blob: 13553e015d7203f244db2e411884c56cbee1bc26 [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 Axboeb0b8d7492007-01-19 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 Axboe53b037442006-07-28 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 Axboe498d3aa22007-04-26 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 Axboe498d3aa22007-04-26 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 Axboe498d3aa22007-04-26 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 Axboe53b037442006-07-28 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 Axboe498d3aa22007-04-26 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 Axboe498d3aa22007-04-26 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 Axboe498d3aa22007-04-26 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 Axboe498d3aa22007-04-26 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
Oleg Nesterov0e7be9e2007-11-07 13:51:35 +0100792static int start_idle_class_timer(struct cfq_data *cfqd)
793{
794 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
795 unsigned long now = jiffies;
796
797 if (time_before(now, end) &&
798 time_after_eq(now, cfqd->last_end_request)) {
799 mod_timer(&cfqd->idle_class_timer, end);
800 return 1;
801 }
802
803 return 0;
804}
805
Jens Axboe498d3aa22007-04-26 12:54:48 +0200806/*
807 * Get next queue for service. Unless we have a queue preemption,
808 * we'll simply select the first cfqq in the service tree.
809 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200810static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200811{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200812 struct cfq_queue *cfqq;
813 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +0200814
Jens Axboeedd75ff2007-04-19 12:03:34 +0200815 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
816 return NULL;
817
818 n = cfq_rb_first(&cfqd->service_tree);
819 cfqq = rb_entry(n, struct cfq_queue, rb_node);
820
821 if (cfq_class_idle(cfqq)) {
Jens Axboe89850f72006-07-22 16:48:31 +0200822 /*
Jens Axboeedd75ff2007-04-19 12:03:34 +0200823 * if we have idle queues and no rt or be queues had
824 * pending requests, either allow immediate service if
825 * the grace period has passed or arm the idle grace
826 * timer
Jens Axboe89850f72006-07-22 16:48:31 +0200827 */
Oleg Nesterov0e7be9e2007-11-07 13:51:35 +0100828 if (start_idle_class_timer(cfqd))
Jens Axboeedd75ff2007-04-19 12:03:34 +0200829 cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +0200830 }
831
Jens Axboe6d048f52007-04-25 12:44:27 +0200832 return cfqq;
833}
834
Jens Axboe498d3aa22007-04-26 12:54:48 +0200835/*
836 * Get and set a new active queue for service.
837 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200838static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
839{
840 struct cfq_queue *cfqq;
841
Jens Axboed9e76202007-04-20 14:27:50 +0200842 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200843 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200844 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200845}
846
Jens Axboed9e76202007-04-20 14:27:50 +0200847static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
848 struct request *rq)
849{
850 if (rq->sector >= cfqd->last_position)
851 return rq->sector - cfqd->last_position;
852 else
853 return cfqd->last_position - rq->sector;
854}
855
Jens Axboe6d048f52007-04-25 12:44:27 +0200856static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
857{
858 struct cfq_io_context *cic = cfqd->active_cic;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200859
Jens Axboe6d048f52007-04-25 12:44:27 +0200860 if (!sample_valid(cic->seek_samples))
861 return 0;
862
863 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
864}
865
Jens Axboed9e76202007-04-20 14:27:50 +0200866static int cfq_close_cooperator(struct cfq_data *cfq_data,
867 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200868{
Jens Axboe6d048f52007-04-25 12:44:27 +0200869 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200870 * We should notice if some of the queues are cooperating, eg
871 * working closely on the same area of the disk. In that case,
872 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +0200873 */
Jens Axboed9e76202007-04-20 14:27:50 +0200874 return 0;
Jens Axboe6d048f52007-04-25 12:44:27 +0200875}
876
877#define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
878
879static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200880{
Jens Axboe17926692007-01-19 11:59:30 +1100881 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +0200882 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100883 unsigned long sl;
884
Jens Axboedd67d052006-06-21 09:36:18 +0200885 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +0200886 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +0200887
888 /*
889 * idle is disabled, either manually or by past process history
890 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200891 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
892 return;
893
Jens Axboe22e2c502005-06-27 10:55:12 +0200894 /*
895 * task has exited, don't wait
896 */
Jens Axboe206dc692006-03-28 13:03:44 +0200897 cic = cfqd->active_cic;
898 if (!cic || !cic->ioc->task)
Jens Axboe6d048f52007-04-25 12:44:27 +0200899 return;
900
901 /*
902 * See if this prio level has a good candidate
903 */
Jens Axboe1afba042007-04-17 12:47:55 +0200904 if (cfq_close_cooperator(cfqd, cfqq) &&
905 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
Jens Axboe6d048f52007-04-25 12:44:27 +0200906 return;
Jens Axboe22e2c502005-06-27 10:55:12 +0200907
Jens Axboe3b181522005-06-27 10:56:24 +0200908 cfq_mark_cfqq_must_dispatch(cfqq);
909 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200910
Jens Axboe206dc692006-03-28 13:03:44 +0200911 /*
912 * we don't want to idle for seeks, but we do want to allow
913 * fair distribution of slice time for a process doing back-to-back
914 * seeks. so allow a little bit of time for him to submit a new rq
915 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200916 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200917 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +0200918 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +0200919
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100920 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921}
922
Jens Axboe498d3aa22007-04-26 12:54:48 +0200923/*
924 * Move request from internal lists to the request queue dispatch list.
925 */
Jens Axboe165125e2007-07-24 09:28:11 +0200926static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927{
Jens Axboe3ed9a292007-04-23 08:33:33 +0200928 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200929 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200930
Jens Axboe5380a102006-07-13 12:37:56 +0200931 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200932 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +0200933 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +0200934
935 if (cfq_cfqq_sync(cfqq))
936 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937}
938
939/*
940 * return expired entry, or NULL to just start from scratch in rbtree
941 */
Jens Axboe5e705372006-07-13 12:39:25 +0200942static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943{
944 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200945 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200946 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947
Jens Axboe3b181522005-06-27 10:56:24 +0200948 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +1100950
951 cfq_mark_cfqq_fifo_expire(cfqq);
952
Jens Axboe89850f72006-07-22 16:48:31 +0200953 if (list_empty(&cfqq->fifo))
954 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955
Jens Axboe6d048f52007-04-25 12:44:27 +0200956 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +0200957 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958
Jens Axboe6d048f52007-04-25 12:44:27 +0200959 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
960 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961
Jens Axboe6d048f52007-04-25 12:44:27 +0200962 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963}
964
Jens Axboe22e2c502005-06-27 10:55:12 +0200965static inline int
966cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
967{
968 const int base_rq = cfqd->cfq_slice_async_rq;
969
970 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
971
972 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
973}
974
975/*
Jens Axboe498d3aa22007-04-26 12:54:48 +0200976 * Select a queue for service. If we have a current active queue,
977 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +0200978 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100979static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200980{
Jens Axboe22e2c502005-06-27 10:55:12 +0200981 struct cfq_queue *cfqq;
982
983 cfqq = cfqd->active_queue;
984 if (!cfqq)
985 goto new_queue;
986
987 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200988 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +0200989 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200990 if (cfq_slice_used(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +0200991 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +0200992
993 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200994 * The active queue has requests and isn't expired, allow it to
995 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +0200996 */
Jens Axboedd67d052006-06-21 09:36:18 +0200997 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200998 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +0200999
1000 /*
1001 * No requests pending. If the active queue still has requests in
1002 * flight or is idling for a new request, allow either of these
1003 * conditions to happen (or time out) before selecting a new queue.
1004 */
Jens Axboecc197472007-04-20 20:45:39 +02001005 if (timer_pending(&cfqd->idle_slice_timer) ||
1006 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001007 cfqq = NULL;
1008 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001009 }
1010
Jens Axboe3b181522005-06-27 10:56:24 +02001011expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001012 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001013new_queue:
1014 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001015keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001016 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001017}
1018
Jens Axboe498d3aa22007-04-26 12:54:48 +02001019/*
1020 * Dispatch some requests from cfqq, moving them to the request queue
1021 * dispatch list.
1022 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001023static int
1024__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1025 int max_dispatch)
1026{
1027 int dispatched = 0;
1028
Jens Axboedd67d052006-06-21 09:36:18 +02001029 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001030
1031 do {
Jens Axboe5e705372006-07-13 12:39:25 +02001032 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001033
1034 /*
1035 * follow expired path, else get first next available
1036 */
Jens Axboe5e705372006-07-13 12:39:25 +02001037 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1038 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001039
1040 /*
1041 * finally, insert request into driver dispatch list
1042 */
Jens Axboe5e705372006-07-13 12:39:25 +02001043 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001044
Jens Axboe22e2c502005-06-27 10:55:12 +02001045 dispatched++;
1046
1047 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +02001048 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1049 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001050 }
1051
Jens Axboedd67d052006-06-21 09:36:18 +02001052 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001053 break;
1054
1055 } while (dispatched < max_dispatch);
1056
1057 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001058 * expire an async queue immediately if it has used up its slice. idle
1059 * queue always expire after 1 dispatch round.
1060 */
Jens Axboea9938002007-04-20 08:55:52 +02001061 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
Jens Axboe20e493a2007-04-23 08:26:36 +02001062 dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboea9938002007-04-20 08:55:52 +02001063 cfq_class_idle(cfqq))) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001064 cfqq->slice_end = jiffies + 1;
Jens Axboe6084cdd2007-04-23 08:25:00 +02001065 cfq_slice_expired(cfqd, 0);
Jens Axboe44f7c162007-01-19 11:51:58 +11001066 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001067
1068 return dispatched;
1069}
1070
Jens Axboed9e76202007-04-20 14:27:50 +02001071static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1072{
1073 int dispatched = 0;
1074
1075 while (cfqq->next_rq) {
1076 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1077 dispatched++;
1078 }
1079
1080 BUG_ON(!list_empty(&cfqq->fifo));
1081 return dispatched;
1082}
1083
Jens Axboe498d3aa22007-04-26 12:54:48 +02001084/*
1085 * Drain our current requests. Used for barriers and when switching
1086 * io schedulers on-the-fly.
1087 */
Jens Axboed9e76202007-04-20 14:27:50 +02001088static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001089{
Jens Axboed9e76202007-04-20 14:27:50 +02001090 int dispatched = 0;
1091 struct rb_node *n;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001092
Jens Axboecc09e292007-04-26 12:53:50 +02001093 while ((n = cfq_rb_first(&cfqd->service_tree)) != NULL) {
Jens Axboed9e76202007-04-20 14:27:50 +02001094 struct cfq_queue *cfqq = rb_entry(n, struct cfq_queue, rb_node);
1095
1096 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1097 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001098
Jens Axboe6084cdd2007-04-23 08:25:00 +02001099 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001100
1101 BUG_ON(cfqd->busy_queues);
1102
1103 return dispatched;
1104}
1105
Jens Axboe165125e2007-07-24 09:28:11 +02001106static int cfq_dispatch_requests(struct request_queue *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107{
1108 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001109 struct cfq_queue *cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001110 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111
Jens Axboe22e2c502005-06-27 10:55:12 +02001112 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 return 0;
1114
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001115 if (unlikely(force))
1116 return cfq_forced_dispatch(cfqd);
1117
Jens Axboecaaa5f92006-06-16 11:23:00 +02001118 dispatched = 0;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001119 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001120 int max_dispatch;
1121
Jens Axboe3ed9a292007-04-23 08:33:33 +02001122 max_dispatch = cfqd->cfq_quantum;
1123 if (cfq_class_idle(cfqq))
1124 max_dispatch = 1;
1125
1126 if (cfqq->dispatched >= max_dispatch) {
1127 if (cfqd->busy_queues > 1)
Jens Axboe6d048f52007-04-25 12:44:27 +02001128 break;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001129 if (cfqq->dispatched >= 4 * max_dispatch)
Jens Axboea9938002007-04-20 08:55:52 +02001130 break;
1131 }
Jens Axboe9ede2092007-01-19 12:11:44 +11001132
Jens Axboe3ed9a292007-04-23 08:33:33 +02001133 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1134 break;
1135
Jens Axboe3b181522005-06-27 10:56:24 +02001136 cfq_clear_cfqq_must_dispatch(cfqq);
1137 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001138 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139
Jens Axboecaaa5f92006-06-16 11:23:00 +02001140 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 }
1142
Jens Axboecaaa5f92006-06-16 11:23:00 +02001143 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144}
1145
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146/*
Jens Axboe5e705372006-07-13 12:39:25 +02001147 * task holds one reference to the queue, dropped when task exits. each rq
1148 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 *
1150 * queue lock must be held here.
1151 */
1152static void cfq_put_queue(struct cfq_queue *cfqq)
1153{
Jens Axboe22e2c502005-06-27 10:55:12 +02001154 struct cfq_data *cfqd = cfqq->cfqd;
1155
1156 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157
1158 if (!atomic_dec_and_test(&cfqq->ref))
1159 return;
1160
1161 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001162 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001163 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001165 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001166 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001167 cfq_schedule_dispatch(cfqd);
1168 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001169
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 kmem_cache_free(cfq_pool, cfqq);
1171}
1172
Jens Axboee2d74ac2006-03-28 08:59:01 +02001173static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174{
Jens Axboe22e2c502005-06-27 10:55:12 +02001175 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001176 struct rb_node *n;
1177 int freed = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001178
Jens Axboe597bc482007-04-24 21:23:53 +02001179 ioc->ioc_data = NULL;
1180
Jens Axboee2d74ac2006-03-28 08:59:01 +02001181 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1182 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1183 rb_erase(&__cic->rb_node, &ioc->cic_root);
Jens Axboe22e2c502005-06-27 10:55:12 +02001184 kmem_cache_free(cfq_ioc_pool, __cic);
Al Viro334e94d2006-03-18 15:05:53 -05001185 freed++;
Jens Axboe22e2c502005-06-27 10:55:12 +02001186 }
1187
Jens Axboe4050cf12006-07-19 05:07:12 +02001188 elv_ioc_count_mod(ioc_count, -freed);
1189
1190 if (ioc_gone && !elv_ioc_count_read(ioc_count))
Al Viro334e94d2006-03-18 15:05:53 -05001191 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192}
1193
Jens Axboe89850f72006-07-22 16:48:31 +02001194static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1195{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001196 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001197 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001198 cfq_schedule_dispatch(cfqd);
1199 }
Jens Axboe89850f72006-07-22 16:48:31 +02001200
1201 cfq_put_queue(cfqq);
1202}
1203
1204static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1205 struct cfq_io_context *cic)
1206{
Jens Axboefc463792006-08-29 09:05:44 +02001207 list_del_init(&cic->queue_list);
1208 smp_wmb();
1209 cic->key = NULL;
1210
Jens Axboe89850f72006-07-22 16:48:31 +02001211 if (cic->cfqq[ASYNC]) {
1212 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1213 cic->cfqq[ASYNC] = NULL;
1214 }
1215
1216 if (cic->cfqq[SYNC]) {
1217 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1218 cic->cfqq[SYNC] = NULL;
1219 }
Jens Axboe89850f72006-07-22 16:48:31 +02001220}
1221
Jens Axboe22e2c502005-06-27 10:55:12 +02001222static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1223{
Al Viro478a82b2006-03-18 13:25:24 -05001224 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001225
Jens Axboe89850f72006-07-22 16:48:31 +02001226 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001227 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001228
Jens Axboefc463792006-08-29 09:05:44 +02001229 spin_lock_irq(q->queue_lock);
Jens Axboe89850f72006-07-22 16:48:31 +02001230 __cfq_exit_single_io_context(cfqd, cic);
Jens Axboefc463792006-08-29 09:05:44 +02001231 spin_unlock_irq(q->queue_lock);
Al Viro12a05732006-03-18 13:38:01 -05001232 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001233}
1234
Jens Axboe498d3aa22007-04-26 12:54:48 +02001235/*
1236 * The process that ioc belongs to has exited, we need to clean up
1237 * and put the internal structures we have that belongs to that process.
1238 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001239static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240{
Jens Axboe22e2c502005-06-27 10:55:12 +02001241 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001242 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +02001243
Jens Axboe597bc482007-04-24 21:23:53 +02001244 ioc->ioc_data = NULL;
1245
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 /*
1247 * put the reference this task is holding to the various queues
1248 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001249 n = rb_first(&ioc->cic_root);
1250 while (n != NULL) {
1251 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1252
Jens Axboe22e2c502005-06-27 10:55:12 +02001253 cfq_exit_single_io_context(__cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001254 n = rb_next(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256}
1257
Jens Axboe22e2c502005-06-27 10:55:12 +02001258static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001259cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260{
Jens Axboeb5deef92006-07-19 23:39:40 +02001261 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262
Christoph Lameter94f60302007-07-17 04:03:29 -07001263 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1264 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001266 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001267 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001268 cic->dtor = cfq_free_io_context;
1269 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001270 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271 }
1272
1273 return cic;
1274}
1275
Jens Axboe22e2c502005-06-27 10:55:12 +02001276static void cfq_init_prio_data(struct cfq_queue *cfqq)
1277{
1278 struct task_struct *tsk = current;
1279 int ioprio_class;
1280
Jens Axboe3b181522005-06-27 10:56:24 +02001281 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001282 return;
1283
1284 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1285 switch (ioprio_class) {
1286 default:
1287 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1288 case IOPRIO_CLASS_NONE:
1289 /*
1290 * no prio set, place us in the middle of the BE classes
1291 */
1292 cfqq->ioprio = task_nice_ioprio(tsk);
1293 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1294 break;
1295 case IOPRIO_CLASS_RT:
1296 cfqq->ioprio = task_ioprio(tsk);
1297 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1298 break;
1299 case IOPRIO_CLASS_BE:
1300 cfqq->ioprio = task_ioprio(tsk);
1301 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1302 break;
1303 case IOPRIO_CLASS_IDLE:
1304 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1305 cfqq->ioprio = 7;
Jens Axboe3b181522005-06-27 10:56:24 +02001306 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001307 break;
1308 }
1309
1310 /*
1311 * keep track of original prio settings in case we have to temporarily
1312 * elevate the priority of this queue
1313 */
1314 cfqq->org_ioprio = cfqq->ioprio;
1315 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001316 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001317}
1318
Al Viro478a82b2006-03-18 13:25:24 -05001319static inline void changed_ioprio(struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001320{
Al Viro478a82b2006-03-18 13:25:24 -05001321 struct cfq_data *cfqd = cic->key;
1322 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001323 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001324
Jens Axboecaaa5f92006-06-16 11:23:00 +02001325 if (unlikely(!cfqd))
1326 return;
1327
Jens Axboec1b707d2006-10-30 19:54:23 +01001328 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001329
1330 cfqq = cic->cfqq[ASYNC];
1331 if (cfqq) {
1332 struct cfq_queue *new_cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001333 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc->task,
Jens Axboecaaa5f92006-06-16 11:23:00 +02001334 GFP_ATOMIC);
1335 if (new_cfqq) {
1336 cic->cfqq[ASYNC] = new_cfqq;
1337 cfq_put_queue(cfqq);
1338 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001339 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001340
1341 cfqq = cic->cfqq[SYNC];
1342 if (cfqq)
1343 cfq_mark_cfqq_prio_changed(cfqq);
1344
Jens Axboec1b707d2006-10-30 19:54:23 +01001345 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001346}
1347
Jens Axboefc463792006-08-29 09:05:44 +02001348static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349{
Al Viroa6a07632006-03-18 13:26:44 -05001350 struct cfq_io_context *cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001351 struct rb_node *n;
Al Viroa6a07632006-03-18 13:26:44 -05001352
Jens Axboefc463792006-08-29 09:05:44 +02001353 ioc->ioprio_changed = 0;
Al Viroa6a07632006-03-18 13:26:44 -05001354
Jens Axboee2d74ac2006-03-28 08:59:01 +02001355 n = rb_first(&ioc->cic_root);
1356 while (n != NULL) {
1357 cic = rb_entry(n, struct cfq_io_context, rb_node);
Jens Axboe3793c652006-05-30 21:11:04 +02001358
Al Viro478a82b2006-03-18 13:25:24 -05001359 changed_ioprio(cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001360 n = rb_next(n);
1361 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362}
1363
1364static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001365cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1366 struct task_struct *tsk, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001369 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370
1371retry:
Vasily Tarasov91fac312007-04-25 12:29:51 +02001372 cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1373 /* cic always exists here */
1374 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375
1376 if (!cfqq) {
1377 if (new_cfqq) {
1378 cfqq = new_cfqq;
1379 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001380 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001381 /*
1382 * Inform the allocator of the fact that we will
1383 * just repeat this allocation if it fails, to allow
1384 * the allocator to do whatever it needs to attempt to
1385 * free memory.
1386 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001388 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1389 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1390 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 spin_lock_irq(cfqd->queue->queue_lock);
1392 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001393 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001394 cfqq = kmem_cache_alloc_node(cfq_pool,
1395 gfp_mask | __GFP_ZERO,
1396 cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001397 if (!cfqq)
1398 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001399 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400
Jens Axboed9e76202007-04-20 14:27:50 +02001401 RB_CLEAR_NODE(&cfqq->rb_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001402 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 atomic_set(&cfqq->ref, 0);
1405 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001406
Vasily Tarasov91fac312007-04-25 12:29:51 +02001407 if (is_sync) {
Jens Axboea9938002007-04-20 08:55:52 +02001408 cfq_mark_cfqq_idle_window(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001409 cfq_mark_cfqq_sync(cfqq);
1410 }
Jens Axboea9938002007-04-20 08:55:52 +02001411
Jens Axboe3b181522005-06-27 10:56:24 +02001412 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b037442006-07-28 09:48:51 +02001413 cfq_mark_cfqq_queue_new(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001414
Jens Axboe3b181522005-06-27 10:56:24 +02001415 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416 }
1417
1418 if (new_cfqq)
1419 kmem_cache_free(cfq_pool, new_cfqq);
1420
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421out:
1422 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1423 return cfqq;
1424}
1425
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001426static struct cfq_queue **
1427cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1428{
1429 switch(ioprio_class) {
1430 case IOPRIO_CLASS_RT:
1431 return &cfqd->async_cfqq[0][ioprio];
1432 case IOPRIO_CLASS_BE:
1433 return &cfqd->async_cfqq[1][ioprio];
1434 case IOPRIO_CLASS_IDLE:
1435 return &cfqd->async_idle_cfqq;
1436 default:
1437 BUG();
1438 }
1439}
1440
Jens Axboe15c31be2007-07-10 13:43:25 +02001441static struct cfq_queue *
1442cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct task_struct *tsk,
1443 gfp_t gfp_mask)
1444{
1445 const int ioprio = task_ioprio(tsk);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001446 const int ioprio_class = task_ioprio_class(tsk);
1447 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001448 struct cfq_queue *cfqq = NULL;
1449
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001450 if (!is_sync) {
1451 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1452 cfqq = *async_cfqq;
1453 }
1454
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001455 if (!cfqq) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001456 cfqq = cfq_find_alloc_queue(cfqd, is_sync, tsk, gfp_mask);
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001457 if (!cfqq)
1458 return NULL;
1459 }
Jens Axboe15c31be2007-07-10 13:43:25 +02001460
1461 /*
1462 * pin the queue now that it's allocated, scheduler exit will prune it
1463 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001464 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001465 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001466 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001467 }
1468
1469 atomic_inc(&cfqq->ref);
1470 return cfqq;
1471}
1472
Jens Axboe498d3aa22007-04-26 12:54:48 +02001473/*
1474 * We drop cfq io contexts lazily, so we may find a dead one.
1475 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001476static void
1477cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1478{
Jens Axboefc463792006-08-29 09:05:44 +02001479 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001480
1481 if (ioc->ioc_data == cic)
1482 ioc->ioc_data = NULL;
1483
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001484 rb_erase(&cic->rb_node, &ioc->cic_root);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001485 kmem_cache_free(cfq_ioc_pool, cic);
Jens Axboe4050cf12006-07-19 05:07:12 +02001486 elv_ioc_count_dec(ioc_count);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001487}
1488
Jens Axboee2d74ac2006-03-28 08:59:01 +02001489static struct cfq_io_context *
1490cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1491{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001492 struct rb_node *n;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001493 struct cfq_io_context *cic;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001494 void *k, *key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001495
Vasily Tarasov91fac312007-04-25 12:29:51 +02001496 if (unlikely(!ioc))
1497 return NULL;
1498
Jens Axboe597bc482007-04-24 21:23:53 +02001499 /*
1500 * we maintain a last-hit cache, to avoid browsing over the tree
1501 */
1502 cic = ioc->ioc_data;
1503 if (cic && cic->key == cfqd)
1504 return cic;
1505
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001506restart:
1507 n = ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001508 while (n) {
1509 cic = rb_entry(n, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001510 /* ->key must be copied to avoid race with cfq_exit_queue() */
1511 k = cic->key;
1512 if (unlikely(!k)) {
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001513 cfq_drop_dead_cic(ioc, cic);
1514 goto restart;
1515 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001516
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001517 if (key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001518 n = n->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001519 else if (key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001520 n = n->rb_right;
Jens Axboe597bc482007-04-24 21:23:53 +02001521 else {
1522 ioc->ioc_data = cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001523 return cic;
Jens Axboe597bc482007-04-24 21:23:53 +02001524 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001525 }
1526
1527 return NULL;
1528}
1529
1530static inline void
1531cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1532 struct cfq_io_context *cic)
1533{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001534 struct rb_node **p;
1535 struct rb_node *parent;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001536 struct cfq_io_context *__cic;
Jens Axboe0261d682006-10-30 19:07:48 +01001537 unsigned long flags;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001538 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001539
Jens Axboee2d74ac2006-03-28 08:59:01 +02001540 cic->ioc = ioc;
1541 cic->key = cfqd;
1542
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001543restart:
1544 parent = NULL;
1545 p = &ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001546 while (*p) {
1547 parent = *p;
1548 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001549 /* ->key must be copied to avoid race with cfq_exit_queue() */
1550 k = __cic->key;
1551 if (unlikely(!k)) {
Oleg Nesterovbe33c3a2006-08-21 08:36:12 +02001552 cfq_drop_dead_cic(ioc, __cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001553 goto restart;
1554 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001555
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001556 if (cic->key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001557 p = &(*p)->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001558 else if (cic->key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001559 p = &(*p)->rb_right;
1560 else
1561 BUG();
1562 }
1563
1564 rb_link_node(&cic->rb_node, parent, p);
1565 rb_insert_color(&cic->rb_node, &ioc->cic_root);
Jens Axboefc463792006-08-29 09:05:44 +02001566
Jens Axboe0261d682006-10-30 19:07:48 +01001567 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001568 list_add(&cic->queue_list, &cfqd->cic_list);
Jens Axboe0261d682006-10-30 19:07:48 +01001569 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001570}
1571
Jens Axboe22e2c502005-06-27 10:55:12 +02001572/*
1573 * Setup general io context and cfq io context. There can be several cfq
1574 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001575 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001576 */
1577static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001578cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579{
Jens Axboe22e2c502005-06-27 10:55:12 +02001580 struct io_context *ioc = NULL;
1581 struct cfq_io_context *cic;
1582
1583 might_sleep_if(gfp_mask & __GFP_WAIT);
1584
Jens Axboeb5deef92006-07-19 23:39:40 +02001585 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001586 if (!ioc)
1587 return NULL;
1588
Jens Axboee2d74ac2006-03-28 08:59:01 +02001589 cic = cfq_cic_rb_lookup(cfqd, ioc);
1590 if (cic)
1591 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001592
Jens Axboee2d74ac2006-03-28 08:59:01 +02001593 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1594 if (cic == NULL)
1595 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001596
Jens Axboee2d74ac2006-03-28 08:59:01 +02001597 cfq_cic_link(cfqd, ioc, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001598out:
Jens Axboefc463792006-08-29 09:05:44 +02001599 smp_read_barrier_depends();
1600 if (unlikely(ioc->ioprio_changed))
1601 cfq_ioc_set_ioprio(ioc);
1602
Jens Axboe22e2c502005-06-27 10:55:12 +02001603 return cic;
1604err:
1605 put_io_context(ioc);
1606 return NULL;
1607}
1608
1609static void
1610cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1611{
Jens Axboeaaf12282007-01-19 11:30:16 +11001612 unsigned long elapsed = jiffies - cic->last_end_request;
1613 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001614
1615 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1616 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1617 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1618}
1619
Jens Axboe206dc692006-03-28 13:03:44 +02001620static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001621cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1622 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001623{
1624 sector_t sdist;
1625 u64 total;
1626
Jens Axboe5e705372006-07-13 12:39:25 +02001627 if (cic->last_request_pos < rq->sector)
1628 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001629 else
Jens Axboe5e705372006-07-13 12:39:25 +02001630 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001631
1632 /*
1633 * Don't allow the seek distance to get too large from the
1634 * odd fragment, pagein, etc
1635 */
1636 if (cic->seek_samples <= 60) /* second&third seek */
1637 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1638 else
1639 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1640
1641 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1642 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1643 total = cic->seek_total + (cic->seek_samples/2);
1644 do_div(total, cic->seek_samples);
1645 cic->seek_mean = (sector_t)total;
1646}
Jens Axboe22e2c502005-06-27 10:55:12 +02001647
1648/*
1649 * Disable idle window if the process thinks too long or seeks so much that
1650 * it doesn't matter
1651 */
1652static void
1653cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1654 struct cfq_io_context *cic)
1655{
Jens Axboe1be92f22007-04-19 14:32:26 +02001656 int enable_idle;
1657
1658 if (!cfq_cfqq_sync(cfqq))
1659 return;
1660
1661 enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001662
Jens Axboecaaa5f92006-06-16 11:23:00 +02001663 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1664 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001665 enable_idle = 0;
1666 else if (sample_valid(cic->ttime_samples)) {
1667 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1668 enable_idle = 0;
1669 else
1670 enable_idle = 1;
1671 }
1672
Jens Axboe3b181522005-06-27 10:56:24 +02001673 if (enable_idle)
1674 cfq_mark_cfqq_idle_window(cfqq);
1675 else
1676 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001677}
1678
Jens Axboe22e2c502005-06-27 10:55:12 +02001679/*
1680 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1681 * no or if we aren't sure, a 1 will cause a preempt.
1682 */
1683static int
1684cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001685 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001686{
Jens Axboe6d048f52007-04-25 12:44:27 +02001687 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001688
Jens Axboe6d048f52007-04-25 12:44:27 +02001689 cfqq = cfqd->active_queue;
1690 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001691 return 0;
1692
Jens Axboe6d048f52007-04-25 12:44:27 +02001693 if (cfq_slice_used(cfqq))
1694 return 1;
1695
1696 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001697 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001698
1699 if (cfq_class_idle(cfqq))
1700 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001701
Jens Axboe22e2c502005-06-27 10:55:12 +02001702 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02001703 * if the new request is sync, but the currently running queue is
1704 * not, let the sync request have priority.
1705 */
Jens Axboe5e705372006-07-13 12:39:25 +02001706 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001707 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001708
Jens Axboe374f84a2006-07-23 01:42:19 +02001709 /*
1710 * So both queues are sync. Let the new request get disk time if
1711 * it's a metadata request and the current queue is doing regular IO.
1712 */
1713 if (rq_is_meta(rq) && !cfqq->meta_pending)
1714 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001715
Jens Axboe1e3335d2007-02-14 19:59:49 +01001716 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1717 return 0;
1718
1719 /*
1720 * if this request is as-good as one we would expect from the
1721 * current cfqq, let it preempt
1722 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001723 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01001724 return 1;
1725
Jens Axboe22e2c502005-06-27 10:55:12 +02001726 return 0;
1727}
1728
1729/*
1730 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1731 * let it have half of its nominal slice.
1732 */
1733static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1734{
Jens Axboe6084cdd2007-04-23 08:25:00 +02001735 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001736
Jens Axboebf572252006-07-19 20:29:12 +02001737 /*
1738 * Put the new queue at the front of the of the current list,
1739 * so we know that it will be selected next.
1740 */
1741 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02001742
1743 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02001744
Jens Axboe44f7c162007-01-19 11:51:58 +11001745 cfqq->slice_end = 0;
1746 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001747}
1748
1749/*
Jens Axboe5e705372006-07-13 12:39:25 +02001750 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001751 * something we should do about it
1752 */
1753static void
Jens Axboe5e705372006-07-13 12:39:25 +02001754cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1755 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001756{
Jens Axboe5e705372006-07-13 12:39:25 +02001757 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001758
Jens Axboe374f84a2006-07-23 01:42:19 +02001759 if (rq_is_meta(rq))
1760 cfqq->meta_pending++;
1761
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001762 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02001763 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001764 cfq_update_idle_window(cfqd, cfqq, cic);
1765
Jens Axboe5e705372006-07-13 12:39:25 +02001766 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001767
1768 if (cfqq == cfqd->active_queue) {
1769 /*
1770 * if we are waiting for a request for this queue, let it rip
1771 * immediately and flag that we must not expire this queue
1772 * just now
1773 */
Jens Axboe3b181522005-06-27 10:56:24 +02001774 if (cfq_cfqq_wait_request(cfqq)) {
1775 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001776 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001777 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001778 }
Jens Axboe5e705372006-07-13 12:39:25 +02001779 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001780 /*
1781 * not the active queue - expire current slice if it is
1782 * idle and has expired it's mean thinktime or this new queue
1783 * has some old slice time left and is of higher priority
1784 */
1785 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001786 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001787 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001788 }
1789}
1790
Jens Axboe165125e2007-07-24 09:28:11 +02001791static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001792{
Jens Axboeb4878f22005-10-20 16:42:29 +02001793 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001794 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001795
1796 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797
Jens Axboe5e705372006-07-13 12:39:25 +02001798 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799
Jens Axboe22e2c502005-06-27 10:55:12 +02001800 list_add_tail(&rq->queuelist, &cfqq->fifo);
1801
Jens Axboe5e705372006-07-13 12:39:25 +02001802 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803}
1804
Jens Axboe165125e2007-07-24 09:28:11 +02001805static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806{
Jens Axboe5e705372006-07-13 12:39:25 +02001807 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001808 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001809 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001810 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811
Jens Axboeb4878f22005-10-20 16:42:29 +02001812 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813
Jens Axboeb4878f22005-10-20 16:42:29 +02001814 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02001815 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02001816 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02001817 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818
Jens Axboe3ed9a292007-04-23 08:33:33 +02001819 if (cfq_cfqq_sync(cfqq))
1820 cfqd->sync_flight--;
1821
Jens Axboeb4878f22005-10-20 16:42:29 +02001822 if (!cfq_class_idle(cfqq))
1823 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001824
Jens Axboecaaa5f92006-06-16 11:23:00 +02001825 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001826 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001827
1828 /*
1829 * If this is the active queue, check if it needs to be expired,
1830 * or if we want to idle in case it has no pending requests.
1831 */
1832 if (cfqd->active_queue == cfqq) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001833 if (cfq_cfqq_slice_new(cfqq)) {
1834 cfq_set_prio_slice(cfqd, cfqq);
1835 cfq_clear_cfqq_slice_new(cfqq);
1836 }
1837 if (cfq_slice_used(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02001838 cfq_slice_expired(cfqd, 1);
Jens Axboe6d048f52007-04-25 12:44:27 +02001839 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1840 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001841 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001842
1843 if (!cfqd->rq_in_driver)
1844 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845}
1846
Jens Axboe22e2c502005-06-27 10:55:12 +02001847/*
1848 * we temporarily boost lower priority queues if they are holding fs exclusive
1849 * resources. they are boosted to normal prio (CLASS_BE/4)
1850 */
1851static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852{
Jens Axboe22e2c502005-06-27 10:55:12 +02001853 if (has_fs_excl()) {
1854 /*
1855 * boost idle prio on transactions that would lock out other
1856 * users of the filesystem
1857 */
1858 if (cfq_class_idle(cfqq))
1859 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1860 if (cfqq->ioprio > IOPRIO_NORM)
1861 cfqq->ioprio = IOPRIO_NORM;
1862 } else {
1863 /*
1864 * check if we need to unboost the queue
1865 */
1866 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1867 cfqq->ioprio_class = cfqq->org_ioprio_class;
1868 if (cfqq->ioprio != cfqq->org_ioprio)
1869 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001871}
1872
Jens Axboe89850f72006-07-22 16:48:31 +02001873static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001874{
Jens Axboe3b181522005-06-27 10:56:24 +02001875 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001876 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001877 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001878 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001879 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001880
1881 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02001882}
1883
Jens Axboe165125e2007-07-24 09:28:11 +02001884static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02001885{
1886 struct cfq_data *cfqd = q->elevator->elevator_data;
1887 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001888 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02001889 struct cfq_queue *cfqq;
1890
1891 /*
1892 * don't force setup of a queue from here, as a call to may_queue
1893 * does not necessarily imply that a request actually will be queued.
1894 * so just lookup a possibly existing queue, or return 'may queue'
1895 * if that fails
1896 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001897 cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1898 if (!cic)
1899 return ELV_MQUEUE_MAY;
1900
1901 cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02001902 if (cfqq) {
1903 cfq_init_prio_data(cfqq);
1904 cfq_prio_boost(cfqq);
1905
Jens Axboe89850f72006-07-22 16:48:31 +02001906 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001907 }
1908
1909 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910}
1911
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912/*
1913 * queue lock held here
1914 */
Jens Axboebb37b942006-12-01 10:42:33 +01001915static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916{
Jens Axboe5e705372006-07-13 12:39:25 +02001917 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918
Jens Axboe5e705372006-07-13 12:39:25 +02001919 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001920 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921
Jens Axboe22e2c502005-06-27 10:55:12 +02001922 BUG_ON(!cfqq->allocated[rw]);
1923 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924
Jens Axboe5e705372006-07-13 12:39:25 +02001925 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02001928 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 cfq_put_queue(cfqq);
1931 }
1932}
1933
1934/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001935 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001937static int
Jens Axboe165125e2007-07-24 09:28:11 +02001938cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939{
1940 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe3b181522005-06-27 10:56:24 +02001941 struct task_struct *tsk = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942 struct cfq_io_context *cic;
1943 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01001944 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001945 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946 unsigned long flags;
1947
1948 might_sleep_if(gfp_mask & __GFP_WAIT);
1949
Jens Axboee2d74ac2006-03-28 08:59:01 +02001950 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001951
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952 spin_lock_irqsave(q->queue_lock, flags);
1953
Jens Axboe22e2c502005-06-27 10:55:12 +02001954 if (!cic)
1955 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956
Vasily Tarasov91fac312007-04-25 12:29:51 +02001957 cfqq = cic_to_cfqq(cic, is_sync);
1958 if (!cfqq) {
1959 cfqq = cfq_get_queue(cfqd, is_sync, tsk, gfp_mask);
1960
Jens Axboe22e2c502005-06-27 10:55:12 +02001961 if (!cfqq)
1962 goto queue_fail;
1963
Vasily Tarasov91fac312007-04-25 12:29:51 +02001964 cic_set_cfqq(cic, cfqq, is_sync);
1965 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966
1967 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02001968 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001969 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02001970
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971 spin_unlock_irqrestore(q->queue_lock, flags);
1972
Jens Axboe5e705372006-07-13 12:39:25 +02001973 rq->elevator_private = cic;
1974 rq->elevator_private2 = cfqq;
1975 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02001976
Jens Axboe22e2c502005-06-27 10:55:12 +02001977queue_fail:
1978 if (cic)
1979 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02001980
Jens Axboe3b181522005-06-27 10:56:24 +02001981 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 spin_unlock_irqrestore(q->queue_lock, flags);
1983 return 1;
1984}
1985
David Howells65f27f32006-11-22 14:55:48 +00001986static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02001987{
David Howells65f27f32006-11-22 14:55:48 +00001988 struct cfq_data *cfqd =
1989 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02001990 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001991 unsigned long flags;
1992
1993 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02001994 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02001995 spin_unlock_irqrestore(q->queue_lock, flags);
1996}
1997
1998/*
1999 * Timer running if the active_queue is currently idling inside its time slice
2000 */
2001static void cfq_idle_slice_timer(unsigned long data)
2002{
2003 struct cfq_data *cfqd = (struct cfq_data *) data;
2004 struct cfq_queue *cfqq;
2005 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002006 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002007
2008 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2009
2010 if ((cfqq = cfqd->active_queue) != NULL) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002011 timed_out = 0;
2012
Jens Axboe22e2c502005-06-27 10:55:12 +02002013 /*
2014 * expired
2015 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002016 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002017 goto expire;
2018
2019 /*
2020 * only expire and reinvoke request handler, if there are
2021 * other queues with pending requests
2022 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002023 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002024 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002025
2026 /*
2027 * not expired and it has a request pending, let it dispatch
2028 */
Jens Axboedd67d052006-06-21 09:36:18 +02002029 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002030 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002031 goto out_kick;
2032 }
2033 }
2034expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002035 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002036out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002037 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002038out_cont:
2039 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2040}
2041
2042/*
2043 * Timer running if an idle class queue is waiting for service
2044 */
2045static void cfq_idle_class_timer(unsigned long data)
2046{
2047 struct cfq_data *cfqd = (struct cfq_data *) data;
Oleg Nesterov0e7be9e2007-11-07 13:51:35 +01002048 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002049
2050 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2051
2052 /*
2053 * race with a non-idle queue, reset timer
2054 */
Oleg Nesterov0e7be9e2007-11-07 13:51:35 +01002055 if (!start_idle_class_timer(cfqd))
Jens Axboe3b181522005-06-27 10:56:24 +02002056 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002057
2058 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2059}
2060
Jens Axboe3b181522005-06-27 10:56:24 +02002061static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2062{
2063 del_timer_sync(&cfqd->idle_slice_timer);
2064 del_timer_sync(&cfqd->idle_class_timer);
Oleg Nesterov43108642007-10-23 15:08:18 +02002065 kblockd_flush_work(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002066}
Jens Axboe22e2c502005-06-27 10:55:12 +02002067
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002068static void cfq_put_async_queues(struct cfq_data *cfqd)
2069{
2070 int i;
2071
2072 for (i = 0; i < IOPRIO_BE_NR; i++) {
2073 if (cfqd->async_cfqq[0][i])
2074 cfq_put_queue(cfqd->async_cfqq[0][i]);
2075 if (cfqd->async_cfqq[1][i])
2076 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002077 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002078
2079 if (cfqd->async_idle_cfqq)
2080 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002081}
2082
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083static void cfq_exit_queue(elevator_t *e)
2084{
Jens Axboe22e2c502005-06-27 10:55:12 +02002085 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002086 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002087
Jens Axboe3b181522005-06-27 10:56:24 +02002088 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002089
Al Virod9ff4182006-03-18 13:51:22 -05002090 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002091
Al Virod9ff4182006-03-18 13:51:22 -05002092 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002093 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002094
2095 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002096 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2097 struct cfq_io_context,
2098 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002099
2100 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002101 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002102
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002103 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002104
Al Virod9ff4182006-03-18 13:51:22 -05002105 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002106
2107 cfq_shutdown_timer_wq(cfqd);
2108
Al Viroa90d7422006-03-18 12:05:37 -05002109 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110}
2111
Jens Axboe165125e2007-07-24 09:28:11 +02002112static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113{
2114 struct cfq_data *cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115
Christoph Lameter94f60302007-07-17 04:03:29 -07002116 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002118 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119
Jens Axboecc09e292007-04-26 12:53:50 +02002120 cfqd->service_tree = CFQ_RB_ROOT;
Al Virod9ff4182006-03-18 13:51:22 -05002121 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124
Jens Axboe22e2c502005-06-27 10:55:12 +02002125 init_timer(&cfqd->idle_slice_timer);
2126 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2127 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2128
2129 init_timer(&cfqd->idle_class_timer);
2130 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2131 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2132
David Howells65f27f32006-11-22 14:55:48 +00002133 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002134
Oleg Nesterovb70c8642007-11-07 09:46:13 +01002135 cfqd->last_end_request = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002137 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2138 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 cfqd->cfq_back_max = cfq_back_max;
2140 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002141 cfqd->cfq_slice[0] = cfq_slice_async;
2142 cfqd->cfq_slice[1] = cfq_slice_sync;
2143 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2144 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02002145
Jens Axboebc1c1162006-06-08 08:49:06 +02002146 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147}
2148
2149static void cfq_slab_kill(void)
2150{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151 if (cfq_pool)
2152 kmem_cache_destroy(cfq_pool);
2153 if (cfq_ioc_pool)
2154 kmem_cache_destroy(cfq_ioc_pool);
2155}
2156
2157static int __init cfq_slab_setup(void)
2158{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002159 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160 if (!cfq_pool)
2161 goto fail;
2162
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002163 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 if (!cfq_ioc_pool)
2165 goto fail;
2166
2167 return 0;
2168fail:
2169 cfq_slab_kill();
2170 return -ENOMEM;
2171}
2172
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173/*
2174 * sysfs parts below -->
2175 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176static ssize_t
2177cfq_var_show(unsigned int var, char *page)
2178{
2179 return sprintf(page, "%d\n", var);
2180}
2181
2182static ssize_t
2183cfq_var_store(unsigned int *var, const char *page, size_t count)
2184{
2185 char *p = (char *) page;
2186
2187 *var = simple_strtoul(p, &p, 10);
2188 return count;
2189}
2190
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002192static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002194 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195 unsigned int __data = __VAR; \
2196 if (__CONV) \
2197 __data = jiffies_to_msecs(__data); \
2198 return cfq_var_show(__data, (page)); \
2199}
2200SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002201SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2202SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002203SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2204SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002205SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2206SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2207SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2208SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209#undef SHOW_FUNCTION
2210
2211#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002212static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002214 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 unsigned int __data; \
2216 int ret = cfq_var_store(&__data, (page), count); \
2217 if (__data < (MIN)) \
2218 __data = (MIN); \
2219 else if (__data > (MAX)) \
2220 __data = (MAX); \
2221 if (__CONV) \
2222 *(__PTR) = msecs_to_jiffies(__data); \
2223 else \
2224 *(__PTR) = __data; \
2225 return ret; \
2226}
2227STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002228STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2229STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002230STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2231STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002232STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2233STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2234STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2235STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236#undef STORE_FUNCTION
2237
Al Viroe572ec72006-03-18 22:27:18 -05002238#define CFQ_ATTR(name) \
2239 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002240
Al Viroe572ec72006-03-18 22:27:18 -05002241static struct elv_fs_entry cfq_attrs[] = {
2242 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002243 CFQ_ATTR(fifo_expire_sync),
2244 CFQ_ATTR(fifo_expire_async),
2245 CFQ_ATTR(back_seek_max),
2246 CFQ_ATTR(back_seek_penalty),
2247 CFQ_ATTR(slice_sync),
2248 CFQ_ATTR(slice_async),
2249 CFQ_ATTR(slice_async_rq),
2250 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002251 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252};
2253
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254static struct elevator_type iosched_cfq = {
2255 .ops = {
2256 .elevator_merge_fn = cfq_merge,
2257 .elevator_merged_fn = cfq_merged_request,
2258 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002259 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002260 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002262 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 .elevator_deactivate_req_fn = cfq_deactivate_request,
2264 .elevator_queue_empty_fn = cfq_queue_empty,
2265 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002266 .elevator_former_req_fn = elv_rb_former_request,
2267 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 .elevator_set_req_fn = cfq_set_request,
2269 .elevator_put_req_fn = cfq_put_request,
2270 .elevator_may_queue_fn = cfq_may_queue,
2271 .elevator_init_fn = cfq_init_queue,
2272 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002273 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 },
Al Viro3d1ab402006-03-18 18:35:43 -05002275 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 .elevator_name = "cfq",
2277 .elevator_owner = THIS_MODULE,
2278};
2279
2280static int __init cfq_init(void)
2281{
Jens Axboe22e2c502005-06-27 10:55:12 +02002282 /*
2283 * could be 0 on HZ < 1000 setups
2284 */
2285 if (!cfq_slice_async)
2286 cfq_slice_async = 1;
2287 if (!cfq_slice_idle)
2288 cfq_slice_idle = 1;
2289
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290 if (cfq_slab_setup())
2291 return -ENOMEM;
2292
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002293 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002295 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296}
2297
2298static void __exit cfq_exit(void)
2299{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002300 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002302 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002303 /* ioc_gone's update must be visible before reading ioc_count */
2304 smp_wmb();
Jens Axboe4050cf12006-07-19 05:07:12 +02002305 if (elv_ioc_count_read(ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002306 wait_for_completion(ioc_gone);
Al Viro334e94d2006-03-18 15:05:53 -05002307 synchronize_rcu();
Christoph Hellwig83521d32005-10-30 15:01:39 -08002308 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309}
2310
2311module_init(cfq_init);
2312module_exit(cfq_exit);
2313
2314MODULE_AUTHOR("Jens Axboe");
2315MODULE_LICENSE("GPL");
2316MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");