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 struct blk_mq_ctx *ctx,
78 gfp_t gfp, bool reserved)
83 tag = blk_mq_get_tag(hctx, &ctx->last_tag, gfp, reserved);
84 if (tag != BLK_MQ_TAG_FAIL) {
85 rq = hctx->tags->rqs[tag];
88 if (blk_mq_tag_busy(hctx)) {
89 rq->cmd_flags = REQ_MQ_INFLIGHT;
90 atomic_inc(&hctx->nr_active);
100 static int blk_mq_queue_enter(struct request_queue *q)
104 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
106 /* we have problems to freeze the queue if it's initializing */
107 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
110 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
112 spin_lock_irq(q->queue_lock);
113 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
114 !blk_queue_bypass(q) || blk_queue_dying(q),
116 /* inc usage with lock hold to avoid freeze_queue runs here */
117 if (!ret && !blk_queue_dying(q))
118 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
119 else if (blk_queue_dying(q))
121 spin_unlock_irq(q->queue_lock);
126 static void blk_mq_queue_exit(struct request_queue *q)
128 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
131 static void __blk_mq_drain_queue(struct request_queue *q)
136 spin_lock_irq(q->queue_lock);
137 count = percpu_counter_sum(&q->mq_usage_counter);
138 spin_unlock_irq(q->queue_lock);
142 blk_mq_run_queues(q, false);
148 * Guarantee no request is in use, so we can change any data structure of
149 * the queue afterward.
151 static void blk_mq_freeze_queue(struct request_queue *q)
155 spin_lock_irq(q->queue_lock);
156 drain = !q->bypass_depth++;
157 queue_flag_set(QUEUE_FLAG_BYPASS, q);
158 spin_unlock_irq(q->queue_lock);
161 __blk_mq_drain_queue(q);
164 void blk_mq_drain_queue(struct request_queue *q)
166 __blk_mq_drain_queue(q);
169 static void blk_mq_unfreeze_queue(struct request_queue *q)
173 spin_lock_irq(q->queue_lock);
174 if (!--q->bypass_depth) {
175 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
178 WARN_ON_ONCE(q->bypass_depth < 0);
179 spin_unlock_irq(q->queue_lock);
181 wake_up_all(&q->mq_freeze_wq);
184 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
186 return blk_mq_has_free_tags(hctx->tags);
188 EXPORT_SYMBOL(blk_mq_can_queue);
190 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
191 struct request *rq, unsigned int rw_flags)
193 if (blk_queue_io_stat(q))
194 rw_flags |= REQ_IO_STAT;
196 INIT_LIST_HEAD(&rq->queuelist);
197 /* csd/requeue_work/fifo_time is initialized before use */
200 rq->cmd_flags |= rw_flags;
202 /* do not touch atomic flags, it needs atomic ops against the timer */
205 rq->__sector = (sector_t) -1;
208 INIT_HLIST_NODE(&rq->hash);
209 RB_CLEAR_NODE(&rq->rb_node);
210 memset(&rq->flush, 0, max(sizeof(rq->flush), sizeof(rq->elv)));
213 rq->start_time = jiffies;
214 #ifdef CONFIG_BLK_CGROUP
216 set_start_time_ns(rq);
217 rq->io_start_time_ns = 0;
219 rq->nr_phys_segments = 0;
220 #if defined(CONFIG_BLK_DEV_INTEGRITY)
221 rq->nr_integrity_segments = 0;
225 /* tag was already set */
227 memset(rq->__cmd, 0, sizeof(rq->__cmd));
229 rq->cmd_len = BLK_MAX_CDB;
237 INIT_LIST_HEAD(&rq->timeout_list);
241 rq->end_io_data = NULL;
244 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
247 static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
254 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
255 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
257 rq = __blk_mq_alloc_request(hctx, ctx, gfp & ~__GFP_WAIT,
260 blk_mq_rq_ctx_init(q, ctx, rq, rw);
264 if (gfp & __GFP_WAIT) {
265 __blk_mq_run_hw_queue(hctx);
272 blk_mq_wait_for_tags(hctx, reserved);
278 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp)
282 if (blk_mq_queue_enter(q))
285 rq = blk_mq_alloc_request_pinned(q, rw, gfp, false);
287 blk_mq_put_ctx(rq->mq_ctx);
290 EXPORT_SYMBOL(blk_mq_alloc_request);
292 struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
297 if (blk_mq_queue_enter(q))
300 rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
302 blk_mq_put_ctx(rq->mq_ctx);
305 EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
307 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
308 struct blk_mq_ctx *ctx, struct request *rq)
310 const int tag = rq->tag;
311 struct request_queue *q = rq->q;
313 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
314 atomic_dec(&hctx->nr_active);
316 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
317 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
318 blk_mq_queue_exit(q);
321 void blk_mq_free_request(struct request *rq)
323 struct blk_mq_ctx *ctx = rq->mq_ctx;
324 struct blk_mq_hw_ctx *hctx;
325 struct request_queue *q = rq->q;
327 ctx->rq_completed[rq_is_sync(rq)]++;
329 hctx = q->mq_ops->map_queue(q, ctx->cpu);
330 __blk_mq_free_request(hctx, ctx, rq);
334 * Clone all relevant state from a request that has been put on hold in
335 * the flush state machine into the preallocated flush request that hangs
336 * off the request queue.
338 * For a driver the flush request should be invisible, that's why we are
339 * impersonating the original request here.
341 void blk_mq_clone_flush_request(struct request *flush_rq,
342 struct request *orig_rq)
344 struct blk_mq_hw_ctx *hctx =
345 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
347 flush_rq->mq_ctx = orig_rq->mq_ctx;
348 flush_rq->tag = orig_rq->tag;
349 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
353 inline void __blk_mq_end_io(struct request *rq, int error)
355 blk_account_io_done(rq);
358 rq->end_io(rq, error);
360 if (unlikely(blk_bidi_rq(rq)))
361 blk_mq_free_request(rq->next_rq);
362 blk_mq_free_request(rq);
365 EXPORT_SYMBOL(__blk_mq_end_io);
367 void blk_mq_end_io(struct request *rq, int error)
369 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
371 __blk_mq_end_io(rq, error);
373 EXPORT_SYMBOL(blk_mq_end_io);
375 static void __blk_mq_complete_request_remote(void *data)
377 struct request *rq = data;
379 rq->q->softirq_done_fn(rq);
382 void __blk_mq_complete_request(struct request *rq)
384 struct blk_mq_ctx *ctx = rq->mq_ctx;
388 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
389 rq->q->softirq_done_fn(rq);
394 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
395 shared = cpus_share_cache(cpu, ctx->cpu);
397 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
398 rq->csd.func = __blk_mq_complete_request_remote;
401 smp_call_function_single_async(ctx->cpu, &rq->csd);
403 rq->q->softirq_done_fn(rq);
409 * blk_mq_complete_request - end I/O on a request
410 * @rq: the request being processed
413 * Ends all I/O on a request. It does not handle partial completions.
414 * The actual completion happens out-of-order, through a IPI handler.
416 void blk_mq_complete_request(struct request *rq)
418 if (unlikely(blk_should_fake_timeout(rq->q)))
420 if (!blk_mark_rq_complete(rq))
421 __blk_mq_complete_request(rq);
423 EXPORT_SYMBOL(blk_mq_complete_request);
425 static void blk_mq_start_request(struct request *rq, bool last)
427 struct request_queue *q = rq->q;
429 trace_block_rq_issue(q, rq);
431 rq->resid_len = blk_rq_bytes(rq);
432 if (unlikely(blk_bidi_rq(rq)))
433 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
436 * Just mark start time and set the started bit. Due to memory
437 * ordering, we know we'll see the correct deadline as long as
438 * REQ_ATOMIC_STARTED is seen.
440 rq->deadline = jiffies + q->rq_timeout;
443 * Mark us as started and clear complete. Complete might have been
444 * set if requeue raced with timeout, which then marked it as
445 * complete. So be sure to clear complete again when we start
446 * the request, otherwise we'll ignore the completion event.
448 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
449 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
451 if (q->dma_drain_size && blk_rq_bytes(rq)) {
453 * Make sure space for the drain appears. We know we can do
454 * this because max_hw_segments has been adjusted to be one
455 * fewer than the device can handle.
457 rq->nr_phys_segments++;
461 * Flag the last request in the series so that drivers know when IO
462 * should be kicked off, if they don't do it on a per-request basis.
464 * Note: the flag isn't the only condition drivers should do kick off.
465 * If drive is busy, the last request might not have the bit set.
468 rq->cmd_flags |= REQ_END;
471 static void __blk_mq_requeue_request(struct request *rq)
473 struct request_queue *q = rq->q;
475 trace_block_rq_requeue(q, rq);
476 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
478 rq->cmd_flags &= ~REQ_END;
480 if (q->dma_drain_size && blk_rq_bytes(rq))
481 rq->nr_phys_segments--;
484 void blk_mq_requeue_request(struct request *rq)
486 __blk_mq_requeue_request(rq);
487 blk_clear_rq_complete(rq);
489 BUG_ON(blk_queued_rq(rq));
490 blk_mq_insert_request(rq, true, true, false);
492 EXPORT_SYMBOL(blk_mq_requeue_request);
494 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
496 return tags->rqs[tag];
498 EXPORT_SYMBOL(blk_mq_tag_to_rq);
500 struct blk_mq_timeout_data {
501 struct blk_mq_hw_ctx *hctx;
503 unsigned int *next_set;
506 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
508 struct blk_mq_timeout_data *data = __data;
509 struct blk_mq_hw_ctx *hctx = data->hctx;
512 /* It may not be in flight yet (this is where
513 * the REQ_ATOMIC_STARTED flag comes in). The requests are
514 * statically allocated, so we know it's always safe to access the
515 * memory associated with a bit offset into ->rqs[].
521 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
522 if (tag >= hctx->tags->nr_tags)
525 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
526 if (rq->q != hctx->queue)
528 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
531 blk_rq_check_expired(rq, data->next, data->next_set);
535 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
537 unsigned int *next_set)
539 struct blk_mq_timeout_data data = {
542 .next_set = next_set,
546 * Ask the tagging code to iterate busy requests, so we can
547 * check them for timeout.
549 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
552 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
554 struct request_queue *q = rq->q;
557 * We know that complete is set at this point. If STARTED isn't set
558 * anymore, then the request isn't active and the "timeout" should
559 * just be ignored. This can happen due to the bitflag ordering.
560 * Timeout first checks if STARTED is set, and if it is, assumes
561 * the request is active. But if we race with completion, then
562 * we both flags will get cleared. So check here again, and ignore
563 * a timeout event with a request that isn't active.
565 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
566 return BLK_EH_NOT_HANDLED;
568 if (!q->mq_ops->timeout)
569 return BLK_EH_RESET_TIMER;
571 return q->mq_ops->timeout(rq);
574 static void blk_mq_rq_timer(unsigned long data)
576 struct request_queue *q = (struct request_queue *) data;
577 struct blk_mq_hw_ctx *hctx;
578 unsigned long next = 0;
581 queue_for_each_hw_ctx(q, hctx, i)
582 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
585 next = blk_rq_timeout(round_jiffies_up(next));
586 mod_timer(&q->timeout, next);
588 queue_for_each_hw_ctx(q, hctx, i)
589 blk_mq_tag_idle(hctx);
594 * Reverse check our software queue for entries that we could potentially
595 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
596 * too much time checking for merges.
598 static bool blk_mq_attempt_merge(struct request_queue *q,
599 struct blk_mq_ctx *ctx, struct bio *bio)
604 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
610 if (!blk_rq_merge_ok(rq, bio))
613 el_ret = blk_try_merge(rq, bio);
614 if (el_ret == ELEVATOR_BACK_MERGE) {
615 if (bio_attempt_back_merge(q, rq, bio)) {
620 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
621 if (bio_attempt_front_merge(q, rq, bio)) {
633 * Run this hardware queue, pulling any software queues mapped to it in.
634 * Note that this function currently has various problems around ordering
635 * of IO. In particular, we'd like FIFO behaviour on handling existing
636 * items on the hctx->dispatch list. Ignore that for now.
638 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
640 struct request_queue *q = hctx->queue;
641 struct blk_mq_ctx *ctx;
646 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
648 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
654 * Touch any software queue that has pending entries.
656 for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
657 clear_bit(bit, hctx->ctx_map);
658 ctx = hctx->ctxs[bit];
660 spin_lock(&ctx->lock);
661 list_splice_tail_init(&ctx->rq_list, &rq_list);
662 spin_unlock(&ctx->lock);
666 * If we have previous entries on our dispatch list, grab them
667 * and stuff them at the front for more fair dispatch.
669 if (!list_empty_careful(&hctx->dispatch)) {
670 spin_lock(&hctx->lock);
671 if (!list_empty(&hctx->dispatch))
672 list_splice_init(&hctx->dispatch, &rq_list);
673 spin_unlock(&hctx->lock);
677 * Delete and return all entries from our dispatch list
682 * Now process all the entries, sending them to the driver.
684 while (!list_empty(&rq_list)) {
687 rq = list_first_entry(&rq_list, struct request, queuelist);
688 list_del_init(&rq->queuelist);
690 blk_mq_start_request(rq, list_empty(&rq_list));
692 ret = q->mq_ops->queue_rq(hctx, rq);
694 case BLK_MQ_RQ_QUEUE_OK:
697 case BLK_MQ_RQ_QUEUE_BUSY:
698 list_add(&rq->queuelist, &rq_list);
699 __blk_mq_requeue_request(rq);
702 pr_err("blk-mq: bad return on queue: %d\n", ret);
703 case BLK_MQ_RQ_QUEUE_ERROR:
705 blk_mq_end_io(rq, rq->errors);
709 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
714 hctx->dispatched[0]++;
715 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
716 hctx->dispatched[ilog2(queued) + 1]++;
719 * Any items that need requeuing? Stuff them into hctx->dispatch,
720 * that is where we will continue on next queue run.
722 if (!list_empty(&rq_list)) {
723 spin_lock(&hctx->lock);
724 list_splice(&rq_list, &hctx->dispatch);
725 spin_unlock(&hctx->lock);
730 * It'd be great if the workqueue API had a way to pass
731 * in a mask and had some smarts for more clever placement.
732 * For now we just round-robin here, switching for every
733 * BLK_MQ_CPU_WORK_BATCH queued items.
735 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
737 int cpu = hctx->next_cpu;
739 if (--hctx->next_cpu_batch <= 0) {
742 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
743 if (next_cpu >= nr_cpu_ids)
744 next_cpu = cpumask_first(hctx->cpumask);
746 hctx->next_cpu = next_cpu;
747 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
753 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
755 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
758 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
759 __blk_mq_run_hw_queue(hctx);
760 else if (hctx->queue->nr_hw_queues == 1)
761 kblockd_schedule_delayed_work(&hctx->run_work, 0);
765 cpu = blk_mq_hctx_next_cpu(hctx);
766 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
770 void blk_mq_run_queues(struct request_queue *q, bool async)
772 struct blk_mq_hw_ctx *hctx;
775 queue_for_each_hw_ctx(q, hctx, i) {
776 if ((!blk_mq_hctx_has_pending(hctx) &&
777 list_empty_careful(&hctx->dispatch)) ||
778 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
782 blk_mq_run_hw_queue(hctx, async);
786 EXPORT_SYMBOL(blk_mq_run_queues);
788 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
790 cancel_delayed_work(&hctx->run_work);
791 cancel_delayed_work(&hctx->delay_work);
792 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
794 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
796 void blk_mq_stop_hw_queues(struct request_queue *q)
798 struct blk_mq_hw_ctx *hctx;
801 queue_for_each_hw_ctx(q, hctx, i)
802 blk_mq_stop_hw_queue(hctx);
804 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
806 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
808 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
811 __blk_mq_run_hw_queue(hctx);
814 EXPORT_SYMBOL(blk_mq_start_hw_queue);
816 void blk_mq_start_hw_queues(struct request_queue *q)
818 struct blk_mq_hw_ctx *hctx;
821 queue_for_each_hw_ctx(q, hctx, i)
822 blk_mq_start_hw_queue(hctx);
824 EXPORT_SYMBOL(blk_mq_start_hw_queues);
827 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
829 struct blk_mq_hw_ctx *hctx;
832 queue_for_each_hw_ctx(q, hctx, i) {
833 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
836 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
838 blk_mq_run_hw_queue(hctx, async);
842 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
844 static void blk_mq_run_work_fn(struct work_struct *work)
846 struct blk_mq_hw_ctx *hctx;
848 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
850 __blk_mq_run_hw_queue(hctx);
853 static void blk_mq_delay_work_fn(struct work_struct *work)
855 struct blk_mq_hw_ctx *hctx;
857 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
859 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
860 __blk_mq_run_hw_queue(hctx);
863 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
865 unsigned long tmo = msecs_to_jiffies(msecs);
867 if (hctx->queue->nr_hw_queues == 1)
868 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
872 cpu = blk_mq_hctx_next_cpu(hctx);
873 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
876 EXPORT_SYMBOL(blk_mq_delay_queue);
878 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
879 struct request *rq, bool at_head)
881 struct blk_mq_ctx *ctx = rq->mq_ctx;
883 trace_block_rq_insert(hctx->queue, rq);
886 list_add(&rq->queuelist, &ctx->rq_list);
888 list_add_tail(&rq->queuelist, &ctx->rq_list);
890 blk_mq_hctx_mark_pending(hctx, ctx);
893 * We do this early, to ensure we are on the right CPU.
898 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
901 struct request_queue *q = rq->q;
902 struct blk_mq_hw_ctx *hctx;
903 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
905 current_ctx = blk_mq_get_ctx(q);
906 if (!cpu_online(ctx->cpu))
907 rq->mq_ctx = ctx = current_ctx;
909 hctx = q->mq_ops->map_queue(q, ctx->cpu);
911 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
912 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
913 blk_insert_flush(rq);
915 spin_lock(&ctx->lock);
916 __blk_mq_insert_request(hctx, rq, at_head);
917 spin_unlock(&ctx->lock);
921 blk_mq_run_hw_queue(hctx, async);
923 blk_mq_put_ctx(current_ctx);
926 static void blk_mq_insert_requests(struct request_queue *q,
927 struct blk_mq_ctx *ctx,
928 struct list_head *list,
933 struct blk_mq_hw_ctx *hctx;
934 struct blk_mq_ctx *current_ctx;
936 trace_block_unplug(q, depth, !from_schedule);
938 current_ctx = blk_mq_get_ctx(q);
940 if (!cpu_online(ctx->cpu))
942 hctx = q->mq_ops->map_queue(q, ctx->cpu);
945 * preemption doesn't flush plug list, so it's possible ctx->cpu is
948 spin_lock(&ctx->lock);
949 while (!list_empty(list)) {
952 rq = list_first_entry(list, struct request, queuelist);
953 list_del_init(&rq->queuelist);
955 __blk_mq_insert_request(hctx, rq, false);
957 spin_unlock(&ctx->lock);
959 blk_mq_run_hw_queue(hctx, from_schedule);
960 blk_mq_put_ctx(current_ctx);
963 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
965 struct request *rqa = container_of(a, struct request, queuelist);
966 struct request *rqb = container_of(b, struct request, queuelist);
968 return !(rqa->mq_ctx < rqb->mq_ctx ||
969 (rqa->mq_ctx == rqb->mq_ctx &&
970 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
973 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
975 struct blk_mq_ctx *this_ctx;
976 struct request_queue *this_q;
982 list_splice_init(&plug->mq_list, &list);
984 list_sort(NULL, &list, plug_ctx_cmp);
990 while (!list_empty(&list)) {
991 rq = list_entry_rq(list.next);
992 list_del_init(&rq->queuelist);
994 if (rq->mq_ctx != this_ctx) {
996 blk_mq_insert_requests(this_q, this_ctx,
1001 this_ctx = rq->mq_ctx;
1007 list_add_tail(&rq->queuelist, &ctx_list);
1011 * If 'this_ctx' is set, we know we have entries to complete
1012 * on 'ctx_list'. Do those.
1015 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1020 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1022 init_request_from_bio(rq, bio);
1023 blk_account_io_start(rq, 1);
1026 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1028 struct blk_mq_hw_ctx *hctx;
1029 struct blk_mq_ctx *ctx;
1030 const int is_sync = rw_is_sync(bio->bi_rw);
1031 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1032 int rw = bio_data_dir(bio);
1034 unsigned int use_plug, request_count = 0;
1037 * If we have multiple hardware queues, just go directly to
1038 * one of those for sync IO.
1040 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
1042 blk_queue_bounce(q, &bio);
1044 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1045 bio_endio(bio, -EIO);
1049 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
1052 if (blk_mq_queue_enter(q)) {
1053 bio_endio(bio, -EIO);
1057 ctx = blk_mq_get_ctx(q);
1058 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1062 trace_block_getrq(q, bio, rw);
1063 rq = __blk_mq_alloc_request(hctx, ctx, GFP_ATOMIC, false);
1065 blk_mq_rq_ctx_init(q, ctx, rq, rw);
1067 blk_mq_put_ctx(ctx);
1068 trace_block_sleeprq(q, bio, rw);
1069 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
1072 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1077 if (unlikely(is_flush_fua)) {
1078 blk_mq_bio_to_request(rq, bio);
1079 blk_insert_flush(rq);
1084 * A task plug currently exists. Since this is completely lockless,
1085 * utilize that to temporarily store requests until the task is
1086 * either done or scheduled away.
1089 struct blk_plug *plug = current->plug;
1092 blk_mq_bio_to_request(rq, bio);
1093 if (list_empty(&plug->mq_list))
1094 trace_block_plug(q);
1095 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1096 blk_flush_plug_list(plug, false);
1097 trace_block_plug(q);
1099 list_add_tail(&rq->queuelist, &plug->mq_list);
1100 blk_mq_put_ctx(ctx);
1105 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1106 init_request_from_bio(rq, bio);
1108 spin_lock(&ctx->lock);
1110 __blk_mq_insert_request(hctx, rq, false);
1111 spin_unlock(&ctx->lock);
1112 blk_account_io_start(rq, 1);
1114 spin_lock(&ctx->lock);
1115 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1116 init_request_from_bio(rq, bio);
1120 spin_unlock(&ctx->lock);
1121 __blk_mq_free_request(hctx, ctx, rq);
1126 * For a SYNC request, send it to the hardware immediately. For an
1127 * ASYNC request, just ensure that we run it later on. The latter
1128 * allows for merging opportunities and more efficient dispatching.
1131 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
1132 blk_mq_put_ctx(ctx);
1136 * Default mapping to a software queue, since we use one per CPU.
1138 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1140 return q->queue_hw_ctx[q->mq_map[cpu]];
1142 EXPORT_SYMBOL(blk_mq_map_queue);
1144 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *set,
1145 unsigned int hctx_index)
1147 return kzalloc_node(sizeof(struct blk_mq_hw_ctx), GFP_KERNEL,
1150 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1152 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1153 unsigned int hctx_index)
1157 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1159 static void blk_mq_hctx_notify(void *data, unsigned long action,
1162 struct blk_mq_hw_ctx *hctx = data;
1163 struct request_queue *q = hctx->queue;
1164 struct blk_mq_ctx *ctx;
1167 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1171 * Move ctx entries to new CPU, if this one is going away.
1173 ctx = __blk_mq_get_ctx(q, cpu);
1175 spin_lock(&ctx->lock);
1176 if (!list_empty(&ctx->rq_list)) {
1177 list_splice_init(&ctx->rq_list, &tmp);
1178 clear_bit(ctx->index_hw, hctx->ctx_map);
1180 spin_unlock(&ctx->lock);
1182 if (list_empty(&tmp))
1185 ctx = blk_mq_get_ctx(q);
1186 spin_lock(&ctx->lock);
1188 while (!list_empty(&tmp)) {
1191 rq = list_first_entry(&tmp, struct request, queuelist);
1193 list_move_tail(&rq->queuelist, &ctx->rq_list);
1196 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1197 blk_mq_hctx_mark_pending(hctx, ctx);
1199 spin_unlock(&ctx->lock);
1201 blk_mq_run_hw_queue(hctx, true);
1202 blk_mq_put_ctx(ctx);
1205 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1206 struct blk_mq_tags *tags, unsigned int hctx_idx)
1210 if (tags->rqs && set->ops->exit_request) {
1213 for (i = 0; i < tags->nr_tags; i++) {
1216 set->ops->exit_request(set->driver_data, tags->rqs[i],
1221 while (!list_empty(&tags->page_list)) {
1222 page = list_first_entry(&tags->page_list, struct page, lru);
1223 list_del_init(&page->lru);
1224 __free_pages(page, page->private);
1229 blk_mq_free_tags(tags);
1232 static size_t order_to_size(unsigned int order)
1234 return (size_t)PAGE_SIZE << order;
1237 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1238 unsigned int hctx_idx)
1240 struct blk_mq_tags *tags;
1241 unsigned int i, j, entries_per_page, max_order = 4;
1242 size_t rq_size, left;
1244 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1249 INIT_LIST_HEAD(&tags->page_list);
1251 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1252 GFP_KERNEL, set->numa_node);
1254 blk_mq_free_tags(tags);
1259 * rq_size is the size of the request plus driver payload, rounded
1260 * to the cacheline size
1262 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1264 left = rq_size * set->queue_depth;
1266 for (i = 0; i < set->queue_depth; ) {
1267 int this_order = max_order;
1272 while (left < order_to_size(this_order - 1) && this_order)
1276 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1282 if (order_to_size(this_order) < rq_size)
1289 page->private = this_order;
1290 list_add_tail(&page->lru, &tags->page_list);
1292 p = page_address(page);
1293 entries_per_page = order_to_size(this_order) / rq_size;
1294 to_do = min(entries_per_page, set->queue_depth - i);
1295 left -= to_do * rq_size;
1296 for (j = 0; j < to_do; j++) {
1298 if (set->ops->init_request) {
1299 if (set->ops->init_request(set->driver_data,
1300 tags->rqs[i], hctx_idx, i,
1313 pr_warn("%s: failed to allocate requests\n", __func__);
1314 blk_mq_free_rq_map(set, tags, hctx_idx);
1318 static int blk_mq_init_hw_queues(struct request_queue *q,
1319 struct blk_mq_tag_set *set)
1321 struct blk_mq_hw_ctx *hctx;
1325 * Initialize hardware queues
1327 queue_for_each_hw_ctx(q, hctx, i) {
1328 unsigned int num_maps;
1331 node = hctx->numa_node;
1332 if (node == NUMA_NO_NODE)
1333 node = hctx->numa_node = set->numa_node;
1335 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1336 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1337 spin_lock_init(&hctx->lock);
1338 INIT_LIST_HEAD(&hctx->dispatch);
1340 hctx->queue_num = i;
1341 hctx->flags = set->flags;
1342 hctx->cmd_size = set->cmd_size;
1344 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1345 blk_mq_hctx_notify, hctx);
1346 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1348 hctx->tags = set->tags[i];
1351 * Allocate space for all possible cpus to avoid allocation in
1354 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1359 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1360 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1365 hctx->nr_ctx_map = num_maps;
1368 if (set->ops->init_hctx &&
1369 set->ops->init_hctx(hctx, set->driver_data, i))
1373 if (i == q->nr_hw_queues)
1379 queue_for_each_hw_ctx(q, hctx, j) {
1383 if (set->ops->exit_hctx)
1384 set->ops->exit_hctx(hctx, j);
1386 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1388 kfree(hctx->ctx_map);
1394 static void blk_mq_init_cpu_queues(struct request_queue *q,
1395 unsigned int nr_hw_queues)
1399 for_each_possible_cpu(i) {
1400 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1401 struct blk_mq_hw_ctx *hctx;
1403 memset(__ctx, 0, sizeof(*__ctx));
1405 spin_lock_init(&__ctx->lock);
1406 INIT_LIST_HEAD(&__ctx->rq_list);
1409 /* If the cpu isn't online, the cpu is mapped to first hctx */
1413 hctx = q->mq_ops->map_queue(q, i);
1414 cpumask_set_cpu(i, hctx->cpumask);
1418 * Set local node, IFF we have more than one hw queue. If
1419 * not, we remain on the home node of the device
1421 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1422 hctx->numa_node = cpu_to_node(i);
1426 static void blk_mq_map_swqueue(struct request_queue *q)
1429 struct blk_mq_hw_ctx *hctx;
1430 struct blk_mq_ctx *ctx;
1432 queue_for_each_hw_ctx(q, hctx, i) {
1433 cpumask_clear(hctx->cpumask);
1438 * Map software to hardware queues
1440 queue_for_each_ctx(q, ctx, i) {
1441 /* If the cpu isn't online, the cpu is mapped to first hctx */
1445 hctx = q->mq_ops->map_queue(q, i);
1446 cpumask_set_cpu(i, hctx->cpumask);
1447 ctx->index_hw = hctx->nr_ctx;
1448 hctx->ctxs[hctx->nr_ctx++] = ctx;
1451 queue_for_each_hw_ctx(q, hctx, i) {
1452 hctx->next_cpu = cpumask_first(hctx->cpumask);
1453 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1457 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1459 struct blk_mq_hw_ctx *hctx;
1460 struct request_queue *q;
1464 if (set->tag_list.next == set->tag_list.prev)
1469 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1470 blk_mq_freeze_queue(q);
1472 queue_for_each_hw_ctx(q, hctx, i) {
1474 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1476 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1478 blk_mq_unfreeze_queue(q);
1482 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1484 struct blk_mq_tag_set *set = q->tag_set;
1486 blk_mq_freeze_queue(q);
1488 mutex_lock(&set->tag_list_lock);
1489 list_del_init(&q->tag_set_list);
1490 blk_mq_update_tag_set_depth(set);
1491 mutex_unlock(&set->tag_list_lock);
1493 blk_mq_unfreeze_queue(q);
1496 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1497 struct request_queue *q)
1501 mutex_lock(&set->tag_list_lock);
1502 list_add_tail(&q->tag_set_list, &set->tag_list);
1503 blk_mq_update_tag_set_depth(set);
1504 mutex_unlock(&set->tag_list_lock);
1507 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1509 struct blk_mq_hw_ctx **hctxs;
1510 struct blk_mq_ctx *ctx;
1511 struct request_queue *q;
1514 ctx = alloc_percpu(struct blk_mq_ctx);
1516 return ERR_PTR(-ENOMEM);
1518 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1524 for (i = 0; i < set->nr_hw_queues; i++) {
1525 hctxs[i] = set->ops->alloc_hctx(set, i);
1529 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1532 atomic_set(&hctxs[i]->nr_active, 0);
1533 hctxs[i]->numa_node = NUMA_NO_NODE;
1534 hctxs[i]->queue_num = i;
1537 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1541 q->mq_map = blk_mq_make_queue_map(set);
1545 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1546 blk_queue_rq_timeout(q, 30000);
1548 q->nr_queues = nr_cpu_ids;
1549 q->nr_hw_queues = set->nr_hw_queues;
1552 q->queue_hw_ctx = hctxs;
1554 q->mq_ops = set->ops;
1555 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1557 q->sg_reserved_size = INT_MAX;
1559 blk_queue_make_request(q, blk_mq_make_request);
1560 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1562 blk_queue_rq_timeout(q, set->timeout);
1564 if (set->ops->complete)
1565 blk_queue_softirq_done(q, set->ops->complete);
1567 blk_mq_init_flush(q);
1568 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1570 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1571 set->cmd_size, cache_line_size()),
1576 if (blk_mq_init_hw_queues(q, set))
1579 blk_mq_map_swqueue(q);
1581 mutex_lock(&all_q_mutex);
1582 list_add_tail(&q->all_q_node, &all_q_list);
1583 mutex_unlock(&all_q_mutex);
1585 blk_mq_add_queue_tag_set(set, q);
1594 blk_cleanup_queue(q);
1596 for (i = 0; i < set->nr_hw_queues; i++) {
1599 free_cpumask_var(hctxs[i]->cpumask);
1600 set->ops->free_hctx(hctxs[i], i);
1605 return ERR_PTR(-ENOMEM);
1607 EXPORT_SYMBOL(blk_mq_init_queue);
1609 void blk_mq_free_queue(struct request_queue *q)
1611 struct blk_mq_hw_ctx *hctx;
1614 blk_mq_del_queue_tag_set(q);
1616 queue_for_each_hw_ctx(q, hctx, i) {
1617 kfree(hctx->ctx_map);
1619 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1620 if (q->mq_ops->exit_hctx)
1621 q->mq_ops->exit_hctx(hctx, i);
1622 free_cpumask_var(hctx->cpumask);
1623 q->mq_ops->free_hctx(hctx, i);
1626 free_percpu(q->queue_ctx);
1627 kfree(q->queue_hw_ctx);
1630 q->queue_ctx = NULL;
1631 q->queue_hw_ctx = NULL;
1634 mutex_lock(&all_q_mutex);
1635 list_del_init(&q->all_q_node);
1636 mutex_unlock(&all_q_mutex);
1639 /* Basically redo blk_mq_init_queue with queue frozen */
1640 static void blk_mq_queue_reinit(struct request_queue *q)
1642 blk_mq_freeze_queue(q);
1644 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1647 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1648 * we should change hctx numa_node according to new topology (this
1649 * involves free and re-allocate memory, worthy doing?)
1652 blk_mq_map_swqueue(q);
1654 blk_mq_unfreeze_queue(q);
1657 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1658 unsigned long action, void *hcpu)
1660 struct request_queue *q;
1663 * Before new mappings are established, hotadded cpu might already
1664 * start handling requests. This doesn't break anything as we map
1665 * offline CPUs to first hardware queue. We will re-init the queue
1666 * below to get optimal settings.
1668 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1669 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1672 mutex_lock(&all_q_mutex);
1673 list_for_each_entry(q, &all_q_list, all_q_node)
1674 blk_mq_queue_reinit(q);
1675 mutex_unlock(&all_q_mutex);
1679 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1683 if (!set->nr_hw_queues)
1685 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1687 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1690 if (!set->nr_hw_queues ||
1691 !set->ops->queue_rq || !set->ops->map_queue ||
1692 !set->ops->alloc_hctx || !set->ops->free_hctx)
1696 set->tags = kmalloc_node(set->nr_hw_queues *
1697 sizeof(struct blk_mq_tags *),
1698 GFP_KERNEL, set->numa_node);
1702 for (i = 0; i < set->nr_hw_queues; i++) {
1703 set->tags[i] = blk_mq_init_rq_map(set, i);
1708 mutex_init(&set->tag_list_lock);
1709 INIT_LIST_HEAD(&set->tag_list);
1715 blk_mq_free_rq_map(set, set->tags[i], i);
1719 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
1721 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
1725 for (i = 0; i < set->nr_hw_queues; i++)
1726 blk_mq_free_rq_map(set, set->tags[i], i);
1729 EXPORT_SYMBOL(blk_mq_free_tag_set);
1731 void blk_mq_disable_hotplug(void)
1733 mutex_lock(&all_q_mutex);
1736 void blk_mq_enable_hotplug(void)
1738 mutex_unlock(&all_q_mutex);
1741 static int __init blk_mq_init(void)
1745 /* Must be called after percpu_counter_hotcpu_callback() */
1746 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1750 subsys_initcall(blk_mq_init);