2 * Deadline i/o scheduler.
4 * Copyright (C) 2002 Jens Axboe <axboe@suse.de>
6 #include <linux/kernel.h>
8 #include <linux/blkdev.h>
9 #include <linux/elevator.h>
10 #include <linux/bio.h>
11 #include <linux/config.h>
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/init.h>
15 #include <linux/compiler.h>
16 #include <linux/hash.h>
17 #include <linux/rbtree.h>
20 * See Documentation/block/deadline-iosched.txt
22 static const int read_expire
= HZ
/ 2; /* max time before a read is submitted. */
23 static const int write_expire
= 5 * HZ
; /* ditto for writes, these limits are SOFT! */
24 static const int writes_starved
= 2; /* max times reads can starve a write */
25 static const int fifo_batch
= 16; /* # of sequential requests treated as one
26 by the above parameters. For throughput. */
28 static const int deadline_hash_shift
= 5;
29 #define DL_HASH_BLOCK(sec) ((sec) >> 3)
30 #define DL_HASH_FN(sec) (hash_long(DL_HASH_BLOCK((sec)), deadline_hash_shift))
31 #define DL_HASH_ENTRIES (1 << deadline_hash_shift)
32 #define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
33 #define list_entry_hash(ptr) list_entry((ptr), struct deadline_rq, hash)
34 #define ON_HASH(drq) (drq)->on_hash
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 deadline_rq
*next_drq
[2];
51 struct list_head
*hash
; /* request hash */
52 unsigned int batching
; /* number of sequential requests made */
53 sector_t last_sector
; /* head position */
54 unsigned int starved
; /* times reads have starved writes */
57 * settings that change how the i/o scheduler behaves
72 * rbtree index, key is the starting offset
74 struct rb_node rb_node
;
77 struct request
*request
;
80 * request hash, key is the ending offset (for back merge lookup)
82 struct list_head hash
;
88 struct list_head fifo
;
89 unsigned long expires
;
92 static void deadline_move_request(struct deadline_data
*dd
, struct deadline_rq
*drq
);
94 static kmem_cache_t
*drq_pool
;
96 #define RQ_DATA(rq) ((struct deadline_rq *) (rq)->elevator_private)
99 * the back merge hash support functions
101 static inline void __deadline_del_drq_hash(struct deadline_rq
*drq
)
104 list_del_init(&drq
->hash
);
107 static inline void deadline_del_drq_hash(struct deadline_rq
*drq
)
110 __deadline_del_drq_hash(drq
);
114 deadline_add_drq_hash(struct deadline_data
*dd
, struct deadline_rq
*drq
)
116 struct request
*rq
= drq
->request
;
118 BUG_ON(ON_HASH(drq
));
121 list_add(&drq
->hash
, &dd
->hash
[DL_HASH_FN(rq_hash_key(rq
))]);
125 * move hot entry to front of chain
128 deadline_hot_drq_hash(struct deadline_data
*dd
, struct deadline_rq
*drq
)
130 struct request
*rq
= drq
->request
;
131 struct list_head
*head
= &dd
->hash
[DL_HASH_FN(rq_hash_key(rq
))];
133 if (ON_HASH(drq
) && drq
->hash
.prev
!= head
) {
134 list_del(&drq
->hash
);
135 list_add(&drq
->hash
, head
);
139 static struct request
*
140 deadline_find_drq_hash(struct deadline_data
*dd
, sector_t offset
)
142 struct list_head
*hash_list
= &dd
->hash
[DL_HASH_FN(offset
)];
143 struct list_head
*entry
, *next
= hash_list
->next
;
145 while ((entry
= next
) != hash_list
) {
146 struct deadline_rq
*drq
= list_entry_hash(entry
);
147 struct request
*__rq
= drq
->request
;
151 BUG_ON(!ON_HASH(drq
));
153 if (!rq_mergeable(__rq
)) {
154 __deadline_del_drq_hash(drq
);
158 if (rq_hash_key(__rq
) == offset
)
166 * rb tree support functions
169 #define RB_EMPTY(root) ((root)->rb_node == NULL)
170 #define ON_RB(node) ((node)->rb_color != RB_NONE)
171 #define RB_CLEAR(node) ((node)->rb_color = RB_NONE)
172 #define rb_entry_drq(node) rb_entry((node), struct deadline_rq, rb_node)
173 #define DRQ_RB_ROOT(dd, drq) (&(dd)->sort_list[rq_data_dir((drq)->request)])
174 #define rq_rb_key(rq) (rq)->sector
176 static struct deadline_rq
*
177 __deadline_add_drq_rb(struct deadline_data
*dd
, struct deadline_rq
*drq
)
179 struct rb_node
**p
= &DRQ_RB_ROOT(dd
, drq
)->rb_node
;
180 struct rb_node
*parent
= NULL
;
181 struct deadline_rq
*__drq
;
185 __drq
= rb_entry_drq(parent
);
187 if (drq
->rb_key
< __drq
->rb_key
)
189 else if (drq
->rb_key
> __drq
->rb_key
)
195 rb_link_node(&drq
->rb_node
, parent
, p
);
200 deadline_add_drq_rb(struct deadline_data
*dd
, struct deadline_rq
*drq
)
202 struct deadline_rq
*__alias
;
204 drq
->rb_key
= rq_rb_key(drq
->request
);
207 __alias
= __deadline_add_drq_rb(dd
, drq
);
209 rb_insert_color(&drq
->rb_node
, DRQ_RB_ROOT(dd
, drq
));
213 deadline_move_request(dd
, __alias
);
218 deadline_del_drq_rb(struct deadline_data
*dd
, struct deadline_rq
*drq
)
220 const int data_dir
= rq_data_dir(drq
->request
);
222 if (dd
->next_drq
[data_dir
] == drq
) {
223 struct rb_node
*rbnext
= rb_next(&drq
->rb_node
);
225 dd
->next_drq
[data_dir
] = NULL
;
227 dd
->next_drq
[data_dir
] = rb_entry_drq(rbnext
);
230 BUG_ON(!ON_RB(&drq
->rb_node
));
231 rb_erase(&drq
->rb_node
, DRQ_RB_ROOT(dd
, drq
));
232 RB_CLEAR(&drq
->rb_node
);
235 static struct request
*
236 deadline_find_drq_rb(struct deadline_data
*dd
, sector_t sector
, int data_dir
)
238 struct rb_node
*n
= dd
->sort_list
[data_dir
].rb_node
;
239 struct deadline_rq
*drq
;
242 drq
= rb_entry_drq(n
);
244 if (sector
< drq
->rb_key
)
246 else if (sector
> drq
->rb_key
)
256 * deadline_find_first_drq finds the first (lowest sector numbered) request
257 * for the specified data_dir. Used to sweep back to the start of the disk
258 * (1-way elevator) after we process the last (highest sector) request.
260 static struct deadline_rq
*
261 deadline_find_first_drq(struct deadline_data
*dd
, int data_dir
)
263 struct rb_node
*n
= dd
->sort_list
[data_dir
].rb_node
;
266 if (n
->rb_left
== NULL
)
267 return rb_entry_drq(n
);
274 * add drq to rbtree and fifo
277 deadline_add_request(struct request_queue
*q
, struct request
*rq
)
279 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
280 struct deadline_rq
*drq
= RQ_DATA(rq
);
282 const int data_dir
= rq_data_dir(drq
->request
);
284 deadline_add_drq_rb(dd
, drq
);
286 * set expire time (only used for reads) and add to fifo list
288 drq
->expires
= jiffies
+ dd
->fifo_expire
[data_dir
];
289 list_add_tail(&drq
->fifo
, &dd
->fifo_list
[data_dir
]);
291 if (rq_mergeable(rq
))
292 deadline_add_drq_hash(dd
, drq
);
296 * remove rq from rbtree, fifo, and hash
298 static void deadline_remove_request(request_queue_t
*q
, struct request
*rq
)
300 struct deadline_rq
*drq
= RQ_DATA(rq
);
301 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
303 list_del_init(&drq
->fifo
);
304 deadline_del_drq_rb(dd
, drq
);
305 deadline_del_drq_hash(drq
);
309 deadline_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
311 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
312 struct request
*__rq
;
316 * see if the merge hash can satisfy a back merge
318 __rq
= deadline_find_drq_hash(dd
, bio
->bi_sector
);
320 BUG_ON(__rq
->sector
+ __rq
->nr_sectors
!= bio
->bi_sector
);
322 if (elv_rq_merge_ok(__rq
, bio
)) {
323 ret
= ELEVATOR_BACK_MERGE
;
329 * check for front merge
331 if (dd
->front_merges
) {
332 sector_t rb_key
= bio
->bi_sector
+ bio_sectors(bio
);
334 __rq
= deadline_find_drq_rb(dd
, rb_key
, bio_data_dir(bio
));
336 BUG_ON(rb_key
!= rq_rb_key(__rq
));
338 if (elv_rq_merge_ok(__rq
, bio
)) {
339 ret
= ELEVATOR_FRONT_MERGE
;
345 return ELEVATOR_NO_MERGE
;
348 deadline_hot_drq_hash(dd
, RQ_DATA(__rq
));
353 static void deadline_merged_request(request_queue_t
*q
, struct request
*req
)
355 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
356 struct deadline_rq
*drq
= RQ_DATA(req
);
359 * hash always needs to be repositioned, key is end sector
361 deadline_del_drq_hash(drq
);
362 deadline_add_drq_hash(dd
, drq
);
365 * if the merge was a front merge, we need to reposition request
367 if (rq_rb_key(req
) != drq
->rb_key
) {
368 deadline_del_drq_rb(dd
, drq
);
369 deadline_add_drq_rb(dd
, drq
);
374 deadline_merged_requests(request_queue_t
*q
, struct request
*req
,
375 struct request
*next
)
377 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
378 struct deadline_rq
*drq
= RQ_DATA(req
);
379 struct deadline_rq
*dnext
= RQ_DATA(next
);
385 * reposition drq (this is the merged request) in hash, and in rbtree
386 * in case of a front merge
388 deadline_del_drq_hash(drq
);
389 deadline_add_drq_hash(dd
, drq
);
391 if (rq_rb_key(req
) != drq
->rb_key
) {
392 deadline_del_drq_rb(dd
, drq
);
393 deadline_add_drq_rb(dd
, drq
);
397 * if dnext expires before drq, assign its expire time to drq
398 * and move into dnext position (dnext will be deleted) in fifo
400 if (!list_empty(&drq
->fifo
) && !list_empty(&dnext
->fifo
)) {
401 if (time_before(dnext
->expires
, drq
->expires
)) {
402 list_move(&drq
->fifo
, &dnext
->fifo
);
403 drq
->expires
= dnext
->expires
;
408 * kill knowledge of next, this one is a goner
410 deadline_remove_request(q
, next
);
414 * move request from sort list to dispatch queue.
417 deadline_move_to_dispatch(struct deadline_data
*dd
, struct deadline_rq
*drq
)
419 request_queue_t
*q
= drq
->request
->q
;
421 deadline_remove_request(q
, drq
->request
);
422 elv_dispatch_add_tail(q
, drq
->request
);
426 * move an entry to dispatch queue
429 deadline_move_request(struct deadline_data
*dd
, struct deadline_rq
*drq
)
431 const int data_dir
= rq_data_dir(drq
->request
);
432 struct rb_node
*rbnext
= rb_next(&drq
->rb_node
);
434 dd
->next_drq
[READ
] = NULL
;
435 dd
->next_drq
[WRITE
] = NULL
;
438 dd
->next_drq
[data_dir
] = rb_entry_drq(rbnext
);
440 dd
->last_sector
= drq
->request
->sector
+ drq
->request
->nr_sectors
;
443 * take it off the sort and fifo list, move
446 deadline_move_to_dispatch(dd
, drq
);
449 #define list_entry_fifo(ptr) list_entry((ptr), struct deadline_rq, fifo)
452 * deadline_check_fifo returns 0 if there are no expired reads on the fifo,
453 * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
455 static inline int deadline_check_fifo(struct deadline_data
*dd
, int ddir
)
457 struct deadline_rq
*drq
= list_entry_fifo(dd
->fifo_list
[ddir
].next
);
462 if (time_after(jiffies
, drq
->expires
))
469 * deadline_dispatch_requests selects the best request according to
470 * read/write expire, fifo_batch, etc
472 static int deadline_dispatch_requests(request_queue_t
*q
, int force
)
474 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
475 const int reads
= !list_empty(&dd
->fifo_list
[READ
]);
476 const int writes
= !list_empty(&dd
->fifo_list
[WRITE
]);
477 struct deadline_rq
*drq
;
481 * batches are currently reads XOR writes
483 if (dd
->next_drq
[WRITE
])
484 drq
= dd
->next_drq
[WRITE
];
486 drq
= dd
->next_drq
[READ
];
489 /* we have a "next request" */
491 if (dd
->last_sector
!= drq
->request
->sector
)
492 /* end the batch on a non sequential request */
493 dd
->batching
+= dd
->fifo_batch
;
495 if (dd
->batching
< dd
->fifo_batch
)
496 /* we are still entitled to batch */
497 goto dispatch_request
;
501 * at this point we are not running a batch. select the appropriate
502 * data direction (read / write)
506 BUG_ON(RB_EMPTY(&dd
->sort_list
[READ
]));
508 if (writes
&& (dd
->starved
++ >= dd
->writes_starved
))
509 goto dispatch_writes
;
513 goto dispatch_find_request
;
517 * there are either no reads or writes have been starved
522 BUG_ON(RB_EMPTY(&dd
->sort_list
[WRITE
]));
528 goto dispatch_find_request
;
533 dispatch_find_request
:
535 * we are not running a batch, find best request for selected data_dir
537 if (deadline_check_fifo(dd
, data_dir
)) {
538 /* An expired request exists - satisfy it */
540 drq
= list_entry_fifo(dd
->fifo_list
[data_dir
].next
);
542 } else if (dd
->next_drq
[data_dir
]) {
544 * The last req was the same dir and we have a next request in
545 * sort order. No expired requests so continue on from here.
547 drq
= dd
->next_drq
[data_dir
];
550 * The last req was the other direction or we have run out of
551 * higher-sectored requests. Go back to the lowest sectored
552 * request (1 way elevator) and start a new batch.
555 drq
= deadline_find_first_drq(dd
, data_dir
);
560 * drq is the selected appropriate request.
563 deadline_move_request(dd
, drq
);
568 static int deadline_queue_empty(request_queue_t
*q
)
570 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
572 return list_empty(&dd
->fifo_list
[WRITE
])
573 && list_empty(&dd
->fifo_list
[READ
]);
576 static struct request
*
577 deadline_former_request(request_queue_t
*q
, struct request
*rq
)
579 struct deadline_rq
*drq
= RQ_DATA(rq
);
580 struct rb_node
*rbprev
= rb_prev(&drq
->rb_node
);
583 return rb_entry_drq(rbprev
)->request
;
588 static struct request
*
589 deadline_latter_request(request_queue_t
*q
, struct request
*rq
)
591 struct deadline_rq
*drq
= RQ_DATA(rq
);
592 struct rb_node
*rbnext
= rb_next(&drq
->rb_node
);
595 return rb_entry_drq(rbnext
)->request
;
600 static void deadline_exit_queue(elevator_t
*e
)
602 struct deadline_data
*dd
= e
->elevator_data
;
604 BUG_ON(!list_empty(&dd
->fifo_list
[READ
]));
605 BUG_ON(!list_empty(&dd
->fifo_list
[WRITE
]));
607 mempool_destroy(dd
->drq_pool
);
613 * initialize elevator private data (deadline_data), and alloc a drq for
614 * each request on the free lists
616 static void *deadline_init_queue(request_queue_t
*q
, elevator_t
*e
)
618 struct deadline_data
*dd
;
624 dd
= kmalloc_node(sizeof(*dd
), GFP_KERNEL
, q
->node
);
627 memset(dd
, 0, sizeof(*dd
));
629 dd
->hash
= kmalloc_node(sizeof(struct list_head
)*DL_HASH_ENTRIES
,
630 GFP_KERNEL
, q
->node
);
636 dd
->drq_pool
= mempool_create_node(BLKDEV_MIN_RQ
, mempool_alloc_slab
,
637 mempool_free_slab
, drq_pool
, q
->node
);
644 for (i
= 0; i
< DL_HASH_ENTRIES
; i
++)
645 INIT_LIST_HEAD(&dd
->hash
[i
]);
647 INIT_LIST_HEAD(&dd
->fifo_list
[READ
]);
648 INIT_LIST_HEAD(&dd
->fifo_list
[WRITE
]);
649 dd
->sort_list
[READ
] = RB_ROOT
;
650 dd
->sort_list
[WRITE
] = RB_ROOT
;
651 dd
->fifo_expire
[READ
] = read_expire
;
652 dd
->fifo_expire
[WRITE
] = write_expire
;
653 dd
->writes_starved
= writes_starved
;
654 dd
->front_merges
= 1;
655 dd
->fifo_batch
= fifo_batch
;
659 static void deadline_put_request(request_queue_t
*q
, struct request
*rq
)
661 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
662 struct deadline_rq
*drq
= RQ_DATA(rq
);
664 mempool_free(drq
, dd
->drq_pool
);
665 rq
->elevator_private
= NULL
;
669 deadline_set_request(request_queue_t
*q
, struct request
*rq
, struct bio
*bio
,
672 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
673 struct deadline_rq
*drq
;
675 drq
= mempool_alloc(dd
->drq_pool
, gfp_mask
);
677 memset(drq
, 0, sizeof(*drq
));
678 RB_CLEAR(&drq
->rb_node
);
681 INIT_LIST_HEAD(&drq
->hash
);
684 INIT_LIST_HEAD(&drq
->fifo
);
686 rq
->elevator_private
= drq
;
698 deadline_var_show(int var
, char *page
)
700 return sprintf(page
, "%d\n", var
);
704 deadline_var_store(int *var
, const char *page
, size_t count
)
706 char *p
= (char *) page
;
708 *var
= simple_strtol(p
, &p
, 10);
712 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
713 static ssize_t __FUNC(elevator_t *e, char *page) \
715 struct deadline_data *dd = e->elevator_data; \
716 int __data = __VAR; \
718 __data = jiffies_to_msecs(__data); \
719 return deadline_var_show(__data, (page)); \
721 SHOW_FUNCTION(deadline_read_expire_show
, dd
->fifo_expire
[READ
], 1);
722 SHOW_FUNCTION(deadline_write_expire_show
, dd
->fifo_expire
[WRITE
], 1);
723 SHOW_FUNCTION(deadline_writes_starved_show
, dd
->writes_starved
, 0);
724 SHOW_FUNCTION(deadline_front_merges_show
, dd
->front_merges
, 0);
725 SHOW_FUNCTION(deadline_fifo_batch_show
, dd
->fifo_batch
, 0);
728 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
729 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
731 struct deadline_data *dd = e->elevator_data; \
733 int ret = deadline_var_store(&__data, (page), count); \
734 if (__data < (MIN)) \
736 else if (__data > (MAX)) \
739 *(__PTR) = msecs_to_jiffies(__data); \
744 STORE_FUNCTION(deadline_read_expire_store
, &dd
->fifo_expire
[READ
], 0, INT_MAX
, 1);
745 STORE_FUNCTION(deadline_write_expire_store
, &dd
->fifo_expire
[WRITE
], 0, INT_MAX
, 1);
746 STORE_FUNCTION(deadline_writes_starved_store
, &dd
->writes_starved
, INT_MIN
, INT_MAX
, 0);
747 STORE_FUNCTION(deadline_front_merges_store
, &dd
->front_merges
, 0, 1, 0);
748 STORE_FUNCTION(deadline_fifo_batch_store
, &dd
->fifo_batch
, 0, INT_MAX
, 0);
749 #undef STORE_FUNCTION
751 #define DD_ATTR(name) \
752 __ATTR(name, S_IRUGO|S_IWUSR, deadline_##name##_show, \
753 deadline_##name##_store)
755 static struct elv_fs_entry deadline_attrs
[] = {
756 DD_ATTR(read_expire
),
757 DD_ATTR(write_expire
),
758 DD_ATTR(writes_starved
),
759 DD_ATTR(front_merges
),
764 static struct elevator_type iosched_deadline
= {
766 .elevator_merge_fn
= deadline_merge
,
767 .elevator_merged_fn
= deadline_merged_request
,
768 .elevator_merge_req_fn
= deadline_merged_requests
,
769 .elevator_dispatch_fn
= deadline_dispatch_requests
,
770 .elevator_add_req_fn
= deadline_add_request
,
771 .elevator_queue_empty_fn
= deadline_queue_empty
,
772 .elevator_former_req_fn
= deadline_former_request
,
773 .elevator_latter_req_fn
= deadline_latter_request
,
774 .elevator_set_req_fn
= deadline_set_request
,
775 .elevator_put_req_fn
= deadline_put_request
,
776 .elevator_init_fn
= deadline_init_queue
,
777 .elevator_exit_fn
= deadline_exit_queue
,
780 .elevator_attrs
= deadline_attrs
,
781 .elevator_name
= "deadline",
782 .elevator_owner
= THIS_MODULE
,
785 static int __init
deadline_init(void)
789 drq_pool
= kmem_cache_create("deadline_drq", sizeof(struct deadline_rq
),
795 ret
= elv_register(&iosched_deadline
);
797 kmem_cache_destroy(drq_pool
);
802 static void __exit
deadline_exit(void)
804 kmem_cache_destroy(drq_pool
);
805 elv_unregister(&iosched_deadline
);
808 module_init(deadline_init
);
809 module_exit(deadline_exit
);
811 MODULE_AUTHOR("Jens Axboe");
812 MODULE_LICENSE("GPL");
813 MODULE_DESCRIPTION("deadline IO scheduler");