blob: baef5fc7cff8b5ca3d184e3d48efdd4be7376203 [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 Axboe22e2c502005-06-27 10:55:12 +020074 request_queue_t *queue;
75
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
95 struct timer_list idle_class_timer;
Linus Torvalds1da177e2005-04-16 15:20:36 -070096
Jens Axboe6d048f52007-04-25 12:44:27 +020097 sector_t last_position;
Jens Axboe22e2c502005-06-27 10:55:12 +020098 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100 /*
101 * tunables, see top of file
102 */
103 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200104 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105 unsigned int cfq_back_penalty;
106 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200107 unsigned int cfq_slice[2];
108 unsigned int cfq_slice_async_rq;
109 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500110
111 struct list_head cic_list;
Jens Axboe6d048f52007-04-25 12:44:27 +0200112
113 sector_t new_seek_mean;
114 u64 new_seek_total;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115};
116
Jens Axboe22e2c502005-06-27 10:55:12 +0200117/*
118 * Per process-grouping structure
119 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120struct cfq_queue {
121 /* reference count */
122 atomic_t ref;
123 /* parent cfq_data */
124 struct cfq_data *cfqd;
Jens Axboed9e76202007-04-20 14:27:50 +0200125 /* service_tree member */
126 struct rb_node rb_node;
127 /* service_tree key */
128 unsigned long rb_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 /* sorted list of pending requests */
130 struct rb_root sort_list;
131 /* if fifo isn't expired, next request to serve */
Jens Axboe5e705372006-07-13 12:39:25 +0200132 struct request *next_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 /* requests queued in sort_list */
134 int queued[2];
135 /* currently allocated requests */
136 int allocated[2];
Jens Axboe374f84a2006-07-23 01:42:19 +0200137 /* pending metadata requests */
138 int meta_pending;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200140 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
Jens Axboe22e2c502005-06-27 10:55:12 +0200142 unsigned long slice_end;
Jens Axboec5b680f2007-01-19 11:56:49 +1100143 long slice_resid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144
Jens Axboe6d048f52007-04-25 12:44:27 +0200145 /* number of requests that are on the dispatch list or inside driver */
146 int dispatched;
Jens Axboe22e2c502005-06-27 10:55:12 +0200147
148 /* io prio of this group */
149 unsigned short ioprio, org_ioprio;
150 unsigned short ioprio_class, org_ioprio_class;
151
Jens Axboe3b181522005-06-27 10:56:24 +0200152 /* various state flags, see below */
153 unsigned int flags;
Jens Axboe6d048f52007-04-25 12:44:27 +0200154
155 sector_t last_request_pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156};
157
Jens Axboe3b181522005-06-27 10:56:24 +0200158enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100159 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
160 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
161 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
162 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
163 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
164 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
165 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
166 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
167 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
Jens Axboe44f7c162007-01-19 11:51:58 +1100168 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200169 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200170};
171
172#define CFQ_CFQQ_FNS(name) \
173static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
174{ \
175 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
176} \
177static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
178{ \
179 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
180} \
181static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
182{ \
183 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
184}
185
186CFQ_CFQQ_FNS(on_rr);
187CFQ_CFQQ_FNS(wait_request);
188CFQ_CFQQ_FNS(must_alloc);
189CFQ_CFQQ_FNS(must_alloc_slice);
190CFQ_CFQQ_FNS(must_dispatch);
191CFQ_CFQQ_FNS(fifo_expire);
192CFQ_CFQQ_FNS(idle_window);
193CFQ_CFQQ_FNS(prio_changed);
Jens Axboe53b03742006-07-28 09:48:51 +0200194CFQ_CFQQ_FNS(queue_new);
Jens Axboe44f7c162007-01-19 11:51:58 +1100195CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200196CFQ_CFQQ_FNS(sync);
Jens Axboe3b181522005-06-27 10:56:24 +0200197#undef CFQ_CFQQ_FNS
198
Jens Axboe5e705372006-07-13 12:39:25 +0200199static void cfq_dispatch_insert(request_queue_t *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200200static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
201 struct task_struct *, gfp_t);
202static struct cfq_io_context *cfq_cic_rb_lookup(struct cfq_data *,
203 struct io_context *);
204
205static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
206 int is_sync)
207{
208 return cic->cfqq[!!is_sync];
209}
210
211static inline void cic_set_cfqq(struct cfq_io_context *cic,
212 struct cfq_queue *cfqq, int is_sync)
213{
214 cic->cfqq[!!is_sync] = cfqq;
215}
216
217/*
218 * We regard a request as SYNC, if it's either a read or has the SYNC bit
219 * set (in which case it could also be direct WRITE).
220 */
221static inline int cfq_bio_sync(struct bio *bio)
222{
223 if (bio_data_dir(bio) == READ || bio_sync(bio))
224 return 1;
225
226 return 0;
227}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700230 * scheduler run of queue, if there are requests pending and no one in the
231 * driver that will restart queueing
232 */
233static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
234{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100235 if (cfqd->busy_queues)
Andrew Morton99f95e52005-06-27 20:14:05 -0700236 kblockd_schedule_work(&cfqd->unplug_work);
237}
238
239static int cfq_queue_empty(request_queue_t *q)
240{
241 struct cfq_data *cfqd = q->elevator->elevator_data;
242
Jens Axboeb4878f22005-10-20 16:42:29 +0200243 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700244}
245
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100247 * Scale schedule slice based on io priority. Use the sync time slice only
248 * if a queue is marked sync and has sync io queued. A sync queue with async
249 * io only, should not get full sync slice length.
250 */
Jens Axboed9e76202007-04-20 14:27:50 +0200251static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
252 unsigned short prio)
253{
254 const int base_slice = cfqd->cfq_slice[sync];
255
256 WARN_ON(prio >= IOPRIO_BE_NR);
257
258 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
259}
260
Jens Axboe44f7c162007-01-19 11:51:58 +1100261static inline int
262cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
263{
Jens Axboed9e76202007-04-20 14:27:50 +0200264 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100265}
266
267static inline void
268cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
269{
270 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
271}
272
273/*
274 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
275 * isn't valid until the first request from the dispatch is activated
276 * and the slice time set.
277 */
278static inline int cfq_slice_used(struct cfq_queue *cfqq)
279{
280 if (cfq_cfqq_slice_new(cfqq))
281 return 0;
282 if (time_before(jiffies, cfqq->slice_end))
283 return 0;
284
285 return 1;
286}
287
288/*
Jens Axboe5e705372006-07-13 12:39:25 +0200289 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200291 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 */
Jens Axboe5e705372006-07-13 12:39:25 +0200293static struct request *
294cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295{
296 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200298#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
299#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
300 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301
Jens Axboe5e705372006-07-13 12:39:25 +0200302 if (rq1 == NULL || rq1 == rq2)
303 return rq2;
304 if (rq2 == NULL)
305 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200306
Jens Axboe5e705372006-07-13 12:39:25 +0200307 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
308 return rq1;
309 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
310 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200311 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
312 return rq1;
313 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
314 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315
Jens Axboe5e705372006-07-13 12:39:25 +0200316 s1 = rq1->sector;
317 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318
Jens Axboe6d048f52007-04-25 12:44:27 +0200319 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321 /*
322 * by definition, 1KiB is 2 sectors
323 */
324 back_max = cfqd->cfq_back_max * 2;
325
326 /*
327 * Strict one way elevator _except_ in the case where we allow
328 * short backward seeks which are biased as twice the cost of a
329 * similar forward seek.
330 */
331 if (s1 >= last)
332 d1 = s1 - last;
333 else if (s1 + back_max >= last)
334 d1 = (last - s1) * cfqd->cfq_back_penalty;
335 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200336 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337
338 if (s2 >= last)
339 d2 = s2 - last;
340 else if (s2 + back_max >= last)
341 d2 = (last - s2) * cfqd->cfq_back_penalty;
342 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200343 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344
345 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346
Andreas Mohre8a99052006-03-28 08:59:49 +0200347 /*
348 * By doing switch() on the bit mask "wrap" we avoid having to
349 * check two variables for all permutations: --> faster!
350 */
351 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200352 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200353 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200354 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200355 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200356 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200357 else {
358 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200359 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200360 else
Jens Axboe5e705372006-07-13 12:39:25 +0200361 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200362 }
363
364 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200365 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200366 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200367 return rq2;
368 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200369 default:
370 /*
371 * Since both rqs are wrapped,
372 * start with the one that's further behind head
373 * (--> only *one* back seek required),
374 * since back seek takes more time than forward.
375 */
376 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200377 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 else
Jens Axboe5e705372006-07-13 12:39:25 +0200379 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 }
381}
382
Jens Axboe498d3aa2007-04-26 12:54:48 +0200383/*
384 * The below is leftmost cache rbtree addon
385 */
Jens Axboecc09e292007-04-26 12:53:50 +0200386static struct rb_node *cfq_rb_first(struct cfq_rb_root *root)
387{
388 if (!root->left)
389 root->left = rb_first(&root->rb);
390
391 return root->left;
392}
393
394static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
395{
396 if (root->left == n)
397 root->left = NULL;
398
399 rb_erase(n, &root->rb);
400 RB_CLEAR_NODE(n);
401}
402
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403/*
404 * would be nice to take fifo expire time into account as well
405 */
Jens Axboe5e705372006-07-13 12:39:25 +0200406static struct request *
407cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
408 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409{
Jens Axboe21183b02006-07-13 12:33:14 +0200410 struct rb_node *rbnext = rb_next(&last->rb_node);
411 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200412 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413
Jens Axboe21183b02006-07-13 12:33:14 +0200414 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415
416 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200417 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200418
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200420 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200421 else {
422 rbnext = rb_first(&cfqq->sort_list);
423 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200424 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200425 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426
Jens Axboe21183b02006-07-13 12:33:14 +0200427 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428}
429
Jens Axboed9e76202007-04-20 14:27:50 +0200430static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
431 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432{
Jens Axboed9e76202007-04-20 14:27:50 +0200433 /*
434 * just an approximation, should be ok.
435 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200436 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
437 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200438}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439
Jens Axboe498d3aa2007-04-26 12:54:48 +0200440/*
441 * The cfqd->service_tree holds all pending cfq_queue's that have
442 * requests waiting to be processed. It is sorted in the order that
443 * we will service the queues.
444 */
Jens Axboed9e76202007-04-20 14:27:50 +0200445static void cfq_service_tree_add(struct cfq_data *cfqd,
Jens Axboeedd75ff2007-04-19 12:03:34 +0200446 struct cfq_queue *cfqq, int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200447{
Jens Axboecc09e292007-04-26 12:53:50 +0200448 struct rb_node **p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200449 struct rb_node *parent = NULL;
Jens Axboed9e76202007-04-20 14:27:50 +0200450 unsigned long rb_key;
Jens Axboe498d3aa2007-04-26 12:54:48 +0200451 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200452
Jens Axboeedd75ff2007-04-19 12:03:34 +0200453 if (!add_front) {
454 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
455 rb_key += cfqq->slice_resid;
456 cfqq->slice_resid = 0;
457 } else
458 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200459
460 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100461 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200462 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100463 */
Jens Axboed9e76202007-04-20 14:27:50 +0200464 if (rb_key == cfqq->rb_key)
465 return;
Jens Axboe53b03742006-07-28 09:48:51 +0200466
Jens Axboecc09e292007-04-26 12:53:50 +0200467 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200468 }
Jens Axboed9e76202007-04-20 14:27:50 +0200469
Jens Axboe498d3aa2007-04-26 12:54:48 +0200470 left = 1;
Jens Axboed9e76202007-04-20 14:27:50 +0200471 while (*p) {
Jens Axboecc09e292007-04-26 12:53:50 +0200472 struct cfq_queue *__cfqq;
Jens Axboe67060e32007-04-18 20:13:32 +0200473 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200474
Jens Axboed9e76202007-04-20 14:27:50 +0200475 parent = *p;
476 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
477
Jens Axboe0c534e02007-04-18 20:01:57 +0200478 /*
479 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200480 * preference to them. IDLE queues goes to the back.
481 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200482 */
483 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200484 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200485 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200486 n = &(*p)->rb_right;
487 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
488 n = &(*p)->rb_left;
489 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
490 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200491 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200492 n = &(*p)->rb_left;
493 else
494 n = &(*p)->rb_right;
495
496 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200497 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200498
499 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200500 }
501
Jens Axboecc09e292007-04-26 12:53:50 +0200502 if (left)
503 cfqd->service_tree.left = &cfqq->rb_node;
504
Jens Axboed9e76202007-04-20 14:27:50 +0200505 cfqq->rb_key = rb_key;
506 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200507 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508}
509
Jens Axboe498d3aa2007-04-26 12:54:48 +0200510/*
511 * Update cfqq's position in the service tree.
512 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200513static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200514{
Jens Axboe6d048f52007-04-25 12:44:27 +0200515 /*
516 * Resorting requires the cfqq to be on the RR list already.
517 */
Jens Axboe498d3aa2007-04-26 12:54:48 +0200518 if (cfq_cfqq_on_rr(cfqq))
Jens Axboeedd75ff2007-04-19 12:03:34 +0200519 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboe6d048f52007-04-25 12:44:27 +0200520}
521
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522/*
523 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200524 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 */
526static inline void
Jens Axboeb4878f22005-10-20 16:42:29 +0200527cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528{
Jens Axboe3b181522005-06-27 10:56:24 +0200529 BUG_ON(cfq_cfqq_on_rr(cfqq));
530 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531 cfqd->busy_queues++;
532
Jens Axboeedd75ff2007-04-19 12:03:34 +0200533 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534}
535
Jens Axboe498d3aa2007-04-26 12:54:48 +0200536/*
537 * Called when the cfqq no longer has requests pending, remove it from
538 * the service tree.
539 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540static inline void
541cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
542{
Jens Axboe3b181522005-06-27 10:56:24 +0200543 BUG_ON(!cfq_cfqq_on_rr(cfqq));
544 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545
Jens Axboecc09e292007-04-26 12:53:50 +0200546 if (!RB_EMPTY_NODE(&cfqq->rb_node))
547 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboed9e76202007-04-20 14:27:50 +0200548
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 BUG_ON(!cfqd->busy_queues);
550 cfqd->busy_queues--;
551}
552
553/*
554 * rb tree support functions
555 */
Jens Axboe5e705372006-07-13 12:39:25 +0200556static inline void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557{
Jens Axboe5e705372006-07-13 12:39:25 +0200558 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200559 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200560 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561
Jens Axboeb4878f22005-10-20 16:42:29 +0200562 BUG_ON(!cfqq->queued[sync]);
563 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564
Jens Axboe5e705372006-07-13 12:39:25 +0200565 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566
Jens Axboedd67d052006-06-21 09:36:18 +0200567 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200568 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569}
570
Jens Axboe5e705372006-07-13 12:39:25 +0200571static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572{
Jens Axboe5e705372006-07-13 12:39:25 +0200573 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe21183b02006-07-13 12:33:14 +0200575 struct request *__alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576
Jens Axboe5380a102006-07-13 12:37:56 +0200577 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578
579 /*
580 * looks a little odd, but the first insert might return an alias.
581 * if that happens, put the alias on the dispatch list
582 */
Jens Axboe21183b02006-07-13 12:33:14 +0200583 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200584 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100585
586 if (!cfq_cfqq_on_rr(cfqq))
587 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200588
589 /*
590 * check if this request is a better next-serve candidate
591 */
592 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
593 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594}
595
596static inline void
Jens Axboe5e705372006-07-13 12:39:25 +0200597cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598{
Jens Axboe5380a102006-07-13 12:37:56 +0200599 elv_rb_del(&cfqq->sort_list, rq);
600 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200601 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602}
603
Jens Axboe206dc692006-03-28 13:03:44 +0200604static struct request *
605cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606{
Jens Axboe206dc692006-03-28 13:03:44 +0200607 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200608 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200609 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610
Vasily Tarasov91fac312007-04-25 12:29:51 +0200611 cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
612 if (!cic)
613 return NULL;
614
615 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200616 if (cfqq) {
617 sector_t sector = bio->bi_sector + bio_sectors(bio);
618
Jens Axboe21183b02006-07-13 12:33:14 +0200619 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200620 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 return NULL;
623}
624
Jens Axboeb4878f22005-10-20 16:42:29 +0200625static void cfq_activate_request(request_queue_t *q, struct request *rq)
626{
627 struct cfq_data *cfqd = q->elevator->elevator_data;
628
629 cfqd->rq_in_driver++;
Jens Axboe25776e32006-06-01 10:12:26 +0200630
631 /*
632 * If the depth is larger 1, it really could be queueing. But lets
633 * make the mark a little higher - idling could still be good for
634 * low queueing, and a low queueing number could also just indicate
635 * a SCSI mid layer like behaviour where limit+1 is often seen.
636 */
637 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
638 cfqd->hw_tag = 1;
Jens Axboe6d048f52007-04-25 12:44:27 +0200639
640 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
Jens Axboeb4878f22005-10-20 16:42:29 +0200641}
642
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
644{
Jens Axboe22e2c502005-06-27 10:55:12 +0200645 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646
Jens Axboeb4878f22005-10-20 16:42:29 +0200647 WARN_ON(!cfqd->rq_in_driver);
648 cfqd->rq_in_driver--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649}
650
Jens Axboeb4878f22005-10-20 16:42:29 +0200651static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652{
Jens Axboe5e705372006-07-13 12:39:25 +0200653 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200654
Jens Axboe5e705372006-07-13 12:39:25 +0200655 if (cfqq->next_rq == rq)
656 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657
Jens Axboeb4878f22005-10-20 16:42:29 +0200658 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200659 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200660
661 if (rq_is_meta(rq)) {
662 WARN_ON(!cfqq->meta_pending);
663 cfqq->meta_pending--;
664 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665}
666
Jens Axboe498d3aa2007-04-26 12:54:48 +0200667static int cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668{
669 struct cfq_data *cfqd = q->elevator->elevator_data;
670 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671
Jens Axboe206dc692006-03-28 13:03:44 +0200672 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200673 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200674 *req = __rq;
675 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 }
677
678 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679}
680
Jens Axboe21183b02006-07-13 12:33:14 +0200681static void cfq_merged_request(request_queue_t *q, struct request *req,
682 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683{
Jens Axboe21183b02006-07-13 12:33:14 +0200684 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200685 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686
Jens Axboe5e705372006-07-13 12:39:25 +0200687 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689}
690
691static void
692cfq_merged_requests(request_queue_t *q, struct request *rq,
693 struct request *next)
694{
Jens Axboe22e2c502005-06-27 10:55:12 +0200695 /*
696 * reposition in fifo if next is older than rq
697 */
698 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
699 time_before(next->start_time, rq->start_time))
700 list_move(&rq->queuelist, &next->queuelist);
701
Jens Axboeb4878f22005-10-20 16:42:29 +0200702 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200703}
704
Jens Axboeda775262006-12-20 11:04:12 +0100705static int cfq_allow_merge(request_queue_t *q, struct request *rq,
706 struct bio *bio)
707{
708 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200709 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100710 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100711
712 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100713 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100714 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200715 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100716 return 0;
717
718 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100719 * Lookup the cfqq that this bio will be queued with. Allow
720 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100721 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200722 cic = cfq_cic_rb_lookup(cfqd, current->io_context);
723 if (!cic)
724 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100725
Vasily Tarasov91fac312007-04-25 12:29:51 +0200726 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100727 if (cfqq == RQ_CFQQ(rq))
728 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100729
Jens Axboeec8acb62007-01-02 18:32:11 +0100730 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100731}
732
Jens Axboe22e2c502005-06-27 10:55:12 +0200733static inline void
734__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
735{
736 if (cfqq) {
737 /*
738 * stop potential idle class queues waiting service
739 */
740 del_timer(&cfqd->idle_class_timer);
741
Jens Axboe22e2c502005-06-27 10:55:12 +0200742 cfqq->slice_end = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200743 cfq_clear_cfqq_must_alloc_slice(cfqq);
744 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100745 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe1afba042007-04-17 12:47:55 +0200746 cfq_clear_cfqq_queue_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200747 }
748
749 cfqd->active_queue = cfqq;
750}
751
752/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100753 * current cfqq expired its slice (or was too idle), select new one
754 */
755static void
756__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200757 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100758{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100759 if (cfq_cfqq_wait_request(cfqq))
760 del_timer(&cfqd->idle_slice_timer);
761
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100762 cfq_clear_cfqq_must_dispatch(cfqq);
763 cfq_clear_cfqq_wait_request(cfqq);
764
765 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200766 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100767 */
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100768 if (timed_out && !cfq_cfqq_slice_new(cfqq))
Jens Axboec5b680f2007-01-19 11:56:49 +1100769 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100770
Jens Axboeedd75ff2007-04-19 12:03:34 +0200771 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100772
773 if (cfqq == cfqd->active_queue)
774 cfqd->active_queue = NULL;
775
776 if (cfqd->active_cic) {
777 put_io_context(cfqd->active_cic->ioc);
778 cfqd->active_cic = NULL;
779 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100780}
781
Jens Axboe6084cdd2007-04-23 08:25:00 +0200782static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100783{
784 struct cfq_queue *cfqq = cfqd->active_queue;
785
786 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200787 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100788}
789
Jens Axboe498d3aa2007-04-26 12:54:48 +0200790/*
791 * Get next queue for service. Unless we have a queue preemption,
792 * we'll simply select the first cfqq in the service tree.
793 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200794static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200795{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200796 struct cfq_queue *cfqq;
797 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +0200798
Jens Axboeedd75ff2007-04-19 12:03:34 +0200799 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
800 return NULL;
801
802 n = cfq_rb_first(&cfqd->service_tree);
803 cfqq = rb_entry(n, struct cfq_queue, rb_node);
804
805 if (cfq_class_idle(cfqq)) {
806 unsigned long end;
807
Jens Axboe89850f72006-07-22 16:48:31 +0200808 /*
Jens Axboeedd75ff2007-04-19 12:03:34 +0200809 * if we have idle queues and no rt or be queues had
810 * pending requests, either allow immediate service if
811 * the grace period has passed or arm the idle grace
812 * timer
Jens Axboe89850f72006-07-22 16:48:31 +0200813 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200814 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
815 if (time_before(jiffies, end)) {
816 mod_timer(&cfqd->idle_class_timer, end);
817 cfqq = NULL;
Jens Axboe67060e32007-04-18 20:13:32 +0200818 }
Jens Axboe22e2c502005-06-27 10:55:12 +0200819 }
820
Jens Axboe6d048f52007-04-25 12:44:27 +0200821 return cfqq;
822}
823
Jens Axboe498d3aa2007-04-26 12:54:48 +0200824/*
825 * Get and set a new active queue for service.
826 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200827static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
828{
829 struct cfq_queue *cfqq;
830
Jens Axboed9e76202007-04-20 14:27:50 +0200831 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200832 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200833 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200834}
835
Jens Axboed9e76202007-04-20 14:27:50 +0200836static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
837 struct request *rq)
838{
839 if (rq->sector >= cfqd->last_position)
840 return rq->sector - cfqd->last_position;
841 else
842 return cfqd->last_position - rq->sector;
843}
844
Jens Axboe6d048f52007-04-25 12:44:27 +0200845static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
846{
847 struct cfq_io_context *cic = cfqd->active_cic;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200848
Jens Axboe6d048f52007-04-25 12:44:27 +0200849 if (!sample_valid(cic->seek_samples))
850 return 0;
851
852 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
853}
854
Jens Axboed9e76202007-04-20 14:27:50 +0200855static int cfq_close_cooperator(struct cfq_data *cfq_data,
856 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200857{
Jens Axboe6d048f52007-04-25 12:44:27 +0200858 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200859 * We should notice if some of the queues are cooperating, eg
860 * working closely on the same area of the disk. In that case,
861 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +0200862 */
Jens Axboed9e76202007-04-20 14:27:50 +0200863 return 0;
Jens Axboe6d048f52007-04-25 12:44:27 +0200864}
865
866#define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
867
868static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200869{
Jens Axboe17926692007-01-19 11:59:30 +1100870 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +0200871 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100872 unsigned long sl;
873
Jens Axboedd67d052006-06-21 09:36:18 +0200874 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +0200875 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +0200876
877 /*
878 * idle is disabled, either manually or by past process history
879 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200880 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
881 return;
882
Jens Axboe22e2c502005-06-27 10:55:12 +0200883 /*
884 * task has exited, don't wait
885 */
Jens Axboe206dc692006-03-28 13:03:44 +0200886 cic = cfqd->active_cic;
887 if (!cic || !cic->ioc->task)
Jens Axboe6d048f52007-04-25 12:44:27 +0200888 return;
889
890 /*
891 * See if this prio level has a good candidate
892 */
Jens Axboe1afba042007-04-17 12:47:55 +0200893 if (cfq_close_cooperator(cfqd, cfqq) &&
894 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
Jens Axboe6d048f52007-04-25 12:44:27 +0200895 return;
Jens Axboe22e2c502005-06-27 10:55:12 +0200896
Jens Axboe3b181522005-06-27 10:56:24 +0200897 cfq_mark_cfqq_must_dispatch(cfqq);
898 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200899
Jens Axboe206dc692006-03-28 13:03:44 +0200900 /*
901 * we don't want to idle for seeks, but we do want to allow
902 * fair distribution of slice time for a process doing back-to-back
903 * seeks. so allow a little bit of time for him to submit a new rq
904 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200905 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200906 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +0200907 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +0200908
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100909 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910}
911
Jens Axboe498d3aa2007-04-26 12:54:48 +0200912/*
913 * Move request from internal lists to the request queue dispatch list.
914 */
Jens Axboe5e705372006-07-13 12:39:25 +0200915static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916{
Jens Axboe3ed9a292007-04-23 08:33:33 +0200917 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200918 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200919
Jens Axboe5380a102006-07-13 12:37:56 +0200920 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200921 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +0200922 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +0200923
924 if (cfq_cfqq_sync(cfqq))
925 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926}
927
928/*
929 * return expired entry, or NULL to just start from scratch in rbtree
930 */
Jens Axboe5e705372006-07-13 12:39:25 +0200931static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932{
933 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200934 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200935 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936
Jens Axboe3b181522005-06-27 10:56:24 +0200937 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +1100939
940 cfq_mark_cfqq_fifo_expire(cfqq);
941
Jens Axboe89850f72006-07-22 16:48:31 +0200942 if (list_empty(&cfqq->fifo))
943 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944
Jens Axboe6d048f52007-04-25 12:44:27 +0200945 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +0200946 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947
Jens Axboe6d048f52007-04-25 12:44:27 +0200948 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
949 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950
Jens Axboe6d048f52007-04-25 12:44:27 +0200951 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952}
953
Jens Axboe22e2c502005-06-27 10:55:12 +0200954static inline int
955cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
956{
957 const int base_rq = cfqd->cfq_slice_async_rq;
958
959 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
960
961 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
962}
963
964/*
Jens Axboe498d3aa2007-04-26 12:54:48 +0200965 * Select a queue for service. If we have a current active queue,
966 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +0200967 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100968static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200969{
Jens Axboe22e2c502005-06-27 10:55:12 +0200970 struct cfq_queue *cfqq;
971
972 cfqq = cfqd->active_queue;
973 if (!cfqq)
974 goto new_queue;
975
976 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200977 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +0200978 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200979 if (cfq_slice_used(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +0200980 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +0200981
982 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200983 * The active queue has requests and isn't expired, allow it to
984 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +0200985 */
Jens Axboedd67d052006-06-21 09:36:18 +0200986 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200987 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +0200988
989 /*
990 * No requests pending. If the active queue still has requests in
991 * flight or is idling for a new request, allow either of these
992 * conditions to happen (or time out) before selecting a new queue.
993 */
Jens Axboecc197472007-04-20 20:45:39 +0200994 if (timer_pending(&cfqd->idle_slice_timer) ||
995 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +0200996 cfqq = NULL;
997 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +0200998 }
999
Jens Axboe3b181522005-06-27 10:56:24 +02001000expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001001 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001002new_queue:
1003 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001004keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001005 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001006}
1007
Jens Axboe498d3aa2007-04-26 12:54:48 +02001008/*
1009 * Dispatch some requests from cfqq, moving them to the request queue
1010 * dispatch list.
1011 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001012static int
1013__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1014 int max_dispatch)
1015{
1016 int dispatched = 0;
1017
Jens Axboedd67d052006-06-21 09:36:18 +02001018 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001019
1020 do {
Jens Axboe5e705372006-07-13 12:39:25 +02001021 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001022
1023 /*
1024 * follow expired path, else get first next available
1025 */
Jens Axboe5e705372006-07-13 12:39:25 +02001026 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1027 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001028
1029 /*
1030 * finally, insert request into driver dispatch list
1031 */
Jens Axboe5e705372006-07-13 12:39:25 +02001032 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001033
Jens Axboe22e2c502005-06-27 10:55:12 +02001034 dispatched++;
1035
1036 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +02001037 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1038 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001039 }
1040
Jens Axboedd67d052006-06-21 09:36:18 +02001041 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001042 break;
1043
1044 } while (dispatched < max_dispatch);
1045
1046 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001047 * expire an async queue immediately if it has used up its slice. idle
1048 * queue always expire after 1 dispatch round.
1049 */
Jens Axboea9938002007-04-20 08:55:52 +02001050 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
Jens Axboe20e493a2007-04-23 08:26:36 +02001051 dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboea9938002007-04-20 08:55:52 +02001052 cfq_class_idle(cfqq))) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001053 cfqq->slice_end = jiffies + 1;
Jens Axboe6084cdd2007-04-23 08:25:00 +02001054 cfq_slice_expired(cfqd, 0);
Jens Axboe44f7c162007-01-19 11:51:58 +11001055 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001056
1057 return dispatched;
1058}
1059
Jens Axboed9e76202007-04-20 14:27:50 +02001060static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1061{
1062 int dispatched = 0;
1063
1064 while (cfqq->next_rq) {
1065 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1066 dispatched++;
1067 }
1068
1069 BUG_ON(!list_empty(&cfqq->fifo));
1070 return dispatched;
1071}
1072
Jens Axboe498d3aa2007-04-26 12:54:48 +02001073/*
1074 * Drain our current requests. Used for barriers and when switching
1075 * io schedulers on-the-fly.
1076 */
Jens Axboed9e76202007-04-20 14:27:50 +02001077static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001078{
Jens Axboed9e76202007-04-20 14:27:50 +02001079 int dispatched = 0;
1080 struct rb_node *n;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001081
Jens Axboecc09e292007-04-26 12:53:50 +02001082 while ((n = cfq_rb_first(&cfqd->service_tree)) != NULL) {
Jens Axboed9e76202007-04-20 14:27:50 +02001083 struct cfq_queue *cfqq = rb_entry(n, struct cfq_queue, rb_node);
1084
1085 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1086 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001087
Jens Axboe6084cdd2007-04-23 08:25:00 +02001088 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001089
1090 BUG_ON(cfqd->busy_queues);
1091
1092 return dispatched;
1093}
1094
Jens Axboed9e76202007-04-20 14:27:50 +02001095static int cfq_dispatch_requests(request_queue_t *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096{
1097 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001098 struct cfq_queue *cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001099 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100
Jens Axboe22e2c502005-06-27 10:55:12 +02001101 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 return 0;
1103
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001104 if (unlikely(force))
1105 return cfq_forced_dispatch(cfqd);
1106
Jens Axboecaaa5f92006-06-16 11:23:00 +02001107 dispatched = 0;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001108 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001109 int max_dispatch;
1110
Jens Axboe3ed9a292007-04-23 08:33:33 +02001111 max_dispatch = cfqd->cfq_quantum;
1112 if (cfq_class_idle(cfqq))
1113 max_dispatch = 1;
1114
1115 if (cfqq->dispatched >= max_dispatch) {
1116 if (cfqd->busy_queues > 1)
Jens Axboe6d048f52007-04-25 12:44:27 +02001117 break;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001118 if (cfqq->dispatched >= 4 * max_dispatch)
Jens Axboea9938002007-04-20 08:55:52 +02001119 break;
1120 }
Jens Axboe9ede2092007-01-19 12:11:44 +11001121
Jens Axboe3ed9a292007-04-23 08:33:33 +02001122 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1123 break;
1124
Jens Axboe3b181522005-06-27 10:56:24 +02001125 cfq_clear_cfqq_must_dispatch(cfqq);
1126 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001127 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128
Jens Axboecaaa5f92006-06-16 11:23:00 +02001129 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 }
1131
Jens Axboecaaa5f92006-06-16 11:23:00 +02001132 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133}
1134
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135/*
Jens Axboe5e705372006-07-13 12:39:25 +02001136 * task holds one reference to the queue, dropped when task exits. each rq
1137 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 *
1139 * queue lock must be held here.
1140 */
1141static void cfq_put_queue(struct cfq_queue *cfqq)
1142{
Jens Axboe22e2c502005-06-27 10:55:12 +02001143 struct cfq_data *cfqd = cfqq->cfqd;
1144
1145 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146
1147 if (!atomic_dec_and_test(&cfqq->ref))
1148 return;
1149
1150 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001151 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001152 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001154 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001155 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001156 cfq_schedule_dispatch(cfqd);
1157 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001158
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 kmem_cache_free(cfq_pool, cfqq);
1160}
1161
Jens Axboee2d74ac2006-03-28 08:59:01 +02001162static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163{
Jens Axboe22e2c502005-06-27 10:55:12 +02001164 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001165 struct rb_node *n;
1166 int freed = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001167
Jens Axboe597bc482007-04-24 21:23:53 +02001168 ioc->ioc_data = NULL;
1169
Jens Axboee2d74ac2006-03-28 08:59:01 +02001170 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1171 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1172 rb_erase(&__cic->rb_node, &ioc->cic_root);
Jens Axboe22e2c502005-06-27 10:55:12 +02001173 kmem_cache_free(cfq_ioc_pool, __cic);
Al Viro334e94d2006-03-18 15:05:53 -05001174 freed++;
Jens Axboe22e2c502005-06-27 10:55:12 +02001175 }
1176
Jens Axboe4050cf12006-07-19 05:07:12 +02001177 elv_ioc_count_mod(ioc_count, -freed);
1178
1179 if (ioc_gone && !elv_ioc_count_read(ioc_count))
Al Viro334e94d2006-03-18 15:05:53 -05001180 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181}
1182
Jens Axboe89850f72006-07-22 16:48:31 +02001183static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1184{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001185 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001186 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001187 cfq_schedule_dispatch(cfqd);
1188 }
Jens Axboe89850f72006-07-22 16:48:31 +02001189
1190 cfq_put_queue(cfqq);
1191}
1192
1193static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1194 struct cfq_io_context *cic)
1195{
Jens Axboefc463792006-08-29 09:05:44 +02001196 list_del_init(&cic->queue_list);
1197 smp_wmb();
1198 cic->key = NULL;
1199
Jens Axboe89850f72006-07-22 16:48:31 +02001200 if (cic->cfqq[ASYNC]) {
1201 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1202 cic->cfqq[ASYNC] = NULL;
1203 }
1204
1205 if (cic->cfqq[SYNC]) {
1206 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1207 cic->cfqq[SYNC] = NULL;
1208 }
Jens Axboe89850f72006-07-22 16:48:31 +02001209}
1210
Jens Axboe22e2c502005-06-27 10:55:12 +02001211static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1212{
Al Viro478a82b2006-03-18 13:25:24 -05001213 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001214
Jens Axboe89850f72006-07-22 16:48:31 +02001215 if (cfqd) {
1216 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001217
Jens Axboefc463792006-08-29 09:05:44 +02001218 spin_lock_irq(q->queue_lock);
Jens Axboe89850f72006-07-22 16:48:31 +02001219 __cfq_exit_single_io_context(cfqd, cic);
Jens Axboefc463792006-08-29 09:05:44 +02001220 spin_unlock_irq(q->queue_lock);
Al Viro12a05732006-03-18 13:38:01 -05001221 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001222}
1223
Jens Axboe498d3aa2007-04-26 12:54:48 +02001224/*
1225 * The process that ioc belongs to has exited, we need to clean up
1226 * and put the internal structures we have that belongs to that process.
1227 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001228static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229{
Jens Axboe22e2c502005-06-27 10:55:12 +02001230 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001231 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +02001232
Jens Axboe597bc482007-04-24 21:23:53 +02001233 ioc->ioc_data = NULL;
1234
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235 /*
1236 * put the reference this task is holding to the various queues
1237 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001238 n = rb_first(&ioc->cic_root);
1239 while (n != NULL) {
1240 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1241
Jens Axboe22e2c502005-06-27 10:55:12 +02001242 cfq_exit_single_io_context(__cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001243 n = rb_next(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001244 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245}
1246
Jens Axboe22e2c502005-06-27 10:55:12 +02001247static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001248cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249{
Jens Axboeb5deef92006-07-19 23:39:40 +02001250 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251
Jens Axboeb5deef92006-07-19 23:39:40 +02001252 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 if (cic) {
Jens Axboe553698f2006-06-14 19:11:57 +02001254 memset(cic, 0, sizeof(*cic));
Jens Axboe22e2c502005-06-27 10:55:12 +02001255 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001256 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001257 cic->dtor = cfq_free_io_context;
1258 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001259 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260 }
1261
1262 return cic;
1263}
1264
Jens Axboe22e2c502005-06-27 10:55:12 +02001265static void cfq_init_prio_data(struct cfq_queue *cfqq)
1266{
1267 struct task_struct *tsk = current;
1268 int ioprio_class;
1269
Jens Axboe3b181522005-06-27 10:56:24 +02001270 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001271 return;
1272
1273 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1274 switch (ioprio_class) {
1275 default:
1276 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1277 case IOPRIO_CLASS_NONE:
1278 /*
1279 * no prio set, place us in the middle of the BE classes
1280 */
1281 cfqq->ioprio = task_nice_ioprio(tsk);
1282 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1283 break;
1284 case IOPRIO_CLASS_RT:
1285 cfqq->ioprio = task_ioprio(tsk);
1286 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1287 break;
1288 case IOPRIO_CLASS_BE:
1289 cfqq->ioprio = task_ioprio(tsk);
1290 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1291 break;
1292 case IOPRIO_CLASS_IDLE:
1293 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1294 cfqq->ioprio = 7;
Jens Axboe3b181522005-06-27 10:56:24 +02001295 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001296 break;
1297 }
1298
1299 /*
1300 * keep track of original prio settings in case we have to temporarily
1301 * elevate the priority of this queue
1302 */
1303 cfqq->org_ioprio = cfqq->ioprio;
1304 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001305 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001306}
1307
Al Viro478a82b2006-03-18 13:25:24 -05001308static inline void changed_ioprio(struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001309{
Al Viro478a82b2006-03-18 13:25:24 -05001310 struct cfq_data *cfqd = cic->key;
1311 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001312 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001313
Jens Axboecaaa5f92006-06-16 11:23:00 +02001314 if (unlikely(!cfqd))
1315 return;
1316
Jens Axboec1b707d2006-10-30 19:54:23 +01001317 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001318
1319 cfqq = cic->cfqq[ASYNC];
1320 if (cfqq) {
1321 struct cfq_queue *new_cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001322 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc->task,
Jens Axboecaaa5f92006-06-16 11:23:00 +02001323 GFP_ATOMIC);
1324 if (new_cfqq) {
1325 cic->cfqq[ASYNC] = new_cfqq;
1326 cfq_put_queue(cfqq);
1327 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001328 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001329
1330 cfqq = cic->cfqq[SYNC];
1331 if (cfqq)
1332 cfq_mark_cfqq_prio_changed(cfqq);
1333
Jens Axboec1b707d2006-10-30 19:54:23 +01001334 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001335}
1336
Jens Axboefc463792006-08-29 09:05:44 +02001337static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338{
Al Viroa6a07632006-03-18 13:26:44 -05001339 struct cfq_io_context *cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001340 struct rb_node *n;
Al Viroa6a07632006-03-18 13:26:44 -05001341
Jens Axboefc463792006-08-29 09:05:44 +02001342 ioc->ioprio_changed = 0;
Al Viroa6a07632006-03-18 13:26:44 -05001343
Jens Axboee2d74ac2006-03-28 08:59:01 +02001344 n = rb_first(&ioc->cic_root);
1345 while (n != NULL) {
1346 cic = rb_entry(n, struct cfq_io_context, rb_node);
Jens Axboe3793c652006-05-30 21:11:04 +02001347
Al Viro478a82b2006-03-18 13:25:24 -05001348 changed_ioprio(cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001349 n = rb_next(n);
1350 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351}
1352
1353static struct cfq_queue *
Vasily Tarasov91fac312007-04-25 12:29:51 +02001354cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct task_struct *tsk,
Al Viro8267e262005-10-21 03:20:53 -04001355 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001358 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359
1360retry:
Vasily Tarasov91fac312007-04-25 12:29:51 +02001361 cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1362 /* cic always exists here */
1363 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364
1365 if (!cfqq) {
1366 if (new_cfqq) {
1367 cfqq = new_cfqq;
1368 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001369 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001370 /*
1371 * Inform the allocator of the fact that we will
1372 * just repeat this allocation if it fails, to allow
1373 * the allocator to do whatever it needs to attempt to
1374 * free memory.
1375 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 spin_unlock_irq(cfqd->queue->queue_lock);
Jens Axboeb5deef92006-07-19 23:39:40 +02001377 new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 spin_lock_irq(cfqd->queue->queue_lock);
1379 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001380 } else {
Jens Axboeb5deef92006-07-19 23:39:40 +02001381 cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001382 if (!cfqq)
1383 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001384 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385
1386 memset(cfqq, 0, sizeof(*cfqq));
1387
Jens Axboed9e76202007-04-20 14:27:50 +02001388 RB_CLEAR_NODE(&cfqq->rb_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001389 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 atomic_set(&cfqq->ref, 0);
1392 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001393
Vasily Tarasov91fac312007-04-25 12:29:51 +02001394 if (is_sync) {
Jens Axboea9938002007-04-20 08:55:52 +02001395 cfq_mark_cfqq_idle_window(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001396 cfq_mark_cfqq_sync(cfqq);
1397 }
Jens Axboea9938002007-04-20 08:55:52 +02001398
Jens Axboe3b181522005-06-27 10:56:24 +02001399 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b03742006-07-28 09:48:51 +02001400 cfq_mark_cfqq_queue_new(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001401
Jens Axboe3b181522005-06-27 10:56:24 +02001402 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403 }
1404
1405 if (new_cfqq)
1406 kmem_cache_free(cfq_pool, new_cfqq);
1407
1408 atomic_inc(&cfqq->ref);
1409out:
1410 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1411 return cfqq;
1412}
1413
Jens Axboe498d3aa2007-04-26 12:54:48 +02001414/*
1415 * We drop cfq io contexts lazily, so we may find a dead one.
1416 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001417static void
1418cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1419{
Jens Axboefc463792006-08-29 09:05:44 +02001420 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001421
1422 if (ioc->ioc_data == cic)
1423 ioc->ioc_data = NULL;
1424
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001425 rb_erase(&cic->rb_node, &ioc->cic_root);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001426 kmem_cache_free(cfq_ioc_pool, cic);
Jens Axboe4050cf12006-07-19 05:07:12 +02001427 elv_ioc_count_dec(ioc_count);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001428}
1429
Jens Axboee2d74ac2006-03-28 08:59:01 +02001430static struct cfq_io_context *
1431cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1432{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001433 struct rb_node *n;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001434 struct cfq_io_context *cic;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001435 void *k, *key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001436
Vasily Tarasov91fac312007-04-25 12:29:51 +02001437 if (unlikely(!ioc))
1438 return NULL;
1439
Jens Axboe597bc482007-04-24 21:23:53 +02001440 /*
1441 * we maintain a last-hit cache, to avoid browsing over the tree
1442 */
1443 cic = ioc->ioc_data;
1444 if (cic && cic->key == cfqd)
1445 return cic;
1446
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001447restart:
1448 n = ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001449 while (n) {
1450 cic = rb_entry(n, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001451 /* ->key must be copied to avoid race with cfq_exit_queue() */
1452 k = cic->key;
1453 if (unlikely(!k)) {
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001454 cfq_drop_dead_cic(ioc, cic);
1455 goto restart;
1456 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001457
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001458 if (key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001459 n = n->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001460 else if (key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001461 n = n->rb_right;
Jens Axboe597bc482007-04-24 21:23:53 +02001462 else {
1463 ioc->ioc_data = cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001464 return cic;
Jens Axboe597bc482007-04-24 21:23:53 +02001465 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001466 }
1467
1468 return NULL;
1469}
1470
1471static inline void
1472cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1473 struct cfq_io_context *cic)
1474{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001475 struct rb_node **p;
1476 struct rb_node *parent;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001477 struct cfq_io_context *__cic;
Jens Axboe0261d682006-10-30 19:07:48 +01001478 unsigned long flags;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001479 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001480
Jens Axboee2d74ac2006-03-28 08:59:01 +02001481 cic->ioc = ioc;
1482 cic->key = cfqd;
1483
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001484restart:
1485 parent = NULL;
1486 p = &ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001487 while (*p) {
1488 parent = *p;
1489 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001490 /* ->key must be copied to avoid race with cfq_exit_queue() */
1491 k = __cic->key;
1492 if (unlikely(!k)) {
Oleg Nesterovbe33c3a2006-08-21 08:36:12 +02001493 cfq_drop_dead_cic(ioc, __cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001494 goto restart;
1495 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001496
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001497 if (cic->key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001498 p = &(*p)->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001499 else if (cic->key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001500 p = &(*p)->rb_right;
1501 else
1502 BUG();
1503 }
1504
1505 rb_link_node(&cic->rb_node, parent, p);
1506 rb_insert_color(&cic->rb_node, &ioc->cic_root);
Jens Axboefc463792006-08-29 09:05:44 +02001507
Jens Axboe0261d682006-10-30 19:07:48 +01001508 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001509 list_add(&cic->queue_list, &cfqd->cic_list);
Jens Axboe0261d682006-10-30 19:07:48 +01001510 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001511}
1512
Jens Axboe22e2c502005-06-27 10:55:12 +02001513/*
1514 * Setup general io context and cfq io context. There can be several cfq
1515 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001516 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001517 */
1518static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001519cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520{
Jens Axboe22e2c502005-06-27 10:55:12 +02001521 struct io_context *ioc = NULL;
1522 struct cfq_io_context *cic;
1523
1524 might_sleep_if(gfp_mask & __GFP_WAIT);
1525
Jens Axboeb5deef92006-07-19 23:39:40 +02001526 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001527 if (!ioc)
1528 return NULL;
1529
Jens Axboee2d74ac2006-03-28 08:59:01 +02001530 cic = cfq_cic_rb_lookup(cfqd, ioc);
1531 if (cic)
1532 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001533
Jens Axboee2d74ac2006-03-28 08:59:01 +02001534 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1535 if (cic == NULL)
1536 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001537
Jens Axboee2d74ac2006-03-28 08:59:01 +02001538 cfq_cic_link(cfqd, ioc, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001539out:
Jens Axboefc463792006-08-29 09:05:44 +02001540 smp_read_barrier_depends();
1541 if (unlikely(ioc->ioprio_changed))
1542 cfq_ioc_set_ioprio(ioc);
1543
Jens Axboe22e2c502005-06-27 10:55:12 +02001544 return cic;
1545err:
1546 put_io_context(ioc);
1547 return NULL;
1548}
1549
1550static void
1551cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1552{
Jens Axboeaaf12282007-01-19 11:30:16 +11001553 unsigned long elapsed = jiffies - cic->last_end_request;
1554 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001555
1556 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1557 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1558 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1559}
1560
Jens Axboe206dc692006-03-28 13:03:44 +02001561static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001562cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1563 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001564{
1565 sector_t sdist;
1566 u64 total;
1567
Jens Axboe5e705372006-07-13 12:39:25 +02001568 if (cic->last_request_pos < rq->sector)
1569 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001570 else
Jens Axboe5e705372006-07-13 12:39:25 +02001571 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001572
Jens Axboe6d048f52007-04-25 12:44:27 +02001573 if (!cic->seek_samples) {
1574 cfqd->new_seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1575 cfqd->new_seek_mean = cfqd->new_seek_total / 256;
1576 }
1577
Jens Axboe206dc692006-03-28 13:03:44 +02001578 /*
1579 * Don't allow the seek distance to get too large from the
1580 * odd fragment, pagein, etc
1581 */
1582 if (cic->seek_samples <= 60) /* second&third seek */
1583 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1584 else
1585 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1586
1587 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1588 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1589 total = cic->seek_total + (cic->seek_samples/2);
1590 do_div(total, cic->seek_samples);
1591 cic->seek_mean = (sector_t)total;
1592}
Jens Axboe22e2c502005-06-27 10:55:12 +02001593
1594/*
1595 * Disable idle window if the process thinks too long or seeks so much that
1596 * it doesn't matter
1597 */
1598static void
1599cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1600 struct cfq_io_context *cic)
1601{
Jens Axboe1be92f22007-04-19 14:32:26 +02001602 int enable_idle;
1603
1604 if (!cfq_cfqq_sync(cfqq))
1605 return;
1606
1607 enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001608
Jens Axboecaaa5f92006-06-16 11:23:00 +02001609 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1610 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001611 enable_idle = 0;
1612 else if (sample_valid(cic->ttime_samples)) {
1613 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1614 enable_idle = 0;
1615 else
1616 enable_idle = 1;
1617 }
1618
Jens Axboe3b181522005-06-27 10:56:24 +02001619 if (enable_idle)
1620 cfq_mark_cfqq_idle_window(cfqq);
1621 else
1622 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001623}
1624
Jens Axboe22e2c502005-06-27 10:55:12 +02001625/*
1626 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1627 * no or if we aren't sure, a 1 will cause a preempt.
1628 */
1629static int
1630cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001631 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001632{
Jens Axboe6d048f52007-04-25 12:44:27 +02001633 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001634
Jens Axboe6d048f52007-04-25 12:44:27 +02001635 cfqq = cfqd->active_queue;
1636 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001637 return 0;
1638
Jens Axboe6d048f52007-04-25 12:44:27 +02001639 if (cfq_slice_used(cfqq))
1640 return 1;
1641
1642 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001643 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001644
1645 if (cfq_class_idle(cfqq))
1646 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001647
Jens Axboe22e2c502005-06-27 10:55:12 +02001648 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02001649 * if the new request is sync, but the currently running queue is
1650 * not, let the sync request have priority.
1651 */
Jens Axboe5e705372006-07-13 12:39:25 +02001652 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001653 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001654
Jens Axboe374f84a2006-07-23 01:42:19 +02001655 /*
1656 * So both queues are sync. Let the new request get disk time if
1657 * it's a metadata request and the current queue is doing regular IO.
1658 */
1659 if (rq_is_meta(rq) && !cfqq->meta_pending)
1660 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001661
Jens Axboe1e3335d2007-02-14 19:59:49 +01001662 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1663 return 0;
1664
1665 /*
1666 * if this request is as-good as one we would expect from the
1667 * current cfqq, let it preempt
1668 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001669 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01001670 return 1;
1671
Jens Axboe22e2c502005-06-27 10:55:12 +02001672 return 0;
1673}
1674
1675/*
1676 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1677 * let it have half of its nominal slice.
1678 */
1679static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1680{
Jens Axboe6084cdd2007-04-23 08:25:00 +02001681 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001682
Jens Axboebf572252006-07-19 20:29:12 +02001683 /*
1684 * Put the new queue at the front of the of the current list,
1685 * so we know that it will be selected next.
1686 */
1687 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02001688
1689 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02001690
Jens Axboe44f7c162007-01-19 11:51:58 +11001691 cfqq->slice_end = 0;
1692 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001693}
1694
1695/*
Jens Axboe5e705372006-07-13 12:39:25 +02001696 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001697 * something we should do about it
1698 */
1699static void
Jens Axboe5e705372006-07-13 12:39:25 +02001700cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1701 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001702{
Jens Axboe5e705372006-07-13 12:39:25 +02001703 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001704
Jens Axboe374f84a2006-07-23 01:42:19 +02001705 if (rq_is_meta(rq))
1706 cfqq->meta_pending++;
1707
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001708 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02001709 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001710 cfq_update_idle_window(cfqd, cfqq, cic);
1711
Jens Axboe5e705372006-07-13 12:39:25 +02001712 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe6d048f52007-04-25 12:44:27 +02001713 cfqq->last_request_pos = cic->last_request_pos;
Jens Axboe22e2c502005-06-27 10:55:12 +02001714
1715 if (cfqq == cfqd->active_queue) {
1716 /*
1717 * if we are waiting for a request for this queue, let it rip
1718 * immediately and flag that we must not expire this queue
1719 * just now
1720 */
Jens Axboe3b181522005-06-27 10:56:24 +02001721 if (cfq_cfqq_wait_request(cfqq)) {
1722 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001723 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001724 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001725 }
Jens Axboe5e705372006-07-13 12:39:25 +02001726 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001727 /*
1728 * not the active queue - expire current slice if it is
1729 * idle and has expired it's mean thinktime or this new queue
1730 * has some old slice time left and is of higher priority
1731 */
1732 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001733 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001734 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001735 }
1736}
1737
Jens Axboeb4878f22005-10-20 16:42:29 +02001738static void cfq_insert_request(request_queue_t *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001739{
Jens Axboeb4878f22005-10-20 16:42:29 +02001740 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001741 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001742
1743 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744
Jens Axboe5e705372006-07-13 12:39:25 +02001745 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746
Jens Axboe22e2c502005-06-27 10:55:12 +02001747 list_add_tail(&rq->queuelist, &cfqq->fifo);
1748
Jens Axboe5e705372006-07-13 12:39:25 +02001749 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750}
1751
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752static void cfq_completed_request(request_queue_t *q, struct request *rq)
1753{
Jens Axboe5e705372006-07-13 12:39:25 +02001754 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001755 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001756 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001757 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758
Jens Axboeb4878f22005-10-20 16:42:29 +02001759 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760
Jens Axboeb4878f22005-10-20 16:42:29 +02001761 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02001762 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02001763 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02001764 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765
Jens Axboe3ed9a292007-04-23 08:33:33 +02001766 if (cfq_cfqq_sync(cfqq))
1767 cfqd->sync_flight--;
1768
Jens Axboeb4878f22005-10-20 16:42:29 +02001769 if (!cfq_class_idle(cfqq))
1770 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001771
Jens Axboecaaa5f92006-06-16 11:23:00 +02001772 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001773 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001774
1775 /*
1776 * If this is the active queue, check if it needs to be expired,
1777 * or if we want to idle in case it has no pending requests.
1778 */
1779 if (cfqd->active_queue == cfqq) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001780 if (cfq_cfqq_slice_new(cfqq)) {
1781 cfq_set_prio_slice(cfqd, cfqq);
1782 cfq_clear_cfqq_slice_new(cfqq);
1783 }
1784 if (cfq_slice_used(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02001785 cfq_slice_expired(cfqd, 1);
Jens Axboe6d048f52007-04-25 12:44:27 +02001786 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1787 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001788 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001789
1790 if (!cfqd->rq_in_driver)
1791 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792}
1793
Jens Axboe22e2c502005-06-27 10:55:12 +02001794/*
1795 * we temporarily boost lower priority queues if they are holding fs exclusive
1796 * resources. they are boosted to normal prio (CLASS_BE/4)
1797 */
1798static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799{
Jens Axboe22e2c502005-06-27 10:55:12 +02001800 if (has_fs_excl()) {
1801 /*
1802 * boost idle prio on transactions that would lock out other
1803 * users of the filesystem
1804 */
1805 if (cfq_class_idle(cfqq))
1806 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1807 if (cfqq->ioprio > IOPRIO_NORM)
1808 cfqq->ioprio = IOPRIO_NORM;
1809 } else {
1810 /*
1811 * check if we need to unboost the queue
1812 */
1813 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1814 cfqq->ioprio_class = cfqq->org_ioprio_class;
1815 if (cfqq->ioprio != cfqq->org_ioprio)
1816 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001818}
1819
Jens Axboe89850f72006-07-22 16:48:31 +02001820static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001821{
Jens Axboe3b181522005-06-27 10:56:24 +02001822 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001823 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001824 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001825 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001826 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001827
1828 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02001829}
1830
Jens Axboecb78b282006-07-28 09:32:57 +02001831static int cfq_may_queue(request_queue_t *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02001832{
1833 struct cfq_data *cfqd = q->elevator->elevator_data;
1834 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001835 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02001836 struct cfq_queue *cfqq;
1837
1838 /*
1839 * don't force setup of a queue from here, as a call to may_queue
1840 * does not necessarily imply that a request actually will be queued.
1841 * so just lookup a possibly existing queue, or return 'may queue'
1842 * if that fails
1843 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001844 cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1845 if (!cic)
1846 return ELV_MQUEUE_MAY;
1847
1848 cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02001849 if (cfqq) {
1850 cfq_init_prio_data(cfqq);
1851 cfq_prio_boost(cfqq);
1852
Jens Axboe89850f72006-07-22 16:48:31 +02001853 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001854 }
1855
1856 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857}
1858
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859/*
1860 * queue lock held here
1861 */
Jens Axboebb37b942006-12-01 10:42:33 +01001862static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863{
Jens Axboe5e705372006-07-13 12:39:25 +02001864 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865
Jens Axboe5e705372006-07-13 12:39:25 +02001866 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001867 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868
Jens Axboe22e2c502005-06-27 10:55:12 +02001869 BUG_ON(!cfqq->allocated[rw]);
1870 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871
Jens Axboe5e705372006-07-13 12:39:25 +02001872 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02001875 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877 cfq_put_queue(cfqq);
1878 }
1879}
1880
1881/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001882 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001884static int
Jens Axboecb78b282006-07-28 09:32:57 +02001885cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886{
1887 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe3b181522005-06-27 10:56:24 +02001888 struct task_struct *tsk = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 struct cfq_io_context *cic;
1890 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01001891 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001892 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893 unsigned long flags;
1894
1895 might_sleep_if(gfp_mask & __GFP_WAIT);
1896
Jens Axboee2d74ac2006-03-28 08:59:01 +02001897 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001898
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 spin_lock_irqsave(q->queue_lock, flags);
1900
Jens Axboe22e2c502005-06-27 10:55:12 +02001901 if (!cic)
1902 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903
Vasily Tarasov91fac312007-04-25 12:29:51 +02001904 cfqq = cic_to_cfqq(cic, is_sync);
1905 if (!cfqq) {
1906 cfqq = cfq_get_queue(cfqd, is_sync, tsk, gfp_mask);
1907
Jens Axboe22e2c502005-06-27 10:55:12 +02001908 if (!cfqq)
1909 goto queue_fail;
1910
Vasily Tarasov91fac312007-04-25 12:29:51 +02001911 cic_set_cfqq(cic, cfqq, is_sync);
1912 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913
1914 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02001915 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001916 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02001917
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918 spin_unlock_irqrestore(q->queue_lock, flags);
1919
Jens Axboe5e705372006-07-13 12:39:25 +02001920 rq->elevator_private = cic;
1921 rq->elevator_private2 = cfqq;
1922 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02001923
Jens Axboe22e2c502005-06-27 10:55:12 +02001924queue_fail:
1925 if (cic)
1926 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02001927
Jens Axboe3b181522005-06-27 10:56:24 +02001928 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 spin_unlock_irqrestore(q->queue_lock, flags);
1930 return 1;
1931}
1932
David Howells65f27f32006-11-22 14:55:48 +00001933static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02001934{
David Howells65f27f32006-11-22 14:55:48 +00001935 struct cfq_data *cfqd =
1936 container_of(work, struct cfq_data, unplug_work);
1937 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001938 unsigned long flags;
1939
1940 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02001941 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02001942 spin_unlock_irqrestore(q->queue_lock, flags);
1943}
1944
1945/*
1946 * Timer running if the active_queue is currently idling inside its time slice
1947 */
1948static void cfq_idle_slice_timer(unsigned long data)
1949{
1950 struct cfq_data *cfqd = (struct cfq_data *) data;
1951 struct cfq_queue *cfqq;
1952 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001953 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001954
1955 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1956
1957 if ((cfqq = cfqd->active_queue) != NULL) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001958 timed_out = 0;
1959
Jens Axboe22e2c502005-06-27 10:55:12 +02001960 /*
1961 * expired
1962 */
Jens Axboe44f7c162007-01-19 11:51:58 +11001963 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001964 goto expire;
1965
1966 /*
1967 * only expire and reinvoke request handler, if there are
1968 * other queues with pending requests
1969 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02001970 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02001971 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02001972
1973 /*
1974 * not expired and it has a request pending, let it dispatch
1975 */
Jens Axboedd67d052006-06-21 09:36:18 +02001976 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001977 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001978 goto out_kick;
1979 }
1980 }
1981expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001982 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02001983out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02001984 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001985out_cont:
1986 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1987}
1988
1989/*
1990 * Timer running if an idle class queue is waiting for service
1991 */
1992static void cfq_idle_class_timer(unsigned long data)
1993{
1994 struct cfq_data *cfqd = (struct cfq_data *) data;
1995 unsigned long flags, end;
1996
1997 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1998
1999 /*
2000 * race with a non-idle queue, reset timer
2001 */
2002 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
Jens Axboeae818a32006-06-01 10:13:43 +02002003 if (!time_after_eq(jiffies, end))
2004 mod_timer(&cfqd->idle_class_timer, end);
2005 else
Jens Axboe3b181522005-06-27 10:56:24 +02002006 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002007
2008 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2009}
2010
Jens Axboe3b181522005-06-27 10:56:24 +02002011static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2012{
2013 del_timer_sync(&cfqd->idle_slice_timer);
2014 del_timer_sync(&cfqd->idle_class_timer);
2015 blk_sync_queue(cfqd->queue);
2016}
Jens Axboe22e2c502005-06-27 10:55:12 +02002017
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018static void cfq_exit_queue(elevator_t *e)
2019{
Jens Axboe22e2c502005-06-27 10:55:12 +02002020 struct cfq_data *cfqd = e->elevator_data;
Al Virod9ff4182006-03-18 13:51:22 -05002021 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002022
Jens Axboe3b181522005-06-27 10:56:24 +02002023 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002024
Al Virod9ff4182006-03-18 13:51:22 -05002025 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002026
Al Virod9ff4182006-03-18 13:51:22 -05002027 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002028 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002029
2030 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002031 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2032 struct cfq_io_context,
2033 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002034
2035 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002036 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002037
Al Virod9ff4182006-03-18 13:51:22 -05002038 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002039
2040 cfq_shutdown_timer_wq(cfqd);
2041
Al Viroa90d7422006-03-18 12:05:37 -05002042 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043}
2044
Jens Axboebb37b942006-12-01 10:42:33 +01002045static void *cfq_init_queue(request_queue_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046{
2047 struct cfq_data *cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048
Jens Axboeb5deef92006-07-19 23:39:40 +02002049 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002051 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052
2053 memset(cfqd, 0, sizeof(*cfqd));
Jens Axboe22e2c502005-06-27 10:55:12 +02002054
Jens Axboecc09e292007-04-26 12:53:50 +02002055 cfqd->service_tree = CFQ_RB_ROOT;
Al Virod9ff4182006-03-18 13:51:22 -05002056 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059
Jens Axboe22e2c502005-06-27 10:55:12 +02002060 init_timer(&cfqd->idle_slice_timer);
2061 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2062 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2063
2064 init_timer(&cfqd->idle_class_timer);
2065 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2066 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2067
David Howells65f27f32006-11-22 14:55:48 +00002068 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002069
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002071 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2072 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073 cfqd->cfq_back_max = cfq_back_max;
2074 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002075 cfqd->cfq_slice[0] = cfq_slice_async;
2076 cfqd->cfq_slice[1] = cfq_slice_sync;
2077 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2078 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02002079
Jens Axboebc1c1162006-06-08 08:49:06 +02002080 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081}
2082
2083static void cfq_slab_kill(void)
2084{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 if (cfq_pool)
2086 kmem_cache_destroy(cfq_pool);
2087 if (cfq_ioc_pool)
2088 kmem_cache_destroy(cfq_ioc_pool);
2089}
2090
2091static int __init cfq_slab_setup(void)
2092{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002093 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 if (!cfq_pool)
2095 goto fail;
2096
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002097 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 if (!cfq_ioc_pool)
2099 goto fail;
2100
2101 return 0;
2102fail:
2103 cfq_slab_kill();
2104 return -ENOMEM;
2105}
2106
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107/*
2108 * sysfs parts below -->
2109 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110static ssize_t
2111cfq_var_show(unsigned int var, char *page)
2112{
2113 return sprintf(page, "%d\n", var);
2114}
2115
2116static ssize_t
2117cfq_var_store(unsigned int *var, const char *page, size_t count)
2118{
2119 char *p = (char *) page;
2120
2121 *var = simple_strtoul(p, &p, 10);
2122 return count;
2123}
2124
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002126static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002128 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 unsigned int __data = __VAR; \
2130 if (__CONV) \
2131 __data = jiffies_to_msecs(__data); \
2132 return cfq_var_show(__data, (page)); \
2133}
2134SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002135SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2136SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002137SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2138SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002139SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2140SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2141SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2142SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143#undef SHOW_FUNCTION
2144
2145#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002146static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002148 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 unsigned int __data; \
2150 int ret = cfq_var_store(&__data, (page), count); \
2151 if (__data < (MIN)) \
2152 __data = (MIN); \
2153 else if (__data > (MAX)) \
2154 __data = (MAX); \
2155 if (__CONV) \
2156 *(__PTR) = msecs_to_jiffies(__data); \
2157 else \
2158 *(__PTR) = __data; \
2159 return ret; \
2160}
2161STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002162STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2163STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002164STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2165STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002166STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2167STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2168STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2169STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170#undef STORE_FUNCTION
2171
Al Viroe572ec72006-03-18 22:27:18 -05002172#define CFQ_ATTR(name) \
2173 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002174
Al Viroe572ec72006-03-18 22:27:18 -05002175static struct elv_fs_entry cfq_attrs[] = {
2176 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002177 CFQ_ATTR(fifo_expire_sync),
2178 CFQ_ATTR(fifo_expire_async),
2179 CFQ_ATTR(back_seek_max),
2180 CFQ_ATTR(back_seek_penalty),
2181 CFQ_ATTR(slice_sync),
2182 CFQ_ATTR(slice_async),
2183 CFQ_ATTR(slice_async_rq),
2184 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002185 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186};
2187
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188static struct elevator_type iosched_cfq = {
2189 .ops = {
2190 .elevator_merge_fn = cfq_merge,
2191 .elevator_merged_fn = cfq_merged_request,
2192 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002193 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002194 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002196 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197 .elevator_deactivate_req_fn = cfq_deactivate_request,
2198 .elevator_queue_empty_fn = cfq_queue_empty,
2199 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002200 .elevator_former_req_fn = elv_rb_former_request,
2201 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 .elevator_set_req_fn = cfq_set_request,
2203 .elevator_put_req_fn = cfq_put_request,
2204 .elevator_may_queue_fn = cfq_may_queue,
2205 .elevator_init_fn = cfq_init_queue,
2206 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002207 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 },
Al Viro3d1ab402006-03-18 18:35:43 -05002209 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 .elevator_name = "cfq",
2211 .elevator_owner = THIS_MODULE,
2212};
2213
2214static int __init cfq_init(void)
2215{
2216 int ret;
2217
Jens Axboe22e2c502005-06-27 10:55:12 +02002218 /*
2219 * could be 0 on HZ < 1000 setups
2220 */
2221 if (!cfq_slice_async)
2222 cfq_slice_async = 1;
2223 if (!cfq_slice_idle)
2224 cfq_slice_idle = 1;
2225
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226 if (cfq_slab_setup())
2227 return -ENOMEM;
2228
2229 ret = elv_register(&iosched_cfq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002230 if (ret)
2231 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 return ret;
2234}
2235
2236static void __exit cfq_exit(void)
2237{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002238 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002240 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002241 /* ioc_gone's update must be visible before reading ioc_count */
2242 smp_wmb();
Jens Axboe4050cf12006-07-19 05:07:12 +02002243 if (elv_ioc_count_read(ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002244 wait_for_completion(ioc_gone);
Al Viro334e94d2006-03-18 15:05:53 -05002245 synchronize_rcu();
Christoph Hellwig83521d32005-10-30 15:01:39 -08002246 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247}
2248
2249module_init(cfq_init);
2250module_exit(cfq_exit);
2251
2252MODULE_AUTHOR("Jens Axboe");
2253MODULE_LICENSE("GPL");
2254MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");