth >>= 1;
2479 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2483 } while (set->queue_depth);
2485 if (!set->queue_depth || err) {
2486 pr_err("blk-mq: failed to allocate request map\n");
2490 if (depth != set->queue_depth)
2491 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2492 depth, set->queue_depth);
2497 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2499 if (set->ops->map_queues)
2500 return set->ops->map_queues(set);
2502 return blk_mq_map_queues(set);
2506 * Alloc a tag set to be associated with one or more request queues.
2507 * May fail with EINVAL for various error conditions. May adjust the
2508 * requested depth down, if if it too large. In that case, the set
2509 * value will be stored in set->queue_depth.
2511 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2515 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2517 if (!set->nr_hw_queues)
2519 if (!set->queue_depth)
2521 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2524 if (!set->ops->queue_rq)
2527 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2528 pr_info("blk-mq: reduced tag depth to %u\n",
2530 set->queue_depth = BLK_MQ_MAX_DEPTH;
2534 * If a crashdump is active, then we are potentially in a very
2535 * memory constrained environment. Limit us to 1 queue and
2536 * 64 tags to prevent using too much memory.
2538 if (is_kdump_kernel()) {
2539 set->nr_hw_queues = 1;
2540 set->queue_depth = min(64U, set->queue_depth);
2543 * There is no use for more h/w queues than cpus.
2545 if (set->nr_hw_queues > nr_cpu_ids)
2546 set->nr_hw_queues = nr_cpu_ids;
2548 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2549 GFP_KERNEL, set->numa_node);
2554 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2555 GFP_KERNEL, set->numa_node);
2559 ret = blk_mq_update_queue_map(set);
2561 goto out_free_mq_map;
2563 ret = blk_mq_alloc_rq_maps(set);
2565 goto out_free_mq_map;
2567 mutex_init(&set->tag_list_lock);
2568 INIT_LIST_HEAD(&set->tag_list);
2580 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2582 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2586 for (i = 0; i < nr_cpu_ids; i++)
2587 blk_mq_free_map_and_requests(set, i);
2595 EXPORT_SYMBOL(blk_mq_free_tag_set);
2597 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2599 struct blk_mq_tag_set *set = q->tag_set;
2600 struct blk_mq_hw_ctx *hctx;
2606 blk_mq_freeze_queue(q);
2609 queue_for_each_hw_ctx(q, hctx, i) {
2613 * If we're using an MQ scheduler, just update the scheduler
2614 * queue depth. This is similar to what the old code would do.
2616 if (!hctx->sched_tags) {
2617 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2618 min(nr, set->queue_depth),
2621 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2629 q->nr_requests = nr;
2631 blk_mq_unfreeze_queue(q);
2636 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2639 struct request_queue *q;
2641 lockdep_assert_held(&set->tag_list_lock);
2643 if (nr_hw_queues > nr_cpu_ids)
2644 nr_hw_queues = nr_cpu_ids;
2645 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2648 list_for_each_entry(q, &set->tag_list, tag_set_list)
2649 blk_mq_freeze_queue(q);
2651 set->nr_hw_queues = nr_hw_queues;
2652 blk_mq_update_queue_map(set);
2653 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2654 blk_mq_realloc_hw_ctxs(set, q);
2655 blk_mq_queue_reinit(q, cpu_online_mask);
2658 list_for_each_entry(q, &set->tag_list, tag_set_list)
2659 blk_mq_unfreeze_queue(q);
2662 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2664 mutex_lock(&set->tag_list_lock);
2665 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2666 mutex_unlock(&set->tag_list_lock);
2668 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2670 /* Enable polling stats and return whether they were already enabled. */
2671 static bool blk_poll_stats_enable(struct request_queue *q)
2673 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2674 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2676 blk_stat_add_callback(q, q->poll_cb);
2680 static void blk_mq_poll_stats_start(struct request_queue *q)
2683 * We don't arm the callback if polling stats are not enabled or the
2684 * callback is already active.
2686 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2687 blk_stat_is_active(q->poll_cb))
2690 blk_stat_activate_msecs(q->poll_cb, 100);
2693 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2695 struct request_queue *q = cb->data;
2698 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2699 if (cb->stat[bucket].nr_samples)
2700 q->poll_stat[bucket] = cb->stat[bucket];
2704 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2705 struct blk_mq_hw_ctx *hctx,
2708 unsigned long ret = 0;
2712 * If stats collection isn't on, don't sleep but turn it on for
2715 if (!blk_poll_stats_enable(q))
2719 * As an optimistic guess, use half of the mean service time
2720 * for this type of request. We can (and should) make this smarter.
2721 * For instance, if the completion latencies are tight, we can
2722 * get closer than just half the mean. This is especially
2723 * important on devices where the completion latencies are longer
2724 * than ~10 usec. We do use the stats for the relevant IO size
2725 * if available which does lead to better estimates.
2727 bucket = blk_mq_poll_stats_bkt(rq);
2731 if (q->poll_stat[bucket].nr_samples)
2732 ret = (q->poll_stat[bucket].mean + 1) / 2;
2737 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2738 struct blk_mq_hw_ctx *hctx,
2741 struct hrtimer_sleeper hs;
2742 enum hrtimer_mode mode;
2746 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2752 * -1: don't ever hybrid sleep
2753 * 0: use half of prev avg
2754 * >0: use this specific value
2756 if (q->poll_nsec == -1)
2758 else if (q->poll_nsec > 0)
2759 nsecs = q->poll_nsec;
2761 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2766 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2769 * This will be replaced with the stats tracking code, using
2770 * 'avg_completion_time / 2' as the pre-sleep target.
2774 mode = HRTIMER_MODE_REL;
2775 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2776 hrtimer_set_expires(&hs.timer, kt);
2778 hrtimer_init_sleeper(&hs, current);
2780 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2782 set_current_state(TASK_UNINTERRUPTIBLE);
2783 hrtimer_start_expires(&hs.timer, mode);
2786 hrtimer_cancel(&hs.timer);
2787 mode = HRTIMER_MODE_ABS;
2788 } while (hs.task && !signal_pending(current));
2790 __set_current_state(TASK_RUNNING);
2791 destroy_hrtimer_on_stack(&hs.timer);
2795 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2797 struct request_queue *q = hctx->queue;
2801 * If we sleep, have the caller restart the poll loop to reset
2802 * the state. Like for the other success return cases, the
2803 * caller is responsible for checking if the IO completed. If
2804 * the IO isn't complete, we'll get called again and will go
2805 * straight to the busy poll loop.
2807 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2810 hctx->poll_considered++;
2812 state = current->state;
2813 while (!need_resched()) {
2816 hctx->poll_invoked++;
2818 ret = q->mq_ops->poll(hctx, rq->tag);
2820 hctx->poll_success++;
2821 set_current_state(TASK_RUNNING);
2825 if (signal_pending_state(state, current))
2826 set_current_state(TASK_RUNNING);
2828 if (current->state == TASK_RUNNING)
2838 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2840 struct blk_mq_hw_ctx *hctx;
2841 struct blk_plug *plug;
2844 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2845 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2848 plug = current->plug;
2850 blk_flush_plug_list(plug, false);
2852 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2853 if (!blk_qc_t_is_internal(cookie))
2854 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2856 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2858 * With scheduling, if the request has completed, we'll
2859 * get a NULL return here, as we clear the sched tag when
2860 * that happens. The request still remains valid, like always,
2861 * so we should be safe with just the NULL check.
2867 return __blk_mq_poll(hctx, rq);
2869 EXPORT_SYMBOL_GPL(blk_mq_poll);
2871 void blk_mq_disable_hotplug(void)
2873 mutex_lock(&all_q_mutex);
2876 void blk_mq_enable_hotplug(void)
2878 mutex_unlock(&all_q_mutex);
2881 static int __init blk_mq_init(void)
2883 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2884 blk_mq_hctx_notify_dead);
2886 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2887 blk_mq_queue_reinit_prepare,
2888 blk_mq_queue_reinit_dead);
2891 subsys_initcall(blk_mq_init);