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/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
34 static DEFINE_MUTEX(all_q_mutex);
35 static LIST_HEAD(all_q_list);
38 * Check if any of the ctx's have pending work in this hardware queue
40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
42 return sbitmap_any_bit_set(&hctx->ctx_map);
46 * Mark this ctx as having pending work in this hardware queue
48 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
49 struct blk_mq_ctx *ctx)
51 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
52 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
55 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
56 struct blk_mq_ctx *ctx)
58 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
61 void blk_mq_freeze_queue_start(struct request_queue *q)
65 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
66 if (freeze_depth == 1) {
67 percpu_ref_kill(&q->q_usage_counter);
68 blk_mq_run_hw_queues(q, false);
71 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
73 static void blk_mq_freeze_queue_wait(struct request_queue *q)
75 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
79 * Guarantee no request is in use, so we can change any data structure of
80 * the queue afterward.
82 void blk_freeze_queue(struct request_queue *q)
85 * In the !blk_mq case we are only calling this to kill the
86 * q_usage_counter, otherwise this increases the freeze depth
87 * and waits for it to return to zero. For this reason there is
88 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
89 * exported to drivers as the only user for unfreeze is blk_mq.
91 blk_mq_freeze_queue_start(q);
92 blk_mq_freeze_queue_wait(q);
95 void blk_mq_freeze_queue(struct request_queue *q)
98 * ...just an alias to keep freeze and unfreeze actions balanced
99 * in the blk_mq_* namespace
103 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
105 void blk_mq_unfreeze_queue(struct request_queue *q)
109 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
110 WARN_ON_ONCE(freeze_depth < 0);
112 percpu_ref_reinit(&q->q_usage_counter);
113 wake_up_all(&q->mq_freeze_wq);
116 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
118 void blk_mq_wake_waiters(struct request_queue *q)
120 struct blk_mq_hw_ctx *hctx;
123 queue_for_each_hw_ctx(q, hctx, i)
124 if (blk_mq_hw_queue_mapped(hctx))
125 blk_mq_tag_wakeup_all(hctx->tags, true);
128 * If we are called because the queue has now been marked as
129 * dying, we need to ensure that processes currently waiting on
130 * the queue are notified as well.
132 wake_up_all(&q->mq_freeze_wq);
135 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
137 return blk_mq_has_free_tags(hctx->tags);
139 EXPORT_SYMBOL(blk_mq_can_queue);
141 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
142 struct request *rq, int op,
143 unsigned int op_flags)
145 if (blk_queue_io_stat(q))
146 op_flags |= REQ_IO_STAT;
148 INIT_LIST_HEAD(&rq->queuelist);
149 /* csd/requeue_work/fifo_time is initialized before use */
152 req_set_op_attrs(rq, op, op_flags);
153 /* do not touch atomic flags, it needs atomic ops against the timer */
155 INIT_HLIST_NODE(&rq->hash);
156 RB_CLEAR_NODE(&rq->rb_node);
159 rq->start_time = jiffies;
160 #ifdef CONFIG_BLK_CGROUP
162 set_start_time_ns(rq);
163 rq->io_start_time_ns = 0;
165 rq->nr_phys_segments = 0;
166 #if defined(CONFIG_BLK_DEV_INTEGRITY)
167 rq->nr_integrity_segments = 0;
170 /* tag was already set */
180 INIT_LIST_HEAD(&rq->timeout_list);
184 rq->end_io_data = NULL;
187 ctx->rq_dispatched[rw_is_sync(op, op_flags)]++;
190 static struct request *
191 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int op, int op_flags)
196 tag = blk_mq_get_tag(data);
197 if (tag != BLK_MQ_TAG_FAIL) {
198 rq = data->hctx->tags->rqs[tag];
200 if (blk_mq_tag_busy(data->hctx)) {
201 rq->cmd_flags = REQ_MQ_INFLIGHT;
202 atomic_inc(&data->hctx->nr_active);
206 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op, op_flags);
213 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
216 struct blk_mq_ctx *ctx;
217 struct blk_mq_hw_ctx *hctx;
219 struct blk_mq_alloc_data alloc_data;
222 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
226 ctx = blk_mq_get_ctx(q);
227 hctx = q->mq_ops->map_queue(q, ctx->cpu);
228 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
229 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
234 return ERR_PTR(-EWOULDBLOCK);
238 rq->__sector = (sector_t) -1;
239 rq->bio = rq->biotail = NULL;
242 EXPORT_SYMBOL(blk_mq_alloc_request);
244 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
245 unsigned int flags, unsigned int hctx_idx)
247 struct blk_mq_hw_ctx *hctx;
248 struct blk_mq_ctx *ctx;
250 struct blk_mq_alloc_data alloc_data;
254 * If the tag allocator sleeps we could get an allocation for a
255 * different hardware context. No need to complicate the low level
256 * allocator for this for the rare use case of a command tied to
259 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
260 return ERR_PTR(-EINVAL);
262 if (hctx_idx >= q->nr_hw_queues)
263 return ERR_PTR(-EIO);
265 ret = blk_queue_enter(q, true);
269 hctx = q->queue_hw_ctx[hctx_idx];
270 ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
272 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
273 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
276 return ERR_PTR(-EWOULDBLOCK);
281 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
283 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
284 struct blk_mq_ctx *ctx, struct request *rq)
286 const int tag = rq->tag;
287 struct request_queue *q = rq->q;
289 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
290 atomic_dec(&hctx->nr_active);
293 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
294 blk_mq_put_tag(hctx, ctx, tag);
298 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
300 struct blk_mq_ctx *ctx = rq->mq_ctx;
302 ctx->rq_completed[rq_is_sync(rq)]++;
303 __blk_mq_free_request(hctx, ctx, rq);
306 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
308 void blk_mq_free_request(struct request *rq)
310 struct blk_mq_hw_ctx *hctx;
311 struct request_queue *q = rq->q;
313 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
314 blk_mq_free_hctx_request(hctx, rq);
316 EXPORT_SYMBOL_GPL(blk_mq_free_request);
318 inline void __blk_mq_end_request(struct request *rq, int error)
320 blk_account_io_done(rq);
323 rq->end_io(rq, error);
325 if (unlikely(blk_bidi_rq(rq)))
326 blk_mq_free_request(rq->next_rq);
327 blk_mq_free_request(rq);
330 EXPORT_SYMBOL(__blk_mq_end_request);
332 void blk_mq_end_request(struct request *rq, int error)
334 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
336 __blk_mq_end_request(rq, error);
338 EXPORT_SYMBOL(blk_mq_end_request);
340 static void __blk_mq_complete_request_remote(void *data)
342 struct request *rq = data;
344 rq->q->softirq_done_fn(rq);
347 static void blk_mq_ipi_complete_request(struct request *rq)
349 struct blk_mq_ctx *ctx = rq->mq_ctx;
353 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
354 rq->q->softirq_done_fn(rq);
359 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
360 shared = cpus_share_cache(cpu, ctx->cpu);
362 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
363 rq->csd.func = __blk_mq_complete_request_remote;
366 smp_call_function_single_async(ctx->cpu, &rq->csd);
368 rq->q->softirq_done_fn(rq);
373 static void __blk_mq_complete_request(struct request *rq)
375 struct request_queue *q = rq->q;
377 if (!q->softirq_done_fn)
378 blk_mq_end_request(rq, rq->errors);
380 blk_mq_ipi_complete_request(rq);
384 * blk_mq_complete_request - end I/O on a request
385 * @rq: the request being processed
388 * Ends all I/O on a request. It does not handle partial completions.
389 * The actual completion happens out-of-order, through a IPI handler.
391 void blk_mq_complete_request(struct request *rq, int error)
393 struct request_queue *q = rq->q;
395 if (unlikely(blk_should_fake_timeout(q)))
397 if (!blk_mark_rq_complete(rq)) {
399 __blk_mq_complete_request(rq);
402 EXPORT_SYMBOL(blk_mq_complete_request);
404 int blk_mq_request_started(struct request *rq)
406 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
408 EXPORT_SYMBOL_GPL(blk_mq_request_started);
410 void blk_mq_start_request(struct request *rq)
412 struct request_queue *q = rq->q;
414 trace_block_rq_issue(q, rq);
416 rq->resid_len = blk_rq_bytes(rq);
417 if (unlikely(blk_bidi_rq(rq)))
418 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
423 * Ensure that ->deadline is visible before set the started
424 * flag and clear the completed flag.
426 smp_mb__before_atomic();
429 * Mark us as started and clear complete. Complete might have been
430 * set if requeue raced with timeout, which then marked it as
431 * complete. So be sure to clear complete again when we start
432 * the request, otherwise we'll ignore the completion event.
434 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
435 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
436 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
437 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
439 if (q->dma_drain_size && blk_rq_bytes(rq)) {
441 * Make sure space for the drain appears. We know we can do
442 * this because max_hw_segments has been adjusted to be one
443 * fewer than the device can handle.
445 rq->nr_phys_segments++;
448 EXPORT_SYMBOL(blk_mq_start_request);
450 static void __blk_mq_requeue_request(struct request *rq)
452 struct request_queue *q = rq->q;
454 trace_block_rq_requeue(q, rq);
456 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
457 if (q->dma_drain_size && blk_rq_bytes(rq))
458 rq->nr_phys_segments--;
462 void blk_mq_requeue_request(struct request *rq)
464 __blk_mq_requeue_request(rq);
466 BUG_ON(blk_queued_rq(rq));
467 blk_mq_add_to_requeue_list(rq, true);
469 EXPORT_SYMBOL(blk_mq_requeue_request);
471 static void blk_mq_requeue_work(struct work_struct *work)
473 struct request_queue *q =
474 container_of(work, struct request_queue, requeue_work.work);
476 struct request *rq, *next;
479 spin_lock_irqsave(&q->requeue_lock, flags);
480 list_splice_init(&q->requeue_list, &rq_list);
481 spin_unlock_irqrestore(&q->requeue_lock, flags);
483 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
484 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
487 rq->cmd_flags &= ~REQ_SOFTBARRIER;
488 list_del_init(&rq->queuelist);
489 blk_mq_insert_request(rq, true, false, false);
492 while (!list_empty(&rq_list)) {
493 rq = list_entry(rq_list.next, struct request, queuelist);
494 list_del_init(&rq->queuelist);
495 blk_mq_insert_request(rq, false, false, false);
499 * Use the start variant of queue running here, so that running
500 * the requeue work will kick stopped queues.
502 blk_mq_start_hw_queues(q);
505 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
507 struct request_queue *q = rq->q;
511 * We abuse this flag that is otherwise used by the I/O scheduler to
512 * request head insertation from the workqueue.
514 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
516 spin_lock_irqsave(&q->requeue_lock, flags);
518 rq->cmd_flags |= REQ_SOFTBARRIER;
519 list_add(&rq->queuelist, &q->requeue_list);
521 list_add_tail(&rq->queuelist, &q->requeue_list);
523 spin_unlock_irqrestore(&q->requeue_lock, flags);
525 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
527 void blk_mq_cancel_requeue_work(struct request_queue *q)
529 cancel_delayed_work_sync(&q->requeue_work);
531 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
533 void blk_mq_kick_requeue_list(struct request_queue *q)
535 kblockd_schedule_delayed_work(&q->requeue_work, 0);
537 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
539 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
542 kblockd_schedule_delayed_work(&q->requeue_work,
543 msecs_to_jiffies(msecs));
545 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
547 void blk_mq_abort_requeue_list(struct request_queue *q)
552 spin_lock_irqsave(&q->requeue_lock, flags);
553 list_splice_init(&q->requeue_list, &rq_list);
554 spin_unlock_irqrestore(&q->requeue_lock, flags);
556 while (!list_empty(&rq_list)) {
559 rq = list_first_entry(&rq_list, struct request, queuelist);
560 list_del_init(&rq->queuelist);
562 blk_mq_end_request(rq, rq->errors);
565 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
567 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
569 if (tag < tags->nr_tags) {
570 prefetch(tags->rqs[tag]);
571 return tags->rqs[tag];
576 EXPORT_SYMBOL(blk_mq_tag_to_rq);
578 struct blk_mq_timeout_data {
580 unsigned int next_set;
583 void blk_mq_rq_timed_out(struct request *req, bool reserved)
585 struct blk_mq_ops *ops = req->q->mq_ops;
586 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
589 * We know that complete is set at this point. If STARTED isn't set
590 * anymore, then the request isn't active and the "timeout" should
591 * just be ignored. This can happen due to the bitflag ordering.
592 * Timeout first checks if STARTED is set, and if it is, assumes
593 * the request is active. But if we race with completion, then
594 * we both flags will get cleared. So check here again, and ignore
595 * a timeout event with a request that isn't active.
597 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
601 ret = ops->timeout(req, reserved);
605 __blk_mq_complete_request(req);
607 case BLK_EH_RESET_TIMER:
609 blk_clear_rq_complete(req);
611 case BLK_EH_NOT_HANDLED:
614 printk(KERN_ERR "block: bad eh return: %d\n", ret);
619 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
620 struct request *rq, void *priv, bool reserved)
622 struct blk_mq_timeout_data *data = priv;
624 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
626 * If a request wasn't started before the queue was
627 * marked dying, kill it here or it'll go unnoticed.
629 if (unlikely(blk_queue_dying(rq->q))) {
631 blk_mq_end_request(rq, rq->errors);
636 if (time_after_eq(jiffies, rq->deadline)) {
637 if (!blk_mark_rq_complete(rq))
638 blk_mq_rq_timed_out(rq, reserved);
639 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
640 data->next = rq->deadline;
645 static void blk_mq_timeout_work(struct work_struct *work)
647 struct request_queue *q =
648 container_of(work, struct request_queue, timeout_work);
649 struct blk_mq_timeout_data data = {
655 /* A deadlock might occur if a request is stuck requiring a
656 * timeout at the same time a queue freeze is waiting
657 * completion, since the timeout code would not be able to
658 * acquire the queue reference here.
660 * That's why we don't use blk_queue_enter here; instead, we use
661 * percpu_ref_tryget directly, because we need to be able to
662 * obtain a reference even in the short window between the queue
663 * starting to freeze, by dropping the first reference in
664 * blk_mq_freeze_queue_start, and the moment the last request is
665 * consumed, marked by the instant q_usage_counter reaches
668 if (!percpu_ref_tryget(&q->q_usage_counter))
671 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
674 data.next = blk_rq_timeout(round_jiffies_up(data.next));
675 mod_timer(&q->timeout, data.next);
677 struct blk_mq_hw_ctx *hctx;
679 queue_for_each_hw_ctx(q, hctx, i) {
680 /* the hctx may be unmapped, so check it here */
681 if (blk_mq_hw_queue_mapped(hctx))
682 blk_mq_tag_idle(hctx);
689 * Reverse check our software queue for entries that we could potentially
690 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
691 * too much time checking for merges.
693 static bool blk_mq_attempt_merge(struct request_queue *q,
694 struct blk_mq_ctx *ctx, struct bio *bio)
699 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
705 if (!blk_rq_merge_ok(rq, bio))
708 el_ret = blk_try_merge(rq, bio);
709 if (el_ret == ELEVATOR_BACK_MERGE) {
710 if (bio_attempt_back_merge(q, rq, bio)) {
715 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
716 if (bio_attempt_front_merge(q, rq, bio)) {
727 struct flush_busy_ctx_data {
728 struct blk_mq_hw_ctx *hctx;
729 struct list_head *list;
732 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
734 struct flush_busy_ctx_data *flush_data = data;
735 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
736 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
738 sbitmap_clear_bit(sb, bitnr);
739 spin_lock(&ctx->lock);
740 list_splice_tail_init(&ctx->rq_list, flush_data->list);
741 spin_unlock(&ctx->lock);
746 * Process software queues that have been marked busy, splicing them
747 * to the for-dispatch
749 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
751 struct flush_busy_ctx_data data = {
756 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
759 static inline unsigned int queued_to_index(unsigned int queued)
764 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
768 * Run this hardware queue, pulling any software queues mapped to it in.
769 * Note that this function currently has various problems around ordering
770 * of IO. In particular, we'd like FIFO behaviour on handling existing
771 * items on the hctx->dispatch list. Ignore that for now.
773 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
775 struct request_queue *q = hctx->queue;
778 LIST_HEAD(driver_list);
779 struct list_head *dptr;
782 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
785 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
786 cpu_online(hctx->next_cpu));
791 * Touch any software queue that has pending entries.
793 flush_busy_ctxs(hctx, &rq_list);
796 * If we have previous entries on our dispatch list, grab them
797 * and stuff them at the front for more fair dispatch.
799 if (!list_empty_careful(&hctx->dispatch)) {
800 spin_lock(&hctx->lock);
801 if (!list_empty(&hctx->dispatch))
802 list_splice_init(&hctx->dispatch, &rq_list);
803 spin_unlock(&hctx->lock);
807 * Start off with dptr being NULL, so we start the first request
808 * immediately, even if we have more pending.
813 * Now process all the entries, sending them to the driver.
816 while (!list_empty(&rq_list)) {
817 struct blk_mq_queue_data bd;
820 rq = list_first_entry(&rq_list, struct request, queuelist);
821 list_del_init(&rq->queuelist);
825 bd.last = list_empty(&rq_list);
827 ret = q->mq_ops->queue_rq(hctx, &bd);
829 case BLK_MQ_RQ_QUEUE_OK:
832 case BLK_MQ_RQ_QUEUE_BUSY:
833 list_add(&rq->queuelist, &rq_list);
834 __blk_mq_requeue_request(rq);
837 pr_err("blk-mq: bad return on queue: %d\n", ret);
838 case BLK_MQ_RQ_QUEUE_ERROR:
840 blk_mq_end_request(rq, rq->errors);
844 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
848 * We've done the first request. If we have more than 1
849 * left in the list, set dptr to defer issue.
851 if (!dptr && rq_list.next != rq_list.prev)
855 hctx->dispatched[queued_to_index(queued)]++;
858 * Any items that need requeuing? Stuff them into hctx->dispatch,
859 * that is where we will continue on next queue run.
861 if (!list_empty(&rq_list)) {
862 spin_lock(&hctx->lock);
863 list_splice(&rq_list, &hctx->dispatch);
864 spin_unlock(&hctx->lock);
866 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
867 * it's possible the queue is stopped and restarted again
868 * before this. Queue restart will dispatch requests. And since
869 * requests in rq_list aren't added into hctx->dispatch yet,
870 * the requests in rq_list might get lost.
872 * blk_mq_run_hw_queue() already checks the STOPPED bit
874 blk_mq_run_hw_queue(hctx, true);
879 * It'd be great if the workqueue API had a way to pass
880 * in a mask and had some smarts for more clever placement.
881 * For now we just round-robin here, switching for every
882 * BLK_MQ_CPU_WORK_BATCH queued items.
884 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
886 if (hctx->queue->nr_hw_queues == 1)
887 return WORK_CPU_UNBOUND;
889 if (--hctx->next_cpu_batch <= 0) {
890 int cpu = hctx->next_cpu, next_cpu;
892 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
893 if (next_cpu >= nr_cpu_ids)
894 next_cpu = cpumask_first(hctx->cpumask);
896 hctx->next_cpu = next_cpu;
897 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
902 return hctx->next_cpu;
905 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
907 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
908 !blk_mq_hw_queue_mapped(hctx)))
911 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
913 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
914 __blk_mq_run_hw_queue(hctx);
922 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
925 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
927 struct blk_mq_hw_ctx *hctx;
930 queue_for_each_hw_ctx(q, hctx, i) {
931 if ((!blk_mq_hctx_has_pending(hctx) &&
932 list_empty_careful(&hctx->dispatch)) ||
933 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
936 blk_mq_run_hw_queue(hctx, async);
939 EXPORT_SYMBOL(blk_mq_run_hw_queues);
941 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
943 cancel_work(&hctx->run_work);
944 cancel_delayed_work(&hctx->delay_work);
945 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
947 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
949 void blk_mq_stop_hw_queues(struct request_queue *q)
951 struct blk_mq_hw_ctx *hctx;
954 queue_for_each_hw_ctx(q, hctx, i)
955 blk_mq_stop_hw_queue(hctx);
957 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
959 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
961 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
963 blk_mq_run_hw_queue(hctx, false);
965 EXPORT_SYMBOL(blk_mq_start_hw_queue);
967 void blk_mq_start_hw_queues(struct request_queue *q)
969 struct blk_mq_hw_ctx *hctx;
972 queue_for_each_hw_ctx(q, hctx, i)
973 blk_mq_start_hw_queue(hctx);
975 EXPORT_SYMBOL(blk_mq_start_hw_queues);
977 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
979 struct blk_mq_hw_ctx *hctx;
982 queue_for_each_hw_ctx(q, hctx, i) {
983 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
986 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
987 blk_mq_run_hw_queue(hctx, async);
990 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
992 static void blk_mq_run_work_fn(struct work_struct *work)
994 struct blk_mq_hw_ctx *hctx;
996 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
998 __blk_mq_run_hw_queue(hctx);
1001 static void blk_mq_delay_work_fn(struct work_struct *work)
1003 struct blk_mq_hw_ctx *hctx;
1005 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1007 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1008 __blk_mq_run_hw_queue(hctx);
1011 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1013 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1016 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1017 &hctx->delay_work, msecs_to_jiffies(msecs));
1019 EXPORT_SYMBOL(blk_mq_delay_queue);
1021 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1025 struct blk_mq_ctx *ctx = rq->mq_ctx;
1027 trace_block_rq_insert(hctx->queue, rq);
1030 list_add(&rq->queuelist, &ctx->rq_list);
1032 list_add_tail(&rq->queuelist, &ctx->rq_list);
1035 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1036 struct request *rq, bool at_head)
1038 struct blk_mq_ctx *ctx = rq->mq_ctx;
1040 __blk_mq_insert_req_list(hctx, rq, at_head);
1041 blk_mq_hctx_mark_pending(hctx, ctx);
1044 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1047 struct blk_mq_ctx *ctx = rq->mq_ctx;
1048 struct request_queue *q = rq->q;
1049 struct blk_mq_hw_ctx *hctx;
1051 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1053 spin_lock(&ctx->lock);
1054 __blk_mq_insert_request(hctx, rq, at_head);
1055 spin_unlock(&ctx->lock);
1058 blk_mq_run_hw_queue(hctx, async);
1061 static void blk_mq_insert_requests(struct request_queue *q,
1062 struct blk_mq_ctx *ctx,
1063 struct list_head *list,
1068 struct blk_mq_hw_ctx *hctx;
1070 trace_block_unplug(q, depth, !from_schedule);
1072 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1075 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1078 spin_lock(&ctx->lock);
1079 while (!list_empty(list)) {
1082 rq = list_first_entry(list, struct request, queuelist);
1083 BUG_ON(rq->mq_ctx != ctx);
1084 list_del_init(&rq->queuelist);
1085 __blk_mq_insert_req_list(hctx, rq, false);
1087 blk_mq_hctx_mark_pending(hctx, ctx);
1088 spin_unlock(&ctx->lock);
1090 blk_mq_run_hw_queue(hctx, from_schedule);
1093 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1095 struct request *rqa = container_of(a, struct request, queuelist);
1096 struct request *rqb = container_of(b, struct request, queuelist);
1098 return !(rqa->mq_ctx < rqb->mq_ctx ||
1099 (rqa->mq_ctx == rqb->mq_ctx &&
1100 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1103 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1105 struct blk_mq_ctx *this_ctx;
1106 struct request_queue *this_q;
1109 LIST_HEAD(ctx_list);
1112 list_splice_init(&plug->mq_list, &list);
1114 list_sort(NULL, &list, plug_ctx_cmp);
1120 while (!list_empty(&list)) {
1121 rq = list_entry_rq(list.next);
1122 list_del_init(&rq->queuelist);
1124 if (rq->mq_ctx != this_ctx) {
1126 blk_mq_insert_requests(this_q, this_ctx,
1131 this_ctx = rq->mq_ctx;
1137 list_add_tail(&rq->queuelist, &ctx_list);
1141 * If 'this_ctx' is set, we know we have entries to complete
1142 * on 'ctx_list'. Do those.
1145 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1150 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1152 init_request_from_bio(rq, bio);
1154 blk_account_io_start(rq, 1);
1157 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1159 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1160 !blk_queue_nomerges(hctx->queue);
1163 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1164 struct blk_mq_ctx *ctx,
1165 struct request *rq, struct bio *bio)
1167 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1168 blk_mq_bio_to_request(rq, bio);
1169 spin_lock(&ctx->lock);
1171 __blk_mq_insert_request(hctx, rq, false);
1172 spin_unlock(&ctx->lock);
1175 struct request_queue *q = hctx->queue;
1177 spin_lock(&ctx->lock);
1178 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1179 blk_mq_bio_to_request(rq, bio);
1183 spin_unlock(&ctx->lock);
1184 __blk_mq_free_request(hctx, ctx, rq);
1189 struct blk_map_ctx {
1190 struct blk_mq_hw_ctx *hctx;
1191 struct blk_mq_ctx *ctx;
1194 static struct request *blk_mq_map_request(struct request_queue *q,
1196 struct blk_map_ctx *data)
1198 struct blk_mq_hw_ctx *hctx;
1199 struct blk_mq_ctx *ctx;
1201 int op = bio_data_dir(bio);
1203 struct blk_mq_alloc_data alloc_data;
1205 blk_queue_enter_live(q);
1206 ctx = blk_mq_get_ctx(q);
1207 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1209 if (rw_is_sync(bio_op(bio), bio->bi_opf))
1210 op_flags |= REQ_SYNC;
1212 trace_block_getrq(q, bio, op);
1213 blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1214 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1222 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1225 struct request_queue *q = rq->q;
1226 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1228 struct blk_mq_queue_data bd = {
1233 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1236 * For OK queue, we are done. For error, kill it. Any other
1237 * error (busy), just add it to our list as we previously
1240 ret = q->mq_ops->queue_rq(hctx, &bd);
1241 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1242 *cookie = new_cookie;
1246 __blk_mq_requeue_request(rq);
1248 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1249 *cookie = BLK_QC_T_NONE;
1251 blk_mq_end_request(rq, rq->errors);
1259 * Multiple hardware queue variant. This will not use per-process plugs,
1260 * but will attempt to bypass the hctx queueing if we can go straight to
1261 * hardware for SYNC IO.
1263 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1265 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1266 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1267 struct blk_map_ctx data;
1269 unsigned int request_count = 0;
1270 struct blk_plug *plug;
1271 struct request *same_queue_rq = NULL;
1274 blk_queue_bounce(q, &bio);
1276 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1278 return BLK_QC_T_NONE;
1281 blk_queue_split(q, &bio, q->bio_split);
1283 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1284 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1285 return BLK_QC_T_NONE;
1287 rq = blk_mq_map_request(q, bio, &data);
1289 return BLK_QC_T_NONE;
1291 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1293 if (unlikely(is_flush_fua)) {
1294 blk_mq_bio_to_request(rq, bio);
1295 blk_insert_flush(rq);
1299 plug = current->plug;
1301 * If the driver supports defer issued based on 'last', then
1302 * queue it up like normal since we can potentially save some
1305 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1306 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1307 struct request *old_rq = NULL;
1309 blk_mq_bio_to_request(rq, bio);
1312 * We do limited pluging. If the bio can be merged, do that.
1313 * Otherwise the existing request in the plug list will be
1314 * issued. So the plug list will have one request at most
1318 * The plug list might get flushed before this. If that
1319 * happens, same_queue_rq is invalid and plug list is
1322 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1323 old_rq = same_queue_rq;
1324 list_del_init(&old_rq->queuelist);
1326 list_add_tail(&rq->queuelist, &plug->mq_list);
1327 } else /* is_sync */
1329 blk_mq_put_ctx(data.ctx);
1332 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1334 blk_mq_insert_request(old_rq, false, true, true);
1338 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1340 * For a SYNC request, send it to the hardware immediately. For
1341 * an ASYNC request, just ensure that we run it later on. The
1342 * latter allows for merging opportunities and more efficient
1346 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1348 blk_mq_put_ctx(data.ctx);
1354 * Single hardware queue variant. This will attempt to use any per-process
1355 * plug for merging and IO deferral.
1357 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1359 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1360 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1361 struct blk_plug *plug;
1362 unsigned int request_count = 0;
1363 struct blk_map_ctx data;
1367 blk_queue_bounce(q, &bio);
1369 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1371 return BLK_QC_T_NONE;
1374 blk_queue_split(q, &bio, q->bio_split);
1376 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1377 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1378 return BLK_QC_T_NONE;
1380 request_count = blk_plug_queued_count(q);
1382 rq = blk_mq_map_request(q, bio, &data);
1384 return BLK_QC_T_NONE;
1386 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1388 if (unlikely(is_flush_fua)) {
1389 blk_mq_bio_to_request(rq, bio);
1390 blk_insert_flush(rq);
1395 * A task plug currently exists. Since this is completely lockless,
1396 * utilize that to temporarily store requests until the task is
1397 * either done or scheduled away.
1399 plug = current->plug;
1401 blk_mq_bio_to_request(rq, bio);
1403 trace_block_plug(q);
1405 blk_mq_put_ctx(data.ctx);
1407 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1408 blk_flush_plug_list(plug, false);
1409 trace_block_plug(q);
1412 list_add_tail(&rq->queuelist, &plug->mq_list);
1416 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1418 * For a SYNC request, send it to the hardware immediately. For
1419 * an ASYNC request, just ensure that we run it later on. The
1420 * latter allows for merging opportunities and more efficient
1424 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1427 blk_mq_put_ctx(data.ctx);
1432 * Default mapping to a software queue, since we use one per CPU.
1434 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1436 return q->queue_hw_ctx[q->mq_map[cpu]];
1438 EXPORT_SYMBOL(blk_mq_map_queue);
1440 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1441 struct blk_mq_tags *tags, unsigned int hctx_idx)
1445 if (tags->rqs && set->ops->exit_request) {
1448 for (i = 0; i < tags->nr_tags; i++) {
1451 set->ops->exit_request(set->driver_data, tags->rqs[i],
1453 tags->rqs[i] = NULL;
1457 while (!list_empty(&tags->page_list)) {
1458 page = list_first_entry(&tags->page_list, struct page, lru);
1459 list_del_init(&page->lru);
1461 * Remove kmemleak object previously allocated in
1462 * blk_mq_init_rq_map().
1464 kmemleak_free(page_address(page));
1465 __free_pages(page, page->private);
1470 blk_mq_free_tags(tags);
1473 static size_t order_to_size(unsigned int order)
1475 return (size_t)PAGE_SIZE << order;
1478 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1479 unsigned int hctx_idx)
1481 struct blk_mq_tags *tags;
1482 unsigned int i, j, entries_per_page, max_order = 4;
1483 size_t rq_size, left;
1485 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1487 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1491 INIT_LIST_HEAD(&tags->page_list);
1493 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1494 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1497 blk_mq_free_tags(tags);
1502 * rq_size is the size of the request plus driver payload, rounded
1503 * to the cacheline size
1505 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1507 left = rq_size * set->queue_depth;
1509 for (i = 0; i < set->queue_depth; ) {
1510 int this_order = max_order;
1515 while (this_order && left < order_to_size(this_order - 1))
1519 page = alloc_pages_node(set->numa_node,
1520 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1526 if (order_to_size(this_order) < rq_size)
1533 page->private = this_order;
1534 list_add_tail(&page->lru, &tags->page_list);
1536 p = page_address(page);
1538 * Allow kmemleak to scan these pages as they contain pointers
1539 * to additional allocations like via ops->init_request().
1541 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1542 entries_per_page = order_to_size(this_order) / rq_size;
1543 to_do = min(entries_per_page, set->queue_depth - i);
1544 left -= to_do * rq_size;
1545 for (j = 0; j < to_do; j++) {
1547 if (set->ops->init_request) {
1548 if (set->ops->init_request(set->driver_data,
1549 tags->rqs[i], hctx_idx, i,
1551 tags->rqs[i] = NULL;
1563 blk_mq_free_rq_map(set, tags, hctx_idx);
1568 * 'cpu' is going away. splice any existing rq_list entries from this
1569 * software queue to the hw queue dispatch list, and ensure that it
1572 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1574 struct blk_mq_ctx *ctx;
1577 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1579 spin_lock(&ctx->lock);
1580 if (!list_empty(&ctx->rq_list)) {
1581 list_splice_init(&ctx->rq_list, &tmp);
1582 blk_mq_hctx_clear_pending(hctx, ctx);
1584 spin_unlock(&ctx->lock);
1586 if (list_empty(&tmp))
1589 spin_lock(&hctx->lock);
1590 list_splice_tail_init(&tmp, &hctx->dispatch);
1591 spin_unlock(&hctx->lock);
1593 blk_mq_run_hw_queue(hctx, true);
1597 static int blk_mq_hctx_notify(void *data, unsigned long action,
1600 struct blk_mq_hw_ctx *hctx = data;
1602 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1603 return blk_mq_hctx_cpu_offline(hctx, cpu);
1606 * In case of CPU online, tags may be reallocated
1607 * in blk_mq_map_swqueue() after mapping is updated.
1613 /* hctx->ctxs will be freed in queue's release handler */
1614 static void blk_mq_exit_hctx(struct request_queue *q,
1615 struct blk_mq_tag_set *set,
1616 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1618 unsigned flush_start_tag = set->queue_depth;
1620 blk_mq_tag_idle(hctx);
1622 if (set->ops->exit_request)
1623 set->ops->exit_request(set->driver_data,
1624 hctx->fq->flush_rq, hctx_idx,
1625 flush_start_tag + hctx_idx);
1627 if (set->ops->exit_hctx)
1628 set->ops->exit_hctx(hctx, hctx_idx);
1630 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1631 blk_free_flush_queue(hctx->fq);
1632 sbitmap_free(&hctx->ctx_map);
1635 static void blk_mq_exit_hw_queues(struct request_queue *q,
1636 struct blk_mq_tag_set *set, int nr_queue)
1638 struct blk_mq_hw_ctx *hctx;
1641 queue_for_each_hw_ctx(q, hctx, i) {
1644 blk_mq_exit_hctx(q, set, hctx, i);
1648 static void blk_mq_free_hw_queues(struct request_queue *q,
1649 struct blk_mq_tag_set *set)
1651 struct blk_mq_hw_ctx *hctx;
1654 queue_for_each_hw_ctx(q, hctx, i)
1655 free_cpumask_var(hctx->cpumask);
1658 static int blk_mq_init_hctx(struct request_queue *q,
1659 struct blk_mq_tag_set *set,
1660 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1663 unsigned flush_start_tag = set->queue_depth;
1665 node = hctx->numa_node;
1666 if (node == NUMA_NO_NODE)
1667 node = hctx->numa_node = set->numa_node;
1669 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1670 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1671 spin_lock_init(&hctx->lock);
1672 INIT_LIST_HEAD(&hctx->dispatch);
1674 hctx->queue_num = hctx_idx;
1675 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1677 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1678 blk_mq_hctx_notify, hctx);
1679 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1681 hctx->tags = set->tags[hctx_idx];
1684 * Allocate space for all possible cpus to avoid allocation at
1687 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1690 goto unregister_cpu_notifier;
1692 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1698 if (set->ops->init_hctx &&
1699 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1702 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1706 if (set->ops->init_request &&
1707 set->ops->init_request(set->driver_data,
1708 hctx->fq->flush_rq, hctx_idx,
1709 flush_start_tag + hctx_idx, node))
1717 if (set->ops->exit_hctx)
1718 set->ops->exit_hctx(hctx, hctx_idx);
1720 sbitmap_free(&hctx->ctx_map);
1723 unregister_cpu_notifier:
1724 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1729 static void blk_mq_init_cpu_queues(struct request_queue *q,
1730 unsigned int nr_hw_queues)
1734 for_each_possible_cpu(i) {
1735 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1736 struct blk_mq_hw_ctx *hctx;
1738 memset(__ctx, 0, sizeof(*__ctx));
1740 spin_lock_init(&__ctx->lock);
1741 INIT_LIST_HEAD(&__ctx->rq_list);
1744 /* If the cpu isn't online, the cpu is mapped to first hctx */
1748 hctx = q->mq_ops->map_queue(q, i);
1751 * Set local node, IFF we have more than one hw queue. If
1752 * not, we remain on the home node of the device
1754 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1755 hctx->numa_node = local_memory_node(cpu_to_node(i));
1759 static void blk_mq_map_swqueue(struct request_queue *q,
1760 const struct cpumask *online_mask)
1763 struct blk_mq_hw_ctx *hctx;
1764 struct blk_mq_ctx *ctx;
1765 struct blk_mq_tag_set *set = q->tag_set;
1768 * Avoid others reading imcomplete hctx->cpumask through sysfs
1770 mutex_lock(&q->sysfs_lock);
1772 queue_for_each_hw_ctx(q, hctx, i) {
1773 cpumask_clear(hctx->cpumask);
1778 * Map software to hardware queues
1780 for_each_possible_cpu(i) {
1781 /* If the cpu isn't online, the cpu is mapped to first hctx */
1782 if (!cpumask_test_cpu(i, online_mask))
1785 ctx = per_cpu_ptr(q->queue_ctx, i);
1786 hctx = q->mq_ops->map_queue(q, i);
1788 cpumask_set_cpu(i, hctx->cpumask);
1789 ctx->index_hw = hctx->nr_ctx;
1790 hctx->ctxs[hctx->nr_ctx++] = ctx;
1793 mutex_unlock(&q->sysfs_lock);
1795 queue_for_each_hw_ctx(q, hctx, i) {
1797 * If no software queues are mapped to this hardware queue,
1798 * disable it and free the request entries.
1800 if (!hctx->nr_ctx) {
1802 blk_mq_free_rq_map(set, set->tags[i], i);
1803 set->tags[i] = NULL;
1809 /* unmapped hw queue can be remapped after CPU topo changed */
1811 set->tags[i] = blk_mq_init_rq_map(set, i);
1812 hctx->tags = set->tags[i];
1813 WARN_ON(!hctx->tags);
1815 cpumask_copy(hctx->tags->cpumask, hctx->cpumask);
1817 * Set the map size to the number of mapped software queues.
1818 * This is more accurate and more efficient than looping
1819 * over all possibly mapped software queues.
1821 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
1824 * Initialize batch roundrobin counts
1826 hctx->next_cpu = cpumask_first(hctx->cpumask);
1827 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1831 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1833 struct blk_mq_hw_ctx *hctx;
1836 queue_for_each_hw_ctx(q, hctx, i) {
1838 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1840 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1844 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1846 struct request_queue *q;
1848 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1849 blk_mq_freeze_queue(q);
1850 queue_set_hctx_shared(q, shared);
1851 blk_mq_unfreeze_queue(q);
1855 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1857 struct blk_mq_tag_set *set = q->tag_set;
1859 mutex_lock(&set->tag_list_lock);
1860 list_del_init(&q->tag_set_list);
1861 if (list_is_singular(&set->tag_list)) {
1862 /* just transitioned to unshared */
1863 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1864 /* update existing queue */
1865 blk_mq_update_tag_set_depth(set, false);
1867 mutex_unlock(&set->tag_list_lock);
1870 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1871 struct request_queue *q)
1875 mutex_lock(&set->tag_list_lock);
1877 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1878 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1879 set->flags |= BLK_MQ_F_TAG_SHARED;
1880 /* update existing queue */
1881 blk_mq_update_tag_set_depth(set, true);
1883 if (set->flags & BLK_MQ_F_TAG_SHARED)
1884 queue_set_hctx_shared(q, true);
1885 list_add_tail(&q->tag_set_list, &set->tag_list);
1887 mutex_unlock(&set->tag_list_lock);
1891 * It is the actual release handler for mq, but we do it from
1892 * request queue's release handler for avoiding use-after-free
1893 * and headache because q->mq_kobj shouldn't have been introduced,
1894 * but we can't group ctx/kctx kobj without it.
1896 void blk_mq_release(struct request_queue *q)
1898 struct blk_mq_hw_ctx *hctx;
1901 /* hctx kobj stays in hctx */
1902 queue_for_each_hw_ctx(q, hctx, i) {
1912 kfree(q->queue_hw_ctx);
1914 /* ctx kobj stays in queue_ctx */
1915 free_percpu(q->queue_ctx);
1918 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1920 struct request_queue *uninit_q, *q;
1922 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1924 return ERR_PTR(-ENOMEM);
1926 q = blk_mq_init_allocated_queue(set, uninit_q);
1928 blk_cleanup_queue(uninit_q);
1932 EXPORT_SYMBOL(blk_mq_init_queue);
1934 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1935 struct request_queue *q)
1938 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
1940 blk_mq_sysfs_unregister(q);
1941 for (i = 0; i < set->nr_hw_queues; i++) {
1947 node = blk_mq_hw_queue_to_node(q->mq_map, i);
1948 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1953 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1960 atomic_set(&hctxs[i]->nr_active, 0);
1961 hctxs[i]->numa_node = node;
1962 hctxs[i]->queue_num = i;
1964 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
1965 free_cpumask_var(hctxs[i]->cpumask);
1970 blk_mq_hctx_kobj_init(hctxs[i]);
1972 for (j = i; j < q->nr_hw_queues; j++) {
1973 struct blk_mq_hw_ctx *hctx = hctxs[j];
1977 blk_mq_free_rq_map(set, hctx->tags, j);
1978 set->tags[j] = NULL;
1980 blk_mq_exit_hctx(q, set, hctx, j);
1981 free_cpumask_var(hctx->cpumask);
1982 kobject_put(&hctx->kobj);
1989 q->nr_hw_queues = i;
1990 blk_mq_sysfs_register(q);
1993 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1994 struct request_queue *q)
1996 /* mark the queue as mq asap */
1997 q->mq_ops = set->ops;
1999 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2003 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2004 GFP_KERNEL, set->numa_node);
2005 if (!q->queue_hw_ctx)
2008 q->mq_map = blk_mq_make_queue_map(set);
2012 blk_mq_realloc_hw_ctxs(set, q);
2013 if (!q->nr_hw_queues)
2016 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2017 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2019 q->nr_queues = nr_cpu_ids;
2021 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2023 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2024 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2026 q->sg_reserved_size = INT_MAX;
2028 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2029 INIT_LIST_HEAD(&q->requeue_list);
2030 spin_lock_init(&q->requeue_lock);
2032 if (q->nr_hw_queues > 1)
2033 blk_queue_make_request(q, blk_mq_make_request);
2035 blk_queue_make_request(q, blk_sq_make_request);
2038 * Do this after blk_queue_make_request() overrides it...
2040 q->nr_requests = set->queue_depth;
2042 if (set->ops->complete)
2043 blk_queue_softirq_done(q, set->ops->complete);
2045 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2048 mutex_lock(&all_q_mutex);
2050 list_add_tail(&q->all_q_node, &all_q_list);
2051 blk_mq_add_queue_tag_set(set, q);
2052 blk_mq_map_swqueue(q, cpu_online_mask);
2054 mutex_unlock(&all_q_mutex);
2062 kfree(q->queue_hw_ctx);
2064 free_percpu(q->queue_ctx);
2067 return ERR_PTR(-ENOMEM);
2069 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2071 void blk_mq_free_queue(struct request_queue *q)
2073 struct blk_mq_tag_set *set = q->tag_set;
2075 mutex_lock(&all_q_mutex);
2076 list_del_init(&q->all_q_node);
2077 mutex_unlock(&all_q_mutex);
2079 blk_mq_del_queue_tag_set(q);
2081 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2082 blk_mq_free_hw_queues(q, set);
2085 /* Basically redo blk_mq_init_queue with queue frozen */
2086 static void blk_mq_queue_reinit(struct request_queue *q,
2087 const struct cpumask *online_mask)
2089 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2091 blk_mq_sysfs_unregister(q);
2093 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2096 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2097 * we should change hctx numa_node according to new topology (this
2098 * involves free and re-allocate memory, worthy doing?)
2101 blk_mq_map_swqueue(q, online_mask);
2103 blk_mq_sysfs_register(q);
2106 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2107 unsigned long action, void *hcpu)
2109 struct request_queue *q;
2110 int cpu = (unsigned long)hcpu;
2112 * New online cpumask which is going to be set in this hotplug event.
2113 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2114 * one-by-one and dynamically allocating this could result in a failure.
2116 static struct cpumask online_new;
2119 * Before hotadded cpu starts handling requests, new mappings must
2120 * be established. Otherwise, these requests in hw queue might
2121 * never be dispatched.
2123 * For example, there is a single hw queue (hctx) and two CPU queues
2124 * (ctx0 for CPU0, and ctx1 for CPU1).
2126 * Now CPU1 is just onlined and a request is inserted into
2127 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2130 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2131 * set in pending bitmap and tries to retrieve requests in
2132 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2133 * so the request in ctx1->rq_list is ignored.
2135 switch (action & ~CPU_TASKS_FROZEN) {
2137 case CPU_UP_CANCELED:
2138 cpumask_copy(&online_new, cpu_online_mask);
2140 case CPU_UP_PREPARE:
2141 cpumask_copy(&online_new, cpu_online_mask);
2142 cpumask_set_cpu(cpu, &online_new);
2148 mutex_lock(&all_q_mutex);
2151 * We need to freeze and reinit all existing queues. Freezing
2152 * involves synchronous wait for an RCU grace period and doing it
2153 * one by one may take a long time. Start freezing all queues in
2154 * one swoop and then wait for the completions so that freezing can
2155 * take place in parallel.
2157 list_for_each_entry(q, &all_q_list, all_q_node)
2158 blk_mq_freeze_queue_start(q);
2159 list_for_each_entry(q, &all_q_list, all_q_node) {
2160 blk_mq_freeze_queue_wait(q);
2163 * timeout handler can't touch hw queue during the
2166 del_timer_sync(&q->timeout);
2169 list_for_each_entry(q, &all_q_list, all_q_node)
2170 blk_mq_queue_reinit(q, &online_new);
2172 list_for_each_entry(q, &all_q_list, all_q_node)
2173 blk_mq_unfreeze_queue(q);
2175 mutex_unlock(&all_q_mutex);
2179 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2183 for (i = 0; i < set->nr_hw_queues; i++) {
2184 set->tags[i] = blk_mq_init_rq_map(set, i);
2193 blk_mq_free_rq_map(set, set->tags[i], i);
2199 * Allocate the request maps associated with this tag_set. Note that this
2200 * may reduce the depth asked for, if memory is tight. set->queue_depth
2201 * will be updated to reflect the allocated depth.
2203 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2208 depth = set->queue_depth;
2210 err = __blk_mq_alloc_rq_maps(set);
2214 set->queue_depth >>= 1;
2215 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2219 } while (set->queue_depth);
2221 if (!set->queue_depth || err) {
2222 pr_err("blk-mq: failed to allocate request map\n");
2226 if (depth != set->queue_depth)
2227 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2228 depth, set->queue_depth);
2233 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2235 return tags->cpumask;
2237 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2240 * Alloc a tag set to be associated with one or more request queues.
2241 * May fail with EINVAL for various error conditions. May adjust the
2242 * requested depth down, if if it too large. In that case, the set
2243 * value will be stored in set->queue_depth.
2245 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2247 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2249 if (!set->nr_hw_queues)
2251 if (!set->queue_depth)
2253 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2256 if (!set->ops->queue_rq || !set->ops->map_queue)
2259 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2260 pr_info("blk-mq: reduced tag depth to %u\n",
2262 set->queue_depth = BLK_MQ_MAX_DEPTH;
2266 * If a crashdump is active, then we are potentially in a very
2267 * memory constrained environment. Limit us to 1 queue and
2268 * 64 tags to prevent using too much memory.
2270 if (is_kdump_kernel()) {
2271 set->nr_hw_queues = 1;
2272 set->queue_depth = min(64U, set->queue_depth);
2275 * There is no use for more h/w queues than cpus.
2277 if (set->nr_hw_queues > nr_cpu_ids)
2278 set->nr_hw_queues = nr_cpu_ids;
2280 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2281 GFP_KERNEL, set->numa_node);
2285 if (blk_mq_alloc_rq_maps(set))
2288 mutex_init(&set->tag_list_lock);
2289 INIT_LIST_HEAD(&set->tag_list);
2297 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2299 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2303 for (i = 0; i < nr_cpu_ids; i++) {
2305 blk_mq_free_rq_map(set, set->tags[i], i);
2311 EXPORT_SYMBOL(blk_mq_free_tag_set);
2313 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2315 struct blk_mq_tag_set *set = q->tag_set;
2316 struct blk_mq_hw_ctx *hctx;
2319 if (!set || nr > set->queue_depth)
2323 queue_for_each_hw_ctx(q, hctx, i) {
2326 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2332 q->nr_requests = nr;
2337 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2339 struct request_queue *q;
2341 if (nr_hw_queues > nr_cpu_ids)
2342 nr_hw_queues = nr_cpu_ids;
2343 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2346 list_for_each_entry(q, &set->tag_list, tag_set_list)
2347 blk_mq_freeze_queue(q);
2349 set->nr_hw_queues = nr_hw_queues;
2350 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2351 blk_mq_realloc_hw_ctxs(set, q);
2353 if (q->nr_hw_queues > 1)
2354 blk_queue_make_request(q, blk_mq_make_request);
2356 blk_queue_make_request(q, blk_sq_make_request);
2358 blk_mq_queue_reinit(q, cpu_online_mask);
2361 list_for_each_entry(q, &set->tag_list, tag_set_list)
2362 blk_mq_unfreeze_queue(q);
2364 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2366 void blk_mq_disable_hotplug(void)
2368 mutex_lock(&all_q_mutex);
2371 void blk_mq_enable_hotplug(void)
2373 mutex_unlock(&all_q_mutex);
2376 static int __init blk_mq_init(void)
2380 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2384 subsys_initcall(blk_mq_init);