blob: d1f2ae1629f786ff0064fcb9e38b8ba1f96e97d1 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * CFQ, or complete fairness queueing, disk scheduler.
3 *
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6 *
7 * Copyright (C) 2003 Jens Axboe <axboe@suse.de>
8 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/config.h>
10#include <linux/module.h>
Al Viro1cc9be62006-03-18 12:29:52 -050011#include <linux/blkdev.h>
12#include <linux/elevator.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/hash.h>
14#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020015#include <linux/ioprio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016
17/*
18 * tunables
19 */
Arjan van de Ven64100092006-01-06 09:46:02 +010020static const int cfq_quantum = 4; /* max queue in one round of service */
21static const int cfq_queued = 8; /* minimum rq allocate limit per-queue*/
22static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
23static const int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */
24static const int cfq_back_penalty = 2; /* penalty of a backwards seek */
Linus Torvalds1da177e2005-04-16 15:20:36 -070025
Arjan van de Ven64100092006-01-06 09:46:02 +010026static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020027static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010028static const int cfq_slice_async_rq = 2;
Jens Axboe3b181522005-06-27 10:56:24 +020029static int cfq_slice_idle = HZ / 100;
Jens Axboe22e2c502005-06-27 10:55:12 +020030
31#define CFQ_IDLE_GRACE (HZ / 10)
32#define CFQ_SLICE_SCALE (5)
33
34#define CFQ_KEY_ASYNC (0)
Jens Axboe3b181522005-06-27 10:56:24 +020035#define CFQ_KEY_ANY (0xffff)
Jens Axboe22e2c502005-06-27 10:55:12 +020036
37/*
38 * disable queueing at the driver/hardware level
39 */
Arjan van de Ven64100092006-01-06 09:46:02 +010040static const int cfq_max_depth = 2;
Jens Axboe22e2c502005-06-27 10:55:12 +020041
Al Viroa6a07632006-03-18 13:26:44 -050042static DEFINE_RWLOCK(cfq_exit_lock);
43
Linus Torvalds1da177e2005-04-16 15:20:36 -070044/*
45 * for the hash of cfqq inside the cfqd
46 */
47#define CFQ_QHASH_SHIFT 6
48#define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
49#define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
50
51/*
52 * for the hash of crq inside the cfqq
53 */
54#define CFQ_MHASH_SHIFT 6
55#define CFQ_MHASH_BLOCK(sec) ((sec) >> 3)
56#define CFQ_MHASH_ENTRIES (1 << CFQ_MHASH_SHIFT)
57#define CFQ_MHASH_FN(sec) hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
58#define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
59#define list_entry_hash(ptr) hlist_entry((ptr), struct cfq_rq, hash)
60
61#define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
Jens Axboe22e2c502005-06-27 10:55:12 +020062#define list_entry_fifo(ptr) list_entry((ptr), struct request, queuelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -070063
64#define RQ_DATA(rq) (rq)->elevator_private
65
66/*
67 * rb-tree defines
68 */
69#define RB_NONE (2)
70#define RB_EMPTY(node) ((node)->rb_node == NULL)
71#define RB_CLEAR_COLOR(node) (node)->rb_color = RB_NONE
72#define RB_CLEAR(node) do { \
73 (node)->rb_parent = NULL; \
74 RB_CLEAR_COLOR((node)); \
75 (node)->rb_right = NULL; \
76 (node)->rb_left = NULL; \
77} while (0)
78#define RB_CLEAR_ROOT(root) ((root)->rb_node = NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -070079#define rb_entry_crq(node) rb_entry((node), struct cfq_rq, rb_node)
80#define rq_rb_key(rq) (rq)->sector
81
Linus Torvalds1da177e2005-04-16 15:20:36 -070082static kmem_cache_t *crq_pool;
83static kmem_cache_t *cfq_pool;
84static kmem_cache_t *cfq_ioc_pool;
85
Al Viro334e94d2006-03-18 15:05:53 -050086static atomic_t ioc_count = ATOMIC_INIT(0);
87static struct completion *ioc_gone;
88
Jens Axboe22e2c502005-06-27 10:55:12 +020089#define CFQ_PRIO_LISTS IOPRIO_BE_NR
90#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
91#define cfq_class_be(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
92#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
93
Jens Axboe3b181522005-06-27 10:56:24 +020094#define ASYNC (0)
95#define SYNC (1)
96
97#define cfq_cfqq_dispatched(cfqq) \
98 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
99
100#define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
101
102#define cfq_cfqq_sync(cfqq) \
103 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
Jens Axboe22e2c502005-06-27 10:55:12 +0200104
105/*
106 * Per block device queue structure
107 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108struct cfq_data {
Jens Axboe22e2c502005-06-27 10:55:12 +0200109 request_queue_t *queue;
110
111 /*
112 * rr list of queues with requests and the count of them
113 */
114 struct list_head rr_list[CFQ_PRIO_LISTS];
115 struct list_head busy_rr;
116 struct list_head cur_rr;
117 struct list_head idle_rr;
118 unsigned int busy_queues;
119
120 /*
121 * non-ordered list of empty cfqq's
122 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123 struct list_head empty_list;
124
Jens Axboe22e2c502005-06-27 10:55:12 +0200125 /*
126 * cfqq lookup hash
127 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128 struct hlist_head *cfq_hash;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129
Jens Axboe22e2c502005-06-27 10:55:12 +0200130 /*
131 * global crq hash for all queues
132 */
133 struct hlist_head *crq_hash;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134
135 unsigned int max_queued;
136
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 mempool_t *crq_pool;
138
Jens Axboe22e2c502005-06-27 10:55:12 +0200139 int rq_in_driver;
140
141 /*
142 * schedule slice state info
143 */
144 /*
145 * idle window management
146 */
147 struct timer_list idle_slice_timer;
148 struct work_struct unplug_work;
149
150 struct cfq_queue *active_queue;
151 struct cfq_io_context *active_cic;
152 int cur_prio, cur_end_prio;
153 unsigned int dispatch_slice;
154
155 struct timer_list idle_class_timer;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156
157 sector_t last_sector;
Jens Axboe22e2c502005-06-27 10:55:12 +0200158 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159
Jens Axboe22e2c502005-06-27 10:55:12 +0200160 unsigned int rq_starved;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161
162 /*
163 * tunables, see top of file
164 */
165 unsigned int cfq_quantum;
166 unsigned int cfq_queued;
Jens Axboe22e2c502005-06-27 10:55:12 +0200167 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 unsigned int cfq_back_penalty;
169 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200170 unsigned int cfq_slice[2];
171 unsigned int cfq_slice_async_rq;
172 unsigned int cfq_slice_idle;
173 unsigned int cfq_max_depth;
Al Virod9ff4182006-03-18 13:51:22 -0500174
175 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176};
177
Jens Axboe22e2c502005-06-27 10:55:12 +0200178/*
179 * Per process-grouping structure
180 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181struct cfq_queue {
182 /* reference count */
183 atomic_t ref;
184 /* parent cfq_data */
185 struct cfq_data *cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200186 /* cfqq lookup hash */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187 struct hlist_node cfq_hash;
188 /* hash key */
Jens Axboe22e2c502005-06-27 10:55:12 +0200189 unsigned int key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 /* on either rr or empty list of cfqd */
191 struct list_head cfq_list;
192 /* sorted list of pending requests */
193 struct rb_root sort_list;
194 /* if fifo isn't expired, next request to serve */
195 struct cfq_rq *next_crq;
196 /* requests queued in sort_list */
197 int queued[2];
198 /* currently allocated requests */
199 int allocated[2];
200 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200201 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202
Jens Axboe22e2c502005-06-27 10:55:12 +0200203 unsigned long slice_start;
204 unsigned long slice_end;
205 unsigned long slice_left;
206 unsigned long service_last;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207
Jens Axboe3b181522005-06-27 10:56:24 +0200208 /* number of requests that are on the dispatch list */
209 int on_dispatch[2];
Jens Axboe22e2c502005-06-27 10:55:12 +0200210
211 /* io prio of this group */
212 unsigned short ioprio, org_ioprio;
213 unsigned short ioprio_class, org_ioprio_class;
214
Jens Axboe3b181522005-06-27 10:56:24 +0200215 /* various state flags, see below */
216 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217};
218
219struct cfq_rq {
220 struct rb_node rb_node;
221 sector_t rb_key;
222 struct request *request;
223 struct hlist_node hash;
224
225 struct cfq_queue *cfq_queue;
226 struct cfq_io_context *io_context;
227
Jens Axboe3b181522005-06-27 10:56:24 +0200228 unsigned int crq_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229};
230
Jens Axboe3b181522005-06-27 10:56:24 +0200231enum cfqq_state_flags {
232 CFQ_CFQQ_FLAG_on_rr = 0,
233 CFQ_CFQQ_FLAG_wait_request,
234 CFQ_CFQQ_FLAG_must_alloc,
235 CFQ_CFQQ_FLAG_must_alloc_slice,
236 CFQ_CFQQ_FLAG_must_dispatch,
237 CFQ_CFQQ_FLAG_fifo_expire,
238 CFQ_CFQQ_FLAG_idle_window,
239 CFQ_CFQQ_FLAG_prio_changed,
Jens Axboe3b181522005-06-27 10:56:24 +0200240};
241
242#define CFQ_CFQQ_FNS(name) \
243static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
244{ \
245 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
246} \
247static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
248{ \
249 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
250} \
251static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
252{ \
253 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
254}
255
256CFQ_CFQQ_FNS(on_rr);
257CFQ_CFQQ_FNS(wait_request);
258CFQ_CFQQ_FNS(must_alloc);
259CFQ_CFQQ_FNS(must_alloc_slice);
260CFQ_CFQQ_FNS(must_dispatch);
261CFQ_CFQQ_FNS(fifo_expire);
262CFQ_CFQQ_FNS(idle_window);
263CFQ_CFQQ_FNS(prio_changed);
Jens Axboe3b181522005-06-27 10:56:24 +0200264#undef CFQ_CFQQ_FNS
265
266enum cfq_rq_state_flags {
Jens Axboeb4878f22005-10-20 16:42:29 +0200267 CFQ_CRQ_FLAG_is_sync = 0,
Jens Axboe3b181522005-06-27 10:56:24 +0200268};
269
270#define CFQ_CRQ_FNS(name) \
271static inline void cfq_mark_crq_##name(struct cfq_rq *crq) \
272{ \
273 crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name); \
274} \
275static inline void cfq_clear_crq_##name(struct cfq_rq *crq) \
276{ \
277 crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name); \
278} \
279static inline int cfq_crq_##name(const struct cfq_rq *crq) \
280{ \
281 return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0; \
282}
283
Jens Axboe3b181522005-06-27 10:56:24 +0200284CFQ_CRQ_FNS(is_sync);
Jens Axboe3b181522005-06-27 10:56:24 +0200285#undef CFQ_CRQ_FNS
286
287static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
Jens Axboeb4878f22005-10-20 16:42:29 +0200288static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
Al Viro6f325a12006-03-18 14:58:37 -0500289static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290
Jens Axboe22e2c502005-06-27 10:55:12 +0200291#define process_sync(tsk) ((tsk)->flags & PF_SYNCWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
293/*
294 * lots of deadline iosched dupes, can be abstracted later...
295 */
296static inline void cfq_del_crq_hash(struct cfq_rq *crq)
297{
298 hlist_del_init(&crq->hash);
299}
300
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
302{
303 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
304
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
306}
307
308static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
309{
310 struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
311 struct hlist_node *entry, *next;
312
313 hlist_for_each_safe(entry, next, hash_list) {
314 struct cfq_rq *crq = list_entry_hash(entry);
315 struct request *__rq = crq->request;
316
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317 if (!rq_mergeable(__rq)) {
318 cfq_del_crq_hash(crq);
319 continue;
320 }
321
322 if (rq_hash_key(__rq) == offset)
323 return __rq;
324 }
325
326 return NULL;
327}
328
Andrew Morton99f95e52005-06-27 20:14:05 -0700329/*
330 * scheduler run of queue, if there are requests pending and no one in the
331 * driver that will restart queueing
332 */
333static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
334{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100335 if (cfqd->busy_queues)
Andrew Morton99f95e52005-06-27 20:14:05 -0700336 kblockd_schedule_work(&cfqd->unplug_work);
337}
338
339static int cfq_queue_empty(request_queue_t *q)
340{
341 struct cfq_data *cfqd = q->elevator->elevator_data;
342
Jens Axboeb4878f22005-10-20 16:42:29 +0200343 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700344}
345
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346/*
347 * Lifted from AS - choose which of crq1 and crq2 that is best served now.
348 * We choose the request that is closest to the head right now. Distance
349 * behind the head are penalized and only allowed to a certain extent.
350 */
351static struct cfq_rq *
352cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
353{
354 sector_t last, s1, s2, d1 = 0, d2 = 0;
355 int r1_wrap = 0, r2_wrap = 0; /* requests are behind the disk head */
356 unsigned long back_max;
357
358 if (crq1 == NULL || crq1 == crq2)
359 return crq2;
360 if (crq2 == NULL)
361 return crq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200362
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200363 if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
364 return crq1;
365 else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
Jens Axboe22e2c502005-06-27 10:55:12 +0200366 return crq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367
368 s1 = crq1->request->sector;
369 s2 = crq2->request->sector;
370
371 last = cfqd->last_sector;
372
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373 /*
374 * by definition, 1KiB is 2 sectors
375 */
376 back_max = cfqd->cfq_back_max * 2;
377
378 /*
379 * Strict one way elevator _except_ in the case where we allow
380 * short backward seeks which are biased as twice the cost of a
381 * similar forward seek.
382 */
383 if (s1 >= last)
384 d1 = s1 - last;
385 else if (s1 + back_max >= last)
386 d1 = (last - s1) * cfqd->cfq_back_penalty;
387 else
388 r1_wrap = 1;
389
390 if (s2 >= last)
391 d2 = s2 - last;
392 else if (s2 + back_max >= last)
393 d2 = (last - s2) * cfqd->cfq_back_penalty;
394 else
395 r2_wrap = 1;
396
397 /* Found required data */
398 if (!r1_wrap && r2_wrap)
399 return crq1;
400 else if (!r2_wrap && r1_wrap)
401 return crq2;
402 else if (r1_wrap && r2_wrap) {
403 /* both behind the head */
404 if (s1 <= s2)
405 return crq1;
406 else
407 return crq2;
408 }
409
410 /* Both requests in front of the head */
411 if (d1 < d2)
412 return crq1;
413 else if (d2 < d1)
414 return crq2;
415 else {
416 if (s1 >= s2)
417 return crq1;
418 else
419 return crq2;
420 }
421}
422
423/*
424 * would be nice to take fifo expire time into account as well
425 */
426static struct cfq_rq *
427cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
428 struct cfq_rq *last)
429{
430 struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
431 struct rb_node *rbnext, *rbprev;
432
Jens Axboeb4878f22005-10-20 16:42:29 +0200433 if (!(rbnext = rb_next(&last->rb_node))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 rbnext = rb_first(&cfqq->sort_list);
Jens Axboe22e2c502005-06-27 10:55:12 +0200435 if (rbnext == &last->rb_node)
436 rbnext = NULL;
437 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438
439 rbprev = rb_prev(&last->rb_node);
440
441 if (rbprev)
442 crq_prev = rb_entry_crq(rbprev);
443 if (rbnext)
444 crq_next = rb_entry_crq(rbnext);
445
446 return cfq_choose_req(cfqd, crq_next, crq_prev);
447}
448
449static void cfq_update_next_crq(struct cfq_rq *crq)
450{
451 struct cfq_queue *cfqq = crq->cfq_queue;
452
453 if (cfqq->next_crq == crq)
454 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
455}
456
Jens Axboe22e2c502005-06-27 10:55:12 +0200457static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458{
Jens Axboe22e2c502005-06-27 10:55:12 +0200459 struct cfq_data *cfqd = cfqq->cfqd;
460 struct list_head *list, *entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461
Jens Axboe3b181522005-06-27 10:56:24 +0200462 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463
464 list_del(&cfqq->cfq_list);
465
Jens Axboe22e2c502005-06-27 10:55:12 +0200466 if (cfq_class_rt(cfqq))
467 list = &cfqd->cur_rr;
468 else if (cfq_class_idle(cfqq))
469 list = &cfqd->idle_rr;
470 else {
471 /*
472 * if cfqq has requests in flight, don't allow it to be
473 * found in cfq_set_active_queue before it has finished them.
474 * this is done to increase fairness between a process that
475 * has lots of io pending vs one that only generates one
476 * sporadically or synchronously
477 */
Jens Axboe3b181522005-06-27 10:56:24 +0200478 if (cfq_cfqq_dispatched(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200479 list = &cfqd->busy_rr;
480 else
481 list = &cfqd->rr_list[cfqq->ioprio];
482 }
483
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200485 * if queue was preempted, just add to front to be fair. busy_rr
486 * isn't sorted.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200488 if (preempted || list == &cfqd->busy_rr) {
489 list_add(&cfqq->cfq_list, list);
490 return;
491 }
492
493 /*
494 * sort by when queue was last serviced
495 */
496 entry = list;
497 while ((entry = entry->prev) != list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
499
Jens Axboe22e2c502005-06-27 10:55:12 +0200500 if (!__cfqq->service_last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 break;
Jens Axboe22e2c502005-06-27 10:55:12 +0200502 if (time_before(__cfqq->service_last, cfqq->service_last))
503 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 }
505
506 list_add(&cfqq->cfq_list, entry);
507}
508
509/*
510 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200511 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 */
513static inline void
Jens Axboeb4878f22005-10-20 16:42:29 +0200514cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515{
Jens Axboe3b181522005-06-27 10:56:24 +0200516 BUG_ON(cfq_cfqq_on_rr(cfqq));
517 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 cfqd->busy_queues++;
519
Jens Axboeb4878f22005-10-20 16:42:29 +0200520 cfq_resort_rr_list(cfqq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521}
522
523static inline void
524cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
525{
Jens Axboe3b181522005-06-27 10:56:24 +0200526 BUG_ON(!cfq_cfqq_on_rr(cfqq));
527 cfq_clear_cfqq_on_rr(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200528 list_move(&cfqq->cfq_list, &cfqd->empty_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529
530 BUG_ON(!cfqd->busy_queues);
531 cfqd->busy_queues--;
532}
533
534/*
535 * rb tree support functions
536 */
537static inline void cfq_del_crq_rb(struct cfq_rq *crq)
538{
539 struct cfq_queue *cfqq = crq->cfq_queue;
Jens Axboeb4878f22005-10-20 16:42:29 +0200540 struct cfq_data *cfqd = cfqq->cfqd;
541 const int sync = cfq_crq_is_sync(crq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542
Jens Axboeb4878f22005-10-20 16:42:29 +0200543 BUG_ON(!cfqq->queued[sync]);
544 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545
Jens Axboeb4878f22005-10-20 16:42:29 +0200546 cfq_update_next_crq(crq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547
Jens Axboeb4878f22005-10-20 16:42:29 +0200548 rb_erase(&crq->rb_node, &cfqq->sort_list);
549 RB_CLEAR_COLOR(&crq->rb_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550
Jens Axboeb4878f22005-10-20 16:42:29 +0200551 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
552 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553}
554
555static struct cfq_rq *
556__cfq_add_crq_rb(struct cfq_rq *crq)
557{
558 struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
559 struct rb_node *parent = NULL;
560 struct cfq_rq *__crq;
561
562 while (*p) {
563 parent = *p;
564 __crq = rb_entry_crq(parent);
565
566 if (crq->rb_key < __crq->rb_key)
567 p = &(*p)->rb_left;
568 else if (crq->rb_key > __crq->rb_key)
569 p = &(*p)->rb_right;
570 else
571 return __crq;
572 }
573
574 rb_link_node(&crq->rb_node, parent, p);
575 return NULL;
576}
577
578static void cfq_add_crq_rb(struct cfq_rq *crq)
579{
580 struct cfq_queue *cfqq = crq->cfq_queue;
581 struct cfq_data *cfqd = cfqq->cfqd;
582 struct request *rq = crq->request;
583 struct cfq_rq *__alias;
584
585 crq->rb_key = rq_rb_key(rq);
Jens Axboe3b181522005-06-27 10:56:24 +0200586 cfqq->queued[cfq_crq_is_sync(crq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587
588 /*
589 * looks a little odd, but the first insert might return an alias.
590 * if that happens, put the alias on the dispatch list
591 */
592 while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
Jens Axboeb4878f22005-10-20 16:42:29 +0200593 cfq_dispatch_insert(cfqd->queue, __alias);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594
595 rb_insert_color(&crq->rb_node, &cfqq->sort_list);
596
Jens Axboe3b181522005-06-27 10:56:24 +0200597 if (!cfq_cfqq_on_rr(cfqq))
Jens Axboeb4878f22005-10-20 16:42:29 +0200598 cfq_add_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599
600 /*
601 * check if this request is a better next-serve candidate
602 */
603 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
604}
605
606static inline void
607cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
608{
Jens Axboeb4878f22005-10-20 16:42:29 +0200609 rb_erase(&crq->rb_node, &cfqq->sort_list);
610 cfqq->queued[cfq_crq_is_sync(crq)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611
612 cfq_add_crq_rb(crq);
613}
614
Jens Axboe22e2c502005-06-27 10:55:12 +0200615static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
616
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617{
Jens Axboe3b181522005-06-27 10:56:24 +0200618 struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 struct rb_node *n;
620
621 if (!cfqq)
622 goto out;
623
624 n = cfqq->sort_list.rb_node;
625 while (n) {
626 struct cfq_rq *crq = rb_entry_crq(n);
627
628 if (sector < crq->rb_key)
629 n = n->rb_left;
630 else if (sector > crq->rb_key)
631 n = n->rb_right;
632 else
633 return crq->request;
634 }
635
636out:
637 return NULL;
638}
639
Jens Axboeb4878f22005-10-20 16:42:29 +0200640static void cfq_activate_request(request_queue_t *q, struct request *rq)
641{
642 struct cfq_data *cfqd = q->elevator->elevator_data;
643
644 cfqd->rq_in_driver++;
645}
646
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
648{
Jens Axboe22e2c502005-06-27 10:55:12 +0200649 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650
Jens Axboeb4878f22005-10-20 16:42:29 +0200651 WARN_ON(!cfqd->rq_in_driver);
652 cfqd->rq_in_driver--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653}
654
Jens Axboeb4878f22005-10-20 16:42:29 +0200655static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656{
657 struct cfq_rq *crq = RQ_DATA(rq);
658
Jens Axboeb4878f22005-10-20 16:42:29 +0200659 list_del_init(&rq->queuelist);
660 cfq_del_crq_rb(crq);
Tejun Heo98b11472005-10-20 16:46:54 +0200661 cfq_del_crq_hash(crq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662}
663
664static int
665cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
666{
667 struct cfq_data *cfqd = q->elevator->elevator_data;
668 struct request *__rq;
669 int ret;
670
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
Jens Axboe22e2c502005-06-27 10:55:12 +0200672 if (__rq && elv_rq_merge_ok(__rq, bio)) {
673 ret = ELEVATOR_BACK_MERGE;
674 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 }
676
677 __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
Jens Axboe22e2c502005-06-27 10:55:12 +0200678 if (__rq && elv_rq_merge_ok(__rq, bio)) {
679 ret = ELEVATOR_FRONT_MERGE;
680 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 }
682
683 return ELEVATOR_NO_MERGE;
684out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 *req = __rq;
686 return ret;
687}
688
689static void cfq_merged_request(request_queue_t *q, struct request *req)
690{
691 struct cfq_data *cfqd = q->elevator->elevator_data;
692 struct cfq_rq *crq = RQ_DATA(req);
693
694 cfq_del_crq_hash(crq);
695 cfq_add_crq_hash(cfqd, crq);
696
Jens Axboeb4878f22005-10-20 16:42:29 +0200697 if (rq_rb_key(req) != crq->rb_key) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 struct cfq_queue *cfqq = crq->cfq_queue;
699
700 cfq_update_next_crq(crq);
701 cfq_reposition_crq_rb(cfqq, crq);
702 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703}
704
705static void
706cfq_merged_requests(request_queue_t *q, struct request *rq,
707 struct request *next)
708{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709 cfq_merged_request(q, rq);
710
Jens Axboe22e2c502005-06-27 10:55:12 +0200711 /*
712 * reposition in fifo if next is older than rq
713 */
714 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
715 time_before(next->start_time, rq->start_time))
716 list_move(&rq->queuelist, &next->queuelist);
717
Jens Axboeb4878f22005-10-20 16:42:29 +0200718 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200719}
720
721static inline void
722__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
723{
724 if (cfqq) {
725 /*
726 * stop potential idle class queues waiting service
727 */
728 del_timer(&cfqd->idle_class_timer);
729
730 cfqq->slice_start = jiffies;
731 cfqq->slice_end = 0;
732 cfqq->slice_left = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200733 cfq_clear_cfqq_must_alloc_slice(cfqq);
734 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200735 }
736
737 cfqd->active_queue = cfqq;
738}
739
740/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100741 * current cfqq expired its slice (or was too idle), select new one
742 */
743static void
744__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
745 int preempted)
746{
747 unsigned long now = jiffies;
748
749 if (cfq_cfqq_wait_request(cfqq))
750 del_timer(&cfqd->idle_slice_timer);
751
752 if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
753 cfqq->service_last = now;
754 cfq_schedule_dispatch(cfqd);
755 }
756
757 cfq_clear_cfqq_must_dispatch(cfqq);
758 cfq_clear_cfqq_wait_request(cfqq);
759
760 /*
761 * store what was left of this slice, if the queue idled out
762 * or was preempted
763 */
764 if (time_after(cfqq->slice_end, now))
765 cfqq->slice_left = cfqq->slice_end - now;
766 else
767 cfqq->slice_left = 0;
768
769 if (cfq_cfqq_on_rr(cfqq))
770 cfq_resort_rr_list(cfqq, preempted);
771
772 if (cfqq == cfqd->active_queue)
773 cfqd->active_queue = NULL;
774
775 if (cfqd->active_cic) {
776 put_io_context(cfqd->active_cic->ioc);
777 cfqd->active_cic = NULL;
778 }
779
780 cfqd->dispatch_slice = 0;
781}
782
783static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
784{
785 struct cfq_queue *cfqq = cfqd->active_queue;
786
787 if (cfqq)
788 __cfq_slice_expired(cfqd, cfqq, preempted);
789}
790
791/*
Jens Axboe22e2c502005-06-27 10:55:12 +0200792 * 0
793 * 0,1
794 * 0,1,2
795 * 0,1,2,3
796 * 0,1,2,3,4
797 * 0,1,2,3,4,5
798 * 0,1,2,3,4,5,6
799 * 0,1,2,3,4,5,6,7
800 */
801static int cfq_get_next_prio_level(struct cfq_data *cfqd)
802{
803 int prio, wrap;
804
805 prio = -1;
806 wrap = 0;
807 do {
808 int p;
809
810 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
811 if (!list_empty(&cfqd->rr_list[p])) {
812 prio = p;
813 break;
814 }
815 }
816
817 if (prio != -1)
818 break;
819 cfqd->cur_prio = 0;
820 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
821 cfqd->cur_end_prio = 0;
822 if (wrap)
823 break;
824 wrap = 1;
825 }
826 } while (1);
827
828 if (unlikely(prio == -1))
829 return -1;
830
831 BUG_ON(prio >= CFQ_PRIO_LISTS);
832
833 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
834
835 cfqd->cur_prio = prio + 1;
836 if (cfqd->cur_prio > cfqd->cur_end_prio) {
837 cfqd->cur_end_prio = cfqd->cur_prio;
838 cfqd->cur_prio = 0;
839 }
840 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
841 cfqd->cur_prio = 0;
842 cfqd->cur_end_prio = 0;
843 }
844
845 return prio;
846}
847
Jens Axboe3b181522005-06-27 10:56:24 +0200848static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200849{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100850 struct cfq_queue *cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +0200851
852 /*
853 * if current list is non-empty, grab first entry. if it is empty,
854 * get next prio level and grab first entry then if any are spliced
855 */
856 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
857 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
858
859 /*
860 * if we have idle queues and no rt or be queues had pending
861 * requests, either allow immediate service if the grace period
862 * has passed or arm the idle grace timer
863 */
864 if (!cfqq && !list_empty(&cfqd->idle_rr)) {
865 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
866
867 if (time_after_eq(jiffies, end))
868 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
869 else
870 mod_timer(&cfqd->idle_class_timer, end);
871 }
872
873 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200874 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200875}
876
Jens Axboe22e2c502005-06-27 10:55:12 +0200877static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
878
879{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100880 unsigned long sl;
881
Jens Axboe22e2c502005-06-27 10:55:12 +0200882 WARN_ON(!RB_EMPTY(&cfqq->sort_list));
883 WARN_ON(cfqq != cfqd->active_queue);
884
885 /*
886 * idle is disabled, either manually or by past process history
887 */
888 if (!cfqd->cfq_slice_idle)
889 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200890 if (!cfq_cfqq_idle_window(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200891 return 0;
892 /*
893 * task has exited, don't wait
894 */
895 if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
896 return 0;
897
Jens Axboe3b181522005-06-27 10:56:24 +0200898 cfq_mark_cfqq_must_dispatch(cfqq);
899 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200900
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100901 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
902 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe22e2c502005-06-27 10:55:12 +0200903 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904}
905
Jens Axboeb4878f22005-10-20 16:42:29 +0200906static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907{
908 struct cfq_data *cfqd = q->elevator->elevator_data;
909 struct cfq_queue *cfqq = crq->cfq_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +0200910
911 cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200912 cfq_remove_request(crq->request);
Jens Axboe3b181522005-06-27 10:56:24 +0200913 cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
Jens Axboeb4878f22005-10-20 16:42:29 +0200914 elv_dispatch_sort(q, crq->request);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915}
916
917/*
918 * return expired entry, or NULL to just start from scratch in rbtree
919 */
920static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
921{
922 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200923 struct request *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 struct cfq_rq *crq;
925
Jens Axboe3b181522005-06-27 10:56:24 +0200926 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 return NULL;
928
Jens Axboe22e2c502005-06-27 10:55:12 +0200929 if (!list_empty(&cfqq->fifo)) {
Jens Axboe3b181522005-06-27 10:56:24 +0200930 int fifo = cfq_cfqq_class_sync(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931
Jens Axboe22e2c502005-06-27 10:55:12 +0200932 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
933 rq = crq->request;
934 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
Jens Axboe3b181522005-06-27 10:56:24 +0200935 cfq_mark_cfqq_fifo_expire(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200936 return crq;
937 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 }
939
940 return NULL;
941}
942
943/*
Jens Axboe3b181522005-06-27 10:56:24 +0200944 * Scale schedule slice based on io priority. Use the sync time slice only
945 * if a queue is marked sync and has sync io queued. A sync queue with async
946 * io only, should not get full sync slice length.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200948static inline int
949cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950{
Jens Axboe22e2c502005-06-27 10:55:12 +0200951 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952
Jens Axboe22e2c502005-06-27 10:55:12 +0200953 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954
Jens Axboe22e2c502005-06-27 10:55:12 +0200955 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956}
957
Jens Axboe22e2c502005-06-27 10:55:12 +0200958static inline void
959cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
960{
961 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
962}
963
964static inline int
965cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
966{
967 const int base_rq = cfqd->cfq_slice_async_rq;
968
969 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
970
971 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
972}
973
974/*
975 * get next queue for service
976 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100977static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200978{
979 unsigned long now = jiffies;
980 struct cfq_queue *cfqq;
981
982 cfqq = cfqd->active_queue;
983 if (!cfqq)
984 goto new_queue;
985
986 /*
987 * slice has expired
988 */
Jens Axboe3b181522005-06-27 10:56:24 +0200989 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
990 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +0200991
992 /*
993 * if queue has requests, dispatch one. if not, check if
994 * enough slice is left to wait for one
995 */
996 if (!RB_EMPTY(&cfqq->sort_list))
997 goto keep_queue;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100998 else if (cfq_cfqq_class_sync(cfqq) &&
Jens Axboe22e2c502005-06-27 10:55:12 +0200999 time_before(now, cfqq->slice_end)) {
1000 if (cfq_arm_slice_timer(cfqd, cfqq))
1001 return NULL;
1002 }
1003
Jens Axboe3b181522005-06-27 10:56:24 +02001004expire:
Jens Axboe22e2c502005-06-27 10:55:12 +02001005 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001006new_queue:
1007 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001008keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001009 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001010}
1011
1012static int
1013__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1014 int max_dispatch)
1015{
1016 int dispatched = 0;
1017
1018 BUG_ON(RB_EMPTY(&cfqq->sort_list));
1019
1020 do {
1021 struct cfq_rq *crq;
1022
1023 /*
1024 * follow expired path, else get first next available
1025 */
1026 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1027 crq = cfqq->next_crq;
1028
1029 /*
1030 * finally, insert request into driver dispatch list
1031 */
Jens Axboeb4878f22005-10-20 16:42:29 +02001032 cfq_dispatch_insert(cfqd->queue, crq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001033
1034 cfqd->dispatch_slice++;
1035 dispatched++;
1036
1037 if (!cfqd->active_cic) {
1038 atomic_inc(&crq->io_context->ioc->refcount);
1039 cfqd->active_cic = crq->io_context;
1040 }
1041
1042 if (RB_EMPTY(&cfqq->sort_list))
1043 break;
1044
1045 } while (dispatched < max_dispatch);
1046
1047 /*
1048 * if slice end isn't set yet, set it. if at least one request was
1049 * sync, use the sync time slice value
1050 */
1051 if (!cfqq->slice_end)
1052 cfq_set_prio_slice(cfqd, cfqq);
1053
1054 /*
1055 * expire an async queue immediately if it has used up its slice. idle
1056 * queue always expire after 1 dispatch round.
1057 */
1058 if ((!cfq_cfqq_sync(cfqq) &&
1059 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1060 cfq_class_idle(cfqq))
1061 cfq_slice_expired(cfqd, 0);
1062
1063 return dispatched;
1064}
1065
1066static int
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001067cfq_forced_dispatch_cfqqs(struct list_head *list)
1068{
1069 int dispatched = 0;
1070 struct cfq_queue *cfqq, *next;
1071 struct cfq_rq *crq;
1072
1073 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1074 while ((crq = cfqq->next_crq)) {
1075 cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1076 dispatched++;
1077 }
1078 BUG_ON(!list_empty(&cfqq->fifo));
1079 }
1080 return dispatched;
1081}
1082
1083static int
1084cfq_forced_dispatch(struct cfq_data *cfqd)
1085{
1086 int i, dispatched = 0;
1087
1088 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1089 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1090
1091 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1092 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1093 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1094
1095 cfq_slice_expired(cfqd, 0);
1096
1097 BUG_ON(cfqd->busy_queues);
1098
1099 return dispatched;
1100}
1101
1102static int
Jens Axboeb4878f22005-10-20 16:42:29 +02001103cfq_dispatch_requests(request_queue_t *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104{
1105 struct cfq_data *cfqd = q->elevator->elevator_data;
1106 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107
Jens Axboe22e2c502005-06-27 10:55:12 +02001108 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 return 0;
1110
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001111 if (unlikely(force))
1112 return cfq_forced_dispatch(cfqd);
1113
1114 cfqq = cfq_select_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001115 if (cfqq) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001116 int max_dispatch;
1117
1118 /*
1119 * if idle window is disabled, allow queue buildup
1120 */
1121 if (!cfq_cfqq_idle_window(cfqq) &&
1122 cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1123 return 0;
1124
Jens Axboe3b181522005-06-27 10:56:24 +02001125 cfq_clear_cfqq_must_dispatch(cfqq);
1126 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001127 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001129 max_dispatch = cfqd->cfq_quantum;
1130 if (cfq_class_idle(cfqq))
1131 max_dispatch = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132
Jens Axboe22e2c502005-06-27 10:55:12 +02001133 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 }
1135
Jens Axboe22e2c502005-06-27 10:55:12 +02001136 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137}
1138
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139/*
1140 * task holds one reference to the queue, dropped when task exits. each crq
1141 * in-flight on this queue also holds a reference, dropped when crq is freed.
1142 *
1143 * queue lock must be held here.
1144 */
1145static void cfq_put_queue(struct cfq_queue *cfqq)
1146{
Jens Axboe22e2c502005-06-27 10:55:12 +02001147 struct cfq_data *cfqd = cfqq->cfqd;
1148
1149 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150
1151 if (!atomic_dec_and_test(&cfqq->ref))
1152 return;
1153
1154 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001155 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001156 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001158 if (unlikely(cfqd->active_queue == cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001159 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02001160
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 /*
1162 * it's on the empty list and still hashed
1163 */
1164 list_del(&cfqq->cfq_list);
1165 hlist_del(&cfqq->cfq_hash);
1166 kmem_cache_free(cfq_pool, cfqq);
1167}
1168
1169static inline struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001170__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1171 const int hashval)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172{
1173 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1174 struct hlist_node *entry, *next;
1175
1176 hlist_for_each_safe(entry, next, hash_list) {
1177 struct cfq_queue *__cfqq = list_entry_qhash(entry);
Al Virob0a69162006-03-14 15:32:50 -05001178 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179
Jens Axboe3b181522005-06-27 10:56:24 +02001180 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 return __cfqq;
1182 }
1183
1184 return NULL;
1185}
1186
1187static struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001188cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189{
Jens Axboe3b181522005-06-27 10:56:24 +02001190 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191}
1192
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193static void cfq_free_io_context(struct cfq_io_context *cic)
1194{
Jens Axboe22e2c502005-06-27 10:55:12 +02001195 struct cfq_io_context *__cic;
1196 struct list_head *entry, *next;
Al Viro334e94d2006-03-18 15:05:53 -05001197 int freed = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001198
1199 list_for_each_safe(entry, next, &cic->list) {
1200 __cic = list_entry(entry, struct cfq_io_context, list);
1201 kmem_cache_free(cfq_ioc_pool, __cic);
Al Viro334e94d2006-03-18 15:05:53 -05001202 freed++;
Jens Axboe22e2c502005-06-27 10:55:12 +02001203 }
1204
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 kmem_cache_free(cfq_ioc_pool, cic);
Al Viro334e94d2006-03-18 15:05:53 -05001206 if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1207 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208}
1209
Al Viroe17a9482006-03-18 13:21:20 -05001210static void cfq_trim(struct io_context *ioc)
1211{
1212 ioc->set_ioprio = NULL;
1213 if (ioc->cic)
1214 cfq_free_io_context(ioc->cic);
1215}
1216
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001218 * Called with interrupts disabled
1219 */
1220static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1221{
Al Viro478a82b2006-03-18 13:25:24 -05001222 struct cfq_data *cfqd = cic->key;
Al Virod9ff4182006-03-18 13:51:22 -05001223 request_queue_t *q;
1224
1225 if (!cfqd)
1226 return;
1227
1228 q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001229
1230 WARN_ON(!irqs_disabled());
1231
1232 spin_lock(q->queue_lock);
1233
Al Viro12a05732006-03-18 13:38:01 -05001234 if (cic->cfqq[ASYNC]) {
1235 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1236 __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1237 cfq_put_queue(cic->cfqq[ASYNC]);
1238 cic->cfqq[ASYNC] = NULL;
1239 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001240
Al Viro12a05732006-03-18 13:38:01 -05001241 if (cic->cfqq[SYNC]) {
1242 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1243 __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1244 cfq_put_queue(cic->cfqq[SYNC]);
1245 cic->cfqq[SYNC] = NULL;
1246 }
1247
Al Viro478a82b2006-03-18 13:25:24 -05001248 cic->key = NULL;
Al Virod9ff4182006-03-18 13:51:22 -05001249 list_del_init(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001250 spin_unlock(q->queue_lock);
1251}
1252
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253static void cfq_exit_io_context(struct cfq_io_context *cic)
1254{
Jens Axboe22e2c502005-06-27 10:55:12 +02001255 struct cfq_io_context *__cic;
1256 struct list_head *entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 unsigned long flags;
1258
Jens Axboe22e2c502005-06-27 10:55:12 +02001259 local_irq_save(flags);
1260
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 /*
1262 * put the reference this task is holding to the various queues
1263 */
Al Virod9ff4182006-03-18 13:51:22 -05001264 read_lock(&cfq_exit_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02001265 list_for_each(entry, &cic->list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 __cic = list_entry(entry, struct cfq_io_context, list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001267 cfq_exit_single_io_context(__cic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 }
1269
Jens Axboe22e2c502005-06-27 10:55:12 +02001270 cfq_exit_single_io_context(cic);
Al Virod9ff4182006-03-18 13:51:22 -05001271 read_unlock(&cfq_exit_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02001272 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273}
1274
Jens Axboe22e2c502005-06-27 10:55:12 +02001275static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001276cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277{
Jens Axboe22e2c502005-06-27 10:55:12 +02001278 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279
1280 if (cic) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281 INIT_LIST_HEAD(&cic->list);
Al Viro12a05732006-03-18 13:38:01 -05001282 cic->cfqq[ASYNC] = NULL;
1283 cic->cfqq[SYNC] = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001284 cic->key = NULL;
1285 cic->last_end_request = jiffies;
1286 cic->ttime_total = 0;
1287 cic->ttime_samples = 0;
1288 cic->ttime_mean = 0;
1289 cic->dtor = cfq_free_io_context;
1290 cic->exit = cfq_exit_io_context;
Al Virod9ff4182006-03-18 13:51:22 -05001291 INIT_LIST_HEAD(&cic->queue_list);
Al Viro334e94d2006-03-18 15:05:53 -05001292 atomic_inc(&ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 }
1294
1295 return cic;
1296}
1297
Jens Axboe22e2c502005-06-27 10:55:12 +02001298static void cfq_init_prio_data(struct cfq_queue *cfqq)
1299{
1300 struct task_struct *tsk = current;
1301 int ioprio_class;
1302
Jens Axboe3b181522005-06-27 10:56:24 +02001303 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001304 return;
1305
1306 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1307 switch (ioprio_class) {
1308 default:
1309 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1310 case IOPRIO_CLASS_NONE:
1311 /*
1312 * no prio set, place us in the middle of the BE classes
1313 */
1314 cfqq->ioprio = task_nice_ioprio(tsk);
1315 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1316 break;
1317 case IOPRIO_CLASS_RT:
1318 cfqq->ioprio = task_ioprio(tsk);
1319 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1320 break;
1321 case IOPRIO_CLASS_BE:
1322 cfqq->ioprio = task_ioprio(tsk);
1323 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1324 break;
1325 case IOPRIO_CLASS_IDLE:
1326 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1327 cfqq->ioprio = 7;
Jens Axboe3b181522005-06-27 10:56:24 +02001328 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001329 break;
1330 }
1331
1332 /*
1333 * keep track of original prio settings in case we have to temporarily
1334 * elevate the priority of this queue
1335 */
1336 cfqq->org_ioprio = cfqq->ioprio;
1337 cfqq->org_ioprio_class = cfqq->ioprio_class;
1338
Jens Axboe3b181522005-06-27 10:56:24 +02001339 if (cfq_cfqq_on_rr(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001340 cfq_resort_rr_list(cfqq, 0);
1341
Jens Axboe3b181522005-06-27 10:56:24 +02001342 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001343}
1344
Al Viro478a82b2006-03-18 13:25:24 -05001345static inline void changed_ioprio(struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001346{
Al Viro478a82b2006-03-18 13:25:24 -05001347 struct cfq_data *cfqd = cic->key;
1348 struct cfq_queue *cfqq;
1349 if (cfqd) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001350 spin_lock(cfqd->queue->queue_lock);
Al Viro12a05732006-03-18 13:38:01 -05001351 cfqq = cic->cfqq[ASYNC];
1352 if (cfqq) {
Al Viro6f325a12006-03-18 14:58:37 -05001353 struct cfq_queue *new_cfqq;
1354 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC,
1355 cic->ioc->task, GFP_ATOMIC);
1356 if (new_cfqq) {
1357 cic->cfqq[ASYNC] = new_cfqq;
1358 cfq_put_queue(cfqq);
1359 }
Al Viro12a05732006-03-18 13:38:01 -05001360 }
1361 cfqq = cic->cfqq[SYNC];
Al Viro478a82b2006-03-18 13:25:24 -05001362 if (cfqq) {
1363 cfq_mark_cfqq_prio_changed(cfqq);
1364 cfq_init_prio_data(cfqq);
1365 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001366 spin_unlock(cfqd->queue->queue_lock);
1367 }
1368}
1369
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001371 * callback from sys_ioprio_set, irqs are disabled
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001373static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374{
Al Viroa6a07632006-03-18 13:26:44 -05001375 struct cfq_io_context *cic;
1376
1377 write_lock(&cfq_exit_lock);
1378
1379 cic = ioc->cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380
Al Viro478a82b2006-03-18 13:25:24 -05001381 changed_ioprio(cic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382
Jens Axboe22e2c502005-06-27 10:55:12 +02001383 list_for_each_entry(cic, &cic->list, list)
Al Viro478a82b2006-03-18 13:25:24 -05001384 changed_ioprio(cic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385
Al Viroa6a07632006-03-18 13:26:44 -05001386 write_unlock(&cfq_exit_lock);
1387
Jens Axboe22e2c502005-06-27 10:55:12 +02001388 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389}
1390
1391static struct cfq_queue *
Al Viro6f325a12006-03-18 14:58:37 -05001392cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
Al Viro8267e262005-10-21 03:20:53 -04001393 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394{
1395 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1396 struct cfq_queue *cfqq, *new_cfqq = NULL;
Al Viro6f325a12006-03-18 14:58:37 -05001397 unsigned short ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398
1399retry:
Al Viro6f325a12006-03-18 14:58:37 -05001400 ioprio = tsk->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02001401 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402
1403 if (!cfqq) {
1404 if (new_cfqq) {
1405 cfqq = new_cfqq;
1406 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001407 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408 spin_unlock_irq(cfqd->queue->queue_lock);
1409 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1410 spin_lock_irq(cfqd->queue->queue_lock);
1411 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001412 } else {
1413 cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1414 if (!cfqq)
1415 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001416 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417
1418 memset(cfqq, 0, sizeof(*cfqq));
1419
1420 INIT_HLIST_NODE(&cfqq->cfq_hash);
1421 INIT_LIST_HEAD(&cfqq->cfq_list);
1422 RB_CLEAR_ROOT(&cfqq->sort_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001423 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424
1425 cfqq->key = key;
1426 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1427 atomic_set(&cfqq->ref, 0);
1428 cfqq->cfqd = cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +02001429 cfqq->service_last = 0;
1430 /*
1431 * set ->slice_left to allow preemption for a new process
1432 */
1433 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02001434 cfq_mark_cfqq_idle_window(cfqq);
1435 cfq_mark_cfqq_prio_changed(cfqq);
1436 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437 }
1438
1439 if (new_cfqq)
1440 kmem_cache_free(cfq_pool, new_cfqq);
1441
1442 atomic_inc(&cfqq->ref);
1443out:
1444 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1445 return cfqq;
1446}
1447
Jens Axboe22e2c502005-06-27 10:55:12 +02001448/*
1449 * Setup general io context and cfq io context. There can be several cfq
1450 * io contexts per general io context, if this process is doing io to more
1451 * than one device managed by cfq. Note that caller is holding a reference to
1452 * cfqq, so we don't need to worry about it disappearing
1453 */
1454static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001455cfq_get_io_context(struct cfq_data *cfqd, pid_t pid, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456{
Jens Axboe22e2c502005-06-27 10:55:12 +02001457 struct io_context *ioc = NULL;
1458 struct cfq_io_context *cic;
1459
1460 might_sleep_if(gfp_mask & __GFP_WAIT);
1461
1462 ioc = get_io_context(gfp_mask);
1463 if (!ioc)
1464 return NULL;
1465
Al Virod9ff4182006-03-18 13:51:22 -05001466restart:
Jens Axboe22e2c502005-06-27 10:55:12 +02001467 if ((cic = ioc->cic) == NULL) {
1468 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1469
1470 if (cic == NULL)
1471 goto err;
1472
1473 /*
1474 * manually increment generic io_context usage count, it
1475 * cannot go away since we are already holding one ref to it
1476 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001477 cic->ioc = ioc;
1478 cic->key = cfqd;
Al Viroa6a07632006-03-18 13:26:44 -05001479 read_lock(&cfq_exit_lock);
Al Viro478a82b2006-03-18 13:25:24 -05001480 ioc->set_ioprio = cfq_ioc_set_ioprio;
1481 ioc->cic = cic;
Al Virod9ff4182006-03-18 13:51:22 -05001482 list_add(&cic->queue_list, &cfqd->cic_list);
Al Viroa6a07632006-03-18 13:26:44 -05001483 read_unlock(&cfq_exit_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02001484 } else {
1485 struct cfq_io_context *__cic;
1486
1487 /*
1488 * the first cic on the list is actually the head itself
1489 */
1490 if (cic->key == cfqd)
1491 goto out;
1492
Al Virod9ff4182006-03-18 13:51:22 -05001493 if (unlikely(!cic->key)) {
1494 read_lock(&cfq_exit_lock);
1495 if (list_empty(&cic->list))
1496 ioc->cic = NULL;
1497 else
1498 ioc->cic = list_entry(cic->list.next,
1499 struct cfq_io_context,
1500 list);
1501 read_unlock(&cfq_exit_lock);
1502 kmem_cache_free(cfq_ioc_pool, cic);
Al Viro334e94d2006-03-18 15:05:53 -05001503 atomic_dec(&ioc_count);
Al Virod9ff4182006-03-18 13:51:22 -05001504 goto restart;
1505 }
1506
Jens Axboe22e2c502005-06-27 10:55:12 +02001507 /*
1508 * cic exists, check if we already are there. linear search
1509 * should be ok here, the list will usually not be more than
1510 * 1 or a few entries long
1511 */
1512 list_for_each_entry(__cic, &cic->list, list) {
1513 /*
1514 * this process is already holding a reference to
1515 * this queue, so no need to get one more
1516 */
1517 if (__cic->key == cfqd) {
1518 cic = __cic;
1519 goto out;
1520 }
Al Virod9ff4182006-03-18 13:51:22 -05001521 if (unlikely(!__cic->key)) {
1522 read_lock(&cfq_exit_lock);
1523 list_del(&__cic->list);
1524 read_unlock(&cfq_exit_lock);
1525 kmem_cache_free(cfq_ioc_pool, __cic);
Al Viro334e94d2006-03-18 15:05:53 -05001526 atomic_dec(&ioc_count);
Al Virod9ff4182006-03-18 13:51:22 -05001527 goto restart;
1528 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001529 }
1530
1531 /*
1532 * nope, process doesn't have a cic assoicated with this
1533 * cfqq yet. get a new one and add to list
1534 */
1535 __cic = cfq_alloc_io_context(cfqd, gfp_mask);
1536 if (__cic == NULL)
1537 goto err;
1538
1539 __cic->ioc = ioc;
1540 __cic->key = cfqd;
Al Viroa6a07632006-03-18 13:26:44 -05001541 read_lock(&cfq_exit_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02001542 list_add(&__cic->list, &cic->list);
Al Virod9ff4182006-03-18 13:51:22 -05001543 list_add(&__cic->queue_list, &cfqd->cic_list);
Al Viroa6a07632006-03-18 13:26:44 -05001544 read_unlock(&cfq_exit_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02001545 cic = __cic;
1546 }
1547
1548out:
1549 return cic;
1550err:
1551 put_io_context(ioc);
1552 return NULL;
1553}
1554
1555static void
1556cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1557{
1558 unsigned long elapsed, ttime;
1559
1560 /*
1561 * if this context already has stuff queued, thinktime is from
1562 * last queue not last end
1563 */
1564#if 0
1565 if (time_after(cic->last_end_request, cic->last_queue))
1566 elapsed = jiffies - cic->last_end_request;
1567 else
1568 elapsed = jiffies - cic->last_queue;
1569#else
1570 elapsed = jiffies - cic->last_end_request;
1571#endif
1572
1573 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1574
1575 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1576 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1577 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1578}
1579
1580#define sample_valid(samples) ((samples) > 80)
1581
1582/*
1583 * Disable idle window if the process thinks too long or seeks so much that
1584 * it doesn't matter
1585 */
1586static void
1587cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1588 struct cfq_io_context *cic)
1589{
Jens Axboe3b181522005-06-27 10:56:24 +02001590 int enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001591
1592 if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1593 enable_idle = 0;
1594 else if (sample_valid(cic->ttime_samples)) {
1595 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1596 enable_idle = 0;
1597 else
1598 enable_idle = 1;
1599 }
1600
Jens Axboe3b181522005-06-27 10:56:24 +02001601 if (enable_idle)
1602 cfq_mark_cfqq_idle_window(cfqq);
1603 else
1604 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001605}
1606
1607
1608/*
1609 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1610 * no or if we aren't sure, a 1 will cause a preempt.
1611 */
1612static int
1613cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1614 struct cfq_rq *crq)
1615{
1616 struct cfq_queue *cfqq = cfqd->active_queue;
1617
1618 if (cfq_class_idle(new_cfqq))
1619 return 0;
1620
1621 if (!cfqq)
1622 return 1;
1623
1624 if (cfq_class_idle(cfqq))
1625 return 1;
Jens Axboe3b181522005-06-27 10:56:24 +02001626 if (!cfq_cfqq_wait_request(new_cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001627 return 0;
1628 /*
1629 * if it doesn't have slice left, forget it
1630 */
1631 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1632 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02001633 if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001634 return 1;
1635
1636 return 0;
1637}
1638
1639/*
1640 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1641 * let it have half of its nominal slice.
1642 */
1643static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1644{
1645 struct cfq_queue *__cfqq, *next;
1646
1647 list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1648 cfq_resort_rr_list(__cfqq, 1);
1649
1650 if (!cfqq->slice_left)
1651 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1652
1653 cfqq->slice_end = cfqq->slice_left + jiffies;
Jens Axboe3b181522005-06-27 10:56:24 +02001654 __cfq_slice_expired(cfqd, cfqq, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001655 __cfq_set_active_queue(cfqd, cfqq);
1656}
1657
1658/*
1659 * should really be a ll_rw_blk.c helper
1660 */
1661static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1662{
1663 request_queue_t *q = cfqd->queue;
1664
1665 if (!blk_queue_plugged(q))
1666 q->request_fn(q);
1667 else
1668 __generic_unplug_device(q);
1669}
1670
1671/*
1672 * Called when a new fs request (crq) is added (to cfqq). Check if there's
1673 * something we should do about it
1674 */
1675static void
1676cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1677 struct cfq_rq *crq)
1678{
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001679 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02001680
1681 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1682
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001683 /*
1684 * we never wait for an async request and we don't allow preemption
1685 * of an async request. so just return early
1686 */
1687 if (!cfq_crq_is_sync(crq))
1688 return;
Jens Axboe22e2c502005-06-27 10:55:12 +02001689
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001690 cic = crq->io_context;
Jens Axboe22e2c502005-06-27 10:55:12 +02001691
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001692 cfq_update_io_thinktime(cfqd, cic);
1693 cfq_update_idle_window(cfqd, cfqq, cic);
1694
1695 cic->last_queue = jiffies;
Jens Axboe22e2c502005-06-27 10:55:12 +02001696
1697 if (cfqq == cfqd->active_queue) {
1698 /*
1699 * if we are waiting for a request for this queue, let it rip
1700 * immediately and flag that we must not expire this queue
1701 * just now
1702 */
Jens Axboe3b181522005-06-27 10:56:24 +02001703 if (cfq_cfqq_wait_request(cfqq)) {
1704 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001705 del_timer(&cfqd->idle_slice_timer);
1706 cfq_start_queueing(cfqd, cfqq);
1707 }
1708 } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1709 /*
1710 * not the active queue - expire current slice if it is
1711 * idle and has expired it's mean thinktime or this new queue
1712 * has some old slice time left and is of higher priority
1713 */
1714 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001715 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001716 cfq_start_queueing(cfqd, cfqq);
1717 }
1718}
1719
Jens Axboeb4878f22005-10-20 16:42:29 +02001720static void cfq_insert_request(request_queue_t *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001721{
Jens Axboeb4878f22005-10-20 16:42:29 +02001722 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe22e2c502005-06-27 10:55:12 +02001723 struct cfq_rq *crq = RQ_DATA(rq);
1724 struct cfq_queue *cfqq = crq->cfq_queue;
1725
1726 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727
1728 cfq_add_crq_rb(crq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729
Jens Axboe22e2c502005-06-27 10:55:12 +02001730 list_add_tail(&rq->queuelist, &cfqq->fifo);
1731
Tejun Heo98b11472005-10-20 16:46:54 +02001732 if (rq_mergeable(rq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001733 cfq_add_crq_hash(cfqd, crq);
1734
Jens Axboe22e2c502005-06-27 10:55:12 +02001735 cfq_crq_enqueued(cfqd, cfqq, crq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736}
1737
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738static void cfq_completed_request(request_queue_t *q, struct request *rq)
1739{
1740 struct cfq_rq *crq = RQ_DATA(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001741 struct cfq_queue *cfqq = crq->cfq_queue;
1742 struct cfq_data *cfqd = cfqq->cfqd;
1743 const int sync = cfq_crq_is_sync(crq);
1744 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745
Jens Axboeb4878f22005-10-20 16:42:29 +02001746 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747
Jens Axboeb4878f22005-10-20 16:42:29 +02001748 WARN_ON(!cfqd->rq_in_driver);
1749 WARN_ON(!cfqq->on_dispatch[sync]);
1750 cfqd->rq_in_driver--;
1751 cfqq->on_dispatch[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752
Jens Axboeb4878f22005-10-20 16:42:29 +02001753 if (!cfq_class_idle(cfqq))
1754 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001755
Jens Axboeb4878f22005-10-20 16:42:29 +02001756 if (!cfq_cfqq_dispatched(cfqq)) {
1757 if (cfq_cfqq_on_rr(cfqq)) {
1758 cfqq->service_last = now;
1759 cfq_resort_rr_list(cfqq, 0);
1760 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001761 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762 }
1763
Jens Axboeb4878f22005-10-20 16:42:29 +02001764 if (cfq_crq_is_sync(crq))
1765 crq->io_context->last_end_request = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766}
1767
1768static struct request *
1769cfq_former_request(request_queue_t *q, struct request *rq)
1770{
1771 struct cfq_rq *crq = RQ_DATA(rq);
1772 struct rb_node *rbprev = rb_prev(&crq->rb_node);
1773
1774 if (rbprev)
1775 return rb_entry_crq(rbprev)->request;
1776
1777 return NULL;
1778}
1779
1780static struct request *
1781cfq_latter_request(request_queue_t *q, struct request *rq)
1782{
1783 struct cfq_rq *crq = RQ_DATA(rq);
1784 struct rb_node *rbnext = rb_next(&crq->rb_node);
1785
1786 if (rbnext)
1787 return rb_entry_crq(rbnext)->request;
1788
1789 return NULL;
1790}
1791
Jens Axboe22e2c502005-06-27 10:55:12 +02001792/*
1793 * we temporarily boost lower priority queues if they are holding fs exclusive
1794 * resources. they are boosted to normal prio (CLASS_BE/4)
1795 */
1796static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797{
Jens Axboe22e2c502005-06-27 10:55:12 +02001798 const int ioprio_class = cfqq->ioprio_class;
1799 const int ioprio = cfqq->ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800
Jens Axboe22e2c502005-06-27 10:55:12 +02001801 if (has_fs_excl()) {
1802 /*
1803 * boost idle prio on transactions that would lock out other
1804 * users of the filesystem
1805 */
1806 if (cfq_class_idle(cfqq))
1807 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1808 if (cfqq->ioprio > IOPRIO_NORM)
1809 cfqq->ioprio = IOPRIO_NORM;
1810 } else {
1811 /*
1812 * check if we need to unboost the queue
1813 */
1814 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1815 cfqq->ioprio_class = cfqq->org_ioprio_class;
1816 if (cfqq->ioprio != cfqq->org_ioprio)
1817 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818 }
1819
Jens Axboe22e2c502005-06-27 10:55:12 +02001820 /*
1821 * refile between round-robin lists if we moved the priority class
1822 */
1823 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
Jens Axboe3b181522005-06-27 10:56:24 +02001824 cfq_cfqq_on_rr(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001825 cfq_resort_rr_list(cfqq, 0);
1826}
1827
1828static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1829{
1830 if (rw == READ || process_sync(task))
1831 return task->pid;
1832
1833 return CFQ_KEY_ASYNC;
1834}
1835
1836static inline int
1837__cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1838 struct task_struct *task, int rw)
1839{
Jens Axboe3b181522005-06-27 10:56:24 +02001840#if 1
1841 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001842 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001843 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001844 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001845 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001846
1847 return ELV_MQUEUE_MAY;
Jens Axboe3b181522005-06-27 10:56:24 +02001848#else
Jens Axboe22e2c502005-06-27 10:55:12 +02001849 if (!cfqq || task->flags & PF_MEMALLOC)
1850 return ELV_MQUEUE_MAY;
Jens Axboe3b181522005-06-27 10:56:24 +02001851 if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1852 if (cfq_cfqq_wait_request(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001853 return ELV_MQUEUE_MUST;
1854
1855 /*
1856 * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1857 * can quickly flood the queue with writes from a single task
1858 */
Andrew Morton99f95e52005-06-27 20:14:05 -07001859 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001860 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001861 return ELV_MQUEUE_MUST;
1862 }
1863
1864 return ELV_MQUEUE_MAY;
1865 }
1866 if (cfq_class_idle(cfqq))
1867 return ELV_MQUEUE_NO;
1868 if (cfqq->allocated[rw] >= cfqd->max_queued) {
1869 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1870 int ret = ELV_MQUEUE_NO;
1871
1872 if (ioc && ioc->nr_batch_requests)
1873 ret = ELV_MQUEUE_MAY;
1874
1875 put_io_context(ioc);
1876 return ret;
1877 }
1878
1879 return ELV_MQUEUE_MAY;
1880#endif
1881}
1882
1883static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1884{
1885 struct cfq_data *cfqd = q->elevator->elevator_data;
1886 struct task_struct *tsk = current;
1887 struct cfq_queue *cfqq;
1888
1889 /*
1890 * don't force setup of a queue from here, as a call to may_queue
1891 * does not necessarily imply that a request actually will be queued.
1892 * so just lookup a possibly existing queue, or return 'may queue'
1893 * if that fails
1894 */
Jens Axboe3b181522005-06-27 10:56:24 +02001895 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001896 if (cfqq) {
1897 cfq_init_prio_data(cfqq);
1898 cfq_prio_boost(cfqq);
1899
1900 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1901 }
1902
1903 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904}
1905
1906static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1907{
Jens Axboe22e2c502005-06-27 10:55:12 +02001908 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909 struct request_list *rl = &q->rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910
Jens Axboe22e2c502005-06-27 10:55:12 +02001911 if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1912 smp_mb();
1913 if (waitqueue_active(&rl->wait[READ]))
1914 wake_up(&rl->wait[READ]);
1915 }
1916
1917 if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1918 smp_mb();
1919 if (waitqueue_active(&rl->wait[WRITE]))
1920 wake_up(&rl->wait[WRITE]);
1921 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922}
1923
1924/*
1925 * queue lock held here
1926 */
1927static void cfq_put_request(request_queue_t *q, struct request *rq)
1928{
1929 struct cfq_data *cfqd = q->elevator->elevator_data;
1930 struct cfq_rq *crq = RQ_DATA(rq);
1931
1932 if (crq) {
1933 struct cfq_queue *cfqq = crq->cfq_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001934 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935
Jens Axboe22e2c502005-06-27 10:55:12 +02001936 BUG_ON(!cfqq->allocated[rw]);
1937 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938
Jens Axboe22e2c502005-06-27 10:55:12 +02001939 put_io_context(crq->io_context->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940
1941 mempool_free(crq, cfqd->crq_pool);
1942 rq->elevator_private = NULL;
1943
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944 cfq_check_waiters(q, cfqq);
1945 cfq_put_queue(cfqq);
1946 }
1947}
1948
1949/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001950 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001952static int
1953cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
Al Viro8267e262005-10-21 03:20:53 -04001954 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955{
1956 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe3b181522005-06-27 10:56:24 +02001957 struct task_struct *tsk = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 struct cfq_io_context *cic;
1959 const int rw = rq_data_dir(rq);
Jens Axboe3b181522005-06-27 10:56:24 +02001960 pid_t key = cfq_queue_pid(tsk, rw);
Jens Axboe22e2c502005-06-27 10:55:12 +02001961 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 struct cfq_rq *crq;
1963 unsigned long flags;
Al Viro12a05732006-03-18 13:38:01 -05001964 int is_sync = key != CFQ_KEY_ASYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965
1966 might_sleep_if(gfp_mask & __GFP_WAIT);
1967
Jens Axboe3b181522005-06-27 10:56:24 +02001968 cic = cfq_get_io_context(cfqd, key, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001969
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970 spin_lock_irqsave(q->queue_lock, flags);
1971
Jens Axboe22e2c502005-06-27 10:55:12 +02001972 if (!cic)
1973 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974
Al Viro12a05732006-03-18 13:38:01 -05001975 if (!cic->cfqq[is_sync]) {
Al Viro6f325a12006-03-18 14:58:37 -05001976 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001977 if (!cfqq)
1978 goto queue_fail;
1979
Al Viro12a05732006-03-18 13:38:01 -05001980 cic->cfqq[is_sync] = cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001981 } else
Al Viro12a05732006-03-18 13:38:01 -05001982 cfqq = cic->cfqq[is_sync];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983
1984 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02001985 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001986 cfqd->rq_starved = 0;
1987 atomic_inc(&cfqq->ref);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 spin_unlock_irqrestore(q->queue_lock, flags);
1989
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990 crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1991 if (crq) {
1992 RB_CLEAR(&crq->rb_node);
1993 crq->rb_key = 0;
1994 crq->request = rq;
1995 INIT_HLIST_NODE(&crq->hash);
1996 crq->cfq_queue = cfqq;
1997 crq->io_context = cic;
Jens Axboe3b181522005-06-27 10:56:24 +02001998
Al Viro12a05732006-03-18 13:38:01 -05001999 if (is_sync)
Jens Axboe3b181522005-06-27 10:56:24 +02002000 cfq_mark_crq_is_sync(crq);
2001 else
2002 cfq_clear_crq_is_sync(crq);
2003
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004 rq->elevator_private = crq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005 return 0;
2006 }
2007
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008 spin_lock_irqsave(q->queue_lock, flags);
2009 cfqq->allocated[rw]--;
Jens Axboe22e2c502005-06-27 10:55:12 +02002010 if (!(cfqq->allocated[0] + cfqq->allocated[1]))
Jens Axboe3b181522005-06-27 10:56:24 +02002011 cfq_mark_cfqq_must_alloc(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012 cfq_put_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002013queue_fail:
2014 if (cic)
2015 put_io_context(cic->ioc);
2016 /*
2017 * mark us rq allocation starved. we need to kickstart the process
2018 * ourselves if there are no pending requests that can do it for us.
2019 * that would be an extremely rare OOM situation
2020 */
2021 cfqd->rq_starved = 1;
Jens Axboe3b181522005-06-27 10:56:24 +02002022 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 spin_unlock_irqrestore(q->queue_lock, flags);
2024 return 1;
2025}
2026
Jens Axboe22e2c502005-06-27 10:55:12 +02002027static void cfq_kick_queue(void *data)
2028{
2029 request_queue_t *q = data;
2030 struct cfq_data *cfqd = q->elevator->elevator_data;
2031 unsigned long flags;
2032
2033 spin_lock_irqsave(q->queue_lock, flags);
2034
2035 if (cfqd->rq_starved) {
2036 struct request_list *rl = &q->rq;
2037
2038 /*
2039 * we aren't guaranteed to get a request after this, but we
2040 * have to be opportunistic
2041 */
2042 smp_mb();
2043 if (waitqueue_active(&rl->wait[READ]))
2044 wake_up(&rl->wait[READ]);
2045 if (waitqueue_active(&rl->wait[WRITE]))
2046 wake_up(&rl->wait[WRITE]);
2047 }
2048
2049 blk_remove_plug(q);
2050 q->request_fn(q);
2051 spin_unlock_irqrestore(q->queue_lock, flags);
2052}
2053
2054/*
2055 * Timer running if the active_queue is currently idling inside its time slice
2056 */
2057static void cfq_idle_slice_timer(unsigned long data)
2058{
2059 struct cfq_data *cfqd = (struct cfq_data *) data;
2060 struct cfq_queue *cfqq;
2061 unsigned long flags;
2062
2063 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2064
2065 if ((cfqq = cfqd->active_queue) != NULL) {
2066 unsigned long now = jiffies;
2067
2068 /*
2069 * expired
2070 */
2071 if (time_after(now, cfqq->slice_end))
2072 goto expire;
2073
2074 /*
2075 * only expire and reinvoke request handler, if there are
2076 * other queues with pending requests
2077 */
Jens Axboeb4878f22005-10-20 16:42:29 +02002078 if (!cfqd->busy_queues) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002079 cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2080 add_timer(&cfqd->idle_slice_timer);
2081 goto out_cont;
2082 }
2083
2084 /*
2085 * not expired and it has a request pending, let it dispatch
2086 */
2087 if (!RB_EMPTY(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002088 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002089 goto out_kick;
2090 }
2091 }
2092expire:
2093 cfq_slice_expired(cfqd, 0);
2094out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002095 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002096out_cont:
2097 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2098}
2099
2100/*
2101 * Timer running if an idle class queue is waiting for service
2102 */
2103static void cfq_idle_class_timer(unsigned long data)
2104{
2105 struct cfq_data *cfqd = (struct cfq_data *) data;
2106 unsigned long flags, end;
2107
2108 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2109
2110 /*
2111 * race with a non-idle queue, reset timer
2112 */
2113 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2114 if (!time_after_eq(jiffies, end)) {
2115 cfqd->idle_class_timer.expires = end;
2116 add_timer(&cfqd->idle_class_timer);
2117 } else
Jens Axboe3b181522005-06-27 10:56:24 +02002118 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002119
2120 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2121}
2122
Jens Axboe3b181522005-06-27 10:56:24 +02002123static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2124{
2125 del_timer_sync(&cfqd->idle_slice_timer);
2126 del_timer_sync(&cfqd->idle_class_timer);
2127 blk_sync_queue(cfqd->queue);
2128}
Jens Axboe22e2c502005-06-27 10:55:12 +02002129
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130static void cfq_exit_queue(elevator_t *e)
2131{
Jens Axboe22e2c502005-06-27 10:55:12 +02002132 struct cfq_data *cfqd = e->elevator_data;
Al Virod9ff4182006-03-18 13:51:22 -05002133 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002134
Jens Axboe3b181522005-06-27 10:56:24 +02002135 cfq_shutdown_timer_wq(cfqd);
Al Virod9ff4182006-03-18 13:51:22 -05002136 write_lock(&cfq_exit_lock);
2137 spin_lock_irq(q->queue_lock);
2138 if (cfqd->active_queue)
2139 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2140 while(!list_empty(&cfqd->cic_list)) {
2141 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2142 struct cfq_io_context,
2143 queue_list);
2144 if (cic->cfqq[ASYNC]) {
2145 cfq_put_queue(cic->cfqq[ASYNC]);
2146 cic->cfqq[ASYNC] = NULL;
2147 }
2148 if (cic->cfqq[SYNC]) {
2149 cfq_put_queue(cic->cfqq[SYNC]);
2150 cic->cfqq[SYNC] = NULL;
2151 }
2152 cic->key = NULL;
2153 list_del_init(&cic->queue_list);
2154 }
2155 spin_unlock_irq(q->queue_lock);
2156 write_unlock(&cfq_exit_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002157
2158 cfq_shutdown_timer_wq(cfqd);
2159
2160 mempool_destroy(cfqd->crq_pool);
2161 kfree(cfqd->crq_hash);
2162 kfree(cfqd->cfq_hash);
2163 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164}
2165
2166static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2167{
2168 struct cfq_data *cfqd;
2169 int i;
2170
2171 cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2172 if (!cfqd)
2173 return -ENOMEM;
2174
2175 memset(cfqd, 0, sizeof(*cfqd));
Jens Axboe22e2c502005-06-27 10:55:12 +02002176
2177 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2178 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2179
2180 INIT_LIST_HEAD(&cfqd->busy_rr);
2181 INIT_LIST_HEAD(&cfqd->cur_rr);
2182 INIT_LIST_HEAD(&cfqd->idle_rr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 INIT_LIST_HEAD(&cfqd->empty_list);
Al Virod9ff4182006-03-18 13:51:22 -05002184 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185
2186 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2187 if (!cfqd->crq_hash)
2188 goto out_crqhash;
2189
2190 cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2191 if (!cfqd->cfq_hash)
2192 goto out_cfqhash;
2193
2194 cfqd->crq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, crq_pool);
2195 if (!cfqd->crq_pool)
2196 goto out_crqpool;
2197
2198 for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2199 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2200 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2201 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2202
2203 e->elevator_data = cfqd;
2204
2205 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206
Jens Axboe22e2c502005-06-27 10:55:12 +02002207 cfqd->max_queued = q->nr_requests / 4;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 q->nr_batching = cfq_queued;
Jens Axboe22e2c502005-06-27 10:55:12 +02002209
2210 init_timer(&cfqd->idle_slice_timer);
2211 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2212 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2213
2214 init_timer(&cfqd->idle_class_timer);
2215 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2216 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2217
2218 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2219
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 cfqd->cfq_queued = cfq_queued;
2221 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002222 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2223 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 cfqd->cfq_back_max = cfq_back_max;
2225 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002226 cfqd->cfq_slice[0] = cfq_slice_async;
2227 cfqd->cfq_slice[1] = cfq_slice_sync;
2228 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2229 cfqd->cfq_slice_idle = cfq_slice_idle;
2230 cfqd->cfq_max_depth = cfq_max_depth;
Jens Axboe3b181522005-06-27 10:56:24 +02002231
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232 return 0;
2233out_crqpool:
2234 kfree(cfqd->cfq_hash);
2235out_cfqhash:
2236 kfree(cfqd->crq_hash);
2237out_crqhash:
2238 kfree(cfqd);
2239 return -ENOMEM;
2240}
2241
2242static void cfq_slab_kill(void)
2243{
2244 if (crq_pool)
2245 kmem_cache_destroy(crq_pool);
2246 if (cfq_pool)
2247 kmem_cache_destroy(cfq_pool);
2248 if (cfq_ioc_pool)
2249 kmem_cache_destroy(cfq_ioc_pool);
2250}
2251
2252static int __init cfq_slab_setup(void)
2253{
2254 crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2255 NULL, NULL);
2256 if (!crq_pool)
2257 goto fail;
2258
2259 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2260 NULL, NULL);
2261 if (!cfq_pool)
2262 goto fail;
2263
2264 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2265 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2266 if (!cfq_ioc_pool)
2267 goto fail;
2268
2269 return 0;
2270fail:
2271 cfq_slab_kill();
2272 return -ENOMEM;
2273}
2274
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275/*
2276 * sysfs parts below -->
2277 */
2278struct cfq_fs_entry {
2279 struct attribute attr;
2280 ssize_t (*show)(struct cfq_data *, char *);
2281 ssize_t (*store)(struct cfq_data *, const char *, size_t);
2282};
2283
2284static ssize_t
2285cfq_var_show(unsigned int var, char *page)
2286{
2287 return sprintf(page, "%d\n", var);
2288}
2289
2290static ssize_t
2291cfq_var_store(unsigned int *var, const char *page, size_t count)
2292{
2293 char *p = (char *) page;
2294
2295 *var = simple_strtoul(p, &p, 10);
2296 return count;
2297}
2298
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2300static ssize_t __FUNC(struct cfq_data *cfqd, char *page) \
2301{ \
2302 unsigned int __data = __VAR; \
2303 if (__CONV) \
2304 __data = jiffies_to_msecs(__data); \
2305 return cfq_var_show(__data, (page)); \
2306}
2307SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2308SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002309SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2310SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
2312SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002313SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2314SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2315SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2316SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2317SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318#undef SHOW_FUNCTION
2319
2320#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2321static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count) \
2322{ \
2323 unsigned int __data; \
2324 int ret = cfq_var_store(&__data, (page), count); \
2325 if (__data < (MIN)) \
2326 __data = (MIN); \
2327 else if (__data > (MAX)) \
2328 __data = (MAX); \
2329 if (__CONV) \
2330 *(__PTR) = msecs_to_jiffies(__data); \
2331 else \
2332 *(__PTR) = __data; \
2333 return ret; \
2334}
2335STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2336STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002337STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2338STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2340STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002341STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2342STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2343STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2344STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2345STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346#undef STORE_FUNCTION
2347
2348static struct cfq_fs_entry cfq_quantum_entry = {
2349 .attr = {.name = "quantum", .mode = S_IRUGO | S_IWUSR },
2350 .show = cfq_quantum_show,
2351 .store = cfq_quantum_store,
2352};
2353static struct cfq_fs_entry cfq_queued_entry = {
2354 .attr = {.name = "queued", .mode = S_IRUGO | S_IWUSR },
2355 .show = cfq_queued_show,
2356 .store = cfq_queued_store,
2357};
Jens Axboe22e2c502005-06-27 10:55:12 +02002358static struct cfq_fs_entry cfq_fifo_expire_sync_entry = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
Jens Axboe22e2c502005-06-27 10:55:12 +02002360 .show = cfq_fifo_expire_sync_show,
2361 .store = cfq_fifo_expire_sync_store,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362};
Jens Axboe22e2c502005-06-27 10:55:12 +02002363static struct cfq_fs_entry cfq_fifo_expire_async_entry = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
Jens Axboe22e2c502005-06-27 10:55:12 +02002365 .show = cfq_fifo_expire_async_show,
2366 .store = cfq_fifo_expire_async_store,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367};
2368static struct cfq_fs_entry cfq_back_max_entry = {
2369 .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
2370 .show = cfq_back_max_show,
2371 .store = cfq_back_max_store,
2372};
2373static struct cfq_fs_entry cfq_back_penalty_entry = {
2374 .attr = {.name = "back_seek_penalty", .mode = S_IRUGO | S_IWUSR },
2375 .show = cfq_back_penalty_show,
2376 .store = cfq_back_penalty_store,
2377};
Jens Axboe22e2c502005-06-27 10:55:12 +02002378static struct cfq_fs_entry cfq_slice_sync_entry = {
2379 .attr = {.name = "slice_sync", .mode = S_IRUGO | S_IWUSR },
2380 .show = cfq_slice_sync_show,
2381 .store = cfq_slice_sync_store,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382};
Jens Axboe22e2c502005-06-27 10:55:12 +02002383static struct cfq_fs_entry cfq_slice_async_entry = {
2384 .attr = {.name = "slice_async", .mode = S_IRUGO | S_IWUSR },
2385 .show = cfq_slice_async_show,
2386 .store = cfq_slice_async_store,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387};
Jens Axboe22e2c502005-06-27 10:55:12 +02002388static struct cfq_fs_entry cfq_slice_async_rq_entry = {
2389 .attr = {.name = "slice_async_rq", .mode = S_IRUGO | S_IWUSR },
2390 .show = cfq_slice_async_rq_show,
2391 .store = cfq_slice_async_rq_store,
2392};
2393static struct cfq_fs_entry cfq_slice_idle_entry = {
2394 .attr = {.name = "slice_idle", .mode = S_IRUGO | S_IWUSR },
2395 .show = cfq_slice_idle_show,
2396 .store = cfq_slice_idle_store,
2397};
2398static struct cfq_fs_entry cfq_max_depth_entry = {
2399 .attr = {.name = "max_depth", .mode = S_IRUGO | S_IWUSR },
2400 .show = cfq_max_depth_show,
2401 .store = cfq_max_depth_store,
2402};
Jens Axboe3b181522005-06-27 10:56:24 +02002403
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404static struct attribute *default_attrs[] = {
2405 &cfq_quantum_entry.attr,
2406 &cfq_queued_entry.attr,
Jens Axboe22e2c502005-06-27 10:55:12 +02002407 &cfq_fifo_expire_sync_entry.attr,
2408 &cfq_fifo_expire_async_entry.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409 &cfq_back_max_entry.attr,
2410 &cfq_back_penalty_entry.attr,
Jens Axboe22e2c502005-06-27 10:55:12 +02002411 &cfq_slice_sync_entry.attr,
2412 &cfq_slice_async_entry.attr,
2413 &cfq_slice_async_rq_entry.attr,
2414 &cfq_slice_idle_entry.attr,
2415 &cfq_max_depth_entry.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416 NULL,
2417};
2418
2419#define to_cfq(atr) container_of((atr), struct cfq_fs_entry, attr)
2420
2421static ssize_t
2422cfq_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
2423{
2424 elevator_t *e = container_of(kobj, elevator_t, kobj);
2425 struct cfq_fs_entry *entry = to_cfq(attr);
2426
2427 if (!entry->show)
Dmitry Torokhov6c1852a2005-04-29 01:26:06 -05002428 return -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429
2430 return entry->show(e->elevator_data, page);
2431}
2432
2433static ssize_t
2434cfq_attr_store(struct kobject *kobj, struct attribute *attr,
2435 const char *page, size_t length)
2436{
2437 elevator_t *e = container_of(kobj, elevator_t, kobj);
2438 struct cfq_fs_entry *entry = to_cfq(attr);
2439
2440 if (!entry->store)
Dmitry Torokhov6c1852a2005-04-29 01:26:06 -05002441 return -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442
2443 return entry->store(e->elevator_data, page, length);
2444}
2445
2446static struct sysfs_ops cfq_sysfs_ops = {
2447 .show = cfq_attr_show,
2448 .store = cfq_attr_store,
2449};
2450
2451static struct kobj_type cfq_ktype = {
2452 .sysfs_ops = &cfq_sysfs_ops,
2453 .default_attrs = default_attrs,
2454};
2455
2456static struct elevator_type iosched_cfq = {
2457 .ops = {
2458 .elevator_merge_fn = cfq_merge,
2459 .elevator_merged_fn = cfq_merged_request,
2460 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeb4878f22005-10-20 16:42:29 +02002461 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002463 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 .elevator_deactivate_req_fn = cfq_deactivate_request,
2465 .elevator_queue_empty_fn = cfq_queue_empty,
2466 .elevator_completed_req_fn = cfq_completed_request,
2467 .elevator_former_req_fn = cfq_former_request,
2468 .elevator_latter_req_fn = cfq_latter_request,
2469 .elevator_set_req_fn = cfq_set_request,
2470 .elevator_put_req_fn = cfq_put_request,
2471 .elevator_may_queue_fn = cfq_may_queue,
2472 .elevator_init_fn = cfq_init_queue,
2473 .elevator_exit_fn = cfq_exit_queue,
Al Viroe17a9482006-03-18 13:21:20 -05002474 .trim = cfq_trim,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475 },
2476 .elevator_ktype = &cfq_ktype,
2477 .elevator_name = "cfq",
2478 .elevator_owner = THIS_MODULE,
2479};
2480
2481static int __init cfq_init(void)
2482{
2483 int ret;
2484
Jens Axboe22e2c502005-06-27 10:55:12 +02002485 /*
2486 * could be 0 on HZ < 1000 setups
2487 */
2488 if (!cfq_slice_async)
2489 cfq_slice_async = 1;
2490 if (!cfq_slice_idle)
2491 cfq_slice_idle = 1;
2492
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 if (cfq_slab_setup())
2494 return -ENOMEM;
2495
2496 ret = elv_register(&iosched_cfq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002497 if (ret)
2498 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500 return ret;
2501}
2502
2503static void __exit cfq_exit(void)
2504{
Al Viro334e94d2006-03-18 15:05:53 -05002505 DECLARE_COMPLETION(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002507 ioc_gone = &all_gone;
2508 barrier();
2509 if (atomic_read(&ioc_count))
2510 complete(ioc_gone);
2511 synchronize_rcu();
Christoph Hellwig83521d32005-10-30 15:01:39 -08002512 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513}
2514
2515module_init(cfq_init);
2516module_exit(cfq_exit);
2517
2518MODULE_AUTHOR("Jens Axboe");
2519MODULE_LICENSE("GPL");
2520MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");