2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
13 #include <linux/init.h>
14 #include <linux/slab.h>
15 #include <linux/workqueue.h>
16 #include <linux/smp.h>
17 #include <linux/llist.h>
18 #include <linux/list_sort.h>
19 #include <linux/cpu.h>
20 #include <linux/cache.h>
21 #include <linux/sched/sysctl.h>
22 #include <linux/delay.h>
24 #include <trace/events/block.h>
26 #include <linux/blk-mq.h>
29 #include "blk-mq-tag.h"
31 static DEFINE_MUTEX(all_q_mutex);
32 static LIST_HEAD(all_q_list);
34 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
37 * Check if any of the ctx's have pending work in this hardware queue
39 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
43 for (i = 0; i < hctx->ctx_map.map_size; i++)
44 if (hctx->ctx_map.map[i].word)
50 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
51 struct blk_mq_ctx *ctx)
53 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
56 #define CTX_TO_BIT(hctx, ctx) \
57 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
60 * Mark this ctx as having pending work in this hardware queue
62 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
63 struct blk_mq_ctx *ctx)
65 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
67 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
68 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
71 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
72 struct blk_mq_ctx *ctx)
74 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
76 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
79 static int blk_mq_queue_enter(struct request_queue *q)
84 if (percpu_ref_tryget_live(&q->mq_usage_counter))
87 ret = wait_event_interruptible(q->mq_freeze_wq,
88 !q->mq_freeze_depth || blk_queue_dying(q));
89 if (blk_queue_dying(q))
96 static void blk_mq_queue_exit(struct request_queue *q)
98 percpu_ref_put(&q->mq_usage_counter);
101 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
103 struct request_queue *q =
104 container_of(ref, struct request_queue, mq_usage_counter);
106 wake_up_all(&q->mq_freeze_wq);
110 * Guarantee no request is in use, so we can change any data structure of
111 * the queue afterward.
113 void blk_mq_freeze_queue(struct request_queue *q)
117 spin_lock_irq(q->queue_lock);
118 freeze = !q->mq_freeze_depth++;
119 spin_unlock_irq(q->queue_lock);
123 * XXX: Temporary kludge to work around SCSI blk-mq stall.
124 * SCSI synchronously creates and destroys many queues
125 * back-to-back during probe leading to lengthy stalls.
126 * This will be fixed by keeping ->mq_usage_counter in
127 * atomic mode until genhd registration, but, for now,
128 * let's work around using expedited synchronization.
130 __percpu_ref_kill_expedited(&q->mq_usage_counter);
132 blk_mq_run_queues(q, false);
134 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
137 static void blk_mq_unfreeze_queue(struct request_queue *q)
141 spin_lock_irq(q->queue_lock);
142 wake = !--q->mq_freeze_depth;
143 WARN_ON_ONCE(q->mq_freeze_depth < 0);
144 spin_unlock_irq(q->queue_lock);
146 percpu_ref_reinit(&q->mq_usage_counter);
147 wake_up_all(&q->mq_freeze_wq);
151 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
153 return blk_mq_has_free_tags(hctx->tags);
155 EXPORT_SYMBOL(blk_mq_can_queue);
157 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
158 struct request *rq, unsigned int rw_flags)
160 if (blk_queue_io_stat(q))
161 rw_flags |= REQ_IO_STAT;
163 INIT_LIST_HEAD(&rq->queuelist);
164 /* csd/requeue_work/fifo_time is initialized before use */
167 rq->cmd_flags |= rw_flags;
168 /* do not touch atomic flags, it needs atomic ops against the timer */
170 INIT_HLIST_NODE(&rq->hash);
171 RB_CLEAR_NODE(&rq->rb_node);
174 rq->start_time = jiffies;
175 #ifdef CONFIG_BLK_CGROUP
177 set_start_time_ns(rq);
178 rq->io_start_time_ns = 0;
180 rq->nr_phys_segments = 0;
181 #if defined(CONFIG_BLK_DEV_INTEGRITY)
182 rq->nr_integrity_segments = 0;
185 /* tag was already set */
195 INIT_LIST_HEAD(&rq->timeout_list);
199 rq->end_io_data = NULL;
202 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
205 static struct request *
206 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
211 tag = blk_mq_get_tag(data);
212 if (tag != BLK_MQ_TAG_FAIL) {
213 rq = data->hctx->tags->rqs[tag];
215 if (blk_mq_tag_busy(data->hctx)) {
216 rq->cmd_flags = REQ_MQ_INFLIGHT;
217 atomic_inc(&data->hctx->nr_active);
221 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
228 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
231 struct blk_mq_ctx *ctx;
232 struct blk_mq_hw_ctx *hctx;
234 struct blk_mq_alloc_data alloc_data;
236 if (blk_mq_queue_enter(q))
239 ctx = blk_mq_get_ctx(q);
240 hctx = q->mq_ops->map_queue(q, ctx->cpu);
241 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
242 reserved, ctx, hctx);
244 rq = __blk_mq_alloc_request(&alloc_data, rw);
245 if (!rq && (gfp & __GFP_WAIT)) {
246 __blk_mq_run_hw_queue(hctx);
249 ctx = blk_mq_get_ctx(q);
250 hctx = q->mq_ops->map_queue(q, ctx->cpu);
251 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
253 rq = __blk_mq_alloc_request(&alloc_data, rw);
254 ctx = alloc_data.ctx;
259 EXPORT_SYMBOL(blk_mq_alloc_request);
261 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
262 struct blk_mq_ctx *ctx, struct request *rq)
264 const int tag = rq->tag;
265 struct request_queue *q = rq->q;
267 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
268 atomic_dec(&hctx->nr_active);
271 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
272 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
273 blk_mq_queue_exit(q);
276 void blk_mq_free_request(struct request *rq)
278 struct blk_mq_ctx *ctx = rq->mq_ctx;
279 struct blk_mq_hw_ctx *hctx;
280 struct request_queue *q = rq->q;
282 ctx->rq_completed[rq_is_sync(rq)]++;
284 hctx = q->mq_ops->map_queue(q, ctx->cpu);
285 __blk_mq_free_request(hctx, ctx, rq);
289 * Clone all relevant state from a request that has been put on hold in
290 * the flush state machine into the preallocated flush request that hangs
291 * off the request queue.
293 * For a driver the flush request should be invisible, that's why we are
294 * impersonating the original request here.
296 void blk_mq_clone_flush_request(struct request *flush_rq,
297 struct request *orig_rq)
299 struct blk_mq_hw_ctx *hctx =
300 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
302 flush_rq->mq_ctx = orig_rq->mq_ctx;
303 flush_rq->tag = orig_rq->tag;
304 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
308 inline void __blk_mq_end_io(struct request *rq, int error)
310 blk_account_io_done(rq);
313 rq->end_io(rq, error);
315 if (unlikely(blk_bidi_rq(rq)))
316 blk_mq_free_request(rq->next_rq);
317 blk_mq_free_request(rq);
320 EXPORT_SYMBOL(__blk_mq_end_io);
322 void blk_mq_end_io(struct request *rq, int error)
324 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
326 __blk_mq_end_io(rq, error);
328 EXPORT_SYMBOL(blk_mq_end_io);
330 static void __blk_mq_complete_request_remote(void *data)
332 struct request *rq = data;
334 rq->q->softirq_done_fn(rq);
337 static void blk_mq_ipi_complete_request(struct request *rq)
339 struct blk_mq_ctx *ctx = rq->mq_ctx;
343 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
344 rq->q->softirq_done_fn(rq);
349 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
350 shared = cpus_share_cache(cpu, ctx->cpu);
352 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
353 rq->csd.func = __blk_mq_complete_request_remote;
356 smp_call_function_single_async(ctx->cpu, &rq->csd);
358 rq->q->softirq_done_fn(rq);
363 void __blk_mq_complete_request(struct request *rq)
365 struct request_queue *q = rq->q;
367 if (!q->softirq_done_fn)
368 blk_mq_end_io(rq, rq->errors);
370 blk_mq_ipi_complete_request(rq);
374 * blk_mq_complete_request - end I/O on a request
375 * @rq: the request being processed
378 * Ends all I/O on a request. It does not handle partial completions.
379 * The actual completion happens out-of-order, through a IPI handler.
381 void blk_mq_complete_request(struct request *rq)
383 struct request_queue *q = rq->q;
385 if (unlikely(blk_should_fake_timeout(q)))
387 if (!blk_mark_rq_complete(rq))
388 __blk_mq_complete_request(rq);
390 EXPORT_SYMBOL(blk_mq_complete_request);
392 static void blk_mq_start_request(struct request *rq, bool last)
394 struct request_queue *q = rq->q;
396 trace_block_rq_issue(q, rq);
398 rq->resid_len = blk_rq_bytes(rq);
399 if (unlikely(blk_bidi_rq(rq)))
400 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
405 * Ensure that ->deadline is visible before set the started
406 * flag and clear the completed flag.
408 smp_mb__before_atomic();
411 * Mark us as started and clear complete. Complete might have been
412 * set if requeue raced with timeout, which then marked it as
413 * complete. So be sure to clear complete again when we start
414 * the request, otherwise we'll ignore the completion event.
416 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
417 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
418 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
419 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
421 if (q->dma_drain_size && blk_rq_bytes(rq)) {
423 * Make sure space for the drain appears. We know we can do
424 * this because max_hw_segments has been adjusted to be one
425 * fewer than the device can handle.
427 rq->nr_phys_segments++;
431 * Flag the last request in the series so that drivers know when IO
432 * should be kicked off, if they don't do it on a per-request basis.
434 * Note: the flag isn't the only condition drivers should do kick off.
435 * If drive is busy, the last request might not have the bit set.
438 rq->cmd_flags |= REQ_END;
441 static void __blk_mq_requeue_request(struct request *rq)
443 struct request_queue *q = rq->q;
445 trace_block_rq_requeue(q, rq);
446 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
448 rq->cmd_flags &= ~REQ_END;
450 if (q->dma_drain_size && blk_rq_bytes(rq))
451 rq->nr_phys_segments--;
454 void blk_mq_requeue_request(struct request *rq)
456 __blk_mq_requeue_request(rq);
457 blk_clear_rq_complete(rq);
459 BUG_ON(blk_queued_rq(rq));
460 blk_mq_add_to_requeue_list(rq, true);
462 EXPORT_SYMBOL(blk_mq_requeue_request);
464 static void blk_mq_requeue_work(struct work_struct *work)
466 struct request_queue *q =
467 container_of(work, struct request_queue, requeue_work);
469 struct request *rq, *next;
472 spin_lock_irqsave(&q->requeue_lock, flags);
473 list_splice_init(&q->requeue_list, &rq_list);
474 spin_unlock_irqrestore(&q->requeue_lock, flags);
476 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
477 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
480 rq->cmd_flags &= ~REQ_SOFTBARRIER;
481 list_del_init(&rq->queuelist);
482 blk_mq_insert_request(rq, true, false, false);
485 while (!list_empty(&rq_list)) {
486 rq = list_entry(rq_list.next, struct request, queuelist);
487 list_del_init(&rq->queuelist);
488 blk_mq_insert_request(rq, false, false, false);
492 * Use the start variant of queue running here, so that running
493 * the requeue work will kick stopped queues.
495 blk_mq_start_hw_queues(q);
498 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
500 struct request_queue *q = rq->q;
504 * We abuse this flag that is otherwise used by the I/O scheduler to
505 * request head insertation from the workqueue.
507 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
509 spin_lock_irqsave(&q->requeue_lock, flags);
511 rq->cmd_flags |= REQ_SOFTBARRIER;
512 list_add(&rq->queuelist, &q->requeue_list);
514 list_add_tail(&rq->queuelist, &q->requeue_list);
516 spin_unlock_irqrestore(&q->requeue_lock, flags);
518 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
520 void blk_mq_kick_requeue_list(struct request_queue *q)
522 kblockd_schedule_work(&q->requeue_work);
524 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
526 static inline bool is_flush_request(struct request *rq, unsigned int tag)
528 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
529 rq->q->flush_rq->tag == tag);
532 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
534 struct request *rq = tags->rqs[tag];
536 if (!is_flush_request(rq, tag))
539 return rq->q->flush_rq;
541 EXPORT_SYMBOL(blk_mq_tag_to_rq);
543 struct blk_mq_timeout_data {
544 struct blk_mq_hw_ctx *hctx;
546 unsigned int *next_set;
549 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
551 struct blk_mq_timeout_data *data = __data;
552 struct blk_mq_hw_ctx *hctx = data->hctx;
555 /* It may not be in flight yet (this is where
556 * the REQ_ATOMIC_STARTED flag comes in). The requests are
557 * statically allocated, so we know it's always safe to access the
558 * memory associated with a bit offset into ->rqs[].
564 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
565 if (tag >= hctx->tags->nr_tags)
568 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
569 if (rq->q != hctx->queue)
571 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
574 blk_rq_check_expired(rq, data->next, data->next_set);
578 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
580 unsigned int *next_set)
582 struct blk_mq_timeout_data data = {
585 .next_set = next_set,
589 * Ask the tagging code to iterate busy requests, so we can
590 * check them for timeout.
592 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
595 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
597 struct request_queue *q = rq->q;
600 * We know that complete is set at this point. If STARTED isn't set
601 * anymore, then the request isn't active and the "timeout" should
602 * just be ignored. This can happen due to the bitflag ordering.
603 * Timeout first checks if STARTED is set, and if it is, assumes
604 * the request is active. But if we race with completion, then
605 * we both flags will get cleared. So check here again, and ignore
606 * a timeout event with a request that isn't active.
608 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
609 return BLK_EH_NOT_HANDLED;
611 if (!q->mq_ops->timeout)
612 return BLK_EH_RESET_TIMER;
614 return q->mq_ops->timeout(rq);
617 static void blk_mq_rq_timer(unsigned long data)
619 struct request_queue *q = (struct request_queue *) data;
620 struct blk_mq_hw_ctx *hctx;
621 unsigned long next = 0;
624 queue_for_each_hw_ctx(q, hctx, i) {
626 * If not software queues are currently mapped to this
627 * hardware queue, there's nothing to check
629 if (!hctx->nr_ctx || !hctx->tags)
632 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
636 next = blk_rq_timeout(round_jiffies_up(next));
637 mod_timer(&q->timeout, next);
639 queue_for_each_hw_ctx(q, hctx, i)
640 blk_mq_tag_idle(hctx);
645 * Reverse check our software queue for entries that we could potentially
646 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
647 * too much time checking for merges.
649 static bool blk_mq_attempt_merge(struct request_queue *q,
650 struct blk_mq_ctx *ctx, struct bio *bio)
655 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
661 if (!blk_rq_merge_ok(rq, bio))
664 el_ret = blk_try_merge(rq, bio);
665 if (el_ret == ELEVATOR_BACK_MERGE) {
666 if (bio_attempt_back_merge(q, rq, bio)) {
671 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
672 if (bio_attempt_front_merge(q, rq, bio)) {
684 * Process software queues that have been marked busy, splicing them
685 * to the for-dispatch
687 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
689 struct blk_mq_ctx *ctx;
692 for (i = 0; i < hctx->ctx_map.map_size; i++) {
693 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
694 unsigned int off, bit;
700 off = i * hctx->ctx_map.bits_per_word;
702 bit = find_next_bit(&bm->word, bm->depth, bit);
703 if (bit >= bm->depth)
706 ctx = hctx->ctxs[bit + off];
707 clear_bit(bit, &bm->word);
708 spin_lock(&ctx->lock);
709 list_splice_tail_init(&ctx->rq_list, list);
710 spin_unlock(&ctx->lock);
718 * Run this hardware queue, pulling any software queues mapped to it in.
719 * Note that this function currently has various problems around ordering
720 * of IO. In particular, we'd like FIFO behaviour on handling existing
721 * items on the hctx->dispatch list. Ignore that for now.
723 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
725 struct request_queue *q = hctx->queue;
730 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
732 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
738 * Touch any software queue that has pending entries.
740 flush_busy_ctxs(hctx, &rq_list);
743 * If we have previous entries on our dispatch list, grab them
744 * and stuff them at the front for more fair dispatch.
746 if (!list_empty_careful(&hctx->dispatch)) {
747 spin_lock(&hctx->lock);
748 if (!list_empty(&hctx->dispatch))
749 list_splice_init(&hctx->dispatch, &rq_list);
750 spin_unlock(&hctx->lock);
754 * Now process all the entries, sending them to the driver.
757 while (!list_empty(&rq_list)) {
760 rq = list_first_entry(&rq_list, struct request, queuelist);
761 list_del_init(&rq->queuelist);
763 blk_mq_start_request(rq, list_empty(&rq_list));
765 ret = q->mq_ops->queue_rq(hctx, rq);
767 case BLK_MQ_RQ_QUEUE_OK:
770 case BLK_MQ_RQ_QUEUE_BUSY:
771 list_add(&rq->queuelist, &rq_list);
772 __blk_mq_requeue_request(rq);
775 pr_err("blk-mq: bad return on queue: %d\n", ret);
776 case BLK_MQ_RQ_QUEUE_ERROR:
778 blk_mq_end_io(rq, rq->errors);
782 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
787 hctx->dispatched[0]++;
788 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
789 hctx->dispatched[ilog2(queued) + 1]++;
792 * Any items that need requeuing? Stuff them into hctx->dispatch,
793 * that is where we will continue on next queue run.
795 if (!list_empty(&rq_list)) {
796 spin_lock(&hctx->lock);
797 list_splice(&rq_list, &hctx->dispatch);
798 spin_unlock(&hctx->lock);
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 * For now we just round-robin here, switching for every
806 * BLK_MQ_CPU_WORK_BATCH queued items.
808 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
810 int cpu = hctx->next_cpu;
812 if (--hctx->next_cpu_batch <= 0) {
815 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
816 if (next_cpu >= nr_cpu_ids)
817 next_cpu = cpumask_first(hctx->cpumask);
819 hctx->next_cpu = next_cpu;
820 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
826 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
828 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
831 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
832 __blk_mq_run_hw_queue(hctx);
833 else if (hctx->queue->nr_hw_queues == 1)
834 kblockd_schedule_delayed_work(&hctx->run_work, 0);
838 cpu = blk_mq_hctx_next_cpu(hctx);
839 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
843 void blk_mq_run_queues(struct request_queue *q, bool async)
845 struct blk_mq_hw_ctx *hctx;
848 queue_for_each_hw_ctx(q, hctx, i) {
849 if ((!blk_mq_hctx_has_pending(hctx) &&
850 list_empty_careful(&hctx->dispatch)) ||
851 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
855 blk_mq_run_hw_queue(hctx, async);
859 EXPORT_SYMBOL(blk_mq_run_queues);
861 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
863 cancel_delayed_work(&hctx->run_work);
864 cancel_delayed_work(&hctx->delay_work);
865 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
867 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
869 void blk_mq_stop_hw_queues(struct request_queue *q)
871 struct blk_mq_hw_ctx *hctx;
874 queue_for_each_hw_ctx(q, hctx, i)
875 blk_mq_stop_hw_queue(hctx);
877 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
879 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
881 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
884 blk_mq_run_hw_queue(hctx, false);
887 EXPORT_SYMBOL(blk_mq_start_hw_queue);
889 void blk_mq_start_hw_queues(struct request_queue *q)
891 struct blk_mq_hw_ctx *hctx;
894 queue_for_each_hw_ctx(q, hctx, i)
895 blk_mq_start_hw_queue(hctx);
897 EXPORT_SYMBOL(blk_mq_start_hw_queues);
900 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
902 struct blk_mq_hw_ctx *hctx;
905 queue_for_each_hw_ctx(q, hctx, i) {
906 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
909 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
911 blk_mq_run_hw_queue(hctx, async);
915 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
917 static void blk_mq_run_work_fn(struct work_struct *work)
919 struct blk_mq_hw_ctx *hctx;
921 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
923 __blk_mq_run_hw_queue(hctx);
926 static void blk_mq_delay_work_fn(struct work_struct *work)
928 struct blk_mq_hw_ctx *hctx;
930 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
932 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
933 __blk_mq_run_hw_queue(hctx);
936 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
938 unsigned long tmo = msecs_to_jiffies(msecs);
940 if (hctx->queue->nr_hw_queues == 1)
941 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
945 cpu = blk_mq_hctx_next_cpu(hctx);
946 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
949 EXPORT_SYMBOL(blk_mq_delay_queue);
951 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
952 struct request *rq, bool at_head)
954 struct blk_mq_ctx *ctx = rq->mq_ctx;
956 trace_block_rq_insert(hctx->queue, rq);
959 list_add(&rq->queuelist, &ctx->rq_list);
961 list_add_tail(&rq->queuelist, &ctx->rq_list);
963 blk_mq_hctx_mark_pending(hctx, ctx);
966 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
969 struct request_queue *q = rq->q;
970 struct blk_mq_hw_ctx *hctx;
971 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
973 current_ctx = blk_mq_get_ctx(q);
974 if (!cpu_online(ctx->cpu))
975 rq->mq_ctx = ctx = current_ctx;
977 hctx = q->mq_ops->map_queue(q, ctx->cpu);
979 spin_lock(&ctx->lock);
980 __blk_mq_insert_request(hctx, rq, at_head);
981 spin_unlock(&ctx->lock);
984 blk_mq_run_hw_queue(hctx, async);
986 blk_mq_put_ctx(current_ctx);
989 static void blk_mq_insert_requests(struct request_queue *q,
990 struct blk_mq_ctx *ctx,
991 struct list_head *list,
996 struct blk_mq_hw_ctx *hctx;
997 struct blk_mq_ctx *current_ctx;
999 trace_block_unplug(q, depth, !from_schedule);
1001 current_ctx = blk_mq_get_ctx(q);
1003 if (!cpu_online(ctx->cpu))
1005 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1008 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1011 spin_lock(&ctx->lock);
1012 while (!list_empty(list)) {
1015 rq = list_first_entry(list, struct request, queuelist);
1016 list_del_init(&rq->queuelist);
1018 __blk_mq_insert_request(hctx, rq, false);
1020 spin_unlock(&ctx->lock);
1022 blk_mq_run_hw_queue(hctx, from_schedule);
1023 blk_mq_put_ctx(current_ctx);
1026 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1028 struct request *rqa = container_of(a, struct request, queuelist);
1029 struct request *rqb = container_of(b, struct request, queuelist);
1031 return !(rqa->mq_ctx < rqb->mq_ctx ||
1032 (rqa->mq_ctx == rqb->mq_ctx &&
1033 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1036 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1038 struct blk_mq_ctx *this_ctx;
1039 struct request_queue *this_q;
1042 LIST_HEAD(ctx_list);
1045 list_splice_init(&plug->mq_list, &list);
1047 list_sort(NULL, &list, plug_ctx_cmp);
1053 while (!list_empty(&list)) {
1054 rq = list_entry_rq(list.next);
1055 list_del_init(&rq->queuelist);
1057 if (rq->mq_ctx != this_ctx) {
1059 blk_mq_insert_requests(this_q, this_ctx,
1064 this_ctx = rq->mq_ctx;
1070 list_add_tail(&rq->queuelist, &ctx_list);
1074 * If 'this_ctx' is set, we know we have entries to complete
1075 * on 'ctx_list'. Do those.
1078 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1083 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1085 init_request_from_bio(rq, bio);
1087 if (blk_do_io_stat(rq))
1088 blk_account_io_start(rq, 1);
1091 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1093 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1094 !blk_queue_nomerges(hctx->queue);
1097 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1098 struct blk_mq_ctx *ctx,
1099 struct request *rq, struct bio *bio)
1101 if (!hctx_allow_merges(hctx)) {
1102 blk_mq_bio_to_request(rq, bio);
1103 spin_lock(&ctx->lock);
1105 __blk_mq_insert_request(hctx, rq, false);
1106 spin_unlock(&ctx->lock);
1109 struct request_queue *q = hctx->queue;
1111 spin_lock(&ctx->lock);
1112 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1113 blk_mq_bio_to_request(rq, bio);
1117 spin_unlock(&ctx->lock);
1118 __blk_mq_free_request(hctx, ctx, rq);
1123 struct blk_map_ctx {
1124 struct blk_mq_hw_ctx *hctx;
1125 struct blk_mq_ctx *ctx;
1128 static struct request *blk_mq_map_request(struct request_queue *q,
1130 struct blk_map_ctx *data)
1132 struct blk_mq_hw_ctx *hctx;
1133 struct blk_mq_ctx *ctx;
1135 int rw = bio_data_dir(bio);
1136 struct blk_mq_alloc_data alloc_data;
1138 if (unlikely(blk_mq_queue_enter(q))) {
1139 bio_endio(bio, -EIO);
1143 ctx = blk_mq_get_ctx(q);
1144 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1146 if (rw_is_sync(bio->bi_rw))
1149 trace_block_getrq(q, bio, rw);
1150 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1152 rq = __blk_mq_alloc_request(&alloc_data, rw);
1153 if (unlikely(!rq)) {
1154 __blk_mq_run_hw_queue(hctx);
1155 blk_mq_put_ctx(ctx);
1156 trace_block_sleeprq(q, bio, rw);
1158 ctx = blk_mq_get_ctx(q);
1159 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1160 blk_mq_set_alloc_data(&alloc_data, q,
1161 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1162 rq = __blk_mq_alloc_request(&alloc_data, rw);
1163 ctx = alloc_data.ctx;
1164 hctx = alloc_data.hctx;
1174 * Multiple hardware queue variant. This will not use per-process plugs,
1175 * but will attempt to bypass the hctx queueing if we can go straight to
1176 * hardware for SYNC IO.
1178 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1180 const int is_sync = rw_is_sync(bio->bi_rw);
1181 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1182 struct blk_map_ctx data;
1185 blk_queue_bounce(q, &bio);
1187 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1188 bio_endio(bio, -EIO);
1192 rq = blk_mq_map_request(q, bio, &data);
1196 if (unlikely(is_flush_fua)) {
1197 blk_mq_bio_to_request(rq, bio);
1198 blk_insert_flush(rq);
1205 blk_mq_bio_to_request(rq, bio);
1206 blk_mq_start_request(rq, true);
1209 * For OK queue, we are done. For error, kill it. Any other
1210 * error (busy), just add it to our list as we previously
1213 ret = q->mq_ops->queue_rq(data.hctx, rq);
1214 if (ret == BLK_MQ_RQ_QUEUE_OK)
1217 __blk_mq_requeue_request(rq);
1219 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1221 blk_mq_end_io(rq, rq->errors);
1227 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1229 * For a SYNC request, send it to the hardware immediately. For
1230 * an ASYNC request, just ensure that we run it later on. The
1231 * latter allows for merging opportunities and more efficient
1235 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1238 blk_mq_put_ctx(data.ctx);
1242 * Single hardware queue variant. This will attempt to use any per-process
1243 * plug for merging and IO deferral.
1245 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1247 const int is_sync = rw_is_sync(bio->bi_rw);
1248 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1249 unsigned int use_plug, request_count = 0;
1250 struct blk_map_ctx data;
1254 * If we have multiple hardware queues, just go directly to
1255 * one of those for sync IO.
1257 use_plug = !is_flush_fua && !is_sync;
1259 blk_queue_bounce(q, &bio);
1261 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1262 bio_endio(bio, -EIO);
1266 if (use_plug && !blk_queue_nomerges(q) &&
1267 blk_attempt_plug_merge(q, bio, &request_count))
1270 rq = blk_mq_map_request(q, bio, &data);
1274 if (unlikely(is_flush_fua)) {
1275 blk_mq_bio_to_request(rq, bio);
1276 blk_insert_flush(rq);
1281 * A task plug currently exists. Since this is completely lockless,
1282 * utilize that to temporarily store requests until the task is
1283 * either done or scheduled away.
1286 struct blk_plug *plug = current->plug;
1289 blk_mq_bio_to_request(rq, bio);
1290 if (list_empty(&plug->mq_list))
1291 trace_block_plug(q);
1292 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1293 blk_flush_plug_list(plug, false);
1294 trace_block_plug(q);
1296 list_add_tail(&rq->queuelist, &plug->mq_list);
1297 blk_mq_put_ctx(data.ctx);
1302 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1304 * For a SYNC request, send it to the hardware immediately. For
1305 * an ASYNC request, just ensure that we run it later on. The
1306 * latter allows for merging opportunities and more efficient
1310 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1313 blk_mq_put_ctx(data.ctx);
1317 * Default mapping to a software queue, since we use one per CPU.
1319 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1321 return q->queue_hw_ctx[q->mq_map[cpu]];
1323 EXPORT_SYMBOL(blk_mq_map_queue);
1325 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1326 struct blk_mq_tags *tags, unsigned int hctx_idx)
1330 if (tags->rqs && set->ops->exit_request) {
1333 for (i = 0; i < tags->nr_tags; i++) {
1336 set->ops->exit_request(set->driver_data, tags->rqs[i],
1338 tags->rqs[i] = NULL;
1342 while (!list_empty(&tags->page_list)) {
1343 page = list_first_entry(&tags->page_list, struct page, lru);
1344 list_del_init(&page->lru);
1345 __free_pages(page, page->private);
1350 blk_mq_free_tags(tags);
1353 static size_t order_to_size(unsigned int order)
1355 return (size_t)PAGE_SIZE << order;
1358 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1359 unsigned int hctx_idx)
1361 struct blk_mq_tags *tags;
1362 unsigned int i, j, entries_per_page, max_order = 4;
1363 size_t rq_size, left;
1365 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1370 INIT_LIST_HEAD(&tags->page_list);
1372 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1373 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1376 blk_mq_free_tags(tags);
1381 * rq_size is the size of the request plus driver payload, rounded
1382 * to the cacheline size
1384 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1386 left = rq_size * set->queue_depth;
1388 for (i = 0; i < set->queue_depth; ) {
1389 int this_order = max_order;
1394 while (left < order_to_size(this_order - 1) && this_order)
1398 page = alloc_pages_node(set->numa_node,
1399 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1405 if (order_to_size(this_order) < rq_size)
1412 page->private = this_order;
1413 list_add_tail(&page->lru, &tags->page_list);
1415 p = page_address(page);
1416 entries_per_page = order_to_size(this_order) / rq_size;
1417 to_do = min(entries_per_page, set->queue_depth - i);
1418 left -= to_do * rq_size;
1419 for (j = 0; j < to_do; j++) {
1421 tags->rqs[i]->atomic_flags = 0;
1422 tags->rqs[i]->cmd_flags = 0;
1423 if (set->ops->init_request) {
1424 if (set->ops->init_request(set->driver_data,
1425 tags->rqs[i], hctx_idx, i,
1427 tags->rqs[i] = NULL;
1440 blk_mq_free_rq_map(set, tags, hctx_idx);
1444 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1449 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1451 unsigned int bpw = 8, total, num_maps, i;
1453 bitmap->bits_per_word = bpw;
1455 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1456 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1461 bitmap->map_size = num_maps;
1464 for (i = 0; i < num_maps; i++) {
1465 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1466 total -= bitmap->map[i].depth;
1472 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1474 struct request_queue *q = hctx->queue;
1475 struct blk_mq_ctx *ctx;
1479 * Move ctx entries to new CPU, if this one is going away.
1481 ctx = __blk_mq_get_ctx(q, cpu);
1483 spin_lock(&ctx->lock);
1484 if (!list_empty(&ctx->rq_list)) {
1485 list_splice_init(&ctx->rq_list, &tmp);
1486 blk_mq_hctx_clear_pending(hctx, ctx);
1488 spin_unlock(&ctx->lock);
1490 if (list_empty(&tmp))
1493 ctx = blk_mq_get_ctx(q);
1494 spin_lock(&ctx->lock);
1496 while (!list_empty(&tmp)) {
1499 rq = list_first_entry(&tmp, struct request, queuelist);
1501 list_move_tail(&rq->queuelist, &ctx->rq_list);
1504 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1505 blk_mq_hctx_mark_pending(hctx, ctx);
1507 spin_unlock(&ctx->lock);
1509 blk_mq_run_hw_queue(hctx, true);
1510 blk_mq_put_ctx(ctx);
1514 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1516 struct request_queue *q = hctx->queue;
1517 struct blk_mq_tag_set *set = q->tag_set;
1519 if (set->tags[hctx->queue_num])
1522 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1523 if (!set->tags[hctx->queue_num])
1526 hctx->tags = set->tags[hctx->queue_num];
1530 static int blk_mq_hctx_notify(void *data, unsigned long action,
1533 struct blk_mq_hw_ctx *hctx = data;
1535 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1536 return blk_mq_hctx_cpu_offline(hctx, cpu);
1537 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1538 return blk_mq_hctx_cpu_online(hctx, cpu);
1543 static void blk_mq_exit_hw_queues(struct request_queue *q,
1544 struct blk_mq_tag_set *set, int nr_queue)
1546 struct blk_mq_hw_ctx *hctx;
1549 queue_for_each_hw_ctx(q, hctx, i) {
1553 blk_mq_tag_idle(hctx);
1555 if (set->ops->exit_hctx)
1556 set->ops->exit_hctx(hctx, i);
1558 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1560 blk_mq_free_bitmap(&hctx->ctx_map);
1565 static void blk_mq_free_hw_queues(struct request_queue *q,
1566 struct blk_mq_tag_set *set)
1568 struct blk_mq_hw_ctx *hctx;
1571 queue_for_each_hw_ctx(q, hctx, i) {
1572 free_cpumask_var(hctx->cpumask);
1577 static int blk_mq_init_hw_queues(struct request_queue *q,
1578 struct blk_mq_tag_set *set)
1580 struct blk_mq_hw_ctx *hctx;
1584 * Initialize hardware queues
1586 queue_for_each_hw_ctx(q, hctx, i) {
1589 node = hctx->numa_node;
1590 if (node == NUMA_NO_NODE)
1591 node = hctx->numa_node = set->numa_node;
1593 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1594 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1595 spin_lock_init(&hctx->lock);
1596 INIT_LIST_HEAD(&hctx->dispatch);
1598 hctx->queue_num = i;
1599 hctx->flags = set->flags;
1600 hctx->cmd_size = set->cmd_size;
1602 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1603 blk_mq_hctx_notify, hctx);
1604 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1606 hctx->tags = set->tags[i];
1609 * Allocate space for all possible cpus to avoid allocation at
1612 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1617 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1622 if (set->ops->init_hctx &&
1623 set->ops->init_hctx(hctx, set->driver_data, i))
1627 if (i == q->nr_hw_queues)
1633 blk_mq_exit_hw_queues(q, set, i);
1638 static void blk_mq_init_cpu_queues(struct request_queue *q,
1639 unsigned int nr_hw_queues)
1643 for_each_possible_cpu(i) {
1644 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1645 struct blk_mq_hw_ctx *hctx;
1647 memset(__ctx, 0, sizeof(*__ctx));
1649 spin_lock_init(&__ctx->lock);
1650 INIT_LIST_HEAD(&__ctx->rq_list);
1653 /* If the cpu isn't online, the cpu is mapped to first hctx */
1657 hctx = q->mq_ops->map_queue(q, i);
1658 cpumask_set_cpu(i, hctx->cpumask);
1662 * Set local node, IFF we have more than one hw queue. If
1663 * not, we remain on the home node of the device
1665 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1666 hctx->numa_node = cpu_to_node(i);
1670 static void blk_mq_map_swqueue(struct request_queue *q)
1673 struct blk_mq_hw_ctx *hctx;
1674 struct blk_mq_ctx *ctx;
1676 queue_for_each_hw_ctx(q, hctx, i) {
1677 cpumask_clear(hctx->cpumask);
1682 * Map software to hardware queues
1684 queue_for_each_ctx(q, ctx, i) {
1685 /* If the cpu isn't online, the cpu is mapped to first hctx */
1689 hctx = q->mq_ops->map_queue(q, i);
1690 cpumask_set_cpu(i, hctx->cpumask);
1691 ctx->index_hw = hctx->nr_ctx;
1692 hctx->ctxs[hctx->nr_ctx++] = ctx;
1695 queue_for_each_hw_ctx(q, hctx, i) {
1697 * If no software queues are mapped to this hardware queue,
1698 * disable it and free the request entries.
1700 if (!hctx->nr_ctx) {
1701 struct blk_mq_tag_set *set = q->tag_set;
1704 blk_mq_free_rq_map(set, set->tags[i], i);
1705 set->tags[i] = NULL;
1712 * Initialize batch roundrobin counts
1714 hctx->next_cpu = cpumask_first(hctx->cpumask);
1715 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1719 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1721 struct blk_mq_hw_ctx *hctx;
1722 struct request_queue *q;
1726 if (set->tag_list.next == set->tag_list.prev)
1731 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1732 blk_mq_freeze_queue(q);
1734 queue_for_each_hw_ctx(q, hctx, i) {
1736 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1738 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1740 blk_mq_unfreeze_queue(q);
1744 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1746 struct blk_mq_tag_set *set = q->tag_set;
1748 mutex_lock(&set->tag_list_lock);
1749 list_del_init(&q->tag_set_list);
1750 blk_mq_update_tag_set_depth(set);
1751 mutex_unlock(&set->tag_list_lock);
1754 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1755 struct request_queue *q)
1759 mutex_lock(&set->tag_list_lock);
1760 list_add_tail(&q->tag_set_list, &set->tag_list);
1761 blk_mq_update_tag_set_depth(set);
1762 mutex_unlock(&set->tag_list_lock);
1765 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1767 struct blk_mq_hw_ctx **hctxs;
1768 struct blk_mq_ctx __percpu *ctx;
1769 struct request_queue *q;
1773 ctx = alloc_percpu(struct blk_mq_ctx);
1775 return ERR_PTR(-ENOMEM);
1777 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1783 map = blk_mq_make_queue_map(set);
1787 for (i = 0; i < set->nr_hw_queues; i++) {
1788 int node = blk_mq_hw_queue_to_node(map, i);
1790 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1795 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1798 atomic_set(&hctxs[i]->nr_active, 0);
1799 hctxs[i]->numa_node = node;
1800 hctxs[i]->queue_num = i;
1803 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1807 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1811 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1812 blk_queue_rq_timeout(q, 30000);
1814 q->nr_queues = nr_cpu_ids;
1815 q->nr_hw_queues = set->nr_hw_queues;
1819 q->queue_hw_ctx = hctxs;
1821 q->mq_ops = set->ops;
1822 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1824 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1825 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1827 q->sg_reserved_size = INT_MAX;
1829 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1830 INIT_LIST_HEAD(&q->requeue_list);
1831 spin_lock_init(&q->requeue_lock);
1833 if (q->nr_hw_queues > 1)
1834 blk_queue_make_request(q, blk_mq_make_request);
1836 blk_queue_make_request(q, blk_sq_make_request);
1838 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1840 blk_queue_rq_timeout(q, set->timeout);
1843 * Do this after blk_queue_make_request() overrides it...
1845 q->nr_requests = set->queue_depth;
1847 if (set->ops->complete)
1848 blk_queue_softirq_done(q, set->ops->complete);
1850 blk_mq_init_flush(q);
1851 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1853 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1854 set->cmd_size, cache_line_size()),
1859 if (blk_mq_init_hw_queues(q, set))
1862 mutex_lock(&all_q_mutex);
1863 list_add_tail(&q->all_q_node, &all_q_list);
1864 mutex_unlock(&all_q_mutex);
1866 blk_mq_add_queue_tag_set(set, q);
1868 blk_mq_map_swqueue(q);
1875 blk_cleanup_queue(q);
1878 for (i = 0; i < set->nr_hw_queues; i++) {
1881 free_cpumask_var(hctxs[i]->cpumask);
1888 return ERR_PTR(-ENOMEM);
1890 EXPORT_SYMBOL(blk_mq_init_queue);
1892 void blk_mq_free_queue(struct request_queue *q)
1894 struct blk_mq_tag_set *set = q->tag_set;
1896 blk_mq_del_queue_tag_set(q);
1898 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1899 blk_mq_free_hw_queues(q, set);
1901 percpu_ref_exit(&q->mq_usage_counter);
1903 free_percpu(q->queue_ctx);
1904 kfree(q->queue_hw_ctx);
1907 q->queue_ctx = NULL;
1908 q->queue_hw_ctx = NULL;
1911 mutex_lock(&all_q_mutex);
1912 list_del_init(&q->all_q_node);
1913 mutex_unlock(&all_q_mutex);
1916 /* Basically redo blk_mq_init_queue with queue frozen */
1917 static void blk_mq_queue_reinit(struct request_queue *q)
1919 blk_mq_freeze_queue(q);
1921 blk_mq_sysfs_unregister(q);
1923 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1926 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1927 * we should change hctx numa_node according to new topology (this
1928 * involves free and re-allocate memory, worthy doing?)
1931 blk_mq_map_swqueue(q);
1933 blk_mq_sysfs_register(q);
1935 blk_mq_unfreeze_queue(q);
1938 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1939 unsigned long action, void *hcpu)
1941 struct request_queue *q;
1944 * Before new mappings are established, hotadded cpu might already
1945 * start handling requests. This doesn't break anything as we map
1946 * offline CPUs to first hardware queue. We will re-init the queue
1947 * below to get optimal settings.
1949 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1950 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1953 mutex_lock(&all_q_mutex);
1954 list_for_each_entry(q, &all_q_list, all_q_node)
1955 blk_mq_queue_reinit(q);
1956 mutex_unlock(&all_q_mutex);
1960 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1964 for (i = 0; i < set->nr_hw_queues; i++) {
1965 set->tags[i] = blk_mq_init_rq_map(set, i);
1974 blk_mq_free_rq_map(set, set->tags[i], i);
1980 * Allocate the request maps associated with this tag_set. Note that this
1981 * may reduce the depth asked for, if memory is tight. set->queue_depth
1982 * will be updated to reflect the allocated depth.
1984 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1989 depth = set->queue_depth;
1991 err = __blk_mq_alloc_rq_maps(set);
1995 set->queue_depth >>= 1;
1996 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2000 } while (set->queue_depth);
2002 if (!set->queue_depth || err) {
2003 pr_err("blk-mq: failed to allocate request map\n");
2007 if (depth != set->queue_depth)
2008 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2009 depth, set->queue_depth);
2015 * Alloc a tag set to be associated with one or more request queues.
2016 * May fail with EINVAL for various error conditions. May adjust the
2017 * requested depth down, if if it too large. In that case, the set
2018 * value will be stored in set->queue_depth.
2020 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2022 if (!set->nr_hw_queues)
2024 if (!set->queue_depth)
2026 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2029 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2032 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2033 pr_info("blk-mq: reduced tag depth to %u\n",
2035 set->queue_depth = BLK_MQ_MAX_DEPTH;
2038 set->tags = kmalloc_node(set->nr_hw_queues *
2039 sizeof(struct blk_mq_tags *),
2040 GFP_KERNEL, set->numa_node);
2044 if (blk_mq_alloc_rq_maps(set))
2047 mutex_init(&set->tag_list_lock);
2048 INIT_LIST_HEAD(&set->tag_list);
2056 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2058 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2062 for (i = 0; i < set->nr_hw_queues; i++) {
2064 blk_mq_free_rq_map(set, set->tags[i], i);
2070 EXPORT_SYMBOL(blk_mq_free_tag_set);
2072 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2074 struct blk_mq_tag_set *set = q->tag_set;
2075 struct blk_mq_hw_ctx *hctx;
2078 if (!set || nr > set->queue_depth)
2082 queue_for_each_hw_ctx(q, hctx, i) {
2083 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2089 q->nr_requests = nr;
2094 void blk_mq_disable_hotplug(void)
2096 mutex_lock(&all_q_mutex);
2099 void blk_mq_enable_hotplug(void)
2101 mutex_unlock(&all_q_mutex);
2104 static int __init blk_mq_init(void)
2108 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2112 subsys_initcall(blk_mq_init);