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) {
84 rq = hctx->tags->rqs[tag];
85 blk_rq_init(hctx->queue, rq);
94 static int blk_mq_queue_enter(struct request_queue *q)
98 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
100 /* we have problems to freeze the queue if it's initializing */
101 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
104 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
106 spin_lock_irq(q->queue_lock);
107 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
108 !blk_queue_bypass(q) || blk_queue_dying(q),
110 /* inc usage with lock hold to avoid freeze_queue runs here */
111 if (!ret && !blk_queue_dying(q))
112 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
113 else if (blk_queue_dying(q))
115 spin_unlock_irq(q->queue_lock);
120 static void blk_mq_queue_exit(struct request_queue *q)
122 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
125 static void __blk_mq_drain_queue(struct request_queue *q)
130 spin_lock_irq(q->queue_lock);
131 count = percpu_counter_sum(&q->mq_usage_counter);
132 spin_unlock_irq(q->queue_lock);
136 blk_mq_run_queues(q, false);
142 * Guarantee no request is in use, so we can change any data structure of
143 * the queue afterward.
145 static void blk_mq_freeze_queue(struct request_queue *q)
149 spin_lock_irq(q->queue_lock);
150 drain = !q->bypass_depth++;
151 queue_flag_set(QUEUE_FLAG_BYPASS, q);
152 spin_unlock_irq(q->queue_lock);
155 __blk_mq_drain_queue(q);
158 void blk_mq_drain_queue(struct request_queue *q)
160 __blk_mq_drain_queue(q);
163 static void blk_mq_unfreeze_queue(struct request_queue *q)
167 spin_lock_irq(q->queue_lock);
168 if (!--q->bypass_depth) {
169 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
172 WARN_ON_ONCE(q->bypass_depth < 0);
173 spin_unlock_irq(q->queue_lock);
175 wake_up_all(&q->mq_freeze_wq);
178 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
180 return blk_mq_has_free_tags(hctx->tags);
182 EXPORT_SYMBOL(blk_mq_can_queue);
184 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
185 struct request *rq, unsigned int rw_flags)
187 if (blk_queue_io_stat(q))
188 rw_flags |= REQ_IO_STAT;
191 rq->cmd_flags = rw_flags;
192 rq->start_time = jiffies;
193 set_start_time_ns(rq);
194 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
197 static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
204 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
205 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
207 rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved);
209 blk_mq_rq_ctx_init(q, ctx, rq, rw);
213 if (gfp & __GFP_WAIT) {
214 __blk_mq_run_hw_queue(hctx);
221 blk_mq_wait_for_tags(hctx->tags);
227 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp)
231 if (blk_mq_queue_enter(q))
234 rq = blk_mq_alloc_request_pinned(q, rw, gfp, false);
236 blk_mq_put_ctx(rq->mq_ctx);
240 struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
245 if (blk_mq_queue_enter(q))
248 rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
250 blk_mq_put_ctx(rq->mq_ctx);
253 EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
255 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
256 struct blk_mq_ctx *ctx, struct request *rq)
258 const int tag = rq->tag;
259 struct request_queue *q = rq->q;
261 blk_mq_put_tag(hctx->tags, tag);
262 blk_mq_queue_exit(q);
265 void blk_mq_free_request(struct request *rq)
267 struct blk_mq_ctx *ctx = rq->mq_ctx;
268 struct blk_mq_hw_ctx *hctx;
269 struct request_queue *q = rq->q;
271 ctx->rq_completed[rq_is_sync(rq)]++;
273 hctx = q->mq_ops->map_queue(q, ctx->cpu);
274 __blk_mq_free_request(hctx, ctx, rq);
278 * Clone all relevant state from a request that has been put on hold in
279 * the flush state machine into the preallocated flush request that hangs
280 * off the request queue.
282 * For a driver the flush request should be invisible, that's why we are
283 * impersonating the original request here.
285 void blk_mq_clone_flush_request(struct request *flush_rq,
286 struct request *orig_rq)
288 struct blk_mq_hw_ctx *hctx =
289 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
291 flush_rq->mq_ctx = orig_rq->mq_ctx;
292 flush_rq->tag = orig_rq->tag;
293 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
297 inline void __blk_mq_end_io(struct request *rq, int error)
299 blk_account_io_done(rq);
302 rq->end_io(rq, error);
304 blk_mq_free_request(rq);
306 EXPORT_SYMBOL(__blk_mq_end_io);
308 void blk_mq_end_io(struct request *rq, int error)
310 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
312 __blk_mq_end_io(rq, error);
314 EXPORT_SYMBOL(blk_mq_end_io);
316 static void __blk_mq_complete_request_remote(void *data)
318 struct request *rq = data;
320 rq->q->softirq_done_fn(rq);
323 void __blk_mq_complete_request(struct request *rq)
325 struct blk_mq_ctx *ctx = rq->mq_ctx;
328 if (!ctx->ipi_redirect) {
329 rq->q->softirq_done_fn(rq);
334 if (cpu != ctx->cpu && cpu_online(ctx->cpu)) {
335 rq->csd.func = __blk_mq_complete_request_remote;
338 smp_call_function_single_async(ctx->cpu, &rq->csd);
340 rq->q->softirq_done_fn(rq);
346 * blk_mq_complete_request - end I/O on a request
347 * @rq: the request being processed
350 * Ends all I/O on a request. It does not handle partial completions.
351 * The actual completion happens out-of-order, through a IPI handler.
353 void blk_mq_complete_request(struct request *rq)
355 if (unlikely(blk_should_fake_timeout(rq->q)))
357 if (!blk_mark_rq_complete(rq))
358 __blk_mq_complete_request(rq);
360 EXPORT_SYMBOL(blk_mq_complete_request);
362 static void blk_mq_start_request(struct request *rq, bool last)
364 struct request_queue *q = rq->q;
366 trace_block_rq_issue(q, rq);
368 rq->resid_len = blk_rq_bytes(rq);
371 * Just mark start time and set the started bit. Due to memory
372 * ordering, we know we'll see the correct deadline as long as
373 * REQ_ATOMIC_STARTED is seen.
375 rq->deadline = jiffies + q->rq_timeout;
376 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
378 if (q->dma_drain_size && blk_rq_bytes(rq)) {
380 * Make sure space for the drain appears. We know we can do
381 * this because max_hw_segments has been adjusted to be one
382 * fewer than the device can handle.
384 rq->nr_phys_segments++;
388 * Flag the last request in the series so that drivers know when IO
389 * should be kicked off, if they don't do it on a per-request basis.
391 * Note: the flag isn't the only condition drivers should do kick off.
392 * If drive is busy, the last request might not have the bit set.
395 rq->cmd_flags |= REQ_END;
398 static void blk_mq_requeue_request(struct request *rq)
400 struct request_queue *q = rq->q;
402 trace_block_rq_requeue(q, rq);
403 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
405 rq->cmd_flags &= ~REQ_END;
407 if (q->dma_drain_size && blk_rq_bytes(rq))
408 rq->nr_phys_segments--;
411 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
413 return tags->rqs[tag];
415 EXPORT_SYMBOL(blk_mq_tag_to_rq);
417 struct blk_mq_timeout_data {
418 struct blk_mq_hw_ctx *hctx;
420 unsigned int *next_set;
423 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
425 struct blk_mq_timeout_data *data = __data;
426 struct blk_mq_hw_ctx *hctx = data->hctx;
429 /* It may not be in flight yet (this is where
430 * the REQ_ATOMIC_STARTED flag comes in). The requests are
431 * statically allocated, so we know it's always safe to access the
432 * memory associated with a bit offset into ->rqs[].
438 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
439 if (tag >= hctx->tags->nr_tags)
442 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
443 if (rq->q != hctx->queue)
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 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
543 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
549 * Touch any software queue that has pending entries.
551 for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
552 clear_bit(bit, hctx->ctx_map);
553 ctx = hctx->ctxs[bit];
554 BUG_ON(bit != ctx->index_hw);
556 spin_lock(&ctx->lock);
557 list_splice_tail_init(&ctx->rq_list, &rq_list);
558 spin_unlock(&ctx->lock);
562 * If we have previous entries on our dispatch list, grab them
563 * and stuff them at the front for more fair dispatch.
565 if (!list_empty_careful(&hctx->dispatch)) {
566 spin_lock(&hctx->lock);
567 if (!list_empty(&hctx->dispatch))
568 list_splice_init(&hctx->dispatch, &rq_list);
569 spin_unlock(&hctx->lock);
573 * Delete and return all entries from our dispatch list
578 * Now process all the entries, sending them to the driver.
580 while (!list_empty(&rq_list)) {
583 rq = list_first_entry(&rq_list, struct request, queuelist);
584 list_del_init(&rq->queuelist);
586 blk_mq_start_request(rq, list_empty(&rq_list));
588 ret = q->mq_ops->queue_rq(hctx, rq);
590 case BLK_MQ_RQ_QUEUE_OK:
593 case BLK_MQ_RQ_QUEUE_BUSY:
595 * FIXME: we should have a mechanism to stop the queue
596 * like blk_stop_queue, otherwise we will waste cpu
599 list_add(&rq->queuelist, &rq_list);
600 blk_mq_requeue_request(rq);
603 pr_err("blk-mq: bad return on queue: %d\n", ret);
604 case BLK_MQ_RQ_QUEUE_ERROR:
606 blk_mq_end_io(rq, rq->errors);
610 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
615 hctx->dispatched[0]++;
616 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
617 hctx->dispatched[ilog2(queued) + 1]++;
620 * Any items that need requeuing? Stuff them into hctx->dispatch,
621 * that is where we will continue on next queue run.
623 if (!list_empty(&rq_list)) {
624 spin_lock(&hctx->lock);
625 list_splice(&rq_list, &hctx->dispatch);
626 spin_unlock(&hctx->lock);
630 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
632 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
635 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
636 __blk_mq_run_hw_queue(hctx);
637 else if (hctx->queue->nr_hw_queues == 1)
638 kblockd_schedule_delayed_work(&hctx->delayed_work, 0);
643 * It'd be great if the workqueue API had a way to pass
644 * in a mask and had some smarts for more clever placement
645 * than the first CPU. Or we could round-robin here. For now,
646 * just queue on the first CPU.
648 cpu = cpumask_first(hctx->cpumask);
649 kblockd_schedule_delayed_work_on(cpu, &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->state))
665 blk_mq_run_hw_queue(hctx, async);
669 EXPORT_SYMBOL(blk_mq_run_queues);
671 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
673 cancel_delayed_work(&hctx->delayed_work);
674 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
676 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
678 void blk_mq_stop_hw_queues(struct request_queue *q)
680 struct blk_mq_hw_ctx *hctx;
683 queue_for_each_hw_ctx(q, hctx, i)
684 blk_mq_stop_hw_queue(hctx);
686 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
688 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
690 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
693 __blk_mq_run_hw_queue(hctx);
696 EXPORT_SYMBOL(blk_mq_start_hw_queue);
698 void blk_mq_start_stopped_hw_queues(struct request_queue *q)
700 struct blk_mq_hw_ctx *hctx;
703 queue_for_each_hw_ctx(q, hctx, i) {
704 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
707 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
709 blk_mq_run_hw_queue(hctx, true);
713 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
715 static void blk_mq_work_fn(struct work_struct *work)
717 struct blk_mq_hw_ctx *hctx;
719 hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work);
721 __blk_mq_run_hw_queue(hctx);
724 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
725 struct request *rq, bool at_head)
727 struct blk_mq_ctx *ctx = rq->mq_ctx;
729 trace_block_rq_insert(hctx->queue, rq);
732 list_add(&rq->queuelist, &ctx->rq_list);
734 list_add_tail(&rq->queuelist, &ctx->rq_list);
735 blk_mq_hctx_mark_pending(hctx, ctx);
738 * We do this early, to ensure we are on the right CPU.
740 blk_mq_add_timer(rq);
743 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
746 struct request_queue *q = rq->q;
747 struct blk_mq_hw_ctx *hctx;
748 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
750 current_ctx = blk_mq_get_ctx(q);
751 if (!cpu_online(ctx->cpu))
752 rq->mq_ctx = ctx = current_ctx;
754 hctx = q->mq_ops->map_queue(q, ctx->cpu);
756 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
757 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
758 blk_insert_flush(rq);
760 spin_lock(&ctx->lock);
761 __blk_mq_insert_request(hctx, rq, at_head);
762 spin_unlock(&ctx->lock);
766 blk_mq_run_hw_queue(hctx, async);
768 blk_mq_put_ctx(current_ctx);
771 static void blk_mq_insert_requests(struct request_queue *q,
772 struct blk_mq_ctx *ctx,
773 struct list_head *list,
778 struct blk_mq_hw_ctx *hctx;
779 struct blk_mq_ctx *current_ctx;
781 trace_block_unplug(q, depth, !from_schedule);
783 current_ctx = blk_mq_get_ctx(q);
785 if (!cpu_online(ctx->cpu))
787 hctx = q->mq_ops->map_queue(q, ctx->cpu);
790 * preemption doesn't flush plug list, so it's possible ctx->cpu is
793 spin_lock(&ctx->lock);
794 while (!list_empty(list)) {
797 rq = list_first_entry(list, struct request, queuelist);
798 list_del_init(&rq->queuelist);
800 __blk_mq_insert_request(hctx, rq, false);
802 spin_unlock(&ctx->lock);
804 blk_mq_run_hw_queue(hctx, from_schedule);
805 blk_mq_put_ctx(current_ctx);
808 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
810 struct request *rqa = container_of(a, struct request, queuelist);
811 struct request *rqb = container_of(b, struct request, queuelist);
813 return !(rqa->mq_ctx < rqb->mq_ctx ||
814 (rqa->mq_ctx == rqb->mq_ctx &&
815 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
818 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
820 struct blk_mq_ctx *this_ctx;
821 struct request_queue *this_q;
827 list_splice_init(&plug->mq_list, &list);
829 list_sort(NULL, &list, plug_ctx_cmp);
835 while (!list_empty(&list)) {
836 rq = list_entry_rq(list.next);
837 list_del_init(&rq->queuelist);
839 if (rq->mq_ctx != this_ctx) {
841 blk_mq_insert_requests(this_q, this_ctx,
846 this_ctx = rq->mq_ctx;
852 list_add_tail(&rq->queuelist, &ctx_list);
856 * If 'this_ctx' is set, we know we have entries to complete
857 * on 'ctx_list'. Do those.
860 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
865 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
867 init_request_from_bio(rq, bio);
868 blk_account_io_start(rq, 1);
871 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
873 struct blk_mq_hw_ctx *hctx;
874 struct blk_mq_ctx *ctx;
875 const int is_sync = rw_is_sync(bio->bi_rw);
876 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
877 int rw = bio_data_dir(bio);
879 unsigned int use_plug, request_count = 0;
882 * If we have multiple hardware queues, just go directly to
883 * one of those for sync IO.
885 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
887 blk_queue_bounce(q, &bio);
889 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
890 bio_endio(bio, -EIO);
894 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
897 if (blk_mq_queue_enter(q)) {
898 bio_endio(bio, -EIO);
902 ctx = blk_mq_get_ctx(q);
903 hctx = q->mq_ops->map_queue(q, ctx->cpu);
907 trace_block_getrq(q, bio, rw);
908 rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
910 blk_mq_rq_ctx_init(q, ctx, rq, rw);
913 trace_block_sleeprq(q, bio, rw);
914 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
917 hctx = q->mq_ops->map_queue(q, ctx->cpu);
922 if (unlikely(is_flush_fua)) {
923 blk_mq_bio_to_request(rq, bio);
924 blk_insert_flush(rq);
929 * A task plug currently exists. Since this is completely lockless,
930 * utilize that to temporarily store requests until the task is
931 * either done or scheduled away.
934 struct blk_plug *plug = current->plug;
937 blk_mq_bio_to_request(rq, bio);
938 if (list_empty(&plug->mq_list))
940 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
941 blk_flush_plug_list(plug, false);
944 list_add_tail(&rq->queuelist, &plug->mq_list);
950 spin_lock(&ctx->lock);
952 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
953 blk_mq_attempt_merge(q, ctx, bio))
954 __blk_mq_free_request(hctx, ctx, rq);
956 blk_mq_bio_to_request(rq, bio);
957 __blk_mq_insert_request(hctx, rq, false);
960 spin_unlock(&ctx->lock);
963 * For a SYNC request, send it to the hardware immediately. For an
964 * ASYNC request, just ensure that we run it later on. The latter
965 * allows for merging opportunities and more efficient dispatching.
968 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
973 * Default mapping to a software queue, since we use one per CPU.
975 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
977 return q->queue_hw_ctx[q->mq_map[cpu]];
979 EXPORT_SYMBOL(blk_mq_map_queue);
981 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *set,
982 unsigned int hctx_index)
984 return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
985 GFP_KERNEL | __GFP_ZERO, set->numa_node);
987 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
989 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
990 unsigned int hctx_index)
994 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
996 static void blk_mq_hctx_notify(void *data, unsigned long action,
999 struct blk_mq_hw_ctx *hctx = data;
1000 struct request_queue *q = hctx->queue;
1001 struct blk_mq_ctx *ctx;
1004 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1008 * Move ctx entries to new CPU, if this one is going away.
1010 ctx = __blk_mq_get_ctx(q, cpu);
1012 spin_lock(&ctx->lock);
1013 if (!list_empty(&ctx->rq_list)) {
1014 list_splice_init(&ctx->rq_list, &tmp);
1015 clear_bit(ctx->index_hw, hctx->ctx_map);
1017 spin_unlock(&ctx->lock);
1019 if (list_empty(&tmp))
1022 ctx = blk_mq_get_ctx(q);
1023 spin_lock(&ctx->lock);
1025 while (!list_empty(&tmp)) {
1028 rq = list_first_entry(&tmp, struct request, queuelist);
1030 list_move_tail(&rq->queuelist, &ctx->rq_list);
1033 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1034 blk_mq_hctx_mark_pending(hctx, ctx);
1036 spin_unlock(&ctx->lock);
1038 blk_mq_run_hw_queue(hctx, true);
1039 blk_mq_put_ctx(ctx);
1042 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1043 struct blk_mq_tags *tags, unsigned int hctx_idx)
1047 if (tags->rqs && set->ops->exit_request) {
1050 for (i = 0; i < tags->nr_tags; i++) {
1053 set->ops->exit_request(set->driver_data, tags->rqs[i],
1058 while (!list_empty(&tags->page_list)) {
1059 page = list_first_entry(&tags->page_list, struct page, lru);
1060 list_del_init(&page->lru);
1061 __free_pages(page, page->private);
1066 blk_mq_free_tags(tags);
1069 static size_t order_to_size(unsigned int order)
1071 size_t ret = PAGE_SIZE;
1079 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1080 unsigned int hctx_idx)
1082 struct blk_mq_tags *tags;
1083 unsigned int i, j, entries_per_page, max_order = 4;
1084 size_t rq_size, left;
1086 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1091 INIT_LIST_HEAD(&tags->page_list);
1093 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1094 GFP_KERNEL, set->numa_node);
1096 blk_mq_free_tags(tags);
1101 * rq_size is the size of the request plus driver payload, rounded
1102 * to the cacheline size
1104 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1106 left = rq_size * set->queue_depth;
1108 for (i = 0; i < set->queue_depth; ) {
1109 int this_order = max_order;
1114 while (left < order_to_size(this_order - 1) && this_order)
1118 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1124 if (order_to_size(this_order) < rq_size)
1131 page->private = this_order;
1132 list_add_tail(&page->lru, &tags->page_list);
1134 p = page_address(page);
1135 entries_per_page = order_to_size(this_order) / rq_size;
1136 to_do = min(entries_per_page, set->queue_depth - i);
1137 left -= to_do * rq_size;
1138 for (j = 0; j < to_do; j++) {
1140 if (set->ops->init_request) {
1141 if (set->ops->init_request(set->driver_data,
1142 tags->rqs[i], hctx_idx, i,
1155 pr_warn("%s: failed to allocate requests\n", __func__);
1156 blk_mq_free_rq_map(set, tags, hctx_idx);
1160 static int blk_mq_init_hw_queues(struct request_queue *q,
1161 struct blk_mq_tag_set *set)
1163 struct blk_mq_hw_ctx *hctx;
1167 * Initialize hardware queues
1169 queue_for_each_hw_ctx(q, hctx, i) {
1170 unsigned int num_maps;
1173 node = hctx->numa_node;
1174 if (node == NUMA_NO_NODE)
1175 node = hctx->numa_node = set->numa_node;
1177 INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
1178 spin_lock_init(&hctx->lock);
1179 INIT_LIST_HEAD(&hctx->dispatch);
1181 hctx->queue_num = i;
1182 hctx->flags = set->flags;
1183 hctx->cmd_size = set->cmd_size;
1185 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1186 blk_mq_hctx_notify, hctx);
1187 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1189 hctx->tags = set->tags[i];
1192 * Allocate space for all possible cpus to avoid allocation in
1195 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1200 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1201 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1206 hctx->nr_ctx_map = num_maps;
1209 if (set->ops->init_hctx &&
1210 set->ops->init_hctx(hctx, set->driver_data, i))
1214 if (i == q->nr_hw_queues)
1220 queue_for_each_hw_ctx(q, hctx, j) {
1224 if (set->ops->exit_hctx)
1225 set->ops->exit_hctx(hctx, j);
1227 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1234 static void blk_mq_init_cpu_queues(struct request_queue *q,
1235 unsigned int nr_hw_queues)
1239 for_each_possible_cpu(i) {
1240 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1241 struct blk_mq_hw_ctx *hctx;
1243 memset(__ctx, 0, sizeof(*__ctx));
1245 spin_lock_init(&__ctx->lock);
1246 INIT_LIST_HEAD(&__ctx->rq_list);
1249 /* If the cpu isn't online, the cpu is mapped to first hctx */
1253 hctx = q->mq_ops->map_queue(q, i);
1254 cpumask_set_cpu(i, hctx->cpumask);
1258 * Set local node, IFF we have more than one hw queue. If
1259 * not, we remain on the home node of the device
1261 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1262 hctx->numa_node = cpu_to_node(i);
1266 static void blk_mq_map_swqueue(struct request_queue *q)
1269 struct blk_mq_hw_ctx *hctx;
1270 struct blk_mq_ctx *ctx;
1272 queue_for_each_hw_ctx(q, hctx, i) {
1273 cpumask_clear(hctx->cpumask);
1278 * Map software to hardware queues
1280 queue_for_each_ctx(q, ctx, i) {
1281 /* If the cpu isn't online, the cpu is mapped to first hctx */
1285 hctx = q->mq_ops->map_queue(q, i);
1286 cpumask_set_cpu(i, hctx->cpumask);
1287 ctx->index_hw = hctx->nr_ctx;
1288 hctx->ctxs[hctx->nr_ctx++] = ctx;
1292 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1294 struct blk_mq_hw_ctx **hctxs;
1295 struct blk_mq_ctx *ctx;
1296 struct request_queue *q;
1299 ctx = alloc_percpu(struct blk_mq_ctx);
1301 return ERR_PTR(-ENOMEM);
1303 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1309 for (i = 0; i < set->nr_hw_queues; i++) {
1310 hctxs[i] = set->ops->alloc_hctx(set, i);
1314 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1317 hctxs[i]->numa_node = NUMA_NO_NODE;
1318 hctxs[i]->queue_num = i;
1321 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1325 q->mq_map = blk_mq_make_queue_map(set);
1329 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1330 blk_queue_rq_timeout(q, 30000);
1332 q->nr_queues = nr_cpu_ids;
1333 q->nr_hw_queues = set->nr_hw_queues;
1336 q->queue_hw_ctx = hctxs;
1338 q->mq_ops = set->ops;
1339 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1341 q->sg_reserved_size = INT_MAX;
1343 blk_queue_make_request(q, blk_mq_make_request);
1344 blk_queue_rq_timed_out(q, set->ops->timeout);
1346 blk_queue_rq_timeout(q, set->timeout);
1348 if (set->ops->complete)
1349 blk_queue_softirq_done(q, set->ops->complete);
1351 blk_mq_init_flush(q);
1352 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1354 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1355 set->cmd_size, cache_line_size()),
1360 if (blk_mq_init_hw_queues(q, set))
1363 blk_mq_map_swqueue(q);
1365 mutex_lock(&all_q_mutex);
1366 list_add_tail(&q->all_q_node, &all_q_list);
1367 mutex_unlock(&all_q_mutex);
1376 blk_cleanup_queue(q);
1378 for (i = 0; i < set->nr_hw_queues; i++) {
1381 free_cpumask_var(hctxs[i]->cpumask);
1382 set->ops->free_hctx(hctxs[i], i);
1387 return ERR_PTR(-ENOMEM);
1389 EXPORT_SYMBOL(blk_mq_init_queue);
1391 void blk_mq_free_queue(struct request_queue *q)
1393 struct blk_mq_hw_ctx *hctx;
1396 queue_for_each_hw_ctx(q, hctx, i) {
1397 kfree(hctx->ctx_map);
1399 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1400 if (q->mq_ops->exit_hctx)
1401 q->mq_ops->exit_hctx(hctx, i);
1402 free_cpumask_var(hctx->cpumask);
1403 q->mq_ops->free_hctx(hctx, i);
1406 free_percpu(q->queue_ctx);
1407 kfree(q->queue_hw_ctx);
1410 q->queue_ctx = NULL;
1411 q->queue_hw_ctx = NULL;
1414 mutex_lock(&all_q_mutex);
1415 list_del_init(&q->all_q_node);
1416 mutex_unlock(&all_q_mutex);
1419 /* Basically redo blk_mq_init_queue with queue frozen */
1420 static void blk_mq_queue_reinit(struct request_queue *q)
1422 blk_mq_freeze_queue(q);
1424 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1427 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1428 * we should change hctx numa_node according to new topology (this
1429 * involves free and re-allocate memory, worthy doing?)
1432 blk_mq_map_swqueue(q);
1434 blk_mq_unfreeze_queue(q);
1437 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1438 unsigned long action, void *hcpu)
1440 struct request_queue *q;
1443 * Before new mapping is established, hotadded cpu might already start
1444 * handling requests. This doesn't break anything as we map offline
1445 * CPUs to first hardware queue. We will re-init queue below to get
1448 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1449 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1452 mutex_lock(&all_q_mutex);
1453 list_for_each_entry(q, &all_q_list, all_q_node)
1454 blk_mq_queue_reinit(q);
1455 mutex_unlock(&all_q_mutex);
1459 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1463 if (!set->nr_hw_queues)
1465 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1467 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1470 if (!set->nr_hw_queues ||
1471 !set->ops->queue_rq || !set->ops->map_queue ||
1472 !set->ops->alloc_hctx || !set->ops->free_hctx)
1476 set->tags = kmalloc_node(set->nr_hw_queues * sizeof(struct blk_mq_tags),
1477 GFP_KERNEL, set->numa_node);
1481 for (i = 0; i < set->nr_hw_queues; i++) {
1482 set->tags[i] = blk_mq_init_rq_map(set, i);
1491 blk_mq_free_rq_map(set, set->tags[i], i);
1495 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
1497 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
1501 for (i = 0; i < set->nr_hw_queues; i++)
1502 blk_mq_free_rq_map(set, set->tags[i], i);
1504 EXPORT_SYMBOL(blk_mq_free_tag_set);
1506 void blk_mq_disable_hotplug(void)
1508 mutex_lock(&all_q_mutex);
1511 void blk_mq_enable_hotplug(void)
1513 mutex_unlock(&all_q_mutex);
1516 static int __init blk_mq_init(void)
1520 /* Must be called after percpu_counter_hotcpu_callback() */
1521 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1525 subsys_initcall(blk_mq_init);