2 * MQ Deadline i/o scheduler - adaptation of the legacy deadline scheduler,
3 * for the blk-mq scheduling framework
5 * Copyright (C) 2016 Jens Axboe <axboe@kernel.dk>
7 #include <linux/kernel.h>
9 #include <linux/blkdev.h>
10 #include <linux/blk-mq.h>
11 #include <linux/elevator.h>
12 #include <linux/bio.h>
13 #include <linux/module.h>
14 #include <linux/slab.h>
15 #include <linux/init.h>
16 #include <linux/compiler.h>
17 #include <linux/rbtree.h>
18 #include <linux/sbitmap.h>
22 #include "blk-mq-debugfs.h"
23 #include "blk-mq-tag.h"
24 #include "blk-mq-sched.h"
27 * See Documentation/block/deadline-iosched.txt
29 static const int read_expire
= HZ
/ 2; /* max time before a read is submitted. */
30 static const int write_expire
= 5 * HZ
; /* ditto for writes, these limits are SOFT! */
31 static const int writes_starved
= 2; /* max times reads can starve a write */
32 static const int fifo_batch
= 16; /* # of sequential requests treated as one
33 by the above parameters. For throughput. */
35 struct deadline_data
{
41 * requests (deadline_rq s) are present on both sort_list and fifo_list
43 struct rb_root sort_list
[2];
44 struct list_head fifo_list
[2];
47 * next in sort order. read, write or both are NULL
49 struct request
*next_rq
[2];
50 unsigned int batching
; /* number of sequential requests made */
51 unsigned int starved
; /* times reads have starved writes */
54 * settings that change how the i/o scheduler behaves
62 struct list_head dispatch
;
65 static inline struct rb_root
*
66 deadline_rb_root(struct deadline_data
*dd
, struct request
*rq
)
68 return &dd
->sort_list
[rq_data_dir(rq
)];
72 * get the request after `rq' in sector-sorted order
74 static inline struct request
*
75 deadline_latter_request(struct request
*rq
)
77 struct rb_node
*node
= rb_next(&rq
->rb_node
);
80 return rb_entry_rq(node
);
86 deadline_add_rq_rb(struct deadline_data
*dd
, struct request
*rq
)
88 struct rb_root
*root
= deadline_rb_root(dd
, rq
);
94 deadline_del_rq_rb(struct deadline_data
*dd
, struct request
*rq
)
96 const int data_dir
= rq_data_dir(rq
);
98 if (dd
->next_rq
[data_dir
] == rq
)
99 dd
->next_rq
[data_dir
] = deadline_latter_request(rq
);
101 elv_rb_del(deadline_rb_root(dd
, rq
), rq
);
105 * remove rq from rbtree and fifo.
107 static void deadline_remove_request(struct request_queue
*q
, struct request
*rq
)
109 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
111 list_del_init(&rq
->queuelist
);
114 * We might not be on the rbtree, if we are doing an insert merge
116 if (!RB_EMPTY_NODE(&rq
->rb_node
))
117 deadline_del_rq_rb(dd
, rq
);
119 elv_rqhash_del(q
, rq
);
120 if (q
->last_merge
== rq
)
121 q
->last_merge
= NULL
;
124 static void dd_request_merged(struct request_queue
*q
, struct request
*req
,
127 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
130 * if the merge was a front merge, we need to reposition request
132 if (type
== ELEVATOR_FRONT_MERGE
) {
133 elv_rb_del(deadline_rb_root(dd
, req
), req
);
134 deadline_add_rq_rb(dd
, req
);
138 static void dd_merged_requests(struct request_queue
*q
, struct request
*req
,
139 struct request
*next
)
142 * if next expires before rq, assign its expire time to rq
143 * and move into next position (next will be deleted) in fifo
145 if (!list_empty(&req
->queuelist
) && !list_empty(&next
->queuelist
)) {
146 if (time_before((unsigned long)next
->fifo_time
,
147 (unsigned long)req
->fifo_time
)) {
148 list_move(&req
->queuelist
, &next
->queuelist
);
149 req
->fifo_time
= next
->fifo_time
;
154 * kill knowledge of next, this one is a goner
156 deadline_remove_request(q
, next
);
160 * move an entry to dispatch queue
163 deadline_move_request(struct deadline_data
*dd
, struct request
*rq
)
165 const int data_dir
= rq_data_dir(rq
);
167 dd
->next_rq
[READ
] = NULL
;
168 dd
->next_rq
[WRITE
] = NULL
;
169 dd
->next_rq
[data_dir
] = deadline_latter_request(rq
);
172 * take it off the sort and fifo list
174 deadline_remove_request(rq
->q
, rq
);
178 * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
179 * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
181 static inline int deadline_check_fifo(struct deadline_data
*dd
, int ddir
)
183 struct request
*rq
= rq_entry_fifo(dd
->fifo_list
[ddir
].next
);
188 if (time_after_eq(jiffies
, (unsigned long)rq
->fifo_time
))
195 * deadline_dispatch_requests selects the best request according to
196 * read/write expire, fifo_batch, etc
198 static struct request
*__dd_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
200 struct deadline_data
*dd
= hctx
->queue
->elevator
->elevator_data
;
205 if (!list_empty(&dd
->dispatch
)) {
206 rq
= list_first_entry(&dd
->dispatch
, struct request
, queuelist
);
207 list_del_init(&rq
->queuelist
);
211 reads
= !list_empty(&dd
->fifo_list
[READ
]);
212 writes
= !list_empty(&dd
->fifo_list
[WRITE
]);
215 * batches are currently reads XOR writes
217 if (dd
->next_rq
[WRITE
])
218 rq
= dd
->next_rq
[WRITE
];
220 rq
= dd
->next_rq
[READ
];
222 if (rq
&& dd
->batching
< dd
->fifo_batch
)
223 /* we have a next request are still entitled to batch */
224 goto dispatch_request
;
227 * at this point we are not running a batch. select the appropriate
228 * data direction (read / write)
232 BUG_ON(RB_EMPTY_ROOT(&dd
->sort_list
[READ
]));
234 if (writes
&& (dd
->starved
++ >= dd
->writes_starved
))
235 goto dispatch_writes
;
239 goto dispatch_find_request
;
243 * there are either no reads or writes have been starved
248 BUG_ON(RB_EMPTY_ROOT(&dd
->sort_list
[WRITE
]));
254 goto dispatch_find_request
;
259 dispatch_find_request
:
261 * we are not running a batch, find best request for selected data_dir
263 if (deadline_check_fifo(dd
, data_dir
) || !dd
->next_rq
[data_dir
]) {
265 * A deadline has expired, the last request was in the other
266 * direction, or we have run out of higher-sectored requests.
267 * Start again from the request with the earliest expiry time.
269 rq
= rq_entry_fifo(dd
->fifo_list
[data_dir
].next
);
272 * The last req was the same dir and we have a next request in
273 * sort order. No expired requests so continue on from here.
275 rq
= dd
->next_rq
[data_dir
];
282 * rq is the selected appropriate request.
285 deadline_move_request(dd
, rq
);
287 rq
->rq_flags
|= RQF_STARTED
;
291 static struct request
*dd_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
293 struct deadline_data
*dd
= hctx
->queue
->elevator
->elevator_data
;
296 spin_lock(&dd
->lock
);
297 rq
= __dd_dispatch_request(hctx
);
298 spin_unlock(&dd
->lock
);
303 static void dd_exit_queue(struct elevator_queue
*e
)
305 struct deadline_data
*dd
= e
->elevator_data
;
307 BUG_ON(!list_empty(&dd
->fifo_list
[READ
]));
308 BUG_ON(!list_empty(&dd
->fifo_list
[WRITE
]));
314 * initialize elevator private data (deadline_data).
316 static int dd_init_queue(struct request_queue
*q
, struct elevator_type
*e
)
318 struct deadline_data
*dd
;
319 struct elevator_queue
*eq
;
321 eq
= elevator_alloc(q
, e
);
325 dd
= kzalloc_node(sizeof(*dd
), GFP_KERNEL
, q
->node
);
327 kobject_put(&eq
->kobj
);
330 eq
->elevator_data
= dd
;
332 INIT_LIST_HEAD(&dd
->fifo_list
[READ
]);
333 INIT_LIST_HEAD(&dd
->fifo_list
[WRITE
]);
334 dd
->sort_list
[READ
] = RB_ROOT
;
335 dd
->sort_list
[WRITE
] = RB_ROOT
;
336 dd
->fifo_expire
[READ
] = read_expire
;
337 dd
->fifo_expire
[WRITE
] = write_expire
;
338 dd
->writes_starved
= writes_starved
;
339 dd
->front_merges
= 1;
340 dd
->fifo_batch
= fifo_batch
;
341 spin_lock_init(&dd
->lock
);
342 INIT_LIST_HEAD(&dd
->dispatch
);
348 static int dd_request_merge(struct request_queue
*q
, struct request
**rq
,
351 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
352 sector_t sector
= bio_end_sector(bio
);
353 struct request
*__rq
;
355 if (!dd
->front_merges
)
356 return ELEVATOR_NO_MERGE
;
358 __rq
= elv_rb_find(&dd
->sort_list
[bio_data_dir(bio
)], sector
);
360 BUG_ON(sector
!= blk_rq_pos(__rq
));
362 if (elv_bio_merge_ok(__rq
, bio
)) {
364 return ELEVATOR_FRONT_MERGE
;
368 return ELEVATOR_NO_MERGE
;
371 static bool dd_bio_merge(struct blk_mq_hw_ctx
*hctx
, struct bio
*bio
)
373 struct request_queue
*q
= hctx
->queue
;
374 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
375 struct request
*free
= NULL
;
378 spin_lock(&dd
->lock
);
379 ret
= blk_mq_sched_try_merge(q
, bio
, &free
);
380 spin_unlock(&dd
->lock
);
383 blk_mq_free_request(free
);
389 * add rq to rbtree and fifo
391 static void dd_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
394 struct request_queue
*q
= hctx
->queue
;
395 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
396 const int data_dir
= rq_data_dir(rq
);
398 if (blk_mq_sched_try_insert_merge(q
, rq
))
401 blk_mq_sched_request_inserted(rq
);
403 if (at_head
|| blk_rq_is_passthrough(rq
)) {
405 list_add(&rq
->queuelist
, &dd
->dispatch
);
407 list_add_tail(&rq
->queuelist
, &dd
->dispatch
);
409 deadline_add_rq_rb(dd
, rq
);
411 if (rq_mergeable(rq
)) {
412 elv_rqhash_add(q
, rq
);
418 * set expire time and add to fifo list
420 rq
->fifo_time
= jiffies
+ dd
->fifo_expire
[data_dir
];
421 list_add_tail(&rq
->queuelist
, &dd
->fifo_list
[data_dir
]);
425 static void dd_insert_requests(struct blk_mq_hw_ctx
*hctx
,
426 struct list_head
*list
, bool at_head
)
428 struct request_queue
*q
= hctx
->queue
;
429 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
431 spin_lock(&dd
->lock
);
432 while (!list_empty(list
)) {
435 rq
= list_first_entry(list
, struct request
, queuelist
);
436 list_del_init(&rq
->queuelist
);
437 dd_insert_request(hctx
, rq
, at_head
);
439 spin_unlock(&dd
->lock
);
442 static bool dd_has_work(struct blk_mq_hw_ctx
*hctx
)
444 struct deadline_data
*dd
= hctx
->queue
->elevator
->elevator_data
;
446 return !list_empty_careful(&dd
->dispatch
) ||
447 !list_empty_careful(&dd
->fifo_list
[0]) ||
448 !list_empty_careful(&dd
->fifo_list
[1]);
455 deadline_var_show(int var
, char *page
)
457 return sprintf(page
, "%d\n", var
);
461 deadline_var_store(int *var
, const char *page
)
463 char *p
= (char *) page
;
465 *var
= simple_strtol(p
, &p
, 10);
468 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
469 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
471 struct deadline_data *dd = e->elevator_data; \
472 int __data = __VAR; \
474 __data = jiffies_to_msecs(__data); \
475 return deadline_var_show(__data, (page)); \
477 SHOW_FUNCTION(deadline_read_expire_show
, dd
->fifo_expire
[READ
], 1);
478 SHOW_FUNCTION(deadline_write_expire_show
, dd
->fifo_expire
[WRITE
], 1);
479 SHOW_FUNCTION(deadline_writes_starved_show
, dd
->writes_starved
, 0);
480 SHOW_FUNCTION(deadline_front_merges_show
, dd
->front_merges
, 0);
481 SHOW_FUNCTION(deadline_fifo_batch_show
, dd
->fifo_batch
, 0);
484 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
485 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
487 struct deadline_data *dd = e->elevator_data; \
489 deadline_var_store(&__data, (page)); \
490 if (__data < (MIN)) \
492 else if (__data > (MAX)) \
495 *(__PTR) = msecs_to_jiffies(__data); \
500 STORE_FUNCTION(deadline_read_expire_store
, &dd
->fifo_expire
[READ
], 0, INT_MAX
, 1);
501 STORE_FUNCTION(deadline_write_expire_store
, &dd
->fifo_expire
[WRITE
], 0, INT_MAX
, 1);
502 STORE_FUNCTION(deadline_writes_starved_store
, &dd
->writes_starved
, INT_MIN
, INT_MAX
, 0);
503 STORE_FUNCTION(deadline_front_merges_store
, &dd
->front_merges
, 0, 1, 0);
504 STORE_FUNCTION(deadline_fifo_batch_store
, &dd
->fifo_batch
, 0, INT_MAX
, 0);
505 #undef STORE_FUNCTION
507 #define DD_ATTR(name) \
508 __ATTR(name, S_IRUGO|S_IWUSR, deadline_##name##_show, \
509 deadline_##name##_store)
511 static struct elv_fs_entry deadline_attrs
[] = {
512 DD_ATTR(read_expire
),
513 DD_ATTR(write_expire
),
514 DD_ATTR(writes_starved
),
515 DD_ATTR(front_merges
),
520 #ifdef CONFIG_BLK_DEBUG_FS
521 #define DEADLINE_DEBUGFS_DDIR_ATTRS(ddir, name) \
522 static void *deadline_##name##_fifo_start(struct seq_file *m, \
524 __acquires(&dd->lock) \
526 struct request_queue *q = m->private; \
527 struct deadline_data *dd = q->elevator->elevator_data; \
529 spin_lock(&dd->lock); \
530 return seq_list_start(&dd->fifo_list[ddir], *pos); \
533 static void *deadline_##name##_fifo_next(struct seq_file *m, void *v, \
536 struct request_queue *q = m->private; \
537 struct deadline_data *dd = q->elevator->elevator_data; \
539 return seq_list_next(v, &dd->fifo_list[ddir], pos); \
542 static void deadline_##name##_fifo_stop(struct seq_file *m, void *v) \
543 __releases(&dd->lock) \
545 struct request_queue *q = m->private; \
546 struct deadline_data *dd = q->elevator->elevator_data; \
548 spin_unlock(&dd->lock); \
551 static const struct seq_operations deadline_##name##_fifo_seq_ops = { \
552 .start = deadline_##name##_fifo_start, \
553 .next = deadline_##name##_fifo_next, \
554 .stop = deadline_##name##_fifo_stop, \
555 .show = blk_mq_debugfs_rq_show, \
558 static int deadline_##name##_next_rq_show(void *data, \
559 struct seq_file *m) \
561 struct request_queue *q = data; \
562 struct deadline_data *dd = q->elevator->elevator_data; \
563 struct request *rq = dd->next_rq[ddir]; \
566 __blk_mq_debugfs_rq_show(m, rq); \
569 DEADLINE_DEBUGFS_DDIR_ATTRS(READ
, read
)
570 DEADLINE_DEBUGFS_DDIR_ATTRS(WRITE
, write
)
571 #undef DEADLINE_DEBUGFS_DDIR_ATTRS
573 static int deadline_batching_show(void *data
, struct seq_file
*m
)
575 struct request_queue
*q
= data
;
576 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
578 seq_printf(m
, "%u\n", dd
->batching
);
582 static int deadline_starved_show(void *data
, struct seq_file
*m
)
584 struct request_queue
*q
= data
;
585 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
587 seq_printf(m
, "%u\n", dd
->starved
);
591 static void *deadline_dispatch_start(struct seq_file
*m
, loff_t
*pos
)
592 __acquires(&dd
->lock
)
594 struct request_queue
*q
= m
->private;
595 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
597 spin_lock(&dd
->lock
);
598 return seq_list_start(&dd
->dispatch
, *pos
);
601 static void *deadline_dispatch_next(struct seq_file
*m
, void *v
, loff_t
*pos
)
603 struct request_queue
*q
= m
->private;
604 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
606 return seq_list_next(v
, &dd
->dispatch
, pos
);
609 static void deadline_dispatch_stop(struct seq_file
*m
, void *v
)
610 __releases(&dd
->lock
)
612 struct request_queue
*q
= m
->private;
613 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
615 spin_unlock(&dd
->lock
);
618 static const struct seq_operations deadline_dispatch_seq_ops
= {
619 .start
= deadline_dispatch_start
,
620 .next
= deadline_dispatch_next
,
621 .stop
= deadline_dispatch_stop
,
622 .show
= blk_mq_debugfs_rq_show
,
625 #define DEADLINE_QUEUE_DDIR_ATTRS(name) \
626 {#name "_fifo_list", 0400, .seq_ops = &deadline_##name##_fifo_seq_ops}, \
627 {#name "_next_rq", 0400, deadline_##name##_next_rq_show}
628 static const struct blk_mq_debugfs_attr deadline_queue_debugfs_attrs
[] = {
629 DEADLINE_QUEUE_DDIR_ATTRS(read
),
630 DEADLINE_QUEUE_DDIR_ATTRS(write
),
631 {"batching", 0400, deadline_batching_show
},
632 {"starved", 0400, deadline_starved_show
},
633 {"dispatch", 0400, .seq_ops
= &deadline_dispatch_seq_ops
},
636 #undef DEADLINE_QUEUE_DDIR_ATTRS
639 static struct elevator_type mq_deadline
= {
641 .insert_requests
= dd_insert_requests
,
642 .dispatch_request
= dd_dispatch_request
,
643 .next_request
= elv_rb_latter_request
,
644 .former_request
= elv_rb_former_request
,
645 .bio_merge
= dd_bio_merge
,
646 .request_merge
= dd_request_merge
,
647 .requests_merged
= dd_merged_requests
,
648 .request_merged
= dd_request_merged
,
649 .has_work
= dd_has_work
,
650 .init_sched
= dd_init_queue
,
651 .exit_sched
= dd_exit_queue
,
655 #ifdef CONFIG_BLK_DEBUG_FS
656 .queue_debugfs_attrs
= deadline_queue_debugfs_attrs
,
658 .elevator_attrs
= deadline_attrs
,
659 .elevator_name
= "mq-deadline",
660 .elevator_owner
= THIS_MODULE
,
662 MODULE_ALIAS("mq-deadline-iosched");
664 static int __init
deadline_init(void)
666 return elv_register(&mq_deadline
);
669 static void __exit
deadline_exit(void)
671 elv_unregister(&mq_deadline
);
674 module_init(deadline_init
);
675 module_exit(deadline_exit
);
677 MODULE_AUTHOR("Jens Axboe");
678 MODULE_LICENSE("GPL");
679 MODULE_DESCRIPTION("MQ deadline IO scheduler");