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 if (unlikely(blk_bidi_rq(rq)))
305 blk_mq_free_request(rq->next_rq);
306 blk_mq_free_request(rq);
309 EXPORT_SYMBOL(__blk_mq_end_io);
311 void blk_mq_end_io(struct request *rq, int error)
313 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
315 __blk_mq_end_io(rq, error);
317 EXPORT_SYMBOL(blk_mq_end_io);
319 static void __blk_mq_complete_request_remote(void *data)
321 struct request *rq = data;
323 rq->q->softirq_done_fn(rq);
326 void __blk_mq_complete_request(struct request *rq)
328 struct blk_mq_ctx *ctx = rq->mq_ctx;
331 if (!ctx->ipi_redirect) {
332 rq->q->softirq_done_fn(rq);
337 if (cpu != ctx->cpu && cpu_online(ctx->cpu)) {
338 rq->csd.func = __blk_mq_complete_request_remote;
341 smp_call_function_single_async(ctx->cpu, &rq->csd);
343 rq->q->softirq_done_fn(rq);
349 * blk_mq_complete_request - end I/O on a request
350 * @rq: the request being processed
353 * Ends all I/O on a request. It does not handle partial completions.
354 * The actual completion happens out-of-order, through a IPI handler.
356 void blk_mq_complete_request(struct request *rq)
358 if (unlikely(blk_should_fake_timeout(rq->q)))
360 if (!blk_mark_rq_complete(rq))
361 __blk_mq_complete_request(rq);
363 EXPORT_SYMBOL(blk_mq_complete_request);
365 static void blk_mq_start_request(struct request *rq, bool last)
367 struct request_queue *q = rq->q;
369 trace_block_rq_issue(q, rq);
371 rq->resid_len = blk_rq_bytes(rq);
372 if (unlikely(blk_bidi_rq(rq)))
373 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
376 * Just mark start time and set the started bit. Due to memory
377 * ordering, we know we'll see the correct deadline as long as
378 * REQ_ATOMIC_STARTED is seen.
380 rq->deadline = jiffies + q->rq_timeout;
381 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
383 if (q->dma_drain_size && blk_rq_bytes(rq)) {
385 * Make sure space for the drain appears. We know we can do
386 * this because max_hw_segments has been adjusted to be one
387 * fewer than the device can handle.
389 rq->nr_phys_segments++;
393 * Flag the last request in the series so that drivers know when IO
394 * should be kicked off, if they don't do it on a per-request basis.
396 * Note: the flag isn't the only condition drivers should do kick off.
397 * If drive is busy, the last request might not have the bit set.
400 rq->cmd_flags |= REQ_END;
403 static void blk_mq_requeue_request(struct request *rq)
405 struct request_queue *q = rq->q;
407 trace_block_rq_requeue(q, rq);
408 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
410 rq->cmd_flags &= ~REQ_END;
412 if (q->dma_drain_size && blk_rq_bytes(rq))
413 rq->nr_phys_segments--;
416 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
418 return tags->rqs[tag];
420 EXPORT_SYMBOL(blk_mq_tag_to_rq);
422 struct blk_mq_timeout_data {
423 struct blk_mq_hw_ctx *hctx;
425 unsigned int *next_set;
428 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
430 struct blk_mq_timeout_data *data = __data;
431 struct blk_mq_hw_ctx *hctx = data->hctx;
434 /* It may not be in flight yet (this is where
435 * the REQ_ATOMIC_STARTED flag comes in). The requests are
436 * statically allocated, so we know it's always safe to access the
437 * memory associated with a bit offset into ->rqs[].
443 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
444 if (tag >= hctx->tags->nr_tags)
447 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
448 if (rq->q != hctx->queue)
450 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
453 blk_rq_check_expired(rq, data->next, data->next_set);
457 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
459 unsigned int *next_set)
461 struct blk_mq_timeout_data data = {
464 .next_set = next_set,
468 * Ask the tagging code to iterate busy requests, so we can
469 * check them for timeout.
471 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
474 static void blk_mq_rq_timer(unsigned long data)
476 struct request_queue *q = (struct request_queue *) data;
477 struct blk_mq_hw_ctx *hctx;
478 unsigned long next = 0;
481 queue_for_each_hw_ctx(q, hctx, i)
482 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
485 mod_timer(&q->timeout, round_jiffies_up(next));
489 * Reverse check our software queue for entries that we could potentially
490 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
491 * too much time checking for merges.
493 static bool blk_mq_attempt_merge(struct request_queue *q,
494 struct blk_mq_ctx *ctx, struct bio *bio)
499 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
505 if (!blk_rq_merge_ok(rq, bio))
508 el_ret = blk_try_merge(rq, bio);
509 if (el_ret == ELEVATOR_BACK_MERGE) {
510 if (bio_attempt_back_merge(q, rq, bio)) {
515 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
516 if (bio_attempt_front_merge(q, rq, bio)) {
527 void blk_mq_add_timer(struct request *rq)
529 __blk_add_timer(rq, NULL);
533 * Run this hardware queue, pulling any software queues mapped to it in.
534 * Note that this function currently has various problems around ordering
535 * of IO. In particular, we'd like FIFO behaviour on handling existing
536 * items on the hctx->dispatch list. Ignore that for now.
538 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
540 struct request_queue *q = hctx->queue;
541 struct blk_mq_ctx *ctx;
546 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
548 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
554 * Touch any software queue that has pending entries.
556 for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
557 clear_bit(bit, hctx->ctx_map);
558 ctx = hctx->ctxs[bit];
559 BUG_ON(bit != ctx->index_hw);
561 spin_lock(&ctx->lock);
562 list_splice_tail_init(&ctx->rq_list, &rq_list);
563 spin_unlock(&ctx->lock);
567 * If we have previous entries on our dispatch list, grab them
568 * and stuff them at the front for more fair dispatch.
570 if (!list_empty_careful(&hctx->dispatch)) {
571 spin_lock(&hctx->lock);
572 if (!list_empty(&hctx->dispatch))
573 list_splice_init(&hctx->dispatch, &rq_list);
574 spin_unlock(&hctx->lock);
578 * Delete and return all entries from our dispatch list
583 * Now process all the entries, sending them to the driver.
585 while (!list_empty(&rq_list)) {
588 rq = list_first_entry(&rq_list, struct request, queuelist);
589 list_del_init(&rq->queuelist);
591 blk_mq_start_request(rq, list_empty(&rq_list));
593 ret = q->mq_ops->queue_rq(hctx, rq);
595 case BLK_MQ_RQ_QUEUE_OK:
598 case BLK_MQ_RQ_QUEUE_BUSY:
600 * FIXME: we should have a mechanism to stop the queue
601 * like blk_stop_queue, otherwise we will waste cpu
604 list_add(&rq->queuelist, &rq_list);
605 blk_mq_requeue_request(rq);
608 pr_err("blk-mq: bad return on queue: %d\n", ret);
609 case BLK_MQ_RQ_QUEUE_ERROR:
611 blk_mq_end_io(rq, rq->errors);
615 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
620 hctx->dispatched[0]++;
621 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
622 hctx->dispatched[ilog2(queued) + 1]++;
625 * Any items that need requeuing? Stuff them into hctx->dispatch,
626 * that is where we will continue on next queue run.
628 if (!list_empty(&rq_list)) {
629 spin_lock(&hctx->lock);
630 list_splice(&rq_list, &hctx->dispatch);
631 spin_unlock(&hctx->lock);
635 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
637 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
640 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
641 __blk_mq_run_hw_queue(hctx);
642 else if (hctx->queue->nr_hw_queues == 1)
643 kblockd_schedule_delayed_work(&hctx->run_work, 0);
648 * It'd be great if the workqueue API had a way to pass
649 * in a mask and had some smarts for more clever placement
650 * than the first CPU. Or we could round-robin here. For now,
651 * just queue on the first CPU.
653 cpu = cpumask_first(hctx->cpumask);
654 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
658 void blk_mq_run_queues(struct request_queue *q, bool async)
660 struct blk_mq_hw_ctx *hctx;
663 queue_for_each_hw_ctx(q, hctx, i) {
664 if ((!blk_mq_hctx_has_pending(hctx) &&
665 list_empty_careful(&hctx->dispatch)) ||
666 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
670 blk_mq_run_hw_queue(hctx, async);
674 EXPORT_SYMBOL(blk_mq_run_queues);
676 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
678 cancel_delayed_work(&hctx->run_work);
679 cancel_delayed_work(&hctx->delay_work);
680 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
682 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
684 void blk_mq_stop_hw_queues(struct request_queue *q)
686 struct blk_mq_hw_ctx *hctx;
689 queue_for_each_hw_ctx(q, hctx, i)
690 blk_mq_stop_hw_queue(hctx);
692 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
694 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
696 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
699 __blk_mq_run_hw_queue(hctx);
702 EXPORT_SYMBOL(blk_mq_start_hw_queue);
704 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
706 struct blk_mq_hw_ctx *hctx;
709 queue_for_each_hw_ctx(q, hctx, i) {
710 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
713 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
715 blk_mq_run_hw_queue(hctx, async);
719 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
721 static void blk_mq_run_work_fn(struct work_struct *work)
723 struct blk_mq_hw_ctx *hctx;
725 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
727 __blk_mq_run_hw_queue(hctx);
730 static void blk_mq_delay_work_fn(struct work_struct *work)
732 struct blk_mq_hw_ctx *hctx;
734 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
736 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
737 __blk_mq_run_hw_queue(hctx);
740 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
742 unsigned long tmo = msecs_to_jiffies(msecs);
744 if (hctx->queue->nr_hw_queues == 1)
745 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
750 * It'd be great if the workqueue API had a way to pass
751 * in a mask and had some smarts for more clever placement
752 * than the first CPU. Or we could round-robin here. For now,
753 * just queue on the first CPU.
755 cpu = cpumask_first(hctx->cpumask);
756 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
759 EXPORT_SYMBOL(blk_mq_delay_queue);
761 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
762 struct request *rq, bool at_head)
764 struct blk_mq_ctx *ctx = rq->mq_ctx;
766 trace_block_rq_insert(hctx->queue, rq);
769 list_add(&rq->queuelist, &ctx->rq_list);
771 list_add_tail(&rq->queuelist, &ctx->rq_list);
772 blk_mq_hctx_mark_pending(hctx, ctx);
775 * We do this early, to ensure we are on the right CPU.
777 blk_mq_add_timer(rq);
780 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
783 struct request_queue *q = rq->q;
784 struct blk_mq_hw_ctx *hctx;
785 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
787 current_ctx = blk_mq_get_ctx(q);
788 if (!cpu_online(ctx->cpu))
789 rq->mq_ctx = ctx = current_ctx;
791 hctx = q->mq_ops->map_queue(q, ctx->cpu);
793 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
794 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
795 blk_insert_flush(rq);
797 spin_lock(&ctx->lock);
798 __blk_mq_insert_request(hctx, rq, at_head);
799 spin_unlock(&ctx->lock);
803 blk_mq_run_hw_queue(hctx, async);
805 blk_mq_put_ctx(current_ctx);
808 static void blk_mq_insert_requests(struct request_queue *q,
809 struct blk_mq_ctx *ctx,
810 struct list_head *list,
815 struct blk_mq_hw_ctx *hctx;
816 struct blk_mq_ctx *current_ctx;
818 trace_block_unplug(q, depth, !from_schedule);
820 current_ctx = blk_mq_get_ctx(q);
822 if (!cpu_online(ctx->cpu))
824 hctx = q->mq_ops->map_queue(q, ctx->cpu);
827 * preemption doesn't flush plug list, so it's possible ctx->cpu is
830 spin_lock(&ctx->lock);
831 while (!list_empty(list)) {
834 rq = list_first_entry(list, struct request, queuelist);
835 list_del_init(&rq->queuelist);
837 __blk_mq_insert_request(hctx, rq, false);
839 spin_unlock(&ctx->lock);
841 blk_mq_run_hw_queue(hctx, from_schedule);
842 blk_mq_put_ctx(current_ctx);
845 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
847 struct request *rqa = container_of(a, struct request, queuelist);
848 struct request *rqb = container_of(b, struct request, queuelist);
850 return !(rqa->mq_ctx < rqb->mq_ctx ||
851 (rqa->mq_ctx == rqb->mq_ctx &&
852 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
855 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
857 struct blk_mq_ctx *this_ctx;
858 struct request_queue *this_q;
864 list_splice_init(&plug->mq_list, &list);
866 list_sort(NULL, &list, plug_ctx_cmp);
872 while (!list_empty(&list)) {
873 rq = list_entry_rq(list.next);
874 list_del_init(&rq->queuelist);
876 if (rq->mq_ctx != this_ctx) {
878 blk_mq_insert_requests(this_q, this_ctx,
883 this_ctx = rq->mq_ctx;
889 list_add_tail(&rq->queuelist, &ctx_list);
893 * If 'this_ctx' is set, we know we have entries to complete
894 * on 'ctx_list'. Do those.
897 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
902 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
904 init_request_from_bio(rq, bio);
905 blk_account_io_start(rq, 1);
908 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
910 struct blk_mq_hw_ctx *hctx;
911 struct blk_mq_ctx *ctx;
912 const int is_sync = rw_is_sync(bio->bi_rw);
913 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
914 int rw = bio_data_dir(bio);
916 unsigned int use_plug, request_count = 0;
919 * If we have multiple hardware queues, just go directly to
920 * one of those for sync IO.
922 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
924 blk_queue_bounce(q, &bio);
926 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
927 bio_endio(bio, -EIO);
931 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
934 if (blk_mq_queue_enter(q)) {
935 bio_endio(bio, -EIO);
939 ctx = blk_mq_get_ctx(q);
940 hctx = q->mq_ops->map_queue(q, ctx->cpu);
944 trace_block_getrq(q, bio, rw);
945 rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
947 blk_mq_rq_ctx_init(q, ctx, rq, rw);
950 trace_block_sleeprq(q, bio, rw);
951 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
954 hctx = q->mq_ops->map_queue(q, ctx->cpu);
959 if (unlikely(is_flush_fua)) {
960 blk_mq_bio_to_request(rq, bio);
961 blk_insert_flush(rq);
966 * A task plug currently exists. Since this is completely lockless,
967 * utilize that to temporarily store requests until the task is
968 * either done or scheduled away.
971 struct blk_plug *plug = current->plug;
974 blk_mq_bio_to_request(rq, bio);
975 if (list_empty(&plug->mq_list))
977 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
978 blk_flush_plug_list(plug, false);
981 list_add_tail(&rq->queuelist, &plug->mq_list);
987 spin_lock(&ctx->lock);
989 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
990 blk_mq_attempt_merge(q, ctx, bio))
991 __blk_mq_free_request(hctx, ctx, rq);
993 blk_mq_bio_to_request(rq, bio);
994 __blk_mq_insert_request(hctx, rq, false);
997 spin_unlock(&ctx->lock);
1000 * For a SYNC request, send it to the hardware immediately. For an
1001 * ASYNC request, just ensure that we run it later on. The latter
1002 * allows for merging opportunities and more efficient dispatching.
1005 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
1006 blk_mq_put_ctx(ctx);
1010 * Default mapping to a software queue, since we use one per CPU.
1012 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1014 return q->queue_hw_ctx[q->mq_map[cpu]];
1016 EXPORT_SYMBOL(blk_mq_map_queue);
1018 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *set,
1019 unsigned int hctx_index)
1021 return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
1022 GFP_KERNEL | __GFP_ZERO, set->numa_node);
1024 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1026 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1027 unsigned int hctx_index)
1031 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1033 static void blk_mq_hctx_notify(void *data, unsigned long action,
1036 struct blk_mq_hw_ctx *hctx = data;
1037 struct request_queue *q = hctx->queue;
1038 struct blk_mq_ctx *ctx;
1041 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1045 * Move ctx entries to new CPU, if this one is going away.
1047 ctx = __blk_mq_get_ctx(q, cpu);
1049 spin_lock(&ctx->lock);
1050 if (!list_empty(&ctx->rq_list)) {
1051 list_splice_init(&ctx->rq_list, &tmp);
1052 clear_bit(ctx->index_hw, hctx->ctx_map);
1054 spin_unlock(&ctx->lock);
1056 if (list_empty(&tmp))
1059 ctx = blk_mq_get_ctx(q);
1060 spin_lock(&ctx->lock);
1062 while (!list_empty(&tmp)) {
1065 rq = list_first_entry(&tmp, struct request, queuelist);
1067 list_move_tail(&rq->queuelist, &ctx->rq_list);
1070 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1071 blk_mq_hctx_mark_pending(hctx, ctx);
1073 spin_unlock(&ctx->lock);
1075 blk_mq_run_hw_queue(hctx, true);
1076 blk_mq_put_ctx(ctx);
1079 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1080 struct blk_mq_tags *tags, unsigned int hctx_idx)
1084 if (tags->rqs && set->ops->exit_request) {
1087 for (i = 0; i < tags->nr_tags; i++) {
1090 set->ops->exit_request(set->driver_data, tags->rqs[i],
1095 while (!list_empty(&tags->page_list)) {
1096 page = list_first_entry(&tags->page_list, struct page, lru);
1097 list_del_init(&page->lru);
1098 __free_pages(page, page->private);
1103 blk_mq_free_tags(tags);
1106 static size_t order_to_size(unsigned int order)
1108 size_t ret = PAGE_SIZE;
1116 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1117 unsigned int hctx_idx)
1119 struct blk_mq_tags *tags;
1120 unsigned int i, j, entries_per_page, max_order = 4;
1121 size_t rq_size, left;
1123 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1128 INIT_LIST_HEAD(&tags->page_list);
1130 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1131 GFP_KERNEL, set->numa_node);
1133 blk_mq_free_tags(tags);
1138 * rq_size is the size of the request plus driver payload, rounded
1139 * to the cacheline size
1141 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1143 left = rq_size * set->queue_depth;
1145 for (i = 0; i < set->queue_depth; ) {
1146 int this_order = max_order;
1151 while (left < order_to_size(this_order - 1) && this_order)
1155 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1161 if (order_to_size(this_order) < rq_size)
1168 page->private = this_order;
1169 list_add_tail(&page->lru, &tags->page_list);
1171 p = page_address(page);
1172 entries_per_page = order_to_size(this_order) / rq_size;
1173 to_do = min(entries_per_page, set->queue_depth - i);
1174 left -= to_do * rq_size;
1175 for (j = 0; j < to_do; j++) {
1177 if (set->ops->init_request) {
1178 if (set->ops->init_request(set->driver_data,
1179 tags->rqs[i], hctx_idx, i,
1192 pr_warn("%s: failed to allocate requests\n", __func__);
1193 blk_mq_free_rq_map(set, tags, hctx_idx);
1197 static int blk_mq_init_hw_queues(struct request_queue *q,
1198 struct blk_mq_tag_set *set)
1200 struct blk_mq_hw_ctx *hctx;
1204 * Initialize hardware queues
1206 queue_for_each_hw_ctx(q, hctx, i) {
1207 unsigned int num_maps;
1210 node = hctx->numa_node;
1211 if (node == NUMA_NO_NODE)
1212 node = hctx->numa_node = set->numa_node;
1214 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1215 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1216 spin_lock_init(&hctx->lock);
1217 INIT_LIST_HEAD(&hctx->dispatch);
1219 hctx->queue_num = i;
1220 hctx->flags = set->flags;
1221 hctx->cmd_size = set->cmd_size;
1223 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1224 blk_mq_hctx_notify, hctx);
1225 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1227 hctx->tags = set->tags[i];
1230 * Allocate space for all possible cpus to avoid allocation in
1233 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1238 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1239 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1244 hctx->nr_ctx_map = num_maps;
1247 if (set->ops->init_hctx &&
1248 set->ops->init_hctx(hctx, set->driver_data, i))
1252 if (i == q->nr_hw_queues)
1258 queue_for_each_hw_ctx(q, hctx, j) {
1262 if (set->ops->exit_hctx)
1263 set->ops->exit_hctx(hctx, j);
1265 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1272 static void blk_mq_init_cpu_queues(struct request_queue *q,
1273 unsigned int nr_hw_queues)
1277 for_each_possible_cpu(i) {
1278 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1279 struct blk_mq_hw_ctx *hctx;
1281 memset(__ctx, 0, sizeof(*__ctx));
1283 spin_lock_init(&__ctx->lock);
1284 INIT_LIST_HEAD(&__ctx->rq_list);
1287 /* If the cpu isn't online, the cpu is mapped to first hctx */
1291 hctx = q->mq_ops->map_queue(q, i);
1292 cpumask_set_cpu(i, hctx->cpumask);
1296 * Set local node, IFF we have more than one hw queue. If
1297 * not, we remain on the home node of the device
1299 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1300 hctx->numa_node = cpu_to_node(i);
1304 static void blk_mq_map_swqueue(struct request_queue *q)
1307 struct blk_mq_hw_ctx *hctx;
1308 struct blk_mq_ctx *ctx;
1310 queue_for_each_hw_ctx(q, hctx, i) {
1311 cpumask_clear(hctx->cpumask);
1316 * Map software to hardware queues
1318 queue_for_each_ctx(q, ctx, i) {
1319 /* If the cpu isn't online, the cpu is mapped to first hctx */
1323 hctx = q->mq_ops->map_queue(q, i);
1324 cpumask_set_cpu(i, hctx->cpumask);
1325 ctx->index_hw = hctx->nr_ctx;
1326 hctx->ctxs[hctx->nr_ctx++] = ctx;
1330 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1332 struct blk_mq_hw_ctx **hctxs;
1333 struct blk_mq_ctx *ctx;
1334 struct request_queue *q;
1337 ctx = alloc_percpu(struct blk_mq_ctx);
1339 return ERR_PTR(-ENOMEM);
1341 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1347 for (i = 0; i < set->nr_hw_queues; i++) {
1348 hctxs[i] = set->ops->alloc_hctx(set, i);
1352 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1355 hctxs[i]->numa_node = NUMA_NO_NODE;
1356 hctxs[i]->queue_num = i;
1359 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1363 q->mq_map = blk_mq_make_queue_map(set);
1367 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1368 blk_queue_rq_timeout(q, 30000);
1370 q->nr_queues = nr_cpu_ids;
1371 q->nr_hw_queues = set->nr_hw_queues;
1374 q->queue_hw_ctx = hctxs;
1376 q->mq_ops = set->ops;
1377 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1379 q->sg_reserved_size = INT_MAX;
1381 blk_queue_make_request(q, blk_mq_make_request);
1382 blk_queue_rq_timed_out(q, set->ops->timeout);
1384 blk_queue_rq_timeout(q, set->timeout);
1386 if (set->ops->complete)
1387 blk_queue_softirq_done(q, set->ops->complete);
1389 blk_mq_init_flush(q);
1390 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1392 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1393 set->cmd_size, cache_line_size()),
1398 if (blk_mq_init_hw_queues(q, set))
1401 blk_mq_map_swqueue(q);
1403 mutex_lock(&all_q_mutex);
1404 list_add_tail(&q->all_q_node, &all_q_list);
1405 mutex_unlock(&all_q_mutex);
1414 blk_cleanup_queue(q);
1416 for (i = 0; i < set->nr_hw_queues; i++) {
1419 free_cpumask_var(hctxs[i]->cpumask);
1420 set->ops->free_hctx(hctxs[i], i);
1425 return ERR_PTR(-ENOMEM);
1427 EXPORT_SYMBOL(blk_mq_init_queue);
1429 void blk_mq_free_queue(struct request_queue *q)
1431 struct blk_mq_hw_ctx *hctx;
1434 queue_for_each_hw_ctx(q, hctx, i) {
1435 kfree(hctx->ctx_map);
1437 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1438 if (q->mq_ops->exit_hctx)
1439 q->mq_ops->exit_hctx(hctx, i);
1440 free_cpumask_var(hctx->cpumask);
1441 q->mq_ops->free_hctx(hctx, i);
1444 free_percpu(q->queue_ctx);
1445 kfree(q->queue_hw_ctx);
1448 q->queue_ctx = NULL;
1449 q->queue_hw_ctx = NULL;
1452 mutex_lock(&all_q_mutex);
1453 list_del_init(&q->all_q_node);
1454 mutex_unlock(&all_q_mutex);
1457 /* Basically redo blk_mq_init_queue with queue frozen */
1458 static void blk_mq_queue_reinit(struct request_queue *q)
1460 blk_mq_freeze_queue(q);
1462 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1465 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1466 * we should change hctx numa_node according to new topology (this
1467 * involves free and re-allocate memory, worthy doing?)
1470 blk_mq_map_swqueue(q);
1472 blk_mq_unfreeze_queue(q);
1475 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1476 unsigned long action, void *hcpu)
1478 struct request_queue *q;
1481 * Before new mapping is established, hotadded cpu might already start
1482 * handling requests. This doesn't break anything as we map offline
1483 * CPUs to first hardware queue. We will re-init queue below to get
1486 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1487 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1490 mutex_lock(&all_q_mutex);
1491 list_for_each_entry(q, &all_q_list, all_q_node)
1492 blk_mq_queue_reinit(q);
1493 mutex_unlock(&all_q_mutex);
1497 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1501 if (!set->nr_hw_queues)
1503 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1505 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1508 if (!set->nr_hw_queues ||
1509 !set->ops->queue_rq || !set->ops->map_queue ||
1510 !set->ops->alloc_hctx || !set->ops->free_hctx)
1514 set->tags = kmalloc_node(set->nr_hw_queues * sizeof(struct blk_mq_tags),
1515 GFP_KERNEL, set->numa_node);
1519 for (i = 0; i < set->nr_hw_queues; i++) {
1520 set->tags[i] = blk_mq_init_rq_map(set, i);
1529 blk_mq_free_rq_map(set, set->tags[i], i);
1533 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
1535 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
1539 for (i = 0; i < set->nr_hw_queues; i++)
1540 blk_mq_free_rq_map(set, set->tags[i], i);
1542 EXPORT_SYMBOL(blk_mq_free_tag_set);
1544 void blk_mq_disable_hotplug(void)
1546 mutex_lock(&all_q_mutex);
1549 void blk_mq_enable_hotplug(void)
1551 mutex_unlock(&all_q_mutex);
1554 static int __init blk_mq_init(void)
1558 /* Must be called after percpu_counter_hotcpu_callback() */
1559 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1563 subsys_initcall(blk_mq_init);