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, unsigned int rw_flags)
164 if (blk_queue_io_stat(q))
165 rw_flags |= REQ_IO_STAT;
167 INIT_LIST_HEAD(&rq->queuelist);
168 /* csd/requeue_work/fifo_time is initialized before use */
171 rq->cmd_flags |= rw_flags;
172 /* do not touch atomic flags, it needs atomic ops against the timer */
174 INIT_HLIST_NODE(&rq->hash);
175 RB_CLEAR_NODE(&rq->rb_node);
178 rq->start_time = jiffies;
179 #ifdef CONFIG_BLK_CGROUP
181 set_start_time_ns(rq);
182 rq->io_start_time_ns = 0;
184 rq->nr_phys_segments = 0;
185 #if defined(CONFIG_BLK_DEV_INTEGRITY)
186 rq->nr_integrity_segments = 0;
189 /* tag was already set */
199 INIT_LIST_HEAD(&rq->timeout_list);
203 rq->end_io_data = NULL;
206 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
209 static struct request *
210 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
215 tag = blk_mq_get_tag(data);
216 if (tag != BLK_MQ_TAG_FAIL) {
217 rq = data->hctx->tags->rqs[tag];
219 if (blk_mq_tag_busy(data->hctx)) {
220 rq->cmd_flags = REQ_MQ_INFLIGHT;
221 atomic_inc(&data->hctx->nr_active);
225 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
232 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
235 struct blk_mq_ctx *ctx;
236 struct blk_mq_hw_ctx *hctx;
238 struct blk_mq_alloc_data alloc_data;
241 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
245 ctx = blk_mq_get_ctx(q);
246 hctx = q->mq_ops->map_queue(q, ctx->cpu);
247 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
249 rq = __blk_mq_alloc_request(&alloc_data, rw);
250 if (!rq && !(flags & BLK_MQ_REQ_NOWAIT)) {
251 __blk_mq_run_hw_queue(hctx);
254 ctx = blk_mq_get_ctx(q);
255 hctx = q->mq_ops->map_queue(q, ctx->cpu);
256 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
257 rq = __blk_mq_alloc_request(&alloc_data, rw);
258 ctx = alloc_data.ctx;
263 return ERR_PTR(-EWOULDBLOCK);
267 EXPORT_SYMBOL(blk_mq_alloc_request);
269 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
270 struct blk_mq_ctx *ctx, struct request *rq)
272 const int tag = rq->tag;
273 struct request_queue *q = rq->q;
275 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
276 atomic_dec(&hctx->nr_active);
279 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
280 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
284 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
286 struct blk_mq_ctx *ctx = rq->mq_ctx;
288 ctx->rq_completed[rq_is_sync(rq)]++;
289 __blk_mq_free_request(hctx, ctx, rq);
292 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
294 void blk_mq_free_request(struct request *rq)
296 struct blk_mq_hw_ctx *hctx;
297 struct request_queue *q = rq->q;
299 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
300 blk_mq_free_hctx_request(hctx, rq);
302 EXPORT_SYMBOL_GPL(blk_mq_free_request);
304 inline void __blk_mq_end_request(struct request *rq, int error)
306 blk_account_io_done(rq);
309 rq->end_io(rq, error);
311 if (unlikely(blk_bidi_rq(rq)))
312 blk_mq_free_request(rq->next_rq);
313 blk_mq_free_request(rq);
316 EXPORT_SYMBOL(__blk_mq_end_request);
318 void blk_mq_end_request(struct request *rq, int error)
320 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
322 __blk_mq_end_request(rq, error);
324 EXPORT_SYMBOL(blk_mq_end_request);
326 static void __blk_mq_complete_request_remote(void *data)
328 struct request *rq = data;
330 rq->q->softirq_done_fn(rq);
333 static void blk_mq_ipi_complete_request(struct request *rq)
335 struct blk_mq_ctx *ctx = rq->mq_ctx;
339 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
340 rq->q->softirq_done_fn(rq);
345 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
346 shared = cpus_share_cache(cpu, ctx->cpu);
348 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
349 rq->csd.func = __blk_mq_complete_request_remote;
352 smp_call_function_single_async(ctx->cpu, &rq->csd);
354 rq->q->softirq_done_fn(rq);
359 static void __blk_mq_complete_request(struct request *rq)
361 struct request_queue *q = rq->q;
363 if (!q->softirq_done_fn)
364 blk_mq_end_request(rq, rq->errors);
366 blk_mq_ipi_complete_request(rq);
370 * blk_mq_complete_request - end I/O on a request
371 * @rq: the request being processed
374 * Ends all I/O on a request. It does not handle partial completions.
375 * The actual completion happens out-of-order, through a IPI handler.
377 void blk_mq_complete_request(struct request *rq, int error)
379 struct request_queue *q = rq->q;
381 if (unlikely(blk_should_fake_timeout(q)))
383 if (!blk_mark_rq_complete(rq)) {
385 __blk_mq_complete_request(rq);
388 EXPORT_SYMBOL(blk_mq_complete_request);
390 int blk_mq_request_started(struct request *rq)
392 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
394 EXPORT_SYMBOL_GPL(blk_mq_request_started);
396 void blk_mq_start_request(struct request *rq)
398 struct request_queue *q = rq->q;
400 trace_block_rq_issue(q, rq);
402 rq->resid_len = blk_rq_bytes(rq);
403 if (unlikely(blk_bidi_rq(rq)))
404 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
409 * Ensure that ->deadline is visible before set the started
410 * flag and clear the completed flag.
412 smp_mb__before_atomic();
415 * Mark us as started and clear complete. Complete might have been
416 * set if requeue raced with timeout, which then marked it as
417 * complete. So be sure to clear complete again when we start
418 * the request, otherwise we'll ignore the completion event.
420 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
421 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
422 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
423 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
425 if (q->dma_drain_size && blk_rq_bytes(rq)) {
427 * Make sure space for the drain appears. We know we can do
428 * this because max_hw_segments has been adjusted to be one
429 * fewer than the device can handle.
431 rq->nr_phys_segments++;
434 EXPORT_SYMBOL(blk_mq_start_request);
436 static void __blk_mq_requeue_request(struct request *rq)
438 struct request_queue *q = rq->q;
440 trace_block_rq_requeue(q, rq);
442 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
443 if (q->dma_drain_size && blk_rq_bytes(rq))
444 rq->nr_phys_segments--;
448 void blk_mq_requeue_request(struct request *rq)
450 __blk_mq_requeue_request(rq);
452 BUG_ON(blk_queued_rq(rq));
453 blk_mq_add_to_requeue_list(rq, true);
455 EXPORT_SYMBOL(blk_mq_requeue_request);
457 static void blk_mq_requeue_work(struct work_struct *work)
459 struct request_queue *q =
460 container_of(work, struct request_queue, requeue_work);
462 struct request *rq, *next;
465 spin_lock_irqsave(&q->requeue_lock, flags);
466 list_splice_init(&q->requeue_list, &rq_list);
467 spin_unlock_irqrestore(&q->requeue_lock, flags);
469 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
470 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
473 rq->cmd_flags &= ~REQ_SOFTBARRIER;
474 list_del_init(&rq->queuelist);
475 blk_mq_insert_request(rq, true, false, false);
478 while (!list_empty(&rq_list)) {
479 rq = list_entry(rq_list.next, struct request, queuelist);
480 list_del_init(&rq->queuelist);
481 blk_mq_insert_request(rq, false, false, false);
485 * Use the start variant of queue running here, so that running
486 * the requeue work will kick stopped queues.
488 blk_mq_start_hw_queues(q);
491 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
493 struct request_queue *q = rq->q;
497 * We abuse this flag that is otherwise used by the I/O scheduler to
498 * request head insertation from the workqueue.
500 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
502 spin_lock_irqsave(&q->requeue_lock, flags);
504 rq->cmd_flags |= REQ_SOFTBARRIER;
505 list_add(&rq->queuelist, &q->requeue_list);
507 list_add_tail(&rq->queuelist, &q->requeue_list);
509 spin_unlock_irqrestore(&q->requeue_lock, flags);
511 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
513 void blk_mq_cancel_requeue_work(struct request_queue *q)
515 cancel_work_sync(&q->requeue_work);
517 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
519 void blk_mq_kick_requeue_list(struct request_queue *q)
521 kblockd_schedule_work(&q->requeue_work);
523 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
525 void blk_mq_abort_requeue_list(struct request_queue *q)
530 spin_lock_irqsave(&q->requeue_lock, flags);
531 list_splice_init(&q->requeue_list, &rq_list);
532 spin_unlock_irqrestore(&q->requeue_lock, flags);
534 while (!list_empty(&rq_list)) {
537 rq = list_first_entry(&rq_list, struct request, queuelist);
538 list_del_init(&rq->queuelist);
540 blk_mq_end_request(rq, rq->errors);
543 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
545 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
547 if (tag < tags->nr_tags)
548 return tags->rqs[tag];
552 EXPORT_SYMBOL(blk_mq_tag_to_rq);
554 struct blk_mq_timeout_data {
556 unsigned int next_set;
559 void blk_mq_rq_timed_out(struct request *req, bool reserved)
561 struct blk_mq_ops *ops = req->q->mq_ops;
562 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
565 * We know that complete is set at this point. If STARTED isn't set
566 * anymore, then the request isn't active and the "timeout" should
567 * just be ignored. This can happen due to the bitflag ordering.
568 * Timeout first checks if STARTED is set, and if it is, assumes
569 * the request is active. But if we race with completion, then
570 * we both flags will get cleared. So check here again, and ignore
571 * a timeout event with a request that isn't active.
573 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
577 ret = ops->timeout(req, reserved);
581 __blk_mq_complete_request(req);
583 case BLK_EH_RESET_TIMER:
585 blk_clear_rq_complete(req);
587 case BLK_EH_NOT_HANDLED:
590 printk(KERN_ERR "block: bad eh return: %d\n", ret);
595 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
596 struct request *rq, void *priv, bool reserved)
598 struct blk_mq_timeout_data *data = priv;
600 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
602 * If a request wasn't started before the queue was
603 * marked dying, kill it here or it'll go unnoticed.
605 if (unlikely(blk_queue_dying(rq->q))) {
607 blk_mq_end_request(rq, rq->errors);
612 if (time_after_eq(jiffies, rq->deadline)) {
613 if (!blk_mark_rq_complete(rq))
614 blk_mq_rq_timed_out(rq, reserved);
615 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
616 data->next = rq->deadline;
621 static void blk_mq_timeout_work(struct work_struct *work)
623 struct request_queue *q =
624 container_of(work, struct request_queue, timeout_work);
625 struct blk_mq_timeout_data data = {
631 if (blk_queue_enter(q, true))
634 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
637 data.next = blk_rq_timeout(round_jiffies_up(data.next));
638 mod_timer(&q->timeout, data.next);
640 struct blk_mq_hw_ctx *hctx;
642 queue_for_each_hw_ctx(q, hctx, i) {
643 /* the hctx may be unmapped, so check it here */
644 if (blk_mq_hw_queue_mapped(hctx))
645 blk_mq_tag_idle(hctx);
652 * Reverse check our software queue for entries that we could potentially
653 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
654 * too much time checking for merges.
656 static bool blk_mq_attempt_merge(struct request_queue *q,
657 struct blk_mq_ctx *ctx, struct bio *bio)
662 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
668 if (!blk_rq_merge_ok(rq, bio))
671 el_ret = blk_try_merge(rq, bio);
672 if (el_ret == ELEVATOR_BACK_MERGE) {
673 if (bio_attempt_back_merge(q, rq, bio)) {
678 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
679 if (bio_attempt_front_merge(q, rq, bio)) {
691 * Process software queues that have been marked busy, splicing them
692 * to the for-dispatch
694 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
696 struct blk_mq_ctx *ctx;
699 for (i = 0; i < hctx->ctx_map.size; i++) {
700 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
701 unsigned int off, bit;
707 off = i * hctx->ctx_map.bits_per_word;
709 bit = find_next_bit(&bm->word, bm->depth, bit);
710 if (bit >= bm->depth)
713 ctx = hctx->ctxs[bit + off];
714 clear_bit(bit, &bm->word);
715 spin_lock(&ctx->lock);
716 list_splice_tail_init(&ctx->rq_list, list);
717 spin_unlock(&ctx->lock);
725 * Run this hardware queue, pulling any software queues mapped to it in.
726 * Note that this function currently has various problems around ordering
727 * of IO. In particular, we'd like FIFO behaviour on handling existing
728 * items on the hctx->dispatch list. Ignore that for now.
730 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
732 struct request_queue *q = hctx->queue;
735 LIST_HEAD(driver_list);
736 struct list_head *dptr;
739 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
741 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
747 * Touch any software queue that has pending entries.
749 flush_busy_ctxs(hctx, &rq_list);
752 * If we have previous entries on our dispatch list, grab them
753 * and stuff them at the front for more fair dispatch.
755 if (!list_empty_careful(&hctx->dispatch)) {
756 spin_lock(&hctx->lock);
757 if (!list_empty(&hctx->dispatch))
758 list_splice_init(&hctx->dispatch, &rq_list);
759 spin_unlock(&hctx->lock);
763 * Start off with dptr being NULL, so we start the first request
764 * immediately, even if we have more pending.
769 * Now process all the entries, sending them to the driver.
772 while (!list_empty(&rq_list)) {
773 struct blk_mq_queue_data bd;
776 rq = list_first_entry(&rq_list, struct request, queuelist);
777 list_del_init(&rq->queuelist);
781 bd.last = list_empty(&rq_list);
783 ret = q->mq_ops->queue_rq(hctx, &bd);
785 case BLK_MQ_RQ_QUEUE_OK:
788 case BLK_MQ_RQ_QUEUE_BUSY:
789 list_add(&rq->queuelist, &rq_list);
790 __blk_mq_requeue_request(rq);
793 pr_err("blk-mq: bad return on queue: %d\n", ret);
794 case BLK_MQ_RQ_QUEUE_ERROR:
796 blk_mq_end_request(rq, rq->errors);
800 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
804 * We've done the first request. If we have more than 1
805 * left in the list, set dptr to defer issue.
807 if (!dptr && rq_list.next != rq_list.prev)
812 hctx->dispatched[0]++;
813 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
814 hctx->dispatched[ilog2(queued) + 1]++;
817 * Any items that need requeuing? Stuff them into hctx->dispatch,
818 * that is where we will continue on next queue run.
820 if (!list_empty(&rq_list)) {
821 spin_lock(&hctx->lock);
822 list_splice(&rq_list, &hctx->dispatch);
823 spin_unlock(&hctx->lock);
825 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
826 * it's possible the queue is stopped and restarted again
827 * before this. Queue restart will dispatch requests. And since
828 * requests in rq_list aren't added into hctx->dispatch yet,
829 * the requests in rq_list might get lost.
831 * blk_mq_run_hw_queue() already checks the STOPPED bit
833 blk_mq_run_hw_queue(hctx, true);
838 * It'd be great if the workqueue API had a way to pass
839 * in a mask and had some smarts for more clever placement.
840 * For now we just round-robin here, switching for every
841 * BLK_MQ_CPU_WORK_BATCH queued items.
843 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
845 if (hctx->queue->nr_hw_queues == 1)
846 return WORK_CPU_UNBOUND;
848 if (--hctx->next_cpu_batch <= 0) {
849 int cpu = hctx->next_cpu, next_cpu;
851 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
852 if (next_cpu >= nr_cpu_ids)
853 next_cpu = cpumask_first(hctx->cpumask);
855 hctx->next_cpu = next_cpu;
856 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
861 return hctx->next_cpu;
864 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
866 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
867 !blk_mq_hw_queue_mapped(hctx)))
872 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
873 __blk_mq_run_hw_queue(hctx);
881 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
885 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
887 struct blk_mq_hw_ctx *hctx;
890 queue_for_each_hw_ctx(q, hctx, i) {
891 if ((!blk_mq_hctx_has_pending(hctx) &&
892 list_empty_careful(&hctx->dispatch)) ||
893 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
896 blk_mq_run_hw_queue(hctx, async);
899 EXPORT_SYMBOL(blk_mq_run_hw_queues);
901 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
903 cancel_delayed_work(&hctx->run_work);
904 cancel_delayed_work(&hctx->delay_work);
905 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
907 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
909 void blk_mq_stop_hw_queues(struct request_queue *q)
911 struct blk_mq_hw_ctx *hctx;
914 queue_for_each_hw_ctx(q, hctx, i)
915 blk_mq_stop_hw_queue(hctx);
917 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
919 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
921 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
923 blk_mq_run_hw_queue(hctx, false);
925 EXPORT_SYMBOL(blk_mq_start_hw_queue);
927 void blk_mq_start_hw_queues(struct request_queue *q)
929 struct blk_mq_hw_ctx *hctx;
932 queue_for_each_hw_ctx(q, hctx, i)
933 blk_mq_start_hw_queue(hctx);
935 EXPORT_SYMBOL(blk_mq_start_hw_queues);
937 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
939 struct blk_mq_hw_ctx *hctx;
942 queue_for_each_hw_ctx(q, hctx, i) {
943 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
946 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
947 blk_mq_run_hw_queue(hctx, async);
950 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
952 static void blk_mq_run_work_fn(struct work_struct *work)
954 struct blk_mq_hw_ctx *hctx;
956 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
958 __blk_mq_run_hw_queue(hctx);
961 static void blk_mq_delay_work_fn(struct work_struct *work)
963 struct blk_mq_hw_ctx *hctx;
965 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
967 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
968 __blk_mq_run_hw_queue(hctx);
971 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
973 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
976 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
977 &hctx->delay_work, msecs_to_jiffies(msecs));
979 EXPORT_SYMBOL(blk_mq_delay_queue);
981 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
982 struct blk_mq_ctx *ctx,
986 trace_block_rq_insert(hctx->queue, rq);
989 list_add(&rq->queuelist, &ctx->rq_list);
991 list_add_tail(&rq->queuelist, &ctx->rq_list);
994 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
995 struct request *rq, bool at_head)
997 struct blk_mq_ctx *ctx = rq->mq_ctx;
999 __blk_mq_insert_req_list(hctx, ctx, rq, at_head);
1000 blk_mq_hctx_mark_pending(hctx, ctx);
1003 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1006 struct request_queue *q = rq->q;
1007 struct blk_mq_hw_ctx *hctx;
1008 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1010 current_ctx = blk_mq_get_ctx(q);
1011 if (!cpu_online(ctx->cpu))
1012 rq->mq_ctx = ctx = current_ctx;
1014 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1016 spin_lock(&ctx->lock);
1017 __blk_mq_insert_request(hctx, rq, at_head);
1018 spin_unlock(&ctx->lock);
1021 blk_mq_run_hw_queue(hctx, async);
1023 blk_mq_put_ctx(current_ctx);
1026 static void blk_mq_insert_requests(struct request_queue *q,
1027 struct blk_mq_ctx *ctx,
1028 struct list_head *list,
1033 struct blk_mq_hw_ctx *hctx;
1034 struct blk_mq_ctx *current_ctx;
1036 trace_block_unplug(q, depth, !from_schedule);
1038 current_ctx = blk_mq_get_ctx(q);
1040 if (!cpu_online(ctx->cpu))
1042 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1045 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1048 spin_lock(&ctx->lock);
1049 while (!list_empty(list)) {
1052 rq = list_first_entry(list, struct request, queuelist);
1053 list_del_init(&rq->queuelist);
1055 __blk_mq_insert_req_list(hctx, ctx, rq, false);
1057 blk_mq_hctx_mark_pending(hctx, ctx);
1058 spin_unlock(&ctx->lock);
1060 blk_mq_run_hw_queue(hctx, from_schedule);
1061 blk_mq_put_ctx(current_ctx);
1064 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1066 struct request *rqa = container_of(a, struct request, queuelist);
1067 struct request *rqb = container_of(b, struct request, queuelist);
1069 return !(rqa->mq_ctx < rqb->mq_ctx ||
1070 (rqa->mq_ctx == rqb->mq_ctx &&
1071 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1074 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1076 struct blk_mq_ctx *this_ctx;
1077 struct request_queue *this_q;
1080 LIST_HEAD(ctx_list);
1083 list_splice_init(&plug->mq_list, &list);
1085 list_sort(NULL, &list, plug_ctx_cmp);
1091 while (!list_empty(&list)) {
1092 rq = list_entry_rq(list.next);
1093 list_del_init(&rq->queuelist);
1095 if (rq->mq_ctx != this_ctx) {
1097 blk_mq_insert_requests(this_q, this_ctx,
1102 this_ctx = rq->mq_ctx;
1108 list_add_tail(&rq->queuelist, &ctx_list);
1112 * If 'this_ctx' is set, we know we have entries to complete
1113 * on 'ctx_list'. Do those.
1116 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1121 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1123 init_request_from_bio(rq, bio);
1125 blk_account_io_start(rq, 1);
1128 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1130 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1131 !blk_queue_nomerges(hctx->queue);
1134 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1135 struct blk_mq_ctx *ctx,
1136 struct request *rq, struct bio *bio)
1138 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1139 blk_mq_bio_to_request(rq, bio);
1140 spin_lock(&ctx->lock);
1142 __blk_mq_insert_request(hctx, rq, false);
1143 spin_unlock(&ctx->lock);
1146 struct request_queue *q = hctx->queue;
1148 spin_lock(&ctx->lock);
1149 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1150 blk_mq_bio_to_request(rq, bio);
1154 spin_unlock(&ctx->lock);
1155 __blk_mq_free_request(hctx, ctx, rq);
1160 struct blk_map_ctx {
1161 struct blk_mq_hw_ctx *hctx;
1162 struct blk_mq_ctx *ctx;
1165 static struct request *blk_mq_map_request(struct request_queue *q,
1167 struct blk_map_ctx *data)
1169 struct blk_mq_hw_ctx *hctx;
1170 struct blk_mq_ctx *ctx;
1172 int rw = bio_data_dir(bio);
1173 struct blk_mq_alloc_data alloc_data;
1175 blk_queue_enter_live(q);
1176 ctx = blk_mq_get_ctx(q);
1177 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1179 if (rw_is_sync(bio->bi_rw))
1182 trace_block_getrq(q, bio, rw);
1183 blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx);
1184 rq = __blk_mq_alloc_request(&alloc_data, rw);
1185 if (unlikely(!rq)) {
1186 __blk_mq_run_hw_queue(hctx);
1187 blk_mq_put_ctx(ctx);
1188 trace_block_sleeprq(q, bio, rw);
1190 ctx = blk_mq_get_ctx(q);
1191 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1192 blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1193 rq = __blk_mq_alloc_request(&alloc_data, rw);
1194 ctx = alloc_data.ctx;
1195 hctx = alloc_data.hctx;
1204 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1207 struct request_queue *q = rq->q;
1208 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1210 struct blk_mq_queue_data bd = {
1215 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1218 * For OK queue, we are done. For error, kill it. Any other
1219 * error (busy), just add it to our list as we previously
1222 ret = q->mq_ops->queue_rq(hctx, &bd);
1223 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1224 *cookie = new_cookie;
1228 __blk_mq_requeue_request(rq);
1230 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1231 *cookie = BLK_QC_T_NONE;
1233 blk_mq_end_request(rq, rq->errors);
1241 * Multiple hardware queue variant. This will not use per-process plugs,
1242 * but will attempt to bypass the hctx queueing if we can go straight to
1243 * hardware for SYNC IO.
1245 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1247 const int is_sync = rw_is_sync(bio->bi_rw);
1248 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1249 struct blk_map_ctx data;
1251 unsigned int request_count = 0;
1252 struct blk_plug *plug;
1253 struct request *same_queue_rq = NULL;
1256 blk_queue_bounce(q, &bio);
1258 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1260 return BLK_QC_T_NONE;
1263 blk_queue_split(q, &bio, q->bio_split);
1265 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1266 if (blk_attempt_plug_merge(q, bio, &request_count,
1268 return BLK_QC_T_NONE;
1270 request_count = blk_plug_queued_count(q);
1272 rq = blk_mq_map_request(q, bio, &data);
1274 return BLK_QC_T_NONE;
1276 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1278 if (unlikely(is_flush_fua)) {
1279 blk_mq_bio_to_request(rq, bio);
1280 blk_insert_flush(rq);
1284 plug = current->plug;
1286 * If the driver supports defer issued based on 'last', then
1287 * queue it up like normal since we can potentially save some
1290 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1291 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1292 struct request *old_rq = NULL;
1294 blk_mq_bio_to_request(rq, bio);
1297 * We do limited pluging. If the bio can be merged, do that.
1298 * Otherwise the existing request in the plug list will be
1299 * issued. So the plug list will have one request at most
1303 * The plug list might get flushed before this. If that
1304 * happens, same_queue_rq is invalid and plug list is
1307 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1308 old_rq = same_queue_rq;
1309 list_del_init(&old_rq->queuelist);
1311 list_add_tail(&rq->queuelist, &plug->mq_list);
1312 } else /* is_sync */
1314 blk_mq_put_ctx(data.ctx);
1317 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1319 blk_mq_insert_request(old_rq, false, true, true);
1323 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1325 * For a SYNC request, send it to the hardware immediately. For
1326 * an ASYNC request, just ensure that we run it later on. The
1327 * latter allows for merging opportunities and more efficient
1331 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1333 blk_mq_put_ctx(data.ctx);
1339 * Single hardware queue variant. This will attempt to use any per-process
1340 * plug for merging and IO deferral.
1342 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1344 const int is_sync = rw_is_sync(bio->bi_rw);
1345 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1346 struct blk_plug *plug;
1347 unsigned int request_count = 0;
1348 struct blk_map_ctx data;
1352 blk_queue_bounce(q, &bio);
1354 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1356 return BLK_QC_T_NONE;
1359 blk_queue_split(q, &bio, q->bio_split);
1361 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1362 blk_attempt_plug_merge(q, bio, &request_count, NULL))
1363 return BLK_QC_T_NONE;
1365 rq = blk_mq_map_request(q, bio, &data);
1367 return BLK_QC_T_NONE;
1369 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1371 if (unlikely(is_flush_fua)) {
1372 blk_mq_bio_to_request(rq, bio);
1373 blk_insert_flush(rq);
1378 * A task plug currently exists. Since this is completely lockless,
1379 * utilize that to temporarily store requests until the task is
1380 * either done or scheduled away.
1382 plug = current->plug;
1384 blk_mq_bio_to_request(rq, bio);
1386 trace_block_plug(q);
1388 blk_mq_put_ctx(data.ctx);
1390 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1391 blk_flush_plug_list(plug, false);
1392 trace_block_plug(q);
1395 list_add_tail(&rq->queuelist, &plug->mq_list);
1399 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1401 * For a SYNC request, send it to the hardware immediately. For
1402 * an ASYNC request, just ensure that we run it later on. The
1403 * latter allows for merging opportunities and more efficient
1407 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1410 blk_mq_put_ctx(data.ctx);
1415 * Default mapping to a software queue, since we use one per CPU.
1417 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1419 return q->queue_hw_ctx[q->mq_map[cpu]];
1421 EXPORT_SYMBOL(blk_mq_map_queue);
1423 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1424 struct blk_mq_tags *tags, unsigned int hctx_idx)
1428 if (tags->rqs && set->ops->exit_request) {
1431 for (i = 0; i < tags->nr_tags; i++) {
1434 set->ops->exit_request(set->driver_data, tags->rqs[i],
1436 tags->rqs[i] = NULL;
1440 while (!list_empty(&tags->page_list)) {
1441 page = list_first_entry(&tags->page_list, struct page, lru);
1442 list_del_init(&page->lru);
1444 * Remove kmemleak object previously allocated in
1445 * blk_mq_init_rq_map().
1447 kmemleak_free(page_address(page));
1448 __free_pages(page, page->private);
1453 blk_mq_free_tags(tags);
1456 static size_t order_to_size(unsigned int order)
1458 return (size_t)PAGE_SIZE << order;
1461 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1462 unsigned int hctx_idx)
1464 struct blk_mq_tags *tags;
1465 unsigned int i, j, entries_per_page, max_order = 4;
1466 size_t rq_size, left;
1468 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1470 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1474 INIT_LIST_HEAD(&tags->page_list);
1476 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1477 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1480 blk_mq_free_tags(tags);
1485 * rq_size is the size of the request plus driver payload, rounded
1486 * to the cacheline size
1488 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1490 left = rq_size * set->queue_depth;
1492 for (i = 0; i < set->queue_depth; ) {
1493 int this_order = max_order;
1498 while (this_order && left < order_to_size(this_order - 1))
1502 page = alloc_pages_node(set->numa_node,
1503 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1509 if (order_to_size(this_order) < rq_size)
1516 page->private = this_order;
1517 list_add_tail(&page->lru, &tags->page_list);
1519 p = page_address(page);
1521 * Allow kmemleak to scan these pages as they contain pointers
1522 * to additional allocations like via ops->init_request().
1524 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1525 entries_per_page = order_to_size(this_order) / rq_size;
1526 to_do = min(entries_per_page, set->queue_depth - i);
1527 left -= to_do * rq_size;
1528 for (j = 0; j < to_do; j++) {
1530 if (set->ops->init_request) {
1531 if (set->ops->init_request(set->driver_data,
1532 tags->rqs[i], hctx_idx, i,
1534 tags->rqs[i] = NULL;
1546 blk_mq_free_rq_map(set, tags, hctx_idx);
1550 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1555 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1557 unsigned int bpw = 8, total, num_maps, i;
1559 bitmap->bits_per_word = bpw;
1561 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1562 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1568 for (i = 0; i < num_maps; i++) {
1569 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1570 total -= bitmap->map[i].depth;
1576 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1578 struct request_queue *q = hctx->queue;
1579 struct blk_mq_ctx *ctx;
1583 * Move ctx entries to new CPU, if this one is going away.
1585 ctx = __blk_mq_get_ctx(q, cpu);
1587 spin_lock(&ctx->lock);
1588 if (!list_empty(&ctx->rq_list)) {
1589 list_splice_init(&ctx->rq_list, &tmp);
1590 blk_mq_hctx_clear_pending(hctx, ctx);
1592 spin_unlock(&ctx->lock);
1594 if (list_empty(&tmp))
1597 ctx = blk_mq_get_ctx(q);
1598 spin_lock(&ctx->lock);
1600 while (!list_empty(&tmp)) {
1603 rq = list_first_entry(&tmp, struct request, queuelist);
1605 list_move_tail(&rq->queuelist, &ctx->rq_list);
1608 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1609 blk_mq_hctx_mark_pending(hctx, ctx);
1611 spin_unlock(&ctx->lock);
1613 blk_mq_run_hw_queue(hctx, true);
1614 blk_mq_put_ctx(ctx);
1618 static int blk_mq_hctx_notify(void *data, unsigned long action,
1621 struct blk_mq_hw_ctx *hctx = data;
1623 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1624 return blk_mq_hctx_cpu_offline(hctx, cpu);
1627 * In case of CPU online, tags may be reallocated
1628 * in blk_mq_map_swqueue() after mapping is updated.
1634 /* hctx->ctxs will be freed in queue's release handler */
1635 static void blk_mq_exit_hctx(struct request_queue *q,
1636 struct blk_mq_tag_set *set,
1637 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1639 unsigned flush_start_tag = set->queue_depth;
1641 blk_mq_tag_idle(hctx);
1643 if (set->ops->exit_request)
1644 set->ops->exit_request(set->driver_data,
1645 hctx->fq->flush_rq, hctx_idx,
1646 flush_start_tag + hctx_idx);
1648 if (set->ops->exit_hctx)
1649 set->ops->exit_hctx(hctx, hctx_idx);
1651 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1652 blk_free_flush_queue(hctx->fq);
1653 blk_mq_free_bitmap(&hctx->ctx_map);
1656 static void blk_mq_exit_hw_queues(struct request_queue *q,
1657 struct blk_mq_tag_set *set, int nr_queue)
1659 struct blk_mq_hw_ctx *hctx;
1662 queue_for_each_hw_ctx(q, hctx, i) {
1665 blk_mq_exit_hctx(q, set, hctx, i);
1669 static void blk_mq_free_hw_queues(struct request_queue *q,
1670 struct blk_mq_tag_set *set)
1672 struct blk_mq_hw_ctx *hctx;
1675 queue_for_each_hw_ctx(q, hctx, i)
1676 free_cpumask_var(hctx->cpumask);
1679 static int blk_mq_init_hctx(struct request_queue *q,
1680 struct blk_mq_tag_set *set,
1681 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1684 unsigned flush_start_tag = set->queue_depth;
1686 node = hctx->numa_node;
1687 if (node == NUMA_NO_NODE)
1688 node = hctx->numa_node = set->numa_node;
1690 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1691 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1692 spin_lock_init(&hctx->lock);
1693 INIT_LIST_HEAD(&hctx->dispatch);
1695 hctx->queue_num = hctx_idx;
1696 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1698 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1699 blk_mq_hctx_notify, hctx);
1700 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1702 hctx->tags = set->tags[hctx_idx];
1705 * Allocate space for all possible cpus to avoid allocation at
1708 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1711 goto unregister_cpu_notifier;
1713 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1718 if (set->ops->init_hctx &&
1719 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1722 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1726 if (set->ops->init_request &&
1727 set->ops->init_request(set->driver_data,
1728 hctx->fq->flush_rq, hctx_idx,
1729 flush_start_tag + hctx_idx, node))
1737 if (set->ops->exit_hctx)
1738 set->ops->exit_hctx(hctx, hctx_idx);
1740 blk_mq_free_bitmap(&hctx->ctx_map);
1743 unregister_cpu_notifier:
1744 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1749 static void blk_mq_init_cpu_queues(struct request_queue *q,
1750 unsigned int nr_hw_queues)
1754 for_each_possible_cpu(i) {
1755 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1756 struct blk_mq_hw_ctx *hctx;
1758 memset(__ctx, 0, sizeof(*__ctx));
1760 spin_lock_init(&__ctx->lock);
1761 INIT_LIST_HEAD(&__ctx->rq_list);
1764 /* If the cpu isn't online, the cpu is mapped to first hctx */
1768 hctx = q->mq_ops->map_queue(q, i);
1771 * Set local node, IFF we have more than one hw queue. If
1772 * not, we remain on the home node of the device
1774 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1775 hctx->numa_node = local_memory_node(cpu_to_node(i));
1779 static void blk_mq_map_swqueue(struct request_queue *q,
1780 const struct cpumask *online_mask)
1783 struct blk_mq_hw_ctx *hctx;
1784 struct blk_mq_ctx *ctx;
1785 struct blk_mq_tag_set *set = q->tag_set;
1788 * Avoid others reading imcomplete hctx->cpumask through sysfs
1790 mutex_lock(&q->sysfs_lock);
1792 queue_for_each_hw_ctx(q, hctx, i) {
1793 cpumask_clear(hctx->cpumask);
1798 * Map software to hardware queues
1800 for_each_possible_cpu(i) {
1801 /* If the cpu isn't online, the cpu is mapped to first hctx */
1802 if (!cpumask_test_cpu(i, online_mask))
1805 ctx = per_cpu_ptr(q->queue_ctx, i);
1806 hctx = q->mq_ops->map_queue(q, i);
1808 cpumask_set_cpu(i, hctx->cpumask);
1809 ctx->index_hw = hctx->nr_ctx;
1810 hctx->ctxs[hctx->nr_ctx++] = ctx;
1813 mutex_unlock(&q->sysfs_lock);
1815 queue_for_each_hw_ctx(q, hctx, i) {
1816 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1819 * If no software queues are mapped to this hardware queue,
1820 * disable it and free the request entries.
1822 if (!hctx->nr_ctx) {
1824 blk_mq_free_rq_map(set, set->tags[i], i);
1825 set->tags[i] = NULL;
1831 /* unmapped hw queue can be remapped after CPU topo changed */
1833 set->tags[i] = blk_mq_init_rq_map(set, i);
1834 hctx->tags = set->tags[i];
1835 WARN_ON(!hctx->tags);
1837 cpumask_copy(hctx->tags->cpumask, hctx->cpumask);
1839 * Set the map size to the number of mapped software queues.
1840 * This is more accurate and more efficient than looping
1841 * over all possibly mapped software queues.
1843 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1846 * Initialize batch roundrobin counts
1848 hctx->next_cpu = cpumask_first(hctx->cpumask);
1849 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1853 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1855 struct blk_mq_hw_ctx *hctx;
1858 queue_for_each_hw_ctx(q, hctx, i) {
1860 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1862 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1866 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1868 struct request_queue *q;
1870 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1871 blk_mq_freeze_queue(q);
1872 queue_set_hctx_shared(q, shared);
1873 blk_mq_unfreeze_queue(q);
1877 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1879 struct blk_mq_tag_set *set = q->tag_set;
1881 mutex_lock(&set->tag_list_lock);
1882 list_del_init(&q->tag_set_list);
1883 if (list_is_singular(&set->tag_list)) {
1884 /* just transitioned to unshared */
1885 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1886 /* update existing queue */
1887 blk_mq_update_tag_set_depth(set, false);
1889 mutex_unlock(&set->tag_list_lock);
1892 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1893 struct request_queue *q)
1897 mutex_lock(&set->tag_list_lock);
1899 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1900 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1901 set->flags |= BLK_MQ_F_TAG_SHARED;
1902 /* update existing queue */
1903 blk_mq_update_tag_set_depth(set, true);
1905 if (set->flags & BLK_MQ_F_TAG_SHARED)
1906 queue_set_hctx_shared(q, true);
1907 list_add_tail(&q->tag_set_list, &set->tag_list);
1909 mutex_unlock(&set->tag_list_lock);
1913 * It is the actual release handler for mq, but we do it from
1914 * request queue's release handler for avoiding use-after-free
1915 * and headache because q->mq_kobj shouldn't have been introduced,
1916 * but we can't group ctx/kctx kobj without it.
1918 void blk_mq_release(struct request_queue *q)
1920 struct blk_mq_hw_ctx *hctx;
1923 /* hctx kobj stays in hctx */
1924 queue_for_each_hw_ctx(q, hctx, i) {
1934 kfree(q->queue_hw_ctx);
1936 /* ctx kobj stays in queue_ctx */
1937 free_percpu(q->queue_ctx);
1940 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1942 struct request_queue *uninit_q, *q;
1944 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1946 return ERR_PTR(-ENOMEM);
1948 q = blk_mq_init_allocated_queue(set, uninit_q);
1950 blk_cleanup_queue(uninit_q);
1954 EXPORT_SYMBOL(blk_mq_init_queue);
1956 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1957 struct request_queue *q)
1960 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
1962 blk_mq_sysfs_unregister(q);
1963 for (i = 0; i < set->nr_hw_queues; i++) {
1969 node = blk_mq_hw_queue_to_node(q->mq_map, i);
1970 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1975 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1982 atomic_set(&hctxs[i]->nr_active, 0);
1983 hctxs[i]->numa_node = node;
1984 hctxs[i]->queue_num = i;
1986 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
1987 free_cpumask_var(hctxs[i]->cpumask);
1992 blk_mq_hctx_kobj_init(hctxs[i]);
1994 for (j = i; j < q->nr_hw_queues; j++) {
1995 struct blk_mq_hw_ctx *hctx = hctxs[j];
1999 blk_mq_free_rq_map(set, hctx->tags, j);
2000 set->tags[j] = NULL;
2002 blk_mq_exit_hctx(q, set, hctx, j);
2003 free_cpumask_var(hctx->cpumask);
2004 kobject_put(&hctx->kobj);
2011 q->nr_hw_queues = i;
2012 blk_mq_sysfs_register(q);
2015 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2016 struct request_queue *q)
2018 /* mark the queue as mq asap */
2019 q->mq_ops = set->ops;
2021 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2025 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2026 GFP_KERNEL, set->numa_node);
2027 if (!q->queue_hw_ctx)
2030 q->mq_map = blk_mq_make_queue_map(set);
2034 blk_mq_realloc_hw_ctxs(set, q);
2035 if (!q->nr_hw_queues)
2038 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2039 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2041 q->nr_queues = nr_cpu_ids;
2043 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2045 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2046 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2048 q->sg_reserved_size = INT_MAX;
2050 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2051 INIT_LIST_HEAD(&q->requeue_list);
2052 spin_lock_init(&q->requeue_lock);
2054 if (q->nr_hw_queues > 1)
2055 blk_queue_make_request(q, blk_mq_make_request);
2057 blk_queue_make_request(q, blk_sq_make_request);
2060 * Do this after blk_queue_make_request() overrides it...
2062 q->nr_requests = set->queue_depth;
2064 if (set->ops->complete)
2065 blk_queue_softirq_done(q, set->ops->complete);
2067 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2070 mutex_lock(&all_q_mutex);
2072 list_add_tail(&q->all_q_node, &all_q_list);
2073 blk_mq_add_queue_tag_set(set, q);
2074 blk_mq_map_swqueue(q, cpu_online_mask);
2076 mutex_unlock(&all_q_mutex);
2084 kfree(q->queue_hw_ctx);
2086 free_percpu(q->queue_ctx);
2089 return ERR_PTR(-ENOMEM);
2091 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2093 void blk_mq_free_queue(struct request_queue *q)
2095 struct blk_mq_tag_set *set = q->tag_set;
2097 mutex_lock(&all_q_mutex);
2098 list_del_init(&q->all_q_node);
2099 mutex_unlock(&all_q_mutex);
2101 blk_mq_del_queue_tag_set(q);
2103 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2104 blk_mq_free_hw_queues(q, set);
2107 /* Basically redo blk_mq_init_queue with queue frozen */
2108 static void blk_mq_queue_reinit(struct request_queue *q,
2109 const struct cpumask *online_mask)
2111 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2113 blk_mq_sysfs_unregister(q);
2115 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2118 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2119 * we should change hctx numa_node according to new topology (this
2120 * involves free and re-allocate memory, worthy doing?)
2123 blk_mq_map_swqueue(q, online_mask);
2125 blk_mq_sysfs_register(q);
2128 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2129 unsigned long action, void *hcpu)
2131 struct request_queue *q;
2132 int cpu = (unsigned long)hcpu;
2134 * New online cpumask which is going to be set in this hotplug event.
2135 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2136 * one-by-one and dynamically allocating this could result in a failure.
2138 static struct cpumask online_new;
2141 * Before hotadded cpu starts handling requests, new mappings must
2142 * be established. Otherwise, these requests in hw queue might
2143 * never be dispatched.
2145 * For example, there is a single hw queue (hctx) and two CPU queues
2146 * (ctx0 for CPU0, and ctx1 for CPU1).
2148 * Now CPU1 is just onlined and a request is inserted into
2149 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2152 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2153 * set in pending bitmap and tries to retrieve requests in
2154 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2155 * so the request in ctx1->rq_list is ignored.
2157 switch (action & ~CPU_TASKS_FROZEN) {
2159 case CPU_UP_CANCELED:
2160 cpumask_copy(&online_new, cpu_online_mask);
2162 case CPU_UP_PREPARE:
2163 cpumask_copy(&online_new, cpu_online_mask);
2164 cpumask_set_cpu(cpu, &online_new);
2170 mutex_lock(&all_q_mutex);
2173 * We need to freeze and reinit all existing queues. Freezing
2174 * involves synchronous wait for an RCU grace period and doing it
2175 * one by one may take a long time. Start freezing all queues in
2176 * one swoop and then wait for the completions so that freezing can
2177 * take place in parallel.
2179 list_for_each_entry(q, &all_q_list, all_q_node)
2180 blk_mq_freeze_queue_start(q);
2181 list_for_each_entry(q, &all_q_list, all_q_node) {
2182 blk_mq_freeze_queue_wait(q);
2185 * timeout handler can't touch hw queue during the
2188 del_timer_sync(&q->timeout);
2191 list_for_each_entry(q, &all_q_list, all_q_node)
2192 blk_mq_queue_reinit(q, &online_new);
2194 list_for_each_entry(q, &all_q_list, all_q_node)
2195 blk_mq_unfreeze_queue(q);
2197 mutex_unlock(&all_q_mutex);
2201 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2205 for (i = 0; i < set->nr_hw_queues; i++) {
2206 set->tags[i] = blk_mq_init_rq_map(set, i);
2215 blk_mq_free_rq_map(set, set->tags[i], i);
2221 * Allocate the request maps associated with this tag_set. Note that this
2222 * may reduce the depth asked for, if memory is tight. set->queue_depth
2223 * will be updated to reflect the allocated depth.
2225 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2230 depth = set->queue_depth;
2232 err = __blk_mq_alloc_rq_maps(set);
2236 set->queue_depth >>= 1;
2237 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2241 } while (set->queue_depth);
2243 if (!set->queue_depth || err) {
2244 pr_err("blk-mq: failed to allocate request map\n");
2248 if (depth != set->queue_depth)
2249 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2250 depth, set->queue_depth);
2255 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2257 return tags->cpumask;
2259 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2262 * Alloc a tag set to be associated with one or more request queues.
2263 * May fail with EINVAL for various error conditions. May adjust the
2264 * requested depth down, if if it too large. In that case, the set
2265 * value will be stored in set->queue_depth.
2267 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2269 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2271 if (!set->nr_hw_queues)
2273 if (!set->queue_depth)
2275 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2278 if (!set->ops->queue_rq || !set->ops->map_queue)
2281 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2282 pr_info("blk-mq: reduced tag depth to %u\n",
2284 set->queue_depth = BLK_MQ_MAX_DEPTH;
2288 * If a crashdump is active, then we are potentially in a very
2289 * memory constrained environment. Limit us to 1 queue and
2290 * 64 tags to prevent using too much memory.
2292 if (is_kdump_kernel()) {
2293 set->nr_hw_queues = 1;
2294 set->queue_depth = min(64U, set->queue_depth);
2297 * There is no use for more h/w queues than cpus.
2299 if (set->nr_hw_queues > nr_cpu_ids)
2300 set->nr_hw_queues = nr_cpu_ids;
2302 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2303 GFP_KERNEL, set->numa_node);
2307 if (blk_mq_alloc_rq_maps(set))
2310 mutex_init(&set->tag_list_lock);
2311 INIT_LIST_HEAD(&set->tag_list);
2319 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2321 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2325 for (i = 0; i < nr_cpu_ids; i++) {
2327 blk_mq_free_rq_map(set, set->tags[i], i);
2333 EXPORT_SYMBOL(blk_mq_free_tag_set);
2335 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2337 struct blk_mq_tag_set *set = q->tag_set;
2338 struct blk_mq_hw_ctx *hctx;
2341 if (!set || nr > set->queue_depth)
2345 queue_for_each_hw_ctx(q, hctx, i) {
2348 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2354 q->nr_requests = nr;
2359 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2361 struct request_queue *q;
2363 if (nr_hw_queues > nr_cpu_ids)
2364 nr_hw_queues = nr_cpu_ids;
2365 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2368 list_for_each_entry(q, &set->tag_list, tag_set_list)
2369 blk_mq_freeze_queue(q);
2371 set->nr_hw_queues = nr_hw_queues;
2372 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2373 blk_mq_realloc_hw_ctxs(set, q);
2375 if (q->nr_hw_queues > 1)
2376 blk_queue_make_request(q, blk_mq_make_request);
2378 blk_queue_make_request(q, blk_sq_make_request);
2380 blk_mq_queue_reinit(q, cpu_online_mask);
2383 list_for_each_entry(q, &set->tag_list, tag_set_list)
2384 blk_mq_unfreeze_queue(q);
2386 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2388 void blk_mq_disable_hotplug(void)
2390 mutex_lock(&all_q_mutex);
2393 void blk_mq_enable_hotplug(void)
2395 mutex_unlock(&all_q_mutex);
2398 static int __init blk_mq_init(void)
2402 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2406 subsys_initcall(blk_mq_init);