aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/block/cfq-iosched.c
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 18:20:36 -0400
committerLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 18:20:36 -0400
commit1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 (patch)
tree0bba044c4ce775e45a88a51686b5d9f90697ea9d /drivers/block/cfq-iosched.c
Linux-2.6.12-rc2v2.6.12-rc2
Initial git repository build. I'm not bothering with the full history, even though we have it. We can create a separate "historical" git archive of that later if we want to, and in the meantime it's about 3.2GB when imported into git - space that would just make the early git days unnecessarily complicated, when we don't have a lot of good infrastructure for it. Let it rip!
Diffstat (limited to 'drivers/block/cfq-iosched.c')
-rw-r--r--drivers/block/cfq-iosched.c1856
1 files changed, 1856 insertions, 0 deletions
diff --git a/drivers/block/cfq-iosched.c b/drivers/block/cfq-iosched.c
new file mode 100644
index 000000000000..0ef7a0065ece
--- /dev/null
+++ b/drivers/block/cfq-iosched.c
@@ -0,0 +1,1856 @@
1/*
2 * linux/drivers/block/cfq-iosched.c
3 *
4 * CFQ, or complete fairness queueing, disk scheduler.
5 *
6 * Based on ideas from a previously unfinished io
7 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
8 *
9 * Copyright (C) 2003 Jens Axboe <axboe@suse.de>
10 */
11#include <linux/kernel.h>
12#include <linux/fs.h>
13#include <linux/blkdev.h>
14#include <linux/elevator.h>
15#include <linux/bio.h>
16#include <linux/config.h>
17#include <linux/module.h>
18#include <linux/slab.h>
19#include <linux/init.h>
20#include <linux/compiler.h>
21#include <linux/hash.h>
22#include <linux/rbtree.h>
23#include <linux/mempool.h>
24
25static unsigned long max_elapsed_crq;
26static unsigned long max_elapsed_dispatch;
27
28/*
29 * tunables
30 */
31static int cfq_quantum = 4; /* max queue in one round of service */
32static int cfq_queued = 8; /* minimum rq allocate limit per-queue*/
33static int cfq_service = HZ; /* period over which service is avg */
34static int cfq_fifo_expire_r = HZ / 2; /* fifo timeout for sync requests */
35static int cfq_fifo_expire_w = 5 * HZ; /* fifo timeout for async requests */
36static int cfq_fifo_rate = HZ / 8; /* fifo expiry rate */
37static int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */
38static int cfq_back_penalty = 2; /* penalty of a backwards seek */
39
40/*
41 * for the hash of cfqq inside the cfqd
42 */
43#define CFQ_QHASH_SHIFT 6
44#define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
45#define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
46
47/*
48 * for the hash of crq inside the cfqq
49 */
50#define CFQ_MHASH_SHIFT 6
51#define CFQ_MHASH_BLOCK(sec) ((sec) >> 3)
52#define CFQ_MHASH_ENTRIES (1 << CFQ_MHASH_SHIFT)
53#define CFQ_MHASH_FN(sec) hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
54#define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
55#define list_entry_hash(ptr) hlist_entry((ptr), struct cfq_rq, hash)
56
57#define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
58
59#define RQ_DATA(rq) (rq)->elevator_private
60
61/*
62 * rb-tree defines
63 */
64#define RB_NONE (2)
65#define RB_EMPTY(node) ((node)->rb_node == NULL)
66#define RB_CLEAR_COLOR(node) (node)->rb_color = RB_NONE
67#define RB_CLEAR(node) do { \
68 (node)->rb_parent = NULL; \
69 RB_CLEAR_COLOR((node)); \
70 (node)->rb_right = NULL; \
71 (node)->rb_left = NULL; \
72} while (0)
73#define RB_CLEAR_ROOT(root) ((root)->rb_node = NULL)
74#define ON_RB(node) ((node)->rb_color != RB_NONE)
75#define rb_entry_crq(node) rb_entry((node), struct cfq_rq, rb_node)
76#define rq_rb_key(rq) (rq)->sector
77
78/*
79 * threshold for switching off non-tag accounting
80 */
81#define CFQ_MAX_TAG (4)
82
83/*
84 * sort key types and names
85 */
86enum {
87 CFQ_KEY_PGID,
88 CFQ_KEY_TGID,
89 CFQ_KEY_UID,
90 CFQ_KEY_GID,
91 CFQ_KEY_LAST,
92};
93
94static char *cfq_key_types[] = { "pgid", "tgid", "uid", "gid", NULL };
95
96static kmem_cache_t *crq_pool;
97static kmem_cache_t *cfq_pool;
98static kmem_cache_t *cfq_ioc_pool;
99
100struct cfq_data {
101 struct list_head rr_list;
102 struct list_head empty_list;
103
104 struct hlist_head *cfq_hash;
105 struct hlist_head *crq_hash;
106
107 /* queues on rr_list (ie they have pending requests */
108 unsigned int busy_queues;
109
110 unsigned int max_queued;
111
112 atomic_t ref;
113
114 int key_type;
115
116 mempool_t *crq_pool;
117
118 request_queue_t *queue;
119
120 sector_t last_sector;
121
122 int rq_in_driver;
123
124 /*
125 * tunables, see top of file
126 */
127 unsigned int cfq_quantum;
128 unsigned int cfq_queued;
129 unsigned int cfq_fifo_expire_r;
130 unsigned int cfq_fifo_expire_w;
131 unsigned int cfq_fifo_batch_expire;
132 unsigned int cfq_back_penalty;
133 unsigned int cfq_back_max;
134 unsigned int find_best_crq;
135
136 unsigned int cfq_tagged;
137};
138
139struct cfq_queue {
140 /* reference count */
141 atomic_t ref;
142 /* parent cfq_data */
143 struct cfq_data *cfqd;
144 /* hash of mergeable requests */
145 struct hlist_node cfq_hash;
146 /* hash key */
147 unsigned long key;
148 /* whether queue is on rr (or empty) list */
149 int on_rr;
150 /* on either rr or empty list of cfqd */
151 struct list_head cfq_list;
152 /* sorted list of pending requests */
153 struct rb_root sort_list;
154 /* if fifo isn't expired, next request to serve */
155 struct cfq_rq *next_crq;
156 /* requests queued in sort_list */
157 int queued[2];
158 /* currently allocated requests */
159 int allocated[2];
160 /* fifo list of requests in sort_list */
161 struct list_head fifo[2];
162 /* last time fifo expired */
163 unsigned long last_fifo_expire;
164
165 int key_type;
166
167 unsigned long service_start;
168 unsigned long service_used;
169
170 unsigned int max_rate;
171
172 /* number of requests that have been handed to the driver */
173 int in_flight;
174 /* number of currently allocated requests */
175 int alloc_limit[2];
176};
177
178struct cfq_rq {
179 struct rb_node rb_node;
180 sector_t rb_key;
181 struct request *request;
182 struct hlist_node hash;
183
184 struct cfq_queue *cfq_queue;
185 struct cfq_io_context *io_context;
186
187 unsigned long service_start;
188 unsigned long queue_start;
189
190 unsigned int in_flight : 1;
191 unsigned int accounted : 1;
192 unsigned int is_sync : 1;
193 unsigned int is_write : 1;
194};
195
196static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned long);
197static void cfq_dispatch_sort(request_queue_t *, struct cfq_rq *);
198static void cfq_update_next_crq(struct cfq_rq *);
199static void cfq_put_cfqd(struct cfq_data *cfqd);
200
201/*
202 * what the fairness is based on (ie how processes are grouped and
203 * differentiated)
204 */
205static inline unsigned long
206cfq_hash_key(struct cfq_data *cfqd, struct task_struct *tsk)
207{
208 /*
209 * optimize this so that ->key_type is the offset into the struct
210 */
211 switch (cfqd->key_type) {
212 case CFQ_KEY_PGID:
213 return process_group(tsk);
214 default:
215 case CFQ_KEY_TGID:
216 return tsk->tgid;
217 case CFQ_KEY_UID:
218 return tsk->uid;
219 case CFQ_KEY_GID:
220 return tsk->gid;
221 }
222}
223
224/*
225 * lots of deadline iosched dupes, can be abstracted later...
226 */
227static inline void cfq_del_crq_hash(struct cfq_rq *crq)
228{
229 hlist_del_init(&crq->hash);
230}
231
232static void cfq_remove_merge_hints(request_queue_t *q, struct cfq_rq *crq)
233{
234 cfq_del_crq_hash(crq);
235
236 if (q->last_merge == crq->request)
237 q->last_merge = NULL;
238
239 cfq_update_next_crq(crq);
240}
241
242static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
243{
244 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
245
246 BUG_ON(!hlist_unhashed(&crq->hash));
247
248 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
249}
250
251static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
252{
253 struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
254 struct hlist_node *entry, *next;
255
256 hlist_for_each_safe(entry, next, hash_list) {
257 struct cfq_rq *crq = list_entry_hash(entry);
258 struct request *__rq = crq->request;
259
260 BUG_ON(hlist_unhashed(&crq->hash));
261
262 if (!rq_mergeable(__rq)) {
263 cfq_del_crq_hash(crq);
264 continue;
265 }
266
267 if (rq_hash_key(__rq) == offset)
268 return __rq;
269 }
270
271 return NULL;
272}
273
274/*
275 * Lifted from AS - choose which of crq1 and crq2 that is best served now.
276 * We choose the request that is closest to the head right now. Distance
277 * behind the head are penalized and only allowed to a certain extent.
278 */
279static struct cfq_rq *
280cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
281{
282 sector_t last, s1, s2, d1 = 0, d2 = 0;
283 int r1_wrap = 0, r2_wrap = 0; /* requests are behind the disk head */
284 unsigned long back_max;
285
286 if (crq1 == NULL || crq1 == crq2)
287 return crq2;
288 if (crq2 == NULL)
289 return crq1;
290
291 s1 = crq1->request->sector;
292 s2 = crq2->request->sector;
293
294 last = cfqd->last_sector;
295
296#if 0
297 if (!list_empty(&cfqd->queue->queue_head)) {
298 struct list_head *entry = &cfqd->queue->queue_head;
299 unsigned long distance = ~0UL;
300 struct request *rq;
301
302 while ((entry = entry->prev) != &cfqd->queue->queue_head) {
303 rq = list_entry_rq(entry);
304
305 if (blk_barrier_rq(rq))
306 break;
307
308 if (distance < abs(s1 - rq->sector + rq->nr_sectors)) {
309 distance = abs(s1 - rq->sector +rq->nr_sectors);
310 last = rq->sector + rq->nr_sectors;
311 }
312 if (distance < abs(s2 - rq->sector + rq->nr_sectors)) {
313 distance = abs(s2 - rq->sector +rq->nr_sectors);
314 last = rq->sector + rq->nr_sectors;
315 }
316 }
317 }
318#endif
319
320 /*
321 * by definition, 1KiB is 2 sectors
322 */
323 back_max = cfqd->cfq_back_max * 2;
324
325 /*
326 * Strict one way elevator _except_ in the case where we allow
327 * short backward seeks which are biased as twice the cost of a
328 * similar forward seek.
329 */
330 if (s1 >= last)
331 d1 = s1 - last;
332 else if (s1 + back_max >= last)
333 d1 = (last - s1) * cfqd->cfq_back_penalty;
334 else
335 r1_wrap = 1;
336
337 if (s2 >= last)
338 d2 = s2 - last;
339 else if (s2 + back_max >= last)
340 d2 = (last - s2) * cfqd->cfq_back_penalty;
341 else
342 r2_wrap = 1;
343
344 /* Found required data */
345 if (!r1_wrap && r2_wrap)
346 return crq1;
347 else if (!r2_wrap && r1_wrap)
348 return crq2;
349 else if (r1_wrap && r2_wrap) {
350 /* both behind the head */
351 if (s1 <= s2)
352 return crq1;
353 else
354 return crq2;
355 }
356
357 /* Both requests in front of the head */
358 if (d1 < d2)
359 return crq1;
360 else if (d2 < d1)
361 return crq2;
362 else {
363 if (s1 >= s2)
364 return crq1;
365 else
366 return crq2;
367 }
368}
369
370/*
371 * would be nice to take fifo expire time into account as well
372 */
373static struct cfq_rq *
374cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
375 struct cfq_rq *last)
376{
377 struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
378 struct rb_node *rbnext, *rbprev;
379
380 if (!ON_RB(&last->rb_node))
381 return NULL;
382
383 if ((rbnext = rb_next(&last->rb_node)) == NULL)
384 rbnext = rb_first(&cfqq->sort_list);
385
386 rbprev = rb_prev(&last->rb_node);
387
388 if (rbprev)
389 crq_prev = rb_entry_crq(rbprev);
390 if (rbnext)
391 crq_next = rb_entry_crq(rbnext);
392
393 return cfq_choose_req(cfqd, crq_next, crq_prev);
394}
395
396static void cfq_update_next_crq(struct cfq_rq *crq)
397{
398 struct cfq_queue *cfqq = crq->cfq_queue;
399
400 if (cfqq->next_crq == crq)
401 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
402}
403
404static int cfq_check_sort_rr_list(struct cfq_queue *cfqq)
405{
406 struct list_head *head = &cfqq->cfqd->rr_list;
407 struct list_head *next, *prev;
408
409 /*
410 * list might still be ordered
411 */
412 next = cfqq->cfq_list.next;
413 if (next != head) {
414 struct cfq_queue *cnext = list_entry_cfqq(next);
415
416 if (cfqq->service_used > cnext->service_used)
417 return 1;
418 }
419
420 prev = cfqq->cfq_list.prev;
421 if (prev != head) {
422 struct cfq_queue *cprev = list_entry_cfqq(prev);
423
424 if (cfqq->service_used < cprev->service_used)
425 return 1;
426 }
427
428 return 0;
429}
430
431static void cfq_sort_rr_list(struct cfq_queue *cfqq, int new_queue)
432{
433 struct list_head *entry = &cfqq->cfqd->rr_list;
434
435 if (!cfqq->on_rr)
436 return;
437 if (!new_queue && !cfq_check_sort_rr_list(cfqq))
438 return;
439
440 list_del(&cfqq->cfq_list);
441
442 /*
443 * sort by our mean service_used, sub-sort by in-flight requests
444 */
445 while ((entry = entry->prev) != &cfqq->cfqd->rr_list) {
446 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
447
448 if (cfqq->service_used > __cfqq->service_used)
449 break;
450 else if (cfqq->service_used == __cfqq->service_used) {
451 struct list_head *prv;
452
453 while ((prv = entry->prev) != &cfqq->cfqd->rr_list) {
454 __cfqq = list_entry_cfqq(prv);
455
456 WARN_ON(__cfqq->service_used > cfqq->service_used);
457 if (cfqq->service_used != __cfqq->service_used)
458 break;
459 if (cfqq->in_flight > __cfqq->in_flight)
460 break;
461
462 entry = prv;
463 }
464 }
465 }
466
467 list_add(&cfqq->cfq_list, entry);
468}
469
470/*
471 * add to busy list of queues for service, trying to be fair in ordering
472 * the pending list according to requests serviced
473 */
474static inline void
475cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
476{
477 /*
478 * it's currently on the empty list
479 */
480 cfqq->on_rr = 1;
481 cfqd->busy_queues++;
482
483 if (time_after(jiffies, cfqq->service_start + cfq_service))
484 cfqq->service_used >>= 3;
485
486 cfq_sort_rr_list(cfqq, 1);
487}
488
489static inline void
490cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
491{
492 list_move(&cfqq->cfq_list, &cfqd->empty_list);
493 cfqq->on_rr = 0;
494
495 BUG_ON(!cfqd->busy_queues);
496 cfqd->busy_queues--;
497}
498
499/*
500 * rb tree support functions
501 */
502static inline void cfq_del_crq_rb(struct cfq_rq *crq)
503{
504 struct cfq_queue *cfqq = crq->cfq_queue;
505
506 if (ON_RB(&crq->rb_node)) {
507 struct cfq_data *cfqd = cfqq->cfqd;
508
509 BUG_ON(!cfqq->queued[crq->is_sync]);
510
511 cfq_update_next_crq(crq);
512
513 cfqq->queued[crq->is_sync]--;
514 rb_erase(&crq->rb_node, &cfqq->sort_list);
515 RB_CLEAR_COLOR(&crq->rb_node);
516
517 if (RB_EMPTY(&cfqq->sort_list) && cfqq->on_rr)
518 cfq_del_cfqq_rr(cfqd, cfqq);
519 }
520}
521
522static struct cfq_rq *
523__cfq_add_crq_rb(struct cfq_rq *crq)
524{
525 struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
526 struct rb_node *parent = NULL;
527 struct cfq_rq *__crq;
528
529 while (*p) {
530 parent = *p;
531 __crq = rb_entry_crq(parent);
532
533 if (crq->rb_key < __crq->rb_key)
534 p = &(*p)->rb_left;
535 else if (crq->rb_key > __crq->rb_key)
536 p = &(*p)->rb_right;
537 else
538 return __crq;
539 }
540
541 rb_link_node(&crq->rb_node, parent, p);
542 return NULL;
543}
544
545static void cfq_add_crq_rb(struct cfq_rq *crq)
546{
547 struct cfq_queue *cfqq = crq->cfq_queue;
548 struct cfq_data *cfqd = cfqq->cfqd;
549 struct request *rq = crq->request;
550 struct cfq_rq *__alias;
551
552 crq->rb_key = rq_rb_key(rq);
553 cfqq->queued[crq->is_sync]++;
554
555 /*
556 * looks a little odd, but the first insert might return an alias.
557 * if that happens, put the alias on the dispatch list
558 */
559 while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
560 cfq_dispatch_sort(cfqd->queue, __alias);
561
562 rb_insert_color(&crq->rb_node, &cfqq->sort_list);
563
564 if (!cfqq->on_rr)
565 cfq_add_cfqq_rr(cfqd, cfqq);
566
567 /*
568 * check if this request is a better next-serve candidate
569 */
570 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
571}
572
573static inline void
574cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
575{
576 if (ON_RB(&crq->rb_node)) {
577 rb_erase(&crq->rb_node, &cfqq->sort_list);
578 cfqq->queued[crq->is_sync]--;
579 }
580
581 cfq_add_crq_rb(crq);
582}
583
584static struct request *
585cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
586{
587 const unsigned long key = cfq_hash_key(cfqd, current);
588 struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, key);
589 struct rb_node *n;
590
591 if (!cfqq)
592 goto out;
593
594 n = cfqq->sort_list.rb_node;
595 while (n) {
596 struct cfq_rq *crq = rb_entry_crq(n);
597
598 if (sector < crq->rb_key)
599 n = n->rb_left;
600 else if (sector > crq->rb_key)
601 n = n->rb_right;
602 else
603 return crq->request;
604 }
605
606out:
607 return NULL;
608}
609
610static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
611{
612 struct cfq_rq *crq = RQ_DATA(rq);
613
614 if (crq) {
615 struct cfq_queue *cfqq = crq->cfq_queue;
616
617 if (cfqq->cfqd->cfq_tagged) {
618 cfqq->service_used--;
619 cfq_sort_rr_list(cfqq, 0);
620 }
621
622 if (crq->accounted) {
623 crq->accounted = 0;
624 cfqq->cfqd->rq_in_driver--;
625 }
626 }
627}
628
629/*
630 * make sure the service time gets corrected on reissue of this request
631 */
632static void cfq_requeue_request(request_queue_t *q, struct request *rq)
633{
634 cfq_deactivate_request(q, rq);
635 list_add(&rq->queuelist, &q->queue_head);
636}
637
638static void cfq_remove_request(request_queue_t *q, struct request *rq)
639{
640 struct cfq_rq *crq = RQ_DATA(rq);
641
642 if (crq) {
643 cfq_remove_merge_hints(q, crq);
644 list_del_init(&rq->queuelist);
645
646 if (crq->cfq_queue)
647 cfq_del_crq_rb(crq);
648 }
649}
650
651static int
652cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
653{
654 struct cfq_data *cfqd = q->elevator->elevator_data;
655 struct request *__rq;
656 int ret;
657
658 ret = elv_try_last_merge(q, bio);
659 if (ret != ELEVATOR_NO_MERGE) {
660 __rq = q->last_merge;
661 goto out_insert;
662 }
663
664 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
665 if (__rq) {
666 BUG_ON(__rq->sector + __rq->nr_sectors != bio->bi_sector);
667
668 if (elv_rq_merge_ok(__rq, bio)) {
669 ret = ELEVATOR_BACK_MERGE;
670 goto out;
671 }
672 }
673
674 __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
675 if (__rq) {
676 if (elv_rq_merge_ok(__rq, bio)) {
677 ret = ELEVATOR_FRONT_MERGE;
678 goto out;
679 }
680 }
681
682 return ELEVATOR_NO_MERGE;
683out:
684 q->last_merge = __rq;
685out_insert:
686 *req = __rq;
687 return ret;
688}
689
690static void cfq_merged_request(request_queue_t *q, struct request *req)
691{
692 struct cfq_data *cfqd = q->elevator->elevator_data;
693 struct cfq_rq *crq = RQ_DATA(req);
694
695 cfq_del_crq_hash(crq);
696 cfq_add_crq_hash(cfqd, crq);
697
698 if (ON_RB(&crq->rb_node) && (rq_rb_key(req) != crq->rb_key)) {
699 struct cfq_queue *cfqq = crq->cfq_queue;
700
701 cfq_update_next_crq(crq);
702 cfq_reposition_crq_rb(cfqq, crq);
703 }
704
705 q->last_merge = req;
706}
707
708static void
709cfq_merged_requests(request_queue_t *q, struct request *rq,
710 struct request *next)
711{
712 struct cfq_rq *crq = RQ_DATA(rq);
713 struct cfq_rq *cnext = RQ_DATA(next);
714
715 cfq_merged_request(q, rq);
716
717 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist)) {
718 if (time_before(cnext->queue_start, crq->queue_start)) {
719 list_move(&rq->queuelist, &next->queuelist);
720 crq->queue_start = cnext->queue_start;
721 }
722 }
723
724 cfq_update_next_crq(cnext);
725 cfq_remove_request(q, next);
726}
727
728/*
729 * we dispatch cfqd->cfq_quantum requests in total from the rr_list queues,
730 * this function sector sorts the selected request to minimize seeks. we start
731 * at cfqd->last_sector, not 0.
732 */
733static void cfq_dispatch_sort(request_queue_t *q, struct cfq_rq *crq)
734{
735 struct cfq_data *cfqd = q->elevator->elevator_data;
736 struct cfq_queue *cfqq = crq->cfq_queue;
737 struct list_head *head = &q->queue_head, *entry = head;
738 struct request *__rq;
739 sector_t last;
740
741 cfq_del_crq_rb(crq);
742 cfq_remove_merge_hints(q, crq);
743 list_del(&crq->request->queuelist);
744
745 last = cfqd->last_sector;
746 while ((entry = entry->prev) != head) {
747 __rq = list_entry_rq(entry);
748
749 if (blk_barrier_rq(crq->request))
750 break;
751 if (!blk_fs_request(crq->request))
752 break;
753
754 if (crq->request->sector > __rq->sector)
755 break;
756 if (__rq->sector > last && crq->request->sector < last) {
757 last = crq->request->sector;
758 break;
759 }
760 }
761
762 cfqd->last_sector = last;
763 crq->in_flight = 1;
764 cfqq->in_flight++;
765 list_add(&crq->request->queuelist, entry);
766}
767
768/*
769 * return expired entry, or NULL to just start from scratch in rbtree
770 */
771static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
772{
773 struct cfq_data *cfqd = cfqq->cfqd;
774 const int reads = !list_empty(&cfqq->fifo[0]);
775 const int writes = !list_empty(&cfqq->fifo[1]);
776 unsigned long now = jiffies;
777 struct cfq_rq *crq;
778
779 if (time_before(now, cfqq->last_fifo_expire + cfqd->cfq_fifo_batch_expire))
780 return NULL;
781
782 crq = RQ_DATA(list_entry(cfqq->fifo[0].next, struct request, queuelist));
783 if (reads && time_after(now, crq->queue_start + cfqd->cfq_fifo_expire_r)) {
784 cfqq->last_fifo_expire = now;
785 return crq;
786 }
787
788 crq = RQ_DATA(list_entry(cfqq->fifo[1].next, struct request, queuelist));
789 if (writes && time_after(now, crq->queue_start + cfqd->cfq_fifo_expire_w)) {
790 cfqq->last_fifo_expire = now;
791 return crq;
792 }
793
794 return NULL;
795}
796
797/*
798 * dispatch a single request from given queue
799 */
800static inline void
801cfq_dispatch_request(request_queue_t *q, struct cfq_data *cfqd,
802 struct cfq_queue *cfqq)
803{
804 struct cfq_rq *crq;
805
806 /*
807 * follow expired path, else get first next available
808 */
809 if ((crq = cfq_check_fifo(cfqq)) == NULL) {
810 if (cfqd->find_best_crq)
811 crq = cfqq->next_crq;
812 else
813 crq = rb_entry_crq(rb_first(&cfqq->sort_list));
814 }
815
816 cfqd->last_sector = crq->request->sector + crq->request->nr_sectors;
817
818 /*
819 * finally, insert request into driver list
820 */
821 cfq_dispatch_sort(q, crq);
822}
823
824static int cfq_dispatch_requests(request_queue_t *q, int max_dispatch)
825{
826 struct cfq_data *cfqd = q->elevator->elevator_data;
827 struct cfq_queue *cfqq;
828 struct list_head *entry, *tmp;
829 int queued, busy_queues, first_round;
830
831 if (list_empty(&cfqd->rr_list))
832 return 0;
833
834 queued = 0;
835 first_round = 1;
836restart:
837 busy_queues = 0;
838 list_for_each_safe(entry, tmp, &cfqd->rr_list) {
839 cfqq = list_entry_cfqq(entry);
840
841 BUG_ON(RB_EMPTY(&cfqq->sort_list));
842
843 /*
844 * first round of queueing, only select from queues that
845 * don't already have io in-flight
846 */
847 if (first_round && cfqq->in_flight)
848 continue;
849
850 cfq_dispatch_request(q, cfqd, cfqq);
851
852 if (!RB_EMPTY(&cfqq->sort_list))
853 busy_queues++;
854
855 queued++;
856 }
857
858 if ((queued < max_dispatch) && (busy_queues || first_round)) {
859 first_round = 0;
860 goto restart;
861 }
862
863 return queued;
864}
865
866static inline void cfq_account_dispatch(struct cfq_rq *crq)
867{
868 struct cfq_queue *cfqq = crq->cfq_queue;
869 struct cfq_data *cfqd = cfqq->cfqd;
870 unsigned long now, elapsed;
871
872 if (!blk_fs_request(crq->request))
873 return;
874
875 /*
876 * accounted bit is necessary since some drivers will call
877 * elv_next_request() many times for the same request (eg ide)
878 */
879 if (crq->accounted)
880 return;
881
882 now = jiffies;
883 if (cfqq->service_start == ~0UL)
884 cfqq->service_start = now;
885
886 /*
887 * on drives with tagged command queueing, command turn-around time
888 * doesn't necessarily reflect the time spent processing this very
889 * command inside the drive. so do the accounting differently there,
890 * by just sorting on the number of requests
891 */
892 if (cfqd->cfq_tagged) {
893 if (time_after(now, cfqq->service_start + cfq_service)) {
894 cfqq->service_start = now;
895 cfqq->service_used /= 10;
896 }
897
898 cfqq->service_used++;
899 cfq_sort_rr_list(cfqq, 0);
900 }
901
902 elapsed = now - crq->queue_start;
903 if (elapsed > max_elapsed_dispatch)
904 max_elapsed_dispatch = elapsed;
905
906 crq->accounted = 1;
907 crq->service_start = now;
908
909 if (++cfqd->rq_in_driver >= CFQ_MAX_TAG && !cfqd->cfq_tagged) {
910 cfqq->cfqd->cfq_tagged = 1;
911 printk("cfq: depth %d reached, tagging now on\n", CFQ_MAX_TAG);
912 }
913}
914
915static inline void
916cfq_account_completion(struct cfq_queue *cfqq, struct cfq_rq *crq)
917{
918 struct cfq_data *cfqd = cfqq->cfqd;
919
920 if (!crq->accounted)
921 return;
922
923 WARN_ON(!cfqd->rq_in_driver);
924 cfqd->rq_in_driver--;
925
926 if (!cfqd->cfq_tagged) {
927 unsigned long now = jiffies;
928 unsigned long duration = now - crq->service_start;
929
930 if (time_after(now, cfqq->service_start + cfq_service)) {
931 cfqq->service_start = now;
932 cfqq->service_used >>= 3;
933 }
934
935 cfqq->service_used += duration;
936 cfq_sort_rr_list(cfqq, 0);
937
938 if (duration > max_elapsed_crq)
939 max_elapsed_crq = duration;
940 }
941}
942
943static struct request *cfq_next_request(request_queue_t *q)
944{
945 struct cfq_data *cfqd = q->elevator->elevator_data;
946 struct request *rq;
947
948 if (!list_empty(&q->queue_head)) {
949 struct cfq_rq *crq;
950dispatch:
951 rq = list_entry_rq(q->queue_head.next);
952
953 if ((crq = RQ_DATA(rq)) != NULL) {
954 cfq_remove_merge_hints(q, crq);
955 cfq_account_dispatch(crq);
956 }
957
958 return rq;
959 }
960
961 if (cfq_dispatch_requests(q, cfqd->cfq_quantum))
962 goto dispatch;
963
964 return NULL;
965}
966
967/*
968 * task holds one reference to the queue, dropped when task exits. each crq
969 * in-flight on this queue also holds a reference, dropped when crq is freed.
970 *
971 * queue lock must be held here.
972 */
973static void cfq_put_queue(struct cfq_queue *cfqq)
974{
975 BUG_ON(!atomic_read(&cfqq->ref));
976
977 if (!atomic_dec_and_test(&cfqq->ref))
978 return;
979
980 BUG_ON(rb_first(&cfqq->sort_list));
981 BUG_ON(cfqq->on_rr);
982
983 cfq_put_cfqd(cfqq->cfqd);
984
985 /*
986 * it's on the empty list and still hashed
987 */
988 list_del(&cfqq->cfq_list);
989 hlist_del(&cfqq->cfq_hash);
990 kmem_cache_free(cfq_pool, cfqq);
991}
992
993static inline struct cfq_queue *
994__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned long key, const int hashval)
995{
996 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
997 struct hlist_node *entry, *next;
998
999 hlist_for_each_safe(entry, next, hash_list) {
1000 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1001
1002 if (__cfqq->key == key)
1003 return __cfqq;
1004 }
1005
1006 return NULL;
1007}
1008
1009static struct cfq_queue *
1010cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned long key)
1011{
1012 return __cfq_find_cfq_hash(cfqd, key, hash_long(key, CFQ_QHASH_SHIFT));
1013}
1014
1015static inline void
1016cfq_rehash_cfqq(struct cfq_data *cfqd, struct cfq_queue **cfqq,
1017 struct cfq_io_context *cic)
1018{
1019 unsigned long hashkey = cfq_hash_key(cfqd, current);
1020 unsigned long hashval = hash_long(hashkey, CFQ_QHASH_SHIFT);
1021 struct cfq_queue *__cfqq;
1022 unsigned long flags;
1023
1024 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1025
1026 hlist_del(&(*cfqq)->cfq_hash);
1027
1028 __cfqq = __cfq_find_cfq_hash(cfqd, hashkey, hashval);
1029 if (!__cfqq || __cfqq == *cfqq) {
1030 __cfqq = *cfqq;
1031 hlist_add_head(&__cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1032 __cfqq->key_type = cfqd->key_type;
1033 } else {
1034 atomic_inc(&__cfqq->ref);
1035 cic->cfqq = __cfqq;
1036 cfq_put_queue(*cfqq);
1037 *cfqq = __cfqq;
1038 }
1039
1040 cic->cfqq = __cfqq;
1041 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1042}
1043
1044static void cfq_free_io_context(struct cfq_io_context *cic)
1045{
1046 kmem_cache_free(cfq_ioc_pool, cic);
1047}
1048
1049/*
1050 * locking hierarchy is: io_context lock -> queue locks
1051 */
1052static void cfq_exit_io_context(struct cfq_io_context *cic)
1053{
1054 struct cfq_queue *cfqq = cic->cfqq;
1055 struct list_head *entry = &cic->list;
1056 request_queue_t *q;
1057 unsigned long flags;
1058
1059 /*
1060 * put the reference this task is holding to the various queues
1061 */
1062 spin_lock_irqsave(&cic->ioc->lock, flags);
1063 while ((entry = cic->list.next) != &cic->list) {
1064 struct cfq_io_context *__cic;
1065
1066 __cic = list_entry(entry, struct cfq_io_context, list);
1067 list_del(entry);
1068
1069 q = __cic->cfqq->cfqd->queue;
1070 spin_lock(q->queue_lock);
1071 cfq_put_queue(__cic->cfqq);
1072 spin_unlock(q->queue_lock);
1073 }
1074
1075 q = cfqq->cfqd->queue;
1076 spin_lock(q->queue_lock);
1077 cfq_put_queue(cfqq);
1078 spin_unlock(q->queue_lock);
1079
1080 cic->cfqq = NULL;
1081 spin_unlock_irqrestore(&cic->ioc->lock, flags);
1082}
1083
1084static struct cfq_io_context *cfq_alloc_io_context(int gfp_flags)
1085{
1086 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_flags);
1087
1088 if (cic) {
1089 cic->dtor = cfq_free_io_context;
1090 cic->exit = cfq_exit_io_context;
1091 INIT_LIST_HEAD(&cic->list);
1092 cic->cfqq = NULL;
1093 }
1094
1095 return cic;
1096}
1097
1098/*
1099 * Setup general io context and cfq io context. There can be several cfq
1100 * io contexts per general io context, if this process is doing io to more
1101 * than one device managed by cfq. Note that caller is holding a reference to
1102 * cfqq, so we don't need to worry about it disappearing
1103 */
1104static struct cfq_io_context *
1105cfq_get_io_context(struct cfq_queue **cfqq, int gfp_flags)
1106{
1107 struct cfq_data *cfqd = (*cfqq)->cfqd;
1108 struct cfq_queue *__cfqq = *cfqq;
1109 struct cfq_io_context *cic;
1110 struct io_context *ioc;
1111
1112 might_sleep_if(gfp_flags & __GFP_WAIT);
1113
1114 ioc = get_io_context(gfp_flags);
1115 if (!ioc)
1116 return NULL;
1117
1118 if ((cic = ioc->cic) == NULL) {
1119 cic = cfq_alloc_io_context(gfp_flags);
1120
1121 if (cic == NULL)
1122 goto err;
1123
1124 ioc->cic = cic;
1125 cic->ioc = ioc;
1126 cic->cfqq = __cfqq;
1127 atomic_inc(&__cfqq->ref);
1128 } else {
1129 struct cfq_io_context *__cic;
1130 unsigned long flags;
1131
1132 /*
1133 * since the first cic on the list is actually the head
1134 * itself, need to check this here or we'll duplicate an
1135 * cic per ioc for no reason
1136 */
1137 if (cic->cfqq == __cfqq)
1138 goto out;
1139
1140 /*
1141 * cic exists, check if we already are there. linear search
1142 * should be ok here, the list will usually not be more than
1143 * 1 or a few entries long
1144 */
1145 spin_lock_irqsave(&ioc->lock, flags);
1146 list_for_each_entry(__cic, &cic->list, list) {
1147 /*
1148 * this process is already holding a reference to
1149 * this queue, so no need to get one more
1150 */
1151 if (__cic->cfqq == __cfqq) {
1152 cic = __cic;
1153 spin_unlock_irqrestore(&ioc->lock, flags);
1154 goto out;
1155 }
1156 }
1157 spin_unlock_irqrestore(&ioc->lock, flags);
1158
1159 /*
1160 * nope, process doesn't have a cic assoicated with this
1161 * cfqq yet. get a new one and add to list
1162 */
1163 __cic = cfq_alloc_io_context(gfp_flags);
1164 if (__cic == NULL)
1165 goto err;
1166
1167 __cic->ioc = ioc;
1168 __cic->cfqq = __cfqq;
1169 atomic_inc(&__cfqq->ref);
1170 spin_lock_irqsave(&ioc->lock, flags);
1171 list_add(&__cic->list, &cic->list);
1172 spin_unlock_irqrestore(&ioc->lock, flags);
1173
1174 cic = __cic;
1175 *cfqq = __cfqq;
1176 }
1177
1178out:
1179 /*
1180 * if key_type has been changed on the fly, we lazily rehash
1181 * each queue at lookup time
1182 */
1183 if ((*cfqq)->key_type != cfqd->key_type)
1184 cfq_rehash_cfqq(cfqd, cfqq, cic);
1185
1186 return cic;
1187err:
1188 put_io_context(ioc);
1189 return NULL;
1190}
1191
1192static struct cfq_queue *
1193__cfq_get_queue(struct cfq_data *cfqd, unsigned long key, int gfp_mask)
1194{
1195 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1196 struct cfq_queue *cfqq, *new_cfqq = NULL;
1197
1198retry:
1199 cfqq = __cfq_find_cfq_hash(cfqd, key, hashval);
1200
1201 if (!cfqq) {
1202 if (new_cfqq) {
1203 cfqq = new_cfqq;
1204 new_cfqq = NULL;
1205 } else if (gfp_mask & __GFP_WAIT) {
1206 spin_unlock_irq(cfqd->queue->queue_lock);
1207 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1208 spin_lock_irq(cfqd->queue->queue_lock);
1209 goto retry;
1210 } else
1211 goto out;
1212
1213 memset(cfqq, 0, sizeof(*cfqq));
1214
1215 INIT_HLIST_NODE(&cfqq->cfq_hash);
1216 INIT_LIST_HEAD(&cfqq->cfq_list);
1217 RB_CLEAR_ROOT(&cfqq->sort_list);
1218 INIT_LIST_HEAD(&cfqq->fifo[0]);
1219 INIT_LIST_HEAD(&cfqq->fifo[1]);
1220
1221 cfqq->key = key;
1222 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1223 atomic_set(&cfqq->ref, 0);
1224 cfqq->cfqd = cfqd;
1225 atomic_inc(&cfqd->ref);
1226 cfqq->key_type = cfqd->key_type;
1227 cfqq->service_start = ~0UL;
1228 }
1229
1230 if (new_cfqq)
1231 kmem_cache_free(cfq_pool, new_cfqq);
1232
1233 atomic_inc(&cfqq->ref);
1234out:
1235 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1236 return cfqq;
1237}
1238
1239static void cfq_enqueue(struct cfq_data *cfqd, struct cfq_rq *crq)
1240{
1241 crq->is_sync = 0;
1242 if (rq_data_dir(crq->request) == READ || current->flags & PF_SYNCWRITE)
1243 crq->is_sync = 1;
1244
1245 cfq_add_crq_rb(crq);
1246 crq->queue_start = jiffies;
1247
1248 list_add_tail(&crq->request->queuelist, &crq->cfq_queue->fifo[crq->is_sync]);
1249}
1250
1251static void
1252cfq_insert_request(request_queue_t *q, struct request *rq, int where)
1253{
1254 struct cfq_data *cfqd = q->elevator->elevator_data;
1255 struct cfq_rq *crq = RQ_DATA(rq);
1256
1257 switch (where) {
1258 case ELEVATOR_INSERT_BACK:
1259 while (cfq_dispatch_requests(q, cfqd->cfq_quantum))
1260 ;
1261 list_add_tail(&rq->queuelist, &q->queue_head);
1262 break;
1263 case ELEVATOR_INSERT_FRONT:
1264 list_add(&rq->queuelist, &q->queue_head);
1265 break;
1266 case ELEVATOR_INSERT_SORT:
1267 BUG_ON(!blk_fs_request(rq));
1268 cfq_enqueue(cfqd, crq);
1269 break;
1270 default:
1271 printk("%s: bad insert point %d\n", __FUNCTION__,where);
1272 return;
1273 }
1274
1275 if (rq_mergeable(rq)) {
1276 cfq_add_crq_hash(cfqd, crq);
1277
1278 if (!q->last_merge)
1279 q->last_merge = rq;
1280 }
1281}
1282
1283static int cfq_queue_empty(request_queue_t *q)
1284{
1285 struct cfq_data *cfqd = q->elevator->elevator_data;
1286
1287 return list_empty(&q->queue_head) && list_empty(&cfqd->rr_list);
1288}
1289
1290static void cfq_completed_request(request_queue_t *q, struct request *rq)
1291{
1292 struct cfq_rq *crq = RQ_DATA(rq);
1293 struct cfq_queue *cfqq;
1294
1295 if (unlikely(!blk_fs_request(rq)))
1296 return;
1297
1298 cfqq = crq->cfq_queue;
1299
1300 if (crq->in_flight) {
1301 WARN_ON(!cfqq->in_flight);
1302 cfqq->in_flight--;
1303 }
1304
1305 cfq_account_completion(cfqq, crq);
1306}
1307
1308static struct request *
1309cfq_former_request(request_queue_t *q, struct request *rq)
1310{
1311 struct cfq_rq *crq = RQ_DATA(rq);
1312 struct rb_node *rbprev = rb_prev(&crq->rb_node);
1313
1314 if (rbprev)
1315 return rb_entry_crq(rbprev)->request;
1316
1317 return NULL;
1318}
1319
1320static struct request *
1321cfq_latter_request(request_queue_t *q, struct request *rq)
1322{
1323 struct cfq_rq *crq = RQ_DATA(rq);
1324 struct rb_node *rbnext = rb_next(&crq->rb_node);
1325
1326 if (rbnext)
1327 return rb_entry_crq(rbnext)->request;
1328
1329 return NULL;
1330}
1331
1332static int cfq_may_queue(request_queue_t *q, int rw)
1333{
1334 struct cfq_data *cfqd = q->elevator->elevator_data;
1335 struct cfq_queue *cfqq;
1336 int ret = ELV_MQUEUE_MAY;
1337
1338 if (current->flags & PF_MEMALLOC)
1339 return ELV_MQUEUE_MAY;
1340
1341 cfqq = cfq_find_cfq_hash(cfqd, cfq_hash_key(cfqd, current));
1342 if (cfqq) {
1343 int limit = cfqd->max_queued;
1344
1345 if (cfqq->allocated[rw] < cfqd->cfq_queued)
1346 return ELV_MQUEUE_MUST;
1347
1348 if (cfqd->busy_queues)
1349 limit = q->nr_requests / cfqd->busy_queues;
1350
1351 if (limit < cfqd->cfq_queued)
1352 limit = cfqd->cfq_queued;
1353 else if (limit > cfqd->max_queued)
1354 limit = cfqd->max_queued;
1355
1356 if (cfqq->allocated[rw] >= limit) {
1357 if (limit > cfqq->alloc_limit[rw])
1358 cfqq->alloc_limit[rw] = limit;
1359
1360 ret = ELV_MQUEUE_NO;
1361 }
1362 }
1363
1364 return ret;
1365}
1366
1367static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1368{
1369 struct request_list *rl = &q->rq;
1370 const int write = waitqueue_active(&rl->wait[WRITE]);
1371 const int read = waitqueue_active(&rl->wait[READ]);
1372
1373 if (read && cfqq->allocated[READ] < cfqq->alloc_limit[READ])
1374 wake_up(&rl->wait[READ]);
1375 if (write && cfqq->allocated[WRITE] < cfqq->alloc_limit[WRITE])
1376 wake_up(&rl->wait[WRITE]);
1377}
1378
1379/*
1380 * queue lock held here
1381 */
1382static void cfq_put_request(request_queue_t *q, struct request *rq)
1383{
1384 struct cfq_data *cfqd = q->elevator->elevator_data;
1385 struct cfq_rq *crq = RQ_DATA(rq);
1386
1387 if (crq) {
1388 struct cfq_queue *cfqq = crq->cfq_queue;
1389
1390 BUG_ON(q->last_merge == rq);
1391 BUG_ON(!hlist_unhashed(&crq->hash));
1392
1393 if (crq->io_context)
1394 put_io_context(crq->io_context->ioc);
1395
1396 BUG_ON(!cfqq->allocated[crq->is_write]);
1397 cfqq->allocated[crq->is_write]--;
1398
1399 mempool_free(crq, cfqd->crq_pool);
1400 rq->elevator_private = NULL;
1401
1402 smp_mb();
1403 cfq_check_waiters(q, cfqq);
1404 cfq_put_queue(cfqq);
1405 }
1406}
1407
1408/*
1409 * Allocate cfq data structures associated with this request. A queue and
1410 */
1411static int cfq_set_request(request_queue_t *q, struct request *rq, int gfp_mask)
1412{
1413 struct cfq_data *cfqd = q->elevator->elevator_data;
1414 struct cfq_io_context *cic;
1415 const int rw = rq_data_dir(rq);
1416 struct cfq_queue *cfqq, *saved_cfqq;
1417 struct cfq_rq *crq;
1418 unsigned long flags;
1419
1420 might_sleep_if(gfp_mask & __GFP_WAIT);
1421
1422 spin_lock_irqsave(q->queue_lock, flags);
1423
1424 cfqq = __cfq_get_queue(cfqd, cfq_hash_key(cfqd, current), gfp_mask);
1425 if (!cfqq)
1426 goto out_lock;
1427
1428repeat:
1429 if (cfqq->allocated[rw] >= cfqd->max_queued)
1430 goto out_lock;
1431
1432 cfqq->allocated[rw]++;
1433 spin_unlock_irqrestore(q->queue_lock, flags);
1434
1435 /*
1436 * if hashing type has changed, the cfq_queue might change here.
1437 */
1438 saved_cfqq = cfqq;
1439 cic = cfq_get_io_context(&cfqq, gfp_mask);
1440 if (!cic)
1441 goto err;
1442
1443 /*
1444 * repeat allocation checks on queue change
1445 */
1446 if (unlikely(saved_cfqq != cfqq)) {
1447 spin_lock_irqsave(q->queue_lock, flags);
1448 saved_cfqq->allocated[rw]--;
1449 goto repeat;
1450 }
1451
1452 crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1453 if (crq) {
1454 RB_CLEAR(&crq->rb_node);
1455 crq->rb_key = 0;
1456 crq->request = rq;
1457 INIT_HLIST_NODE(&crq->hash);
1458 crq->cfq_queue = cfqq;
1459 crq->io_context = cic;
1460 crq->service_start = crq->queue_start = 0;
1461 crq->in_flight = crq->accounted = crq->is_sync = 0;
1462 crq->is_write = rw;
1463 rq->elevator_private = crq;
1464 cfqq->alloc_limit[rw] = 0;
1465 return 0;
1466 }
1467
1468 put_io_context(cic->ioc);
1469err:
1470 spin_lock_irqsave(q->queue_lock, flags);
1471 cfqq->allocated[rw]--;
1472 cfq_put_queue(cfqq);
1473out_lock:
1474 spin_unlock_irqrestore(q->queue_lock, flags);
1475 return 1;
1476}
1477
1478static void cfq_put_cfqd(struct cfq_data *cfqd)
1479{
1480 request_queue_t *q = cfqd->queue;
1481
1482 if (!atomic_dec_and_test(&cfqd->ref))
1483 return;
1484
1485 blk_put_queue(q);
1486
1487 mempool_destroy(cfqd->crq_pool);
1488 kfree(cfqd->crq_hash);
1489 kfree(cfqd->cfq_hash);
1490 kfree(cfqd);
1491}
1492
1493static void cfq_exit_queue(elevator_t *e)
1494{
1495 cfq_put_cfqd(e->elevator_data);
1496}
1497
1498static int cfq_init_queue(request_queue_t *q, elevator_t *e)
1499{
1500 struct cfq_data *cfqd;
1501 int i;
1502
1503 cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
1504 if (!cfqd)
1505 return -ENOMEM;
1506
1507 memset(cfqd, 0, sizeof(*cfqd));
1508 INIT_LIST_HEAD(&cfqd->rr_list);
1509 INIT_LIST_HEAD(&cfqd->empty_list);
1510
1511 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
1512 if (!cfqd->crq_hash)
1513 goto out_crqhash;
1514
1515 cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
1516 if (!cfqd->cfq_hash)
1517 goto out_cfqhash;
1518
1519 cfqd->crq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, crq_pool);
1520 if (!cfqd->crq_pool)
1521 goto out_crqpool;
1522
1523 for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
1524 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
1525 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
1526 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
1527
1528 e->elevator_data = cfqd;
1529
1530 cfqd->queue = q;
1531 atomic_inc(&q->refcnt);
1532
1533 /*
1534 * just set it to some high value, we want anyone to be able to queue
1535 * some requests. fairness is handled differently
1536 */
1537 q->nr_requests = 1024;
1538 cfqd->max_queued = q->nr_requests / 16;
1539 q->nr_batching = cfq_queued;
1540 cfqd->key_type = CFQ_KEY_TGID;
1541 cfqd->find_best_crq = 1;
1542 atomic_set(&cfqd->ref, 1);
1543
1544 cfqd->cfq_queued = cfq_queued;
1545 cfqd->cfq_quantum = cfq_quantum;
1546 cfqd->cfq_fifo_expire_r = cfq_fifo_expire_r;
1547 cfqd->cfq_fifo_expire_w = cfq_fifo_expire_w;
1548 cfqd->cfq_fifo_batch_expire = cfq_fifo_rate;
1549 cfqd->cfq_back_max = cfq_back_max;
1550 cfqd->cfq_back_penalty = cfq_back_penalty;
1551
1552 return 0;
1553out_crqpool:
1554 kfree(cfqd->cfq_hash);
1555out_cfqhash:
1556 kfree(cfqd->crq_hash);
1557out_crqhash:
1558 kfree(cfqd);
1559 return -ENOMEM;
1560}
1561
1562static void cfq_slab_kill(void)
1563{
1564 if (crq_pool)
1565 kmem_cache_destroy(crq_pool);
1566 if (cfq_pool)
1567 kmem_cache_destroy(cfq_pool);
1568 if (cfq_ioc_pool)
1569 kmem_cache_destroy(cfq_ioc_pool);
1570}
1571
1572static int __init cfq_slab_setup(void)
1573{
1574 crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
1575 NULL, NULL);
1576 if (!crq_pool)
1577 goto fail;
1578
1579 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
1580 NULL, NULL);
1581 if (!cfq_pool)
1582 goto fail;
1583
1584 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
1585 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
1586 if (!cfq_ioc_pool)
1587 goto fail;
1588
1589 return 0;
1590fail:
1591 cfq_slab_kill();
1592 return -ENOMEM;
1593}
1594
1595
1596/*
1597 * sysfs parts below -->
1598 */
1599struct cfq_fs_entry {
1600 struct attribute attr;
1601 ssize_t (*show)(struct cfq_data *, char *);
1602 ssize_t (*store)(struct cfq_data *, const char *, size_t);
1603};
1604
1605static ssize_t
1606cfq_var_show(unsigned int var, char *page)
1607{
1608 return sprintf(page, "%d\n", var);
1609}
1610
1611static ssize_t
1612cfq_var_store(unsigned int *var, const char *page, size_t count)
1613{
1614 char *p = (char *) page;
1615
1616 *var = simple_strtoul(p, &p, 10);
1617 return count;
1618}
1619
1620static ssize_t
1621cfq_clear_elapsed(struct cfq_data *cfqd, const char *page, size_t count)
1622{
1623 max_elapsed_dispatch = max_elapsed_crq = 0;
1624 return count;
1625}
1626
1627static ssize_t
1628cfq_set_key_type(struct cfq_data *cfqd, const char *page, size_t count)
1629{
1630 spin_lock_irq(cfqd->queue->queue_lock);
1631 if (!strncmp(page, "pgid", 4))
1632 cfqd->key_type = CFQ_KEY_PGID;
1633 else if (!strncmp(page, "tgid", 4))
1634 cfqd->key_type = CFQ_KEY_TGID;
1635 else if (!strncmp(page, "uid", 3))
1636 cfqd->key_type = CFQ_KEY_UID;
1637 else if (!strncmp(page, "gid", 3))
1638 cfqd->key_type = CFQ_KEY_GID;
1639 spin_unlock_irq(cfqd->queue->queue_lock);
1640 return count;
1641}
1642
1643static ssize_t
1644cfq_read_key_type(struct cfq_data *cfqd, char *page)
1645{
1646 ssize_t len = 0;
1647 int i;
1648
1649 for (i = CFQ_KEY_PGID; i < CFQ_KEY_LAST; i++) {
1650 if (cfqd->key_type == i)
1651 len += sprintf(page+len, "[%s] ", cfq_key_types[i]);
1652 else
1653 len += sprintf(page+len, "%s ", cfq_key_types[i]);
1654 }
1655 len += sprintf(page+len, "\n");
1656 return len;
1657}
1658
1659#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
1660static ssize_t __FUNC(struct cfq_data *cfqd, char *page) \
1661{ \
1662 unsigned int __data = __VAR; \
1663 if (__CONV) \
1664 __data = jiffies_to_msecs(__data); \
1665 return cfq_var_show(__data, (page)); \
1666}
1667SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
1668SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
1669SHOW_FUNCTION(cfq_fifo_expire_r_show, cfqd->cfq_fifo_expire_r, 1);
1670SHOW_FUNCTION(cfq_fifo_expire_w_show, cfqd->cfq_fifo_expire_w, 1);
1671SHOW_FUNCTION(cfq_fifo_batch_expire_show, cfqd->cfq_fifo_batch_expire, 1);
1672SHOW_FUNCTION(cfq_find_best_show, cfqd->find_best_crq, 0);
1673SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
1674SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
1675#undef SHOW_FUNCTION
1676
1677#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
1678static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count) \
1679{ \
1680 unsigned int __data; \
1681 int ret = cfq_var_store(&__data, (page), count); \
1682 if (__data < (MIN)) \
1683 __data = (MIN); \
1684 else if (__data > (MAX)) \
1685 __data = (MAX); \
1686 if (__CONV) \
1687 *(__PTR) = msecs_to_jiffies(__data); \
1688 else \
1689 *(__PTR) = __data; \
1690 return ret; \
1691}
1692STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
1693STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
1694STORE_FUNCTION(cfq_fifo_expire_r_store, &cfqd->cfq_fifo_expire_r, 1, UINT_MAX, 1);
1695STORE_FUNCTION(cfq_fifo_expire_w_store, &cfqd->cfq_fifo_expire_w, 1, UINT_MAX, 1);
1696STORE_FUNCTION(cfq_fifo_batch_expire_store, &cfqd->cfq_fifo_batch_expire, 0, UINT_MAX, 1);
1697STORE_FUNCTION(cfq_find_best_store, &cfqd->find_best_crq, 0, 1, 0);
1698STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
1699STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
1700#undef STORE_FUNCTION
1701
1702static struct cfq_fs_entry cfq_quantum_entry = {
1703 .attr = {.name = "quantum", .mode = S_IRUGO | S_IWUSR },
1704 .show = cfq_quantum_show,
1705 .store = cfq_quantum_store,
1706};
1707static struct cfq_fs_entry cfq_queued_entry = {
1708 .attr = {.name = "queued", .mode = S_IRUGO | S_IWUSR },
1709 .show = cfq_queued_show,
1710 .store = cfq_queued_store,
1711};
1712static struct cfq_fs_entry cfq_fifo_expire_r_entry = {
1713 .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
1714 .show = cfq_fifo_expire_r_show,
1715 .store = cfq_fifo_expire_r_store,
1716};
1717static struct cfq_fs_entry cfq_fifo_expire_w_entry = {
1718 .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
1719 .show = cfq_fifo_expire_w_show,
1720 .store = cfq_fifo_expire_w_store,
1721};
1722static struct cfq_fs_entry cfq_fifo_batch_expire_entry = {
1723 .attr = {.name = "fifo_batch_expire", .mode = S_IRUGO | S_IWUSR },
1724 .show = cfq_fifo_batch_expire_show,
1725 .store = cfq_fifo_batch_expire_store,
1726};
1727static struct cfq_fs_entry cfq_find_best_entry = {
1728 .attr = {.name = "find_best_crq", .mode = S_IRUGO | S_IWUSR },
1729 .show = cfq_find_best_show,
1730 .store = cfq_find_best_store,
1731};
1732static struct cfq_fs_entry cfq_back_max_entry = {
1733 .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
1734 .show = cfq_back_max_show,
1735 .store = cfq_back_max_store,
1736};
1737static struct cfq_fs_entry cfq_back_penalty_entry = {
1738 .attr = {.name = "back_seek_penalty", .mode = S_IRUGO | S_IWUSR },
1739 .show = cfq_back_penalty_show,
1740 .store = cfq_back_penalty_store,
1741};
1742static struct cfq_fs_entry cfq_clear_elapsed_entry = {
1743 .attr = {.name = "clear_elapsed", .mode = S_IWUSR },
1744 .store = cfq_clear_elapsed,
1745};
1746static struct cfq_fs_entry cfq_key_type_entry = {
1747 .attr = {.name = "key_type", .mode = S_IRUGO | S_IWUSR },
1748 .show = cfq_read_key_type,
1749 .store = cfq_set_key_type,
1750};
1751
1752static struct attribute *default_attrs[] = {
1753 &cfq_quantum_entry.attr,
1754 &cfq_queued_entry.attr,
1755 &cfq_fifo_expire_r_entry.attr,
1756 &cfq_fifo_expire_w_entry.attr,
1757 &cfq_fifo_batch_expire_entry.attr,
1758 &cfq_key_type_entry.attr,
1759 &cfq_find_best_entry.attr,
1760 &cfq_back_max_entry.attr,
1761 &cfq_back_penalty_entry.attr,
1762 &cfq_clear_elapsed_entry.attr,
1763 NULL,
1764};
1765
1766#define to_cfq(atr) container_of((atr), struct cfq_fs_entry, attr)
1767
1768static ssize_t
1769cfq_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
1770{
1771 elevator_t *e = container_of(kobj, elevator_t, kobj);
1772 struct cfq_fs_entry *entry = to_cfq(attr);
1773
1774 if (!entry->show)
1775 return 0;
1776
1777 return entry->show(e->elevator_data, page);
1778}
1779
1780static ssize_t
1781cfq_attr_store(struct kobject *kobj, struct attribute *attr,
1782 const char *page, size_t length)
1783{
1784 elevator_t *e = container_of(kobj, elevator_t, kobj);
1785 struct cfq_fs_entry *entry = to_cfq(attr);
1786
1787 if (!entry->store)
1788 return -EINVAL;
1789
1790 return entry->store(e->elevator_data, page, length);
1791}
1792
1793static struct sysfs_ops cfq_sysfs_ops = {
1794 .show = cfq_attr_show,
1795 .store = cfq_attr_store,
1796};
1797
1798static struct kobj_type cfq_ktype = {
1799 .sysfs_ops = &cfq_sysfs_ops,
1800 .default_attrs = default_attrs,
1801};
1802
1803static struct elevator_type iosched_cfq = {
1804 .ops = {
1805 .elevator_merge_fn = cfq_merge,
1806 .elevator_merged_fn = cfq_merged_request,
1807 .elevator_merge_req_fn = cfq_merged_requests,
1808 .elevator_next_req_fn = cfq_next_request,
1809 .elevator_add_req_fn = cfq_insert_request,
1810 .elevator_remove_req_fn = cfq_remove_request,
1811 .elevator_requeue_req_fn = cfq_requeue_request,
1812 .elevator_deactivate_req_fn = cfq_deactivate_request,
1813 .elevator_queue_empty_fn = cfq_queue_empty,
1814 .elevator_completed_req_fn = cfq_completed_request,
1815 .elevator_former_req_fn = cfq_former_request,
1816 .elevator_latter_req_fn = cfq_latter_request,
1817 .elevator_set_req_fn = cfq_set_request,
1818 .elevator_put_req_fn = cfq_put_request,
1819 .elevator_may_queue_fn = cfq_may_queue,
1820 .elevator_init_fn = cfq_init_queue,
1821 .elevator_exit_fn = cfq_exit_queue,
1822 },
1823 .elevator_ktype = &cfq_ktype,
1824 .elevator_name = "cfq",
1825 .elevator_owner = THIS_MODULE,
1826};
1827
1828static int __init cfq_init(void)
1829{
1830 int ret;
1831
1832 if (cfq_slab_setup())
1833 return -ENOMEM;
1834
1835 ret = elv_register(&iosched_cfq);
1836 if (!ret) {
1837 __module_get(THIS_MODULE);
1838 return 0;
1839 }
1840
1841 cfq_slab_kill();
1842 return ret;
1843}
1844
1845static void __exit cfq_exit(void)
1846{
1847 cfq_slab_kill();
1848 elv_unregister(&iosched_cfq);
1849}
1850
1851module_init(cfq_init);
1852module_exit(cfq_exit);
1853
1854MODULE_AUTHOR("Jens Axboe");
1855MODULE_LICENSE("GPL");
1856MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");