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 DEFINE_PER_CPU(struct llist_head, ipi_lists);
32 static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
35 return per_cpu_ptr(q->queue_ctx, cpu);
39 * This assumes per-cpu software queueing queues. They could be per-node
40 * as well, for instance. For now this is hardcoded as-is. Note that we don't
41 * care about preemption, since we know the ctx's are persistent. This does
42 * mean that we can't rely on ctx always matching the currently running CPU.
44 static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
46 return __blk_mq_get_ctx(q, get_cpu());
49 static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
55 * Check if any of the ctx's have pending work in this hardware queue
57 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
61 for (i = 0; i < hctx->nr_ctx_map; i++)
69 * Mark this ctx as having pending work in this hardware queue
71 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
72 struct blk_mq_ctx *ctx)
74 if (!test_bit(ctx->index_hw, hctx->ctx_map))
75 set_bit(ctx->index_hw, hctx->ctx_map);
78 static struct request *blk_mq_alloc_rq(struct blk_mq_hw_ctx *hctx, gfp_t gfp,
84 tag = blk_mq_get_tag(hctx->tags, gfp, reserved);
85 if (tag != BLK_MQ_TAG_FAIL) {
95 static int blk_mq_queue_enter(struct request_queue *q)
99 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
101 /* we have problems to freeze the queue if it's initializing */
102 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
105 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
107 spin_lock_irq(q->queue_lock);
108 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
109 !blk_queue_bypass(q), *q->queue_lock);
110 /* inc usage with lock hold to avoid freeze_queue runs here */
112 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
113 spin_unlock_irq(q->queue_lock);
118 static void blk_mq_queue_exit(struct request_queue *q)
120 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
124 * Guarantee no request is in use, so we can change any data structure of
125 * the queue afterward.
127 static void blk_mq_freeze_queue(struct request_queue *q)
131 spin_lock_irq(q->queue_lock);
132 drain = !q->bypass_depth++;
133 queue_flag_set(QUEUE_FLAG_BYPASS, q);
134 spin_unlock_irq(q->queue_lock);
142 spin_lock_irq(q->queue_lock);
143 count = percpu_counter_sum(&q->mq_usage_counter);
144 spin_unlock_irq(q->queue_lock);
148 blk_mq_run_queues(q, false);
153 static void blk_mq_unfreeze_queue(struct request_queue *q)
157 spin_lock_irq(q->queue_lock);
158 if (!--q->bypass_depth) {
159 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
162 WARN_ON_ONCE(q->bypass_depth < 0);
163 spin_unlock_irq(q->queue_lock);
165 wake_up_all(&q->mq_freeze_wq);
168 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
170 return blk_mq_has_free_tags(hctx->tags);
172 EXPORT_SYMBOL(blk_mq_can_queue);
174 static void blk_mq_rq_ctx_init(struct blk_mq_ctx *ctx, struct request *rq,
175 unsigned int rw_flags)
178 rq->cmd_flags = rw_flags;
179 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
182 static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx,
183 gfp_t gfp, bool reserved)
185 return blk_mq_alloc_rq(hctx, gfp, reserved);
188 static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
195 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
196 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
198 rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved);
200 blk_mq_rq_ctx_init(ctx, rq, rw);
202 } else if (!(gfp & __GFP_WAIT))
206 __blk_mq_run_hw_queue(hctx);
207 blk_mq_wait_for_tags(hctx->tags);
213 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
214 gfp_t gfp, bool reserved)
218 if (blk_mq_queue_enter(q))
221 rq = blk_mq_alloc_request_pinned(q, rw, gfp, reserved);
222 blk_mq_put_ctx(rq->mq_ctx);
226 struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
231 if (blk_mq_queue_enter(q))
234 rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
235 blk_mq_put_ctx(rq->mq_ctx);
238 EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
241 * Re-init and set pdu, if we have it
243 static void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq)
245 blk_rq_init(hctx->queue, rq);
248 rq->special = blk_mq_rq_to_pdu(rq);
251 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
252 struct blk_mq_ctx *ctx, struct request *rq)
254 const int tag = rq->tag;
255 struct request_queue *q = rq->q;
257 blk_mq_rq_init(hctx, rq);
258 blk_mq_put_tag(hctx->tags, tag);
260 blk_mq_queue_exit(q);
263 void blk_mq_free_request(struct request *rq)
265 struct blk_mq_ctx *ctx = rq->mq_ctx;
266 struct blk_mq_hw_ctx *hctx;
267 struct request_queue *q = rq->q;
269 ctx->rq_completed[rq_is_sync(rq)]++;
271 hctx = q->mq_ops->map_queue(q, ctx->cpu);
272 __blk_mq_free_request(hctx, ctx, rq);
275 static void blk_mq_bio_endio(struct request *rq, struct bio *bio, int error)
278 clear_bit(BIO_UPTODATE, &bio->bi_flags);
279 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
282 if (unlikely(rq->cmd_flags & REQ_QUIET))
283 set_bit(BIO_QUIET, &bio->bi_flags);
285 /* don't actually finish bio if it's part of flush sequence */
286 if (!(rq->cmd_flags & REQ_FLUSH_SEQ))
287 bio_endio(bio, error);
290 void blk_mq_complete_request(struct request *rq, int error)
292 struct bio *bio = rq->bio;
293 unsigned int bytes = 0;
295 trace_block_rq_complete(rq->q, rq);
298 struct bio *next = bio->bi_next;
301 bytes += bio->bi_size;
302 blk_mq_bio_endio(rq, bio, error);
306 blk_account_io_completion(rq, bytes);
309 rq->end_io(rq, error);
311 blk_mq_free_request(rq);
313 blk_account_io_done(rq);
316 void __blk_mq_end_io(struct request *rq, int error)
318 if (!blk_mark_rq_complete(rq))
319 blk_mq_complete_request(rq, error);
322 #if defined(CONFIG_SMP)
325 * Called with interrupts disabled.
327 static void ipi_end_io(void *data)
329 struct llist_head *list = &per_cpu(ipi_lists, smp_processor_id());
330 struct llist_node *entry, *next;
333 entry = llist_del_all(list);
337 rq = llist_entry(entry, struct request, ll_list);
338 __blk_mq_end_io(rq, rq->errors);
343 static int ipi_remote_cpu(struct blk_mq_ctx *ctx, const int cpu,
344 struct request *rq, const int error)
346 struct call_single_data *data = &rq->csd;
349 rq->ll_list.next = NULL;
352 * If the list is non-empty, an existing IPI must already
353 * be "in flight". If that is the case, we need not schedule
356 if (llist_add(&rq->ll_list, &per_cpu(ipi_lists, ctx->cpu))) {
357 data->func = ipi_end_io;
359 __smp_call_function_single(ctx->cpu, data, 0);
364 #else /* CONFIG_SMP */
365 static int ipi_remote_cpu(struct blk_mq_ctx *ctx, const int cpu,
366 struct request *rq, const int error)
373 * End IO on this request on a multiqueue enabled driver. We'll either do
374 * it directly inline, or punt to a local IPI handler on the matching
377 void blk_mq_end_io(struct request *rq, int error)
379 struct blk_mq_ctx *ctx = rq->mq_ctx;
382 if (!ctx->ipi_redirect)
383 return __blk_mq_end_io(rq, error);
387 if (cpu == ctx->cpu || !cpu_online(ctx->cpu) ||
388 !ipi_remote_cpu(ctx, cpu, rq, error))
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 list_add_tail(&rq->queuelist, &ctx->rq_list);
722 blk_mq_hctx_mark_pending(hctx, ctx);
725 * We do this early, to ensure we are on the right CPU.
727 blk_mq_add_timer(rq);
730 void blk_mq_insert_request(struct request_queue *q, struct request *rq,
733 struct blk_mq_hw_ctx *hctx;
734 struct blk_mq_ctx *ctx, *current_ctx;
737 hctx = q->mq_ops->map_queue(q, ctx->cpu);
739 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA)) {
740 blk_insert_flush(rq);
742 current_ctx = blk_mq_get_ctx(q);
744 if (!cpu_online(ctx->cpu)) {
746 hctx = q->mq_ops->map_queue(q, ctx->cpu);
749 spin_lock(&ctx->lock);
750 __blk_mq_insert_request(hctx, rq);
751 spin_unlock(&ctx->lock);
753 blk_mq_put_ctx(current_ctx);
757 __blk_mq_run_hw_queue(hctx);
759 EXPORT_SYMBOL(blk_mq_insert_request);
762 * This is a special version of blk_mq_insert_request to bypass FLUSH request
763 * check. Should only be used internally.
765 void blk_mq_run_request(struct request *rq, bool run_queue, bool async)
767 struct request_queue *q = rq->q;
768 struct blk_mq_hw_ctx *hctx;
769 struct blk_mq_ctx *ctx, *current_ctx;
771 current_ctx = blk_mq_get_ctx(q);
774 if (!cpu_online(ctx->cpu)) {
778 hctx = q->mq_ops->map_queue(q, ctx->cpu);
780 /* ctx->cpu might be offline */
781 spin_lock(&ctx->lock);
782 __blk_mq_insert_request(hctx, rq);
783 spin_unlock(&ctx->lock);
785 blk_mq_put_ctx(current_ctx);
788 blk_mq_run_hw_queue(hctx, async);
791 static void blk_mq_insert_requests(struct request_queue *q,
792 struct blk_mq_ctx *ctx,
793 struct list_head *list,
798 struct blk_mq_hw_ctx *hctx;
799 struct blk_mq_ctx *current_ctx;
801 trace_block_unplug(q, depth, !from_schedule);
803 current_ctx = blk_mq_get_ctx(q);
805 if (!cpu_online(ctx->cpu))
807 hctx = q->mq_ops->map_queue(q, ctx->cpu);
810 * preemption doesn't flush plug list, so it's possible ctx->cpu is
813 spin_lock(&ctx->lock);
814 while (!list_empty(list)) {
817 rq = list_first_entry(list, struct request, queuelist);
818 list_del_init(&rq->queuelist);
820 __blk_mq_insert_request(hctx, rq);
822 spin_unlock(&ctx->lock);
824 blk_mq_put_ctx(current_ctx);
826 blk_mq_run_hw_queue(hctx, from_schedule);
829 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
831 struct request *rqa = container_of(a, struct request, queuelist);
832 struct request *rqb = container_of(b, struct request, queuelist);
834 return !(rqa->mq_ctx < rqb->mq_ctx ||
835 (rqa->mq_ctx == rqb->mq_ctx &&
836 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
839 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
841 struct blk_mq_ctx *this_ctx;
842 struct request_queue *this_q;
848 list_splice_init(&plug->mq_list, &list);
850 list_sort(NULL, &list, plug_ctx_cmp);
856 while (!list_empty(&list)) {
857 rq = list_entry_rq(list.next);
858 list_del_init(&rq->queuelist);
860 if (rq->mq_ctx != this_ctx) {
862 blk_mq_insert_requests(this_q, this_ctx,
867 this_ctx = rq->mq_ctx;
873 list_add_tail(&rq->queuelist, &ctx_list);
877 * If 'this_ctx' is set, we know we have entries to complete
878 * on 'ctx_list'. Do those.
881 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
886 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
888 init_request_from_bio(rq, bio);
889 blk_account_io_start(rq, 1);
892 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
894 struct blk_mq_hw_ctx *hctx;
895 struct blk_mq_ctx *ctx;
896 const int is_sync = rw_is_sync(bio->bi_rw);
897 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
898 int rw = bio_data_dir(bio);
900 unsigned int use_plug, request_count = 0;
903 * If we have multiple hardware queues, just go directly to
904 * one of those for sync IO.
906 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
908 blk_queue_bounce(q, &bio);
910 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
913 if (blk_mq_queue_enter(q)) {
914 bio_endio(bio, -EIO);
918 ctx = blk_mq_get_ctx(q);
919 hctx = q->mq_ops->map_queue(q, ctx->cpu);
921 trace_block_getrq(q, bio, rw);
922 rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
924 blk_mq_rq_ctx_init(ctx, rq, rw);
927 trace_block_sleeprq(q, bio, rw);
928 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
931 hctx = q->mq_ops->map_queue(q, ctx->cpu);
936 if (unlikely(is_flush_fua)) {
937 blk_mq_bio_to_request(rq, bio);
939 blk_insert_flush(rq);
944 * A task plug currently exists. Since this is completely lockless,
945 * utilize that to temporarily store requests until the task is
946 * either done or scheduled away.
949 struct blk_plug *plug = current->plug;
952 blk_mq_bio_to_request(rq, bio);
953 if (list_empty(&plug->mq_list))
955 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
956 blk_flush_plug_list(plug, false);
959 list_add_tail(&rq->queuelist, &plug->mq_list);
965 spin_lock(&ctx->lock);
967 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
968 blk_mq_attempt_merge(q, ctx, bio))
969 __blk_mq_free_request(hctx, ctx, rq);
971 blk_mq_bio_to_request(rq, bio);
972 __blk_mq_insert_request(hctx, rq);
975 spin_unlock(&ctx->lock);
979 * For a SYNC request, send it to the hardware immediately. For an
980 * ASYNC request, just ensure that we run it later on. The latter
981 * allows for merging opportunities and more efficient dispatching.
984 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
988 * Default mapping to a software queue, since we use one per CPU.
990 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
992 return q->queue_hw_ctx[q->mq_map[cpu]];
994 EXPORT_SYMBOL(blk_mq_map_queue);
996 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg,
997 unsigned int hctx_index)
999 return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
1000 GFP_KERNEL | __GFP_ZERO, reg->numa_node);
1002 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1004 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1005 unsigned int hctx_index)
1009 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1011 static void blk_mq_hctx_notify(void *data, unsigned long action,
1014 struct blk_mq_hw_ctx *hctx = data;
1015 struct blk_mq_ctx *ctx;
1018 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1022 * Move ctx entries to new CPU, if this one is going away.
1024 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1026 spin_lock(&ctx->lock);
1027 if (!list_empty(&ctx->rq_list)) {
1028 list_splice_init(&ctx->rq_list, &tmp);
1029 clear_bit(ctx->index_hw, hctx->ctx_map);
1031 spin_unlock(&ctx->lock);
1033 if (list_empty(&tmp))
1036 ctx = blk_mq_get_ctx(hctx->queue);
1037 spin_lock(&ctx->lock);
1039 while (!list_empty(&tmp)) {
1042 rq = list_first_entry(&tmp, struct request, queuelist);
1044 list_move_tail(&rq->queuelist, &ctx->rq_list);
1047 blk_mq_hctx_mark_pending(hctx, ctx);
1049 spin_unlock(&ctx->lock);
1050 blk_mq_put_ctx(ctx);
1053 static void blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx,
1054 void (*init)(void *, struct blk_mq_hw_ctx *,
1055 struct request *, unsigned int),
1060 for (i = 0; i < hctx->queue_depth; i++) {
1061 struct request *rq = hctx->rqs[i];
1063 init(data, hctx, rq, i);
1067 void blk_mq_init_commands(struct request_queue *q,
1068 void (*init)(void *, struct blk_mq_hw_ctx *,
1069 struct request *, unsigned int),
1072 struct blk_mq_hw_ctx *hctx;
1075 queue_for_each_hw_ctx(q, hctx, i)
1076 blk_mq_init_hw_commands(hctx, init, data);
1078 EXPORT_SYMBOL(blk_mq_init_commands);
1080 static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx)
1084 while (!list_empty(&hctx->page_list)) {
1085 page = list_first_entry(&hctx->page_list, struct page, list);
1086 list_del_init(&page->list);
1087 __free_pages(page, page->private);
1093 blk_mq_free_tags(hctx->tags);
1096 static size_t order_to_size(unsigned int order)
1098 size_t ret = PAGE_SIZE;
1106 static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
1107 unsigned int reserved_tags, int node)
1109 unsigned int i, j, entries_per_page, max_order = 4;
1110 size_t rq_size, left;
1112 INIT_LIST_HEAD(&hctx->page_list);
1114 hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *),
1120 * rq_size is the size of the request plus driver payload, rounded
1121 * to the cacheline size
1123 rq_size = round_up(sizeof(struct request) + hctx->cmd_size,
1125 left = rq_size * hctx->queue_depth;
1127 for (i = 0; i < hctx->queue_depth;) {
1128 int this_order = max_order;
1133 while (left < order_to_size(this_order - 1) && this_order)
1137 page = alloc_pages_node(node, GFP_KERNEL, this_order);
1142 if (order_to_size(this_order) < rq_size)
1149 page->private = this_order;
1150 list_add_tail(&page->list, &hctx->page_list);
1152 p = page_address(page);
1153 entries_per_page = order_to_size(this_order) / rq_size;
1154 to_do = min(entries_per_page, hctx->queue_depth - i);
1155 left -= to_do * rq_size;
1156 for (j = 0; j < to_do; j++) {
1158 blk_mq_rq_init(hctx, hctx->rqs[i]);
1164 if (i < (reserved_tags + BLK_MQ_TAG_MIN))
1166 else if (i != hctx->queue_depth) {
1167 hctx->queue_depth = i;
1168 pr_warn("%s: queue depth set to %u because of low memory\n",
1172 hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
1175 blk_mq_free_rq_map(hctx);
1182 static int blk_mq_init_hw_queues(struct request_queue *q,
1183 struct blk_mq_reg *reg, void *driver_data)
1185 struct blk_mq_hw_ctx *hctx;
1189 * Initialize hardware queues
1191 queue_for_each_hw_ctx(q, hctx, i) {
1192 unsigned int num_maps;
1195 node = hctx->numa_node;
1196 if (node == NUMA_NO_NODE)
1197 node = hctx->numa_node = reg->numa_node;
1199 INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
1200 spin_lock_init(&hctx->lock);
1201 INIT_LIST_HEAD(&hctx->dispatch);
1203 hctx->queue_num = i;
1204 hctx->flags = reg->flags;
1205 hctx->queue_depth = reg->queue_depth;
1206 hctx->cmd_size = reg->cmd_size;
1208 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1209 blk_mq_hctx_notify, hctx);
1210 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1212 if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
1216 * Allocate space for all possible cpus to avoid allocation in
1219 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1224 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1225 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1230 hctx->nr_ctx_map = num_maps;
1233 if (reg->ops->init_hctx &&
1234 reg->ops->init_hctx(hctx, driver_data, i))
1238 if (i == q->nr_hw_queues)
1244 queue_for_each_hw_ctx(q, hctx, j) {
1248 if (reg->ops->exit_hctx)
1249 reg->ops->exit_hctx(hctx, j);
1251 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1252 blk_mq_free_rq_map(hctx);
1259 static void blk_mq_init_cpu_queues(struct request_queue *q,
1260 unsigned int nr_hw_queues)
1264 for_each_possible_cpu(i) {
1265 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1266 struct blk_mq_hw_ctx *hctx;
1268 memset(__ctx, 0, sizeof(*__ctx));
1270 spin_lock_init(&__ctx->lock);
1271 INIT_LIST_HEAD(&__ctx->rq_list);
1274 /* If the cpu isn't online, the cpu is mapped to first hctx */
1275 hctx = q->mq_ops->map_queue(q, i);
1282 * Set local node, IFF we have more than one hw queue. If
1283 * not, we remain on the home node of the device
1285 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1286 hctx->numa_node = cpu_to_node(i);
1290 static void blk_mq_map_swqueue(struct request_queue *q)
1293 struct blk_mq_hw_ctx *hctx;
1294 struct blk_mq_ctx *ctx;
1296 queue_for_each_hw_ctx(q, hctx, i) {
1301 * Map software to hardware queues
1303 queue_for_each_ctx(q, ctx, i) {
1304 /* If the cpu isn't online, the cpu is mapped to first hctx */
1305 hctx = q->mq_ops->map_queue(q, i);
1306 ctx->index_hw = hctx->nr_ctx;
1307 hctx->ctxs[hctx->nr_ctx++] = ctx;
1311 struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg,
1314 struct blk_mq_hw_ctx **hctxs;
1315 struct blk_mq_ctx *ctx;
1316 struct request_queue *q;
1319 if (!reg->nr_hw_queues ||
1320 !reg->ops->queue_rq || !reg->ops->map_queue ||
1321 !reg->ops->alloc_hctx || !reg->ops->free_hctx)
1322 return ERR_PTR(-EINVAL);
1324 if (!reg->queue_depth)
1325 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1326 else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) {
1327 pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth);
1328 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1332 * Set aside a tag for flush requests. It will only be used while
1333 * another flush request is in progress but outside the driver.
1335 * TODO: only allocate if flushes are supported
1338 reg->reserved_tags++;
1340 if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN))
1341 return ERR_PTR(-EINVAL);
1343 ctx = alloc_percpu(struct blk_mq_ctx);
1345 return ERR_PTR(-ENOMEM);
1347 hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1353 for (i = 0; i < reg->nr_hw_queues; i++) {
1354 hctxs[i] = reg->ops->alloc_hctx(reg, i);
1358 hctxs[i]->numa_node = NUMA_NO_NODE;
1359 hctxs[i]->queue_num = i;
1362 q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node);
1366 q->mq_map = blk_mq_make_queue_map(reg);
1370 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1371 blk_queue_rq_timeout(q, 30000);
1373 q->nr_queues = nr_cpu_ids;
1374 q->nr_hw_queues = reg->nr_hw_queues;
1377 q->queue_hw_ctx = hctxs;
1379 q->mq_ops = reg->ops;
1381 blk_queue_make_request(q, blk_mq_make_request);
1382 blk_queue_rq_timed_out(q, reg->ops->timeout);
1384 blk_queue_rq_timeout(q, reg->timeout);
1386 blk_mq_init_flush(q);
1387 blk_mq_init_cpu_queues(q, reg->nr_hw_queues);
1389 if (blk_mq_init_hw_queues(q, reg, driver_data))
1392 blk_mq_map_swqueue(q);
1394 mutex_lock(&all_q_mutex);
1395 list_add_tail(&q->all_q_node, &all_q_list);
1396 mutex_unlock(&all_q_mutex);
1402 blk_cleanup_queue(q);
1404 for (i = 0; i < reg->nr_hw_queues; i++) {
1407 reg->ops->free_hctx(hctxs[i], i);
1412 return ERR_PTR(-ENOMEM);
1414 EXPORT_SYMBOL(blk_mq_init_queue);
1416 void blk_mq_free_queue(struct request_queue *q)
1418 struct blk_mq_hw_ctx *hctx;
1421 queue_for_each_hw_ctx(q, hctx, i) {
1422 cancel_delayed_work_sync(&hctx->delayed_work);
1423 kfree(hctx->ctx_map);
1425 blk_mq_free_rq_map(hctx);
1426 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1427 if (q->mq_ops->exit_hctx)
1428 q->mq_ops->exit_hctx(hctx, i);
1429 q->mq_ops->free_hctx(hctx, i);
1432 free_percpu(q->queue_ctx);
1433 kfree(q->queue_hw_ctx);
1436 q->queue_ctx = NULL;
1437 q->queue_hw_ctx = NULL;
1440 mutex_lock(&all_q_mutex);
1441 list_del_init(&q->all_q_node);
1442 mutex_unlock(&all_q_mutex);
1444 EXPORT_SYMBOL(blk_mq_free_queue);
1446 /* Basically redo blk_mq_init_queue with queue frozen */
1447 static void __cpuinit blk_mq_queue_reinit(struct request_queue *q)
1449 blk_mq_freeze_queue(q);
1451 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1454 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1455 * we should change hctx numa_node according to new topology (this
1456 * involves free and re-allocate memory, worthy doing?)
1459 blk_mq_map_swqueue(q);
1461 blk_mq_unfreeze_queue(q);
1464 static int __cpuinit blk_mq_queue_reinit_notify(struct notifier_block *nb,
1465 unsigned long action, void *hcpu)
1467 struct request_queue *q;
1470 * Before new mapping is established, hotadded cpu might already start
1471 * handling requests. This doesn't break anything as we map offline
1472 * CPUs to first hardware queue. We will re-init queue below to get
1475 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1476 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1479 mutex_lock(&all_q_mutex);
1480 list_for_each_entry(q, &all_q_list, all_q_node)
1481 blk_mq_queue_reinit(q);
1482 mutex_unlock(&all_q_mutex);
1486 static int __init blk_mq_init(void)
1490 for_each_possible_cpu(i)
1491 init_llist_head(&per_cpu(ipi_lists, i));
1495 /* Must be called after percpu_counter_hotcpu_callback() */
1496 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1500 subsys_initcall(blk_mq_init);