1 // SPDX-License-Identifier: GPL-2.0
3 * MQ Deadline i/o scheduler - adaptation of the legacy deadline scheduler,
4 * for the blk-mq scheduling framework
6 * Copyright (C) 2016 Jens Axboe <axboe@kernel.dk>
8 #include <linux/kernel.h>
10 #include <linux/blkdev.h>
11 #include <linux/blk-mq.h>
12 #include <linux/elevator.h>
13 #include <linux/bio.h>
14 #include <linux/module.h>
15 #include <linux/slab.h>
16 #include <linux/init.h>
17 #include <linux/compiler.h>
18 #include <linux/rbtree.h>
19 #include <linux/sbitmap.h>
23 #include "blk-mq-debugfs.h"
24 #include "blk-mq-tag.h"
25 #include "blk-mq-sched.h"
28 * See Documentation/block/deadline-iosched.rst
30 static const int read_expire = HZ / 2; /* max time before a read is submitted. */
31 static const int write_expire = 5 * HZ; /* ditto for writes, these limits are SOFT! */
32 static const int writes_starved = 2; /* max times reads can starve a write */
33 static const int fifo_batch = 16; /* # of sequential requests treated as one
34 by the above parameters. For throughput. */
36 struct deadline_data {
42 * requests (deadline_rq s) are present on both sort_list and fifo_list
44 struct rb_root sort_list[2];
45 struct list_head fifo_list[2];
48 * next in sort order. read, write or both are NULL
50 struct request *next_rq[2];
51 unsigned int batching; /* number of sequential requests made */
52 unsigned int starved; /* times reads have starved writes */
55 * settings that change how the i/o scheduler behaves
64 struct list_head dispatch;
67 static inline struct rb_root *
68 deadline_rb_root(struct deadline_data *dd, struct request *rq)
70 return &dd->sort_list[rq_data_dir(rq)];
74 * get the request after `rq' in sector-sorted order
76 static inline struct request *
77 deadline_latter_request(struct request *rq)
79 struct rb_node *node = rb_next(&rq->rb_node);
82 return rb_entry_rq(node);
88 deadline_add_rq_rb(struct deadline_data *dd, struct request *rq)
90 struct rb_root *root = deadline_rb_root(dd, rq);
96 deadline_del_rq_rb(struct deadline_data *dd, struct request *rq)
98 const int data_dir = rq_data_dir(rq);
100 if (dd->next_rq[data_dir] == rq)
101 dd->next_rq[data_dir] = deadline_latter_request(rq);
103 elv_rb_del(deadline_rb_root(dd, rq), rq);
107 * remove rq from rbtree and fifo.
109 static void deadline_remove_request(struct request_queue *q, struct request *rq)
111 struct deadline_data *dd = q->elevator->elevator_data;
113 list_del_init(&rq->queuelist);
116 * We might not be on the rbtree, if we are doing an insert merge
118 if (!RB_EMPTY_NODE(&rq->rb_node))
119 deadline_del_rq_rb(dd, rq);
121 elv_rqhash_del(q, rq);
122 if (q->last_merge == rq)
123 q->last_merge = NULL;
126 static void dd_request_merged(struct request_queue *q, struct request *req,
129 struct deadline_data *dd = q->elevator->elevator_data;
132 * if the merge was a front merge, we need to reposition request
134 if (type == ELEVATOR_FRONT_MERGE) {
135 elv_rb_del(deadline_rb_root(dd, req), req);
136 deadline_add_rq_rb(dd, req);
140 static void dd_merged_requests(struct request_queue *q, struct request *req,
141 struct request *next)
144 * if next expires before rq, assign its expire time to rq
145 * and move into next position (next will be deleted) in fifo
147 if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
148 if (time_before((unsigned long)next->fifo_time,
149 (unsigned long)req->fifo_time)) {
150 list_move(&req->queuelist, &next->queuelist);
151 req->fifo_time = next->fifo_time;
156 * kill knowledge of next, this one is a goner
158 deadline_remove_request(q, next);
162 * move an entry to dispatch queue
165 deadline_move_request(struct deadline_data *dd, struct request *rq)
167 const int data_dir = rq_data_dir(rq);
169 dd->next_rq[READ] = NULL;
170 dd->next_rq[WRITE] = NULL;
171 dd->next_rq[data_dir] = deadline_latter_request(rq);
174 * take it off the sort and fifo list
176 deadline_remove_request(rq->q, rq);
180 * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
181 * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
183 static inline int deadline_check_fifo(struct deadline_data *dd, int ddir)
185 struct request *rq = rq_entry_fifo(dd->fifo_list[ddir].next);
190 if (time_after_eq(jiffies, (unsigned long)rq->fifo_time))
197 * For the specified data direction, return the next request to
198 * dispatch using arrival ordered lists.
200 static struct request *
201 deadline_fifo_request(struct deadline_data *dd, int data_dir)
206 if (WARN_ON_ONCE(data_dir != READ && data_dir != WRITE))
209 if (list_empty(&dd->fifo_list[data_dir]))
212 rq = rq_entry_fifo(dd->fifo_list[data_dir].next);
213 if (data_dir == READ || !blk_queue_is_zoned(rq->q))
217 * Look for a write request that can be dispatched, that is one with
218 * an unlocked target zone.
220 spin_lock_irqsave(&dd->zone_lock, flags);
221 list_for_each_entry(rq, &dd->fifo_list[WRITE], queuelist) {
222 if (blk_req_can_dispatch_to_zone(rq))
227 spin_unlock_irqrestore(&dd->zone_lock, flags);
233 * For the specified data direction, return the next request to
234 * dispatch using sector position sorted lists.
236 static struct request *
237 deadline_next_request(struct deadline_data *dd, int data_dir)
242 if (WARN_ON_ONCE(data_dir != READ && data_dir != WRITE))
245 rq = dd->next_rq[data_dir];
249 if (data_dir == READ || !blk_queue_is_zoned(rq->q))
253 * Look for a write request that can be dispatched, that is one with
254 * an unlocked target zone.
256 spin_lock_irqsave(&dd->zone_lock, flags);
258 if (blk_req_can_dispatch_to_zone(rq))
260 rq = deadline_latter_request(rq);
262 spin_unlock_irqrestore(&dd->zone_lock, flags);
268 * deadline_dispatch_requests selects the best request according to
269 * read/write expire, fifo_batch, etc
271 static struct request *__dd_dispatch_request(struct deadline_data *dd)
273 struct request *rq, *next_rq;
277 if (!list_empty(&dd->dispatch)) {
278 rq = list_first_entry(&dd->dispatch, struct request, queuelist);
279 list_del_init(&rq->queuelist);
283 reads = !list_empty(&dd->fifo_list[READ]);
284 writes = !list_empty(&dd->fifo_list[WRITE]);
287 * batches are currently reads XOR writes
289 rq = deadline_next_request(dd, WRITE);
291 rq = deadline_next_request(dd, READ);
293 if (rq && dd->batching < dd->fifo_batch)
294 /* we have a next request are still entitled to batch */
295 goto dispatch_request;
298 * at this point we are not running a batch. select the appropriate
299 * data direction (read / write)
303 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[READ]));
305 if (deadline_fifo_request(dd, WRITE) &&
306 (dd->starved++ >= dd->writes_starved))
307 goto dispatch_writes;
311 goto dispatch_find_request;
315 * there are either no reads or writes have been starved
320 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[WRITE]));
326 goto dispatch_find_request;
331 dispatch_find_request:
333 * we are not running a batch, find best request for selected data_dir
335 next_rq = deadline_next_request(dd, data_dir);
336 if (deadline_check_fifo(dd, data_dir) || !next_rq) {
338 * A deadline has expired, the last request was in the other
339 * direction, or we have run out of higher-sectored requests.
340 * Start again from the request with the earliest expiry time.
342 rq = deadline_fifo_request(dd, data_dir);
345 * The last req was the same dir and we have a next request in
346 * sort order. No expired requests so continue on from here.
352 * For a zoned block device, if we only have writes queued and none of
353 * them can be dispatched, rq will be NULL.
362 * rq is the selected appropriate request.
365 deadline_move_request(dd, rq);
368 * If the request needs its target zone locked, do it.
370 blk_req_zone_write_lock(rq);
371 rq->rq_flags |= RQF_STARTED;
376 * One confusing aspect here is that we get called for a specific
377 * hardware queue, but we may return a request that is for a
378 * different hardware queue. This is because mq-deadline has shared
379 * state for all hardware queues, in terms of sorting, FIFOs, etc.
381 * For a zoned block device, __dd_dispatch_request() may return NULL
382 * if all the queued write requests are directed at zones that are already
383 * locked due to on-going write requests. In this case, make sure to mark
384 * the queue as needing a restart to ensure that the queue is run again
385 * and the pending writes dispatched once the target zones for the ongoing
386 * write requests are unlocked in dd_finish_request().
388 static struct request *dd_dispatch_request(struct blk_mq_hw_ctx *hctx)
390 struct deadline_data *dd = hctx->queue->elevator->elevator_data;
393 spin_lock(&dd->lock);
394 rq = __dd_dispatch_request(dd);
395 if (!rq && blk_queue_is_zoned(hctx->queue) &&
396 !list_empty(&dd->fifo_list[WRITE]))
397 blk_mq_sched_mark_restart_hctx(hctx);
398 spin_unlock(&dd->lock);
403 static void dd_exit_queue(struct elevator_queue *e)
405 struct deadline_data *dd = e->elevator_data;
407 BUG_ON(!list_empty(&dd->fifo_list[READ]));
408 BUG_ON(!list_empty(&dd->fifo_list[WRITE]));
414 * initialize elevator private data (deadline_data).
416 static int dd_init_queue(struct request_queue *q, struct elevator_type *e)
418 struct deadline_data *dd;
419 struct elevator_queue *eq;
421 eq = elevator_alloc(q, e);
425 dd = kzalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
427 kobject_put(&eq->kobj);
430 eq->elevator_data = dd;
432 INIT_LIST_HEAD(&dd->fifo_list[READ]);
433 INIT_LIST_HEAD(&dd->fifo_list[WRITE]);
434 dd->sort_list[READ] = RB_ROOT;
435 dd->sort_list[WRITE] = RB_ROOT;
436 dd->fifo_expire[READ] = read_expire;
437 dd->fifo_expire[WRITE] = write_expire;
438 dd->writes_starved = writes_starved;
439 dd->front_merges = 1;
440 dd->fifo_batch = fifo_batch;
441 spin_lock_init(&dd->lock);
442 spin_lock_init(&dd->zone_lock);
443 INIT_LIST_HEAD(&dd->dispatch);
449 static int dd_request_merge(struct request_queue *q, struct request **rq,
452 struct deadline_data *dd = q->elevator->elevator_data;
453 sector_t sector = bio_end_sector(bio);
454 struct request *__rq;
456 if (!dd->front_merges)
457 return ELEVATOR_NO_MERGE;
459 __rq = elv_rb_find(&dd->sort_list[bio_data_dir(bio)], sector);
461 BUG_ON(sector != blk_rq_pos(__rq));
463 if (elv_bio_merge_ok(__rq, bio)) {
465 return ELEVATOR_FRONT_MERGE;
469 return ELEVATOR_NO_MERGE;
472 static bool dd_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio,
473 unsigned int nr_segs)
475 struct request_queue *q = hctx->queue;
476 struct deadline_data *dd = q->elevator->elevator_data;
477 struct request *free = NULL;
480 spin_lock(&dd->lock);
481 ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
482 spin_unlock(&dd->lock);
485 blk_mq_free_request(free);
491 * add rq to rbtree and fifo
493 static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
496 struct request_queue *q = hctx->queue;
497 struct deadline_data *dd = q->elevator->elevator_data;
498 const int data_dir = rq_data_dir(rq);
501 * This may be a requeue of a write request that has locked its
502 * target zone. If it is the case, this releases the zone lock.
504 blk_req_zone_write_unlock(rq);
506 if (blk_mq_sched_try_insert_merge(q, rq))
509 blk_mq_sched_request_inserted(rq);
511 if (at_head || blk_rq_is_passthrough(rq)) {
513 list_add(&rq->queuelist, &dd->dispatch);
515 list_add_tail(&rq->queuelist, &dd->dispatch);
517 deadline_add_rq_rb(dd, rq);
519 if (rq_mergeable(rq)) {
520 elv_rqhash_add(q, rq);
526 * set expire time and add to fifo list
528 rq->fifo_time = jiffies + dd->fifo_expire[data_dir];
529 list_add_tail(&rq->queuelist, &dd->fifo_list[data_dir]);
533 static void dd_insert_requests(struct blk_mq_hw_ctx *hctx,
534 struct list_head *list, bool at_head)
536 struct request_queue *q = hctx->queue;
537 struct deadline_data *dd = q->elevator->elevator_data;
539 spin_lock(&dd->lock);
540 while (!list_empty(list)) {
543 rq = list_first_entry(list, struct request, queuelist);
544 list_del_init(&rq->queuelist);
545 dd_insert_request(hctx, rq, at_head);
547 spin_unlock(&dd->lock);
551 * Nothing to do here. This is defined only to ensure that .finish_request
552 * method is called upon request completion.
554 static void dd_prepare_request(struct request *rq, struct bio *bio)
559 * For zoned block devices, write unlock the target zone of
560 * completed write requests. Do this while holding the zone lock
561 * spinlock so that the zone is never unlocked while deadline_fifo_request()
562 * or deadline_next_request() are executing. This function is called for
563 * all requests, whether or not these requests complete successfully.
565 static void dd_finish_request(struct request *rq)
567 struct request_queue *q = rq->q;
569 if (blk_queue_is_zoned(q)) {
570 struct deadline_data *dd = q->elevator->elevator_data;
573 spin_lock_irqsave(&dd->zone_lock, flags);
574 blk_req_zone_write_unlock(rq);
575 spin_unlock_irqrestore(&dd->zone_lock, flags);
579 static bool dd_has_work(struct blk_mq_hw_ctx *hctx)
581 struct deadline_data *dd = hctx->queue->elevator->elevator_data;
583 return !list_empty_careful(&dd->dispatch) ||
584 !list_empty_careful(&dd->fifo_list[0]) ||
585 !list_empty_careful(&dd->fifo_list[1]);
592 deadline_var_show(int var, char *page)
594 return sprintf(page, "%d\n", var);
598 deadline_var_store(int *var, const char *page)
600 char *p = (char *) page;
602 *var = simple_strtol(p, &p, 10);
605 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
606 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
608 struct deadline_data *dd = e->elevator_data; \
609 int __data = __VAR; \
611 __data = jiffies_to_msecs(__data); \
612 return deadline_var_show(__data, (page)); \
614 SHOW_FUNCTION(deadline_read_expire_show, dd->fifo_expire[READ], 1);
615 SHOW_FUNCTION(deadline_write_expire_show, dd->fifo_expire[WRITE], 1);
616 SHOW_FUNCTION(deadline_writes_starved_show, dd->writes_starved, 0);
617 SHOW_FUNCTION(deadline_front_merges_show, dd->front_merges, 0);
618 SHOW_FUNCTION(deadline_fifo_batch_show, dd->fifo_batch, 0);
621 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
622 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
624 struct deadline_data *dd = e->elevator_data; \
626 deadline_var_store(&__data, (page)); \
627 if (__data < (MIN)) \
629 else if (__data > (MAX)) \
632 *(__PTR) = msecs_to_jiffies(__data); \
637 STORE_FUNCTION(deadline_read_expire_store, &dd->fifo_expire[READ], 0, INT_MAX, 1);
638 STORE_FUNCTION(deadline_write_expire_store, &dd->fifo_expire[WRITE], 0, INT_MAX, 1);
639 STORE_FUNCTION(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX, 0);
640 STORE_FUNCTION(deadline_front_merges_store, &dd->front_merges, 0, 1, 0);
641 STORE_FUNCTION(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX, 0);
642 #undef STORE_FUNCTION
644 #define DD_ATTR(name) \
645 __ATTR(name, 0644, deadline_##name##_show, deadline_##name##_store)
647 static struct elv_fs_entry deadline_attrs[] = {
648 DD_ATTR(read_expire),
649 DD_ATTR(write_expire),
650 DD_ATTR(writes_starved),
651 DD_ATTR(front_merges),
656 #ifdef CONFIG_BLK_DEBUG_FS
657 #define DEADLINE_DEBUGFS_DDIR_ATTRS(ddir, name) \
658 static void *deadline_##name##_fifo_start(struct seq_file *m, \
660 __acquires(&dd->lock) \
662 struct request_queue *q = m->private; \
663 struct deadline_data *dd = q->elevator->elevator_data; \
665 spin_lock(&dd->lock); \
666 return seq_list_start(&dd->fifo_list[ddir], *pos); \
669 static void *deadline_##name##_fifo_next(struct seq_file *m, void *v, \
672 struct request_queue *q = m->private; \
673 struct deadline_data *dd = q->elevator->elevator_data; \
675 return seq_list_next(v, &dd->fifo_list[ddir], pos); \
678 static void deadline_##name##_fifo_stop(struct seq_file *m, void *v) \
679 __releases(&dd->lock) \
681 struct request_queue *q = m->private; \
682 struct deadline_data *dd = q->elevator->elevator_data; \
684 spin_unlock(&dd->lock); \
687 static const struct seq_operations deadline_##name##_fifo_seq_ops = { \
688 .start = deadline_##name##_fifo_start, \
689 .next = deadline_##name##_fifo_next, \
690 .stop = deadline_##name##_fifo_stop, \
691 .show = blk_mq_debugfs_rq_show, \
694 static int deadline_##name##_next_rq_show(void *data, \
695 struct seq_file *m) \
697 struct request_queue *q = data; \
698 struct deadline_data *dd = q->elevator->elevator_data; \
699 struct request *rq = dd->next_rq[ddir]; \
702 __blk_mq_debugfs_rq_show(m, rq); \
705 DEADLINE_DEBUGFS_DDIR_ATTRS(READ, read)
706 DEADLINE_DEBUGFS_DDIR_ATTRS(WRITE, write)
707 #undef DEADLINE_DEBUGFS_DDIR_ATTRS
709 static int deadline_batching_show(void *data, struct seq_file *m)
711 struct request_queue *q = data;
712 struct deadline_data *dd = q->elevator->elevator_data;
714 seq_printf(m, "%u\n", dd->batching);
718 static int deadline_starved_show(void *data, struct seq_file *m)
720 struct request_queue *q = data;
721 struct deadline_data *dd = q->elevator->elevator_data;
723 seq_printf(m, "%u\n", dd->starved);
727 static void *deadline_dispatch_start(struct seq_file *m, loff_t *pos)
728 __acquires(&dd->lock)
730 struct request_queue *q = m->private;
731 struct deadline_data *dd = q->elevator->elevator_data;
733 spin_lock(&dd->lock);
734 return seq_list_start(&dd->dispatch, *pos);
737 static void *deadline_dispatch_next(struct seq_file *m, void *v, loff_t *pos)
739 struct request_queue *q = m->private;
740 struct deadline_data *dd = q->elevator->elevator_data;
742 return seq_list_next(v, &dd->dispatch, pos);
745 static void deadline_dispatch_stop(struct seq_file *m, void *v)
746 __releases(&dd->lock)
748 struct request_queue *q = m->private;
749 struct deadline_data *dd = q->elevator->elevator_data;
751 spin_unlock(&dd->lock);
754 static const struct seq_operations deadline_dispatch_seq_ops = {
755 .start = deadline_dispatch_start,
756 .next = deadline_dispatch_next,
757 .stop = deadline_dispatch_stop,
758 .show = blk_mq_debugfs_rq_show,
761 #define DEADLINE_QUEUE_DDIR_ATTRS(name) \
762 {#name "_fifo_list", 0400, .seq_ops = &deadline_##name##_fifo_seq_ops}, \
763 {#name "_next_rq", 0400, deadline_##name##_next_rq_show}
764 static const struct blk_mq_debugfs_attr deadline_queue_debugfs_attrs[] = {
765 DEADLINE_QUEUE_DDIR_ATTRS(read),
766 DEADLINE_QUEUE_DDIR_ATTRS(write),
767 {"batching", 0400, deadline_batching_show},
768 {"starved", 0400, deadline_starved_show},
769 {"dispatch", 0400, .seq_ops = &deadline_dispatch_seq_ops},
772 #undef DEADLINE_QUEUE_DDIR_ATTRS
775 static struct elevator_type mq_deadline = {
777 .insert_requests = dd_insert_requests,
778 .dispatch_request = dd_dispatch_request,
779 .prepare_request = dd_prepare_request,
780 .finish_request = dd_finish_request,
781 .next_request = elv_rb_latter_request,
782 .former_request = elv_rb_former_request,
783 .bio_merge = dd_bio_merge,
784 .request_merge = dd_request_merge,
785 .requests_merged = dd_merged_requests,
786 .request_merged = dd_request_merged,
787 .has_work = dd_has_work,
788 .init_sched = dd_init_queue,
789 .exit_sched = dd_exit_queue,
792 #ifdef CONFIG_BLK_DEBUG_FS
793 .queue_debugfs_attrs = deadline_queue_debugfs_attrs,
795 .elevator_attrs = deadline_attrs,
796 .elevator_name = "mq-deadline",
797 .elevator_alias = "deadline",
798 .elevator_owner = THIS_MODULE,
800 MODULE_ALIAS("mq-deadline-iosched");
802 static int __init deadline_init(void)
804 return elv_register(&mq_deadline);
807 static void __exit deadline_exit(void)
809 elv_unregister(&mq_deadline);
812 module_init(deadline_init);
813 module_exit(deadline_exit);
815 MODULE_AUTHOR("Jens Axboe");
816 MODULE_LICENSE("GPL");
817 MODULE_DESCRIPTION("MQ deadline IO scheduler");