blob: b6491c020f26ec53dedcbed84e20aea0a7aea299 [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 *
Jens Axboe0fe23472006-09-04 15:41:16 +02007 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/module.h>
Al Viro1cc9be62006-03-18 12:29:52 -050010#include <linux/blkdev.h>
11#include <linux/elevator.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070012#include <linux/hash.h>
13#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020014#include <linux/ioprio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070015
16/*
17 * tunables
18 */
Arjan van de Ven64100092006-01-06 09:46:02 +010019static const int cfq_quantum = 4; /* max queue in one round of service */
Arjan van de Ven64100092006-01-06 09:46:02 +010020static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
21static const int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */
22static const int cfq_back_penalty = 2; /* penalty of a backwards seek */
Linus Torvalds1da177e2005-04-16 15:20:36 -070023
Arjan van de Ven64100092006-01-06 09:46:02 +010024static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020025static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010026static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020027static int cfq_slice_idle = HZ / 125;
Jens Axboe22e2c502005-06-27 10:55:12 +020028
29#define CFQ_IDLE_GRACE (HZ / 10)
30#define CFQ_SLICE_SCALE (5)
31
32#define CFQ_KEY_ASYNC (0)
Jens Axboe22e2c502005-06-27 10:55:12 +020033
Linus Torvalds1da177e2005-04-16 15:20:36 -070034/*
35 * for the hash of cfqq inside the cfqd
36 */
37#define CFQ_QHASH_SHIFT 6
38#define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
39#define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
40
Linus Torvalds1da177e2005-04-16 15:20:36 -070041#define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
42
Jens Axboe5e705372006-07-13 12:39:25 +020043#define RQ_CIC(rq) ((struct cfq_io_context*)(rq)->elevator_private)
44#define RQ_CFQQ(rq) ((rq)->elevator_private2)
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
Christoph Lametere18b8902006-12-06 20:33:20 -080046static struct kmem_cache *cfq_pool;
47static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070048
Jens Axboe4050cf12006-07-19 05:07:12 +020049static DEFINE_PER_CPU(unsigned long, ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050050static struct completion *ioc_gone;
51
Jens Axboe22e2c502005-06-27 10:55:12 +020052#define CFQ_PRIO_LISTS IOPRIO_BE_NR
53#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020054#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
55
Jens Axboe3b181522005-06-27 10:56:24 +020056#define ASYNC (0)
57#define SYNC (1)
58
59#define cfq_cfqq_dispatched(cfqq) \
60 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
61
62#define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
63
64#define cfq_cfqq_sync(cfqq) \
65 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
Jens Axboe22e2c502005-06-27 10:55:12 +020066
Jens Axboe206dc692006-03-28 13:03:44 +020067#define sample_valid(samples) ((samples) > 80)
68
Jens Axboe22e2c502005-06-27 10:55:12 +020069/*
70 * Per block device queue structure
71 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070072struct cfq_data {
Jens Axboe22e2c502005-06-27 10:55:12 +020073 request_queue_t *queue;
74
75 /*
76 * rr list of queues with requests and the count of them
77 */
78 struct list_head rr_list[CFQ_PRIO_LISTS];
79 struct list_head busy_rr;
80 struct list_head cur_rr;
81 struct list_head idle_rr;
82 unsigned int busy_queues;
83
84 /*
Jens Axboe22e2c502005-06-27 10:55:12 +020085 * cfqq lookup hash
86 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 struct hlist_head *cfq_hash;
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
Jens Axboe22e2c502005-06-27 10:55:12 +020089 int rq_in_driver;
Jens Axboe25776e32006-06-01 10:12:26 +020090 int hw_tag;
Jens Axboe22e2c502005-06-27 10:55:12 +020091
92 /*
Jens Axboe22e2c502005-06-27 10:55:12 +020093 * idle window management
94 */
95 struct timer_list idle_slice_timer;
96 struct work_struct unplug_work;
97
98 struct cfq_queue *active_queue;
99 struct cfq_io_context *active_cic;
100 int cur_prio, cur_end_prio;
101 unsigned int dispatch_slice;
102
103 struct timer_list idle_class_timer;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104
105 sector_t last_sector;
Jens Axboe22e2c502005-06-27 10:55:12 +0200106 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108 /*
109 * tunables, see top of file
110 */
111 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200112 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 unsigned int cfq_back_penalty;
114 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200115 unsigned int cfq_slice[2];
116 unsigned int cfq_slice_async_rq;
117 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500118
119 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120};
121
Jens Axboe22e2c502005-06-27 10:55:12 +0200122/*
123 * Per process-grouping structure
124 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125struct cfq_queue {
126 /* reference count */
127 atomic_t ref;
128 /* parent cfq_data */
129 struct cfq_data *cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200130 /* cfqq lookup hash */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 struct hlist_node cfq_hash;
132 /* hash key */
Jens Axboe22e2c502005-06-27 10:55:12 +0200133 unsigned int key;
Jens Axboe981a7972006-07-19 14:56:28 +0200134 /* member of the rr/busy/cur/idle cfqd list */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 struct list_head cfq_list;
136 /* sorted list of pending requests */
137 struct rb_root sort_list;
138 /* if fifo isn't expired, next request to serve */
Jens Axboe5e705372006-07-13 12:39:25 +0200139 struct request *next_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140 /* requests queued in sort_list */
141 int queued[2];
142 /* currently allocated requests */
143 int allocated[2];
Jens Axboe374f84a2006-07-23 01:42:19 +0200144 /* pending metadata requests */
145 int meta_pending;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200147 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148
Jens Axboe22e2c502005-06-27 10:55:12 +0200149 unsigned long slice_end;
Jens Axboe99f96282007-02-05 11:56:25 +0100150 unsigned long service_last;
Jens Axboec5b680f2007-01-19 11:56:49 +1100151 long slice_resid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152
Jens Axboe3b181522005-06-27 10:56:24 +0200153 /* number of requests that are on the dispatch list */
154 int on_dispatch[2];
Jens Axboe22e2c502005-06-27 10:55:12 +0200155
156 /* io prio of this group */
157 unsigned short ioprio, org_ioprio;
158 unsigned short ioprio_class, org_ioprio_class;
159
Jens Axboe3b181522005-06-27 10:56:24 +0200160 /* various state flags, see below */
161 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162};
163
Jens Axboe3b181522005-06-27 10:56:24 +0200164enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100165 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
166 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
167 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
168 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
169 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
170 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
171 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
172 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
173 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
Jens Axboe44f7c162007-01-19 11:51:58 +1100174 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Jens Axboe3b181522005-06-27 10:56:24 +0200175};
176
177#define CFQ_CFQQ_FNS(name) \
178static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
179{ \
180 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
181} \
182static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
183{ \
184 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
185} \
186static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
187{ \
188 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
189}
190
191CFQ_CFQQ_FNS(on_rr);
192CFQ_CFQQ_FNS(wait_request);
193CFQ_CFQQ_FNS(must_alloc);
194CFQ_CFQQ_FNS(must_alloc_slice);
195CFQ_CFQQ_FNS(must_dispatch);
196CFQ_CFQQ_FNS(fifo_expire);
197CFQ_CFQQ_FNS(idle_window);
198CFQ_CFQQ_FNS(prio_changed);
Jens Axboe53b037442006-07-28 09:48:51 +0200199CFQ_CFQQ_FNS(queue_new);
Jens Axboe44f7c162007-01-19 11:51:58 +1100200CFQ_CFQQ_FNS(slice_new);
Jens Axboe3b181522005-06-27 10:56:24 +0200201#undef CFQ_CFQQ_FNS
202
Jens Axboe3b181522005-06-27 10:56:24 +0200203static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
Jens Axboe5e705372006-07-13 12:39:25 +0200204static void cfq_dispatch_insert(request_queue_t *, struct request *);
Al Viro6f325a12006-03-18 14:58:37 -0500205static 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 -0700206
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700208 * scheduler run of queue, if there are requests pending and no one in the
209 * driver that will restart queueing
210 */
211static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
212{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100213 if (cfqd->busy_queues)
Andrew Morton99f95e52005-06-27 20:14:05 -0700214 kblockd_schedule_work(&cfqd->unplug_work);
215}
216
217static int cfq_queue_empty(request_queue_t *q)
218{
219 struct cfq_data *cfqd = q->elevator->elevator_data;
220
Jens Axboeb4878f22005-10-20 16:42:29 +0200221 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700222}
223
Jens Axboe7749a8d2006-12-13 13:02:26 +0100224static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync)
Jens Axboe206dc692006-03-28 13:03:44 +0200225{
Jens Axboe7749a8d2006-12-13 13:02:26 +0100226 /*
227 * Use the per-process queue, for read requests and syncronous writes
228 */
229 if (!(rw & REQ_RW) || is_sync)
Jens Axboe206dc692006-03-28 13:03:44 +0200230 return task->pid;
231
232 return CFQ_KEY_ASYNC;
233}
234
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100236 * Scale schedule slice based on io priority. Use the sync time slice only
237 * if a queue is marked sync and has sync io queued. A sync queue with async
238 * io only, should not get full sync slice length.
239 */
240static inline int
241cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
242{
243 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
244
245 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
246
247 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
248}
249
250static inline void
251cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
252{
253 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
Jens Axboec5b680f2007-01-19 11:56:49 +1100254 cfqq->slice_end += cfqq->slice_resid;
255
256 /*
257 * Don't carry over residual for more than one slice, we only want
258 * to slightly correct the fairness. Carrying over forever would
259 * easily introduce oscillations.
260 */
261 cfqq->slice_resid = 0;
Jens Axboe44f7c162007-01-19 11:51:58 +1100262}
263
264/*
265 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
266 * isn't valid until the first request from the dispatch is activated
267 * and the slice time set.
268 */
269static inline int cfq_slice_used(struct cfq_queue *cfqq)
270{
271 if (cfq_cfqq_slice_new(cfqq))
272 return 0;
273 if (time_before(jiffies, cfqq->slice_end))
274 return 0;
275
276 return 1;
277}
278
279/*
Jens Axboe5e705372006-07-13 12:39:25 +0200280 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200282 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283 */
Jens Axboe5e705372006-07-13 12:39:25 +0200284static struct request *
285cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286{
287 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200289#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
290#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
291 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
Jens Axboe5e705372006-07-13 12:39:25 +0200293 if (rq1 == NULL || rq1 == rq2)
294 return rq2;
295 if (rq2 == NULL)
296 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200297
Jens Axboe5e705372006-07-13 12:39:25 +0200298 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
299 return rq1;
300 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
301 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200302 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
303 return rq1;
304 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
305 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306
Jens Axboe5e705372006-07-13 12:39:25 +0200307 s1 = rq1->sector;
308 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309
310 last = cfqd->last_sector;
311
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312 /*
313 * by definition, 1KiB is 2 sectors
314 */
315 back_max = cfqd->cfq_back_max * 2;
316
317 /*
318 * Strict one way elevator _except_ in the case where we allow
319 * short backward seeks which are biased as twice the cost of a
320 * similar forward seek.
321 */
322 if (s1 >= last)
323 d1 = s1 - last;
324 else if (s1 + back_max >= last)
325 d1 = (last - s1) * cfqd->cfq_back_penalty;
326 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200327 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328
329 if (s2 >= last)
330 d2 = s2 - last;
331 else if (s2 + back_max >= last)
332 d2 = (last - s2) * cfqd->cfq_back_penalty;
333 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200334 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335
336 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337
Andreas Mohre8a99052006-03-28 08:59:49 +0200338 /*
339 * By doing switch() on the bit mask "wrap" we avoid having to
340 * check two variables for all permutations: --> faster!
341 */
342 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200343 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200344 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200345 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200346 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200347 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200348 else {
349 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200350 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200351 else
Jens Axboe5e705372006-07-13 12:39:25 +0200352 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200353 }
354
355 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200356 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200357 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200358 return rq2;
359 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200360 default:
361 /*
362 * Since both rqs are wrapped,
363 * start with the one that's further behind head
364 * (--> only *one* back seek required),
365 * since back seek takes more time than forward.
366 */
367 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200368 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 else
Jens Axboe5e705372006-07-13 12:39:25 +0200370 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371 }
372}
373
374/*
375 * would be nice to take fifo expire time into account as well
376 */
Jens Axboe5e705372006-07-13 12:39:25 +0200377static struct request *
378cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
379 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380{
Jens Axboe21183b02006-07-13 12:33:14 +0200381 struct rb_node *rbnext = rb_next(&last->rb_node);
382 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200383 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384
Jens Axboe21183b02006-07-13 12:33:14 +0200385 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386
387 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200388 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200389
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200391 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200392 else {
393 rbnext = rb_first(&cfqq->sort_list);
394 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200395 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200396 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
Jens Axboe21183b02006-07-13 12:33:14 +0200398 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399}
400
Jens Axboe22e2c502005-06-27 10:55:12 +0200401static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402{
Jens Axboe22e2c502005-06-27 10:55:12 +0200403 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe99f96282007-02-05 11:56:25 +0100404 struct list_head *list, *n;
405 struct cfq_queue *__cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406
Jens Axboe98e41c72007-02-05 11:55:35 +0100407 /*
408 * Resorting requires the cfqq to be on the RR list already.
409 */
410 if (!cfq_cfqq_on_rr(cfqq))
411 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412
413 list_del(&cfqq->cfq_list);
414
Jens Axboe22e2c502005-06-27 10:55:12 +0200415 if (cfq_class_rt(cfqq))
416 list = &cfqd->cur_rr;
417 else if (cfq_class_idle(cfqq))
418 list = &cfqd->idle_rr;
419 else {
420 /*
421 * if cfqq has requests in flight, don't allow it to be
422 * found in cfq_set_active_queue before it has finished them.
423 * this is done to increase fairness between a process that
424 * has lots of io pending vs one that only generates one
425 * sporadically or synchronously
426 */
Jens Axboe3b181522005-06-27 10:56:24 +0200427 if (cfq_cfqq_dispatched(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200428 list = &cfqd->busy_rr;
429 else
430 list = &cfqd->rr_list[cfqq->ioprio];
431 }
432
Jens Axboe53b037442006-07-28 09:48:51 +0200433 if (preempted || cfq_cfqq_queue_new(cfqq)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100434 /*
435 * If this queue was preempted or is new (never been serviced),
436 * let it be added first for fairness but beind other new
437 * queues.
438 */
439 n = list;
Jens Axboe53b037442006-07-28 09:48:51 +0200440 while (n->next != list) {
441 __cfqq = list_entry_cfqq(n->next);
442 if (!cfq_cfqq_queue_new(__cfqq))
443 break;
444
445 n = n->next;
446 }
Jens Axboe99f96282007-02-05 11:56:25 +0100447 list_add_tail(&cfqq->cfq_list, n);
448 } else if (!cfq_cfqq_class_sync(cfqq)) {
449 /*
450 * async queue always goes to the end. this wont be overly
451 * unfair to writes, as the sort of the sync queue wont be
452 * allowed to pass the async queue again.
453 */
454 list_add_tail(&cfqq->cfq_list, list);
455 } else {
456 /*
457 * sort by last service, but don't cross a new or async
458 * queue. we don't cross a new queue because it hasn't been
459 * service before, and we don't cross an async queue because
460 * it gets added to the end on expire.
461 */
462 n = list;
463 while ((n = n->prev) != list) {
464 struct cfq_queue *__cfqq = list_entry_cfqq(n);
Jens Axboe53b037442006-07-28 09:48:51 +0200465
Jens Axboe99f96282007-02-05 11:56:25 +0100466 if (!cfq_cfqq_class_sync(cfqq) || !__cfqq->service_last)
467 break;
468 if (time_before(__cfqq->service_last, cfqq->service_last))
469 break;
470 }
471 list_add(&cfqq->cfq_list, n);
Jens Axboe22e2c502005-06-27 10:55:12 +0200472 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473}
474
475/*
476 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200477 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 */
479static inline void
Jens Axboeb4878f22005-10-20 16:42:29 +0200480cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481{
Jens Axboe3b181522005-06-27 10:56:24 +0200482 BUG_ON(cfq_cfqq_on_rr(cfqq));
483 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 cfqd->busy_queues++;
485
Jens Axboeb4878f22005-10-20 16:42:29 +0200486 cfq_resort_rr_list(cfqq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487}
488
489static inline void
490cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
491{
Jens Axboe3b181522005-06-27 10:56:24 +0200492 BUG_ON(!cfq_cfqq_on_rr(cfqq));
493 cfq_clear_cfqq_on_rr(cfqq);
Jens Axboe981a7972006-07-19 14:56:28 +0200494 list_del_init(&cfqq->cfq_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495
496 BUG_ON(!cfqd->busy_queues);
497 cfqd->busy_queues--;
498}
499
500/*
501 * rb tree support functions
502 */
Jens Axboe5e705372006-07-13 12:39:25 +0200503static inline void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504{
Jens Axboe5e705372006-07-13 12:39:25 +0200505 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200506 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200507 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508
Jens Axboeb4878f22005-10-20 16:42:29 +0200509 BUG_ON(!cfqq->queued[sync]);
510 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511
Jens Axboe5e705372006-07-13 12:39:25 +0200512 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513
Jens Axboedd67d052006-06-21 09:36:18 +0200514 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200515 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516}
517
Jens Axboe5e705372006-07-13 12:39:25 +0200518static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519{
Jens Axboe5e705372006-07-13 12:39:25 +0200520 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe21183b02006-07-13 12:33:14 +0200522 struct request *__alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523
Jens Axboe5380a102006-07-13 12:37:56 +0200524 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525
526 /*
527 * looks a little odd, but the first insert might return an alias.
528 * if that happens, put the alias on the dispatch list
529 */
Jens Axboe21183b02006-07-13 12:33:14 +0200530 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200531 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100532
533 if (!cfq_cfqq_on_rr(cfqq))
534 cfq_add_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535}
536
537static inline void
Jens Axboe5e705372006-07-13 12:39:25 +0200538cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539{
Jens Axboe5380a102006-07-13 12:37:56 +0200540 elv_rb_del(&cfqq->sort_list, rq);
541 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200542 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543}
544
Jens Axboe206dc692006-03-28 13:03:44 +0200545static struct request *
546cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547{
Jens Axboe206dc692006-03-28 13:03:44 +0200548 struct task_struct *tsk = current;
Jens Axboe7749a8d2006-12-13 13:02:26 +0100549 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
Jens Axboe206dc692006-03-28 13:03:44 +0200550 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551
Jens Axboe206dc692006-03-28 13:03:44 +0200552 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
Jens Axboe89850f72006-07-22 16:48:31 +0200553 if (cfqq) {
554 sector_t sector = bio->bi_sector + bio_sectors(bio);
555
Jens Axboe21183b02006-07-13 12:33:14 +0200556 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200557 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 return NULL;
560}
561
Jens Axboeb4878f22005-10-20 16:42:29 +0200562static void cfq_activate_request(request_queue_t *q, struct request *rq)
563{
564 struct cfq_data *cfqd = q->elevator->elevator_data;
565
566 cfqd->rq_in_driver++;
Jens Axboe25776e32006-06-01 10:12:26 +0200567
568 /*
569 * If the depth is larger 1, it really could be queueing. But lets
570 * make the mark a little higher - idling could still be good for
571 * low queueing, and a low queueing number could also just indicate
572 * a SCSI mid layer like behaviour where limit+1 is often seen.
573 */
574 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
575 cfqd->hw_tag = 1;
Jens Axboeb4878f22005-10-20 16:42:29 +0200576}
577
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
579{
Jens Axboe22e2c502005-06-27 10:55:12 +0200580 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581
Jens Axboeb4878f22005-10-20 16:42:29 +0200582 WARN_ON(!cfqd->rq_in_driver);
583 cfqd->rq_in_driver--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584}
585
Jens Axboeb4878f22005-10-20 16:42:29 +0200586static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587{
Jens Axboe5e705372006-07-13 12:39:25 +0200588 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200589
Jens Axboe5e705372006-07-13 12:39:25 +0200590 if (cfqq->next_rq == rq)
591 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592
Jens Axboeb4878f22005-10-20 16:42:29 +0200593 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200594 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200595
596 if (rq_is_meta(rq)) {
597 WARN_ON(!cfqq->meta_pending);
598 cfqq->meta_pending--;
599 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600}
601
602static int
603cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
604{
605 struct cfq_data *cfqd = q->elevator->elevator_data;
606 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607
Jens Axboe206dc692006-03-28 13:03:44 +0200608 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200609 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200610 *req = __rq;
611 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 }
613
614 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615}
616
Jens Axboe21183b02006-07-13 12:33:14 +0200617static void cfq_merged_request(request_queue_t *q, struct request *req,
618 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619{
Jens Axboe21183b02006-07-13 12:33:14 +0200620 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200621 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622
Jens Axboe5e705372006-07-13 12:39:25 +0200623 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625}
626
627static void
628cfq_merged_requests(request_queue_t *q, struct request *rq,
629 struct request *next)
630{
Jens Axboe22e2c502005-06-27 10:55:12 +0200631 /*
632 * reposition in fifo if next is older than rq
633 */
634 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
635 time_before(next->start_time, rq->start_time))
636 list_move(&rq->queuelist, &next->queuelist);
637
Jens Axboeb4878f22005-10-20 16:42:29 +0200638 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200639}
640
Jens Axboeda775262006-12-20 11:04:12 +0100641static int cfq_allow_merge(request_queue_t *q, struct request *rq,
642 struct bio *bio)
643{
644 struct cfq_data *cfqd = q->elevator->elevator_data;
645 const int rw = bio_data_dir(bio);
646 struct cfq_queue *cfqq;
647 pid_t key;
648
649 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100650 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100651 */
Jens Axboeec8acb62007-01-02 18:32:11 +0100652 if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100653 return 0;
654
655 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100656 * Lookup the cfqq that this bio will be queued with. Allow
657 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100658 */
Jens Axboe719d3402006-12-22 09:38:53 +0100659 key = cfq_queue_pid(current, rw, bio_sync(bio));
Jens Axboeda775262006-12-20 11:04:12 +0100660 cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
Jens Axboe719d3402006-12-22 09:38:53 +0100661
662 if (cfqq == RQ_CFQQ(rq))
663 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100664
Jens Axboeec8acb62007-01-02 18:32:11 +0100665 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100666}
667
Jens Axboe22e2c502005-06-27 10:55:12 +0200668static inline void
669__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
670{
671 if (cfqq) {
672 /*
673 * stop potential idle class queues waiting service
674 */
675 del_timer(&cfqd->idle_class_timer);
676
Jens Axboe22e2c502005-06-27 10:55:12 +0200677 cfqq->slice_end = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200678 cfq_clear_cfqq_must_alloc_slice(cfqq);
679 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100680 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200681 }
682
683 cfqd->active_queue = cfqq;
684}
685
686/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100687 * current cfqq expired its slice (or was too idle), select new one
688 */
689static void
690__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100691 int preempted, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100692{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100693 if (cfq_cfqq_wait_request(cfqq))
694 del_timer(&cfqd->idle_slice_timer);
695
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100696 cfq_clear_cfqq_must_dispatch(cfqq);
697 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe53b037442006-07-28 09:48:51 +0200698 cfq_clear_cfqq_queue_new(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100699
700 /*
701 * store what was left of this slice, if the queue idled out
702 * or was preempted
703 */
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100704 if (timed_out && !cfq_cfqq_slice_new(cfqq))
Jens Axboec5b680f2007-01-19 11:56:49 +1100705 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100706
Jens Axboe98e41c72007-02-05 11:55:35 +0100707 cfq_resort_rr_list(cfqq, preempted);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100708
709 if (cfqq == cfqd->active_queue)
710 cfqd->active_queue = NULL;
711
712 if (cfqd->active_cic) {
713 put_io_context(cfqd->active_cic->ioc);
714 cfqd->active_cic = NULL;
715 }
716
717 cfqd->dispatch_slice = 0;
718}
719
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100720static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted,
721 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100722{
723 struct cfq_queue *cfqq = cfqd->active_queue;
724
725 if (cfqq)
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100726 __cfq_slice_expired(cfqd, cfqq, preempted, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100727}
728
729/*
Jens Axboe22e2c502005-06-27 10:55:12 +0200730 * 0
731 * 0,1
732 * 0,1,2
733 * 0,1,2,3
734 * 0,1,2,3,4
735 * 0,1,2,3,4,5
736 * 0,1,2,3,4,5,6
737 * 0,1,2,3,4,5,6,7
738 */
739static int cfq_get_next_prio_level(struct cfq_data *cfqd)
740{
741 int prio, wrap;
742
743 prio = -1;
744 wrap = 0;
745 do {
746 int p;
747
748 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
749 if (!list_empty(&cfqd->rr_list[p])) {
750 prio = p;
751 break;
752 }
753 }
754
755 if (prio != -1)
756 break;
757 cfqd->cur_prio = 0;
758 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
759 cfqd->cur_end_prio = 0;
760 if (wrap)
761 break;
762 wrap = 1;
763 }
764 } while (1);
765
766 if (unlikely(prio == -1))
767 return -1;
768
769 BUG_ON(prio >= CFQ_PRIO_LISTS);
770
771 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
772
773 cfqd->cur_prio = prio + 1;
774 if (cfqd->cur_prio > cfqd->cur_end_prio) {
775 cfqd->cur_end_prio = cfqd->cur_prio;
776 cfqd->cur_prio = 0;
777 }
778 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
779 cfqd->cur_prio = 0;
780 cfqd->cur_end_prio = 0;
781 }
782
783 return prio;
784}
785
Jens Axboe3b181522005-06-27 10:56:24 +0200786static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200787{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100788 struct cfq_queue *cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +0200789
Jens Axboe89850f72006-07-22 16:48:31 +0200790 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
791 /*
792 * if current list is non-empty, grab first entry. if it is
793 * empty, get next prio level and grab first entry then if any
794 * are spliced
795 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200796 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
Jens Axboe89850f72006-07-22 16:48:31 +0200797 } else if (!list_empty(&cfqd->busy_rr)) {
798 /*
799 * If no new queues are available, check if the busy list has
800 * some before falling back to idle io.
801 */
Jens Axboee0de0202006-06-01 10:07:26 +0200802 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
Jens Axboe89850f72006-07-22 16:48:31 +0200803 } else if (!list_empty(&cfqd->idle_rr)) {
804 /*
805 * if we have idle queues and no rt or be queues had pending
806 * requests, either allow immediate service if the grace period
807 * has passed or arm the idle grace timer
808 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200809 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
810
811 if (time_after_eq(jiffies, end))
812 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
813 else
814 mod_timer(&cfqd->idle_class_timer, end);
815 }
816
817 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200818 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200819}
820
Jens Axboecaaa5f92006-06-16 11:23:00 +0200821#define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
822
Jens Axboe17926692007-01-19 11:59:30 +1100823static int cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200824{
Jens Axboe17926692007-01-19 11:59:30 +1100825 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +0200826 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100827 unsigned long sl;
828
Jens Axboedd67d052006-06-21 09:36:18 +0200829 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +0200830
831 /*
832 * idle is disabled, either manually or by past process history
833 */
834 if (!cfqd->cfq_slice_idle)
835 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200836 if (!cfq_cfqq_idle_window(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200837 return 0;
838 /*
839 * task has exited, don't wait
840 */
Jens Axboe206dc692006-03-28 13:03:44 +0200841 cic = cfqd->active_cic;
842 if (!cic || !cic->ioc->task)
Jens Axboe22e2c502005-06-27 10:55:12 +0200843 return 0;
844
Jens Axboe3b181522005-06-27 10:56:24 +0200845 cfq_mark_cfqq_must_dispatch(cfqq);
846 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200847
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100848 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
Jens Axboe206dc692006-03-28 13:03:44 +0200849
850 /*
851 * we don't want to idle for seeks, but we do want to allow
852 * fair distribution of slice time for a process doing back-to-back
853 * seeks. so allow a little bit of time for him to submit a new rq
854 */
Jens Axboecaaa5f92006-06-16 11:23:00 +0200855 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboe44eb1232006-07-25 15:05:21 +0200856 sl = min(sl, msecs_to_jiffies(2));
Jens Axboe206dc692006-03-28 13:03:44 +0200857
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100858 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe22e2c502005-06-27 10:55:12 +0200859 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860}
861
Jens Axboe5e705372006-07-13 12:39:25 +0200862static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863{
864 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200865 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200866
Jens Axboe5380a102006-07-13 12:37:56 +0200867 cfq_remove_request(rq);
868 cfqq->on_dispatch[rq_is_sync(rq)]++;
869 elv_dispatch_sort(q, rq);
Jens Axboefd61af02006-06-16 15:35:39 +0200870
871 rq = list_entry(q->queue_head.prev, struct request, queuelist);
872 cfqd->last_sector = rq->sector + rq->nr_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873}
874
875/*
876 * return expired entry, or NULL to just start from scratch in rbtree
877 */
Jens Axboe5e705372006-07-13 12:39:25 +0200878static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879{
880 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200881 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200882 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883
Jens Axboe3b181522005-06-27 10:56:24 +0200884 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +1100886
887 cfq_mark_cfqq_fifo_expire(cfqq);
888
Jens Axboe89850f72006-07-22 16:48:31 +0200889 if (list_empty(&cfqq->fifo))
890 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891
Jens Axboe89850f72006-07-22 16:48:31 +0200892 fifo = cfq_cfqq_class_sync(cfqq);
893 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894
Jens Axboecb887412007-01-19 12:01:16 +1100895 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
Jens Axboe89850f72006-07-22 16:48:31 +0200896 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897
898 return NULL;
899}
900
Jens Axboe22e2c502005-06-27 10:55:12 +0200901static inline int
902cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
903{
904 const int base_rq = cfqd->cfq_slice_async_rq;
905
906 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
907
908 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
909}
910
911/*
912 * get next queue for service
913 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100914static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200915{
Jens Axboe22e2c502005-06-27 10:55:12 +0200916 struct cfq_queue *cfqq;
917
918 cfqq = cfqd->active_queue;
919 if (!cfqq)
920 goto new_queue;
921
922 /*
923 * slice has expired
924 */
Jens Axboe44f7c162007-01-19 11:51:58 +1100925 if (!cfq_cfqq_must_dispatch(cfqq) && cfq_slice_used(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +0200926 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +0200927
928 /*
929 * if queue has requests, dispatch one. if not, check if
930 * enough slice is left to wait for one
931 */
Jens Axboedd67d052006-06-21 09:36:18 +0200932 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200933 goto keep_queue;
Jens Axboe44f7c162007-01-19 11:51:58 +1100934 else if (cfq_cfqq_slice_new(cfqq) || cfq_cfqq_dispatched(cfqq)) {
Jens Axboecaaa5f92006-06-16 11:23:00 +0200935 cfqq = NULL;
936 goto keep_queue;
937 } else if (cfq_cfqq_class_sync(cfqq)) {
Jens Axboe17926692007-01-19 11:59:30 +1100938 if (cfq_arm_slice_timer(cfqd))
Jens Axboe22e2c502005-06-27 10:55:12 +0200939 return NULL;
940 }
941
Jens Axboe3b181522005-06-27 10:56:24 +0200942expire:
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100943 cfq_slice_expired(cfqd, 0, 0);
Jens Axboe3b181522005-06-27 10:56:24 +0200944new_queue:
945 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200946keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +0200947 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200948}
949
950static int
951__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
952 int max_dispatch)
953{
954 int dispatched = 0;
955
Jens Axboedd67d052006-06-21 09:36:18 +0200956 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +0200957
958 do {
Jens Axboe5e705372006-07-13 12:39:25 +0200959 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200960
961 /*
962 * follow expired path, else get first next available
963 */
Jens Axboe5e705372006-07-13 12:39:25 +0200964 if ((rq = cfq_check_fifo(cfqq)) == NULL)
965 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200966
967 /*
968 * finally, insert request into driver dispatch list
969 */
Jens Axboe5e705372006-07-13 12:39:25 +0200970 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200971
972 cfqd->dispatch_slice++;
973 dispatched++;
974
975 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +0200976 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
977 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200978 }
979
Jens Axboedd67d052006-06-21 09:36:18 +0200980 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200981 break;
982
983 } while (dispatched < max_dispatch);
984
985 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200986 * expire an async queue immediately if it has used up its slice. idle
987 * queue always expire after 1 dispatch round.
988 */
989 if ((!cfq_cfqq_sync(cfqq) &&
990 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboe44f7c162007-01-19 11:51:58 +1100991 cfq_class_idle(cfqq)) {
992 cfqq->slice_end = jiffies + 1;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100993 cfq_slice_expired(cfqd, 0, 0);
Jens Axboe44f7c162007-01-19 11:51:58 +1100994 }
Jens Axboe22e2c502005-06-27 10:55:12 +0200995
996 return dispatched;
997}
998
999static int
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001000cfq_forced_dispatch_cfqqs(struct list_head *list)
1001{
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001002 struct cfq_queue *cfqq, *next;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001003 int dispatched;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001004
Jens Axboecaaa5f92006-06-16 11:23:00 +02001005 dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001006 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
Jens Axboe5e705372006-07-13 12:39:25 +02001007 while (cfqq->next_rq) {
1008 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001009 dispatched++;
1010 }
1011 BUG_ON(!list_empty(&cfqq->fifo));
1012 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001013
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001014 return dispatched;
1015}
1016
1017static int
1018cfq_forced_dispatch(struct cfq_data *cfqd)
1019{
1020 int i, dispatched = 0;
1021
1022 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1023 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1024
1025 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1026 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1027 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1028
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001029 cfq_slice_expired(cfqd, 0, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001030
1031 BUG_ON(cfqd->busy_queues);
1032
1033 return dispatched;
1034}
1035
1036static int
Jens Axboeb4878f22005-10-20 16:42:29 +02001037cfq_dispatch_requests(request_queue_t *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038{
1039 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001040 struct cfq_queue *cfqq, *prev_cfqq;
1041 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042
Jens Axboe22e2c502005-06-27 10:55:12 +02001043 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 return 0;
1045
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001046 if (unlikely(force))
1047 return cfq_forced_dispatch(cfqd);
1048
Jens Axboecaaa5f92006-06-16 11:23:00 +02001049 dispatched = 0;
1050 prev_cfqq = NULL;
1051 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001052 int max_dispatch;
1053
Jens Axboecaaa5f92006-06-16 11:23:00 +02001054 /*
1055 * Don't repeat dispatch from the previous queue.
1056 */
1057 if (prev_cfqq == cfqq)
1058 break;
1059
Jens Axboe9ede2092007-01-19 12:11:44 +11001060 /*
1061 * So we have dispatched before in this round, if the
1062 * next queue has idling enabled (must be sync), don't
1063 * allow it service until the previous have continued.
1064 */
1065 if (cfqd->rq_in_driver && cfq_cfqq_idle_window(cfqq))
1066 break;
1067
Jens Axboe3b181522005-06-27 10:56:24 +02001068 cfq_clear_cfqq_must_dispatch(cfqq);
1069 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001070 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001072 max_dispatch = cfqd->cfq_quantum;
1073 if (cfq_class_idle(cfqq))
1074 max_dispatch = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075
Jens Axboecaaa5f92006-06-16 11:23:00 +02001076 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001077 prev_cfqq = cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078 }
1079
Jens Axboecaaa5f92006-06-16 11:23:00 +02001080 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081}
1082
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083/*
Jens Axboe5e705372006-07-13 12:39:25 +02001084 * task holds one reference to the queue, dropped when task exits. each rq
1085 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 *
1087 * queue lock must be held here.
1088 */
1089static void cfq_put_queue(struct cfq_queue *cfqq)
1090{
Jens Axboe22e2c502005-06-27 10:55:12 +02001091 struct cfq_data *cfqd = cfqq->cfqd;
1092
1093 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094
1095 if (!atomic_dec_and_test(&cfqq->ref))
1096 return;
1097
1098 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001099 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001100 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001102 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001103 __cfq_slice_expired(cfqd, cfqq, 0, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001104 cfq_schedule_dispatch(cfqd);
1105 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001106
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 /*
1108 * it's on the empty list and still hashed
1109 */
1110 list_del(&cfqq->cfq_list);
1111 hlist_del(&cfqq->cfq_hash);
1112 kmem_cache_free(cfq_pool, cfqq);
1113}
1114
Jens Axboe1ea25ecb2006-07-18 22:24:11 +02001115static struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001116__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1117 const int hashval)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118{
1119 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
Jens Axboe206dc692006-03-28 13:03:44 +02001120 struct hlist_node *entry;
1121 struct cfq_queue *__cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122
Jens Axboe206dc692006-03-28 13:03:44 +02001123 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
Al Virob0a69162006-03-14 15:32:50 -05001124 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125
Jens Axboe206dc692006-03-28 13:03:44 +02001126 if (__cfqq->key == key && (__p == prio || !prio))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 return __cfqq;
1128 }
1129
1130 return NULL;
1131}
1132
1133static struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001134cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135{
Jens Axboe3b181522005-06-27 10:56:24 +02001136 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137}
1138
Jens Axboee2d74ac2006-03-28 08:59:01 +02001139static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140{
Jens Axboe22e2c502005-06-27 10:55:12 +02001141 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001142 struct rb_node *n;
1143 int freed = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001144
Jens Axboee2d74ac2006-03-28 08:59:01 +02001145 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1146 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1147 rb_erase(&__cic->rb_node, &ioc->cic_root);
Jens Axboe22e2c502005-06-27 10:55:12 +02001148 kmem_cache_free(cfq_ioc_pool, __cic);
Al Viro334e94d2006-03-18 15:05:53 -05001149 freed++;
Jens Axboe22e2c502005-06-27 10:55:12 +02001150 }
1151
Jens Axboe4050cf12006-07-19 05:07:12 +02001152 elv_ioc_count_mod(ioc_count, -freed);
1153
1154 if (ioc_gone && !elv_ioc_count_read(ioc_count))
Al Viro334e94d2006-03-18 15:05:53 -05001155 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156}
1157
Jens Axboe89850f72006-07-22 16:48:31 +02001158static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1159{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001160 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001161 __cfq_slice_expired(cfqd, cfqq, 0, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001162 cfq_schedule_dispatch(cfqd);
1163 }
Jens Axboe89850f72006-07-22 16:48:31 +02001164
1165 cfq_put_queue(cfqq);
1166}
1167
1168static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1169 struct cfq_io_context *cic)
1170{
Jens Axboefc463792006-08-29 09:05:44 +02001171 list_del_init(&cic->queue_list);
1172 smp_wmb();
1173 cic->key = NULL;
1174
Jens Axboe89850f72006-07-22 16:48:31 +02001175 if (cic->cfqq[ASYNC]) {
1176 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1177 cic->cfqq[ASYNC] = NULL;
1178 }
1179
1180 if (cic->cfqq[SYNC]) {
1181 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1182 cic->cfqq[SYNC] = NULL;
1183 }
Jens Axboe89850f72006-07-22 16:48:31 +02001184}
1185
1186
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001188 * Called with interrupts disabled
1189 */
1190static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1191{
Al Viro478a82b2006-03-18 13:25:24 -05001192 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001193
Jens Axboe89850f72006-07-22 16:48:31 +02001194 if (cfqd) {
1195 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001196
Jens Axboefc463792006-08-29 09:05:44 +02001197 spin_lock_irq(q->queue_lock);
Jens Axboe89850f72006-07-22 16:48:31 +02001198 __cfq_exit_single_io_context(cfqd, cic);
Jens Axboefc463792006-08-29 09:05:44 +02001199 spin_unlock_irq(q->queue_lock);
Al Viro12a05732006-03-18 13:38:01 -05001200 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001201}
1202
Jens Axboee2d74ac2006-03-28 08:59:01 +02001203static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204{
Jens Axboe22e2c502005-06-27 10:55:12 +02001205 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001206 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +02001207
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 /*
1209 * put the reference this task is holding to the various queues
1210 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001211
1212 n = rb_first(&ioc->cic_root);
1213 while (n != NULL) {
1214 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1215
Jens Axboe22e2c502005-06-27 10:55:12 +02001216 cfq_exit_single_io_context(__cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001217 n = rb_next(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219}
1220
Jens Axboe22e2c502005-06-27 10:55:12 +02001221static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001222cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223{
Jens Axboeb5deef92006-07-19 23:39:40 +02001224 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225
Jens Axboeb5deef92006-07-19 23:39:40 +02001226 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 if (cic) {
Jens Axboe553698f2006-06-14 19:11:57 +02001228 memset(cic, 0, sizeof(*cic));
Jens Axboe22e2c502005-06-27 10:55:12 +02001229 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001230 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001231 cic->dtor = cfq_free_io_context;
1232 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001233 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 }
1235
1236 return cic;
1237}
1238
Jens Axboe22e2c502005-06-27 10:55:12 +02001239static void cfq_init_prio_data(struct cfq_queue *cfqq)
1240{
1241 struct task_struct *tsk = current;
1242 int ioprio_class;
1243
Jens Axboe3b181522005-06-27 10:56:24 +02001244 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001245 return;
1246
1247 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1248 switch (ioprio_class) {
1249 default:
1250 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1251 case IOPRIO_CLASS_NONE:
1252 /*
1253 * no prio set, place us in the middle of the BE classes
1254 */
1255 cfqq->ioprio = task_nice_ioprio(tsk);
1256 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1257 break;
1258 case IOPRIO_CLASS_RT:
1259 cfqq->ioprio = task_ioprio(tsk);
1260 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1261 break;
1262 case IOPRIO_CLASS_BE:
1263 cfqq->ioprio = task_ioprio(tsk);
1264 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1265 break;
1266 case IOPRIO_CLASS_IDLE:
1267 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1268 cfqq->ioprio = 7;
Jens Axboe3b181522005-06-27 10:56:24 +02001269 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001270 break;
1271 }
1272
1273 /*
1274 * keep track of original prio settings in case we have to temporarily
1275 * elevate the priority of this queue
1276 */
1277 cfqq->org_ioprio = cfqq->ioprio;
1278 cfqq->org_ioprio_class = cfqq->ioprio_class;
1279
Jens Axboe98e41c72007-02-05 11:55:35 +01001280 cfq_resort_rr_list(cfqq, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001281 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001282}
1283
Al Viro478a82b2006-03-18 13:25:24 -05001284static inline void changed_ioprio(struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001285{
Al Viro478a82b2006-03-18 13:25:24 -05001286 struct cfq_data *cfqd = cic->key;
1287 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001288 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001289
Jens Axboecaaa5f92006-06-16 11:23:00 +02001290 if (unlikely(!cfqd))
1291 return;
1292
Jens Axboec1b707d2006-10-30 19:54:23 +01001293 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001294
1295 cfqq = cic->cfqq[ASYNC];
1296 if (cfqq) {
1297 struct cfq_queue *new_cfqq;
1298 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1299 GFP_ATOMIC);
1300 if (new_cfqq) {
1301 cic->cfqq[ASYNC] = new_cfqq;
1302 cfq_put_queue(cfqq);
1303 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001304 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001305
1306 cfqq = cic->cfqq[SYNC];
1307 if (cfqq)
1308 cfq_mark_cfqq_prio_changed(cfqq);
1309
Jens Axboec1b707d2006-10-30 19:54:23 +01001310 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001311}
1312
Jens Axboefc463792006-08-29 09:05:44 +02001313static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314{
Al Viroa6a07632006-03-18 13:26:44 -05001315 struct cfq_io_context *cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001316 struct rb_node *n;
Al Viroa6a07632006-03-18 13:26:44 -05001317
Jens Axboefc463792006-08-29 09:05:44 +02001318 ioc->ioprio_changed = 0;
Al Viroa6a07632006-03-18 13:26:44 -05001319
Jens Axboee2d74ac2006-03-28 08:59:01 +02001320 n = rb_first(&ioc->cic_root);
1321 while (n != NULL) {
1322 cic = rb_entry(n, struct cfq_io_context, rb_node);
Jens Axboe3793c652006-05-30 21:11:04 +02001323
Al Viro478a82b2006-03-18 13:25:24 -05001324 changed_ioprio(cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001325 n = rb_next(n);
1326 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327}
1328
1329static struct cfq_queue *
Al Viro6f325a12006-03-18 14:58:37 -05001330cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
Al Viro8267e262005-10-21 03:20:53 -04001331 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332{
1333 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1334 struct cfq_queue *cfqq, *new_cfqq = NULL;
Al Viro6f325a12006-03-18 14:58:37 -05001335 unsigned short ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336
1337retry:
Al Viro6f325a12006-03-18 14:58:37 -05001338 ioprio = tsk->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02001339 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340
1341 if (!cfqq) {
1342 if (new_cfqq) {
1343 cfqq = new_cfqq;
1344 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001345 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001346 /*
1347 * Inform the allocator of the fact that we will
1348 * just repeat this allocation if it fails, to allow
1349 * the allocator to do whatever it needs to attempt to
1350 * free memory.
1351 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 spin_unlock_irq(cfqd->queue->queue_lock);
Jens Axboeb5deef92006-07-19 23:39:40 +02001353 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 spin_lock_irq(cfqd->queue->queue_lock);
1355 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001356 } else {
Jens Axboeb5deef92006-07-19 23:39:40 +02001357 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001358 if (!cfqq)
1359 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001360 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361
1362 memset(cfqq, 0, sizeof(*cfqq));
1363
1364 INIT_HLIST_NODE(&cfqq->cfq_hash);
1365 INIT_LIST_HEAD(&cfqq->cfq_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001366 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367
1368 cfqq->key = key;
1369 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1370 atomic_set(&cfqq->ref, 0);
1371 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001372
Jens Axboecaaa5f92006-06-16 11:23:00 +02001373 cfq_mark_cfqq_idle_window(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001374 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b037442006-07-28 09:48:51 +02001375 cfq_mark_cfqq_queue_new(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001376 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 }
1378
1379 if (new_cfqq)
1380 kmem_cache_free(cfq_pool, new_cfqq);
1381
1382 atomic_inc(&cfqq->ref);
1383out:
1384 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1385 return cfqq;
1386}
1387
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001388static void
1389cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1390{
Jens Axboefc463792006-08-29 09:05:44 +02001391 WARN_ON(!list_empty(&cic->queue_list));
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001392 rb_erase(&cic->rb_node, &ioc->cic_root);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001393 kmem_cache_free(cfq_ioc_pool, cic);
Jens Axboe4050cf12006-07-19 05:07:12 +02001394 elv_ioc_count_dec(ioc_count);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001395}
1396
Jens Axboee2d74ac2006-03-28 08:59:01 +02001397static struct cfq_io_context *
1398cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1399{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001400 struct rb_node *n;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001401 struct cfq_io_context *cic;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001402 void *k, *key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001403
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001404restart:
1405 n = ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001406 while (n) {
1407 cic = rb_entry(n, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001408 /* ->key must be copied to avoid race with cfq_exit_queue() */
1409 k = cic->key;
1410 if (unlikely(!k)) {
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001411 cfq_drop_dead_cic(ioc, cic);
1412 goto restart;
1413 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001414
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001415 if (key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001416 n = n->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001417 else if (key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001418 n = n->rb_right;
1419 else
1420 return cic;
1421 }
1422
1423 return NULL;
1424}
1425
1426static inline void
1427cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1428 struct cfq_io_context *cic)
1429{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001430 struct rb_node **p;
1431 struct rb_node *parent;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001432 struct cfq_io_context *__cic;
Jens Axboe0261d682006-10-30 19:07:48 +01001433 unsigned long flags;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001434 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001435
Jens Axboee2d74ac2006-03-28 08:59:01 +02001436 cic->ioc = ioc;
1437 cic->key = cfqd;
1438
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001439restart:
1440 parent = NULL;
1441 p = &ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001442 while (*p) {
1443 parent = *p;
1444 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001445 /* ->key must be copied to avoid race with cfq_exit_queue() */
1446 k = __cic->key;
1447 if (unlikely(!k)) {
Oleg Nesterovbe33c3a2006-08-21 08:36:12 +02001448 cfq_drop_dead_cic(ioc, __cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001449 goto restart;
1450 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001451
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001452 if (cic->key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001453 p = &(*p)->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001454 else if (cic->key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001455 p = &(*p)->rb_right;
1456 else
1457 BUG();
1458 }
1459
1460 rb_link_node(&cic->rb_node, parent, p);
1461 rb_insert_color(&cic->rb_node, &ioc->cic_root);
Jens Axboefc463792006-08-29 09:05:44 +02001462
Jens Axboe0261d682006-10-30 19:07:48 +01001463 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001464 list_add(&cic->queue_list, &cfqd->cic_list);
Jens Axboe0261d682006-10-30 19:07:48 +01001465 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001466}
1467
Jens Axboe22e2c502005-06-27 10:55:12 +02001468/*
1469 * Setup general io context and cfq io context. There can be several cfq
1470 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001471 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001472 */
1473static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001474cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475{
Jens Axboe22e2c502005-06-27 10:55:12 +02001476 struct io_context *ioc = NULL;
1477 struct cfq_io_context *cic;
1478
1479 might_sleep_if(gfp_mask & __GFP_WAIT);
1480
Jens Axboeb5deef92006-07-19 23:39:40 +02001481 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001482 if (!ioc)
1483 return NULL;
1484
Jens Axboee2d74ac2006-03-28 08:59:01 +02001485 cic = cfq_cic_rb_lookup(cfqd, ioc);
1486 if (cic)
1487 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001488
Jens Axboee2d74ac2006-03-28 08:59:01 +02001489 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1490 if (cic == NULL)
1491 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001492
Jens Axboee2d74ac2006-03-28 08:59:01 +02001493 cfq_cic_link(cfqd, ioc, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001494out:
Jens Axboefc463792006-08-29 09:05:44 +02001495 smp_read_barrier_depends();
1496 if (unlikely(ioc->ioprio_changed))
1497 cfq_ioc_set_ioprio(ioc);
1498
Jens Axboe22e2c502005-06-27 10:55:12 +02001499 return cic;
1500err:
1501 put_io_context(ioc);
1502 return NULL;
1503}
1504
1505static void
1506cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1507{
Jens Axboeaaf12282007-01-19 11:30:16 +11001508 unsigned long elapsed = jiffies - cic->last_end_request;
1509 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001510
1511 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1512 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1513 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1514}
1515
Jens Axboe206dc692006-03-28 13:03:44 +02001516static void
Jens Axboebb37b942006-12-01 10:42:33 +01001517cfq_update_io_seektime(struct cfq_io_context *cic, struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001518{
1519 sector_t sdist;
1520 u64 total;
1521
Jens Axboe5e705372006-07-13 12:39:25 +02001522 if (cic->last_request_pos < rq->sector)
1523 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001524 else
Jens Axboe5e705372006-07-13 12:39:25 +02001525 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001526
1527 /*
1528 * Don't allow the seek distance to get too large from the
1529 * odd fragment, pagein, etc
1530 */
1531 if (cic->seek_samples <= 60) /* second&third seek */
1532 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1533 else
1534 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1535
1536 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1537 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1538 total = cic->seek_total + (cic->seek_samples/2);
1539 do_div(total, cic->seek_samples);
1540 cic->seek_mean = (sector_t)total;
1541}
Jens Axboe22e2c502005-06-27 10:55:12 +02001542
1543/*
1544 * Disable idle window if the process thinks too long or seeks so much that
1545 * it doesn't matter
1546 */
1547static void
1548cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1549 struct cfq_io_context *cic)
1550{
Jens Axboe3b181522005-06-27 10:56:24 +02001551 int enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001552
Jens Axboecaaa5f92006-06-16 11:23:00 +02001553 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1554 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001555 enable_idle = 0;
1556 else if (sample_valid(cic->ttime_samples)) {
1557 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1558 enable_idle = 0;
1559 else
1560 enable_idle = 1;
1561 }
1562
Jens Axboe3b181522005-06-27 10:56:24 +02001563 if (enable_idle)
1564 cfq_mark_cfqq_idle_window(cfqq);
1565 else
1566 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001567}
1568
Jens Axboe22e2c502005-06-27 10:55:12 +02001569/*
1570 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1571 * no or if we aren't sure, a 1 will cause a preempt.
1572 */
1573static int
1574cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001575 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001576{
1577 struct cfq_queue *cfqq = cfqd->active_queue;
1578
1579 if (cfq_class_idle(new_cfqq))
1580 return 0;
1581
1582 if (!cfqq)
Jens Axboecaaa5f92006-06-16 11:23:00 +02001583 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001584
1585 if (cfq_class_idle(cfqq))
1586 return 1;
Jens Axboe3b181522005-06-27 10:56:24 +02001587 if (!cfq_cfqq_wait_request(new_cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001588 return 0;
1589 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02001590 * if the new request is sync, but the currently running queue is
1591 * not, let the sync request have priority.
1592 */
Jens Axboe5e705372006-07-13 12:39:25 +02001593 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001594 return 1;
Jens Axboe374f84a2006-07-23 01:42:19 +02001595 /*
1596 * So both queues are sync. Let the new request get disk time if
1597 * it's a metadata request and the current queue is doing regular IO.
1598 */
1599 if (rq_is_meta(rq) && !cfqq->meta_pending)
1600 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001601
1602 return 0;
1603}
1604
1605/*
1606 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1607 * let it have half of its nominal slice.
1608 */
1609static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1610{
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001611 cfq_slice_expired(cfqd, 1, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001612
Jens Axboebf572252006-07-19 20:29:12 +02001613 /*
1614 * Put the new queue at the front of the of the current list,
1615 * so we know that it will be selected next.
1616 */
1617 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1618 list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1619
Jens Axboe44f7c162007-01-19 11:51:58 +11001620 cfqq->slice_end = 0;
1621 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001622}
1623
1624/*
Jens Axboe5e705372006-07-13 12:39:25 +02001625 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001626 * something we should do about it
1627 */
1628static void
Jens Axboe5e705372006-07-13 12:39:25 +02001629cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1630 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001631{
Jens Axboe5e705372006-07-13 12:39:25 +02001632 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001633
Jens Axboe374f84a2006-07-23 01:42:19 +02001634 if (rq_is_meta(rq))
1635 cfqq->meta_pending++;
1636
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001637 /*
Jens Axboe5380a102006-07-13 12:37:56 +02001638 * check if this request is a better next-serve candidate)) {
Jens Axboe21183b02006-07-13 12:33:14 +02001639 */
Jens Axboe5e705372006-07-13 12:39:25 +02001640 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1641 BUG_ON(!cfqq->next_rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001642
1643 /*
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001644 * we never wait for an async request and we don't allow preemption
1645 * of an async request. so just return early
1646 */
Jens Axboe5e705372006-07-13 12:39:25 +02001647 if (!rq_is_sync(rq)) {
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001648 /*
1649 * sync process issued an async request, if it's waiting
1650 * then expire it and kick rq handling.
1651 */
1652 if (cic == cfqd->active_cic &&
1653 del_timer(&cfqd->idle_slice_timer)) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001654 cfq_slice_expired(cfqd, 0, 0);
Jens Axboedc72ef42006-07-20 14:54:05 +02001655 blk_start_queueing(cfqd->queue);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001656 }
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001657 return;
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001658 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001659
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001660 cfq_update_io_thinktime(cfqd, cic);
Jens Axboebb37b942006-12-01 10:42:33 +01001661 cfq_update_io_seektime(cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001662 cfq_update_idle_window(cfqd, cfqq, cic);
1663
Jens Axboe5e705372006-07-13 12:39:25 +02001664 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001665
1666 if (cfqq == cfqd->active_queue) {
1667 /*
1668 * if we are waiting for a request for this queue, let it rip
1669 * immediately and flag that we must not expire this queue
1670 * just now
1671 */
Jens Axboe3b181522005-06-27 10:56:24 +02001672 if (cfq_cfqq_wait_request(cfqq)) {
1673 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001674 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001675 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001676 }
Jens Axboe5e705372006-07-13 12:39:25 +02001677 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001678 /*
1679 * not the active queue - expire current slice if it is
1680 * idle and has expired it's mean thinktime or this new queue
1681 * has some old slice time left and is of higher priority
1682 */
1683 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001684 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001685 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001686 }
1687}
1688
Jens Axboeb4878f22005-10-20 16:42:29 +02001689static void cfq_insert_request(request_queue_t *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001690{
Jens Axboeb4878f22005-10-20 16:42:29 +02001691 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001692 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001693
1694 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695
Jens Axboe5e705372006-07-13 12:39:25 +02001696 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697
Jens Axboe22e2c502005-06-27 10:55:12 +02001698 list_add_tail(&rq->queuelist, &cfqq->fifo);
1699
Jens Axboe5e705372006-07-13 12:39:25 +02001700 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701}
1702
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703static void cfq_completed_request(request_queue_t *q, struct request *rq)
1704{
Jens Axboe5e705372006-07-13 12:39:25 +02001705 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001706 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001707 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001708 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709
Jens Axboeb4878f22005-10-20 16:42:29 +02001710 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711
Jens Axboeb4878f22005-10-20 16:42:29 +02001712 WARN_ON(!cfqd->rq_in_driver);
1713 WARN_ON(!cfqq->on_dispatch[sync]);
1714 cfqd->rq_in_driver--;
1715 cfqq->on_dispatch[sync]--;
Jens Axboe99f96282007-02-05 11:56:25 +01001716 cfqq->service_last = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717
Jens Axboeb4878f22005-10-20 16:42:29 +02001718 if (!cfq_class_idle(cfqq))
1719 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001720
Jens Axboe98e41c72007-02-05 11:55:35 +01001721 cfq_resort_rr_list(cfqq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722
Jens Axboecaaa5f92006-06-16 11:23:00 +02001723 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001724 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001725
1726 /*
1727 * If this is the active queue, check if it needs to be expired,
1728 * or if we want to idle in case it has no pending requests.
1729 */
1730 if (cfqd->active_queue == cfqq) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001731 if (cfq_cfqq_slice_new(cfqq)) {
1732 cfq_set_prio_slice(cfqd, cfqq);
1733 cfq_clear_cfqq_slice_new(cfqq);
1734 }
1735 if (cfq_slice_used(cfqq))
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001736 cfq_slice_expired(cfqd, 0, 1);
Jens Axboedd67d052006-06-21 09:36:18 +02001737 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe17926692007-01-19 11:59:30 +11001738 if (!cfq_arm_slice_timer(cfqd))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001739 cfq_schedule_dispatch(cfqd);
1740 }
1741 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742}
1743
Jens Axboe22e2c502005-06-27 10:55:12 +02001744/*
1745 * we temporarily boost lower priority queues if they are holding fs exclusive
1746 * resources. they are boosted to normal prio (CLASS_BE/4)
1747 */
1748static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749{
Jens Axboe22e2c502005-06-27 10:55:12 +02001750 const int ioprio_class = cfqq->ioprio_class;
1751 const int ioprio = cfqq->ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752
Jens Axboe22e2c502005-06-27 10:55:12 +02001753 if (has_fs_excl()) {
1754 /*
1755 * boost idle prio on transactions that would lock out other
1756 * users of the filesystem
1757 */
1758 if (cfq_class_idle(cfqq))
1759 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1760 if (cfqq->ioprio > IOPRIO_NORM)
1761 cfqq->ioprio = IOPRIO_NORM;
1762 } else {
1763 /*
1764 * check if we need to unboost the queue
1765 */
1766 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1767 cfqq->ioprio_class = cfqq->org_ioprio_class;
1768 if (cfqq->ioprio != cfqq->org_ioprio)
1769 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 }
1771
Jens Axboe22e2c502005-06-27 10:55:12 +02001772 /*
1773 * refile between round-robin lists if we moved the priority class
1774 */
Jens Axboe98e41c72007-02-05 11:55:35 +01001775 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio))
Jens Axboe22e2c502005-06-27 10:55:12 +02001776 cfq_resort_rr_list(cfqq, 0);
1777}
1778
Jens Axboe89850f72006-07-22 16:48:31 +02001779static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001780{
Jens Axboe3b181522005-06-27 10:56:24 +02001781 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001782 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001783 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001784 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001785 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001786
1787 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02001788}
1789
Jens Axboecb78b282006-07-28 09:32:57 +02001790static int cfq_may_queue(request_queue_t *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02001791{
1792 struct cfq_data *cfqd = q->elevator->elevator_data;
1793 struct task_struct *tsk = current;
1794 struct cfq_queue *cfqq;
Jens Axboe7749a8d2006-12-13 13:02:26 +01001795 unsigned int key;
1796
1797 key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02001798
1799 /*
1800 * don't force setup of a queue from here, as a call to may_queue
1801 * does not necessarily imply that a request actually will be queued.
1802 * so just lookup a possibly existing queue, or return 'may queue'
1803 * if that fails
1804 */
Jens Axboe7749a8d2006-12-13 13:02:26 +01001805 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001806 if (cfqq) {
1807 cfq_init_prio_data(cfqq);
1808 cfq_prio_boost(cfqq);
1809
Jens Axboe89850f72006-07-22 16:48:31 +02001810 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001811 }
1812
1813 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814}
1815
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816/*
1817 * queue lock held here
1818 */
Jens Axboebb37b942006-12-01 10:42:33 +01001819static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820{
Jens Axboe5e705372006-07-13 12:39:25 +02001821 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822
Jens Axboe5e705372006-07-13 12:39:25 +02001823 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001824 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825
Jens Axboe22e2c502005-06-27 10:55:12 +02001826 BUG_ON(!cfqq->allocated[rw]);
1827 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828
Jens Axboe5e705372006-07-13 12:39:25 +02001829 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02001832 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 cfq_put_queue(cfqq);
1835 }
1836}
1837
1838/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001839 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001841static int
Jens Axboecb78b282006-07-28 09:32:57 +02001842cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843{
1844 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe3b181522005-06-27 10:56:24 +02001845 struct task_struct *tsk = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 struct cfq_io_context *cic;
1847 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01001848 const int is_sync = rq_is_sync(rq);
1849 pid_t key = cfq_queue_pid(tsk, rw, is_sync);
Jens Axboe22e2c502005-06-27 10:55:12 +02001850 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851 unsigned long flags;
1852
1853 might_sleep_if(gfp_mask & __GFP_WAIT);
1854
Jens Axboee2d74ac2006-03-28 08:59:01 +02001855 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001856
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 spin_lock_irqsave(q->queue_lock, flags);
1858
Jens Axboe22e2c502005-06-27 10:55:12 +02001859 if (!cic)
1860 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861
Al Viro12a05732006-03-18 13:38:01 -05001862 if (!cic->cfqq[is_sync]) {
Al Viro6f325a12006-03-18 14:58:37 -05001863 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001864 if (!cfqq)
1865 goto queue_fail;
1866
Al Viro12a05732006-03-18 13:38:01 -05001867 cic->cfqq[is_sync] = cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001868 } else
Al Viro12a05732006-03-18 13:38:01 -05001869 cfqq = cic->cfqq[is_sync];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870
1871 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02001872 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001873 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02001874
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 spin_unlock_irqrestore(q->queue_lock, flags);
1876
Jens Axboe5e705372006-07-13 12:39:25 +02001877 rq->elevator_private = cic;
1878 rq->elevator_private2 = cfqq;
1879 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02001880
Jens Axboe22e2c502005-06-27 10:55:12 +02001881queue_fail:
1882 if (cic)
1883 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02001884
Jens Axboe3b181522005-06-27 10:56:24 +02001885 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 spin_unlock_irqrestore(q->queue_lock, flags);
1887 return 1;
1888}
1889
David Howells65f27f32006-11-22 14:55:48 +00001890static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02001891{
David Howells65f27f32006-11-22 14:55:48 +00001892 struct cfq_data *cfqd =
1893 container_of(work, struct cfq_data, unplug_work);
1894 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001895 unsigned long flags;
1896
1897 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02001898 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02001899 spin_unlock_irqrestore(q->queue_lock, flags);
1900}
1901
1902/*
1903 * Timer running if the active_queue is currently idling inside its time slice
1904 */
1905static void cfq_idle_slice_timer(unsigned long data)
1906{
1907 struct cfq_data *cfqd = (struct cfq_data *) data;
1908 struct cfq_queue *cfqq;
1909 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001910 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001911
1912 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1913
1914 if ((cfqq = cfqd->active_queue) != NULL) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001915 timed_out = 0;
1916
Jens Axboe22e2c502005-06-27 10:55:12 +02001917 /*
1918 * expired
1919 */
Jens Axboe44f7c162007-01-19 11:51:58 +11001920 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001921 goto expire;
1922
1923 /*
1924 * only expire and reinvoke request handler, if there are
1925 * other queues with pending requests
1926 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02001927 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02001928 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02001929
1930 /*
1931 * not expired and it has a request pending, let it dispatch
1932 */
Jens Axboedd67d052006-06-21 09:36:18 +02001933 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001934 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001935 goto out_kick;
1936 }
1937 }
1938expire:
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001939 cfq_slice_expired(cfqd, 0, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02001940out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02001941 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001942out_cont:
1943 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1944}
1945
1946/*
1947 * Timer running if an idle class queue is waiting for service
1948 */
1949static void cfq_idle_class_timer(unsigned long data)
1950{
1951 struct cfq_data *cfqd = (struct cfq_data *) data;
1952 unsigned long flags, end;
1953
1954 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1955
1956 /*
1957 * race with a non-idle queue, reset timer
1958 */
1959 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
Jens Axboeae818a32006-06-01 10:13:43 +02001960 if (!time_after_eq(jiffies, end))
1961 mod_timer(&cfqd->idle_class_timer, end);
1962 else
Jens Axboe3b181522005-06-27 10:56:24 +02001963 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001964
1965 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1966}
1967
Jens Axboe3b181522005-06-27 10:56:24 +02001968static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1969{
1970 del_timer_sync(&cfqd->idle_slice_timer);
1971 del_timer_sync(&cfqd->idle_class_timer);
1972 blk_sync_queue(cfqd->queue);
1973}
Jens Axboe22e2c502005-06-27 10:55:12 +02001974
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975static void cfq_exit_queue(elevator_t *e)
1976{
Jens Axboe22e2c502005-06-27 10:55:12 +02001977 struct cfq_data *cfqd = e->elevator_data;
Al Virod9ff4182006-03-18 13:51:22 -05001978 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001979
Jens Axboe3b181522005-06-27 10:56:24 +02001980 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001981
Al Virod9ff4182006-03-18 13:51:22 -05001982 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001983
Al Virod9ff4182006-03-18 13:51:22 -05001984 if (cfqd->active_queue)
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001985 __cfq_slice_expired(cfqd, cfqd->active_queue, 0, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001986
1987 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05001988 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1989 struct cfq_io_context,
1990 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02001991
1992 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05001993 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001994
Al Virod9ff4182006-03-18 13:51:22 -05001995 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05001996
1997 cfq_shutdown_timer_wq(cfqd);
1998
Al Viroa90d7422006-03-18 12:05:37 -05001999 kfree(cfqd->cfq_hash);
2000 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001}
2002
Jens Axboebb37b942006-12-01 10:42:33 +01002003static void *cfq_init_queue(request_queue_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004{
2005 struct cfq_data *cfqd;
2006 int i;
2007
Jens Axboeb5deef92006-07-19 23:39:40 +02002008 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002010 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011
2012 memset(cfqd, 0, sizeof(*cfqd));
Jens Axboe22e2c502005-06-27 10:55:12 +02002013
2014 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2015 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2016
2017 INIT_LIST_HEAD(&cfqd->busy_rr);
2018 INIT_LIST_HEAD(&cfqd->cur_rr);
2019 INIT_LIST_HEAD(&cfqd->idle_rr);
Al Virod9ff4182006-03-18 13:51:22 -05002020 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021
Jens Axboeb5deef92006-07-19 23:39:40 +02002022 cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 if (!cfqd->cfq_hash)
Jens Axboe5e705372006-07-13 12:39:25 +02002024 goto out_free;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2027 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2028
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030
Jens Axboe22e2c502005-06-27 10:55:12 +02002031 init_timer(&cfqd->idle_slice_timer);
2032 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2033 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2034
2035 init_timer(&cfqd->idle_class_timer);
2036 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2037 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2038
David Howells65f27f32006-11-22 14:55:48 +00002039 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002040
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002042 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2043 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044 cfqd->cfq_back_max = cfq_back_max;
2045 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002046 cfqd->cfq_slice[0] = cfq_slice_async;
2047 cfqd->cfq_slice[1] = cfq_slice_sync;
2048 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2049 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02002050
Jens Axboebc1c1162006-06-08 08:49:06 +02002051 return cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +02002052out_free:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053 kfree(cfqd);
Jens Axboebc1c1162006-06-08 08:49:06 +02002054 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055}
2056
2057static void cfq_slab_kill(void)
2058{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 if (cfq_pool)
2060 kmem_cache_destroy(cfq_pool);
2061 if (cfq_ioc_pool)
2062 kmem_cache_destroy(cfq_ioc_pool);
2063}
2064
2065static int __init cfq_slab_setup(void)
2066{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2068 NULL, NULL);
2069 if (!cfq_pool)
2070 goto fail;
2071
2072 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2073 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2074 if (!cfq_ioc_pool)
2075 goto fail;
2076
2077 return 0;
2078fail:
2079 cfq_slab_kill();
2080 return -ENOMEM;
2081}
2082
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083/*
2084 * sysfs parts below -->
2085 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086
2087static ssize_t
2088cfq_var_show(unsigned int var, char *page)
2089{
2090 return sprintf(page, "%d\n", var);
2091}
2092
2093static ssize_t
2094cfq_var_store(unsigned int *var, const char *page, size_t count)
2095{
2096 char *p = (char *) page;
2097
2098 *var = simple_strtoul(p, &p, 10);
2099 return count;
2100}
2101
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002103static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002105 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106 unsigned int __data = __VAR; \
2107 if (__CONV) \
2108 __data = jiffies_to_msecs(__data); \
2109 return cfq_var_show(__data, (page)); \
2110}
2111SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002112SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2113SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002114SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2115SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002116SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2117SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2118SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2119SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120#undef SHOW_FUNCTION
2121
2122#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002123static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002125 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 unsigned int __data; \
2127 int ret = cfq_var_store(&__data, (page), count); \
2128 if (__data < (MIN)) \
2129 __data = (MIN); \
2130 else if (__data > (MAX)) \
2131 __data = (MAX); \
2132 if (__CONV) \
2133 *(__PTR) = msecs_to_jiffies(__data); \
2134 else \
2135 *(__PTR) = __data; \
2136 return ret; \
2137}
2138STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002139STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2140STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002141STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2142STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002143STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2144STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2145STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2146STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147#undef STORE_FUNCTION
2148
Al Viroe572ec72006-03-18 22:27:18 -05002149#define CFQ_ATTR(name) \
2150 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002151
Al Viroe572ec72006-03-18 22:27:18 -05002152static struct elv_fs_entry cfq_attrs[] = {
2153 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002154 CFQ_ATTR(fifo_expire_sync),
2155 CFQ_ATTR(fifo_expire_async),
2156 CFQ_ATTR(back_seek_max),
2157 CFQ_ATTR(back_seek_penalty),
2158 CFQ_ATTR(slice_sync),
2159 CFQ_ATTR(slice_async),
2160 CFQ_ATTR(slice_async_rq),
2161 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002162 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163};
2164
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165static struct elevator_type iosched_cfq = {
2166 .ops = {
2167 .elevator_merge_fn = cfq_merge,
2168 .elevator_merged_fn = cfq_merged_request,
2169 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002170 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002171 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002173 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 .elevator_deactivate_req_fn = cfq_deactivate_request,
2175 .elevator_queue_empty_fn = cfq_queue_empty,
2176 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002177 .elevator_former_req_fn = elv_rb_former_request,
2178 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 .elevator_set_req_fn = cfq_set_request,
2180 .elevator_put_req_fn = cfq_put_request,
2181 .elevator_may_queue_fn = cfq_may_queue,
2182 .elevator_init_fn = cfq_init_queue,
2183 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002184 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 },
Al Viro3d1ab402006-03-18 18:35:43 -05002186 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187 .elevator_name = "cfq",
2188 .elevator_owner = THIS_MODULE,
2189};
2190
2191static int __init cfq_init(void)
2192{
2193 int ret;
2194
Jens Axboe22e2c502005-06-27 10:55:12 +02002195 /*
2196 * could be 0 on HZ < 1000 setups
2197 */
2198 if (!cfq_slice_async)
2199 cfq_slice_async = 1;
2200 if (!cfq_slice_idle)
2201 cfq_slice_idle = 1;
2202
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 if (cfq_slab_setup())
2204 return -ENOMEM;
2205
2206 ret = elv_register(&iosched_cfq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002207 if (ret)
2208 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 return ret;
2211}
2212
2213static void __exit cfq_exit(void)
2214{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002215 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002217 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002218 /* ioc_gone's update must be visible before reading ioc_count */
2219 smp_wmb();
Jens Axboe4050cf12006-07-19 05:07:12 +02002220 if (elv_ioc_count_read(ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002221 wait_for_completion(ioc_gone);
Al Viro334e94d2006-03-18 15:05:53 -05002222 synchronize_rcu();
Christoph Hellwig83521d32005-10-30 15:01:39 -08002223 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224}
2225
2226module_init(cfq_init);
2227module_exit(cfq_exit);
2228
2229MODULE_AUTHOR("Jens Axboe");
2230MODULE_LICENSE("GPL");
2231MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");