2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
34 static DEFINE_MUTEX(all_q_mutex);
35 static LIST_HEAD(all_q_list);
38 * Check if any of the ctx's have pending work in this hardware queue
40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
42 return sbitmap_any_bit_set(&hctx->ctx_map);
46 * Mark this ctx as having pending work in this hardware queue
48 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
49 struct blk_mq_ctx *ctx)
51 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
52 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
55 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
56 struct blk_mq_ctx *ctx)
58 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
61 void blk_mq_freeze_queue_start(struct request_queue *q)
65 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
66 if (freeze_depth == 1) {
67 percpu_ref_kill(&q->q_usage_counter);
68 blk_mq_run_hw_queues(q, false);
71 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
73 static void blk_mq_freeze_queue_wait(struct request_queue *q)
75 swait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
79 * Guarantee no request is in use, so we can change any data structure of
80 * the queue afterward.
82 void blk_freeze_queue(struct request_queue *q)
85 * In the !blk_mq case we are only calling this to kill the
86 * q_usage_counter, otherwise this increases the freeze depth
87 * and waits for it to return to zero. For this reason there is
88 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
89 * exported to drivers as the only user for unfreeze is blk_mq.
91 blk_mq_freeze_queue_start(q);
92 blk_mq_freeze_queue_wait(q);
95 void blk_mq_freeze_queue(struct request_queue *q)
98 * ...just an alias to keep freeze and unfreeze actions balanced
99 * in the blk_mq_* namespace
103 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
105 void blk_mq_unfreeze_queue(struct request_queue *q)
109 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
110 WARN_ON_ONCE(freeze_depth < 0);
112 percpu_ref_reinit(&q->q_usage_counter);
113 swake_up_all(&q->mq_freeze_wq);
116 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
118 void blk_mq_wake_waiters(struct request_queue *q)
120 struct blk_mq_hw_ctx *hctx;
123 queue_for_each_hw_ctx(q, hctx, i)
124 if (blk_mq_hw_queue_mapped(hctx))
125 blk_mq_tag_wakeup_all(hctx->tags, true);
128 * If we are called because the queue has now been marked as
129 * dying, we need to ensure that processes currently waiting on
130 * the queue are notified as well.
132 swake_up_all(&q->mq_freeze_wq);
135 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
137 return blk_mq_has_free_tags(hctx->tags);
139 EXPORT_SYMBOL(blk_mq_can_queue);
141 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
142 struct request *rq, int op,
143 unsigned int op_flags)
145 if (blk_queue_io_stat(q))
146 op_flags |= REQ_IO_STAT;
148 INIT_LIST_HEAD(&rq->queuelist);
149 /* csd/requeue_work/fifo_time is initialized before use */
152 req_set_op_attrs(rq, op, op_flags);
153 /* do not touch atomic flags, it needs atomic ops against the timer */
155 INIT_HLIST_NODE(&rq->hash);
156 RB_CLEAR_NODE(&rq->rb_node);
159 rq->start_time = jiffies;
160 #ifdef CONFIG_BLK_CGROUP
162 set_start_time_ns(rq);
163 rq->io_start_time_ns = 0;
165 rq->nr_phys_segments = 0;
166 #if defined(CONFIG_BLK_DEV_INTEGRITY)
167 rq->nr_integrity_segments = 0;
170 /* tag was already set */
180 #ifdef CONFIG_PREEMPT_RT_FULL
181 INIT_WORK(&rq->work, __blk_mq_complete_request_remote_work);
183 INIT_LIST_HEAD(&rq->timeout_list);
187 rq->end_io_data = NULL;
190 ctx->rq_dispatched[rw_is_sync(op, op_flags)]++;
193 static struct request *
194 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int op, int op_flags)
199 tag = blk_mq_get_tag(data);
200 if (tag != BLK_MQ_TAG_FAIL) {
201 rq = data->hctx->tags->rqs[tag];
203 if (blk_mq_tag_busy(data->hctx)) {
204 rq->cmd_flags = REQ_MQ_INFLIGHT;
205 atomic_inc(&data->hctx->nr_active);
209 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op, op_flags);
216 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
219 struct blk_mq_ctx *ctx;
220 struct blk_mq_hw_ctx *hctx;
222 struct blk_mq_alloc_data alloc_data;
225 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
229 ctx = blk_mq_get_ctx(q);
230 hctx = blk_mq_map_queue(q, ctx->cpu);
231 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
232 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
237 return ERR_PTR(-EWOULDBLOCK);
241 rq->__sector = (sector_t) -1;
242 rq->bio = rq->biotail = NULL;
245 EXPORT_SYMBOL(blk_mq_alloc_request);
247 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
248 unsigned int flags, unsigned int hctx_idx)
250 struct blk_mq_hw_ctx *hctx;
251 struct blk_mq_ctx *ctx;
253 struct blk_mq_alloc_data alloc_data;
257 * If the tag allocator sleeps we could get an allocation for a
258 * different hardware context. No need to complicate the low level
259 * allocator for this for the rare use case of a command tied to
262 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
263 return ERR_PTR(-EINVAL);
265 if (hctx_idx >= q->nr_hw_queues)
266 return ERR_PTR(-EIO);
268 ret = blk_queue_enter(q, true);
273 * Check if the hardware context is actually mapped to anything.
274 * If not tell the caller that it should skip this queue.
276 hctx = q->queue_hw_ctx[hctx_idx];
277 if (!blk_mq_hw_queue_mapped(hctx)) {
281 ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
283 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
284 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
296 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
298 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
299 struct blk_mq_ctx *ctx, struct request *rq)
301 const int tag = rq->tag;
302 struct request_queue *q = rq->q;
304 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
305 atomic_dec(&hctx->nr_active);
308 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
309 blk_mq_put_tag(hctx, ctx, tag);
313 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
315 struct blk_mq_ctx *ctx = rq->mq_ctx;
317 ctx->rq_completed[rq_is_sync(rq)]++;
318 __blk_mq_free_request(hctx, ctx, rq);
321 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
323 void blk_mq_free_request(struct request *rq)
325 blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
327 EXPORT_SYMBOL_GPL(blk_mq_free_request);
329 inline void __blk_mq_end_request(struct request *rq, int error)
331 blk_account_io_done(rq);
334 rq->end_io(rq, error);
336 if (unlikely(blk_bidi_rq(rq)))
337 blk_mq_free_request(rq->next_rq);
338 blk_mq_free_request(rq);
341 EXPORT_SYMBOL(__blk_mq_end_request);
343 void blk_mq_end_request(struct request *rq, int error)
345 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
347 __blk_mq_end_request(rq, error);
349 EXPORT_SYMBOL(blk_mq_end_request);
351 #ifdef CONFIG_PREEMPT_RT_FULL
353 void __blk_mq_complete_request_remote_work(struct work_struct *work)
355 struct request *rq = container_of(work, struct request, work);
357 rq->q->softirq_done_fn(rq);
362 static void __blk_mq_complete_request_remote(void *data)
364 struct request *rq = data;
366 rq->q->softirq_done_fn(rq);
371 static void blk_mq_ipi_complete_request(struct request *rq)
373 struct blk_mq_ctx *ctx = rq->mq_ctx;
377 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
378 rq->q->softirq_done_fn(rq);
382 cpu = get_cpu_light();
383 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
384 shared = cpus_share_cache(cpu, ctx->cpu);
386 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
387 #ifdef CONFIG_PREEMPT_RT_FULL
388 schedule_work_on(ctx->cpu, &rq->work);
390 rq->csd.func = __blk_mq_complete_request_remote;
393 smp_call_function_single_async(ctx->cpu, &rq->csd);
396 rq->q->softirq_done_fn(rq);
401 static void __blk_mq_complete_request(struct request *rq)
403 struct request_queue *q = rq->q;
405 if (!q->softirq_done_fn)
406 blk_mq_end_request(rq, rq->errors);
408 blk_mq_ipi_complete_request(rq);
412 * blk_mq_complete_request - end I/O on a request
413 * @rq: the request being processed
416 * Ends all I/O on a request. It does not handle partial completions.
417 * The actual completion happens out-of-order, through a IPI handler.
419 void blk_mq_complete_request(struct request *rq, int error)
421 struct request_queue *q = rq->q;
423 if (unlikely(blk_should_fake_timeout(q)))
425 if (!blk_mark_rq_complete(rq)) {
427 __blk_mq_complete_request(rq);
430 EXPORT_SYMBOL(blk_mq_complete_request);
432 int blk_mq_request_started(struct request *rq)
434 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
436 EXPORT_SYMBOL_GPL(blk_mq_request_started);
438 void blk_mq_start_request(struct request *rq)
440 struct request_queue *q = rq->q;
442 trace_block_rq_issue(q, rq);
444 rq->resid_len = blk_rq_bytes(rq);
445 if (unlikely(blk_bidi_rq(rq)))
446 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
451 * Ensure that ->deadline is visible before set the started
452 * flag and clear the completed flag.
454 smp_mb__before_atomic();
457 * Mark us as started and clear complete. Complete might have been
458 * set if requeue raced with timeout, which then marked it as
459 * complete. So be sure to clear complete again when we start
460 * the request, otherwise we'll ignore the completion event.
462 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
463 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
464 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
465 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
467 if (q->dma_drain_size && blk_rq_bytes(rq)) {
469 * Make sure space for the drain appears. We know we can do
470 * this because max_hw_segments has been adjusted to be one
471 * fewer than the device can handle.
473 rq->nr_phys_segments++;
476 EXPORT_SYMBOL(blk_mq_start_request);
478 static void __blk_mq_requeue_request(struct request *rq)
480 struct request_queue *q = rq->q;
482 trace_block_rq_requeue(q, rq);
484 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
485 if (q->dma_drain_size && blk_rq_bytes(rq))
486 rq->nr_phys_segments--;
490 void blk_mq_requeue_request(struct request *rq)
492 __blk_mq_requeue_request(rq);
494 BUG_ON(blk_queued_rq(rq));
495 blk_mq_add_to_requeue_list(rq, true);
497 EXPORT_SYMBOL(blk_mq_requeue_request);
499 static void blk_mq_requeue_work(struct work_struct *work)
501 struct request_queue *q =
502 container_of(work, struct request_queue, requeue_work.work);
504 struct request *rq, *next;
507 spin_lock_irqsave(&q->requeue_lock, flags);
508 list_splice_init(&q->requeue_list, &rq_list);
509 spin_unlock_irqrestore(&q->requeue_lock, flags);
511 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
512 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
515 rq->cmd_flags &= ~REQ_SOFTBARRIER;
516 list_del_init(&rq->queuelist);
517 blk_mq_insert_request(rq, true, false, false);
520 while (!list_empty(&rq_list)) {
521 rq = list_entry(rq_list.next, struct request, queuelist);
522 list_del_init(&rq->queuelist);
523 blk_mq_insert_request(rq, false, false, false);
527 * Use the start variant of queue running here, so that running
528 * the requeue work will kick stopped queues.
530 blk_mq_start_hw_queues(q);
533 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
535 struct request_queue *q = rq->q;
539 * We abuse this flag that is otherwise used by the I/O scheduler to
540 * request head insertation from the workqueue.
542 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
544 spin_lock_irqsave(&q->requeue_lock, flags);
546 rq->cmd_flags |= REQ_SOFTBARRIER;
547 list_add(&rq->queuelist, &q->requeue_list);
549 list_add_tail(&rq->queuelist, &q->requeue_list);
551 spin_unlock_irqrestore(&q->requeue_lock, flags);
553 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
555 void blk_mq_cancel_requeue_work(struct request_queue *q)
557 cancel_delayed_work_sync(&q->requeue_work);
559 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
561 void blk_mq_kick_requeue_list(struct request_queue *q)
563 kblockd_schedule_delayed_work(&q->requeue_work, 0);
565 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
567 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
570 kblockd_schedule_delayed_work(&q->requeue_work,
571 msecs_to_jiffies(msecs));
573 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
575 void blk_mq_abort_requeue_list(struct request_queue *q)
580 spin_lock_irqsave(&q->requeue_lock, flags);
581 list_splice_init(&q->requeue_list, &rq_list);
582 spin_unlock_irqrestore(&q->requeue_lock, flags);
584 while (!list_empty(&rq_list)) {
587 rq = list_first_entry(&rq_list, struct request, queuelist);
588 list_del_init(&rq->queuelist);
590 blk_mq_end_request(rq, rq->errors);
593 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
595 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
597 if (tag < tags->nr_tags) {
598 prefetch(tags->rqs[tag]);
599 return tags->rqs[tag];
604 EXPORT_SYMBOL(blk_mq_tag_to_rq);
606 struct blk_mq_timeout_data {
608 unsigned int next_set;
611 void blk_mq_rq_timed_out(struct request *req, bool reserved)
613 struct blk_mq_ops *ops = req->q->mq_ops;
614 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
617 * We know that complete is set at this point. If STARTED isn't set
618 * anymore, then the request isn't active and the "timeout" should
619 * just be ignored. This can happen due to the bitflag ordering.
620 * Timeout first checks if STARTED is set, and if it is, assumes
621 * the request is active. But if we race with completion, then
622 * we both flags will get cleared. So check here again, and ignore
623 * a timeout event with a request that isn't active.
625 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
629 ret = ops->timeout(req, reserved);
633 __blk_mq_complete_request(req);
635 case BLK_EH_RESET_TIMER:
637 blk_clear_rq_complete(req);
639 case BLK_EH_NOT_HANDLED:
642 printk(KERN_ERR "block: bad eh return: %d\n", ret);
647 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
648 struct request *rq, void *priv, bool reserved)
650 struct blk_mq_timeout_data *data = priv;
652 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
654 * If a request wasn't started before the queue was
655 * marked dying, kill it here or it'll go unnoticed.
657 if (unlikely(blk_queue_dying(rq->q))) {
659 blk_mq_end_request(rq, rq->errors);
664 if (time_after_eq(jiffies, rq->deadline)) {
665 if (!blk_mark_rq_complete(rq))
666 blk_mq_rq_timed_out(rq, reserved);
667 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
668 data->next = rq->deadline;
673 static void blk_mq_timeout_work(struct work_struct *work)
675 struct request_queue *q =
676 container_of(work, struct request_queue, timeout_work);
677 struct blk_mq_timeout_data data = {
683 /* A deadlock might occur if a request is stuck requiring a
684 * timeout at the same time a queue freeze is waiting
685 * completion, since the timeout code would not be able to
686 * acquire the queue reference here.
688 * That's why we don't use blk_queue_enter here; instead, we use
689 * percpu_ref_tryget directly, because we need to be able to
690 * obtain a reference even in the short window between the queue
691 * starting to freeze, by dropping the first reference in
692 * blk_mq_freeze_queue_start, and the moment the last request is
693 * consumed, marked by the instant q_usage_counter reaches
696 if (!percpu_ref_tryget(&q->q_usage_counter))
699 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
702 data.next = blk_rq_timeout(round_jiffies_up(data.next));
703 mod_timer(&q->timeout, data.next);
705 struct blk_mq_hw_ctx *hctx;
707 queue_for_each_hw_ctx(q, hctx, i) {
708 /* the hctx may be unmapped, so check it here */
709 if (blk_mq_hw_queue_mapped(hctx))
710 blk_mq_tag_idle(hctx);
717 * Reverse check our software queue for entries that we could potentially
718 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
719 * too much time checking for merges.
721 static bool blk_mq_attempt_merge(struct request_queue *q,
722 struct blk_mq_ctx *ctx, struct bio *bio)
727 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
733 if (!blk_rq_merge_ok(rq, bio))
736 el_ret = blk_try_merge(rq, bio);
737 if (el_ret == ELEVATOR_BACK_MERGE) {
738 if (bio_attempt_back_merge(q, rq, bio)) {
743 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
744 if (bio_attempt_front_merge(q, rq, bio)) {
755 struct flush_busy_ctx_data {
756 struct blk_mq_hw_ctx *hctx;
757 struct list_head *list;
760 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
762 struct flush_busy_ctx_data *flush_data = data;
763 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
764 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
766 sbitmap_clear_bit(sb, bitnr);
767 spin_lock(&ctx->lock);
768 list_splice_tail_init(&ctx->rq_list, flush_data->list);
769 spin_unlock(&ctx->lock);
774 * Process software queues that have been marked busy, splicing them
775 * to the for-dispatch
777 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
779 struct flush_busy_ctx_data data = {
784 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
787 static inline unsigned int queued_to_index(unsigned int queued)
792 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
796 * Run this hardware queue, pulling any software queues mapped to it in.
797 * Note that this function currently has various problems around ordering
798 * of IO. In particular, we'd like FIFO behaviour on handling existing
799 * items on the hctx->dispatch list. Ignore that for now.
801 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
803 struct request_queue *q = hctx->queue;
806 LIST_HEAD(driver_list);
807 struct list_head *dptr;
810 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
813 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
814 cpu_online(hctx->next_cpu));
819 * Touch any software queue that has pending entries.
821 flush_busy_ctxs(hctx, &rq_list);
824 * If we have previous entries on our dispatch list, grab them
825 * and stuff them at the front for more fair dispatch.
827 if (!list_empty_careful(&hctx->dispatch)) {
828 spin_lock(&hctx->lock);
829 if (!list_empty(&hctx->dispatch))
830 list_splice_init(&hctx->dispatch, &rq_list);
831 spin_unlock(&hctx->lock);
835 * Start off with dptr being NULL, so we start the first request
836 * immediately, even if we have more pending.
841 * Now process all the entries, sending them to the driver.
844 while (!list_empty(&rq_list)) {
845 struct blk_mq_queue_data bd;
848 rq = list_first_entry(&rq_list, struct request, queuelist);
849 list_del_init(&rq->queuelist);
853 bd.last = list_empty(&rq_list);
855 ret = q->mq_ops->queue_rq(hctx, &bd);
857 case BLK_MQ_RQ_QUEUE_OK:
860 case BLK_MQ_RQ_QUEUE_BUSY:
861 list_add(&rq->queuelist, &rq_list);
862 __blk_mq_requeue_request(rq);
865 pr_err("blk-mq: bad return on queue: %d\n", ret);
866 case BLK_MQ_RQ_QUEUE_ERROR:
868 blk_mq_end_request(rq, rq->errors);
872 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
876 * We've done the first request. If we have more than 1
877 * left in the list, set dptr to defer issue.
879 if (!dptr && rq_list.next != rq_list.prev)
883 hctx->dispatched[queued_to_index(queued)]++;
886 * Any items that need requeuing? Stuff them into hctx->dispatch,
887 * that is where we will continue on next queue run.
889 if (!list_empty(&rq_list)) {
890 spin_lock(&hctx->lock);
891 list_splice(&rq_list, &hctx->dispatch);
892 spin_unlock(&hctx->lock);
894 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
895 * it's possible the queue is stopped and restarted again
896 * before this. Queue restart will dispatch requests. And since
897 * requests in rq_list aren't added into hctx->dispatch yet,
898 * the requests in rq_list might get lost.
900 * blk_mq_run_hw_queue() already checks the STOPPED bit
902 blk_mq_run_hw_queue(hctx, true);
907 * It'd be great if the workqueue API had a way to pass
908 * in a mask and had some smarts for more clever placement.
909 * For now we just round-robin here, switching for every
910 * BLK_MQ_CPU_WORK_BATCH queued items.
912 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
914 if (hctx->queue->nr_hw_queues == 1)
915 return WORK_CPU_UNBOUND;
917 if (--hctx->next_cpu_batch <= 0) {
918 int cpu = hctx->next_cpu, next_cpu;
920 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
921 if (next_cpu >= nr_cpu_ids)
922 next_cpu = cpumask_first(hctx->cpumask);
924 hctx->next_cpu = next_cpu;
925 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
930 return hctx->next_cpu;
933 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
935 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
936 !blk_mq_hw_queue_mapped(hctx)))
939 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
940 int cpu = get_cpu_light();
941 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
942 __blk_mq_run_hw_queue(hctx);
950 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
953 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
955 struct blk_mq_hw_ctx *hctx;
958 queue_for_each_hw_ctx(q, hctx, i) {
959 if ((!blk_mq_hctx_has_pending(hctx) &&
960 list_empty_careful(&hctx->dispatch)) ||
961 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
964 blk_mq_run_hw_queue(hctx, async);
967 EXPORT_SYMBOL(blk_mq_run_hw_queues);
969 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
971 cancel_work(&hctx->run_work);
972 cancel_delayed_work(&hctx->delay_work);
973 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
975 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
977 void blk_mq_stop_hw_queues(struct request_queue *q)
979 struct blk_mq_hw_ctx *hctx;
982 queue_for_each_hw_ctx(q, hctx, i)
983 blk_mq_stop_hw_queue(hctx);
985 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
987 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
989 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
991 blk_mq_run_hw_queue(hctx, false);
993 EXPORT_SYMBOL(blk_mq_start_hw_queue);
995 void blk_mq_start_hw_queues(struct request_queue *q)
997 struct blk_mq_hw_ctx *hctx;
1000 queue_for_each_hw_ctx(q, hctx, i)
1001 blk_mq_start_hw_queue(hctx);
1003 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1005 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1007 struct blk_mq_hw_ctx *hctx;
1010 queue_for_each_hw_ctx(q, hctx, i) {
1011 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1014 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1015 blk_mq_run_hw_queue(hctx, async);
1018 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1020 static void blk_mq_run_work_fn(struct work_struct *work)
1022 struct blk_mq_hw_ctx *hctx;
1024 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1026 __blk_mq_run_hw_queue(hctx);
1029 static void blk_mq_delay_work_fn(struct work_struct *work)
1031 struct blk_mq_hw_ctx *hctx;
1033 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1035 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1036 __blk_mq_run_hw_queue(hctx);
1039 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1041 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1044 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1045 &hctx->delay_work, msecs_to_jiffies(msecs));
1047 EXPORT_SYMBOL(blk_mq_delay_queue);
1049 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1053 struct blk_mq_ctx *ctx = rq->mq_ctx;
1055 trace_block_rq_insert(hctx->queue, rq);
1058 list_add(&rq->queuelist, &ctx->rq_list);
1060 list_add_tail(&rq->queuelist, &ctx->rq_list);
1063 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1064 struct request *rq, bool at_head)
1066 struct blk_mq_ctx *ctx = rq->mq_ctx;
1068 __blk_mq_insert_req_list(hctx, rq, at_head);
1069 blk_mq_hctx_mark_pending(hctx, ctx);
1072 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1075 struct blk_mq_ctx *ctx = rq->mq_ctx;
1076 struct request_queue *q = rq->q;
1077 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
1079 spin_lock(&ctx->lock);
1080 __blk_mq_insert_request(hctx, rq, at_head);
1081 spin_unlock(&ctx->lock);
1084 blk_mq_run_hw_queue(hctx, async);
1087 static void blk_mq_insert_requests(struct request_queue *q,
1088 struct blk_mq_ctx *ctx,
1089 struct list_head *list,
1094 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
1096 trace_block_unplug(q, depth, !from_schedule);
1099 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1102 spin_lock(&ctx->lock);
1103 while (!list_empty(list)) {
1106 rq = list_first_entry(list, struct request, queuelist);
1107 BUG_ON(rq->mq_ctx != ctx);
1108 list_del_init(&rq->queuelist);
1109 __blk_mq_insert_req_list(hctx, rq, false);
1111 blk_mq_hctx_mark_pending(hctx, ctx);
1112 spin_unlock(&ctx->lock);
1114 blk_mq_run_hw_queue(hctx, from_schedule);
1117 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1119 struct request *rqa = container_of(a, struct request, queuelist);
1120 struct request *rqb = container_of(b, struct request, queuelist);
1122 return !(rqa->mq_ctx < rqb->mq_ctx ||
1123 (rqa->mq_ctx == rqb->mq_ctx &&
1124 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1127 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1129 struct blk_mq_ctx *this_ctx;
1130 struct request_queue *this_q;
1133 LIST_HEAD(ctx_list);
1136 list_splice_init(&plug->mq_list, &list);
1138 list_sort(NULL, &list, plug_ctx_cmp);
1144 while (!list_empty(&list)) {
1145 rq = list_entry_rq(list.next);
1146 list_del_init(&rq->queuelist);
1148 if (rq->mq_ctx != this_ctx) {
1150 blk_mq_insert_requests(this_q, this_ctx,
1155 this_ctx = rq->mq_ctx;
1161 list_add_tail(&rq->queuelist, &ctx_list);
1165 * If 'this_ctx' is set, we know we have entries to complete
1166 * on 'ctx_list'. Do those.
1169 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1174 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1176 init_request_from_bio(rq, bio);
1178 blk_account_io_start(rq, 1);
1181 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1183 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1184 !blk_queue_nomerges(hctx->queue);
1187 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1188 struct blk_mq_ctx *ctx,
1189 struct request *rq, struct bio *bio)
1191 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1192 blk_mq_bio_to_request(rq, bio);
1193 spin_lock(&ctx->lock);
1195 __blk_mq_insert_request(hctx, rq, false);
1196 spin_unlock(&ctx->lock);
1199 struct request_queue *q = hctx->queue;
1201 spin_lock(&ctx->lock);
1202 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1203 blk_mq_bio_to_request(rq, bio);
1207 spin_unlock(&ctx->lock);
1208 __blk_mq_free_request(hctx, ctx, rq);
1213 struct blk_map_ctx {
1214 struct blk_mq_hw_ctx *hctx;
1215 struct blk_mq_ctx *ctx;
1218 static struct request *blk_mq_map_request(struct request_queue *q,
1220 struct blk_map_ctx *data)
1222 struct blk_mq_hw_ctx *hctx;
1223 struct blk_mq_ctx *ctx;
1225 int op = bio_data_dir(bio);
1227 struct blk_mq_alloc_data alloc_data;
1229 blk_queue_enter_live(q);
1230 ctx = blk_mq_get_ctx(q);
1231 hctx = blk_mq_map_queue(q, ctx->cpu);
1233 if (rw_is_sync(bio_op(bio), bio->bi_opf))
1234 op_flags |= REQ_SYNC;
1236 trace_block_getrq(q, bio, op);
1237 blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1238 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1240 data->hctx = alloc_data.hctx;
1241 data->ctx = alloc_data.ctx;
1242 data->hctx->queued++;
1246 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1249 struct request_queue *q = rq->q;
1250 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
1251 struct blk_mq_queue_data bd = {
1256 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1259 * For OK queue, we are done. For error, kill it. Any other
1260 * error (busy), just add it to our list as we previously
1263 ret = q->mq_ops->queue_rq(hctx, &bd);
1264 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1265 *cookie = new_cookie;
1269 __blk_mq_requeue_request(rq);
1271 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1272 *cookie = BLK_QC_T_NONE;
1274 blk_mq_end_request(rq, rq->errors);
1282 * Multiple hardware queue variant. This will not use per-process plugs,
1283 * but will attempt to bypass the hctx queueing if we can go straight to
1284 * hardware for SYNC IO.
1286 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1288 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1289 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1290 struct blk_map_ctx data;
1292 unsigned int request_count = 0;
1293 struct blk_plug *plug;
1294 struct request *same_queue_rq = NULL;
1297 blk_queue_bounce(q, &bio);
1299 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1301 return BLK_QC_T_NONE;
1304 blk_queue_split(q, &bio, q->bio_split);
1306 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1307 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1308 return BLK_QC_T_NONE;
1310 rq = blk_mq_map_request(q, bio, &data);
1312 return BLK_QC_T_NONE;
1314 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1316 if (unlikely(is_flush_fua)) {
1317 blk_mq_bio_to_request(rq, bio);
1318 blk_insert_flush(rq);
1322 plug = current->plug;
1324 * If the driver supports defer issued based on 'last', then
1325 * queue it up like normal since we can potentially save some
1328 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1329 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1330 struct request *old_rq = NULL;
1332 blk_mq_bio_to_request(rq, bio);
1335 * We do limited pluging. If the bio can be merged, do that.
1336 * Otherwise the existing request in the plug list will be
1337 * issued. So the plug list will have one request at most
1341 * The plug list might get flushed before this. If that
1342 * happens, same_queue_rq is invalid and plug list is
1345 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1346 old_rq = same_queue_rq;
1347 list_del_init(&old_rq->queuelist);
1349 list_add_tail(&rq->queuelist, &plug->mq_list);
1350 } else /* is_sync */
1352 blk_mq_put_ctx(data.ctx);
1355 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1357 blk_mq_insert_request(old_rq, false, true, true);
1361 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1363 * For a SYNC request, send it to the hardware immediately. For
1364 * an ASYNC request, just ensure that we run it later on. The
1365 * latter allows for merging opportunities and more efficient
1369 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1371 blk_mq_put_ctx(data.ctx);
1377 * Single hardware queue variant. This will attempt to use any per-process
1378 * plug for merging and IO deferral.
1380 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1382 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1383 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1384 struct blk_plug *plug;
1385 unsigned int request_count = 0;
1386 struct blk_map_ctx data;
1390 blk_queue_bounce(q, &bio);
1392 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1394 return BLK_QC_T_NONE;
1397 blk_queue_split(q, &bio, q->bio_split);
1399 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1400 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1401 return BLK_QC_T_NONE;
1403 request_count = blk_plug_queued_count(q);
1405 rq = blk_mq_map_request(q, bio, &data);
1407 return BLK_QC_T_NONE;
1409 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1411 if (unlikely(is_flush_fua)) {
1412 blk_mq_bio_to_request(rq, bio);
1413 blk_insert_flush(rq);
1418 * A task plug currently exists. Since this is completely lockless,
1419 * utilize that to temporarily store requests until the task is
1420 * either done or scheduled away.
1422 plug = current->plug;
1424 blk_mq_bio_to_request(rq, bio);
1426 trace_block_plug(q);
1428 blk_mq_put_ctx(data.ctx);
1430 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1431 blk_flush_plug_list(plug, false);
1432 trace_block_plug(q);
1435 list_add_tail(&rq->queuelist, &plug->mq_list);
1439 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1441 * For a SYNC request, send it to the hardware immediately. For
1442 * an ASYNC request, just ensure that we run it later on. The
1443 * latter allows for merging opportunities and more efficient
1447 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1450 blk_mq_put_ctx(data.ctx);
1454 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1455 struct blk_mq_tags *tags, unsigned int hctx_idx)
1459 if (tags->rqs && set->ops->exit_request) {
1462 for (i = 0; i < tags->nr_tags; i++) {
1465 set->ops->exit_request(set->driver_data, tags->rqs[i],
1467 tags->rqs[i] = NULL;
1471 while (!list_empty(&tags->page_list)) {
1472 page = list_first_entry(&tags->page_list, struct page, lru);
1473 list_del_init(&page->lru);
1475 * Remove kmemleak object previously allocated in
1476 * blk_mq_init_rq_map().
1478 kmemleak_free(page_address(page));
1479 __free_pages(page, page->private);
1484 blk_mq_free_tags(tags);
1487 static size_t order_to_size(unsigned int order)
1489 return (size_t)PAGE_SIZE << order;
1492 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1493 unsigned int hctx_idx)
1495 struct blk_mq_tags *tags;
1496 unsigned int i, j, entries_per_page, max_order = 4;
1497 size_t rq_size, left;
1499 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1501 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1505 INIT_LIST_HEAD(&tags->page_list);
1507 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1508 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1511 blk_mq_free_tags(tags);
1516 * rq_size is the size of the request plus driver payload, rounded
1517 * to the cacheline size
1519 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1521 left = rq_size * set->queue_depth;
1523 for (i = 0; i < set->queue_depth; ) {
1524 int this_order = max_order;
1529 while (this_order && left < order_to_size(this_order - 1))
1533 page = alloc_pages_node(set->numa_node,
1534 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1540 if (order_to_size(this_order) < rq_size)
1547 page->private = this_order;
1548 list_add_tail(&page->lru, &tags->page_list);
1550 p = page_address(page);
1552 * Allow kmemleak to scan these pages as they contain pointers
1553 * to additional allocations like via ops->init_request().
1555 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1556 entries_per_page = order_to_size(this_order) / rq_size;
1557 to_do = min(entries_per_page, set->queue_depth - i);
1558 left -= to_do * rq_size;
1559 for (j = 0; j < to_do; j++) {
1561 if (set->ops->init_request) {
1562 if (set->ops->init_request(set->driver_data,
1563 tags->rqs[i], hctx_idx, i,
1565 tags->rqs[i] = NULL;
1577 blk_mq_free_rq_map(set, tags, hctx_idx);
1582 * 'cpu' is going away. splice any existing rq_list entries from this
1583 * software queue to the hw queue dispatch list, and ensure that it
1586 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1588 struct blk_mq_hw_ctx *hctx;
1589 struct blk_mq_ctx *ctx;
1592 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1593 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1595 spin_lock(&ctx->lock);
1596 if (!list_empty(&ctx->rq_list)) {
1597 list_splice_init(&ctx->rq_list, &tmp);
1598 blk_mq_hctx_clear_pending(hctx, ctx);
1600 spin_unlock(&ctx->lock);
1602 if (list_empty(&tmp))
1605 spin_lock(&hctx->lock);
1606 list_splice_tail_init(&tmp, &hctx->dispatch);
1607 spin_unlock(&hctx->lock);
1609 blk_mq_run_hw_queue(hctx, true);
1613 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1615 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1619 /* hctx->ctxs will be freed in queue's release handler */
1620 static void blk_mq_exit_hctx(struct request_queue *q,
1621 struct blk_mq_tag_set *set,
1622 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1624 unsigned flush_start_tag = set->queue_depth;
1626 blk_mq_tag_idle(hctx);
1628 if (set->ops->exit_request)
1629 set->ops->exit_request(set->driver_data,
1630 hctx->fq->flush_rq, hctx_idx,
1631 flush_start_tag + hctx_idx);
1633 if (set->ops->exit_hctx)
1634 set->ops->exit_hctx(hctx, hctx_idx);
1636 blk_mq_remove_cpuhp(hctx);
1637 blk_free_flush_queue(hctx->fq);
1638 sbitmap_free(&hctx->ctx_map);
1641 static void blk_mq_exit_hw_queues(struct request_queue *q,
1642 struct blk_mq_tag_set *set, int nr_queue)
1644 struct blk_mq_hw_ctx *hctx;
1647 queue_for_each_hw_ctx(q, hctx, i) {
1650 blk_mq_exit_hctx(q, set, hctx, i);
1654 static void blk_mq_free_hw_queues(struct request_queue *q,
1655 struct blk_mq_tag_set *set)
1657 struct blk_mq_hw_ctx *hctx;
1660 queue_for_each_hw_ctx(q, hctx, i)
1661 free_cpumask_var(hctx->cpumask);
1664 static int blk_mq_init_hctx(struct request_queue *q,
1665 struct blk_mq_tag_set *set,
1666 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1669 unsigned flush_start_tag = set->queue_depth;
1671 node = hctx->numa_node;
1672 if (node == NUMA_NO_NODE)
1673 node = hctx->numa_node = set->numa_node;
1675 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1676 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1677 spin_lock_init(&hctx->lock);
1678 INIT_LIST_HEAD(&hctx->dispatch);
1680 hctx->queue_num = hctx_idx;
1681 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1683 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1685 hctx->tags = set->tags[hctx_idx];
1688 * Allocate space for all possible cpus to avoid allocation at
1691 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1694 goto unregister_cpu_notifier;
1696 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1702 if (set->ops->init_hctx &&
1703 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1706 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1710 if (set->ops->init_request &&
1711 set->ops->init_request(set->driver_data,
1712 hctx->fq->flush_rq, hctx_idx,
1713 flush_start_tag + hctx_idx, node))
1721 if (set->ops->exit_hctx)
1722 set->ops->exit_hctx(hctx, hctx_idx);
1724 sbitmap_free(&hctx->ctx_map);
1727 unregister_cpu_notifier:
1728 blk_mq_remove_cpuhp(hctx);
1732 static void blk_mq_init_cpu_queues(struct request_queue *q,
1733 unsigned int nr_hw_queues)
1737 for_each_possible_cpu(i) {
1738 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1739 struct blk_mq_hw_ctx *hctx;
1741 memset(__ctx, 0, sizeof(*__ctx));
1743 spin_lock_init(&__ctx->lock);
1744 INIT_LIST_HEAD(&__ctx->rq_list);
1747 /* If the cpu isn't online, the cpu is mapped to first hctx */
1751 hctx = blk_mq_map_queue(q, i);
1754 * Set local node, IFF we have more than one hw queue. If
1755 * not, we remain on the home node of the device
1757 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1758 hctx->numa_node = local_memory_node(cpu_to_node(i));
1762 static void blk_mq_map_swqueue(struct request_queue *q,
1763 const struct cpumask *online_mask)
1766 struct blk_mq_hw_ctx *hctx;
1767 struct blk_mq_ctx *ctx;
1768 struct blk_mq_tag_set *set = q->tag_set;
1771 * Avoid others reading imcomplete hctx->cpumask through sysfs
1773 mutex_lock(&q->sysfs_lock);
1775 queue_for_each_hw_ctx(q, hctx, i) {
1776 cpumask_clear(hctx->cpumask);
1781 * Map software to hardware queues
1783 for_each_possible_cpu(i) {
1784 /* If the cpu isn't online, the cpu is mapped to first hctx */
1785 if (!cpumask_test_cpu(i, online_mask))
1788 ctx = per_cpu_ptr(q->queue_ctx, i);
1789 hctx = blk_mq_map_queue(q, i);
1791 cpumask_set_cpu(i, hctx->cpumask);
1792 ctx->index_hw = hctx->nr_ctx;
1793 hctx->ctxs[hctx->nr_ctx++] = ctx;
1796 mutex_unlock(&q->sysfs_lock);
1798 queue_for_each_hw_ctx(q, hctx, i) {
1800 * If no software queues are mapped to this hardware queue,
1801 * disable it and free the request entries.
1803 if (!hctx->nr_ctx) {
1805 blk_mq_free_rq_map(set, set->tags[i], i);
1806 set->tags[i] = NULL;
1812 /* unmapped hw queue can be remapped after CPU topo changed */
1814 set->tags[i] = blk_mq_init_rq_map(set, i);
1815 hctx->tags = set->tags[i];
1816 WARN_ON(!hctx->tags);
1819 * Set the map size to the number of mapped software queues.
1820 * This is more accurate and more efficient than looping
1821 * over all possibly mapped software queues.
1823 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
1826 * Initialize batch roundrobin counts
1828 hctx->next_cpu = cpumask_first(hctx->cpumask);
1829 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1833 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1835 struct blk_mq_hw_ctx *hctx;
1838 queue_for_each_hw_ctx(q, hctx, i) {
1840 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1842 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1846 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1848 struct request_queue *q;
1850 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1851 blk_mq_freeze_queue(q);
1852 queue_set_hctx_shared(q, shared);
1853 blk_mq_unfreeze_queue(q);
1857 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1859 struct blk_mq_tag_set *set = q->tag_set;
1861 mutex_lock(&set->tag_list_lock);
1862 list_del_init(&q->tag_set_list);
1863 if (list_is_singular(&set->tag_list)) {
1864 /* just transitioned to unshared */
1865 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1866 /* update existing queue */
1867 blk_mq_update_tag_set_depth(set, false);
1869 mutex_unlock(&set->tag_list_lock);
1872 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1873 struct request_queue *q)
1877 mutex_lock(&set->tag_list_lock);
1879 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1880 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1881 set->flags |= BLK_MQ_F_TAG_SHARED;
1882 /* update existing queue */
1883 blk_mq_update_tag_set_depth(set, true);
1885 if (set->flags & BLK_MQ_F_TAG_SHARED)
1886 queue_set_hctx_shared(q, true);
1887 list_add_tail(&q->tag_set_list, &set->tag_list);
1889 mutex_unlock(&set->tag_list_lock);
1893 * It is the actual release handler for mq, but we do it from
1894 * request queue's release handler for avoiding use-after-free
1895 * and headache because q->mq_kobj shouldn't have been introduced,
1896 * but we can't group ctx/kctx kobj without it.
1898 void blk_mq_release(struct request_queue *q)
1900 struct blk_mq_hw_ctx *hctx;
1903 /* hctx kobj stays in hctx */
1904 queue_for_each_hw_ctx(q, hctx, i) {
1913 kfree(q->queue_hw_ctx);
1915 /* ctx kobj stays in queue_ctx */
1916 free_percpu(q->queue_ctx);
1919 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1921 struct request_queue *uninit_q, *q;
1923 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1925 return ERR_PTR(-ENOMEM);
1927 q = blk_mq_init_allocated_queue(set, uninit_q);
1929 blk_cleanup_queue(uninit_q);
1933 EXPORT_SYMBOL(blk_mq_init_queue);
1935 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1936 struct request_queue *q)
1939 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
1941 blk_mq_sysfs_unregister(q);
1942 for (i = 0; i < set->nr_hw_queues; i++) {
1948 node = blk_mq_hw_queue_to_node(q->mq_map, i);
1949 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1954 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1961 atomic_set(&hctxs[i]->nr_active, 0);
1962 hctxs[i]->numa_node = node;
1963 hctxs[i]->queue_num = i;
1965 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
1966 free_cpumask_var(hctxs[i]->cpumask);
1971 blk_mq_hctx_kobj_init(hctxs[i]);
1973 for (j = i; j < q->nr_hw_queues; j++) {
1974 struct blk_mq_hw_ctx *hctx = hctxs[j];
1978 blk_mq_free_rq_map(set, hctx->tags, j);
1979 set->tags[j] = NULL;
1981 blk_mq_exit_hctx(q, set, hctx, j);
1982 free_cpumask_var(hctx->cpumask);
1983 kobject_put(&hctx->kobj);
1990 q->nr_hw_queues = i;
1991 blk_mq_sysfs_register(q);
1994 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1995 struct request_queue *q)
1997 /* mark the queue as mq asap */
1998 q->mq_ops = set->ops;
2000 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2004 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2005 GFP_KERNEL, set->numa_node);
2006 if (!q->queue_hw_ctx)
2009 q->mq_map = set->mq_map;
2011 blk_mq_realloc_hw_ctxs(set, q);
2012 if (!q->nr_hw_queues)
2015 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2016 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2018 q->nr_queues = nr_cpu_ids;
2020 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2022 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2023 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2025 q->sg_reserved_size = INT_MAX;
2027 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2028 INIT_LIST_HEAD(&q->requeue_list);
2029 spin_lock_init(&q->requeue_lock);
2031 if (q->nr_hw_queues > 1)
2032 blk_queue_make_request(q, blk_mq_make_request);
2034 blk_queue_make_request(q, blk_sq_make_request);
2037 * Do this after blk_queue_make_request() overrides it...
2039 q->nr_requests = set->queue_depth;
2041 if (set->ops->complete)
2042 blk_queue_softirq_done(q, set->ops->complete);
2044 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2047 mutex_lock(&all_q_mutex);
2049 list_add_tail(&q->all_q_node, &all_q_list);
2050 blk_mq_add_queue_tag_set(set, q);
2051 blk_mq_map_swqueue(q, cpu_online_mask);
2053 mutex_unlock(&all_q_mutex);
2059 kfree(q->queue_hw_ctx);
2061 free_percpu(q->queue_ctx);
2064 return ERR_PTR(-ENOMEM);
2066 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2068 void blk_mq_free_queue(struct request_queue *q)
2070 struct blk_mq_tag_set *set = q->tag_set;
2072 mutex_lock(&all_q_mutex);
2073 list_del_init(&q->all_q_node);
2074 mutex_unlock(&all_q_mutex);
2076 blk_mq_del_queue_tag_set(q);
2078 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2079 blk_mq_free_hw_queues(q, set);
2082 /* Basically redo blk_mq_init_queue with queue frozen */
2083 static void blk_mq_queue_reinit(struct request_queue *q,
2084 const struct cpumask *online_mask)
2086 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2088 blk_mq_sysfs_unregister(q);
2091 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2092 * we should change hctx numa_node according to new topology (this
2093 * involves free and re-allocate memory, worthy doing?)
2096 blk_mq_map_swqueue(q, online_mask);
2098 blk_mq_sysfs_register(q);
2102 * New online cpumask which is going to be set in this hotplug event.
2103 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2104 * one-by-one and dynamically allocating this could result in a failure.
2106 static struct cpumask cpuhp_online_new;
2108 static void blk_mq_queue_reinit_work(void)
2110 struct request_queue *q;
2112 mutex_lock(&all_q_mutex);
2114 * We need to freeze and reinit all existing queues. Freezing
2115 * involves synchronous wait for an RCU grace period and doing it
2116 * one by one may take a long time. Start freezing all queues in
2117 * one swoop and then wait for the completions so that freezing can
2118 * take place in parallel.
2120 list_for_each_entry(q, &all_q_list, all_q_node)
2121 blk_mq_freeze_queue_start(q);
2122 list_for_each_entry(q, &all_q_list, all_q_node) {
2123 blk_mq_freeze_queue_wait(q);
2126 * timeout handler can't touch hw queue during the
2129 del_timer_sync(&q->timeout);
2132 list_for_each_entry(q, &all_q_list, all_q_node)
2133 blk_mq_queue_reinit(q, &cpuhp_online_new);
2135 list_for_each_entry(q, &all_q_list, all_q_node)
2136 blk_mq_unfreeze_queue(q);
2138 mutex_unlock(&all_q_mutex);
2141 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2143 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2144 blk_mq_queue_reinit_work();
2149 * Before hotadded cpu starts handling requests, new mappings must be
2150 * established. Otherwise, these requests in hw queue might never be
2153 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2154 * for CPU0, and ctx1 for CPU1).
2156 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2157 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2159 * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in
2160 * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2161 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list
2164 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2166 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2167 cpumask_set_cpu(cpu, &cpuhp_online_new);
2168 blk_mq_queue_reinit_work();
2172 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2176 for (i = 0; i < set->nr_hw_queues; i++) {
2177 set->tags[i] = blk_mq_init_rq_map(set, i);
2186 blk_mq_free_rq_map(set, set->tags[i], i);
2192 * Allocate the request maps associated with this tag_set. Note that this
2193 * may reduce the depth asked for, if memory is tight. set->queue_depth
2194 * will be updated to reflect the allocated depth.
2196 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2201 depth = set->queue_depth;
2203 err = __blk_mq_alloc_rq_maps(set);
2207 set->queue_depth >>= 1;
2208 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2212 } while (set->queue_depth);
2214 if (!set->queue_depth || err) {
2215 pr_err("blk-mq: failed to allocate request map\n");
2219 if (depth != set->queue_depth)
2220 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2221 depth, set->queue_depth);
2227 * Alloc a tag set to be associated with one or more request queues.
2228 * May fail with EINVAL for various error conditions. May adjust the
2229 * requested depth down, if if it too large. In that case, the set
2230 * value will be stored in set->queue_depth.
2232 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2236 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2238 if (!set->nr_hw_queues)
2240 if (!set->queue_depth)
2242 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2245 if (!set->ops->queue_rq)
2248 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2249 pr_info("blk-mq: reduced tag depth to %u\n",
2251 set->queue_depth = BLK_MQ_MAX_DEPTH;
2255 * If a crashdump is active, then we are potentially in a very
2256 * memory constrained environment. Limit us to 1 queue and
2257 * 64 tags to prevent using too much memory.
2259 if (is_kdump_kernel()) {
2260 set->nr_hw_queues = 1;
2261 set->queue_depth = min(64U, set->queue_depth);
2264 * There is no use for more h/w queues than cpus.
2266 if (set->nr_hw_queues > nr_cpu_ids)
2267 set->nr_hw_queues = nr_cpu_ids;
2269 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2270 GFP_KERNEL, set->numa_node);
2275 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2276 GFP_KERNEL, set->numa_node);
2280 if (set->ops->map_queues)
2281 ret = set->ops->map_queues(set);
2283 ret = blk_mq_map_queues(set);
2285 goto out_free_mq_map;
2287 ret = blk_mq_alloc_rq_maps(set);
2289 goto out_free_mq_map;
2291 mutex_init(&set->tag_list_lock);
2292 INIT_LIST_HEAD(&set->tag_list);
2304 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2306 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2310 for (i = 0; i < nr_cpu_ids; i++) {
2312 blk_mq_free_rq_map(set, set->tags[i], i);
2321 EXPORT_SYMBOL(blk_mq_free_tag_set);
2323 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2325 struct blk_mq_tag_set *set = q->tag_set;
2326 struct blk_mq_hw_ctx *hctx;
2329 if (!set || nr > set->queue_depth)
2333 queue_for_each_hw_ctx(q, hctx, i) {
2336 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2342 q->nr_requests = nr;
2347 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2349 struct request_queue *q;
2351 if (nr_hw_queues > nr_cpu_ids)
2352 nr_hw_queues = nr_cpu_ids;
2353 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2356 list_for_each_entry(q, &set->tag_list, tag_set_list)
2357 blk_mq_freeze_queue(q);
2359 set->nr_hw_queues = nr_hw_queues;
2360 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2361 blk_mq_realloc_hw_ctxs(set, q);
2363 if (q->nr_hw_queues > 1)
2364 blk_queue_make_request(q, blk_mq_make_request);
2366 blk_queue_make_request(q, blk_sq_make_request);
2368 blk_mq_queue_reinit(q, cpu_online_mask);
2371 list_for_each_entry(q, &set->tag_list, tag_set_list)
2372 blk_mq_unfreeze_queue(q);
2374 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2376 void blk_mq_disable_hotplug(void)
2378 mutex_lock(&all_q_mutex);
2381 void blk_mq_enable_hotplug(void)
2383 mutex_unlock(&all_q_mutex);
2386 static int __init blk_mq_init(void)
2388 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2389 blk_mq_hctx_notify_dead);
2391 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2392 blk_mq_queue_reinit_prepare,
2393 blk_mq_queue_reinit_dead);
2396 subsys_initcall(blk_mq_init);