blob: 78c6b312bd3030d3c86a9c35e3facc842b73e3eb [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 Axboe206dc692006-03-28 13:03:44 +0200222static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
223{
Jens Axboeb31dc662006-06-13 08:26:10 +0200224 if (rw == READ || rw == WRITE_SYNC)
Jens Axboe206dc692006-03-28 13:03:44 +0200225 return task->pid;
226
227 return CFQ_KEY_ASYNC;
228}
229
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230/*
Jens Axboe5e705372006-07-13 12:39:25 +0200231 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200233 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 */
Jens Axboe5e705372006-07-13 12:39:25 +0200235static struct request *
236cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237{
238 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200240#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
241#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
242 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243
Jens Axboe5e705372006-07-13 12:39:25 +0200244 if (rq1 == NULL || rq1 == rq2)
245 return rq2;
246 if (rq2 == NULL)
247 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200248
Jens Axboe5e705372006-07-13 12:39:25 +0200249 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
250 return rq1;
251 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
252 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200253 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
254 return rq1;
255 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
256 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257
Jens Axboe5e705372006-07-13 12:39:25 +0200258 s1 = rq1->sector;
259 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260
261 last = cfqd->last_sector;
262
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 /*
264 * by definition, 1KiB is 2 sectors
265 */
266 back_max = cfqd->cfq_back_max * 2;
267
268 /*
269 * Strict one way elevator _except_ in the case where we allow
270 * short backward seeks which are biased as twice the cost of a
271 * similar forward seek.
272 */
273 if (s1 >= last)
274 d1 = s1 - last;
275 else if (s1 + back_max >= last)
276 d1 = (last - s1) * cfqd->cfq_back_penalty;
277 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200278 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
280 if (s2 >= last)
281 d2 = s2 - last;
282 else if (s2 + back_max >= last)
283 d2 = (last - s2) * cfqd->cfq_back_penalty;
284 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200285 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286
287 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288
Andreas Mohre8a99052006-03-28 08:59:49 +0200289 /*
290 * By doing switch() on the bit mask "wrap" we avoid having to
291 * check two variables for all permutations: --> faster!
292 */
293 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200294 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200295 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200296 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200297 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200298 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200299 else {
300 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200301 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200302 else
Jens Axboe5e705372006-07-13 12:39:25 +0200303 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200304 }
305
306 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200307 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200308 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200309 return rq2;
310 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200311 default:
312 /*
313 * Since both rqs are wrapped,
314 * start with the one that's further behind head
315 * (--> only *one* back seek required),
316 * since back seek takes more time than forward.
317 */
318 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200319 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 else
Jens Axboe5e705372006-07-13 12:39:25 +0200321 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322 }
323}
324
325/*
326 * would be nice to take fifo expire time into account as well
327 */
Jens Axboe5e705372006-07-13 12:39:25 +0200328static struct request *
329cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
330 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331{
Jens Axboe21183b02006-07-13 12:33:14 +0200332 struct rb_node *rbnext = rb_next(&last->rb_node);
333 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200334 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335
Jens Axboe21183b02006-07-13 12:33:14 +0200336 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337
338 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200339 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200340
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200342 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200343 else {
344 rbnext = rb_first(&cfqq->sort_list);
345 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200346 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200347 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348
Jens Axboe21183b02006-07-13 12:33:14 +0200349 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350}
351
Jens Axboe22e2c502005-06-27 10:55:12 +0200352static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353{
Jens Axboe22e2c502005-06-27 10:55:12 +0200354 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe53b03742006-07-28 09:48:51 +0200355 struct list_head *list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356
Jens Axboe3b181522005-06-27 10:56:24 +0200357 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358
359 list_del(&cfqq->cfq_list);
360
Jens Axboe22e2c502005-06-27 10:55:12 +0200361 if (cfq_class_rt(cfqq))
362 list = &cfqd->cur_rr;
363 else if (cfq_class_idle(cfqq))
364 list = &cfqd->idle_rr;
365 else {
366 /*
367 * if cfqq has requests in flight, don't allow it to be
368 * found in cfq_set_active_queue before it has finished them.
369 * this is done to increase fairness between a process that
370 * has lots of io pending vs one that only generates one
371 * sporadically or synchronously
372 */
Jens Axboe3b181522005-06-27 10:56:24 +0200373 if (cfq_cfqq_dispatched(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200374 list = &cfqd->busy_rr;
375 else
376 list = &cfqd->rr_list[cfqq->ioprio];
377 }
378
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 /*
Jens Axboe53b03742006-07-28 09:48:51 +0200380 * If this queue was preempted or is new (never been serviced), let
381 * it be added first for fairness but beind other new queues.
382 * Otherwise, just add to the back of the list.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 */
Jens Axboe53b03742006-07-28 09:48:51 +0200384 if (preempted || cfq_cfqq_queue_new(cfqq)) {
385 struct list_head *n = list;
386 struct cfq_queue *__cfqq;
Jens Axboeb52a8342006-06-01 18:53:43 +0200387
Jens Axboe53b03742006-07-28 09:48:51 +0200388 while (n->next != list) {
389 __cfqq = list_entry_cfqq(n->next);
390 if (!cfq_cfqq_queue_new(__cfqq))
391 break;
392
393 n = n->next;
394 }
395
396 list = n;
Jens Axboe22e2c502005-06-27 10:55:12 +0200397 }
398
Jens Axboe53b03742006-07-28 09:48:51 +0200399 list_add_tail(&cfqq->cfq_list, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400}
401
402/*
403 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200404 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 */
406static inline void
Jens Axboeb4878f22005-10-20 16:42:29 +0200407cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408{
Jens Axboe3b181522005-06-27 10:56:24 +0200409 BUG_ON(cfq_cfqq_on_rr(cfqq));
410 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 cfqd->busy_queues++;
412
Jens Axboeb4878f22005-10-20 16:42:29 +0200413 cfq_resort_rr_list(cfqq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414}
415
416static inline void
417cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
418{
Jens Axboe3b181522005-06-27 10:56:24 +0200419 BUG_ON(!cfq_cfqq_on_rr(cfqq));
420 cfq_clear_cfqq_on_rr(cfqq);
Jens Axboe981a7972006-07-19 14:56:28 +0200421 list_del_init(&cfqq->cfq_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422
423 BUG_ON(!cfqd->busy_queues);
424 cfqd->busy_queues--;
425}
426
427/*
428 * rb tree support functions
429 */
Jens Axboe5e705372006-07-13 12:39:25 +0200430static inline void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431{
Jens Axboe5e705372006-07-13 12:39:25 +0200432 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200433 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200434 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435
Jens Axboeb4878f22005-10-20 16:42:29 +0200436 BUG_ON(!cfqq->queued[sync]);
437 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438
Jens Axboe5e705372006-07-13 12:39:25 +0200439 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440
Jens Axboedd67d052006-06-21 09:36:18 +0200441 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200442 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443}
444
Jens Axboe5e705372006-07-13 12:39:25 +0200445static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446{
Jens Axboe5e705372006-07-13 12:39:25 +0200447 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe21183b02006-07-13 12:33:14 +0200449 struct request *__alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
Jens Axboe5380a102006-07-13 12:37:56 +0200451 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452
453 /*
454 * looks a little odd, but the first insert might return an alias.
455 * if that happens, put the alias on the dispatch list
456 */
Jens Axboe21183b02006-07-13 12:33:14 +0200457 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200458 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100459
460 if (!cfq_cfqq_on_rr(cfqq))
461 cfq_add_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462}
463
464static inline void
Jens Axboe5e705372006-07-13 12:39:25 +0200465cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466{
Jens Axboe5380a102006-07-13 12:37:56 +0200467 elv_rb_del(&cfqq->sort_list, rq);
468 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200469 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470}
471
Jens Axboe206dc692006-03-28 13:03:44 +0200472static struct request *
473cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474{
Jens Axboe206dc692006-03-28 13:03:44 +0200475 struct task_struct *tsk = current;
476 pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
477 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478
Jens Axboe206dc692006-03-28 13:03:44 +0200479 cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
Jens Axboe89850f72006-07-22 16:48:31 +0200480 if (cfqq) {
481 sector_t sector = bio->bi_sector + bio_sectors(bio);
482
Jens Axboe21183b02006-07-13 12:33:14 +0200483 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200484 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 return NULL;
487}
488
Jens Axboeb4878f22005-10-20 16:42:29 +0200489static void cfq_activate_request(request_queue_t *q, struct request *rq)
490{
491 struct cfq_data *cfqd = q->elevator->elevator_data;
492
493 cfqd->rq_in_driver++;
Jens Axboe25776e32006-06-01 10:12:26 +0200494
495 /*
496 * If the depth is larger 1, it really could be queueing. But lets
497 * make the mark a little higher - idling could still be good for
498 * low queueing, and a low queueing number could also just indicate
499 * a SCSI mid layer like behaviour where limit+1 is often seen.
500 */
501 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
502 cfqd->hw_tag = 1;
Jens Axboeb4878f22005-10-20 16:42:29 +0200503}
504
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
506{
Jens Axboe22e2c502005-06-27 10:55:12 +0200507 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508
Jens Axboeb4878f22005-10-20 16:42:29 +0200509 WARN_ON(!cfqd->rq_in_driver);
510 cfqd->rq_in_driver--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511}
512
Jens Axboeb4878f22005-10-20 16:42:29 +0200513static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514{
Jens Axboe5e705372006-07-13 12:39:25 +0200515 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200516
Jens Axboe5e705372006-07-13 12:39:25 +0200517 if (cfqq->next_rq == rq)
518 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519
Jens Axboeb4878f22005-10-20 16:42:29 +0200520 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200521 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200522
523 if (rq_is_meta(rq)) {
524 WARN_ON(!cfqq->meta_pending);
525 cfqq->meta_pending--;
526 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527}
528
529static int
530cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
531{
532 struct cfq_data *cfqd = q->elevator->elevator_data;
533 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534
Jens Axboe206dc692006-03-28 13:03:44 +0200535 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200536 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200537 *req = __rq;
538 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539 }
540
541 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542}
543
Jens Axboe21183b02006-07-13 12:33:14 +0200544static void cfq_merged_request(request_queue_t *q, struct request *req,
545 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546{
Jens Axboe21183b02006-07-13 12:33:14 +0200547 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200548 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549
Jens Axboe5e705372006-07-13 12:39:25 +0200550 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552}
553
554static void
555cfq_merged_requests(request_queue_t *q, struct request *rq,
556 struct request *next)
557{
Jens Axboe22e2c502005-06-27 10:55:12 +0200558 /*
559 * reposition in fifo if next is older than rq
560 */
561 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
562 time_before(next->start_time, rq->start_time))
563 list_move(&rq->queuelist, &next->queuelist);
564
Jens Axboeb4878f22005-10-20 16:42:29 +0200565 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200566}
567
568static inline void
569__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
570{
571 if (cfqq) {
572 /*
573 * stop potential idle class queues waiting service
574 */
575 del_timer(&cfqd->idle_class_timer);
576
577 cfqq->slice_start = jiffies;
578 cfqq->slice_end = 0;
579 cfqq->slice_left = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200580 cfq_clear_cfqq_must_alloc_slice(cfqq);
581 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200582 }
583
584 cfqd->active_queue = cfqq;
585}
586
587/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100588 * current cfqq expired its slice (or was too idle), select new one
589 */
590static void
591__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
592 int preempted)
593{
594 unsigned long now = jiffies;
595
596 if (cfq_cfqq_wait_request(cfqq))
597 del_timer(&cfqd->idle_slice_timer);
598
Jens Axboe53b03742006-07-28 09:48:51 +0200599 if (!preempted && !cfq_cfqq_dispatched(cfqq))
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100600 cfq_schedule_dispatch(cfqd);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100601
602 cfq_clear_cfqq_must_dispatch(cfqq);
603 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe53b03742006-07-28 09:48:51 +0200604 cfq_clear_cfqq_queue_new(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100605
606 /*
607 * store what was left of this slice, if the queue idled out
608 * or was preempted
609 */
610 if (time_after(cfqq->slice_end, now))
611 cfqq->slice_left = cfqq->slice_end - now;
612 else
613 cfqq->slice_left = 0;
614
615 if (cfq_cfqq_on_rr(cfqq))
616 cfq_resort_rr_list(cfqq, preempted);
617
618 if (cfqq == cfqd->active_queue)
619 cfqd->active_queue = NULL;
620
621 if (cfqd->active_cic) {
622 put_io_context(cfqd->active_cic->ioc);
623 cfqd->active_cic = NULL;
624 }
625
626 cfqd->dispatch_slice = 0;
627}
628
629static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
630{
631 struct cfq_queue *cfqq = cfqd->active_queue;
632
633 if (cfqq)
634 __cfq_slice_expired(cfqd, cfqq, preempted);
635}
636
637/*
Jens Axboe22e2c502005-06-27 10:55:12 +0200638 * 0
639 * 0,1
640 * 0,1,2
641 * 0,1,2,3
642 * 0,1,2,3,4
643 * 0,1,2,3,4,5
644 * 0,1,2,3,4,5,6
645 * 0,1,2,3,4,5,6,7
646 */
647static int cfq_get_next_prio_level(struct cfq_data *cfqd)
648{
649 int prio, wrap;
650
651 prio = -1;
652 wrap = 0;
653 do {
654 int p;
655
656 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
657 if (!list_empty(&cfqd->rr_list[p])) {
658 prio = p;
659 break;
660 }
661 }
662
663 if (prio != -1)
664 break;
665 cfqd->cur_prio = 0;
666 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
667 cfqd->cur_end_prio = 0;
668 if (wrap)
669 break;
670 wrap = 1;
671 }
672 } while (1);
673
674 if (unlikely(prio == -1))
675 return -1;
676
677 BUG_ON(prio >= CFQ_PRIO_LISTS);
678
679 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
680
681 cfqd->cur_prio = prio + 1;
682 if (cfqd->cur_prio > cfqd->cur_end_prio) {
683 cfqd->cur_end_prio = cfqd->cur_prio;
684 cfqd->cur_prio = 0;
685 }
686 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
687 cfqd->cur_prio = 0;
688 cfqd->cur_end_prio = 0;
689 }
690
691 return prio;
692}
693
Jens Axboe3b181522005-06-27 10:56:24 +0200694static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200695{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100696 struct cfq_queue *cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +0200697
Jens Axboe89850f72006-07-22 16:48:31 +0200698 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
699 /*
700 * if current list is non-empty, grab first entry. if it is
701 * empty, get next prio level and grab first entry then if any
702 * are spliced
703 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200704 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
Jens Axboe89850f72006-07-22 16:48:31 +0200705 } else if (!list_empty(&cfqd->busy_rr)) {
706 /*
707 * If no new queues are available, check if the busy list has
708 * some before falling back to idle io.
709 */
Jens Axboee0de0202006-06-01 10:07:26 +0200710 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
Jens Axboe89850f72006-07-22 16:48:31 +0200711 } else if (!list_empty(&cfqd->idle_rr)) {
712 /*
713 * if we have idle queues and no rt or be queues had pending
714 * requests, either allow immediate service if the grace period
715 * has passed or arm the idle grace timer
716 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200717 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
718
719 if (time_after_eq(jiffies, end))
720 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
721 else
722 mod_timer(&cfqd->idle_class_timer, end);
723 }
724
725 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200726 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200727}
728
Jens Axboecaaa5f92006-06-16 11:23:00 +0200729#define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
730
Jens Axboe22e2c502005-06-27 10:55:12 +0200731static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
732
733{
Jens Axboe206dc692006-03-28 13:03:44 +0200734 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100735 unsigned long sl;
736
Jens Axboedd67d052006-06-21 09:36:18 +0200737 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +0200738 WARN_ON(cfqq != cfqd->active_queue);
739
740 /*
741 * idle is disabled, either manually or by past process history
742 */
743 if (!cfqd->cfq_slice_idle)
744 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200745 if (!cfq_cfqq_idle_window(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200746 return 0;
747 /*
748 * task has exited, don't wait
749 */
Jens Axboe206dc692006-03-28 13:03:44 +0200750 cic = cfqd->active_cic;
751 if (!cic || !cic->ioc->task)
Jens Axboe22e2c502005-06-27 10:55:12 +0200752 return 0;
753
Jens Axboe3b181522005-06-27 10:56:24 +0200754 cfq_mark_cfqq_must_dispatch(cfqq);
755 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200756
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100757 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
Jens Axboe206dc692006-03-28 13:03:44 +0200758
759 /*
760 * we don't want to idle for seeks, but we do want to allow
761 * fair distribution of slice time for a process doing back-to-back
762 * seeks. so allow a little bit of time for him to submit a new rq
763 */
Jens Axboecaaa5f92006-06-16 11:23:00 +0200764 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboe44eb1232006-07-25 15:05:21 +0200765 sl = min(sl, msecs_to_jiffies(2));
Jens Axboe206dc692006-03-28 13:03:44 +0200766
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100767 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe22e2c502005-06-27 10:55:12 +0200768 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769}
770
Jens Axboe5e705372006-07-13 12:39:25 +0200771static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772{
773 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200774 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200775
Jens Axboe5380a102006-07-13 12:37:56 +0200776 cfq_remove_request(rq);
777 cfqq->on_dispatch[rq_is_sync(rq)]++;
778 elv_dispatch_sort(q, rq);
Jens Axboefd61af02006-06-16 15:35:39 +0200779
780 rq = list_entry(q->queue_head.prev, struct request, queuelist);
781 cfqd->last_sector = rq->sector + rq->nr_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782}
783
784/*
785 * return expired entry, or NULL to just start from scratch in rbtree
786 */
Jens Axboe5e705372006-07-13 12:39:25 +0200787static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788{
789 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200790 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200791 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792
Jens Axboe3b181522005-06-27 10:56:24 +0200793 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 return NULL;
Jens Axboe89850f72006-07-22 16:48:31 +0200795 if (list_empty(&cfqq->fifo))
796 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797
Jens Axboe89850f72006-07-22 16:48:31 +0200798 fifo = cfq_cfqq_class_sync(cfqq);
799 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800
Jens Axboe89850f72006-07-22 16:48:31 +0200801 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
802 cfq_mark_cfqq_fifo_expire(cfqq);
803 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 }
805
806 return NULL;
807}
808
809/*
Jens Axboe3b181522005-06-27 10:56:24 +0200810 * Scale schedule slice based on io priority. Use the sync time slice only
811 * if a queue is marked sync and has sync io queued. A sync queue with async
812 * io only, should not get full sync slice length.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200814static inline int
815cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816{
Jens Axboe22e2c502005-06-27 10:55:12 +0200817 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818
Jens Axboe22e2c502005-06-27 10:55:12 +0200819 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820
Jens Axboe22e2c502005-06-27 10:55:12 +0200821 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822}
823
Jens Axboe22e2c502005-06-27 10:55:12 +0200824static inline void
825cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
826{
827 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
828}
829
830static inline int
831cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
832{
833 const int base_rq = cfqd->cfq_slice_async_rq;
834
835 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
836
837 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
838}
839
840/*
841 * get next queue for service
842 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100843static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200844{
845 unsigned long now = jiffies;
846 struct cfq_queue *cfqq;
847
848 cfqq = cfqd->active_queue;
849 if (!cfqq)
850 goto new_queue;
851
852 /*
853 * slice has expired
854 */
Jens Axboe3b181522005-06-27 10:56:24 +0200855 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
856 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +0200857
858 /*
859 * if queue has requests, dispatch one. if not, check if
860 * enough slice is left to wait for one
861 */
Jens Axboedd67d052006-06-21 09:36:18 +0200862 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200863 goto keep_queue;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200864 else if (cfq_cfqq_dispatched(cfqq)) {
865 cfqq = NULL;
866 goto keep_queue;
867 } else if (cfq_cfqq_class_sync(cfqq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +0200868 if (cfq_arm_slice_timer(cfqd, cfqq))
869 return NULL;
870 }
871
Jens Axboe3b181522005-06-27 10:56:24 +0200872expire:
Jens Axboe22e2c502005-06-27 10:55:12 +0200873 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +0200874new_queue:
875 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200876keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +0200877 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200878}
879
880static int
881__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
882 int max_dispatch)
883{
884 int dispatched = 0;
885
Jens Axboedd67d052006-06-21 09:36:18 +0200886 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +0200887
888 do {
Jens Axboe5e705372006-07-13 12:39:25 +0200889 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200890
891 /*
892 * follow expired path, else get first next available
893 */
Jens Axboe5e705372006-07-13 12:39:25 +0200894 if ((rq = cfq_check_fifo(cfqq)) == NULL)
895 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200896
897 /*
898 * finally, insert request into driver dispatch list
899 */
Jens Axboe5e705372006-07-13 12:39:25 +0200900 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200901
902 cfqd->dispatch_slice++;
903 dispatched++;
904
905 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +0200906 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
907 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200908 }
909
Jens Axboedd67d052006-06-21 09:36:18 +0200910 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200911 break;
912
913 } while (dispatched < max_dispatch);
914
915 /*
Jens Axboecaaa5f92006-06-16 11:23:00 +0200916 * if slice end isn't set yet, set it.
Jens Axboe22e2c502005-06-27 10:55:12 +0200917 */
918 if (!cfqq->slice_end)
919 cfq_set_prio_slice(cfqd, cfqq);
920
921 /*
922 * expire an async queue immediately if it has used up its slice. idle
923 * queue always expire after 1 dispatch round.
924 */
925 if ((!cfq_cfqq_sync(cfqq) &&
926 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboecaaa5f92006-06-16 11:23:00 +0200927 cfq_class_idle(cfqq) ||
928 !cfq_cfqq_idle_window(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +0200929 cfq_slice_expired(cfqd, 0);
930
931 return dispatched;
932}
933
934static int
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100935cfq_forced_dispatch_cfqqs(struct list_head *list)
936{
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100937 struct cfq_queue *cfqq, *next;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200938 int dispatched;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100939
Jens Axboecaaa5f92006-06-16 11:23:00 +0200940 dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100941 list_for_each_entry_safe(cfqq, next, list, cfq_list) {
Jens Axboe5e705372006-07-13 12:39:25 +0200942 while (cfqq->next_rq) {
943 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100944 dispatched++;
945 }
946 BUG_ON(!list_empty(&cfqq->fifo));
947 }
Jens Axboecaaa5f92006-06-16 11:23:00 +0200948
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100949 return dispatched;
950}
951
952static int
953cfq_forced_dispatch(struct cfq_data *cfqd)
954{
955 int i, dispatched = 0;
956
957 for (i = 0; i < CFQ_PRIO_LISTS; i++)
958 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
959
960 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
961 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
962 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
963
964 cfq_slice_expired(cfqd, 0);
965
966 BUG_ON(cfqd->busy_queues);
967
968 return dispatched;
969}
970
971static int
Jens Axboeb4878f22005-10-20 16:42:29 +0200972cfq_dispatch_requests(request_queue_t *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973{
974 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200975 struct cfq_queue *cfqq, *prev_cfqq;
976 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977
Jens Axboe22e2c502005-06-27 10:55:12 +0200978 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979 return 0;
980
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100981 if (unlikely(force))
982 return cfq_forced_dispatch(cfqd);
983
Jens Axboecaaa5f92006-06-16 11:23:00 +0200984 dispatched = 0;
985 prev_cfqq = NULL;
986 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +0200987 int max_dispatch;
988
Jens Axboecaaa5f92006-06-16 11:23:00 +0200989 /*
990 * Don't repeat dispatch from the previous queue.
991 */
992 if (prev_cfqq == cfqq)
993 break;
994
Jens Axboe3b181522005-06-27 10:56:24 +0200995 cfq_clear_cfqq_must_dispatch(cfqq);
996 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200997 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100999 max_dispatch = cfqd->cfq_quantum;
1000 if (cfq_class_idle(cfqq))
1001 max_dispatch = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002
Jens Axboecaaa5f92006-06-16 11:23:00 +02001003 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1004
1005 /*
1006 * If the dispatch cfqq has idling enabled and is still
1007 * the active queue, break out.
1008 */
1009 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1010 break;
1011
1012 prev_cfqq = cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 }
1014
Jens Axboecaaa5f92006-06-16 11:23:00 +02001015 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016}
1017
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018/*
Jens Axboe5e705372006-07-13 12:39:25 +02001019 * task holds one reference to the queue, dropped when task exits. each rq
1020 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 *
1022 * queue lock must be held here.
1023 */
1024static void cfq_put_queue(struct cfq_queue *cfqq)
1025{
Jens Axboe22e2c502005-06-27 10:55:12 +02001026 struct cfq_data *cfqd = cfqq->cfqd;
1027
1028 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029
1030 if (!atomic_dec_and_test(&cfqq->ref))
1031 return;
1032
1033 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001034 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001035 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001037 if (unlikely(cfqd->active_queue == cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001038 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02001039
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 /*
1041 * it's on the empty list and still hashed
1042 */
1043 list_del(&cfqq->cfq_list);
1044 hlist_del(&cfqq->cfq_hash);
1045 kmem_cache_free(cfq_pool, cfqq);
1046}
1047
Jens Axboe1ea25ec2006-07-18 22:24:11 +02001048static struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001049__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1050 const int hashval)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051{
1052 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
Jens Axboe206dc692006-03-28 13:03:44 +02001053 struct hlist_node *entry;
1054 struct cfq_queue *__cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055
Jens Axboe206dc692006-03-28 13:03:44 +02001056 hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
Al Virob0a69162006-03-14 15:32:50 -05001057 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058
Jens Axboe206dc692006-03-28 13:03:44 +02001059 if (__cfqq->key == key && (__p == prio || !prio))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060 return __cfqq;
1061 }
1062
1063 return NULL;
1064}
1065
1066static struct cfq_queue *
Jens Axboe3b181522005-06-27 10:56:24 +02001067cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068{
Jens Axboe3b181522005-06-27 10:56:24 +02001069 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070}
1071
Jens Axboee2d74ac2006-03-28 08:59:01 +02001072static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073{
Jens Axboe22e2c502005-06-27 10:55:12 +02001074 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001075 struct rb_node *n;
1076 int freed = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001077
Jens Axboee2d74ac2006-03-28 08:59:01 +02001078 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1079 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1080 rb_erase(&__cic->rb_node, &ioc->cic_root);
Jens Axboe22e2c502005-06-27 10:55:12 +02001081 kmem_cache_free(cfq_ioc_pool, __cic);
Al Viro334e94d2006-03-18 15:05:53 -05001082 freed++;
Jens Axboe22e2c502005-06-27 10:55:12 +02001083 }
1084
Jens Axboe4050cf12006-07-19 05:07:12 +02001085 elv_ioc_count_mod(ioc_count, -freed);
1086
1087 if (ioc_gone && !elv_ioc_count_read(ioc_count))
Al Viro334e94d2006-03-18 15:05:53 -05001088 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089}
1090
Jens Axboe89850f72006-07-22 16:48:31 +02001091static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1092{
1093 if (unlikely(cfqq == cfqd->active_queue))
1094 __cfq_slice_expired(cfqd, cfqq, 0);
1095
1096 cfq_put_queue(cfqq);
1097}
1098
1099static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1100 struct cfq_io_context *cic)
1101{
Jens Axboefc463792006-08-29 09:05:44 +02001102 list_del_init(&cic->queue_list);
1103 smp_wmb();
1104 cic->key = NULL;
1105
Jens Axboe89850f72006-07-22 16:48:31 +02001106 if (cic->cfqq[ASYNC]) {
1107 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1108 cic->cfqq[ASYNC] = NULL;
1109 }
1110
1111 if (cic->cfqq[SYNC]) {
1112 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1113 cic->cfqq[SYNC] = NULL;
1114 }
Jens Axboe89850f72006-07-22 16:48:31 +02001115}
1116
1117
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001119 * Called with interrupts disabled
1120 */
1121static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1122{
Al Viro478a82b2006-03-18 13:25:24 -05001123 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001124
Jens Axboe89850f72006-07-22 16:48:31 +02001125 if (cfqd) {
1126 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001127
Jens Axboefc463792006-08-29 09:05:44 +02001128 spin_lock_irq(q->queue_lock);
Jens Axboe89850f72006-07-22 16:48:31 +02001129 __cfq_exit_single_io_context(cfqd, cic);
Jens Axboefc463792006-08-29 09:05:44 +02001130 spin_unlock_irq(q->queue_lock);
Al Viro12a05732006-03-18 13:38:01 -05001131 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001132}
1133
Jens Axboee2d74ac2006-03-28 08:59:01 +02001134static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135{
Jens Axboe22e2c502005-06-27 10:55:12 +02001136 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001137 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +02001138
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 /*
1140 * put the reference this task is holding to the various queues
1141 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001142
1143 n = rb_first(&ioc->cic_root);
1144 while (n != NULL) {
1145 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1146
Jens Axboe22e2c502005-06-27 10:55:12 +02001147 cfq_exit_single_io_context(__cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001148 n = rb_next(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150}
1151
Jens Axboe22e2c502005-06-27 10:55:12 +02001152static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001153cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154{
Jens Axboeb5deef92006-07-19 23:39:40 +02001155 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156
Jens Axboeb5deef92006-07-19 23:39:40 +02001157 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 if (cic) {
Jens Axboe553698f2006-06-14 19:11:57 +02001159 memset(cic, 0, sizeof(*cic));
Jens Axboe22e2c502005-06-27 10:55:12 +02001160 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001161 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001162 cic->dtor = cfq_free_io_context;
1163 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001164 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165 }
1166
1167 return cic;
1168}
1169
Jens Axboe22e2c502005-06-27 10:55:12 +02001170static void cfq_init_prio_data(struct cfq_queue *cfqq)
1171{
1172 struct task_struct *tsk = current;
1173 int ioprio_class;
1174
Jens Axboe3b181522005-06-27 10:56:24 +02001175 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001176 return;
1177
1178 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1179 switch (ioprio_class) {
1180 default:
1181 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1182 case IOPRIO_CLASS_NONE:
1183 /*
1184 * no prio set, place us in the middle of the BE classes
1185 */
1186 cfqq->ioprio = task_nice_ioprio(tsk);
1187 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1188 break;
1189 case IOPRIO_CLASS_RT:
1190 cfqq->ioprio = task_ioprio(tsk);
1191 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1192 break;
1193 case IOPRIO_CLASS_BE:
1194 cfqq->ioprio = task_ioprio(tsk);
1195 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1196 break;
1197 case IOPRIO_CLASS_IDLE:
1198 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1199 cfqq->ioprio = 7;
Jens Axboe3b181522005-06-27 10:56:24 +02001200 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001201 break;
1202 }
1203
1204 /*
1205 * keep track of original prio settings in case we have to temporarily
1206 * elevate the priority of this queue
1207 */
1208 cfqq->org_ioprio = cfqq->ioprio;
1209 cfqq->org_ioprio_class = cfqq->ioprio_class;
1210
Jens Axboe3b181522005-06-27 10:56:24 +02001211 if (cfq_cfqq_on_rr(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001212 cfq_resort_rr_list(cfqq, 0);
1213
Jens Axboe3b181522005-06-27 10:56:24 +02001214 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001215}
1216
Al Viro478a82b2006-03-18 13:25:24 -05001217static inline void changed_ioprio(struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001218{
Al Viro478a82b2006-03-18 13:25:24 -05001219 struct cfq_data *cfqd = cic->key;
1220 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001221 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001222
Jens Axboecaaa5f92006-06-16 11:23:00 +02001223 if (unlikely(!cfqd))
1224 return;
1225
Jens Axboec1b707d2006-10-30 19:54:23 +01001226 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001227
1228 cfqq = cic->cfqq[ASYNC];
1229 if (cfqq) {
1230 struct cfq_queue *new_cfqq;
1231 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1232 GFP_ATOMIC);
1233 if (new_cfqq) {
1234 cic->cfqq[ASYNC] = new_cfqq;
1235 cfq_put_queue(cfqq);
1236 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001237 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001238
1239 cfqq = cic->cfqq[SYNC];
1240 if (cfqq)
1241 cfq_mark_cfqq_prio_changed(cfqq);
1242
Jens Axboec1b707d2006-10-30 19:54:23 +01001243 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001244}
1245
Jens Axboefc463792006-08-29 09:05:44 +02001246static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247{
Al Viroa6a07632006-03-18 13:26:44 -05001248 struct cfq_io_context *cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001249 struct rb_node *n;
Al Viroa6a07632006-03-18 13:26:44 -05001250
Jens Axboefc463792006-08-29 09:05:44 +02001251 ioc->ioprio_changed = 0;
Al Viroa6a07632006-03-18 13:26:44 -05001252
Jens Axboee2d74ac2006-03-28 08:59:01 +02001253 n = rb_first(&ioc->cic_root);
1254 while (n != NULL) {
1255 cic = rb_entry(n, struct cfq_io_context, rb_node);
Jens Axboe3793c652006-05-30 21:11:04 +02001256
Al Viro478a82b2006-03-18 13:25:24 -05001257 changed_ioprio(cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001258 n = rb_next(n);
1259 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260}
1261
1262static struct cfq_queue *
Al Viro6f325a12006-03-18 14:58:37 -05001263cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
Al Viro8267e262005-10-21 03:20:53 -04001264 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265{
1266 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1267 struct cfq_queue *cfqq, *new_cfqq = NULL;
Al Viro6f325a12006-03-18 14:58:37 -05001268 unsigned short ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269
1270retry:
Al Viro6f325a12006-03-18 14:58:37 -05001271 ioprio = tsk->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02001272 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273
1274 if (!cfqq) {
1275 if (new_cfqq) {
1276 cfqq = new_cfqq;
1277 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001278 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001279 /*
1280 * Inform the allocator of the fact that we will
1281 * just repeat this allocation if it fails, to allow
1282 * the allocator to do whatever it needs to attempt to
1283 * free memory.
1284 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 spin_unlock_irq(cfqd->queue->queue_lock);
Jens Axboeb5deef92006-07-19 23:39:40 +02001286 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 spin_lock_irq(cfqd->queue->queue_lock);
1288 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001289 } else {
Jens Axboeb5deef92006-07-19 23:39:40 +02001290 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001291 if (!cfqq)
1292 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001293 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294
1295 memset(cfqq, 0, sizeof(*cfqq));
1296
1297 INIT_HLIST_NODE(&cfqq->cfq_hash);
1298 INIT_LIST_HEAD(&cfqq->cfq_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001299 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300
1301 cfqq->key = key;
1302 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1303 atomic_set(&cfqq->ref, 0);
1304 cfqq->cfqd = cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +02001305 /*
1306 * set ->slice_left to allow preemption for a new process
1307 */
1308 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001309 cfq_mark_cfqq_idle_window(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001310 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b03742006-07-28 09:48:51 +02001311 cfq_mark_cfqq_queue_new(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001312 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313 }
1314
1315 if (new_cfqq)
1316 kmem_cache_free(cfq_pool, new_cfqq);
1317
1318 atomic_inc(&cfqq->ref);
1319out:
1320 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1321 return cfqq;
1322}
1323
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001324static void
1325cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1326{
Jens Axboefc463792006-08-29 09:05:44 +02001327 WARN_ON(!list_empty(&cic->queue_list));
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001328 rb_erase(&cic->rb_node, &ioc->cic_root);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001329 kmem_cache_free(cfq_ioc_pool, cic);
Jens Axboe4050cf12006-07-19 05:07:12 +02001330 elv_ioc_count_dec(ioc_count);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001331}
1332
Jens Axboee2d74ac2006-03-28 08:59:01 +02001333static struct cfq_io_context *
1334cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1335{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001336 struct rb_node *n;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001337 struct cfq_io_context *cic;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001338 void *k, *key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001339
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001340restart:
1341 n = ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001342 while (n) {
1343 cic = rb_entry(n, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001344 /* ->key must be copied to avoid race with cfq_exit_queue() */
1345 k = cic->key;
1346 if (unlikely(!k)) {
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001347 cfq_drop_dead_cic(ioc, cic);
1348 goto restart;
1349 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001350
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001351 if (key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001352 n = n->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001353 else if (key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001354 n = n->rb_right;
1355 else
1356 return cic;
1357 }
1358
1359 return NULL;
1360}
1361
1362static inline void
1363cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1364 struct cfq_io_context *cic)
1365{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001366 struct rb_node **p;
1367 struct rb_node *parent;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001368 struct cfq_io_context *__cic;
Jens Axboe0261d682006-10-30 19:07:48 +01001369 unsigned long flags;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001370 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001371
Jens Axboee2d74ac2006-03-28 08:59:01 +02001372 cic->ioc = ioc;
1373 cic->key = cfqd;
1374
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001375restart:
1376 parent = NULL;
1377 p = &ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001378 while (*p) {
1379 parent = *p;
1380 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001381 /* ->key must be copied to avoid race with cfq_exit_queue() */
1382 k = __cic->key;
1383 if (unlikely(!k)) {
Oleg Nesterovbe33c3a2006-08-21 08:36:12 +02001384 cfq_drop_dead_cic(ioc, __cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001385 goto restart;
1386 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001387
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001388 if (cic->key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001389 p = &(*p)->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001390 else if (cic->key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001391 p = &(*p)->rb_right;
1392 else
1393 BUG();
1394 }
1395
1396 rb_link_node(&cic->rb_node, parent, p);
1397 rb_insert_color(&cic->rb_node, &ioc->cic_root);
Jens Axboefc463792006-08-29 09:05:44 +02001398
Jens Axboe0261d682006-10-30 19:07:48 +01001399 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001400 list_add(&cic->queue_list, &cfqd->cic_list);
Jens Axboe0261d682006-10-30 19:07:48 +01001401 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001402}
1403
Jens Axboe22e2c502005-06-27 10:55:12 +02001404/*
1405 * Setup general io context and cfq io context. There can be several cfq
1406 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001407 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001408 */
1409static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001410cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411{
Jens Axboe22e2c502005-06-27 10:55:12 +02001412 struct io_context *ioc = NULL;
1413 struct cfq_io_context *cic;
1414
1415 might_sleep_if(gfp_mask & __GFP_WAIT);
1416
Jens Axboeb5deef92006-07-19 23:39:40 +02001417 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001418 if (!ioc)
1419 return NULL;
1420
Jens Axboee2d74ac2006-03-28 08:59:01 +02001421 cic = cfq_cic_rb_lookup(cfqd, ioc);
1422 if (cic)
1423 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001424
Jens Axboee2d74ac2006-03-28 08:59:01 +02001425 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1426 if (cic == NULL)
1427 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001428
Jens Axboee2d74ac2006-03-28 08:59:01 +02001429 cfq_cic_link(cfqd, ioc, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001430out:
Jens Axboefc463792006-08-29 09:05:44 +02001431 smp_read_barrier_depends();
1432 if (unlikely(ioc->ioprio_changed))
1433 cfq_ioc_set_ioprio(ioc);
1434
Jens Axboe22e2c502005-06-27 10:55:12 +02001435 return cic;
1436err:
1437 put_io_context(ioc);
1438 return NULL;
1439}
1440
1441static void
1442cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1443{
1444 unsigned long elapsed, ttime;
1445
1446 /*
1447 * if this context already has stuff queued, thinktime is from
1448 * last queue not last end
1449 */
1450#if 0
1451 if (time_after(cic->last_end_request, cic->last_queue))
1452 elapsed = jiffies - cic->last_end_request;
1453 else
1454 elapsed = jiffies - cic->last_queue;
1455#else
1456 elapsed = jiffies - cic->last_end_request;
1457#endif
1458
1459 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1460
1461 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1462 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1463 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1464}
1465
Jens Axboe206dc692006-03-28 13:03:44 +02001466static void
Jens Axboebb37b942006-12-01 10:42:33 +01001467cfq_update_io_seektime(struct cfq_io_context *cic, struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001468{
1469 sector_t sdist;
1470 u64 total;
1471
Jens Axboe5e705372006-07-13 12:39:25 +02001472 if (cic->last_request_pos < rq->sector)
1473 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001474 else
Jens Axboe5e705372006-07-13 12:39:25 +02001475 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001476
1477 /*
1478 * Don't allow the seek distance to get too large from the
1479 * odd fragment, pagein, etc
1480 */
1481 if (cic->seek_samples <= 60) /* second&third seek */
1482 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1483 else
1484 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1485
1486 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1487 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1488 total = cic->seek_total + (cic->seek_samples/2);
1489 do_div(total, cic->seek_samples);
1490 cic->seek_mean = (sector_t)total;
1491}
Jens Axboe22e2c502005-06-27 10:55:12 +02001492
1493/*
1494 * Disable idle window if the process thinks too long or seeks so much that
1495 * it doesn't matter
1496 */
1497static void
1498cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1499 struct cfq_io_context *cic)
1500{
Jens Axboe3b181522005-06-27 10:56:24 +02001501 int enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001502
Jens Axboecaaa5f92006-06-16 11:23:00 +02001503 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1504 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001505 enable_idle = 0;
1506 else if (sample_valid(cic->ttime_samples)) {
1507 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1508 enable_idle = 0;
1509 else
1510 enable_idle = 1;
1511 }
1512
Jens Axboe3b181522005-06-27 10:56:24 +02001513 if (enable_idle)
1514 cfq_mark_cfqq_idle_window(cfqq);
1515 else
1516 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001517}
1518
1519
1520/*
1521 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1522 * no or if we aren't sure, a 1 will cause a preempt.
1523 */
1524static int
1525cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001526 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001527{
1528 struct cfq_queue *cfqq = cfqd->active_queue;
1529
1530 if (cfq_class_idle(new_cfqq))
1531 return 0;
1532
1533 if (!cfqq)
Jens Axboecaaa5f92006-06-16 11:23:00 +02001534 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001535
1536 if (cfq_class_idle(cfqq))
1537 return 1;
Jens Axboe3b181522005-06-27 10:56:24 +02001538 if (!cfq_cfqq_wait_request(new_cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001539 return 0;
1540 /*
1541 * if it doesn't have slice left, forget it
1542 */
1543 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1544 return 0;
Jens Axboe374f84a2006-07-23 01:42:19 +02001545 /*
1546 * if the new request is sync, but the currently running queue is
1547 * not, let the sync request have priority.
1548 */
Jens Axboe5e705372006-07-13 12:39:25 +02001549 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001550 return 1;
Jens Axboe374f84a2006-07-23 01:42:19 +02001551 /*
1552 * So both queues are sync. Let the new request get disk time if
1553 * it's a metadata request and the current queue is doing regular IO.
1554 */
1555 if (rq_is_meta(rq) && !cfqq->meta_pending)
1556 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001557
1558 return 0;
1559}
1560
1561/*
1562 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1563 * let it have half of its nominal slice.
1564 */
1565static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1566{
Jens Axboebf572252006-07-19 20:29:12 +02001567 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001568
1569 if (!cfqq->slice_left)
1570 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1571
Jens Axboebf572252006-07-19 20:29:12 +02001572 /*
1573 * Put the new queue at the front of the of the current list,
1574 * so we know that it will be selected next.
1575 */
1576 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1577 list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1578
Jens Axboe22e2c502005-06-27 10:55:12 +02001579 cfqq->slice_end = cfqq->slice_left + jiffies;
Jens Axboe22e2c502005-06-27 10:55:12 +02001580}
1581
1582/*
Jens Axboe5e705372006-07-13 12:39:25 +02001583 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001584 * something we should do about it
1585 */
1586static void
Jens Axboe5e705372006-07-13 12:39:25 +02001587cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1588 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001589{
Jens Axboe5e705372006-07-13 12:39:25 +02001590 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001591
Jens Axboe374f84a2006-07-23 01:42:19 +02001592 if (rq_is_meta(rq))
1593 cfqq->meta_pending++;
1594
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001595 /*
Jens Axboe5380a102006-07-13 12:37:56 +02001596 * check if this request is a better next-serve candidate)) {
Jens Axboe21183b02006-07-13 12:33:14 +02001597 */
Jens Axboe5e705372006-07-13 12:39:25 +02001598 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1599 BUG_ON(!cfqq->next_rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001600
1601 /*
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001602 * we never wait for an async request and we don't allow preemption
1603 * of an async request. so just return early
1604 */
Jens Axboe5e705372006-07-13 12:39:25 +02001605 if (!rq_is_sync(rq)) {
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001606 /*
1607 * sync process issued an async request, if it's waiting
1608 * then expire it and kick rq handling.
1609 */
1610 if (cic == cfqd->active_cic &&
1611 del_timer(&cfqd->idle_slice_timer)) {
1612 cfq_slice_expired(cfqd, 0);
Jens Axboedc72ef42006-07-20 14:54:05 +02001613 blk_start_queueing(cfqd->queue);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001614 }
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001615 return;
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001616 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001617
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001618 cfq_update_io_thinktime(cfqd, cic);
Jens Axboebb37b942006-12-01 10:42:33 +01001619 cfq_update_io_seektime(cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001620 cfq_update_idle_window(cfqd, cfqq, cic);
1621
1622 cic->last_queue = jiffies;
Jens Axboe5e705372006-07-13 12:39:25 +02001623 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001624
1625 if (cfqq == cfqd->active_queue) {
1626 /*
1627 * if we are waiting for a request for this queue, let it rip
1628 * immediately and flag that we must not expire this queue
1629 * just now
1630 */
Jens Axboe3b181522005-06-27 10:56:24 +02001631 if (cfq_cfqq_wait_request(cfqq)) {
1632 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001633 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001634 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001635 }
Jens Axboe5e705372006-07-13 12:39:25 +02001636 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001637 /*
1638 * not the active queue - expire current slice if it is
1639 * idle and has expired it's mean thinktime or this new queue
1640 * has some old slice time left and is of higher priority
1641 */
1642 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001643 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001644 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001645 }
1646}
1647
Jens Axboeb4878f22005-10-20 16:42:29 +02001648static void cfq_insert_request(request_queue_t *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001649{
Jens Axboeb4878f22005-10-20 16:42:29 +02001650 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001651 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001652
1653 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654
Jens Axboe5e705372006-07-13 12:39:25 +02001655 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656
Jens Axboe22e2c502005-06-27 10:55:12 +02001657 list_add_tail(&rq->queuelist, &cfqq->fifo);
1658
Jens Axboe5e705372006-07-13 12:39:25 +02001659 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660}
1661
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662static void cfq_completed_request(request_queue_t *q, struct request *rq)
1663{
Jens Axboe5e705372006-07-13 12:39:25 +02001664 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001665 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001666 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001667 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668
Jens Axboeb4878f22005-10-20 16:42:29 +02001669 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670
Jens Axboeb4878f22005-10-20 16:42:29 +02001671 WARN_ON(!cfqd->rq_in_driver);
1672 WARN_ON(!cfqq->on_dispatch[sync]);
1673 cfqd->rq_in_driver--;
1674 cfqq->on_dispatch[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675
Jens Axboeb4878f22005-10-20 16:42:29 +02001676 if (!cfq_class_idle(cfqq))
1677 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001678
Jens Axboe53b03742006-07-28 09:48:51 +02001679 if (!cfq_cfqq_dispatched(cfqq) && cfq_cfqq_on_rr(cfqq))
1680 cfq_resort_rr_list(cfqq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681
Jens Axboecaaa5f92006-06-16 11:23:00 +02001682 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001683 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001684
1685 /*
1686 * If this is the active queue, check if it needs to be expired,
1687 * or if we want to idle in case it has no pending requests.
1688 */
1689 if (cfqd->active_queue == cfqq) {
1690 if (time_after(now, cfqq->slice_end))
1691 cfq_slice_expired(cfqd, 0);
Jens Axboedd67d052006-06-21 09:36:18 +02001692 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001693 if (!cfq_arm_slice_timer(cfqd, cfqq))
1694 cfq_schedule_dispatch(cfqd);
1695 }
1696 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697}
1698
Jens Axboe22e2c502005-06-27 10:55:12 +02001699/*
1700 * we temporarily boost lower priority queues if they are holding fs exclusive
1701 * resources. they are boosted to normal prio (CLASS_BE/4)
1702 */
1703static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704{
Jens Axboe22e2c502005-06-27 10:55:12 +02001705 const int ioprio_class = cfqq->ioprio_class;
1706 const int ioprio = cfqq->ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707
Jens Axboe22e2c502005-06-27 10:55:12 +02001708 if (has_fs_excl()) {
1709 /*
1710 * boost idle prio on transactions that would lock out other
1711 * users of the filesystem
1712 */
1713 if (cfq_class_idle(cfqq))
1714 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1715 if (cfqq->ioprio > IOPRIO_NORM)
1716 cfqq->ioprio = IOPRIO_NORM;
1717 } else {
1718 /*
1719 * check if we need to unboost the queue
1720 */
1721 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1722 cfqq->ioprio_class = cfqq->org_ioprio_class;
1723 if (cfqq->ioprio != cfqq->org_ioprio)
1724 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 }
1726
Jens Axboe22e2c502005-06-27 10:55:12 +02001727 /*
1728 * refile between round-robin lists if we moved the priority class
1729 */
1730 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
Jens Axboe3b181522005-06-27 10:56:24 +02001731 cfq_cfqq_on_rr(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001732 cfq_resort_rr_list(cfqq, 0);
1733}
1734
Jens Axboe89850f72006-07-22 16:48:31 +02001735static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001736{
Jens Axboe3b181522005-06-27 10:56:24 +02001737 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001738 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001739 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001740 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001741 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001742
1743 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02001744}
1745
Jens Axboecb78b282006-07-28 09:32:57 +02001746static int cfq_may_queue(request_queue_t *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02001747{
1748 struct cfq_data *cfqd = q->elevator->elevator_data;
1749 struct task_struct *tsk = current;
1750 struct cfq_queue *cfqq;
1751
1752 /*
1753 * don't force setup of a queue from here, as a call to may_queue
1754 * does not necessarily imply that a request actually will be queued.
1755 * so just lookup a possibly existing queue, or return 'may queue'
1756 * if that fails
1757 */
Jens Axboe3b181522005-06-27 10:56:24 +02001758 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001759 if (cfqq) {
1760 cfq_init_prio_data(cfqq);
1761 cfq_prio_boost(cfqq);
1762
Jens Axboe89850f72006-07-22 16:48:31 +02001763 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001764 }
1765
1766 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767}
1768
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769/*
1770 * queue lock held here
1771 */
Jens Axboebb37b942006-12-01 10:42:33 +01001772static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773{
Jens Axboe5e705372006-07-13 12:39:25 +02001774 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775
Jens Axboe5e705372006-07-13 12:39:25 +02001776 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001777 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778
Jens Axboe22e2c502005-06-27 10:55:12 +02001779 BUG_ON(!cfqq->allocated[rw]);
1780 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781
Jens Axboe5e705372006-07-13 12:39:25 +02001782 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02001785 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 cfq_put_queue(cfqq);
1788 }
1789}
1790
1791/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001792 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001794static int
Jens Axboecb78b282006-07-28 09:32:57 +02001795cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796{
1797 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe3b181522005-06-27 10:56:24 +02001798 struct task_struct *tsk = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 struct cfq_io_context *cic;
1800 const int rw = rq_data_dir(rq);
Jens Axboe3b181522005-06-27 10:56:24 +02001801 pid_t key = cfq_queue_pid(tsk, rw);
Jens Axboe22e2c502005-06-27 10:55:12 +02001802 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 unsigned long flags;
Al Viro12a05732006-03-18 13:38:01 -05001804 int is_sync = key != CFQ_KEY_ASYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805
1806 might_sleep_if(gfp_mask & __GFP_WAIT);
1807
Jens Axboee2d74ac2006-03-28 08:59:01 +02001808 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001809
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 spin_lock_irqsave(q->queue_lock, flags);
1811
Jens Axboe22e2c502005-06-27 10:55:12 +02001812 if (!cic)
1813 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814
Al Viro12a05732006-03-18 13:38:01 -05001815 if (!cic->cfqq[is_sync]) {
Al Viro6f325a12006-03-18 14:58:37 -05001816 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001817 if (!cfqq)
1818 goto queue_fail;
1819
Al Viro12a05732006-03-18 13:38:01 -05001820 cic->cfqq[is_sync] = cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001821 } else
Al Viro12a05732006-03-18 13:38:01 -05001822 cfqq = cic->cfqq[is_sync];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823
1824 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02001825 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001826 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02001827
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 spin_unlock_irqrestore(q->queue_lock, flags);
1829
Jens Axboe5e705372006-07-13 12:39:25 +02001830 rq->elevator_private = cic;
1831 rq->elevator_private2 = cfqq;
1832 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02001833
Jens Axboe22e2c502005-06-27 10:55:12 +02001834queue_fail:
1835 if (cic)
1836 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02001837
Jens Axboe3b181522005-06-27 10:56:24 +02001838 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839 spin_unlock_irqrestore(q->queue_lock, flags);
1840 return 1;
1841}
1842
David Howells65f27f32006-11-22 14:55:48 +00001843static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02001844{
David Howells65f27f32006-11-22 14:55:48 +00001845 struct cfq_data *cfqd =
1846 container_of(work, struct cfq_data, unplug_work);
1847 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001848 unsigned long flags;
1849
1850 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02001851 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02001852 spin_unlock_irqrestore(q->queue_lock, flags);
1853}
1854
1855/*
1856 * Timer running if the active_queue is currently idling inside its time slice
1857 */
1858static void cfq_idle_slice_timer(unsigned long data)
1859{
1860 struct cfq_data *cfqd = (struct cfq_data *) data;
1861 struct cfq_queue *cfqq;
1862 unsigned long flags;
1863
1864 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1865
1866 if ((cfqq = cfqd->active_queue) != NULL) {
1867 unsigned long now = jiffies;
1868
1869 /*
1870 * expired
1871 */
1872 if (time_after(now, cfqq->slice_end))
1873 goto expire;
1874
1875 /*
1876 * only expire and reinvoke request handler, if there are
1877 * other queues with pending requests
1878 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02001879 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02001880 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02001881
1882 /*
1883 * not expired and it has a request pending, let it dispatch
1884 */
Jens Axboedd67d052006-06-21 09:36:18 +02001885 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001886 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001887 goto out_kick;
1888 }
1889 }
1890expire:
1891 cfq_slice_expired(cfqd, 0);
1892out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02001893 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001894out_cont:
1895 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1896}
1897
1898/*
1899 * Timer running if an idle class queue is waiting for service
1900 */
1901static void cfq_idle_class_timer(unsigned long data)
1902{
1903 struct cfq_data *cfqd = (struct cfq_data *) data;
1904 unsigned long flags, end;
1905
1906 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1907
1908 /*
1909 * race with a non-idle queue, reset timer
1910 */
1911 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
Jens Axboeae818a32006-06-01 10:13:43 +02001912 if (!time_after_eq(jiffies, end))
1913 mod_timer(&cfqd->idle_class_timer, end);
1914 else
Jens Axboe3b181522005-06-27 10:56:24 +02001915 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001916
1917 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1918}
1919
Jens Axboe3b181522005-06-27 10:56:24 +02001920static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1921{
1922 del_timer_sync(&cfqd->idle_slice_timer);
1923 del_timer_sync(&cfqd->idle_class_timer);
1924 blk_sync_queue(cfqd->queue);
1925}
Jens Axboe22e2c502005-06-27 10:55:12 +02001926
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927static void cfq_exit_queue(elevator_t *e)
1928{
Jens Axboe22e2c502005-06-27 10:55:12 +02001929 struct cfq_data *cfqd = e->elevator_data;
Al Virod9ff4182006-03-18 13:51:22 -05001930 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001931
Jens Axboe3b181522005-06-27 10:56:24 +02001932 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001933
Al Virod9ff4182006-03-18 13:51:22 -05001934 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001935
Al Virod9ff4182006-03-18 13:51:22 -05001936 if (cfqd->active_queue)
1937 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001938
1939 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05001940 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1941 struct cfq_io_context,
1942 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02001943
1944 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05001945 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001946
Al Virod9ff4182006-03-18 13:51:22 -05001947 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05001948
1949 cfq_shutdown_timer_wq(cfqd);
1950
Al Viroa90d7422006-03-18 12:05:37 -05001951 kfree(cfqd->cfq_hash);
1952 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953}
1954
Jens Axboebb37b942006-12-01 10:42:33 +01001955static void *cfq_init_queue(request_queue_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956{
1957 struct cfq_data *cfqd;
1958 int i;
1959
Jens Axboeb5deef92006-07-19 23:39:40 +02001960 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02001962 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963
1964 memset(cfqd, 0, sizeof(*cfqd));
Jens Axboe22e2c502005-06-27 10:55:12 +02001965
1966 for (i = 0; i < CFQ_PRIO_LISTS; i++)
1967 INIT_LIST_HEAD(&cfqd->rr_list[i]);
1968
1969 INIT_LIST_HEAD(&cfqd->busy_rr);
1970 INIT_LIST_HEAD(&cfqd->cur_rr);
1971 INIT_LIST_HEAD(&cfqd->idle_rr);
Al Virod9ff4182006-03-18 13:51:22 -05001972 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973
Jens Axboeb5deef92006-07-19 23:39:40 +02001974 cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 if (!cfqd->cfq_hash)
Jens Axboe5e705372006-07-13 12:39:25 +02001976 goto out_free;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977
Linus Torvalds1da177e2005-04-16 15:20:36 -07001978 for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
1979 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
1980
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982
Jens Axboe22e2c502005-06-27 10:55:12 +02001983 init_timer(&cfqd->idle_slice_timer);
1984 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
1985 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
1986
1987 init_timer(&cfqd->idle_class_timer);
1988 cfqd->idle_class_timer.function = cfq_idle_class_timer;
1989 cfqd->idle_class_timer.data = (unsigned long) cfqd;
1990
David Howells65f27f32006-11-22 14:55:48 +00001991 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001992
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02001994 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
1995 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 cfqd->cfq_back_max = cfq_back_max;
1997 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02001998 cfqd->cfq_slice[0] = cfq_slice_async;
1999 cfqd->cfq_slice[1] = cfq_slice_sync;
2000 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2001 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02002002
Jens Axboebc1c1162006-06-08 08:49:06 +02002003 return cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +02002004out_free:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005 kfree(cfqd);
Jens Axboebc1c1162006-06-08 08:49:06 +02002006 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007}
2008
2009static void cfq_slab_kill(void)
2010{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011 if (cfq_pool)
2012 kmem_cache_destroy(cfq_pool);
2013 if (cfq_ioc_pool)
2014 kmem_cache_destroy(cfq_ioc_pool);
2015}
2016
2017static int __init cfq_slab_setup(void)
2018{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2020 NULL, NULL);
2021 if (!cfq_pool)
2022 goto fail;
2023
2024 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2025 sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2026 if (!cfq_ioc_pool)
2027 goto fail;
2028
2029 return 0;
2030fail:
2031 cfq_slab_kill();
2032 return -ENOMEM;
2033}
2034
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035/*
2036 * sysfs parts below -->
2037 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038
2039static ssize_t
2040cfq_var_show(unsigned int var, char *page)
2041{
2042 return sprintf(page, "%d\n", var);
2043}
2044
2045static ssize_t
2046cfq_var_store(unsigned int *var, const char *page, size_t count)
2047{
2048 char *p = (char *) page;
2049
2050 *var = simple_strtoul(p, &p, 10);
2051 return count;
2052}
2053
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002055static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002057 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 unsigned int __data = __VAR; \
2059 if (__CONV) \
2060 __data = jiffies_to_msecs(__data); \
2061 return cfq_var_show(__data, (page)); \
2062}
2063SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002064SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2065SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002066SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2067SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002068SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2069SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2070SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2071SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072#undef SHOW_FUNCTION
2073
2074#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002075static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002077 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078 unsigned int __data; \
2079 int ret = cfq_var_store(&__data, (page), count); \
2080 if (__data < (MIN)) \
2081 __data = (MIN); \
2082 else if (__data > (MAX)) \
2083 __data = (MAX); \
2084 if (__CONV) \
2085 *(__PTR) = msecs_to_jiffies(__data); \
2086 else \
2087 *(__PTR) = __data; \
2088 return ret; \
2089}
2090STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002091STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2092STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002093STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2094STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002095STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2096STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2097STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2098STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099#undef STORE_FUNCTION
2100
Al Viroe572ec72006-03-18 22:27:18 -05002101#define CFQ_ATTR(name) \
2102 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002103
Al Viroe572ec72006-03-18 22:27:18 -05002104static struct elv_fs_entry cfq_attrs[] = {
2105 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002106 CFQ_ATTR(fifo_expire_sync),
2107 CFQ_ATTR(fifo_expire_async),
2108 CFQ_ATTR(back_seek_max),
2109 CFQ_ATTR(back_seek_penalty),
2110 CFQ_ATTR(slice_sync),
2111 CFQ_ATTR(slice_async),
2112 CFQ_ATTR(slice_async_rq),
2113 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002114 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115};
2116
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117static struct elevator_type iosched_cfq = {
2118 .ops = {
2119 .elevator_merge_fn = cfq_merge,
2120 .elevator_merged_fn = cfq_merged_request,
2121 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeb4878f22005-10-20 16:42:29 +02002122 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002124 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125 .elevator_deactivate_req_fn = cfq_deactivate_request,
2126 .elevator_queue_empty_fn = cfq_queue_empty,
2127 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002128 .elevator_former_req_fn = elv_rb_former_request,
2129 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130 .elevator_set_req_fn = cfq_set_request,
2131 .elevator_put_req_fn = cfq_put_request,
2132 .elevator_may_queue_fn = cfq_may_queue,
2133 .elevator_init_fn = cfq_init_queue,
2134 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002135 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136 },
Al Viro3d1ab402006-03-18 18:35:43 -05002137 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 .elevator_name = "cfq",
2139 .elevator_owner = THIS_MODULE,
2140};
2141
2142static int __init cfq_init(void)
2143{
2144 int ret;
2145
Jens Axboe22e2c502005-06-27 10:55:12 +02002146 /*
2147 * could be 0 on HZ < 1000 setups
2148 */
2149 if (!cfq_slice_async)
2150 cfq_slice_async = 1;
2151 if (!cfq_slice_idle)
2152 cfq_slice_idle = 1;
2153
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154 if (cfq_slab_setup())
2155 return -ENOMEM;
2156
2157 ret = elv_register(&iosched_cfq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002158 if (ret)
2159 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161 return ret;
2162}
2163
2164static void __exit cfq_exit(void)
2165{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002166 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002168 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002169 /* ioc_gone's update must be visible before reading ioc_count */
2170 smp_wmb();
Jens Axboe4050cf12006-07-19 05:07:12 +02002171 if (elv_ioc_count_read(ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002172 wait_for_completion(ioc_gone);
Al Viro334e94d2006-03-18 15:05:53 -05002173 synchronize_rcu();
Christoph Hellwig83521d32005-10-30 15:01:39 -08002174 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175}
2176
2177module_init(cfq_init);
2178module_exit(cfq_exit);
2179
2180MODULE_AUTHOR("Jens Axboe");
2181MODULE_LICENSE("GPL");
2182MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");