1 #include <linux/kernel.h>
2 #include <linux/module.h>
3 #include <linux/backing-dev.h>
5 #include <linux/blkdev.h>
7 #include <linux/init.h>
8 #include <linux/slab.h>
9 #include <linux/workqueue.h>
10 #include <linux/smp.h>
11 #include <linux/llist.h>
12 #include <linux/list_sort.h>
13 #include <linux/cpu.h>
14 #include <linux/cache.h>
15 #include <linux/sched/sysctl.h>
16 #include <linux/delay.h>
18 #include <trace/events/block.h>
20 #include <linux/blk-mq.h>
23 #include "blk-mq-tag.h"
25 static DEFINE_MUTEX(all_q_mutex);
26 static LIST_HEAD(all_q_list);
28 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
30 static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
33 return per_cpu_ptr(q->queue_ctx, cpu);
37 * This assumes per-cpu software queueing queues. They could be per-node
38 * as well, for instance. For now this is hardcoded as-is. Note that we don't
39 * care about preemption, since we know the ctx's are persistent. This does
40 * mean that we can't rely on ctx always matching the currently running CPU.
42 static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
44 return __blk_mq_get_ctx(q, get_cpu());
47 static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
53 * Check if any of the ctx's have pending work in this hardware queue
55 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
59 for (i = 0; i < hctx->nr_ctx_map; i++)
67 * Mark this ctx as having pending work in this hardware queue
69 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
70 struct blk_mq_ctx *ctx)
72 if (!test_bit(ctx->index_hw, hctx->ctx_map))
73 set_bit(ctx->index_hw, hctx->ctx_map);
76 static struct request *blk_mq_alloc_rq(struct blk_mq_hw_ctx *hctx, gfp_t gfp,
82 tag = blk_mq_get_tag(hctx->tags, gfp, reserved);
83 if (tag != BLK_MQ_TAG_FAIL) {
93 static int blk_mq_queue_enter(struct request_queue *q)
97 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
99 /* we have problems to freeze the queue if it's initializing */
100 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
103 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
105 spin_lock_irq(q->queue_lock);
106 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
107 !blk_queue_bypass(q) || blk_queue_dying(q),
109 /* inc usage with lock hold to avoid freeze_queue runs here */
110 if (!ret && !blk_queue_dying(q))
111 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
112 else if (blk_queue_dying(q))
114 spin_unlock_irq(q->queue_lock);
119 static void blk_mq_queue_exit(struct request_queue *q)
121 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
124 static void __blk_mq_drain_queue(struct request_queue *q)
129 spin_lock_irq(q->queue_lock);
130 count = percpu_counter_sum(&q->mq_usage_counter);
131 spin_unlock_irq(q->queue_lock);
135 blk_mq_run_queues(q, false);
141 * Guarantee no request is in use, so we can change any data structure of
142 * the queue afterward.
144 static void blk_mq_freeze_queue(struct request_queue *q)
148 spin_lock_irq(q->queue_lock);
149 drain = !q->bypass_depth++;
150 queue_flag_set(QUEUE_FLAG_BYPASS, q);
151 spin_unlock_irq(q->queue_lock);
154 __blk_mq_drain_queue(q);
157 void blk_mq_drain_queue(struct request_queue *q)
159 __blk_mq_drain_queue(q);
162 static void blk_mq_unfreeze_queue(struct request_queue *q)
166 spin_lock_irq(q->queue_lock);
167 if (!--q->bypass_depth) {
168 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
171 WARN_ON_ONCE(q->bypass_depth < 0);
172 spin_unlock_irq(q->queue_lock);
174 wake_up_all(&q->mq_freeze_wq);
177 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
179 return blk_mq_has_free_tags(hctx->tags);
181 EXPORT_SYMBOL(blk_mq_can_queue);
183 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
184 struct request *rq, unsigned int rw_flags)
186 if (blk_queue_io_stat(q))
187 rw_flags |= REQ_IO_STAT;
190 rq->cmd_flags = rw_flags;
191 rq->start_time = jiffies;
192 set_start_time_ns(rq);
193 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
196 static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx,
197 gfp_t gfp, bool reserved,
201 bool is_flush = false;
203 * flush need allocate a request, leave at least one request for
204 * non-flush IO to avoid deadlock
206 if ((rw & REQ_FLUSH) && !(rw & REQ_FLUSH_SEQ)) {
207 if (atomic_inc_return(&hctx->pending_flush) >=
208 hctx->queue_depth - hctx->reserved_tags - 1) {
209 atomic_dec(&hctx->pending_flush);
214 req = blk_mq_alloc_rq(hctx, gfp, reserved);
215 if (!req && is_flush)
216 atomic_dec(&hctx->pending_flush);
220 static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
227 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
228 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
230 rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved, rw);
232 blk_mq_rq_ctx_init(q, ctx, rq, rw);
237 if (!(gfp & __GFP_WAIT))
240 __blk_mq_run_hw_queue(hctx);
241 blk_mq_wait_for_tags(hctx->tags);
247 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
248 gfp_t gfp, bool reserved)
252 if (blk_mq_queue_enter(q))
255 rq = blk_mq_alloc_request_pinned(q, rw, gfp, reserved);
257 blk_mq_put_ctx(rq->mq_ctx);
261 struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
266 if (blk_mq_queue_enter(q))
269 rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
271 blk_mq_put_ctx(rq->mq_ctx);
274 EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
277 * Re-init and set pdu, if we have it
279 static void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq)
281 blk_rq_init(hctx->queue, rq);
284 rq->special = blk_mq_rq_to_pdu(rq);
287 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
288 struct blk_mq_ctx *ctx, struct request *rq)
290 const int tag = rq->tag;
291 struct request_queue *q = rq->q;
293 if ((rq->cmd_flags & REQ_FLUSH) && !(rq->cmd_flags & REQ_FLUSH_SEQ))
294 atomic_dec(&hctx->pending_flush);
296 blk_mq_rq_init(hctx, rq);
297 blk_mq_put_tag(hctx->tags, tag);
299 blk_mq_queue_exit(q);
302 void blk_mq_free_request(struct request *rq)
304 struct blk_mq_ctx *ctx = rq->mq_ctx;
305 struct blk_mq_hw_ctx *hctx;
306 struct request_queue *q = rq->q;
308 ctx->rq_completed[rq_is_sync(rq)]++;
310 hctx = q->mq_ops->map_queue(q, ctx->cpu);
311 __blk_mq_free_request(hctx, ctx, rq);
314 static void blk_mq_bio_endio(struct request *rq, struct bio *bio, int error)
317 clear_bit(BIO_UPTODATE, &bio->bi_flags);
318 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
321 if (unlikely(rq->cmd_flags & REQ_QUIET))
322 set_bit(BIO_QUIET, &bio->bi_flags);
324 /* don't actually finish bio if it's part of flush sequence */
325 if (!(rq->cmd_flags & REQ_FLUSH_SEQ))
326 bio_endio(bio, error);
329 void blk_mq_complete_request(struct request *rq, int error)
331 struct bio *bio = rq->bio;
332 unsigned int bytes = 0;
334 trace_block_rq_complete(rq->q, rq);
337 struct bio *next = bio->bi_next;
340 bytes += bio->bi_iter.bi_size;
341 blk_mq_bio_endio(rq, bio, error);
345 blk_account_io_completion(rq, bytes);
347 blk_account_io_done(rq);
350 rq->end_io(rq, error);
352 blk_mq_free_request(rq);
355 void __blk_mq_end_io(struct request *rq, int error)
357 if (!blk_mark_rq_complete(rq))
358 blk_mq_complete_request(rq, error);
361 static void blk_mq_end_io_remote(void *data)
363 struct request *rq = data;
365 __blk_mq_end_io(rq, rq->errors);
369 * End IO on this request on a multiqueue enabled driver. We'll either do
370 * it directly inline, or punt to a local IPI handler on the matching
373 void blk_mq_end_io(struct request *rq, int error)
375 struct blk_mq_ctx *ctx = rq->mq_ctx;
378 if (!ctx->ipi_redirect)
379 return __blk_mq_end_io(rq, error);
382 if (cpu != ctx->cpu && cpu_online(ctx->cpu)) {
384 rq->csd.func = blk_mq_end_io_remote;
387 __smp_call_function_single(ctx->cpu, &rq->csd, 0);
389 __blk_mq_end_io(rq, error);
393 EXPORT_SYMBOL(blk_mq_end_io);
395 static void blk_mq_start_request(struct request *rq)
397 struct request_queue *q = rq->q;
399 trace_block_rq_issue(q, rq);
402 * Just mark start time and set the started bit. Due to memory
403 * ordering, we know we'll see the correct deadline as long as
404 * REQ_ATOMIC_STARTED is seen.
406 rq->deadline = jiffies + q->rq_timeout;
407 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
410 static void blk_mq_requeue_request(struct request *rq)
412 struct request_queue *q = rq->q;
414 trace_block_rq_requeue(q, rq);
415 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
418 struct blk_mq_timeout_data {
419 struct blk_mq_hw_ctx *hctx;
421 unsigned int *next_set;
424 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
426 struct blk_mq_timeout_data *data = __data;
427 struct blk_mq_hw_ctx *hctx = data->hctx;
430 /* It may not be in flight yet (this is where
431 * the REQ_ATOMIC_STARTED flag comes in). The requests are
432 * statically allocated, so we know it's always safe to access the
433 * memory associated with a bit offset into ->rqs[].
439 tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag);
440 if (tag >= hctx->queue_depth)
443 rq = hctx->rqs[tag++];
445 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
448 blk_rq_check_expired(rq, data->next, data->next_set);
452 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
454 unsigned int *next_set)
456 struct blk_mq_timeout_data data = {
459 .next_set = next_set,
463 * Ask the tagging code to iterate busy requests, so we can
464 * check them for timeout.
466 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
469 static void blk_mq_rq_timer(unsigned long data)
471 struct request_queue *q = (struct request_queue *) data;
472 struct blk_mq_hw_ctx *hctx;
473 unsigned long next = 0;
476 queue_for_each_hw_ctx(q, hctx, i)
477 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
480 mod_timer(&q->timeout, round_jiffies_up(next));
484 * Reverse check our software queue for entries that we could potentially
485 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
486 * too much time checking for merges.
488 static bool blk_mq_attempt_merge(struct request_queue *q,
489 struct blk_mq_ctx *ctx, struct bio *bio)
494 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
500 if (!blk_rq_merge_ok(rq, bio))
503 el_ret = blk_try_merge(rq, bio);
504 if (el_ret == ELEVATOR_BACK_MERGE) {
505 if (bio_attempt_back_merge(q, rq, bio)) {
510 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
511 if (bio_attempt_front_merge(q, rq, bio)) {
522 void blk_mq_add_timer(struct request *rq)
524 __blk_add_timer(rq, NULL);
528 * Run this hardware queue, pulling any software queues mapped to it in.
529 * Note that this function currently has various problems around ordering
530 * of IO. In particular, we'd like FIFO behaviour on handling existing
531 * items on the hctx->dispatch list. Ignore that for now.
533 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
535 struct request_queue *q = hctx->queue;
536 struct blk_mq_ctx *ctx;
541 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
547 * Touch any software queue that has pending entries.
549 for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
550 clear_bit(bit, hctx->ctx_map);
551 ctx = hctx->ctxs[bit];
552 BUG_ON(bit != ctx->index_hw);
554 spin_lock(&ctx->lock);
555 list_splice_tail_init(&ctx->rq_list, &rq_list);
556 spin_unlock(&ctx->lock);
560 * If we have previous entries on our dispatch list, grab them
561 * and stuff them at the front for more fair dispatch.
563 if (!list_empty_careful(&hctx->dispatch)) {
564 spin_lock(&hctx->lock);
565 if (!list_empty(&hctx->dispatch))
566 list_splice_init(&hctx->dispatch, &rq_list);
567 spin_unlock(&hctx->lock);
571 * Delete and return all entries from our dispatch list
576 * Now process all the entries, sending them to the driver.
578 while (!list_empty(&rq_list)) {
581 rq = list_first_entry(&rq_list, struct request, queuelist);
582 list_del_init(&rq->queuelist);
583 blk_mq_start_request(rq);
586 * Last request in the series. Flag it as such, this
587 * enables drivers to know when IO should be kicked off,
588 * if they don't do it on a per-request basis.
590 * Note: the flag isn't the only condition drivers
591 * should do kick off. If drive is busy, the last
592 * request might not have the bit set.
594 if (list_empty(&rq_list))
595 rq->cmd_flags |= REQ_END;
597 ret = q->mq_ops->queue_rq(hctx, rq);
599 case BLK_MQ_RQ_QUEUE_OK:
602 case BLK_MQ_RQ_QUEUE_BUSY:
604 * FIXME: we should have a mechanism to stop the queue
605 * like blk_stop_queue, otherwise we will waste cpu
608 list_add(&rq->queuelist, &rq_list);
609 blk_mq_requeue_request(rq);
612 pr_err("blk-mq: bad return on queue: %d\n", ret);
614 case BLK_MQ_RQ_QUEUE_ERROR:
615 blk_mq_end_io(rq, rq->errors);
619 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
624 hctx->dispatched[0]++;
625 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
626 hctx->dispatched[ilog2(queued) + 1]++;
629 * Any items that need requeuing? Stuff them into hctx->dispatch,
630 * that is where we will continue on next queue run.
632 if (!list_empty(&rq_list)) {
633 spin_lock(&hctx->lock);
634 list_splice(&rq_list, &hctx->dispatch);
635 spin_unlock(&hctx->lock);
639 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
641 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
645 __blk_mq_run_hw_queue(hctx);
647 struct request_queue *q = hctx->queue;
649 kblockd_schedule_delayed_work(q, &hctx->delayed_work, 0);
653 void blk_mq_run_queues(struct request_queue *q, bool async)
655 struct blk_mq_hw_ctx *hctx;
658 queue_for_each_hw_ctx(q, hctx, i) {
659 if ((!blk_mq_hctx_has_pending(hctx) &&
660 list_empty_careful(&hctx->dispatch)) ||
661 test_bit(BLK_MQ_S_STOPPED, &hctx->flags))
664 blk_mq_run_hw_queue(hctx, async);
667 EXPORT_SYMBOL(blk_mq_run_queues);
669 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
671 cancel_delayed_work(&hctx->delayed_work);
672 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
674 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
676 void blk_mq_stop_hw_queues(struct request_queue *q)
678 struct blk_mq_hw_ctx *hctx;
681 queue_for_each_hw_ctx(q, hctx, i)
682 blk_mq_stop_hw_queue(hctx);
684 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
686 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
688 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
689 __blk_mq_run_hw_queue(hctx);
691 EXPORT_SYMBOL(blk_mq_start_hw_queue);
693 void blk_mq_start_stopped_hw_queues(struct request_queue *q)
695 struct blk_mq_hw_ctx *hctx;
698 queue_for_each_hw_ctx(q, hctx, i) {
699 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
702 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
703 blk_mq_run_hw_queue(hctx, true);
706 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
708 static void blk_mq_work_fn(struct work_struct *work)
710 struct blk_mq_hw_ctx *hctx;
712 hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work);
713 __blk_mq_run_hw_queue(hctx);
716 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
719 struct blk_mq_ctx *ctx = rq->mq_ctx;
721 trace_block_rq_insert(hctx->queue, rq);
723 list_add_tail(&rq->queuelist, &ctx->rq_list);
724 blk_mq_hctx_mark_pending(hctx, ctx);
727 * We do this early, to ensure we are on the right CPU.
729 blk_mq_add_timer(rq);
732 void blk_mq_insert_request(struct request_queue *q, struct request *rq,
735 struct blk_mq_hw_ctx *hctx;
736 struct blk_mq_ctx *ctx, *current_ctx;
739 hctx = q->mq_ops->map_queue(q, ctx->cpu);
741 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA)) {
742 blk_insert_flush(rq);
744 current_ctx = blk_mq_get_ctx(q);
746 if (!cpu_online(ctx->cpu)) {
748 hctx = q->mq_ops->map_queue(q, ctx->cpu);
751 spin_lock(&ctx->lock);
752 __blk_mq_insert_request(hctx, rq);
753 spin_unlock(&ctx->lock);
755 blk_mq_put_ctx(current_ctx);
759 __blk_mq_run_hw_queue(hctx);
761 EXPORT_SYMBOL(blk_mq_insert_request);
764 * This is a special version of blk_mq_insert_request to bypass FLUSH request
765 * check. Should only be used internally.
767 void blk_mq_run_request(struct request *rq, bool run_queue, bool async)
769 struct request_queue *q = rq->q;
770 struct blk_mq_hw_ctx *hctx;
771 struct blk_mq_ctx *ctx, *current_ctx;
773 current_ctx = blk_mq_get_ctx(q);
776 if (!cpu_online(ctx->cpu)) {
780 hctx = q->mq_ops->map_queue(q, ctx->cpu);
782 /* ctx->cpu might be offline */
783 spin_lock(&ctx->lock);
784 __blk_mq_insert_request(hctx, rq);
785 spin_unlock(&ctx->lock);
787 blk_mq_put_ctx(current_ctx);
790 blk_mq_run_hw_queue(hctx, async);
793 static void blk_mq_insert_requests(struct request_queue *q,
794 struct blk_mq_ctx *ctx,
795 struct list_head *list,
800 struct blk_mq_hw_ctx *hctx;
801 struct blk_mq_ctx *current_ctx;
803 trace_block_unplug(q, depth, !from_schedule);
805 current_ctx = blk_mq_get_ctx(q);
807 if (!cpu_online(ctx->cpu))
809 hctx = q->mq_ops->map_queue(q, ctx->cpu);
812 * preemption doesn't flush plug list, so it's possible ctx->cpu is
815 spin_lock(&ctx->lock);
816 while (!list_empty(list)) {
819 rq = list_first_entry(list, struct request, queuelist);
820 list_del_init(&rq->queuelist);
822 __blk_mq_insert_request(hctx, rq);
824 spin_unlock(&ctx->lock);
826 blk_mq_put_ctx(current_ctx);
828 blk_mq_run_hw_queue(hctx, from_schedule);
831 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
833 struct request *rqa = container_of(a, struct request, queuelist);
834 struct request *rqb = container_of(b, struct request, queuelist);
836 return !(rqa->mq_ctx < rqb->mq_ctx ||
837 (rqa->mq_ctx == rqb->mq_ctx &&
838 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
841 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
843 struct blk_mq_ctx *this_ctx;
844 struct request_queue *this_q;
850 list_splice_init(&plug->mq_list, &list);
852 list_sort(NULL, &list, plug_ctx_cmp);
858 while (!list_empty(&list)) {
859 rq = list_entry_rq(list.next);
860 list_del_init(&rq->queuelist);
862 if (rq->mq_ctx != this_ctx) {
864 blk_mq_insert_requests(this_q, this_ctx,
869 this_ctx = rq->mq_ctx;
875 list_add_tail(&rq->queuelist, &ctx_list);
879 * If 'this_ctx' is set, we know we have entries to complete
880 * on 'ctx_list'. Do those.
883 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
888 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
890 init_request_from_bio(rq, bio);
891 blk_account_io_start(rq, 1);
894 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
896 struct blk_mq_hw_ctx *hctx;
897 struct blk_mq_ctx *ctx;
898 const int is_sync = rw_is_sync(bio->bi_rw);
899 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
900 int rw = bio_data_dir(bio);
902 unsigned int use_plug, request_count = 0;
905 * If we have multiple hardware queues, just go directly to
906 * one of those for sync IO.
908 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
910 blk_queue_bounce(q, &bio);
912 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
915 if (blk_mq_queue_enter(q)) {
916 bio_endio(bio, -EIO);
920 ctx = blk_mq_get_ctx(q);
921 hctx = q->mq_ops->map_queue(q, ctx->cpu);
923 trace_block_getrq(q, bio, rw);
924 rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false, bio->bi_rw);
926 blk_mq_rq_ctx_init(q, ctx, rq, bio->bi_rw);
929 trace_block_sleeprq(q, bio, rw);
930 rq = blk_mq_alloc_request_pinned(q, bio->bi_rw,
931 __GFP_WAIT|GFP_ATOMIC, false);
933 hctx = q->mq_ops->map_queue(q, ctx->cpu);
938 if (unlikely(is_flush_fua)) {
939 blk_mq_bio_to_request(rq, bio);
941 blk_insert_flush(rq);
946 * A task plug currently exists. Since this is completely lockless,
947 * utilize that to temporarily store requests until the task is
948 * either done or scheduled away.
951 struct blk_plug *plug = current->plug;
954 blk_mq_bio_to_request(rq, bio);
955 if (list_empty(&plug->mq_list))
957 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
958 blk_flush_plug_list(plug, false);
961 list_add_tail(&rq->queuelist, &plug->mq_list);
967 spin_lock(&ctx->lock);
969 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
970 blk_mq_attempt_merge(q, ctx, bio))
971 __blk_mq_free_request(hctx, ctx, rq);
973 blk_mq_bio_to_request(rq, bio);
974 __blk_mq_insert_request(hctx, rq);
977 spin_unlock(&ctx->lock);
981 * For a SYNC request, send it to the hardware immediately. For an
982 * ASYNC request, just ensure that we run it later on. The latter
983 * allows for merging opportunities and more efficient dispatching.
986 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
990 * Default mapping to a software queue, since we use one per CPU.
992 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
994 return q->queue_hw_ctx[q->mq_map[cpu]];
996 EXPORT_SYMBOL(blk_mq_map_queue);
998 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg,
999 unsigned int hctx_index)
1001 return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
1002 GFP_KERNEL | __GFP_ZERO, reg->numa_node);
1004 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1006 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1007 unsigned int hctx_index)
1011 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1013 static void blk_mq_hctx_notify(void *data, unsigned long action,
1016 struct blk_mq_hw_ctx *hctx = data;
1017 struct blk_mq_ctx *ctx;
1020 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1024 * Move ctx entries to new CPU, if this one is going away.
1026 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1028 spin_lock(&ctx->lock);
1029 if (!list_empty(&ctx->rq_list)) {
1030 list_splice_init(&ctx->rq_list, &tmp);
1031 clear_bit(ctx->index_hw, hctx->ctx_map);
1033 spin_unlock(&ctx->lock);
1035 if (list_empty(&tmp))
1038 ctx = blk_mq_get_ctx(hctx->queue);
1039 spin_lock(&ctx->lock);
1041 while (!list_empty(&tmp)) {
1044 rq = list_first_entry(&tmp, struct request, queuelist);
1046 list_move_tail(&rq->queuelist, &ctx->rq_list);
1049 blk_mq_hctx_mark_pending(hctx, ctx);
1051 spin_unlock(&ctx->lock);
1052 blk_mq_put_ctx(ctx);
1055 static void blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx,
1056 void (*init)(void *, struct blk_mq_hw_ctx *,
1057 struct request *, unsigned int),
1062 for (i = 0; i < hctx->queue_depth; i++) {
1063 struct request *rq = hctx->rqs[i];
1065 init(data, hctx, rq, i);
1069 void blk_mq_init_commands(struct request_queue *q,
1070 void (*init)(void *, struct blk_mq_hw_ctx *,
1071 struct request *, unsigned int),
1074 struct blk_mq_hw_ctx *hctx;
1077 queue_for_each_hw_ctx(q, hctx, i)
1078 blk_mq_init_hw_commands(hctx, init, data);
1080 EXPORT_SYMBOL(blk_mq_init_commands);
1082 static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx)
1086 while (!list_empty(&hctx->page_list)) {
1087 page = list_first_entry(&hctx->page_list, struct page, lru);
1088 list_del_init(&page->lru);
1089 __free_pages(page, page->private);
1095 blk_mq_free_tags(hctx->tags);
1098 static size_t order_to_size(unsigned int order)
1100 size_t ret = PAGE_SIZE;
1108 static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
1109 unsigned int reserved_tags, int node)
1111 unsigned int i, j, entries_per_page, max_order = 4;
1112 size_t rq_size, left;
1114 INIT_LIST_HEAD(&hctx->page_list);
1116 hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *),
1122 * rq_size is the size of the request plus driver payload, rounded
1123 * to the cacheline size
1125 rq_size = round_up(sizeof(struct request) + hctx->cmd_size,
1127 left = rq_size * hctx->queue_depth;
1129 for (i = 0; i < hctx->queue_depth;) {
1130 int this_order = max_order;
1135 while (left < order_to_size(this_order - 1) && this_order)
1139 page = alloc_pages_node(node, GFP_KERNEL, this_order);
1144 if (order_to_size(this_order) < rq_size)
1151 page->private = this_order;
1152 list_add_tail(&page->lru, &hctx->page_list);
1154 p = page_address(page);
1155 entries_per_page = order_to_size(this_order) / rq_size;
1156 to_do = min(entries_per_page, hctx->queue_depth - i);
1157 left -= to_do * rq_size;
1158 for (j = 0; j < to_do; j++) {
1160 blk_mq_rq_init(hctx, hctx->rqs[i]);
1166 if (i < (reserved_tags + BLK_MQ_TAG_MIN))
1168 else if (i != hctx->queue_depth) {
1169 hctx->queue_depth = i;
1170 pr_warn("%s: queue depth set to %u because of low memory\n",
1174 hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
1177 blk_mq_free_rq_map(hctx);
1184 static int blk_mq_init_hw_queues(struct request_queue *q,
1185 struct blk_mq_reg *reg, void *driver_data)
1187 struct blk_mq_hw_ctx *hctx;
1191 * Initialize hardware queues
1193 queue_for_each_hw_ctx(q, hctx, i) {
1194 unsigned int num_maps;
1197 node = hctx->numa_node;
1198 if (node == NUMA_NO_NODE)
1199 node = hctx->numa_node = reg->numa_node;
1201 INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
1202 spin_lock_init(&hctx->lock);
1203 INIT_LIST_HEAD(&hctx->dispatch);
1205 hctx->queue_num = i;
1206 hctx->flags = reg->flags;
1207 hctx->queue_depth = reg->queue_depth;
1208 hctx->reserved_tags = reg->reserved_tags;
1209 hctx->cmd_size = reg->cmd_size;
1210 atomic_set(&hctx->pending_flush, 0);
1212 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1213 blk_mq_hctx_notify, hctx);
1214 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1216 if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
1220 * Allocate space for all possible cpus to avoid allocation in
1223 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1228 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1229 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1234 hctx->nr_ctx_map = num_maps;
1237 if (reg->ops->init_hctx &&
1238 reg->ops->init_hctx(hctx, driver_data, i))
1242 if (i == q->nr_hw_queues)
1248 queue_for_each_hw_ctx(q, hctx, j) {
1252 if (reg->ops->exit_hctx)
1253 reg->ops->exit_hctx(hctx, j);
1255 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1256 blk_mq_free_rq_map(hctx);
1263 static void blk_mq_init_cpu_queues(struct request_queue *q,
1264 unsigned int nr_hw_queues)
1268 for_each_possible_cpu(i) {
1269 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1270 struct blk_mq_hw_ctx *hctx;
1272 memset(__ctx, 0, sizeof(*__ctx));
1274 spin_lock_init(&__ctx->lock);
1275 INIT_LIST_HEAD(&__ctx->rq_list);
1278 /* If the cpu isn't online, the cpu is mapped to first hctx */
1279 hctx = q->mq_ops->map_queue(q, i);
1286 * Set local node, IFF we have more than one hw queue. If
1287 * not, we remain on the home node of the device
1289 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1290 hctx->numa_node = cpu_to_node(i);
1294 static void blk_mq_map_swqueue(struct request_queue *q)
1297 struct blk_mq_hw_ctx *hctx;
1298 struct blk_mq_ctx *ctx;
1300 queue_for_each_hw_ctx(q, hctx, i) {
1305 * Map software to hardware queues
1307 queue_for_each_ctx(q, ctx, i) {
1308 /* If the cpu isn't online, the cpu is mapped to first hctx */
1309 hctx = q->mq_ops->map_queue(q, i);
1310 ctx->index_hw = hctx->nr_ctx;
1311 hctx->ctxs[hctx->nr_ctx++] = ctx;
1315 struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg,
1318 struct blk_mq_hw_ctx **hctxs;
1319 struct blk_mq_ctx *ctx;
1320 struct request_queue *q;
1323 if (!reg->nr_hw_queues ||
1324 !reg->ops->queue_rq || !reg->ops->map_queue ||
1325 !reg->ops->alloc_hctx || !reg->ops->free_hctx)
1326 return ERR_PTR(-EINVAL);
1328 if (!reg->queue_depth)
1329 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1330 else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) {
1331 pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth);
1332 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1335 if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN))
1336 return ERR_PTR(-EINVAL);
1338 ctx = alloc_percpu(struct blk_mq_ctx);
1340 return ERR_PTR(-ENOMEM);
1342 hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1348 for (i = 0; i < reg->nr_hw_queues; i++) {
1349 hctxs[i] = reg->ops->alloc_hctx(reg, i);
1353 hctxs[i]->numa_node = NUMA_NO_NODE;
1354 hctxs[i]->queue_num = i;
1357 q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node);
1361 q->mq_map = blk_mq_make_queue_map(reg);
1365 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1366 blk_queue_rq_timeout(q, 30000);
1368 q->nr_queues = nr_cpu_ids;
1369 q->nr_hw_queues = reg->nr_hw_queues;
1372 q->queue_hw_ctx = hctxs;
1374 q->mq_ops = reg->ops;
1375 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1377 blk_queue_make_request(q, blk_mq_make_request);
1378 blk_queue_rq_timed_out(q, reg->ops->timeout);
1380 blk_queue_rq_timeout(q, reg->timeout);
1382 blk_mq_init_flush(q);
1383 blk_mq_init_cpu_queues(q, reg->nr_hw_queues);
1385 if (blk_mq_init_hw_queues(q, reg, driver_data))
1388 blk_mq_map_swqueue(q);
1390 mutex_lock(&all_q_mutex);
1391 list_add_tail(&q->all_q_node, &all_q_list);
1392 mutex_unlock(&all_q_mutex);
1398 blk_cleanup_queue(q);
1400 for (i = 0; i < reg->nr_hw_queues; i++) {
1403 reg->ops->free_hctx(hctxs[i], i);
1408 return ERR_PTR(-ENOMEM);
1410 EXPORT_SYMBOL(blk_mq_init_queue);
1412 void blk_mq_free_queue(struct request_queue *q)
1414 struct blk_mq_hw_ctx *hctx;
1417 queue_for_each_hw_ctx(q, hctx, i) {
1418 kfree(hctx->ctx_map);
1420 blk_mq_free_rq_map(hctx);
1421 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1422 if (q->mq_ops->exit_hctx)
1423 q->mq_ops->exit_hctx(hctx, i);
1424 q->mq_ops->free_hctx(hctx, i);
1427 free_percpu(q->queue_ctx);
1428 kfree(q->queue_hw_ctx);
1431 q->queue_ctx = NULL;
1432 q->queue_hw_ctx = NULL;
1435 mutex_lock(&all_q_mutex);
1436 list_del_init(&q->all_q_node);
1437 mutex_unlock(&all_q_mutex);
1440 /* Basically redo blk_mq_init_queue with queue frozen */
1441 static void blk_mq_queue_reinit(struct request_queue *q)
1443 blk_mq_freeze_queue(q);
1445 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1448 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1449 * we should change hctx numa_node according to new topology (this
1450 * involves free and re-allocate memory, worthy doing?)
1453 blk_mq_map_swqueue(q);
1455 blk_mq_unfreeze_queue(q);
1458 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1459 unsigned long action, void *hcpu)
1461 struct request_queue *q;
1464 * Before new mapping is established, hotadded cpu might already start
1465 * handling requests. This doesn't break anything as we map offline
1466 * CPUs to first hardware queue. We will re-init queue below to get
1469 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1470 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1473 mutex_lock(&all_q_mutex);
1474 list_for_each_entry(q, &all_q_list, all_q_node)
1475 blk_mq_queue_reinit(q);
1476 mutex_unlock(&all_q_mutex);
1480 static int __init blk_mq_init(void)
1484 /* Must be called after percpu_counter_hotcpu_callback() */
1485 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1489 subsys_initcall(blk_mq_init);