blob: 54dc05439009d09968572e6f12ce30d0fef151e8 [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/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020013#include <linux/ioprio.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014
15/*
16 * tunables
17 */
Arjan van de Ven64100092006-01-06 09:46:02 +010018static const int cfq_quantum = 4; /* max queue in one round of service */
Arjan van de Ven64100092006-01-06 09:46:02 +010019static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
20static const int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */
21static const int cfq_back_penalty = 2; /* penalty of a backwards seek */
Linus Torvalds1da177e2005-04-16 15:20:36 -070022
Arjan van de Ven64100092006-01-06 09:46:02 +010023static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020024static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010025static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020026static int cfq_slice_idle = HZ / 125;
Jens Axboe22e2c502005-06-27 10:55:12 +020027
Jens Axboed9e76202007-04-20 14:27:50 +020028/*
29 * grace period before allowing idle class to get disk access
30 */
Jens Axboe22e2c502005-06-27 10:55:12 +020031#define CFQ_IDLE_GRACE (HZ / 10)
Jens Axboed9e76202007-04-20 14:27:50 +020032
33/*
34 * below this threshold, we consider thinktime immediate
35 */
36#define CFQ_MIN_TT (2)
37
Jens Axboe22e2c502005-06-27 10:55:12 +020038#define CFQ_SLICE_SCALE (5)
39
Jens Axboe5e705372006-07-13 12:39:25 +020040#define RQ_CIC(rq) ((struct cfq_io_context*)(rq)->elevator_private)
41#define RQ_CFQQ(rq) ((rq)->elevator_private2)
Linus Torvalds1da177e2005-04-16 15:20:36 -070042
Christoph Lametere18b8902006-12-06 20:33:20 -080043static struct kmem_cache *cfq_pool;
44static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
Jens Axboe4050cf12006-07-19 05:07:12 +020046static DEFINE_PER_CPU(unsigned long, ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050047static struct completion *ioc_gone;
48
Jens Axboe22e2c502005-06-27 10:55:12 +020049#define CFQ_PRIO_LISTS IOPRIO_BE_NR
50#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020051#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
52
Jens Axboe3b181522005-06-27 10:56:24 +020053#define ASYNC (0)
54#define SYNC (1)
55
Jens Axboe206dc692006-03-28 13:03:44 +020056#define sample_valid(samples) ((samples) > 80)
57
Jens Axboe22e2c502005-06-27 10:55:12 +020058/*
Jens Axboecc09e292007-04-26 12:53:50 +020059 * Most of our rbtree usage is for sorting with min extraction, so
60 * if we cache the leftmost node we don't have to walk down the tree
61 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
62 * move this into the elevator for the rq sorting as well.
63 */
64struct cfq_rb_root {
65 struct rb_root rb;
66 struct rb_node *left;
67};
68#define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
69
70/*
Jens Axboe22e2c502005-06-27 10:55:12 +020071 * Per block device queue structure
72 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070073struct cfq_data {
Jens Axboe165125e2007-07-24 09:28:11 +020074 struct request_queue *queue;
Jens Axboe22e2c502005-06-27 10:55:12 +020075
76 /*
77 * rr list of queues with requests and the count of them
78 */
Jens Axboecc09e292007-04-26 12:53:50 +020079 struct cfq_rb_root service_tree;
Jens Axboe22e2c502005-06-27 10:55:12 +020080 unsigned int busy_queues;
81
Jens Axboe22e2c502005-06-27 10:55:12 +020082 int rq_in_driver;
Jens Axboe3ed9a292007-04-23 08:33:33 +020083 int sync_flight;
Jens Axboe25776e32006-06-01 10:12:26 +020084 int hw_tag;
Jens Axboe22e2c502005-06-27 10:55:12 +020085
86 /*
Jens Axboe22e2c502005-06-27 10:55:12 +020087 * idle window management
88 */
89 struct timer_list idle_slice_timer;
90 struct work_struct unplug_work;
91
92 struct cfq_queue *active_queue;
93 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +020094
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +020095 /*
96 * async queue for each priority case
97 */
98 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
99 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200100
Jens Axboe22e2c502005-06-27 10:55:12 +0200101 struct timer_list idle_class_timer;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102
Jens Axboe6d048f52007-04-25 12:44:27 +0200103 sector_t last_position;
Jens Axboe22e2c502005-06-27 10:55:12 +0200104 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106 /*
107 * tunables, see top of file
108 */
109 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200110 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111 unsigned int cfq_back_penalty;
112 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200113 unsigned int cfq_slice[2];
114 unsigned int cfq_slice_async_rq;
115 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500116
117 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118};
119
Jens Axboe22e2c502005-06-27 10:55:12 +0200120/*
121 * Per process-grouping structure
122 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123struct cfq_queue {
124 /* reference count */
125 atomic_t ref;
126 /* parent cfq_data */
127 struct cfq_data *cfqd;
Jens Axboed9e76202007-04-20 14:27:50 +0200128 /* service_tree member */
129 struct rb_node rb_node;
130 /* service_tree key */
131 unsigned long rb_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 /* sorted list of pending requests */
133 struct rb_root sort_list;
134 /* if fifo isn't expired, next request to serve */
Jens Axboe5e705372006-07-13 12:39:25 +0200135 struct request *next_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136 /* requests queued in sort_list */
137 int queued[2];
138 /* currently allocated requests */
139 int allocated[2];
Jens Axboe374f84a2006-07-23 01:42:19 +0200140 /* pending metadata requests */
141 int meta_pending;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200143 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144
Jens Axboe22e2c502005-06-27 10:55:12 +0200145 unsigned long slice_end;
Jens Axboec5b680f2007-01-19 11:56:49 +1100146 long slice_resid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
Jens Axboe6d048f52007-04-25 12:44:27 +0200148 /* number of requests that are on the dispatch list or inside driver */
149 int dispatched;
Jens Axboe22e2c502005-06-27 10:55:12 +0200150
151 /* io prio of this group */
152 unsigned short ioprio, org_ioprio;
153 unsigned short ioprio_class, org_ioprio_class;
154
Jens Axboe3b181522005-06-27 10:56:24 +0200155 /* various state flags, see below */
156 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157};
158
Jens Axboe3b181522005-06-27 10:56:24 +0200159enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100160 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
161 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
162 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
163 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
164 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
165 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
166 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
167 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
168 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
Jens Axboe44f7c162007-01-19 11:51:58 +1100169 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200170 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200171};
172
173#define CFQ_CFQQ_FNS(name) \
174static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
175{ \
176 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
177} \
178static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
179{ \
180 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
181} \
182static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
183{ \
184 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
185}
186
187CFQ_CFQQ_FNS(on_rr);
188CFQ_CFQQ_FNS(wait_request);
189CFQ_CFQQ_FNS(must_alloc);
190CFQ_CFQQ_FNS(must_alloc_slice);
191CFQ_CFQQ_FNS(must_dispatch);
192CFQ_CFQQ_FNS(fifo_expire);
193CFQ_CFQQ_FNS(idle_window);
194CFQ_CFQQ_FNS(prio_changed);
Jens Axboe53b03742006-07-28 09:48:51 +0200195CFQ_CFQQ_FNS(queue_new);
Jens Axboe44f7c162007-01-19 11:51:58 +1100196CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200197CFQ_CFQQ_FNS(sync);
Jens Axboe3b181522005-06-27 10:56:24 +0200198#undef CFQ_CFQQ_FNS
199
Jens Axboe165125e2007-07-24 09:28:11 +0200200static void cfq_dispatch_insert(struct request_queue *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200201static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
202 struct task_struct *, gfp_t);
203static struct cfq_io_context *cfq_cic_rb_lookup(struct cfq_data *,
204 struct io_context *);
205
206static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
207 int is_sync)
208{
209 return cic->cfqq[!!is_sync];
210}
211
212static inline void cic_set_cfqq(struct cfq_io_context *cic,
213 struct cfq_queue *cfqq, int is_sync)
214{
215 cic->cfqq[!!is_sync] = cfqq;
216}
217
218/*
219 * We regard a request as SYNC, if it's either a read or has the SYNC bit
220 * set (in which case it could also be direct WRITE).
221 */
222static inline int cfq_bio_sync(struct bio *bio)
223{
224 if (bio_data_dir(bio) == READ || bio_sync(bio))
225 return 1;
226
227 return 0;
228}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700231 * scheduler run of queue, if there are requests pending and no one in the
232 * driver that will restart queueing
233 */
234static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
235{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100236 if (cfqd->busy_queues)
Andrew Morton99f95e52005-06-27 20:14:05 -0700237 kblockd_schedule_work(&cfqd->unplug_work);
238}
239
Jens Axboe165125e2007-07-24 09:28:11 +0200240static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700241{
242 struct cfq_data *cfqd = q->elevator->elevator_data;
243
Jens Axboeb4878f22005-10-20 16:42:29 +0200244 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700245}
246
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100248 * Scale schedule slice based on io priority. Use the sync time slice only
249 * if a queue is marked sync and has sync io queued. A sync queue with async
250 * io only, should not get full sync slice length.
251 */
Jens Axboed9e76202007-04-20 14:27:50 +0200252static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
253 unsigned short prio)
254{
255 const int base_slice = cfqd->cfq_slice[sync];
256
257 WARN_ON(prio >= IOPRIO_BE_NR);
258
259 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
260}
261
Jens Axboe44f7c162007-01-19 11:51:58 +1100262static inline int
263cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
264{
Jens Axboed9e76202007-04-20 14:27:50 +0200265 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100266}
267
268static inline void
269cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
270{
271 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
272}
273
274/*
275 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
276 * isn't valid until the first request from the dispatch is activated
277 * and the slice time set.
278 */
279static inline int cfq_slice_used(struct cfq_queue *cfqq)
280{
281 if (cfq_cfqq_slice_new(cfqq))
282 return 0;
283 if (time_before(jiffies, cfqq->slice_end))
284 return 0;
285
286 return 1;
287}
288
289/*
Jens Axboe5e705372006-07-13 12:39:25 +0200290 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200292 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293 */
Jens Axboe5e705372006-07-13 12:39:25 +0200294static struct request *
295cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296{
297 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200299#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
300#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
301 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302
Jens Axboe5e705372006-07-13 12:39:25 +0200303 if (rq1 == NULL || rq1 == rq2)
304 return rq2;
305 if (rq2 == NULL)
306 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200307
Jens Axboe5e705372006-07-13 12:39:25 +0200308 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
309 return rq1;
310 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
311 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200312 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
313 return rq1;
314 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
315 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316
Jens Axboe5e705372006-07-13 12:39:25 +0200317 s1 = rq1->sector;
318 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319
Jens Axboe6d048f52007-04-25 12:44:27 +0200320 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322 /*
323 * by definition, 1KiB is 2 sectors
324 */
325 back_max = cfqd->cfq_back_max * 2;
326
327 /*
328 * Strict one way elevator _except_ in the case where we allow
329 * short backward seeks which are biased as twice the cost of a
330 * similar forward seek.
331 */
332 if (s1 >= last)
333 d1 = s1 - last;
334 else if (s1 + back_max >= last)
335 d1 = (last - s1) * cfqd->cfq_back_penalty;
336 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200337 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338
339 if (s2 >= last)
340 d2 = s2 - last;
341 else if (s2 + back_max >= last)
342 d2 = (last - s2) * cfqd->cfq_back_penalty;
343 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200344 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345
346 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347
Andreas Mohre8a99052006-03-28 08:59:49 +0200348 /*
349 * By doing switch() on the bit mask "wrap" we avoid having to
350 * check two variables for all permutations: --> faster!
351 */
352 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200353 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200354 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200355 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200356 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200357 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200358 else {
359 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200360 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200361 else
Jens Axboe5e705372006-07-13 12:39:25 +0200362 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200363 }
364
365 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200366 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200367 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200368 return rq2;
369 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200370 default:
371 /*
372 * Since both rqs are wrapped,
373 * start with the one that's further behind head
374 * (--> only *one* back seek required),
375 * since back seek takes more time than forward.
376 */
377 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200378 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 else
Jens Axboe5e705372006-07-13 12:39:25 +0200380 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 }
382}
383
Jens Axboe498d3aa2007-04-26 12:54:48 +0200384/*
385 * The below is leftmost cache rbtree addon
386 */
Jens Axboecc09e292007-04-26 12:53:50 +0200387static struct rb_node *cfq_rb_first(struct cfq_rb_root *root)
388{
389 if (!root->left)
390 root->left = rb_first(&root->rb);
391
392 return root->left;
393}
394
395static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
396{
397 if (root->left == n)
398 root->left = NULL;
399
400 rb_erase(n, &root->rb);
401 RB_CLEAR_NODE(n);
402}
403
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404/*
405 * would be nice to take fifo expire time into account as well
406 */
Jens Axboe5e705372006-07-13 12:39:25 +0200407static struct request *
408cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
409 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410{
Jens Axboe21183b02006-07-13 12:33:14 +0200411 struct rb_node *rbnext = rb_next(&last->rb_node);
412 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200413 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414
Jens Axboe21183b02006-07-13 12:33:14 +0200415 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
417 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200418 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200419
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200421 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200422 else {
423 rbnext = rb_first(&cfqq->sort_list);
424 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200425 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200426 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427
Jens Axboe21183b02006-07-13 12:33:14 +0200428 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429}
430
Jens Axboed9e76202007-04-20 14:27:50 +0200431static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
432 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433{
Jens Axboed9e76202007-04-20 14:27:50 +0200434 /*
435 * just an approximation, should be ok.
436 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200437 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
438 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200439}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440
Jens Axboe498d3aa2007-04-26 12:54:48 +0200441/*
442 * The cfqd->service_tree holds all pending cfq_queue's that have
443 * requests waiting to be processed. It is sorted in the order that
444 * we will service the queues.
445 */
Jens Axboed9e76202007-04-20 14:27:50 +0200446static void cfq_service_tree_add(struct cfq_data *cfqd,
Jens Axboeedd75ff2007-04-19 12:03:34 +0200447 struct cfq_queue *cfqq, int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200448{
Jens Axboecc09e292007-04-26 12:53:50 +0200449 struct rb_node **p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200450 struct rb_node *parent = NULL;
Jens Axboed9e76202007-04-20 14:27:50 +0200451 unsigned long rb_key;
Jens Axboe498d3aa2007-04-26 12:54:48 +0200452 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200453
Jens Axboeedd75ff2007-04-19 12:03:34 +0200454 if (!add_front) {
455 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
456 rb_key += cfqq->slice_resid;
457 cfqq->slice_resid = 0;
458 } else
459 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200460
461 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100462 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200463 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100464 */
Jens Axboed9e76202007-04-20 14:27:50 +0200465 if (rb_key == cfqq->rb_key)
466 return;
Jens Axboe53b03742006-07-28 09:48:51 +0200467
Jens Axboecc09e292007-04-26 12:53:50 +0200468 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200469 }
Jens Axboed9e76202007-04-20 14:27:50 +0200470
Jens Axboe498d3aa2007-04-26 12:54:48 +0200471 left = 1;
Jens Axboed9e76202007-04-20 14:27:50 +0200472 while (*p) {
Jens Axboecc09e292007-04-26 12:53:50 +0200473 struct cfq_queue *__cfqq;
Jens Axboe67060e32007-04-18 20:13:32 +0200474 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200475
Jens Axboed9e76202007-04-20 14:27:50 +0200476 parent = *p;
477 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
478
Jens Axboe0c534e02007-04-18 20:01:57 +0200479 /*
480 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200481 * preference to them. IDLE queues goes to the back.
482 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200483 */
484 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200485 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200486 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200487 n = &(*p)->rb_right;
488 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
489 n = &(*p)->rb_left;
490 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
491 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200492 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200493 n = &(*p)->rb_left;
494 else
495 n = &(*p)->rb_right;
496
497 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200498 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200499
500 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200501 }
502
Jens Axboecc09e292007-04-26 12:53:50 +0200503 if (left)
504 cfqd->service_tree.left = &cfqq->rb_node;
505
Jens Axboed9e76202007-04-20 14:27:50 +0200506 cfqq->rb_key = rb_key;
507 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200508 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509}
510
Jens Axboe498d3aa2007-04-26 12:54:48 +0200511/*
512 * Update cfqq's position in the service tree.
513 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200514static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200515{
Jens Axboe6d048f52007-04-25 12:44:27 +0200516 /*
517 * Resorting requires the cfqq to be on the RR list already.
518 */
Jens Axboe498d3aa2007-04-26 12:54:48 +0200519 if (cfq_cfqq_on_rr(cfqq))
Jens Axboeedd75ff2007-04-19 12:03:34 +0200520 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboe6d048f52007-04-25 12:44:27 +0200521}
522
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523/*
524 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200525 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 */
527static inline void
Jens Axboeb4878f22005-10-20 16:42:29 +0200528cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
Jens Axboe3b181522005-06-27 10:56:24 +0200530 BUG_ON(cfq_cfqq_on_rr(cfqq));
531 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 cfqd->busy_queues++;
533
Jens Axboeedd75ff2007-04-19 12:03:34 +0200534 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535}
536
Jens Axboe498d3aa2007-04-26 12:54:48 +0200537/*
538 * Called when the cfqq no longer has requests pending, remove it from
539 * the service tree.
540 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541static inline void
542cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
543{
Jens Axboe3b181522005-06-27 10:56:24 +0200544 BUG_ON(!cfq_cfqq_on_rr(cfqq));
545 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546
Jens Axboecc09e292007-04-26 12:53:50 +0200547 if (!RB_EMPTY_NODE(&cfqq->rb_node))
548 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboed9e76202007-04-20 14:27:50 +0200549
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 BUG_ON(!cfqd->busy_queues);
551 cfqd->busy_queues--;
552}
553
554/*
555 * rb tree support functions
556 */
Jens Axboe5e705372006-07-13 12:39:25 +0200557static inline void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558{
Jens Axboe5e705372006-07-13 12:39:25 +0200559 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200560 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200561 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562
Jens Axboeb4878f22005-10-20 16:42:29 +0200563 BUG_ON(!cfqq->queued[sync]);
564 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565
Jens Axboe5e705372006-07-13 12:39:25 +0200566 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567
Jens Axboedd67d052006-06-21 09:36:18 +0200568 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200569 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570}
571
Jens Axboe5e705372006-07-13 12:39:25 +0200572static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573{
Jens Axboe5e705372006-07-13 12:39:25 +0200574 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe21183b02006-07-13 12:33:14 +0200576 struct request *__alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577
Jens Axboe5380a102006-07-13 12:37:56 +0200578 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579
580 /*
581 * looks a little odd, but the first insert might return an alias.
582 * if that happens, put the alias on the dispatch list
583 */
Jens Axboe21183b02006-07-13 12:33:14 +0200584 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200585 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100586
587 if (!cfq_cfqq_on_rr(cfqq))
588 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200589
590 /*
591 * check if this request is a better next-serve candidate
592 */
593 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
594 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595}
596
597static inline void
Jens Axboe5e705372006-07-13 12:39:25 +0200598cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599{
Jens Axboe5380a102006-07-13 12:37:56 +0200600 elv_rb_del(&cfqq->sort_list, rq);
601 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200602 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603}
604
Jens Axboe206dc692006-03-28 13:03:44 +0200605static struct request *
606cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607{
Jens Axboe206dc692006-03-28 13:03:44 +0200608 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200609 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200610 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611
Vasily Tarasov91fac312007-04-25 12:29:51 +0200612 cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
613 if (!cic)
614 return NULL;
615
616 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200617 if (cfqq) {
618 sector_t sector = bio->bi_sector + bio_sectors(bio);
619
Jens Axboe21183b02006-07-13 12:33:14 +0200620 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200621 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623 return NULL;
624}
625
Jens Axboe165125e2007-07-24 09:28:11 +0200626static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200627{
628 struct cfq_data *cfqd = q->elevator->elevator_data;
629
630 cfqd->rq_in_driver++;
Jens Axboe25776e32006-06-01 10:12:26 +0200631
632 /*
633 * If the depth is larger 1, it really could be queueing. But lets
634 * make the mark a little higher - idling could still be good for
635 * low queueing, and a low queueing number could also just indicate
636 * a SCSI mid layer like behaviour where limit+1 is often seen.
637 */
638 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
639 cfqd->hw_tag = 1;
Jens Axboe6d048f52007-04-25 12:44:27 +0200640
641 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
Jens Axboeb4878f22005-10-20 16:42:29 +0200642}
643
Jens Axboe165125e2007-07-24 09:28:11 +0200644static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645{
Jens Axboe22e2c502005-06-27 10:55:12 +0200646 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647
Jens Axboeb4878f22005-10-20 16:42:29 +0200648 WARN_ON(!cfqd->rq_in_driver);
649 cfqd->rq_in_driver--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650}
651
Jens Axboeb4878f22005-10-20 16:42:29 +0200652static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653{
Jens Axboe5e705372006-07-13 12:39:25 +0200654 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200655
Jens Axboe5e705372006-07-13 12:39:25 +0200656 if (cfqq->next_rq == rq)
657 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658
Jens Axboeb4878f22005-10-20 16:42:29 +0200659 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200660 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200661
662 if (rq_is_meta(rq)) {
663 WARN_ON(!cfqq->meta_pending);
664 cfqq->meta_pending--;
665 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666}
667
Jens Axboe165125e2007-07-24 09:28:11 +0200668static int cfq_merge(struct request_queue *q, struct request **req,
669 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670{
671 struct cfq_data *cfqd = q->elevator->elevator_data;
672 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673
Jens Axboe206dc692006-03-28 13:03:44 +0200674 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200675 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200676 *req = __rq;
677 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 }
679
680 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681}
682
Jens Axboe165125e2007-07-24 09:28:11 +0200683static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200684 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685{
Jens Axboe21183b02006-07-13 12:33:14 +0200686 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200687 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688
Jens Axboe5e705372006-07-13 12:39:25 +0200689 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691}
692
693static void
Jens Axboe165125e2007-07-24 09:28:11 +0200694cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 struct request *next)
696{
Jens Axboe22e2c502005-06-27 10:55:12 +0200697 /*
698 * reposition in fifo if next is older than rq
699 */
700 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
701 time_before(next->start_time, rq->start_time))
702 list_move(&rq->queuelist, &next->queuelist);
703
Jens Axboeb4878f22005-10-20 16:42:29 +0200704 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200705}
706
Jens Axboe165125e2007-07-24 09:28:11 +0200707static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100708 struct bio *bio)
709{
710 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200711 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100712 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100713
714 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100715 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100716 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200717 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100718 return 0;
719
720 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100721 * Lookup the cfqq that this bio will be queued with. Allow
722 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100723 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200724 cic = cfq_cic_rb_lookup(cfqd, current->io_context);
725 if (!cic)
726 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100727
Vasily Tarasov91fac312007-04-25 12:29:51 +0200728 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100729 if (cfqq == RQ_CFQQ(rq))
730 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100731
Jens Axboeec8acb62007-01-02 18:32:11 +0100732 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100733}
734
Jens Axboe22e2c502005-06-27 10:55:12 +0200735static inline void
736__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
737{
738 if (cfqq) {
739 /*
740 * stop potential idle class queues waiting service
741 */
742 del_timer(&cfqd->idle_class_timer);
743
Jens Axboe22e2c502005-06-27 10:55:12 +0200744 cfqq->slice_end = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200745 cfq_clear_cfqq_must_alloc_slice(cfqq);
746 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100747 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe1afba042007-04-17 12:47:55 +0200748 cfq_clear_cfqq_queue_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200749 }
750
751 cfqd->active_queue = cfqq;
752}
753
754/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100755 * current cfqq expired its slice (or was too idle), select new one
756 */
757static void
758__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200759 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100760{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100761 if (cfq_cfqq_wait_request(cfqq))
762 del_timer(&cfqd->idle_slice_timer);
763
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100764 cfq_clear_cfqq_must_dispatch(cfqq);
765 cfq_clear_cfqq_wait_request(cfqq);
766
767 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200768 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100769 */
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100770 if (timed_out && !cfq_cfqq_slice_new(cfqq))
Jens Axboec5b680f2007-01-19 11:56:49 +1100771 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100772
Jens Axboeedd75ff2007-04-19 12:03:34 +0200773 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100774
775 if (cfqq == cfqd->active_queue)
776 cfqd->active_queue = NULL;
777
778 if (cfqd->active_cic) {
779 put_io_context(cfqd->active_cic->ioc);
780 cfqd->active_cic = NULL;
781 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100782}
783
Jens Axboe6084cdd2007-04-23 08:25:00 +0200784static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100785{
786 struct cfq_queue *cfqq = cfqd->active_queue;
787
788 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200789 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100790}
791
Jens Axboe498d3aa2007-04-26 12:54:48 +0200792/*
793 * Get next queue for service. Unless we have a queue preemption,
794 * we'll simply select the first cfqq in the service tree.
795 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200796static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200797{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200798 struct cfq_queue *cfqq;
799 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +0200800
Jens Axboeedd75ff2007-04-19 12:03:34 +0200801 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
802 return NULL;
803
804 n = cfq_rb_first(&cfqd->service_tree);
805 cfqq = rb_entry(n, struct cfq_queue, rb_node);
806
807 if (cfq_class_idle(cfqq)) {
808 unsigned long end;
809
Jens Axboe89850f72006-07-22 16:48:31 +0200810 /*
Jens Axboeedd75ff2007-04-19 12:03:34 +0200811 * if we have idle queues and no rt or be queues had
812 * pending requests, either allow immediate service if
813 * the grace period has passed or arm the idle grace
814 * timer
Jens Axboe89850f72006-07-22 16:48:31 +0200815 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200816 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
817 if (time_before(jiffies, end)) {
818 mod_timer(&cfqd->idle_class_timer, end);
819 cfqq = NULL;
Jens Axboe67060e32007-04-18 20:13:32 +0200820 }
Jens Axboe22e2c502005-06-27 10:55:12 +0200821 }
822
Jens Axboe6d048f52007-04-25 12:44:27 +0200823 return cfqq;
824}
825
Jens Axboe498d3aa2007-04-26 12:54:48 +0200826/*
827 * Get and set a new active queue for service.
828 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200829static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
830{
831 struct cfq_queue *cfqq;
832
Jens Axboed9e76202007-04-20 14:27:50 +0200833 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200834 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200835 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200836}
837
Jens Axboed9e76202007-04-20 14:27:50 +0200838static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
839 struct request *rq)
840{
841 if (rq->sector >= cfqd->last_position)
842 return rq->sector - cfqd->last_position;
843 else
844 return cfqd->last_position - rq->sector;
845}
846
Jens Axboe6d048f52007-04-25 12:44:27 +0200847static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
848{
849 struct cfq_io_context *cic = cfqd->active_cic;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200850
Jens Axboe6d048f52007-04-25 12:44:27 +0200851 if (!sample_valid(cic->seek_samples))
852 return 0;
853
854 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
855}
856
Jens Axboed9e76202007-04-20 14:27:50 +0200857static int cfq_close_cooperator(struct cfq_data *cfq_data,
858 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200859{
Jens Axboe6d048f52007-04-25 12:44:27 +0200860 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200861 * We should notice if some of the queues are cooperating, eg
862 * working closely on the same area of the disk. In that case,
863 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +0200864 */
Jens Axboed9e76202007-04-20 14:27:50 +0200865 return 0;
Jens Axboe6d048f52007-04-25 12:44:27 +0200866}
867
868#define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
869
870static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200871{
Jens Axboe17926692007-01-19 11:59:30 +1100872 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +0200873 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100874 unsigned long sl;
875
Jens Axboedd67d052006-06-21 09:36:18 +0200876 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +0200877 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +0200878
879 /*
880 * idle is disabled, either manually or by past process history
881 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200882 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
883 return;
884
Jens Axboe22e2c502005-06-27 10:55:12 +0200885 /*
886 * task has exited, don't wait
887 */
Jens Axboe206dc692006-03-28 13:03:44 +0200888 cic = cfqd->active_cic;
889 if (!cic || !cic->ioc->task)
Jens Axboe6d048f52007-04-25 12:44:27 +0200890 return;
891
892 /*
893 * See if this prio level has a good candidate
894 */
Jens Axboe1afba042007-04-17 12:47:55 +0200895 if (cfq_close_cooperator(cfqd, cfqq) &&
896 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
Jens Axboe6d048f52007-04-25 12:44:27 +0200897 return;
Jens Axboe22e2c502005-06-27 10:55:12 +0200898
Jens Axboe3b181522005-06-27 10:56:24 +0200899 cfq_mark_cfqq_must_dispatch(cfqq);
900 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200901
Jens Axboe206dc692006-03-28 13:03:44 +0200902 /*
903 * we don't want to idle for seeks, but we do want to allow
904 * fair distribution of slice time for a process doing back-to-back
905 * seeks. so allow a little bit of time for him to submit a new rq
906 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200907 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200908 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +0200909 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +0200910
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100911 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912}
913
Jens Axboe498d3aa2007-04-26 12:54:48 +0200914/*
915 * Move request from internal lists to the request queue dispatch list.
916 */
Jens Axboe165125e2007-07-24 09:28:11 +0200917static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918{
Jens Axboe3ed9a292007-04-23 08:33:33 +0200919 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200920 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200921
Jens Axboe5380a102006-07-13 12:37:56 +0200922 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200923 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +0200924 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +0200925
926 if (cfq_cfqq_sync(cfqq))
927 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928}
929
930/*
931 * return expired entry, or NULL to just start from scratch in rbtree
932 */
Jens Axboe5e705372006-07-13 12:39:25 +0200933static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934{
935 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200936 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200937 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938
Jens Axboe3b181522005-06-27 10:56:24 +0200939 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +1100941
942 cfq_mark_cfqq_fifo_expire(cfqq);
943
Jens Axboe89850f72006-07-22 16:48:31 +0200944 if (list_empty(&cfqq->fifo))
945 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946
Jens Axboe6d048f52007-04-25 12:44:27 +0200947 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +0200948 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949
Jens Axboe6d048f52007-04-25 12:44:27 +0200950 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
951 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952
Jens Axboe6d048f52007-04-25 12:44:27 +0200953 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954}
955
Jens Axboe22e2c502005-06-27 10:55:12 +0200956static inline int
957cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
958{
959 const int base_rq = cfqd->cfq_slice_async_rq;
960
961 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
962
963 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
964}
965
966/*
Jens Axboe498d3aa2007-04-26 12:54:48 +0200967 * Select a queue for service. If we have a current active queue,
968 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +0200969 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100970static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200971{
Jens Axboe22e2c502005-06-27 10:55:12 +0200972 struct cfq_queue *cfqq;
973
974 cfqq = cfqd->active_queue;
975 if (!cfqq)
976 goto new_queue;
977
978 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200979 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +0200980 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200981 if (cfq_slice_used(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +0200982 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +0200983
984 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200985 * The active queue has requests and isn't expired, allow it to
986 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +0200987 */
Jens Axboedd67d052006-06-21 09:36:18 +0200988 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200989 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +0200990
991 /*
992 * No requests pending. If the active queue still has requests in
993 * flight or is idling for a new request, allow either of these
994 * conditions to happen (or time out) before selecting a new queue.
995 */
Jens Axboecc197472007-04-20 20:45:39 +0200996 if (timer_pending(&cfqd->idle_slice_timer) ||
997 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +0200998 cfqq = NULL;
999 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001000 }
1001
Jens Axboe3b181522005-06-27 10:56:24 +02001002expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001003 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001004new_queue:
1005 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001006keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001007 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001008}
1009
Jens Axboe498d3aa2007-04-26 12:54:48 +02001010/*
1011 * Dispatch some requests from cfqq, moving them to the request queue
1012 * dispatch list.
1013 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001014static int
1015__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1016 int max_dispatch)
1017{
1018 int dispatched = 0;
1019
Jens Axboedd67d052006-06-21 09:36:18 +02001020 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001021
1022 do {
Jens Axboe5e705372006-07-13 12:39:25 +02001023 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001024
1025 /*
1026 * follow expired path, else get first next available
1027 */
Jens Axboe5e705372006-07-13 12:39:25 +02001028 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1029 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001030
1031 /*
1032 * finally, insert request into driver dispatch list
1033 */
Jens Axboe5e705372006-07-13 12:39:25 +02001034 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001035
Jens Axboe22e2c502005-06-27 10:55:12 +02001036 dispatched++;
1037
1038 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +02001039 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1040 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001041 }
1042
Jens Axboedd67d052006-06-21 09:36:18 +02001043 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001044 break;
1045
1046 } while (dispatched < max_dispatch);
1047
1048 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001049 * expire an async queue immediately if it has used up its slice. idle
1050 * queue always expire after 1 dispatch round.
1051 */
Jens Axboea9938002007-04-20 08:55:52 +02001052 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
Jens Axboe20e493a2007-04-23 08:26:36 +02001053 dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboea9938002007-04-20 08:55:52 +02001054 cfq_class_idle(cfqq))) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001055 cfqq->slice_end = jiffies + 1;
Jens Axboe6084cdd2007-04-23 08:25:00 +02001056 cfq_slice_expired(cfqd, 0);
Jens Axboe44f7c162007-01-19 11:51:58 +11001057 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001058
1059 return dispatched;
1060}
1061
Jens Axboed9e76202007-04-20 14:27:50 +02001062static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1063{
1064 int dispatched = 0;
1065
1066 while (cfqq->next_rq) {
1067 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1068 dispatched++;
1069 }
1070
1071 BUG_ON(!list_empty(&cfqq->fifo));
1072 return dispatched;
1073}
1074
Jens Axboe498d3aa2007-04-26 12:54:48 +02001075/*
1076 * Drain our current requests. Used for barriers and when switching
1077 * io schedulers on-the-fly.
1078 */
Jens Axboed9e76202007-04-20 14:27:50 +02001079static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001080{
Jens Axboed9e76202007-04-20 14:27:50 +02001081 int dispatched = 0;
1082 struct rb_node *n;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001083
Jens Axboecc09e292007-04-26 12:53:50 +02001084 while ((n = cfq_rb_first(&cfqd->service_tree)) != NULL) {
Jens Axboed9e76202007-04-20 14:27:50 +02001085 struct cfq_queue *cfqq = rb_entry(n, struct cfq_queue, rb_node);
1086
1087 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1088 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001089
Jens Axboe6084cdd2007-04-23 08:25:00 +02001090 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001091
1092 BUG_ON(cfqd->busy_queues);
1093
1094 return dispatched;
1095}
1096
Jens Axboe165125e2007-07-24 09:28:11 +02001097static int cfq_dispatch_requests(struct request_queue *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098{
1099 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001100 struct cfq_queue *cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001101 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102
Jens Axboe22e2c502005-06-27 10:55:12 +02001103 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104 return 0;
1105
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001106 if (unlikely(force))
1107 return cfq_forced_dispatch(cfqd);
1108
Jens Axboecaaa5f92006-06-16 11:23:00 +02001109 dispatched = 0;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001110 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001111 int max_dispatch;
1112
Jens Axboe3ed9a292007-04-23 08:33:33 +02001113 max_dispatch = cfqd->cfq_quantum;
1114 if (cfq_class_idle(cfqq))
1115 max_dispatch = 1;
1116
1117 if (cfqq->dispatched >= max_dispatch) {
1118 if (cfqd->busy_queues > 1)
Jens Axboe6d048f52007-04-25 12:44:27 +02001119 break;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001120 if (cfqq->dispatched >= 4 * max_dispatch)
Jens Axboea9938002007-04-20 08:55:52 +02001121 break;
1122 }
Jens Axboe9ede2092007-01-19 12:11:44 +11001123
Jens Axboe3ed9a292007-04-23 08:33:33 +02001124 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1125 break;
1126
Jens Axboe3b181522005-06-27 10:56:24 +02001127 cfq_clear_cfqq_must_dispatch(cfqq);
1128 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001129 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
Jens Axboecaaa5f92006-06-16 11:23:00 +02001131 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 }
1133
Jens Axboecaaa5f92006-06-16 11:23:00 +02001134 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135}
1136
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137/*
Jens Axboe5e705372006-07-13 12:39:25 +02001138 * task holds one reference to the queue, dropped when task exits. each rq
1139 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 *
1141 * queue lock must be held here.
1142 */
1143static void cfq_put_queue(struct cfq_queue *cfqq)
1144{
Jens Axboe22e2c502005-06-27 10:55:12 +02001145 struct cfq_data *cfqd = cfqq->cfqd;
1146
1147 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148
1149 if (!atomic_dec_and_test(&cfqq->ref))
1150 return;
1151
1152 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001153 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001154 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001156 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001157 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001158 cfq_schedule_dispatch(cfqd);
1159 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001160
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 kmem_cache_free(cfq_pool, cfqq);
1162}
1163
Jens Axboee2d74ac2006-03-28 08:59:01 +02001164static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165{
Jens Axboe22e2c502005-06-27 10:55:12 +02001166 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001167 struct rb_node *n;
1168 int freed = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001169
Jens Axboe597bc482007-04-24 21:23:53 +02001170 ioc->ioc_data = NULL;
1171
Jens Axboee2d74ac2006-03-28 08:59:01 +02001172 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1173 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1174 rb_erase(&__cic->rb_node, &ioc->cic_root);
Jens Axboe22e2c502005-06-27 10:55:12 +02001175 kmem_cache_free(cfq_ioc_pool, __cic);
Al Viro334e94d2006-03-18 15:05:53 -05001176 freed++;
Jens Axboe22e2c502005-06-27 10:55:12 +02001177 }
1178
Jens Axboe4050cf12006-07-19 05:07:12 +02001179 elv_ioc_count_mod(ioc_count, -freed);
1180
1181 if (ioc_gone && !elv_ioc_count_read(ioc_count))
Al Viro334e94d2006-03-18 15:05:53 -05001182 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183}
1184
Jens Axboe89850f72006-07-22 16:48:31 +02001185static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1186{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001187 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001188 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001189 cfq_schedule_dispatch(cfqd);
1190 }
Jens Axboe89850f72006-07-22 16:48:31 +02001191
1192 cfq_put_queue(cfqq);
1193}
1194
1195static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1196 struct cfq_io_context *cic)
1197{
Jens Axboefc463792006-08-29 09:05:44 +02001198 list_del_init(&cic->queue_list);
1199 smp_wmb();
1200 cic->key = NULL;
1201
Jens Axboe89850f72006-07-22 16:48:31 +02001202 if (cic->cfqq[ASYNC]) {
1203 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1204 cic->cfqq[ASYNC] = NULL;
1205 }
1206
1207 if (cic->cfqq[SYNC]) {
1208 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1209 cic->cfqq[SYNC] = NULL;
1210 }
Jens Axboe89850f72006-07-22 16:48:31 +02001211}
1212
Jens Axboe22e2c502005-06-27 10:55:12 +02001213static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1214{
Al Viro478a82b2006-03-18 13:25:24 -05001215 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001216
Jens Axboe89850f72006-07-22 16:48:31 +02001217 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001218 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001219
Jens Axboefc463792006-08-29 09:05:44 +02001220 spin_lock_irq(q->queue_lock);
Jens Axboe89850f72006-07-22 16:48:31 +02001221 __cfq_exit_single_io_context(cfqd, cic);
Jens Axboefc463792006-08-29 09:05:44 +02001222 spin_unlock_irq(q->queue_lock);
Al Viro12a05732006-03-18 13:38:01 -05001223 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001224}
1225
Jens Axboe498d3aa2007-04-26 12:54:48 +02001226/*
1227 * The process that ioc belongs to has exited, we need to clean up
1228 * and put the internal structures we have that belongs to that process.
1229 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001230static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231{
Jens Axboe22e2c502005-06-27 10:55:12 +02001232 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001233 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +02001234
Jens Axboe597bc482007-04-24 21:23:53 +02001235 ioc->ioc_data = NULL;
1236
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 /*
1238 * put the reference this task is holding to the various queues
1239 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001240 n = rb_first(&ioc->cic_root);
1241 while (n != NULL) {
1242 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1243
Jens Axboe22e2c502005-06-27 10:55:12 +02001244 cfq_exit_single_io_context(__cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001245 n = rb_next(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247}
1248
Jens Axboe22e2c502005-06-27 10:55:12 +02001249static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001250cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251{
Jens Axboeb5deef92006-07-19 23:39:40 +02001252 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253
Christoph Lameter94f60302007-07-17 04:03:29 -07001254 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1255 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001257 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001258 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001259 cic->dtor = cfq_free_io_context;
1260 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001261 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262 }
1263
1264 return cic;
1265}
1266
Jens Axboe22e2c502005-06-27 10:55:12 +02001267static void cfq_init_prio_data(struct cfq_queue *cfqq)
1268{
1269 struct task_struct *tsk = current;
1270 int ioprio_class;
1271
Jens Axboe3b181522005-06-27 10:56:24 +02001272 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001273 return;
1274
1275 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1276 switch (ioprio_class) {
1277 default:
1278 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1279 case IOPRIO_CLASS_NONE:
1280 /*
1281 * no prio set, place us in the middle of the BE classes
1282 */
1283 cfqq->ioprio = task_nice_ioprio(tsk);
1284 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1285 break;
1286 case IOPRIO_CLASS_RT:
1287 cfqq->ioprio = task_ioprio(tsk);
1288 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1289 break;
1290 case IOPRIO_CLASS_BE:
1291 cfqq->ioprio = task_ioprio(tsk);
1292 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1293 break;
1294 case IOPRIO_CLASS_IDLE:
1295 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1296 cfqq->ioprio = 7;
Jens Axboe3b181522005-06-27 10:56:24 +02001297 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001298 break;
1299 }
1300
1301 /*
1302 * keep track of original prio settings in case we have to temporarily
1303 * elevate the priority of this queue
1304 */
1305 cfqq->org_ioprio = cfqq->ioprio;
1306 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001307 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001308}
1309
Al Viro478a82b2006-03-18 13:25:24 -05001310static inline void changed_ioprio(struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001311{
Al Viro478a82b2006-03-18 13:25:24 -05001312 struct cfq_data *cfqd = cic->key;
1313 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001314 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001315
Jens Axboecaaa5f92006-06-16 11:23:00 +02001316 if (unlikely(!cfqd))
1317 return;
1318
Jens Axboec1b707d2006-10-30 19:54:23 +01001319 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001320
1321 cfqq = cic->cfqq[ASYNC];
1322 if (cfqq) {
1323 struct cfq_queue *new_cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001324 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc->task,
Jens Axboecaaa5f92006-06-16 11:23:00 +02001325 GFP_ATOMIC);
1326 if (new_cfqq) {
1327 cic->cfqq[ASYNC] = new_cfqq;
1328 cfq_put_queue(cfqq);
1329 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001330 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001331
1332 cfqq = cic->cfqq[SYNC];
1333 if (cfqq)
1334 cfq_mark_cfqq_prio_changed(cfqq);
1335
Jens Axboec1b707d2006-10-30 19:54:23 +01001336 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001337}
1338
Jens Axboefc463792006-08-29 09:05:44 +02001339static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340{
Al Viroa6a07632006-03-18 13:26:44 -05001341 struct cfq_io_context *cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001342 struct rb_node *n;
Al Viroa6a07632006-03-18 13:26:44 -05001343
Jens Axboefc463792006-08-29 09:05:44 +02001344 ioc->ioprio_changed = 0;
Al Viroa6a07632006-03-18 13:26:44 -05001345
Jens Axboee2d74ac2006-03-28 08:59:01 +02001346 n = rb_first(&ioc->cic_root);
1347 while (n != NULL) {
1348 cic = rb_entry(n, struct cfq_io_context, rb_node);
Jens Axboe3793c652006-05-30 21:11:04 +02001349
Al Viro478a82b2006-03-18 13:25:24 -05001350 changed_ioprio(cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001351 n = rb_next(n);
1352 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353}
1354
1355static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001356cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1357 struct task_struct *tsk, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001360 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361
1362retry:
Vasily Tarasov91fac312007-04-25 12:29:51 +02001363 cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1364 /* cic always exists here */
1365 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366
1367 if (!cfqq) {
1368 if (new_cfqq) {
1369 cfqq = new_cfqq;
1370 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001371 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001372 /*
1373 * Inform the allocator of the fact that we will
1374 * just repeat this allocation if it fails, to allow
1375 * the allocator to do whatever it needs to attempt to
1376 * free memory.
1377 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001379 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1380 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1381 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 spin_lock_irq(cfqd->queue->queue_lock);
1383 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001384 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001385 cfqq = kmem_cache_alloc_node(cfq_pool,
1386 gfp_mask | __GFP_ZERO,
1387 cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001388 if (!cfqq)
1389 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001390 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391
Jens Axboed9e76202007-04-20 14:27:50 +02001392 RB_CLEAR_NODE(&cfqq->rb_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001393 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 atomic_set(&cfqq->ref, 0);
1396 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001397
Vasily Tarasov91fac312007-04-25 12:29:51 +02001398 if (is_sync) {
Jens Axboea9938002007-04-20 08:55:52 +02001399 cfq_mark_cfqq_idle_window(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001400 cfq_mark_cfqq_sync(cfqq);
1401 }
Jens Axboea9938002007-04-20 08:55:52 +02001402
Jens Axboe3b181522005-06-27 10:56:24 +02001403 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b03742006-07-28 09:48:51 +02001404 cfq_mark_cfqq_queue_new(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001405
Jens Axboe3b181522005-06-27 10:56:24 +02001406 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407 }
1408
1409 if (new_cfqq)
1410 kmem_cache_free(cfq_pool, new_cfqq);
1411
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412out:
1413 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1414 return cfqq;
1415}
1416
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001417static struct cfq_queue **
1418cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1419{
1420 switch(ioprio_class) {
1421 case IOPRIO_CLASS_RT:
1422 return &cfqd->async_cfqq[0][ioprio];
1423 case IOPRIO_CLASS_BE:
1424 return &cfqd->async_cfqq[1][ioprio];
1425 case IOPRIO_CLASS_IDLE:
1426 return &cfqd->async_idle_cfqq;
1427 default:
1428 BUG();
1429 }
1430}
1431
Jens Axboe15c31be2007-07-10 13:43:25 +02001432static struct cfq_queue *
1433cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct task_struct *tsk,
1434 gfp_t gfp_mask)
1435{
1436 const int ioprio = task_ioprio(tsk);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001437 const int ioprio_class = task_ioprio_class(tsk);
1438 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001439 struct cfq_queue *cfqq = NULL;
1440
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001441 if (!is_sync) {
1442 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1443 cfqq = *async_cfqq;
1444 }
1445
Jens Axboe15c31be2007-07-10 13:43:25 +02001446 if (!cfqq)
1447 cfqq = cfq_find_alloc_queue(cfqd, is_sync, tsk, gfp_mask);
1448
1449 /*
1450 * pin the queue now that it's allocated, scheduler exit will prune it
1451 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001452 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001453 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001454 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001455 }
1456
1457 atomic_inc(&cfqq->ref);
1458 return cfqq;
1459}
1460
Jens Axboe498d3aa2007-04-26 12:54:48 +02001461/*
1462 * We drop cfq io contexts lazily, so we may find a dead one.
1463 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001464static void
1465cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1466{
Jens Axboefc463792006-08-29 09:05:44 +02001467 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001468
1469 if (ioc->ioc_data == cic)
1470 ioc->ioc_data = NULL;
1471
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001472 rb_erase(&cic->rb_node, &ioc->cic_root);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001473 kmem_cache_free(cfq_ioc_pool, cic);
Jens Axboe4050cf12006-07-19 05:07:12 +02001474 elv_ioc_count_dec(ioc_count);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001475}
1476
Jens Axboee2d74ac2006-03-28 08:59:01 +02001477static struct cfq_io_context *
1478cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1479{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001480 struct rb_node *n;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001481 struct cfq_io_context *cic;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001482 void *k, *key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001483
Vasily Tarasov91fac312007-04-25 12:29:51 +02001484 if (unlikely(!ioc))
1485 return NULL;
1486
Jens Axboe597bc482007-04-24 21:23:53 +02001487 /*
1488 * we maintain a last-hit cache, to avoid browsing over the tree
1489 */
1490 cic = ioc->ioc_data;
1491 if (cic && cic->key == cfqd)
1492 return cic;
1493
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001494restart:
1495 n = ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001496 while (n) {
1497 cic = rb_entry(n, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001498 /* ->key must be copied to avoid race with cfq_exit_queue() */
1499 k = cic->key;
1500 if (unlikely(!k)) {
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001501 cfq_drop_dead_cic(ioc, cic);
1502 goto restart;
1503 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001504
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001505 if (key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001506 n = n->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001507 else if (key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001508 n = n->rb_right;
Jens Axboe597bc482007-04-24 21:23:53 +02001509 else {
1510 ioc->ioc_data = cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001511 return cic;
Jens Axboe597bc482007-04-24 21:23:53 +02001512 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001513 }
1514
1515 return NULL;
1516}
1517
1518static inline void
1519cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1520 struct cfq_io_context *cic)
1521{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001522 struct rb_node **p;
1523 struct rb_node *parent;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001524 struct cfq_io_context *__cic;
Jens Axboe0261d682006-10-30 19:07:48 +01001525 unsigned long flags;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001526 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001527
Jens Axboee2d74ac2006-03-28 08:59:01 +02001528 cic->ioc = ioc;
1529 cic->key = cfqd;
1530
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001531restart:
1532 parent = NULL;
1533 p = &ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001534 while (*p) {
1535 parent = *p;
1536 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001537 /* ->key must be copied to avoid race with cfq_exit_queue() */
1538 k = __cic->key;
1539 if (unlikely(!k)) {
Oleg Nesterovbe33c3a2006-08-21 08:36:12 +02001540 cfq_drop_dead_cic(ioc, __cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001541 goto restart;
1542 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001543
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001544 if (cic->key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001545 p = &(*p)->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001546 else if (cic->key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001547 p = &(*p)->rb_right;
1548 else
1549 BUG();
1550 }
1551
1552 rb_link_node(&cic->rb_node, parent, p);
1553 rb_insert_color(&cic->rb_node, &ioc->cic_root);
Jens Axboefc463792006-08-29 09:05:44 +02001554
Jens Axboe0261d682006-10-30 19:07:48 +01001555 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001556 list_add(&cic->queue_list, &cfqd->cic_list);
Jens Axboe0261d682006-10-30 19:07:48 +01001557 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001558}
1559
Jens Axboe22e2c502005-06-27 10:55:12 +02001560/*
1561 * Setup general io context and cfq io context. There can be several cfq
1562 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001563 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001564 */
1565static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001566cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567{
Jens Axboe22e2c502005-06-27 10:55:12 +02001568 struct io_context *ioc = NULL;
1569 struct cfq_io_context *cic;
1570
1571 might_sleep_if(gfp_mask & __GFP_WAIT);
1572
Jens Axboeb5deef92006-07-19 23:39:40 +02001573 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001574 if (!ioc)
1575 return NULL;
1576
Jens Axboee2d74ac2006-03-28 08:59:01 +02001577 cic = cfq_cic_rb_lookup(cfqd, ioc);
1578 if (cic)
1579 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001580
Jens Axboee2d74ac2006-03-28 08:59:01 +02001581 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1582 if (cic == NULL)
1583 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001584
Jens Axboee2d74ac2006-03-28 08:59:01 +02001585 cfq_cic_link(cfqd, ioc, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001586out:
Jens Axboefc463792006-08-29 09:05:44 +02001587 smp_read_barrier_depends();
1588 if (unlikely(ioc->ioprio_changed))
1589 cfq_ioc_set_ioprio(ioc);
1590
Jens Axboe22e2c502005-06-27 10:55:12 +02001591 return cic;
1592err:
1593 put_io_context(ioc);
1594 return NULL;
1595}
1596
1597static void
1598cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1599{
Jens Axboeaaf12282007-01-19 11:30:16 +11001600 unsigned long elapsed = jiffies - cic->last_end_request;
1601 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001602
1603 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1604 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1605 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1606}
1607
Jens Axboe206dc692006-03-28 13:03:44 +02001608static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001609cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1610 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001611{
1612 sector_t sdist;
1613 u64 total;
1614
Jens Axboe5e705372006-07-13 12:39:25 +02001615 if (cic->last_request_pos < rq->sector)
1616 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001617 else
Jens Axboe5e705372006-07-13 12:39:25 +02001618 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001619
1620 /*
1621 * Don't allow the seek distance to get too large from the
1622 * odd fragment, pagein, etc
1623 */
1624 if (cic->seek_samples <= 60) /* second&third seek */
1625 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1626 else
1627 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1628
1629 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1630 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1631 total = cic->seek_total + (cic->seek_samples/2);
1632 do_div(total, cic->seek_samples);
1633 cic->seek_mean = (sector_t)total;
1634}
Jens Axboe22e2c502005-06-27 10:55:12 +02001635
1636/*
1637 * Disable idle window if the process thinks too long or seeks so much that
1638 * it doesn't matter
1639 */
1640static void
1641cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1642 struct cfq_io_context *cic)
1643{
Jens Axboe1be92f22007-04-19 14:32:26 +02001644 int enable_idle;
1645
1646 if (!cfq_cfqq_sync(cfqq))
1647 return;
1648
1649 enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001650
Jens Axboecaaa5f92006-06-16 11:23:00 +02001651 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1652 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001653 enable_idle = 0;
1654 else if (sample_valid(cic->ttime_samples)) {
1655 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1656 enable_idle = 0;
1657 else
1658 enable_idle = 1;
1659 }
1660
Jens Axboe3b181522005-06-27 10:56:24 +02001661 if (enable_idle)
1662 cfq_mark_cfqq_idle_window(cfqq);
1663 else
1664 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001665}
1666
Jens Axboe22e2c502005-06-27 10:55:12 +02001667/*
1668 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1669 * no or if we aren't sure, a 1 will cause a preempt.
1670 */
1671static int
1672cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001673 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001674{
Jens Axboe6d048f52007-04-25 12:44:27 +02001675 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001676
Jens Axboe6d048f52007-04-25 12:44:27 +02001677 cfqq = cfqd->active_queue;
1678 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001679 return 0;
1680
Jens Axboe6d048f52007-04-25 12:44:27 +02001681 if (cfq_slice_used(cfqq))
1682 return 1;
1683
1684 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001685 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001686
1687 if (cfq_class_idle(cfqq))
1688 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001689
Jens Axboe22e2c502005-06-27 10:55:12 +02001690 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02001691 * if the new request is sync, but the currently running queue is
1692 * not, let the sync request have priority.
1693 */
Jens Axboe5e705372006-07-13 12:39:25 +02001694 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001695 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001696
Jens Axboe374f84a2006-07-23 01:42:19 +02001697 /*
1698 * So both queues are sync. Let the new request get disk time if
1699 * it's a metadata request and the current queue is doing regular IO.
1700 */
1701 if (rq_is_meta(rq) && !cfqq->meta_pending)
1702 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001703
Jens Axboe1e3335d2007-02-14 19:59:49 +01001704 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1705 return 0;
1706
1707 /*
1708 * if this request is as-good as one we would expect from the
1709 * current cfqq, let it preempt
1710 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001711 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01001712 return 1;
1713
Jens Axboe22e2c502005-06-27 10:55:12 +02001714 return 0;
1715}
1716
1717/*
1718 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1719 * let it have half of its nominal slice.
1720 */
1721static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1722{
Jens Axboe6084cdd2007-04-23 08:25:00 +02001723 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001724
Jens Axboebf572252006-07-19 20:29:12 +02001725 /*
1726 * Put the new queue at the front of the of the current list,
1727 * so we know that it will be selected next.
1728 */
1729 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02001730
1731 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02001732
Jens Axboe44f7c162007-01-19 11:51:58 +11001733 cfqq->slice_end = 0;
1734 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001735}
1736
1737/*
Jens Axboe5e705372006-07-13 12:39:25 +02001738 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001739 * something we should do about it
1740 */
1741static void
Jens Axboe5e705372006-07-13 12:39:25 +02001742cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1743 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001744{
Jens Axboe5e705372006-07-13 12:39:25 +02001745 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001746
Jens Axboe374f84a2006-07-23 01:42:19 +02001747 if (rq_is_meta(rq))
1748 cfqq->meta_pending++;
1749
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001750 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02001751 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001752 cfq_update_idle_window(cfqd, cfqq, cic);
1753
Jens Axboe5e705372006-07-13 12:39:25 +02001754 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001755
1756 if (cfqq == cfqd->active_queue) {
1757 /*
1758 * if we are waiting for a request for this queue, let it rip
1759 * immediately and flag that we must not expire this queue
1760 * just now
1761 */
Jens Axboe3b181522005-06-27 10:56:24 +02001762 if (cfq_cfqq_wait_request(cfqq)) {
1763 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001764 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001765 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001766 }
Jens Axboe5e705372006-07-13 12:39:25 +02001767 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001768 /*
1769 * not the active queue - expire current slice if it is
1770 * idle and has expired it's mean thinktime or this new queue
1771 * has some old slice time left and is of higher priority
1772 */
1773 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001774 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001775 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001776 }
1777}
1778
Jens Axboe165125e2007-07-24 09:28:11 +02001779static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001780{
Jens Axboeb4878f22005-10-20 16:42:29 +02001781 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001782 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001783
1784 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785
Jens Axboe5e705372006-07-13 12:39:25 +02001786 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787
Jens Axboe22e2c502005-06-27 10:55:12 +02001788 list_add_tail(&rq->queuelist, &cfqq->fifo);
1789
Jens Axboe5e705372006-07-13 12:39:25 +02001790 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791}
1792
Jens Axboe165125e2007-07-24 09:28:11 +02001793static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794{
Jens Axboe5e705372006-07-13 12:39:25 +02001795 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001796 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001797 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001798 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799
Jens Axboeb4878f22005-10-20 16:42:29 +02001800 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801
Jens Axboeb4878f22005-10-20 16:42:29 +02001802 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02001803 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02001804 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02001805 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806
Jens Axboe3ed9a292007-04-23 08:33:33 +02001807 if (cfq_cfqq_sync(cfqq))
1808 cfqd->sync_flight--;
1809
Jens Axboeb4878f22005-10-20 16:42:29 +02001810 if (!cfq_class_idle(cfqq))
1811 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001812
Jens Axboecaaa5f92006-06-16 11:23:00 +02001813 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001814 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001815
1816 /*
1817 * If this is the active queue, check if it needs to be expired,
1818 * or if we want to idle in case it has no pending requests.
1819 */
1820 if (cfqd->active_queue == cfqq) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001821 if (cfq_cfqq_slice_new(cfqq)) {
1822 cfq_set_prio_slice(cfqd, cfqq);
1823 cfq_clear_cfqq_slice_new(cfqq);
1824 }
1825 if (cfq_slice_used(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02001826 cfq_slice_expired(cfqd, 1);
Jens Axboe6d048f52007-04-25 12:44:27 +02001827 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1828 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001829 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001830
1831 if (!cfqd->rq_in_driver)
1832 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833}
1834
Jens Axboe22e2c502005-06-27 10:55:12 +02001835/*
1836 * we temporarily boost lower priority queues if they are holding fs exclusive
1837 * resources. they are boosted to normal prio (CLASS_BE/4)
1838 */
1839static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840{
Jens Axboe22e2c502005-06-27 10:55:12 +02001841 if (has_fs_excl()) {
1842 /*
1843 * boost idle prio on transactions that would lock out other
1844 * users of the filesystem
1845 */
1846 if (cfq_class_idle(cfqq))
1847 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1848 if (cfqq->ioprio > IOPRIO_NORM)
1849 cfqq->ioprio = IOPRIO_NORM;
1850 } else {
1851 /*
1852 * check if we need to unboost the queue
1853 */
1854 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1855 cfqq->ioprio_class = cfqq->org_ioprio_class;
1856 if (cfqq->ioprio != cfqq->org_ioprio)
1857 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001859}
1860
Jens Axboe89850f72006-07-22 16:48:31 +02001861static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001862{
Jens Axboe3b181522005-06-27 10:56:24 +02001863 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001864 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001865 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001866 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001867 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001868
1869 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02001870}
1871
Jens Axboe165125e2007-07-24 09:28:11 +02001872static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02001873{
1874 struct cfq_data *cfqd = q->elevator->elevator_data;
1875 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001876 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02001877 struct cfq_queue *cfqq;
1878
1879 /*
1880 * don't force setup of a queue from here, as a call to may_queue
1881 * does not necessarily imply that a request actually will be queued.
1882 * so just lookup a possibly existing queue, or return 'may queue'
1883 * if that fails
1884 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001885 cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1886 if (!cic)
1887 return ELV_MQUEUE_MAY;
1888
1889 cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02001890 if (cfqq) {
1891 cfq_init_prio_data(cfqq);
1892 cfq_prio_boost(cfqq);
1893
Jens Axboe89850f72006-07-22 16:48:31 +02001894 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001895 }
1896
1897 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898}
1899
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900/*
1901 * queue lock held here
1902 */
Jens Axboebb37b942006-12-01 10:42:33 +01001903static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904{
Jens Axboe5e705372006-07-13 12:39:25 +02001905 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906
Jens Axboe5e705372006-07-13 12:39:25 +02001907 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001908 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909
Jens Axboe22e2c502005-06-27 10:55:12 +02001910 BUG_ON(!cfqq->allocated[rw]);
1911 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912
Jens Axboe5e705372006-07-13 12:39:25 +02001913 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02001916 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918 cfq_put_queue(cfqq);
1919 }
1920}
1921
1922/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001923 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001925static int
Jens Axboe165125e2007-07-24 09:28:11 +02001926cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927{
1928 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe3b181522005-06-27 10:56:24 +02001929 struct task_struct *tsk = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 struct cfq_io_context *cic;
1931 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01001932 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001933 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934 unsigned long flags;
1935
1936 might_sleep_if(gfp_mask & __GFP_WAIT);
1937
Jens Axboee2d74ac2006-03-28 08:59:01 +02001938 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001939
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 spin_lock_irqsave(q->queue_lock, flags);
1941
Jens Axboe22e2c502005-06-27 10:55:12 +02001942 if (!cic)
1943 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944
Vasily Tarasov91fac312007-04-25 12:29:51 +02001945 cfqq = cic_to_cfqq(cic, is_sync);
1946 if (!cfqq) {
1947 cfqq = cfq_get_queue(cfqd, is_sync, tsk, gfp_mask);
1948
Jens Axboe22e2c502005-06-27 10:55:12 +02001949 if (!cfqq)
1950 goto queue_fail;
1951
Vasily Tarasov91fac312007-04-25 12:29:51 +02001952 cic_set_cfqq(cic, cfqq, is_sync);
1953 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954
1955 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02001956 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001957 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02001958
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 spin_unlock_irqrestore(q->queue_lock, flags);
1960
Jens Axboe5e705372006-07-13 12:39:25 +02001961 rq->elevator_private = cic;
1962 rq->elevator_private2 = cfqq;
1963 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02001964
Jens Axboe22e2c502005-06-27 10:55:12 +02001965queue_fail:
1966 if (cic)
1967 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02001968
Jens Axboe3b181522005-06-27 10:56:24 +02001969 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970 spin_unlock_irqrestore(q->queue_lock, flags);
1971 return 1;
1972}
1973
David Howells65f27f32006-11-22 14:55:48 +00001974static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02001975{
David Howells65f27f32006-11-22 14:55:48 +00001976 struct cfq_data *cfqd =
1977 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02001978 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001979 unsigned long flags;
1980
1981 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02001982 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02001983 spin_unlock_irqrestore(q->queue_lock, flags);
1984}
1985
1986/*
1987 * Timer running if the active_queue is currently idling inside its time slice
1988 */
1989static void cfq_idle_slice_timer(unsigned long data)
1990{
1991 struct cfq_data *cfqd = (struct cfq_data *) data;
1992 struct cfq_queue *cfqq;
1993 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001994 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001995
1996 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1997
1998 if ((cfqq = cfqd->active_queue) != NULL) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001999 timed_out = 0;
2000
Jens Axboe22e2c502005-06-27 10:55:12 +02002001 /*
2002 * expired
2003 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002004 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002005 goto expire;
2006
2007 /*
2008 * only expire and reinvoke request handler, if there are
2009 * other queues with pending requests
2010 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002011 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002012 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002013
2014 /*
2015 * not expired and it has a request pending, let it dispatch
2016 */
Jens Axboedd67d052006-06-21 09:36:18 +02002017 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002018 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002019 goto out_kick;
2020 }
2021 }
2022expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002023 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002024out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002025 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002026out_cont:
2027 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2028}
2029
2030/*
2031 * Timer running if an idle class queue is waiting for service
2032 */
2033static void cfq_idle_class_timer(unsigned long data)
2034{
2035 struct cfq_data *cfqd = (struct cfq_data *) data;
2036 unsigned long flags, end;
2037
2038 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2039
2040 /*
2041 * race with a non-idle queue, reset timer
2042 */
2043 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
Jens Axboeae818a32006-06-01 10:13:43 +02002044 if (!time_after_eq(jiffies, end))
2045 mod_timer(&cfqd->idle_class_timer, end);
2046 else
Jens Axboe3b181522005-06-27 10:56:24 +02002047 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002048
2049 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2050}
2051
Jens Axboe3b181522005-06-27 10:56:24 +02002052static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2053{
2054 del_timer_sync(&cfqd->idle_slice_timer);
2055 del_timer_sync(&cfqd->idle_class_timer);
2056 blk_sync_queue(cfqd->queue);
2057}
Jens Axboe22e2c502005-06-27 10:55:12 +02002058
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002059static void cfq_put_async_queues(struct cfq_data *cfqd)
2060{
2061 int i;
2062
2063 for (i = 0; i < IOPRIO_BE_NR; i++) {
2064 if (cfqd->async_cfqq[0][i])
2065 cfq_put_queue(cfqd->async_cfqq[0][i]);
2066 if (cfqd->async_cfqq[1][i])
2067 cfq_put_queue(cfqd->async_cfqq[1][i]);
2068 if (cfqd->async_idle_cfqq)
2069 cfq_put_queue(cfqd->async_idle_cfqq);
2070 }
2071}
2072
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073static void cfq_exit_queue(elevator_t *e)
2074{
Jens Axboe22e2c502005-06-27 10:55:12 +02002075 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002076 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002077
Jens Axboe3b181522005-06-27 10:56:24 +02002078 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002079
Al Virod9ff4182006-03-18 13:51:22 -05002080 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002081
Al Virod9ff4182006-03-18 13:51:22 -05002082 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002083 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002084
2085 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002086 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2087 struct cfq_io_context,
2088 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002089
2090 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002091 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002092
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002093 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002094
Al Virod9ff4182006-03-18 13:51:22 -05002095 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002096
2097 cfq_shutdown_timer_wq(cfqd);
2098
Al Viroa90d7422006-03-18 12:05:37 -05002099 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100}
2101
Jens Axboe165125e2007-07-24 09:28:11 +02002102static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103{
2104 struct cfq_data *cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105
Christoph Lameter94f60302007-07-17 04:03:29 -07002106 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002108 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109
Jens Axboecc09e292007-04-26 12:53:50 +02002110 cfqd->service_tree = CFQ_RB_ROOT;
Al Virod9ff4182006-03-18 13:51:22 -05002111 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114
Jens Axboe22e2c502005-06-27 10:55:12 +02002115 init_timer(&cfqd->idle_slice_timer);
2116 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2117 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2118
2119 init_timer(&cfqd->idle_class_timer);
2120 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2121 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2122
David Howells65f27f32006-11-22 14:55:48 +00002123 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002124
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002126 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2127 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 cfqd->cfq_back_max = cfq_back_max;
2129 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002130 cfqd->cfq_slice[0] = cfq_slice_async;
2131 cfqd->cfq_slice[1] = cfq_slice_sync;
2132 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2133 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02002134
Jens Axboebc1c1162006-06-08 08:49:06 +02002135 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136}
2137
2138static void cfq_slab_kill(void)
2139{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140 if (cfq_pool)
2141 kmem_cache_destroy(cfq_pool);
2142 if (cfq_ioc_pool)
2143 kmem_cache_destroy(cfq_ioc_pool);
2144}
2145
2146static int __init cfq_slab_setup(void)
2147{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002148 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 if (!cfq_pool)
2150 goto fail;
2151
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002152 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153 if (!cfq_ioc_pool)
2154 goto fail;
2155
2156 return 0;
2157fail:
2158 cfq_slab_kill();
2159 return -ENOMEM;
2160}
2161
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162/*
2163 * sysfs parts below -->
2164 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165static ssize_t
2166cfq_var_show(unsigned int var, char *page)
2167{
2168 return sprintf(page, "%d\n", var);
2169}
2170
2171static ssize_t
2172cfq_var_store(unsigned int *var, const char *page, size_t count)
2173{
2174 char *p = (char *) page;
2175
2176 *var = simple_strtoul(p, &p, 10);
2177 return count;
2178}
2179
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002181static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002183 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184 unsigned int __data = __VAR; \
2185 if (__CONV) \
2186 __data = jiffies_to_msecs(__data); \
2187 return cfq_var_show(__data, (page)); \
2188}
2189SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002190SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2191SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002192SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2193SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002194SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2195SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2196SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2197SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198#undef SHOW_FUNCTION
2199
2200#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002201static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002203 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 unsigned int __data; \
2205 int ret = cfq_var_store(&__data, (page), count); \
2206 if (__data < (MIN)) \
2207 __data = (MIN); \
2208 else if (__data > (MAX)) \
2209 __data = (MAX); \
2210 if (__CONV) \
2211 *(__PTR) = msecs_to_jiffies(__data); \
2212 else \
2213 *(__PTR) = __data; \
2214 return ret; \
2215}
2216STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002217STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2218STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002219STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2220STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002221STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2222STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2223STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2224STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225#undef STORE_FUNCTION
2226
Al Viroe572ec72006-03-18 22:27:18 -05002227#define CFQ_ATTR(name) \
2228 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002229
Al Viroe572ec72006-03-18 22:27:18 -05002230static struct elv_fs_entry cfq_attrs[] = {
2231 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002232 CFQ_ATTR(fifo_expire_sync),
2233 CFQ_ATTR(fifo_expire_async),
2234 CFQ_ATTR(back_seek_max),
2235 CFQ_ATTR(back_seek_penalty),
2236 CFQ_ATTR(slice_sync),
2237 CFQ_ATTR(slice_async),
2238 CFQ_ATTR(slice_async_rq),
2239 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002240 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241};
2242
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243static struct elevator_type iosched_cfq = {
2244 .ops = {
2245 .elevator_merge_fn = cfq_merge,
2246 .elevator_merged_fn = cfq_merged_request,
2247 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002248 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002249 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002251 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 .elevator_deactivate_req_fn = cfq_deactivate_request,
2253 .elevator_queue_empty_fn = cfq_queue_empty,
2254 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002255 .elevator_former_req_fn = elv_rb_former_request,
2256 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 .elevator_set_req_fn = cfq_set_request,
2258 .elevator_put_req_fn = cfq_put_request,
2259 .elevator_may_queue_fn = cfq_may_queue,
2260 .elevator_init_fn = cfq_init_queue,
2261 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002262 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 },
Al Viro3d1ab402006-03-18 18:35:43 -05002264 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 .elevator_name = "cfq",
2266 .elevator_owner = THIS_MODULE,
2267};
2268
2269static int __init cfq_init(void)
2270{
2271 int ret;
2272
Jens Axboe22e2c502005-06-27 10:55:12 +02002273 /*
2274 * could be 0 on HZ < 1000 setups
2275 */
2276 if (!cfq_slice_async)
2277 cfq_slice_async = 1;
2278 if (!cfq_slice_idle)
2279 cfq_slice_idle = 1;
2280
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 if (cfq_slab_setup())
2282 return -ENOMEM;
2283
2284 ret = elv_register(&iosched_cfq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002285 if (ret)
2286 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 return ret;
2289}
2290
2291static void __exit cfq_exit(void)
2292{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002293 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002295 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002296 /* ioc_gone's update must be visible before reading ioc_count */
2297 smp_wmb();
Jens Axboe4050cf12006-07-19 05:07:12 +02002298 if (elv_ioc_count_read(ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002299 wait_for_completion(ioc_gone);
Al Viro334e94d2006-03-18 15:05:53 -05002300 synchronize_rcu();
Christoph Hellwig83521d32005-10-30 15:01:39 -08002301 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302}
2303
2304module_init(cfq_init);
2305module_exit(cfq_exit);
2306
2307MODULE_AUTHOR("Jens Axboe");
2308MODULE_LICENSE("GPL");
2309MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");