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>
26 #include <trace/events/block.h>
28 #include <linux/blk-mq.h>
31 #include "blk-mq-tag.h"
33 static DEFINE_MUTEX(all_q_mutex);
34 static LIST_HEAD(all_q_list);
36 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
39 * Check if any of the ctx's have pending work in this hardware queue
41 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
45 for (i = 0; i < hctx->ctx_map.size; i++)
46 if (hctx->ctx_map.map[i].word)
52 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
53 struct blk_mq_ctx *ctx)
55 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
58 #define CTX_TO_BIT(hctx, ctx) \
59 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
62 * Mark this ctx as having pending work in this hardware queue
64 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
65 struct blk_mq_ctx *ctx)
67 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
69 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
70 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
73 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
74 struct blk_mq_ctx *ctx)
76 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
78 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
81 void blk_mq_freeze_queue_start(struct request_queue *q)
85 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
86 if (freeze_depth == 1) {
87 percpu_ref_kill(&q->q_usage_counter);
88 blk_mq_run_hw_queues(q, false);
91 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
93 static void blk_mq_freeze_queue_wait(struct request_queue *q)
95 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
102 void blk_freeze_queue(struct request_queue *q)
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
111 blk_mq_freeze_queue_start(q);
112 blk_mq_freeze_queue_wait(q);
115 void blk_mq_freeze_queue(struct request_queue *q)
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
125 void blk_mq_unfreeze_queue(struct request_queue *q)
129 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
130 WARN_ON_ONCE(freeze_depth < 0);
132 percpu_ref_reinit(&q->q_usage_counter);
133 wake_up_all(&q->mq_freeze_wq);
136 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
138 void blk_mq_wake_waiters(struct request_queue *q)
140 struct blk_mq_hw_ctx *hctx;
143 queue_for_each_hw_ctx(q, hctx, i)
144 if (blk_mq_hw_queue_mapped(hctx))
145 blk_mq_tag_wakeup_all(hctx->tags, true);
148 * If we are called because the queue has now been marked as
149 * dying, we need to ensure that processes currently waiting on
150 * the queue are notified as well.
152 wake_up_all(&q->mq_freeze_wq);
155 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
157 return blk_mq_has_free_tags(hctx->tags);
159 EXPORT_SYMBOL(blk_mq_can_queue);
161 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
162 struct request *rq, int op,
163 unsigned int op_flags)
165 if (blk_queue_io_stat(q))
166 op_flags |= REQ_IO_STAT;
168 INIT_LIST_HEAD(&rq->queuelist);
169 /* csd/requeue_work/fifo_time is initialized before use */
172 req_set_op_attrs(rq, op, op_flags);
173 /* do not touch atomic flags, it needs atomic ops against the timer */
175 INIT_HLIST_NODE(&rq->hash);
176 RB_CLEAR_NODE(&rq->rb_node);
179 rq->start_time = jiffies;
180 #ifdef CONFIG_BLK_CGROUP
182 set_start_time_ns(rq);
183 rq->io_start_time_ns = 0;
185 rq->nr_phys_segments = 0;
186 #if defined(CONFIG_BLK_DEV_INTEGRITY)
187 rq->nr_integrity_segments = 0;
190 /* tag was already set */
200 INIT_LIST_HEAD(&rq->timeout_list);
204 rq->end_io_data = NULL;
207 ctx->rq_dispatched[rw_is_sync(op, op_flags)]++;
210 static struct request *
211 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int op, int op_flags)
216 tag = blk_mq_get_tag(data);
217 if (tag != BLK_MQ_TAG_FAIL) {
218 rq = data->hctx->tags->rqs[tag];
220 if (blk_mq_tag_busy(data->hctx)) {
221 rq->cmd_flags = REQ_MQ_INFLIGHT;
222 atomic_inc(&data->hctx->nr_active);
226 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op, op_flags);
233 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
236 struct blk_mq_ctx *ctx;
237 struct blk_mq_hw_ctx *hctx;
239 struct blk_mq_alloc_data alloc_data;
242 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
246 ctx = blk_mq_get_ctx(q);
247 hctx = q->mq_ops->map_queue(q, ctx->cpu);
248 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
250 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
251 if (!rq && !(flags & BLK_MQ_REQ_NOWAIT)) {
252 __blk_mq_run_hw_queue(hctx);
255 ctx = blk_mq_get_ctx(q);
256 hctx = q->mq_ops->map_queue(q, ctx->cpu);
257 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
258 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
259 ctx = alloc_data.ctx;
264 return ERR_PTR(-EWOULDBLOCK);
268 rq->__sector = (sector_t) -1;
269 rq->bio = rq->biotail = NULL;
272 EXPORT_SYMBOL(blk_mq_alloc_request);
274 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
275 unsigned int flags, unsigned int hctx_idx)
277 struct blk_mq_hw_ctx *hctx;
278 struct blk_mq_ctx *ctx;
280 struct blk_mq_alloc_data alloc_data;
284 * If the tag allocator sleeps we could get an allocation for a
285 * different hardware context. No need to complicate the low level
286 * allocator for this for the rare use case of a command tied to
289 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
290 return ERR_PTR(-EINVAL);
292 if (hctx_idx >= q->nr_hw_queues)
293 return ERR_PTR(-EIO);
295 ret = blk_queue_enter(q, true);
299 hctx = q->queue_hw_ctx[hctx_idx];
300 ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
302 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
303 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
306 return ERR_PTR(-EWOULDBLOCK);
311 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
313 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
314 struct blk_mq_ctx *ctx, struct request *rq)
316 const int tag = rq->tag;
317 struct request_queue *q = rq->q;
319 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
320 atomic_dec(&hctx->nr_active);
323 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
324 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
328 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
330 struct blk_mq_ctx *ctx = rq->mq_ctx;
332 ctx->rq_completed[rq_is_sync(rq)]++;
333 __blk_mq_free_request(hctx, ctx, rq);
336 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
338 void blk_mq_free_request(struct request *rq)
340 struct blk_mq_hw_ctx *hctx;
341 struct request_queue *q = rq->q;
343 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
344 blk_mq_free_hctx_request(hctx, rq);
346 EXPORT_SYMBOL_GPL(blk_mq_free_request);
348 inline void __blk_mq_end_request(struct request *rq, int error)
350 blk_account_io_done(rq);
353 rq->end_io(rq, error);
355 if (unlikely(blk_bidi_rq(rq)))
356 blk_mq_free_request(rq->next_rq);
357 blk_mq_free_request(rq);
360 EXPORT_SYMBOL(__blk_mq_end_request);
362 void blk_mq_end_request(struct request *rq, int error)
364 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
366 __blk_mq_end_request(rq, error);
368 EXPORT_SYMBOL(blk_mq_end_request);
370 static void __blk_mq_complete_request_remote(void *data)
372 struct request *rq = data;
374 rq->q->softirq_done_fn(rq);
377 static void blk_mq_ipi_complete_request(struct request *rq)
379 struct blk_mq_ctx *ctx = rq->mq_ctx;
383 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
384 rq->q->softirq_done_fn(rq);
389 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
390 shared = cpus_share_cache(cpu, ctx->cpu);
392 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
393 rq->csd.func = __blk_mq_complete_request_remote;
396 smp_call_function_single_async(ctx->cpu, &rq->csd);
398 rq->q->softirq_done_fn(rq);
403 static void __blk_mq_complete_request(struct request *rq)
405 struct request_queue *q = rq->q;
407 if (!q->softirq_done_fn)
408 blk_mq_end_request(rq, rq->errors);
410 blk_mq_ipi_complete_request(rq);
414 * blk_mq_complete_request - end I/O on a request
415 * @rq: the request being processed
418 * Ends all I/O on a request. It does not handle partial completions.
419 * The actual completion happens out-of-order, through a IPI handler.
421 void blk_mq_complete_request(struct request *rq, int error)
423 struct request_queue *q = rq->q;
425 if (unlikely(blk_should_fake_timeout(q)))
427 if (!blk_mark_rq_complete(rq)) {
429 __blk_mq_complete_request(rq);
432 EXPORT_SYMBOL(blk_mq_complete_request);
434 int blk_mq_request_started(struct request *rq)
436 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
438 EXPORT_SYMBOL_GPL(blk_mq_request_started);
440 void blk_mq_start_request(struct request *rq)
442 struct request_queue *q = rq->q;
444 trace_block_rq_issue(q, rq);
446 rq->resid_len = blk_rq_bytes(rq);
447 if (unlikely(blk_bidi_rq(rq)))
448 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
453 * Ensure that ->deadline is visible before set the started
454 * flag and clear the completed flag.
456 smp_mb__before_atomic();
459 * Mark us as started and clear complete. Complete might have been
460 * set if requeue raced with timeout, which then marked it as
461 * complete. So be sure to clear complete again when we start
462 * the request, otherwise we'll ignore the completion event.
464 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
465 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
466 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
467 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
469 if (q->dma_drain_size && blk_rq_bytes(rq)) {
471 * Make sure space for the drain appears. We know we can do
472 * this because max_hw_segments has been adjusted to be one
473 * fewer than the device can handle.
475 rq->nr_phys_segments++;
478 EXPORT_SYMBOL(blk_mq_start_request);
480 static void __blk_mq_requeue_request(struct request *rq)
482 struct request_queue *q = rq->q;
484 trace_block_rq_requeue(q, rq);
486 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
487 if (q->dma_drain_size && blk_rq_bytes(rq))
488 rq->nr_phys_segments--;
492 void blk_mq_requeue_request(struct request *rq)
494 __blk_mq_requeue_request(rq);
496 BUG_ON(blk_queued_rq(rq));
497 blk_mq_add_to_requeue_list(rq, true);
499 EXPORT_SYMBOL(blk_mq_requeue_request);
501 static void blk_mq_requeue_work(struct work_struct *work)
503 struct request_queue *q =
504 container_of(work, struct request_queue, requeue_work);
506 struct request *rq, *next;
509 spin_lock_irqsave(&q->requeue_lock, flags);
510 list_splice_init(&q->requeue_list, &rq_list);
511 spin_unlock_irqrestore(&q->requeue_lock, flags);
513 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
514 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
517 rq->cmd_flags &= ~REQ_SOFTBARRIER;
518 list_del_init(&rq->queuelist);
519 blk_mq_insert_request(rq, true, false, false);
522 while (!list_empty(&rq_list)) {
523 rq = list_entry(rq_list.next, struct request, queuelist);
524 list_del_init(&rq->queuelist);
525 blk_mq_insert_request(rq, false, false, false);
529 * Use the start variant of queue running here, so that running
530 * the requeue work will kick stopped queues.
532 blk_mq_start_hw_queues(q);
535 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
537 struct request_queue *q = rq->q;
541 * We abuse this flag that is otherwise used by the I/O scheduler to
542 * request head insertation from the workqueue.
544 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
546 spin_lock_irqsave(&q->requeue_lock, flags);
548 rq->cmd_flags |= REQ_SOFTBARRIER;
549 list_add(&rq->queuelist, &q->requeue_list);
551 list_add_tail(&rq->queuelist, &q->requeue_list);
553 spin_unlock_irqrestore(&q->requeue_lock, flags);
555 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
557 void blk_mq_cancel_requeue_work(struct request_queue *q)
559 cancel_work_sync(&q->requeue_work);
561 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
563 void blk_mq_kick_requeue_list(struct request_queue *q)
565 kblockd_schedule_work(&q->requeue_work);
567 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
569 void blk_mq_abort_requeue_list(struct request_queue *q)
574 spin_lock_irqsave(&q->requeue_lock, flags);
575 list_splice_init(&q->requeue_list, &rq_list);
576 spin_unlock_irqrestore(&q->requeue_lock, flags);
578 while (!list_empty(&rq_list)) {
581 rq = list_first_entry(&rq_list, struct request, queuelist);
582 list_del_init(&rq->queuelist);
584 blk_mq_end_request(rq, rq->errors);
587 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
589 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
591 if (tag < tags->nr_tags)
592 return tags->rqs[tag];
596 EXPORT_SYMBOL(blk_mq_tag_to_rq);
598 struct blk_mq_timeout_data {
600 unsigned int next_set;
603 void blk_mq_rq_timed_out(struct request *req, bool reserved)
605 struct blk_mq_ops *ops = req->q->mq_ops;
606 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
609 * We know that complete is set at this point. If STARTED isn't set
610 * anymore, then the request isn't active and the "timeout" should
611 * just be ignored. This can happen due to the bitflag ordering.
612 * Timeout first checks if STARTED is set, and if it is, assumes
613 * the request is active. But if we race with completion, then
614 * we both flags will get cleared. So check here again, and ignore
615 * a timeout event with a request that isn't active.
617 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
621 ret = ops->timeout(req, reserved);
625 __blk_mq_complete_request(req);
627 case BLK_EH_RESET_TIMER:
629 blk_clear_rq_complete(req);
631 case BLK_EH_NOT_HANDLED:
634 printk(KERN_ERR "block: bad eh return: %d\n", ret);
639 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
640 struct request *rq, void *priv, bool reserved)
642 struct blk_mq_timeout_data *data = priv;
644 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
646 * If a request wasn't started before the queue was
647 * marked dying, kill it here or it'll go unnoticed.
649 if (unlikely(blk_queue_dying(rq->q))) {
651 blk_mq_end_request(rq, rq->errors);
656 if (time_after_eq(jiffies, rq->deadline)) {
657 if (!blk_mark_rq_complete(rq))
658 blk_mq_rq_timed_out(rq, reserved);
659 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
660 data->next = rq->deadline;
665 static void blk_mq_timeout_work(struct work_struct *work)
667 struct request_queue *q =
668 container_of(work, struct request_queue, timeout_work);
669 struct blk_mq_timeout_data data = {
675 if (blk_queue_enter(q, true))
678 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
681 data.next = blk_rq_timeout(round_jiffies_up(data.next));
682 mod_timer(&q->timeout, data.next);
684 struct blk_mq_hw_ctx *hctx;
686 queue_for_each_hw_ctx(q, hctx, i) {
687 /* the hctx may be unmapped, so check it here */
688 if (blk_mq_hw_queue_mapped(hctx))
689 blk_mq_tag_idle(hctx);
696 * Reverse check our software queue for entries that we could potentially
697 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
698 * too much time checking for merges.
700 static bool blk_mq_attempt_merge(struct request_queue *q,
701 struct blk_mq_ctx *ctx, struct bio *bio)
706 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
712 if (!blk_rq_merge_ok(rq, bio))
715 el_ret = blk_try_merge(rq, bio);
716 if (el_ret == ELEVATOR_BACK_MERGE) {
717 if (bio_attempt_back_merge(q, rq, bio)) {
722 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
723 if (bio_attempt_front_merge(q, rq, bio)) {
735 * Process software queues that have been marked busy, splicing them
736 * to the for-dispatch
738 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
740 struct blk_mq_ctx *ctx;
743 for (i = 0; i < hctx->ctx_map.size; i++) {
744 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
745 unsigned int off, bit;
751 off = i * hctx->ctx_map.bits_per_word;
753 bit = find_next_bit(&bm->word, bm->depth, bit);
754 if (bit >= bm->depth)
757 ctx = hctx->ctxs[bit + off];
758 clear_bit(bit, &bm->word);
759 spin_lock(&ctx->lock);
760 list_splice_tail_init(&ctx->rq_list, list);
761 spin_unlock(&ctx->lock);
769 * Run this hardware queue, pulling any software queues mapped to it in.
770 * Note that this function currently has various problems around ordering
771 * of IO. In particular, we'd like FIFO behaviour on handling existing
772 * items on the hctx->dispatch list. Ignore that for now.
774 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
776 struct request_queue *q = hctx->queue;
779 LIST_HEAD(driver_list);
780 struct list_head *dptr;
783 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
785 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
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)
856 hctx->dispatched[0]++;
857 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
858 hctx->dispatched[ilog2(queued) + 1]++;
861 * Any items that need requeuing? Stuff them into hctx->dispatch,
862 * that is where we will continue on next queue run.
864 if (!list_empty(&rq_list)) {
865 spin_lock(&hctx->lock);
866 list_splice(&rq_list, &hctx->dispatch);
867 spin_unlock(&hctx->lock);
869 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
870 * it's possible the queue is stopped and restarted again
871 * before this. Queue restart will dispatch requests. And since
872 * requests in rq_list aren't added into hctx->dispatch yet,
873 * the requests in rq_list might get lost.
875 * blk_mq_run_hw_queue() already checks the STOPPED bit
877 blk_mq_run_hw_queue(hctx, true);
882 * It'd be great if the workqueue API had a way to pass
883 * in a mask and had some smarts for more clever placement.
884 * For now we just round-robin here, switching for every
885 * BLK_MQ_CPU_WORK_BATCH queued items.
887 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
889 if (hctx->queue->nr_hw_queues == 1)
890 return WORK_CPU_UNBOUND;
892 if (--hctx->next_cpu_batch <= 0) {
893 int cpu = hctx->next_cpu, next_cpu;
895 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
896 if (next_cpu >= nr_cpu_ids)
897 next_cpu = cpumask_first(hctx->cpumask);
899 hctx->next_cpu = next_cpu;
900 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
905 return hctx->next_cpu;
908 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
910 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
911 !blk_mq_hw_queue_mapped(hctx)))
916 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
917 __blk_mq_run_hw_queue(hctx);
925 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
929 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
931 struct blk_mq_hw_ctx *hctx;
934 queue_for_each_hw_ctx(q, hctx, i) {
935 if ((!blk_mq_hctx_has_pending(hctx) &&
936 list_empty_careful(&hctx->dispatch)) ||
937 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
940 blk_mq_run_hw_queue(hctx, async);
943 EXPORT_SYMBOL(blk_mq_run_hw_queues);
945 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
947 cancel_delayed_work(&hctx->run_work);
948 cancel_delayed_work(&hctx->delay_work);
949 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
951 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
953 void blk_mq_stop_hw_queues(struct request_queue *q)
955 struct blk_mq_hw_ctx *hctx;
958 queue_for_each_hw_ctx(q, hctx, i)
959 blk_mq_stop_hw_queue(hctx);
961 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
963 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
965 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
967 blk_mq_run_hw_queue(hctx, false);
969 EXPORT_SYMBOL(blk_mq_start_hw_queue);
971 void blk_mq_start_hw_queues(struct request_queue *q)
973 struct blk_mq_hw_ctx *hctx;
976 queue_for_each_hw_ctx(q, hctx, i)
977 blk_mq_start_hw_queue(hctx);
979 EXPORT_SYMBOL(blk_mq_start_hw_queues);
981 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
983 struct blk_mq_hw_ctx *hctx;
986 queue_for_each_hw_ctx(q, hctx, i) {
987 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
990 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
991 blk_mq_run_hw_queue(hctx, async);
994 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
996 static void blk_mq_run_work_fn(struct work_struct *work)
998 struct blk_mq_hw_ctx *hctx;
1000 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1002 __blk_mq_run_hw_queue(hctx);
1005 static void blk_mq_delay_work_fn(struct work_struct *work)
1007 struct blk_mq_hw_ctx *hctx;
1009 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1011 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1012 __blk_mq_run_hw_queue(hctx);
1015 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1017 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1020 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1021 &hctx->delay_work, msecs_to_jiffies(msecs));
1023 EXPORT_SYMBOL(blk_mq_delay_queue);
1025 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1026 struct blk_mq_ctx *ctx,
1030 trace_block_rq_insert(hctx->queue, rq);
1033 list_add(&rq->queuelist, &ctx->rq_list);
1035 list_add_tail(&rq->queuelist, &ctx->rq_list);
1038 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1039 struct request *rq, bool at_head)
1041 struct blk_mq_ctx *ctx = rq->mq_ctx;
1043 __blk_mq_insert_req_list(hctx, ctx, rq, at_head);
1044 blk_mq_hctx_mark_pending(hctx, ctx);
1047 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1050 struct request_queue *q = rq->q;
1051 struct blk_mq_hw_ctx *hctx;
1052 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1054 current_ctx = blk_mq_get_ctx(q);
1055 if (!cpu_online(ctx->cpu))
1056 rq->mq_ctx = ctx = current_ctx;
1058 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1060 spin_lock(&ctx->lock);
1061 __blk_mq_insert_request(hctx, rq, at_head);
1062 spin_unlock(&ctx->lock);
1065 blk_mq_run_hw_queue(hctx, async);
1067 blk_mq_put_ctx(current_ctx);
1070 static void blk_mq_insert_requests(struct request_queue *q,
1071 struct blk_mq_ctx *ctx,
1072 struct list_head *list,
1077 struct blk_mq_hw_ctx *hctx;
1078 struct blk_mq_ctx *current_ctx;
1080 trace_block_unplug(q, depth, !from_schedule);
1082 current_ctx = blk_mq_get_ctx(q);
1084 if (!cpu_online(ctx->cpu))
1086 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1089 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1092 spin_lock(&ctx->lock);
1093 while (!list_empty(list)) {
1096 rq = list_first_entry(list, struct request, queuelist);
1097 list_del_init(&rq->queuelist);
1099 __blk_mq_insert_req_list(hctx, ctx, rq, false);
1101 blk_mq_hctx_mark_pending(hctx, ctx);
1102 spin_unlock(&ctx->lock);
1104 blk_mq_run_hw_queue(hctx, from_schedule);
1105 blk_mq_put_ctx(current_ctx);
1108 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1110 struct request *rqa = container_of(a, struct request, queuelist);
1111 struct request *rqb = container_of(b, struct request, queuelist);
1113 return !(rqa->mq_ctx < rqb->mq_ctx ||
1114 (rqa->mq_ctx == rqb->mq_ctx &&
1115 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1118 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1120 struct blk_mq_ctx *this_ctx;
1121 struct request_queue *this_q;
1124 LIST_HEAD(ctx_list);
1127 list_splice_init(&plug->mq_list, &list);
1129 list_sort(NULL, &list, plug_ctx_cmp);
1135 while (!list_empty(&list)) {
1136 rq = list_entry_rq(list.next);
1137 list_del_init(&rq->queuelist);
1139 if (rq->mq_ctx != this_ctx) {
1141 blk_mq_insert_requests(this_q, this_ctx,
1146 this_ctx = rq->mq_ctx;
1152 list_add_tail(&rq->queuelist, &ctx_list);
1156 * If 'this_ctx' is set, we know we have entries to complete
1157 * on 'ctx_list'. Do those.
1160 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1165 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1167 init_request_from_bio(rq, bio);
1169 blk_account_io_start(rq, 1);
1172 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1174 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1175 !blk_queue_nomerges(hctx->queue);
1178 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1179 struct blk_mq_ctx *ctx,
1180 struct request *rq, struct bio *bio)
1182 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1183 blk_mq_bio_to_request(rq, bio);
1184 spin_lock(&ctx->lock);
1186 __blk_mq_insert_request(hctx, rq, false);
1187 spin_unlock(&ctx->lock);
1190 struct request_queue *q = hctx->queue;
1192 spin_lock(&ctx->lock);
1193 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1194 blk_mq_bio_to_request(rq, bio);
1198 spin_unlock(&ctx->lock);
1199 __blk_mq_free_request(hctx, ctx, rq);
1204 struct blk_map_ctx {
1205 struct blk_mq_hw_ctx *hctx;
1206 struct blk_mq_ctx *ctx;
1209 static struct request *blk_mq_map_request(struct request_queue *q,
1211 struct blk_map_ctx *data)
1213 struct blk_mq_hw_ctx *hctx;
1214 struct blk_mq_ctx *ctx;
1216 int op = bio_data_dir(bio);
1218 struct blk_mq_alloc_data alloc_data;
1220 blk_queue_enter_live(q);
1221 ctx = blk_mq_get_ctx(q);
1222 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1224 if (rw_is_sync(bio_op(bio), bio->bi_rw))
1225 op_flags |= REQ_SYNC;
1227 trace_block_getrq(q, bio, op);
1228 blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx);
1229 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1230 if (unlikely(!rq)) {
1231 __blk_mq_run_hw_queue(hctx);
1232 blk_mq_put_ctx(ctx);
1233 trace_block_sleeprq(q, bio, op);
1235 ctx = blk_mq_get_ctx(q);
1236 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1237 blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1238 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1239 ctx = alloc_data.ctx;
1240 hctx = alloc_data.hctx;
1249 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1252 struct request_queue *q = rq->q;
1253 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1255 struct blk_mq_queue_data bd = {
1260 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1263 * For OK queue, we are done. For error, kill it. Any other
1264 * error (busy), just add it to our list as we previously
1267 ret = q->mq_ops->queue_rq(hctx, &bd);
1268 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1269 *cookie = new_cookie;
1273 __blk_mq_requeue_request(rq);
1275 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1276 *cookie = BLK_QC_T_NONE;
1278 blk_mq_end_request(rq, rq->errors);
1286 * Multiple hardware queue variant. This will not use per-process plugs,
1287 * but will attempt to bypass the hctx queueing if we can go straight to
1288 * hardware for SYNC IO.
1290 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1292 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_rw);
1293 const int is_flush_fua = bio->bi_rw & (REQ_PREFLUSH | REQ_FUA);
1294 struct blk_map_ctx data;
1296 unsigned int request_count = 0;
1297 struct blk_plug *plug;
1298 struct request *same_queue_rq = NULL;
1301 blk_queue_bounce(q, &bio);
1303 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1305 return BLK_QC_T_NONE;
1308 blk_queue_split(q, &bio, q->bio_split);
1310 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1311 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1312 return BLK_QC_T_NONE;
1314 rq = blk_mq_map_request(q, bio, &data);
1316 return BLK_QC_T_NONE;
1318 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1320 if (unlikely(is_flush_fua)) {
1321 blk_mq_bio_to_request(rq, bio);
1322 blk_insert_flush(rq);
1326 plug = current->plug;
1328 * If the driver supports defer issued based on 'last', then
1329 * queue it up like normal since we can potentially save some
1332 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1333 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1334 struct request *old_rq = NULL;
1336 blk_mq_bio_to_request(rq, bio);
1339 * We do limited pluging. If the bio can be merged, do that.
1340 * Otherwise the existing request in the plug list will be
1341 * issued. So the plug list will have one request at most
1345 * The plug list might get flushed before this. If that
1346 * happens, same_queue_rq is invalid and plug list is
1349 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1350 old_rq = same_queue_rq;
1351 list_del_init(&old_rq->queuelist);
1353 list_add_tail(&rq->queuelist, &plug->mq_list);
1354 } else /* is_sync */
1356 blk_mq_put_ctx(data.ctx);
1359 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1361 blk_mq_insert_request(old_rq, false, true, true);
1365 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1367 * For a SYNC request, send it to the hardware immediately. For
1368 * an ASYNC request, just ensure that we run it later on. The
1369 * latter allows for merging opportunities and more efficient
1373 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1375 blk_mq_put_ctx(data.ctx);
1381 * Single hardware queue variant. This will attempt to use any per-process
1382 * plug for merging and IO deferral.
1384 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1386 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_rw);
1387 const int is_flush_fua = bio->bi_rw & (REQ_PREFLUSH | REQ_FUA);
1388 struct blk_plug *plug;
1389 unsigned int request_count = 0;
1390 struct blk_map_ctx data;
1394 blk_queue_bounce(q, &bio);
1396 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1398 return BLK_QC_T_NONE;
1401 blk_queue_split(q, &bio, q->bio_split);
1403 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1404 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1405 return BLK_QC_T_NONE;
1407 request_count = blk_plug_queued_count(q);
1409 rq = blk_mq_map_request(q, bio, &data);
1411 return BLK_QC_T_NONE;
1413 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1415 if (unlikely(is_flush_fua)) {
1416 blk_mq_bio_to_request(rq, bio);
1417 blk_insert_flush(rq);
1422 * A task plug currently exists. Since this is completely lockless,
1423 * utilize that to temporarily store requests until the task is
1424 * either done or scheduled away.
1426 plug = current->plug;
1428 blk_mq_bio_to_request(rq, bio);
1430 trace_block_plug(q);
1432 blk_mq_put_ctx(data.ctx);
1434 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1435 blk_flush_plug_list(plug, false);
1436 trace_block_plug(q);
1439 list_add_tail(&rq->queuelist, &plug->mq_list);
1443 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1445 * For a SYNC request, send it to the hardware immediately. For
1446 * an ASYNC request, just ensure that we run it later on. The
1447 * latter allows for merging opportunities and more efficient
1451 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1454 blk_mq_put_ctx(data.ctx);
1459 * Default mapping to a software queue, since we use one per CPU.
1461 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1463 return q->queue_hw_ctx[q->mq_map[cpu]];
1465 EXPORT_SYMBOL(blk_mq_map_queue);
1467 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1468 struct blk_mq_tags *tags, unsigned int hctx_idx)
1472 if (tags->rqs && set->ops->exit_request) {
1475 for (i = 0; i < tags->nr_tags; i++) {
1478 set->ops->exit_request(set->driver_data, tags->rqs[i],
1480 tags->rqs[i] = NULL;
1484 while (!list_empty(&tags->page_list)) {
1485 page = list_first_entry(&tags->page_list, struct page, lru);
1486 list_del_init(&page->lru);
1488 * Remove kmemleak object previously allocated in
1489 * blk_mq_init_rq_map().
1491 kmemleak_free(page_address(page));
1492 __free_pages(page, page->private);
1497 blk_mq_free_tags(tags);
1500 static size_t order_to_size(unsigned int order)
1502 return (size_t)PAGE_SIZE << order;
1505 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1506 unsigned int hctx_idx)
1508 struct blk_mq_tags *tags;
1509 unsigned int i, j, entries_per_page, max_order = 4;
1510 size_t rq_size, left;
1512 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1514 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1518 INIT_LIST_HEAD(&tags->page_list);
1520 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1521 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1524 blk_mq_free_tags(tags);
1529 * rq_size is the size of the request plus driver payload, rounded
1530 * to the cacheline size
1532 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1534 left = rq_size * set->queue_depth;
1536 for (i = 0; i < set->queue_depth; ) {
1537 int this_order = max_order;
1542 while (this_order && left < order_to_size(this_order - 1))
1546 page = alloc_pages_node(set->numa_node,
1547 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1553 if (order_to_size(this_order) < rq_size)
1560 page->private = this_order;
1561 list_add_tail(&page->lru, &tags->page_list);
1563 p = page_address(page);
1565 * Allow kmemleak to scan these pages as they contain pointers
1566 * to additional allocations like via ops->init_request().
1568 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1569 entries_per_page = order_to_size(this_order) / rq_size;
1570 to_do = min(entries_per_page, set->queue_depth - i);
1571 left -= to_do * rq_size;
1572 for (j = 0; j < to_do; j++) {
1574 if (set->ops->init_request) {
1575 if (set->ops->init_request(set->driver_data,
1576 tags->rqs[i], hctx_idx, i,
1578 tags->rqs[i] = NULL;
1590 blk_mq_free_rq_map(set, tags, hctx_idx);
1594 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1599 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1601 unsigned int bpw = 8, total, num_maps, i;
1603 bitmap->bits_per_word = bpw;
1605 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1606 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1612 for (i = 0; i < num_maps; i++) {
1613 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1614 total -= bitmap->map[i].depth;
1620 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1622 struct request_queue *q = hctx->queue;
1623 struct blk_mq_ctx *ctx;
1627 * Move ctx entries to new CPU, if this one is going away.
1629 ctx = __blk_mq_get_ctx(q, cpu);
1631 spin_lock(&ctx->lock);
1632 if (!list_empty(&ctx->rq_list)) {
1633 list_splice_init(&ctx->rq_list, &tmp);
1634 blk_mq_hctx_clear_pending(hctx, ctx);
1636 spin_unlock(&ctx->lock);
1638 if (list_empty(&tmp))
1641 ctx = blk_mq_get_ctx(q);
1642 spin_lock(&ctx->lock);
1644 while (!list_empty(&tmp)) {
1647 rq = list_first_entry(&tmp, struct request, queuelist);
1649 list_move_tail(&rq->queuelist, &ctx->rq_list);
1652 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1653 blk_mq_hctx_mark_pending(hctx, ctx);
1655 spin_unlock(&ctx->lock);
1657 blk_mq_run_hw_queue(hctx, true);
1658 blk_mq_put_ctx(ctx);
1662 static int blk_mq_hctx_notify(void *data, unsigned long action,
1665 struct blk_mq_hw_ctx *hctx = data;
1667 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1668 return blk_mq_hctx_cpu_offline(hctx, cpu);
1671 * In case of CPU online, tags may be reallocated
1672 * in blk_mq_map_swqueue() after mapping is updated.
1678 /* hctx->ctxs will be freed in queue's release handler */
1679 static void blk_mq_exit_hctx(struct request_queue *q,
1680 struct blk_mq_tag_set *set,
1681 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1683 unsigned flush_start_tag = set->queue_depth;
1685 blk_mq_tag_idle(hctx);
1687 if (set->ops->exit_request)
1688 set->ops->exit_request(set->driver_data,
1689 hctx->fq->flush_rq, hctx_idx,
1690 flush_start_tag + hctx_idx);
1692 if (set->ops->exit_hctx)
1693 set->ops->exit_hctx(hctx, hctx_idx);
1695 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1696 blk_free_flush_queue(hctx->fq);
1697 blk_mq_free_bitmap(&hctx->ctx_map);
1700 static void blk_mq_exit_hw_queues(struct request_queue *q,
1701 struct blk_mq_tag_set *set, int nr_queue)
1703 struct blk_mq_hw_ctx *hctx;
1706 queue_for_each_hw_ctx(q, hctx, i) {
1709 blk_mq_exit_hctx(q, set, hctx, i);
1713 static void blk_mq_free_hw_queues(struct request_queue *q,
1714 struct blk_mq_tag_set *set)
1716 struct blk_mq_hw_ctx *hctx;
1719 queue_for_each_hw_ctx(q, hctx, i)
1720 free_cpumask_var(hctx->cpumask);
1723 static int blk_mq_init_hctx(struct request_queue *q,
1724 struct blk_mq_tag_set *set,
1725 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1728 unsigned flush_start_tag = set->queue_depth;
1730 node = hctx->numa_node;
1731 if (node == NUMA_NO_NODE)
1732 node = hctx->numa_node = set->numa_node;
1734 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1735 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1736 spin_lock_init(&hctx->lock);
1737 INIT_LIST_HEAD(&hctx->dispatch);
1739 hctx->queue_num = hctx_idx;
1740 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1742 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1743 blk_mq_hctx_notify, hctx);
1744 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1746 hctx->tags = set->tags[hctx_idx];
1749 * Allocate space for all possible cpus to avoid allocation at
1752 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1755 goto unregister_cpu_notifier;
1757 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1762 if (set->ops->init_hctx &&
1763 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1766 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1770 if (set->ops->init_request &&
1771 set->ops->init_request(set->driver_data,
1772 hctx->fq->flush_rq, hctx_idx,
1773 flush_start_tag + hctx_idx, node))
1781 if (set->ops->exit_hctx)
1782 set->ops->exit_hctx(hctx, hctx_idx);
1784 blk_mq_free_bitmap(&hctx->ctx_map);
1787 unregister_cpu_notifier:
1788 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1793 static void blk_mq_init_cpu_queues(struct request_queue *q,
1794 unsigned int nr_hw_queues)
1798 for_each_possible_cpu(i) {
1799 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1800 struct blk_mq_hw_ctx *hctx;
1802 memset(__ctx, 0, sizeof(*__ctx));
1804 spin_lock_init(&__ctx->lock);
1805 INIT_LIST_HEAD(&__ctx->rq_list);
1808 /* If the cpu isn't online, the cpu is mapped to first hctx */
1812 hctx = q->mq_ops->map_queue(q, i);
1815 * Set local node, IFF we have more than one hw queue. If
1816 * not, we remain on the home node of the device
1818 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1819 hctx->numa_node = local_memory_node(cpu_to_node(i));
1823 static void blk_mq_map_swqueue(struct request_queue *q,
1824 const struct cpumask *online_mask)
1827 struct blk_mq_hw_ctx *hctx;
1828 struct blk_mq_ctx *ctx;
1829 struct blk_mq_tag_set *set = q->tag_set;
1832 * Avoid others reading imcomplete hctx->cpumask through sysfs
1834 mutex_lock(&q->sysfs_lock);
1836 queue_for_each_hw_ctx(q, hctx, i) {
1837 cpumask_clear(hctx->cpumask);
1842 * Map software to hardware queues
1844 for_each_possible_cpu(i) {
1845 /* If the cpu isn't online, the cpu is mapped to first hctx */
1846 if (!cpumask_test_cpu(i, online_mask))
1849 ctx = per_cpu_ptr(q->queue_ctx, i);
1850 hctx = q->mq_ops->map_queue(q, i);
1852 cpumask_set_cpu(i, hctx->cpumask);
1853 ctx->index_hw = hctx->nr_ctx;
1854 hctx->ctxs[hctx->nr_ctx++] = ctx;
1857 mutex_unlock(&q->sysfs_lock);
1859 queue_for_each_hw_ctx(q, hctx, i) {
1860 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1863 * If no software queues are mapped to this hardware queue,
1864 * disable it and free the request entries.
1866 if (!hctx->nr_ctx) {
1868 blk_mq_free_rq_map(set, set->tags[i], i);
1869 set->tags[i] = NULL;
1875 /* unmapped hw queue can be remapped after CPU topo changed */
1877 set->tags[i] = blk_mq_init_rq_map(set, i);
1878 hctx->tags = set->tags[i];
1879 WARN_ON(!hctx->tags);
1881 cpumask_copy(hctx->tags->cpumask, hctx->cpumask);
1883 * Set the map size to the number of mapped software queues.
1884 * This is more accurate and more efficient than looping
1885 * over all possibly mapped software queues.
1887 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1890 * Initialize batch roundrobin counts
1892 hctx->next_cpu = cpumask_first(hctx->cpumask);
1893 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1897 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1899 struct blk_mq_hw_ctx *hctx;
1902 queue_for_each_hw_ctx(q, hctx, i) {
1904 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1906 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1910 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1912 struct request_queue *q;
1914 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1915 blk_mq_freeze_queue(q);
1916 queue_set_hctx_shared(q, shared);
1917 blk_mq_unfreeze_queue(q);
1921 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1923 struct blk_mq_tag_set *set = q->tag_set;
1925 mutex_lock(&set->tag_list_lock);
1926 list_del_init(&q->tag_set_list);
1927 if (list_is_singular(&set->tag_list)) {
1928 /* just transitioned to unshared */
1929 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1930 /* update existing queue */
1931 blk_mq_update_tag_set_depth(set, false);
1933 mutex_unlock(&set->tag_list_lock);
1936 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1937 struct request_queue *q)
1941 mutex_lock(&set->tag_list_lock);
1943 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1944 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1945 set->flags |= BLK_MQ_F_TAG_SHARED;
1946 /* update existing queue */
1947 blk_mq_update_tag_set_depth(set, true);
1949 if (set->flags & BLK_MQ_F_TAG_SHARED)
1950 queue_set_hctx_shared(q, true);
1951 list_add_tail(&q->tag_set_list, &set->tag_list);
1953 mutex_unlock(&set->tag_list_lock);
1957 * It is the actual release handler for mq, but we do it from
1958 * request queue's release handler for avoiding use-after-free
1959 * and headache because q->mq_kobj shouldn't have been introduced,
1960 * but we can't group ctx/kctx kobj without it.
1962 void blk_mq_release(struct request_queue *q)
1964 struct blk_mq_hw_ctx *hctx;
1967 /* hctx kobj stays in hctx */
1968 queue_for_each_hw_ctx(q, hctx, i) {
1978 kfree(q->queue_hw_ctx);
1980 /* ctx kobj stays in queue_ctx */
1981 free_percpu(q->queue_ctx);
1984 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1986 struct request_queue *uninit_q, *q;
1988 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1990 return ERR_PTR(-ENOMEM);
1992 q = blk_mq_init_allocated_queue(set, uninit_q);
1994 blk_cleanup_queue(uninit_q);
1998 EXPORT_SYMBOL(blk_mq_init_queue);
2000 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2001 struct request_queue *q)
2004 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2006 blk_mq_sysfs_unregister(q);
2007 for (i = 0; i < set->nr_hw_queues; i++) {
2013 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2014 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2019 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2026 atomic_set(&hctxs[i]->nr_active, 0);
2027 hctxs[i]->numa_node = node;
2028 hctxs[i]->queue_num = i;
2030 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2031 free_cpumask_var(hctxs[i]->cpumask);
2036 blk_mq_hctx_kobj_init(hctxs[i]);
2038 for (j = i; j < q->nr_hw_queues; j++) {
2039 struct blk_mq_hw_ctx *hctx = hctxs[j];
2043 blk_mq_free_rq_map(set, hctx->tags, j);
2044 set->tags[j] = NULL;
2046 blk_mq_exit_hctx(q, set, hctx, j);
2047 free_cpumask_var(hctx->cpumask);
2048 kobject_put(&hctx->kobj);
2055 q->nr_hw_queues = i;
2056 blk_mq_sysfs_register(q);
2059 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2060 struct request_queue *q)
2062 /* mark the queue as mq asap */
2063 q->mq_ops = set->ops;
2065 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2069 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2070 GFP_KERNEL, set->numa_node);
2071 if (!q->queue_hw_ctx)
2074 q->mq_map = blk_mq_make_queue_map(set);
2078 blk_mq_realloc_hw_ctxs(set, q);
2079 if (!q->nr_hw_queues)
2082 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2083 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2085 q->nr_queues = nr_cpu_ids;
2087 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2089 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2090 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2092 q->sg_reserved_size = INT_MAX;
2094 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2095 INIT_LIST_HEAD(&q->requeue_list);
2096 spin_lock_init(&q->requeue_lock);
2098 if (q->nr_hw_queues > 1)
2099 blk_queue_make_request(q, blk_mq_make_request);
2101 blk_queue_make_request(q, blk_sq_make_request);
2104 * Do this after blk_queue_make_request() overrides it...
2106 q->nr_requests = set->queue_depth;
2108 if (set->ops->complete)
2109 blk_queue_softirq_done(q, set->ops->complete);
2111 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2114 mutex_lock(&all_q_mutex);
2116 list_add_tail(&q->all_q_node, &all_q_list);
2117 blk_mq_add_queue_tag_set(set, q);
2118 blk_mq_map_swqueue(q, cpu_online_mask);
2120 mutex_unlock(&all_q_mutex);
2128 kfree(q->queue_hw_ctx);
2130 free_percpu(q->queue_ctx);
2133 return ERR_PTR(-ENOMEM);
2135 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2137 void blk_mq_free_queue(struct request_queue *q)
2139 struct blk_mq_tag_set *set = q->tag_set;
2141 mutex_lock(&all_q_mutex);
2142 list_del_init(&q->all_q_node);
2143 mutex_unlock(&all_q_mutex);
2145 blk_mq_del_queue_tag_set(q);
2147 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2148 blk_mq_free_hw_queues(q, set);
2151 /* Basically redo blk_mq_init_queue with queue frozen */
2152 static void blk_mq_queue_reinit(struct request_queue *q,
2153 const struct cpumask *online_mask)
2155 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2157 blk_mq_sysfs_unregister(q);
2159 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2162 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2163 * we should change hctx numa_node according to new topology (this
2164 * involves free and re-allocate memory, worthy doing?)
2167 blk_mq_map_swqueue(q, online_mask);
2169 blk_mq_sysfs_register(q);
2172 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2173 unsigned long action, void *hcpu)
2175 struct request_queue *q;
2176 int cpu = (unsigned long)hcpu;
2178 * New online cpumask which is going to be set in this hotplug event.
2179 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2180 * one-by-one and dynamically allocating this could result in a failure.
2182 static struct cpumask online_new;
2185 * Before hotadded cpu starts handling requests, new mappings must
2186 * be established. Otherwise, these requests in hw queue might
2187 * never be dispatched.
2189 * For example, there is a single hw queue (hctx) and two CPU queues
2190 * (ctx0 for CPU0, and ctx1 for CPU1).
2192 * Now CPU1 is just onlined and a request is inserted into
2193 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2196 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2197 * set in pending bitmap and tries to retrieve requests in
2198 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2199 * so the request in ctx1->rq_list is ignored.
2201 switch (action & ~CPU_TASKS_FROZEN) {
2203 case CPU_UP_CANCELED:
2204 cpumask_copy(&online_new, cpu_online_mask);
2206 case CPU_UP_PREPARE:
2207 cpumask_copy(&online_new, cpu_online_mask);
2208 cpumask_set_cpu(cpu, &online_new);
2214 mutex_lock(&all_q_mutex);
2217 * We need to freeze and reinit all existing queues. Freezing
2218 * involves synchronous wait for an RCU grace period and doing it
2219 * one by one may take a long time. Start freezing all queues in
2220 * one swoop and then wait for the completions so that freezing can
2221 * take place in parallel.
2223 list_for_each_entry(q, &all_q_list, all_q_node)
2224 blk_mq_freeze_queue_start(q);
2225 list_for_each_entry(q, &all_q_list, all_q_node) {
2226 blk_mq_freeze_queue_wait(q);
2229 * timeout handler can't touch hw queue during the
2232 del_timer_sync(&q->timeout);
2235 list_for_each_entry(q, &all_q_list, all_q_node)
2236 blk_mq_queue_reinit(q, &online_new);
2238 list_for_each_entry(q, &all_q_list, all_q_node)
2239 blk_mq_unfreeze_queue(q);
2241 mutex_unlock(&all_q_mutex);
2245 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2249 for (i = 0; i < set->nr_hw_queues; i++) {
2250 set->tags[i] = blk_mq_init_rq_map(set, i);
2259 blk_mq_free_rq_map(set, set->tags[i], i);
2265 * Allocate the request maps associated with this tag_set. Note that this
2266 * may reduce the depth asked for, if memory is tight. set->queue_depth
2267 * will be updated to reflect the allocated depth.
2269 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2274 depth = set->queue_depth;
2276 err = __blk_mq_alloc_rq_maps(set);
2280 set->queue_depth >>= 1;
2281 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2285 } while (set->queue_depth);
2287 if (!set->queue_depth || err) {
2288 pr_err("blk-mq: failed to allocate request map\n");
2292 if (depth != set->queue_depth)
2293 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2294 depth, set->queue_depth);
2299 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2301 return tags->cpumask;
2303 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2306 * Alloc a tag set to be associated with one or more request queues.
2307 * May fail with EINVAL for various error conditions. May adjust the
2308 * requested depth down, if if it too large. In that case, the set
2309 * value will be stored in set->queue_depth.
2311 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2313 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2315 if (!set->nr_hw_queues)
2317 if (!set->queue_depth)
2319 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2322 if (!set->ops->queue_rq || !set->ops->map_queue)
2325 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2326 pr_info("blk-mq: reduced tag depth to %u\n",
2328 set->queue_depth = BLK_MQ_MAX_DEPTH;
2332 * If a crashdump is active, then we are potentially in a very
2333 * memory constrained environment. Limit us to 1 queue and
2334 * 64 tags to prevent using too much memory.
2336 if (is_kdump_kernel()) {
2337 set->nr_hw_queues = 1;
2338 set->queue_depth = min(64U, set->queue_depth);
2341 * There is no use for more h/w queues than cpus.
2343 if (set->nr_hw_queues > nr_cpu_ids)
2344 set->nr_hw_queues = nr_cpu_ids;
2346 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2347 GFP_KERNEL, set->numa_node);
2351 if (blk_mq_alloc_rq_maps(set))
2354 mutex_init(&set->tag_list_lock);
2355 INIT_LIST_HEAD(&set->tag_list);
2363 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2365 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2369 for (i = 0; i < nr_cpu_ids; i++) {
2371 blk_mq_free_rq_map(set, set->tags[i], i);
2377 EXPORT_SYMBOL(blk_mq_free_tag_set);
2379 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2381 struct blk_mq_tag_set *set = q->tag_set;
2382 struct blk_mq_hw_ctx *hctx;
2385 if (!set || nr > set->queue_depth)
2389 queue_for_each_hw_ctx(q, hctx, i) {
2392 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2398 q->nr_requests = nr;
2403 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2405 struct request_queue *q;
2407 if (nr_hw_queues > nr_cpu_ids)
2408 nr_hw_queues = nr_cpu_ids;
2409 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2412 list_for_each_entry(q, &set->tag_list, tag_set_list)
2413 blk_mq_freeze_queue(q);
2415 set->nr_hw_queues = nr_hw_queues;
2416 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2417 blk_mq_realloc_hw_ctxs(set, q);
2419 if (q->nr_hw_queues > 1)
2420 blk_queue_make_request(q, blk_mq_make_request);
2422 blk_queue_make_request(q, blk_sq_make_request);
2424 blk_mq_queue_reinit(q, cpu_online_mask);
2427 list_for_each_entry(q, &set->tag_list, tag_set_list)
2428 blk_mq_unfreeze_queue(q);
2430 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2432 void blk_mq_disable_hotplug(void)
2434 mutex_lock(&all_q_mutex);
2437 void blk_mq_enable_hotplug(void)
2439 mutex_unlock(&all_q_mutex);
2442 static int __init blk_mq_init(void)
2446 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2450 subsys_initcall(blk_mq_init);