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, reserved);
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;
332 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
333 rq->q->softirq_done_fn(rq);
338 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
339 shared = cpus_share_cache(cpu, ctx->cpu);
341 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
342 rq->csd.func = __blk_mq_complete_request_remote;
345 smp_call_function_single_async(ctx->cpu, &rq->csd);
347 rq->q->softirq_done_fn(rq);
353 * blk_mq_complete_request - end I/O on a request
354 * @rq: the request being processed
357 * Ends all I/O on a request. It does not handle partial completions.
358 * The actual completion happens out-of-order, through a IPI handler.
360 void blk_mq_complete_request(struct request *rq)
362 if (unlikely(blk_should_fake_timeout(rq->q)))
364 if (!blk_mark_rq_complete(rq))
365 __blk_mq_complete_request(rq);
367 EXPORT_SYMBOL(blk_mq_complete_request);
369 static void blk_mq_start_request(struct request *rq, bool last)
371 struct request_queue *q = rq->q;
373 trace_block_rq_issue(q, rq);
375 rq->resid_len = blk_rq_bytes(rq);
376 if (unlikely(blk_bidi_rq(rq)))
377 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
380 * Just mark start time and set the started bit. Due to memory
381 * ordering, we know we'll see the correct deadline as long as
382 * REQ_ATOMIC_STARTED is seen.
384 rq->deadline = jiffies + q->rq_timeout;
387 * Mark us as started and clear complete. Complete might have been
388 * set if requeue raced with timeout, which then marked it as
389 * complete. So be sure to clear complete again when we start
390 * the request, otherwise we'll ignore the completion event.
392 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
393 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
395 if (q->dma_drain_size && blk_rq_bytes(rq)) {
397 * Make sure space for the drain appears. We know we can do
398 * this because max_hw_segments has been adjusted to be one
399 * fewer than the device can handle.
401 rq->nr_phys_segments++;
405 * Flag the last request in the series so that drivers know when IO
406 * should be kicked off, if they don't do it on a per-request basis.
408 * Note: the flag isn't the only condition drivers should do kick off.
409 * If drive is busy, the last request might not have the bit set.
412 rq->cmd_flags |= REQ_END;
415 static void __blk_mq_requeue_request(struct request *rq)
417 struct request_queue *q = rq->q;
419 trace_block_rq_requeue(q, rq);
420 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
422 rq->cmd_flags &= ~REQ_END;
424 if (q->dma_drain_size && blk_rq_bytes(rq))
425 rq->nr_phys_segments--;
428 void blk_mq_requeue_request(struct request *rq)
430 struct request_queue *q = rq->q;
432 __blk_mq_requeue_request(rq);
433 blk_clear_rq_complete(rq);
435 trace_block_rq_requeue(q, rq);
437 BUG_ON(blk_queued_rq(rq));
438 blk_mq_insert_request(rq, true, true, false);
440 EXPORT_SYMBOL(blk_mq_requeue_request);
442 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
444 return tags->rqs[tag];
446 EXPORT_SYMBOL(blk_mq_tag_to_rq);
448 struct blk_mq_timeout_data {
449 struct blk_mq_hw_ctx *hctx;
451 unsigned int *next_set;
454 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
456 struct blk_mq_timeout_data *data = __data;
457 struct blk_mq_hw_ctx *hctx = data->hctx;
460 /* It may not be in flight yet (this is where
461 * the REQ_ATOMIC_STARTED flag comes in). The requests are
462 * statically allocated, so we know it's always safe to access the
463 * memory associated with a bit offset into ->rqs[].
469 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
470 if (tag >= hctx->tags->nr_tags)
473 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
474 if (rq->q != hctx->queue)
476 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
479 blk_rq_check_expired(rq, data->next, data->next_set);
483 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
485 unsigned int *next_set)
487 struct blk_mq_timeout_data data = {
490 .next_set = next_set,
494 * Ask the tagging code to iterate busy requests, so we can
495 * check them for timeout.
497 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
500 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
502 struct request_queue *q = rq->q;
505 * We know that complete is set at this point. If STARTED isn't set
506 * anymore, then the request isn't active and the "timeout" should
507 * just be ignored. This can happen due to the bitflag ordering.
508 * Timeout first checks if STARTED is set, and if it is, assumes
509 * the request is active. But if we race with completion, then
510 * we both flags will get cleared. So check here again, and ignore
511 * a timeout event with a request that isn't active.
513 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
514 return BLK_EH_NOT_HANDLED;
516 if (!q->mq_ops->timeout)
517 return BLK_EH_RESET_TIMER;
519 return q->mq_ops->timeout(rq);
522 static void blk_mq_rq_timer(unsigned long data)
524 struct request_queue *q = (struct request_queue *) data;
525 struct blk_mq_hw_ctx *hctx;
526 unsigned long next = 0;
529 queue_for_each_hw_ctx(q, hctx, i)
530 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
533 mod_timer(&q->timeout, round_jiffies_up(next));
537 * Reverse check our software queue for entries that we could potentially
538 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
539 * too much time checking for merges.
541 static bool blk_mq_attempt_merge(struct request_queue *q,
542 struct blk_mq_ctx *ctx, struct bio *bio)
547 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
553 if (!blk_rq_merge_ok(rq, bio))
556 el_ret = blk_try_merge(rq, bio);
557 if (el_ret == ELEVATOR_BACK_MERGE) {
558 if (bio_attempt_back_merge(q, rq, bio)) {
563 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
564 if (bio_attempt_front_merge(q, rq, bio)) {
576 * Run this hardware queue, pulling any software queues mapped to it in.
577 * Note that this function currently has various problems around ordering
578 * of IO. In particular, we'd like FIFO behaviour on handling existing
579 * items on the hctx->dispatch list. Ignore that for now.
581 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
583 struct request_queue *q = hctx->queue;
584 struct blk_mq_ctx *ctx;
589 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
591 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
597 * Touch any software queue that has pending entries.
599 for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
600 clear_bit(bit, hctx->ctx_map);
601 ctx = hctx->ctxs[bit];
603 spin_lock(&ctx->lock);
604 list_splice_tail_init(&ctx->rq_list, &rq_list);
605 spin_unlock(&ctx->lock);
609 * If we have previous entries on our dispatch list, grab them
610 * and stuff them at the front for more fair dispatch.
612 if (!list_empty_careful(&hctx->dispatch)) {
613 spin_lock(&hctx->lock);
614 if (!list_empty(&hctx->dispatch))
615 list_splice_init(&hctx->dispatch, &rq_list);
616 spin_unlock(&hctx->lock);
620 * Delete and return all entries from our dispatch list
625 * Now process all the entries, sending them to the driver.
627 while (!list_empty(&rq_list)) {
630 rq = list_first_entry(&rq_list, struct request, queuelist);
631 list_del_init(&rq->queuelist);
633 blk_mq_start_request(rq, list_empty(&rq_list));
635 ret = q->mq_ops->queue_rq(hctx, rq);
637 case BLK_MQ_RQ_QUEUE_OK:
640 case BLK_MQ_RQ_QUEUE_BUSY:
642 * FIXME: we should have a mechanism to stop the queue
643 * like blk_stop_queue, otherwise we will waste cpu
646 list_add(&rq->queuelist, &rq_list);
647 __blk_mq_requeue_request(rq);
650 pr_err("blk-mq: bad return on queue: %d\n", ret);
651 case BLK_MQ_RQ_QUEUE_ERROR:
653 blk_mq_end_io(rq, rq->errors);
657 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
662 hctx->dispatched[0]++;
663 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
664 hctx->dispatched[ilog2(queued) + 1]++;
667 * Any items that need requeuing? Stuff them into hctx->dispatch,
668 * that is where we will continue on next queue run.
670 if (!list_empty(&rq_list)) {
671 spin_lock(&hctx->lock);
672 list_splice(&rq_list, &hctx->dispatch);
673 spin_unlock(&hctx->lock);
677 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
679 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
682 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
683 __blk_mq_run_hw_queue(hctx);
684 else if (hctx->queue->nr_hw_queues == 1)
685 kblockd_schedule_delayed_work(&hctx->run_work, 0);
690 * It'd be great if the workqueue API had a way to pass
691 * in a mask and had some smarts for more clever placement
692 * than the first CPU. Or we could round-robin here. For now,
693 * just queue on the first CPU.
695 cpu = cpumask_first(hctx->cpumask);
696 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
700 void blk_mq_run_queues(struct request_queue *q, bool async)
702 struct blk_mq_hw_ctx *hctx;
705 queue_for_each_hw_ctx(q, hctx, i) {
706 if ((!blk_mq_hctx_has_pending(hctx) &&
707 list_empty_careful(&hctx->dispatch)) ||
708 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
712 blk_mq_run_hw_queue(hctx, async);
716 EXPORT_SYMBOL(blk_mq_run_queues);
718 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
720 cancel_delayed_work(&hctx->run_work);
721 cancel_delayed_work(&hctx->delay_work);
722 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
724 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
726 void blk_mq_stop_hw_queues(struct request_queue *q)
728 struct blk_mq_hw_ctx *hctx;
731 queue_for_each_hw_ctx(q, hctx, i)
732 blk_mq_stop_hw_queue(hctx);
734 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
736 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
738 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
741 __blk_mq_run_hw_queue(hctx);
744 EXPORT_SYMBOL(blk_mq_start_hw_queue);
746 void blk_mq_start_hw_queues(struct request_queue *q)
748 struct blk_mq_hw_ctx *hctx;
751 queue_for_each_hw_ctx(q, hctx, i)
752 blk_mq_start_hw_queue(hctx);
754 EXPORT_SYMBOL(blk_mq_start_hw_queues);
757 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
759 struct blk_mq_hw_ctx *hctx;
762 queue_for_each_hw_ctx(q, hctx, i) {
763 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
766 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
768 blk_mq_run_hw_queue(hctx, async);
772 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
774 static void blk_mq_run_work_fn(struct work_struct *work)
776 struct blk_mq_hw_ctx *hctx;
778 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
780 __blk_mq_run_hw_queue(hctx);
783 static void blk_mq_delay_work_fn(struct work_struct *work)
785 struct blk_mq_hw_ctx *hctx;
787 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
789 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
790 __blk_mq_run_hw_queue(hctx);
793 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
795 unsigned long tmo = msecs_to_jiffies(msecs);
797 if (hctx->queue->nr_hw_queues == 1)
798 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
803 * It'd be great if the workqueue API had a way to pass
804 * in a mask and had some smarts for more clever placement
805 * than the first CPU. Or we could round-robin here. For now,
806 * just queue on the first CPU.
808 cpu = cpumask_first(hctx->cpumask);
809 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
812 EXPORT_SYMBOL(blk_mq_delay_queue);
814 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
815 struct request *rq, bool at_head)
817 struct blk_mq_ctx *ctx = rq->mq_ctx;
819 trace_block_rq_insert(hctx->queue, rq);
822 list_add(&rq->queuelist, &ctx->rq_list);
824 list_add_tail(&rq->queuelist, &ctx->rq_list);
825 blk_mq_hctx_mark_pending(hctx, ctx);
828 * We do this early, to ensure we are on the right CPU.
833 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
836 struct request_queue *q = rq->q;
837 struct blk_mq_hw_ctx *hctx;
838 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
840 current_ctx = blk_mq_get_ctx(q);
841 if (!cpu_online(ctx->cpu))
842 rq->mq_ctx = ctx = current_ctx;
844 hctx = q->mq_ops->map_queue(q, ctx->cpu);
846 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
847 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
848 blk_insert_flush(rq);
850 spin_lock(&ctx->lock);
851 __blk_mq_insert_request(hctx, rq, at_head);
852 spin_unlock(&ctx->lock);
856 blk_mq_run_hw_queue(hctx, async);
858 blk_mq_put_ctx(current_ctx);
861 static void blk_mq_insert_requests(struct request_queue *q,
862 struct blk_mq_ctx *ctx,
863 struct list_head *list,
868 struct blk_mq_hw_ctx *hctx;
869 struct blk_mq_ctx *current_ctx;
871 trace_block_unplug(q, depth, !from_schedule);
873 current_ctx = blk_mq_get_ctx(q);
875 if (!cpu_online(ctx->cpu))
877 hctx = q->mq_ops->map_queue(q, ctx->cpu);
880 * preemption doesn't flush plug list, so it's possible ctx->cpu is
883 spin_lock(&ctx->lock);
884 while (!list_empty(list)) {
887 rq = list_first_entry(list, struct request, queuelist);
888 list_del_init(&rq->queuelist);
890 __blk_mq_insert_request(hctx, rq, false);
892 spin_unlock(&ctx->lock);
894 blk_mq_run_hw_queue(hctx, from_schedule);
895 blk_mq_put_ctx(current_ctx);
898 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
900 struct request *rqa = container_of(a, struct request, queuelist);
901 struct request *rqb = container_of(b, struct request, queuelist);
903 return !(rqa->mq_ctx < rqb->mq_ctx ||
904 (rqa->mq_ctx == rqb->mq_ctx &&
905 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
908 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
910 struct blk_mq_ctx *this_ctx;
911 struct request_queue *this_q;
917 list_splice_init(&plug->mq_list, &list);
919 list_sort(NULL, &list, plug_ctx_cmp);
925 while (!list_empty(&list)) {
926 rq = list_entry_rq(list.next);
927 list_del_init(&rq->queuelist);
929 if (rq->mq_ctx != this_ctx) {
931 blk_mq_insert_requests(this_q, this_ctx,
936 this_ctx = rq->mq_ctx;
942 list_add_tail(&rq->queuelist, &ctx_list);
946 * If 'this_ctx' is set, we know we have entries to complete
947 * on 'ctx_list'. Do those.
950 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
955 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
957 init_request_from_bio(rq, bio);
958 blk_account_io_start(rq, 1);
961 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
963 struct blk_mq_hw_ctx *hctx;
964 struct blk_mq_ctx *ctx;
965 const int is_sync = rw_is_sync(bio->bi_rw);
966 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
967 int rw = bio_data_dir(bio);
969 unsigned int use_plug, request_count = 0;
972 * If we have multiple hardware queues, just go directly to
973 * one of those for sync IO.
975 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
977 blk_queue_bounce(q, &bio);
979 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
980 bio_endio(bio, -EIO);
984 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
987 if (blk_mq_queue_enter(q)) {
988 bio_endio(bio, -EIO);
992 ctx = blk_mq_get_ctx(q);
993 hctx = q->mq_ops->map_queue(q, ctx->cpu);
997 trace_block_getrq(q, bio, rw);
998 rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
1000 blk_mq_rq_ctx_init(q, ctx, rq, rw);
1002 blk_mq_put_ctx(ctx);
1003 trace_block_sleeprq(q, bio, rw);
1004 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
1007 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1012 if (unlikely(is_flush_fua)) {
1013 blk_mq_bio_to_request(rq, bio);
1014 blk_insert_flush(rq);
1019 * A task plug currently exists. Since this is completely lockless,
1020 * utilize that to temporarily store requests until the task is
1021 * either done or scheduled away.
1024 struct blk_plug *plug = current->plug;
1027 blk_mq_bio_to_request(rq, bio);
1028 if (list_empty(&plug->mq_list))
1029 trace_block_plug(q);
1030 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1031 blk_flush_plug_list(plug, false);
1032 trace_block_plug(q);
1034 list_add_tail(&rq->queuelist, &plug->mq_list);
1035 blk_mq_put_ctx(ctx);
1040 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1041 init_request_from_bio(rq, bio);
1043 spin_lock(&ctx->lock);
1045 __blk_mq_insert_request(hctx, rq, false);
1046 spin_unlock(&ctx->lock);
1047 blk_account_io_start(rq, 1);
1049 spin_lock(&ctx->lock);
1050 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1051 init_request_from_bio(rq, bio);
1055 spin_unlock(&ctx->lock);
1056 __blk_mq_free_request(hctx, ctx, rq);
1061 * For a SYNC request, send it to the hardware immediately. For an
1062 * ASYNC request, just ensure that we run it later on. The latter
1063 * allows for merging opportunities and more efficient dispatching.
1066 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
1067 blk_mq_put_ctx(ctx);
1071 * Default mapping to a software queue, since we use one per CPU.
1073 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1075 return q->queue_hw_ctx[q->mq_map[cpu]];
1077 EXPORT_SYMBOL(blk_mq_map_queue);
1079 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *set,
1080 unsigned int hctx_index)
1082 return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
1083 GFP_KERNEL | __GFP_ZERO, set->numa_node);
1085 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1087 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1088 unsigned int hctx_index)
1092 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1094 static void blk_mq_hctx_notify(void *data, unsigned long action,
1097 struct blk_mq_hw_ctx *hctx = data;
1098 struct request_queue *q = hctx->queue;
1099 struct blk_mq_ctx *ctx;
1102 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1106 * Move ctx entries to new CPU, if this one is going away.
1108 ctx = __blk_mq_get_ctx(q, cpu);
1110 spin_lock(&ctx->lock);
1111 if (!list_empty(&ctx->rq_list)) {
1112 list_splice_init(&ctx->rq_list, &tmp);
1113 clear_bit(ctx->index_hw, hctx->ctx_map);
1115 spin_unlock(&ctx->lock);
1117 if (list_empty(&tmp))
1120 ctx = blk_mq_get_ctx(q);
1121 spin_lock(&ctx->lock);
1123 while (!list_empty(&tmp)) {
1126 rq = list_first_entry(&tmp, struct request, queuelist);
1128 list_move_tail(&rq->queuelist, &ctx->rq_list);
1131 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1132 blk_mq_hctx_mark_pending(hctx, ctx);
1134 spin_unlock(&ctx->lock);
1136 blk_mq_run_hw_queue(hctx, true);
1137 blk_mq_put_ctx(ctx);
1140 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1141 struct blk_mq_tags *tags, unsigned int hctx_idx)
1145 if (tags->rqs && set->ops->exit_request) {
1148 for (i = 0; i < tags->nr_tags; i++) {
1151 set->ops->exit_request(set->driver_data, tags->rqs[i],
1156 while (!list_empty(&tags->page_list)) {
1157 page = list_first_entry(&tags->page_list, struct page, lru);
1158 list_del_init(&page->lru);
1159 __free_pages(page, page->private);
1164 blk_mq_free_tags(tags);
1167 static size_t order_to_size(unsigned int order)
1169 return (size_t)PAGE_SIZE << order;
1172 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1173 unsigned int hctx_idx)
1175 struct blk_mq_tags *tags;
1176 unsigned int i, j, entries_per_page, max_order = 4;
1177 size_t rq_size, left;
1179 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1184 INIT_LIST_HEAD(&tags->page_list);
1186 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1187 GFP_KERNEL, set->numa_node);
1189 blk_mq_free_tags(tags);
1194 * rq_size is the size of the request plus driver payload, rounded
1195 * to the cacheline size
1197 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1199 left = rq_size * set->queue_depth;
1201 for (i = 0; i < set->queue_depth; ) {
1202 int this_order = max_order;
1207 while (left < order_to_size(this_order - 1) && this_order)
1211 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1217 if (order_to_size(this_order) < rq_size)
1224 page->private = this_order;
1225 list_add_tail(&page->lru, &tags->page_list);
1227 p = page_address(page);
1228 entries_per_page = order_to_size(this_order) / rq_size;
1229 to_do = min(entries_per_page, set->queue_depth - i);
1230 left -= to_do * rq_size;
1231 for (j = 0; j < to_do; j++) {
1233 if (set->ops->init_request) {
1234 if (set->ops->init_request(set->driver_data,
1235 tags->rqs[i], hctx_idx, i,
1248 pr_warn("%s: failed to allocate requests\n", __func__);
1249 blk_mq_free_rq_map(set, tags, hctx_idx);
1253 static int blk_mq_init_hw_queues(struct request_queue *q,
1254 struct blk_mq_tag_set *set)
1256 struct blk_mq_hw_ctx *hctx;
1260 * Initialize hardware queues
1262 queue_for_each_hw_ctx(q, hctx, i) {
1263 unsigned int num_maps;
1266 node = hctx->numa_node;
1267 if (node == NUMA_NO_NODE)
1268 node = hctx->numa_node = set->numa_node;
1270 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1271 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1272 spin_lock_init(&hctx->lock);
1273 INIT_LIST_HEAD(&hctx->dispatch);
1275 hctx->queue_num = i;
1276 hctx->flags = set->flags;
1277 hctx->cmd_size = set->cmd_size;
1279 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1280 blk_mq_hctx_notify, hctx);
1281 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1283 hctx->tags = set->tags[i];
1286 * Allocate space for all possible cpus to avoid allocation in
1289 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1294 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1295 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1300 hctx->nr_ctx_map = num_maps;
1303 if (set->ops->init_hctx &&
1304 set->ops->init_hctx(hctx, set->driver_data, i))
1308 if (i == q->nr_hw_queues)
1314 queue_for_each_hw_ctx(q, hctx, j) {
1318 if (set->ops->exit_hctx)
1319 set->ops->exit_hctx(hctx, j);
1321 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1323 kfree(hctx->ctx_map);
1329 static void blk_mq_init_cpu_queues(struct request_queue *q,
1330 unsigned int nr_hw_queues)
1334 for_each_possible_cpu(i) {
1335 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1336 struct blk_mq_hw_ctx *hctx;
1338 memset(__ctx, 0, sizeof(*__ctx));
1340 spin_lock_init(&__ctx->lock);
1341 INIT_LIST_HEAD(&__ctx->rq_list);
1344 /* If the cpu isn't online, the cpu is mapped to first hctx */
1348 hctx = q->mq_ops->map_queue(q, i);
1349 cpumask_set_cpu(i, hctx->cpumask);
1353 * Set local node, IFF we have more than one hw queue. If
1354 * not, we remain on the home node of the device
1356 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1357 hctx->numa_node = cpu_to_node(i);
1361 static void blk_mq_map_swqueue(struct request_queue *q)
1364 struct blk_mq_hw_ctx *hctx;
1365 struct blk_mq_ctx *ctx;
1367 queue_for_each_hw_ctx(q, hctx, i) {
1368 cpumask_clear(hctx->cpumask);
1373 * Map software to hardware queues
1375 queue_for_each_ctx(q, ctx, i) {
1376 /* If the cpu isn't online, the cpu is mapped to first hctx */
1380 hctx = q->mq_ops->map_queue(q, i);
1381 cpumask_set_cpu(i, hctx->cpumask);
1382 ctx->index_hw = hctx->nr_ctx;
1383 hctx->ctxs[hctx->nr_ctx++] = ctx;
1387 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1389 struct blk_mq_hw_ctx **hctxs;
1390 struct blk_mq_ctx *ctx;
1391 struct request_queue *q;
1394 ctx = alloc_percpu(struct blk_mq_ctx);
1396 return ERR_PTR(-ENOMEM);
1398 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1404 for (i = 0; i < set->nr_hw_queues; i++) {
1405 hctxs[i] = set->ops->alloc_hctx(set, i);
1409 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1412 hctxs[i]->numa_node = NUMA_NO_NODE;
1413 hctxs[i]->queue_num = i;
1416 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1420 q->mq_map = blk_mq_make_queue_map(set);
1424 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1425 blk_queue_rq_timeout(q, 30000);
1427 q->nr_queues = nr_cpu_ids;
1428 q->nr_hw_queues = set->nr_hw_queues;
1431 q->queue_hw_ctx = hctxs;
1433 q->mq_ops = set->ops;
1434 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1436 q->sg_reserved_size = INT_MAX;
1438 blk_queue_make_request(q, blk_mq_make_request);
1439 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1441 blk_queue_rq_timeout(q, set->timeout);
1443 if (set->ops->complete)
1444 blk_queue_softirq_done(q, set->ops->complete);
1446 blk_mq_init_flush(q);
1447 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1449 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1450 set->cmd_size, cache_line_size()),
1455 if (blk_mq_init_hw_queues(q, set))
1458 blk_mq_map_swqueue(q);
1460 mutex_lock(&all_q_mutex);
1461 list_add_tail(&q->all_q_node, &all_q_list);
1462 mutex_unlock(&all_q_mutex);
1471 blk_cleanup_queue(q);
1473 for (i = 0; i < set->nr_hw_queues; i++) {
1476 free_cpumask_var(hctxs[i]->cpumask);
1477 set->ops->free_hctx(hctxs[i], i);
1482 return ERR_PTR(-ENOMEM);
1484 EXPORT_SYMBOL(blk_mq_init_queue);
1486 void blk_mq_free_queue(struct request_queue *q)
1488 struct blk_mq_hw_ctx *hctx;
1491 queue_for_each_hw_ctx(q, hctx, i) {
1492 kfree(hctx->ctx_map);
1494 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1495 if (q->mq_ops->exit_hctx)
1496 q->mq_ops->exit_hctx(hctx, i);
1497 free_cpumask_var(hctx->cpumask);
1498 q->mq_ops->free_hctx(hctx, i);
1501 free_percpu(q->queue_ctx);
1502 kfree(q->queue_hw_ctx);
1505 q->queue_ctx = NULL;
1506 q->queue_hw_ctx = NULL;
1509 mutex_lock(&all_q_mutex);
1510 list_del_init(&q->all_q_node);
1511 mutex_unlock(&all_q_mutex);
1514 /* Basically redo blk_mq_init_queue with queue frozen */
1515 static void blk_mq_queue_reinit(struct request_queue *q)
1517 blk_mq_freeze_queue(q);
1519 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1522 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1523 * we should change hctx numa_node according to new topology (this
1524 * involves free and re-allocate memory, worthy doing?)
1527 blk_mq_map_swqueue(q);
1529 blk_mq_unfreeze_queue(q);
1532 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1533 unsigned long action, void *hcpu)
1535 struct request_queue *q;
1538 * Before new mapping is established, hotadded cpu might already start
1539 * handling requests. This doesn't break anything as we map offline
1540 * CPUs to first hardware queue. We will re-init queue below to get
1543 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1544 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1547 mutex_lock(&all_q_mutex);
1548 list_for_each_entry(q, &all_q_list, all_q_node)
1549 blk_mq_queue_reinit(q);
1550 mutex_unlock(&all_q_mutex);
1554 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1558 if (!set->nr_hw_queues)
1560 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1562 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1565 if (!set->nr_hw_queues ||
1566 !set->ops->queue_rq || !set->ops->map_queue ||
1567 !set->ops->alloc_hctx || !set->ops->free_hctx)
1571 set->tags = kmalloc_node(set->nr_hw_queues *
1572 sizeof(struct blk_mq_tags *),
1573 GFP_KERNEL, set->numa_node);
1577 for (i = 0; i < set->nr_hw_queues; i++) {
1578 set->tags[i] = blk_mq_init_rq_map(set, i);
1587 blk_mq_free_rq_map(set, set->tags[i], i);
1591 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
1593 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
1597 for (i = 0; i < set->nr_hw_queues; i++)
1598 blk_mq_free_rq_map(set, set->tags[i], i);
1601 EXPORT_SYMBOL(blk_mq_free_tag_set);
1603 void blk_mq_disable_hotplug(void)
1605 mutex_lock(&all_q_mutex);
1608 void blk_mq_enable_hotplug(void)
1610 mutex_unlock(&all_q_mutex);
1613 static int __init blk_mq_init(void)
1617 /* Must be called after percpu_counter_hotcpu_callback() */
1618 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1622 subsys_initcall(blk_mq_init);