1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/kmemleak.h>
15 #include <linux/init.h>
16 #include <linux/slab.h>
17 #include <linux/workqueue.h>
18 #include <linux/smp.h>
19 #include <linux/llist.h>
20 #include <linux/list_sort.h>
21 #include <linux/cpu.h>
22 #include <linux/cache.h>
23 #include <linux/sched/sysctl.h>
24 #include <linux/sched/topology.h>
25 #include <linux/sched/signal.h>
26 #include <linux/delay.h>
27 #include <linux/crash_dump.h>
28 #include <linux/prefetch.h>
30 #include <trace/events/block.h>
32 #include <linux/blk-mq.h>
33 #include <linux/t10-pi.h>
36 #include "blk-mq-debugfs.h"
37 #include "blk-mq-tag.h"
40 #include "blk-mq-sched.h"
41 #include "blk-rq-qos.h"
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, sectors, bucket;
50 ddir = rq_data_dir(rq);
51 sectors = blk_rq_stats_sectors(rq);
53 bucket = ddir + 2 * ilog2(sectors);
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, dispatch list or elevator
65 * have pending work in this hardware queue.
67 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
69 return !list_empty_careful(&hctx->dispatch) ||
70 sbitmap_any_bit_set(&hctx->ctx_map) ||
71 blk_mq_sched_has_work(hctx);
75 * Mark this ctx as having pending work in this hardware queue
77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
78 struct blk_mq_ctx *ctx)
80 const int bit = ctx->index_hw[hctx->type];
82 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
83 sbitmap_set_bit(&hctx->ctx_map, bit);
86 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
87 struct blk_mq_ctx *ctx)
89 const int bit = ctx->index_hw[hctx->type];
91 sbitmap_clear_bit(&hctx->ctx_map, bit);
95 struct hd_struct *part;
96 unsigned int inflight[2];
99 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
100 struct request *rq, void *priv,
103 struct mq_inflight *mi = priv;
105 if (rq->part == mi->part)
106 mi->inflight[rq_data_dir(rq)]++;
111 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
113 struct mq_inflight mi = { .part = part };
115 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
117 return mi.inflight[0] + mi.inflight[1];
120 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
121 unsigned int inflight[2])
123 struct mq_inflight mi = { .part = part };
125 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
126 inflight[0] = mi.inflight[0];
127 inflight[1] = mi.inflight[1];
130 void blk_freeze_queue_start(struct request_queue *q)
132 mutex_lock(&q->mq_freeze_lock);
133 if (++q->mq_freeze_depth == 1) {
134 percpu_ref_kill(&q->q_usage_counter);
135 mutex_unlock(&q->mq_freeze_lock);
137 blk_mq_run_hw_queues(q, false);
139 mutex_unlock(&q->mq_freeze_lock);
142 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
144 void blk_mq_freeze_queue_wait(struct request_queue *q)
146 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
148 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
150 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
151 unsigned long timeout)
153 return wait_event_timeout(q->mq_freeze_wq,
154 percpu_ref_is_zero(&q->q_usage_counter),
157 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
160 * Guarantee no request is in use, so we can change any data structure of
161 * the queue afterward.
163 void blk_freeze_queue(struct request_queue *q)
166 * In the !blk_mq case we are only calling this to kill the
167 * q_usage_counter, otherwise this increases the freeze depth
168 * and waits for it to return to zero. For this reason there is
169 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
170 * exported to drivers as the only user for unfreeze is blk_mq.
172 blk_freeze_queue_start(q);
173 blk_mq_freeze_queue_wait(q);
176 void blk_mq_freeze_queue(struct request_queue *q)
179 * ...just an alias to keep freeze and unfreeze actions balanced
180 * in the blk_mq_* namespace
184 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
186 void blk_mq_unfreeze_queue(struct request_queue *q)
188 mutex_lock(&q->mq_freeze_lock);
189 q->mq_freeze_depth--;
190 WARN_ON_ONCE(q->mq_freeze_depth < 0);
191 if (!q->mq_freeze_depth) {
192 percpu_ref_resurrect(&q->q_usage_counter);
193 wake_up_all(&q->mq_freeze_wq);
195 mutex_unlock(&q->mq_freeze_lock);
197 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
200 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
201 * mpt3sas driver such that this function can be removed.
203 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
205 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
207 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
210 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
213 * Note: this function does not prevent that the struct request end_io()
214 * callback function is invoked. Once this function is returned, we make
215 * sure no dispatch can happen until the queue is unquiesced via
216 * blk_mq_unquiesce_queue().
218 void blk_mq_quiesce_queue(struct request_queue *q)
220 struct blk_mq_hw_ctx *hctx;
224 blk_mq_quiesce_queue_nowait(q);
226 queue_for_each_hw_ctx(q, hctx, i) {
227 if (hctx->flags & BLK_MQ_F_BLOCKING)
228 synchronize_srcu(hctx->srcu);
235 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
238 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
241 * This function recovers queue into the state before quiescing
242 * which is done by blk_mq_quiesce_queue.
244 void blk_mq_unquiesce_queue(struct request_queue *q)
246 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
248 /* dispatch requests which are inserted during quiescing */
249 blk_mq_run_hw_queues(q, true);
251 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
253 void blk_mq_wake_waiters(struct request_queue *q)
255 struct blk_mq_hw_ctx *hctx;
258 queue_for_each_hw_ctx(q, hctx, i)
259 if (blk_mq_hw_queue_mapped(hctx))
260 blk_mq_tag_wakeup_all(hctx->tags, true);
264 * Only need start/end time stamping if we have iostat or
265 * blk stats enabled, or using an IO scheduler.
267 static inline bool blk_mq_need_time_stamp(struct request *rq)
269 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
272 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
273 unsigned int tag, unsigned int op, u64 alloc_time_ns)
275 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
276 struct request *rq = tags->static_rqs[tag];
277 req_flags_t rq_flags = 0;
279 if (data->flags & BLK_MQ_REQ_INTERNAL) {
281 rq->internal_tag = tag;
283 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
284 rq_flags = RQF_MQ_INFLIGHT;
285 atomic_inc(&data->hctx->nr_active);
288 rq->internal_tag = -1;
289 data->hctx->tags->rqs[rq->tag] = rq;
292 /* csd/requeue_work/fifo_time is initialized before use */
294 rq->mq_ctx = data->ctx;
295 rq->mq_hctx = data->hctx;
296 rq->rq_flags = rq_flags;
298 if (data->flags & BLK_MQ_REQ_PREEMPT)
299 rq->rq_flags |= RQF_PREEMPT;
300 if (blk_queue_io_stat(data->q))
301 rq->rq_flags |= RQF_IO_STAT;
302 INIT_LIST_HEAD(&rq->queuelist);
303 INIT_HLIST_NODE(&rq->hash);
304 RB_CLEAR_NODE(&rq->rb_node);
307 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
308 rq->alloc_time_ns = alloc_time_ns;
310 if (blk_mq_need_time_stamp(rq))
311 rq->start_time_ns = ktime_get_ns();
313 rq->start_time_ns = 0;
314 rq->io_start_time_ns = 0;
315 rq->stats_sectors = 0;
316 rq->nr_phys_segments = 0;
317 #if defined(CONFIG_BLK_DEV_INTEGRITY)
318 rq->nr_integrity_segments = 0;
320 /* tag was already set */
322 WRITE_ONCE(rq->deadline, 0);
327 rq->end_io_data = NULL;
329 data->ctx->rq_dispatched[op_is_sync(op)]++;
330 refcount_set(&rq->ref, 1);
334 static struct request *blk_mq_get_request(struct request_queue *q,
336 struct blk_mq_alloc_data *data)
338 struct elevator_queue *e = q->elevator;
341 bool clear_ctx_on_error = false;
342 u64 alloc_time_ns = 0;
344 blk_queue_enter_live(q);
346 /* alloc_time includes depth and tag waits */
347 if (blk_queue_rq_alloc_time(q))
348 alloc_time_ns = ktime_get_ns();
351 if (likely(!data->ctx)) {
352 data->ctx = blk_mq_get_ctx(q);
353 clear_ctx_on_error = true;
355 if (likely(!data->hctx))
356 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
358 if (data->cmd_flags & REQ_NOWAIT)
359 data->flags |= BLK_MQ_REQ_NOWAIT;
362 data->flags |= BLK_MQ_REQ_INTERNAL;
365 * Flush requests are special and go directly to the
366 * dispatch list. Don't include reserved tags in the
367 * limiting, as it isn't useful.
369 if (!op_is_flush(data->cmd_flags) &&
370 e->type->ops.limit_depth &&
371 !(data->flags & BLK_MQ_REQ_RESERVED))
372 e->type->ops.limit_depth(data->cmd_flags, data);
374 blk_mq_tag_busy(data->hctx);
377 tag = blk_mq_get_tag(data);
378 if (tag == BLK_MQ_TAG_FAIL) {
379 if (clear_ctx_on_error)
385 rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
386 if (!op_is_flush(data->cmd_flags)) {
388 if (e && e->type->ops.prepare_request) {
389 if (e->type->icq_cache)
390 blk_mq_sched_assign_ioc(rq);
392 e->type->ops.prepare_request(rq, bio);
393 rq->rq_flags |= RQF_ELVPRIV;
396 data->hctx->queued++;
400 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
401 blk_mq_req_flags_t flags)
403 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
407 ret = blk_queue_enter(q, flags);
411 rq = blk_mq_get_request(q, NULL, &alloc_data);
415 return ERR_PTR(-EWOULDBLOCK);
418 rq->__sector = (sector_t) -1;
419 rq->bio = rq->biotail = NULL;
422 EXPORT_SYMBOL(blk_mq_alloc_request);
424 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
425 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
427 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
433 * If the tag allocator sleeps we could get an allocation for a
434 * different hardware context. No need to complicate the low level
435 * allocator for this for the rare use case of a command tied to
438 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
439 return ERR_PTR(-EINVAL);
441 if (hctx_idx >= q->nr_hw_queues)
442 return ERR_PTR(-EIO);
444 ret = blk_queue_enter(q, flags);
449 * Check if the hardware context is actually mapped to anything.
450 * If not tell the caller that it should skip this queue.
452 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
453 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
455 return ERR_PTR(-EXDEV);
457 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
458 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
460 rq = blk_mq_get_request(q, NULL, &alloc_data);
464 return ERR_PTR(-EWOULDBLOCK);
468 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
470 static void __blk_mq_free_request(struct request *rq)
472 struct request_queue *q = rq->q;
473 struct blk_mq_ctx *ctx = rq->mq_ctx;
474 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
475 const int sched_tag = rq->internal_tag;
477 blk_pm_mark_last_busy(rq);
480 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
482 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
483 blk_mq_sched_restart(hctx);
487 void blk_mq_free_request(struct request *rq)
489 struct request_queue *q = rq->q;
490 struct elevator_queue *e = q->elevator;
491 struct blk_mq_ctx *ctx = rq->mq_ctx;
492 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
494 if (rq->rq_flags & RQF_ELVPRIV) {
495 if (e && e->type->ops.finish_request)
496 e->type->ops.finish_request(rq);
498 put_io_context(rq->elv.icq->ioc);
503 ctx->rq_completed[rq_is_sync(rq)]++;
504 if (rq->rq_flags & RQF_MQ_INFLIGHT)
505 atomic_dec(&hctx->nr_active);
507 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
508 laptop_io_completion(q->backing_dev_info);
512 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
513 if (refcount_dec_and_test(&rq->ref))
514 __blk_mq_free_request(rq);
516 EXPORT_SYMBOL_GPL(blk_mq_free_request);
518 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
522 if (blk_mq_need_time_stamp(rq))
523 now = ktime_get_ns();
525 if (rq->rq_flags & RQF_STATS) {
526 blk_mq_poll_stats_start(rq->q);
527 blk_stat_add(rq, now);
530 if (rq->internal_tag != -1)
531 blk_mq_sched_completed_request(rq, now);
533 blk_account_io_done(rq, now);
536 rq_qos_done(rq->q, rq);
537 rq->end_io(rq, error);
539 blk_mq_free_request(rq);
542 EXPORT_SYMBOL(__blk_mq_end_request);
544 void blk_mq_end_request(struct request *rq, blk_status_t error)
546 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
548 __blk_mq_end_request(rq, error);
550 EXPORT_SYMBOL(blk_mq_end_request);
552 static void __blk_mq_complete_request_remote(void *data)
554 struct request *rq = data;
555 struct request_queue *q = rq->q;
557 q->mq_ops->complete(rq);
560 static void __blk_mq_complete_request(struct request *rq)
562 struct blk_mq_ctx *ctx = rq->mq_ctx;
563 struct request_queue *q = rq->q;
567 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
569 * Most of single queue controllers, there is only one irq vector
570 * for handling IO completion, and the only irq's affinity is set
571 * as all possible CPUs. On most of ARCHs, this affinity means the
572 * irq is handled on one specific CPU.
574 * So complete IO reqeust in softirq context in case of single queue
575 * for not degrading IO performance by irqsoff latency.
577 if (q->nr_hw_queues == 1) {
578 __blk_complete_request(rq);
583 * For a polled request, always complete locallly, it's pointless
584 * to redirect the completion.
586 if ((rq->cmd_flags & REQ_HIPRI) ||
587 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
588 q->mq_ops->complete(rq);
593 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
594 shared = cpus_share_cache(cpu, ctx->cpu);
596 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
597 rq->csd.func = __blk_mq_complete_request_remote;
600 smp_call_function_single_async(ctx->cpu, &rq->csd);
602 q->mq_ops->complete(rq);
607 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
608 __releases(hctx->srcu)
610 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
613 srcu_read_unlock(hctx->srcu, srcu_idx);
616 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
617 __acquires(hctx->srcu)
619 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
620 /* shut up gcc false positive */
624 *srcu_idx = srcu_read_lock(hctx->srcu);
628 * blk_mq_complete_request - end I/O on a request
629 * @rq: the request being processed
632 * Ends all I/O on a request. It does not handle partial completions.
633 * The actual completion happens out-of-order, through a IPI handler.
635 bool blk_mq_complete_request(struct request *rq)
637 if (unlikely(blk_should_fake_timeout(rq->q)))
639 __blk_mq_complete_request(rq);
642 EXPORT_SYMBOL(blk_mq_complete_request);
645 * blk_mq_start_request - Start processing a request
646 * @rq: Pointer to request to be started
648 * Function used by device drivers to notify the block layer that a request
649 * is going to be processed now, so blk layer can do proper initializations
650 * such as starting the timeout timer.
652 void blk_mq_start_request(struct request *rq)
654 struct request_queue *q = rq->q;
656 trace_block_rq_issue(q, rq);
658 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
659 rq->io_start_time_ns = ktime_get_ns();
660 rq->stats_sectors = blk_rq_sectors(rq);
661 rq->rq_flags |= RQF_STATS;
665 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
668 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
670 if (q->dma_drain_size && blk_rq_bytes(rq)) {
672 * Make sure space for the drain appears. We know we can do
673 * this because max_hw_segments has been adjusted to be one
674 * fewer than the device can handle.
676 rq->nr_phys_segments++;
679 #ifdef CONFIG_BLK_DEV_INTEGRITY
680 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
681 q->integrity.profile->prepare_fn(rq);
684 EXPORT_SYMBOL(blk_mq_start_request);
686 static void __blk_mq_requeue_request(struct request *rq)
688 struct request_queue *q = rq->q;
690 blk_mq_put_driver_tag(rq);
692 trace_block_rq_requeue(q, rq);
693 rq_qos_requeue(q, rq);
695 if (blk_mq_request_started(rq)) {
696 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
697 rq->rq_flags &= ~RQF_TIMED_OUT;
698 if (q->dma_drain_size && blk_rq_bytes(rq))
699 rq->nr_phys_segments--;
703 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
705 __blk_mq_requeue_request(rq);
707 /* this request will be re-inserted to io scheduler queue */
708 blk_mq_sched_requeue_request(rq);
710 BUG_ON(!list_empty(&rq->queuelist));
711 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
713 EXPORT_SYMBOL(blk_mq_requeue_request);
715 static void blk_mq_requeue_work(struct work_struct *work)
717 struct request_queue *q =
718 container_of(work, struct request_queue, requeue_work.work);
720 struct request *rq, *next;
722 spin_lock_irq(&q->requeue_lock);
723 list_splice_init(&q->requeue_list, &rq_list);
724 spin_unlock_irq(&q->requeue_lock);
726 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
727 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
730 rq->rq_flags &= ~RQF_SOFTBARRIER;
731 list_del_init(&rq->queuelist);
733 * If RQF_DONTPREP, rq has contained some driver specific
734 * data, so insert it to hctx dispatch list to avoid any
737 if (rq->rq_flags & RQF_DONTPREP)
738 blk_mq_request_bypass_insert(rq, false, false);
740 blk_mq_sched_insert_request(rq, true, false, false);
743 while (!list_empty(&rq_list)) {
744 rq = list_entry(rq_list.next, struct request, queuelist);
745 list_del_init(&rq->queuelist);
746 blk_mq_sched_insert_request(rq, false, false, false);
749 blk_mq_run_hw_queues(q, false);
752 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
753 bool kick_requeue_list)
755 struct request_queue *q = rq->q;
759 * We abuse this flag that is otherwise used by the I/O scheduler to
760 * request head insertion from the workqueue.
762 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
764 spin_lock_irqsave(&q->requeue_lock, flags);
766 rq->rq_flags |= RQF_SOFTBARRIER;
767 list_add(&rq->queuelist, &q->requeue_list);
769 list_add_tail(&rq->queuelist, &q->requeue_list);
771 spin_unlock_irqrestore(&q->requeue_lock, flags);
773 if (kick_requeue_list)
774 blk_mq_kick_requeue_list(q);
777 void blk_mq_kick_requeue_list(struct request_queue *q)
779 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
781 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
783 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
786 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
787 msecs_to_jiffies(msecs));
789 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
791 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
793 if (tag < tags->nr_tags) {
794 prefetch(tags->rqs[tag]);
795 return tags->rqs[tag];
800 EXPORT_SYMBOL(blk_mq_tag_to_rq);
802 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
803 void *priv, bool reserved)
806 * If we find a request that is inflight and the queue matches,
807 * we know the queue is busy. Return false to stop the iteration.
809 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
819 bool blk_mq_queue_inflight(struct request_queue *q)
823 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
826 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
828 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
830 req->rq_flags |= RQF_TIMED_OUT;
831 if (req->q->mq_ops->timeout) {
832 enum blk_eh_timer_return ret;
834 ret = req->q->mq_ops->timeout(req, reserved);
835 if (ret == BLK_EH_DONE)
837 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
843 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
845 unsigned long deadline;
847 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
849 if (rq->rq_flags & RQF_TIMED_OUT)
852 deadline = READ_ONCE(rq->deadline);
853 if (time_after_eq(jiffies, deadline))
858 else if (time_after(*next, deadline))
863 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
864 struct request *rq, void *priv, bool reserved)
866 unsigned long *next = priv;
869 * Just do a quick check if it is expired before locking the request in
870 * so we're not unnecessarilly synchronizing across CPUs.
872 if (!blk_mq_req_expired(rq, next))
876 * We have reason to believe the request may be expired. Take a
877 * reference on the request to lock this request lifetime into its
878 * currently allocated context to prevent it from being reallocated in
879 * the event the completion by-passes this timeout handler.
881 * If the reference was already released, then the driver beat the
882 * timeout handler to posting a natural completion.
884 if (!refcount_inc_not_zero(&rq->ref))
888 * The request is now locked and cannot be reallocated underneath the
889 * timeout handler's processing. Re-verify this exact request is truly
890 * expired; if it is not expired, then the request was completed and
891 * reallocated as a new request.
893 if (blk_mq_req_expired(rq, next))
894 blk_mq_rq_timed_out(rq, reserved);
896 if (is_flush_rq(rq, hctx))
898 else if (refcount_dec_and_test(&rq->ref))
899 __blk_mq_free_request(rq);
904 static void blk_mq_timeout_work(struct work_struct *work)
906 struct request_queue *q =
907 container_of(work, struct request_queue, timeout_work);
908 unsigned long next = 0;
909 struct blk_mq_hw_ctx *hctx;
912 /* A deadlock might occur if a request is stuck requiring a
913 * timeout at the same time a queue freeze is waiting
914 * completion, since the timeout code would not be able to
915 * acquire the queue reference here.
917 * That's why we don't use blk_queue_enter here; instead, we use
918 * percpu_ref_tryget directly, because we need to be able to
919 * obtain a reference even in the short window between the queue
920 * starting to freeze, by dropping the first reference in
921 * blk_freeze_queue_start, and the moment the last request is
922 * consumed, marked by the instant q_usage_counter reaches
925 if (!percpu_ref_tryget(&q->q_usage_counter))
928 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
931 mod_timer(&q->timeout, next);
934 * Request timeouts are handled as a forward rolling timer. If
935 * we end up here it means that no requests are pending and
936 * also that no request has been pending for a while. Mark
939 queue_for_each_hw_ctx(q, hctx, i) {
940 /* the hctx may be unmapped, so check it here */
941 if (blk_mq_hw_queue_mapped(hctx))
942 blk_mq_tag_idle(hctx);
948 struct flush_busy_ctx_data {
949 struct blk_mq_hw_ctx *hctx;
950 struct list_head *list;
953 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
955 struct flush_busy_ctx_data *flush_data = data;
956 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
957 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
958 enum hctx_type type = hctx->type;
960 spin_lock(&ctx->lock);
961 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
962 sbitmap_clear_bit(sb, bitnr);
963 spin_unlock(&ctx->lock);
968 * Process software queues that have been marked busy, splicing them
969 * to the for-dispatch
971 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
973 struct flush_busy_ctx_data data = {
978 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
980 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
982 struct dispatch_rq_data {
983 struct blk_mq_hw_ctx *hctx;
987 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
990 struct dispatch_rq_data *dispatch_data = data;
991 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
992 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
993 enum hctx_type type = hctx->type;
995 spin_lock(&ctx->lock);
996 if (!list_empty(&ctx->rq_lists[type])) {
997 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
998 list_del_init(&dispatch_data->rq->queuelist);
999 if (list_empty(&ctx->rq_lists[type]))
1000 sbitmap_clear_bit(sb, bitnr);
1002 spin_unlock(&ctx->lock);
1004 return !dispatch_data->rq;
1007 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1008 struct blk_mq_ctx *start)
1010 unsigned off = start ? start->index_hw[hctx->type] : 0;
1011 struct dispatch_rq_data data = {
1016 __sbitmap_for_each_set(&hctx->ctx_map, off,
1017 dispatch_rq_from_ctx, &data);
1022 static inline unsigned int queued_to_index(unsigned int queued)
1027 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1030 bool blk_mq_get_driver_tag(struct request *rq)
1032 struct blk_mq_alloc_data data = {
1034 .hctx = rq->mq_hctx,
1035 .flags = BLK_MQ_REQ_NOWAIT,
1036 .cmd_flags = rq->cmd_flags,
1043 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1044 data.flags |= BLK_MQ_REQ_RESERVED;
1046 shared = blk_mq_tag_busy(data.hctx);
1047 rq->tag = blk_mq_get_tag(&data);
1050 rq->rq_flags |= RQF_MQ_INFLIGHT;
1051 atomic_inc(&data.hctx->nr_active);
1053 data.hctx->tags->rqs[rq->tag] = rq;
1056 return rq->tag != -1;
1059 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1060 int flags, void *key)
1062 struct blk_mq_hw_ctx *hctx;
1064 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1066 spin_lock(&hctx->dispatch_wait_lock);
1067 if (!list_empty(&wait->entry)) {
1068 struct sbitmap_queue *sbq;
1070 list_del_init(&wait->entry);
1071 sbq = &hctx->tags->bitmap_tags;
1072 atomic_dec(&sbq->ws_active);
1074 spin_unlock(&hctx->dispatch_wait_lock);
1076 blk_mq_run_hw_queue(hctx, true);
1081 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1082 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1083 * restart. For both cases, take care to check the condition again after
1084 * marking us as waiting.
1086 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1089 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1090 struct wait_queue_head *wq;
1091 wait_queue_entry_t *wait;
1094 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1095 blk_mq_sched_mark_restart_hctx(hctx);
1098 * It's possible that a tag was freed in the window between the
1099 * allocation failure and adding the hardware queue to the wait
1102 * Don't clear RESTART here, someone else could have set it.
1103 * At most this will cost an extra queue run.
1105 return blk_mq_get_driver_tag(rq);
1108 wait = &hctx->dispatch_wait;
1109 if (!list_empty_careful(&wait->entry))
1112 wq = &bt_wait_ptr(sbq, hctx)->wait;
1114 spin_lock_irq(&wq->lock);
1115 spin_lock(&hctx->dispatch_wait_lock);
1116 if (!list_empty(&wait->entry)) {
1117 spin_unlock(&hctx->dispatch_wait_lock);
1118 spin_unlock_irq(&wq->lock);
1122 atomic_inc(&sbq->ws_active);
1123 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1124 __add_wait_queue(wq, wait);
1127 * It's possible that a tag was freed in the window between the
1128 * allocation failure and adding the hardware queue to the wait
1131 ret = blk_mq_get_driver_tag(rq);
1133 spin_unlock(&hctx->dispatch_wait_lock);
1134 spin_unlock_irq(&wq->lock);
1139 * We got a tag, remove ourselves from the wait queue to ensure
1140 * someone else gets the wakeup.
1142 list_del_init(&wait->entry);
1143 atomic_dec(&sbq->ws_active);
1144 spin_unlock(&hctx->dispatch_wait_lock);
1145 spin_unlock_irq(&wq->lock);
1150 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1151 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1153 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1154 * - EWMA is one simple way to compute running average value
1155 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1156 * - take 4 as factor for avoiding to get too small(0) result, and this
1157 * factor doesn't matter because EWMA decreases exponentially
1159 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1163 if (hctx->queue->elevator)
1166 ewma = hctx->dispatch_busy;
1171 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1173 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1174 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1176 hctx->dispatch_busy = ewma;
1179 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1181 static void blk_mq_handle_dev_resource(struct request *rq,
1182 struct list_head *list)
1184 struct request *next =
1185 list_first_entry_or_null(list, struct request, queuelist);
1188 * If an I/O scheduler has been configured and we got a driver tag for
1189 * the next request already, free it.
1192 blk_mq_put_driver_tag(next);
1194 list_add(&rq->queuelist, list);
1195 __blk_mq_requeue_request(rq);
1199 * Returns true if we did some work AND can potentially do more.
1201 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1204 struct blk_mq_hw_ctx *hctx;
1205 struct request *rq, *nxt;
1206 bool no_tag = false;
1208 blk_status_t ret = BLK_STS_OK;
1210 if (list_empty(list))
1213 WARN_ON(!list_is_singular(list) && got_budget);
1216 * Now process all the entries, sending them to the driver.
1218 errors = queued = 0;
1220 struct blk_mq_queue_data bd;
1222 rq = list_first_entry(list, struct request, queuelist);
1225 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1228 if (!blk_mq_get_driver_tag(rq)) {
1230 * The initial allocation attempt failed, so we need to
1231 * rerun the hardware queue when a tag is freed. The
1232 * waitqueue takes care of that. If the queue is run
1233 * before we add this entry back on the dispatch list,
1234 * we'll re-run it below.
1236 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1237 blk_mq_put_dispatch_budget(hctx);
1239 * For non-shared tags, the RESTART check
1242 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1248 list_del_init(&rq->queuelist);
1253 * Flag last if we have no more requests, or if we have more
1254 * but can't assign a driver tag to it.
1256 if (list_empty(list))
1259 nxt = list_first_entry(list, struct request, queuelist);
1260 bd.last = !blk_mq_get_driver_tag(nxt);
1263 ret = q->mq_ops->queue_rq(hctx, &bd);
1264 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1265 blk_mq_handle_dev_resource(rq, list);
1269 if (unlikely(ret != BLK_STS_OK)) {
1271 blk_mq_end_request(rq, BLK_STS_IOERR);
1276 } while (!list_empty(list));
1278 hctx->dispatched[queued_to_index(queued)]++;
1281 * Any items that need requeuing? Stuff them into hctx->dispatch,
1282 * that is where we will continue on next queue run.
1284 if (!list_empty(list)) {
1288 * If we didn't flush the entire list, we could have told
1289 * the driver there was more coming, but that turned out to
1292 if (q->mq_ops->commit_rqs && queued)
1293 q->mq_ops->commit_rqs(hctx);
1295 spin_lock(&hctx->lock);
1296 list_splice_tail_init(list, &hctx->dispatch);
1297 spin_unlock(&hctx->lock);
1300 * If SCHED_RESTART was set by the caller of this function and
1301 * it is no longer set that means that it was cleared by another
1302 * thread and hence that a queue rerun is needed.
1304 * If 'no_tag' is set, that means that we failed getting
1305 * a driver tag with an I/O scheduler attached. If our dispatch
1306 * waitqueue is no longer active, ensure that we run the queue
1307 * AFTER adding our entries back to the list.
1309 * If no I/O scheduler has been configured it is possible that
1310 * the hardware queue got stopped and restarted before requests
1311 * were pushed back onto the dispatch list. Rerun the queue to
1312 * avoid starvation. Notes:
1313 * - blk_mq_run_hw_queue() checks whether or not a queue has
1314 * been stopped before rerunning a queue.
1315 * - Some but not all block drivers stop a queue before
1316 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1319 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1320 * bit is set, run queue after a delay to avoid IO stalls
1321 * that could otherwise occur if the queue is idle.
1323 needs_restart = blk_mq_sched_needs_restart(hctx);
1324 if (!needs_restart ||
1325 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1326 blk_mq_run_hw_queue(hctx, true);
1327 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1328 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1330 blk_mq_update_dispatch_busy(hctx, true);
1333 blk_mq_update_dispatch_busy(hctx, false);
1336 * If the host/device is unable to accept more work, inform the
1339 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1342 return (queued + errors) != 0;
1346 * __blk_mq_run_hw_queue - Run a hardware queue.
1347 * @hctx: Pointer to the hardware queue to run.
1349 * Send pending requests to the hardware.
1351 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1356 * We should be running this queue from one of the CPUs that
1359 * There are at least two related races now between setting
1360 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1361 * __blk_mq_run_hw_queue():
1363 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1364 * but later it becomes online, then this warning is harmless
1367 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1368 * but later it becomes offline, then the warning can't be
1369 * triggered, and we depend on blk-mq timeout handler to
1370 * handle dispatched requests to this hctx
1372 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1373 cpu_online(hctx->next_cpu)) {
1374 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1375 raw_smp_processor_id(),
1376 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1381 * We can't run the queue inline with ints disabled. Ensure that
1382 * we catch bad users of this early.
1384 WARN_ON_ONCE(in_interrupt());
1386 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1388 hctx_lock(hctx, &srcu_idx);
1389 blk_mq_sched_dispatch_requests(hctx);
1390 hctx_unlock(hctx, srcu_idx);
1393 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1395 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1397 if (cpu >= nr_cpu_ids)
1398 cpu = cpumask_first(hctx->cpumask);
1403 * It'd be great if the workqueue API had a way to pass
1404 * in a mask and had some smarts for more clever placement.
1405 * For now we just round-robin here, switching for every
1406 * BLK_MQ_CPU_WORK_BATCH queued items.
1408 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1411 int next_cpu = hctx->next_cpu;
1413 if (hctx->queue->nr_hw_queues == 1)
1414 return WORK_CPU_UNBOUND;
1416 if (--hctx->next_cpu_batch <= 0) {
1418 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1420 if (next_cpu >= nr_cpu_ids)
1421 next_cpu = blk_mq_first_mapped_cpu(hctx);
1422 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1426 * Do unbound schedule if we can't find a online CPU for this hctx,
1427 * and it should only happen in the path of handling CPU DEAD.
1429 if (!cpu_online(next_cpu)) {
1436 * Make sure to re-select CPU next time once after CPUs
1437 * in hctx->cpumask become online again.
1439 hctx->next_cpu = next_cpu;
1440 hctx->next_cpu_batch = 1;
1441 return WORK_CPU_UNBOUND;
1444 hctx->next_cpu = next_cpu;
1449 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1450 * @hctx: Pointer to the hardware queue to run.
1451 * @async: If we want to run the queue asynchronously.
1452 * @msecs: Microseconds of delay to wait before running the queue.
1454 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1455 * with a delay of @msecs.
1457 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1458 unsigned long msecs)
1460 if (unlikely(blk_mq_hctx_stopped(hctx)))
1463 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1464 int cpu = get_cpu();
1465 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1466 __blk_mq_run_hw_queue(hctx);
1474 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1475 msecs_to_jiffies(msecs));
1479 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1480 * @hctx: Pointer to the hardware queue to run.
1481 * @msecs: Microseconds of delay to wait before running the queue.
1483 * Run a hardware queue asynchronously with a delay of @msecs.
1485 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1487 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1489 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1492 * blk_mq_run_hw_queue - Start to run a hardware queue.
1493 * @hctx: Pointer to the hardware queue to run.
1494 * @async: If we want to run the queue asynchronously.
1496 * Check if the request queue is not in a quiesced state and if there are
1497 * pending requests to be sent. If this is true, run the queue to send requests
1500 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1506 * When queue is quiesced, we may be switching io scheduler, or
1507 * updating nr_hw_queues, or other things, and we can't run queue
1508 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1510 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1513 hctx_lock(hctx, &srcu_idx);
1514 need_run = !blk_queue_quiesced(hctx->queue) &&
1515 blk_mq_hctx_has_pending(hctx);
1516 hctx_unlock(hctx, srcu_idx);
1519 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1521 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1524 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1525 * @q: Pointer to the request queue to run.
1526 * @async: If we want to run the queue asynchronously.
1528 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1530 struct blk_mq_hw_ctx *hctx;
1533 queue_for_each_hw_ctx(q, hctx, i) {
1534 if (blk_mq_hctx_stopped(hctx))
1537 blk_mq_run_hw_queue(hctx, async);
1540 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1543 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1544 * @q: request queue.
1546 * The caller is responsible for serializing this function against
1547 * blk_mq_{start,stop}_hw_queue().
1549 bool blk_mq_queue_stopped(struct request_queue *q)
1551 struct blk_mq_hw_ctx *hctx;
1554 queue_for_each_hw_ctx(q, hctx, i)
1555 if (blk_mq_hctx_stopped(hctx))
1560 EXPORT_SYMBOL(blk_mq_queue_stopped);
1563 * This function is often used for pausing .queue_rq() by driver when
1564 * there isn't enough resource or some conditions aren't satisfied, and
1565 * BLK_STS_RESOURCE is usually returned.
1567 * We do not guarantee that dispatch can be drained or blocked
1568 * after blk_mq_stop_hw_queue() returns. Please use
1569 * blk_mq_quiesce_queue() for that requirement.
1571 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1573 cancel_delayed_work(&hctx->run_work);
1575 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1577 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1580 * This function is often used for pausing .queue_rq() by driver when
1581 * there isn't enough resource or some conditions aren't satisfied, and
1582 * BLK_STS_RESOURCE is usually returned.
1584 * We do not guarantee that dispatch can be drained or blocked
1585 * after blk_mq_stop_hw_queues() returns. Please use
1586 * blk_mq_quiesce_queue() for that requirement.
1588 void blk_mq_stop_hw_queues(struct request_queue *q)
1590 struct blk_mq_hw_ctx *hctx;
1593 queue_for_each_hw_ctx(q, hctx, i)
1594 blk_mq_stop_hw_queue(hctx);
1596 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1598 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1600 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1602 blk_mq_run_hw_queue(hctx, false);
1604 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1606 void blk_mq_start_hw_queues(struct request_queue *q)
1608 struct blk_mq_hw_ctx *hctx;
1611 queue_for_each_hw_ctx(q, hctx, i)
1612 blk_mq_start_hw_queue(hctx);
1614 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1616 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1618 if (!blk_mq_hctx_stopped(hctx))
1621 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1622 blk_mq_run_hw_queue(hctx, async);
1624 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1626 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1628 struct blk_mq_hw_ctx *hctx;
1631 queue_for_each_hw_ctx(q, hctx, i)
1632 blk_mq_start_stopped_hw_queue(hctx, async);
1634 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1636 static void blk_mq_run_work_fn(struct work_struct *work)
1638 struct blk_mq_hw_ctx *hctx;
1640 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1643 * If we are stopped, don't run the queue.
1645 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1648 __blk_mq_run_hw_queue(hctx);
1651 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1655 struct blk_mq_ctx *ctx = rq->mq_ctx;
1656 enum hctx_type type = hctx->type;
1658 lockdep_assert_held(&ctx->lock);
1660 trace_block_rq_insert(hctx->queue, rq);
1663 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1665 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1668 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1671 struct blk_mq_ctx *ctx = rq->mq_ctx;
1673 lockdep_assert_held(&ctx->lock);
1675 __blk_mq_insert_req_list(hctx, rq, at_head);
1676 blk_mq_hctx_mark_pending(hctx, ctx);
1680 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1681 * @rq: Pointer to request to be inserted.
1682 * @run_queue: If we should run the hardware queue after inserting the request.
1684 * Should only be used carefully, when the caller knows we want to
1685 * bypass a potential IO scheduler on the target device.
1687 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1690 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1692 spin_lock(&hctx->lock);
1694 list_add(&rq->queuelist, &hctx->dispatch);
1696 list_add_tail(&rq->queuelist, &hctx->dispatch);
1697 spin_unlock(&hctx->lock);
1700 blk_mq_run_hw_queue(hctx, false);
1703 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1704 struct list_head *list)
1708 enum hctx_type type = hctx->type;
1711 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1714 list_for_each_entry(rq, list, queuelist) {
1715 BUG_ON(rq->mq_ctx != ctx);
1716 trace_block_rq_insert(hctx->queue, rq);
1719 spin_lock(&ctx->lock);
1720 list_splice_tail_init(list, &ctx->rq_lists[type]);
1721 blk_mq_hctx_mark_pending(hctx, ctx);
1722 spin_unlock(&ctx->lock);
1725 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1727 struct request *rqa = container_of(a, struct request, queuelist);
1728 struct request *rqb = container_of(b, struct request, queuelist);
1730 if (rqa->mq_ctx != rqb->mq_ctx)
1731 return rqa->mq_ctx > rqb->mq_ctx;
1732 if (rqa->mq_hctx != rqb->mq_hctx)
1733 return rqa->mq_hctx > rqb->mq_hctx;
1735 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1738 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1742 if (list_empty(&plug->mq_list))
1744 list_splice_init(&plug->mq_list, &list);
1746 if (plug->rq_count > 2 && plug->multiple_queues)
1747 list_sort(NULL, &list, plug_rq_cmp);
1752 struct list_head rq_list;
1753 struct request *rq, *head_rq = list_entry_rq(list.next);
1754 struct list_head *pos = &head_rq->queuelist; /* skip first */
1755 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1756 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1757 unsigned int depth = 1;
1759 list_for_each_continue(pos, &list) {
1760 rq = list_entry_rq(pos);
1762 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1767 list_cut_before(&rq_list, &list, pos);
1768 trace_block_unplug(head_rq->q, depth, !from_schedule);
1769 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1771 } while(!list_empty(&list));
1774 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1775 unsigned int nr_segs)
1777 if (bio->bi_opf & REQ_RAHEAD)
1778 rq->cmd_flags |= REQ_FAILFAST_MASK;
1780 rq->__sector = bio->bi_iter.bi_sector;
1781 rq->write_hint = bio->bi_write_hint;
1782 blk_rq_bio_prep(rq, bio, nr_segs);
1784 blk_account_io_start(rq, true);
1787 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1789 blk_qc_t *cookie, bool last)
1791 struct request_queue *q = rq->q;
1792 struct blk_mq_queue_data bd = {
1796 blk_qc_t new_cookie;
1799 new_cookie = request_to_qc_t(hctx, rq);
1802 * For OK queue, we are done. For error, caller may kill it.
1803 * Any other error (busy), just add it to our list as we
1804 * previously would have done.
1806 ret = q->mq_ops->queue_rq(hctx, &bd);
1809 blk_mq_update_dispatch_busy(hctx, false);
1810 *cookie = new_cookie;
1812 case BLK_STS_RESOURCE:
1813 case BLK_STS_DEV_RESOURCE:
1814 blk_mq_update_dispatch_busy(hctx, true);
1815 __blk_mq_requeue_request(rq);
1818 blk_mq_update_dispatch_busy(hctx, false);
1819 *cookie = BLK_QC_T_NONE;
1826 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1829 bool bypass_insert, bool last)
1831 struct request_queue *q = rq->q;
1832 bool run_queue = true;
1835 * RCU or SRCU read lock is needed before checking quiesced flag.
1837 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1838 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1839 * and avoid driver to try to dispatch again.
1841 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1843 bypass_insert = false;
1847 if (q->elevator && !bypass_insert)
1850 if (!blk_mq_get_dispatch_budget(hctx))
1853 if (!blk_mq_get_driver_tag(rq)) {
1854 blk_mq_put_dispatch_budget(hctx);
1858 return __blk_mq_issue_directly(hctx, rq, cookie, last);
1861 return BLK_STS_RESOURCE;
1863 blk_mq_request_bypass_insert(rq, false, run_queue);
1868 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
1869 * @hctx: Pointer of the associated hardware queue.
1870 * @rq: Pointer to request to be sent.
1871 * @cookie: Request queue cookie.
1873 * If the device has enough resources to accept a new request now, send the
1874 * request directly to device driver. Else, insert at hctx->dispatch queue, so
1875 * we can try send it another time in the future. Requests inserted at this
1876 * queue have higher priority.
1878 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1879 struct request *rq, blk_qc_t *cookie)
1884 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1886 hctx_lock(hctx, &srcu_idx);
1888 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1889 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1890 blk_mq_request_bypass_insert(rq, false, true);
1891 else if (ret != BLK_STS_OK)
1892 blk_mq_end_request(rq, ret);
1894 hctx_unlock(hctx, srcu_idx);
1897 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1901 blk_qc_t unused_cookie;
1902 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1904 hctx_lock(hctx, &srcu_idx);
1905 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1906 hctx_unlock(hctx, srcu_idx);
1911 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1912 struct list_head *list)
1916 while (!list_empty(list)) {
1918 struct request *rq = list_first_entry(list, struct request,
1921 list_del_init(&rq->queuelist);
1922 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1923 if (ret != BLK_STS_OK) {
1924 if (ret == BLK_STS_RESOURCE ||
1925 ret == BLK_STS_DEV_RESOURCE) {
1926 blk_mq_request_bypass_insert(rq, false,
1930 blk_mq_end_request(rq, ret);
1936 * If we didn't flush the entire list, we could have told
1937 * the driver there was more coming, but that turned out to
1940 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs && queued)
1941 hctx->queue->mq_ops->commit_rqs(hctx);
1944 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1946 list_add_tail(&rq->queuelist, &plug->mq_list);
1948 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1949 struct request *tmp;
1951 tmp = list_first_entry(&plug->mq_list, struct request,
1953 if (tmp->q != rq->q)
1954 plug->multiple_queues = true;
1959 * blk_mq_make_request - Create and send a request to block device.
1960 * @q: Request queue pointer.
1961 * @bio: Bio pointer.
1963 * Builds up a request structure from @q and @bio and send to the device. The
1964 * request may not be queued directly to hardware if:
1965 * * This request can be merged with another one
1966 * * We want to place request at plug queue for possible future merging
1967 * * There is an IO scheduler active at this queue
1969 * It will not queue the request if there is an error with the bio, or at the
1972 * Returns: Request queue cookie.
1974 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1976 const int is_sync = op_is_sync(bio->bi_opf);
1977 const int is_flush_fua = op_is_flush(bio->bi_opf);
1978 struct blk_mq_alloc_data data = { .flags = 0};
1980 struct blk_plug *plug;
1981 struct request *same_queue_rq = NULL;
1982 unsigned int nr_segs;
1985 blk_queue_bounce(q, &bio);
1986 __blk_queue_split(q, &bio, &nr_segs);
1988 if (!bio_integrity_prep(bio))
1989 return BLK_QC_T_NONE;
1991 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1992 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1993 return BLK_QC_T_NONE;
1995 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1996 return BLK_QC_T_NONE;
1998 rq_qos_throttle(q, bio);
2000 data.cmd_flags = bio->bi_opf;
2001 rq = blk_mq_get_request(q, bio, &data);
2002 if (unlikely(!rq)) {
2003 rq_qos_cleanup(q, bio);
2004 if (bio->bi_opf & REQ_NOWAIT)
2005 bio_wouldblock_error(bio);
2006 return BLK_QC_T_NONE;
2009 trace_block_getrq(q, bio, bio->bi_opf);
2011 rq_qos_track(q, rq, bio);
2013 cookie = request_to_qc_t(data.hctx, rq);
2015 blk_mq_bio_to_request(rq, bio, nr_segs);
2017 plug = blk_mq_plug(q, bio);
2018 if (unlikely(is_flush_fua)) {
2019 /* Bypass scheduler for flush requests */
2020 blk_insert_flush(rq);
2021 blk_mq_run_hw_queue(data.hctx, true);
2022 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2023 !blk_queue_nonrot(q))) {
2025 * Use plugging if we have a ->commit_rqs() hook as well, as
2026 * we know the driver uses bd->last in a smart fashion.
2028 * Use normal plugging if this disk is slow HDD, as sequential
2029 * IO may benefit a lot from plug merging.
2031 unsigned int request_count = plug->rq_count;
2032 struct request *last = NULL;
2035 trace_block_plug(q);
2037 last = list_entry_rq(plug->mq_list.prev);
2039 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2040 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2041 blk_flush_plug_list(plug, false);
2042 trace_block_plug(q);
2045 blk_add_rq_to_plug(plug, rq);
2046 } else if (q->elevator) {
2047 /* Insert the request at the IO scheduler queue */
2048 blk_mq_sched_insert_request(rq, false, true, true);
2049 } else if (plug && !blk_queue_nomerges(q)) {
2051 * We do limited plugging. If the bio can be merged, do that.
2052 * Otherwise the existing request in the plug list will be
2053 * issued. So the plug list will have one request at most
2054 * The plug list might get flushed before this. If that happens,
2055 * the plug list is empty, and same_queue_rq is invalid.
2057 if (list_empty(&plug->mq_list))
2058 same_queue_rq = NULL;
2059 if (same_queue_rq) {
2060 list_del_init(&same_queue_rq->queuelist);
2063 blk_add_rq_to_plug(plug, rq);
2064 trace_block_plug(q);
2066 if (same_queue_rq) {
2067 data.hctx = same_queue_rq->mq_hctx;
2068 trace_block_unplug(q, 1, true);
2069 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2072 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2073 !data.hctx->dispatch_busy) {
2075 * There is no scheduler and we can try to send directly
2078 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2081 blk_mq_sched_insert_request(rq, false, true, true);
2087 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2088 unsigned int hctx_idx)
2092 if (tags->rqs && set->ops->exit_request) {
2095 for (i = 0; i < tags->nr_tags; i++) {
2096 struct request *rq = tags->static_rqs[i];
2100 set->ops->exit_request(set, rq, hctx_idx);
2101 tags->static_rqs[i] = NULL;
2105 while (!list_empty(&tags->page_list)) {
2106 page = list_first_entry(&tags->page_list, struct page, lru);
2107 list_del_init(&page->lru);
2109 * Remove kmemleak object previously allocated in
2110 * blk_mq_alloc_rqs().
2112 kmemleak_free(page_address(page));
2113 __free_pages(page, page->private);
2117 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2121 kfree(tags->static_rqs);
2122 tags->static_rqs = NULL;
2124 blk_mq_free_tags(tags);
2127 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2128 unsigned int hctx_idx,
2129 unsigned int nr_tags,
2130 unsigned int reserved_tags)
2132 struct blk_mq_tags *tags;
2135 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2136 if (node == NUMA_NO_NODE)
2137 node = set->numa_node;
2139 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2140 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2144 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2145 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2148 blk_mq_free_tags(tags);
2152 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2153 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2155 if (!tags->static_rqs) {
2157 blk_mq_free_tags(tags);
2164 static size_t order_to_size(unsigned int order)
2166 return (size_t)PAGE_SIZE << order;
2169 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2170 unsigned int hctx_idx, int node)
2174 if (set->ops->init_request) {
2175 ret = set->ops->init_request(set, rq, hctx_idx, node);
2180 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2184 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2185 unsigned int hctx_idx, unsigned int depth)
2187 unsigned int i, j, entries_per_page, max_order = 4;
2188 size_t rq_size, left;
2191 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2192 if (node == NUMA_NO_NODE)
2193 node = set->numa_node;
2195 INIT_LIST_HEAD(&tags->page_list);
2198 * rq_size is the size of the request plus driver payload, rounded
2199 * to the cacheline size
2201 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2203 left = rq_size * depth;
2205 for (i = 0; i < depth; ) {
2206 int this_order = max_order;
2211 while (this_order && left < order_to_size(this_order - 1))
2215 page = alloc_pages_node(node,
2216 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2222 if (order_to_size(this_order) < rq_size)
2229 page->private = this_order;
2230 list_add_tail(&page->lru, &tags->page_list);
2232 p = page_address(page);
2234 * Allow kmemleak to scan these pages as they contain pointers
2235 * to additional allocations like via ops->init_request().
2237 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2238 entries_per_page = order_to_size(this_order) / rq_size;
2239 to_do = min(entries_per_page, depth - i);
2240 left -= to_do * rq_size;
2241 for (j = 0; j < to_do; j++) {
2242 struct request *rq = p;
2244 tags->static_rqs[i] = rq;
2245 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2246 tags->static_rqs[i] = NULL;
2257 blk_mq_free_rqs(set, tags, hctx_idx);
2262 * 'cpu' is going away. splice any existing rq_list entries from this
2263 * software queue to the hw queue dispatch list, and ensure that it
2266 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2268 struct blk_mq_hw_ctx *hctx;
2269 struct blk_mq_ctx *ctx;
2271 enum hctx_type type;
2273 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2274 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2277 spin_lock(&ctx->lock);
2278 if (!list_empty(&ctx->rq_lists[type])) {
2279 list_splice_init(&ctx->rq_lists[type], &tmp);
2280 blk_mq_hctx_clear_pending(hctx, ctx);
2282 spin_unlock(&ctx->lock);
2284 if (list_empty(&tmp))
2287 spin_lock(&hctx->lock);
2288 list_splice_tail_init(&tmp, &hctx->dispatch);
2289 spin_unlock(&hctx->lock);
2291 blk_mq_run_hw_queue(hctx, true);
2295 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2297 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2301 /* hctx->ctxs will be freed in queue's release handler */
2302 static void blk_mq_exit_hctx(struct request_queue *q,
2303 struct blk_mq_tag_set *set,
2304 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2306 if (blk_mq_hw_queue_mapped(hctx))
2307 blk_mq_tag_idle(hctx);
2309 if (set->ops->exit_request)
2310 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2312 if (set->ops->exit_hctx)
2313 set->ops->exit_hctx(hctx, hctx_idx);
2315 blk_mq_remove_cpuhp(hctx);
2317 spin_lock(&q->unused_hctx_lock);
2318 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2319 spin_unlock(&q->unused_hctx_lock);
2322 static void blk_mq_exit_hw_queues(struct request_queue *q,
2323 struct blk_mq_tag_set *set, int nr_queue)
2325 struct blk_mq_hw_ctx *hctx;
2328 queue_for_each_hw_ctx(q, hctx, i) {
2331 blk_mq_debugfs_unregister_hctx(hctx);
2332 blk_mq_exit_hctx(q, set, hctx, i);
2336 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2338 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2340 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2341 __alignof__(struct blk_mq_hw_ctx)) !=
2342 sizeof(struct blk_mq_hw_ctx));
2344 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2345 hw_ctx_size += sizeof(struct srcu_struct);
2350 static int blk_mq_init_hctx(struct request_queue *q,
2351 struct blk_mq_tag_set *set,
2352 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2354 hctx->queue_num = hctx_idx;
2356 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2358 hctx->tags = set->tags[hctx_idx];
2360 if (set->ops->init_hctx &&
2361 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2362 goto unregister_cpu_notifier;
2364 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2370 if (set->ops->exit_hctx)
2371 set->ops->exit_hctx(hctx, hctx_idx);
2372 unregister_cpu_notifier:
2373 blk_mq_remove_cpuhp(hctx);
2377 static struct blk_mq_hw_ctx *
2378 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2381 struct blk_mq_hw_ctx *hctx;
2382 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2384 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2386 goto fail_alloc_hctx;
2388 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2391 atomic_set(&hctx->nr_active, 0);
2392 if (node == NUMA_NO_NODE)
2393 node = set->numa_node;
2394 hctx->numa_node = node;
2396 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2397 spin_lock_init(&hctx->lock);
2398 INIT_LIST_HEAD(&hctx->dispatch);
2400 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2402 INIT_LIST_HEAD(&hctx->hctx_list);
2405 * Allocate space for all possible cpus to avoid allocation at
2408 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2413 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2418 spin_lock_init(&hctx->dispatch_wait_lock);
2419 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2420 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2422 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2426 if (hctx->flags & BLK_MQ_F_BLOCKING)
2427 init_srcu_struct(hctx->srcu);
2428 blk_mq_hctx_kobj_init(hctx);
2433 sbitmap_free(&hctx->ctx_map);
2437 free_cpumask_var(hctx->cpumask);
2444 static void blk_mq_init_cpu_queues(struct request_queue *q,
2445 unsigned int nr_hw_queues)
2447 struct blk_mq_tag_set *set = q->tag_set;
2450 for_each_possible_cpu(i) {
2451 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2452 struct blk_mq_hw_ctx *hctx;
2456 spin_lock_init(&__ctx->lock);
2457 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2458 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2463 * Set local node, IFF we have more than one hw queue. If
2464 * not, we remain on the home node of the device
2466 for (j = 0; j < set->nr_maps; j++) {
2467 hctx = blk_mq_map_queue_type(q, j, i);
2468 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2469 hctx->numa_node = local_memory_node(cpu_to_node(i));
2474 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2478 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2479 set->queue_depth, set->reserved_tags);
2480 if (!set->tags[hctx_idx])
2483 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2488 blk_mq_free_rq_map(set->tags[hctx_idx]);
2489 set->tags[hctx_idx] = NULL;
2493 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2494 unsigned int hctx_idx)
2496 if (set->tags && set->tags[hctx_idx]) {
2497 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2498 blk_mq_free_rq_map(set->tags[hctx_idx]);
2499 set->tags[hctx_idx] = NULL;
2503 static void blk_mq_map_swqueue(struct request_queue *q)
2505 unsigned int i, j, hctx_idx;
2506 struct blk_mq_hw_ctx *hctx;
2507 struct blk_mq_ctx *ctx;
2508 struct blk_mq_tag_set *set = q->tag_set;
2510 queue_for_each_hw_ctx(q, hctx, i) {
2511 cpumask_clear(hctx->cpumask);
2513 hctx->dispatch_from = NULL;
2517 * Map software to hardware queues.
2519 * If the cpu isn't present, the cpu is mapped to first hctx.
2521 for_each_possible_cpu(i) {
2522 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2523 /* unmapped hw queue can be remapped after CPU topo changed */
2524 if (!set->tags[hctx_idx] &&
2525 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2527 * If tags initialization fail for some hctx,
2528 * that hctx won't be brought online. In this
2529 * case, remap the current ctx to hctx[0] which
2530 * is guaranteed to always have tags allocated
2532 set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2535 ctx = per_cpu_ptr(q->queue_ctx, i);
2536 for (j = 0; j < set->nr_maps; j++) {
2537 if (!set->map[j].nr_queues) {
2538 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2539 HCTX_TYPE_DEFAULT, i);
2543 hctx = blk_mq_map_queue_type(q, j, i);
2544 ctx->hctxs[j] = hctx;
2546 * If the CPU is already set in the mask, then we've
2547 * mapped this one already. This can happen if
2548 * devices share queues across queue maps.
2550 if (cpumask_test_cpu(i, hctx->cpumask))
2553 cpumask_set_cpu(i, hctx->cpumask);
2555 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2556 hctx->ctxs[hctx->nr_ctx++] = ctx;
2559 * If the nr_ctx type overflows, we have exceeded the
2560 * amount of sw queues we can support.
2562 BUG_ON(!hctx->nr_ctx);
2565 for (; j < HCTX_MAX_TYPES; j++)
2566 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2567 HCTX_TYPE_DEFAULT, i);
2570 queue_for_each_hw_ctx(q, hctx, i) {
2572 * If no software queues are mapped to this hardware queue,
2573 * disable it and free the request entries.
2575 if (!hctx->nr_ctx) {
2576 /* Never unmap queue 0. We need it as a
2577 * fallback in case of a new remap fails
2580 if (i && set->tags[i])
2581 blk_mq_free_map_and_requests(set, i);
2587 hctx->tags = set->tags[i];
2588 WARN_ON(!hctx->tags);
2591 * Set the map size to the number of mapped software queues.
2592 * This is more accurate and more efficient than looping
2593 * over all possibly mapped software queues.
2595 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2598 * Initialize batch roundrobin counts
2600 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2601 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2606 * Caller needs to ensure that we're either frozen/quiesced, or that
2607 * the queue isn't live yet.
2609 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2611 struct blk_mq_hw_ctx *hctx;
2614 queue_for_each_hw_ctx(q, hctx, i) {
2616 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2618 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2622 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2625 struct request_queue *q;
2627 lockdep_assert_held(&set->tag_list_lock);
2629 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2630 blk_mq_freeze_queue(q);
2631 queue_set_hctx_shared(q, shared);
2632 blk_mq_unfreeze_queue(q);
2636 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2638 struct blk_mq_tag_set *set = q->tag_set;
2640 mutex_lock(&set->tag_list_lock);
2641 list_del_rcu(&q->tag_set_list);
2642 if (list_is_singular(&set->tag_list)) {
2643 /* just transitioned to unshared */
2644 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2645 /* update existing queue */
2646 blk_mq_update_tag_set_depth(set, false);
2648 mutex_unlock(&set->tag_list_lock);
2649 INIT_LIST_HEAD(&q->tag_set_list);
2652 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2653 struct request_queue *q)
2655 mutex_lock(&set->tag_list_lock);
2658 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2660 if (!list_empty(&set->tag_list) &&
2661 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2662 set->flags |= BLK_MQ_F_TAG_SHARED;
2663 /* update existing queue */
2664 blk_mq_update_tag_set_depth(set, true);
2666 if (set->flags & BLK_MQ_F_TAG_SHARED)
2667 queue_set_hctx_shared(q, true);
2668 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2670 mutex_unlock(&set->tag_list_lock);
2673 /* All allocations will be freed in release handler of q->mq_kobj */
2674 static int blk_mq_alloc_ctxs(struct request_queue *q)
2676 struct blk_mq_ctxs *ctxs;
2679 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2683 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2684 if (!ctxs->queue_ctx)
2687 for_each_possible_cpu(cpu) {
2688 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2692 q->mq_kobj = &ctxs->kobj;
2693 q->queue_ctx = ctxs->queue_ctx;
2702 * It is the actual release handler for mq, but we do it from
2703 * request queue's release handler for avoiding use-after-free
2704 * and headache because q->mq_kobj shouldn't have been introduced,
2705 * but we can't group ctx/kctx kobj without it.
2707 void blk_mq_release(struct request_queue *q)
2709 struct blk_mq_hw_ctx *hctx, *next;
2712 queue_for_each_hw_ctx(q, hctx, i)
2713 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2715 /* all hctx are in .unused_hctx_list now */
2716 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2717 list_del_init(&hctx->hctx_list);
2718 kobject_put(&hctx->kobj);
2721 kfree(q->queue_hw_ctx);
2724 * release .mq_kobj and sw queue's kobject now because
2725 * both share lifetime with request queue.
2727 blk_mq_sysfs_deinit(q);
2730 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
2733 struct request_queue *uninit_q, *q;
2735 uninit_q = __blk_alloc_queue(set->numa_node);
2737 return ERR_PTR(-ENOMEM);
2738 uninit_q->queuedata = queuedata;
2741 * Initialize the queue without an elevator. device_add_disk() will do
2742 * the initialization.
2744 q = blk_mq_init_allocated_queue(set, uninit_q, false);
2746 blk_cleanup_queue(uninit_q);
2750 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
2752 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2754 return blk_mq_init_queue_data(set, NULL);
2756 EXPORT_SYMBOL(blk_mq_init_queue);
2759 * Helper for setting up a queue with mq ops, given queue depth, and
2760 * the passed in mq ops flags.
2762 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2763 const struct blk_mq_ops *ops,
2764 unsigned int queue_depth,
2765 unsigned int set_flags)
2767 struct request_queue *q;
2770 memset(set, 0, sizeof(*set));
2772 set->nr_hw_queues = 1;
2774 set->queue_depth = queue_depth;
2775 set->numa_node = NUMA_NO_NODE;
2776 set->flags = set_flags;
2778 ret = blk_mq_alloc_tag_set(set);
2780 return ERR_PTR(ret);
2782 q = blk_mq_init_queue(set);
2784 blk_mq_free_tag_set(set);
2790 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2792 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2793 struct blk_mq_tag_set *set, struct request_queue *q,
2794 int hctx_idx, int node)
2796 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2798 /* reuse dead hctx first */
2799 spin_lock(&q->unused_hctx_lock);
2800 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2801 if (tmp->numa_node == node) {
2807 list_del_init(&hctx->hctx_list);
2808 spin_unlock(&q->unused_hctx_lock);
2811 hctx = blk_mq_alloc_hctx(q, set, node);
2815 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2821 kobject_put(&hctx->kobj);
2826 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2827 struct request_queue *q)
2830 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2832 if (q->nr_hw_queues < set->nr_hw_queues) {
2833 struct blk_mq_hw_ctx **new_hctxs;
2835 new_hctxs = kcalloc_node(set->nr_hw_queues,
2836 sizeof(*new_hctxs), GFP_KERNEL,
2841 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
2843 q->queue_hw_ctx = new_hctxs;
2848 /* protect against switching io scheduler */
2849 mutex_lock(&q->sysfs_lock);
2850 for (i = 0; i < set->nr_hw_queues; i++) {
2852 struct blk_mq_hw_ctx *hctx;
2854 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2856 * If the hw queue has been mapped to another numa node,
2857 * we need to realloc the hctx. If allocation fails, fallback
2858 * to use the previous one.
2860 if (hctxs[i] && (hctxs[i]->numa_node == node))
2863 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2866 blk_mq_exit_hctx(q, set, hctxs[i], i);
2870 pr_warn("Allocate new hctx on node %d fails,\
2871 fallback to previous one on node %d\n",
2872 node, hctxs[i]->numa_node);
2878 * Increasing nr_hw_queues fails. Free the newly allocated
2879 * hctxs and keep the previous q->nr_hw_queues.
2881 if (i != set->nr_hw_queues) {
2882 j = q->nr_hw_queues;
2886 end = q->nr_hw_queues;
2887 q->nr_hw_queues = set->nr_hw_queues;
2890 for (; j < end; j++) {
2891 struct blk_mq_hw_ctx *hctx = hctxs[j];
2895 blk_mq_free_map_and_requests(set, j);
2896 blk_mq_exit_hctx(q, set, hctx, j);
2900 mutex_unlock(&q->sysfs_lock);
2903 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2904 struct request_queue *q,
2907 /* mark the queue as mq asap */
2908 q->mq_ops = set->ops;
2910 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2911 blk_mq_poll_stats_bkt,
2912 BLK_MQ_POLL_STATS_BKTS, q);
2916 if (blk_mq_alloc_ctxs(q))
2919 /* init q->mq_kobj and sw queues' kobjects */
2920 blk_mq_sysfs_init(q);
2922 INIT_LIST_HEAD(&q->unused_hctx_list);
2923 spin_lock_init(&q->unused_hctx_lock);
2925 blk_mq_realloc_hw_ctxs(set, q);
2926 if (!q->nr_hw_queues)
2929 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2930 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2934 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2935 if (set->nr_maps > HCTX_TYPE_POLL &&
2936 set->map[HCTX_TYPE_POLL].nr_queues)
2937 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2939 q->sg_reserved_size = INT_MAX;
2941 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2942 INIT_LIST_HEAD(&q->requeue_list);
2943 spin_lock_init(&q->requeue_lock);
2945 q->make_request_fn = blk_mq_make_request;
2946 q->nr_requests = set->queue_depth;
2949 * Default to classic polling
2951 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2953 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2954 blk_mq_add_queue_tag_set(set, q);
2955 blk_mq_map_swqueue(q);
2958 elevator_init_mq(q);
2963 kfree(q->queue_hw_ctx);
2964 q->nr_hw_queues = 0;
2965 blk_mq_sysfs_deinit(q);
2967 blk_stat_free_callback(q->poll_cb);
2971 return ERR_PTR(-ENOMEM);
2973 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2975 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2976 void blk_mq_exit_queue(struct request_queue *q)
2978 struct blk_mq_tag_set *set = q->tag_set;
2980 blk_mq_del_queue_tag_set(q);
2981 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2984 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2988 for (i = 0; i < set->nr_hw_queues; i++)
2989 if (!__blk_mq_alloc_rq_map(set, i))
2996 blk_mq_free_rq_map(set->tags[i]);
3002 * Allocate the request maps associated with this tag_set. Note that this
3003 * may reduce the depth asked for, if memory is tight. set->queue_depth
3004 * will be updated to reflect the allocated depth.
3006 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3011 depth = set->queue_depth;
3013 err = __blk_mq_alloc_rq_maps(set);
3017 set->queue_depth >>= 1;
3018 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3022 } while (set->queue_depth);
3024 if (!set->queue_depth || err) {
3025 pr_err("blk-mq: failed to allocate request map\n");
3029 if (depth != set->queue_depth)
3030 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3031 depth, set->queue_depth);
3036 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3039 * blk_mq_map_queues() and multiple .map_queues() implementations
3040 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3041 * number of hardware queues.
3043 if (set->nr_maps == 1)
3044 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3046 if (set->ops->map_queues && !is_kdump_kernel()) {
3050 * transport .map_queues is usually done in the following
3053 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3054 * mask = get_cpu_mask(queue)
3055 * for_each_cpu(cpu, mask)
3056 * set->map[x].mq_map[cpu] = queue;
3059 * When we need to remap, the table has to be cleared for
3060 * killing stale mapping since one CPU may not be mapped
3063 for (i = 0; i < set->nr_maps; i++)
3064 blk_mq_clear_mq_map(&set->map[i]);
3066 return set->ops->map_queues(set);
3068 BUG_ON(set->nr_maps > 1);
3069 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3073 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3074 int cur_nr_hw_queues, int new_nr_hw_queues)
3076 struct blk_mq_tags **new_tags;
3078 if (cur_nr_hw_queues >= new_nr_hw_queues)
3081 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3082 GFP_KERNEL, set->numa_node);
3087 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3088 sizeof(*set->tags));
3090 set->tags = new_tags;
3091 set->nr_hw_queues = new_nr_hw_queues;
3097 * Alloc a tag set to be associated with one or more request queues.
3098 * May fail with EINVAL for various error conditions. May adjust the
3099 * requested depth down, if it's too large. In that case, the set
3100 * value will be stored in set->queue_depth.
3102 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3106 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3108 if (!set->nr_hw_queues)
3110 if (!set->queue_depth)
3112 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3115 if (!set->ops->queue_rq)
3118 if (!set->ops->get_budget ^ !set->ops->put_budget)
3121 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3122 pr_info("blk-mq: reduced tag depth to %u\n",
3124 set->queue_depth = BLK_MQ_MAX_DEPTH;
3129 else if (set->nr_maps > HCTX_MAX_TYPES)
3133 * If a crashdump is active, then we are potentially in a very
3134 * memory constrained environment. Limit us to 1 queue and
3135 * 64 tags to prevent using too much memory.
3137 if (is_kdump_kernel()) {
3138 set->nr_hw_queues = 1;
3140 set->queue_depth = min(64U, set->queue_depth);
3143 * There is no use for more h/w queues than cpus if we just have
3146 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3147 set->nr_hw_queues = nr_cpu_ids;
3149 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3153 for (i = 0; i < set->nr_maps; i++) {
3154 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3155 sizeof(set->map[i].mq_map[0]),
3156 GFP_KERNEL, set->numa_node);
3157 if (!set->map[i].mq_map)
3158 goto out_free_mq_map;
3159 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3162 ret = blk_mq_update_queue_map(set);
3164 goto out_free_mq_map;
3166 ret = blk_mq_alloc_rq_maps(set);
3168 goto out_free_mq_map;
3170 mutex_init(&set->tag_list_lock);
3171 INIT_LIST_HEAD(&set->tag_list);
3176 for (i = 0; i < set->nr_maps; i++) {
3177 kfree(set->map[i].mq_map);
3178 set->map[i].mq_map = NULL;
3184 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3186 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3190 for (i = 0; i < set->nr_hw_queues; i++)
3191 blk_mq_free_map_and_requests(set, i);
3193 for (j = 0; j < set->nr_maps; j++) {
3194 kfree(set->map[j].mq_map);
3195 set->map[j].mq_map = NULL;
3201 EXPORT_SYMBOL(blk_mq_free_tag_set);
3203 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3205 struct blk_mq_tag_set *set = q->tag_set;
3206 struct blk_mq_hw_ctx *hctx;
3212 if (q->nr_requests == nr)
3215 blk_mq_freeze_queue(q);
3216 blk_mq_quiesce_queue(q);
3219 queue_for_each_hw_ctx(q, hctx, i) {
3223 * If we're using an MQ scheduler, just update the scheduler
3224 * queue depth. This is similar to what the old code would do.
3226 if (!hctx->sched_tags) {
3227 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3230 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3235 if (q->elevator && q->elevator->type->ops.depth_updated)
3236 q->elevator->type->ops.depth_updated(hctx);
3240 q->nr_requests = nr;
3242 blk_mq_unquiesce_queue(q);
3243 blk_mq_unfreeze_queue(q);
3249 * request_queue and elevator_type pair.
3250 * It is just used by __blk_mq_update_nr_hw_queues to cache
3251 * the elevator_type associated with a request_queue.
3253 struct blk_mq_qe_pair {
3254 struct list_head node;
3255 struct request_queue *q;
3256 struct elevator_type *type;
3260 * Cache the elevator_type in qe pair list and switch the
3261 * io scheduler to 'none'
3263 static bool blk_mq_elv_switch_none(struct list_head *head,
3264 struct request_queue *q)
3266 struct blk_mq_qe_pair *qe;
3271 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3275 INIT_LIST_HEAD(&qe->node);
3277 qe->type = q->elevator->type;
3278 list_add(&qe->node, head);
3280 mutex_lock(&q->sysfs_lock);
3282 * After elevator_switch_mq, the previous elevator_queue will be
3283 * released by elevator_release. The reference of the io scheduler
3284 * module get by elevator_get will also be put. So we need to get
3285 * a reference of the io scheduler module here to prevent it to be
3288 __module_get(qe->type->elevator_owner);
3289 elevator_switch_mq(q, NULL);
3290 mutex_unlock(&q->sysfs_lock);
3295 static void blk_mq_elv_switch_back(struct list_head *head,
3296 struct request_queue *q)
3298 struct blk_mq_qe_pair *qe;
3299 struct elevator_type *t = NULL;
3301 list_for_each_entry(qe, head, node)
3310 list_del(&qe->node);
3313 mutex_lock(&q->sysfs_lock);
3314 elevator_switch_mq(q, t);
3315 mutex_unlock(&q->sysfs_lock);
3318 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3321 struct request_queue *q;
3323 int prev_nr_hw_queues;
3325 lockdep_assert_held(&set->tag_list_lock);
3327 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3328 nr_hw_queues = nr_cpu_ids;
3329 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3332 list_for_each_entry(q, &set->tag_list, tag_set_list)
3333 blk_mq_freeze_queue(q);
3335 * Switch IO scheduler to 'none', cleaning up the data associated
3336 * with the previous scheduler. We will switch back once we are done
3337 * updating the new sw to hw queue mappings.
3339 list_for_each_entry(q, &set->tag_list, tag_set_list)
3340 if (!blk_mq_elv_switch_none(&head, q))
3343 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3344 blk_mq_debugfs_unregister_hctxs(q);
3345 blk_mq_sysfs_unregister(q);
3348 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3352 prev_nr_hw_queues = set->nr_hw_queues;
3353 set->nr_hw_queues = nr_hw_queues;
3354 blk_mq_update_queue_map(set);
3356 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3357 blk_mq_realloc_hw_ctxs(set, q);
3358 if (q->nr_hw_queues != set->nr_hw_queues) {
3359 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3360 nr_hw_queues, prev_nr_hw_queues);
3361 set->nr_hw_queues = prev_nr_hw_queues;
3362 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3365 blk_mq_map_swqueue(q);
3369 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3370 blk_mq_sysfs_register(q);
3371 blk_mq_debugfs_register_hctxs(q);
3375 list_for_each_entry(q, &set->tag_list, tag_set_list)
3376 blk_mq_elv_switch_back(&head, q);
3378 list_for_each_entry(q, &set->tag_list, tag_set_list)
3379 blk_mq_unfreeze_queue(q);
3382 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3384 mutex_lock(&set->tag_list_lock);
3385 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3386 mutex_unlock(&set->tag_list_lock);
3388 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3390 /* Enable polling stats and return whether they were already enabled. */
3391 static bool blk_poll_stats_enable(struct request_queue *q)
3393 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3394 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3396 blk_stat_add_callback(q, q->poll_cb);
3400 static void blk_mq_poll_stats_start(struct request_queue *q)
3403 * We don't arm the callback if polling stats are not enabled or the
3404 * callback is already active.
3406 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3407 blk_stat_is_active(q->poll_cb))
3410 blk_stat_activate_msecs(q->poll_cb, 100);
3413 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3415 struct request_queue *q = cb->data;
3418 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3419 if (cb->stat[bucket].nr_samples)
3420 q->poll_stat[bucket] = cb->stat[bucket];
3424 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3427 unsigned long ret = 0;
3431 * If stats collection isn't on, don't sleep but turn it on for
3434 if (!blk_poll_stats_enable(q))
3438 * As an optimistic guess, use half of the mean service time
3439 * for this type of request. We can (and should) make this smarter.
3440 * For instance, if the completion latencies are tight, we can
3441 * get closer than just half the mean. This is especially
3442 * important on devices where the completion latencies are longer
3443 * than ~10 usec. We do use the stats for the relevant IO size
3444 * if available which does lead to better estimates.
3446 bucket = blk_mq_poll_stats_bkt(rq);
3450 if (q->poll_stat[bucket].nr_samples)
3451 ret = (q->poll_stat[bucket].mean + 1) / 2;
3456 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3459 struct hrtimer_sleeper hs;
3460 enum hrtimer_mode mode;
3464 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3468 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3470 * 0: use half of prev avg
3471 * >0: use this specific value
3473 if (q->poll_nsec > 0)
3474 nsecs = q->poll_nsec;
3476 nsecs = blk_mq_poll_nsecs(q, rq);
3481 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3484 * This will be replaced with the stats tracking code, using
3485 * 'avg_completion_time / 2' as the pre-sleep target.
3489 mode = HRTIMER_MODE_REL;
3490 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3491 hrtimer_set_expires(&hs.timer, kt);
3494 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3496 set_current_state(TASK_UNINTERRUPTIBLE);
3497 hrtimer_sleeper_start_expires(&hs, mode);
3500 hrtimer_cancel(&hs.timer);
3501 mode = HRTIMER_MODE_ABS;
3502 } while (hs.task && !signal_pending(current));
3504 __set_current_state(TASK_RUNNING);
3505 destroy_hrtimer_on_stack(&hs.timer);
3509 static bool blk_mq_poll_hybrid(struct request_queue *q,
3510 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3514 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3517 if (!blk_qc_t_is_internal(cookie))
3518 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3520 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3522 * With scheduling, if the request has completed, we'll
3523 * get a NULL return here, as we clear the sched tag when
3524 * that happens. The request still remains valid, like always,
3525 * so we should be safe with just the NULL check.
3531 return blk_mq_poll_hybrid_sleep(q, rq);
3535 * blk_poll - poll for IO completions
3537 * @cookie: cookie passed back at IO submission time
3538 * @spin: whether to spin for completions
3541 * Poll for completions on the passed in queue. Returns number of
3542 * completed entries found. If @spin is true, then blk_poll will continue
3543 * looping until at least one completion is found, unless the task is
3544 * otherwise marked running (or we need to reschedule).
3546 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3548 struct blk_mq_hw_ctx *hctx;
3551 if (!blk_qc_t_valid(cookie) ||
3552 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3556 blk_flush_plug_list(current->plug, false);
3558 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3561 * If we sleep, have the caller restart the poll loop to reset
3562 * the state. Like for the other success return cases, the
3563 * caller is responsible for checking if the IO completed. If
3564 * the IO isn't complete, we'll get called again and will go
3565 * straight to the busy poll loop.
3567 if (blk_mq_poll_hybrid(q, hctx, cookie))
3570 hctx->poll_considered++;
3572 state = current->state;
3576 hctx->poll_invoked++;
3578 ret = q->mq_ops->poll(hctx);
3580 hctx->poll_success++;
3581 __set_current_state(TASK_RUNNING);
3585 if (signal_pending_state(state, current))
3586 __set_current_state(TASK_RUNNING);
3588 if (current->state == TASK_RUNNING)
3590 if (ret < 0 || !spin)
3593 } while (!need_resched());
3595 __set_current_state(TASK_RUNNING);
3598 EXPORT_SYMBOL_GPL(blk_poll);
3600 unsigned int blk_mq_rq_cpu(struct request *rq)
3602 return rq->mq_ctx->cpu;
3604 EXPORT_SYMBOL(blk_mq_rq_cpu);
3606 static int __init blk_mq_init(void)
3608 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3609 blk_mq_hctx_notify_dead);
3612 subsys_initcall(blk_mq_init);