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"
39 #include "blk-rq-qos.h"
41 static void blk_mq_poll_stats_start(struct request_queue *q);
42 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
44 static int blk_mq_poll_stats_bkt(const struct request *rq)
46 int ddir, bytes, bucket;
48 ddir = rq_data_dir(rq);
49 bytes = blk_rq_bytes(rq);
51 bucket = ddir + 2*(ilog2(bytes) - 9);
55 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
56 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
62 * Check if any of the ctx, dispatch list or elevator
63 * have pending work in this hardware queue.
65 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
67 return !list_empty_careful(&hctx->dispatch) ||
68 sbitmap_any_bit_set(&hctx->ctx_map) ||
69 blk_mq_sched_has_work(hctx);
73 * Mark this ctx as having pending work in this hardware queue
75 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
76 struct blk_mq_ctx *ctx)
78 const int bit = ctx->index_hw[hctx->type];
80 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
81 sbitmap_set_bit(&hctx->ctx_map, bit);
84 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
85 struct blk_mq_ctx *ctx)
87 const int bit = ctx->index_hw[hctx->type];
89 sbitmap_clear_bit(&hctx->ctx_map, bit);
93 struct hd_struct *part;
94 unsigned int *inflight;
97 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
98 struct request *rq, void *priv,
101 struct mq_inflight *mi = priv;
104 * index[0] counts the specific partition that was asked for.
106 if (rq->part == mi->part)
112 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
114 unsigned inflight[2];
115 struct mq_inflight mi = { .part = part, .inflight = inflight, };
117 inflight[0] = inflight[1] = 0;
118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
123 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
124 struct request *rq, void *priv,
127 struct mq_inflight *mi = priv;
129 if (rq->part == mi->part)
130 mi->inflight[rq_data_dir(rq)]++;
135 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
136 unsigned int inflight[2])
138 struct mq_inflight mi = { .part = part, .inflight = inflight, };
140 inflight[0] = inflight[1] = 0;
141 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
144 void blk_freeze_queue_start(struct request_queue *q)
148 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
149 if (freeze_depth == 1) {
150 percpu_ref_kill(&q->q_usage_counter);
152 blk_mq_run_hw_queues(q, false);
155 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
157 void blk_mq_freeze_queue_wait(struct request_queue *q)
159 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
161 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
163 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
164 unsigned long timeout)
166 return wait_event_timeout(q->mq_freeze_wq,
167 percpu_ref_is_zero(&q->q_usage_counter),
170 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
173 * Guarantee no request is in use, so we can change any data structure of
174 * the queue afterward.
176 void blk_freeze_queue(struct request_queue *q)
179 * In the !blk_mq case we are only calling this to kill the
180 * q_usage_counter, otherwise this increases the freeze depth
181 * and waits for it to return to zero. For this reason there is
182 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
183 * exported to drivers as the only user for unfreeze is blk_mq.
185 blk_freeze_queue_start(q);
186 blk_mq_freeze_queue_wait(q);
189 void blk_mq_freeze_queue(struct request_queue *q)
192 * ...just an alias to keep freeze and unfreeze actions balanced
193 * in the blk_mq_* namespace
197 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
199 void blk_mq_unfreeze_queue(struct request_queue *q)
203 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
204 WARN_ON_ONCE(freeze_depth < 0);
206 percpu_ref_resurrect(&q->q_usage_counter);
207 wake_up_all(&q->mq_freeze_wq);
210 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
213 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
214 * mpt3sas driver such that this function can be removed.
216 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
218 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
220 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
223 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
226 * Note: this function does not prevent that the struct request end_io()
227 * callback function is invoked. Once this function is returned, we make
228 * sure no dispatch can happen until the queue is unquiesced via
229 * blk_mq_unquiesce_queue().
231 void blk_mq_quiesce_queue(struct request_queue *q)
233 struct blk_mq_hw_ctx *hctx;
237 blk_mq_quiesce_queue_nowait(q);
239 queue_for_each_hw_ctx(q, hctx, i) {
240 if (hctx->flags & BLK_MQ_F_BLOCKING)
241 synchronize_srcu(hctx->srcu);
248 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
251 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
254 * This function recovers queue into the state before quiescing
255 * which is done by blk_mq_quiesce_queue.
257 void blk_mq_unquiesce_queue(struct request_queue *q)
259 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
261 /* dispatch requests which are inserted during quiescing */
262 blk_mq_run_hw_queues(q, true);
264 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
266 void blk_mq_wake_waiters(struct request_queue *q)
268 struct blk_mq_hw_ctx *hctx;
271 queue_for_each_hw_ctx(q, hctx, i)
272 if (blk_mq_hw_queue_mapped(hctx))
273 blk_mq_tag_wakeup_all(hctx->tags, true);
276 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
278 return blk_mq_has_free_tags(hctx->tags);
280 EXPORT_SYMBOL(blk_mq_can_queue);
283 * Only need start/end time stamping if we have stats enabled, or using
286 static inline bool blk_mq_need_time_stamp(struct request *rq)
288 return (rq->rq_flags & RQF_IO_STAT) || rq->q->elevator;
291 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
292 unsigned int tag, unsigned int op)
294 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
295 struct request *rq = tags->static_rqs[tag];
296 req_flags_t rq_flags = 0;
298 if (data->flags & BLK_MQ_REQ_INTERNAL) {
300 rq->internal_tag = tag;
302 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
303 rq_flags = RQF_MQ_INFLIGHT;
304 atomic_inc(&data->hctx->nr_active);
307 rq->internal_tag = -1;
308 data->hctx->tags->rqs[rq->tag] = rq;
311 /* csd/requeue_work/fifo_time is initialized before use */
313 rq->mq_ctx = data->ctx;
314 rq->mq_hctx = data->hctx;
315 rq->rq_flags = rq_flags;
317 if (data->flags & BLK_MQ_REQ_PREEMPT)
318 rq->rq_flags |= RQF_PREEMPT;
319 if (blk_queue_io_stat(data->q))
320 rq->rq_flags |= RQF_IO_STAT;
321 INIT_LIST_HEAD(&rq->queuelist);
322 INIT_HLIST_NODE(&rq->hash);
323 RB_CLEAR_NODE(&rq->rb_node);
326 if (blk_mq_need_time_stamp(rq))
327 rq->start_time_ns = ktime_get_ns();
329 rq->start_time_ns = 0;
330 rq->io_start_time_ns = 0;
331 rq->nr_phys_segments = 0;
332 #if defined(CONFIG_BLK_DEV_INTEGRITY)
333 rq->nr_integrity_segments = 0;
335 /* tag was already set */
337 WRITE_ONCE(rq->deadline, 0);
342 rq->end_io_data = NULL;
344 data->ctx->rq_dispatched[op_is_sync(op)]++;
345 refcount_set(&rq->ref, 1);
349 static struct request *blk_mq_get_request(struct request_queue *q,
351 struct blk_mq_alloc_data *data)
353 struct elevator_queue *e = q->elevator;
356 bool put_ctx_on_error = false;
358 blk_queue_enter_live(q);
360 if (likely(!data->ctx)) {
361 data->ctx = blk_mq_get_ctx(q);
362 put_ctx_on_error = true;
364 if (likely(!data->hctx))
365 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
367 if (data->cmd_flags & REQ_NOWAIT)
368 data->flags |= BLK_MQ_REQ_NOWAIT;
371 data->flags |= BLK_MQ_REQ_INTERNAL;
374 * Flush requests are special and go directly to the
375 * dispatch list. Don't include reserved tags in the
376 * limiting, as it isn't useful.
378 if (!op_is_flush(data->cmd_flags) &&
379 e->type->ops.limit_depth &&
380 !(data->flags & BLK_MQ_REQ_RESERVED))
381 e->type->ops.limit_depth(data->cmd_flags, data);
383 blk_mq_tag_busy(data->hctx);
386 tag = blk_mq_get_tag(data);
387 if (tag == BLK_MQ_TAG_FAIL) {
388 if (put_ctx_on_error) {
389 blk_mq_put_ctx(data->ctx);
396 rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags);
397 if (!op_is_flush(data->cmd_flags)) {
399 if (e && e->type->ops.prepare_request) {
400 if (e->type->icq_cache)
401 blk_mq_sched_assign_ioc(rq);
403 e->type->ops.prepare_request(rq, bio);
404 rq->rq_flags |= RQF_ELVPRIV;
407 data->hctx->queued++;
411 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
412 blk_mq_req_flags_t flags)
414 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
418 ret = blk_queue_enter(q, flags);
422 rq = blk_mq_get_request(q, NULL, &alloc_data);
426 return ERR_PTR(-EWOULDBLOCK);
428 blk_mq_put_ctx(alloc_data.ctx);
431 rq->__sector = (sector_t) -1;
432 rq->bio = rq->biotail = NULL;
435 EXPORT_SYMBOL(blk_mq_alloc_request);
437 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
438 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
440 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
446 * If the tag allocator sleeps we could get an allocation for a
447 * different hardware context. No need to complicate the low level
448 * allocator for this for the rare use case of a command tied to
451 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
452 return ERR_PTR(-EINVAL);
454 if (hctx_idx >= q->nr_hw_queues)
455 return ERR_PTR(-EIO);
457 ret = blk_queue_enter(q, flags);
462 * Check if the hardware context is actually mapped to anything.
463 * If not tell the caller that it should skip this queue.
465 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
466 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
468 return ERR_PTR(-EXDEV);
470 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
471 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
473 rq = blk_mq_get_request(q, NULL, &alloc_data);
477 return ERR_PTR(-EWOULDBLOCK);
481 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
483 static void __blk_mq_free_request(struct request *rq)
485 struct request_queue *q = rq->q;
486 struct blk_mq_ctx *ctx = rq->mq_ctx;
487 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
488 const int sched_tag = rq->internal_tag;
490 blk_pm_mark_last_busy(rq);
493 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
495 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
496 blk_mq_sched_restart(hctx);
500 void blk_mq_free_request(struct request *rq)
502 struct request_queue *q = rq->q;
503 struct elevator_queue *e = q->elevator;
504 struct blk_mq_ctx *ctx = rq->mq_ctx;
505 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
507 if (rq->rq_flags & RQF_ELVPRIV) {
508 if (e && e->type->ops.finish_request)
509 e->type->ops.finish_request(rq);
511 put_io_context(rq->elv.icq->ioc);
516 ctx->rq_completed[rq_is_sync(rq)]++;
517 if (rq->rq_flags & RQF_MQ_INFLIGHT)
518 atomic_dec(&hctx->nr_active);
520 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
521 laptop_io_completion(q->backing_dev_info);
525 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
526 if (refcount_dec_and_test(&rq->ref))
527 __blk_mq_free_request(rq);
529 EXPORT_SYMBOL_GPL(blk_mq_free_request);
531 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
535 if (blk_mq_need_time_stamp(rq))
536 now = ktime_get_ns();
538 if (rq->rq_flags & RQF_STATS) {
539 blk_mq_poll_stats_start(rq->q);
540 blk_stat_add(rq, now);
543 if (rq->internal_tag != -1)
544 blk_mq_sched_completed_request(rq, now);
546 blk_account_io_done(rq, now);
549 rq_qos_done(rq->q, rq);
550 rq->end_io(rq, error);
552 blk_mq_free_request(rq);
555 EXPORT_SYMBOL(__blk_mq_end_request);
557 void blk_mq_end_request(struct request *rq, blk_status_t error)
559 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
561 __blk_mq_end_request(rq, error);
563 EXPORT_SYMBOL(blk_mq_end_request);
565 static void __blk_mq_complete_request_remote(void *data)
567 struct request *rq = data;
568 struct request_queue *q = rq->q;
570 q->mq_ops->complete(rq);
573 static void __blk_mq_complete_request(struct request *rq)
575 struct blk_mq_ctx *ctx = rq->mq_ctx;
576 struct request_queue *q = rq->q;
580 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
582 * Most of single queue controllers, there is only one irq vector
583 * for handling IO completion, and the only irq's affinity is set
584 * as all possible CPUs. On most of ARCHs, this affinity means the
585 * irq is handled on one specific CPU.
587 * So complete IO reqeust in softirq context in case of single queue
588 * for not degrading IO performance by irqsoff latency.
590 if (q->nr_hw_queues == 1) {
591 __blk_complete_request(rq);
596 * For a polled request, always complete locallly, it's pointless
597 * to redirect the completion.
599 if ((rq->cmd_flags & REQ_HIPRI) ||
600 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
601 q->mq_ops->complete(rq);
606 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
607 shared = cpus_share_cache(cpu, ctx->cpu);
609 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
610 rq->csd.func = __blk_mq_complete_request_remote;
613 smp_call_function_single_async(ctx->cpu, &rq->csd);
615 q->mq_ops->complete(rq);
620 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
621 __releases(hctx->srcu)
623 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
626 srcu_read_unlock(hctx->srcu, srcu_idx);
629 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
630 __acquires(hctx->srcu)
632 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
633 /* shut up gcc false positive */
637 *srcu_idx = srcu_read_lock(hctx->srcu);
641 * blk_mq_complete_request - end I/O on a request
642 * @rq: the request being processed
645 * Ends all I/O on a request. It does not handle partial completions.
646 * The actual completion happens out-of-order, through a IPI handler.
648 bool blk_mq_complete_request(struct request *rq)
650 if (unlikely(blk_should_fake_timeout(rq->q)))
652 __blk_mq_complete_request(rq);
655 EXPORT_SYMBOL(blk_mq_complete_request);
657 int blk_mq_request_started(struct request *rq)
659 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
661 EXPORT_SYMBOL_GPL(blk_mq_request_started);
663 void blk_mq_start_request(struct request *rq)
665 struct request_queue *q = rq->q;
667 blk_mq_sched_started_request(rq);
669 trace_block_rq_issue(q, rq);
671 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
672 rq->io_start_time_ns = ktime_get_ns();
673 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
674 rq->throtl_size = blk_rq_sectors(rq);
676 rq->rq_flags |= RQF_STATS;
680 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
683 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
685 if (q->dma_drain_size && blk_rq_bytes(rq)) {
687 * Make sure space for the drain appears. We know we can do
688 * this because max_hw_segments has been adjusted to be one
689 * fewer than the device can handle.
691 rq->nr_phys_segments++;
694 EXPORT_SYMBOL(blk_mq_start_request);
696 static void __blk_mq_requeue_request(struct request *rq)
698 struct request_queue *q = rq->q;
700 blk_mq_put_driver_tag(rq);
702 trace_block_rq_requeue(q, rq);
703 rq_qos_requeue(q, rq);
705 if (blk_mq_request_started(rq)) {
706 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
707 rq->rq_flags &= ~RQF_TIMED_OUT;
708 if (q->dma_drain_size && blk_rq_bytes(rq))
709 rq->nr_phys_segments--;
713 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
715 __blk_mq_requeue_request(rq);
717 /* this request will be re-inserted to io scheduler queue */
718 blk_mq_sched_requeue_request(rq);
720 BUG_ON(!list_empty(&rq->queuelist));
721 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
723 EXPORT_SYMBOL(blk_mq_requeue_request);
725 static void blk_mq_requeue_work(struct work_struct *work)
727 struct request_queue *q =
728 container_of(work, struct request_queue, requeue_work.work);
730 struct request *rq, *next;
732 spin_lock_irq(&q->requeue_lock);
733 list_splice_init(&q->requeue_list, &rq_list);
734 spin_unlock_irq(&q->requeue_lock);
736 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
737 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
740 rq->rq_flags &= ~RQF_SOFTBARRIER;
741 list_del_init(&rq->queuelist);
743 * If RQF_DONTPREP, rq has contained some driver specific
744 * data, so insert it to hctx dispatch list to avoid any
747 if (rq->rq_flags & RQF_DONTPREP)
748 blk_mq_request_bypass_insert(rq, false);
750 blk_mq_sched_insert_request(rq, true, false, false);
753 while (!list_empty(&rq_list)) {
754 rq = list_entry(rq_list.next, struct request, queuelist);
755 list_del_init(&rq->queuelist);
756 blk_mq_sched_insert_request(rq, false, false, false);
759 blk_mq_run_hw_queues(q, false);
762 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
763 bool kick_requeue_list)
765 struct request_queue *q = rq->q;
769 * We abuse this flag that is otherwise used by the I/O scheduler to
770 * request head insertion from the workqueue.
772 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
774 spin_lock_irqsave(&q->requeue_lock, flags);
776 rq->rq_flags |= RQF_SOFTBARRIER;
777 list_add(&rq->queuelist, &q->requeue_list);
779 list_add_tail(&rq->queuelist, &q->requeue_list);
781 spin_unlock_irqrestore(&q->requeue_lock, flags);
783 if (kick_requeue_list)
784 blk_mq_kick_requeue_list(q);
787 void blk_mq_kick_requeue_list(struct request_queue *q)
789 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
791 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
793 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
796 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
797 msecs_to_jiffies(msecs));
799 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
801 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
803 if (tag < tags->nr_tags) {
804 prefetch(tags->rqs[tag]);
805 return tags->rqs[tag];
810 EXPORT_SYMBOL(blk_mq_tag_to_rq);
812 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
813 void *priv, bool reserved)
816 * If we find a request that is inflight and the queue matches,
817 * we know the queue is busy. Return false to stop the iteration.
819 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
829 bool blk_mq_queue_inflight(struct request_queue *q)
833 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
836 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
838 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
840 req->rq_flags |= RQF_TIMED_OUT;
841 if (req->q->mq_ops->timeout) {
842 enum blk_eh_timer_return ret;
844 ret = req->q->mq_ops->timeout(req, reserved);
845 if (ret == BLK_EH_DONE)
847 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
853 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
855 unsigned long deadline;
857 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
859 if (rq->rq_flags & RQF_TIMED_OUT)
862 deadline = READ_ONCE(rq->deadline);
863 if (time_after_eq(jiffies, deadline))
868 else if (time_after(*next, deadline))
873 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
874 struct request *rq, void *priv, bool reserved)
876 unsigned long *next = priv;
879 * Just do a quick check if it is expired before locking the request in
880 * so we're not unnecessarilly synchronizing across CPUs.
882 if (!blk_mq_req_expired(rq, next))
886 * We have reason to believe the request may be expired. Take a
887 * reference on the request to lock this request lifetime into its
888 * currently allocated context to prevent it from being reallocated in
889 * the event the completion by-passes this timeout handler.
891 * If the reference was already released, then the driver beat the
892 * timeout handler to posting a natural completion.
894 if (!refcount_inc_not_zero(&rq->ref))
898 * The request is now locked and cannot be reallocated underneath the
899 * timeout handler's processing. Re-verify this exact request is truly
900 * expired; if it is not expired, then the request was completed and
901 * reallocated as a new request.
903 if (blk_mq_req_expired(rq, next))
904 blk_mq_rq_timed_out(rq, reserved);
905 if (refcount_dec_and_test(&rq->ref))
906 __blk_mq_free_request(rq);
911 static void blk_mq_timeout_work(struct work_struct *work)
913 struct request_queue *q =
914 container_of(work, struct request_queue, timeout_work);
915 unsigned long next = 0;
916 struct blk_mq_hw_ctx *hctx;
919 /* A deadlock might occur if a request is stuck requiring a
920 * timeout at the same time a queue freeze is waiting
921 * completion, since the timeout code would not be able to
922 * acquire the queue reference here.
924 * That's why we don't use blk_queue_enter here; instead, we use
925 * percpu_ref_tryget directly, because we need to be able to
926 * obtain a reference even in the short window between the queue
927 * starting to freeze, by dropping the first reference in
928 * blk_freeze_queue_start, and the moment the last request is
929 * consumed, marked by the instant q_usage_counter reaches
932 if (!percpu_ref_tryget(&q->q_usage_counter))
935 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
938 mod_timer(&q->timeout, next);
941 * Request timeouts are handled as a forward rolling timer. If
942 * we end up here it means that no requests are pending and
943 * also that no request has been pending for a while. Mark
946 queue_for_each_hw_ctx(q, hctx, i) {
947 /* the hctx may be unmapped, so check it here */
948 if (blk_mq_hw_queue_mapped(hctx))
949 blk_mq_tag_idle(hctx);
955 struct flush_busy_ctx_data {
956 struct blk_mq_hw_ctx *hctx;
957 struct list_head *list;
960 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
962 struct flush_busy_ctx_data *flush_data = data;
963 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
964 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
965 enum hctx_type type = hctx->type;
967 spin_lock(&ctx->lock);
968 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
969 sbitmap_clear_bit(sb, bitnr);
970 spin_unlock(&ctx->lock);
975 * Process software queues that have been marked busy, splicing them
976 * to the for-dispatch
978 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
980 struct flush_busy_ctx_data data = {
985 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
987 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
989 struct dispatch_rq_data {
990 struct blk_mq_hw_ctx *hctx;
994 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
997 struct dispatch_rq_data *dispatch_data = data;
998 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
999 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1000 enum hctx_type type = hctx->type;
1002 spin_lock(&ctx->lock);
1003 if (!list_empty(&ctx->rq_lists[type])) {
1004 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1005 list_del_init(&dispatch_data->rq->queuelist);
1006 if (list_empty(&ctx->rq_lists[type]))
1007 sbitmap_clear_bit(sb, bitnr);
1009 spin_unlock(&ctx->lock);
1011 return !dispatch_data->rq;
1014 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1015 struct blk_mq_ctx *start)
1017 unsigned off = start ? start->index_hw[hctx->type] : 0;
1018 struct dispatch_rq_data data = {
1023 __sbitmap_for_each_set(&hctx->ctx_map, off,
1024 dispatch_rq_from_ctx, &data);
1029 static inline unsigned int queued_to_index(unsigned int queued)
1034 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1037 bool blk_mq_get_driver_tag(struct request *rq)
1039 struct blk_mq_alloc_data data = {
1041 .hctx = rq->mq_hctx,
1042 .flags = BLK_MQ_REQ_NOWAIT,
1043 .cmd_flags = rq->cmd_flags,
1050 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1051 data.flags |= BLK_MQ_REQ_RESERVED;
1053 shared = blk_mq_tag_busy(data.hctx);
1054 rq->tag = blk_mq_get_tag(&data);
1057 rq->rq_flags |= RQF_MQ_INFLIGHT;
1058 atomic_inc(&data.hctx->nr_active);
1060 data.hctx->tags->rqs[rq->tag] = rq;
1064 return rq->tag != -1;
1067 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1068 int flags, void *key)
1070 struct blk_mq_hw_ctx *hctx;
1072 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1074 spin_lock(&hctx->dispatch_wait_lock);
1075 if (!list_empty(&wait->entry)) {
1076 struct sbitmap_queue *sbq;
1078 list_del_init(&wait->entry);
1079 sbq = &hctx->tags->bitmap_tags;
1080 atomic_dec(&sbq->ws_active);
1082 spin_unlock(&hctx->dispatch_wait_lock);
1084 blk_mq_run_hw_queue(hctx, true);
1089 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1090 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1091 * restart. For both cases, take care to check the condition again after
1092 * marking us as waiting.
1094 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1097 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1098 struct wait_queue_head *wq;
1099 wait_queue_entry_t *wait;
1102 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1103 blk_mq_sched_mark_restart_hctx(hctx);
1106 * It's possible that a tag was freed in the window between the
1107 * allocation failure and adding the hardware queue to the wait
1110 * Don't clear RESTART here, someone else could have set it.
1111 * At most this will cost an extra queue run.
1113 return blk_mq_get_driver_tag(rq);
1116 wait = &hctx->dispatch_wait;
1117 if (!list_empty_careful(&wait->entry))
1120 wq = &bt_wait_ptr(sbq, hctx)->wait;
1122 spin_lock_irq(&wq->lock);
1123 spin_lock(&hctx->dispatch_wait_lock);
1124 if (!list_empty(&wait->entry)) {
1125 spin_unlock(&hctx->dispatch_wait_lock);
1126 spin_unlock_irq(&wq->lock);
1130 atomic_inc(&sbq->ws_active);
1131 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1132 __add_wait_queue(wq, wait);
1135 * It's possible that a tag was freed in the window between the
1136 * allocation failure and adding the hardware queue to the wait
1139 ret = blk_mq_get_driver_tag(rq);
1141 spin_unlock(&hctx->dispatch_wait_lock);
1142 spin_unlock_irq(&wq->lock);
1147 * We got a tag, remove ourselves from the wait queue to ensure
1148 * someone else gets the wakeup.
1150 list_del_init(&wait->entry);
1151 atomic_dec(&sbq->ws_active);
1152 spin_unlock(&hctx->dispatch_wait_lock);
1153 spin_unlock_irq(&wq->lock);
1158 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1159 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1161 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1162 * - EWMA is one simple way to compute running average value
1163 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1164 * - take 4 as factor for avoiding to get too small(0) result, and this
1165 * factor doesn't matter because EWMA decreases exponentially
1167 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1171 if (hctx->queue->elevator)
1174 ewma = hctx->dispatch_busy;
1179 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1181 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1182 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1184 hctx->dispatch_busy = ewma;
1187 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1190 * Returns true if we did some work AND can potentially do more.
1192 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1195 struct blk_mq_hw_ctx *hctx;
1196 struct request *rq, *nxt;
1197 bool no_tag = false;
1199 blk_status_t ret = BLK_STS_OK;
1201 if (list_empty(list))
1204 WARN_ON(!list_is_singular(list) && got_budget);
1207 * Now process all the entries, sending them to the driver.
1209 errors = queued = 0;
1211 struct blk_mq_queue_data bd;
1213 rq = list_first_entry(list, struct request, queuelist);
1216 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1219 if (!blk_mq_get_driver_tag(rq)) {
1221 * The initial allocation attempt failed, so we need to
1222 * rerun the hardware queue when a tag is freed. The
1223 * waitqueue takes care of that. If the queue is run
1224 * before we add this entry back on the dispatch list,
1225 * we'll re-run it below.
1227 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1228 blk_mq_put_dispatch_budget(hctx);
1230 * For non-shared tags, the RESTART check
1233 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1239 list_del_init(&rq->queuelist);
1244 * Flag last if we have no more requests, or if we have more
1245 * but can't assign a driver tag to it.
1247 if (list_empty(list))
1250 nxt = list_first_entry(list, struct request, queuelist);
1251 bd.last = !blk_mq_get_driver_tag(nxt);
1254 ret = q->mq_ops->queue_rq(hctx, &bd);
1255 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1257 * If an I/O scheduler has been configured and we got a
1258 * driver tag for the next request already, free it
1261 if (!list_empty(list)) {
1262 nxt = list_first_entry(list, struct request, queuelist);
1263 blk_mq_put_driver_tag(nxt);
1265 list_add(&rq->queuelist, list);
1266 __blk_mq_requeue_request(rq);
1270 if (unlikely(ret != BLK_STS_OK)) {
1272 blk_mq_end_request(rq, BLK_STS_IOERR);
1277 } while (!list_empty(list));
1279 hctx->dispatched[queued_to_index(queued)]++;
1282 * Any items that need requeuing? Stuff them into hctx->dispatch,
1283 * that is where we will continue on next queue run.
1285 if (!list_empty(list)) {
1289 * If we didn't flush the entire list, we could have told
1290 * the driver there was more coming, but that turned out to
1293 if (q->mq_ops->commit_rqs)
1294 q->mq_ops->commit_rqs(hctx);
1296 spin_lock(&hctx->lock);
1297 list_splice_init(list, &hctx->dispatch);
1298 spin_unlock(&hctx->lock);
1301 * If SCHED_RESTART was set by the caller of this function and
1302 * it is no longer set that means that it was cleared by another
1303 * thread and hence that a queue rerun is needed.
1305 * If 'no_tag' is set, that means that we failed getting
1306 * a driver tag with an I/O scheduler attached. If our dispatch
1307 * waitqueue is no longer active, ensure that we run the queue
1308 * AFTER adding our entries back to the list.
1310 * If no I/O scheduler has been configured it is possible that
1311 * the hardware queue got stopped and restarted before requests
1312 * were pushed back onto the dispatch list. Rerun the queue to
1313 * avoid starvation. Notes:
1314 * - blk_mq_run_hw_queue() checks whether or not a queue has
1315 * been stopped before rerunning a queue.
1316 * - Some but not all block drivers stop a queue before
1317 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1320 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1321 * bit is set, run queue after a delay to avoid IO stalls
1322 * that could otherwise occur if the queue is idle.
1324 needs_restart = blk_mq_sched_needs_restart(hctx);
1325 if (!needs_restart ||
1326 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1327 blk_mq_run_hw_queue(hctx, true);
1328 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1329 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1331 blk_mq_update_dispatch_busy(hctx, true);
1334 blk_mq_update_dispatch_busy(hctx, false);
1337 * If the host/device is unable to accept more work, inform the
1340 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1343 return (queued + errors) != 0;
1346 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1351 * We should be running this queue from one of the CPUs that
1354 * There are at least two related races now between setting
1355 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1356 * __blk_mq_run_hw_queue():
1358 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1359 * but later it becomes online, then this warning is harmless
1362 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1363 * but later it becomes offline, then the warning can't be
1364 * triggered, and we depend on blk-mq timeout handler to
1365 * handle dispatched requests to this hctx
1367 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1368 cpu_online(hctx->next_cpu)) {
1369 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1370 raw_smp_processor_id(),
1371 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1376 * We can't run the queue inline with ints disabled. Ensure that
1377 * we catch bad users of this early.
1379 WARN_ON_ONCE(in_interrupt());
1381 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1383 hctx_lock(hctx, &srcu_idx);
1384 blk_mq_sched_dispatch_requests(hctx);
1385 hctx_unlock(hctx, srcu_idx);
1388 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1390 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1392 if (cpu >= nr_cpu_ids)
1393 cpu = cpumask_first(hctx->cpumask);
1398 * It'd be great if the workqueue API had a way to pass
1399 * in a mask and had some smarts for more clever placement.
1400 * For now we just round-robin here, switching for every
1401 * BLK_MQ_CPU_WORK_BATCH queued items.
1403 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1406 int next_cpu = hctx->next_cpu;
1408 if (hctx->queue->nr_hw_queues == 1)
1409 return WORK_CPU_UNBOUND;
1411 if (--hctx->next_cpu_batch <= 0) {
1413 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1415 if (next_cpu >= nr_cpu_ids)
1416 next_cpu = blk_mq_first_mapped_cpu(hctx);
1417 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1421 * Do unbound schedule if we can't find a online CPU for this hctx,
1422 * and it should only happen in the path of handling CPU DEAD.
1424 if (!cpu_online(next_cpu)) {
1431 * Make sure to re-select CPU next time once after CPUs
1432 * in hctx->cpumask become online again.
1434 hctx->next_cpu = next_cpu;
1435 hctx->next_cpu_batch = 1;
1436 return WORK_CPU_UNBOUND;
1439 hctx->next_cpu = next_cpu;
1443 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1444 unsigned long msecs)
1446 if (unlikely(blk_mq_hctx_stopped(hctx)))
1449 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1450 int cpu = get_cpu();
1451 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1452 __blk_mq_run_hw_queue(hctx);
1460 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1461 msecs_to_jiffies(msecs));
1464 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1466 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1468 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1470 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1476 * When queue is quiesced, we may be switching io scheduler, or
1477 * updating nr_hw_queues, or other things, and we can't run queue
1478 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1480 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1483 hctx_lock(hctx, &srcu_idx);
1484 need_run = !blk_queue_quiesced(hctx->queue) &&
1485 blk_mq_hctx_has_pending(hctx);
1486 hctx_unlock(hctx, srcu_idx);
1489 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1495 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1497 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1499 struct blk_mq_hw_ctx *hctx;
1502 queue_for_each_hw_ctx(q, hctx, i) {
1503 if (blk_mq_hctx_stopped(hctx))
1506 blk_mq_run_hw_queue(hctx, async);
1509 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1512 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1513 * @q: request queue.
1515 * The caller is responsible for serializing this function against
1516 * blk_mq_{start,stop}_hw_queue().
1518 bool blk_mq_queue_stopped(struct request_queue *q)
1520 struct blk_mq_hw_ctx *hctx;
1523 queue_for_each_hw_ctx(q, hctx, i)
1524 if (blk_mq_hctx_stopped(hctx))
1529 EXPORT_SYMBOL(blk_mq_queue_stopped);
1532 * This function is often used for pausing .queue_rq() by driver when
1533 * there isn't enough resource or some conditions aren't satisfied, and
1534 * BLK_STS_RESOURCE is usually returned.
1536 * We do not guarantee that dispatch can be drained or blocked
1537 * after blk_mq_stop_hw_queue() returns. Please use
1538 * blk_mq_quiesce_queue() for that requirement.
1540 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1542 cancel_delayed_work(&hctx->run_work);
1544 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1546 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1549 * This function is often used for pausing .queue_rq() by driver when
1550 * there isn't enough resource or some conditions aren't satisfied, and
1551 * BLK_STS_RESOURCE is usually returned.
1553 * We do not guarantee that dispatch can be drained or blocked
1554 * after blk_mq_stop_hw_queues() returns. Please use
1555 * blk_mq_quiesce_queue() for that requirement.
1557 void blk_mq_stop_hw_queues(struct request_queue *q)
1559 struct blk_mq_hw_ctx *hctx;
1562 queue_for_each_hw_ctx(q, hctx, i)
1563 blk_mq_stop_hw_queue(hctx);
1565 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1567 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1569 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1571 blk_mq_run_hw_queue(hctx, false);
1573 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1575 void blk_mq_start_hw_queues(struct request_queue *q)
1577 struct blk_mq_hw_ctx *hctx;
1580 queue_for_each_hw_ctx(q, hctx, i)
1581 blk_mq_start_hw_queue(hctx);
1583 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1585 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1587 if (!blk_mq_hctx_stopped(hctx))
1590 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1591 blk_mq_run_hw_queue(hctx, async);
1593 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1595 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1597 struct blk_mq_hw_ctx *hctx;
1600 queue_for_each_hw_ctx(q, hctx, i)
1601 blk_mq_start_stopped_hw_queue(hctx, async);
1603 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1605 static void blk_mq_run_work_fn(struct work_struct *work)
1607 struct blk_mq_hw_ctx *hctx;
1609 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1612 * If we are stopped, don't run the queue.
1614 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1617 __blk_mq_run_hw_queue(hctx);
1620 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1624 struct blk_mq_ctx *ctx = rq->mq_ctx;
1625 enum hctx_type type = hctx->type;
1627 lockdep_assert_held(&ctx->lock);
1629 trace_block_rq_insert(hctx->queue, rq);
1632 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1634 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1637 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1640 struct blk_mq_ctx *ctx = rq->mq_ctx;
1642 lockdep_assert_held(&ctx->lock);
1644 __blk_mq_insert_req_list(hctx, rq, at_head);
1645 blk_mq_hctx_mark_pending(hctx, ctx);
1649 * Should only be used carefully, when the caller knows we want to
1650 * bypass a potential IO scheduler on the target device.
1652 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1654 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1656 spin_lock(&hctx->lock);
1657 list_add_tail(&rq->queuelist, &hctx->dispatch);
1658 spin_unlock(&hctx->lock);
1661 blk_mq_run_hw_queue(hctx, false);
1664 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1665 struct list_head *list)
1669 enum hctx_type type = hctx->type;
1672 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1675 list_for_each_entry(rq, list, queuelist) {
1676 BUG_ON(rq->mq_ctx != ctx);
1677 trace_block_rq_insert(hctx->queue, rq);
1680 spin_lock(&ctx->lock);
1681 list_splice_tail_init(list, &ctx->rq_lists[type]);
1682 blk_mq_hctx_mark_pending(hctx, ctx);
1683 spin_unlock(&ctx->lock);
1686 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1688 struct request *rqa = container_of(a, struct request, queuelist);
1689 struct request *rqb = container_of(b, struct request, queuelist);
1691 if (rqa->mq_ctx < rqb->mq_ctx)
1693 else if (rqa->mq_ctx > rqb->mq_ctx)
1695 else if (rqa->mq_hctx < rqb->mq_hctx)
1697 else if (rqa->mq_hctx > rqb->mq_hctx)
1700 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1703 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1705 struct blk_mq_hw_ctx *this_hctx;
1706 struct blk_mq_ctx *this_ctx;
1707 struct request_queue *this_q;
1713 list_splice_init(&plug->mq_list, &list);
1716 if (plug->rq_count > 2 && plug->multiple_queues)
1717 list_sort(NULL, &list, plug_rq_cmp);
1724 while (!list_empty(&list)) {
1725 rq = list_entry_rq(list.next);
1726 list_del_init(&rq->queuelist);
1728 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1730 trace_block_unplug(this_q, depth, !from_schedule);
1731 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1737 this_ctx = rq->mq_ctx;
1738 this_hctx = rq->mq_hctx;
1743 list_add_tail(&rq->queuelist, &rq_list);
1747 * If 'this_hctx' is set, we know we have entries to complete
1748 * on 'rq_list'. Do those.
1751 trace_block_unplug(this_q, depth, !from_schedule);
1752 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1757 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1759 blk_init_request_from_bio(rq, bio);
1761 blk_account_io_start(rq, true);
1764 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1766 blk_qc_t *cookie, bool last)
1768 struct request_queue *q = rq->q;
1769 struct blk_mq_queue_data bd = {
1773 blk_qc_t new_cookie;
1776 new_cookie = request_to_qc_t(hctx, rq);
1779 * For OK queue, we are done. For error, caller may kill it.
1780 * Any other error (busy), just add it to our list as we
1781 * previously would have done.
1783 ret = q->mq_ops->queue_rq(hctx, &bd);
1786 blk_mq_update_dispatch_busy(hctx, false);
1787 *cookie = new_cookie;
1789 case BLK_STS_RESOURCE:
1790 case BLK_STS_DEV_RESOURCE:
1791 blk_mq_update_dispatch_busy(hctx, true);
1792 __blk_mq_requeue_request(rq);
1795 blk_mq_update_dispatch_busy(hctx, false);
1796 *cookie = BLK_QC_T_NONE;
1803 blk_status_t blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1806 bool bypass, bool last)
1808 struct request_queue *q = rq->q;
1809 bool run_queue = true;
1810 blk_status_t ret = BLK_STS_RESOURCE;
1814 hctx_lock(hctx, &srcu_idx);
1816 * hctx_lock is needed before checking quiesced flag.
1818 * When queue is stopped or quiesced, ignore 'bypass', insert
1819 * and return BLK_STS_OK to caller, and avoid driver to try to
1822 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) {
1828 if (unlikely(q->elevator && !bypass))
1831 if (!blk_mq_get_dispatch_budget(hctx))
1834 if (!blk_mq_get_driver_tag(rq)) {
1835 blk_mq_put_dispatch_budget(hctx);
1840 * Always add a request that has been through
1841 *.queue_rq() to the hardware dispatch list.
1844 ret = __blk_mq_issue_directly(hctx, rq, cookie, last);
1846 hctx_unlock(hctx, srcu_idx);
1850 case BLK_STS_DEV_RESOURCE:
1851 case BLK_STS_RESOURCE:
1853 blk_mq_request_bypass_insert(rq, run_queue);
1855 * We have to return BLK_STS_OK for the DM
1856 * to avoid livelock. Otherwise, we return
1857 * the real result to indicate whether the
1858 * request is direct-issued successfully.
1860 ret = bypass ? BLK_STS_OK : ret;
1861 } else if (!bypass) {
1862 blk_mq_sched_insert_request(rq, false,
1868 blk_mq_end_request(rq, ret);
1875 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1876 struct list_head *list)
1879 blk_status_t ret = BLK_STS_OK;
1881 while (!list_empty(list)) {
1882 struct request *rq = list_first_entry(list, struct request,
1885 list_del_init(&rq->queuelist);
1886 if (ret == BLK_STS_OK)
1887 ret = blk_mq_try_issue_directly(hctx, rq, &unused,
1891 blk_mq_sched_insert_request(rq, false, true, false);
1895 * If we didn't flush the entire list, we could have told
1896 * the driver there was more coming, but that turned out to
1899 if (ret != BLK_STS_OK && hctx->queue->mq_ops->commit_rqs)
1900 hctx->queue->mq_ops->commit_rqs(hctx);
1903 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1905 list_add_tail(&rq->queuelist, &plug->mq_list);
1907 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1908 struct request *tmp;
1910 tmp = list_first_entry(&plug->mq_list, struct request,
1912 if (tmp->q != rq->q)
1913 plug->multiple_queues = true;
1917 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1919 const int is_sync = op_is_sync(bio->bi_opf);
1920 const int is_flush_fua = op_is_flush(bio->bi_opf);
1921 struct blk_mq_alloc_data data = { .flags = 0};
1923 struct blk_plug *plug;
1924 struct request *same_queue_rq = NULL;
1927 blk_queue_bounce(q, &bio);
1929 blk_queue_split(q, &bio);
1931 if (!bio_integrity_prep(bio))
1932 return BLK_QC_T_NONE;
1934 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1935 blk_attempt_plug_merge(q, bio, &same_queue_rq))
1936 return BLK_QC_T_NONE;
1938 if (blk_mq_sched_bio_merge(q, bio))
1939 return BLK_QC_T_NONE;
1941 rq_qos_throttle(q, bio);
1943 data.cmd_flags = bio->bi_opf;
1944 rq = blk_mq_get_request(q, bio, &data);
1945 if (unlikely(!rq)) {
1946 rq_qos_cleanup(q, bio);
1947 if (bio->bi_opf & REQ_NOWAIT)
1948 bio_wouldblock_error(bio);
1949 return BLK_QC_T_NONE;
1952 trace_block_getrq(q, bio, bio->bi_opf);
1954 rq_qos_track(q, rq, bio);
1956 cookie = request_to_qc_t(data.hctx, rq);
1958 plug = current->plug;
1959 if (unlikely(is_flush_fua)) {
1960 blk_mq_put_ctx(data.ctx);
1961 blk_mq_bio_to_request(rq, bio);
1963 /* bypass scheduler for flush rq */
1964 blk_insert_flush(rq);
1965 blk_mq_run_hw_queue(data.hctx, true);
1966 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1968 * Use plugging if we have a ->commit_rqs() hook as well, as
1969 * we know the driver uses bd->last in a smart fashion.
1971 unsigned int request_count = plug->rq_count;
1972 struct request *last = NULL;
1974 blk_mq_put_ctx(data.ctx);
1975 blk_mq_bio_to_request(rq, bio);
1978 trace_block_plug(q);
1980 last = list_entry_rq(plug->mq_list.prev);
1982 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1983 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1984 blk_flush_plug_list(plug, false);
1985 trace_block_plug(q);
1988 blk_add_rq_to_plug(plug, rq);
1989 } else if (plug && !blk_queue_nomerges(q)) {
1990 blk_mq_bio_to_request(rq, bio);
1993 * We do limited plugging. If the bio can be merged, do that.
1994 * Otherwise the existing request in the plug list will be
1995 * issued. So the plug list will have one request at most
1996 * The plug list might get flushed before this. If that happens,
1997 * the plug list is empty, and same_queue_rq is invalid.
1999 if (list_empty(&plug->mq_list))
2000 same_queue_rq = NULL;
2001 if (same_queue_rq) {
2002 list_del_init(&same_queue_rq->queuelist);
2005 blk_add_rq_to_plug(plug, rq);
2007 blk_mq_put_ctx(data.ctx);
2009 if (same_queue_rq) {
2010 data.hctx = same_queue_rq->mq_hctx;
2011 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2012 &cookie, false, true);
2014 } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2015 !data.hctx->dispatch_busy)) {
2016 blk_mq_put_ctx(data.ctx);
2017 blk_mq_bio_to_request(rq, bio);
2018 blk_mq_try_issue_directly(data.hctx, rq, &cookie, false, true);
2020 blk_mq_put_ctx(data.ctx);
2021 blk_mq_bio_to_request(rq, bio);
2022 blk_mq_sched_insert_request(rq, false, true, true);
2028 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2029 unsigned int hctx_idx)
2033 if (tags->rqs && set->ops->exit_request) {
2036 for (i = 0; i < tags->nr_tags; i++) {
2037 struct request *rq = tags->static_rqs[i];
2041 set->ops->exit_request(set, rq, hctx_idx);
2042 tags->static_rqs[i] = NULL;
2046 while (!list_empty(&tags->page_list)) {
2047 page = list_first_entry(&tags->page_list, struct page, lru);
2048 list_del_init(&page->lru);
2050 * Remove kmemleak object previously allocated in
2051 * blk_mq_init_rq_map().
2053 kmemleak_free(page_address(page));
2054 __free_pages(page, page->private);
2058 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2062 kfree(tags->static_rqs);
2063 tags->static_rqs = NULL;
2065 blk_mq_free_tags(tags);
2068 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2069 unsigned int hctx_idx,
2070 unsigned int nr_tags,
2071 unsigned int reserved_tags)
2073 struct blk_mq_tags *tags;
2076 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2077 if (node == NUMA_NO_NODE)
2078 node = set->numa_node;
2080 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2081 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2085 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2086 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2089 blk_mq_free_tags(tags);
2093 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2094 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2096 if (!tags->static_rqs) {
2098 blk_mq_free_tags(tags);
2105 static size_t order_to_size(unsigned int order)
2107 return (size_t)PAGE_SIZE << order;
2110 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2111 unsigned int hctx_idx, int node)
2115 if (set->ops->init_request) {
2116 ret = set->ops->init_request(set, rq, hctx_idx, node);
2121 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2125 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2126 unsigned int hctx_idx, unsigned int depth)
2128 unsigned int i, j, entries_per_page, max_order = 4;
2129 size_t rq_size, left;
2132 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2133 if (node == NUMA_NO_NODE)
2134 node = set->numa_node;
2136 INIT_LIST_HEAD(&tags->page_list);
2139 * rq_size is the size of the request plus driver payload, rounded
2140 * to the cacheline size
2142 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2144 left = rq_size * depth;
2146 for (i = 0; i < depth; ) {
2147 int this_order = max_order;
2152 while (this_order && left < order_to_size(this_order - 1))
2156 page = alloc_pages_node(node,
2157 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2163 if (order_to_size(this_order) < rq_size)
2170 page->private = this_order;
2171 list_add_tail(&page->lru, &tags->page_list);
2173 p = page_address(page);
2175 * Allow kmemleak to scan these pages as they contain pointers
2176 * to additional allocations like via ops->init_request().
2178 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2179 entries_per_page = order_to_size(this_order) / rq_size;
2180 to_do = min(entries_per_page, depth - i);
2181 left -= to_do * rq_size;
2182 for (j = 0; j < to_do; j++) {
2183 struct request *rq = p;
2185 tags->static_rqs[i] = rq;
2186 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2187 tags->static_rqs[i] = NULL;
2198 blk_mq_free_rqs(set, tags, hctx_idx);
2203 * 'cpu' is going away. splice any existing rq_list entries from this
2204 * software queue to the hw queue dispatch list, and ensure that it
2207 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2209 struct blk_mq_hw_ctx *hctx;
2210 struct blk_mq_ctx *ctx;
2212 enum hctx_type type;
2214 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2215 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2218 spin_lock(&ctx->lock);
2219 if (!list_empty(&ctx->rq_lists[type])) {
2220 list_splice_init(&ctx->rq_lists[type], &tmp);
2221 blk_mq_hctx_clear_pending(hctx, ctx);
2223 spin_unlock(&ctx->lock);
2225 if (list_empty(&tmp))
2228 spin_lock(&hctx->lock);
2229 list_splice_tail_init(&tmp, &hctx->dispatch);
2230 spin_unlock(&hctx->lock);
2232 blk_mq_run_hw_queue(hctx, true);
2236 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2238 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2242 /* hctx->ctxs will be freed in queue's release handler */
2243 static void blk_mq_exit_hctx(struct request_queue *q,
2244 struct blk_mq_tag_set *set,
2245 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2247 if (blk_mq_hw_queue_mapped(hctx))
2248 blk_mq_tag_idle(hctx);
2250 if (set->ops->exit_request)
2251 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2253 if (set->ops->exit_hctx)
2254 set->ops->exit_hctx(hctx, hctx_idx);
2256 if (hctx->flags & BLK_MQ_F_BLOCKING)
2257 cleanup_srcu_struct(hctx->srcu);
2259 blk_mq_remove_cpuhp(hctx);
2260 blk_free_flush_queue(hctx->fq);
2261 sbitmap_free(&hctx->ctx_map);
2264 static void blk_mq_exit_hw_queues(struct request_queue *q,
2265 struct blk_mq_tag_set *set, int nr_queue)
2267 struct blk_mq_hw_ctx *hctx;
2270 queue_for_each_hw_ctx(q, hctx, i) {
2273 blk_mq_debugfs_unregister_hctx(hctx);
2274 blk_mq_exit_hctx(q, set, hctx, i);
2278 static int blk_mq_init_hctx(struct request_queue *q,
2279 struct blk_mq_tag_set *set,
2280 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2284 node = hctx->numa_node;
2285 if (node == NUMA_NO_NODE)
2286 node = hctx->numa_node = set->numa_node;
2288 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2289 spin_lock_init(&hctx->lock);
2290 INIT_LIST_HEAD(&hctx->dispatch);
2292 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2294 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2296 hctx->tags = set->tags[hctx_idx];
2299 * Allocate space for all possible cpus to avoid allocation at
2302 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2303 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node);
2305 goto unregister_cpu_notifier;
2307 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2308 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node))
2313 spin_lock_init(&hctx->dispatch_wait_lock);
2314 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2315 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2317 if (set->ops->init_hctx &&
2318 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2321 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2322 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
2326 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2329 if (hctx->flags & BLK_MQ_F_BLOCKING)
2330 init_srcu_struct(hctx->srcu);
2337 if (set->ops->exit_hctx)
2338 set->ops->exit_hctx(hctx, hctx_idx);
2340 sbitmap_free(&hctx->ctx_map);
2343 unregister_cpu_notifier:
2344 blk_mq_remove_cpuhp(hctx);
2348 static void blk_mq_init_cpu_queues(struct request_queue *q,
2349 unsigned int nr_hw_queues)
2351 struct blk_mq_tag_set *set = q->tag_set;
2354 for_each_possible_cpu(i) {
2355 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2356 struct blk_mq_hw_ctx *hctx;
2360 spin_lock_init(&__ctx->lock);
2361 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2362 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2367 * Set local node, IFF we have more than one hw queue. If
2368 * not, we remain on the home node of the device
2370 for (j = 0; j < set->nr_maps; j++) {
2371 hctx = blk_mq_map_queue_type(q, j, i);
2372 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2373 hctx->numa_node = local_memory_node(cpu_to_node(i));
2378 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2382 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2383 set->queue_depth, set->reserved_tags);
2384 if (!set->tags[hctx_idx])
2387 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2392 blk_mq_free_rq_map(set->tags[hctx_idx]);
2393 set->tags[hctx_idx] = NULL;
2397 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2398 unsigned int hctx_idx)
2400 if (set->tags && set->tags[hctx_idx]) {
2401 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2402 blk_mq_free_rq_map(set->tags[hctx_idx]);
2403 set->tags[hctx_idx] = NULL;
2407 static void blk_mq_map_swqueue(struct request_queue *q)
2409 unsigned int i, j, hctx_idx;
2410 struct blk_mq_hw_ctx *hctx;
2411 struct blk_mq_ctx *ctx;
2412 struct blk_mq_tag_set *set = q->tag_set;
2415 * Avoid others reading imcomplete hctx->cpumask through sysfs
2417 mutex_lock(&q->sysfs_lock);
2419 queue_for_each_hw_ctx(q, hctx, i) {
2420 cpumask_clear(hctx->cpumask);
2422 hctx->dispatch_from = NULL;
2426 * Map software to hardware queues.
2428 * If the cpu isn't present, the cpu is mapped to first hctx.
2430 for_each_possible_cpu(i) {
2431 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2432 /* unmapped hw queue can be remapped after CPU topo changed */
2433 if (!set->tags[hctx_idx] &&
2434 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2436 * If tags initialization fail for some hctx,
2437 * that hctx won't be brought online. In this
2438 * case, remap the current ctx to hctx[0] which
2439 * is guaranteed to always have tags allocated
2441 set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2444 ctx = per_cpu_ptr(q->queue_ctx, i);
2445 for (j = 0; j < set->nr_maps; j++) {
2446 if (!set->map[j].nr_queues) {
2447 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2448 HCTX_TYPE_DEFAULT, i);
2452 hctx = blk_mq_map_queue_type(q, j, i);
2453 ctx->hctxs[j] = hctx;
2455 * If the CPU is already set in the mask, then we've
2456 * mapped this one already. This can happen if
2457 * devices share queues across queue maps.
2459 if (cpumask_test_cpu(i, hctx->cpumask))
2462 cpumask_set_cpu(i, hctx->cpumask);
2464 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2465 hctx->ctxs[hctx->nr_ctx++] = ctx;
2468 * If the nr_ctx type overflows, we have exceeded the
2469 * amount of sw queues we can support.
2471 BUG_ON(!hctx->nr_ctx);
2474 for (; j < HCTX_MAX_TYPES; j++)
2475 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2476 HCTX_TYPE_DEFAULT, i);
2479 mutex_unlock(&q->sysfs_lock);
2481 queue_for_each_hw_ctx(q, hctx, i) {
2483 * If no software queues are mapped to this hardware queue,
2484 * disable it and free the request entries.
2486 if (!hctx->nr_ctx) {
2487 /* Never unmap queue 0. We need it as a
2488 * fallback in case of a new remap fails
2491 if (i && set->tags[i])
2492 blk_mq_free_map_and_requests(set, i);
2498 hctx->tags = set->tags[i];
2499 WARN_ON(!hctx->tags);
2502 * Set the map size to the number of mapped software queues.
2503 * This is more accurate and more efficient than looping
2504 * over all possibly mapped software queues.
2506 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2509 * Initialize batch roundrobin counts
2511 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2512 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2517 * Caller needs to ensure that we're either frozen/quiesced, or that
2518 * the queue isn't live yet.
2520 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2522 struct blk_mq_hw_ctx *hctx;
2525 queue_for_each_hw_ctx(q, hctx, i) {
2527 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2529 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2533 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2536 struct request_queue *q;
2538 lockdep_assert_held(&set->tag_list_lock);
2540 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2541 blk_mq_freeze_queue(q);
2542 queue_set_hctx_shared(q, shared);
2543 blk_mq_unfreeze_queue(q);
2547 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2549 struct blk_mq_tag_set *set = q->tag_set;
2551 mutex_lock(&set->tag_list_lock);
2552 list_del_rcu(&q->tag_set_list);
2553 if (list_is_singular(&set->tag_list)) {
2554 /* just transitioned to unshared */
2555 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2556 /* update existing queue */
2557 blk_mq_update_tag_set_depth(set, false);
2559 mutex_unlock(&set->tag_list_lock);
2560 INIT_LIST_HEAD(&q->tag_set_list);
2563 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2564 struct request_queue *q)
2566 mutex_lock(&set->tag_list_lock);
2569 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2571 if (!list_empty(&set->tag_list) &&
2572 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2573 set->flags |= BLK_MQ_F_TAG_SHARED;
2574 /* update existing queue */
2575 blk_mq_update_tag_set_depth(set, true);
2577 if (set->flags & BLK_MQ_F_TAG_SHARED)
2578 queue_set_hctx_shared(q, true);
2579 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2581 mutex_unlock(&set->tag_list_lock);
2584 /* All allocations will be freed in release handler of q->mq_kobj */
2585 static int blk_mq_alloc_ctxs(struct request_queue *q)
2587 struct blk_mq_ctxs *ctxs;
2590 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2594 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2595 if (!ctxs->queue_ctx)
2598 for_each_possible_cpu(cpu) {
2599 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2603 q->mq_kobj = &ctxs->kobj;
2604 q->queue_ctx = ctxs->queue_ctx;
2613 * It is the actual release handler for mq, but we do it from
2614 * request queue's release handler for avoiding use-after-free
2615 * and headache because q->mq_kobj shouldn't have been introduced,
2616 * but we can't group ctx/kctx kobj without it.
2618 void blk_mq_release(struct request_queue *q)
2620 struct blk_mq_hw_ctx *hctx;
2623 /* hctx kobj stays in hctx */
2624 queue_for_each_hw_ctx(q, hctx, i) {
2627 kobject_put(&hctx->kobj);
2630 kfree(q->queue_hw_ctx);
2633 * release .mq_kobj and sw queue's kobject now because
2634 * both share lifetime with request queue.
2636 blk_mq_sysfs_deinit(q);
2639 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2641 struct request_queue *uninit_q, *q;
2643 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2645 return ERR_PTR(-ENOMEM);
2647 q = blk_mq_init_allocated_queue(set, uninit_q);
2649 blk_cleanup_queue(uninit_q);
2653 EXPORT_SYMBOL(blk_mq_init_queue);
2656 * Helper for setting up a queue with mq ops, given queue depth, and
2657 * the passed in mq ops flags.
2659 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2660 const struct blk_mq_ops *ops,
2661 unsigned int queue_depth,
2662 unsigned int set_flags)
2664 struct request_queue *q;
2667 memset(set, 0, sizeof(*set));
2669 set->nr_hw_queues = 1;
2671 set->queue_depth = queue_depth;
2672 set->numa_node = NUMA_NO_NODE;
2673 set->flags = set_flags;
2675 ret = blk_mq_alloc_tag_set(set);
2677 return ERR_PTR(ret);
2679 q = blk_mq_init_queue(set);
2681 blk_mq_free_tag_set(set);
2687 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2689 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2691 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2693 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2694 __alignof__(struct blk_mq_hw_ctx)) !=
2695 sizeof(struct blk_mq_hw_ctx));
2697 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2698 hw_ctx_size += sizeof(struct srcu_struct);
2703 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2704 struct blk_mq_tag_set *set, struct request_queue *q,
2705 int hctx_idx, int node)
2707 struct blk_mq_hw_ctx *hctx;
2709 hctx = kzalloc_node(blk_mq_hw_ctx_size(set),
2710 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2715 if (!zalloc_cpumask_var_node(&hctx->cpumask,
2716 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2722 atomic_set(&hctx->nr_active, 0);
2723 hctx->numa_node = node;
2724 hctx->queue_num = hctx_idx;
2726 if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) {
2727 free_cpumask_var(hctx->cpumask);
2731 blk_mq_hctx_kobj_init(hctx);
2736 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2737 struct request_queue *q)
2740 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2742 /* protect against switching io scheduler */
2743 mutex_lock(&q->sysfs_lock);
2744 for (i = 0; i < set->nr_hw_queues; i++) {
2746 struct blk_mq_hw_ctx *hctx;
2748 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2750 * If the hw queue has been mapped to another numa node,
2751 * we need to realloc the hctx. If allocation fails, fallback
2752 * to use the previous one.
2754 if (hctxs[i] && (hctxs[i]->numa_node == node))
2757 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2760 blk_mq_exit_hctx(q, set, hctxs[i], i);
2761 kobject_put(&hctxs[i]->kobj);
2766 pr_warn("Allocate new hctx on node %d fails,\
2767 fallback to previous one on node %d\n",
2768 node, hctxs[i]->numa_node);
2774 * Increasing nr_hw_queues fails. Free the newly allocated
2775 * hctxs and keep the previous q->nr_hw_queues.
2777 if (i != set->nr_hw_queues) {
2778 j = q->nr_hw_queues;
2782 end = q->nr_hw_queues;
2783 q->nr_hw_queues = set->nr_hw_queues;
2786 for (; j < end; j++) {
2787 struct blk_mq_hw_ctx *hctx = hctxs[j];
2791 blk_mq_free_map_and_requests(set, j);
2792 blk_mq_exit_hctx(q, set, hctx, j);
2793 kobject_put(&hctx->kobj);
2798 mutex_unlock(&q->sysfs_lock);
2802 * Maximum number of hardware queues we support. For single sets, we'll never
2803 * have more than the CPUs (software queues). For multiple sets, the tag_set
2804 * user may have set ->nr_hw_queues larger.
2806 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2808 if (set->nr_maps == 1)
2811 return max(set->nr_hw_queues, nr_cpu_ids);
2814 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2815 struct request_queue *q)
2817 /* mark the queue as mq asap */
2818 q->mq_ops = set->ops;
2820 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2821 blk_mq_poll_stats_bkt,
2822 BLK_MQ_POLL_STATS_BKTS, q);
2826 if (blk_mq_alloc_ctxs(q))
2829 /* init q->mq_kobj and sw queues' kobjects */
2830 blk_mq_sysfs_init(q);
2832 q->nr_queues = nr_hw_queues(set);
2833 q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2834 GFP_KERNEL, set->numa_node);
2835 if (!q->queue_hw_ctx)
2838 blk_mq_realloc_hw_ctxs(set, q);
2839 if (!q->nr_hw_queues)
2842 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2843 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2847 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2848 if (set->nr_maps > HCTX_TYPE_POLL &&
2849 set->map[HCTX_TYPE_POLL].nr_queues)
2850 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2852 q->sg_reserved_size = INT_MAX;
2854 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2855 INIT_LIST_HEAD(&q->requeue_list);
2856 spin_lock_init(&q->requeue_lock);
2858 blk_queue_make_request(q, blk_mq_make_request);
2861 * Do this after blk_queue_make_request() overrides it...
2863 q->nr_requests = set->queue_depth;
2866 * Default to classic polling
2868 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2870 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2871 blk_mq_add_queue_tag_set(set, q);
2872 blk_mq_map_swqueue(q);
2874 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2877 ret = elevator_init_mq(q);
2879 return ERR_PTR(ret);
2885 kfree(q->queue_hw_ctx);
2887 blk_mq_sysfs_deinit(q);
2890 return ERR_PTR(-ENOMEM);
2892 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2894 void blk_mq_free_queue(struct request_queue *q)
2896 struct blk_mq_tag_set *set = q->tag_set;
2898 blk_mq_del_queue_tag_set(q);
2899 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2902 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2906 for (i = 0; i < set->nr_hw_queues; i++)
2907 if (!__blk_mq_alloc_rq_map(set, i))
2914 blk_mq_free_rq_map(set->tags[i]);
2920 * Allocate the request maps associated with this tag_set. Note that this
2921 * may reduce the depth asked for, if memory is tight. set->queue_depth
2922 * will be updated to reflect the allocated depth.
2924 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2929 depth = set->queue_depth;
2931 err = __blk_mq_alloc_rq_maps(set);
2935 set->queue_depth >>= 1;
2936 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2940 } while (set->queue_depth);
2942 if (!set->queue_depth || err) {
2943 pr_err("blk-mq: failed to allocate request map\n");
2947 if (depth != set->queue_depth)
2948 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2949 depth, set->queue_depth);
2954 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2956 if (set->ops->map_queues && !is_kdump_kernel()) {
2960 * transport .map_queues is usually done in the following
2963 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2964 * mask = get_cpu_mask(queue)
2965 * for_each_cpu(cpu, mask)
2966 * set->map[x].mq_map[cpu] = queue;
2969 * When we need to remap, the table has to be cleared for
2970 * killing stale mapping since one CPU may not be mapped
2973 for (i = 0; i < set->nr_maps; i++)
2974 blk_mq_clear_mq_map(&set->map[i]);
2976 return set->ops->map_queues(set);
2978 BUG_ON(set->nr_maps > 1);
2979 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2984 * Alloc a tag set to be associated with one or more request queues.
2985 * May fail with EINVAL for various error conditions. May adjust the
2986 * requested depth down, if it's too large. In that case, the set
2987 * value will be stored in set->queue_depth.
2989 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2993 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2995 if (!set->nr_hw_queues)
2997 if (!set->queue_depth)
2999 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3002 if (!set->ops->queue_rq)
3005 if (!set->ops->get_budget ^ !set->ops->put_budget)
3008 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3009 pr_info("blk-mq: reduced tag depth to %u\n",
3011 set->queue_depth = BLK_MQ_MAX_DEPTH;
3016 else if (set->nr_maps > HCTX_MAX_TYPES)
3020 * If a crashdump is active, then we are potentially in a very
3021 * memory constrained environment. Limit us to 1 queue and
3022 * 64 tags to prevent using too much memory.
3024 if (is_kdump_kernel()) {
3025 set->nr_hw_queues = 1;
3027 set->queue_depth = min(64U, set->queue_depth);
3030 * There is no use for more h/w queues than cpus if we just have
3033 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3034 set->nr_hw_queues = nr_cpu_ids;
3036 set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3037 GFP_KERNEL, set->numa_node);
3042 for (i = 0; i < set->nr_maps; i++) {
3043 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3044 sizeof(set->map[i].mq_map[0]),
3045 GFP_KERNEL, set->numa_node);
3046 if (!set->map[i].mq_map)
3047 goto out_free_mq_map;
3048 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3051 ret = blk_mq_update_queue_map(set);
3053 goto out_free_mq_map;
3055 ret = blk_mq_alloc_rq_maps(set);
3057 goto out_free_mq_map;
3059 mutex_init(&set->tag_list_lock);
3060 INIT_LIST_HEAD(&set->tag_list);
3065 for (i = 0; i < set->nr_maps; i++) {
3066 kfree(set->map[i].mq_map);
3067 set->map[i].mq_map = NULL;
3073 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3075 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3079 for (i = 0; i < nr_hw_queues(set); i++)
3080 blk_mq_free_map_and_requests(set, i);
3082 for (j = 0; j < set->nr_maps; j++) {
3083 kfree(set->map[j].mq_map);
3084 set->map[j].mq_map = NULL;
3090 EXPORT_SYMBOL(blk_mq_free_tag_set);
3092 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3094 struct blk_mq_tag_set *set = q->tag_set;
3095 struct blk_mq_hw_ctx *hctx;
3101 if (q->nr_requests == nr)
3104 blk_mq_freeze_queue(q);
3105 blk_mq_quiesce_queue(q);
3108 queue_for_each_hw_ctx(q, hctx, i) {
3112 * If we're using an MQ scheduler, just update the scheduler
3113 * queue depth. This is similar to what the old code would do.
3115 if (!hctx->sched_tags) {
3116 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3119 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3127 q->nr_requests = nr;
3129 blk_mq_unquiesce_queue(q);
3130 blk_mq_unfreeze_queue(q);
3136 * request_queue and elevator_type pair.
3137 * It is just used by __blk_mq_update_nr_hw_queues to cache
3138 * the elevator_type associated with a request_queue.
3140 struct blk_mq_qe_pair {
3141 struct list_head node;
3142 struct request_queue *q;
3143 struct elevator_type *type;
3147 * Cache the elevator_type in qe pair list and switch the
3148 * io scheduler to 'none'
3150 static bool blk_mq_elv_switch_none(struct list_head *head,
3151 struct request_queue *q)
3153 struct blk_mq_qe_pair *qe;
3158 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3162 INIT_LIST_HEAD(&qe->node);
3164 qe->type = q->elevator->type;
3165 list_add(&qe->node, head);
3167 mutex_lock(&q->sysfs_lock);
3169 * After elevator_switch_mq, the previous elevator_queue will be
3170 * released by elevator_release. The reference of the io scheduler
3171 * module get by elevator_get will also be put. So we need to get
3172 * a reference of the io scheduler module here to prevent it to be
3175 __module_get(qe->type->elevator_owner);
3176 elevator_switch_mq(q, NULL);
3177 mutex_unlock(&q->sysfs_lock);
3182 static void blk_mq_elv_switch_back(struct list_head *head,
3183 struct request_queue *q)
3185 struct blk_mq_qe_pair *qe;
3186 struct elevator_type *t = NULL;
3188 list_for_each_entry(qe, head, node)
3197 list_del(&qe->node);
3200 mutex_lock(&q->sysfs_lock);
3201 elevator_switch_mq(q, t);
3202 mutex_unlock(&q->sysfs_lock);
3205 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3208 struct request_queue *q;
3210 int prev_nr_hw_queues;
3212 lockdep_assert_held(&set->tag_list_lock);
3214 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3215 nr_hw_queues = nr_cpu_ids;
3216 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3219 list_for_each_entry(q, &set->tag_list, tag_set_list)
3220 blk_mq_freeze_queue(q);
3222 * Sync with blk_mq_queue_tag_busy_iter.
3226 * Switch IO scheduler to 'none', cleaning up the data associated
3227 * with the previous scheduler. We will switch back once we are done
3228 * updating the new sw to hw queue mappings.
3230 list_for_each_entry(q, &set->tag_list, tag_set_list)
3231 if (!blk_mq_elv_switch_none(&head, q))
3234 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3235 blk_mq_debugfs_unregister_hctxs(q);
3236 blk_mq_sysfs_unregister(q);
3239 prev_nr_hw_queues = set->nr_hw_queues;
3240 set->nr_hw_queues = nr_hw_queues;
3241 blk_mq_update_queue_map(set);
3243 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3244 blk_mq_realloc_hw_ctxs(set, q);
3245 if (q->nr_hw_queues != set->nr_hw_queues) {
3246 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3247 nr_hw_queues, prev_nr_hw_queues);
3248 set->nr_hw_queues = prev_nr_hw_queues;
3249 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3252 blk_mq_map_swqueue(q);
3255 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3256 blk_mq_sysfs_register(q);
3257 blk_mq_debugfs_register_hctxs(q);
3261 list_for_each_entry(q, &set->tag_list, tag_set_list)
3262 blk_mq_elv_switch_back(&head, q);
3264 list_for_each_entry(q, &set->tag_list, tag_set_list)
3265 blk_mq_unfreeze_queue(q);
3268 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3270 mutex_lock(&set->tag_list_lock);
3271 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3272 mutex_unlock(&set->tag_list_lock);
3274 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3276 /* Enable polling stats and return whether they were already enabled. */
3277 static bool blk_poll_stats_enable(struct request_queue *q)
3279 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3280 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3282 blk_stat_add_callback(q, q->poll_cb);
3286 static void blk_mq_poll_stats_start(struct request_queue *q)
3289 * We don't arm the callback if polling stats are not enabled or the
3290 * callback is already active.
3292 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3293 blk_stat_is_active(q->poll_cb))
3296 blk_stat_activate_msecs(q->poll_cb, 100);
3299 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3301 struct request_queue *q = cb->data;
3304 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3305 if (cb->stat[bucket].nr_samples)
3306 q->poll_stat[bucket] = cb->stat[bucket];
3310 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3311 struct blk_mq_hw_ctx *hctx,
3314 unsigned long ret = 0;
3318 * If stats collection isn't on, don't sleep but turn it on for
3321 if (!blk_poll_stats_enable(q))
3325 * As an optimistic guess, use half of the mean service time
3326 * for this type of request. We can (and should) make this smarter.
3327 * For instance, if the completion latencies are tight, we can
3328 * get closer than just half the mean. This is especially
3329 * important on devices where the completion latencies are longer
3330 * than ~10 usec. We do use the stats for the relevant IO size
3331 * if available which does lead to better estimates.
3333 bucket = blk_mq_poll_stats_bkt(rq);
3337 if (q->poll_stat[bucket].nr_samples)
3338 ret = (q->poll_stat[bucket].mean + 1) / 2;
3343 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3344 struct blk_mq_hw_ctx *hctx,
3347 struct hrtimer_sleeper hs;
3348 enum hrtimer_mode mode;
3352 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3356 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3358 * 0: use half of prev avg
3359 * >0: use this specific value
3361 if (q->poll_nsec > 0)
3362 nsecs = q->poll_nsec;
3364 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3369 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3372 * This will be replaced with the stats tracking code, using
3373 * 'avg_completion_time / 2' as the pre-sleep target.
3377 mode = HRTIMER_MODE_REL;
3378 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3379 hrtimer_set_expires(&hs.timer, kt);
3381 hrtimer_init_sleeper(&hs, current);
3383 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3385 set_current_state(TASK_UNINTERRUPTIBLE);
3386 hrtimer_start_expires(&hs.timer, mode);
3389 hrtimer_cancel(&hs.timer);
3390 mode = HRTIMER_MODE_ABS;
3391 } while (hs.task && !signal_pending(current));
3393 __set_current_state(TASK_RUNNING);
3394 destroy_hrtimer_on_stack(&hs.timer);
3398 static bool blk_mq_poll_hybrid(struct request_queue *q,
3399 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3403 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3406 if (!blk_qc_t_is_internal(cookie))
3407 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3409 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3411 * With scheduling, if the request has completed, we'll
3412 * get a NULL return here, as we clear the sched tag when
3413 * that happens. The request still remains valid, like always,
3414 * so we should be safe with just the NULL check.
3420 return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3424 * blk_poll - poll for IO completions
3426 * @cookie: cookie passed back at IO submission time
3427 * @spin: whether to spin for completions
3430 * Poll for completions on the passed in queue. Returns number of
3431 * completed entries found. If @spin is true, then blk_poll will continue
3432 * looping until at least one completion is found, unless the task is
3433 * otherwise marked running (or we need to reschedule).
3435 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3437 struct blk_mq_hw_ctx *hctx;
3440 if (!blk_qc_t_valid(cookie) ||
3441 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3445 blk_flush_plug_list(current->plug, false);
3447 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3450 * If we sleep, have the caller restart the poll loop to reset
3451 * the state. Like for the other success return cases, the
3452 * caller is responsible for checking if the IO completed. If
3453 * the IO isn't complete, we'll get called again and will go
3454 * straight to the busy poll loop.
3456 if (blk_mq_poll_hybrid(q, hctx, cookie))
3459 hctx->poll_considered++;
3461 state = current->state;
3465 hctx->poll_invoked++;
3467 ret = q->mq_ops->poll(hctx);
3469 hctx->poll_success++;
3470 __set_current_state(TASK_RUNNING);
3474 if (signal_pending_state(state, current))
3475 __set_current_state(TASK_RUNNING);
3477 if (current->state == TASK_RUNNING)
3479 if (ret < 0 || !spin)
3482 } while (!need_resched());
3484 __set_current_state(TASK_RUNNING);
3487 EXPORT_SYMBOL_GPL(blk_poll);
3489 unsigned int blk_mq_rq_cpu(struct request *rq)
3491 return rq->mq_ctx->cpu;
3493 EXPORT_SYMBOL(blk_mq_rq_cpu);
3495 static int __init blk_mq_init(void)
3497 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3498 blk_mq_hctx_notify_dead);
3501 subsys_initcall(blk_mq_init);