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 ON_HASH(drq) (!hlist_unhashed(&(drq)->hash))
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 deadline_rq *next_drq[2];
50 struct hlist_head *hash; /* request hash */
51 unsigned int batching; /* number of sequential requests made */
52 sector_t last_sector; /* head position */
53 unsigned int starved; /* times reads have starved writes */
56 * settings that change how the i/o scheduler behaves
71 * rbtree index, key is the starting offset
73 struct rb_node rb_node;
76 struct request *request;
79 * request hash, key is the ending offset (for back merge lookup)
81 struct hlist_node hash;
86 struct list_head fifo;
87 unsigned long expires;
90 static void deadline_move_request(struct deadline_data *dd, struct deadline_rq *drq);
92 static kmem_cache_t *drq_pool;
94 #define RQ_DATA(rq) ((struct deadline_rq *) (rq)->elevator_private)
97 * the back merge hash support functions
99 static inline void __deadline_del_drq_hash(struct deadline_rq *drq)
101 hlist_del_init(&drq->hash);
104 static inline void deadline_del_drq_hash(struct deadline_rq *drq)
107 __deadline_del_drq_hash(drq);
111 deadline_add_drq_hash(struct deadline_data *dd, struct deadline_rq *drq)
113 struct request *rq = drq->request;
115 BUG_ON(ON_HASH(drq));
117 hlist_add_head(&drq->hash, &dd->hash[DL_HASH_FN(rq_hash_key(rq))]);
121 * move hot entry to front of chain
124 deadline_hot_drq_hash(struct deadline_data *dd, struct deadline_rq *drq)
126 struct request *rq = drq->request;
127 struct hlist_head *head = &dd->hash[DL_HASH_FN(rq_hash_key(rq))];
129 if (ON_HASH(drq) && &drq->hash != head->first) {
130 hlist_del(&drq->hash);
131 hlist_add_head(&drq->hash, head);
135 static struct request *
136 deadline_find_drq_hash(struct deadline_data *dd, sector_t offset)
138 struct hlist_head *hash_list = &dd->hash[DL_HASH_FN(offset)];
139 struct hlist_node *entry, *next;
140 struct deadline_rq *drq;
142 hlist_for_each_entry_safe(drq, entry, next, hash_list, hash) {
143 struct request *__rq = drq->request;
145 BUG_ON(!ON_HASH(drq));
147 if (!rq_mergeable(__rq)) {
148 __deadline_del_drq_hash(drq);
152 if (rq_hash_key(__rq) == offset)
160 * rb tree support functions
162 #define RB_EMPTY(root) ((root)->rb_node == NULL)
163 #define ON_RB(node) (rb_parent(node) != node)
164 #define RB_CLEAR(node) (rb_set_parent(node, node))
165 #define rb_entry_drq(node) rb_entry((node), struct deadline_rq, rb_node)
166 #define DRQ_RB_ROOT(dd, drq) (&(dd)->sort_list[rq_data_dir((drq)->request)])
167 #define rq_rb_key(rq) (rq)->sector
169 static struct deadline_rq *
170 __deadline_add_drq_rb(struct deadline_data *dd, struct deadline_rq *drq)
172 struct rb_node **p = &DRQ_RB_ROOT(dd, drq)->rb_node;
173 struct rb_node *parent = NULL;
174 struct deadline_rq *__drq;
178 __drq = rb_entry_drq(parent);
180 if (drq->rb_key < __drq->rb_key)
182 else if (drq->rb_key > __drq->rb_key)
188 rb_link_node(&drq->rb_node, parent, p);
193 deadline_add_drq_rb(struct deadline_data *dd, struct deadline_rq *drq)
195 struct deadline_rq *__alias;
197 drq->rb_key = rq_rb_key(drq->request);
200 __alias = __deadline_add_drq_rb(dd, drq);
202 rb_insert_color(&drq->rb_node, DRQ_RB_ROOT(dd, drq));
206 deadline_move_request(dd, __alias);
211 deadline_del_drq_rb(struct deadline_data *dd, struct deadline_rq *drq)
213 const int data_dir = rq_data_dir(drq->request);
215 if (dd->next_drq[data_dir] == drq) {
216 struct rb_node *rbnext = rb_next(&drq->rb_node);
218 dd->next_drq[data_dir] = NULL;
220 dd->next_drq[data_dir] = rb_entry_drq(rbnext);
223 BUG_ON(!ON_RB(&drq->rb_node));
224 rb_erase(&drq->rb_node, DRQ_RB_ROOT(dd, drq));
225 RB_CLEAR(&drq->rb_node);
228 static struct request *
229 deadline_find_drq_rb(struct deadline_data *dd, sector_t sector, int data_dir)
231 struct rb_node *n = dd->sort_list[data_dir].rb_node;
232 struct deadline_rq *drq;
235 drq = rb_entry_drq(n);
237 if (sector < drq->rb_key)
239 else if (sector > drq->rb_key)
249 * deadline_find_first_drq finds the first (lowest sector numbered) request
250 * for the specified data_dir. Used to sweep back to the start of the disk
251 * (1-way elevator) after we process the last (highest sector) request.
253 static struct deadline_rq *
254 deadline_find_first_drq(struct deadline_data *dd, int data_dir)
256 struct rb_node *n = dd->sort_list[data_dir].rb_node;
259 if (n->rb_left == NULL)
260 return rb_entry_drq(n);
267 * add drq to rbtree and fifo
270 deadline_add_request(struct request_queue *q, struct request *rq)
272 struct deadline_data *dd = q->elevator->elevator_data;
273 struct deadline_rq *drq = RQ_DATA(rq);
275 const int data_dir = rq_data_dir(drq->request);
277 deadline_add_drq_rb(dd, drq);
279 * set expire time (only used for reads) and add to fifo list
281 drq->expires = jiffies + dd->fifo_expire[data_dir];
282 list_add_tail(&drq->fifo, &dd->fifo_list[data_dir]);
284 if (rq_mergeable(rq))
285 deadline_add_drq_hash(dd, drq);
289 * remove rq from rbtree, fifo, and hash
291 static void deadline_remove_request(request_queue_t *q, struct request *rq)
293 struct deadline_rq *drq = RQ_DATA(rq);
294 struct deadline_data *dd = q->elevator->elevator_data;
296 list_del_init(&drq->fifo);
297 deadline_del_drq_rb(dd, drq);
298 deadline_del_drq_hash(drq);
302 deadline_merge(request_queue_t *q, struct request **req, struct bio *bio)
304 struct deadline_data *dd = q->elevator->elevator_data;
305 struct request *__rq;
309 * see if the merge hash can satisfy a back merge
311 __rq = deadline_find_drq_hash(dd, bio->bi_sector);
313 BUG_ON(__rq->sector + __rq->nr_sectors != bio->bi_sector);
315 if (elv_rq_merge_ok(__rq, bio)) {
316 ret = ELEVATOR_BACK_MERGE;
322 * check for front merge
324 if (dd->front_merges) {
325 sector_t rb_key = bio->bi_sector + bio_sectors(bio);
327 __rq = deadline_find_drq_rb(dd, rb_key, bio_data_dir(bio));
329 BUG_ON(rb_key != rq_rb_key(__rq));
331 if (elv_rq_merge_ok(__rq, bio)) {
332 ret = ELEVATOR_FRONT_MERGE;
338 return ELEVATOR_NO_MERGE;
341 deadline_hot_drq_hash(dd, RQ_DATA(__rq));
346 static void deadline_merged_request(request_queue_t *q, struct request *req)
348 struct deadline_data *dd = q->elevator->elevator_data;
349 struct deadline_rq *drq = RQ_DATA(req);
352 * hash always needs to be repositioned, key is end sector
354 deadline_del_drq_hash(drq);
355 deadline_add_drq_hash(dd, drq);
358 * if the merge was a front merge, we need to reposition request
360 if (rq_rb_key(req) != drq->rb_key) {
361 deadline_del_drq_rb(dd, drq);
362 deadline_add_drq_rb(dd, drq);
367 deadline_merged_requests(request_queue_t *q, struct request *req,
368 struct request *next)
370 struct deadline_data *dd = q->elevator->elevator_data;
371 struct deadline_rq *drq = RQ_DATA(req);
372 struct deadline_rq *dnext = RQ_DATA(next);
378 * reposition drq (this is the merged request) in hash, and in rbtree
379 * in case of a front merge
381 deadline_del_drq_hash(drq);
382 deadline_add_drq_hash(dd, drq);
384 if (rq_rb_key(req) != drq->rb_key) {
385 deadline_del_drq_rb(dd, drq);
386 deadline_add_drq_rb(dd, drq);
390 * if dnext expires before drq, assign its expire time to drq
391 * and move into dnext position (dnext will be deleted) in fifo
393 if (!list_empty(&drq->fifo) && !list_empty(&dnext->fifo)) {
394 if (time_before(dnext->expires, drq->expires)) {
395 list_move(&drq->fifo, &dnext->fifo);
396 drq->expires = dnext->expires;
401 * kill knowledge of next, this one is a goner
403 deadline_remove_request(q, next);
407 * move request from sort list to dispatch queue.
410 deadline_move_to_dispatch(struct deadline_data *dd, struct deadline_rq *drq)
412 request_queue_t *q = drq->request->q;
414 deadline_remove_request(q, drq->request);
415 elv_dispatch_add_tail(q, drq->request);
419 * move an entry to dispatch queue
422 deadline_move_request(struct deadline_data *dd, struct deadline_rq *drq)
424 const int data_dir = rq_data_dir(drq->request);
425 struct rb_node *rbnext = rb_next(&drq->rb_node);
427 dd->next_drq[READ] = NULL;
428 dd->next_drq[WRITE] = NULL;
431 dd->next_drq[data_dir] = rb_entry_drq(rbnext);
433 dd->last_sector = drq->request->sector + drq->request->nr_sectors;
436 * take it off the sort and fifo list, move
439 deadline_move_to_dispatch(dd, drq);
442 #define list_entry_fifo(ptr) list_entry((ptr), struct deadline_rq, fifo)
445 * deadline_check_fifo returns 0 if there are no expired reads on the fifo,
446 * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
448 static inline int deadline_check_fifo(struct deadline_data *dd, int ddir)
450 struct deadline_rq *drq = list_entry_fifo(dd->fifo_list[ddir].next);
455 if (time_after(jiffies, drq->expires))
462 * deadline_dispatch_requests selects the best request according to
463 * read/write expire, fifo_batch, etc
465 static int deadline_dispatch_requests(request_queue_t *q, int force)
467 struct deadline_data *dd = q->elevator->elevator_data;
468 const int reads = !list_empty(&dd->fifo_list[READ]);
469 const int writes = !list_empty(&dd->fifo_list[WRITE]);
470 struct deadline_rq *drq;
474 * batches are currently reads XOR writes
476 if (dd->next_drq[WRITE])
477 drq = dd->next_drq[WRITE];
479 drq = dd->next_drq[READ];
482 /* we have a "next request" */
484 if (dd->last_sector != drq->request->sector)
485 /* end the batch on a non sequential request */
486 dd->batching += dd->fifo_batch;
488 if (dd->batching < dd->fifo_batch)
489 /* we are still entitled to batch */
490 goto dispatch_request;
494 * at this point we are not running a batch. select the appropriate
495 * data direction (read / write)
499 BUG_ON(RB_EMPTY(&dd->sort_list[READ]));
501 if (writes && (dd->starved++ >= dd->writes_starved))
502 goto dispatch_writes;
506 goto dispatch_find_request;
510 * there are either no reads or writes have been starved
515 BUG_ON(RB_EMPTY(&dd->sort_list[WRITE]));
521 goto dispatch_find_request;
526 dispatch_find_request:
528 * we are not running a batch, find best request for selected data_dir
530 if (deadline_check_fifo(dd, data_dir)) {
531 /* An expired request exists - satisfy it */
533 drq = list_entry_fifo(dd->fifo_list[data_dir].next);
535 } else if (dd->next_drq[data_dir]) {
537 * The last req was the same dir and we have a next request in
538 * sort order. No expired requests so continue on from here.
540 drq = dd->next_drq[data_dir];
543 * The last req was the other direction or we have run out of
544 * higher-sectored requests. Go back to the lowest sectored
545 * request (1 way elevator) and start a new batch.
548 drq = deadline_find_first_drq(dd, data_dir);
553 * drq is the selected appropriate request.
556 deadline_move_request(dd, drq);
561 static int deadline_queue_empty(request_queue_t *q)
563 struct deadline_data *dd = q->elevator->elevator_data;
565 return list_empty(&dd->fifo_list[WRITE])
566 && list_empty(&dd->fifo_list[READ]);
569 static struct request *
570 deadline_former_request(request_queue_t *q, struct request *rq)
572 struct deadline_rq *drq = RQ_DATA(rq);
573 struct rb_node *rbprev = rb_prev(&drq->rb_node);
576 return rb_entry_drq(rbprev)->request;
581 static struct request *
582 deadline_latter_request(request_queue_t *q, struct request *rq)
584 struct deadline_rq *drq = RQ_DATA(rq);
585 struct rb_node *rbnext = rb_next(&drq->rb_node);
588 return rb_entry_drq(rbnext)->request;
593 static void deadline_exit_queue(elevator_t *e)
595 struct deadline_data *dd = e->elevator_data;
597 BUG_ON(!list_empty(&dd->fifo_list[READ]));
598 BUG_ON(!list_empty(&dd->fifo_list[WRITE]));
600 mempool_destroy(dd->drq_pool);
606 * initialize elevator private data (deadline_data), and alloc a drq for
607 * each request on the free lists
609 static void *deadline_init_queue(request_queue_t *q, elevator_t *e)
611 struct deadline_data *dd;
617 dd = kmalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
620 memset(dd, 0, sizeof(*dd));
622 dd->hash = kmalloc_node(sizeof(struct hlist_head)*DL_HASH_ENTRIES,
623 GFP_KERNEL, q->node);
629 dd->drq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
630 mempool_free_slab, drq_pool, q->node);
637 for (i = 0; i < DL_HASH_ENTRIES; i++)
638 INIT_HLIST_HEAD(&dd->hash[i]);
640 INIT_LIST_HEAD(&dd->fifo_list[READ]);
641 INIT_LIST_HEAD(&dd->fifo_list[WRITE]);
642 dd->sort_list[READ] = RB_ROOT;
643 dd->sort_list[WRITE] = RB_ROOT;
644 dd->fifo_expire[READ] = read_expire;
645 dd->fifo_expire[WRITE] = write_expire;
646 dd->writes_starved = writes_starved;
647 dd->front_merges = 1;
648 dd->fifo_batch = fifo_batch;
652 static void deadline_put_request(request_queue_t *q, struct request *rq)
654 struct deadline_data *dd = q->elevator->elevator_data;
655 struct deadline_rq *drq = RQ_DATA(rq);
657 mempool_free(drq, dd->drq_pool);
658 rq->elevator_private = NULL;
662 deadline_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
665 struct deadline_data *dd = q->elevator->elevator_data;
666 struct deadline_rq *drq;
668 drq = mempool_alloc(dd->drq_pool, gfp_mask);
670 memset(drq, 0, sizeof(*drq));
671 RB_CLEAR(&drq->rb_node);
674 INIT_HLIST_NODE(&drq->hash);
676 INIT_LIST_HEAD(&drq->fifo);
678 rq->elevator_private = drq;
690 deadline_var_show(int var, char *page)
692 return sprintf(page, "%d\n", var);
696 deadline_var_store(int *var, const char *page, size_t count)
698 char *p = (char *) page;
700 *var = simple_strtol(p, &p, 10);
704 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
705 static ssize_t __FUNC(elevator_t *e, char *page) \
707 struct deadline_data *dd = e->elevator_data; \
708 int __data = __VAR; \
710 __data = jiffies_to_msecs(__data); \
711 return deadline_var_show(__data, (page)); \
713 SHOW_FUNCTION(deadline_read_expire_show, dd->fifo_expire[READ], 1);
714 SHOW_FUNCTION(deadline_write_expire_show, dd->fifo_expire[WRITE], 1);
715 SHOW_FUNCTION(deadline_writes_starved_show, dd->writes_starved, 0);
716 SHOW_FUNCTION(deadline_front_merges_show, dd->front_merges, 0);
717 SHOW_FUNCTION(deadline_fifo_batch_show, dd->fifo_batch, 0);
720 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
721 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
723 struct deadline_data *dd = e->elevator_data; \
725 int ret = deadline_var_store(&__data, (page), count); \
726 if (__data < (MIN)) \
728 else if (__data > (MAX)) \
731 *(__PTR) = msecs_to_jiffies(__data); \
736 STORE_FUNCTION(deadline_read_expire_store, &dd->fifo_expire[READ], 0, INT_MAX, 1);
737 STORE_FUNCTION(deadline_write_expire_store, &dd->fifo_expire[WRITE], 0, INT_MAX, 1);
738 STORE_FUNCTION(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX, 0);
739 STORE_FUNCTION(deadline_front_merges_store, &dd->front_merges, 0, 1, 0);
740 STORE_FUNCTION(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX, 0);
741 #undef STORE_FUNCTION
743 #define DD_ATTR(name) \
744 __ATTR(name, S_IRUGO|S_IWUSR, deadline_##name##_show, \
745 deadline_##name##_store)
747 static struct elv_fs_entry deadline_attrs[] = {
748 DD_ATTR(read_expire),
749 DD_ATTR(write_expire),
750 DD_ATTR(writes_starved),
751 DD_ATTR(front_merges),
756 static struct elevator_type iosched_deadline = {
758 .elevator_merge_fn = deadline_merge,
759 .elevator_merged_fn = deadline_merged_request,
760 .elevator_merge_req_fn = deadline_merged_requests,
761 .elevator_dispatch_fn = deadline_dispatch_requests,
762 .elevator_add_req_fn = deadline_add_request,
763 .elevator_queue_empty_fn = deadline_queue_empty,
764 .elevator_former_req_fn = deadline_former_request,
765 .elevator_latter_req_fn = deadline_latter_request,
766 .elevator_set_req_fn = deadline_set_request,
767 .elevator_put_req_fn = deadline_put_request,
768 .elevator_init_fn = deadline_init_queue,
769 .elevator_exit_fn = deadline_exit_queue,
772 .elevator_attrs = deadline_attrs,
773 .elevator_name = "deadline",
774 .elevator_owner = THIS_MODULE,
777 static int __init deadline_init(void)
781 drq_pool = kmem_cache_create("deadline_drq", sizeof(struct deadline_rq),
787 ret = elv_register(&iosched_deadline);
789 kmem_cache_destroy(drq_pool);
794 static void __exit deadline_exit(void)
796 kmem_cache_destroy(drq_pool);
797 elv_unregister(&iosched_deadline);
800 module_init(deadline_init);
801 module_exit(deadline_exit);
803 MODULE_AUTHOR("Jens Axboe");
804 MODULE_LICENSE("GPL");
805 MODULE_DESCRIPTION("deadline IO scheduler");