blob: d148ccbc36d17efa03e93d5af68340a77e4db532 [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
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +020095 /*
96 * async queue for each priority case
97 */
98 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
99 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200100
Jens Axboe22e2c502005-06-27 10:55:12 +0200101 struct timer_list idle_class_timer;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102
Jens Axboe6d048f52007-04-25 12:44:27 +0200103 sector_t last_position;
Jens Axboe22e2c502005-06-27 10:55:12 +0200104 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106 /*
107 * tunables, see top of file
108 */
109 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200110 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111 unsigned int cfq_back_penalty;
112 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200113 unsigned int cfq_slice[2];
114 unsigned int cfq_slice_async_rq;
115 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500116
117 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118};
119
Jens Axboe22e2c502005-06-27 10:55:12 +0200120/*
121 * Per process-grouping structure
122 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123struct cfq_queue {
124 /* reference count */
125 atomic_t ref;
126 /* parent cfq_data */
127 struct cfq_data *cfqd;
Jens Axboed9e76202007-04-20 14:27:50 +0200128 /* service_tree member */
129 struct rb_node rb_node;
130 /* service_tree key */
131 unsigned long rb_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 /* sorted list of pending requests */
133 struct rb_root sort_list;
134 /* if fifo isn't expired, next request to serve */
Jens Axboe5e705372006-07-13 12:39:25 +0200135 struct request *next_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136 /* requests queued in sort_list */
137 int queued[2];
138 /* currently allocated requests */
139 int allocated[2];
Jens Axboe374f84a2006-07-23 01:42:19 +0200140 /* pending metadata requests */
141 int meta_pending;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200143 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144
Jens Axboe22e2c502005-06-27 10:55:12 +0200145 unsigned long slice_end;
Jens Axboec5b680f2007-01-19 11:56:49 +1100146 long slice_resid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
Jens Axboe6d048f52007-04-25 12:44:27 +0200148 /* number of requests that are on the dispatch list or inside driver */
149 int dispatched;
Jens Axboe22e2c502005-06-27 10:55:12 +0200150
151 /* io prio of this group */
152 unsigned short ioprio, org_ioprio;
153 unsigned short ioprio_class, org_ioprio_class;
154
Jens Axboe3b181522005-06-27 10:56:24 +0200155 /* various state flags, see below */
156 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157};
158
Jens Axboe3b181522005-06-27 10:56:24 +0200159enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100160 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
161 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
162 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
163 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
164 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
165 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
166 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
167 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
168 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
Jens Axboe44f7c162007-01-19 11:51:58 +1100169 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200170 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200171};
172
173#define CFQ_CFQQ_FNS(name) \
174static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
175{ \
176 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
177} \
178static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
179{ \
180 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
181} \
182static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
183{ \
184 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
185}
186
187CFQ_CFQQ_FNS(on_rr);
188CFQ_CFQQ_FNS(wait_request);
189CFQ_CFQQ_FNS(must_alloc);
190CFQ_CFQQ_FNS(must_alloc_slice);
191CFQ_CFQQ_FNS(must_dispatch);
192CFQ_CFQQ_FNS(fifo_expire);
193CFQ_CFQQ_FNS(idle_window);
194CFQ_CFQQ_FNS(prio_changed);
Jens Axboe53b037442006-07-28 09:48:51 +0200195CFQ_CFQQ_FNS(queue_new);
Jens Axboe44f7c162007-01-19 11:51:58 +1100196CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200197CFQ_CFQQ_FNS(sync);
Jens Axboe3b181522005-06-27 10:56:24 +0200198#undef CFQ_CFQQ_FNS
199
Jens Axboe5e705372006-07-13 12:39:25 +0200200static void cfq_dispatch_insert(request_queue_t *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200201static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
202 struct task_struct *, gfp_t);
203static struct cfq_io_context *cfq_cic_rb_lookup(struct cfq_data *,
204 struct io_context *);
205
206static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
207 int is_sync)
208{
209 return cic->cfqq[!!is_sync];
210}
211
212static inline void cic_set_cfqq(struct cfq_io_context *cic,
213 struct cfq_queue *cfqq, int is_sync)
214{
215 cic->cfqq[!!is_sync] = cfqq;
216}
217
218/*
219 * We regard a request as SYNC, if it's either a read or has the SYNC bit
220 * set (in which case it could also be direct WRITE).
221 */
222static inline int cfq_bio_sync(struct bio *bio)
223{
224 if (bio_data_dir(bio) == READ || bio_sync(bio))
225 return 1;
226
227 return 0;
228}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700231 * scheduler run of queue, if there are requests pending and no one in the
232 * driver that will restart queueing
233 */
234static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
235{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100236 if (cfqd->busy_queues)
Andrew Morton99f95e52005-06-27 20:14:05 -0700237 kblockd_schedule_work(&cfqd->unplug_work);
238}
239
240static int cfq_queue_empty(request_queue_t *q)
241{
242 struct cfq_data *cfqd = q->elevator->elevator_data;
243
Jens Axboeb4878f22005-10-20 16:42:29 +0200244 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700245}
246
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100248 * Scale schedule slice based on io priority. Use the sync time slice only
249 * if a queue is marked sync and has sync io queued. A sync queue with async
250 * io only, should not get full sync slice length.
251 */
Jens Axboed9e76202007-04-20 14:27:50 +0200252static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
253 unsigned short prio)
254{
255 const int base_slice = cfqd->cfq_slice[sync];
256
257 WARN_ON(prio >= IOPRIO_BE_NR);
258
259 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
260}
261
Jens Axboe44f7c162007-01-19 11:51:58 +1100262static inline int
263cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
264{
Jens Axboed9e76202007-04-20 14:27:50 +0200265 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100266}
267
268static inline void
269cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
270{
271 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
272}
273
274/*
275 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
276 * isn't valid until the first request from the dispatch is activated
277 * and the slice time set.
278 */
279static inline int cfq_slice_used(struct cfq_queue *cfqq)
280{
281 if (cfq_cfqq_slice_new(cfqq))
282 return 0;
283 if (time_before(jiffies, cfqq->slice_end))
284 return 0;
285
286 return 1;
287}
288
289/*
Jens Axboe5e705372006-07-13 12:39:25 +0200290 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200292 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293 */
Jens Axboe5e705372006-07-13 12:39:25 +0200294static struct request *
295cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296{
297 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200299#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
300#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
301 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302
Jens Axboe5e705372006-07-13 12:39:25 +0200303 if (rq1 == NULL || rq1 == rq2)
304 return rq2;
305 if (rq2 == NULL)
306 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200307
Jens Axboe5e705372006-07-13 12:39:25 +0200308 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
309 return rq1;
310 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
311 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200312 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
313 return rq1;
314 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
315 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316
Jens Axboe5e705372006-07-13 12:39:25 +0200317 s1 = rq1->sector;
318 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319
Jens Axboe6d048f52007-04-25 12:44:27 +0200320 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322 /*
323 * by definition, 1KiB is 2 sectors
324 */
325 back_max = cfqd->cfq_back_max * 2;
326
327 /*
328 * Strict one way elevator _except_ in the case where we allow
329 * short backward seeks which are biased as twice the cost of a
330 * similar forward seek.
331 */
332 if (s1 >= last)
333 d1 = s1 - last;
334 else if (s1 + back_max >= last)
335 d1 = (last - s1) * cfqd->cfq_back_penalty;
336 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200337 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338
339 if (s2 >= last)
340 d2 = s2 - last;
341 else if (s2 + back_max >= last)
342 d2 = (last - s2) * cfqd->cfq_back_penalty;
343 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200344 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345
346 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347
Andreas Mohre8a99052006-03-28 08:59:49 +0200348 /*
349 * By doing switch() on the bit mask "wrap" we avoid having to
350 * check two variables for all permutations: --> faster!
351 */
352 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200353 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200354 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200355 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200356 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200357 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200358 else {
359 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200360 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200361 else
Jens Axboe5e705372006-07-13 12:39:25 +0200362 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200363 }
364
365 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200366 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200367 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200368 return rq2;
369 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200370 default:
371 /*
372 * Since both rqs are wrapped,
373 * start with the one that's further behind head
374 * (--> only *one* back seek required),
375 * since back seek takes more time than forward.
376 */
377 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200378 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 else
Jens Axboe5e705372006-07-13 12:39:25 +0200380 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 }
382}
383
Jens Axboe498d3aa22007-04-26 12:54:48 +0200384/*
385 * The below is leftmost cache rbtree addon
386 */
Jens Axboecc09e292007-04-26 12:53:50 +0200387static struct rb_node *cfq_rb_first(struct cfq_rb_root *root)
388{
389 if (!root->left)
390 root->left = rb_first(&root->rb);
391
392 return root->left;
393}
394
395static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
396{
397 if (root->left == n)
398 root->left = NULL;
399
400 rb_erase(n, &root->rb);
401 RB_CLEAR_NODE(n);
402}
403
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404/*
405 * would be nice to take fifo expire time into account as well
406 */
Jens Axboe5e705372006-07-13 12:39:25 +0200407static struct request *
408cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
409 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410{
Jens Axboe21183b02006-07-13 12:33:14 +0200411 struct rb_node *rbnext = rb_next(&last->rb_node);
412 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200413 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414
Jens Axboe21183b02006-07-13 12:33:14 +0200415 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
417 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200418 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200419
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200421 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200422 else {
423 rbnext = rb_first(&cfqq->sort_list);
424 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200425 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200426 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427
Jens Axboe21183b02006-07-13 12:33:14 +0200428 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429}
430
Jens Axboed9e76202007-04-20 14:27:50 +0200431static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
432 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433{
Jens Axboed9e76202007-04-20 14:27:50 +0200434 /*
435 * just an approximation, should be ok.
436 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200437 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
438 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200439}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440
Jens Axboe498d3aa22007-04-26 12:54:48 +0200441/*
442 * The cfqd->service_tree holds all pending cfq_queue's that have
443 * requests waiting to be processed. It is sorted in the order that
444 * we will service the queues.
445 */
Jens Axboed9e76202007-04-20 14:27:50 +0200446static void cfq_service_tree_add(struct cfq_data *cfqd,
Jens Axboeedd75ff2007-04-19 12:03:34 +0200447 struct cfq_queue *cfqq, int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200448{
Jens Axboecc09e292007-04-26 12:53:50 +0200449 struct rb_node **p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200450 struct rb_node *parent = NULL;
Jens Axboed9e76202007-04-20 14:27:50 +0200451 unsigned long rb_key;
Jens Axboe498d3aa22007-04-26 12:54:48 +0200452 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200453
Jens Axboeedd75ff2007-04-19 12:03:34 +0200454 if (!add_front) {
455 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
456 rb_key += cfqq->slice_resid;
457 cfqq->slice_resid = 0;
458 } else
459 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200460
461 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100462 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200463 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100464 */
Jens Axboed9e76202007-04-20 14:27:50 +0200465 if (rb_key == cfqq->rb_key)
466 return;
Jens Axboe53b037442006-07-28 09:48:51 +0200467
Jens Axboecc09e292007-04-26 12:53:50 +0200468 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200469 }
Jens Axboed9e76202007-04-20 14:27:50 +0200470
Jens Axboe498d3aa22007-04-26 12:54:48 +0200471 left = 1;
Jens Axboed9e76202007-04-20 14:27:50 +0200472 while (*p) {
Jens Axboecc09e292007-04-26 12:53:50 +0200473 struct cfq_queue *__cfqq;
Jens Axboe67060e32007-04-18 20:13:32 +0200474 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200475
Jens Axboed9e76202007-04-20 14:27:50 +0200476 parent = *p;
477 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
478
Jens Axboe0c534e02007-04-18 20:01:57 +0200479 /*
480 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200481 * preference to them. IDLE queues goes to the back.
482 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200483 */
484 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200485 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200486 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200487 n = &(*p)->rb_right;
488 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
489 n = &(*p)->rb_left;
490 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
491 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200492 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200493 n = &(*p)->rb_left;
494 else
495 n = &(*p)->rb_right;
496
497 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200498 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200499
500 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200501 }
502
Jens Axboecc09e292007-04-26 12:53:50 +0200503 if (left)
504 cfqd->service_tree.left = &cfqq->rb_node;
505
Jens Axboed9e76202007-04-20 14:27:50 +0200506 cfqq->rb_key = rb_key;
507 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200508 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509}
510
Jens Axboe498d3aa22007-04-26 12:54:48 +0200511/*
512 * Update cfqq's position in the service tree.
513 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200514static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200515{
Jens Axboe6d048f52007-04-25 12:44:27 +0200516 /*
517 * Resorting requires the cfqq to be on the RR list already.
518 */
Jens Axboe498d3aa22007-04-26 12:54:48 +0200519 if (cfq_cfqq_on_rr(cfqq))
Jens Axboeedd75ff2007-04-19 12:03:34 +0200520 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboe6d048f52007-04-25 12:44:27 +0200521}
522
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523/*
524 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200525 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 */
527static inline void
Jens Axboeb4878f22005-10-20 16:42:29 +0200528cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
Jens Axboe3b181522005-06-27 10:56:24 +0200530 BUG_ON(cfq_cfqq_on_rr(cfqq));
531 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 cfqd->busy_queues++;
533
Jens Axboeedd75ff2007-04-19 12:03:34 +0200534 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535}
536
Jens Axboe498d3aa22007-04-26 12:54:48 +0200537/*
538 * Called when the cfqq no longer has requests pending, remove it from
539 * the service tree.
540 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541static inline void
542cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
543{
Jens Axboe3b181522005-06-27 10:56:24 +0200544 BUG_ON(!cfq_cfqq_on_rr(cfqq));
545 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546
Jens Axboecc09e292007-04-26 12:53:50 +0200547 if (!RB_EMPTY_NODE(&cfqq->rb_node))
548 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboed9e76202007-04-20 14:27:50 +0200549
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 BUG_ON(!cfqd->busy_queues);
551 cfqd->busy_queues--;
552}
553
554/*
555 * rb tree support functions
556 */
Jens Axboe5e705372006-07-13 12:39:25 +0200557static inline void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558{
Jens Axboe5e705372006-07-13 12:39:25 +0200559 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200560 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200561 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562
Jens Axboeb4878f22005-10-20 16:42:29 +0200563 BUG_ON(!cfqq->queued[sync]);
564 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565
Jens Axboe5e705372006-07-13 12:39:25 +0200566 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567
Jens Axboedd67d052006-06-21 09:36:18 +0200568 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200569 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570}
571
Jens Axboe5e705372006-07-13 12:39:25 +0200572static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573{
Jens Axboe5e705372006-07-13 12:39:25 +0200574 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe21183b02006-07-13 12:33:14 +0200576 struct request *__alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577
Jens Axboe5380a102006-07-13 12:37:56 +0200578 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579
580 /*
581 * looks a little odd, but the first insert might return an alias.
582 * if that happens, put the alias on the dispatch list
583 */
Jens Axboe21183b02006-07-13 12:33:14 +0200584 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200585 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100586
587 if (!cfq_cfqq_on_rr(cfqq))
588 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200589
590 /*
591 * check if this request is a better next-serve candidate
592 */
593 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
594 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595}
596
597static inline void
Jens Axboe5e705372006-07-13 12:39:25 +0200598cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599{
Jens Axboe5380a102006-07-13 12:37:56 +0200600 elv_rb_del(&cfqq->sort_list, rq);
601 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200602 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603}
604
Jens Axboe206dc692006-03-28 13:03:44 +0200605static struct request *
606cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607{
Jens Axboe206dc692006-03-28 13:03:44 +0200608 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200609 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200610 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611
Vasily Tarasov91fac312007-04-25 12:29:51 +0200612 cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
613 if (!cic)
614 return NULL;
615
616 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200617 if (cfqq) {
618 sector_t sector = bio->bi_sector + bio_sectors(bio);
619
Jens Axboe21183b02006-07-13 12:33:14 +0200620 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200621 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623 return NULL;
624}
625
Jens Axboeb4878f22005-10-20 16:42:29 +0200626static void cfq_activate_request(request_queue_t *q, struct request *rq)
627{
628 struct cfq_data *cfqd = q->elevator->elevator_data;
629
630 cfqd->rq_in_driver++;
Jens Axboe25776e32006-06-01 10:12:26 +0200631
632 /*
633 * If the depth is larger 1, it really could be queueing. But lets
634 * make the mark a little higher - idling could still be good for
635 * low queueing, and a low queueing number could also just indicate
636 * a SCSI mid layer like behaviour where limit+1 is often seen.
637 */
638 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
639 cfqd->hw_tag = 1;
Jens Axboe6d048f52007-04-25 12:44:27 +0200640
641 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
Jens Axboeb4878f22005-10-20 16:42:29 +0200642}
643
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
645{
Jens Axboe22e2c502005-06-27 10:55:12 +0200646 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647
Jens Axboeb4878f22005-10-20 16:42:29 +0200648 WARN_ON(!cfqd->rq_in_driver);
649 cfqd->rq_in_driver--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650}
651
Jens Axboeb4878f22005-10-20 16:42:29 +0200652static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653{
Jens Axboe5e705372006-07-13 12:39:25 +0200654 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200655
Jens Axboe5e705372006-07-13 12:39:25 +0200656 if (cfqq->next_rq == rq)
657 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658
Jens Axboeb4878f22005-10-20 16:42:29 +0200659 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200660 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200661
662 if (rq_is_meta(rq)) {
663 WARN_ON(!cfqq->meta_pending);
664 cfqq->meta_pending--;
665 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666}
667
Jens Axboe498d3aa22007-04-26 12:54:48 +0200668static int cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669{
670 struct cfq_data *cfqd = q->elevator->elevator_data;
671 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672
Jens Axboe206dc692006-03-28 13:03:44 +0200673 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200674 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200675 *req = __rq;
676 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 }
678
679 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680}
681
Jens Axboe21183b02006-07-13 12:33:14 +0200682static void cfq_merged_request(request_queue_t *q, struct request *req,
683 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684{
Jens Axboe21183b02006-07-13 12:33:14 +0200685 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200686 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687
Jens Axboe5e705372006-07-13 12:39:25 +0200688 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690}
691
692static void
693cfq_merged_requests(request_queue_t *q, struct request *rq,
694 struct request *next)
695{
Jens Axboe22e2c502005-06-27 10:55:12 +0200696 /*
697 * reposition in fifo if next is older than rq
698 */
699 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
700 time_before(next->start_time, rq->start_time))
701 list_move(&rq->queuelist, &next->queuelist);
702
Jens Axboeb4878f22005-10-20 16:42:29 +0200703 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200704}
705
Jens Axboeda775262006-12-20 11:04:12 +0100706static int cfq_allow_merge(request_queue_t *q, struct request *rq,
707 struct bio *bio)
708{
709 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200710 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100711 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100712
713 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100714 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100715 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200716 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100717 return 0;
718
719 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100720 * Lookup the cfqq that this bio will be queued with. Allow
721 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100722 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200723 cic = cfq_cic_rb_lookup(cfqd, current->io_context);
724 if (!cic)
725 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100726
Vasily Tarasov91fac312007-04-25 12:29:51 +0200727 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100728 if (cfqq == RQ_CFQQ(rq))
729 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100730
Jens Axboeec8acb62007-01-02 18:32:11 +0100731 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100732}
733
Jens Axboe22e2c502005-06-27 10:55:12 +0200734static inline void
735__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
736{
737 if (cfqq) {
738 /*
739 * stop potential idle class queues waiting service
740 */
741 del_timer(&cfqd->idle_class_timer);
742
Jens Axboe22e2c502005-06-27 10:55:12 +0200743 cfqq->slice_end = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200744 cfq_clear_cfqq_must_alloc_slice(cfqq);
745 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100746 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe1afba042007-04-17 12:47:55 +0200747 cfq_clear_cfqq_queue_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200748 }
749
750 cfqd->active_queue = cfqq;
751}
752
753/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100754 * current cfqq expired its slice (or was too idle), select new one
755 */
756static void
757__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200758 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100759{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100760 if (cfq_cfqq_wait_request(cfqq))
761 del_timer(&cfqd->idle_slice_timer);
762
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100763 cfq_clear_cfqq_must_dispatch(cfqq);
764 cfq_clear_cfqq_wait_request(cfqq);
765
766 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200767 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100768 */
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100769 if (timed_out && !cfq_cfqq_slice_new(cfqq))
Jens Axboec5b680f2007-01-19 11:56:49 +1100770 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100771
Jens Axboeedd75ff2007-04-19 12:03:34 +0200772 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100773
774 if (cfqq == cfqd->active_queue)
775 cfqd->active_queue = NULL;
776
777 if (cfqd->active_cic) {
778 put_io_context(cfqd->active_cic->ioc);
779 cfqd->active_cic = NULL;
780 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100781}
782
Jens Axboe6084cdd2007-04-23 08:25:00 +0200783static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100784{
785 struct cfq_queue *cfqq = cfqd->active_queue;
786
787 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200788 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100789}
790
Jens Axboe498d3aa22007-04-26 12:54:48 +0200791/*
792 * Get next queue for service. Unless we have a queue preemption,
793 * we'll simply select the first cfqq in the service tree.
794 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200795static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200796{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200797 struct cfq_queue *cfqq;
798 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +0200799
Jens Axboeedd75ff2007-04-19 12:03:34 +0200800 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
801 return NULL;
802
803 n = cfq_rb_first(&cfqd->service_tree);
804 cfqq = rb_entry(n, struct cfq_queue, rb_node);
805
806 if (cfq_class_idle(cfqq)) {
807 unsigned long end;
808
Jens Axboe89850f72006-07-22 16:48:31 +0200809 /*
Jens Axboeedd75ff2007-04-19 12:03:34 +0200810 * if we have idle queues and no rt or be queues had
811 * pending requests, either allow immediate service if
812 * the grace period has passed or arm the idle grace
813 * timer
Jens Axboe89850f72006-07-22 16:48:31 +0200814 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200815 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
816 if (time_before(jiffies, end)) {
817 mod_timer(&cfqd->idle_class_timer, end);
818 cfqq = NULL;
Jens Axboe67060e32007-04-18 20:13:32 +0200819 }
Jens Axboe22e2c502005-06-27 10:55:12 +0200820 }
821
Jens Axboe6d048f52007-04-25 12:44:27 +0200822 return cfqq;
823}
824
Jens Axboe498d3aa22007-04-26 12:54:48 +0200825/*
826 * Get and set a new active queue for service.
827 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200828static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
829{
830 struct cfq_queue *cfqq;
831
Jens Axboed9e76202007-04-20 14:27:50 +0200832 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200833 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200834 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200835}
836
Jens Axboed9e76202007-04-20 14:27:50 +0200837static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
838 struct request *rq)
839{
840 if (rq->sector >= cfqd->last_position)
841 return rq->sector - cfqd->last_position;
842 else
843 return cfqd->last_position - rq->sector;
844}
845
Jens Axboe6d048f52007-04-25 12:44:27 +0200846static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
847{
848 struct cfq_io_context *cic = cfqd->active_cic;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200849
Jens Axboe6d048f52007-04-25 12:44:27 +0200850 if (!sample_valid(cic->seek_samples))
851 return 0;
852
853 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
854}
855
Jens Axboed9e76202007-04-20 14:27:50 +0200856static int cfq_close_cooperator(struct cfq_data *cfq_data,
857 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200858{
Jens Axboe6d048f52007-04-25 12:44:27 +0200859 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200860 * We should notice if some of the queues are cooperating, eg
861 * working closely on the same area of the disk. In that case,
862 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +0200863 */
Jens Axboed9e76202007-04-20 14:27:50 +0200864 return 0;
Jens Axboe6d048f52007-04-25 12:44:27 +0200865}
866
867#define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
868
869static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200870{
Jens Axboe17926692007-01-19 11:59:30 +1100871 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +0200872 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100873 unsigned long sl;
874
Jens Axboedd67d052006-06-21 09:36:18 +0200875 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +0200876 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +0200877
878 /*
879 * idle is disabled, either manually or by past process history
880 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200881 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
882 return;
883
Jens Axboe22e2c502005-06-27 10:55:12 +0200884 /*
885 * task has exited, don't wait
886 */
Jens Axboe206dc692006-03-28 13:03:44 +0200887 cic = cfqd->active_cic;
888 if (!cic || !cic->ioc->task)
Jens Axboe6d048f52007-04-25 12:44:27 +0200889 return;
890
891 /*
892 * See if this prio level has a good candidate
893 */
Jens Axboe1afba042007-04-17 12:47:55 +0200894 if (cfq_close_cooperator(cfqd, cfqq) &&
895 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
Jens Axboe6d048f52007-04-25 12:44:27 +0200896 return;
Jens Axboe22e2c502005-06-27 10:55:12 +0200897
Jens Axboe3b181522005-06-27 10:56:24 +0200898 cfq_mark_cfqq_must_dispatch(cfqq);
899 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200900
Jens Axboe206dc692006-03-28 13:03:44 +0200901 /*
902 * we don't want to idle for seeks, but we do want to allow
903 * fair distribution of slice time for a process doing back-to-back
904 * seeks. so allow a little bit of time for him to submit a new rq
905 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200906 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200907 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +0200908 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +0200909
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100910 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911}
912
Jens Axboe498d3aa22007-04-26 12:54:48 +0200913/*
914 * Move request from internal lists to the request queue dispatch list.
915 */
Jens Axboe5e705372006-07-13 12:39:25 +0200916static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917{
Jens Axboe3ed9a292007-04-23 08:33:33 +0200918 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200919 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200920
Jens Axboe5380a102006-07-13 12:37:56 +0200921 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200922 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +0200923 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +0200924
925 if (cfq_cfqq_sync(cfqq))
926 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927}
928
929/*
930 * return expired entry, or NULL to just start from scratch in rbtree
931 */
Jens Axboe5e705372006-07-13 12:39:25 +0200932static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933{
934 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200935 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200936 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937
Jens Axboe3b181522005-06-27 10:56:24 +0200938 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +1100940
941 cfq_mark_cfqq_fifo_expire(cfqq);
942
Jens Axboe89850f72006-07-22 16:48:31 +0200943 if (list_empty(&cfqq->fifo))
944 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945
Jens Axboe6d048f52007-04-25 12:44:27 +0200946 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +0200947 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948
Jens Axboe6d048f52007-04-25 12:44:27 +0200949 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
950 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951
Jens Axboe6d048f52007-04-25 12:44:27 +0200952 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953}
954
Jens Axboe22e2c502005-06-27 10:55:12 +0200955static inline int
956cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
957{
958 const int base_rq = cfqd->cfq_slice_async_rq;
959
960 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
961
962 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
963}
964
965/*
Jens Axboe498d3aa22007-04-26 12:54:48 +0200966 * Select a queue for service. If we have a current active queue,
967 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +0200968 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100969static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200970{
Jens Axboe22e2c502005-06-27 10:55:12 +0200971 struct cfq_queue *cfqq;
972
973 cfqq = cfqd->active_queue;
974 if (!cfqq)
975 goto new_queue;
976
977 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200978 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +0200979 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200980 if (cfq_slice_used(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +0200981 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +0200982
983 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200984 * The active queue has requests and isn't expired, allow it to
985 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +0200986 */
Jens Axboedd67d052006-06-21 09:36:18 +0200987 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200988 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +0200989
990 /*
991 * No requests pending. If the active queue still has requests in
992 * flight or is idling for a new request, allow either of these
993 * conditions to happen (or time out) before selecting a new queue.
994 */
Jens Axboecc197472007-04-20 20:45:39 +0200995 if (timer_pending(&cfqd->idle_slice_timer) ||
996 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +0200997 cfqq = NULL;
998 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +0200999 }
1000
Jens Axboe3b181522005-06-27 10:56:24 +02001001expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001002 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001003new_queue:
1004 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001005keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001006 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001007}
1008
Jens Axboe498d3aa22007-04-26 12:54:48 +02001009/*
1010 * Dispatch some requests from cfqq, moving them to the request queue
1011 * dispatch list.
1012 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001013static int
1014__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1015 int max_dispatch)
1016{
1017 int dispatched = 0;
1018
Jens Axboedd67d052006-06-21 09:36:18 +02001019 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001020
1021 do {
Jens Axboe5e705372006-07-13 12:39:25 +02001022 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001023
1024 /*
1025 * follow expired path, else get first next available
1026 */
Jens Axboe5e705372006-07-13 12:39:25 +02001027 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1028 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001029
1030 /*
1031 * finally, insert request into driver dispatch list
1032 */
Jens Axboe5e705372006-07-13 12:39:25 +02001033 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001034
Jens Axboe22e2c502005-06-27 10:55:12 +02001035 dispatched++;
1036
1037 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +02001038 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1039 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001040 }
1041
Jens Axboedd67d052006-06-21 09:36:18 +02001042 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001043 break;
1044
1045 } while (dispatched < max_dispatch);
1046
1047 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001048 * expire an async queue immediately if it has used up its slice. idle
1049 * queue always expire after 1 dispatch round.
1050 */
Jens Axboea9938002007-04-20 08:55:52 +02001051 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
Jens Axboe20e493a2007-04-23 08:26:36 +02001052 dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboea9938002007-04-20 08:55:52 +02001053 cfq_class_idle(cfqq))) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001054 cfqq->slice_end = jiffies + 1;
Jens Axboe6084cdd2007-04-23 08:25:00 +02001055 cfq_slice_expired(cfqd, 0);
Jens Axboe44f7c162007-01-19 11:51:58 +11001056 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001057
1058 return dispatched;
1059}
1060
Jens Axboed9e76202007-04-20 14:27:50 +02001061static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1062{
1063 int dispatched = 0;
1064
1065 while (cfqq->next_rq) {
1066 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1067 dispatched++;
1068 }
1069
1070 BUG_ON(!list_empty(&cfqq->fifo));
1071 return dispatched;
1072}
1073
Jens Axboe498d3aa22007-04-26 12:54:48 +02001074/*
1075 * Drain our current requests. Used for barriers and when switching
1076 * io schedulers on-the-fly.
1077 */
Jens Axboed9e76202007-04-20 14:27:50 +02001078static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001079{
Jens Axboed9e76202007-04-20 14:27:50 +02001080 int dispatched = 0;
1081 struct rb_node *n;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001082
Jens Axboecc09e292007-04-26 12:53:50 +02001083 while ((n = cfq_rb_first(&cfqd->service_tree)) != NULL) {
Jens Axboed9e76202007-04-20 14:27:50 +02001084 struct cfq_queue *cfqq = rb_entry(n, struct cfq_queue, rb_node);
1085
1086 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1087 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001088
Jens Axboe6084cdd2007-04-23 08:25:00 +02001089 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001090
1091 BUG_ON(cfqd->busy_queues);
1092
1093 return dispatched;
1094}
1095
Jens Axboed9e76202007-04-20 14:27:50 +02001096static int cfq_dispatch_requests(request_queue_t *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097{
1098 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001099 struct cfq_queue *cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001100 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101
Jens Axboe22e2c502005-06-27 10:55:12 +02001102 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 return 0;
1104
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001105 if (unlikely(force))
1106 return cfq_forced_dispatch(cfqd);
1107
Jens Axboecaaa5f92006-06-16 11:23:00 +02001108 dispatched = 0;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001109 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001110 int max_dispatch;
1111
Jens Axboe3ed9a292007-04-23 08:33:33 +02001112 max_dispatch = cfqd->cfq_quantum;
1113 if (cfq_class_idle(cfqq))
1114 max_dispatch = 1;
1115
1116 if (cfqq->dispatched >= max_dispatch) {
1117 if (cfqd->busy_queues > 1)
Jens Axboe6d048f52007-04-25 12:44:27 +02001118 break;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001119 if (cfqq->dispatched >= 4 * max_dispatch)
Jens Axboea9938002007-04-20 08:55:52 +02001120 break;
1121 }
Jens Axboe9ede2092007-01-19 12:11:44 +11001122
Jens Axboe3ed9a292007-04-23 08:33:33 +02001123 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1124 break;
1125
Jens Axboe3b181522005-06-27 10:56:24 +02001126 cfq_clear_cfqq_must_dispatch(cfqq);
1127 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001128 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129
Jens Axboecaaa5f92006-06-16 11:23:00 +02001130 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 }
1132
Jens Axboecaaa5f92006-06-16 11:23:00 +02001133 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134}
1135
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136/*
Jens Axboe5e705372006-07-13 12:39:25 +02001137 * task holds one reference to the queue, dropped when task exits. each rq
1138 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 *
1140 * queue lock must be held here.
1141 */
1142static void cfq_put_queue(struct cfq_queue *cfqq)
1143{
Jens Axboe22e2c502005-06-27 10:55:12 +02001144 struct cfq_data *cfqd = cfqq->cfqd;
1145
1146 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147
1148 if (!atomic_dec_and_test(&cfqq->ref))
1149 return;
1150
1151 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001152 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001153 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001155 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001156 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001157 cfq_schedule_dispatch(cfqd);
1158 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001159
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160 kmem_cache_free(cfq_pool, cfqq);
1161}
1162
Jens Axboee2d74ac2006-03-28 08:59:01 +02001163static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164{
Jens Axboe22e2c502005-06-27 10:55:12 +02001165 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001166 struct rb_node *n;
1167 int freed = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001168
Jens Axboe597bc482007-04-24 21:23:53 +02001169 ioc->ioc_data = NULL;
1170
Jens Axboee2d74ac2006-03-28 08:59:01 +02001171 while ((n = rb_first(&ioc->cic_root)) != NULL) {
1172 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1173 rb_erase(&__cic->rb_node, &ioc->cic_root);
Jens Axboe22e2c502005-06-27 10:55:12 +02001174 kmem_cache_free(cfq_ioc_pool, __cic);
Al Viro334e94d2006-03-18 15:05:53 -05001175 freed++;
Jens Axboe22e2c502005-06-27 10:55:12 +02001176 }
1177
Jens Axboe4050cf12006-07-19 05:07:12 +02001178 elv_ioc_count_mod(ioc_count, -freed);
1179
1180 if (ioc_gone && !elv_ioc_count_read(ioc_count))
Al Viro334e94d2006-03-18 15:05:53 -05001181 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182}
1183
Jens Axboe89850f72006-07-22 16:48:31 +02001184static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1185{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001186 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001187 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001188 cfq_schedule_dispatch(cfqd);
1189 }
Jens Axboe89850f72006-07-22 16:48:31 +02001190
1191 cfq_put_queue(cfqq);
1192}
1193
1194static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1195 struct cfq_io_context *cic)
1196{
Jens Axboefc463792006-08-29 09:05:44 +02001197 list_del_init(&cic->queue_list);
1198 smp_wmb();
1199 cic->key = NULL;
1200
Jens Axboe89850f72006-07-22 16:48:31 +02001201 if (cic->cfqq[ASYNC]) {
1202 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1203 cic->cfqq[ASYNC] = NULL;
1204 }
1205
1206 if (cic->cfqq[SYNC]) {
1207 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1208 cic->cfqq[SYNC] = NULL;
1209 }
Jens Axboe89850f72006-07-22 16:48:31 +02001210}
1211
Jens Axboe22e2c502005-06-27 10:55:12 +02001212static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1213{
Al Viro478a82b2006-03-18 13:25:24 -05001214 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001215
Jens Axboe89850f72006-07-22 16:48:31 +02001216 if (cfqd) {
1217 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001218
Jens Axboefc463792006-08-29 09:05:44 +02001219 spin_lock_irq(q->queue_lock);
Jens Axboe89850f72006-07-22 16:48:31 +02001220 __cfq_exit_single_io_context(cfqd, cic);
Jens Axboefc463792006-08-29 09:05:44 +02001221 spin_unlock_irq(q->queue_lock);
Al Viro12a05732006-03-18 13:38:01 -05001222 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001223}
1224
Jens Axboe498d3aa22007-04-26 12:54:48 +02001225/*
1226 * The process that ioc belongs to has exited, we need to clean up
1227 * and put the internal structures we have that belongs to that process.
1228 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001229static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230{
Jens Axboe22e2c502005-06-27 10:55:12 +02001231 struct cfq_io_context *__cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001232 struct rb_node *n;
Jens Axboe22e2c502005-06-27 10:55:12 +02001233
Jens Axboe597bc482007-04-24 21:23:53 +02001234 ioc->ioc_data = NULL;
1235
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236 /*
1237 * put the reference this task is holding to the various queues
1238 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001239 n = rb_first(&ioc->cic_root);
1240 while (n != NULL) {
1241 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1242
Jens Axboe22e2c502005-06-27 10:55:12 +02001243 cfq_exit_single_io_context(__cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001244 n = rb_next(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246}
1247
Jens Axboe22e2c502005-06-27 10:55:12 +02001248static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001249cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250{
Jens Axboeb5deef92006-07-19 23:39:40 +02001251 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252
Christoph Lameter94f60302007-07-17 04:03:29 -07001253 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1254 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001256 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001257 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001258 cic->dtor = cfq_free_io_context;
1259 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001260 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 }
1262
1263 return cic;
1264}
1265
Jens Axboe22e2c502005-06-27 10:55:12 +02001266static void cfq_init_prio_data(struct cfq_queue *cfqq)
1267{
1268 struct task_struct *tsk = current;
1269 int ioprio_class;
1270
Jens Axboe3b181522005-06-27 10:56:24 +02001271 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001272 return;
1273
1274 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1275 switch (ioprio_class) {
1276 default:
1277 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1278 case IOPRIO_CLASS_NONE:
1279 /*
1280 * no prio set, place us in the middle of the BE classes
1281 */
1282 cfqq->ioprio = task_nice_ioprio(tsk);
1283 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1284 break;
1285 case IOPRIO_CLASS_RT:
1286 cfqq->ioprio = task_ioprio(tsk);
1287 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1288 break;
1289 case IOPRIO_CLASS_BE:
1290 cfqq->ioprio = task_ioprio(tsk);
1291 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1292 break;
1293 case IOPRIO_CLASS_IDLE:
1294 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1295 cfqq->ioprio = 7;
Jens Axboe3b181522005-06-27 10:56:24 +02001296 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001297 break;
1298 }
1299
1300 /*
1301 * keep track of original prio settings in case we have to temporarily
1302 * elevate the priority of this queue
1303 */
1304 cfqq->org_ioprio = cfqq->ioprio;
1305 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001306 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001307}
1308
Al Viro478a82b2006-03-18 13:25:24 -05001309static inline void changed_ioprio(struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001310{
Al Viro478a82b2006-03-18 13:25:24 -05001311 struct cfq_data *cfqd = cic->key;
1312 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001313 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001314
Jens Axboecaaa5f92006-06-16 11:23:00 +02001315 if (unlikely(!cfqd))
1316 return;
1317
Jens Axboec1b707d2006-10-30 19:54:23 +01001318 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001319
1320 cfqq = cic->cfqq[ASYNC];
1321 if (cfqq) {
1322 struct cfq_queue *new_cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001323 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc->task,
Jens Axboecaaa5f92006-06-16 11:23:00 +02001324 GFP_ATOMIC);
1325 if (new_cfqq) {
1326 cic->cfqq[ASYNC] = new_cfqq;
1327 cfq_put_queue(cfqq);
1328 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001329 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001330
1331 cfqq = cic->cfqq[SYNC];
1332 if (cfqq)
1333 cfq_mark_cfqq_prio_changed(cfqq);
1334
Jens Axboec1b707d2006-10-30 19:54:23 +01001335 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001336}
1337
Jens Axboefc463792006-08-29 09:05:44 +02001338static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339{
Al Viroa6a07632006-03-18 13:26:44 -05001340 struct cfq_io_context *cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001341 struct rb_node *n;
Al Viroa6a07632006-03-18 13:26:44 -05001342
Jens Axboefc463792006-08-29 09:05:44 +02001343 ioc->ioprio_changed = 0;
Al Viroa6a07632006-03-18 13:26:44 -05001344
Jens Axboee2d74ac2006-03-28 08:59:01 +02001345 n = rb_first(&ioc->cic_root);
1346 while (n != NULL) {
1347 cic = rb_entry(n, struct cfq_io_context, rb_node);
Jens Axboe3793c652006-05-30 21:11:04 +02001348
Al Viro478a82b2006-03-18 13:25:24 -05001349 changed_ioprio(cic);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001350 n = rb_next(n);
1351 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352}
1353
1354static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001355cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1356 struct task_struct *tsk, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001359 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360
1361retry:
Vasily Tarasov91fac312007-04-25 12:29:51 +02001362 cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1363 /* cic always exists here */
1364 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365
1366 if (!cfqq) {
1367 if (new_cfqq) {
1368 cfqq = new_cfqq;
1369 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001370 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001371 /*
1372 * Inform the allocator of the fact that we will
1373 * just repeat this allocation if it fails, to allow
1374 * the allocator to do whatever it needs to attempt to
1375 * free memory.
1376 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001378 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1379 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1380 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 spin_lock_irq(cfqd->queue->queue_lock);
1382 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001383 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001384 cfqq = kmem_cache_alloc_node(cfq_pool,
1385 gfp_mask | __GFP_ZERO,
1386 cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001387 if (!cfqq)
1388 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001389 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390
Jens Axboed9e76202007-04-20 14:27:50 +02001391 RB_CLEAR_NODE(&cfqq->rb_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001392 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 atomic_set(&cfqq->ref, 0);
1395 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001396
Vasily Tarasov91fac312007-04-25 12:29:51 +02001397 if (is_sync) {
Jens Axboea9938002007-04-20 08:55:52 +02001398 cfq_mark_cfqq_idle_window(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001399 cfq_mark_cfqq_sync(cfqq);
1400 }
Jens Axboea9938002007-04-20 08:55:52 +02001401
Jens Axboe3b181522005-06-27 10:56:24 +02001402 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b037442006-07-28 09:48:51 +02001403 cfq_mark_cfqq_queue_new(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001404
Jens Axboe3b181522005-06-27 10:56:24 +02001405 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 }
1407
1408 if (new_cfqq)
1409 kmem_cache_free(cfq_pool, new_cfqq);
1410
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411out:
1412 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1413 return cfqq;
1414}
1415
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001416static struct cfq_queue **
1417cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1418{
1419 switch(ioprio_class) {
1420 case IOPRIO_CLASS_RT:
1421 return &cfqd->async_cfqq[0][ioprio];
1422 case IOPRIO_CLASS_BE:
1423 return &cfqd->async_cfqq[1][ioprio];
1424 case IOPRIO_CLASS_IDLE:
1425 return &cfqd->async_idle_cfqq;
1426 default:
1427 BUG();
1428 }
1429}
1430
Jens Axboe15c31be2007-07-10 13:43:25 +02001431static struct cfq_queue *
1432cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct task_struct *tsk,
1433 gfp_t gfp_mask)
1434{
1435 const int ioprio = task_ioprio(tsk);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001436 const int ioprio_class = task_ioprio_class(tsk);
1437 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001438 struct cfq_queue *cfqq = NULL;
1439
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001440 if (!is_sync) {
1441 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1442 cfqq = *async_cfqq;
1443 }
1444
Jens Axboe15c31be2007-07-10 13:43:25 +02001445 if (!cfqq)
1446 cfqq = cfq_find_alloc_queue(cfqd, is_sync, tsk, gfp_mask);
1447
1448 /*
1449 * pin the queue now that it's allocated, scheduler exit will prune it
1450 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001451 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001452 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001453 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001454 }
1455
1456 atomic_inc(&cfqq->ref);
1457 return cfqq;
1458}
1459
Jens Axboe498d3aa22007-04-26 12:54:48 +02001460/*
1461 * We drop cfq io contexts lazily, so we may find a dead one.
1462 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001463static void
1464cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1465{
Jens Axboefc463792006-08-29 09:05:44 +02001466 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001467
1468 if (ioc->ioc_data == cic)
1469 ioc->ioc_data = NULL;
1470
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001471 rb_erase(&cic->rb_node, &ioc->cic_root);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001472 kmem_cache_free(cfq_ioc_pool, cic);
Jens Axboe4050cf12006-07-19 05:07:12 +02001473 elv_ioc_count_dec(ioc_count);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001474}
1475
Jens Axboee2d74ac2006-03-28 08:59:01 +02001476static struct cfq_io_context *
1477cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1478{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001479 struct rb_node *n;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001480 struct cfq_io_context *cic;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001481 void *k, *key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001482
Vasily Tarasov91fac312007-04-25 12:29:51 +02001483 if (unlikely(!ioc))
1484 return NULL;
1485
Jens Axboe597bc482007-04-24 21:23:53 +02001486 /*
1487 * we maintain a last-hit cache, to avoid browsing over the tree
1488 */
1489 cic = ioc->ioc_data;
1490 if (cic && cic->key == cfqd)
1491 return cic;
1492
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001493restart:
1494 n = ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001495 while (n) {
1496 cic = rb_entry(n, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001497 /* ->key must be copied to avoid race with cfq_exit_queue() */
1498 k = cic->key;
1499 if (unlikely(!k)) {
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001500 cfq_drop_dead_cic(ioc, cic);
1501 goto restart;
1502 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001503
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001504 if (key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001505 n = n->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001506 else if (key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001507 n = n->rb_right;
Jens Axboe597bc482007-04-24 21:23:53 +02001508 else {
1509 ioc->ioc_data = cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001510 return cic;
Jens Axboe597bc482007-04-24 21:23:53 +02001511 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001512 }
1513
1514 return NULL;
1515}
1516
1517static inline void
1518cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1519 struct cfq_io_context *cic)
1520{
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001521 struct rb_node **p;
1522 struct rb_node *parent;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001523 struct cfq_io_context *__cic;
Jens Axboe0261d682006-10-30 19:07:48 +01001524 unsigned long flags;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001525 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001526
Jens Axboee2d74ac2006-03-28 08:59:01 +02001527 cic->ioc = ioc;
1528 cic->key = cfqd;
1529
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001530restart:
1531 parent = NULL;
1532 p = &ioc->cic_root.rb_node;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001533 while (*p) {
1534 parent = *p;
1535 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001536 /* ->key must be copied to avoid race with cfq_exit_queue() */
1537 k = __cic->key;
1538 if (unlikely(!k)) {
Oleg Nesterovbe33c3a2006-08-21 08:36:12 +02001539 cfq_drop_dead_cic(ioc, __cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001540 goto restart;
1541 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001542
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001543 if (cic->key < k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001544 p = &(*p)->rb_left;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001545 else if (cic->key > k)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001546 p = &(*p)->rb_right;
1547 else
1548 BUG();
1549 }
1550
1551 rb_link_node(&cic->rb_node, parent, p);
1552 rb_insert_color(&cic->rb_node, &ioc->cic_root);
Jens Axboefc463792006-08-29 09:05:44 +02001553
Jens Axboe0261d682006-10-30 19:07:48 +01001554 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001555 list_add(&cic->queue_list, &cfqd->cic_list);
Jens Axboe0261d682006-10-30 19:07:48 +01001556 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001557}
1558
Jens Axboe22e2c502005-06-27 10:55:12 +02001559/*
1560 * Setup general io context and cfq io context. There can be several cfq
1561 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001562 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001563 */
1564static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001565cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566{
Jens Axboe22e2c502005-06-27 10:55:12 +02001567 struct io_context *ioc = NULL;
1568 struct cfq_io_context *cic;
1569
1570 might_sleep_if(gfp_mask & __GFP_WAIT);
1571
Jens Axboeb5deef92006-07-19 23:39:40 +02001572 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001573 if (!ioc)
1574 return NULL;
1575
Jens Axboee2d74ac2006-03-28 08:59:01 +02001576 cic = cfq_cic_rb_lookup(cfqd, ioc);
1577 if (cic)
1578 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001579
Jens Axboee2d74ac2006-03-28 08:59:01 +02001580 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1581 if (cic == NULL)
1582 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001583
Jens Axboee2d74ac2006-03-28 08:59:01 +02001584 cfq_cic_link(cfqd, ioc, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001585out:
Jens Axboefc463792006-08-29 09:05:44 +02001586 smp_read_barrier_depends();
1587 if (unlikely(ioc->ioprio_changed))
1588 cfq_ioc_set_ioprio(ioc);
1589
Jens Axboe22e2c502005-06-27 10:55:12 +02001590 return cic;
1591err:
1592 put_io_context(ioc);
1593 return NULL;
1594}
1595
1596static void
1597cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1598{
Jens Axboeaaf12282007-01-19 11:30:16 +11001599 unsigned long elapsed = jiffies - cic->last_end_request;
1600 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001601
1602 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1603 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1604 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1605}
1606
Jens Axboe206dc692006-03-28 13:03:44 +02001607static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001608cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1609 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001610{
1611 sector_t sdist;
1612 u64 total;
1613
Jens Axboe5e705372006-07-13 12:39:25 +02001614 if (cic->last_request_pos < rq->sector)
1615 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001616 else
Jens Axboe5e705372006-07-13 12:39:25 +02001617 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001618
1619 /*
1620 * Don't allow the seek distance to get too large from the
1621 * odd fragment, pagein, etc
1622 */
1623 if (cic->seek_samples <= 60) /* second&third seek */
1624 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1625 else
1626 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1627
1628 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1629 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1630 total = cic->seek_total + (cic->seek_samples/2);
1631 do_div(total, cic->seek_samples);
1632 cic->seek_mean = (sector_t)total;
1633}
Jens Axboe22e2c502005-06-27 10:55:12 +02001634
1635/*
1636 * Disable idle window if the process thinks too long or seeks so much that
1637 * it doesn't matter
1638 */
1639static void
1640cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1641 struct cfq_io_context *cic)
1642{
Jens Axboe1be92f2f2007-04-19 14:32:26 +02001643 int enable_idle;
1644
1645 if (!cfq_cfqq_sync(cfqq))
1646 return;
1647
1648 enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001649
Jens Axboecaaa5f92006-06-16 11:23:00 +02001650 if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1651 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001652 enable_idle = 0;
1653 else if (sample_valid(cic->ttime_samples)) {
1654 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1655 enable_idle = 0;
1656 else
1657 enable_idle = 1;
1658 }
1659
Jens Axboe3b181522005-06-27 10:56:24 +02001660 if (enable_idle)
1661 cfq_mark_cfqq_idle_window(cfqq);
1662 else
1663 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001664}
1665
Jens Axboe22e2c502005-06-27 10:55:12 +02001666/*
1667 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1668 * no or if we aren't sure, a 1 will cause a preempt.
1669 */
1670static int
1671cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001672 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001673{
Jens Axboe6d048f52007-04-25 12:44:27 +02001674 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001675
Jens Axboe6d048f52007-04-25 12:44:27 +02001676 cfqq = cfqd->active_queue;
1677 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001678 return 0;
1679
Jens Axboe6d048f52007-04-25 12:44:27 +02001680 if (cfq_slice_used(cfqq))
1681 return 1;
1682
1683 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001684 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001685
1686 if (cfq_class_idle(cfqq))
1687 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001688
Jens Axboe22e2c502005-06-27 10:55:12 +02001689 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02001690 * if the new request is sync, but the currently running queue is
1691 * not, let the sync request have priority.
1692 */
Jens Axboe5e705372006-07-13 12:39:25 +02001693 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001694 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001695
Jens Axboe374f84a2006-07-23 01:42:19 +02001696 /*
1697 * So both queues are sync. Let the new request get disk time if
1698 * it's a metadata request and the current queue is doing regular IO.
1699 */
1700 if (rq_is_meta(rq) && !cfqq->meta_pending)
1701 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001702
Jens Axboe1e3335d2007-02-14 19:59:49 +01001703 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1704 return 0;
1705
1706 /*
1707 * if this request is as-good as one we would expect from the
1708 * current cfqq, let it preempt
1709 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001710 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01001711 return 1;
1712
Jens Axboe22e2c502005-06-27 10:55:12 +02001713 return 0;
1714}
1715
1716/*
1717 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1718 * let it have half of its nominal slice.
1719 */
1720static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1721{
Jens Axboe6084cdd2007-04-23 08:25:00 +02001722 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001723
Jens Axboebf572252006-07-19 20:29:12 +02001724 /*
1725 * Put the new queue at the front of the of the current list,
1726 * so we know that it will be selected next.
1727 */
1728 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02001729
1730 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02001731
Jens Axboe44f7c162007-01-19 11:51:58 +11001732 cfqq->slice_end = 0;
1733 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001734}
1735
1736/*
Jens Axboe5e705372006-07-13 12:39:25 +02001737 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001738 * something we should do about it
1739 */
1740static void
Jens Axboe5e705372006-07-13 12:39:25 +02001741cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1742 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001743{
Jens Axboe5e705372006-07-13 12:39:25 +02001744 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001745
Jens Axboe374f84a2006-07-23 01:42:19 +02001746 if (rq_is_meta(rq))
1747 cfqq->meta_pending++;
1748
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001749 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02001750 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001751 cfq_update_idle_window(cfqd, cfqq, cic);
1752
Jens Axboe5e705372006-07-13 12:39:25 +02001753 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001754
1755 if (cfqq == cfqd->active_queue) {
1756 /*
1757 * if we are waiting for a request for this queue, let it rip
1758 * immediately and flag that we must not expire this queue
1759 * just now
1760 */
Jens Axboe3b181522005-06-27 10:56:24 +02001761 if (cfq_cfqq_wait_request(cfqq)) {
1762 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001763 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001764 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001765 }
Jens Axboe5e705372006-07-13 12:39:25 +02001766 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001767 /*
1768 * not the active queue - expire current slice if it is
1769 * idle and has expired it's mean thinktime or this new queue
1770 * has some old slice time left and is of higher priority
1771 */
1772 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001773 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001774 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001775 }
1776}
1777
Jens Axboeb4878f22005-10-20 16:42:29 +02001778static void cfq_insert_request(request_queue_t *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001779{
Jens Axboeb4878f22005-10-20 16:42:29 +02001780 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001781 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001782
1783 cfq_init_prio_data(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784
Jens Axboe5e705372006-07-13 12:39:25 +02001785 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786
Jens Axboe22e2c502005-06-27 10:55:12 +02001787 list_add_tail(&rq->queuelist, &cfqq->fifo);
1788
Jens Axboe5e705372006-07-13 12:39:25 +02001789 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790}
1791
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792static void cfq_completed_request(request_queue_t *q, struct request *rq)
1793{
Jens Axboe5e705372006-07-13 12:39:25 +02001794 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001795 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001796 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001797 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798
Jens Axboeb4878f22005-10-20 16:42:29 +02001799 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800
Jens Axboeb4878f22005-10-20 16:42:29 +02001801 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02001802 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02001803 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02001804 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805
Jens Axboe3ed9a292007-04-23 08:33:33 +02001806 if (cfq_cfqq_sync(cfqq))
1807 cfqd->sync_flight--;
1808
Jens Axboeb4878f22005-10-20 16:42:29 +02001809 if (!cfq_class_idle(cfqq))
1810 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001811
Jens Axboecaaa5f92006-06-16 11:23:00 +02001812 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001813 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001814
1815 /*
1816 * If this is the active queue, check if it needs to be expired,
1817 * or if we want to idle in case it has no pending requests.
1818 */
1819 if (cfqd->active_queue == cfqq) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001820 if (cfq_cfqq_slice_new(cfqq)) {
1821 cfq_set_prio_slice(cfqd, cfqq);
1822 cfq_clear_cfqq_slice_new(cfqq);
1823 }
1824 if (cfq_slice_used(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02001825 cfq_slice_expired(cfqd, 1);
Jens Axboe6d048f52007-04-25 12:44:27 +02001826 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1827 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001828 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001829
1830 if (!cfqd->rq_in_driver)
1831 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832}
1833
Jens Axboe22e2c502005-06-27 10:55:12 +02001834/*
1835 * we temporarily boost lower priority queues if they are holding fs exclusive
1836 * resources. they are boosted to normal prio (CLASS_BE/4)
1837 */
1838static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839{
Jens Axboe22e2c502005-06-27 10:55:12 +02001840 if (has_fs_excl()) {
1841 /*
1842 * boost idle prio on transactions that would lock out other
1843 * users of the filesystem
1844 */
1845 if (cfq_class_idle(cfqq))
1846 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1847 if (cfqq->ioprio > IOPRIO_NORM)
1848 cfqq->ioprio = IOPRIO_NORM;
1849 } else {
1850 /*
1851 * check if we need to unboost the queue
1852 */
1853 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1854 cfqq->ioprio_class = cfqq->org_ioprio_class;
1855 if (cfqq->ioprio != cfqq->org_ioprio)
1856 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001858}
1859
Jens Axboe89850f72006-07-22 16:48:31 +02001860static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001861{
Jens Axboe3b181522005-06-27 10:56:24 +02001862 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001863 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001864 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001865 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001866 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001867
1868 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02001869}
1870
Jens Axboecb78b282006-07-28 09:32:57 +02001871static int cfq_may_queue(request_queue_t *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02001872{
1873 struct cfq_data *cfqd = q->elevator->elevator_data;
1874 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001875 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02001876 struct cfq_queue *cfqq;
1877
1878 /*
1879 * don't force setup of a queue from here, as a call to may_queue
1880 * does not necessarily imply that a request actually will be queued.
1881 * so just lookup a possibly existing queue, or return 'may queue'
1882 * if that fails
1883 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001884 cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1885 if (!cic)
1886 return ELV_MQUEUE_MAY;
1887
1888 cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02001889 if (cfqq) {
1890 cfq_init_prio_data(cfqq);
1891 cfq_prio_boost(cfqq);
1892
Jens Axboe89850f72006-07-22 16:48:31 +02001893 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001894 }
1895
1896 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897}
1898
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899/*
1900 * queue lock held here
1901 */
Jens Axboebb37b942006-12-01 10:42:33 +01001902static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903{
Jens Axboe5e705372006-07-13 12:39:25 +02001904 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905
Jens Axboe5e705372006-07-13 12:39:25 +02001906 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001907 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908
Jens Axboe22e2c502005-06-27 10:55:12 +02001909 BUG_ON(!cfqq->allocated[rw]);
1910 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911
Jens Axboe5e705372006-07-13 12:39:25 +02001912 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02001915 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917 cfq_put_queue(cfqq);
1918 }
1919}
1920
1921/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001922 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001924static int
Jens Axboecb78b282006-07-28 09:32:57 +02001925cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926{
1927 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe3b181522005-06-27 10:56:24 +02001928 struct task_struct *tsk = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 struct cfq_io_context *cic;
1930 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01001931 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001932 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 unsigned long flags;
1934
1935 might_sleep_if(gfp_mask & __GFP_WAIT);
1936
Jens Axboee2d74ac2006-03-28 08:59:01 +02001937 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001938
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 spin_lock_irqsave(q->queue_lock, flags);
1940
Jens Axboe22e2c502005-06-27 10:55:12 +02001941 if (!cic)
1942 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943
Vasily Tarasov91fac312007-04-25 12:29:51 +02001944 cfqq = cic_to_cfqq(cic, is_sync);
1945 if (!cfqq) {
1946 cfqq = cfq_get_queue(cfqd, is_sync, tsk, gfp_mask);
1947
Jens Axboe22e2c502005-06-27 10:55:12 +02001948 if (!cfqq)
1949 goto queue_fail;
1950
Vasily Tarasov91fac312007-04-25 12:29:51 +02001951 cic_set_cfqq(cic, cfqq, is_sync);
1952 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953
1954 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02001955 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001956 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02001957
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 spin_unlock_irqrestore(q->queue_lock, flags);
1959
Jens Axboe5e705372006-07-13 12:39:25 +02001960 rq->elevator_private = cic;
1961 rq->elevator_private2 = cfqq;
1962 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02001963
Jens Axboe22e2c502005-06-27 10:55:12 +02001964queue_fail:
1965 if (cic)
1966 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02001967
Jens Axboe3b181522005-06-27 10:56:24 +02001968 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 spin_unlock_irqrestore(q->queue_lock, flags);
1970 return 1;
1971}
1972
David Howells65f27f32006-11-22 14:55:48 +00001973static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02001974{
David Howells65f27f32006-11-22 14:55:48 +00001975 struct cfq_data *cfqd =
1976 container_of(work, struct cfq_data, unplug_work);
1977 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001978 unsigned long flags;
1979
1980 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02001981 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02001982 spin_unlock_irqrestore(q->queue_lock, flags);
1983}
1984
1985/*
1986 * Timer running if the active_queue is currently idling inside its time slice
1987 */
1988static void cfq_idle_slice_timer(unsigned long data)
1989{
1990 struct cfq_data *cfqd = (struct cfq_data *) data;
1991 struct cfq_queue *cfqq;
1992 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001993 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001994
1995 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1996
1997 if ((cfqq = cfqd->active_queue) != NULL) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11001998 timed_out = 0;
1999
Jens Axboe22e2c502005-06-27 10:55:12 +02002000 /*
2001 * expired
2002 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002003 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002004 goto expire;
2005
2006 /*
2007 * only expire and reinvoke request handler, if there are
2008 * other queues with pending requests
2009 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002010 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002011 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002012
2013 /*
2014 * not expired and it has a request pending, let it dispatch
2015 */
Jens Axboedd67d052006-06-21 09:36:18 +02002016 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002017 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002018 goto out_kick;
2019 }
2020 }
2021expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002022 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002023out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002024 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002025out_cont:
2026 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2027}
2028
2029/*
2030 * Timer running if an idle class queue is waiting for service
2031 */
2032static void cfq_idle_class_timer(unsigned long data)
2033{
2034 struct cfq_data *cfqd = (struct cfq_data *) data;
2035 unsigned long flags, end;
2036
2037 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2038
2039 /*
2040 * race with a non-idle queue, reset timer
2041 */
2042 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
Jens Axboeae818a32006-06-01 10:13:43 +02002043 if (!time_after_eq(jiffies, end))
2044 mod_timer(&cfqd->idle_class_timer, end);
2045 else
Jens Axboe3b181522005-06-27 10:56:24 +02002046 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002047
2048 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2049}
2050
Jens Axboe3b181522005-06-27 10:56:24 +02002051static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2052{
2053 del_timer_sync(&cfqd->idle_slice_timer);
2054 del_timer_sync(&cfqd->idle_class_timer);
2055 blk_sync_queue(cfqd->queue);
2056}
Jens Axboe22e2c502005-06-27 10:55:12 +02002057
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002058static void cfq_put_async_queues(struct cfq_data *cfqd)
2059{
2060 int i;
2061
2062 for (i = 0; i < IOPRIO_BE_NR; i++) {
2063 if (cfqd->async_cfqq[0][i])
2064 cfq_put_queue(cfqd->async_cfqq[0][i]);
2065 if (cfqd->async_cfqq[1][i])
2066 cfq_put_queue(cfqd->async_cfqq[1][i]);
2067 if (cfqd->async_idle_cfqq)
2068 cfq_put_queue(cfqd->async_idle_cfqq);
2069 }
2070}
2071
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072static void cfq_exit_queue(elevator_t *e)
2073{
Jens Axboe22e2c502005-06-27 10:55:12 +02002074 struct cfq_data *cfqd = e->elevator_data;
Al Virod9ff4182006-03-18 13:51:22 -05002075 request_queue_t *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002076
Jens Axboe3b181522005-06-27 10:56:24 +02002077 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002078
Al Virod9ff4182006-03-18 13:51:22 -05002079 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002080
Al Virod9ff4182006-03-18 13:51:22 -05002081 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002082 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002083
2084 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002085 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2086 struct cfq_io_context,
2087 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002088
2089 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002090 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002091
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002092 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002093
Al Virod9ff4182006-03-18 13:51:22 -05002094 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002095
2096 cfq_shutdown_timer_wq(cfqd);
2097
Al Viroa90d7422006-03-18 12:05:37 -05002098 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099}
2100
Jens Axboebb37b942006-12-01 10:42:33 +01002101static void *cfq_init_queue(request_queue_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102{
2103 struct cfq_data *cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104
Christoph Lameter94f60302007-07-17 04:03:29 -07002105 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002107 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108
Jens Axboecc09e292007-04-26 12:53:50 +02002109 cfqd->service_tree = CFQ_RB_ROOT;
Al Virod9ff4182006-03-18 13:51:22 -05002110 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113
Jens Axboe22e2c502005-06-27 10:55:12 +02002114 init_timer(&cfqd->idle_slice_timer);
2115 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2116 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2117
2118 init_timer(&cfqd->idle_class_timer);
2119 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2120 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2121
David Howells65f27f32006-11-22 14:55:48 +00002122 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002123
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002125 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2126 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127 cfqd->cfq_back_max = cfq_back_max;
2128 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002129 cfqd->cfq_slice[0] = cfq_slice_async;
2130 cfqd->cfq_slice[1] = cfq_slice_sync;
2131 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2132 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02002133
Jens Axboebc1c1162006-06-08 08:49:06 +02002134 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135}
2136
2137static void cfq_slab_kill(void)
2138{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 if (cfq_pool)
2140 kmem_cache_destroy(cfq_pool);
2141 if (cfq_ioc_pool)
2142 kmem_cache_destroy(cfq_ioc_pool);
2143}
2144
2145static int __init cfq_slab_setup(void)
2146{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002147 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 if (!cfq_pool)
2149 goto fail;
2150
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002151 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 if (!cfq_ioc_pool)
2153 goto fail;
2154
2155 return 0;
2156fail:
2157 cfq_slab_kill();
2158 return -ENOMEM;
2159}
2160
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161/*
2162 * sysfs parts below -->
2163 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164static ssize_t
2165cfq_var_show(unsigned int var, char *page)
2166{
2167 return sprintf(page, "%d\n", var);
2168}
2169
2170static ssize_t
2171cfq_var_store(unsigned int *var, const char *page, size_t count)
2172{
2173 char *p = (char *) page;
2174
2175 *var = simple_strtoul(p, &p, 10);
2176 return count;
2177}
2178
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002180static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002182 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 unsigned int __data = __VAR; \
2184 if (__CONV) \
2185 __data = jiffies_to_msecs(__data); \
2186 return cfq_var_show(__data, (page)); \
2187}
2188SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002189SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2190SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002191SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2192SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002193SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2194SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2195SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2196SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197#undef SHOW_FUNCTION
2198
2199#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002200static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002202 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 unsigned int __data; \
2204 int ret = cfq_var_store(&__data, (page), count); \
2205 if (__data < (MIN)) \
2206 __data = (MIN); \
2207 else if (__data > (MAX)) \
2208 __data = (MAX); \
2209 if (__CONV) \
2210 *(__PTR) = msecs_to_jiffies(__data); \
2211 else \
2212 *(__PTR) = __data; \
2213 return ret; \
2214}
2215STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002216STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2217STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002218STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2219STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002220STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2221STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2222STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2223STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224#undef STORE_FUNCTION
2225
Al Viroe572ec72006-03-18 22:27:18 -05002226#define CFQ_ATTR(name) \
2227 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002228
Al Viroe572ec72006-03-18 22:27:18 -05002229static struct elv_fs_entry cfq_attrs[] = {
2230 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002231 CFQ_ATTR(fifo_expire_sync),
2232 CFQ_ATTR(fifo_expire_async),
2233 CFQ_ATTR(back_seek_max),
2234 CFQ_ATTR(back_seek_penalty),
2235 CFQ_ATTR(slice_sync),
2236 CFQ_ATTR(slice_async),
2237 CFQ_ATTR(slice_async_rq),
2238 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002239 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240};
2241
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242static struct elevator_type iosched_cfq = {
2243 .ops = {
2244 .elevator_merge_fn = cfq_merge,
2245 .elevator_merged_fn = cfq_merged_request,
2246 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002247 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002248 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002250 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251 .elevator_deactivate_req_fn = cfq_deactivate_request,
2252 .elevator_queue_empty_fn = cfq_queue_empty,
2253 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002254 .elevator_former_req_fn = elv_rb_former_request,
2255 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 .elevator_set_req_fn = cfq_set_request,
2257 .elevator_put_req_fn = cfq_put_request,
2258 .elevator_may_queue_fn = cfq_may_queue,
2259 .elevator_init_fn = cfq_init_queue,
2260 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002261 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 },
Al Viro3d1ab402006-03-18 18:35:43 -05002263 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264 .elevator_name = "cfq",
2265 .elevator_owner = THIS_MODULE,
2266};
2267
2268static int __init cfq_init(void)
2269{
2270 int ret;
2271
Jens Axboe22e2c502005-06-27 10:55:12 +02002272 /*
2273 * could be 0 on HZ < 1000 setups
2274 */
2275 if (!cfq_slice_async)
2276 cfq_slice_async = 1;
2277 if (!cfq_slice_idle)
2278 cfq_slice_idle = 1;
2279
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 if (cfq_slab_setup())
2281 return -ENOMEM;
2282
2283 ret = elv_register(&iosched_cfq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002284 if (ret)
2285 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 return ret;
2288}
2289
2290static void __exit cfq_exit(void)
2291{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002292 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002294 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002295 /* ioc_gone's update must be visible before reading ioc_count */
2296 smp_wmb();
Jens Axboe4050cf12006-07-19 05:07:12 +02002297 if (elv_ioc_count_read(ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002298 wait_for_completion(ioc_gone);
Al Viro334e94d2006-03-18 15:05:53 -05002299 synchronize_rcu();
Christoph Hellwig83521d32005-10-30 15:01:39 -08002300 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301}
2302
2303module_init(cfq_init);
2304module_exit(cfq_exit);
2305
2306MODULE_AUTHOR("Jens Axboe");
2307MODULE_LICENSE("GPL");
2308MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");