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 EXPORT_SYMBOL(blk_mq_alloc_request);
270 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
271 struct blk_mq_ctx *ctx, struct request *rq)
273 const int tag = rq->tag;
274 struct request_queue *q = rq->q;
276 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
277 atomic_dec(&hctx->nr_active);
280 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
281 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
285 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
287 struct blk_mq_ctx *ctx = rq->mq_ctx;
289 ctx->rq_completed[rq_is_sync(rq)]++;
290 __blk_mq_free_request(hctx, ctx, rq);
293 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
295 void blk_mq_free_request(struct request *rq)
297 struct blk_mq_hw_ctx *hctx;
298 struct request_queue *q = rq->q;
300 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
301 blk_mq_free_hctx_request(hctx, rq);
303 EXPORT_SYMBOL_GPL(blk_mq_free_request);
305 inline void __blk_mq_end_request(struct request *rq, int error)
307 blk_account_io_done(rq);
310 rq->end_io(rq, error);
312 if (unlikely(blk_bidi_rq(rq)))
313 blk_mq_free_request(rq->next_rq);
314 blk_mq_free_request(rq);
317 EXPORT_SYMBOL(__blk_mq_end_request);
319 void blk_mq_end_request(struct request *rq, int error)
321 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
323 __blk_mq_end_request(rq, error);
325 EXPORT_SYMBOL(blk_mq_end_request);
327 static void __blk_mq_complete_request_remote(void *data)
329 struct request *rq = data;
331 rq->q->softirq_done_fn(rq);
334 static void blk_mq_ipi_complete_request(struct request *rq)
336 struct blk_mq_ctx *ctx = rq->mq_ctx;
340 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
341 rq->q->softirq_done_fn(rq);
346 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
347 shared = cpus_share_cache(cpu, ctx->cpu);
349 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
350 rq->csd.func = __blk_mq_complete_request_remote;
353 smp_call_function_single_async(ctx->cpu, &rq->csd);
355 rq->q->softirq_done_fn(rq);
360 static void __blk_mq_complete_request(struct request *rq)
362 struct request_queue *q = rq->q;
364 if (!q->softirq_done_fn)
365 blk_mq_end_request(rq, rq->errors);
367 blk_mq_ipi_complete_request(rq);
371 * blk_mq_complete_request - end I/O on a request
372 * @rq: the request being processed
375 * Ends all I/O on a request. It does not handle partial completions.
376 * The actual completion happens out-of-order, through a IPI handler.
378 void blk_mq_complete_request(struct request *rq, int error)
380 struct request_queue *q = rq->q;
382 if (unlikely(blk_should_fake_timeout(q)))
384 if (!blk_mark_rq_complete(rq)) {
386 __blk_mq_complete_request(rq);
389 EXPORT_SYMBOL(blk_mq_complete_request);
391 int blk_mq_request_started(struct request *rq)
393 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
395 EXPORT_SYMBOL_GPL(blk_mq_request_started);
397 void blk_mq_start_request(struct request *rq)
399 struct request_queue *q = rq->q;
401 trace_block_rq_issue(q, rq);
403 rq->resid_len = blk_rq_bytes(rq);
404 if (unlikely(blk_bidi_rq(rq)))
405 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
410 * Ensure that ->deadline is visible before set the started
411 * flag and clear the completed flag.
413 smp_mb__before_atomic();
416 * Mark us as started and clear complete. Complete might have been
417 * set if requeue raced with timeout, which then marked it as
418 * complete. So be sure to clear complete again when we start
419 * the request, otherwise we'll ignore the completion event.
421 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
422 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
423 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
424 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
426 if (q->dma_drain_size && blk_rq_bytes(rq)) {
428 * Make sure space for the drain appears. We know we can do
429 * this because max_hw_segments has been adjusted to be one
430 * fewer than the device can handle.
432 rq->nr_phys_segments++;
435 EXPORT_SYMBOL(blk_mq_start_request);
437 static void __blk_mq_requeue_request(struct request *rq)
439 struct request_queue *q = rq->q;
441 trace_block_rq_requeue(q, rq);
443 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
444 if (q->dma_drain_size && blk_rq_bytes(rq))
445 rq->nr_phys_segments--;
449 void blk_mq_requeue_request(struct request *rq)
451 __blk_mq_requeue_request(rq);
453 BUG_ON(blk_queued_rq(rq));
454 blk_mq_add_to_requeue_list(rq, true);
456 EXPORT_SYMBOL(blk_mq_requeue_request);
458 static void blk_mq_requeue_work(struct work_struct *work)
460 struct request_queue *q =
461 container_of(work, struct request_queue, requeue_work);
463 struct request *rq, *next;
466 spin_lock_irqsave(&q->requeue_lock, flags);
467 list_splice_init(&q->requeue_list, &rq_list);
468 spin_unlock_irqrestore(&q->requeue_lock, flags);
470 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
471 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
474 rq->cmd_flags &= ~REQ_SOFTBARRIER;
475 list_del_init(&rq->queuelist);
476 blk_mq_insert_request(rq, true, false, false);
479 while (!list_empty(&rq_list)) {
480 rq = list_entry(rq_list.next, struct request, queuelist);
481 list_del_init(&rq->queuelist);
482 blk_mq_insert_request(rq, false, false, false);
486 * Use the start variant of queue running here, so that running
487 * the requeue work will kick stopped queues.
489 blk_mq_start_hw_queues(q);
492 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
494 struct request_queue *q = rq->q;
498 * We abuse this flag that is otherwise used by the I/O scheduler to
499 * request head insertation from the workqueue.
501 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
503 spin_lock_irqsave(&q->requeue_lock, flags);
505 rq->cmd_flags |= REQ_SOFTBARRIER;
506 list_add(&rq->queuelist, &q->requeue_list);
508 list_add_tail(&rq->queuelist, &q->requeue_list);
510 spin_unlock_irqrestore(&q->requeue_lock, flags);
512 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
514 void blk_mq_cancel_requeue_work(struct request_queue *q)
516 cancel_work_sync(&q->requeue_work);
518 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
520 void blk_mq_kick_requeue_list(struct request_queue *q)
522 kblockd_schedule_work(&q->requeue_work);
524 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
526 void blk_mq_abort_requeue_list(struct request_queue *q)
531 spin_lock_irqsave(&q->requeue_lock, flags);
532 list_splice_init(&q->requeue_list, &rq_list);
533 spin_unlock_irqrestore(&q->requeue_lock, flags);
535 while (!list_empty(&rq_list)) {
538 rq = list_first_entry(&rq_list, struct request, queuelist);
539 list_del_init(&rq->queuelist);
541 blk_mq_end_request(rq, rq->errors);
544 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
546 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
548 if (tag < tags->nr_tags)
549 return tags->rqs[tag];
553 EXPORT_SYMBOL(blk_mq_tag_to_rq);
555 struct blk_mq_timeout_data {
557 unsigned int next_set;
560 void blk_mq_rq_timed_out(struct request *req, bool reserved)
562 struct blk_mq_ops *ops = req->q->mq_ops;
563 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
566 * We know that complete is set at this point. If STARTED isn't set
567 * anymore, then the request isn't active and the "timeout" should
568 * just be ignored. This can happen due to the bitflag ordering.
569 * Timeout first checks if STARTED is set, and if it is, assumes
570 * the request is active. But if we race with completion, then
571 * we both flags will get cleared. So check here again, and ignore
572 * a timeout event with a request that isn't active.
574 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
578 ret = ops->timeout(req, reserved);
582 __blk_mq_complete_request(req);
584 case BLK_EH_RESET_TIMER:
586 blk_clear_rq_complete(req);
588 case BLK_EH_NOT_HANDLED:
591 printk(KERN_ERR "block: bad eh return: %d\n", ret);
596 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
597 struct request *rq, void *priv, bool reserved)
599 struct blk_mq_timeout_data *data = priv;
601 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
603 * If a request wasn't started before the queue was
604 * marked dying, kill it here or it'll go unnoticed.
606 if (unlikely(blk_queue_dying(rq->q))) {
608 blk_mq_end_request(rq, rq->errors);
613 if (time_after_eq(jiffies, rq->deadline)) {
614 if (!blk_mark_rq_complete(rq))
615 blk_mq_rq_timed_out(rq, reserved);
616 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
617 data->next = rq->deadline;
622 static void blk_mq_timeout_work(struct work_struct *work)
624 struct request_queue *q =
625 container_of(work, struct request_queue, timeout_work);
626 struct blk_mq_timeout_data data = {
632 if (blk_queue_enter(q, true))
635 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
638 data.next = blk_rq_timeout(round_jiffies_up(data.next));
639 mod_timer(&q->timeout, data.next);
641 struct blk_mq_hw_ctx *hctx;
643 queue_for_each_hw_ctx(q, hctx, i) {
644 /* the hctx may be unmapped, so check it here */
645 if (blk_mq_hw_queue_mapped(hctx))
646 blk_mq_tag_idle(hctx);
653 * Reverse check our software queue for entries that we could potentially
654 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
655 * too much time checking for merges.
657 static bool blk_mq_attempt_merge(struct request_queue *q,
658 struct blk_mq_ctx *ctx, struct bio *bio)
663 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
669 if (!blk_rq_merge_ok(rq, bio))
672 el_ret = blk_try_merge(rq, bio);
673 if (el_ret == ELEVATOR_BACK_MERGE) {
674 if (bio_attempt_back_merge(q, rq, bio)) {
679 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
680 if (bio_attempt_front_merge(q, rq, bio)) {
692 * Process software queues that have been marked busy, splicing them
693 * to the for-dispatch
695 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
697 struct blk_mq_ctx *ctx;
700 for (i = 0; i < hctx->ctx_map.size; i++) {
701 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
702 unsigned int off, bit;
708 off = i * hctx->ctx_map.bits_per_word;
710 bit = find_next_bit(&bm->word, bm->depth, bit);
711 if (bit >= bm->depth)
714 ctx = hctx->ctxs[bit + off];
715 clear_bit(bit, &bm->word);
716 spin_lock(&ctx->lock);
717 list_splice_tail_init(&ctx->rq_list, list);
718 spin_unlock(&ctx->lock);
726 * Run this hardware queue, pulling any software queues mapped to it in.
727 * Note that this function currently has various problems around ordering
728 * of IO. In particular, we'd like FIFO behaviour on handling existing
729 * items on the hctx->dispatch list. Ignore that for now.
731 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
733 struct request_queue *q = hctx->queue;
736 LIST_HEAD(driver_list);
737 struct list_head *dptr;
740 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
742 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
748 * Touch any software queue that has pending entries.
750 flush_busy_ctxs(hctx, &rq_list);
753 * If we have previous entries on our dispatch list, grab them
754 * and stuff them at the front for more fair dispatch.
756 if (!list_empty_careful(&hctx->dispatch)) {
757 spin_lock(&hctx->lock);
758 if (!list_empty(&hctx->dispatch))
759 list_splice_init(&hctx->dispatch, &rq_list);
760 spin_unlock(&hctx->lock);
764 * Start off with dptr being NULL, so we start the first request
765 * immediately, even if we have more pending.
770 * Now process all the entries, sending them to the driver.
773 while (!list_empty(&rq_list)) {
774 struct blk_mq_queue_data bd;
777 rq = list_first_entry(&rq_list, struct request, queuelist);
778 list_del_init(&rq->queuelist);
782 bd.last = list_empty(&rq_list);
784 ret = q->mq_ops->queue_rq(hctx, &bd);
786 case BLK_MQ_RQ_QUEUE_OK:
789 case BLK_MQ_RQ_QUEUE_BUSY:
790 list_add(&rq->queuelist, &rq_list);
791 __blk_mq_requeue_request(rq);
794 pr_err("blk-mq: bad return on queue: %d\n", ret);
795 case BLK_MQ_RQ_QUEUE_ERROR:
797 blk_mq_end_request(rq, rq->errors);
801 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
805 * We've done the first request. If we have more than 1
806 * left in the list, set dptr to defer issue.
808 if (!dptr && rq_list.next != rq_list.prev)
813 hctx->dispatched[0]++;
814 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
815 hctx->dispatched[ilog2(queued) + 1]++;
818 * Any items that need requeuing? Stuff them into hctx->dispatch,
819 * that is where we will continue on next queue run.
821 if (!list_empty(&rq_list)) {
822 spin_lock(&hctx->lock);
823 list_splice(&rq_list, &hctx->dispatch);
824 spin_unlock(&hctx->lock);
826 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
827 * it's possible the queue is stopped and restarted again
828 * before this. Queue restart will dispatch requests. And since
829 * requests in rq_list aren't added into hctx->dispatch yet,
830 * the requests in rq_list might get lost.
832 * blk_mq_run_hw_queue() already checks the STOPPED bit
834 blk_mq_run_hw_queue(hctx, true);
839 * It'd be great if the workqueue API had a way to pass
840 * in a mask and had some smarts for more clever placement.
841 * For now we just round-robin here, switching for every
842 * BLK_MQ_CPU_WORK_BATCH queued items.
844 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
846 if (hctx->queue->nr_hw_queues == 1)
847 return WORK_CPU_UNBOUND;
849 if (--hctx->next_cpu_batch <= 0) {
850 int cpu = hctx->next_cpu, next_cpu;
852 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
853 if (next_cpu >= nr_cpu_ids)
854 next_cpu = cpumask_first(hctx->cpumask);
856 hctx->next_cpu = next_cpu;
857 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
862 return hctx->next_cpu;
865 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
867 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
868 !blk_mq_hw_queue_mapped(hctx)))
873 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
874 __blk_mq_run_hw_queue(hctx);
882 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
886 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
888 struct blk_mq_hw_ctx *hctx;
891 queue_for_each_hw_ctx(q, hctx, i) {
892 if ((!blk_mq_hctx_has_pending(hctx) &&
893 list_empty_careful(&hctx->dispatch)) ||
894 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
897 blk_mq_run_hw_queue(hctx, async);
900 EXPORT_SYMBOL(blk_mq_run_hw_queues);
902 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
904 cancel_delayed_work(&hctx->run_work);
905 cancel_delayed_work(&hctx->delay_work);
906 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
908 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
910 void blk_mq_stop_hw_queues(struct request_queue *q)
912 struct blk_mq_hw_ctx *hctx;
915 queue_for_each_hw_ctx(q, hctx, i)
916 blk_mq_stop_hw_queue(hctx);
918 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
920 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
922 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
924 blk_mq_run_hw_queue(hctx, false);
926 EXPORT_SYMBOL(blk_mq_start_hw_queue);
928 void blk_mq_start_hw_queues(struct request_queue *q)
930 struct blk_mq_hw_ctx *hctx;
933 queue_for_each_hw_ctx(q, hctx, i)
934 blk_mq_start_hw_queue(hctx);
936 EXPORT_SYMBOL(blk_mq_start_hw_queues);
938 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
940 struct blk_mq_hw_ctx *hctx;
943 queue_for_each_hw_ctx(q, hctx, i) {
944 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
947 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
948 blk_mq_run_hw_queue(hctx, async);
951 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
953 static void blk_mq_run_work_fn(struct work_struct *work)
955 struct blk_mq_hw_ctx *hctx;
957 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
959 __blk_mq_run_hw_queue(hctx);
962 static void blk_mq_delay_work_fn(struct work_struct *work)
964 struct blk_mq_hw_ctx *hctx;
966 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
968 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
969 __blk_mq_run_hw_queue(hctx);
972 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
974 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
977 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
978 &hctx->delay_work, msecs_to_jiffies(msecs));
980 EXPORT_SYMBOL(blk_mq_delay_queue);
982 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
983 struct blk_mq_ctx *ctx,
987 trace_block_rq_insert(hctx->queue, rq);
990 list_add(&rq->queuelist, &ctx->rq_list);
992 list_add_tail(&rq->queuelist, &ctx->rq_list);
995 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
996 struct request *rq, bool at_head)
998 struct blk_mq_ctx *ctx = rq->mq_ctx;
1000 __blk_mq_insert_req_list(hctx, ctx, rq, at_head);
1001 blk_mq_hctx_mark_pending(hctx, ctx);
1004 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1007 struct request_queue *q = rq->q;
1008 struct blk_mq_hw_ctx *hctx;
1009 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1011 current_ctx = blk_mq_get_ctx(q);
1012 if (!cpu_online(ctx->cpu))
1013 rq->mq_ctx = ctx = current_ctx;
1015 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1017 spin_lock(&ctx->lock);
1018 __blk_mq_insert_request(hctx, rq, at_head);
1019 spin_unlock(&ctx->lock);
1022 blk_mq_run_hw_queue(hctx, async);
1024 blk_mq_put_ctx(current_ctx);
1027 static void blk_mq_insert_requests(struct request_queue *q,
1028 struct blk_mq_ctx *ctx,
1029 struct list_head *list,
1034 struct blk_mq_hw_ctx *hctx;
1035 struct blk_mq_ctx *current_ctx;
1037 trace_block_unplug(q, depth, !from_schedule);
1039 current_ctx = blk_mq_get_ctx(q);
1041 if (!cpu_online(ctx->cpu))
1043 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1046 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1049 spin_lock(&ctx->lock);
1050 while (!list_empty(list)) {
1053 rq = list_first_entry(list, struct request, queuelist);
1054 list_del_init(&rq->queuelist);
1056 __blk_mq_insert_req_list(hctx, ctx, rq, false);
1058 blk_mq_hctx_mark_pending(hctx, ctx);
1059 spin_unlock(&ctx->lock);
1061 blk_mq_run_hw_queue(hctx, from_schedule);
1062 blk_mq_put_ctx(current_ctx);
1065 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1067 struct request *rqa = container_of(a, struct request, queuelist);
1068 struct request *rqb = container_of(b, struct request, queuelist);
1070 return !(rqa->mq_ctx < rqb->mq_ctx ||
1071 (rqa->mq_ctx == rqb->mq_ctx &&
1072 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1075 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1077 struct blk_mq_ctx *this_ctx;
1078 struct request_queue *this_q;
1081 LIST_HEAD(ctx_list);
1084 list_splice_init(&plug->mq_list, &list);
1086 list_sort(NULL, &list, plug_ctx_cmp);
1092 while (!list_empty(&list)) {
1093 rq = list_entry_rq(list.next);
1094 list_del_init(&rq->queuelist);
1096 if (rq->mq_ctx != this_ctx) {
1098 blk_mq_insert_requests(this_q, this_ctx,
1103 this_ctx = rq->mq_ctx;
1109 list_add_tail(&rq->queuelist, &ctx_list);
1113 * If 'this_ctx' is set, we know we have entries to complete
1114 * on 'ctx_list'. Do those.
1117 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1122 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1124 init_request_from_bio(rq, bio);
1126 blk_account_io_start(rq, 1);
1129 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1131 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1132 !blk_queue_nomerges(hctx->queue);
1135 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1136 struct blk_mq_ctx *ctx,
1137 struct request *rq, struct bio *bio)
1139 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1140 blk_mq_bio_to_request(rq, bio);
1141 spin_lock(&ctx->lock);
1143 __blk_mq_insert_request(hctx, rq, false);
1144 spin_unlock(&ctx->lock);
1147 struct request_queue *q = hctx->queue;
1149 spin_lock(&ctx->lock);
1150 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1151 blk_mq_bio_to_request(rq, bio);
1155 spin_unlock(&ctx->lock);
1156 __blk_mq_free_request(hctx, ctx, rq);
1161 struct blk_map_ctx {
1162 struct blk_mq_hw_ctx *hctx;
1163 struct blk_mq_ctx *ctx;
1166 static struct request *blk_mq_map_request(struct request_queue *q,
1168 struct blk_map_ctx *data)
1170 struct blk_mq_hw_ctx *hctx;
1171 struct blk_mq_ctx *ctx;
1173 int op = bio_data_dir(bio);
1175 struct blk_mq_alloc_data alloc_data;
1177 blk_queue_enter_live(q);
1178 ctx = blk_mq_get_ctx(q);
1179 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1181 if (rw_is_sync(bio->bi_rw))
1182 op_flags |= REQ_SYNC;
1184 trace_block_getrq(q, bio, op);
1185 blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx);
1186 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1187 if (unlikely(!rq)) {
1188 __blk_mq_run_hw_queue(hctx);
1189 blk_mq_put_ctx(ctx);
1190 trace_block_sleeprq(q, bio, op);
1192 ctx = blk_mq_get_ctx(q);
1193 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1194 blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1195 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1196 ctx = alloc_data.ctx;
1197 hctx = alloc_data.hctx;
1206 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1209 struct request_queue *q = rq->q;
1210 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1212 struct blk_mq_queue_data bd = {
1217 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1220 * For OK queue, we are done. For error, kill it. Any other
1221 * error (busy), just add it to our list as we previously
1224 ret = q->mq_ops->queue_rq(hctx, &bd);
1225 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1226 *cookie = new_cookie;
1230 __blk_mq_requeue_request(rq);
1232 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1233 *cookie = BLK_QC_T_NONE;
1235 blk_mq_end_request(rq, rq->errors);
1243 * Multiple hardware queue variant. This will not use per-process plugs,
1244 * but will attempt to bypass the hctx queueing if we can go straight to
1245 * hardware for SYNC IO.
1247 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1249 const int is_sync = rw_is_sync(bio->bi_rw);
1250 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1251 struct blk_map_ctx data;
1253 unsigned int request_count = 0;
1254 struct blk_plug *plug;
1255 struct request *same_queue_rq = NULL;
1258 blk_queue_bounce(q, &bio);
1260 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1262 return BLK_QC_T_NONE;
1265 blk_queue_split(q, &bio, q->bio_split);
1267 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1268 if (blk_attempt_plug_merge(q, bio, &request_count,
1270 return BLK_QC_T_NONE;
1272 request_count = blk_plug_queued_count(q);
1274 rq = blk_mq_map_request(q, bio, &data);
1276 return BLK_QC_T_NONE;
1278 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1280 if (unlikely(is_flush_fua)) {
1281 blk_mq_bio_to_request(rq, bio);
1282 blk_insert_flush(rq);
1286 plug = current->plug;
1288 * If the driver supports defer issued based on 'last', then
1289 * queue it up like normal since we can potentially save some
1292 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1293 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1294 struct request *old_rq = NULL;
1296 blk_mq_bio_to_request(rq, bio);
1299 * We do limited pluging. If the bio can be merged, do that.
1300 * Otherwise the existing request in the plug list will be
1301 * issued. So the plug list will have one request at most
1305 * The plug list might get flushed before this. If that
1306 * happens, same_queue_rq is invalid and plug list is
1309 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1310 old_rq = same_queue_rq;
1311 list_del_init(&old_rq->queuelist);
1313 list_add_tail(&rq->queuelist, &plug->mq_list);
1314 } else /* is_sync */
1316 blk_mq_put_ctx(data.ctx);
1319 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1321 blk_mq_insert_request(old_rq, false, true, true);
1325 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1327 * For a SYNC request, send it to the hardware immediately. For
1328 * an ASYNC request, just ensure that we run it later on. The
1329 * latter allows for merging opportunities and more efficient
1333 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1335 blk_mq_put_ctx(data.ctx);
1341 * Single hardware queue variant. This will attempt to use any per-process
1342 * plug for merging and IO deferral.
1344 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1346 const int is_sync = rw_is_sync(bio->bi_rw);
1347 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1348 struct blk_plug *plug;
1349 unsigned int request_count = 0;
1350 struct blk_map_ctx data;
1354 blk_queue_bounce(q, &bio);
1356 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1358 return BLK_QC_T_NONE;
1361 blk_queue_split(q, &bio, q->bio_split);
1363 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1364 blk_attempt_plug_merge(q, bio, &request_count, NULL))
1365 return BLK_QC_T_NONE;
1367 rq = blk_mq_map_request(q, bio, &data);
1369 return BLK_QC_T_NONE;
1371 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1373 if (unlikely(is_flush_fua)) {
1374 blk_mq_bio_to_request(rq, bio);
1375 blk_insert_flush(rq);
1380 * A task plug currently exists. Since this is completely lockless,
1381 * utilize that to temporarily store requests until the task is
1382 * either done or scheduled away.
1384 plug = current->plug;
1386 blk_mq_bio_to_request(rq, bio);
1388 trace_block_plug(q);
1390 blk_mq_put_ctx(data.ctx);
1392 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1393 blk_flush_plug_list(plug, false);
1394 trace_block_plug(q);
1397 list_add_tail(&rq->queuelist, &plug->mq_list);
1401 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1403 * For a SYNC request, send it to the hardware immediately. For
1404 * an ASYNC request, just ensure that we run it later on. The
1405 * latter allows for merging opportunities and more efficient
1409 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1412 blk_mq_put_ctx(data.ctx);
1417 * Default mapping to a software queue, since we use one per CPU.
1419 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1421 return q->queue_hw_ctx[q->mq_map[cpu]];
1423 EXPORT_SYMBOL(blk_mq_map_queue);
1425 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1426 struct blk_mq_tags *tags, unsigned int hctx_idx)
1430 if (tags->rqs && set->ops->exit_request) {
1433 for (i = 0; i < tags->nr_tags; i++) {
1436 set->ops->exit_request(set->driver_data, tags->rqs[i],
1438 tags->rqs[i] = NULL;
1442 while (!list_empty(&tags->page_list)) {
1443 page = list_first_entry(&tags->page_list, struct page, lru);
1444 list_del_init(&page->lru);
1446 * Remove kmemleak object previously allocated in
1447 * blk_mq_init_rq_map().
1449 kmemleak_free(page_address(page));
1450 __free_pages(page, page->private);
1455 blk_mq_free_tags(tags);
1458 static size_t order_to_size(unsigned int order)
1460 return (size_t)PAGE_SIZE << order;
1463 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1464 unsigned int hctx_idx)
1466 struct blk_mq_tags *tags;
1467 unsigned int i, j, entries_per_page, max_order = 4;
1468 size_t rq_size, left;
1470 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1472 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1476 INIT_LIST_HEAD(&tags->page_list);
1478 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1479 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1482 blk_mq_free_tags(tags);
1487 * rq_size is the size of the request plus driver payload, rounded
1488 * to the cacheline size
1490 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1492 left = rq_size * set->queue_depth;
1494 for (i = 0; i < set->queue_depth; ) {
1495 int this_order = max_order;
1500 while (this_order && left < order_to_size(this_order - 1))
1504 page = alloc_pages_node(set->numa_node,
1505 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1511 if (order_to_size(this_order) < rq_size)
1518 page->private = this_order;
1519 list_add_tail(&page->lru, &tags->page_list);
1521 p = page_address(page);
1523 * Allow kmemleak to scan these pages as they contain pointers
1524 * to additional allocations like via ops->init_request().
1526 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1527 entries_per_page = order_to_size(this_order) / rq_size;
1528 to_do = min(entries_per_page, set->queue_depth - i);
1529 left -= to_do * rq_size;
1530 for (j = 0; j < to_do; j++) {
1532 if (set->ops->init_request) {
1533 if (set->ops->init_request(set->driver_data,
1534 tags->rqs[i], hctx_idx, i,
1536 tags->rqs[i] = NULL;
1548 blk_mq_free_rq_map(set, tags, hctx_idx);
1552 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1557 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1559 unsigned int bpw = 8, total, num_maps, i;
1561 bitmap->bits_per_word = bpw;
1563 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1564 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1570 for (i = 0; i < num_maps; i++) {
1571 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1572 total -= bitmap->map[i].depth;
1578 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1580 struct request_queue *q = hctx->queue;
1581 struct blk_mq_ctx *ctx;
1585 * Move ctx entries to new CPU, if this one is going away.
1587 ctx = __blk_mq_get_ctx(q, cpu);
1589 spin_lock(&ctx->lock);
1590 if (!list_empty(&ctx->rq_list)) {
1591 list_splice_init(&ctx->rq_list, &tmp);
1592 blk_mq_hctx_clear_pending(hctx, ctx);
1594 spin_unlock(&ctx->lock);
1596 if (list_empty(&tmp))
1599 ctx = blk_mq_get_ctx(q);
1600 spin_lock(&ctx->lock);
1602 while (!list_empty(&tmp)) {
1605 rq = list_first_entry(&tmp, struct request, queuelist);
1607 list_move_tail(&rq->queuelist, &ctx->rq_list);
1610 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1611 blk_mq_hctx_mark_pending(hctx, ctx);
1613 spin_unlock(&ctx->lock);
1615 blk_mq_run_hw_queue(hctx, true);
1616 blk_mq_put_ctx(ctx);
1620 static int blk_mq_hctx_notify(void *data, unsigned long action,
1623 struct blk_mq_hw_ctx *hctx = data;
1625 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1626 return blk_mq_hctx_cpu_offline(hctx, cpu);
1629 * In case of CPU online, tags may be reallocated
1630 * in blk_mq_map_swqueue() after mapping is updated.
1636 /* hctx->ctxs will be freed in queue's release handler */
1637 static void blk_mq_exit_hctx(struct request_queue *q,
1638 struct blk_mq_tag_set *set,
1639 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1641 unsigned flush_start_tag = set->queue_depth;
1643 blk_mq_tag_idle(hctx);
1645 if (set->ops->exit_request)
1646 set->ops->exit_request(set->driver_data,
1647 hctx->fq->flush_rq, hctx_idx,
1648 flush_start_tag + hctx_idx);
1650 if (set->ops->exit_hctx)
1651 set->ops->exit_hctx(hctx, hctx_idx);
1653 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1654 blk_free_flush_queue(hctx->fq);
1655 blk_mq_free_bitmap(&hctx->ctx_map);
1658 static void blk_mq_exit_hw_queues(struct request_queue *q,
1659 struct blk_mq_tag_set *set, int nr_queue)
1661 struct blk_mq_hw_ctx *hctx;
1664 queue_for_each_hw_ctx(q, hctx, i) {
1667 blk_mq_exit_hctx(q, set, hctx, i);
1671 static void blk_mq_free_hw_queues(struct request_queue *q,
1672 struct blk_mq_tag_set *set)
1674 struct blk_mq_hw_ctx *hctx;
1677 queue_for_each_hw_ctx(q, hctx, i)
1678 free_cpumask_var(hctx->cpumask);
1681 static int blk_mq_init_hctx(struct request_queue *q,
1682 struct blk_mq_tag_set *set,
1683 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1686 unsigned flush_start_tag = set->queue_depth;
1688 node = hctx->numa_node;
1689 if (node == NUMA_NO_NODE)
1690 node = hctx->numa_node = set->numa_node;
1692 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1693 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1694 spin_lock_init(&hctx->lock);
1695 INIT_LIST_HEAD(&hctx->dispatch);
1697 hctx->queue_num = hctx_idx;
1698 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1700 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1701 blk_mq_hctx_notify, hctx);
1702 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1704 hctx->tags = set->tags[hctx_idx];
1707 * Allocate space for all possible cpus to avoid allocation at
1710 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1713 goto unregister_cpu_notifier;
1715 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1720 if (set->ops->init_hctx &&
1721 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1724 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1728 if (set->ops->init_request &&
1729 set->ops->init_request(set->driver_data,
1730 hctx->fq->flush_rq, hctx_idx,
1731 flush_start_tag + hctx_idx, node))
1739 if (set->ops->exit_hctx)
1740 set->ops->exit_hctx(hctx, hctx_idx);
1742 blk_mq_free_bitmap(&hctx->ctx_map);
1745 unregister_cpu_notifier:
1746 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1751 static void blk_mq_init_cpu_queues(struct request_queue *q,
1752 unsigned int nr_hw_queues)
1756 for_each_possible_cpu(i) {
1757 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1758 struct blk_mq_hw_ctx *hctx;
1760 memset(__ctx, 0, sizeof(*__ctx));
1762 spin_lock_init(&__ctx->lock);
1763 INIT_LIST_HEAD(&__ctx->rq_list);
1766 /* If the cpu isn't online, the cpu is mapped to first hctx */
1770 hctx = q->mq_ops->map_queue(q, i);
1773 * Set local node, IFF we have more than one hw queue. If
1774 * not, we remain on the home node of the device
1776 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1777 hctx->numa_node = local_memory_node(cpu_to_node(i));
1781 static void blk_mq_map_swqueue(struct request_queue *q,
1782 const struct cpumask *online_mask)
1785 struct blk_mq_hw_ctx *hctx;
1786 struct blk_mq_ctx *ctx;
1787 struct blk_mq_tag_set *set = q->tag_set;
1790 * Avoid others reading imcomplete hctx->cpumask through sysfs
1792 mutex_lock(&q->sysfs_lock);
1794 queue_for_each_hw_ctx(q, hctx, i) {
1795 cpumask_clear(hctx->cpumask);
1800 * Map software to hardware queues
1802 for_each_possible_cpu(i) {
1803 /* If the cpu isn't online, the cpu is mapped to first hctx */
1804 if (!cpumask_test_cpu(i, online_mask))
1807 ctx = per_cpu_ptr(q->queue_ctx, i);
1808 hctx = q->mq_ops->map_queue(q, i);
1810 cpumask_set_cpu(i, hctx->cpumask);
1811 ctx->index_hw = hctx->nr_ctx;
1812 hctx->ctxs[hctx->nr_ctx++] = ctx;
1815 mutex_unlock(&q->sysfs_lock);
1817 queue_for_each_hw_ctx(q, hctx, i) {
1818 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1821 * If no software queues are mapped to this hardware queue,
1822 * disable it and free the request entries.
1824 if (!hctx->nr_ctx) {
1826 blk_mq_free_rq_map(set, set->tags[i], i);
1827 set->tags[i] = NULL;
1833 /* unmapped hw queue can be remapped after CPU topo changed */
1835 set->tags[i] = blk_mq_init_rq_map(set, i);
1836 hctx->tags = set->tags[i];
1837 WARN_ON(!hctx->tags);
1839 cpumask_copy(hctx->tags->cpumask, hctx->cpumask);
1841 * Set the map size to the number of mapped software queues.
1842 * This is more accurate and more efficient than looping
1843 * over all possibly mapped software queues.
1845 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1848 * Initialize batch roundrobin counts
1850 hctx->next_cpu = cpumask_first(hctx->cpumask);
1851 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1855 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1857 struct blk_mq_hw_ctx *hctx;
1860 queue_for_each_hw_ctx(q, hctx, i) {
1862 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1864 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1868 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1870 struct request_queue *q;
1872 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1873 blk_mq_freeze_queue(q);
1874 queue_set_hctx_shared(q, shared);
1875 blk_mq_unfreeze_queue(q);
1879 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1881 struct blk_mq_tag_set *set = q->tag_set;
1883 mutex_lock(&set->tag_list_lock);
1884 list_del_init(&q->tag_set_list);
1885 if (list_is_singular(&set->tag_list)) {
1886 /* just transitioned to unshared */
1887 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1888 /* update existing queue */
1889 blk_mq_update_tag_set_depth(set, false);
1891 mutex_unlock(&set->tag_list_lock);
1894 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1895 struct request_queue *q)
1899 mutex_lock(&set->tag_list_lock);
1901 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1902 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1903 set->flags |= BLK_MQ_F_TAG_SHARED;
1904 /* update existing queue */
1905 blk_mq_update_tag_set_depth(set, true);
1907 if (set->flags & BLK_MQ_F_TAG_SHARED)
1908 queue_set_hctx_shared(q, true);
1909 list_add_tail(&q->tag_set_list, &set->tag_list);
1911 mutex_unlock(&set->tag_list_lock);
1915 * It is the actual release handler for mq, but we do it from
1916 * request queue's release handler for avoiding use-after-free
1917 * and headache because q->mq_kobj shouldn't have been introduced,
1918 * but we can't group ctx/kctx kobj without it.
1920 void blk_mq_release(struct request_queue *q)
1922 struct blk_mq_hw_ctx *hctx;
1925 /* hctx kobj stays in hctx */
1926 queue_for_each_hw_ctx(q, hctx, i) {
1936 kfree(q->queue_hw_ctx);
1938 /* ctx kobj stays in queue_ctx */
1939 free_percpu(q->queue_ctx);
1942 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1944 struct request_queue *uninit_q, *q;
1946 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1948 return ERR_PTR(-ENOMEM);
1950 q = blk_mq_init_allocated_queue(set, uninit_q);
1952 blk_cleanup_queue(uninit_q);
1956 EXPORT_SYMBOL(blk_mq_init_queue);
1958 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1959 struct request_queue *q)
1962 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
1964 blk_mq_sysfs_unregister(q);
1965 for (i = 0; i < set->nr_hw_queues; i++) {
1971 node = blk_mq_hw_queue_to_node(q->mq_map, i);
1972 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1977 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1984 atomic_set(&hctxs[i]->nr_active, 0);
1985 hctxs[i]->numa_node = node;
1986 hctxs[i]->queue_num = i;
1988 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
1989 free_cpumask_var(hctxs[i]->cpumask);
1994 blk_mq_hctx_kobj_init(hctxs[i]);
1996 for (j = i; j < q->nr_hw_queues; j++) {
1997 struct blk_mq_hw_ctx *hctx = hctxs[j];
2001 blk_mq_free_rq_map(set, hctx->tags, j);
2002 set->tags[j] = NULL;
2004 blk_mq_exit_hctx(q, set, hctx, j);
2005 free_cpumask_var(hctx->cpumask);
2006 kobject_put(&hctx->kobj);
2013 q->nr_hw_queues = i;
2014 blk_mq_sysfs_register(q);
2017 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2018 struct request_queue *q)
2020 /* mark the queue as mq asap */
2021 q->mq_ops = set->ops;
2023 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2027 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2028 GFP_KERNEL, set->numa_node);
2029 if (!q->queue_hw_ctx)
2032 q->mq_map = blk_mq_make_queue_map(set);
2036 blk_mq_realloc_hw_ctxs(set, q);
2037 if (!q->nr_hw_queues)
2040 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2041 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2043 q->nr_queues = nr_cpu_ids;
2045 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2047 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2048 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2050 q->sg_reserved_size = INT_MAX;
2052 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2053 INIT_LIST_HEAD(&q->requeue_list);
2054 spin_lock_init(&q->requeue_lock);
2056 if (q->nr_hw_queues > 1)
2057 blk_queue_make_request(q, blk_mq_make_request);
2059 blk_queue_make_request(q, blk_sq_make_request);
2062 * Do this after blk_queue_make_request() overrides it...
2064 q->nr_requests = set->queue_depth;
2066 if (set->ops->complete)
2067 blk_queue_softirq_done(q, set->ops->complete);
2069 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2072 mutex_lock(&all_q_mutex);
2074 list_add_tail(&q->all_q_node, &all_q_list);
2075 blk_mq_add_queue_tag_set(set, q);
2076 blk_mq_map_swqueue(q, cpu_online_mask);
2078 mutex_unlock(&all_q_mutex);
2086 kfree(q->queue_hw_ctx);
2088 free_percpu(q->queue_ctx);
2091 return ERR_PTR(-ENOMEM);
2093 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2095 void blk_mq_free_queue(struct request_queue *q)
2097 struct blk_mq_tag_set *set = q->tag_set;
2099 mutex_lock(&all_q_mutex);
2100 list_del_init(&q->all_q_node);
2101 mutex_unlock(&all_q_mutex);
2103 blk_mq_del_queue_tag_set(q);
2105 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2106 blk_mq_free_hw_queues(q, set);
2109 /* Basically redo blk_mq_init_queue with queue frozen */
2110 static void blk_mq_queue_reinit(struct request_queue *q,
2111 const struct cpumask *online_mask)
2113 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2115 blk_mq_sysfs_unregister(q);
2117 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2120 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2121 * we should change hctx numa_node according to new topology (this
2122 * involves free and re-allocate memory, worthy doing?)
2125 blk_mq_map_swqueue(q, online_mask);
2127 blk_mq_sysfs_register(q);
2130 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2131 unsigned long action, void *hcpu)
2133 struct request_queue *q;
2134 int cpu = (unsigned long)hcpu;
2136 * New online cpumask which is going to be set in this hotplug event.
2137 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2138 * one-by-one and dynamically allocating this could result in a failure.
2140 static struct cpumask online_new;
2143 * Before hotadded cpu starts handling requests, new mappings must
2144 * be established. Otherwise, these requests in hw queue might
2145 * never be dispatched.
2147 * For example, there is a single hw queue (hctx) and two CPU queues
2148 * (ctx0 for CPU0, and ctx1 for CPU1).
2150 * Now CPU1 is just onlined and a request is inserted into
2151 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2154 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2155 * set in pending bitmap and tries to retrieve requests in
2156 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2157 * so the request in ctx1->rq_list is ignored.
2159 switch (action & ~CPU_TASKS_FROZEN) {
2161 case CPU_UP_CANCELED:
2162 cpumask_copy(&online_new, cpu_online_mask);
2164 case CPU_UP_PREPARE:
2165 cpumask_copy(&online_new, cpu_online_mask);
2166 cpumask_set_cpu(cpu, &online_new);
2172 mutex_lock(&all_q_mutex);
2175 * We need to freeze and reinit all existing queues. Freezing
2176 * involves synchronous wait for an RCU grace period and doing it
2177 * one by one may take a long time. Start freezing all queues in
2178 * one swoop and then wait for the completions so that freezing can
2179 * take place in parallel.
2181 list_for_each_entry(q, &all_q_list, all_q_node)
2182 blk_mq_freeze_queue_start(q);
2183 list_for_each_entry(q, &all_q_list, all_q_node) {
2184 blk_mq_freeze_queue_wait(q);
2187 * timeout handler can't touch hw queue during the
2190 del_timer_sync(&q->timeout);
2193 list_for_each_entry(q, &all_q_list, all_q_node)
2194 blk_mq_queue_reinit(q, &online_new);
2196 list_for_each_entry(q, &all_q_list, all_q_node)
2197 blk_mq_unfreeze_queue(q);
2199 mutex_unlock(&all_q_mutex);
2203 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2207 for (i = 0; i < set->nr_hw_queues; i++) {
2208 set->tags[i] = blk_mq_init_rq_map(set, i);
2217 blk_mq_free_rq_map(set, set->tags[i], i);
2223 * Allocate the request maps associated with this tag_set. Note that this
2224 * may reduce the depth asked for, if memory is tight. set->queue_depth
2225 * will be updated to reflect the allocated depth.
2227 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2232 depth = set->queue_depth;
2234 err = __blk_mq_alloc_rq_maps(set);
2238 set->queue_depth >>= 1;
2239 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2243 } while (set->queue_depth);
2245 if (!set->queue_depth || err) {
2246 pr_err("blk-mq: failed to allocate request map\n");
2250 if (depth != set->queue_depth)
2251 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2252 depth, set->queue_depth);
2257 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2259 return tags->cpumask;
2261 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2264 * Alloc a tag set to be associated with one or more request queues.
2265 * May fail with EINVAL for various error conditions. May adjust the
2266 * requested depth down, if if it too large. In that case, the set
2267 * value will be stored in set->queue_depth.
2269 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2271 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2273 if (!set->nr_hw_queues)
2275 if (!set->queue_depth)
2277 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2280 if (!set->ops->queue_rq || !set->ops->map_queue)
2283 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2284 pr_info("blk-mq: reduced tag depth to %u\n",
2286 set->queue_depth = BLK_MQ_MAX_DEPTH;
2290 * If a crashdump is active, then we are potentially in a very
2291 * memory constrained environment. Limit us to 1 queue and
2292 * 64 tags to prevent using too much memory.
2294 if (is_kdump_kernel()) {
2295 set->nr_hw_queues = 1;
2296 set->queue_depth = min(64U, set->queue_depth);
2299 * There is no use for more h/w queues than cpus.
2301 if (set->nr_hw_queues > nr_cpu_ids)
2302 set->nr_hw_queues = nr_cpu_ids;
2304 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2305 GFP_KERNEL, set->numa_node);
2309 if (blk_mq_alloc_rq_maps(set))
2312 mutex_init(&set->tag_list_lock);
2313 INIT_LIST_HEAD(&set->tag_list);
2321 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2323 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2327 for (i = 0; i < nr_cpu_ids; i++) {
2329 blk_mq_free_rq_map(set, set->tags[i], i);
2335 EXPORT_SYMBOL(blk_mq_free_tag_set);
2337 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2339 struct blk_mq_tag_set *set = q->tag_set;
2340 struct blk_mq_hw_ctx *hctx;
2343 if (!set || nr > set->queue_depth)
2347 queue_for_each_hw_ctx(q, hctx, i) {
2350 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2356 q->nr_requests = nr;
2361 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2363 struct request_queue *q;
2365 if (nr_hw_queues > nr_cpu_ids)
2366 nr_hw_queues = nr_cpu_ids;
2367 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2370 list_for_each_entry(q, &set->tag_list, tag_set_list)
2371 blk_mq_freeze_queue(q);
2373 set->nr_hw_queues = nr_hw_queues;
2374 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2375 blk_mq_realloc_hw_ctxs(set, q);
2377 if (q->nr_hw_queues > 1)
2378 blk_queue_make_request(q, blk_mq_make_request);
2380 blk_queue_make_request(q, blk_sq_make_request);
2382 blk_mq_queue_reinit(q, cpu_online_mask);
2385 list_for_each_entry(q, &set->tag_list, tag_set_list)
2386 blk_mq_unfreeze_queue(q);
2388 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2390 void blk_mq_disable_hotplug(void)
2392 mutex_lock(&all_q_mutex);
2395 void blk_mq_enable_hotplug(void)
2397 mutex_unlock(&all_q_mutex);
2400 static int __init blk_mq_init(void)
2404 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2408 subsys_initcall(blk_mq_init);