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_request(struct blk_mq_hw_ctx *hctx,
77 gfp_t gfp, bool reserved)
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_pinned(struct request_queue *q,
203 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
204 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
206 rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved);
208 blk_mq_rq_ctx_init(q, ctx, rq, rw);
213 if (!(gfp & __GFP_WAIT))
216 __blk_mq_run_hw_queue(hctx);
217 blk_mq_wait_for_tags(hctx->tags);
223 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp)
227 if (blk_mq_queue_enter(q))
230 rq = blk_mq_alloc_request_pinned(q, rw, gfp, false);
232 blk_mq_put_ctx(rq->mq_ctx);
236 struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
241 if (blk_mq_queue_enter(q))
244 rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
246 blk_mq_put_ctx(rq->mq_ctx);
249 EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
252 * Re-init and set pdu, if we have it
254 void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq)
256 blk_rq_init(hctx->queue, rq);
259 rq->special = blk_mq_rq_to_pdu(rq);
262 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
263 struct blk_mq_ctx *ctx, struct request *rq)
265 const int tag = rq->tag;
266 struct request_queue *q = rq->q;
268 blk_mq_rq_init(hctx, rq);
269 blk_mq_put_tag(hctx->tags, tag);
271 blk_mq_queue_exit(q);
274 void blk_mq_free_request(struct request *rq)
276 struct blk_mq_ctx *ctx = rq->mq_ctx;
277 struct blk_mq_hw_ctx *hctx;
278 struct request_queue *q = rq->q;
280 ctx->rq_completed[rq_is_sync(rq)]++;
282 hctx = q->mq_ops->map_queue(q, ctx->cpu);
283 __blk_mq_free_request(hctx, ctx, rq);
286 bool blk_mq_end_io_partial(struct request *rq, int error, unsigned int nr_bytes)
288 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
291 blk_account_io_done(rq);
294 rq->end_io(rq, error);
296 blk_mq_free_request(rq);
299 EXPORT_SYMBOL(blk_mq_end_io_partial);
301 static void __blk_mq_complete_request_remote(void *data)
303 struct request *rq = data;
305 rq->q->softirq_done_fn(rq);
308 void __blk_mq_complete_request(struct request *rq)
310 struct blk_mq_ctx *ctx = rq->mq_ctx;
313 if (!ctx->ipi_redirect) {
314 rq->q->softirq_done_fn(rq);
319 if (cpu != ctx->cpu && cpu_online(ctx->cpu)) {
320 rq->csd.func = __blk_mq_complete_request_remote;
323 smp_call_function_single_async(ctx->cpu, &rq->csd);
325 rq->q->softirq_done_fn(rq);
331 * blk_mq_complete_request - end I/O on a request
332 * @rq: the request being processed
335 * Ends all I/O on a request. It does not handle partial completions.
336 * The actual completion happens out-of-order, through a IPI handler.
338 void blk_mq_complete_request(struct request *rq)
340 if (unlikely(blk_should_fake_timeout(rq->q)))
342 if (!blk_mark_rq_complete(rq))
343 __blk_mq_complete_request(rq);
345 EXPORT_SYMBOL(blk_mq_complete_request);
347 static void blk_mq_start_request(struct request *rq, bool last)
349 struct request_queue *q = rq->q;
351 trace_block_rq_issue(q, rq);
354 * Just mark start time and set the started bit. Due to memory
355 * ordering, we know we'll see the correct deadline as long as
356 * REQ_ATOMIC_STARTED is seen.
358 rq->deadline = jiffies + q->rq_timeout;
359 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
361 if (q->dma_drain_size && blk_rq_bytes(rq)) {
363 * Make sure space for the drain appears. We know we can do
364 * this because max_hw_segments has been adjusted to be one
365 * fewer than the device can handle.
367 rq->nr_phys_segments++;
371 * Flag the last request in the series so that drivers know when IO
372 * should be kicked off, if they don't do it on a per-request basis.
374 * Note: the flag isn't the only condition drivers should do kick off.
375 * If drive is busy, the last request might not have the bit set.
378 rq->cmd_flags |= REQ_END;
381 static void blk_mq_requeue_request(struct request *rq)
383 struct request_queue *q = rq->q;
385 trace_block_rq_requeue(q, rq);
386 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
388 rq->cmd_flags &= ~REQ_END;
390 if (q->dma_drain_size && blk_rq_bytes(rq))
391 rq->nr_phys_segments--;
394 struct blk_mq_timeout_data {
395 struct blk_mq_hw_ctx *hctx;
397 unsigned int *next_set;
400 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
402 struct blk_mq_timeout_data *data = __data;
403 struct blk_mq_hw_ctx *hctx = data->hctx;
406 /* It may not be in flight yet (this is where
407 * the REQ_ATOMIC_STARTED flag comes in). The requests are
408 * statically allocated, so we know it's always safe to access the
409 * memory associated with a bit offset into ->rqs[].
415 tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag);
416 if (tag >= hctx->queue_depth)
419 rq = hctx->rqs[tag++];
421 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
424 blk_rq_check_expired(rq, data->next, data->next_set);
428 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
430 unsigned int *next_set)
432 struct blk_mq_timeout_data data = {
435 .next_set = next_set,
439 * Ask the tagging code to iterate busy requests, so we can
440 * check them for timeout.
442 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
445 static void blk_mq_rq_timer(unsigned long data)
447 struct request_queue *q = (struct request_queue *) data;
448 struct blk_mq_hw_ctx *hctx;
449 unsigned long next = 0;
452 queue_for_each_hw_ctx(q, hctx, i)
453 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
456 mod_timer(&q->timeout, round_jiffies_up(next));
460 * Reverse check our software queue for entries that we could potentially
461 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
462 * too much time checking for merges.
464 static bool blk_mq_attempt_merge(struct request_queue *q,
465 struct blk_mq_ctx *ctx, struct bio *bio)
470 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
476 if (!blk_rq_merge_ok(rq, bio))
479 el_ret = blk_try_merge(rq, bio);
480 if (el_ret == ELEVATOR_BACK_MERGE) {
481 if (bio_attempt_back_merge(q, rq, bio)) {
486 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
487 if (bio_attempt_front_merge(q, rq, bio)) {
498 void blk_mq_add_timer(struct request *rq)
500 __blk_add_timer(rq, NULL);
504 * Run this hardware queue, pulling any software queues mapped to it in.
505 * Note that this function currently has various problems around ordering
506 * of IO. In particular, we'd like FIFO behaviour on handling existing
507 * items on the hctx->dispatch list. Ignore that for now.
509 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
511 struct request_queue *q = hctx->queue;
512 struct blk_mq_ctx *ctx;
517 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
523 * Touch any software queue that has pending entries.
525 for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
526 clear_bit(bit, hctx->ctx_map);
527 ctx = hctx->ctxs[bit];
528 BUG_ON(bit != ctx->index_hw);
530 spin_lock(&ctx->lock);
531 list_splice_tail_init(&ctx->rq_list, &rq_list);
532 spin_unlock(&ctx->lock);
536 * If we have previous entries on our dispatch list, grab them
537 * and stuff them at the front for more fair dispatch.
539 if (!list_empty_careful(&hctx->dispatch)) {
540 spin_lock(&hctx->lock);
541 if (!list_empty(&hctx->dispatch))
542 list_splice_init(&hctx->dispatch, &rq_list);
543 spin_unlock(&hctx->lock);
547 * Delete and return all entries from our dispatch list
552 * Now process all the entries, sending them to the driver.
554 while (!list_empty(&rq_list)) {
557 rq = list_first_entry(&rq_list, struct request, queuelist);
558 list_del_init(&rq->queuelist);
560 blk_mq_start_request(rq, list_empty(&rq_list));
562 ret = q->mq_ops->queue_rq(hctx, rq);
564 case BLK_MQ_RQ_QUEUE_OK:
567 case BLK_MQ_RQ_QUEUE_BUSY:
569 * FIXME: we should have a mechanism to stop the queue
570 * like blk_stop_queue, otherwise we will waste cpu
573 list_add(&rq->queuelist, &rq_list);
574 blk_mq_requeue_request(rq);
577 pr_err("blk-mq: bad return on queue: %d\n", ret);
578 case BLK_MQ_RQ_QUEUE_ERROR:
580 blk_mq_end_io(rq, rq->errors);
584 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
589 hctx->dispatched[0]++;
590 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
591 hctx->dispatched[ilog2(queued) + 1]++;
594 * Any items that need requeuing? Stuff them into hctx->dispatch,
595 * that is where we will continue on next queue run.
597 if (!list_empty(&rq_list)) {
598 spin_lock(&hctx->lock);
599 list_splice(&rq_list, &hctx->dispatch);
600 spin_unlock(&hctx->lock);
604 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
606 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
610 __blk_mq_run_hw_queue(hctx);
612 kblockd_schedule_delayed_work(&hctx->delayed_work, 0);
615 void blk_mq_run_queues(struct request_queue *q, bool async)
617 struct blk_mq_hw_ctx *hctx;
620 queue_for_each_hw_ctx(q, hctx, i) {
621 if ((!blk_mq_hctx_has_pending(hctx) &&
622 list_empty_careful(&hctx->dispatch)) ||
623 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
626 blk_mq_run_hw_queue(hctx, async);
629 EXPORT_SYMBOL(blk_mq_run_queues);
631 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
633 cancel_delayed_work(&hctx->delayed_work);
634 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
636 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
638 void blk_mq_stop_hw_queues(struct request_queue *q)
640 struct blk_mq_hw_ctx *hctx;
643 queue_for_each_hw_ctx(q, hctx, i)
644 blk_mq_stop_hw_queue(hctx);
646 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
648 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
650 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
651 __blk_mq_run_hw_queue(hctx);
653 EXPORT_SYMBOL(blk_mq_start_hw_queue);
655 void blk_mq_start_stopped_hw_queues(struct request_queue *q)
657 struct blk_mq_hw_ctx *hctx;
660 queue_for_each_hw_ctx(q, hctx, i) {
661 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
664 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
665 blk_mq_run_hw_queue(hctx, true);
668 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
670 static void blk_mq_work_fn(struct work_struct *work)
672 struct blk_mq_hw_ctx *hctx;
674 hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work);
675 __blk_mq_run_hw_queue(hctx);
678 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
679 struct request *rq, bool at_head)
681 struct blk_mq_ctx *ctx = rq->mq_ctx;
683 trace_block_rq_insert(hctx->queue, rq);
686 list_add(&rq->queuelist, &ctx->rq_list);
688 list_add_tail(&rq->queuelist, &ctx->rq_list);
689 blk_mq_hctx_mark_pending(hctx, ctx);
692 * We do this early, to ensure we are on the right CPU.
694 blk_mq_add_timer(rq);
697 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
700 struct request_queue *q = rq->q;
701 struct blk_mq_hw_ctx *hctx;
702 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
704 current_ctx = blk_mq_get_ctx(q);
705 if (!cpu_online(ctx->cpu))
706 rq->mq_ctx = ctx = current_ctx;
708 hctx = q->mq_ops->map_queue(q, ctx->cpu);
710 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
711 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
712 blk_insert_flush(rq);
714 spin_lock(&ctx->lock);
715 __blk_mq_insert_request(hctx, rq, at_head);
716 spin_unlock(&ctx->lock);
719 blk_mq_put_ctx(current_ctx);
722 blk_mq_run_hw_queue(hctx, async);
725 static void blk_mq_insert_requests(struct request_queue *q,
726 struct blk_mq_ctx *ctx,
727 struct list_head *list,
732 struct blk_mq_hw_ctx *hctx;
733 struct blk_mq_ctx *current_ctx;
735 trace_block_unplug(q, depth, !from_schedule);
737 current_ctx = blk_mq_get_ctx(q);
739 if (!cpu_online(ctx->cpu))
741 hctx = q->mq_ops->map_queue(q, ctx->cpu);
744 * preemption doesn't flush plug list, so it's possible ctx->cpu is
747 spin_lock(&ctx->lock);
748 while (!list_empty(list)) {
751 rq = list_first_entry(list, struct request, queuelist);
752 list_del_init(&rq->queuelist);
754 __blk_mq_insert_request(hctx, rq, false);
756 spin_unlock(&ctx->lock);
758 blk_mq_put_ctx(current_ctx);
760 blk_mq_run_hw_queue(hctx, from_schedule);
763 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
765 struct request *rqa = container_of(a, struct request, queuelist);
766 struct request *rqb = container_of(b, struct request, queuelist);
768 return !(rqa->mq_ctx < rqb->mq_ctx ||
769 (rqa->mq_ctx == rqb->mq_ctx &&
770 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
773 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
775 struct blk_mq_ctx *this_ctx;
776 struct request_queue *this_q;
782 list_splice_init(&plug->mq_list, &list);
784 list_sort(NULL, &list, plug_ctx_cmp);
790 while (!list_empty(&list)) {
791 rq = list_entry_rq(list.next);
792 list_del_init(&rq->queuelist);
794 if (rq->mq_ctx != this_ctx) {
796 blk_mq_insert_requests(this_q, this_ctx,
801 this_ctx = rq->mq_ctx;
807 list_add_tail(&rq->queuelist, &ctx_list);
811 * If 'this_ctx' is set, we know we have entries to complete
812 * on 'ctx_list'. Do those.
815 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
820 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
822 init_request_from_bio(rq, bio);
823 blk_account_io_start(rq, 1);
826 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
828 struct blk_mq_hw_ctx *hctx;
829 struct blk_mq_ctx *ctx;
830 const int is_sync = rw_is_sync(bio->bi_rw);
831 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
832 int rw = bio_data_dir(bio);
834 unsigned int use_plug, request_count = 0;
837 * If we have multiple hardware queues, just go directly to
838 * one of those for sync IO.
840 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
842 blk_queue_bounce(q, &bio);
844 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
845 bio_endio(bio, -EIO);
849 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
852 if (blk_mq_queue_enter(q)) {
853 bio_endio(bio, -EIO);
857 ctx = blk_mq_get_ctx(q);
858 hctx = q->mq_ops->map_queue(q, ctx->cpu);
862 trace_block_getrq(q, bio, rw);
863 rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
865 blk_mq_rq_ctx_init(q, ctx, rq, rw);
868 trace_block_sleeprq(q, bio, rw);
869 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
872 hctx = q->mq_ops->map_queue(q, ctx->cpu);
877 if (unlikely(is_flush_fua)) {
878 blk_mq_bio_to_request(rq, bio);
880 blk_insert_flush(rq);
885 * A task plug currently exists. Since this is completely lockless,
886 * utilize that to temporarily store requests until the task is
887 * either done or scheduled away.
890 struct blk_plug *plug = current->plug;
893 blk_mq_bio_to_request(rq, bio);
894 if (list_empty(&plug->mq_list))
896 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
897 blk_flush_plug_list(plug, false);
900 list_add_tail(&rq->queuelist, &plug->mq_list);
906 spin_lock(&ctx->lock);
908 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
909 blk_mq_attempt_merge(q, ctx, bio))
910 __blk_mq_free_request(hctx, ctx, rq);
912 blk_mq_bio_to_request(rq, bio);
913 __blk_mq_insert_request(hctx, rq, false);
916 spin_unlock(&ctx->lock);
920 * For a SYNC request, send it to the hardware immediately. For an
921 * ASYNC request, just ensure that we run it later on. The latter
922 * allows for merging opportunities and more efficient dispatching.
925 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
929 * Default mapping to a software queue, since we use one per CPU.
931 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
933 return q->queue_hw_ctx[q->mq_map[cpu]];
935 EXPORT_SYMBOL(blk_mq_map_queue);
937 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg,
938 unsigned int hctx_index)
940 return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
941 GFP_KERNEL | __GFP_ZERO, reg->numa_node);
943 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
945 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
946 unsigned int hctx_index)
950 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
952 static void blk_mq_hctx_notify(void *data, unsigned long action,
955 struct blk_mq_hw_ctx *hctx = data;
956 struct request_queue *q = hctx->queue;
957 struct blk_mq_ctx *ctx;
960 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
964 * Move ctx entries to new CPU, if this one is going away.
966 ctx = __blk_mq_get_ctx(q, cpu);
968 spin_lock(&ctx->lock);
969 if (!list_empty(&ctx->rq_list)) {
970 list_splice_init(&ctx->rq_list, &tmp);
971 clear_bit(ctx->index_hw, hctx->ctx_map);
973 spin_unlock(&ctx->lock);
975 if (list_empty(&tmp))
978 ctx = blk_mq_get_ctx(q);
979 spin_lock(&ctx->lock);
981 while (!list_empty(&tmp)) {
984 rq = list_first_entry(&tmp, struct request, queuelist);
986 list_move_tail(&rq->queuelist, &ctx->rq_list);
989 hctx = q->mq_ops->map_queue(q, ctx->cpu);
990 blk_mq_hctx_mark_pending(hctx, ctx);
992 spin_unlock(&ctx->lock);
995 blk_mq_run_hw_queue(hctx, true);
998 static int blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx,
999 int (*init)(void *, struct blk_mq_hw_ctx *,
1000 struct request *, unsigned int),
1006 for (i = 0; i < hctx->queue_depth; i++) {
1007 struct request *rq = hctx->rqs[i];
1009 ret = init(data, hctx, rq, i);
1017 int blk_mq_init_commands(struct request_queue *q,
1018 int (*init)(void *, struct blk_mq_hw_ctx *,
1019 struct request *, unsigned int),
1022 struct blk_mq_hw_ctx *hctx;
1026 queue_for_each_hw_ctx(q, hctx, i) {
1027 ret = blk_mq_init_hw_commands(hctx, init, data);
1034 EXPORT_SYMBOL(blk_mq_init_commands);
1036 static void blk_mq_free_hw_commands(struct blk_mq_hw_ctx *hctx,
1037 void (*free)(void *, struct blk_mq_hw_ctx *,
1038 struct request *, unsigned int),
1043 for (i = 0; i < hctx->queue_depth; i++) {
1044 struct request *rq = hctx->rqs[i];
1046 free(data, hctx, rq, i);
1050 void blk_mq_free_commands(struct request_queue *q,
1051 void (*free)(void *, struct blk_mq_hw_ctx *,
1052 struct request *, unsigned int),
1055 struct blk_mq_hw_ctx *hctx;
1058 queue_for_each_hw_ctx(q, hctx, i)
1059 blk_mq_free_hw_commands(hctx, free, data);
1061 EXPORT_SYMBOL(blk_mq_free_commands);
1063 static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx)
1067 while (!list_empty(&hctx->page_list)) {
1068 page = list_first_entry(&hctx->page_list, struct page, lru);
1069 list_del_init(&page->lru);
1070 __free_pages(page, page->private);
1076 blk_mq_free_tags(hctx->tags);
1079 static size_t order_to_size(unsigned int order)
1081 size_t ret = PAGE_SIZE;
1089 static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
1090 unsigned int reserved_tags, int node)
1092 unsigned int i, j, entries_per_page, max_order = 4;
1093 size_t rq_size, left;
1095 INIT_LIST_HEAD(&hctx->page_list);
1097 hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *),
1103 * rq_size is the size of the request plus driver payload, rounded
1104 * to the cacheline size
1106 rq_size = round_up(sizeof(struct request) + hctx->cmd_size,
1108 left = rq_size * hctx->queue_depth;
1110 for (i = 0; i < hctx->queue_depth;) {
1111 int this_order = max_order;
1116 while (left < order_to_size(this_order - 1) && this_order)
1120 page = alloc_pages_node(node, GFP_KERNEL, this_order);
1125 if (order_to_size(this_order) < rq_size)
1132 page->private = this_order;
1133 list_add_tail(&page->lru, &hctx->page_list);
1135 p = page_address(page);
1136 entries_per_page = order_to_size(this_order) / rq_size;
1137 to_do = min(entries_per_page, hctx->queue_depth - i);
1138 left -= to_do * rq_size;
1139 for (j = 0; j < to_do; j++) {
1141 blk_mq_rq_init(hctx, hctx->rqs[i]);
1147 if (i < (reserved_tags + BLK_MQ_TAG_MIN))
1149 else if (i != hctx->queue_depth) {
1150 hctx->queue_depth = i;
1151 pr_warn("%s: queue depth set to %u because of low memory\n",
1155 hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
1158 blk_mq_free_rq_map(hctx);
1165 static int blk_mq_init_hw_queues(struct request_queue *q,
1166 struct blk_mq_reg *reg, void *driver_data)
1168 struct blk_mq_hw_ctx *hctx;
1172 * Initialize hardware queues
1174 queue_for_each_hw_ctx(q, hctx, i) {
1175 unsigned int num_maps;
1178 node = hctx->numa_node;
1179 if (node == NUMA_NO_NODE)
1180 node = hctx->numa_node = reg->numa_node;
1182 INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
1183 spin_lock_init(&hctx->lock);
1184 INIT_LIST_HEAD(&hctx->dispatch);
1186 hctx->queue_num = i;
1187 hctx->flags = reg->flags;
1188 hctx->queue_depth = reg->queue_depth;
1189 hctx->cmd_size = reg->cmd_size;
1191 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1192 blk_mq_hctx_notify, hctx);
1193 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1195 if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
1199 * Allocate space for all possible cpus to avoid allocation in
1202 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1207 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1208 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1213 hctx->nr_ctx_map = num_maps;
1216 if (reg->ops->init_hctx &&
1217 reg->ops->init_hctx(hctx, driver_data, i))
1221 if (i == q->nr_hw_queues)
1227 queue_for_each_hw_ctx(q, hctx, j) {
1231 if (reg->ops->exit_hctx)
1232 reg->ops->exit_hctx(hctx, j);
1234 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1235 blk_mq_free_rq_map(hctx);
1242 static void blk_mq_init_cpu_queues(struct request_queue *q,
1243 unsigned int nr_hw_queues)
1247 for_each_possible_cpu(i) {
1248 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1249 struct blk_mq_hw_ctx *hctx;
1251 memset(__ctx, 0, sizeof(*__ctx));
1253 spin_lock_init(&__ctx->lock);
1254 INIT_LIST_HEAD(&__ctx->rq_list);
1257 /* If the cpu isn't online, the cpu is mapped to first hctx */
1258 hctx = q->mq_ops->map_queue(q, i);
1265 * Set local node, IFF we have more than one hw queue. If
1266 * not, we remain on the home node of the device
1268 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1269 hctx->numa_node = cpu_to_node(i);
1273 static void blk_mq_map_swqueue(struct request_queue *q)
1276 struct blk_mq_hw_ctx *hctx;
1277 struct blk_mq_ctx *ctx;
1279 queue_for_each_hw_ctx(q, hctx, i) {
1284 * Map software to hardware queues
1286 queue_for_each_ctx(q, ctx, i) {
1287 /* If the cpu isn't online, the cpu is mapped to first hctx */
1288 hctx = q->mq_ops->map_queue(q, i);
1289 ctx->index_hw = hctx->nr_ctx;
1290 hctx->ctxs[hctx->nr_ctx++] = ctx;
1294 struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg,
1297 struct blk_mq_hw_ctx **hctxs;
1298 struct blk_mq_ctx *ctx;
1299 struct request_queue *q;
1302 if (!reg->nr_hw_queues ||
1303 !reg->ops->queue_rq || !reg->ops->map_queue ||
1304 !reg->ops->alloc_hctx || !reg->ops->free_hctx)
1305 return ERR_PTR(-EINVAL);
1307 if (!reg->queue_depth)
1308 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1309 else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) {
1310 pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth);
1311 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1314 if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN))
1315 return ERR_PTR(-EINVAL);
1317 ctx = alloc_percpu(struct blk_mq_ctx);
1319 return ERR_PTR(-ENOMEM);
1321 hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1327 for (i = 0; i < reg->nr_hw_queues; i++) {
1328 hctxs[i] = reg->ops->alloc_hctx(reg, i);
1332 hctxs[i]->numa_node = NUMA_NO_NODE;
1333 hctxs[i]->queue_num = i;
1336 q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node);
1340 q->mq_map = blk_mq_make_queue_map(reg);
1344 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1345 blk_queue_rq_timeout(q, 30000);
1347 q->nr_queues = nr_cpu_ids;
1348 q->nr_hw_queues = reg->nr_hw_queues;
1351 q->queue_hw_ctx = hctxs;
1353 q->mq_ops = reg->ops;
1354 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1356 q->sg_reserved_size = INT_MAX;
1358 blk_queue_make_request(q, blk_mq_make_request);
1359 blk_queue_rq_timed_out(q, reg->ops->timeout);
1361 blk_queue_rq_timeout(q, reg->timeout);
1363 if (reg->ops->complete)
1364 blk_queue_softirq_done(q, reg->ops->complete);
1366 blk_mq_init_flush(q);
1367 blk_mq_init_cpu_queues(q, reg->nr_hw_queues);
1369 q->flush_rq = kzalloc(round_up(sizeof(struct request) + reg->cmd_size,
1370 cache_line_size()), GFP_KERNEL);
1374 if (blk_mq_init_hw_queues(q, reg, driver_data))
1377 blk_mq_map_swqueue(q);
1379 mutex_lock(&all_q_mutex);
1380 list_add_tail(&q->all_q_node, &all_q_list);
1381 mutex_unlock(&all_q_mutex);
1390 blk_cleanup_queue(q);
1392 for (i = 0; i < reg->nr_hw_queues; i++) {
1395 reg->ops->free_hctx(hctxs[i], i);
1400 return ERR_PTR(-ENOMEM);
1402 EXPORT_SYMBOL(blk_mq_init_queue);
1404 void blk_mq_free_queue(struct request_queue *q)
1406 struct blk_mq_hw_ctx *hctx;
1409 queue_for_each_hw_ctx(q, hctx, i) {
1410 kfree(hctx->ctx_map);
1412 blk_mq_free_rq_map(hctx);
1413 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1414 if (q->mq_ops->exit_hctx)
1415 q->mq_ops->exit_hctx(hctx, i);
1416 q->mq_ops->free_hctx(hctx, i);
1419 free_percpu(q->queue_ctx);
1420 kfree(q->queue_hw_ctx);
1423 q->queue_ctx = NULL;
1424 q->queue_hw_ctx = NULL;
1427 mutex_lock(&all_q_mutex);
1428 list_del_init(&q->all_q_node);
1429 mutex_unlock(&all_q_mutex);
1432 /* Basically redo blk_mq_init_queue with queue frozen */
1433 static void blk_mq_queue_reinit(struct request_queue *q)
1435 blk_mq_freeze_queue(q);
1437 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1440 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1441 * we should change hctx numa_node according to new topology (this
1442 * involves free and re-allocate memory, worthy doing?)
1445 blk_mq_map_swqueue(q);
1447 blk_mq_unfreeze_queue(q);
1450 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1451 unsigned long action, void *hcpu)
1453 struct request_queue *q;
1456 * Before new mapping is established, hotadded cpu might already start
1457 * handling requests. This doesn't break anything as we map offline
1458 * CPUs to first hardware queue. We will re-init queue below to get
1461 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1462 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1465 mutex_lock(&all_q_mutex);
1466 list_for_each_entry(q, &all_q_list, all_q_node)
1467 blk_mq_queue_reinit(q);
1468 mutex_unlock(&all_q_mutex);
1472 void blk_mq_disable_hotplug(void)
1474 mutex_lock(&all_q_mutex);
1477 void blk_mq_enable_hotplug(void)
1479 mutex_unlock(&all_q_mutex);
1482 static int __init blk_mq_init(void)
1486 /* Must be called after percpu_counter_hotcpu_callback() */
1487 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1491 subsys_initcall(blk_mq_init);