2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
34 static DEFINE_MUTEX(all_q_mutex);
35 static LIST_HEAD(all_q_list);
37 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
40 * Check if any of the ctx's have pending work in this hardware queue
42 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
46 for (i = 0; i < hctx->ctx_map.size; i++)
47 if (hctx->ctx_map.map[i].word)
53 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
54 struct blk_mq_ctx *ctx)
56 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
59 #define CTX_TO_BIT(hctx, ctx) \
60 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
63 * Mark this ctx as having pending work in this hardware queue
65 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
66 struct blk_mq_ctx *ctx)
68 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
70 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
71 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
74 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
75 struct blk_mq_ctx *ctx)
77 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
79 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
82 void blk_mq_freeze_queue_start(struct request_queue *q)
86 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
87 if (freeze_depth == 1) {
88 percpu_ref_kill(&q->q_usage_counter);
89 blk_mq_run_hw_queues(q, false);
92 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
94 static void blk_mq_freeze_queue_wait(struct request_queue *q)
96 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
100 * Guarantee no request is in use, so we can change any data structure of
101 * the queue afterward.
103 void blk_freeze_queue(struct request_queue *q)
106 * In the !blk_mq case we are only calling this to kill the
107 * q_usage_counter, otherwise this increases the freeze depth
108 * and waits for it to return to zero. For this reason there is
109 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
110 * exported to drivers as the only user for unfreeze is blk_mq.
112 blk_mq_freeze_queue_start(q);
113 blk_mq_freeze_queue_wait(q);
116 void blk_mq_freeze_queue(struct request_queue *q)
119 * ...just an alias to keep freeze and unfreeze actions balanced
120 * in the blk_mq_* namespace
124 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
126 void blk_mq_unfreeze_queue(struct request_queue *q)
130 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
131 WARN_ON_ONCE(freeze_depth < 0);
133 percpu_ref_reinit(&q->q_usage_counter);
134 wake_up_all(&q->mq_freeze_wq);
137 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
139 void blk_mq_wake_waiters(struct request_queue *q)
141 struct blk_mq_hw_ctx *hctx;
144 queue_for_each_hw_ctx(q, hctx, i)
145 if (blk_mq_hw_queue_mapped(hctx))
146 blk_mq_tag_wakeup_all(hctx->tags, true);
149 * If we are called because the queue has now been marked as
150 * dying, we need to ensure that processes currently waiting on
151 * the queue are notified as well.
153 wake_up_all(&q->mq_freeze_wq);
156 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
158 return blk_mq_has_free_tags(hctx->tags);
160 EXPORT_SYMBOL(blk_mq_can_queue);
162 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
163 struct request *rq, int op,
164 unsigned int op_flags)
166 if (blk_queue_io_stat(q))
167 op_flags |= REQ_IO_STAT;
169 INIT_LIST_HEAD(&rq->queuelist);
170 /* csd/requeue_work/fifo_time is initialized before use */
173 req_set_op_attrs(rq, op, op_flags);
174 /* do not touch atomic flags, it needs atomic ops against the timer */
176 INIT_HLIST_NODE(&rq->hash);
177 RB_CLEAR_NODE(&rq->rb_node);
180 rq->start_time = jiffies;
181 #ifdef CONFIG_BLK_CGROUP
183 set_start_time_ns(rq);
184 rq->io_start_time_ns = 0;
186 rq->nr_phys_segments = 0;
187 #if defined(CONFIG_BLK_DEV_INTEGRITY)
188 rq->nr_integrity_segments = 0;
191 /* tag was already set */
201 INIT_LIST_HEAD(&rq->timeout_list);
205 rq->end_io_data = NULL;
208 ctx->rq_dispatched[rw_is_sync(op, op_flags)]++;
211 static struct request *
212 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int op, int op_flags)
217 tag = blk_mq_get_tag(data);
218 if (tag != BLK_MQ_TAG_FAIL) {
219 rq = data->hctx->tags->rqs[tag];
221 if (blk_mq_tag_busy(data->hctx)) {
222 rq->cmd_flags = REQ_MQ_INFLIGHT;
223 atomic_inc(&data->hctx->nr_active);
227 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op, op_flags);
234 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
237 struct blk_mq_ctx *ctx;
238 struct blk_mq_hw_ctx *hctx;
240 struct blk_mq_alloc_data alloc_data;
243 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
247 ctx = blk_mq_get_ctx(q);
248 hctx = q->mq_ops->map_queue(q, ctx->cpu);
249 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
251 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
252 if (!rq && !(flags & BLK_MQ_REQ_NOWAIT)) {
253 __blk_mq_run_hw_queue(hctx);
256 ctx = blk_mq_get_ctx(q);
257 hctx = q->mq_ops->map_queue(q, ctx->cpu);
258 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
259 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
260 ctx = alloc_data.ctx;
265 return ERR_PTR(-EWOULDBLOCK);
269 rq->__sector = (sector_t) -1;
270 rq->bio = rq->biotail = NULL;
273 EXPORT_SYMBOL(blk_mq_alloc_request);
275 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
276 unsigned int flags, unsigned int hctx_idx)
278 struct blk_mq_hw_ctx *hctx;
279 struct blk_mq_ctx *ctx;
281 struct blk_mq_alloc_data alloc_data;
285 * If the tag allocator sleeps we could get an allocation for a
286 * different hardware context. No need to complicate the low level
287 * allocator for this for the rare use case of a command tied to
290 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
291 return ERR_PTR(-EINVAL);
293 if (hctx_idx >= q->nr_hw_queues)
294 return ERR_PTR(-EIO);
296 ret = blk_queue_enter(q, true);
300 hctx = q->queue_hw_ctx[hctx_idx];
301 ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
303 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
304 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
307 return ERR_PTR(-EWOULDBLOCK);
312 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
314 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
315 struct blk_mq_ctx *ctx, struct request *rq)
317 const int tag = rq->tag;
318 struct request_queue *q = rq->q;
320 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
321 atomic_dec(&hctx->nr_active);
324 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
325 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
329 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
331 struct blk_mq_ctx *ctx = rq->mq_ctx;
333 ctx->rq_completed[rq_is_sync(rq)]++;
334 __blk_mq_free_request(hctx, ctx, rq);
337 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
339 void blk_mq_free_request(struct request *rq)
341 struct blk_mq_hw_ctx *hctx;
342 struct request_queue *q = rq->q;
344 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
345 blk_mq_free_hctx_request(hctx, rq);
347 EXPORT_SYMBOL_GPL(blk_mq_free_request);
349 inline void __blk_mq_end_request(struct request *rq, int error)
351 blk_account_io_done(rq);
354 rq->end_io(rq, error);
356 if (unlikely(blk_bidi_rq(rq)))
357 blk_mq_free_request(rq->next_rq);
358 blk_mq_free_request(rq);
361 EXPORT_SYMBOL(__blk_mq_end_request);
363 void blk_mq_end_request(struct request *rq, int error)
365 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
367 __blk_mq_end_request(rq, error);
369 EXPORT_SYMBOL(blk_mq_end_request);
371 static void __blk_mq_complete_request_remote(void *data)
373 struct request *rq = data;
375 rq->q->softirq_done_fn(rq);
378 static void blk_mq_ipi_complete_request(struct request *rq)
380 struct blk_mq_ctx *ctx = rq->mq_ctx;
384 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
385 rq->q->softirq_done_fn(rq);
390 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
391 shared = cpus_share_cache(cpu, ctx->cpu);
393 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
394 rq->csd.func = __blk_mq_complete_request_remote;
397 smp_call_function_single_async(ctx->cpu, &rq->csd);
399 rq->q->softirq_done_fn(rq);
404 static void __blk_mq_complete_request(struct request *rq)
406 struct request_queue *q = rq->q;
408 if (!q->softirq_done_fn)
409 blk_mq_end_request(rq, rq->errors);
411 blk_mq_ipi_complete_request(rq);
415 * blk_mq_complete_request - end I/O on a request
416 * @rq: the request being processed
419 * Ends all I/O on a request. It does not handle partial completions.
420 * The actual completion happens out-of-order, through a IPI handler.
422 void blk_mq_complete_request(struct request *rq, int error)
424 struct request_queue *q = rq->q;
426 if (unlikely(blk_should_fake_timeout(q)))
428 if (!blk_mark_rq_complete(rq)) {
430 __blk_mq_complete_request(rq);
433 EXPORT_SYMBOL(blk_mq_complete_request);
435 int blk_mq_request_started(struct request *rq)
437 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
439 EXPORT_SYMBOL_GPL(blk_mq_request_started);
441 void blk_mq_start_request(struct request *rq)
443 struct request_queue *q = rq->q;
445 trace_block_rq_issue(q, rq);
447 rq->resid_len = blk_rq_bytes(rq);
448 if (unlikely(blk_bidi_rq(rq)))
449 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
454 * Ensure that ->deadline is visible before set the started
455 * flag and clear the completed flag.
457 smp_mb__before_atomic();
460 * Mark us as started and clear complete. Complete might have been
461 * set if requeue raced with timeout, which then marked it as
462 * complete. So be sure to clear complete again when we start
463 * the request, otherwise we'll ignore the completion event.
465 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
466 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
467 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
468 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
470 if (q->dma_drain_size && blk_rq_bytes(rq)) {
472 * Make sure space for the drain appears. We know we can do
473 * this because max_hw_segments has been adjusted to be one
474 * fewer than the device can handle.
476 rq->nr_phys_segments++;
479 EXPORT_SYMBOL(blk_mq_start_request);
481 static void __blk_mq_requeue_request(struct request *rq)
483 struct request_queue *q = rq->q;
485 trace_block_rq_requeue(q, rq);
487 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
488 if (q->dma_drain_size && blk_rq_bytes(rq))
489 rq->nr_phys_segments--;
493 void blk_mq_requeue_request(struct request *rq)
495 __blk_mq_requeue_request(rq);
497 BUG_ON(blk_queued_rq(rq));
498 blk_mq_add_to_requeue_list(rq, true);
500 EXPORT_SYMBOL(blk_mq_requeue_request);
502 static void blk_mq_requeue_work(struct work_struct *work)
504 struct request_queue *q =
505 container_of(work, struct request_queue, requeue_work);
507 struct request *rq, *next;
510 spin_lock_irqsave(&q->requeue_lock, flags);
511 list_splice_init(&q->requeue_list, &rq_list);
512 spin_unlock_irqrestore(&q->requeue_lock, flags);
514 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
515 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
518 rq->cmd_flags &= ~REQ_SOFTBARRIER;
519 list_del_init(&rq->queuelist);
520 blk_mq_insert_request(rq, true, false, false);
523 while (!list_empty(&rq_list)) {
524 rq = list_entry(rq_list.next, struct request, queuelist);
525 list_del_init(&rq->queuelist);
526 blk_mq_insert_request(rq, false, false, false);
530 * Use the start variant of queue running here, so that running
531 * the requeue work will kick stopped queues.
533 blk_mq_start_hw_queues(q);
536 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
538 struct request_queue *q = rq->q;
542 * We abuse this flag that is otherwise used by the I/O scheduler to
543 * request head insertation from the workqueue.
545 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
547 spin_lock_irqsave(&q->requeue_lock, flags);
549 rq->cmd_flags |= REQ_SOFTBARRIER;
550 list_add(&rq->queuelist, &q->requeue_list);
552 list_add_tail(&rq->queuelist, &q->requeue_list);
554 spin_unlock_irqrestore(&q->requeue_lock, flags);
556 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
558 void blk_mq_cancel_requeue_work(struct request_queue *q)
560 cancel_work_sync(&q->requeue_work);
562 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
564 void blk_mq_kick_requeue_list(struct request_queue *q)
566 kblockd_schedule_work(&q->requeue_work);
568 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
570 void blk_mq_abort_requeue_list(struct request_queue *q)
575 spin_lock_irqsave(&q->requeue_lock, flags);
576 list_splice_init(&q->requeue_list, &rq_list);
577 spin_unlock_irqrestore(&q->requeue_lock, flags);
579 while (!list_empty(&rq_list)) {
582 rq = list_first_entry(&rq_list, struct request, queuelist);
583 list_del_init(&rq->queuelist);
585 blk_mq_end_request(rq, rq->errors);
588 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
590 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
592 if (tag < tags->nr_tags) {
593 prefetch(tags->rqs[tag]);
594 return tags->rqs[tag];
599 EXPORT_SYMBOL(blk_mq_tag_to_rq);
601 struct blk_mq_timeout_data {
603 unsigned int next_set;
606 void blk_mq_rq_timed_out(struct request *req, bool reserved)
608 struct blk_mq_ops *ops = req->q->mq_ops;
609 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
612 * We know that complete is set at this point. If STARTED isn't set
613 * anymore, then the request isn't active and the "timeout" should
614 * just be ignored. This can happen due to the bitflag ordering.
615 * Timeout first checks if STARTED is set, and if it is, assumes
616 * the request is active. But if we race with completion, then
617 * we both flags will get cleared. So check here again, and ignore
618 * a timeout event with a request that isn't active.
620 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
624 ret = ops->timeout(req, reserved);
628 __blk_mq_complete_request(req);
630 case BLK_EH_RESET_TIMER:
632 blk_clear_rq_complete(req);
634 case BLK_EH_NOT_HANDLED:
637 printk(KERN_ERR "block: bad eh return: %d\n", ret);
642 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
643 struct request *rq, void *priv, bool reserved)
645 struct blk_mq_timeout_data *data = priv;
647 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
649 * If a request wasn't started before the queue was
650 * marked dying, kill it here or it'll go unnoticed.
652 if (unlikely(blk_queue_dying(rq->q))) {
654 blk_mq_end_request(rq, rq->errors);
659 if (time_after_eq(jiffies, rq->deadline)) {
660 if (!blk_mark_rq_complete(rq))
661 blk_mq_rq_timed_out(rq, reserved);
662 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
663 data->next = rq->deadline;
668 static void blk_mq_timeout_work(struct work_struct *work)
670 struct request_queue *q =
671 container_of(work, struct request_queue, timeout_work);
672 struct blk_mq_timeout_data data = {
678 /* A deadlock might occur if a request is stuck requiring a
679 * timeout at the same time a queue freeze is waiting
680 * completion, since the timeout code would not be able to
681 * acquire the queue reference here.
683 * That's why we don't use blk_queue_enter here; instead, we use
684 * percpu_ref_tryget directly, because we need to be able to
685 * obtain a reference even in the short window between the queue
686 * starting to freeze, by dropping the first reference in
687 * blk_mq_freeze_queue_start, and the moment the last request is
688 * consumed, marked by the instant q_usage_counter reaches
691 if (!percpu_ref_tryget(&q->q_usage_counter))
694 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
697 data.next = blk_rq_timeout(round_jiffies_up(data.next));
698 mod_timer(&q->timeout, data.next);
700 struct blk_mq_hw_ctx *hctx;
702 queue_for_each_hw_ctx(q, hctx, i) {
703 /* the hctx may be unmapped, so check it here */
704 if (blk_mq_hw_queue_mapped(hctx))
705 blk_mq_tag_idle(hctx);
712 * Reverse check our software queue for entries that we could potentially
713 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
714 * too much time checking for merges.
716 static bool blk_mq_attempt_merge(struct request_queue *q,
717 struct blk_mq_ctx *ctx, struct bio *bio)
722 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
728 if (!blk_rq_merge_ok(rq, bio))
731 el_ret = blk_try_merge(rq, bio);
732 if (el_ret == ELEVATOR_BACK_MERGE) {
733 if (bio_attempt_back_merge(q, rq, bio)) {
738 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
739 if (bio_attempt_front_merge(q, rq, bio)) {
751 * Process software queues that have been marked busy, splicing them
752 * to the for-dispatch
754 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
756 struct blk_mq_ctx *ctx;
759 for (i = 0; i < hctx->ctx_map.size; i++) {
760 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
761 unsigned int off, bit;
767 off = i * hctx->ctx_map.bits_per_word;
769 bit = find_next_bit(&bm->word, bm->depth, bit);
770 if (bit >= bm->depth)
773 ctx = hctx->ctxs[bit + off];
774 clear_bit(bit, &bm->word);
775 spin_lock(&ctx->lock);
776 list_splice_tail_init(&ctx->rq_list, list);
777 spin_unlock(&ctx->lock);
785 * Run this hardware queue, pulling any software queues mapped to it in.
786 * Note that this function currently has various problems around ordering
787 * of IO. In particular, we'd like FIFO behaviour on handling existing
788 * items on the hctx->dispatch list. Ignore that for now.
790 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
792 struct request_queue *q = hctx->queue;
795 LIST_HEAD(driver_list);
796 struct list_head *dptr;
799 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
802 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
803 cpu_online(hctx->next_cpu));
808 * Touch any software queue that has pending entries.
810 flush_busy_ctxs(hctx, &rq_list);
813 * If we have previous entries on our dispatch list, grab them
814 * and stuff them at the front for more fair dispatch.
816 if (!list_empty_careful(&hctx->dispatch)) {
817 spin_lock(&hctx->lock);
818 if (!list_empty(&hctx->dispatch))
819 list_splice_init(&hctx->dispatch, &rq_list);
820 spin_unlock(&hctx->lock);
824 * Start off with dptr being NULL, so we start the first request
825 * immediately, even if we have more pending.
830 * Now process all the entries, sending them to the driver.
833 while (!list_empty(&rq_list)) {
834 struct blk_mq_queue_data bd;
837 rq = list_first_entry(&rq_list, struct request, queuelist);
838 list_del_init(&rq->queuelist);
842 bd.last = list_empty(&rq_list);
844 ret = q->mq_ops->queue_rq(hctx, &bd);
846 case BLK_MQ_RQ_QUEUE_OK:
849 case BLK_MQ_RQ_QUEUE_BUSY:
850 list_add(&rq->queuelist, &rq_list);
851 __blk_mq_requeue_request(rq);
854 pr_err("blk-mq: bad return on queue: %d\n", ret);
855 case BLK_MQ_RQ_QUEUE_ERROR:
857 blk_mq_end_request(rq, rq->errors);
861 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
865 * We've done the first request. If we have more than 1
866 * left in the list, set dptr to defer issue.
868 if (!dptr && rq_list.next != rq_list.prev)
873 hctx->dispatched[0]++;
874 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
875 hctx->dispatched[ilog2(queued) + 1]++;
878 * Any items that need requeuing? Stuff them into hctx->dispatch,
879 * that is where we will continue on next queue run.
881 if (!list_empty(&rq_list)) {
882 spin_lock(&hctx->lock);
883 list_splice(&rq_list, &hctx->dispatch);
884 spin_unlock(&hctx->lock);
886 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
887 * it's possible the queue is stopped and restarted again
888 * before this. Queue restart will dispatch requests. And since
889 * requests in rq_list aren't added into hctx->dispatch yet,
890 * the requests in rq_list might get lost.
892 * blk_mq_run_hw_queue() already checks the STOPPED bit
894 blk_mq_run_hw_queue(hctx, true);
899 * It'd be great if the workqueue API had a way to pass
900 * in a mask and had some smarts for more clever placement.
901 * For now we just round-robin here, switching for every
902 * BLK_MQ_CPU_WORK_BATCH queued items.
904 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
906 if (hctx->queue->nr_hw_queues == 1)
907 return WORK_CPU_UNBOUND;
909 if (--hctx->next_cpu_batch <= 0) {
910 int cpu = hctx->next_cpu, next_cpu;
912 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
913 if (next_cpu >= nr_cpu_ids)
914 next_cpu = cpumask_first(hctx->cpumask);
916 hctx->next_cpu = next_cpu;
917 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
922 return hctx->next_cpu;
925 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
927 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
928 !blk_mq_hw_queue_mapped(hctx)))
933 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
934 __blk_mq_run_hw_queue(hctx);
942 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
945 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
947 struct blk_mq_hw_ctx *hctx;
950 queue_for_each_hw_ctx(q, hctx, i) {
951 if ((!blk_mq_hctx_has_pending(hctx) &&
952 list_empty_careful(&hctx->dispatch)) ||
953 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
956 blk_mq_run_hw_queue(hctx, async);
959 EXPORT_SYMBOL(blk_mq_run_hw_queues);
961 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
963 cancel_work(&hctx->run_work);
964 cancel_delayed_work(&hctx->delay_work);
965 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
967 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
969 void blk_mq_stop_hw_queues(struct request_queue *q)
971 struct blk_mq_hw_ctx *hctx;
974 queue_for_each_hw_ctx(q, hctx, i)
975 blk_mq_stop_hw_queue(hctx);
977 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
979 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
981 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
983 blk_mq_run_hw_queue(hctx, false);
985 EXPORT_SYMBOL(blk_mq_start_hw_queue);
987 void blk_mq_start_hw_queues(struct request_queue *q)
989 struct blk_mq_hw_ctx *hctx;
992 queue_for_each_hw_ctx(q, hctx, i)
993 blk_mq_start_hw_queue(hctx);
995 EXPORT_SYMBOL(blk_mq_start_hw_queues);
997 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
999 struct blk_mq_hw_ctx *hctx;
1002 queue_for_each_hw_ctx(q, hctx, i) {
1003 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1006 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1007 blk_mq_run_hw_queue(hctx, async);
1010 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1012 static void blk_mq_run_work_fn(struct work_struct *work)
1014 struct blk_mq_hw_ctx *hctx;
1016 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1018 __blk_mq_run_hw_queue(hctx);
1021 static void blk_mq_delay_work_fn(struct work_struct *work)
1023 struct blk_mq_hw_ctx *hctx;
1025 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1027 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1028 __blk_mq_run_hw_queue(hctx);
1031 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1033 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1036 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1037 &hctx->delay_work, msecs_to_jiffies(msecs));
1039 EXPORT_SYMBOL(blk_mq_delay_queue);
1041 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1045 struct blk_mq_ctx *ctx = rq->mq_ctx;
1047 trace_block_rq_insert(hctx->queue, rq);
1050 list_add(&rq->queuelist, &ctx->rq_list);
1052 list_add_tail(&rq->queuelist, &ctx->rq_list);
1055 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1056 struct request *rq, bool at_head)
1058 struct blk_mq_ctx *ctx = rq->mq_ctx;
1060 __blk_mq_insert_req_list(hctx, rq, at_head);
1061 blk_mq_hctx_mark_pending(hctx, ctx);
1064 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1067 struct blk_mq_ctx *ctx = rq->mq_ctx;
1068 struct request_queue *q = rq->q;
1069 struct blk_mq_hw_ctx *hctx;
1071 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1073 spin_lock(&ctx->lock);
1074 __blk_mq_insert_request(hctx, rq, at_head);
1075 spin_unlock(&ctx->lock);
1078 blk_mq_run_hw_queue(hctx, async);
1081 static void blk_mq_insert_requests(struct request_queue *q,
1082 struct blk_mq_ctx *ctx,
1083 struct list_head *list,
1088 struct blk_mq_hw_ctx *hctx;
1090 trace_block_unplug(q, depth, !from_schedule);
1092 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1095 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1098 spin_lock(&ctx->lock);
1099 while (!list_empty(list)) {
1102 rq = list_first_entry(list, struct request, queuelist);
1103 BUG_ON(rq->mq_ctx != ctx);
1104 list_del_init(&rq->queuelist);
1105 __blk_mq_insert_req_list(hctx, rq, false);
1107 blk_mq_hctx_mark_pending(hctx, ctx);
1108 spin_unlock(&ctx->lock);
1110 blk_mq_run_hw_queue(hctx, from_schedule);
1113 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1115 struct request *rqa = container_of(a, struct request, queuelist);
1116 struct request *rqb = container_of(b, struct request, queuelist);
1118 return !(rqa->mq_ctx < rqb->mq_ctx ||
1119 (rqa->mq_ctx == rqb->mq_ctx &&
1120 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1123 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1125 struct blk_mq_ctx *this_ctx;
1126 struct request_queue *this_q;
1129 LIST_HEAD(ctx_list);
1132 list_splice_init(&plug->mq_list, &list);
1134 list_sort(NULL, &list, plug_ctx_cmp);
1140 while (!list_empty(&list)) {
1141 rq = list_entry_rq(list.next);
1142 list_del_init(&rq->queuelist);
1144 if (rq->mq_ctx != this_ctx) {
1146 blk_mq_insert_requests(this_q, this_ctx,
1151 this_ctx = rq->mq_ctx;
1157 list_add_tail(&rq->queuelist, &ctx_list);
1161 * If 'this_ctx' is set, we know we have entries to complete
1162 * on 'ctx_list'. Do those.
1165 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1170 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1172 init_request_from_bio(rq, bio);
1174 blk_account_io_start(rq, 1);
1177 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1179 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1180 !blk_queue_nomerges(hctx->queue);
1183 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1184 struct blk_mq_ctx *ctx,
1185 struct request *rq, struct bio *bio)
1187 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1188 blk_mq_bio_to_request(rq, bio);
1189 spin_lock(&ctx->lock);
1191 __blk_mq_insert_request(hctx, rq, false);
1192 spin_unlock(&ctx->lock);
1195 struct request_queue *q = hctx->queue;
1197 spin_lock(&ctx->lock);
1198 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1199 blk_mq_bio_to_request(rq, bio);
1203 spin_unlock(&ctx->lock);
1204 __blk_mq_free_request(hctx, ctx, rq);
1209 struct blk_map_ctx {
1210 struct blk_mq_hw_ctx *hctx;
1211 struct blk_mq_ctx *ctx;
1214 static struct request *blk_mq_map_request(struct request_queue *q,
1216 struct blk_map_ctx *data)
1218 struct blk_mq_hw_ctx *hctx;
1219 struct blk_mq_ctx *ctx;
1221 int op = bio_data_dir(bio);
1223 struct blk_mq_alloc_data alloc_data;
1225 blk_queue_enter_live(q);
1226 ctx = blk_mq_get_ctx(q);
1227 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1229 if (rw_is_sync(bio_op(bio), bio->bi_opf))
1230 op_flags |= REQ_SYNC;
1232 trace_block_getrq(q, bio, op);
1233 blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx);
1234 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1235 if (unlikely(!rq)) {
1236 __blk_mq_run_hw_queue(hctx);
1237 blk_mq_put_ctx(ctx);
1238 trace_block_sleeprq(q, bio, op);
1240 ctx = blk_mq_get_ctx(q);
1241 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1242 blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1243 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1244 ctx = alloc_data.ctx;
1245 hctx = alloc_data.hctx;
1254 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1257 struct request_queue *q = rq->q;
1258 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1260 struct blk_mq_queue_data bd = {
1265 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1268 * For OK queue, we are done. For error, kill it. Any other
1269 * error (busy), just add it to our list as we previously
1272 ret = q->mq_ops->queue_rq(hctx, &bd);
1273 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1274 *cookie = new_cookie;
1278 __blk_mq_requeue_request(rq);
1280 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1281 *cookie = BLK_QC_T_NONE;
1283 blk_mq_end_request(rq, rq->errors);
1291 * Multiple hardware queue variant. This will not use per-process plugs,
1292 * but will attempt to bypass the hctx queueing if we can go straight to
1293 * hardware for SYNC IO.
1295 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1297 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1298 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1299 struct blk_map_ctx data;
1301 unsigned int request_count = 0;
1302 struct blk_plug *plug;
1303 struct request *same_queue_rq = NULL;
1306 blk_queue_bounce(q, &bio);
1308 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1310 return BLK_QC_T_NONE;
1313 blk_queue_split(q, &bio, q->bio_split);
1315 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1316 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1317 return BLK_QC_T_NONE;
1319 rq = blk_mq_map_request(q, bio, &data);
1321 return BLK_QC_T_NONE;
1323 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1325 if (unlikely(is_flush_fua)) {
1326 blk_mq_bio_to_request(rq, bio);
1327 blk_insert_flush(rq);
1331 plug = current->plug;
1333 * If the driver supports defer issued based on 'last', then
1334 * queue it up like normal since we can potentially save some
1337 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1338 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1339 struct request *old_rq = NULL;
1341 blk_mq_bio_to_request(rq, bio);
1344 * We do limited pluging. If the bio can be merged, do that.
1345 * Otherwise the existing request in the plug list will be
1346 * issued. So the plug list will have one request at most
1350 * The plug list might get flushed before this. If that
1351 * happens, same_queue_rq is invalid and plug list is
1354 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1355 old_rq = same_queue_rq;
1356 list_del_init(&old_rq->queuelist);
1358 list_add_tail(&rq->queuelist, &plug->mq_list);
1359 } else /* is_sync */
1361 blk_mq_put_ctx(data.ctx);
1364 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1366 blk_mq_insert_request(old_rq, false, true, true);
1370 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1372 * For a SYNC request, send it to the hardware immediately. For
1373 * an ASYNC request, just ensure that we run it later on. The
1374 * latter allows for merging opportunities and more efficient
1378 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1380 blk_mq_put_ctx(data.ctx);
1386 * Single hardware queue variant. This will attempt to use any per-process
1387 * plug for merging and IO deferral.
1389 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1391 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1392 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1393 struct blk_plug *plug;
1394 unsigned int request_count = 0;
1395 struct blk_map_ctx data;
1399 blk_queue_bounce(q, &bio);
1401 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1403 return BLK_QC_T_NONE;
1406 blk_queue_split(q, &bio, q->bio_split);
1408 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1409 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1410 return BLK_QC_T_NONE;
1412 request_count = blk_plug_queued_count(q);
1414 rq = blk_mq_map_request(q, bio, &data);
1416 return BLK_QC_T_NONE;
1418 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1420 if (unlikely(is_flush_fua)) {
1421 blk_mq_bio_to_request(rq, bio);
1422 blk_insert_flush(rq);
1427 * A task plug currently exists. Since this is completely lockless,
1428 * utilize that to temporarily store requests until the task is
1429 * either done or scheduled away.
1431 plug = current->plug;
1433 blk_mq_bio_to_request(rq, bio);
1435 trace_block_plug(q);
1437 blk_mq_put_ctx(data.ctx);
1439 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1440 blk_flush_plug_list(plug, false);
1441 trace_block_plug(q);
1444 list_add_tail(&rq->queuelist, &plug->mq_list);
1448 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1450 * For a SYNC request, send it to the hardware immediately. For
1451 * an ASYNC request, just ensure that we run it later on. The
1452 * latter allows for merging opportunities and more efficient
1456 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1459 blk_mq_put_ctx(data.ctx);
1464 * Default mapping to a software queue, since we use one per CPU.
1466 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1468 return q->queue_hw_ctx[q->mq_map[cpu]];
1470 EXPORT_SYMBOL(blk_mq_map_queue);
1472 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1473 struct blk_mq_tags *tags, unsigned int hctx_idx)
1477 if (tags->rqs && set->ops->exit_request) {
1480 for (i = 0; i < tags->nr_tags; i++) {
1483 set->ops->exit_request(set->driver_data, tags->rqs[i],
1485 tags->rqs[i] = NULL;
1489 while (!list_empty(&tags->page_list)) {
1490 page = list_first_entry(&tags->page_list, struct page, lru);
1491 list_del_init(&page->lru);
1493 * Remove kmemleak object previously allocated in
1494 * blk_mq_init_rq_map().
1496 kmemleak_free(page_address(page));
1497 __free_pages(page, page->private);
1502 blk_mq_free_tags(tags);
1505 static size_t order_to_size(unsigned int order)
1507 return (size_t)PAGE_SIZE << order;
1510 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1511 unsigned int hctx_idx)
1513 struct blk_mq_tags *tags;
1514 unsigned int i, j, entries_per_page, max_order = 4;
1515 size_t rq_size, left;
1517 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1519 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1523 INIT_LIST_HEAD(&tags->page_list);
1525 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1526 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1529 blk_mq_free_tags(tags);
1534 * rq_size is the size of the request plus driver payload, rounded
1535 * to the cacheline size
1537 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1539 left = rq_size * set->queue_depth;
1541 for (i = 0; i < set->queue_depth; ) {
1542 int this_order = max_order;
1547 while (this_order && left < order_to_size(this_order - 1))
1551 page = alloc_pages_node(set->numa_node,
1552 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1558 if (order_to_size(this_order) < rq_size)
1565 page->private = this_order;
1566 list_add_tail(&page->lru, &tags->page_list);
1568 p = page_address(page);
1570 * Allow kmemleak to scan these pages as they contain pointers
1571 * to additional allocations like via ops->init_request().
1573 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1574 entries_per_page = order_to_size(this_order) / rq_size;
1575 to_do = min(entries_per_page, set->queue_depth - i);
1576 left -= to_do * rq_size;
1577 for (j = 0; j < to_do; j++) {
1579 if (set->ops->init_request) {
1580 if (set->ops->init_request(set->driver_data,
1581 tags->rqs[i], hctx_idx, i,
1583 tags->rqs[i] = NULL;
1595 blk_mq_free_rq_map(set, tags, hctx_idx);
1599 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1604 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1606 unsigned int bpw = 8, total, num_maps, i;
1608 bitmap->bits_per_word = bpw;
1610 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1611 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1617 for (i = 0; i < num_maps; i++) {
1618 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1619 total -= bitmap->map[i].depth;
1626 * 'cpu' is going away. splice any existing rq_list entries from this
1627 * software queue to the hw queue dispatch list, and ensure that it
1630 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1632 struct blk_mq_ctx *ctx;
1635 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1637 spin_lock(&ctx->lock);
1638 if (!list_empty(&ctx->rq_list)) {
1639 list_splice_init(&ctx->rq_list, &tmp);
1640 blk_mq_hctx_clear_pending(hctx, ctx);
1642 spin_unlock(&ctx->lock);
1644 if (list_empty(&tmp))
1647 spin_lock(&hctx->lock);
1648 list_splice_tail_init(&tmp, &hctx->dispatch);
1649 spin_unlock(&hctx->lock);
1651 blk_mq_run_hw_queue(hctx, true);
1655 static int blk_mq_hctx_notify(void *data, unsigned long action,
1658 struct blk_mq_hw_ctx *hctx = data;
1660 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1661 return blk_mq_hctx_cpu_offline(hctx, cpu);
1664 * In case of CPU online, tags may be reallocated
1665 * in blk_mq_map_swqueue() after mapping is updated.
1671 /* hctx->ctxs will be freed in queue's release handler */
1672 static void blk_mq_exit_hctx(struct request_queue *q,
1673 struct blk_mq_tag_set *set,
1674 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1676 unsigned flush_start_tag = set->queue_depth;
1678 blk_mq_tag_idle(hctx);
1680 if (set->ops->exit_request)
1681 set->ops->exit_request(set->driver_data,
1682 hctx->fq->flush_rq, hctx_idx,
1683 flush_start_tag + hctx_idx);
1685 if (set->ops->exit_hctx)
1686 set->ops->exit_hctx(hctx, hctx_idx);
1688 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1689 blk_free_flush_queue(hctx->fq);
1690 blk_mq_free_bitmap(&hctx->ctx_map);
1693 static void blk_mq_exit_hw_queues(struct request_queue *q,
1694 struct blk_mq_tag_set *set, int nr_queue)
1696 struct blk_mq_hw_ctx *hctx;
1699 queue_for_each_hw_ctx(q, hctx, i) {
1702 blk_mq_exit_hctx(q, set, hctx, i);
1706 static void blk_mq_free_hw_queues(struct request_queue *q,
1707 struct blk_mq_tag_set *set)
1709 struct blk_mq_hw_ctx *hctx;
1712 queue_for_each_hw_ctx(q, hctx, i)
1713 free_cpumask_var(hctx->cpumask);
1716 static int blk_mq_init_hctx(struct request_queue *q,
1717 struct blk_mq_tag_set *set,
1718 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1721 unsigned flush_start_tag = set->queue_depth;
1723 node = hctx->numa_node;
1724 if (node == NUMA_NO_NODE)
1725 node = hctx->numa_node = set->numa_node;
1727 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1728 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1729 spin_lock_init(&hctx->lock);
1730 INIT_LIST_HEAD(&hctx->dispatch);
1732 hctx->queue_num = hctx_idx;
1733 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1735 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1736 blk_mq_hctx_notify, hctx);
1737 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1739 hctx->tags = set->tags[hctx_idx];
1742 * Allocate space for all possible cpus to avoid allocation at
1745 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1748 goto unregister_cpu_notifier;
1750 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1755 if (set->ops->init_hctx &&
1756 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1759 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1763 if (set->ops->init_request &&
1764 set->ops->init_request(set->driver_data,
1765 hctx->fq->flush_rq, hctx_idx,
1766 flush_start_tag + hctx_idx, node))
1774 if (set->ops->exit_hctx)
1775 set->ops->exit_hctx(hctx, hctx_idx);
1777 blk_mq_free_bitmap(&hctx->ctx_map);
1780 unregister_cpu_notifier:
1781 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1786 static void blk_mq_init_cpu_queues(struct request_queue *q,
1787 unsigned int nr_hw_queues)
1791 for_each_possible_cpu(i) {
1792 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1793 struct blk_mq_hw_ctx *hctx;
1795 memset(__ctx, 0, sizeof(*__ctx));
1797 spin_lock_init(&__ctx->lock);
1798 INIT_LIST_HEAD(&__ctx->rq_list);
1801 /* If the cpu isn't online, the cpu is mapped to first hctx */
1805 hctx = q->mq_ops->map_queue(q, i);
1808 * Set local node, IFF we have more than one hw queue. If
1809 * not, we remain on the home node of the device
1811 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1812 hctx->numa_node = local_memory_node(cpu_to_node(i));
1816 static void blk_mq_map_swqueue(struct request_queue *q,
1817 const struct cpumask *online_mask)
1820 struct blk_mq_hw_ctx *hctx;
1821 struct blk_mq_ctx *ctx;
1822 struct blk_mq_tag_set *set = q->tag_set;
1825 * Avoid others reading imcomplete hctx->cpumask through sysfs
1827 mutex_lock(&q->sysfs_lock);
1829 queue_for_each_hw_ctx(q, hctx, i) {
1830 cpumask_clear(hctx->cpumask);
1835 * Map software to hardware queues
1837 for_each_possible_cpu(i) {
1838 /* If the cpu isn't online, the cpu is mapped to first hctx */
1839 if (!cpumask_test_cpu(i, online_mask))
1842 ctx = per_cpu_ptr(q->queue_ctx, i);
1843 hctx = q->mq_ops->map_queue(q, i);
1845 cpumask_set_cpu(i, hctx->cpumask);
1846 ctx->index_hw = hctx->nr_ctx;
1847 hctx->ctxs[hctx->nr_ctx++] = ctx;
1850 mutex_unlock(&q->sysfs_lock);
1852 queue_for_each_hw_ctx(q, hctx, i) {
1853 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1856 * If no software queues are mapped to this hardware queue,
1857 * disable it and free the request entries.
1859 if (!hctx->nr_ctx) {
1861 blk_mq_free_rq_map(set, set->tags[i], i);
1862 set->tags[i] = NULL;
1868 /* unmapped hw queue can be remapped after CPU topo changed */
1870 set->tags[i] = blk_mq_init_rq_map(set, i);
1871 hctx->tags = set->tags[i];
1872 WARN_ON(!hctx->tags);
1874 cpumask_copy(hctx->tags->cpumask, hctx->cpumask);
1876 * Set the map size to the number of mapped software queues.
1877 * This is more accurate and more efficient than looping
1878 * over all possibly mapped software queues.
1880 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1883 * Initialize batch roundrobin counts
1885 hctx->next_cpu = cpumask_first(hctx->cpumask);
1886 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1890 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1892 struct blk_mq_hw_ctx *hctx;
1895 queue_for_each_hw_ctx(q, hctx, i) {
1897 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1899 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1903 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1905 struct request_queue *q;
1907 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1908 blk_mq_freeze_queue(q);
1909 queue_set_hctx_shared(q, shared);
1910 blk_mq_unfreeze_queue(q);
1914 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1916 struct blk_mq_tag_set *set = q->tag_set;
1918 mutex_lock(&set->tag_list_lock);
1919 list_del_init(&q->tag_set_list);
1920 if (list_is_singular(&set->tag_list)) {
1921 /* just transitioned to unshared */
1922 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1923 /* update existing queue */
1924 blk_mq_update_tag_set_depth(set, false);
1926 mutex_unlock(&set->tag_list_lock);
1929 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1930 struct request_queue *q)
1934 mutex_lock(&set->tag_list_lock);
1936 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1937 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1938 set->flags |= BLK_MQ_F_TAG_SHARED;
1939 /* update existing queue */
1940 blk_mq_update_tag_set_depth(set, true);
1942 if (set->flags & BLK_MQ_F_TAG_SHARED)
1943 queue_set_hctx_shared(q, true);
1944 list_add_tail(&q->tag_set_list, &set->tag_list);
1946 mutex_unlock(&set->tag_list_lock);
1950 * It is the actual release handler for mq, but we do it from
1951 * request queue's release handler for avoiding use-after-free
1952 * and headache because q->mq_kobj shouldn't have been introduced,
1953 * but we can't group ctx/kctx kobj without it.
1955 void blk_mq_release(struct request_queue *q)
1957 struct blk_mq_hw_ctx *hctx;
1960 /* hctx kobj stays in hctx */
1961 queue_for_each_hw_ctx(q, hctx, i) {
1971 kfree(q->queue_hw_ctx);
1973 /* ctx kobj stays in queue_ctx */
1974 free_percpu(q->queue_ctx);
1977 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1979 struct request_queue *uninit_q, *q;
1981 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1983 return ERR_PTR(-ENOMEM);
1985 q = blk_mq_init_allocated_queue(set, uninit_q);
1987 blk_cleanup_queue(uninit_q);
1991 EXPORT_SYMBOL(blk_mq_init_queue);
1993 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1994 struct request_queue *q)
1997 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
1999 blk_mq_sysfs_unregister(q);
2000 for (i = 0; i < set->nr_hw_queues; i++) {
2006 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2007 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2012 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2019 atomic_set(&hctxs[i]->nr_active, 0);
2020 hctxs[i]->numa_node = node;
2021 hctxs[i]->queue_num = i;
2023 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2024 free_cpumask_var(hctxs[i]->cpumask);
2029 blk_mq_hctx_kobj_init(hctxs[i]);
2031 for (j = i; j < q->nr_hw_queues; j++) {
2032 struct blk_mq_hw_ctx *hctx = hctxs[j];
2036 blk_mq_free_rq_map(set, hctx->tags, j);
2037 set->tags[j] = NULL;
2039 blk_mq_exit_hctx(q, set, hctx, j);
2040 free_cpumask_var(hctx->cpumask);
2041 kobject_put(&hctx->kobj);
2048 q->nr_hw_queues = i;
2049 blk_mq_sysfs_register(q);
2052 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2053 struct request_queue *q)
2055 /* mark the queue as mq asap */
2056 q->mq_ops = set->ops;
2058 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2062 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2063 GFP_KERNEL, set->numa_node);
2064 if (!q->queue_hw_ctx)
2067 q->mq_map = blk_mq_make_queue_map(set);
2071 blk_mq_realloc_hw_ctxs(set, q);
2072 if (!q->nr_hw_queues)
2075 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2076 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2078 q->nr_queues = nr_cpu_ids;
2080 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2082 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2083 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2085 q->sg_reserved_size = INT_MAX;
2087 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2088 INIT_LIST_HEAD(&q->requeue_list);
2089 spin_lock_init(&q->requeue_lock);
2091 if (q->nr_hw_queues > 1)
2092 blk_queue_make_request(q, blk_mq_make_request);
2094 blk_queue_make_request(q, blk_sq_make_request);
2097 * Do this after blk_queue_make_request() overrides it...
2099 q->nr_requests = set->queue_depth;
2101 if (set->ops->complete)
2102 blk_queue_softirq_done(q, set->ops->complete);
2104 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2107 mutex_lock(&all_q_mutex);
2109 list_add_tail(&q->all_q_node, &all_q_list);
2110 blk_mq_add_queue_tag_set(set, q);
2111 blk_mq_map_swqueue(q, cpu_online_mask);
2113 mutex_unlock(&all_q_mutex);
2121 kfree(q->queue_hw_ctx);
2123 free_percpu(q->queue_ctx);
2126 return ERR_PTR(-ENOMEM);
2128 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2130 void blk_mq_free_queue(struct request_queue *q)
2132 struct blk_mq_tag_set *set = q->tag_set;
2134 mutex_lock(&all_q_mutex);
2135 list_del_init(&q->all_q_node);
2136 mutex_unlock(&all_q_mutex);
2138 blk_mq_del_queue_tag_set(q);
2140 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2141 blk_mq_free_hw_queues(q, set);
2144 /* Basically redo blk_mq_init_queue with queue frozen */
2145 static void blk_mq_queue_reinit(struct request_queue *q,
2146 const struct cpumask *online_mask)
2148 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2150 blk_mq_sysfs_unregister(q);
2152 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2155 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2156 * we should change hctx numa_node according to new topology (this
2157 * involves free and re-allocate memory, worthy doing?)
2160 blk_mq_map_swqueue(q, online_mask);
2162 blk_mq_sysfs_register(q);
2165 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2166 unsigned long action, void *hcpu)
2168 struct request_queue *q;
2169 int cpu = (unsigned long)hcpu;
2171 * New online cpumask which is going to be set in this hotplug event.
2172 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2173 * one-by-one and dynamically allocating this could result in a failure.
2175 static struct cpumask online_new;
2178 * Before hotadded cpu starts handling requests, new mappings must
2179 * be established. Otherwise, these requests in hw queue might
2180 * never be dispatched.
2182 * For example, there is a single hw queue (hctx) and two CPU queues
2183 * (ctx0 for CPU0, and ctx1 for CPU1).
2185 * Now CPU1 is just onlined and a request is inserted into
2186 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2189 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2190 * set in pending bitmap and tries to retrieve requests in
2191 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2192 * so the request in ctx1->rq_list is ignored.
2194 switch (action & ~CPU_TASKS_FROZEN) {
2196 case CPU_UP_CANCELED:
2197 cpumask_copy(&online_new, cpu_online_mask);
2199 case CPU_UP_PREPARE:
2200 cpumask_copy(&online_new, cpu_online_mask);
2201 cpumask_set_cpu(cpu, &online_new);
2207 mutex_lock(&all_q_mutex);
2210 * We need to freeze and reinit all existing queues. Freezing
2211 * involves synchronous wait for an RCU grace period and doing it
2212 * one by one may take a long time. Start freezing all queues in
2213 * one swoop and then wait for the completions so that freezing can
2214 * take place in parallel.
2216 list_for_each_entry(q, &all_q_list, all_q_node)
2217 blk_mq_freeze_queue_start(q);
2218 list_for_each_entry(q, &all_q_list, all_q_node) {
2219 blk_mq_freeze_queue_wait(q);
2222 * timeout handler can't touch hw queue during the
2225 del_timer_sync(&q->timeout);
2228 list_for_each_entry(q, &all_q_list, all_q_node)
2229 blk_mq_queue_reinit(q, &online_new);
2231 list_for_each_entry(q, &all_q_list, all_q_node)
2232 blk_mq_unfreeze_queue(q);
2234 mutex_unlock(&all_q_mutex);
2238 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2242 for (i = 0; i < set->nr_hw_queues; i++) {
2243 set->tags[i] = blk_mq_init_rq_map(set, i);
2252 blk_mq_free_rq_map(set, set->tags[i], i);
2258 * Allocate the request maps associated with this tag_set. Note that this
2259 * may reduce the depth asked for, if memory is tight. set->queue_depth
2260 * will be updated to reflect the allocated depth.
2262 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2267 depth = set->queue_depth;
2269 err = __blk_mq_alloc_rq_maps(set);
2273 set->queue_depth >>= 1;
2274 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2278 } while (set->queue_depth);
2280 if (!set->queue_depth || err) {
2281 pr_err("blk-mq: failed to allocate request map\n");
2285 if (depth != set->queue_depth)
2286 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2287 depth, set->queue_depth);
2292 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2294 return tags->cpumask;
2296 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2299 * Alloc a tag set to be associated with one or more request queues.
2300 * May fail with EINVAL for various error conditions. May adjust the
2301 * requested depth down, if if it too large. In that case, the set
2302 * value will be stored in set->queue_depth.
2304 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2306 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2308 if (!set->nr_hw_queues)
2310 if (!set->queue_depth)
2312 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2315 if (!set->ops->queue_rq || !set->ops->map_queue)
2318 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2319 pr_info("blk-mq: reduced tag depth to %u\n",
2321 set->queue_depth = BLK_MQ_MAX_DEPTH;
2325 * If a crashdump is active, then we are potentially in a very
2326 * memory constrained environment. Limit us to 1 queue and
2327 * 64 tags to prevent using too much memory.
2329 if (is_kdump_kernel()) {
2330 set->nr_hw_queues = 1;
2331 set->queue_depth = min(64U, set->queue_depth);
2334 * There is no use for more h/w queues than cpus.
2336 if (set->nr_hw_queues > nr_cpu_ids)
2337 set->nr_hw_queues = nr_cpu_ids;
2339 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2340 GFP_KERNEL, set->numa_node);
2344 if (blk_mq_alloc_rq_maps(set))
2347 mutex_init(&set->tag_list_lock);
2348 INIT_LIST_HEAD(&set->tag_list);
2356 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2358 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2362 for (i = 0; i < nr_cpu_ids; i++) {
2364 blk_mq_free_rq_map(set, set->tags[i], i);
2370 EXPORT_SYMBOL(blk_mq_free_tag_set);
2372 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2374 struct blk_mq_tag_set *set = q->tag_set;
2375 struct blk_mq_hw_ctx *hctx;
2378 if (!set || nr > set->queue_depth)
2382 queue_for_each_hw_ctx(q, hctx, i) {
2385 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2391 q->nr_requests = nr;
2396 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2398 struct request_queue *q;
2400 if (nr_hw_queues > nr_cpu_ids)
2401 nr_hw_queues = nr_cpu_ids;
2402 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2405 list_for_each_entry(q, &set->tag_list, tag_set_list)
2406 blk_mq_freeze_queue(q);
2408 set->nr_hw_queues = nr_hw_queues;
2409 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2410 blk_mq_realloc_hw_ctxs(set, q);
2412 if (q->nr_hw_queues > 1)
2413 blk_queue_make_request(q, blk_mq_make_request);
2415 blk_queue_make_request(q, blk_sq_make_request);
2417 blk_mq_queue_reinit(q, cpu_online_mask);
2420 list_for_each_entry(q, &set->tag_list, tag_set_list)
2421 blk_mq_unfreeze_queue(q);
2423 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2425 void blk_mq_disable_hotplug(void)
2427 mutex_lock(&all_q_mutex);
2430 void blk_mq_enable_hotplug(void)
2432 mutex_unlock(&all_q_mutex);
2435 static int __init blk_mq_init(void)
2439 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2443 subsys_initcall(blk_mq_init);