blob: 052b17487625f0edf905feb1beadaec7c8e5fa96 [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 Axboe206dc692006-03-28 13:03:44 +020029static int cfq_slice_idle = HZ / 70;
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 Axboe22e2c502005-06-27 10:55:12 +020035
Jens Axboe3793c652006-05-30 21:11:04 +020036static DEFINE_SPINLOCK(cfq_exit_lock);
Al Viroa6a07632006-03-18 13:26:44 -050037
Linus Torvalds1da177e2005-04-16 15:20:36 -070038/*
39 * for the hash of cfqq inside the cfqd
40 */
41#define CFQ_QHASH_SHIFT 6
42#define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
43#define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
44
45/*
46 * for the hash of crq inside the cfqq
47 */
48#define CFQ_MHASH_SHIFT 6
49#define CFQ_MHASH_BLOCK(sec) ((sec) >> 3)
50#define CFQ_MHASH_ENTRIES (1 << CFQ_MHASH_SHIFT)
51#define CFQ_MHASH_FN(sec) hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
52#define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
53#define list_entry_hash(ptr) hlist_entry((ptr), struct cfq_rq, hash)
54
55#define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
Jens Axboe22e2c502005-06-27 10:55:12 +020056#define list_entry_fifo(ptr) list_entry((ptr), struct request, queuelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -070057
58#define RQ_DATA(rq) (rq)->elevator_private
59
60/*
61 * rb-tree defines
62 */
63#define RB_NONE (2)
64#define RB_EMPTY(node) ((node)->rb_node == NULL)
65#define RB_CLEAR_COLOR(node) (node)->rb_color = RB_NONE
66#define RB_CLEAR(node) do { \
67 (node)->rb_parent = NULL; \
68 RB_CLEAR_COLOR((node)); \
69 (node)->rb_right = NULL; \
70 (node)->rb_left = NULL; \
71} while (0)
72#define RB_CLEAR_ROOT(root) ((root)->rb_node = NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -070073#define rb_entry_crq(node) rb_entry((node), struct cfq_rq, rb_node)
74#define rq_rb_key(rq) (rq)->sector
75
Linus Torvalds1da177e2005-04-16 15:20:36 -070076static kmem_cache_t *crq_pool;
77static kmem_cache_t *cfq_pool;
78static kmem_cache_t *cfq_ioc_pool;
79
Al Viro334e94d2006-03-18 15:05:53 -050080static atomic_t ioc_count = ATOMIC_INIT(0);
81static struct completion *ioc_gone;
82
Jens Axboe22e2c502005-06-27 10:55:12 +020083#define CFQ_PRIO_LISTS IOPRIO_BE_NR
84#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
85#define cfq_class_be(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
86#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
87
Jens Axboe3b181522005-06-27 10:56:24 +020088#define ASYNC (0)
89#define SYNC (1)
90
91#define cfq_cfqq_dispatched(cfqq) \
92 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
93
94#define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
95
96#define cfq_cfqq_sync(cfqq) \
97 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
Jens Axboe22e2c502005-06-27 10:55:12 +020098
Jens Axboe206dc692006-03-28 13:03:44 +020099#define sample_valid(samples) ((samples) > 80)
100
Jens Axboe22e2c502005-06-27 10:55:12 +0200101/*
102 * Per block device queue structure
103 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104struct cfq_data {
Jens Axboe22e2c502005-06-27 10:55:12 +0200105 request_queue_t *queue;
106
107 /*
108 * rr list of queues with requests and the count of them
109 */
110 struct list_head rr_list[CFQ_PRIO_LISTS];
111 struct list_head busy_rr;
112 struct list_head cur_rr;
113 struct list_head idle_rr;
114 unsigned int busy_queues;
115
116 /*
117 * non-ordered list of empty cfqq's
118 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119 struct list_head empty_list;
120
Jens Axboe22e2c502005-06-27 10:55:12 +0200121 /*
122 * cfqq lookup hash
123 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 struct hlist_head *cfq_hash;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125
Jens Axboe22e2c502005-06-27 10:55:12 +0200126 /*
127 * global crq hash for all queues
128 */
129 struct hlist_head *crq_hash;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130
131 unsigned int max_queued;
132
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 mempool_t *crq_pool;
134
Jens Axboe22e2c502005-06-27 10:55:12 +0200135 int rq_in_driver;
Jens Axboe25776e32006-06-01 10:12:26 +0200136 int hw_tag;
Jens Axboe22e2c502005-06-27 10:55:12 +0200137
138 /*
139 * schedule slice state info
140 */
141 /*
142 * idle window management
143 */
144 struct timer_list idle_slice_timer;
145 struct work_struct unplug_work;
146
147 struct cfq_queue *active_queue;
148 struct cfq_io_context *active_cic;
149 int cur_prio, cur_end_prio;
150 unsigned int dispatch_slice;
151
152 struct timer_list idle_class_timer;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153
154 sector_t last_sector;
Jens Axboe22e2c502005-06-27 10:55:12 +0200155 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156
Jens Axboe22e2c502005-06-27 10:55:12 +0200157 unsigned int rq_starved;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158
159 /*
160 * tunables, see top of file
161 */
162 unsigned int cfq_quantum;
163 unsigned int cfq_queued;
Jens Axboe22e2c502005-06-27 10:55:12 +0200164 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165 unsigned int cfq_back_penalty;
166 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200167 unsigned int cfq_slice[2];
168 unsigned int cfq_slice_async_rq;
169 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500170
171 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172};
173
Jens Axboe22e2c502005-06-27 10:55:12 +0200174/*
175 * Per process-grouping structure
176 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177struct cfq_queue {
178 /* reference count */
179 atomic_t ref;
180 /* parent cfq_data */
181 struct cfq_data *cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200182 /* cfqq lookup hash */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 struct hlist_node cfq_hash;
184 /* hash key */
Jens Axboe22e2c502005-06-27 10:55:12 +0200185 unsigned int key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186 /* on either rr or empty list of cfqd */
187 struct list_head cfq_list;
188 /* sorted list of pending requests */
189 struct rb_root sort_list;
190 /* if fifo isn't expired, next request to serve */
191 struct cfq_rq *next_crq;
192 /* requests queued in sort_list */
193 int queued[2];
194 /* currently allocated requests */
195 int allocated[2];
196 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200197 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198
Jens Axboe22e2c502005-06-27 10:55:12 +0200199 unsigned long slice_start;
200 unsigned long slice_end;
201 unsigned long slice_left;
202 unsigned long service_last;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203
Jens Axboe3b181522005-06-27 10:56:24 +0200204 /* number of requests that are on the dispatch list */
205 int on_dispatch[2];
Jens Axboe22e2c502005-06-27 10:55:12 +0200206
207 /* io prio of this group */
208 unsigned short ioprio, org_ioprio;
209 unsigned short ioprio_class, org_ioprio_class;
210
Jens Axboe3b181522005-06-27 10:56:24 +0200211 /* various state flags, see below */
212 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213};
214
215struct cfq_rq {
216 struct rb_node rb_node;
217 sector_t rb_key;
218 struct request *request;
219 struct hlist_node hash;
220
221 struct cfq_queue *cfq_queue;
222 struct cfq_io_context *io_context;
223
Jens Axboe3b181522005-06-27 10:56:24 +0200224 unsigned int crq_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225};
226
Jens Axboe3b181522005-06-27 10:56:24 +0200227enum cfqq_state_flags {
228 CFQ_CFQQ_FLAG_on_rr = 0,
229 CFQ_CFQQ_FLAG_wait_request,
230 CFQ_CFQQ_FLAG_must_alloc,
231 CFQ_CFQQ_FLAG_must_alloc_slice,
232 CFQ_CFQQ_FLAG_must_dispatch,
233 CFQ_CFQQ_FLAG_fifo_expire,
234 CFQ_CFQQ_FLAG_idle_window,
235 CFQ_CFQQ_FLAG_prio_changed,
Jens Axboe3b181522005-06-27 10:56:24 +0200236};
237
238#define CFQ_CFQQ_FNS(name) \
239static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
240{ \
241 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
242} \
243static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
244{ \
245 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
246} \
247static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
248{ \
249 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
250}
251
252CFQ_CFQQ_FNS(on_rr);
253CFQ_CFQQ_FNS(wait_request);
254CFQ_CFQQ_FNS(must_alloc);
255CFQ_CFQQ_FNS(must_alloc_slice);
256CFQ_CFQQ_FNS(must_dispatch);
257CFQ_CFQQ_FNS(fifo_expire);
258CFQ_CFQQ_FNS(idle_window);
259CFQ_CFQQ_FNS(prio_changed);
Jens Axboe3b181522005-06-27 10:56:24 +0200260#undef CFQ_CFQQ_FNS
261
262enum cfq_rq_state_flags {
Jens Axboeb4878f22005-10-20 16:42:29 +0200263 CFQ_CRQ_FLAG_is_sync = 0,
Jens Axboe3b181522005-06-27 10:56:24 +0200264};
265
266#define CFQ_CRQ_FNS(name) \
267static inline void cfq_mark_crq_##name(struct cfq_rq *crq) \
268{ \
269 crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name); \
270} \
271static inline void cfq_clear_crq_##name(struct cfq_rq *crq) \
272{ \
273 crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name); \
274} \
275static inline int cfq_crq_##name(const struct cfq_rq *crq) \
276{ \
277 return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0; \
278}
279
Jens Axboe3b181522005-06-27 10:56:24 +0200280CFQ_CRQ_FNS(is_sync);
Jens Axboe3b181522005-06-27 10:56:24 +0200281#undef CFQ_CRQ_FNS
282
283static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
Jens Axboeb4878f22005-10-20 16:42:29 +0200284static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
Al Viro6f325a12006-03-18 14:58:37 -0500285static 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 -0700286
Jens Axboe22e2c502005-06-27 10:55:12 +0200287#define process_sync(tsk) ((tsk)->flags & PF_SYNCWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288
289/*
290 * lots of deadline iosched dupes, can be abstracted later...
291 */
292static inline void cfq_del_crq_hash(struct cfq_rq *crq)
293{
294 hlist_del_init(&crq->hash);
295}
296
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
298{
299 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
300
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
302}
303
304static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
305{
306 struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
307 struct hlist_node *entry, *next;
308
309 hlist_for_each_safe(entry, next, hash_list) {
310 struct cfq_rq *crq = list_entry_hash(entry);
311 struct request *__rq = crq->request;
312
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 if (!rq_mergeable(__rq)) {
314 cfq_del_crq_hash(crq);
315 continue;
316 }
317
318 if (rq_hash_key(__rq) == offset)
319 return __rq;
320 }
321
322 return NULL;
323}
324
Andrew Morton99f95e52005-06-27 20:14:05 -0700325/*
326 * scheduler run of queue, if there are requests pending and no one in the
327 * driver that will restart queueing
328 */
329static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
330{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100331 if (cfqd->busy_queues)
Andrew Morton99f95e52005-06-27 20:14:05 -0700332 kblockd_schedule_work(&cfqd->unplug_work);
333}
334
335static int cfq_queue_empty(request_queue_t *q)
336{
337 struct cfq_data *cfqd = q->elevator->elevator_data;
338
Jens Axboeb4878f22005-10-20 16:42:29 +0200339 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700340}
341
Jens Axboe206dc692006-03-28 13:03:44 +0200342static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
343{
344 if (rw == READ || process_sync(task))
345 return task->pid;
346
347 return CFQ_KEY_ASYNC;
348}
349
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350/*
351 * Lifted from AS - choose which of crq1 and crq2 that is best served now.
352 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200353 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 */
355static struct cfq_rq *
356cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
357{
358 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200360#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
361#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
362 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363
364 if (crq1 == NULL || crq1 == crq2)
365 return crq2;
366 if (crq2 == NULL)
367 return crq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200368
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200369 if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
370 return crq1;
371 else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
Jens Axboe22e2c502005-06-27 10:55:12 +0200372 return crq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
374 s1 = crq1->request->sector;
375 s2 = crq2->request->sector;
376
377 last = cfqd->last_sector;
378
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 /*
380 * by definition, 1KiB is 2 sectors
381 */
382 back_max = cfqd->cfq_back_max * 2;
383
384 /*
385 * Strict one way elevator _except_ in the case where we allow
386 * short backward seeks which are biased as twice the cost of a
387 * similar forward seek.
388 */
389 if (s1 >= last)
390 d1 = s1 - last;
391 else if (s1 + back_max >= last)
392 d1 = (last - s1) * cfqd->cfq_back_penalty;
393 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200394 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395
396 if (s2 >= last)
397 d2 = s2 - last;
398 else if (s2 + back_max >= last)
399 d2 = (last - s2) * cfqd->cfq_back_penalty;
400 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200401 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
403 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404
Andreas Mohre8a99052006-03-28 08:59:49 +0200405 /*
406 * By doing switch() on the bit mask "wrap" we avoid having to
407 * check two variables for all permutations: --> faster!
408 */
409 switch (wrap) {
410 case 0: /* common case for CFQ: crq1 and crq2 not wrapped */
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 case CFQ_RQ2_WRAP:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 return crq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200424 case CFQ_RQ1_WRAP:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 return crq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200426 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both crqs wrapped */
427 default:
428 /*
429 * Since both rqs are wrapped,
430 * start with the one that's further behind head
431 * (--> only *one* back seek required),
432 * since back seek takes more time than forward.
433 */
434 if (s1 <= s2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 return crq1;
436 else
437 return crq2;
438 }
439}
440
441/*
442 * would be nice to take fifo expire time into account as well
443 */
444static struct cfq_rq *
445cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
446 struct cfq_rq *last)
447{
448 struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
449 struct rb_node *rbnext, *rbprev;
450
Jens Axboeb4878f22005-10-20 16:42:29 +0200451 if (!(rbnext = rb_next(&last->rb_node))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 rbnext = rb_first(&cfqq->sort_list);
Jens Axboe22e2c502005-06-27 10:55:12 +0200453 if (rbnext == &last->rb_node)
454 rbnext = NULL;
455 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456
457 rbprev = rb_prev(&last->rb_node);
458
459 if (rbprev)
460 crq_prev = rb_entry_crq(rbprev);
461 if (rbnext)
462 crq_next = rb_entry_crq(rbnext);
463
464 return cfq_choose_req(cfqd, crq_next, crq_prev);
465}
466
467static void cfq_update_next_crq(struct cfq_rq *crq)
468{
469 struct cfq_queue *cfqq = crq->cfq_queue;
470
471 if (cfqq->next_crq == crq)
472 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
473}
474
Jens Axboe22e2c502005-06-27 10:55:12 +0200475static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476{
Jens Axboe22e2c502005-06-27 10:55:12 +0200477 struct cfq_data *cfqd = cfqq->cfqd;
478 struct list_head *list, *entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479
Jens Axboe3b181522005-06-27 10:56:24 +0200480 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481
482 list_del(&cfqq->cfq_list);
483
Jens Axboe22e2c502005-06-27 10:55:12 +0200484 if (cfq_class_rt(cfqq))
485 list = &cfqd->cur_rr;
486 else if (cfq_class_idle(cfqq))
487 list = &cfqd->idle_rr;
488 else {
489 /*
490 * if cfqq has requests in flight, don't allow it to be
491 * found in cfq_set_active_queue before it has finished them.
492 * this is done to increase fairness between a process that
493 * has lots of io pending vs one that only generates one
494 * sporadically or synchronously
495 */
Jens Axboe3b181522005-06-27 10:56:24 +0200496 if (cfq_cfqq_dispatched(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200497 list = &cfqd->busy_rr;
498 else
499 list = &cfqd->rr_list[cfqq->ioprio];
500 }
501
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200503 * if queue was preempted, just add to front to be fair. busy_rr
Jens Axboeb52a8342006-06-01 18:53:43 +0200504 * isn't sorted, but insert at the back for fairness.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200506 if (preempted || list == &cfqd->busy_rr) {
Jens Axboeb52a8342006-06-01 18:53:43 +0200507 if (preempted)
508 list = list->prev;
509
510 list_add_tail(&cfqq->cfq_list, list);
Jens Axboe22e2c502005-06-27 10:55:12 +0200511 return;
512 }
513
514 /*
515 * sort by when queue was last serviced
516 */
517 entry = list;
518 while ((entry = entry->prev) != list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
520
Jens Axboe22e2c502005-06-27 10:55:12 +0200521 if (!__cfqq->service_last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 break;
Jens Axboe22e2c502005-06-27 10:55:12 +0200523 if (time_before(__cfqq->service_last, cfqq->service_last))
524 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 }
526
527 list_add(&cfqq->cfq_list, entry);
528}
529
530/*
531 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200532 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 */
534static inline void
Jens Axboeb4878f22005-10-20 16:42:29 +0200535cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536{
Jens Axboe3b181522005-06-27 10:56:24 +0200537 BUG_ON(cfq_cfqq_on_rr(cfqq));
538 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539 cfqd->busy_queues++;
540
Jens Axboeb4878f22005-10-20 16:42:29 +0200541 cfq_resort_rr_list(cfqq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542}
543
544static inline void
545cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
546{
Jens Axboe3b181522005-06-27 10:56:24 +0200547 BUG_ON(!cfq_cfqq_on_rr(cfqq));
548 cfq_clear_cfqq_on_rr(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200549 list_move(&cfqq->cfq_list, &cfqd->empty_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550
551 BUG_ON(!cfqd->busy_queues);
552 cfqd->busy_queues--;
553}
554
555/*
556 * rb tree support functions
557 */
558static inline void cfq_del_crq_rb(struct cfq_rq *crq)
559{
560 struct cfq_queue *cfqq = crq->cfq_queue;
Jens Axboeb4878f22005-10-20 16:42:29 +0200561 struct cfq_data *cfqd = cfqq->cfqd;
562 const int sync = cfq_crq_is_sync(crq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563
Jens Axboeb4878f22005-10-20 16:42:29 +0200564 BUG_ON(!cfqq->queued[sync]);
565 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566
Jens Axboeb4878f22005-10-20 16:42:29 +0200567 cfq_update_next_crq(crq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568
Jens Axboeb4878f22005-10-20 16:42:29 +0200569 rb_erase(&crq->rb_node, &cfqq->sort_list);
570 RB_CLEAR_COLOR(&crq->rb_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571
Jens Axboeb4878f22005-10-20 16:42:29 +0200572 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
573 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574}
575
576static struct cfq_rq *
577__cfq_add_crq_rb(struct cfq_rq *crq)
578{
579 struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
580 struct rb_node *parent = NULL;
581 struct cfq_rq *__crq;
582
583 while (*p) {
584 parent = *p;
585 __crq = rb_entry_crq(parent);
586
587 if (crq->rb_key < __crq->rb_key)
588 p = &(*p)->rb_left;
589 else if (crq->rb_key > __crq->rb_key)
590 p = &(*p)->rb_right;
591 else
592 return __crq;
593 }
594
595 rb_link_node(&crq->rb_node, parent, p);
596 return NULL;
597}
598
599static void cfq_add_crq_rb(struct cfq_rq *crq)
600{
601 struct cfq_queue *cfqq = crq->cfq_queue;
602 struct cfq_data *cfqd = cfqq->cfqd;
603 struct request *rq = crq->request;
604 struct cfq_rq *__alias;
605
606 crq->rb_key = rq_rb_key(rq);
Jens Axboe3b181522005-06-27 10:56:24 +0200607 cfqq->queued[cfq_crq_is_sync(crq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608
609 /*
610 * looks a little odd, but the first insert might return an alias.
611 * if that happens, put the alias on the dispatch list
612 */
613 while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
Jens Axboeb4878f22005-10-20 16:42:29 +0200614 cfq_dispatch_insert(cfqd->queue, __alias);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615
616 rb_insert_color(&crq->rb_node, &cfqq->sort_list);
617
Jens Axboe3b181522005-06-27 10:56:24 +0200618 if (!cfq_cfqq_on_rr(cfqq))
Jens Axboeb4878f22005-10-20 16:42:29 +0200619 cfq_add_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620
621 /*
622 * check if this request is a better next-serve candidate
623 */
624 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
625}
626
627static inline void
628cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
629{
Jens Axboeb4878f22005-10-20 16:42:29 +0200630 rb_erase(&crq->rb_node, &cfqq->sort_list);
631 cfqq->queued[cfq_crq_is_sync(crq)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632
633 cfq_add_crq_rb(crq);
634}
635
Jens Axboe206dc692006-03-28 13:03:44 +0200636static struct request *
637cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638{
Jens Axboe206dc692006-03-28 13:03:44 +0200639 struct task_struct *tsk = current;
640 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
641 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642 struct rb_node *n;
Jens Axboe206dc692006-03-28 13:03:44 +0200643 sector_t sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644
Jens Axboe206dc692006-03-28 13:03:44 +0200645 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646 if (!cfqq)
647 goto out;
648
Jens Axboe206dc692006-03-28 13:03:44 +0200649 sector = bio->bi_sector + bio_sectors(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650 n = cfqq->sort_list.rb_node;
651 while (n) {
652 struct cfq_rq *crq = rb_entry_crq(n);
653
654 if (sector < crq->rb_key)
655 n = n->rb_left;
656 else if (sector > crq->rb_key)
657 n = n->rb_right;
658 else
659 return crq->request;
660 }
661
662out:
663 return NULL;
664}
665
Jens Axboeb4878f22005-10-20 16:42:29 +0200666static void cfq_activate_request(request_queue_t *q, struct request *rq)
667{
668 struct cfq_data *cfqd = q->elevator->elevator_data;
669
670 cfqd->rq_in_driver++;
Jens Axboe25776e32006-06-01 10:12:26 +0200671
672 /*
673 * If the depth is larger 1, it really could be queueing. But lets
674 * make the mark a little higher - idling could still be good for
675 * low queueing, and a low queueing number could also just indicate
676 * a SCSI mid layer like behaviour where limit+1 is often seen.
677 */
678 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
679 cfqd->hw_tag = 1;
Jens Axboeb4878f22005-10-20 16:42:29 +0200680}
681
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
683{
Jens Axboe22e2c502005-06-27 10:55:12 +0200684 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685
Jens Axboeb4878f22005-10-20 16:42:29 +0200686 WARN_ON(!cfqd->rq_in_driver);
687 cfqd->rq_in_driver--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688}
689
Jens Axboeb4878f22005-10-20 16:42:29 +0200690static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691{
692 struct cfq_rq *crq = RQ_DATA(rq);
693
Jens Axboeb4878f22005-10-20 16:42:29 +0200694 list_del_init(&rq->queuelist);
695 cfq_del_crq_rb(crq);
Tejun Heo98b11472005-10-20 16:46:54 +0200696 cfq_del_crq_hash(crq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697}
698
699static int
700cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
701{
702 struct cfq_data *cfqd = q->elevator->elevator_data;
703 struct request *__rq;
704 int ret;
705
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
Jens Axboe22e2c502005-06-27 10:55:12 +0200707 if (__rq && elv_rq_merge_ok(__rq, bio)) {
708 ret = ELEVATOR_BACK_MERGE;
709 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 }
711
Jens Axboe206dc692006-03-28 13:03:44 +0200712 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200713 if (__rq && elv_rq_merge_ok(__rq, bio)) {
714 ret = ELEVATOR_FRONT_MERGE;
715 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 }
717
718 return ELEVATOR_NO_MERGE;
719out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 *req = __rq;
721 return ret;
722}
723
724static void cfq_merged_request(request_queue_t *q, struct request *req)
725{
726 struct cfq_data *cfqd = q->elevator->elevator_data;
727 struct cfq_rq *crq = RQ_DATA(req);
728
729 cfq_del_crq_hash(crq);
730 cfq_add_crq_hash(cfqd, crq);
731
Jens Axboeb4878f22005-10-20 16:42:29 +0200732 if (rq_rb_key(req) != crq->rb_key) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 struct cfq_queue *cfqq = crq->cfq_queue;
734
735 cfq_update_next_crq(crq);
736 cfq_reposition_crq_rb(cfqq, crq);
737 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738}
739
740static void
741cfq_merged_requests(request_queue_t *q, struct request *rq,
742 struct request *next)
743{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 cfq_merged_request(q, rq);
745
Jens Axboe22e2c502005-06-27 10:55:12 +0200746 /*
747 * reposition in fifo if next is older than rq
748 */
749 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
750 time_before(next->start_time, rq->start_time))
751 list_move(&rq->queuelist, &next->queuelist);
752
Jens Axboeb4878f22005-10-20 16:42:29 +0200753 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200754}
755
756static inline void
757__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
758{
759 if (cfqq) {
760 /*
761 * stop potential idle class queues waiting service
762 */
763 del_timer(&cfqd->idle_class_timer);
764
765 cfqq->slice_start = jiffies;
766 cfqq->slice_end = 0;
767 cfqq->slice_left = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200768 cfq_clear_cfqq_must_alloc_slice(cfqq);
769 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200770 }
771
772 cfqd->active_queue = cfqq;
773}
774
775/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100776 * current cfqq expired its slice (or was too idle), select new one
777 */
778static void
779__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
780 int preempted)
781{
782 unsigned long now = jiffies;
783
784 if (cfq_cfqq_wait_request(cfqq))
785 del_timer(&cfqd->idle_slice_timer);
786
787 if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
788 cfqq->service_last = now;
789 cfq_schedule_dispatch(cfqd);
790 }
791
792 cfq_clear_cfqq_must_dispatch(cfqq);
793 cfq_clear_cfqq_wait_request(cfqq);
794
795 /*
796 * store what was left of this slice, if the queue idled out
797 * or was preempted
798 */
799 if (time_after(cfqq->slice_end, now))
800 cfqq->slice_left = cfqq->slice_end - now;
801 else
802 cfqq->slice_left = 0;
803
804 if (cfq_cfqq_on_rr(cfqq))
805 cfq_resort_rr_list(cfqq, preempted);
806
807 if (cfqq == cfqd->active_queue)
808 cfqd->active_queue = NULL;
809
810 if (cfqd->active_cic) {
811 put_io_context(cfqd->active_cic->ioc);
812 cfqd->active_cic = NULL;
813 }
814
815 cfqd->dispatch_slice = 0;
816}
817
818static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
819{
820 struct cfq_queue *cfqq = cfqd->active_queue;
821
822 if (cfqq)
823 __cfq_slice_expired(cfqd, cfqq, preempted);
824}
825
826/*
Jens Axboe22e2c502005-06-27 10:55:12 +0200827 * 0
828 * 0,1
829 * 0,1,2
830 * 0,1,2,3
831 * 0,1,2,3,4
832 * 0,1,2,3,4,5
833 * 0,1,2,3,4,5,6
834 * 0,1,2,3,4,5,6,7
835 */
836static int cfq_get_next_prio_level(struct cfq_data *cfqd)
837{
838 int prio, wrap;
839
840 prio = -1;
841 wrap = 0;
842 do {
843 int p;
844
845 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
846 if (!list_empty(&cfqd->rr_list[p])) {
847 prio = p;
848 break;
849 }
850 }
851
852 if (prio != -1)
853 break;
854 cfqd->cur_prio = 0;
855 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
856 cfqd->cur_end_prio = 0;
857 if (wrap)
858 break;
859 wrap = 1;
860 }
861 } while (1);
862
863 if (unlikely(prio == -1))
864 return -1;
865
866 BUG_ON(prio >= CFQ_PRIO_LISTS);
867
868 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
869
870 cfqd->cur_prio = prio + 1;
871 if (cfqd->cur_prio > cfqd->cur_end_prio) {
872 cfqd->cur_end_prio = cfqd->cur_prio;
873 cfqd->cur_prio = 0;
874 }
875 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
876 cfqd->cur_prio = 0;
877 cfqd->cur_end_prio = 0;
878 }
879
880 return prio;
881}
882
Jens Axboe3b181522005-06-27 10:56:24 +0200883static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200884{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100885 struct cfq_queue *cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +0200886
887 /*
888 * if current list is non-empty, grab first entry. if it is empty,
889 * get next prio level and grab first entry then if any are spliced
890 */
891 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
892 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
893
894 /*
Jens Axboee0de0202006-06-01 10:07:26 +0200895 * If no new queues are available, check if the busy list has some
896 * before falling back to idle io.
897 */
898 if (!cfqq && !list_empty(&cfqd->busy_rr))
899 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
900
901 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200902 * if we have idle queues and no rt or be queues had pending
903 * requests, either allow immediate service if the grace period
904 * has passed or arm the idle grace timer
905 */
906 if (!cfqq && !list_empty(&cfqd->idle_rr)) {
907 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
908
909 if (time_after_eq(jiffies, end))
910 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
911 else
912 mod_timer(&cfqd->idle_class_timer, end);
913 }
914
915 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200916 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200917}
918
Jens Axboe22e2c502005-06-27 10:55:12 +0200919static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
920
921{
Jens Axboe206dc692006-03-28 13:03:44 +0200922 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100923 unsigned long sl;
924
Jens Axboe22e2c502005-06-27 10:55:12 +0200925 WARN_ON(!RB_EMPTY(&cfqq->sort_list));
926 WARN_ON(cfqq != cfqd->active_queue);
927
928 /*
929 * idle is disabled, either manually or by past process history
930 */
931 if (!cfqd->cfq_slice_idle)
932 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200933 if (!cfq_cfqq_idle_window(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200934 return 0;
935 /*
936 * task has exited, don't wait
937 */
Jens Axboe206dc692006-03-28 13:03:44 +0200938 cic = cfqd->active_cic;
939 if (!cic || !cic->ioc->task)
Jens Axboe22e2c502005-06-27 10:55:12 +0200940 return 0;
941
Jens Axboe3b181522005-06-27 10:56:24 +0200942 cfq_mark_cfqq_must_dispatch(cfqq);
943 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200944
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100945 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
Jens Axboe206dc692006-03-28 13:03:44 +0200946
947 /*
948 * we don't want to idle for seeks, but we do want to allow
949 * fair distribution of slice time for a process doing back-to-back
950 * seeks. so allow a little bit of time for him to submit a new rq
951 */
952 if (sample_valid(cic->seek_samples) && cic->seek_mean > 131072)
953 sl = 2;
954
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100955 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe22e2c502005-06-27 10:55:12 +0200956 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957}
958
Jens Axboeb4878f22005-10-20 16:42:29 +0200959static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960{
961 struct cfq_data *cfqd = q->elevator->elevator_data;
962 struct cfq_queue *cfqq = crq->cfq_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +0200963
964 cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200965 cfq_remove_request(crq->request);
Jens Axboe3b181522005-06-27 10:56:24 +0200966 cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
Jens Axboeb4878f22005-10-20 16:42:29 +0200967 elv_dispatch_sort(q, crq->request);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968}
969
970/*
971 * return expired entry, or NULL to just start from scratch in rbtree
972 */
973static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
974{
975 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200976 struct request *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 struct cfq_rq *crq;
978
Jens Axboe3b181522005-06-27 10:56:24 +0200979 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980 return NULL;
981
Jens Axboe22e2c502005-06-27 10:55:12 +0200982 if (!list_empty(&cfqq->fifo)) {
Jens Axboe3b181522005-06-27 10:56:24 +0200983 int fifo = cfq_cfqq_class_sync(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984
Jens Axboe22e2c502005-06-27 10:55:12 +0200985 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
986 rq = crq->request;
987 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
Jens Axboe3b181522005-06-27 10:56:24 +0200988 cfq_mark_cfqq_fifo_expire(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200989 return crq;
990 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 }
992
993 return NULL;
994}
995
996/*
Jens Axboe3b181522005-06-27 10:56:24 +0200997 * Scale schedule slice based on io priority. Use the sync time slice only
998 * if a queue is marked sync and has sync io queued. A sync queue with async
999 * io only, should not get full sync slice length.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001001static inline int
1002cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003{
Jens Axboe22e2c502005-06-27 10:55:12 +02001004 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005
Jens Axboe22e2c502005-06-27 10:55:12 +02001006 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007
Jens Axboe22e2c502005-06-27 10:55:12 +02001008 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009}
1010
Jens Axboe22e2c502005-06-27 10:55:12 +02001011static inline void
1012cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1013{
1014 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
1015}
1016
1017static inline int
1018cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1019{
1020 const int base_rq = cfqd->cfq_slice_async_rq;
1021
1022 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1023
1024 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1025}
1026
1027/*
1028 * get next queue for service
1029 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001030static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001031{
1032 unsigned long now = jiffies;
1033 struct cfq_queue *cfqq;
1034
1035 cfqq = cfqd->active_queue;
1036 if (!cfqq)
1037 goto new_queue;
1038
1039 /*
1040 * slice has expired
1041 */
Jens Axboe3b181522005-06-27 10:56:24 +02001042 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1043 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +02001044
1045 /*
1046 * if queue has requests, dispatch one. if not, check if
1047 * enough slice is left to wait for one
1048 */
1049 if (!RB_EMPTY(&cfqq->sort_list))
1050 goto keep_queue;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001051 else if (cfq_cfqq_class_sync(cfqq) &&
Jens Axboe22e2c502005-06-27 10:55:12 +02001052 time_before(now, cfqq->slice_end)) {
1053 if (cfq_arm_slice_timer(cfqd, cfqq))
1054 return NULL;
1055 }
1056
Jens Axboe3b181522005-06-27 10:56:24 +02001057expire:
Jens Axboe22e2c502005-06-27 10:55:12 +02001058 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001059new_queue:
1060 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001061keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001062 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001063}
1064
1065static int
1066__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1067 int max_dispatch)
1068{
1069 int dispatched = 0;
1070
1071 BUG_ON(RB_EMPTY(&cfqq->sort_list));
1072
1073 do {
1074 struct cfq_rq *crq;
1075
1076 /*
1077 * follow expired path, else get first next available
1078 */
1079 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1080 crq = cfqq->next_crq;
1081
1082 /*
1083 * finally, insert request into driver dispatch list
1084 */
Jens Axboeb4878f22005-10-20 16:42:29 +02001085 cfq_dispatch_insert(cfqd->queue, crq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001086
1087 cfqd->dispatch_slice++;
1088 dispatched++;
1089
1090 if (!cfqd->active_cic) {
1091 atomic_inc(&crq->io_context->ioc->refcount);
1092 cfqd->active_cic = crq->io_context;
1093 }
1094
1095 if (RB_EMPTY(&cfqq->sort_list))
1096 break;
1097
1098 } while (dispatched < max_dispatch);
1099
1100 /*
1101 * if slice end isn't set yet, set it. if at least one request was
1102 * sync, use the sync time slice value
1103 */
1104 if (!cfqq->slice_end)
1105 cfq_set_prio_slice(cfqd, cfqq);
1106
1107 /*
1108 * expire an async queue immediately if it has used up its slice. idle
1109 * queue always expire after 1 dispatch round.
1110 */
1111 if ((!cfq_cfqq_sync(cfqq) &&
1112 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1113 cfq_class_idle(cfqq))
1114 cfq_slice_expired(cfqd, 0);
1115
1116 return dispatched;
1117}
1118
1119static int
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001120cfq_forced_dispatch_cfqqs(struct list_head *list)
1121{
1122 int dispatched = 0;
1123 struct cfq_queue *cfqq, *next;
1124 struct cfq_rq *crq;
1125
1126 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1127 while ((crq = cfqq->next_crq)) {
1128 cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1129 dispatched++;
1130 }
1131 BUG_ON(!list_empty(&cfqq->fifo));
1132 }
1133 return dispatched;
1134}
1135
1136static int
1137cfq_forced_dispatch(struct cfq_data *cfqd)
1138{
1139 int i, dispatched = 0;
1140
1141 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1142 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1143
1144 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1145 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1146 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1147
1148 cfq_slice_expired(cfqd, 0);
1149
1150 BUG_ON(cfqd->busy_queues);
1151
1152 return dispatched;
1153}
1154
1155static int
Jens Axboeb4878f22005-10-20 16:42:29 +02001156cfq_dispatch_requests(request_queue_t *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157{
1158 struct cfq_data *cfqd = q->elevator->elevator_data;
1159 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160
Jens Axboe22e2c502005-06-27 10:55:12 +02001161 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 return 0;
1163
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001164 if (unlikely(force))
1165 return cfq_forced_dispatch(cfqd);
1166
1167 cfqq = cfq_select_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001168 if (cfqq) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001169 int max_dispatch;
1170
Jens Axboe3b181522005-06-27 10:56:24 +02001171 cfq_clear_cfqq_must_dispatch(cfqq);
1172 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001173 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001175 max_dispatch = cfqd->cfq_quantum;
1176 if (cfq_class_idle(cfqq))
1177 max_dispatch = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178
Jens Axboe22e2c502005-06-27 10:55:12 +02001179 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 }
1181
Jens Axboe22e2c502005-06-27 10:55:12 +02001182 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183}
1184
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185/*
1186 * task holds one reference to the queue, dropped when task exits. each crq
1187 * in-flight on this queue also holds a reference, dropped when crq is freed.
1188 *
1189 * queue lock must be held here.
1190 */
1191static void cfq_put_queue(struct cfq_queue *cfqq)
1192{
Jens Axboe22e2c502005-06-27 10:55:12 +02001193 struct cfq_data *cfqd = cfqq->cfqd;
1194
1195 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196
1197 if (!atomic_dec_and_test(&cfqq->ref))
1198 return;
1199
1200 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001201 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001202 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001204 if (unlikely(cfqd->active_queue == cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001205 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02001206
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207 /*
1208 * it's on the empty list and still hashed
1209 */
1210 list_del(&cfqq->cfq_list);
1211 hlist_del(&cfqq->cfq_hash);
1212 kmem_cache_free(cfq_pool, cfqq);
1213}
1214
1215static inline struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001216__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1217 const int hashval)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218{
1219 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
Jens Axboe206dc692006-03-28 13:03:44 +02001220 struct hlist_node *entry;
1221 struct cfq_queue *__cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222
Jens Axboe206dc692006-03-28 13:03:44 +02001223 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
Al Virob0a69162006-03-14 15:32:50 -05001224 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225
Jens Axboe206dc692006-03-28 13:03:44 +02001226 if (__cfqq->key == key && (__p == prio || !prio))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 return __cfqq;
1228 }
1229
1230 return NULL;
1231}
1232
1233static struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001234cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235{
Jens Axboe3b181522005-06-27 10:56:24 +02001236 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237}
1238
Jens Axboee2d74ac2006-03-28 08:59:01 +02001239static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240{
Jens Axboe22e2c502005-06-27 10:55:12 +02001241 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001242 struct rb_node *n;
1243 int freed = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001244
Jens Axboee2d74ac2006-03-28 08:59:01 +02001245 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1246 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1247 rb_erase(&__cic->rb_node, &ioc->cic_root);
Jens Axboe22e2c502005-06-27 10:55:12 +02001248 kmem_cache_free(cfq_ioc_pool, __cic);
Al Viro334e94d2006-03-18 15:05:53 -05001249 freed++;
Jens Axboe22e2c502005-06-27 10:55:12 +02001250 }
1251
Al Viro334e94d2006-03-18 15:05:53 -05001252 if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1253 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254}
1255
Al Viroe17a9482006-03-18 13:21:20 -05001256static void cfq_trim(struct io_context *ioc)
1257{
1258 ioc->set_ioprio = NULL;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001259 cfq_free_io_context(ioc);
Al Viroe17a9482006-03-18 13:21:20 -05001260}
1261
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001263 * Called with interrupts disabled
1264 */
1265static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1266{
Al Viro478a82b2006-03-18 13:25:24 -05001267 struct cfq_data *cfqd = cic->key;
Al Virod9ff4182006-03-18 13:51:22 -05001268 request_queue_t *q;
1269
1270 if (!cfqd)
1271 return;
1272
1273 q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001274
1275 WARN_ON(!irqs_disabled());
1276
1277 spin_lock(q->queue_lock);
1278
Al Viro12a05732006-03-18 13:38:01 -05001279 if (cic->cfqq[ASYNC]) {
1280 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1281 __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1282 cfq_put_queue(cic->cfqq[ASYNC]);
1283 cic->cfqq[ASYNC] = NULL;
1284 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001285
Al Viro12a05732006-03-18 13:38:01 -05001286 if (cic->cfqq[SYNC]) {
1287 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1288 __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1289 cfq_put_queue(cic->cfqq[SYNC]);
1290 cic->cfqq[SYNC] = NULL;
1291 }
1292
Al Viro478a82b2006-03-18 13:25:24 -05001293 cic->key = NULL;
Al Virod9ff4182006-03-18 13:51:22 -05001294 list_del_init(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001295 spin_unlock(q->queue_lock);
1296}
1297
Jens Axboee2d74ac2006-03-28 08:59:01 +02001298static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299{
Jens Axboe22e2c502005-06-27 10:55:12 +02001300 struct cfq_io_context *__cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 unsigned long flags;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001302 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +02001303
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 /*
1305 * put the reference this task is holding to the various queues
1306 */
Jens Axboe3793c652006-05-30 21:11:04 +02001307 spin_lock_irqsave(&cfq_exit_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001308
1309 n = rb_first(&ioc->cic_root);
1310 while (n != NULL) {
1311 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1312
Jens Axboe22e2c502005-06-27 10:55:12 +02001313 cfq_exit_single_io_context(__cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001314 n = rb_next(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 }
1316
Jens Axboe3793c652006-05-30 21:11:04 +02001317 spin_unlock_irqrestore(&cfq_exit_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318}
1319
Jens Axboe22e2c502005-06-27 10:55:12 +02001320static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001321cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322{
Jens Axboe22e2c502005-06-27 10:55:12 +02001323 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324
1325 if (cic) {
Jens Axboe553698f2006-06-14 19:11:57 +02001326 memset(cic, 0, sizeof(*cic));
1327 RB_CLEAR_COLOR(&cic->rb_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001328 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001329 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001330 cic->dtor = cfq_free_io_context;
1331 cic->exit = cfq_exit_io_context;
Al Viro334e94d2006-03-18 15:05:53 -05001332 atomic_inc(&ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 }
1334
1335 return cic;
1336}
1337
Jens Axboe22e2c502005-06-27 10:55:12 +02001338static void cfq_init_prio_data(struct cfq_queue *cfqq)
1339{
1340 struct task_struct *tsk = current;
1341 int ioprio_class;
1342
Jens Axboe3b181522005-06-27 10:56:24 +02001343 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001344 return;
1345
1346 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1347 switch (ioprio_class) {
1348 default:
1349 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1350 case IOPRIO_CLASS_NONE:
1351 /*
1352 * no prio set, place us in the middle of the BE classes
1353 */
1354 cfqq->ioprio = task_nice_ioprio(tsk);
1355 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1356 break;
1357 case IOPRIO_CLASS_RT:
1358 cfqq->ioprio = task_ioprio(tsk);
1359 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1360 break;
1361 case IOPRIO_CLASS_BE:
1362 cfqq->ioprio = task_ioprio(tsk);
1363 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1364 break;
1365 case IOPRIO_CLASS_IDLE:
1366 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1367 cfqq->ioprio = 7;
Jens Axboe3b181522005-06-27 10:56:24 +02001368 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001369 break;
1370 }
1371
1372 /*
1373 * keep track of original prio settings in case we have to temporarily
1374 * elevate the priority of this queue
1375 */
1376 cfqq->org_ioprio = cfqq->ioprio;
1377 cfqq->org_ioprio_class = cfqq->ioprio_class;
1378
Jens Axboe3b181522005-06-27 10:56:24 +02001379 if (cfq_cfqq_on_rr(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001380 cfq_resort_rr_list(cfqq, 0);
1381
Jens Axboe3b181522005-06-27 10:56:24 +02001382 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001383}
1384
Al Viro478a82b2006-03-18 13:25:24 -05001385static inline void changed_ioprio(struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001386{
Al Viro478a82b2006-03-18 13:25:24 -05001387 struct cfq_data *cfqd = cic->key;
1388 struct cfq_queue *cfqq;
1389 if (cfqd) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001390 spin_lock(cfqd->queue->queue_lock);
Al Viro12a05732006-03-18 13:38:01 -05001391 cfqq = cic->cfqq[ASYNC];
1392 if (cfqq) {
Al Viro6f325a12006-03-18 14:58:37 -05001393 struct cfq_queue *new_cfqq;
1394 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC,
1395 cic->ioc->task, GFP_ATOMIC);
1396 if (new_cfqq) {
1397 cic->cfqq[ASYNC] = new_cfqq;
1398 cfq_put_queue(cfqq);
1399 }
Al Viro12a05732006-03-18 13:38:01 -05001400 }
1401 cfqq = cic->cfqq[SYNC];
Al Viro478a82b2006-03-18 13:25:24 -05001402 if (cfqq) {
1403 cfq_mark_cfqq_prio_changed(cfqq);
1404 cfq_init_prio_data(cfqq);
1405 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001406 spin_unlock(cfqd->queue->queue_lock);
1407 }
1408}
1409
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001411 * callback from sys_ioprio_set, irqs are disabled
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001413static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414{
Al Viroa6a07632006-03-18 13:26:44 -05001415 struct cfq_io_context *cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001416 struct rb_node *n;
Al Viroa6a07632006-03-18 13:26:44 -05001417
Jens Axboe3793c652006-05-30 21:11:04 +02001418 spin_lock(&cfq_exit_lock);
Al Viroa6a07632006-03-18 13:26:44 -05001419
Jens Axboee2d74ac2006-03-28 08:59:01 +02001420 n = rb_first(&ioc->cic_root);
1421 while (n != NULL) {
1422 cic = rb_entry(n, struct cfq_io_context, rb_node);
Jens Axboe3793c652006-05-30 21:11:04 +02001423
Al Viro478a82b2006-03-18 13:25:24 -05001424 changed_ioprio(cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001425 n = rb_next(n);
1426 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427
Jens Axboe3793c652006-05-30 21:11:04 +02001428 spin_unlock(&cfq_exit_lock);
Al Viroa6a07632006-03-18 13:26:44 -05001429
Jens Axboe22e2c502005-06-27 10:55:12 +02001430 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431}
1432
1433static struct cfq_queue *
Al Viro6f325a12006-03-18 14:58:37 -05001434cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
Al Viro8267e262005-10-21 03:20:53 -04001435 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436{
1437 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1438 struct cfq_queue *cfqq, *new_cfqq = NULL;
Al Viro6f325a12006-03-18 14:58:37 -05001439 unsigned short ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440
1441retry:
Al Viro6f325a12006-03-18 14:58:37 -05001442 ioprio = tsk->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02001443 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444
1445 if (!cfqq) {
1446 if (new_cfqq) {
1447 cfqq = new_cfqq;
1448 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001449 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 spin_unlock_irq(cfqd->queue->queue_lock);
1451 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1452 spin_lock_irq(cfqd->queue->queue_lock);
1453 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001454 } else {
1455 cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1456 if (!cfqq)
1457 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001458 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459
1460 memset(cfqq, 0, sizeof(*cfqq));
1461
1462 INIT_HLIST_NODE(&cfqq->cfq_hash);
1463 INIT_LIST_HEAD(&cfqq->cfq_list);
1464 RB_CLEAR_ROOT(&cfqq->sort_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001465 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466
1467 cfqq->key = key;
1468 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1469 atomic_set(&cfqq->ref, 0);
1470 cfqq->cfqd = cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +02001471 cfqq->service_last = 0;
1472 /*
1473 * set ->slice_left to allow preemption for a new process
1474 */
1475 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
Jens Axboe25776e32006-06-01 10:12:26 +02001476 if (!cfqd->hw_tag)
1477 cfq_mark_cfqq_idle_window(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001478 cfq_mark_cfqq_prio_changed(cfqq);
1479 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 }
1481
1482 if (new_cfqq)
1483 kmem_cache_free(cfq_pool, new_cfqq);
1484
1485 atomic_inc(&cfqq->ref);
1486out:
1487 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1488 return cfqq;
1489}
1490
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001491static void
1492cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1493{
Jens Axboe3793c652006-05-30 21:11:04 +02001494 spin_lock(&cfq_exit_lock);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001495 rb_erase(&cic->rb_node, &ioc->cic_root);
Jens Axboe3793c652006-05-30 21:11:04 +02001496 list_del_init(&cic->queue_list);
1497 spin_unlock(&cfq_exit_lock);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001498 kmem_cache_free(cfq_ioc_pool, cic);
1499 atomic_dec(&ioc_count);
1500}
1501
Jens Axboee2d74ac2006-03-28 08:59:01 +02001502static struct cfq_io_context *
1503cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1504{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001505 struct rb_node *n;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001506 struct cfq_io_context *cic;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001507 void *k, *key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001508
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001509restart:
1510 n = ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001511 while (n) {
1512 cic = rb_entry(n, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001513 /* ->key must be copied to avoid race with cfq_exit_queue() */
1514 k = cic->key;
1515 if (unlikely(!k)) {
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001516 cfq_drop_dead_cic(ioc, cic);
1517 goto restart;
1518 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001519
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001520 if (key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001521 n = n->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001522 else if (key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001523 n = n->rb_right;
1524 else
1525 return cic;
1526 }
1527
1528 return NULL;
1529}
1530
1531static inline void
1532cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1533 struct cfq_io_context *cic)
1534{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001535 struct rb_node **p;
1536 struct rb_node *parent;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001537 struct cfq_io_context *__cic;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001538 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001539
Jens Axboee2d74ac2006-03-28 08:59:01 +02001540 cic->ioc = ioc;
1541 cic->key = cfqd;
1542
1543 ioc->set_ioprio = cfq_ioc_set_ioprio;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001544restart:
1545 parent = NULL;
1546 p = &ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001547 while (*p) {
1548 parent = *p;
1549 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001550 /* ->key must be copied to avoid race with cfq_exit_queue() */
1551 k = __cic->key;
1552 if (unlikely(!k)) {
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001553 cfq_drop_dead_cic(ioc, cic);
1554 goto restart;
1555 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001556
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001557 if (cic->key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001558 p = &(*p)->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001559 else if (cic->key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001560 p = &(*p)->rb_right;
1561 else
1562 BUG();
1563 }
1564
Jens Axboe3793c652006-05-30 21:11:04 +02001565 spin_lock(&cfq_exit_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001566 rb_link_node(&cic->rb_node, parent, p);
1567 rb_insert_color(&cic->rb_node, &ioc->cic_root);
1568 list_add(&cic->queue_list, &cfqd->cic_list);
Jens Axboe3793c652006-05-30 21:11:04 +02001569 spin_unlock(&cfq_exit_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001570}
1571
Jens Axboe22e2c502005-06-27 10:55:12 +02001572/*
1573 * Setup general io context and cfq io context. There can be several cfq
1574 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001575 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001576 */
1577static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001578cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579{
Jens Axboe22e2c502005-06-27 10:55:12 +02001580 struct io_context *ioc = NULL;
1581 struct cfq_io_context *cic;
1582
1583 might_sleep_if(gfp_mask & __GFP_WAIT);
1584
1585 ioc = get_io_context(gfp_mask);
1586 if (!ioc)
1587 return NULL;
1588
Jens Axboee2d74ac2006-03-28 08:59:01 +02001589 cic = cfq_cic_rb_lookup(cfqd, ioc);
1590 if (cic)
1591 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001592
Jens Axboee2d74ac2006-03-28 08:59:01 +02001593 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1594 if (cic == NULL)
1595 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001596
Jens Axboee2d74ac2006-03-28 08:59:01 +02001597 cfq_cic_link(cfqd, ioc, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001598out:
1599 return cic;
1600err:
1601 put_io_context(ioc);
1602 return NULL;
1603}
1604
1605static void
1606cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1607{
1608 unsigned long elapsed, ttime;
1609
1610 /*
1611 * if this context already has stuff queued, thinktime is from
1612 * last queue not last end
1613 */
1614#if 0
1615 if (time_after(cic->last_end_request, cic->last_queue))
1616 elapsed = jiffies - cic->last_end_request;
1617 else
1618 elapsed = jiffies - cic->last_queue;
1619#else
1620 elapsed = jiffies - cic->last_end_request;
1621#endif
1622
1623 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1624
1625 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1626 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1627 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1628}
1629
Jens Axboe206dc692006-03-28 13:03:44 +02001630static void
1631cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1632 struct cfq_rq *crq)
1633{
1634 sector_t sdist;
1635 u64 total;
1636
1637 if (cic->last_request_pos < crq->request->sector)
1638 sdist = crq->request->sector - cic->last_request_pos;
1639 else
1640 sdist = cic->last_request_pos - crq->request->sector;
1641
1642 /*
1643 * Don't allow the seek distance to get too large from the
1644 * odd fragment, pagein, etc
1645 */
1646 if (cic->seek_samples <= 60) /* second&third seek */
1647 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1648 else
1649 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1650
1651 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1652 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1653 total = cic->seek_total + (cic->seek_samples/2);
1654 do_div(total, cic->seek_samples);
1655 cic->seek_mean = (sector_t)total;
1656}
Jens Axboe22e2c502005-06-27 10:55:12 +02001657
1658/*
1659 * Disable idle window if the process thinks too long or seeks so much that
1660 * it doesn't matter
1661 */
1662static void
1663cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1664 struct cfq_io_context *cic)
1665{
Jens Axboe3b181522005-06-27 10:56:24 +02001666 int enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001667
Jens Axboe25776e32006-06-01 10:12:26 +02001668 if (!cic->ioc->task || !cfqd->cfq_slice_idle || cfqd->hw_tag)
Jens Axboe22e2c502005-06-27 10:55:12 +02001669 enable_idle = 0;
1670 else if (sample_valid(cic->ttime_samples)) {
1671 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1672 enable_idle = 0;
1673 else
1674 enable_idle = 1;
1675 }
1676
Jens Axboe3b181522005-06-27 10:56:24 +02001677 if (enable_idle)
1678 cfq_mark_cfqq_idle_window(cfqq);
1679 else
1680 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001681}
1682
1683
1684/*
1685 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1686 * no or if we aren't sure, a 1 will cause a preempt.
1687 */
1688static int
1689cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1690 struct cfq_rq *crq)
1691{
1692 struct cfq_queue *cfqq = cfqd->active_queue;
1693
1694 if (cfq_class_idle(new_cfqq))
1695 return 0;
1696
1697 if (!cfqq)
1698 return 1;
1699
1700 if (cfq_class_idle(cfqq))
1701 return 1;
Jens Axboe3b181522005-06-27 10:56:24 +02001702 if (!cfq_cfqq_wait_request(new_cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001703 return 0;
1704 /*
1705 * if it doesn't have slice left, forget it
1706 */
1707 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1708 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02001709 if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001710 return 1;
1711
1712 return 0;
1713}
1714
1715/*
1716 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1717 * let it have half of its nominal slice.
1718 */
1719static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1720{
1721 struct cfq_queue *__cfqq, *next;
1722
1723 list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1724 cfq_resort_rr_list(__cfqq, 1);
1725
1726 if (!cfqq->slice_left)
1727 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1728
1729 cfqq->slice_end = cfqq->slice_left + jiffies;
Jens Axboe3b181522005-06-27 10:56:24 +02001730 __cfq_slice_expired(cfqd, cfqq, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001731 __cfq_set_active_queue(cfqd, cfqq);
1732}
1733
1734/*
1735 * should really be a ll_rw_blk.c helper
1736 */
1737static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1738{
1739 request_queue_t *q = cfqd->queue;
1740
1741 if (!blk_queue_plugged(q))
1742 q->request_fn(q);
1743 else
1744 __generic_unplug_device(q);
1745}
1746
1747/*
1748 * Called when a new fs request (crq) is added (to cfqq). Check if there's
1749 * something we should do about it
1750 */
1751static void
1752cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1753 struct cfq_rq *crq)
1754{
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001755 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02001756
1757 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1758
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001759 cic = crq->io_context;
1760
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001761 /*
1762 * we never wait for an async request and we don't allow preemption
1763 * of an async request. so just return early
1764 */
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001765 if (!cfq_crq_is_sync(crq)) {
1766 /*
1767 * sync process issued an async request, if it's waiting
1768 * then expire it and kick rq handling.
1769 */
1770 if (cic == cfqd->active_cic &&
1771 del_timer(&cfqd->idle_slice_timer)) {
1772 cfq_slice_expired(cfqd, 0);
1773 cfq_start_queueing(cfqd, cfqq);
1774 }
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001775 return;
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001776 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001777
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001778 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe206dc692006-03-28 13:03:44 +02001779 cfq_update_io_seektime(cfqd, cic, crq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001780 cfq_update_idle_window(cfqd, cfqq, cic);
1781
1782 cic->last_queue = jiffies;
Jens Axboe206dc692006-03-28 13:03:44 +02001783 cic->last_request_pos = crq->request->sector + crq->request->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001784
1785 if (cfqq == cfqd->active_queue) {
1786 /*
1787 * if we are waiting for a request for this queue, let it rip
1788 * immediately and flag that we must not expire this queue
1789 * just now
1790 */
Jens Axboe3b181522005-06-27 10:56:24 +02001791 if (cfq_cfqq_wait_request(cfqq)) {
1792 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001793 del_timer(&cfqd->idle_slice_timer);
1794 cfq_start_queueing(cfqd, cfqq);
1795 }
1796 } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1797 /*
1798 * not the active queue - expire current slice if it is
1799 * idle and has expired it's mean thinktime or this new queue
1800 * has some old slice time left and is of higher priority
1801 */
1802 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001803 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001804 cfq_start_queueing(cfqd, cfqq);
1805 }
1806}
1807
Jens Axboeb4878f22005-10-20 16:42:29 +02001808static void cfq_insert_request(request_queue_t *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001809{
Jens Axboeb4878f22005-10-20 16:42:29 +02001810 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe22e2c502005-06-27 10:55:12 +02001811 struct cfq_rq *crq = RQ_DATA(rq);
1812 struct cfq_queue *cfqq = crq->cfq_queue;
1813
1814 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815
1816 cfq_add_crq_rb(crq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817
Jens Axboe22e2c502005-06-27 10:55:12 +02001818 list_add_tail(&rq->queuelist, &cfqq->fifo);
1819
Tejun Heo98b11472005-10-20 16:46:54 +02001820 if (rq_mergeable(rq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001821 cfq_add_crq_hash(cfqd, crq);
1822
Jens Axboe22e2c502005-06-27 10:55:12 +02001823 cfq_crq_enqueued(cfqd, cfqq, crq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824}
1825
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826static void cfq_completed_request(request_queue_t *q, struct request *rq)
1827{
1828 struct cfq_rq *crq = RQ_DATA(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001829 struct cfq_queue *cfqq = crq->cfq_queue;
1830 struct cfq_data *cfqd = cfqq->cfqd;
1831 const int sync = cfq_crq_is_sync(crq);
1832 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833
Jens Axboeb4878f22005-10-20 16:42:29 +02001834 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835
Jens Axboeb4878f22005-10-20 16:42:29 +02001836 WARN_ON(!cfqd->rq_in_driver);
1837 WARN_ON(!cfqq->on_dispatch[sync]);
1838 cfqd->rq_in_driver--;
1839 cfqq->on_dispatch[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840
Jens Axboeb4878f22005-10-20 16:42:29 +02001841 if (!cfq_class_idle(cfqq))
1842 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001843
Jens Axboeb4878f22005-10-20 16:42:29 +02001844 if (!cfq_cfqq_dispatched(cfqq)) {
1845 if (cfq_cfqq_on_rr(cfqq)) {
1846 cfqq->service_last = now;
1847 cfq_resort_rr_list(cfqq, 0);
1848 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001849 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 }
1851
Jens Axboeb4878f22005-10-20 16:42:29 +02001852 if (cfq_crq_is_sync(crq))
1853 crq->io_context->last_end_request = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854}
1855
1856static struct request *
1857cfq_former_request(request_queue_t *q, struct request *rq)
1858{
1859 struct cfq_rq *crq = RQ_DATA(rq);
1860 struct rb_node *rbprev = rb_prev(&crq->rb_node);
1861
1862 if (rbprev)
1863 return rb_entry_crq(rbprev)->request;
1864
1865 return NULL;
1866}
1867
1868static struct request *
1869cfq_latter_request(request_queue_t *q, struct request *rq)
1870{
1871 struct cfq_rq *crq = RQ_DATA(rq);
1872 struct rb_node *rbnext = rb_next(&crq->rb_node);
1873
1874 if (rbnext)
1875 return rb_entry_crq(rbnext)->request;
1876
1877 return NULL;
1878}
1879
Jens Axboe22e2c502005-06-27 10:55:12 +02001880/*
1881 * we temporarily boost lower priority queues if they are holding fs exclusive
1882 * resources. they are boosted to normal prio (CLASS_BE/4)
1883 */
1884static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885{
Jens Axboe22e2c502005-06-27 10:55:12 +02001886 const int ioprio_class = cfqq->ioprio_class;
1887 const int ioprio = cfqq->ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888
Jens Axboe22e2c502005-06-27 10:55:12 +02001889 if (has_fs_excl()) {
1890 /*
1891 * boost idle prio on transactions that would lock out other
1892 * users of the filesystem
1893 */
1894 if (cfq_class_idle(cfqq))
1895 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1896 if (cfqq->ioprio > IOPRIO_NORM)
1897 cfqq->ioprio = IOPRIO_NORM;
1898 } else {
1899 /*
1900 * check if we need to unboost the queue
1901 */
1902 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1903 cfqq->ioprio_class = cfqq->org_ioprio_class;
1904 if (cfqq->ioprio != cfqq->org_ioprio)
1905 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906 }
1907
Jens Axboe22e2c502005-06-27 10:55:12 +02001908 /*
1909 * refile between round-robin lists if we moved the priority class
1910 */
1911 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
Jens Axboe3b181522005-06-27 10:56:24 +02001912 cfq_cfqq_on_rr(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001913 cfq_resort_rr_list(cfqq, 0);
1914}
1915
Jens Axboe22e2c502005-06-27 10:55:12 +02001916static inline int
1917__cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1918 struct task_struct *task, int rw)
1919{
Jens Axboe3b181522005-06-27 10:56:24 +02001920#if 1
1921 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001922 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001923 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001924 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001925 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001926
1927 return ELV_MQUEUE_MAY;
Jens Axboe3b181522005-06-27 10:56:24 +02001928#else
Jens Axboe22e2c502005-06-27 10:55:12 +02001929 if (!cfqq || task->flags & PF_MEMALLOC)
1930 return ELV_MQUEUE_MAY;
Jens Axboe3b181522005-06-27 10:56:24 +02001931 if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1932 if (cfq_cfqq_wait_request(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001933 return ELV_MQUEUE_MUST;
1934
1935 /*
1936 * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1937 * can quickly flood the queue with writes from a single task
1938 */
Andrew Morton99f95e52005-06-27 20:14:05 -07001939 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001940 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001941 return ELV_MQUEUE_MUST;
1942 }
1943
1944 return ELV_MQUEUE_MAY;
1945 }
1946 if (cfq_class_idle(cfqq))
1947 return ELV_MQUEUE_NO;
1948 if (cfqq->allocated[rw] >= cfqd->max_queued) {
1949 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1950 int ret = ELV_MQUEUE_NO;
1951
1952 if (ioc && ioc->nr_batch_requests)
1953 ret = ELV_MQUEUE_MAY;
1954
1955 put_io_context(ioc);
1956 return ret;
1957 }
1958
1959 return ELV_MQUEUE_MAY;
1960#endif
1961}
1962
1963static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1964{
1965 struct cfq_data *cfqd = q->elevator->elevator_data;
1966 struct task_struct *tsk = current;
1967 struct cfq_queue *cfqq;
1968
1969 /*
1970 * don't force setup of a queue from here, as a call to may_queue
1971 * does not necessarily imply that a request actually will be queued.
1972 * so just lookup a possibly existing queue, or return 'may queue'
1973 * if that fails
1974 */
Jens Axboe3b181522005-06-27 10:56:24 +02001975 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001976 if (cfqq) {
1977 cfq_init_prio_data(cfqq);
1978 cfq_prio_boost(cfqq);
1979
1980 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1981 }
1982
1983 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984}
1985
1986static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1987{
Jens Axboe22e2c502005-06-27 10:55:12 +02001988 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 struct request_list *rl = &q->rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990
Jens Axboe22e2c502005-06-27 10:55:12 +02001991 if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1992 smp_mb();
1993 if (waitqueue_active(&rl->wait[READ]))
1994 wake_up(&rl->wait[READ]);
1995 }
1996
1997 if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1998 smp_mb();
1999 if (waitqueue_active(&rl->wait[WRITE]))
2000 wake_up(&rl->wait[WRITE]);
2001 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002}
2003
2004/*
2005 * queue lock held here
2006 */
2007static void cfq_put_request(request_queue_t *q, struct request *rq)
2008{
2009 struct cfq_data *cfqd = q->elevator->elevator_data;
2010 struct cfq_rq *crq = RQ_DATA(rq);
2011
2012 if (crq) {
2013 struct cfq_queue *cfqq = crq->cfq_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002014 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015
Jens Axboe22e2c502005-06-27 10:55:12 +02002016 BUG_ON(!cfqq->allocated[rw]);
2017 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018
Jens Axboe22e2c502005-06-27 10:55:12 +02002019 put_io_context(crq->io_context->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020
2021 mempool_free(crq, cfqd->crq_pool);
2022 rq->elevator_private = NULL;
2023
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024 cfq_check_waiters(q, cfqq);
2025 cfq_put_queue(cfqq);
2026 }
2027}
2028
2029/*
Jens Axboe22e2c502005-06-27 10:55:12 +02002030 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031 */
Jens Axboe22e2c502005-06-27 10:55:12 +02002032static int
2033cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
Al Viro8267e262005-10-21 03:20:53 -04002034 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035{
2036 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe3b181522005-06-27 10:56:24 +02002037 struct task_struct *tsk = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038 struct cfq_io_context *cic;
2039 const int rw = rq_data_dir(rq);
Jens Axboe3b181522005-06-27 10:56:24 +02002040 pid_t key = cfq_queue_pid(tsk, rw);
Jens Axboe22e2c502005-06-27 10:55:12 +02002041 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042 struct cfq_rq *crq;
2043 unsigned long flags;
Al Viro12a05732006-03-18 13:38:01 -05002044 int is_sync = key != CFQ_KEY_ASYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045
2046 might_sleep_if(gfp_mask & __GFP_WAIT);
2047
Jens Axboee2d74ac2006-03-28 08:59:01 +02002048 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02002049
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050 spin_lock_irqsave(q->queue_lock, flags);
2051
Jens Axboe22e2c502005-06-27 10:55:12 +02002052 if (!cic)
2053 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054
Al Viro12a05732006-03-18 13:38:01 -05002055 if (!cic->cfqq[is_sync]) {
Al Viro6f325a12006-03-18 14:58:37 -05002056 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02002057 if (!cfqq)
2058 goto queue_fail;
2059
Al Viro12a05732006-03-18 13:38:01 -05002060 cic->cfqq[is_sync] = cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002061 } else
Al Viro12a05732006-03-18 13:38:01 -05002062 cfqq = cic->cfqq[is_sync];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063
2064 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02002065 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002066 cfqd->rq_starved = 0;
2067 atomic_inc(&cfqq->ref);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068 spin_unlock_irqrestore(q->queue_lock, flags);
2069
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070 crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
2071 if (crq) {
2072 RB_CLEAR(&crq->rb_node);
2073 crq->rb_key = 0;
2074 crq->request = rq;
2075 INIT_HLIST_NODE(&crq->hash);
2076 crq->cfq_queue = cfqq;
2077 crq->io_context = cic;
Jens Axboe3b181522005-06-27 10:56:24 +02002078
Al Viro12a05732006-03-18 13:38:01 -05002079 if (is_sync)
Jens Axboe3b181522005-06-27 10:56:24 +02002080 cfq_mark_crq_is_sync(crq);
2081 else
2082 cfq_clear_crq_is_sync(crq);
2083
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084 rq->elevator_private = crq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 return 0;
2086 }
2087
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088 spin_lock_irqsave(q->queue_lock, flags);
2089 cfqq->allocated[rw]--;
Jens Axboe22e2c502005-06-27 10:55:12 +02002090 if (!(cfqq->allocated[0] + cfqq->allocated[1]))
Jens Axboe3b181522005-06-27 10:56:24 +02002091 cfq_mark_cfqq_must_alloc(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 cfq_put_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002093queue_fail:
2094 if (cic)
2095 put_io_context(cic->ioc);
2096 /*
2097 * mark us rq allocation starved. we need to kickstart the process
2098 * ourselves if there are no pending requests that can do it for us.
2099 * that would be an extremely rare OOM situation
2100 */
2101 cfqd->rq_starved = 1;
Jens Axboe3b181522005-06-27 10:56:24 +02002102 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 spin_unlock_irqrestore(q->queue_lock, flags);
2104 return 1;
2105}
2106
Jens Axboe22e2c502005-06-27 10:55:12 +02002107static void cfq_kick_queue(void *data)
2108{
2109 request_queue_t *q = data;
2110 struct cfq_data *cfqd = q->elevator->elevator_data;
2111 unsigned long flags;
2112
2113 spin_lock_irqsave(q->queue_lock, flags);
2114
2115 if (cfqd->rq_starved) {
2116 struct request_list *rl = &q->rq;
2117
2118 /*
2119 * we aren't guaranteed to get a request after this, but we
2120 * have to be opportunistic
2121 */
2122 smp_mb();
2123 if (waitqueue_active(&rl->wait[READ]))
2124 wake_up(&rl->wait[READ]);
2125 if (waitqueue_active(&rl->wait[WRITE]))
2126 wake_up(&rl->wait[WRITE]);
2127 }
2128
2129 blk_remove_plug(q);
2130 q->request_fn(q);
2131 spin_unlock_irqrestore(q->queue_lock, flags);
2132}
2133
2134/*
2135 * Timer running if the active_queue is currently idling inside its time slice
2136 */
2137static void cfq_idle_slice_timer(unsigned long data)
2138{
2139 struct cfq_data *cfqd = (struct cfq_data *) data;
2140 struct cfq_queue *cfqq;
2141 unsigned long flags;
2142
2143 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2144
2145 if ((cfqq = cfqd->active_queue) != NULL) {
2146 unsigned long now = jiffies;
2147
2148 /*
2149 * expired
2150 */
2151 if (time_after(now, cfqq->slice_end))
2152 goto expire;
2153
2154 /*
2155 * only expire and reinvoke request handler, if there are
2156 * other queues with pending requests
2157 */
Jens Axboeb4878f22005-10-20 16:42:29 +02002158 if (!cfqd->busy_queues) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002159 cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2160 add_timer(&cfqd->idle_slice_timer);
2161 goto out_cont;
2162 }
2163
2164 /*
2165 * not expired and it has a request pending, let it dispatch
2166 */
2167 if (!RB_EMPTY(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002168 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002169 goto out_kick;
2170 }
2171 }
2172expire:
2173 cfq_slice_expired(cfqd, 0);
2174out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002175 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002176out_cont:
2177 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2178}
2179
2180/*
2181 * Timer running if an idle class queue is waiting for service
2182 */
2183static void cfq_idle_class_timer(unsigned long data)
2184{
2185 struct cfq_data *cfqd = (struct cfq_data *) data;
2186 unsigned long flags, end;
2187
2188 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2189
2190 /*
2191 * race with a non-idle queue, reset timer
2192 */
2193 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
Jens Axboeae818a32006-06-01 10:13:43 +02002194 if (!time_after_eq(jiffies, end))
2195 mod_timer(&cfqd->idle_class_timer, end);
2196 else
Jens Axboe3b181522005-06-27 10:56:24 +02002197 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002198
2199 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2200}
2201
Jens Axboe3b181522005-06-27 10:56:24 +02002202static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2203{
2204 del_timer_sync(&cfqd->idle_slice_timer);
2205 del_timer_sync(&cfqd->idle_class_timer);
2206 blk_sync_queue(cfqd->queue);
2207}
Jens Axboe22e2c502005-06-27 10:55:12 +02002208
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209static void cfq_exit_queue(elevator_t *e)
2210{
Jens Axboe22e2c502005-06-27 10:55:12 +02002211 struct cfq_data *cfqd = e->elevator_data;
Al Virod9ff4182006-03-18 13:51:22 -05002212 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002213
Jens Axboe3b181522005-06-27 10:56:24 +02002214 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002215
Jens Axboe3793c652006-05-30 21:11:04 +02002216 spin_lock(&cfq_exit_lock);
Al Virod9ff4182006-03-18 13:51:22 -05002217 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002218
Al Virod9ff4182006-03-18 13:51:22 -05002219 if (cfqd->active_queue)
2220 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002221
2222 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002223 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2224 struct cfq_io_context,
2225 queue_list);
2226 if (cic->cfqq[ASYNC]) {
2227 cfq_put_queue(cic->cfqq[ASYNC]);
2228 cic->cfqq[ASYNC] = NULL;
2229 }
2230 if (cic->cfqq[SYNC]) {
2231 cfq_put_queue(cic->cfqq[SYNC]);
2232 cic->cfqq[SYNC] = NULL;
2233 }
2234 cic->key = NULL;
2235 list_del_init(&cic->queue_list);
2236 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002237
Al Virod9ff4182006-03-18 13:51:22 -05002238 spin_unlock_irq(q->queue_lock);
Jens Axboe3793c652006-05-30 21:11:04 +02002239 spin_unlock(&cfq_exit_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002240
2241 cfq_shutdown_timer_wq(cfqd);
2242
2243 mempool_destroy(cfqd->crq_pool);
2244 kfree(cfqd->crq_hash);
2245 kfree(cfqd->cfq_hash);
2246 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247}
2248
Jens Axboebc1c1162006-06-08 08:49:06 +02002249static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250{
2251 struct cfq_data *cfqd;
2252 int i;
2253
2254 cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2255 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002256 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257
2258 memset(cfqd, 0, sizeof(*cfqd));
Jens Axboe22e2c502005-06-27 10:55:12 +02002259
2260 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2261 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2262
2263 INIT_LIST_HEAD(&cfqd->busy_rr);
2264 INIT_LIST_HEAD(&cfqd->cur_rr);
2265 INIT_LIST_HEAD(&cfqd->idle_rr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 INIT_LIST_HEAD(&cfqd->empty_list);
Al Virod9ff4182006-03-18 13:51:22 -05002267 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268
2269 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2270 if (!cfqd->crq_hash)
2271 goto out_crqhash;
2272
2273 cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2274 if (!cfqd->cfq_hash)
2275 goto out_cfqhash;
2276
Matthew Dobson93d23412006-03-26 01:37:50 -08002277 cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278 if (!cfqd->crq_pool)
2279 goto out_crqpool;
2280
2281 for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2282 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2283 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2284 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2285
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287
Jens Axboe22e2c502005-06-27 10:55:12 +02002288 cfqd->max_queued = q->nr_requests / 4;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289 q->nr_batching = cfq_queued;
Jens Axboe22e2c502005-06-27 10:55:12 +02002290
2291 init_timer(&cfqd->idle_slice_timer);
2292 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2293 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2294
2295 init_timer(&cfqd->idle_class_timer);
2296 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2297 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2298
2299 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2300
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301 cfqd->cfq_queued = cfq_queued;
2302 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002303 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2304 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 cfqd->cfq_back_max = cfq_back_max;
2306 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002307 cfqd->cfq_slice[0] = cfq_slice_async;
2308 cfqd->cfq_slice[1] = cfq_slice_sync;
2309 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2310 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02002311
Jens Axboebc1c1162006-06-08 08:49:06 +02002312 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313out_crqpool:
2314 kfree(cfqd->cfq_hash);
2315out_cfqhash:
2316 kfree(cfqd->crq_hash);
2317out_crqhash:
2318 kfree(cfqd);
Jens Axboebc1c1162006-06-08 08:49:06 +02002319 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320}
2321
2322static void cfq_slab_kill(void)
2323{
2324 if (crq_pool)
2325 kmem_cache_destroy(crq_pool);
2326 if (cfq_pool)
2327 kmem_cache_destroy(cfq_pool);
2328 if (cfq_ioc_pool)
2329 kmem_cache_destroy(cfq_ioc_pool);
2330}
2331
2332static int __init cfq_slab_setup(void)
2333{
2334 crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2335 NULL, NULL);
2336 if (!crq_pool)
2337 goto fail;
2338
2339 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2340 NULL, NULL);
2341 if (!cfq_pool)
2342 goto fail;
2343
2344 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2345 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2346 if (!cfq_ioc_pool)
2347 goto fail;
2348
2349 return 0;
2350fail:
2351 cfq_slab_kill();
2352 return -ENOMEM;
2353}
2354
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355/*
2356 * sysfs parts below -->
2357 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358
2359static ssize_t
2360cfq_var_show(unsigned int var, char *page)
2361{
2362 return sprintf(page, "%d\n", var);
2363}
2364
2365static ssize_t
2366cfq_var_store(unsigned int *var, const char *page, size_t count)
2367{
2368 char *p = (char *) page;
2369
2370 *var = simple_strtoul(p, &p, 10);
2371 return count;
2372}
2373
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002375static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002377 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378 unsigned int __data = __VAR; \
2379 if (__CONV) \
2380 __data = jiffies_to_msecs(__data); \
2381 return cfq_var_show(__data, (page)); \
2382}
2383SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2384SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002385SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2386SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002387SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2388SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002389SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2390SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2391SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2392SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393#undef SHOW_FUNCTION
2394
2395#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002396static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002398 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 unsigned int __data; \
2400 int ret = cfq_var_store(&__data, (page), count); \
2401 if (__data < (MIN)) \
2402 __data = (MIN); \
2403 else if (__data > (MAX)) \
2404 __data = (MAX); \
2405 if (__CONV) \
2406 *(__PTR) = msecs_to_jiffies(__data); \
2407 else \
2408 *(__PTR) = __data; \
2409 return ret; \
2410}
2411STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2412STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002413STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2414STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002415STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2416STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002417STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2418STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2419STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2420STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421#undef STORE_FUNCTION
2422
Al Viroe572ec72006-03-18 22:27:18 -05002423#define CFQ_ATTR(name) \
2424 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002425
Al Viroe572ec72006-03-18 22:27:18 -05002426static struct elv_fs_entry cfq_attrs[] = {
2427 CFQ_ATTR(quantum),
2428 CFQ_ATTR(queued),
2429 CFQ_ATTR(fifo_expire_sync),
2430 CFQ_ATTR(fifo_expire_async),
2431 CFQ_ATTR(back_seek_max),
2432 CFQ_ATTR(back_seek_penalty),
2433 CFQ_ATTR(slice_sync),
2434 CFQ_ATTR(slice_async),
2435 CFQ_ATTR(slice_async_rq),
2436 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002437 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438};
2439
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440static struct elevator_type iosched_cfq = {
2441 .ops = {
2442 .elevator_merge_fn = cfq_merge,
2443 .elevator_merged_fn = cfq_merged_request,
2444 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeb4878f22005-10-20 16:42:29 +02002445 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002447 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 .elevator_deactivate_req_fn = cfq_deactivate_request,
2449 .elevator_queue_empty_fn = cfq_queue_empty,
2450 .elevator_completed_req_fn = cfq_completed_request,
2451 .elevator_former_req_fn = cfq_former_request,
2452 .elevator_latter_req_fn = cfq_latter_request,
2453 .elevator_set_req_fn = cfq_set_request,
2454 .elevator_put_req_fn = cfq_put_request,
2455 .elevator_may_queue_fn = cfq_may_queue,
2456 .elevator_init_fn = cfq_init_queue,
2457 .elevator_exit_fn = cfq_exit_queue,
Al Viroe17a9482006-03-18 13:21:20 -05002458 .trim = cfq_trim,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 },
Al Viro3d1ab402006-03-18 18:35:43 -05002460 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461 .elevator_name = "cfq",
2462 .elevator_owner = THIS_MODULE,
2463};
2464
2465static int __init cfq_init(void)
2466{
2467 int ret;
2468
Jens Axboe22e2c502005-06-27 10:55:12 +02002469 /*
2470 * could be 0 on HZ < 1000 setups
2471 */
2472 if (!cfq_slice_async)
2473 cfq_slice_async = 1;
2474 if (!cfq_slice_idle)
2475 cfq_slice_idle = 1;
2476
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477 if (cfq_slab_setup())
2478 return -ENOMEM;
2479
2480 ret = elv_register(&iosched_cfq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002481 if (ret)
2482 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 return ret;
2485}
2486
2487static void __exit cfq_exit(void)
2488{
Al Viro334e94d2006-03-18 15:05:53 -05002489 DECLARE_COMPLETION(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002491 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002492 /* ioc_gone's update must be visible before reading ioc_count */
2493 smp_wmb();
Al Viro334e94d2006-03-18 15:05:53 -05002494 if (atomic_read(&ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002495 wait_for_completion(ioc_gone);
Al Viro334e94d2006-03-18 15:05:53 -05002496 synchronize_rcu();
Christoph Hellwig83521d32005-10-30 15:01:39 -08002497 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498}
2499
2500module_init(cfq_init);
2501module_exit(cfq_exit);
2502
2503MODULE_AUTHOR("Jens Axboe");
2504MODULE_LICENSE("GPL");
2505MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");