blob: f92ba2a869b4fe32348002cf00260ce6ac943542 [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 Axboeb0b8d742007-01-19 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 Axboe53b03742006-07-28 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 Axboe53b03742006-07-28 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 Axboe53b03742006-07-28 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 Axboe53b03742006-07-28 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);
Jens Axboe5044eed2007-04-25 11:53:48 +0200535
536 /*
537 * check if this request is a better next-serve candidate
538 */
539 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
540 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541}
542
543static inline void
Jens Axboe5e705372006-07-13 12:39:25 +0200544cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545{
Jens Axboe5380a102006-07-13 12:37:56 +0200546 elv_rb_del(&cfqq->sort_list, rq);
547 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200548 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549}
550
Jens Axboe206dc692006-03-28 13:03:44 +0200551static struct request *
552cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553{
Jens Axboe206dc692006-03-28 13:03:44 +0200554 struct task_struct *tsk = current;
Jens Axboe7749a8d2006-12-13 13:02:26 +0100555 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
Jens Axboe206dc692006-03-28 13:03:44 +0200556 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557
Jens Axboe206dc692006-03-28 13:03:44 +0200558 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
Jens Axboe89850f72006-07-22 16:48:31 +0200559 if (cfqq) {
560 sector_t sector = bio->bi_sector + bio_sectors(bio);
561
Jens Axboe21183b02006-07-13 12:33:14 +0200562 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200563 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565 return NULL;
566}
567
Jens Axboeb4878f22005-10-20 16:42:29 +0200568static void cfq_activate_request(request_queue_t *q, struct request *rq)
569{
570 struct cfq_data *cfqd = q->elevator->elevator_data;
571
572 cfqd->rq_in_driver++;
Jens Axboe25776e32006-06-01 10:12:26 +0200573
574 /*
575 * If the depth is larger 1, it really could be queueing. But lets
576 * make the mark a little higher - idling could still be good for
577 * low queueing, and a low queueing number could also just indicate
578 * a SCSI mid layer like behaviour where limit+1 is often seen.
579 */
580 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
581 cfqd->hw_tag = 1;
Jens Axboeb4878f22005-10-20 16:42:29 +0200582}
583
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
585{
Jens Axboe22e2c502005-06-27 10:55:12 +0200586 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587
Jens Axboeb4878f22005-10-20 16:42:29 +0200588 WARN_ON(!cfqd->rq_in_driver);
589 cfqd->rq_in_driver--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590}
591
Jens Axboeb4878f22005-10-20 16:42:29 +0200592static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593{
Jens Axboe5e705372006-07-13 12:39:25 +0200594 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200595
Jens Axboe5e705372006-07-13 12:39:25 +0200596 if (cfqq->next_rq == rq)
597 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598
Jens Axboeb4878f22005-10-20 16:42:29 +0200599 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200600 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200601
602 if (rq_is_meta(rq)) {
603 WARN_ON(!cfqq->meta_pending);
604 cfqq->meta_pending--;
605 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606}
607
608static int
609cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
610{
611 struct cfq_data *cfqd = q->elevator->elevator_data;
612 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613
Jens Axboe206dc692006-03-28 13:03:44 +0200614 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200615 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200616 *req = __rq;
617 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 }
619
620 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621}
622
Jens Axboe21183b02006-07-13 12:33:14 +0200623static void cfq_merged_request(request_queue_t *q, struct request *req,
624 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625{
Jens Axboe21183b02006-07-13 12:33:14 +0200626 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200627 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628
Jens Axboe5e705372006-07-13 12:39:25 +0200629 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631}
632
633static void
634cfq_merged_requests(request_queue_t *q, struct request *rq,
635 struct request *next)
636{
Jens Axboe22e2c502005-06-27 10:55:12 +0200637 /*
638 * reposition in fifo if next is older than rq
639 */
640 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
641 time_before(next->start_time, rq->start_time))
642 list_move(&rq->queuelist, &next->queuelist);
643
Jens Axboeb4878f22005-10-20 16:42:29 +0200644 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200645}
646
Jens Axboeda775262006-12-20 11:04:12 +0100647static int cfq_allow_merge(request_queue_t *q, struct request *rq,
648 struct bio *bio)
649{
650 struct cfq_data *cfqd = q->elevator->elevator_data;
651 const int rw = bio_data_dir(bio);
652 struct cfq_queue *cfqq;
653 pid_t key;
654
655 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100656 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100657 */
Jens Axboeec8acb62007-01-02 18:32:11 +0100658 if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100659 return 0;
660
661 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100662 * Lookup the cfqq that this bio will be queued with. Allow
663 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100664 */
Jens Axboe719d3402006-12-22 09:38:53 +0100665 key = cfq_queue_pid(current, rw, bio_sync(bio));
Jens Axboeda775262006-12-20 11:04:12 +0100666 cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
Jens Axboe719d3402006-12-22 09:38:53 +0100667
668 if (cfqq == RQ_CFQQ(rq))
669 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100670
Jens Axboeec8acb62007-01-02 18:32:11 +0100671 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100672}
673
Jens Axboe22e2c502005-06-27 10:55:12 +0200674static inline void
675__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
676{
677 if (cfqq) {
678 /*
679 * stop potential idle class queues waiting service
680 */
681 del_timer(&cfqd->idle_class_timer);
682
Jens Axboe22e2c502005-06-27 10:55:12 +0200683 cfqq->slice_end = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200684 cfq_clear_cfqq_must_alloc_slice(cfqq);
685 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100686 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200687 }
688
689 cfqd->active_queue = cfqq;
690}
691
692/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100693 * current cfqq expired its slice (or was too idle), select new one
694 */
695static void
696__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100697 int preempted, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100698{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100699 if (cfq_cfqq_wait_request(cfqq))
700 del_timer(&cfqd->idle_slice_timer);
701
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100702 cfq_clear_cfqq_must_dispatch(cfqq);
703 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe53b03742006-07-28 09:48:51 +0200704 cfq_clear_cfqq_queue_new(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100705
706 /*
707 * store what was left of this slice, if the queue idled out
708 * or was preempted
709 */
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100710 if (timed_out && !cfq_cfqq_slice_new(cfqq))
Jens Axboec5b680f2007-01-19 11:56:49 +1100711 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100712
Jens Axboe98e41c72007-02-05 11:55:35 +0100713 cfq_resort_rr_list(cfqq, preempted);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100714
715 if (cfqq == cfqd->active_queue)
716 cfqd->active_queue = NULL;
717
718 if (cfqd->active_cic) {
719 put_io_context(cfqd->active_cic->ioc);
720 cfqd->active_cic = NULL;
721 }
722
723 cfqd->dispatch_slice = 0;
724}
725
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100726static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted,
727 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100728{
729 struct cfq_queue *cfqq = cfqd->active_queue;
730
731 if (cfqq)
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100732 __cfq_slice_expired(cfqd, cfqq, preempted, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100733}
734
735/*
Jens Axboe22e2c502005-06-27 10:55:12 +0200736 * 0
737 * 0,1
738 * 0,1,2
739 * 0,1,2,3
740 * 0,1,2,3,4
741 * 0,1,2,3,4,5
742 * 0,1,2,3,4,5,6
743 * 0,1,2,3,4,5,6,7
744 */
745static int cfq_get_next_prio_level(struct cfq_data *cfqd)
746{
747 int prio, wrap;
748
749 prio = -1;
750 wrap = 0;
751 do {
752 int p;
753
754 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
755 if (!list_empty(&cfqd->rr_list[p])) {
756 prio = p;
757 break;
758 }
759 }
760
761 if (prio != -1)
762 break;
763 cfqd->cur_prio = 0;
764 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
765 cfqd->cur_end_prio = 0;
766 if (wrap)
767 break;
768 wrap = 1;
769 }
770 } while (1);
771
772 if (unlikely(prio == -1))
773 return -1;
774
775 BUG_ON(prio >= CFQ_PRIO_LISTS);
776
777 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
778
779 cfqd->cur_prio = prio + 1;
780 if (cfqd->cur_prio > cfqd->cur_end_prio) {
781 cfqd->cur_end_prio = cfqd->cur_prio;
782 cfqd->cur_prio = 0;
783 }
784 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
785 cfqd->cur_prio = 0;
786 cfqd->cur_end_prio = 0;
787 }
788
789 return prio;
790}
791
Jens Axboe3b181522005-06-27 10:56:24 +0200792static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200793{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100794 struct cfq_queue *cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +0200795
Jens Axboe89850f72006-07-22 16:48:31 +0200796 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
797 /*
798 * if current list is non-empty, grab first entry. if it is
799 * empty, get next prio level and grab first entry then if any
800 * are spliced
801 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200802 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
Jens Axboe89850f72006-07-22 16:48:31 +0200803 } else if (!list_empty(&cfqd->busy_rr)) {
804 /*
805 * If no new queues are available, check if the busy list has
806 * some before falling back to idle io.
807 */
Jens Axboee0de0202006-06-01 10:07:26 +0200808 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
Jens Axboe89850f72006-07-22 16:48:31 +0200809 } else if (!list_empty(&cfqd->idle_rr)) {
810 /*
811 * if we have idle queues and no rt or be queues had pending
812 * requests, either allow immediate service if the grace period
813 * has passed or arm the idle grace timer
814 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200815 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
816
817 if (time_after_eq(jiffies, end))
818 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
819 else
820 mod_timer(&cfqd->idle_class_timer, end);
821 }
822
823 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200824 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200825}
826
Jens Axboecaaa5f92006-06-16 11:23:00 +0200827#define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
828
Jens Axboe17926692007-01-19 11:59:30 +1100829static int cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200830{
Jens Axboe17926692007-01-19 11:59:30 +1100831 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +0200832 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100833 unsigned long sl;
834
Jens Axboedd67d052006-06-21 09:36:18 +0200835 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +0200836
837 /*
838 * idle is disabled, either manually or by past process history
839 */
840 if (!cfqd->cfq_slice_idle)
841 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200842 if (!cfq_cfqq_idle_window(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200843 return 0;
844 /*
845 * task has exited, don't wait
846 */
Jens Axboe206dc692006-03-28 13:03:44 +0200847 cic = cfqd->active_cic;
848 if (!cic || !cic->ioc->task)
Jens Axboe22e2c502005-06-27 10:55:12 +0200849 return 0;
850
Jens Axboe3b181522005-06-27 10:56:24 +0200851 cfq_mark_cfqq_must_dispatch(cfqq);
852 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200853
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100854 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
Jens Axboe206dc692006-03-28 13:03:44 +0200855
856 /*
857 * we don't want to idle for seeks, but we do want to allow
858 * fair distribution of slice time for a process doing back-to-back
859 * seeks. so allow a little bit of time for him to submit a new rq
860 */
Jens Axboecaaa5f92006-06-16 11:23:00 +0200861 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboe44eb1232006-07-25 15:05:21 +0200862 sl = min(sl, msecs_to_jiffies(2));
Jens Axboe206dc692006-03-28 13:03:44 +0200863
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100864 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe22e2c502005-06-27 10:55:12 +0200865 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866}
867
Jens Axboe5e705372006-07-13 12:39:25 +0200868static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869{
870 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200871 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200872
Jens Axboe5380a102006-07-13 12:37:56 +0200873 cfq_remove_request(rq);
874 cfqq->on_dispatch[rq_is_sync(rq)]++;
875 elv_dispatch_sort(q, rq);
Jens Axboefd61af02006-06-16 15:35:39 +0200876
877 rq = list_entry(q->queue_head.prev, struct request, queuelist);
878 cfqd->last_sector = rq->sector + rq->nr_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879}
880
881/*
882 * return expired entry, or NULL to just start from scratch in rbtree
883 */
Jens Axboe5e705372006-07-13 12:39:25 +0200884static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885{
886 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200887 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200888 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889
Jens Axboe3b181522005-06-27 10:56:24 +0200890 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +1100892
893 cfq_mark_cfqq_fifo_expire(cfqq);
894
Jens Axboe89850f72006-07-22 16:48:31 +0200895 if (list_empty(&cfqq->fifo))
896 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897
Jens Axboe89850f72006-07-22 16:48:31 +0200898 fifo = cfq_cfqq_class_sync(cfqq);
899 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900
Jens Axboecb887412007-01-19 12:01:16 +1100901 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
Jens Axboe89850f72006-07-22 16:48:31 +0200902 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903
904 return NULL;
905}
906
Jens Axboe22e2c502005-06-27 10:55:12 +0200907static inline int
908cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
909{
910 const int base_rq = cfqd->cfq_slice_async_rq;
911
912 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
913
914 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
915}
916
917/*
918 * get next queue for service
919 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100920static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200921{
Jens Axboe22e2c502005-06-27 10:55:12 +0200922 struct cfq_queue *cfqq;
923
924 cfqq = cfqd->active_queue;
925 if (!cfqq)
926 goto new_queue;
927
928 /*
929 * slice has expired
930 */
Jens Axboe44f7c162007-01-19 11:51:58 +1100931 if (!cfq_cfqq_must_dispatch(cfqq) && cfq_slice_used(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +0200932 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +0200933
934 /*
935 * if queue has requests, dispatch one. if not, check if
936 * enough slice is left to wait for one
937 */
Jens Axboedd67d052006-06-21 09:36:18 +0200938 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200939 goto keep_queue;
Jens Axboe44f7c162007-01-19 11:51:58 +1100940 else if (cfq_cfqq_slice_new(cfqq) || cfq_cfqq_dispatched(cfqq)) {
Jens Axboecaaa5f92006-06-16 11:23:00 +0200941 cfqq = NULL;
942 goto keep_queue;
943 } else if (cfq_cfqq_class_sync(cfqq)) {
Jens Axboe17926692007-01-19 11:59:30 +1100944 if (cfq_arm_slice_timer(cfqd))
Jens Axboe22e2c502005-06-27 10:55:12 +0200945 return NULL;
946 }
947
Jens Axboe3b181522005-06-27 10:56:24 +0200948expire:
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100949 cfq_slice_expired(cfqd, 0, 0);
Jens Axboe3b181522005-06-27 10:56:24 +0200950new_queue:
951 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200952keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +0200953 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200954}
955
956static int
957__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
958 int max_dispatch)
959{
960 int dispatched = 0;
961
Jens Axboedd67d052006-06-21 09:36:18 +0200962 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +0200963
964 do {
Jens Axboe5e705372006-07-13 12:39:25 +0200965 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200966
967 /*
968 * follow expired path, else get first next available
969 */
Jens Axboe5e705372006-07-13 12:39:25 +0200970 if ((rq = cfq_check_fifo(cfqq)) == NULL)
971 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200972
973 /*
974 * finally, insert request into driver dispatch list
975 */
Jens Axboe5e705372006-07-13 12:39:25 +0200976 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200977
978 cfqd->dispatch_slice++;
979 dispatched++;
980
981 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +0200982 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
983 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200984 }
985
Jens Axboedd67d052006-06-21 09:36:18 +0200986 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200987 break;
988
989 } while (dispatched < max_dispatch);
990
991 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200992 * expire an async queue immediately if it has used up its slice. idle
993 * queue always expire after 1 dispatch round.
994 */
Jens Axboea9938002007-04-20 08:55:52 +0200995 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
Jens Axboe22e2c502005-06-27 10:55:12 +0200996 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboea9938002007-04-20 08:55:52 +0200997 cfq_class_idle(cfqq))) {
Jens Axboe44f7c162007-01-19 11:51:58 +1100998 cfqq->slice_end = jiffies + 1;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100999 cfq_slice_expired(cfqd, 0, 0);
Jens Axboe44f7c162007-01-19 11:51:58 +11001000 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001001
1002 return dispatched;
1003}
1004
1005static int
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001006cfq_forced_dispatch_cfqqs(struct list_head *list)
1007{
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001008 struct cfq_queue *cfqq, *next;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001009 int dispatched;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001010
Jens Axboecaaa5f92006-06-16 11:23:00 +02001011 dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001012 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
Jens Axboe5e705372006-07-13 12:39:25 +02001013 while (cfqq->next_rq) {
1014 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001015 dispatched++;
1016 }
1017 BUG_ON(!list_empty(&cfqq->fifo));
1018 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001019
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001020 return dispatched;
1021}
1022
1023static int
1024cfq_forced_dispatch(struct cfq_data *cfqd)
1025{
1026 int i, dispatched = 0;
1027
1028 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1029 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1030
1031 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1032 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1033 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1034
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001035 cfq_slice_expired(cfqd, 0, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001036
1037 BUG_ON(cfqd->busy_queues);
1038
1039 return dispatched;
1040}
1041
1042static int
Jens Axboeb4878f22005-10-20 16:42:29 +02001043cfq_dispatch_requests(request_queue_t *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044{
1045 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001046 struct cfq_queue *cfqq, *prev_cfqq;
1047 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048
Jens Axboe22e2c502005-06-27 10:55:12 +02001049 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050 return 0;
1051
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001052 if (unlikely(force))
1053 return cfq_forced_dispatch(cfqd);
1054
Jens Axboecaaa5f92006-06-16 11:23:00 +02001055 dispatched = 0;
1056 prev_cfqq = NULL;
1057 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001058 int max_dispatch;
1059
Jens Axboea9938002007-04-20 08:55:52 +02001060 if (cfqd->busy_queues > 1) {
1061 /*
1062 * Don't repeat dispatch from the previous queue.
1063 */
1064 if (prev_cfqq == cfqq)
1065 break;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001066
Jens Axboea9938002007-04-20 08:55:52 +02001067 /*
1068 * So we have dispatched before in this round, if the
1069 * next queue has idling enabled (must be sync), don't
1070 * allow it service until the previous have continued.
1071 */
1072 if (cfqd->rq_in_driver && cfq_cfqq_idle_window(cfqq))
1073 break;
1074 }
Jens Axboe9ede2092007-01-19 12:11:44 +11001075
Jens Axboe3b181522005-06-27 10:56:24 +02001076 cfq_clear_cfqq_must_dispatch(cfqq);
1077 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001078 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001080 max_dispatch = cfqd->cfq_quantum;
1081 if (cfq_class_idle(cfqq))
1082 max_dispatch = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083
Jens Axboecaaa5f92006-06-16 11:23:00 +02001084 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001085 prev_cfqq = cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 }
1087
Jens Axboecaaa5f92006-06-16 11:23:00 +02001088 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089}
1090
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091/*
Jens Axboe5e705372006-07-13 12:39:25 +02001092 * task holds one reference to the queue, dropped when task exits. each rq
1093 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094 *
1095 * queue lock must be held here.
1096 */
1097static void cfq_put_queue(struct cfq_queue *cfqq)
1098{
Jens Axboe22e2c502005-06-27 10:55:12 +02001099 struct cfq_data *cfqd = cfqq->cfqd;
1100
1101 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102
1103 if (!atomic_dec_and_test(&cfqq->ref))
1104 return;
1105
1106 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001107 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001108 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001110 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001111 __cfq_slice_expired(cfqd, cfqq, 0, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001112 cfq_schedule_dispatch(cfqd);
1113 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001114
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 /*
1116 * it's on the empty list and still hashed
1117 */
1118 list_del(&cfqq->cfq_list);
1119 hlist_del(&cfqq->cfq_hash);
1120 kmem_cache_free(cfq_pool, cfqq);
1121}
1122
Jens Axboe1ea25ec2006-07-18 22:24:11 +02001123static struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001124__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1125 const int hashval)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126{
1127 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
Jens Axboe206dc692006-03-28 13:03:44 +02001128 struct hlist_node *entry;
1129 struct cfq_queue *__cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
Jens Axboe206dc692006-03-28 13:03:44 +02001131 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
Al Virob0a69162006-03-14 15:32:50 -05001132 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133
Jens Axboe206dc692006-03-28 13:03:44 +02001134 if (__cfqq->key == key && (__p == prio || !prio))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 return __cfqq;
1136 }
1137
1138 return NULL;
1139}
1140
1141static struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001142cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143{
Jens Axboe3b181522005-06-27 10:56:24 +02001144 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145}
1146
Jens Axboee2d74ac2006-03-28 08:59:01 +02001147static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148{
Jens Axboe22e2c502005-06-27 10:55:12 +02001149 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001150 struct rb_node *n;
1151 int freed = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001152
Jens Axboee2d74ac2006-03-28 08:59:01 +02001153 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1154 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1155 rb_erase(&__cic->rb_node, &ioc->cic_root);
Jens Axboe22e2c502005-06-27 10:55:12 +02001156 kmem_cache_free(cfq_ioc_pool, __cic);
Al Viro334e94d2006-03-18 15:05:53 -05001157 freed++;
Jens Axboe22e2c502005-06-27 10:55:12 +02001158 }
1159
Jens Axboe4050cf12006-07-19 05:07:12 +02001160 elv_ioc_count_mod(ioc_count, -freed);
1161
1162 if (ioc_gone && !elv_ioc_count_read(ioc_count))
Al Viro334e94d2006-03-18 15:05:53 -05001163 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164}
1165
Jens Axboe89850f72006-07-22 16:48:31 +02001166static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1167{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001168 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001169 __cfq_slice_expired(cfqd, cfqq, 0, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001170 cfq_schedule_dispatch(cfqd);
1171 }
Jens Axboe89850f72006-07-22 16:48:31 +02001172
1173 cfq_put_queue(cfqq);
1174}
1175
1176static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1177 struct cfq_io_context *cic)
1178{
Jens Axboefc463792006-08-29 09:05:44 +02001179 list_del_init(&cic->queue_list);
1180 smp_wmb();
1181 cic->key = NULL;
1182
Jens Axboe89850f72006-07-22 16:48:31 +02001183 if (cic->cfqq[ASYNC]) {
1184 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1185 cic->cfqq[ASYNC] = NULL;
1186 }
1187
1188 if (cic->cfqq[SYNC]) {
1189 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1190 cic->cfqq[SYNC] = NULL;
1191 }
Jens Axboe89850f72006-07-22 16:48:31 +02001192}
1193
1194
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001196 * Called with interrupts disabled
1197 */
1198static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1199{
Al Viro478a82b2006-03-18 13:25:24 -05001200 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001201
Jens Axboe89850f72006-07-22 16:48:31 +02001202 if (cfqd) {
1203 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001204
Jens Axboefc463792006-08-29 09:05:44 +02001205 spin_lock_irq(q->queue_lock);
Jens Axboe89850f72006-07-22 16:48:31 +02001206 __cfq_exit_single_io_context(cfqd, cic);
Jens Axboefc463792006-08-29 09:05:44 +02001207 spin_unlock_irq(q->queue_lock);
Al Viro12a05732006-03-18 13:38:01 -05001208 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001209}
1210
Jens Axboee2d74ac2006-03-28 08:59:01 +02001211static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212{
Jens Axboe22e2c502005-06-27 10:55:12 +02001213 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001214 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +02001215
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 /*
1217 * put the reference this task is holding to the various queues
1218 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001219
1220 n = rb_first(&ioc->cic_root);
1221 while (n != NULL) {
1222 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1223
Jens Axboe22e2c502005-06-27 10:55:12 +02001224 cfq_exit_single_io_context(__cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001225 n = rb_next(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227}
1228
Jens Axboe22e2c502005-06-27 10:55:12 +02001229static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001230cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231{
Jens Axboeb5deef92006-07-19 23:39:40 +02001232 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233
Jens Axboeb5deef92006-07-19 23:39:40 +02001234 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235 if (cic) {
Jens Axboe553698f2006-06-14 19:11:57 +02001236 memset(cic, 0, sizeof(*cic));
Jens Axboe22e2c502005-06-27 10:55:12 +02001237 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001238 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001239 cic->dtor = cfq_free_io_context;
1240 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001241 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242 }
1243
1244 return cic;
1245}
1246
Jens Axboe22e2c502005-06-27 10:55:12 +02001247static void cfq_init_prio_data(struct cfq_queue *cfqq)
1248{
1249 struct task_struct *tsk = current;
1250 int ioprio_class;
1251
Jens Axboe3b181522005-06-27 10:56:24 +02001252 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001253 return;
1254
1255 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1256 switch (ioprio_class) {
1257 default:
1258 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1259 case IOPRIO_CLASS_NONE:
1260 /*
1261 * no prio set, place us in the middle of the BE classes
1262 */
1263 cfqq->ioprio = task_nice_ioprio(tsk);
1264 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1265 break;
1266 case IOPRIO_CLASS_RT:
1267 cfqq->ioprio = task_ioprio(tsk);
1268 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1269 break;
1270 case IOPRIO_CLASS_BE:
1271 cfqq->ioprio = task_ioprio(tsk);
1272 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1273 break;
1274 case IOPRIO_CLASS_IDLE:
1275 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1276 cfqq->ioprio = 7;
Jens Axboe3b181522005-06-27 10:56:24 +02001277 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001278 break;
1279 }
1280
1281 /*
1282 * keep track of original prio settings in case we have to temporarily
1283 * elevate the priority of this queue
1284 */
1285 cfqq->org_ioprio = cfqq->ioprio;
1286 cfqq->org_ioprio_class = cfqq->ioprio_class;
1287
Jens Axboe98e41c72007-02-05 11:55:35 +01001288 cfq_resort_rr_list(cfqq, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001289 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001290}
1291
Al Viro478a82b2006-03-18 13:25:24 -05001292static inline void changed_ioprio(struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001293{
Al Viro478a82b2006-03-18 13:25:24 -05001294 struct cfq_data *cfqd = cic->key;
1295 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001296 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001297
Jens Axboecaaa5f92006-06-16 11:23:00 +02001298 if (unlikely(!cfqd))
1299 return;
1300
Jens Axboec1b707d2006-10-30 19:54:23 +01001301 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001302
1303 cfqq = cic->cfqq[ASYNC];
1304 if (cfqq) {
1305 struct cfq_queue *new_cfqq;
1306 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1307 GFP_ATOMIC);
1308 if (new_cfqq) {
1309 cic->cfqq[ASYNC] = new_cfqq;
1310 cfq_put_queue(cfqq);
1311 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001312 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001313
1314 cfqq = cic->cfqq[SYNC];
1315 if (cfqq)
1316 cfq_mark_cfqq_prio_changed(cfqq);
1317
Jens Axboec1b707d2006-10-30 19:54:23 +01001318 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001319}
1320
Jens Axboefc463792006-08-29 09:05:44 +02001321static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322{
Al Viroa6a07632006-03-18 13:26:44 -05001323 struct cfq_io_context *cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001324 struct rb_node *n;
Al Viroa6a07632006-03-18 13:26:44 -05001325
Jens Axboefc463792006-08-29 09:05:44 +02001326 ioc->ioprio_changed = 0;
Al Viroa6a07632006-03-18 13:26:44 -05001327
Jens Axboee2d74ac2006-03-28 08:59:01 +02001328 n = rb_first(&ioc->cic_root);
1329 while (n != NULL) {
1330 cic = rb_entry(n, struct cfq_io_context, rb_node);
Jens Axboe3793c652006-05-30 21:11:04 +02001331
Al Viro478a82b2006-03-18 13:25:24 -05001332 changed_ioprio(cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001333 n = rb_next(n);
1334 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335}
1336
1337static struct cfq_queue *
Al Viro6f325a12006-03-18 14:58:37 -05001338cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
Al Viro8267e262005-10-21 03:20:53 -04001339 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340{
1341 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1342 struct cfq_queue *cfqq, *new_cfqq = NULL;
Al Viro6f325a12006-03-18 14:58:37 -05001343 unsigned short ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344
1345retry:
Al Viro6f325a12006-03-18 14:58:37 -05001346 ioprio = tsk->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02001347 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348
1349 if (!cfqq) {
1350 if (new_cfqq) {
1351 cfqq = new_cfqq;
1352 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001353 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001354 /*
1355 * Inform the allocator of the fact that we will
1356 * just repeat this allocation if it fails, to allow
1357 * the allocator to do whatever it needs to attempt to
1358 * free memory.
1359 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 spin_unlock_irq(cfqd->queue->queue_lock);
Jens Axboeb5deef92006-07-19 23:39:40 +02001361 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 spin_lock_irq(cfqd->queue->queue_lock);
1363 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001364 } else {
Jens Axboeb5deef92006-07-19 23:39:40 +02001365 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001366 if (!cfqq)
1367 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001368 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369
1370 memset(cfqq, 0, sizeof(*cfqq));
1371
1372 INIT_HLIST_NODE(&cfqq->cfq_hash);
1373 INIT_LIST_HEAD(&cfqq->cfq_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001374 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375
1376 cfqq->key = key;
1377 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1378 atomic_set(&cfqq->ref, 0);
1379 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001380
Jens Axboea9938002007-04-20 08:55:52 +02001381 if (key != CFQ_KEY_ASYNC)
1382 cfq_mark_cfqq_idle_window(cfqq);
1383
Jens Axboe3b181522005-06-27 10:56:24 +02001384 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b03742006-07-28 09:48:51 +02001385 cfq_mark_cfqq_queue_new(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001386 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 }
1388
1389 if (new_cfqq)
1390 kmem_cache_free(cfq_pool, new_cfqq);
1391
1392 atomic_inc(&cfqq->ref);
1393out:
1394 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1395 return cfqq;
1396}
1397
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001398static void
1399cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1400{
Jens Axboefc463792006-08-29 09:05:44 +02001401 WARN_ON(!list_empty(&cic->queue_list));
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001402 rb_erase(&cic->rb_node, &ioc->cic_root);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001403 kmem_cache_free(cfq_ioc_pool, cic);
Jens Axboe4050cf12006-07-19 05:07:12 +02001404 elv_ioc_count_dec(ioc_count);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001405}
1406
Jens Axboee2d74ac2006-03-28 08:59:01 +02001407static struct cfq_io_context *
1408cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1409{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001410 struct rb_node *n;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001411 struct cfq_io_context *cic;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001412 void *k, *key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001413
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001414restart:
1415 n = ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001416 while (n) {
1417 cic = rb_entry(n, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001418 /* ->key must be copied to avoid race with cfq_exit_queue() */
1419 k = cic->key;
1420 if (unlikely(!k)) {
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001421 cfq_drop_dead_cic(ioc, cic);
1422 goto restart;
1423 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001424
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001425 if (key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001426 n = n->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001427 else if (key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001428 n = n->rb_right;
1429 else
1430 return cic;
1431 }
1432
1433 return NULL;
1434}
1435
1436static inline void
1437cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1438 struct cfq_io_context *cic)
1439{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001440 struct rb_node **p;
1441 struct rb_node *parent;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001442 struct cfq_io_context *__cic;
Jens Axboe0261d682006-10-30 19:07:48 +01001443 unsigned long flags;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001444 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001445
Jens Axboee2d74ac2006-03-28 08:59:01 +02001446 cic->ioc = ioc;
1447 cic->key = cfqd;
1448
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001449restart:
1450 parent = NULL;
1451 p = &ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001452 while (*p) {
1453 parent = *p;
1454 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001455 /* ->key must be copied to avoid race with cfq_exit_queue() */
1456 k = __cic->key;
1457 if (unlikely(!k)) {
Oleg Nesterovbe33c3a2006-08-21 08:36:12 +02001458 cfq_drop_dead_cic(ioc, __cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001459 goto restart;
1460 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001461
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001462 if (cic->key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001463 p = &(*p)->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001464 else if (cic->key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001465 p = &(*p)->rb_right;
1466 else
1467 BUG();
1468 }
1469
1470 rb_link_node(&cic->rb_node, parent, p);
1471 rb_insert_color(&cic->rb_node, &ioc->cic_root);
Jens Axboefc463792006-08-29 09:05:44 +02001472
Jens Axboe0261d682006-10-30 19:07:48 +01001473 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001474 list_add(&cic->queue_list, &cfqd->cic_list);
Jens Axboe0261d682006-10-30 19:07:48 +01001475 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001476}
1477
Jens Axboe22e2c502005-06-27 10:55:12 +02001478/*
1479 * Setup general io context and cfq io context. There can be several cfq
1480 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001481 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001482 */
1483static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001484cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485{
Jens Axboe22e2c502005-06-27 10:55:12 +02001486 struct io_context *ioc = NULL;
1487 struct cfq_io_context *cic;
1488
1489 might_sleep_if(gfp_mask & __GFP_WAIT);
1490
Jens Axboeb5deef92006-07-19 23:39:40 +02001491 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001492 if (!ioc)
1493 return NULL;
1494
Jens Axboee2d74ac2006-03-28 08:59:01 +02001495 cic = cfq_cic_rb_lookup(cfqd, ioc);
1496 if (cic)
1497 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001498
Jens Axboee2d74ac2006-03-28 08:59:01 +02001499 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1500 if (cic == NULL)
1501 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001502
Jens Axboee2d74ac2006-03-28 08:59:01 +02001503 cfq_cic_link(cfqd, ioc, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001504out:
Jens Axboefc463792006-08-29 09:05:44 +02001505 smp_read_barrier_depends();
1506 if (unlikely(ioc->ioprio_changed))
1507 cfq_ioc_set_ioprio(ioc);
1508
Jens Axboe22e2c502005-06-27 10:55:12 +02001509 return cic;
1510err:
1511 put_io_context(ioc);
1512 return NULL;
1513}
1514
1515static void
1516cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1517{
Jens Axboeaaf12282007-01-19 11:30:16 +11001518 unsigned long elapsed = jiffies - cic->last_end_request;
1519 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001520
1521 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1522 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1523 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1524}
1525
Jens Axboe206dc692006-03-28 13:03:44 +02001526static void
Jens Axboebb37b942006-12-01 10:42:33 +01001527cfq_update_io_seektime(struct cfq_io_context *cic, struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001528{
1529 sector_t sdist;
1530 u64 total;
1531
Jens Axboe5e705372006-07-13 12:39:25 +02001532 if (cic->last_request_pos < rq->sector)
1533 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001534 else
Jens Axboe5e705372006-07-13 12:39:25 +02001535 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001536
1537 /*
1538 * Don't allow the seek distance to get too large from the
1539 * odd fragment, pagein, etc
1540 */
1541 if (cic->seek_samples <= 60) /* second&third seek */
1542 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1543 else
1544 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1545
1546 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1547 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1548 total = cic->seek_total + (cic->seek_samples/2);
1549 do_div(total, cic->seek_samples);
1550 cic->seek_mean = (sector_t)total;
1551}
Jens Axboe22e2c502005-06-27 10:55:12 +02001552
1553/*
1554 * Disable idle window if the process thinks too long or seeks so much that
1555 * it doesn't matter
1556 */
1557static void
1558cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1559 struct cfq_io_context *cic)
1560{
Jens Axboe3b181522005-06-27 10:56:24 +02001561 int enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001562
Jens Axboecaaa5f92006-06-16 11:23:00 +02001563 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1564 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001565 enable_idle = 0;
1566 else if (sample_valid(cic->ttime_samples)) {
1567 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1568 enable_idle = 0;
1569 else
1570 enable_idle = 1;
1571 }
1572
Jens Axboe3b181522005-06-27 10:56:24 +02001573 if (enable_idle)
1574 cfq_mark_cfqq_idle_window(cfqq);
1575 else
1576 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001577}
1578
Jens Axboe22e2c502005-06-27 10:55:12 +02001579/*
1580 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1581 * no or if we aren't sure, a 1 will cause a preempt.
1582 */
1583static int
1584cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001585 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001586{
1587 struct cfq_queue *cfqq = cfqd->active_queue;
1588
1589 if (cfq_class_idle(new_cfqq))
1590 return 0;
1591
1592 if (!cfqq)
Jens Axboecaaa5f92006-06-16 11:23:00 +02001593 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001594
1595 if (cfq_class_idle(cfqq))
1596 return 1;
Jens Axboe3b181522005-06-27 10:56:24 +02001597 if (!cfq_cfqq_wait_request(new_cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001598 return 0;
1599 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02001600 * if the new request is sync, but the currently running queue is
1601 * not, let the sync request have priority.
1602 */
Jens Axboe5e705372006-07-13 12:39:25 +02001603 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001604 return 1;
Jens Axboe374f84a2006-07-23 01:42:19 +02001605 /*
1606 * So both queues are sync. Let the new request get disk time if
1607 * it's a metadata request and the current queue is doing regular IO.
1608 */
1609 if (rq_is_meta(rq) && !cfqq->meta_pending)
1610 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001611
1612 return 0;
1613}
1614
1615/*
1616 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1617 * let it have half of its nominal slice.
1618 */
1619static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1620{
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001621 cfq_slice_expired(cfqd, 1, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001622
Jens Axboebf572252006-07-19 20:29:12 +02001623 /*
1624 * Put the new queue at the front of the of the current list,
1625 * so we know that it will be selected next.
1626 */
1627 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1628 list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1629
Jens Axboe44f7c162007-01-19 11:51:58 +11001630 cfqq->slice_end = 0;
1631 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001632}
1633
1634/*
Jens Axboe5e705372006-07-13 12:39:25 +02001635 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001636 * something we should do about it
1637 */
1638static void
Jens Axboe5e705372006-07-13 12:39:25 +02001639cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1640 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001641{
Jens Axboe5e705372006-07-13 12:39:25 +02001642 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001643
Jens Axboe374f84a2006-07-23 01:42:19 +02001644 if (rq_is_meta(rq))
1645 cfqq->meta_pending++;
1646
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001647 /*
1648 * we never wait for an async request and we don't allow preemption
1649 * of an async request. so just return early
1650 */
Jens Axboe5e705372006-07-13 12:39:25 +02001651 if (!rq_is_sync(rq)) {
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001652 /*
1653 * sync process issued an async request, if it's waiting
1654 * then expire it and kick rq handling.
1655 */
1656 if (cic == cfqd->active_cic &&
1657 del_timer(&cfqd->idle_slice_timer)) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001658 cfq_slice_expired(cfqd, 0, 0);
Jens Axboedc72ef42006-07-20 14:54:05 +02001659 blk_start_queueing(cfqd->queue);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001660 }
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001661 return;
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001662 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001663
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001664 cfq_update_io_thinktime(cfqd, cic);
Jens Axboebb37b942006-12-01 10:42:33 +01001665 cfq_update_io_seektime(cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001666 cfq_update_idle_window(cfqd, cfqq, cic);
1667
Jens Axboe5e705372006-07-13 12:39:25 +02001668 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001669
1670 if (cfqq == cfqd->active_queue) {
1671 /*
1672 * if we are waiting for a request for this queue, let it rip
1673 * immediately and flag that we must not expire this queue
1674 * just now
1675 */
Jens Axboe3b181522005-06-27 10:56:24 +02001676 if (cfq_cfqq_wait_request(cfqq)) {
1677 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001678 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001679 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001680 }
Jens Axboe5e705372006-07-13 12:39:25 +02001681 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001682 /*
1683 * not the active queue - expire current slice if it is
1684 * idle and has expired it's mean thinktime or this new queue
1685 * has some old slice time left and is of higher priority
1686 */
1687 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001688 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001689 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001690 }
1691}
1692
Jens Axboeb4878f22005-10-20 16:42:29 +02001693static void cfq_insert_request(request_queue_t *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001694{
Jens Axboeb4878f22005-10-20 16:42:29 +02001695 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001696 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001697
1698 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699
Jens Axboe5e705372006-07-13 12:39:25 +02001700 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701
Jens Axboe22e2c502005-06-27 10:55:12 +02001702 list_add_tail(&rq->queuelist, &cfqq->fifo);
1703
Jens Axboe5e705372006-07-13 12:39:25 +02001704 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705}
1706
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707static void cfq_completed_request(request_queue_t *q, struct request *rq)
1708{
Jens Axboe5e705372006-07-13 12:39:25 +02001709 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001710 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001711 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001712 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713
Jens Axboeb4878f22005-10-20 16:42:29 +02001714 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715
Jens Axboeb4878f22005-10-20 16:42:29 +02001716 WARN_ON(!cfqd->rq_in_driver);
1717 WARN_ON(!cfqq->on_dispatch[sync]);
1718 cfqd->rq_in_driver--;
1719 cfqq->on_dispatch[sync]--;
Jens Axboe99f96282007-02-05 11:56:25 +01001720 cfqq->service_last = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721
Jens Axboeb4878f22005-10-20 16:42:29 +02001722 if (!cfq_class_idle(cfqq))
1723 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001724
Jens Axboe98e41c72007-02-05 11:55:35 +01001725 cfq_resort_rr_list(cfqq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726
Jens Axboecaaa5f92006-06-16 11:23:00 +02001727 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001728 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001729
1730 /*
1731 * If this is the active queue, check if it needs to be expired,
1732 * or if we want to idle in case it has no pending requests.
1733 */
1734 if (cfqd->active_queue == cfqq) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001735 if (cfq_cfqq_slice_new(cfqq)) {
1736 cfq_set_prio_slice(cfqd, cfqq);
1737 cfq_clear_cfqq_slice_new(cfqq);
1738 }
1739 if (cfq_slice_used(cfqq))
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001740 cfq_slice_expired(cfqd, 0, 1);
Jens Axboedd67d052006-06-21 09:36:18 +02001741 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe17926692007-01-19 11:59:30 +11001742 if (!cfq_arm_slice_timer(cfqd))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001743 cfq_schedule_dispatch(cfqd);
1744 }
1745 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746}
1747
Jens Axboe22e2c502005-06-27 10:55:12 +02001748/*
1749 * we temporarily boost lower priority queues if they are holding fs exclusive
1750 * resources. they are boosted to normal prio (CLASS_BE/4)
1751 */
1752static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753{
Jens Axboe22e2c502005-06-27 10:55:12 +02001754 const int ioprio_class = cfqq->ioprio_class;
1755 const int ioprio = cfqq->ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756
Jens Axboe22e2c502005-06-27 10:55:12 +02001757 if (has_fs_excl()) {
1758 /*
1759 * boost idle prio on transactions that would lock out other
1760 * users of the filesystem
1761 */
1762 if (cfq_class_idle(cfqq))
1763 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1764 if (cfqq->ioprio > IOPRIO_NORM)
1765 cfqq->ioprio = IOPRIO_NORM;
1766 } else {
1767 /*
1768 * check if we need to unboost the queue
1769 */
1770 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1771 cfqq->ioprio_class = cfqq->org_ioprio_class;
1772 if (cfqq->ioprio != cfqq->org_ioprio)
1773 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 }
1775
Jens Axboe22e2c502005-06-27 10:55:12 +02001776 /*
1777 * refile between round-robin lists if we moved the priority class
1778 */
Jens Axboe98e41c72007-02-05 11:55:35 +01001779 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio))
Jens Axboe22e2c502005-06-27 10:55:12 +02001780 cfq_resort_rr_list(cfqq, 0);
1781}
1782
Jens Axboe89850f72006-07-22 16:48:31 +02001783static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001784{
Jens Axboe3b181522005-06-27 10:56:24 +02001785 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001786 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001787 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001788 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001789 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001790
1791 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02001792}
1793
Jens Axboecb78b282006-07-28 09:32:57 +02001794static int cfq_may_queue(request_queue_t *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02001795{
1796 struct cfq_data *cfqd = q->elevator->elevator_data;
1797 struct task_struct *tsk = current;
1798 struct cfq_queue *cfqq;
Jens Axboe7749a8d2006-12-13 13:02:26 +01001799 unsigned int key;
1800
1801 key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02001802
1803 /*
1804 * don't force setup of a queue from here, as a call to may_queue
1805 * does not necessarily imply that a request actually will be queued.
1806 * so just lookup a possibly existing queue, or return 'may queue'
1807 * if that fails
1808 */
Jens Axboe7749a8d2006-12-13 13:02:26 +01001809 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001810 if (cfqq) {
1811 cfq_init_prio_data(cfqq);
1812 cfq_prio_boost(cfqq);
1813
Jens Axboe89850f72006-07-22 16:48:31 +02001814 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001815 }
1816
1817 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818}
1819
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820/*
1821 * queue lock held here
1822 */
Jens Axboebb37b942006-12-01 10:42:33 +01001823static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824{
Jens Axboe5e705372006-07-13 12:39:25 +02001825 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826
Jens Axboe5e705372006-07-13 12:39:25 +02001827 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001828 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829
Jens Axboe22e2c502005-06-27 10:55:12 +02001830 BUG_ON(!cfqq->allocated[rw]);
1831 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832
Jens Axboe5e705372006-07-13 12:39:25 +02001833 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02001836 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 cfq_put_queue(cfqq);
1839 }
1840}
1841
1842/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001843 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001845static int
Jens Axboecb78b282006-07-28 09:32:57 +02001846cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847{
1848 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe3b181522005-06-27 10:56:24 +02001849 struct task_struct *tsk = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 struct cfq_io_context *cic;
1851 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01001852 const int is_sync = rq_is_sync(rq);
1853 pid_t key = cfq_queue_pid(tsk, rw, is_sync);
Jens Axboe22e2c502005-06-27 10:55:12 +02001854 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 unsigned long flags;
1856
1857 might_sleep_if(gfp_mask & __GFP_WAIT);
1858
Jens Axboee2d74ac2006-03-28 08:59:01 +02001859 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001860
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861 spin_lock_irqsave(q->queue_lock, flags);
1862
Jens Axboe22e2c502005-06-27 10:55:12 +02001863 if (!cic)
1864 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865
Al Viro12a05732006-03-18 13:38:01 -05001866 if (!cic->cfqq[is_sync]) {
Al Viro6f325a12006-03-18 14:58:37 -05001867 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001868 if (!cfqq)
1869 goto queue_fail;
1870
Al Viro12a05732006-03-18 13:38:01 -05001871 cic->cfqq[is_sync] = cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001872 } else
Al Viro12a05732006-03-18 13:38:01 -05001873 cfqq = cic->cfqq[is_sync];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874
1875 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02001876 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001877 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02001878
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879 spin_unlock_irqrestore(q->queue_lock, flags);
1880
Jens Axboe5e705372006-07-13 12:39:25 +02001881 rq->elevator_private = cic;
1882 rq->elevator_private2 = cfqq;
1883 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02001884
Jens Axboe22e2c502005-06-27 10:55:12 +02001885queue_fail:
1886 if (cic)
1887 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02001888
Jens Axboe3b181522005-06-27 10:56:24 +02001889 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 spin_unlock_irqrestore(q->queue_lock, flags);
1891 return 1;
1892}
1893
David Howells65f27f32006-11-22 14:55:48 +00001894static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02001895{
David Howells65f27f32006-11-22 14:55:48 +00001896 struct cfq_data *cfqd =
1897 container_of(work, struct cfq_data, unplug_work);
1898 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001899 unsigned long flags;
1900
1901 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02001902 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02001903 spin_unlock_irqrestore(q->queue_lock, flags);
1904}
1905
1906/*
1907 * Timer running if the active_queue is currently idling inside its time slice
1908 */
1909static void cfq_idle_slice_timer(unsigned long data)
1910{
1911 struct cfq_data *cfqd = (struct cfq_data *) data;
1912 struct cfq_queue *cfqq;
1913 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001914 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001915
1916 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1917
1918 if ((cfqq = cfqd->active_queue) != NULL) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001919 timed_out = 0;
1920
Jens Axboe22e2c502005-06-27 10:55:12 +02001921 /*
1922 * expired
1923 */
Jens Axboe44f7c162007-01-19 11:51:58 +11001924 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001925 goto expire;
1926
1927 /*
1928 * only expire and reinvoke request handler, if there are
1929 * other queues with pending requests
1930 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02001931 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02001932 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02001933
1934 /*
1935 * not expired and it has a request pending, let it dispatch
1936 */
Jens Axboedd67d052006-06-21 09:36:18 +02001937 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001938 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001939 goto out_kick;
1940 }
1941 }
1942expire:
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001943 cfq_slice_expired(cfqd, 0, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02001944out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02001945 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001946out_cont:
1947 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1948}
1949
1950/*
1951 * Timer running if an idle class queue is waiting for service
1952 */
1953static void cfq_idle_class_timer(unsigned long data)
1954{
1955 struct cfq_data *cfqd = (struct cfq_data *) data;
1956 unsigned long flags, end;
1957
1958 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1959
1960 /*
1961 * race with a non-idle queue, reset timer
1962 */
1963 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
Jens Axboeae818a32006-06-01 10:13:43 +02001964 if (!time_after_eq(jiffies, end))
1965 mod_timer(&cfqd->idle_class_timer, end);
1966 else
Jens Axboe3b181522005-06-27 10:56:24 +02001967 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001968
1969 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1970}
1971
Jens Axboe3b181522005-06-27 10:56:24 +02001972static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1973{
1974 del_timer_sync(&cfqd->idle_slice_timer);
1975 del_timer_sync(&cfqd->idle_class_timer);
1976 blk_sync_queue(cfqd->queue);
1977}
Jens Axboe22e2c502005-06-27 10:55:12 +02001978
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979static void cfq_exit_queue(elevator_t *e)
1980{
Jens Axboe22e2c502005-06-27 10:55:12 +02001981 struct cfq_data *cfqd = e->elevator_data;
Al Virod9ff4182006-03-18 13:51:22 -05001982 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001983
Jens Axboe3b181522005-06-27 10:56:24 +02001984 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001985
Al Virod9ff4182006-03-18 13:51:22 -05001986 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001987
Al Virod9ff4182006-03-18 13:51:22 -05001988 if (cfqd->active_queue)
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001989 __cfq_slice_expired(cfqd, cfqd->active_queue, 0, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001990
1991 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05001992 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1993 struct cfq_io_context,
1994 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02001995
1996 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05001997 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001998
Al Virod9ff4182006-03-18 13:51:22 -05001999 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002000
2001 cfq_shutdown_timer_wq(cfqd);
2002
Al Viroa90d7422006-03-18 12:05:37 -05002003 kfree(cfqd->cfq_hash);
2004 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005}
2006
Jens Axboebb37b942006-12-01 10:42:33 +01002007static void *cfq_init_queue(request_queue_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008{
2009 struct cfq_data *cfqd;
2010 int i;
2011
Jens Axboeb5deef92006-07-19 23:39:40 +02002012 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002014 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015
2016 memset(cfqd, 0, sizeof(*cfqd));
Jens Axboe22e2c502005-06-27 10:55:12 +02002017
2018 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2019 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2020
2021 INIT_LIST_HEAD(&cfqd->busy_rr);
2022 INIT_LIST_HEAD(&cfqd->cur_rr);
2023 INIT_LIST_HEAD(&cfqd->idle_rr);
Al Virod9ff4182006-03-18 13:51:22 -05002024 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025
Jens Axboeb5deef92006-07-19 23:39:40 +02002026 cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 if (!cfqd->cfq_hash)
Jens Axboe5e705372006-07-13 12:39:25 +02002028 goto out_free;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2031 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2032
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034
Jens Axboe22e2c502005-06-27 10:55:12 +02002035 init_timer(&cfqd->idle_slice_timer);
2036 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2037 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2038
2039 init_timer(&cfqd->idle_class_timer);
2040 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2041 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2042
David Howells65f27f32006-11-22 14:55:48 +00002043 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002044
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002046 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2047 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048 cfqd->cfq_back_max = cfq_back_max;
2049 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002050 cfqd->cfq_slice[0] = cfq_slice_async;
2051 cfqd->cfq_slice[1] = cfq_slice_sync;
2052 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2053 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02002054
Jens Axboebc1c1162006-06-08 08:49:06 +02002055 return cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +02002056out_free:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057 kfree(cfqd);
Jens Axboebc1c1162006-06-08 08:49:06 +02002058 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059}
2060
2061static void cfq_slab_kill(void)
2062{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063 if (cfq_pool)
2064 kmem_cache_destroy(cfq_pool);
2065 if (cfq_ioc_pool)
2066 kmem_cache_destroy(cfq_ioc_pool);
2067}
2068
2069static int __init cfq_slab_setup(void)
2070{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2072 NULL, NULL);
2073 if (!cfq_pool)
2074 goto fail;
2075
2076 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2077 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2078 if (!cfq_ioc_pool)
2079 goto fail;
2080
2081 return 0;
2082fail:
2083 cfq_slab_kill();
2084 return -ENOMEM;
2085}
2086
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087/*
2088 * sysfs parts below -->
2089 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090
2091static ssize_t
2092cfq_var_show(unsigned int var, char *page)
2093{
2094 return sprintf(page, "%d\n", var);
2095}
2096
2097static ssize_t
2098cfq_var_store(unsigned int *var, const char *page, size_t count)
2099{
2100 char *p = (char *) page;
2101
2102 *var = simple_strtoul(p, &p, 10);
2103 return count;
2104}
2105
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002107static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002109 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110 unsigned int __data = __VAR; \
2111 if (__CONV) \
2112 __data = jiffies_to_msecs(__data); \
2113 return cfq_var_show(__data, (page)); \
2114}
2115SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002116SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2117SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002118SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2119SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002120SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2121SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2122SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2123SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124#undef SHOW_FUNCTION
2125
2126#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002127static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002129 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130 unsigned int __data; \
2131 int ret = cfq_var_store(&__data, (page), count); \
2132 if (__data < (MIN)) \
2133 __data = (MIN); \
2134 else if (__data > (MAX)) \
2135 __data = (MAX); \
2136 if (__CONV) \
2137 *(__PTR) = msecs_to_jiffies(__data); \
2138 else \
2139 *(__PTR) = __data; \
2140 return ret; \
2141}
2142STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002143STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2144STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002145STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2146STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002147STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2148STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2149STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2150STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151#undef STORE_FUNCTION
2152
Al Viroe572ec72006-03-18 22:27:18 -05002153#define CFQ_ATTR(name) \
2154 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002155
Al Viroe572ec72006-03-18 22:27:18 -05002156static struct elv_fs_entry cfq_attrs[] = {
2157 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002158 CFQ_ATTR(fifo_expire_sync),
2159 CFQ_ATTR(fifo_expire_async),
2160 CFQ_ATTR(back_seek_max),
2161 CFQ_ATTR(back_seek_penalty),
2162 CFQ_ATTR(slice_sync),
2163 CFQ_ATTR(slice_async),
2164 CFQ_ATTR(slice_async_rq),
2165 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002166 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167};
2168
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169static struct elevator_type iosched_cfq = {
2170 .ops = {
2171 .elevator_merge_fn = cfq_merge,
2172 .elevator_merged_fn = cfq_merged_request,
2173 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002174 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002175 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002177 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 .elevator_deactivate_req_fn = cfq_deactivate_request,
2179 .elevator_queue_empty_fn = cfq_queue_empty,
2180 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002181 .elevator_former_req_fn = elv_rb_former_request,
2182 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 .elevator_set_req_fn = cfq_set_request,
2184 .elevator_put_req_fn = cfq_put_request,
2185 .elevator_may_queue_fn = cfq_may_queue,
2186 .elevator_init_fn = cfq_init_queue,
2187 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002188 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189 },
Al Viro3d1ab402006-03-18 18:35:43 -05002190 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 .elevator_name = "cfq",
2192 .elevator_owner = THIS_MODULE,
2193};
2194
2195static int __init cfq_init(void)
2196{
2197 int ret;
2198
Jens Axboe22e2c502005-06-27 10:55:12 +02002199 /*
2200 * could be 0 on HZ < 1000 setups
2201 */
2202 if (!cfq_slice_async)
2203 cfq_slice_async = 1;
2204 if (!cfq_slice_idle)
2205 cfq_slice_idle = 1;
2206
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 if (cfq_slab_setup())
2208 return -ENOMEM;
2209
2210 ret = elv_register(&iosched_cfq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002211 if (ret)
2212 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214 return ret;
2215}
2216
2217static void __exit cfq_exit(void)
2218{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002219 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002221 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002222 /* ioc_gone's update must be visible before reading ioc_count */
2223 smp_wmb();
Jens Axboe4050cf12006-07-19 05:07:12 +02002224 if (elv_ioc_count_read(ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002225 wait_for_completion(ioc_gone);
Al Viro334e94d2006-03-18 15:05:53 -05002226 synchronize_rcu();
Christoph Hellwig83521d32005-10-30 15:01:39 -08002227 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228}
2229
2230module_init(cfq_init);
2231module_exit(cfq_exit);
2232
2233MODULE_AUTHOR("Jens Axboe");
2234MODULE_LICENSE("GPL");
2235MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");