blob: f28d1fb30608cd5667b6b6f9c67e545bdb4802de [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/*
Jens Axboe08717142008-01-28 11:38:15 +010029 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020030 */
Jens Axboe08717142008-01-28 11:38:15 +010031#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020032
33/*
34 * below this threshold, we consider thinktime immediate
35 */
36#define CFQ_MIN_TT (2)
37
Jens Axboe22e2c502005-06-27 10:55:12 +020038#define CFQ_SLICE_SCALE (5)
39
Jens Axboe5e705372006-07-13 12:39:25 +020040#define RQ_CIC(rq) ((struct cfq_io_context*)(rq)->elevator_private)
41#define RQ_CFQQ(rq) ((rq)->elevator_private2)
Linus Torvalds1da177e2005-04-16 15:20:36 -070042
Christoph Lametere18b8902006-12-06 20:33:20 -080043static struct kmem_cache *cfq_pool;
44static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
Jens Axboe4050cf12006-07-19 05:07:12 +020046static DEFINE_PER_CPU(unsigned long, ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050047static struct completion *ioc_gone;
48
Jens Axboe22e2c502005-06-27 10:55:12 +020049#define CFQ_PRIO_LISTS IOPRIO_BE_NR
50#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020051#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
52
Jens Axboe3b181522005-06-27 10:56:24 +020053#define ASYNC (0)
54#define SYNC (1)
55
Jens Axboe206dc692006-03-28 13:03:44 +020056#define sample_valid(samples) ((samples) > 80)
57
Jens Axboe22e2c502005-06-27 10:55:12 +020058/*
Jens Axboecc09e292007-04-26 12:53:50 +020059 * Most of our rbtree usage is for sorting with min extraction, so
60 * if we cache the leftmost node we don't have to walk down the tree
61 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
62 * move this into the elevator for the rq sorting as well.
63 */
64struct cfq_rb_root {
65 struct rb_root rb;
66 struct rb_node *left;
67};
68#define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
69
70/*
Jens Axboe22e2c502005-06-27 10:55:12 +020071 * Per block device queue structure
72 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070073struct cfq_data {
Jens Axboe165125e2007-07-24 09:28:11 +020074 struct request_queue *queue;
Jens Axboe22e2c502005-06-27 10:55:12 +020075
76 /*
77 * rr list of queues with requests and the count of them
78 */
Jens Axboecc09e292007-04-26 12:53:50 +020079 struct cfq_rb_root service_tree;
Jens Axboe22e2c502005-06-27 10:55:12 +020080 unsigned int busy_queues;
81
Jens Axboe22e2c502005-06-27 10:55:12 +020082 int rq_in_driver;
Jens Axboe3ed9a292007-04-23 08:33:33 +020083 int sync_flight;
Jens Axboe25776e32006-06-01 10:12:26 +020084 int hw_tag;
Jens Axboe22e2c502005-06-27 10:55:12 +020085
86 /*
Jens Axboe22e2c502005-06-27 10:55:12 +020087 * idle window management
88 */
89 struct timer_list idle_slice_timer;
90 struct work_struct unplug_work;
91
92 struct cfq_queue *active_queue;
93 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +020094
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +020095 /*
96 * async queue for each priority case
97 */
98 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
99 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200100
Jens Axboe6d048f52007-04-25 12:44:27 +0200101 sector_t last_position;
Jens Axboe22e2c502005-06-27 10:55:12 +0200102 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104 /*
105 * tunables, see top of file
106 */
107 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200108 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109 unsigned int cfq_back_penalty;
110 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200111 unsigned int cfq_slice[2];
112 unsigned int cfq_slice_async_rq;
113 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500114
115 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116};
117
Jens Axboe22e2c502005-06-27 10:55:12 +0200118/*
119 * Per process-grouping structure
120 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121struct cfq_queue {
122 /* reference count */
123 atomic_t ref;
124 /* parent cfq_data */
125 struct cfq_data *cfqd;
Jens Axboed9e76202007-04-20 14:27:50 +0200126 /* service_tree member */
127 struct rb_node rb_node;
128 /* service_tree key */
129 unsigned long rb_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130 /* sorted list of pending requests */
131 struct rb_root sort_list;
132 /* if fifo isn't expired, next request to serve */
Jens Axboe5e705372006-07-13 12:39:25 +0200133 struct request *next_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 /* requests queued in sort_list */
135 int queued[2];
136 /* currently allocated requests */
137 int allocated[2];
Jens Axboe374f84a2006-07-23 01:42:19 +0200138 /* pending metadata requests */
139 int meta_pending;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200141 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142
Jens Axboe22e2c502005-06-27 10:55:12 +0200143 unsigned long slice_end;
Jens Axboec5b680f2007-01-19 11:56:49 +1100144 long slice_resid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
Jens Axboe6d048f52007-04-25 12:44:27 +0200146 /* number of requests that are on the dispatch list or inside driver */
147 int dispatched;
Jens Axboe22e2c502005-06-27 10:55:12 +0200148
149 /* io prio of this group */
150 unsigned short ioprio, org_ioprio;
151 unsigned short ioprio_class, org_ioprio_class;
152
Jens Axboe3b181522005-06-27 10:56:24 +0200153 /* various state flags, see below */
154 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155};
156
Jens Axboe3b181522005-06-27 10:56:24 +0200157enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100158 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
159 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
160 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
161 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
162 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
163 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
164 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
165 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
166 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
Jens Axboe44f7c162007-01-19 11:51:58 +1100167 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200168 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200169};
170
171#define CFQ_CFQQ_FNS(name) \
172static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
173{ \
174 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
175} \
176static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
177{ \
178 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
179} \
180static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
181{ \
182 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
183}
184
185CFQ_CFQQ_FNS(on_rr);
186CFQ_CFQQ_FNS(wait_request);
187CFQ_CFQQ_FNS(must_alloc);
188CFQ_CFQQ_FNS(must_alloc_slice);
189CFQ_CFQQ_FNS(must_dispatch);
190CFQ_CFQQ_FNS(fifo_expire);
191CFQ_CFQQ_FNS(idle_window);
192CFQ_CFQQ_FNS(prio_changed);
Jens Axboe53b03742006-07-28 09:48:51 +0200193CFQ_CFQQ_FNS(queue_new);
Jens Axboe44f7c162007-01-19 11:51:58 +1100194CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200195CFQ_CFQQ_FNS(sync);
Jens Axboe3b181522005-06-27 10:56:24 +0200196#undef CFQ_CFQQ_FNS
197
Jens Axboe165125e2007-07-24 09:28:11 +0200198static void cfq_dispatch_insert(struct request_queue *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200199static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
Jens Axboefd0928d2008-01-24 08:52:45 +0100200 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100201static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200202 struct io_context *);
203
204static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
205 int is_sync)
206{
207 return cic->cfqq[!!is_sync];
208}
209
210static inline void cic_set_cfqq(struct cfq_io_context *cic,
211 struct cfq_queue *cfqq, int is_sync)
212{
213 cic->cfqq[!!is_sync] = cfqq;
214}
215
216/*
217 * We regard a request as SYNC, if it's either a read or has the SYNC bit
218 * set (in which case it could also be direct WRITE).
219 */
220static inline int cfq_bio_sync(struct bio *bio)
221{
222 if (bio_data_dir(bio) == READ || bio_sync(bio))
223 return 1;
224
225 return 0;
226}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700229 * scheduler run of queue, if there are requests pending and no one in the
230 * driver that will restart queueing
231 */
232static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
233{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100234 if (cfqd->busy_queues)
Andrew Morton99f95e52005-06-27 20:14:05 -0700235 kblockd_schedule_work(&cfqd->unplug_work);
236}
237
Jens Axboe165125e2007-07-24 09:28:11 +0200238static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700239{
240 struct cfq_data *cfqd = q->elevator->elevator_data;
241
Jens Axboeb4878f22005-10-20 16:42:29 +0200242 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700243}
244
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100246 * Scale schedule slice based on io priority. Use the sync time slice only
247 * if a queue is marked sync and has sync io queued. A sync queue with async
248 * io only, should not get full sync slice length.
249 */
Jens Axboed9e76202007-04-20 14:27:50 +0200250static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
251 unsigned short prio)
252{
253 const int base_slice = cfqd->cfq_slice[sync];
254
255 WARN_ON(prio >= IOPRIO_BE_NR);
256
257 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
258}
259
Jens Axboe44f7c162007-01-19 11:51:58 +1100260static inline int
261cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
262{
Jens Axboed9e76202007-04-20 14:27:50 +0200263 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100264}
265
266static inline void
267cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
268{
269 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
270}
271
272/*
273 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
274 * isn't valid until the first request from the dispatch is activated
275 * and the slice time set.
276 */
277static inline int cfq_slice_used(struct cfq_queue *cfqq)
278{
279 if (cfq_cfqq_slice_new(cfqq))
280 return 0;
281 if (time_before(jiffies, cfqq->slice_end))
282 return 0;
283
284 return 1;
285}
286
287/*
Jens Axboe5e705372006-07-13 12:39:25 +0200288 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200290 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 */
Jens Axboe5e705372006-07-13 12:39:25 +0200292static struct request *
293cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294{
295 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200297#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
298#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
299 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300
Jens Axboe5e705372006-07-13 12:39:25 +0200301 if (rq1 == NULL || rq1 == rq2)
302 return rq2;
303 if (rq2 == NULL)
304 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200305
Jens Axboe5e705372006-07-13 12:39:25 +0200306 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
307 return rq1;
308 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
309 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200310 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
311 return rq1;
312 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
313 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314
Jens Axboe5e705372006-07-13 12:39:25 +0200315 s1 = rq1->sector;
316 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317
Jens Axboe6d048f52007-04-25 12:44:27 +0200318 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 /*
321 * by definition, 1KiB is 2 sectors
322 */
323 back_max = cfqd->cfq_back_max * 2;
324
325 /*
326 * Strict one way elevator _except_ in the case where we allow
327 * short backward seeks which are biased as twice the cost of a
328 * similar forward seek.
329 */
330 if (s1 >= last)
331 d1 = s1 - last;
332 else if (s1 + back_max >= last)
333 d1 = (last - s1) * cfqd->cfq_back_penalty;
334 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200335 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336
337 if (s2 >= last)
338 d2 = s2 - last;
339 else if (s2 + back_max >= last)
340 d2 = (last - s2) * cfqd->cfq_back_penalty;
341 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200342 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343
344 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345
Andreas Mohre8a99052006-03-28 08:59:49 +0200346 /*
347 * By doing switch() on the bit mask "wrap" we avoid having to
348 * check two variables for all permutations: --> faster!
349 */
350 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200351 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200352 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200353 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200354 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200355 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200356 else {
357 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200358 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200359 else
Jens Axboe5e705372006-07-13 12:39:25 +0200360 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200361 }
362
363 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200364 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200365 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200366 return rq2;
367 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200368 default:
369 /*
370 * Since both rqs are wrapped,
371 * start with the one that's further behind head
372 * (--> only *one* back seek required),
373 * since back seek takes more time than forward.
374 */
375 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200376 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 else
Jens Axboe5e705372006-07-13 12:39:25 +0200378 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 }
380}
381
Jens Axboe498d3aa2007-04-26 12:54:48 +0200382/*
383 * The below is leftmost cache rbtree addon
384 */
Jens Axboe08717142008-01-28 11:38:15 +0100385static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200386{
387 if (!root->left)
388 root->left = rb_first(&root->rb);
389
Jens Axboe08717142008-01-28 11:38:15 +0100390 if (root->left)
391 return rb_entry(root->left, struct cfq_queue, rb_node);
392
393 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200394}
395
396static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
397{
398 if (root->left == n)
399 root->left = NULL;
400
401 rb_erase(n, &root->rb);
402 RB_CLEAR_NODE(n);
403}
404
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405/*
406 * would be nice to take fifo expire time into account as well
407 */
Jens Axboe5e705372006-07-13 12:39:25 +0200408static struct request *
409cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
410 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411{
Jens Axboe21183b02006-07-13 12:33:14 +0200412 struct rb_node *rbnext = rb_next(&last->rb_node);
413 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200414 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415
Jens Axboe21183b02006-07-13 12:33:14 +0200416 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417
418 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200419 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200420
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200422 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200423 else {
424 rbnext = rb_first(&cfqq->sort_list);
425 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200426 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200427 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428
Jens Axboe21183b02006-07-13 12:33:14 +0200429 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430}
431
Jens Axboed9e76202007-04-20 14:27:50 +0200432static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
433 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434{
Jens Axboed9e76202007-04-20 14:27:50 +0200435 /*
436 * just an approximation, should be ok.
437 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200438 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
439 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200440}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441
Jens Axboe498d3aa2007-04-26 12:54:48 +0200442/*
443 * The cfqd->service_tree holds all pending cfq_queue's that have
444 * requests waiting to be processed. It is sorted in the order that
445 * we will service the queues.
446 */
Jens Axboed9e76202007-04-20 14:27:50 +0200447static void cfq_service_tree_add(struct cfq_data *cfqd,
Jens Axboeedd75ff2007-04-19 12:03:34 +0200448 struct cfq_queue *cfqq, int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200449{
Jens Axboe08717142008-01-28 11:38:15 +0100450 struct rb_node **p, *parent;
451 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +0200452 unsigned long rb_key;
Jens Axboe498d3aa2007-04-26 12:54:48 +0200453 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200454
Jens Axboe08717142008-01-28 11:38:15 +0100455 if (cfq_class_idle(cfqq)) {
456 rb_key = CFQ_IDLE_DELAY;
457 parent = rb_last(&cfqd->service_tree.rb);
458 if (parent && parent != &cfqq->rb_node) {
459 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
460 rb_key += __cfqq->rb_key;
461 } else
462 rb_key += jiffies;
463 } else if (!add_front) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200464 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
465 rb_key += cfqq->slice_resid;
466 cfqq->slice_resid = 0;
467 } else
468 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200469
470 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100471 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200472 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100473 */
Jens Axboed9e76202007-04-20 14:27:50 +0200474 if (rb_key == cfqq->rb_key)
475 return;
Jens Axboe53b03742006-07-28 09:48:51 +0200476
Jens Axboecc09e292007-04-26 12:53:50 +0200477 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200478 }
Jens Axboed9e76202007-04-20 14:27:50 +0200479
Jens Axboe498d3aa2007-04-26 12:54:48 +0200480 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +0100481 parent = NULL;
482 p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200483 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +0200484 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200485
Jens Axboed9e76202007-04-20 14:27:50 +0200486 parent = *p;
487 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
488
Jens Axboe0c534e02007-04-18 20:01:57 +0200489 /*
490 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200491 * preference to them. IDLE queues goes to the back.
492 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200493 */
494 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200495 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200496 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200497 n = &(*p)->rb_right;
498 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
499 n = &(*p)->rb_left;
500 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
501 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200502 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200503 n = &(*p)->rb_left;
504 else
505 n = &(*p)->rb_right;
506
507 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200508 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200509
510 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200511 }
512
Jens Axboecc09e292007-04-26 12:53:50 +0200513 if (left)
514 cfqd->service_tree.left = &cfqq->rb_node;
515
Jens Axboed9e76202007-04-20 14:27:50 +0200516 cfqq->rb_key = rb_key;
517 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200518 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519}
520
Jens Axboe498d3aa2007-04-26 12:54:48 +0200521/*
522 * Update cfqq's position in the service tree.
523 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200524static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200525{
Jens Axboe6d048f52007-04-25 12:44:27 +0200526 /*
527 * Resorting requires the cfqq to be on the RR list already.
528 */
Jens Axboe498d3aa2007-04-26 12:54:48 +0200529 if (cfq_cfqq_on_rr(cfqq))
Jens Axboeedd75ff2007-04-19 12:03:34 +0200530 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboe6d048f52007-04-25 12:44:27 +0200531}
532
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533/*
534 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200535 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 */
Jens Axboefebffd62008-01-28 13:19:43 +0100537static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538{
Jens Axboe3b181522005-06-27 10:56:24 +0200539 BUG_ON(cfq_cfqq_on_rr(cfqq));
540 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541 cfqd->busy_queues++;
542
Jens Axboeedd75ff2007-04-19 12:03:34 +0200543 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544}
545
Jens Axboe498d3aa2007-04-26 12:54:48 +0200546/*
547 * Called when the cfqq no longer has requests pending, remove it from
548 * the service tree.
549 */
Jens Axboefebffd62008-01-28 13:19:43 +0100550static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551{
Jens Axboe3b181522005-06-27 10:56:24 +0200552 BUG_ON(!cfq_cfqq_on_rr(cfqq));
553 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554
Jens Axboecc09e292007-04-26 12:53:50 +0200555 if (!RB_EMPTY_NODE(&cfqq->rb_node))
556 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboed9e76202007-04-20 14:27:50 +0200557
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558 BUG_ON(!cfqd->busy_queues);
559 cfqd->busy_queues--;
560}
561
562/*
563 * rb tree support functions
564 */
Jens Axboefebffd62008-01-28 13:19:43 +0100565static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566{
Jens Axboe5e705372006-07-13 12:39:25 +0200567 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200568 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200569 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570
Jens Axboeb4878f22005-10-20 16:42:29 +0200571 BUG_ON(!cfqq->queued[sync]);
572 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573
Jens Axboe5e705372006-07-13 12:39:25 +0200574 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575
Jens Axboedd67d052006-06-21 09:36:18 +0200576 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200577 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578}
579
Jens Axboe5e705372006-07-13 12:39:25 +0200580static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581{
Jens Axboe5e705372006-07-13 12:39:25 +0200582 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe21183b02006-07-13 12:33:14 +0200584 struct request *__alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585
Jens Axboe5380a102006-07-13 12:37:56 +0200586 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587
588 /*
589 * looks a little odd, but the first insert might return an alias.
590 * if that happens, put the alias on the dispatch list
591 */
Jens Axboe21183b02006-07-13 12:33:14 +0200592 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200593 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100594
595 if (!cfq_cfqq_on_rr(cfqq))
596 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200597
598 /*
599 * check if this request is a better next-serve candidate
600 */
601 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
602 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603}
604
Jens Axboefebffd62008-01-28 13:19:43 +0100605static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606{
Jens Axboe5380a102006-07-13 12:37:56 +0200607 elv_rb_del(&cfqq->sort_list, rq);
608 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200609 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610}
611
Jens Axboe206dc692006-03-28 13:03:44 +0200612static struct request *
613cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614{
Jens Axboe206dc692006-03-28 13:03:44 +0200615 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200616 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200617 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618
Jens Axboe4ac845a2008-01-24 08:44:49 +0100619 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200620 if (!cic)
621 return NULL;
622
623 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200624 if (cfqq) {
625 sector_t sector = bio->bi_sector + bio_sectors(bio);
626
Jens Axboe21183b02006-07-13 12:33:14 +0200627 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200628 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 return NULL;
631}
632
Jens Axboe165125e2007-07-24 09:28:11 +0200633static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200634{
635 struct cfq_data *cfqd = q->elevator->elevator_data;
636
637 cfqd->rq_in_driver++;
Jens Axboe25776e32006-06-01 10:12:26 +0200638
639 /*
640 * If the depth is larger 1, it really could be queueing. But lets
641 * make the mark a little higher - idling could still be good for
642 * low queueing, and a low queueing number could also just indicate
643 * a SCSI mid layer like behaviour where limit+1 is often seen.
644 */
645 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
646 cfqd->hw_tag = 1;
Jens Axboe6d048f52007-04-25 12:44:27 +0200647
648 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
Jens Axboeb4878f22005-10-20 16:42:29 +0200649}
650
Jens Axboe165125e2007-07-24 09:28:11 +0200651static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652{
Jens Axboe22e2c502005-06-27 10:55:12 +0200653 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654
Jens Axboeb4878f22005-10-20 16:42:29 +0200655 WARN_ON(!cfqd->rq_in_driver);
656 cfqd->rq_in_driver--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657}
658
Jens Axboeb4878f22005-10-20 16:42:29 +0200659static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660{
Jens Axboe5e705372006-07-13 12:39:25 +0200661 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200662
Jens Axboe5e705372006-07-13 12:39:25 +0200663 if (cfqq->next_rq == rq)
664 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665
Jens Axboeb4878f22005-10-20 16:42:29 +0200666 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200667 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200668
669 if (rq_is_meta(rq)) {
670 WARN_ON(!cfqq->meta_pending);
671 cfqq->meta_pending--;
672 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673}
674
Jens Axboe165125e2007-07-24 09:28:11 +0200675static int cfq_merge(struct request_queue *q, struct request **req,
676 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677{
678 struct cfq_data *cfqd = q->elevator->elevator_data;
679 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680
Jens Axboe206dc692006-03-28 13:03:44 +0200681 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200682 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200683 *req = __rq;
684 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 }
686
687 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688}
689
Jens Axboe165125e2007-07-24 09:28:11 +0200690static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200691 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692{
Jens Axboe21183b02006-07-13 12:33:14 +0200693 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200694 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695
Jens Axboe5e705372006-07-13 12:39:25 +0200696 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698}
699
700static void
Jens Axboe165125e2007-07-24 09:28:11 +0200701cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702 struct request *next)
703{
Jens Axboe22e2c502005-06-27 10:55:12 +0200704 /*
705 * reposition in fifo if next is older than rq
706 */
707 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
708 time_before(next->start_time, rq->start_time))
709 list_move(&rq->queuelist, &next->queuelist);
710
Jens Axboeb4878f22005-10-20 16:42:29 +0200711 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200712}
713
Jens Axboe165125e2007-07-24 09:28:11 +0200714static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100715 struct bio *bio)
716{
717 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200718 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100719 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100720
721 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100722 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100723 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200724 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100725 return 0;
726
727 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100728 * Lookup the cfqq that this bio will be queued with. Allow
729 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100730 */
Jens Axboe4ac845a2008-01-24 08:44:49 +0100731 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200732 if (!cic)
733 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100734
Vasily Tarasov91fac312007-04-25 12:29:51 +0200735 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100736 if (cfqq == RQ_CFQQ(rq))
737 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100738
Jens Axboeec8acb62007-01-02 18:32:11 +0100739 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100740}
741
Jens Axboefebffd62008-01-28 13:19:43 +0100742static void __cfq_set_active_queue(struct cfq_data *cfqd,
743 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +0200744{
745 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +0200746 cfqq->slice_end = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200747 cfq_clear_cfqq_must_alloc_slice(cfqq);
748 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100749 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe1afba042007-04-17 12:47:55 +0200750 cfq_clear_cfqq_queue_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200751 }
752
753 cfqd->active_queue = cfqq;
754}
755
756/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100757 * current cfqq expired its slice (or was too idle), select new one
758 */
759static void
760__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200761 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100762{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100763 if (cfq_cfqq_wait_request(cfqq))
764 del_timer(&cfqd->idle_slice_timer);
765
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100766 cfq_clear_cfqq_must_dispatch(cfqq);
767 cfq_clear_cfqq_wait_request(cfqq);
768
769 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200770 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100771 */
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100772 if (timed_out && !cfq_cfqq_slice_new(cfqq))
Jens Axboec5b680f2007-01-19 11:56:49 +1100773 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100774
Jens Axboeedd75ff2007-04-19 12:03:34 +0200775 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100776
777 if (cfqq == cfqd->active_queue)
778 cfqd->active_queue = NULL;
779
780 if (cfqd->active_cic) {
781 put_io_context(cfqd->active_cic->ioc);
782 cfqd->active_cic = NULL;
783 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100784}
785
Jens Axboe6084cdd2007-04-23 08:25:00 +0200786static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100787{
788 struct cfq_queue *cfqq = cfqd->active_queue;
789
790 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200791 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100792}
793
Jens Axboe498d3aa2007-04-26 12:54:48 +0200794/*
795 * Get next queue for service. Unless we have a queue preemption,
796 * we'll simply select the first cfqq in the service tree.
797 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200798static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
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
Jens Axboe08717142008-01-28 11:38:15 +0100803 return cfq_rb_first(&cfqd->service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +0200804}
805
Jens Axboe498d3aa2007-04-26 12:54:48 +0200806/*
807 * Get and set a new active queue for service.
808 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200809static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
810{
811 struct cfq_queue *cfqq;
812
Jens Axboed9e76202007-04-20 14:27:50 +0200813 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200814 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200815 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200816}
817
Jens Axboed9e76202007-04-20 14:27:50 +0200818static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
819 struct request *rq)
820{
821 if (rq->sector >= cfqd->last_position)
822 return rq->sector - cfqd->last_position;
823 else
824 return cfqd->last_position - rq->sector;
825}
826
Jens Axboe6d048f52007-04-25 12:44:27 +0200827static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
828{
829 struct cfq_io_context *cic = cfqd->active_cic;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200830
Jens Axboe6d048f52007-04-25 12:44:27 +0200831 if (!sample_valid(cic->seek_samples))
832 return 0;
833
834 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
835}
836
Jens Axboed9e76202007-04-20 14:27:50 +0200837static int cfq_close_cooperator(struct cfq_data *cfq_data,
838 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200839{
Jens Axboe6d048f52007-04-25 12:44:27 +0200840 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200841 * We should notice if some of the queues are cooperating, eg
842 * working closely on the same area of the disk. In that case,
843 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +0200844 */
Jens Axboed9e76202007-04-20 14:27:50 +0200845 return 0;
Jens Axboe6d048f52007-04-25 12:44:27 +0200846}
847
848#define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
849
850static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200851{
Jens Axboe17926692007-01-19 11:59:30 +1100852 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +0200853 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100854 unsigned long sl;
855
Jens Axboedd67d052006-06-21 09:36:18 +0200856 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +0200857 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +0200858
859 /*
860 * idle is disabled, either manually or by past process history
861 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200862 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
863 return;
864
Jens Axboe22e2c502005-06-27 10:55:12 +0200865 /*
866 * task has exited, don't wait
867 */
Jens Axboe206dc692006-03-28 13:03:44 +0200868 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +0100869 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +0200870 return;
871
872 /*
873 * See if this prio level has a good candidate
874 */
Jens Axboe1afba042007-04-17 12:47:55 +0200875 if (cfq_close_cooperator(cfqd, cfqq) &&
876 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
Jens Axboe6d048f52007-04-25 12:44:27 +0200877 return;
Jens Axboe22e2c502005-06-27 10:55:12 +0200878
Jens Axboe3b181522005-06-27 10:56:24 +0200879 cfq_mark_cfqq_must_dispatch(cfqq);
880 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200881
Jens Axboe206dc692006-03-28 13:03:44 +0200882 /*
883 * we don't want to idle for seeks, but we do want to allow
884 * fair distribution of slice time for a process doing back-to-back
885 * seeks. so allow a little bit of time for him to submit a new rq
886 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200887 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200888 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +0200889 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +0200890
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100891 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892}
893
Jens Axboe498d3aa2007-04-26 12:54:48 +0200894/*
895 * Move request from internal lists to the request queue dispatch list.
896 */
Jens Axboe165125e2007-07-24 09:28:11 +0200897static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898{
Jens Axboe3ed9a292007-04-23 08:33:33 +0200899 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200900 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200901
Jens Axboe5380a102006-07-13 12:37:56 +0200902 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200903 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +0200904 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +0200905
906 if (cfq_cfqq_sync(cfqq))
907 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908}
909
910/*
911 * return expired entry, or NULL to just start from scratch in rbtree
912 */
Jens Axboefebffd62008-01-28 13:19:43 +0100913static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914{
915 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200916 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200917 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918
Jens Axboe3b181522005-06-27 10:56:24 +0200919 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +1100921
922 cfq_mark_cfqq_fifo_expire(cfqq);
923
Jens Axboe89850f72006-07-22 16:48:31 +0200924 if (list_empty(&cfqq->fifo))
925 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926
Jens Axboe6d048f52007-04-25 12:44:27 +0200927 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +0200928 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929
Jens Axboe6d048f52007-04-25 12:44:27 +0200930 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
931 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932
Jens Axboe6d048f52007-04-25 12:44:27 +0200933 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934}
935
Jens Axboe22e2c502005-06-27 10:55:12 +0200936static inline int
937cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
938{
939 const int base_rq = cfqd->cfq_slice_async_rq;
940
941 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
942
943 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
944}
945
946/*
Jens Axboe498d3aa2007-04-26 12:54:48 +0200947 * Select a queue for service. If we have a current active queue,
948 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +0200949 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100950static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200951{
Jens Axboe22e2c502005-06-27 10:55:12 +0200952 struct cfq_queue *cfqq;
953
954 cfqq = cfqd->active_queue;
955 if (!cfqq)
956 goto new_queue;
957
958 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200959 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +0200960 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200961 if (cfq_slice_used(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +0200962 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +0200963
964 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200965 * The active queue has requests and isn't expired, allow it to
966 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +0200967 */
Jens Axboedd67d052006-06-21 09:36:18 +0200968 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200969 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +0200970
971 /*
972 * No requests pending. If the active queue still has requests in
973 * flight or is idling for a new request, allow either of these
974 * conditions to happen (or time out) before selecting a new queue.
975 */
Jens Axboecc197472007-04-20 20:45:39 +0200976 if (timer_pending(&cfqd->idle_slice_timer) ||
977 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +0200978 cfqq = NULL;
979 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +0200980 }
981
Jens Axboe3b181522005-06-27 10:56:24 +0200982expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +0200983 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +0200984new_queue:
985 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200986keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +0200987 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200988}
989
Jens Axboe498d3aa2007-04-26 12:54:48 +0200990/*
991 * Dispatch some requests from cfqq, moving them to the request queue
992 * dispatch list.
993 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200994static int
995__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
996 int max_dispatch)
997{
998 int dispatched = 0;
999
Jens Axboedd67d052006-06-21 09:36:18 +02001000 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001001
1002 do {
Jens Axboe5e705372006-07-13 12:39:25 +02001003 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001004
1005 /*
1006 * follow expired path, else get first next available
1007 */
Jens Axboe5e705372006-07-13 12:39:25 +02001008 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1009 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001010
1011 /*
1012 * finally, insert request into driver dispatch list
1013 */
Jens Axboe5e705372006-07-13 12:39:25 +02001014 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001015
Jens Axboe22e2c502005-06-27 10:55:12 +02001016 dispatched++;
1017
1018 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +02001019 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1020 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001021 }
1022
Jens Axboedd67d052006-06-21 09:36:18 +02001023 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001024 break;
1025
1026 } while (dispatched < max_dispatch);
1027
1028 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001029 * expire an async queue immediately if it has used up its slice. idle
1030 * queue always expire after 1 dispatch round.
1031 */
Jens Axboea9938002007-04-20 08:55:52 +02001032 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
Jens Axboe20e493a2007-04-23 08:26:36 +02001033 dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboea9938002007-04-20 08:55:52 +02001034 cfq_class_idle(cfqq))) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001035 cfqq->slice_end = jiffies + 1;
Jens Axboe6084cdd2007-04-23 08:25:00 +02001036 cfq_slice_expired(cfqd, 0);
Jens Axboe44f7c162007-01-19 11:51:58 +11001037 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001038
1039 return dispatched;
1040}
1041
Jens Axboefebffd62008-01-28 13:19:43 +01001042static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02001043{
1044 int dispatched = 0;
1045
1046 while (cfqq->next_rq) {
1047 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1048 dispatched++;
1049 }
1050
1051 BUG_ON(!list_empty(&cfqq->fifo));
1052 return dispatched;
1053}
1054
Jens Axboe498d3aa2007-04-26 12:54:48 +02001055/*
1056 * Drain our current requests. Used for barriers and when switching
1057 * io schedulers on-the-fly.
1058 */
Jens Axboed9e76202007-04-20 14:27:50 +02001059static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001060{
Jens Axboe08717142008-01-28 11:38:15 +01001061 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001062 int dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001063
Jens Axboe08717142008-01-28 11:38:15 +01001064 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
Jens Axboed9e76202007-04-20 14:27:50 +02001065 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001066
Jens Axboe6084cdd2007-04-23 08:25:00 +02001067 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001068
1069 BUG_ON(cfqd->busy_queues);
1070
1071 return dispatched;
1072}
1073
Jens Axboe165125e2007-07-24 09:28:11 +02001074static int cfq_dispatch_requests(struct request_queue *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075{
1076 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001077 struct cfq_queue *cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001078 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079
Jens Axboe22e2c502005-06-27 10:55:12 +02001080 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081 return 0;
1082
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001083 if (unlikely(force))
1084 return cfq_forced_dispatch(cfqd);
1085
Jens Axboecaaa5f92006-06-16 11:23:00 +02001086 dispatched = 0;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001087 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001088 int max_dispatch;
1089
Jens Axboe3ed9a292007-04-23 08:33:33 +02001090 max_dispatch = cfqd->cfq_quantum;
1091 if (cfq_class_idle(cfqq))
1092 max_dispatch = 1;
1093
1094 if (cfqq->dispatched >= max_dispatch) {
1095 if (cfqd->busy_queues > 1)
Jens Axboe6d048f52007-04-25 12:44:27 +02001096 break;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001097 if (cfqq->dispatched >= 4 * max_dispatch)
Jens Axboea9938002007-04-20 08:55:52 +02001098 break;
1099 }
Jens Axboe9ede2092007-01-19 12:11:44 +11001100
Jens Axboe3ed9a292007-04-23 08:33:33 +02001101 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1102 break;
1103
Jens Axboe3b181522005-06-27 10:56:24 +02001104 cfq_clear_cfqq_must_dispatch(cfqq);
1105 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001106 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107
Jens Axboecaaa5f92006-06-16 11:23:00 +02001108 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 }
1110
Jens Axboecaaa5f92006-06-16 11:23:00 +02001111 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112}
1113
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114/*
Jens Axboe5e705372006-07-13 12:39:25 +02001115 * task holds one reference to the queue, dropped when task exits. each rq
1116 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 *
1118 * queue lock must be held here.
1119 */
1120static void cfq_put_queue(struct cfq_queue *cfqq)
1121{
Jens Axboe22e2c502005-06-27 10:55:12 +02001122 struct cfq_data *cfqd = cfqq->cfqd;
1123
1124 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125
1126 if (!atomic_dec_and_test(&cfqq->ref))
1127 return;
1128
1129 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001130 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001131 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001133 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001134 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001135 cfq_schedule_dispatch(cfqd);
1136 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001137
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 kmem_cache_free(cfq_pool, cfqq);
1139}
1140
Jens Axboe4ac845a2008-01-24 08:44:49 +01001141/*
1142 * Call func for each cic attached to this ioc. Returns number of cic's seen.
1143 */
1144#define CIC_GANG_NR 16
1145static unsigned int
1146call_for_each_cic(struct io_context *ioc,
1147 void (*func)(struct io_context *, struct cfq_io_context *))
1148{
1149 struct cfq_io_context *cics[CIC_GANG_NR];
1150 unsigned long index = 0;
1151 unsigned int called = 0;
1152 int nr;
1153
1154 rcu_read_lock();
1155
1156 do {
1157 int i;
1158
1159 /*
1160 * Perhaps there's a better way - this just gang lookups from
1161 * 0 to the end, restarting after each CIC_GANG_NR from the
1162 * last key + 1.
1163 */
1164 nr = radix_tree_gang_lookup(&ioc->radix_root, (void **) cics,
1165 index, CIC_GANG_NR);
1166 if (!nr)
1167 break;
1168
1169 called += nr;
1170 index = 1 + (unsigned long) cics[nr - 1]->key;
1171
1172 for (i = 0; i < nr; i++)
1173 func(ioc, cics[i]);
1174 } while (nr == CIC_GANG_NR);
1175
1176 rcu_read_unlock();
1177
1178 return called;
1179}
1180
1181static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1182{
1183 unsigned long flags;
1184
1185 BUG_ON(!cic->dead_key);
1186
1187 spin_lock_irqsave(&ioc->lock, flags);
1188 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1189 spin_unlock_irqrestore(&ioc->lock, flags);
1190
1191 kmem_cache_free(cfq_ioc_pool, cic);
1192}
1193
Jens Axboee2d74ac2006-03-28 08:59:01 +02001194static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001196 int freed;
Jens Axboe22e2c502005-06-27 10:55:12 +02001197
Jens Axboe4ac845a2008-01-24 08:44:49 +01001198 /*
1199 * ioc->refcount is zero here, so no more cic's are allowed to be
1200 * linked into this ioc. So it should be ok to iterate over the known
1201 * list, we will see all cic's since no new ones are added.
1202 */
1203 freed = call_for_each_cic(ioc, cic_free_func);
Jens Axboe22e2c502005-06-27 10:55:12 +02001204
Jens Axboe4050cf12006-07-19 05:07:12 +02001205 elv_ioc_count_mod(ioc_count, -freed);
1206
1207 if (ioc_gone && !elv_ioc_count_read(ioc_count))
Al Viro334e94d2006-03-18 15:05:53 -05001208 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209}
1210
Jens Axboe89850f72006-07-22 16:48:31 +02001211static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1212{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001213 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001214 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001215 cfq_schedule_dispatch(cfqd);
1216 }
Jens Axboe89850f72006-07-22 16:48:31 +02001217
1218 cfq_put_queue(cfqq);
1219}
1220
1221static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1222 struct cfq_io_context *cic)
1223{
Jens Axboefc463792006-08-29 09:05:44 +02001224 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001225
1226 /*
1227 * Make sure key == NULL is seen for dead queues
1228 */
Jens Axboefc463792006-08-29 09:05:44 +02001229 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001230 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02001231 cic->key = NULL;
1232
Jens Axboe89850f72006-07-22 16:48:31 +02001233 if (cic->cfqq[ASYNC]) {
1234 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1235 cic->cfqq[ASYNC] = NULL;
1236 }
1237
1238 if (cic->cfqq[SYNC]) {
1239 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1240 cic->cfqq[SYNC] = NULL;
1241 }
Jens Axboe89850f72006-07-22 16:48:31 +02001242}
1243
Jens Axboe4ac845a2008-01-24 08:44:49 +01001244static void cfq_exit_single_io_context(struct io_context *ioc,
1245 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001246{
Al Viro478a82b2006-03-18 13:25:24 -05001247 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001248
Jens Axboe89850f72006-07-22 16:48:31 +02001249 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001250 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001251 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02001252
Jens Axboe4ac845a2008-01-24 08:44:49 +01001253 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe89850f72006-07-22 16:48:31 +02001254 __cfq_exit_single_io_context(cfqd, cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001255 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05001256 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001257}
1258
Jens Axboe498d3aa2007-04-26 12:54:48 +02001259/*
1260 * The process that ioc belongs to has exited, we need to clean up
1261 * and put the internal structures we have that belongs to that process.
1262 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001263static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001265 rcu_assign_pointer(ioc->ioc_data, NULL);
1266 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267}
1268
Jens Axboe22e2c502005-06-27 10:55:12 +02001269static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001270cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271{
Jens Axboeb5deef92006-07-19 23:39:40 +02001272 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273
Christoph Lameter94f60302007-07-17 04:03:29 -07001274 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1275 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001277 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001278 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001279 cic->dtor = cfq_free_io_context;
1280 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001281 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 }
1283
1284 return cic;
1285}
1286
Jens Axboefd0928d2008-01-24 08:52:45 +01001287static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02001288{
1289 struct task_struct *tsk = current;
1290 int ioprio_class;
1291
Jens Axboe3b181522005-06-27 10:56:24 +02001292 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001293 return;
1294
Jens Axboefd0928d2008-01-24 08:52:45 +01001295 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001296 switch (ioprio_class) {
1297 default:
1298 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1299 case IOPRIO_CLASS_NONE:
1300 /*
1301 * no prio set, place us in the middle of the BE classes
1302 */
1303 cfqq->ioprio = task_nice_ioprio(tsk);
1304 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1305 break;
1306 case IOPRIO_CLASS_RT:
Jens Axboefd0928d2008-01-24 08:52:45 +01001307 cfqq->ioprio = task_ioprio(ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02001308 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1309 break;
1310 case IOPRIO_CLASS_BE:
Jens Axboefd0928d2008-01-24 08:52:45 +01001311 cfqq->ioprio = task_ioprio(ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02001312 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1313 break;
1314 case IOPRIO_CLASS_IDLE:
1315 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1316 cfqq->ioprio = 7;
Jens Axboe3b181522005-06-27 10:56:24 +02001317 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001318 break;
1319 }
1320
1321 /*
1322 * keep track of original prio settings in case we have to temporarily
1323 * elevate the priority of this queue
1324 */
1325 cfqq->org_ioprio = cfqq->ioprio;
1326 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001327 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001328}
1329
Jens Axboefebffd62008-01-28 13:19:43 +01001330static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001331{
Al Viro478a82b2006-03-18 13:25:24 -05001332 struct cfq_data *cfqd = cic->key;
1333 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001334 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001335
Jens Axboecaaa5f92006-06-16 11:23:00 +02001336 if (unlikely(!cfqd))
1337 return;
1338
Jens Axboec1b707d2006-10-30 19:54:23 +01001339 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001340
1341 cfqq = cic->cfqq[ASYNC];
1342 if (cfqq) {
1343 struct cfq_queue *new_cfqq;
Jens Axboefd0928d2008-01-24 08:52:45 +01001344 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001345 if (new_cfqq) {
1346 cic->cfqq[ASYNC] = new_cfqq;
1347 cfq_put_queue(cfqq);
1348 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001349 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001350
1351 cfqq = cic->cfqq[SYNC];
1352 if (cfqq)
1353 cfq_mark_cfqq_prio_changed(cfqq);
1354
Jens Axboec1b707d2006-10-30 19:54:23 +01001355 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001356}
1357
Jens Axboefc463792006-08-29 09:05:44 +02001358static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001360 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02001361 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362}
1363
1364static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001365cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01001366 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001369 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370
1371retry:
Jens Axboe4ac845a2008-01-24 08:44:49 +01001372 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001373 /* cic always exists here */
1374 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375
1376 if (!cfqq) {
1377 if (new_cfqq) {
1378 cfqq = new_cfqq;
1379 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001380 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001381 /*
1382 * Inform the allocator of the fact that we will
1383 * just repeat this allocation if it fails, to allow
1384 * the allocator to do whatever it needs to attempt to
1385 * free memory.
1386 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001388 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1389 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1390 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 spin_lock_irq(cfqd->queue->queue_lock);
1392 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001393 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001394 cfqq = kmem_cache_alloc_node(cfq_pool,
1395 gfp_mask | __GFP_ZERO,
1396 cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001397 if (!cfqq)
1398 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001399 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400
Jens Axboed9e76202007-04-20 14:27:50 +02001401 RB_CLEAR_NODE(&cfqq->rb_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001402 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 atomic_set(&cfqq->ref, 0);
1405 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001406
Jens Axboe3b181522005-06-27 10:56:24 +02001407 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b03742006-07-28 09:48:51 +02001408 cfq_mark_cfqq_queue_new(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001409
Jens Axboefd0928d2008-01-24 08:52:45 +01001410 cfq_init_prio_data(cfqq, ioc);
Jens Axboe08717142008-01-28 11:38:15 +01001411
1412 if (is_sync) {
1413 if (!cfq_class_idle(cfqq))
1414 cfq_mark_cfqq_idle_window(cfqq);
1415 cfq_mark_cfqq_sync(cfqq);
1416 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417 }
1418
1419 if (new_cfqq)
1420 kmem_cache_free(cfq_pool, new_cfqq);
1421
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422out:
1423 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1424 return cfqq;
1425}
1426
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001427static struct cfq_queue **
1428cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1429{
1430 switch(ioprio_class) {
1431 case IOPRIO_CLASS_RT:
1432 return &cfqd->async_cfqq[0][ioprio];
1433 case IOPRIO_CLASS_BE:
1434 return &cfqd->async_cfqq[1][ioprio];
1435 case IOPRIO_CLASS_IDLE:
1436 return &cfqd->async_idle_cfqq;
1437 default:
1438 BUG();
1439 }
1440}
1441
Jens Axboe15c31be2007-07-10 13:43:25 +02001442static struct cfq_queue *
Jens Axboefd0928d2008-01-24 08:52:45 +01001443cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02001444 gfp_t gfp_mask)
1445{
Jens Axboefd0928d2008-01-24 08:52:45 +01001446 const int ioprio = task_ioprio(ioc);
1447 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001448 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001449 struct cfq_queue *cfqq = NULL;
1450
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001451 if (!is_sync) {
1452 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1453 cfqq = *async_cfqq;
1454 }
1455
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001456 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01001457 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001458 if (!cfqq)
1459 return NULL;
1460 }
Jens Axboe15c31be2007-07-10 13:43:25 +02001461
1462 /*
1463 * pin the queue now that it's allocated, scheduler exit will prune it
1464 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001465 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001466 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001467 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001468 }
1469
1470 atomic_inc(&cfqq->ref);
1471 return cfqq;
1472}
1473
Jens Axboe4ac845a2008-01-24 08:44:49 +01001474static void cfq_cic_free(struct cfq_io_context *cic)
1475{
1476 kmem_cache_free(cfq_ioc_pool, cic);
1477 elv_ioc_count_dec(ioc_count);
1478
1479 if (ioc_gone && !elv_ioc_count_read(ioc_count))
1480 complete(ioc_gone);
1481}
1482
Jens Axboe498d3aa2007-04-26 12:54:48 +02001483/*
1484 * We drop cfq io contexts lazily, so we may find a dead one.
1485 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001486static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001487cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1488 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001489{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001490 unsigned long flags;
1491
Jens Axboefc463792006-08-29 09:05:44 +02001492 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001493
Jens Axboe4ac845a2008-01-24 08:44:49 +01001494 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02001495
Jens Axboe4ac845a2008-01-24 08:44:49 +01001496 if (ioc->ioc_data == cic)
1497 rcu_assign_pointer(ioc->ioc_data, NULL);
1498
1499 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1500 spin_unlock_irqrestore(&ioc->lock, flags);
1501
1502 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001503}
1504
Jens Axboee2d74ac2006-03-28 08:59:01 +02001505static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01001506cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001507{
Jens Axboee2d74ac2006-03-28 08:59:01 +02001508 struct cfq_io_context *cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001509 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001510
Vasily Tarasov91fac312007-04-25 12:29:51 +02001511 if (unlikely(!ioc))
1512 return NULL;
1513
Jens Axboe597bc482007-04-24 21:23:53 +02001514 /*
1515 * we maintain a last-hit cache, to avoid browsing over the tree
1516 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001517 cic = rcu_dereference(ioc->ioc_data);
Jens Axboe597bc482007-04-24 21:23:53 +02001518 if (cic && cic->key == cfqd)
1519 return cic;
1520
Jens Axboe4ac845a2008-01-24 08:44:49 +01001521 do {
1522 rcu_read_lock();
1523 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1524 rcu_read_unlock();
1525 if (!cic)
1526 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001527 /* ->key must be copied to avoid race with cfq_exit_queue() */
1528 k = cic->key;
1529 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001530 cfq_drop_dead_cic(cfqd, ioc, cic);
1531 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001532 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001533
Jens Axboe4ac845a2008-01-24 08:44:49 +01001534 rcu_assign_pointer(ioc->ioc_data, cic);
1535 break;
1536 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001537
Jens Axboe4ac845a2008-01-24 08:44:49 +01001538 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001539}
1540
Jens Axboe4ac845a2008-01-24 08:44:49 +01001541/*
1542 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1543 * the process specific cfq io context when entered from the block layer.
1544 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1545 */
Jens Axboefebffd62008-01-28 13:19:43 +01001546static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1547 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001548{
Jens Axboe0261d682006-10-30 19:07:48 +01001549 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001550 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001551
Jens Axboe4ac845a2008-01-24 08:44:49 +01001552 ret = radix_tree_preload(gfp_mask);
1553 if (!ret) {
1554 cic->ioc = ioc;
1555 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001556
Jens Axboe4ac845a2008-01-24 08:44:49 +01001557 spin_lock_irqsave(&ioc->lock, flags);
1558 ret = radix_tree_insert(&ioc->radix_root,
1559 (unsigned long) cfqd, cic);
1560 spin_unlock_irqrestore(&ioc->lock, flags);
1561
1562 radix_tree_preload_end();
1563
1564 if (!ret) {
1565 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1566 list_add(&cic->queue_list, &cfqd->cic_list);
1567 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001568 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001569 }
1570
Jens Axboe4ac845a2008-01-24 08:44:49 +01001571 if (ret)
1572 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02001573
Jens Axboe4ac845a2008-01-24 08:44:49 +01001574 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001575}
1576
Jens Axboe22e2c502005-06-27 10:55:12 +02001577/*
1578 * Setup general io context and cfq io context. There can be several cfq
1579 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001580 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001581 */
1582static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001583cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584{
Jens Axboe22e2c502005-06-27 10:55:12 +02001585 struct io_context *ioc = NULL;
1586 struct cfq_io_context *cic;
1587
1588 might_sleep_if(gfp_mask & __GFP_WAIT);
1589
Jens Axboeb5deef92006-07-19 23:39:40 +02001590 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001591 if (!ioc)
1592 return NULL;
1593
Jens Axboe4ac845a2008-01-24 08:44:49 +01001594 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001595 if (cic)
1596 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001597
Jens Axboee2d74ac2006-03-28 08:59:01 +02001598 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1599 if (cic == NULL)
1600 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001601
Jens Axboe4ac845a2008-01-24 08:44:49 +01001602 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1603 goto err_free;
1604
Jens Axboe22e2c502005-06-27 10:55:12 +02001605out:
Jens Axboefc463792006-08-29 09:05:44 +02001606 smp_read_barrier_depends();
1607 if (unlikely(ioc->ioprio_changed))
1608 cfq_ioc_set_ioprio(ioc);
1609
Jens Axboe22e2c502005-06-27 10:55:12 +02001610 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001611err_free:
1612 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001613err:
1614 put_io_context(ioc);
1615 return NULL;
1616}
1617
1618static void
1619cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1620{
Jens Axboeaaf12282007-01-19 11:30:16 +11001621 unsigned long elapsed = jiffies - cic->last_end_request;
1622 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001623
1624 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1625 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1626 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1627}
1628
Jens Axboe206dc692006-03-28 13:03:44 +02001629static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001630cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1631 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001632{
1633 sector_t sdist;
1634 u64 total;
1635
Jens Axboe5e705372006-07-13 12:39:25 +02001636 if (cic->last_request_pos < rq->sector)
1637 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001638 else
Jens Axboe5e705372006-07-13 12:39:25 +02001639 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001640
1641 /*
1642 * Don't allow the seek distance to get too large from the
1643 * odd fragment, pagein, etc
1644 */
1645 if (cic->seek_samples <= 60) /* second&third seek */
1646 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1647 else
1648 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1649
1650 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1651 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1652 total = cic->seek_total + (cic->seek_samples/2);
1653 do_div(total, cic->seek_samples);
1654 cic->seek_mean = (sector_t)total;
1655}
Jens Axboe22e2c502005-06-27 10:55:12 +02001656
1657/*
1658 * Disable idle window if the process thinks too long or seeks so much that
1659 * it doesn't matter
1660 */
1661static void
1662cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1663 struct cfq_io_context *cic)
1664{
Jens Axboe1be92f22007-04-19 14:32:26 +02001665 int enable_idle;
1666
Jens Axboe08717142008-01-28 11:38:15 +01001667 /*
1668 * Don't idle for async or idle io prio class
1669 */
1670 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02001671 return;
1672
1673 enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001674
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001675 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Jens Axboecaaa5f92006-06-16 11:23:00 +02001676 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001677 enable_idle = 0;
1678 else if (sample_valid(cic->ttime_samples)) {
1679 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1680 enable_idle = 0;
1681 else
1682 enable_idle = 1;
1683 }
1684
Jens Axboe3b181522005-06-27 10:56:24 +02001685 if (enable_idle)
1686 cfq_mark_cfqq_idle_window(cfqq);
1687 else
1688 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001689}
1690
Jens Axboe22e2c502005-06-27 10:55:12 +02001691/*
1692 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1693 * no or if we aren't sure, a 1 will cause a preempt.
1694 */
1695static int
1696cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001697 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001698{
Jens Axboe6d048f52007-04-25 12:44:27 +02001699 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001700
Jens Axboe6d048f52007-04-25 12:44:27 +02001701 cfqq = cfqd->active_queue;
1702 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001703 return 0;
1704
Jens Axboe6d048f52007-04-25 12:44:27 +02001705 if (cfq_slice_used(cfqq))
1706 return 1;
1707
1708 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001709 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001710
1711 if (cfq_class_idle(cfqq))
1712 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001713
Jens Axboe22e2c502005-06-27 10:55:12 +02001714 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02001715 * if the new request is sync, but the currently running queue is
1716 * not, let the sync request have priority.
1717 */
Jens Axboe5e705372006-07-13 12:39:25 +02001718 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001719 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001720
Jens Axboe374f84a2006-07-23 01:42:19 +02001721 /*
1722 * So both queues are sync. Let the new request get disk time if
1723 * it's a metadata request and the current queue is doing regular IO.
1724 */
1725 if (rq_is_meta(rq) && !cfqq->meta_pending)
1726 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001727
Jens Axboe1e3335d2007-02-14 19:59:49 +01001728 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1729 return 0;
1730
1731 /*
1732 * if this request is as-good as one we would expect from the
1733 * current cfqq, let it preempt
1734 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001735 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01001736 return 1;
1737
Jens Axboe22e2c502005-06-27 10:55:12 +02001738 return 0;
1739}
1740
1741/*
1742 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1743 * let it have half of its nominal slice.
1744 */
1745static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1746{
Jens Axboe6084cdd2007-04-23 08:25:00 +02001747 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001748
Jens Axboebf572252006-07-19 20:29:12 +02001749 /*
1750 * Put the new queue at the front of the of the current list,
1751 * so we know that it will be selected next.
1752 */
1753 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02001754
1755 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02001756
Jens Axboe44f7c162007-01-19 11:51:58 +11001757 cfqq->slice_end = 0;
1758 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001759}
1760
1761/*
Jens Axboe5e705372006-07-13 12:39:25 +02001762 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001763 * something we should do about it
1764 */
1765static void
Jens Axboe5e705372006-07-13 12:39:25 +02001766cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1767 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001768{
Jens Axboe5e705372006-07-13 12:39:25 +02001769 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001770
Jens Axboe374f84a2006-07-23 01:42:19 +02001771 if (rq_is_meta(rq))
1772 cfqq->meta_pending++;
1773
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001774 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02001775 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001776 cfq_update_idle_window(cfqd, cfqq, cic);
1777
Jens Axboe5e705372006-07-13 12:39:25 +02001778 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001779
1780 if (cfqq == cfqd->active_queue) {
1781 /*
1782 * if we are waiting for a request for this queue, let it rip
1783 * immediately and flag that we must not expire this queue
1784 * just now
1785 */
Jens Axboe3b181522005-06-27 10:56:24 +02001786 if (cfq_cfqq_wait_request(cfqq)) {
1787 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001788 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001789 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001790 }
Jens Axboe5e705372006-07-13 12:39:25 +02001791 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001792 /*
1793 * not the active queue - expire current slice if it is
1794 * idle and has expired it's mean thinktime or this new queue
1795 * has some old slice time left and is of higher priority
1796 */
1797 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001798 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001799 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001800 }
1801}
1802
Jens Axboe165125e2007-07-24 09:28:11 +02001803static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001804{
Jens Axboeb4878f22005-10-20 16:42:29 +02001805 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001806 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001807
Jens Axboefd0928d2008-01-24 08:52:45 +01001808 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809
Jens Axboe5e705372006-07-13 12:39:25 +02001810 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811
Jens Axboe22e2c502005-06-27 10:55:12 +02001812 list_add_tail(&rq->queuelist, &cfqq->fifo);
1813
Jens Axboe5e705372006-07-13 12:39:25 +02001814 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815}
1816
Jens Axboe165125e2007-07-24 09:28:11 +02001817static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818{
Jens Axboe5e705372006-07-13 12:39:25 +02001819 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001820 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001821 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001822 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823
Jens Axboeb4878f22005-10-20 16:42:29 +02001824 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825
Jens Axboeb4878f22005-10-20 16:42:29 +02001826 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02001827 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02001828 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02001829 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830
Jens Axboe3ed9a292007-04-23 08:33:33 +02001831 if (cfq_cfqq_sync(cfqq))
1832 cfqd->sync_flight--;
1833
Jens Axboeb4878f22005-10-20 16:42:29 +02001834 if (!cfq_class_idle(cfqq))
1835 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001836
Jens Axboecaaa5f92006-06-16 11:23:00 +02001837 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001838 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001839
1840 /*
1841 * If this is the active queue, check if it needs to be expired,
1842 * or if we want to idle in case it has no pending requests.
1843 */
1844 if (cfqd->active_queue == cfqq) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001845 if (cfq_cfqq_slice_new(cfqq)) {
1846 cfq_set_prio_slice(cfqd, cfqq);
1847 cfq_clear_cfqq_slice_new(cfqq);
1848 }
Jens Axboe08717142008-01-28 11:38:15 +01001849 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02001850 cfq_slice_expired(cfqd, 1);
Jens Axboe6d048f52007-04-25 12:44:27 +02001851 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1852 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001853 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001854
1855 if (!cfqd->rq_in_driver)
1856 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857}
1858
Jens Axboe22e2c502005-06-27 10:55:12 +02001859/*
1860 * we temporarily boost lower priority queues if they are holding fs exclusive
1861 * resources. they are boosted to normal prio (CLASS_BE/4)
1862 */
1863static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864{
Jens Axboe22e2c502005-06-27 10:55:12 +02001865 if (has_fs_excl()) {
1866 /*
1867 * boost idle prio on transactions that would lock out other
1868 * users of the filesystem
1869 */
1870 if (cfq_class_idle(cfqq))
1871 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1872 if (cfqq->ioprio > IOPRIO_NORM)
1873 cfqq->ioprio = IOPRIO_NORM;
1874 } else {
1875 /*
1876 * check if we need to unboost the queue
1877 */
1878 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1879 cfqq->ioprio_class = cfqq->org_ioprio_class;
1880 if (cfqq->ioprio != cfqq->org_ioprio)
1881 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001883}
1884
Jens Axboe89850f72006-07-22 16:48:31 +02001885static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001886{
Jens Axboe3b181522005-06-27 10:56:24 +02001887 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001888 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001889 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001890 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001891 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001892
1893 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02001894}
1895
Jens Axboe165125e2007-07-24 09:28:11 +02001896static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02001897{
1898 struct cfq_data *cfqd = q->elevator->elevator_data;
1899 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001900 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02001901 struct cfq_queue *cfqq;
1902
1903 /*
1904 * don't force setup of a queue from here, as a call to may_queue
1905 * does not necessarily imply that a request actually will be queued.
1906 * so just lookup a possibly existing queue, or return 'may queue'
1907 * if that fails
1908 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001909 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001910 if (!cic)
1911 return ELV_MQUEUE_MAY;
1912
1913 cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02001914 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01001915 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02001916 cfq_prio_boost(cfqq);
1917
Jens Axboe89850f72006-07-22 16:48:31 +02001918 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001919 }
1920
1921 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922}
1923
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924/*
1925 * queue lock held here
1926 */
Jens Axboebb37b942006-12-01 10:42:33 +01001927static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928{
Jens Axboe5e705372006-07-13 12:39:25 +02001929 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930
Jens Axboe5e705372006-07-13 12:39:25 +02001931 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001932 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933
Jens Axboe22e2c502005-06-27 10:55:12 +02001934 BUG_ON(!cfqq->allocated[rw]);
1935 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936
Jens Axboe5e705372006-07-13 12:39:25 +02001937 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02001940 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942 cfq_put_queue(cfqq);
1943 }
1944}
1945
1946/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001947 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001949static int
Jens Axboe165125e2007-07-24 09:28:11 +02001950cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951{
1952 struct cfq_data *cfqd = q->elevator->elevator_data;
1953 struct cfq_io_context *cic;
1954 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01001955 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001956 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957 unsigned long flags;
1958
1959 might_sleep_if(gfp_mask & __GFP_WAIT);
1960
Jens Axboee2d74ac2006-03-28 08:59:01 +02001961 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001962
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963 spin_lock_irqsave(q->queue_lock, flags);
1964
Jens Axboe22e2c502005-06-27 10:55:12 +02001965 if (!cic)
1966 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967
Vasily Tarasov91fac312007-04-25 12:29:51 +02001968 cfqq = cic_to_cfqq(cic, is_sync);
1969 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01001970 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001971
Jens Axboe22e2c502005-06-27 10:55:12 +02001972 if (!cfqq)
1973 goto queue_fail;
1974
Vasily Tarasov91fac312007-04-25 12:29:51 +02001975 cic_set_cfqq(cic, cfqq, is_sync);
1976 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977
1978 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02001979 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001980 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02001981
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 spin_unlock_irqrestore(q->queue_lock, flags);
1983
Jens Axboe5e705372006-07-13 12:39:25 +02001984 rq->elevator_private = cic;
1985 rq->elevator_private2 = cfqq;
1986 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02001987
Jens Axboe22e2c502005-06-27 10:55:12 +02001988queue_fail:
1989 if (cic)
1990 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02001991
Jens Axboe3b181522005-06-27 10:56:24 +02001992 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993 spin_unlock_irqrestore(q->queue_lock, flags);
1994 return 1;
1995}
1996
David Howells65f27f32006-11-22 14:55:48 +00001997static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02001998{
David Howells65f27f32006-11-22 14:55:48 +00001999 struct cfq_data *cfqd =
2000 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02002001 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002002 unsigned long flags;
2003
2004 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02002005 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02002006 spin_unlock_irqrestore(q->queue_lock, flags);
2007}
2008
2009/*
2010 * Timer running if the active_queue is currently idling inside its time slice
2011 */
2012static void cfq_idle_slice_timer(unsigned long data)
2013{
2014 struct cfq_data *cfqd = (struct cfq_data *) data;
2015 struct cfq_queue *cfqq;
2016 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002017 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002018
2019 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2020
2021 if ((cfqq = cfqd->active_queue) != NULL) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002022 timed_out = 0;
2023
Jens Axboe22e2c502005-06-27 10:55:12 +02002024 /*
2025 * expired
2026 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002027 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002028 goto expire;
2029
2030 /*
2031 * only expire and reinvoke request handler, if there are
2032 * other queues with pending requests
2033 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002034 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002035 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002036
2037 /*
2038 * not expired and it has a request pending, let it dispatch
2039 */
Jens Axboedd67d052006-06-21 09:36:18 +02002040 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002041 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002042 goto out_kick;
2043 }
2044 }
2045expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002046 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002047out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002048 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002049out_cont:
2050 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2051}
2052
Jens Axboe3b181522005-06-27 10:56:24 +02002053static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2054{
2055 del_timer_sync(&cfqd->idle_slice_timer);
Oleg Nesterov43108642007-10-23 15:08:18 +02002056 kblockd_flush_work(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002057}
Jens Axboe22e2c502005-06-27 10:55:12 +02002058
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002059static void cfq_put_async_queues(struct cfq_data *cfqd)
2060{
2061 int i;
2062
2063 for (i = 0; i < IOPRIO_BE_NR; i++) {
2064 if (cfqd->async_cfqq[0][i])
2065 cfq_put_queue(cfqd->async_cfqq[0][i]);
2066 if (cfqd->async_cfqq[1][i])
2067 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002068 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002069
2070 if (cfqd->async_idle_cfqq)
2071 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002072}
2073
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074static void cfq_exit_queue(elevator_t *e)
2075{
Jens Axboe22e2c502005-06-27 10:55:12 +02002076 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002077 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002078
Jens Axboe3b181522005-06-27 10:56:24 +02002079 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002080
Al Virod9ff4182006-03-18 13:51:22 -05002081 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002082
Al Virod9ff4182006-03-18 13:51:22 -05002083 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002084 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002085
2086 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002087 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2088 struct cfq_io_context,
2089 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002090
2091 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002092 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002093
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002094 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002095
Al Virod9ff4182006-03-18 13:51:22 -05002096 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002097
2098 cfq_shutdown_timer_wq(cfqd);
2099
Al Viroa90d7422006-03-18 12:05:37 -05002100 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101}
2102
Jens Axboe165125e2007-07-24 09:28:11 +02002103static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104{
2105 struct cfq_data *cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106
Christoph Lameter94f60302007-07-17 04:03:29 -07002107 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002109 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110
Jens Axboecc09e292007-04-26 12:53:50 +02002111 cfqd->service_tree = CFQ_RB_ROOT;
Al Virod9ff4182006-03-18 13:51:22 -05002112 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115
Jens Axboe22e2c502005-06-27 10:55:12 +02002116 init_timer(&cfqd->idle_slice_timer);
2117 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2118 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2119
David Howells65f27f32006-11-22 14:55:48 +00002120 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002121
Oleg Nesterovb70c8642007-11-07 09:46:13 +01002122 cfqd->last_end_request = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002124 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2125 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 cfqd->cfq_back_max = cfq_back_max;
2127 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002128 cfqd->cfq_slice[0] = cfq_slice_async;
2129 cfqd->cfq_slice[1] = cfq_slice_sync;
2130 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2131 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02002132
Jens Axboebc1c1162006-06-08 08:49:06 +02002133 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134}
2135
2136static void cfq_slab_kill(void)
2137{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 if (cfq_pool)
2139 kmem_cache_destroy(cfq_pool);
2140 if (cfq_ioc_pool)
2141 kmem_cache_destroy(cfq_ioc_pool);
2142}
2143
2144static int __init cfq_slab_setup(void)
2145{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002146 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 if (!cfq_pool)
2148 goto fail;
2149
Jens Axboe4ac845a2008-01-24 08:44:49 +01002150 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, SLAB_DESTROY_BY_RCU);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151 if (!cfq_ioc_pool)
2152 goto fail;
2153
2154 return 0;
2155fail:
2156 cfq_slab_kill();
2157 return -ENOMEM;
2158}
2159
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160/*
2161 * sysfs parts below -->
2162 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163static ssize_t
2164cfq_var_show(unsigned int var, char *page)
2165{
2166 return sprintf(page, "%d\n", var);
2167}
2168
2169static ssize_t
2170cfq_var_store(unsigned int *var, const char *page, size_t count)
2171{
2172 char *p = (char *) page;
2173
2174 *var = simple_strtoul(p, &p, 10);
2175 return count;
2176}
2177
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002179static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002181 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182 unsigned int __data = __VAR; \
2183 if (__CONV) \
2184 __data = jiffies_to_msecs(__data); \
2185 return cfq_var_show(__data, (page)); \
2186}
2187SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002188SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2189SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002190SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2191SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002192SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2193SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2194SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2195SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196#undef SHOW_FUNCTION
2197
2198#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002199static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002201 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 unsigned int __data; \
2203 int ret = cfq_var_store(&__data, (page), count); \
2204 if (__data < (MIN)) \
2205 __data = (MIN); \
2206 else if (__data > (MAX)) \
2207 __data = (MAX); \
2208 if (__CONV) \
2209 *(__PTR) = msecs_to_jiffies(__data); \
2210 else \
2211 *(__PTR) = __data; \
2212 return ret; \
2213}
2214STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002215STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2216STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002217STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2218STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002219STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2220STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2221STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2222STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223#undef STORE_FUNCTION
2224
Al Viroe572ec72006-03-18 22:27:18 -05002225#define CFQ_ATTR(name) \
2226 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002227
Al Viroe572ec72006-03-18 22:27:18 -05002228static struct elv_fs_entry cfq_attrs[] = {
2229 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002230 CFQ_ATTR(fifo_expire_sync),
2231 CFQ_ATTR(fifo_expire_async),
2232 CFQ_ATTR(back_seek_max),
2233 CFQ_ATTR(back_seek_penalty),
2234 CFQ_ATTR(slice_sync),
2235 CFQ_ATTR(slice_async),
2236 CFQ_ATTR(slice_async_rq),
2237 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002238 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239};
2240
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241static struct elevator_type iosched_cfq = {
2242 .ops = {
2243 .elevator_merge_fn = cfq_merge,
2244 .elevator_merged_fn = cfq_merged_request,
2245 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002246 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002247 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002249 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 .elevator_deactivate_req_fn = cfq_deactivate_request,
2251 .elevator_queue_empty_fn = cfq_queue_empty,
2252 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002253 .elevator_former_req_fn = elv_rb_former_request,
2254 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 .elevator_set_req_fn = cfq_set_request,
2256 .elevator_put_req_fn = cfq_put_request,
2257 .elevator_may_queue_fn = cfq_may_queue,
2258 .elevator_init_fn = cfq_init_queue,
2259 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002260 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261 },
Al Viro3d1ab402006-03-18 18:35:43 -05002262 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 .elevator_name = "cfq",
2264 .elevator_owner = THIS_MODULE,
2265};
2266
2267static int __init cfq_init(void)
2268{
Jens Axboe22e2c502005-06-27 10:55:12 +02002269 /*
2270 * could be 0 on HZ < 1000 setups
2271 */
2272 if (!cfq_slice_async)
2273 cfq_slice_async = 1;
2274 if (!cfq_slice_idle)
2275 cfq_slice_idle = 1;
2276
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 if (cfq_slab_setup())
2278 return -ENOMEM;
2279
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002280 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002282 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283}
2284
2285static void __exit cfq_exit(void)
2286{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002287 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002289 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002290 /* ioc_gone's update must be visible before reading ioc_count */
2291 smp_wmb();
Jens Axboe4050cf12006-07-19 05:07:12 +02002292 if (elv_ioc_count_read(ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002293 wait_for_completion(ioc_gone);
Al Viro334e94d2006-03-18 15:05:53 -05002294 synchronize_rcu();
Christoph Hellwig83521d32005-10-30 15:01:39 -08002295 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296}
2297
2298module_init(cfq_init);
2299module_exit(cfq_exit);
2300
2301MODULE_AUTHOR("Jens Axboe");
2302MODULE_LICENSE("GPL");
2303MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");