blob: 07b706243772e46fe2434248bf0e4fba938fca0a [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_start;
150 unsigned long slice_end;
151 unsigned long slice_left;
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 {
165 CFQ_CFQQ_FLAG_on_rr = 0,
166 CFQ_CFQQ_FLAG_wait_request,
167 CFQ_CFQQ_FLAG_must_alloc,
168 CFQ_CFQQ_FLAG_must_alloc_slice,
169 CFQ_CFQQ_FLAG_must_dispatch,
170 CFQ_CFQQ_FLAG_fifo_expire,
171 CFQ_CFQQ_FLAG_idle_window,
172 CFQ_CFQQ_FLAG_prio_changed,
Jens Axboe53b03742006-07-28 09:48:51 +0200173 CFQ_CFQQ_FLAG_queue_new,
Jens Axboe3b181522005-06-27 10:56:24 +0200174};
175
176#define CFQ_CFQQ_FNS(name) \
177static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
178{ \
179 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
180} \
181static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
182{ \
183 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
184} \
185static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
186{ \
187 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
188}
189
190CFQ_CFQQ_FNS(on_rr);
191CFQ_CFQQ_FNS(wait_request);
192CFQ_CFQQ_FNS(must_alloc);
193CFQ_CFQQ_FNS(must_alloc_slice);
194CFQ_CFQQ_FNS(must_dispatch);
195CFQ_CFQQ_FNS(fifo_expire);
196CFQ_CFQQ_FNS(idle_window);
197CFQ_CFQQ_FNS(prio_changed);
Jens Axboe53b03742006-07-28 09:48:51 +0200198CFQ_CFQQ_FNS(queue_new);
Jens Axboe3b181522005-06-27 10:56:24 +0200199#undef CFQ_CFQQ_FNS
200
Jens Axboe3b181522005-06-27 10:56:24 +0200201static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
Jens Axboe5e705372006-07-13 12:39:25 +0200202static void cfq_dispatch_insert(request_queue_t *, struct request *);
Al Viro6f325a12006-03-18 14:58:37 -0500203static 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 -0700204
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700206 * scheduler run of queue, if there are requests pending and no one in the
207 * driver that will restart queueing
208 */
209static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
210{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100211 if (cfqd->busy_queues)
Andrew Morton99f95e52005-06-27 20:14:05 -0700212 kblockd_schedule_work(&cfqd->unplug_work);
213}
214
215static int cfq_queue_empty(request_queue_t *q)
216{
217 struct cfq_data *cfqd = q->elevator->elevator_data;
218
Jens Axboeb4878f22005-10-20 16:42:29 +0200219 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700220}
221
Jens Axboe7749a8d2006-12-13 13:02:26 +0100222static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync)
Jens Axboe206dc692006-03-28 13:03:44 +0200223{
Jens Axboe7749a8d2006-12-13 13:02:26 +0100224 /*
225 * Use the per-process queue, for read requests and syncronous writes
226 */
227 if (!(rw & REQ_RW) || is_sync)
Jens Axboe206dc692006-03-28 13:03:44 +0200228 return task->pid;
229
230 return CFQ_KEY_ASYNC;
231}
232
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233/*
Jens Axboe5e705372006-07-13 12:39:25 +0200234 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200236 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 */
Jens Axboe5e705372006-07-13 12:39:25 +0200238static struct request *
239cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240{
241 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200243#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
244#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
245 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246
Jens Axboe5e705372006-07-13 12:39:25 +0200247 if (rq1 == NULL || rq1 == rq2)
248 return rq2;
249 if (rq2 == NULL)
250 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200251
Jens Axboe5e705372006-07-13 12:39:25 +0200252 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
253 return rq1;
254 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
255 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200256 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
257 return rq1;
258 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
259 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260
Jens Axboe5e705372006-07-13 12:39:25 +0200261 s1 = rq1->sector;
262 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263
264 last = cfqd->last_sector;
265
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266 /*
267 * by definition, 1KiB is 2 sectors
268 */
269 back_max = cfqd->cfq_back_max * 2;
270
271 /*
272 * Strict one way elevator _except_ in the case where we allow
273 * short backward seeks which are biased as twice the cost of a
274 * similar forward seek.
275 */
276 if (s1 >= last)
277 d1 = s1 - last;
278 else if (s1 + back_max >= last)
279 d1 = (last - s1) * cfqd->cfq_back_penalty;
280 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200281 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282
283 if (s2 >= last)
284 d2 = s2 - last;
285 else if (s2 + back_max >= last)
286 d2 = (last - s2) * cfqd->cfq_back_penalty;
287 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200288 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289
290 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291
Andreas Mohre8a99052006-03-28 08:59:49 +0200292 /*
293 * By doing switch() on the bit mask "wrap" we avoid having to
294 * check two variables for all permutations: --> faster!
295 */
296 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200297 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200298 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200299 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200300 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200301 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200302 else {
303 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200304 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200305 else
Jens Axboe5e705372006-07-13 12:39:25 +0200306 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200307 }
308
309 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200310 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200311 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200312 return rq2;
313 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200314 default:
315 /*
316 * Since both rqs are wrapped,
317 * start with the one that's further behind head
318 * (--> only *one* back seek required),
319 * since back seek takes more time than forward.
320 */
321 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200322 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323 else
Jens Axboe5e705372006-07-13 12:39:25 +0200324 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 }
326}
327
328/*
329 * would be nice to take fifo expire time into account as well
330 */
Jens Axboe5e705372006-07-13 12:39:25 +0200331static struct request *
332cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
333 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334{
Jens Axboe21183b02006-07-13 12:33:14 +0200335 struct rb_node *rbnext = rb_next(&last->rb_node);
336 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200337 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338
Jens Axboe21183b02006-07-13 12:33:14 +0200339 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340
341 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200342 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200343
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200345 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200346 else {
347 rbnext = rb_first(&cfqq->sort_list);
348 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200349 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200350 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351
Jens Axboe21183b02006-07-13 12:33:14 +0200352 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353}
354
Jens Axboe22e2c502005-06-27 10:55:12 +0200355static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356{
Jens Axboe22e2c502005-06-27 10:55:12 +0200357 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe53b03742006-07-28 09:48:51 +0200358 struct list_head *list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359
Jens Axboe3b181522005-06-27 10:56:24 +0200360 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361
362 list_del(&cfqq->cfq_list);
363
Jens Axboe22e2c502005-06-27 10:55:12 +0200364 if (cfq_class_rt(cfqq))
365 list = &cfqd->cur_rr;
366 else if (cfq_class_idle(cfqq))
367 list = &cfqd->idle_rr;
368 else {
369 /*
370 * if cfqq has requests in flight, don't allow it to be
371 * found in cfq_set_active_queue before it has finished them.
372 * this is done to increase fairness between a process that
373 * has lots of io pending vs one that only generates one
374 * sporadically or synchronously
375 */
Jens Axboe3b181522005-06-27 10:56:24 +0200376 if (cfq_cfqq_dispatched(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200377 list = &cfqd->busy_rr;
378 else
379 list = &cfqd->rr_list[cfqq->ioprio];
380 }
381
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 /*
Jens Axboe53b03742006-07-28 09:48:51 +0200383 * If this queue was preempted or is new (never been serviced), let
384 * it be added first for fairness but beind other new queues.
385 * Otherwise, just add to the back of the list.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386 */
Jens Axboe53b03742006-07-28 09:48:51 +0200387 if (preempted || cfq_cfqq_queue_new(cfqq)) {
388 struct list_head *n = list;
389 struct cfq_queue *__cfqq;
Jens Axboeb52a8342006-06-01 18:53:43 +0200390
Jens Axboe53b03742006-07-28 09:48:51 +0200391 while (n->next != list) {
392 __cfqq = list_entry_cfqq(n->next);
393 if (!cfq_cfqq_queue_new(__cfqq))
394 break;
395
396 n = n->next;
397 }
398
399 list = n;
Jens Axboe22e2c502005-06-27 10:55:12 +0200400 }
401
Jens Axboe53b03742006-07-28 09:48:51 +0200402 list_add_tail(&cfqq->cfq_list, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403}
404
405/*
406 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200407 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408 */
409static inline void
Jens Axboeb4878f22005-10-20 16:42:29 +0200410cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411{
Jens Axboe3b181522005-06-27 10:56:24 +0200412 BUG_ON(cfq_cfqq_on_rr(cfqq));
413 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 cfqd->busy_queues++;
415
Jens Axboeb4878f22005-10-20 16:42:29 +0200416 cfq_resort_rr_list(cfqq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417}
418
419static inline void
420cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
421{
Jens Axboe3b181522005-06-27 10:56:24 +0200422 BUG_ON(!cfq_cfqq_on_rr(cfqq));
423 cfq_clear_cfqq_on_rr(cfqq);
Jens Axboe981a7972006-07-19 14:56:28 +0200424 list_del_init(&cfqq->cfq_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425
426 BUG_ON(!cfqd->busy_queues);
427 cfqd->busy_queues--;
428}
429
430/*
431 * rb tree support functions
432 */
Jens Axboe5e705372006-07-13 12:39:25 +0200433static inline void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434{
Jens Axboe5e705372006-07-13 12:39:25 +0200435 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200436 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200437 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438
Jens Axboeb4878f22005-10-20 16:42:29 +0200439 BUG_ON(!cfqq->queued[sync]);
440 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441
Jens Axboe5e705372006-07-13 12:39:25 +0200442 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
Jens Axboedd67d052006-06-21 09:36:18 +0200444 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200445 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446}
447
Jens Axboe5e705372006-07-13 12:39:25 +0200448static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449{
Jens Axboe5e705372006-07-13 12:39:25 +0200450 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe21183b02006-07-13 12:33:14 +0200452 struct request *__alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453
Jens Axboe5380a102006-07-13 12:37:56 +0200454 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
456 /*
457 * looks a little odd, but the first insert might return an alias.
458 * if that happens, put the alias on the dispatch list
459 */
Jens Axboe21183b02006-07-13 12:33:14 +0200460 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200461 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100462
463 if (!cfq_cfqq_on_rr(cfqq))
464 cfq_add_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465}
466
467static inline void
Jens Axboe5e705372006-07-13 12:39:25 +0200468cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469{
Jens Axboe5380a102006-07-13 12:37:56 +0200470 elv_rb_del(&cfqq->sort_list, rq);
471 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200472 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473}
474
Jens Axboe206dc692006-03-28 13:03:44 +0200475static struct request *
476cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477{
Jens Axboe206dc692006-03-28 13:03:44 +0200478 struct task_struct *tsk = current;
Jens Axboe7749a8d2006-12-13 13:02:26 +0100479 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
Jens Axboe206dc692006-03-28 13:03:44 +0200480 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481
Jens Axboe206dc692006-03-28 13:03:44 +0200482 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
Jens Axboe89850f72006-07-22 16:48:31 +0200483 if (cfqq) {
484 sector_t sector = bio->bi_sector + bio_sectors(bio);
485
Jens Axboe21183b02006-07-13 12:33:14 +0200486 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200487 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 return NULL;
490}
491
Jens Axboeb4878f22005-10-20 16:42:29 +0200492static void cfq_activate_request(request_queue_t *q, struct request *rq)
493{
494 struct cfq_data *cfqd = q->elevator->elevator_data;
495
496 cfqd->rq_in_driver++;
Jens Axboe25776e32006-06-01 10:12:26 +0200497
498 /*
499 * If the depth is larger 1, it really could be queueing. But lets
500 * make the mark a little higher - idling could still be good for
501 * low queueing, and a low queueing number could also just indicate
502 * a SCSI mid layer like behaviour where limit+1 is often seen.
503 */
504 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
505 cfqd->hw_tag = 1;
Jens Axboeb4878f22005-10-20 16:42:29 +0200506}
507
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
509{
Jens Axboe22e2c502005-06-27 10:55:12 +0200510 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511
Jens Axboeb4878f22005-10-20 16:42:29 +0200512 WARN_ON(!cfqd->rq_in_driver);
513 cfqd->rq_in_driver--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514}
515
Jens Axboeb4878f22005-10-20 16:42:29 +0200516static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517{
Jens Axboe5e705372006-07-13 12:39:25 +0200518 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200519
Jens Axboe5e705372006-07-13 12:39:25 +0200520 if (cfqq->next_rq == rq)
521 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522
Jens Axboeb4878f22005-10-20 16:42:29 +0200523 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200524 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200525
526 if (rq_is_meta(rq)) {
527 WARN_ON(!cfqq->meta_pending);
528 cfqq->meta_pending--;
529 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530}
531
532static int
533cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
534{
535 struct cfq_data *cfqd = q->elevator->elevator_data;
536 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537
Jens Axboe206dc692006-03-28 13:03:44 +0200538 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200539 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200540 *req = __rq;
541 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 }
543
544 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545}
546
Jens Axboe21183b02006-07-13 12:33:14 +0200547static void cfq_merged_request(request_queue_t *q, struct request *req,
548 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549{
Jens Axboe21183b02006-07-13 12:33:14 +0200550 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200551 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552
Jens Axboe5e705372006-07-13 12:39:25 +0200553 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555}
556
557static void
558cfq_merged_requests(request_queue_t *q, struct request *rq,
559 struct request *next)
560{
Jens Axboe22e2c502005-06-27 10:55:12 +0200561 /*
562 * reposition in fifo if next is older than rq
563 */
564 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
565 time_before(next->start_time, rq->start_time))
566 list_move(&rq->queuelist, &next->queuelist);
567
Jens Axboeb4878f22005-10-20 16:42:29 +0200568 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200569}
570
Jens Axboeda775262006-12-20 11:04:12 +0100571static int cfq_allow_merge(request_queue_t *q, struct request *rq,
572 struct bio *bio)
573{
574 struct cfq_data *cfqd = q->elevator->elevator_data;
575 const int rw = bio_data_dir(bio);
576 struct cfq_queue *cfqq;
577 pid_t key;
578
579 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100580 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100581 */
Jens Axboeec8acb62007-01-02 18:32:11 +0100582 if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100583 return 0;
584
585 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100586 * Lookup the cfqq that this bio will be queued with. Allow
587 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100588 */
Jens Axboe719d3402006-12-22 09:38:53 +0100589 key = cfq_queue_pid(current, rw, bio_sync(bio));
Jens Axboeda775262006-12-20 11:04:12 +0100590 cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
Jens Axboe719d3402006-12-22 09:38:53 +0100591
592 if (cfqq == RQ_CFQQ(rq))
593 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100594
Jens Axboeec8acb62007-01-02 18:32:11 +0100595 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100596}
597
Jens Axboe22e2c502005-06-27 10:55:12 +0200598static inline void
599__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
600{
601 if (cfqq) {
602 /*
603 * stop potential idle class queues waiting service
604 */
605 del_timer(&cfqd->idle_class_timer);
606
607 cfqq->slice_start = jiffies;
608 cfqq->slice_end = 0;
609 cfqq->slice_left = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200610 cfq_clear_cfqq_must_alloc_slice(cfqq);
611 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200612 }
613
614 cfqd->active_queue = cfqq;
615}
616
617/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100618 * current cfqq expired its slice (or was too idle), select new one
619 */
620static void
621__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
622 int preempted)
623{
624 unsigned long now = jiffies;
625
626 if (cfq_cfqq_wait_request(cfqq))
627 del_timer(&cfqd->idle_slice_timer);
628
Jens Axboe53b03742006-07-28 09:48:51 +0200629 if (!preempted && !cfq_cfqq_dispatched(cfqq))
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100630 cfq_schedule_dispatch(cfqd);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100631
632 cfq_clear_cfqq_must_dispatch(cfqq);
633 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe53b03742006-07-28 09:48:51 +0200634 cfq_clear_cfqq_queue_new(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100635
636 /*
637 * store what was left of this slice, if the queue idled out
638 * or was preempted
639 */
640 if (time_after(cfqq->slice_end, now))
641 cfqq->slice_left = cfqq->slice_end - now;
642 else
643 cfqq->slice_left = 0;
644
645 if (cfq_cfqq_on_rr(cfqq))
646 cfq_resort_rr_list(cfqq, preempted);
647
648 if (cfqq == cfqd->active_queue)
649 cfqd->active_queue = NULL;
650
651 if (cfqd->active_cic) {
652 put_io_context(cfqd->active_cic->ioc);
653 cfqd->active_cic = NULL;
654 }
655
656 cfqd->dispatch_slice = 0;
657}
658
659static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
660{
661 struct cfq_queue *cfqq = cfqd->active_queue;
662
663 if (cfqq)
664 __cfq_slice_expired(cfqd, cfqq, preempted);
665}
666
667/*
Jens Axboe22e2c502005-06-27 10:55:12 +0200668 * 0
669 * 0,1
670 * 0,1,2
671 * 0,1,2,3
672 * 0,1,2,3,4
673 * 0,1,2,3,4,5
674 * 0,1,2,3,4,5,6
675 * 0,1,2,3,4,5,6,7
676 */
677static int cfq_get_next_prio_level(struct cfq_data *cfqd)
678{
679 int prio, wrap;
680
681 prio = -1;
682 wrap = 0;
683 do {
684 int p;
685
686 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
687 if (!list_empty(&cfqd->rr_list[p])) {
688 prio = p;
689 break;
690 }
691 }
692
693 if (prio != -1)
694 break;
695 cfqd->cur_prio = 0;
696 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
697 cfqd->cur_end_prio = 0;
698 if (wrap)
699 break;
700 wrap = 1;
701 }
702 } while (1);
703
704 if (unlikely(prio == -1))
705 return -1;
706
707 BUG_ON(prio >= CFQ_PRIO_LISTS);
708
709 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
710
711 cfqd->cur_prio = prio + 1;
712 if (cfqd->cur_prio > cfqd->cur_end_prio) {
713 cfqd->cur_end_prio = cfqd->cur_prio;
714 cfqd->cur_prio = 0;
715 }
716 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
717 cfqd->cur_prio = 0;
718 cfqd->cur_end_prio = 0;
719 }
720
721 return prio;
722}
723
Jens Axboe3b181522005-06-27 10:56:24 +0200724static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200725{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100726 struct cfq_queue *cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +0200727
Jens Axboe89850f72006-07-22 16:48:31 +0200728 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
729 /*
730 * if current list is non-empty, grab first entry. if it is
731 * empty, get next prio level and grab first entry then if any
732 * are spliced
733 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200734 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
Jens Axboe89850f72006-07-22 16:48:31 +0200735 } else if (!list_empty(&cfqd->busy_rr)) {
736 /*
737 * If no new queues are available, check if the busy list has
738 * some before falling back to idle io.
739 */
Jens Axboee0de0202006-06-01 10:07:26 +0200740 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
Jens Axboe89850f72006-07-22 16:48:31 +0200741 } else if (!list_empty(&cfqd->idle_rr)) {
742 /*
743 * if we have idle queues and no rt or be queues had pending
744 * requests, either allow immediate service if the grace period
745 * has passed or arm the idle grace timer
746 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200747 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
748
749 if (time_after_eq(jiffies, end))
750 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
751 else
752 mod_timer(&cfqd->idle_class_timer, end);
753 }
754
755 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200756 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200757}
758
Jens Axboecaaa5f92006-06-16 11:23:00 +0200759#define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
760
Jens Axboe22e2c502005-06-27 10:55:12 +0200761static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
762
763{
Jens Axboe206dc692006-03-28 13:03:44 +0200764 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100765 unsigned long sl;
766
Jens Axboedd67d052006-06-21 09:36:18 +0200767 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +0200768 WARN_ON(cfqq != cfqd->active_queue);
769
770 /*
771 * idle is disabled, either manually or by past process history
772 */
773 if (!cfqd->cfq_slice_idle)
774 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200775 if (!cfq_cfqq_idle_window(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200776 return 0;
777 /*
778 * task has exited, don't wait
779 */
Jens Axboe206dc692006-03-28 13:03:44 +0200780 cic = cfqd->active_cic;
781 if (!cic || !cic->ioc->task)
Jens Axboe22e2c502005-06-27 10:55:12 +0200782 return 0;
783
Jens Axboe3b181522005-06-27 10:56:24 +0200784 cfq_mark_cfqq_must_dispatch(cfqq);
785 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200786
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100787 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
Jens Axboe206dc692006-03-28 13:03:44 +0200788
789 /*
790 * we don't want to idle for seeks, but we do want to allow
791 * fair distribution of slice time for a process doing back-to-back
792 * seeks. so allow a little bit of time for him to submit a new rq
793 */
Jens Axboecaaa5f92006-06-16 11:23:00 +0200794 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboe44eb1232006-07-25 15:05:21 +0200795 sl = min(sl, msecs_to_jiffies(2));
Jens Axboe206dc692006-03-28 13:03:44 +0200796
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100797 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe22e2c502005-06-27 10:55:12 +0200798 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799}
800
Jens Axboe5e705372006-07-13 12:39:25 +0200801static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802{
803 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200804 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200805
Jens Axboe5380a102006-07-13 12:37:56 +0200806 cfq_remove_request(rq);
807 cfqq->on_dispatch[rq_is_sync(rq)]++;
808 elv_dispatch_sort(q, rq);
Jens Axboefd61af02006-06-16 15:35:39 +0200809
810 rq = list_entry(q->queue_head.prev, struct request, queuelist);
811 cfqd->last_sector = rq->sector + rq->nr_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812}
813
814/*
815 * return expired entry, or NULL to just start from scratch in rbtree
816 */
Jens Axboe5e705372006-07-13 12:39:25 +0200817static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818{
819 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200820 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200821 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822
Jens Axboe3b181522005-06-27 10:56:24 +0200823 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 return NULL;
Jens Axboe89850f72006-07-22 16:48:31 +0200825 if (list_empty(&cfqq->fifo))
826 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827
Jens Axboe89850f72006-07-22 16:48:31 +0200828 fifo = cfq_cfqq_class_sync(cfqq);
829 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830
Jens Axboe89850f72006-07-22 16:48:31 +0200831 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
832 cfq_mark_cfqq_fifo_expire(cfqq);
833 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 }
835
836 return NULL;
837}
838
839/*
Jens Axboe3b181522005-06-27 10:56:24 +0200840 * Scale schedule slice based on io priority. Use the sync time slice only
841 * if a queue is marked sync and has sync io queued. A sync queue with async
842 * io only, should not get full sync slice length.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200844static inline int
845cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846{
Jens Axboe22e2c502005-06-27 10:55:12 +0200847 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848
Jens Axboe22e2c502005-06-27 10:55:12 +0200849 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850
Jens Axboe22e2c502005-06-27 10:55:12 +0200851 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852}
853
Jens Axboe22e2c502005-06-27 10:55:12 +0200854static inline void
855cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
856{
857 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
858}
859
860static inline int
861cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
862{
863 const int base_rq = cfqd->cfq_slice_async_rq;
864
865 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
866
867 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
868}
869
870/*
871 * get next queue for service
872 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100873static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200874{
875 unsigned long now = jiffies;
876 struct cfq_queue *cfqq;
877
878 cfqq = cfqd->active_queue;
879 if (!cfqq)
880 goto new_queue;
881
882 /*
883 * slice has expired
884 */
Jens Axboe3b181522005-06-27 10:56:24 +0200885 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
886 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +0200887
888 /*
889 * if queue has requests, dispatch one. if not, check if
890 * enough slice is left to wait for one
891 */
Jens Axboedd67d052006-06-21 09:36:18 +0200892 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200893 goto keep_queue;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200894 else if (cfq_cfqq_dispatched(cfqq)) {
895 cfqq = NULL;
896 goto keep_queue;
897 } else if (cfq_cfqq_class_sync(cfqq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +0200898 if (cfq_arm_slice_timer(cfqd, cfqq))
899 return NULL;
900 }
901
Jens Axboe3b181522005-06-27 10:56:24 +0200902expire:
Jens Axboe22e2c502005-06-27 10:55:12 +0200903 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +0200904new_queue:
905 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200906keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +0200907 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200908}
909
910static int
911__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
912 int max_dispatch)
913{
914 int dispatched = 0;
915
Jens Axboedd67d052006-06-21 09:36:18 +0200916 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +0200917
918 do {
Jens Axboe5e705372006-07-13 12:39:25 +0200919 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200920
921 /*
922 * follow expired path, else get first next available
923 */
Jens Axboe5e705372006-07-13 12:39:25 +0200924 if ((rq = cfq_check_fifo(cfqq)) == NULL)
925 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200926
927 /*
928 * finally, insert request into driver dispatch list
929 */
Jens Axboe5e705372006-07-13 12:39:25 +0200930 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200931
932 cfqd->dispatch_slice++;
933 dispatched++;
934
935 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +0200936 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
937 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200938 }
939
Jens Axboedd67d052006-06-21 09:36:18 +0200940 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200941 break;
942
943 } while (dispatched < max_dispatch);
944
945 /*
Jens Axboecaaa5f92006-06-16 11:23:00 +0200946 * if slice end isn't set yet, set it.
Jens Axboe22e2c502005-06-27 10:55:12 +0200947 */
948 if (!cfqq->slice_end)
949 cfq_set_prio_slice(cfqd, cfqq);
950
951 /*
952 * expire an async queue immediately if it has used up its slice. idle
953 * queue always expire after 1 dispatch round.
954 */
955 if ((!cfq_cfqq_sync(cfqq) &&
956 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboecaaa5f92006-06-16 11:23:00 +0200957 cfq_class_idle(cfqq) ||
958 !cfq_cfqq_idle_window(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200959 cfq_slice_expired(cfqd, 0);
960
961 return dispatched;
962}
963
964static int
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100965cfq_forced_dispatch_cfqqs(struct list_head *list)
966{
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100967 struct cfq_queue *cfqq, *next;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200968 int dispatched;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100969
Jens Axboecaaa5f92006-06-16 11:23:00 +0200970 dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100971 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
Jens Axboe5e705372006-07-13 12:39:25 +0200972 while (cfqq->next_rq) {
973 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100974 dispatched++;
975 }
976 BUG_ON(!list_empty(&cfqq->fifo));
977 }
Jens Axboecaaa5f92006-06-16 11:23:00 +0200978
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100979 return dispatched;
980}
981
982static int
983cfq_forced_dispatch(struct cfq_data *cfqd)
984{
985 int i, dispatched = 0;
986
987 for (i = 0; i < CFQ_PRIO_LISTS; i++)
988 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
989
990 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
991 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
992 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
993
994 cfq_slice_expired(cfqd, 0);
995
996 BUG_ON(cfqd->busy_queues);
997
998 return dispatched;
999}
1000
1001static int
Jens Axboeb4878f22005-10-20 16:42:29 +02001002cfq_dispatch_requests(request_queue_t *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003{
1004 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001005 struct cfq_queue *cfqq, *prev_cfqq;
1006 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007
Jens Axboe22e2c502005-06-27 10:55:12 +02001008 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009 return 0;
1010
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001011 if (unlikely(force))
1012 return cfq_forced_dispatch(cfqd);
1013
Jens Axboecaaa5f92006-06-16 11:23:00 +02001014 dispatched = 0;
1015 prev_cfqq = NULL;
1016 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001017 int max_dispatch;
1018
Jens Axboecaaa5f92006-06-16 11:23:00 +02001019 /*
1020 * Don't repeat dispatch from the previous queue.
1021 */
1022 if (prev_cfqq == cfqq)
1023 break;
1024
Jens Axboe3b181522005-06-27 10:56:24 +02001025 cfq_clear_cfqq_must_dispatch(cfqq);
1026 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001027 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001029 max_dispatch = cfqd->cfq_quantum;
1030 if (cfq_class_idle(cfqq))
1031 max_dispatch = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032
Jens Axboecaaa5f92006-06-16 11:23:00 +02001033 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1034
1035 /*
1036 * If the dispatch cfqq has idling enabled and is still
1037 * the active queue, break out.
1038 */
1039 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1040 break;
1041
1042 prev_cfqq = cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 }
1044
Jens Axboecaaa5f92006-06-16 11:23:00 +02001045 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046}
1047
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048/*
Jens Axboe5e705372006-07-13 12:39:25 +02001049 * task holds one reference to the queue, dropped when task exits. each rq
1050 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051 *
1052 * queue lock must be held here.
1053 */
1054static void cfq_put_queue(struct cfq_queue *cfqq)
1055{
Jens Axboe22e2c502005-06-27 10:55:12 +02001056 struct cfq_data *cfqd = cfqq->cfqd;
1057
1058 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059
1060 if (!atomic_dec_and_test(&cfqq->ref))
1061 return;
1062
1063 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001064 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001065 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001067 if (unlikely(cfqd->active_queue == cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001068 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02001069
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070 /*
1071 * it's on the empty list and still hashed
1072 */
1073 list_del(&cfqq->cfq_list);
1074 hlist_del(&cfqq->cfq_hash);
1075 kmem_cache_free(cfq_pool, cfqq);
1076}
1077
Jens Axboe1ea25ec2006-07-18 22:24:11 +02001078static struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001079__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1080 const int hashval)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081{
1082 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
Jens Axboe206dc692006-03-28 13:03:44 +02001083 struct hlist_node *entry;
1084 struct cfq_queue *__cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085
Jens Axboe206dc692006-03-28 13:03:44 +02001086 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
Al Virob0a69162006-03-14 15:32:50 -05001087 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088
Jens Axboe206dc692006-03-28 13:03:44 +02001089 if (__cfqq->key == key && (__p == prio || !prio))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 return __cfqq;
1091 }
1092
1093 return NULL;
1094}
1095
1096static struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001097cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098{
Jens Axboe3b181522005-06-27 10:56:24 +02001099 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100}
1101
Jens Axboee2d74ac2006-03-28 08:59:01 +02001102static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103{
Jens Axboe22e2c502005-06-27 10:55:12 +02001104 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001105 struct rb_node *n;
1106 int freed = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001107
Jens Axboee2d74ac2006-03-28 08:59:01 +02001108 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1109 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1110 rb_erase(&__cic->rb_node, &ioc->cic_root);
Jens Axboe22e2c502005-06-27 10:55:12 +02001111 kmem_cache_free(cfq_ioc_pool, __cic);
Al Viro334e94d2006-03-18 15:05:53 -05001112 freed++;
Jens Axboe22e2c502005-06-27 10:55:12 +02001113 }
1114
Jens Axboe4050cf12006-07-19 05:07:12 +02001115 elv_ioc_count_mod(ioc_count, -freed);
1116
1117 if (ioc_gone && !elv_ioc_count_read(ioc_count))
Al Viro334e94d2006-03-18 15:05:53 -05001118 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119}
1120
Jens Axboe89850f72006-07-22 16:48:31 +02001121static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1122{
1123 if (unlikely(cfqq == cfqd->active_queue))
1124 __cfq_slice_expired(cfqd, cfqq, 0);
1125
1126 cfq_put_queue(cfqq);
1127}
1128
1129static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1130 struct cfq_io_context *cic)
1131{
Jens Axboefc463792006-08-29 09:05:44 +02001132 list_del_init(&cic->queue_list);
1133 smp_wmb();
1134 cic->key = NULL;
1135
Jens Axboe89850f72006-07-22 16:48:31 +02001136 if (cic->cfqq[ASYNC]) {
1137 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1138 cic->cfqq[ASYNC] = NULL;
1139 }
1140
1141 if (cic->cfqq[SYNC]) {
1142 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1143 cic->cfqq[SYNC] = NULL;
1144 }
Jens Axboe89850f72006-07-22 16:48:31 +02001145}
1146
1147
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001149 * Called with interrupts disabled
1150 */
1151static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1152{
Al Viro478a82b2006-03-18 13:25:24 -05001153 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001154
Jens Axboe89850f72006-07-22 16:48:31 +02001155 if (cfqd) {
1156 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001157
Jens Axboefc463792006-08-29 09:05:44 +02001158 spin_lock_irq(q->queue_lock);
Jens Axboe89850f72006-07-22 16:48:31 +02001159 __cfq_exit_single_io_context(cfqd, cic);
Jens Axboefc463792006-08-29 09:05:44 +02001160 spin_unlock_irq(q->queue_lock);
Al Viro12a05732006-03-18 13:38:01 -05001161 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001162}
1163
Jens Axboee2d74ac2006-03-28 08:59:01 +02001164static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165{
Jens Axboe22e2c502005-06-27 10:55:12 +02001166 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001167 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +02001168
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 /*
1170 * put the reference this task is holding to the various queues
1171 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001172
1173 n = rb_first(&ioc->cic_root);
1174 while (n != NULL) {
1175 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1176
Jens Axboe22e2c502005-06-27 10:55:12 +02001177 cfq_exit_single_io_context(__cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001178 n = rb_next(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180}
1181
Jens Axboe22e2c502005-06-27 10:55:12 +02001182static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001183cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184{
Jens Axboeb5deef92006-07-19 23:39:40 +02001185 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186
Jens Axboeb5deef92006-07-19 23:39:40 +02001187 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 if (cic) {
Jens Axboe553698f2006-06-14 19:11:57 +02001189 memset(cic, 0, sizeof(*cic));
Jens Axboe22e2c502005-06-27 10:55:12 +02001190 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001191 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001192 cic->dtor = cfq_free_io_context;
1193 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001194 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 }
1196
1197 return cic;
1198}
1199
Jens Axboe22e2c502005-06-27 10:55:12 +02001200static void cfq_init_prio_data(struct cfq_queue *cfqq)
1201{
1202 struct task_struct *tsk = current;
1203 int ioprio_class;
1204
Jens Axboe3b181522005-06-27 10:56:24 +02001205 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001206 return;
1207
1208 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1209 switch (ioprio_class) {
1210 default:
1211 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1212 case IOPRIO_CLASS_NONE:
1213 /*
1214 * no prio set, place us in the middle of the BE classes
1215 */
1216 cfqq->ioprio = task_nice_ioprio(tsk);
1217 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1218 break;
1219 case IOPRIO_CLASS_RT:
1220 cfqq->ioprio = task_ioprio(tsk);
1221 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1222 break;
1223 case IOPRIO_CLASS_BE:
1224 cfqq->ioprio = task_ioprio(tsk);
1225 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1226 break;
1227 case IOPRIO_CLASS_IDLE:
1228 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1229 cfqq->ioprio = 7;
Jens Axboe3b181522005-06-27 10:56:24 +02001230 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001231 break;
1232 }
1233
1234 /*
1235 * keep track of original prio settings in case we have to temporarily
1236 * elevate the priority of this queue
1237 */
1238 cfqq->org_ioprio = cfqq->ioprio;
1239 cfqq->org_ioprio_class = cfqq->ioprio_class;
1240
Jens Axboe3b181522005-06-27 10:56:24 +02001241 if (cfq_cfqq_on_rr(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001242 cfq_resort_rr_list(cfqq, 0);
1243
Jens Axboe3b181522005-06-27 10:56:24 +02001244 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001245}
1246
Al Viro478a82b2006-03-18 13:25:24 -05001247static inline void changed_ioprio(struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001248{
Al Viro478a82b2006-03-18 13:25:24 -05001249 struct cfq_data *cfqd = cic->key;
1250 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001251 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001252
Jens Axboecaaa5f92006-06-16 11:23:00 +02001253 if (unlikely(!cfqd))
1254 return;
1255
Jens Axboec1b707d2006-10-30 19:54:23 +01001256 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001257
1258 cfqq = cic->cfqq[ASYNC];
1259 if (cfqq) {
1260 struct cfq_queue *new_cfqq;
1261 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1262 GFP_ATOMIC);
1263 if (new_cfqq) {
1264 cic->cfqq[ASYNC] = new_cfqq;
1265 cfq_put_queue(cfqq);
1266 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001267 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001268
1269 cfqq = cic->cfqq[SYNC];
1270 if (cfqq)
1271 cfq_mark_cfqq_prio_changed(cfqq);
1272
Jens Axboec1b707d2006-10-30 19:54:23 +01001273 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001274}
1275
Jens Axboefc463792006-08-29 09:05:44 +02001276static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277{
Al Viroa6a07632006-03-18 13:26:44 -05001278 struct cfq_io_context *cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001279 struct rb_node *n;
Al Viroa6a07632006-03-18 13:26:44 -05001280
Jens Axboefc463792006-08-29 09:05:44 +02001281 ioc->ioprio_changed = 0;
Al Viroa6a07632006-03-18 13:26:44 -05001282
Jens Axboee2d74ac2006-03-28 08:59:01 +02001283 n = rb_first(&ioc->cic_root);
1284 while (n != NULL) {
1285 cic = rb_entry(n, struct cfq_io_context, rb_node);
Jens Axboe3793c652006-05-30 21:11:04 +02001286
Al Viro478a82b2006-03-18 13:25:24 -05001287 changed_ioprio(cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001288 n = rb_next(n);
1289 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290}
1291
1292static struct cfq_queue *
Al Viro6f325a12006-03-18 14:58:37 -05001293cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
Al Viro8267e262005-10-21 03:20:53 -04001294 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001295{
1296 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1297 struct cfq_queue *cfqq, *new_cfqq = NULL;
Al Viro6f325a12006-03-18 14:58:37 -05001298 unsigned short ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299
1300retry:
Al Viro6f325a12006-03-18 14:58:37 -05001301 ioprio = tsk->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02001302 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303
1304 if (!cfqq) {
1305 if (new_cfqq) {
1306 cfqq = new_cfqq;
1307 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001308 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001309 /*
1310 * Inform the allocator of the fact that we will
1311 * just repeat this allocation if it fails, to allow
1312 * the allocator to do whatever it needs to attempt to
1313 * free memory.
1314 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 spin_unlock_irq(cfqd->queue->queue_lock);
Jens Axboeb5deef92006-07-19 23:39:40 +02001316 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 spin_lock_irq(cfqd->queue->queue_lock);
1318 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001319 } else {
Jens Axboeb5deef92006-07-19 23:39:40 +02001320 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001321 if (!cfqq)
1322 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001323 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324
1325 memset(cfqq, 0, sizeof(*cfqq));
1326
1327 INIT_HLIST_NODE(&cfqq->cfq_hash);
1328 INIT_LIST_HEAD(&cfqq->cfq_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001329 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330
1331 cfqq->key = key;
1332 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1333 atomic_set(&cfqq->ref, 0);
1334 cfqq->cfqd = cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +02001335 /*
1336 * set ->slice_left to allow preemption for a new process
1337 */
1338 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001339 cfq_mark_cfqq_idle_window(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001340 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b03742006-07-28 09:48:51 +02001341 cfq_mark_cfqq_queue_new(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001342 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 }
1344
1345 if (new_cfqq)
1346 kmem_cache_free(cfq_pool, new_cfqq);
1347
1348 atomic_inc(&cfqq->ref);
1349out:
1350 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1351 return cfqq;
1352}
1353
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001354static void
1355cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1356{
Jens Axboefc463792006-08-29 09:05:44 +02001357 WARN_ON(!list_empty(&cic->queue_list));
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001358 rb_erase(&cic->rb_node, &ioc->cic_root);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001359 kmem_cache_free(cfq_ioc_pool, cic);
Jens Axboe4050cf12006-07-19 05:07:12 +02001360 elv_ioc_count_dec(ioc_count);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001361}
1362
Jens Axboee2d74ac2006-03-28 08:59:01 +02001363static struct cfq_io_context *
1364cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1365{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001366 struct rb_node *n;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001367 struct cfq_io_context *cic;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001368 void *k, *key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001369
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001370restart:
1371 n = ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001372 while (n) {
1373 cic = rb_entry(n, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001374 /* ->key must be copied to avoid race with cfq_exit_queue() */
1375 k = cic->key;
1376 if (unlikely(!k)) {
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001377 cfq_drop_dead_cic(ioc, cic);
1378 goto restart;
1379 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001380
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001381 if (key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001382 n = n->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001383 else if (key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001384 n = n->rb_right;
1385 else
1386 return cic;
1387 }
1388
1389 return NULL;
1390}
1391
1392static inline void
1393cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1394 struct cfq_io_context *cic)
1395{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001396 struct rb_node **p;
1397 struct rb_node *parent;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001398 struct cfq_io_context *__cic;
Jens Axboe0261d682006-10-30 19:07:48 +01001399 unsigned long flags;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001400 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001401
Jens Axboee2d74ac2006-03-28 08:59:01 +02001402 cic->ioc = ioc;
1403 cic->key = cfqd;
1404
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001405restart:
1406 parent = NULL;
1407 p = &ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001408 while (*p) {
1409 parent = *p;
1410 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001411 /* ->key must be copied to avoid race with cfq_exit_queue() */
1412 k = __cic->key;
1413 if (unlikely(!k)) {
Oleg Nesterovbe33c3a2006-08-21 08:36:12 +02001414 cfq_drop_dead_cic(ioc, __cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001415 goto restart;
1416 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001417
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001418 if (cic->key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001419 p = &(*p)->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001420 else if (cic->key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001421 p = &(*p)->rb_right;
1422 else
1423 BUG();
1424 }
1425
1426 rb_link_node(&cic->rb_node, parent, p);
1427 rb_insert_color(&cic->rb_node, &ioc->cic_root);
Jens Axboefc463792006-08-29 09:05:44 +02001428
Jens Axboe0261d682006-10-30 19:07:48 +01001429 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001430 list_add(&cic->queue_list, &cfqd->cic_list);
Jens Axboe0261d682006-10-30 19:07:48 +01001431 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001432}
1433
Jens Axboe22e2c502005-06-27 10:55:12 +02001434/*
1435 * Setup general io context and cfq io context. There can be several cfq
1436 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001437 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001438 */
1439static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001440cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441{
Jens Axboe22e2c502005-06-27 10:55:12 +02001442 struct io_context *ioc = NULL;
1443 struct cfq_io_context *cic;
1444
1445 might_sleep_if(gfp_mask & __GFP_WAIT);
1446
Jens Axboeb5deef92006-07-19 23:39:40 +02001447 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001448 if (!ioc)
1449 return NULL;
1450
Jens Axboee2d74ac2006-03-28 08:59:01 +02001451 cic = cfq_cic_rb_lookup(cfqd, ioc);
1452 if (cic)
1453 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001454
Jens Axboee2d74ac2006-03-28 08:59:01 +02001455 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1456 if (cic == NULL)
1457 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001458
Jens Axboee2d74ac2006-03-28 08:59:01 +02001459 cfq_cic_link(cfqd, ioc, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001460out:
Jens Axboefc463792006-08-29 09:05:44 +02001461 smp_read_barrier_depends();
1462 if (unlikely(ioc->ioprio_changed))
1463 cfq_ioc_set_ioprio(ioc);
1464
Jens Axboe22e2c502005-06-27 10:55:12 +02001465 return cic;
1466err:
1467 put_io_context(ioc);
1468 return NULL;
1469}
1470
1471static void
1472cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1473{
1474 unsigned long elapsed, ttime;
1475
1476 /*
1477 * if this context already has stuff queued, thinktime is from
1478 * last queue not last end
1479 */
1480#if 0
1481 if (time_after(cic->last_end_request, cic->last_queue))
1482 elapsed = jiffies - cic->last_end_request;
1483 else
1484 elapsed = jiffies - cic->last_queue;
1485#else
1486 elapsed = jiffies - cic->last_end_request;
1487#endif
1488
1489 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1490
1491 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1492 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1493 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1494}
1495
Jens Axboe206dc692006-03-28 13:03:44 +02001496static void
Jens Axboebb37b942006-12-01 10:42:33 +01001497cfq_update_io_seektime(struct cfq_io_context *cic, struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001498{
1499 sector_t sdist;
1500 u64 total;
1501
Jens Axboe5e705372006-07-13 12:39:25 +02001502 if (cic->last_request_pos < rq->sector)
1503 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001504 else
Jens Axboe5e705372006-07-13 12:39:25 +02001505 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001506
1507 /*
1508 * Don't allow the seek distance to get too large from the
1509 * odd fragment, pagein, etc
1510 */
1511 if (cic->seek_samples <= 60) /* second&third seek */
1512 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1513 else
1514 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1515
1516 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1517 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1518 total = cic->seek_total + (cic->seek_samples/2);
1519 do_div(total, cic->seek_samples);
1520 cic->seek_mean = (sector_t)total;
1521}
Jens Axboe22e2c502005-06-27 10:55:12 +02001522
1523/*
1524 * Disable idle window if the process thinks too long or seeks so much that
1525 * it doesn't matter
1526 */
1527static void
1528cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1529 struct cfq_io_context *cic)
1530{
Jens Axboe3b181522005-06-27 10:56:24 +02001531 int enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001532
Jens Axboecaaa5f92006-06-16 11:23:00 +02001533 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1534 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001535 enable_idle = 0;
1536 else if (sample_valid(cic->ttime_samples)) {
1537 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1538 enable_idle = 0;
1539 else
1540 enable_idle = 1;
1541 }
1542
Jens Axboe3b181522005-06-27 10:56:24 +02001543 if (enable_idle)
1544 cfq_mark_cfqq_idle_window(cfqq);
1545 else
1546 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001547}
1548
1549
1550/*
1551 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1552 * no or if we aren't sure, a 1 will cause a preempt.
1553 */
1554static int
1555cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001556 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001557{
1558 struct cfq_queue *cfqq = cfqd->active_queue;
1559
1560 if (cfq_class_idle(new_cfqq))
1561 return 0;
1562
1563 if (!cfqq)
Jens Axboecaaa5f92006-06-16 11:23:00 +02001564 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001565
1566 if (cfq_class_idle(cfqq))
1567 return 1;
Jens Axboe3b181522005-06-27 10:56:24 +02001568 if (!cfq_cfqq_wait_request(new_cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001569 return 0;
1570 /*
1571 * if it doesn't have slice left, forget it
1572 */
1573 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1574 return 0;
Jens Axboe374f84a2006-07-23 01:42:19 +02001575 /*
1576 * if the new request is sync, but the currently running queue is
1577 * not, let the sync request have priority.
1578 */
Jens Axboe5e705372006-07-13 12:39:25 +02001579 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001580 return 1;
Jens Axboe374f84a2006-07-23 01:42:19 +02001581 /*
1582 * So both queues are sync. Let the new request get disk time if
1583 * it's a metadata request and the current queue is doing regular IO.
1584 */
1585 if (rq_is_meta(rq) && !cfqq->meta_pending)
1586 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001587
1588 return 0;
1589}
1590
1591/*
1592 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1593 * let it have half of its nominal slice.
1594 */
1595static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1596{
Jens Axboebf572252006-07-19 20:29:12 +02001597 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001598
1599 if (!cfqq->slice_left)
1600 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1601
Jens Axboebf572252006-07-19 20:29:12 +02001602 /*
1603 * Put the new queue at the front of the of the current list,
1604 * so we know that it will be selected next.
1605 */
1606 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1607 list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1608
Jens Axboe22e2c502005-06-27 10:55:12 +02001609 cfqq->slice_end = cfqq->slice_left + jiffies;
Jens Axboe22e2c502005-06-27 10:55:12 +02001610}
1611
1612/*
Jens Axboe5e705372006-07-13 12:39:25 +02001613 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001614 * something we should do about it
1615 */
1616static void
Jens Axboe5e705372006-07-13 12:39:25 +02001617cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1618 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001619{
Jens Axboe5e705372006-07-13 12:39:25 +02001620 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001621
Jens Axboe374f84a2006-07-23 01:42:19 +02001622 if (rq_is_meta(rq))
1623 cfqq->meta_pending++;
1624
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001625 /*
Jens Axboe5380a102006-07-13 12:37:56 +02001626 * check if this request is a better next-serve candidate)) {
Jens Axboe21183b02006-07-13 12:33:14 +02001627 */
Jens Axboe5e705372006-07-13 12:39:25 +02001628 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1629 BUG_ON(!cfqq->next_rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001630
1631 /*
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001632 * we never wait for an async request and we don't allow preemption
1633 * of an async request. so just return early
1634 */
Jens Axboe5e705372006-07-13 12:39:25 +02001635 if (!rq_is_sync(rq)) {
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001636 /*
1637 * sync process issued an async request, if it's waiting
1638 * then expire it and kick rq handling.
1639 */
1640 if (cic == cfqd->active_cic &&
1641 del_timer(&cfqd->idle_slice_timer)) {
1642 cfq_slice_expired(cfqd, 0);
Jens Axboedc72ef42006-07-20 14:54:05 +02001643 blk_start_queueing(cfqd->queue);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001644 }
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001645 return;
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001646 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001647
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001648 cfq_update_io_thinktime(cfqd, cic);
Jens Axboebb37b942006-12-01 10:42:33 +01001649 cfq_update_io_seektime(cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001650 cfq_update_idle_window(cfqd, cfqq, cic);
1651
1652 cic->last_queue = jiffies;
Jens Axboe5e705372006-07-13 12:39:25 +02001653 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001654
1655 if (cfqq == cfqd->active_queue) {
1656 /*
1657 * if we are waiting for a request for this queue, let it rip
1658 * immediately and flag that we must not expire this queue
1659 * just now
1660 */
Jens Axboe3b181522005-06-27 10:56:24 +02001661 if (cfq_cfqq_wait_request(cfqq)) {
1662 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001663 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001664 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001665 }
Jens Axboe5e705372006-07-13 12:39:25 +02001666 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001667 /*
1668 * not the active queue - expire current slice if it is
1669 * idle and has expired it's mean thinktime or this new queue
1670 * has some old slice time left and is of higher priority
1671 */
1672 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001673 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001674 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001675 }
1676}
1677
Jens Axboeb4878f22005-10-20 16:42:29 +02001678static void cfq_insert_request(request_queue_t *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001679{
Jens Axboeb4878f22005-10-20 16:42:29 +02001680 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001681 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001682
1683 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684
Jens Axboe5e705372006-07-13 12:39:25 +02001685 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
Jens Axboe22e2c502005-06-27 10:55:12 +02001687 list_add_tail(&rq->queuelist, &cfqq->fifo);
1688
Jens Axboe5e705372006-07-13 12:39:25 +02001689 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690}
1691
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692static void cfq_completed_request(request_queue_t *q, struct request *rq)
1693{
Jens Axboe5e705372006-07-13 12:39:25 +02001694 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001695 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001696 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001697 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698
Jens Axboeb4878f22005-10-20 16:42:29 +02001699 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700
Jens Axboeb4878f22005-10-20 16:42:29 +02001701 WARN_ON(!cfqd->rq_in_driver);
1702 WARN_ON(!cfqq->on_dispatch[sync]);
1703 cfqd->rq_in_driver--;
1704 cfqq->on_dispatch[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705
Jens Axboeb4878f22005-10-20 16:42:29 +02001706 if (!cfq_class_idle(cfqq))
1707 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001708
Jens Axboe53b03742006-07-28 09:48:51 +02001709 if (!cfq_cfqq_dispatched(cfqq) && cfq_cfqq_on_rr(cfqq))
1710 cfq_resort_rr_list(cfqq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711
Jens Axboecaaa5f92006-06-16 11:23:00 +02001712 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001713 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001714
1715 /*
1716 * If this is the active queue, check if it needs to be expired,
1717 * or if we want to idle in case it has no pending requests.
1718 */
1719 if (cfqd->active_queue == cfqq) {
1720 if (time_after(now, cfqq->slice_end))
1721 cfq_slice_expired(cfqd, 0);
Jens Axboedd67d052006-06-21 09:36:18 +02001722 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001723 if (!cfq_arm_slice_timer(cfqd, cfqq))
1724 cfq_schedule_dispatch(cfqd);
1725 }
1726 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727}
1728
Jens Axboe22e2c502005-06-27 10:55:12 +02001729/*
1730 * we temporarily boost lower priority queues if they are holding fs exclusive
1731 * resources. they are boosted to normal prio (CLASS_BE/4)
1732 */
1733static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734{
Jens Axboe22e2c502005-06-27 10:55:12 +02001735 const int ioprio_class = cfqq->ioprio_class;
1736 const int ioprio = cfqq->ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737
Jens Axboe22e2c502005-06-27 10:55:12 +02001738 if (has_fs_excl()) {
1739 /*
1740 * boost idle prio on transactions that would lock out other
1741 * users of the filesystem
1742 */
1743 if (cfq_class_idle(cfqq))
1744 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1745 if (cfqq->ioprio > IOPRIO_NORM)
1746 cfqq->ioprio = IOPRIO_NORM;
1747 } else {
1748 /*
1749 * check if we need to unboost the queue
1750 */
1751 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1752 cfqq->ioprio_class = cfqq->org_ioprio_class;
1753 if (cfqq->ioprio != cfqq->org_ioprio)
1754 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755 }
1756
Jens Axboe22e2c502005-06-27 10:55:12 +02001757 /*
1758 * refile between round-robin lists if we moved the priority class
1759 */
1760 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
Jens Axboe3b181522005-06-27 10:56:24 +02001761 cfq_cfqq_on_rr(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001762 cfq_resort_rr_list(cfqq, 0);
1763}
1764
Jens Axboe89850f72006-07-22 16:48:31 +02001765static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001766{
Jens Axboe3b181522005-06-27 10:56:24 +02001767 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001768 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001769 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001770 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001771 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001772
1773 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02001774}
1775
Jens Axboecb78b282006-07-28 09:32:57 +02001776static int cfq_may_queue(request_queue_t *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02001777{
1778 struct cfq_data *cfqd = q->elevator->elevator_data;
1779 struct task_struct *tsk = current;
1780 struct cfq_queue *cfqq;
Jens Axboe7749a8d2006-12-13 13:02:26 +01001781 unsigned int key;
1782
1783 key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02001784
1785 /*
1786 * don't force setup of a queue from here, as a call to may_queue
1787 * does not necessarily imply that a request actually will be queued.
1788 * so just lookup a possibly existing queue, or return 'may queue'
1789 * if that fails
1790 */
Jens Axboe7749a8d2006-12-13 13:02:26 +01001791 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001792 if (cfqq) {
1793 cfq_init_prio_data(cfqq);
1794 cfq_prio_boost(cfqq);
1795
Jens Axboe89850f72006-07-22 16:48:31 +02001796 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001797 }
1798
1799 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800}
1801
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802/*
1803 * queue lock held here
1804 */
Jens Axboebb37b942006-12-01 10:42:33 +01001805static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806{
Jens Axboe5e705372006-07-13 12:39:25 +02001807 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808
Jens Axboe5e705372006-07-13 12:39:25 +02001809 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001810 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811
Jens Axboe22e2c502005-06-27 10:55:12 +02001812 BUG_ON(!cfqq->allocated[rw]);
1813 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814
Jens Axboe5e705372006-07-13 12:39:25 +02001815 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02001818 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820 cfq_put_queue(cfqq);
1821 }
1822}
1823
1824/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001825 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001827static int
Jens Axboecb78b282006-07-28 09:32:57 +02001828cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829{
1830 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe3b181522005-06-27 10:56:24 +02001831 struct task_struct *tsk = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 struct cfq_io_context *cic;
1833 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01001834 const int is_sync = rq_is_sync(rq);
1835 pid_t key = cfq_queue_pid(tsk, rw, is_sync);
Jens Axboe22e2c502005-06-27 10:55:12 +02001836 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837 unsigned long flags;
1838
1839 might_sleep_if(gfp_mask & __GFP_WAIT);
1840
Jens Axboee2d74ac2006-03-28 08:59:01 +02001841 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001842
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 spin_lock_irqsave(q->queue_lock, flags);
1844
Jens Axboe22e2c502005-06-27 10:55:12 +02001845 if (!cic)
1846 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847
Al Viro12a05732006-03-18 13:38:01 -05001848 if (!cic->cfqq[is_sync]) {
Al Viro6f325a12006-03-18 14:58:37 -05001849 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001850 if (!cfqq)
1851 goto queue_fail;
1852
Al Viro12a05732006-03-18 13:38:01 -05001853 cic->cfqq[is_sync] = cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001854 } else
Al Viro12a05732006-03-18 13:38:01 -05001855 cfqq = cic->cfqq[is_sync];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856
1857 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02001858 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001859 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02001860
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861 spin_unlock_irqrestore(q->queue_lock, flags);
1862
Jens Axboe5e705372006-07-13 12:39:25 +02001863 rq->elevator_private = cic;
1864 rq->elevator_private2 = cfqq;
1865 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02001866
Jens Axboe22e2c502005-06-27 10:55:12 +02001867queue_fail:
1868 if (cic)
1869 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02001870
Jens Axboe3b181522005-06-27 10:56:24 +02001871 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 spin_unlock_irqrestore(q->queue_lock, flags);
1873 return 1;
1874}
1875
David Howells65f27f32006-11-22 14:55:48 +00001876static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02001877{
David Howells65f27f32006-11-22 14:55:48 +00001878 struct cfq_data *cfqd =
1879 container_of(work, struct cfq_data, unplug_work);
1880 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001881 unsigned long flags;
1882
1883 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02001884 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02001885 spin_unlock_irqrestore(q->queue_lock, flags);
1886}
1887
1888/*
1889 * Timer running if the active_queue is currently idling inside its time slice
1890 */
1891static void cfq_idle_slice_timer(unsigned long data)
1892{
1893 struct cfq_data *cfqd = (struct cfq_data *) data;
1894 struct cfq_queue *cfqq;
1895 unsigned long flags;
1896
1897 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1898
1899 if ((cfqq = cfqd->active_queue) != NULL) {
1900 unsigned long now = jiffies;
1901
1902 /*
1903 * expired
1904 */
1905 if (time_after(now, cfqq->slice_end))
1906 goto expire;
1907
1908 /*
1909 * only expire and reinvoke request handler, if there are
1910 * other queues with pending requests
1911 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02001912 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02001913 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02001914
1915 /*
1916 * not expired and it has a request pending, let it dispatch
1917 */
Jens Axboedd67d052006-06-21 09:36:18 +02001918 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001919 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001920 goto out_kick;
1921 }
1922 }
1923expire:
1924 cfq_slice_expired(cfqd, 0);
1925out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02001926 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001927out_cont:
1928 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1929}
1930
1931/*
1932 * Timer running if an idle class queue is waiting for service
1933 */
1934static void cfq_idle_class_timer(unsigned long data)
1935{
1936 struct cfq_data *cfqd = (struct cfq_data *) data;
1937 unsigned long flags, end;
1938
1939 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1940
1941 /*
1942 * race with a non-idle queue, reset timer
1943 */
1944 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
Jens Axboeae818a32006-06-01 10:13:43 +02001945 if (!time_after_eq(jiffies, end))
1946 mod_timer(&cfqd->idle_class_timer, end);
1947 else
Jens Axboe3b181522005-06-27 10:56:24 +02001948 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001949
1950 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1951}
1952
Jens Axboe3b181522005-06-27 10:56:24 +02001953static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1954{
1955 del_timer_sync(&cfqd->idle_slice_timer);
1956 del_timer_sync(&cfqd->idle_class_timer);
1957 blk_sync_queue(cfqd->queue);
1958}
Jens Axboe22e2c502005-06-27 10:55:12 +02001959
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960static void cfq_exit_queue(elevator_t *e)
1961{
Jens Axboe22e2c502005-06-27 10:55:12 +02001962 struct cfq_data *cfqd = e->elevator_data;
Al Virod9ff4182006-03-18 13:51:22 -05001963 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001964
Jens Axboe3b181522005-06-27 10:56:24 +02001965 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001966
Al Virod9ff4182006-03-18 13:51:22 -05001967 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001968
Al Virod9ff4182006-03-18 13:51:22 -05001969 if (cfqd->active_queue)
1970 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001971
1972 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05001973 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1974 struct cfq_io_context,
1975 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02001976
1977 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05001978 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001979
Al Virod9ff4182006-03-18 13:51:22 -05001980 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05001981
1982 cfq_shutdown_timer_wq(cfqd);
1983
Al Viroa90d7422006-03-18 12:05:37 -05001984 kfree(cfqd->cfq_hash);
1985 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986}
1987
Jens Axboebb37b942006-12-01 10:42:33 +01001988static void *cfq_init_queue(request_queue_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989{
1990 struct cfq_data *cfqd;
1991 int i;
1992
Jens Axboeb5deef92006-07-19 23:39:40 +02001993 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02001995 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996
1997 memset(cfqd, 0, sizeof(*cfqd));
Jens Axboe22e2c502005-06-27 10:55:12 +02001998
1999 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2000 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2001
2002 INIT_LIST_HEAD(&cfqd->busy_rr);
2003 INIT_LIST_HEAD(&cfqd->cur_rr);
2004 INIT_LIST_HEAD(&cfqd->idle_rr);
Al Virod9ff4182006-03-18 13:51:22 -05002005 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006
Jens Axboeb5deef92006-07-19 23:39:40 +02002007 cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008 if (!cfqd->cfq_hash)
Jens Axboe5e705372006-07-13 12:39:25 +02002009 goto out_free;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2012 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2013
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015
Jens Axboe22e2c502005-06-27 10:55:12 +02002016 init_timer(&cfqd->idle_slice_timer);
2017 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2018 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2019
2020 init_timer(&cfqd->idle_class_timer);
2021 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2022 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2023
David Howells65f27f32006-11-22 14:55:48 +00002024 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002025
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002027 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2028 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029 cfqd->cfq_back_max = cfq_back_max;
2030 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002031 cfqd->cfq_slice[0] = cfq_slice_async;
2032 cfqd->cfq_slice[1] = cfq_slice_sync;
2033 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2034 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02002035
Jens Axboebc1c1162006-06-08 08:49:06 +02002036 return cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +02002037out_free:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038 kfree(cfqd);
Jens Axboebc1c1162006-06-08 08:49:06 +02002039 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040}
2041
2042static void cfq_slab_kill(void)
2043{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044 if (cfq_pool)
2045 kmem_cache_destroy(cfq_pool);
2046 if (cfq_ioc_pool)
2047 kmem_cache_destroy(cfq_ioc_pool);
2048}
2049
2050static int __init cfq_slab_setup(void)
2051{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2053 NULL, NULL);
2054 if (!cfq_pool)
2055 goto fail;
2056
2057 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2058 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2059 if (!cfq_ioc_pool)
2060 goto fail;
2061
2062 return 0;
2063fail:
2064 cfq_slab_kill();
2065 return -ENOMEM;
2066}
2067
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068/*
2069 * sysfs parts below -->
2070 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071
2072static ssize_t
2073cfq_var_show(unsigned int var, char *page)
2074{
2075 return sprintf(page, "%d\n", var);
2076}
2077
2078static ssize_t
2079cfq_var_store(unsigned int *var, const char *page, size_t count)
2080{
2081 char *p = (char *) page;
2082
2083 *var = simple_strtoul(p, &p, 10);
2084 return count;
2085}
2086
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002088static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002090 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 unsigned int __data = __VAR; \
2092 if (__CONV) \
2093 __data = jiffies_to_msecs(__data); \
2094 return cfq_var_show(__data, (page)); \
2095}
2096SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002097SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2098SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002099SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2100SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002101SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2102SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2103SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2104SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105#undef SHOW_FUNCTION
2106
2107#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002108static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002110 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111 unsigned int __data; \
2112 int ret = cfq_var_store(&__data, (page), count); \
2113 if (__data < (MIN)) \
2114 __data = (MIN); \
2115 else if (__data > (MAX)) \
2116 __data = (MAX); \
2117 if (__CONV) \
2118 *(__PTR) = msecs_to_jiffies(__data); \
2119 else \
2120 *(__PTR) = __data; \
2121 return ret; \
2122}
2123STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002124STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2125STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002126STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2127STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002128STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2129STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2130STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2131STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132#undef STORE_FUNCTION
2133
Al Viroe572ec72006-03-18 22:27:18 -05002134#define CFQ_ATTR(name) \
2135 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002136
Al Viroe572ec72006-03-18 22:27:18 -05002137static struct elv_fs_entry cfq_attrs[] = {
2138 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002139 CFQ_ATTR(fifo_expire_sync),
2140 CFQ_ATTR(fifo_expire_async),
2141 CFQ_ATTR(back_seek_max),
2142 CFQ_ATTR(back_seek_penalty),
2143 CFQ_ATTR(slice_sync),
2144 CFQ_ATTR(slice_async),
2145 CFQ_ATTR(slice_async_rq),
2146 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002147 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148};
2149
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150static struct elevator_type iosched_cfq = {
2151 .ops = {
2152 .elevator_merge_fn = cfq_merge,
2153 .elevator_merged_fn = cfq_merged_request,
2154 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002155 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002156 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002158 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 .elevator_deactivate_req_fn = cfq_deactivate_request,
2160 .elevator_queue_empty_fn = cfq_queue_empty,
2161 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002162 .elevator_former_req_fn = elv_rb_former_request,
2163 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 .elevator_set_req_fn = cfq_set_request,
2165 .elevator_put_req_fn = cfq_put_request,
2166 .elevator_may_queue_fn = cfq_may_queue,
2167 .elevator_init_fn = cfq_init_queue,
2168 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002169 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170 },
Al Viro3d1ab402006-03-18 18:35:43 -05002171 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172 .elevator_name = "cfq",
2173 .elevator_owner = THIS_MODULE,
2174};
2175
2176static int __init cfq_init(void)
2177{
2178 int ret;
2179
Jens Axboe22e2c502005-06-27 10:55:12 +02002180 /*
2181 * could be 0 on HZ < 1000 setups
2182 */
2183 if (!cfq_slice_async)
2184 cfq_slice_async = 1;
2185 if (!cfq_slice_idle)
2186 cfq_slice_idle = 1;
2187
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188 if (cfq_slab_setup())
2189 return -ENOMEM;
2190
2191 ret = elv_register(&iosched_cfq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002192 if (ret)
2193 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195 return ret;
2196}
2197
2198static void __exit cfq_exit(void)
2199{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002200 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002202 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002203 /* ioc_gone's update must be visible before reading ioc_count */
2204 smp_wmb();
Jens Axboe4050cf12006-07-19 05:07:12 +02002205 if (elv_ioc_count_read(ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002206 wait_for_completion(ioc_gone);
Al Viro334e94d2006-03-18 15:05:53 -05002207 synchronize_rcu();
Christoph Hellwig83521d32005-10-30 15:01:39 -08002208 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209}
2210
2211module_init(cfq_init);
2212module_exit(cfq_exit);
2213
2214MODULE_AUTHOR("Jens Axboe");
2215MODULE_LICENSE("GPL");
2216MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");