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/module.h>
12 #include <linux/slab.h>
13 #include <linux/init.h>
14 #include <linux/compiler.h>
15 #include <linux/hash.h>
16 #include <linux/rbtree.h>
19 * See Documentation/block/deadline-iosched.txt
21 static const int read_expire = HZ / 2; /* max time before a read is submitted. */
22 static const int write_expire = 5 * HZ; /* ditto for writes, these limits are SOFT! */
23 static const int writes_starved = 2; /* max times reads can starve a write */
24 static const int fifo_batch = 16; /* # of sequential requests treated as one
25 by the above parameters. For throughput. */
27 static const int deadline_hash_shift = 5;
28 #define DL_HASH_BLOCK(sec) ((sec) >> 3)
29 #define DL_HASH_FN(sec) (hash_long(DL_HASH_BLOCK((sec)), deadline_hash_shift))
30 #define DL_HASH_ENTRIES (1 << deadline_hash_shift)
31 #define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
32 #define ON_HASH(drq) (!hlist_unhashed(&(drq)->hash))
34 struct deadline_data {
40 * requests (deadline_rq s) are present on both sort_list and fifo_list
42 struct rb_root sort_list[2];
43 struct list_head fifo_list[2];
46 * next in sort order. read, write or both are NULL
48 struct deadline_rq *next_drq[2];
49 struct hlist_head *hash; /* request hash */
50 unsigned int batching; /* number of sequential requests made */
51 sector_t last_sector; /* head position */
52 unsigned int starved; /* times reads have starved writes */
55 * settings that change how the i/o scheduler behaves
70 * rbtree index, key is the starting offset
72 struct rb_node rb_node;
75 struct request *request;
78 * request hash, key is the ending offset (for back merge lookup)
80 struct hlist_node hash;
85 struct list_head fifo;
86 unsigned long expires;
89 static void deadline_move_request(struct deadline_data *dd, struct deadline_rq *drq);
91 static kmem_cache_t *drq_pool;
93 #define RQ_DATA(rq) ((struct deadline_rq *) (rq)->elevator_private)
96 * the back merge hash support functions
98 static inline void __deadline_del_drq_hash(struct deadline_rq *drq)
100 hlist_del_init(&drq->hash);
103 static inline void deadline_del_drq_hash(struct deadline_rq *drq)
106 __deadline_del_drq_hash(drq);
110 deadline_add_drq_hash(struct deadline_data *dd, struct deadline_rq *drq)
112 struct request *rq = drq->request;
114 BUG_ON(ON_HASH(drq));
116 hlist_add_head(&drq->hash, &dd->hash[DL_HASH_FN(rq_hash_key(rq))]);
120 * move hot entry to front of chain
123 deadline_hot_drq_hash(struct deadline_data *dd, struct deadline_rq *drq)
125 struct request *rq = drq->request;
126 struct hlist_head *head = &dd->hash[DL_HASH_FN(rq_hash_key(rq))];
128 if (ON_HASH(drq) && &drq->hash != head->first) {
129 hlist_del(&drq->hash);
130 hlist_add_head(&drq->hash, head);
134 static struct request *
135 deadline_find_drq_hash(struct deadline_data *dd, sector_t offset)
137 struct hlist_head *hash_list = &dd->hash[DL_HASH_FN(offset)];
138 struct hlist_node *entry, *next;
139 struct deadline_rq *drq;
141 hlist_for_each_entry_safe(drq, entry, next, hash_list, hash) {
142 struct request *__rq = drq->request;
144 BUG_ON(!ON_HASH(drq));
146 if (!rq_mergeable(__rq)) {
147 __deadline_del_drq_hash(drq);
151 if (rq_hash_key(__rq) == offset)
159 * rb tree support functions
161 #define rb_entry_drq(node) rb_entry((node), struct deadline_rq, rb_node)
162 #define DRQ_RB_ROOT(dd, drq) (&(dd)->sort_list[rq_data_dir((drq)->request)])
163 #define rq_rb_key(rq) (rq)->sector
165 static struct deadline_rq *
166 __deadline_add_drq_rb(struct deadline_data *dd, struct deadline_rq *drq)
168 struct rb_node **p = &DRQ_RB_ROOT(dd, drq)->rb_node;
169 struct rb_node *parent = NULL;
170 struct deadline_rq *__drq;
174 __drq = rb_entry_drq(parent);
176 if (drq->rb_key < __drq->rb_key)
178 else if (drq->rb_key > __drq->rb_key)
184 rb_link_node(&drq->rb_node, parent, p);
189 deadline_add_drq_rb(struct deadline_data *dd, struct deadline_rq *drq)
191 struct deadline_rq *__alias;
193 drq->rb_key = rq_rb_key(drq->request);
196 __alias = __deadline_add_drq_rb(dd, drq);
198 rb_insert_color(&drq->rb_node, DRQ_RB_ROOT(dd, drq));
202 deadline_move_request(dd, __alias);
207 deadline_del_drq_rb(struct deadline_data *dd, struct deadline_rq *drq)
209 const int data_dir = rq_data_dir(drq->request);
211 if (dd->next_drq[data_dir] == drq) {
212 struct rb_node *rbnext = rb_next(&drq->rb_node);
214 dd->next_drq[data_dir] = NULL;
216 dd->next_drq[data_dir] = rb_entry_drq(rbnext);
219 BUG_ON(!RB_EMPTY_NODE(&drq->rb_node));
220 rb_erase(&drq->rb_node, DRQ_RB_ROOT(dd, drq));
221 RB_CLEAR_NODE(&drq->rb_node);
224 static struct request *
225 deadline_find_drq_rb(struct deadline_data *dd, sector_t sector, int data_dir)
227 struct rb_node *n = dd->sort_list[data_dir].rb_node;
228 struct deadline_rq *drq;
231 drq = rb_entry_drq(n);
233 if (sector < drq->rb_key)
235 else if (sector > drq->rb_key)
245 * deadline_find_first_drq finds the first (lowest sector numbered) request
246 * for the specified data_dir. Used to sweep back to the start of the disk
247 * (1-way elevator) after we process the last (highest sector) request.
249 static struct deadline_rq *
250 deadline_find_first_drq(struct deadline_data *dd, int data_dir)
252 struct rb_node *n = dd->sort_list[data_dir].rb_node;
255 if (n->rb_left == NULL)
256 return rb_entry_drq(n);
263 * add drq to rbtree and fifo
266 deadline_add_request(struct request_queue *q, struct request *rq)
268 struct deadline_data *dd = q->elevator->elevator_data;
269 struct deadline_rq *drq = RQ_DATA(rq);
271 const int data_dir = rq_data_dir(drq->request);
273 deadline_add_drq_rb(dd, drq);
275 * set expire time (only used for reads) and add to fifo list
277 drq->expires = jiffies + dd->fifo_expire[data_dir];
278 list_add_tail(&drq->fifo, &dd->fifo_list[data_dir]);
280 if (rq_mergeable(rq))
281 deadline_add_drq_hash(dd, drq);
285 * remove rq from rbtree, fifo, and hash
287 static void deadline_remove_request(request_queue_t *q, struct request *rq)
289 struct deadline_rq *drq = RQ_DATA(rq);
290 struct deadline_data *dd = q->elevator->elevator_data;
292 list_del_init(&drq->fifo);
293 deadline_del_drq_rb(dd, drq);
294 deadline_del_drq_hash(drq);
298 deadline_merge(request_queue_t *q, struct request **req, struct bio *bio)
300 struct deadline_data *dd = q->elevator->elevator_data;
301 struct request *__rq;
305 * see if the merge hash can satisfy a back merge
307 __rq = deadline_find_drq_hash(dd, bio->bi_sector);
309 BUG_ON(__rq->sector + __rq->nr_sectors != bio->bi_sector);
311 if (elv_rq_merge_ok(__rq, bio)) {
312 ret = ELEVATOR_BACK_MERGE;
318 * check for front merge
320 if (dd->front_merges) {
321 sector_t rb_key = bio->bi_sector + bio_sectors(bio);
323 __rq = deadline_find_drq_rb(dd, rb_key, bio_data_dir(bio));
325 BUG_ON(rb_key != rq_rb_key(__rq));
327 if (elv_rq_merge_ok(__rq, bio)) {
328 ret = ELEVATOR_FRONT_MERGE;
334 return ELEVATOR_NO_MERGE;
337 deadline_hot_drq_hash(dd, RQ_DATA(__rq));
342 static void deadline_merged_request(request_queue_t *q, struct request *req)
344 struct deadline_data *dd = q->elevator->elevator_data;
345 struct deadline_rq *drq = RQ_DATA(req);
348 * hash always needs to be repositioned, key is end sector
350 deadline_del_drq_hash(drq);
351 deadline_add_drq_hash(dd, drq);
354 * if the merge was a front merge, we need to reposition request
356 if (rq_rb_key(req) != drq->rb_key) {
357 deadline_del_drq_rb(dd, drq);
358 deadline_add_drq_rb(dd, drq);
363 deadline_merged_requests(request_queue_t *q, struct request *req,
364 struct request *next)
366 struct deadline_data *dd = q->elevator->elevator_data;
367 struct deadline_rq *drq = RQ_DATA(req);
368 struct deadline_rq *dnext = RQ_DATA(next);
374 * reposition drq (this is the merged request) in hash, and in rbtree
375 * in case of a front merge
377 deadline_del_drq_hash(drq);
378 deadline_add_drq_hash(dd, drq);
380 if (rq_rb_key(req) != drq->rb_key) {
381 deadline_del_drq_rb(dd, drq);
382 deadline_add_drq_rb(dd, drq);
386 * if dnext expires before drq, assign its expire time to drq
387 * and move into dnext position (dnext will be deleted) in fifo
389 if (!list_empty(&drq->fifo) && !list_empty(&dnext->fifo)) {
390 if (time_before(dnext->expires, drq->expires)) {
391 list_move(&drq->fifo, &dnext->fifo);
392 drq->expires = dnext->expires;
397 * kill knowledge of next, this one is a goner
399 deadline_remove_request(q, next);
403 * move request from sort list to dispatch queue.
406 deadline_move_to_dispatch(struct deadline_data *dd, struct deadline_rq *drq)
408 request_queue_t *q = drq->request->q;
410 deadline_remove_request(q, drq->request);
411 elv_dispatch_add_tail(q, drq->request);
415 * move an entry to dispatch queue
418 deadline_move_request(struct deadline_data *dd, struct deadline_rq *drq)
420 const int data_dir = rq_data_dir(drq->request);
421 struct rb_node *rbnext = rb_next(&drq->rb_node);
423 dd->next_drq[READ] = NULL;
424 dd->next_drq[WRITE] = NULL;
427 dd->next_drq[data_dir] = rb_entry_drq(rbnext);
429 dd->last_sector = drq->request->sector + drq->request->nr_sectors;
432 * take it off the sort and fifo list, move
435 deadline_move_to_dispatch(dd, drq);
438 #define list_entry_fifo(ptr) list_entry((ptr), struct deadline_rq, fifo)
441 * deadline_check_fifo returns 0 if there are no expired reads on the fifo,
442 * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
444 static inline int deadline_check_fifo(struct deadline_data *dd, int ddir)
446 struct deadline_rq *drq = list_entry_fifo(dd->fifo_list[ddir].next);
451 if (time_after(jiffies, drq->expires))
458 * deadline_dispatch_requests selects the best request according to
459 * read/write expire, fifo_batch, etc
461 static int deadline_dispatch_requests(request_queue_t *q, int force)
463 struct deadline_data *dd = q->elevator->elevator_data;
464 const int reads = !list_empty(&dd->fifo_list[READ]);
465 const int writes = !list_empty(&dd->fifo_list[WRITE]);
466 struct deadline_rq *drq;
470 * batches are currently reads XOR writes
472 if (dd->next_drq[WRITE])
473 drq = dd->next_drq[WRITE];
475 drq = dd->next_drq[READ];
478 /* we have a "next request" */
480 if (dd->last_sector != drq->request->sector)
481 /* end the batch on a non sequential request */
482 dd->batching += dd->fifo_batch;
484 if (dd->batching < dd->fifo_batch)
485 /* we are still entitled to batch */
486 goto dispatch_request;
490 * at this point we are not running a batch. select the appropriate
491 * data direction (read / write)
495 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[READ]));
497 if (writes && (dd->starved++ >= dd->writes_starved))
498 goto dispatch_writes;
502 goto dispatch_find_request;
506 * there are either no reads or writes have been starved
511 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[WRITE]));
517 goto dispatch_find_request;
522 dispatch_find_request:
524 * we are not running a batch, find best request for selected data_dir
526 if (deadline_check_fifo(dd, data_dir)) {
527 /* An expired request exists - satisfy it */
529 drq = list_entry_fifo(dd->fifo_list[data_dir].next);
531 } else if (dd->next_drq[data_dir]) {
533 * The last req was the same dir and we have a next request in
534 * sort order. No expired requests so continue on from here.
536 drq = dd->next_drq[data_dir];
539 * The last req was the other direction or we have run out of
540 * higher-sectored requests. Go back to the lowest sectored
541 * request (1 way elevator) and start a new batch.
544 drq = deadline_find_first_drq(dd, data_dir);
549 * drq is the selected appropriate request.
552 deadline_move_request(dd, drq);
557 static int deadline_queue_empty(request_queue_t *q)
559 struct deadline_data *dd = q->elevator->elevator_data;
561 return list_empty(&dd->fifo_list[WRITE])
562 && list_empty(&dd->fifo_list[READ]);
565 static struct request *
566 deadline_former_request(request_queue_t *q, struct request *rq)
568 struct deadline_rq *drq = RQ_DATA(rq);
569 struct rb_node *rbprev = rb_prev(&drq->rb_node);
572 return rb_entry_drq(rbprev)->request;
577 static struct request *
578 deadline_latter_request(request_queue_t *q, struct request *rq)
580 struct deadline_rq *drq = RQ_DATA(rq);
581 struct rb_node *rbnext = rb_next(&drq->rb_node);
584 return rb_entry_drq(rbnext)->request;
589 static void deadline_exit_queue(elevator_t *e)
591 struct deadline_data *dd = e->elevator_data;
593 BUG_ON(!list_empty(&dd->fifo_list[READ]));
594 BUG_ON(!list_empty(&dd->fifo_list[WRITE]));
596 mempool_destroy(dd->drq_pool);
602 * initialize elevator private data (deadline_data), and alloc a drq for
603 * each request on the free lists
605 static void *deadline_init_queue(request_queue_t *q, elevator_t *e)
607 struct deadline_data *dd;
613 dd = kmalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
616 memset(dd, 0, sizeof(*dd));
618 dd->hash = kmalloc_node(sizeof(struct hlist_head)*DL_HASH_ENTRIES,
619 GFP_KERNEL, q->node);
625 dd->drq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
626 mempool_free_slab, drq_pool, q->node);
633 for (i = 0; i < DL_HASH_ENTRIES; i++)
634 INIT_HLIST_HEAD(&dd->hash[i]);
636 INIT_LIST_HEAD(&dd->fifo_list[READ]);
637 INIT_LIST_HEAD(&dd->fifo_list[WRITE]);
638 dd->sort_list[READ] = RB_ROOT;
639 dd->sort_list[WRITE] = RB_ROOT;
640 dd->fifo_expire[READ] = read_expire;
641 dd->fifo_expire[WRITE] = write_expire;
642 dd->writes_starved = writes_starved;
643 dd->front_merges = 1;
644 dd->fifo_batch = fifo_batch;
648 static void deadline_put_request(request_queue_t *q, struct request *rq)
650 struct deadline_data *dd = q->elevator->elevator_data;
651 struct deadline_rq *drq = RQ_DATA(rq);
653 mempool_free(drq, dd->drq_pool);
654 rq->elevator_private = NULL;
658 deadline_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
661 struct deadline_data *dd = q->elevator->elevator_data;
662 struct deadline_rq *drq;
664 drq = mempool_alloc(dd->drq_pool, gfp_mask);
666 memset(drq, 0, sizeof(*drq));
667 RB_CLEAR_NODE(&drq->rb_node);
670 INIT_HLIST_NODE(&drq->hash);
672 INIT_LIST_HEAD(&drq->fifo);
674 rq->elevator_private = drq;
686 deadline_var_show(int var, char *page)
688 return sprintf(page, "%d\n", var);
692 deadline_var_store(int *var, const char *page, size_t count)
694 char *p = (char *) page;
696 *var = simple_strtol(p, &p, 10);
700 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
701 static ssize_t __FUNC(elevator_t *e, char *page) \
703 struct deadline_data *dd = e->elevator_data; \
704 int __data = __VAR; \
706 __data = jiffies_to_msecs(__data); \
707 return deadline_var_show(__data, (page)); \
709 SHOW_FUNCTION(deadline_read_expire_show, dd->fifo_expire[READ], 1);
710 SHOW_FUNCTION(deadline_write_expire_show, dd->fifo_expire[WRITE], 1);
711 SHOW_FUNCTION(deadline_writes_starved_show, dd->writes_starved, 0);
712 SHOW_FUNCTION(deadline_front_merges_show, dd->front_merges, 0);
713 SHOW_FUNCTION(deadline_fifo_batch_show, dd->fifo_batch, 0);
716 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
717 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
719 struct deadline_data *dd = e->elevator_data; \
721 int ret = deadline_var_store(&__data, (page), count); \
722 if (__data < (MIN)) \
724 else if (__data > (MAX)) \
727 *(__PTR) = msecs_to_jiffies(__data); \
732 STORE_FUNCTION(deadline_read_expire_store, &dd->fifo_expire[READ], 0, INT_MAX, 1);
733 STORE_FUNCTION(deadline_write_expire_store, &dd->fifo_expire[WRITE], 0, INT_MAX, 1);
734 STORE_FUNCTION(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX, 0);
735 STORE_FUNCTION(deadline_front_merges_store, &dd->front_merges, 0, 1, 0);
736 STORE_FUNCTION(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX, 0);
737 #undef STORE_FUNCTION
739 #define DD_ATTR(name) \
740 __ATTR(name, S_IRUGO|S_IWUSR, deadline_##name##_show, \
741 deadline_##name##_store)
743 static struct elv_fs_entry deadline_attrs[] = {
744 DD_ATTR(read_expire),
745 DD_ATTR(write_expire),
746 DD_ATTR(writes_starved),
747 DD_ATTR(front_merges),
752 static struct elevator_type iosched_deadline = {
754 .elevator_merge_fn = deadline_merge,
755 .elevator_merged_fn = deadline_merged_request,
756 .elevator_merge_req_fn = deadline_merged_requests,
757 .elevator_dispatch_fn = deadline_dispatch_requests,
758 .elevator_add_req_fn = deadline_add_request,
759 .elevator_queue_empty_fn = deadline_queue_empty,
760 .elevator_former_req_fn = deadline_former_request,
761 .elevator_latter_req_fn = deadline_latter_request,
762 .elevator_set_req_fn = deadline_set_request,
763 .elevator_put_req_fn = deadline_put_request,
764 .elevator_init_fn = deadline_init_queue,
765 .elevator_exit_fn = deadline_exit_queue,
768 .elevator_attrs = deadline_attrs,
769 .elevator_name = "deadline",
770 .elevator_owner = THIS_MODULE,
773 static int __init deadline_init(void)
777 drq_pool = kmem_cache_create("deadline_drq", sizeof(struct deadline_rq),
783 ret = elv_register(&iosched_deadline);
785 kmem_cache_destroy(drq_pool);
790 static void __exit deadline_exit(void)
792 kmem_cache_destroy(drq_pool);
793 elv_unregister(&iosched_deadline);
796 module_init(deadline_init);
797 module_exit(deadline_exit);
799 MODULE_AUTHOR("Jens Axboe");
800 MODULE_LICENSE("GPL");
801 MODULE_DESCRIPTION("deadline IO scheduler");