blob: d01b411c72f053225c2faee733e4299afeeec51b [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 */
Jens Axboefe094d92008-01-31 13:08:54 +010018/* max queue in one round of service */
19static const int cfq_quantum = 4;
Arjan van de Ven64100092006-01-06 09:46:02 +010020static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010021/* maximum backwards seek, in KiB */
22static const int cfq_back_max = 16 * 1024;
23/* penalty of a backwards seek */
24static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010025static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020026static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010027static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020028static int cfq_slice_idle = HZ / 125;
Jens Axboe22e2c502005-06-27 10:55:12 +020029
Jens Axboed9e76202007-04-20 14:27:50 +020030/*
Jens Axboe08717142008-01-28 11:38:15 +010031 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020032 */
Jens Axboe08717142008-01-28 11:38:15 +010033#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020034
35/*
36 * below this threshold, we consider thinktime immediate
37 */
38#define CFQ_MIN_TT (2)
39
Jens Axboe22e2c502005-06-27 10:55:12 +020040#define CFQ_SLICE_SCALE (5)
41
Jens Axboefe094d92008-01-31 13:08:54 +010042#define RQ_CIC(rq) \
43 ((struct cfq_io_context *) (rq)->elevator_private)
Jens Axboe5e705372006-07-13 12:39:25 +020044#define RQ_CFQQ(rq) ((rq)->elevator_private2)
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
Christoph Lametere18b8902006-12-06 20:33:20 -080046static struct kmem_cache *cfq_pool;
47static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070048
Jens Axboe4050cf12006-07-19 05:07:12 +020049static DEFINE_PER_CPU(unsigned long, ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050050static struct completion *ioc_gone;
51
Jens Axboe22e2c502005-06-27 10:55:12 +020052#define CFQ_PRIO_LISTS IOPRIO_BE_NR
53#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020054#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
55
Jens Axboe3b181522005-06-27 10:56:24 +020056#define ASYNC (0)
57#define SYNC (1)
58
Jens Axboe206dc692006-03-28 13:03:44 +020059#define sample_valid(samples) ((samples) > 80)
60
Jens Axboe22e2c502005-06-27 10:55:12 +020061/*
Jens Axboecc09e292007-04-26 12:53:50 +020062 * Most of our rbtree usage is for sorting with min extraction, so
63 * if we cache the leftmost node we don't have to walk down the tree
64 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
65 * move this into the elevator for the rq sorting as well.
66 */
67struct cfq_rb_root {
68 struct rb_root rb;
69 struct rb_node *left;
70};
71#define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
72
73/*
Jens Axboe22e2c502005-06-27 10:55:12 +020074 * Per block device queue structure
75 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070076struct cfq_data {
Jens Axboe165125e2007-07-24 09:28:11 +020077 struct request_queue *queue;
Jens Axboe22e2c502005-06-27 10:55:12 +020078
79 /*
80 * rr list of queues with requests and the count of them
81 */
Jens Axboecc09e292007-04-26 12:53:50 +020082 struct cfq_rb_root service_tree;
Jens Axboe22e2c502005-06-27 10:55:12 +020083 unsigned int busy_queues;
84
Jens Axboe22e2c502005-06-27 10:55:12 +020085 int rq_in_driver;
Jens Axboe3ed9a292007-04-23 08:33:33 +020086 int sync_flight;
Jens Axboe25776e32006-06-01 10:12:26 +020087 int hw_tag;
Jens Axboe22e2c502005-06-27 10:55:12 +020088
89 /*
Jens Axboe22e2c502005-06-27 10:55:12 +020090 * idle window management
91 */
92 struct timer_list idle_slice_timer;
93 struct work_struct unplug_work;
94
95 struct cfq_queue *active_queue;
96 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +020097
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +020098 /*
99 * async queue for each priority case
100 */
101 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
102 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200103
Jens Axboe6d048f52007-04-25 12:44:27 +0200104 sector_t last_position;
Jens Axboe22e2c502005-06-27 10:55:12 +0200105 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107 /*
108 * tunables, see top of file
109 */
110 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200111 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112 unsigned int cfq_back_penalty;
113 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200114 unsigned int cfq_slice[2];
115 unsigned int cfq_slice_async_rq;
116 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500117
118 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119};
120
Jens Axboe22e2c502005-06-27 10:55:12 +0200121/*
122 * Per process-grouping structure
123 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124struct cfq_queue {
125 /* reference count */
126 atomic_t ref;
Richard Kennedybe754d22008-05-23 06:52:00 +0200127 /* various state flags, see below */
128 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 /* parent cfq_data */
130 struct cfq_data *cfqd;
Jens Axboed9e76202007-04-20 14:27:50 +0200131 /* service_tree member */
132 struct rb_node rb_node;
133 /* service_tree key */
134 unsigned long rb_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 /* sorted list of pending requests */
136 struct rb_root sort_list;
137 /* if fifo isn't expired, next request to serve */
Jens Axboe5e705372006-07-13 12:39:25 +0200138 struct request *next_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 /* requests queued in sort_list */
140 int queued[2];
141 /* currently allocated requests */
142 int allocated[2];
143 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200144 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
Jens Axboe22e2c502005-06-27 10:55:12 +0200146 unsigned long slice_end;
Jens Axboec5b680f2007-01-19 11:56:49 +1100147 long slice_resid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148
Richard Kennedybe754d22008-05-23 06:52:00 +0200149 /* pending metadata requests */
150 int meta_pending;
Jens Axboe6d048f52007-04-25 12:44:27 +0200151 /* number of requests that are on the dispatch list or inside driver */
152 int dispatched;
Jens Axboe22e2c502005-06-27 10:55:12 +0200153
154 /* io prio of this group */
155 unsigned short ioprio, org_ioprio;
156 unsigned short ioprio_class, org_ioprio_class;
157
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158};
159
Jens Axboe3b181522005-06-27 10:56:24 +0200160enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100161 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
162 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
163 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
164 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
165 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
166 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
167 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
168 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
169 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
Jens Axboe44f7c162007-01-19 11:51:58 +1100170 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200171 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200172};
173
174#define CFQ_CFQQ_FNS(name) \
175static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
176{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100177 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200178} \
179static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
180{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100181 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200182} \
183static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
184{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100185 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200186}
187
188CFQ_CFQQ_FNS(on_rr);
189CFQ_CFQQ_FNS(wait_request);
190CFQ_CFQQ_FNS(must_alloc);
191CFQ_CFQQ_FNS(must_alloc_slice);
192CFQ_CFQQ_FNS(must_dispatch);
193CFQ_CFQQ_FNS(fifo_expire);
194CFQ_CFQQ_FNS(idle_window);
195CFQ_CFQQ_FNS(prio_changed);
Jens Axboe53b037442006-07-28 09:48:51 +0200196CFQ_CFQQ_FNS(queue_new);
Jens Axboe44f7c162007-01-19 11:51:58 +1100197CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200198CFQ_CFQQ_FNS(sync);
Jens Axboe3b181522005-06-27 10:56:24 +0200199#undef CFQ_CFQQ_FNS
200
Jens Axboe165125e2007-07-24 09:28:11 +0200201static void cfq_dispatch_insert(struct request_queue *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200202static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
Jens Axboefd0928d2008-01-24 08:52:45 +0100203 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100204static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200205 struct io_context *);
206
207static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
208 int is_sync)
209{
210 return cic->cfqq[!!is_sync];
211}
212
213static inline void cic_set_cfqq(struct cfq_io_context *cic,
214 struct cfq_queue *cfqq, int is_sync)
215{
216 cic->cfqq[!!is_sync] = cfqq;
217}
218
219/*
220 * We regard a request as SYNC, if it's either a read or has the SYNC bit
221 * set (in which case it could also be direct WRITE).
222 */
223static inline int cfq_bio_sync(struct bio *bio)
224{
225 if (bio_data_dir(bio) == READ || bio_sync(bio))
226 return 1;
227
228 return 0;
229}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700232 * scheduler run of queue, if there are requests pending and no one in the
233 * driver that will restart queueing
234 */
235static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
236{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100237 if (cfqd->busy_queues)
Andrew Morton99f95e52005-06-27 20:14:05 -0700238 kblockd_schedule_work(&cfqd->unplug_work);
239}
240
Jens Axboe165125e2007-07-24 09:28:11 +0200241static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700242{
243 struct cfq_data *cfqd = q->elevator->elevator_data;
244
Jens Axboeb4878f22005-10-20 16:42:29 +0200245 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700246}
247
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100249 * Scale schedule slice based on io priority. Use the sync time slice only
250 * if a queue is marked sync and has sync io queued. A sync queue with async
251 * io only, should not get full sync slice length.
252 */
Jens Axboed9e76202007-04-20 14:27:50 +0200253static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
254 unsigned short prio)
255{
256 const int base_slice = cfqd->cfq_slice[sync];
257
258 WARN_ON(prio >= IOPRIO_BE_NR);
259
260 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
261}
262
Jens Axboe44f7c162007-01-19 11:51:58 +1100263static inline int
264cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
265{
Jens Axboed9e76202007-04-20 14:27:50 +0200266 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100267}
268
269static inline void
270cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
271{
272 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
273}
274
275/*
276 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
277 * isn't valid until the first request from the dispatch is activated
278 * and the slice time set.
279 */
280static inline int cfq_slice_used(struct cfq_queue *cfqq)
281{
282 if (cfq_cfqq_slice_new(cfqq))
283 return 0;
284 if (time_before(jiffies, cfqq->slice_end))
285 return 0;
286
287 return 1;
288}
289
290/*
Jens Axboe5e705372006-07-13 12:39:25 +0200291 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200293 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294 */
Jens Axboe5e705372006-07-13 12:39:25 +0200295static struct request *
296cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297{
298 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200300#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
301#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
302 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303
Jens Axboe5e705372006-07-13 12:39:25 +0200304 if (rq1 == NULL || rq1 == rq2)
305 return rq2;
306 if (rq2 == NULL)
307 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200308
Jens Axboe5e705372006-07-13 12:39:25 +0200309 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
310 return rq1;
311 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
312 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200313 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
314 return rq1;
315 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
316 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317
Jens Axboe5e705372006-07-13 12:39:25 +0200318 s1 = rq1->sector;
319 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320
Jens Axboe6d048f52007-04-25 12:44:27 +0200321 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323 /*
324 * by definition, 1KiB is 2 sectors
325 */
326 back_max = cfqd->cfq_back_max * 2;
327
328 /*
329 * Strict one way elevator _except_ in the case where we allow
330 * short backward seeks which are biased as twice the cost of a
331 * similar forward seek.
332 */
333 if (s1 >= last)
334 d1 = s1 - last;
335 else if (s1 + back_max >= last)
336 d1 = (last - s1) * cfqd->cfq_back_penalty;
337 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200338 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339
340 if (s2 >= last)
341 d2 = s2 - last;
342 else if (s2 + back_max >= last)
343 d2 = (last - s2) * cfqd->cfq_back_penalty;
344 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200345 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346
347 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348
Andreas Mohre8a99052006-03-28 08:59:49 +0200349 /*
350 * By doing switch() on the bit mask "wrap" we avoid having to
351 * check two variables for all permutations: --> faster!
352 */
353 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200354 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200355 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200356 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200357 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200358 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200359 else {
360 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200361 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200362 else
Jens Axboe5e705372006-07-13 12:39:25 +0200363 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200364 }
365
366 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200367 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200368 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200369 return rq2;
370 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200371 default:
372 /*
373 * Since both rqs are wrapped,
374 * start with the one that's further behind head
375 * (--> only *one* back seek required),
376 * since back seek takes more time than forward.
377 */
378 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200379 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 else
Jens Axboe5e705372006-07-13 12:39:25 +0200381 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 }
383}
384
Jens Axboe498d3aa22007-04-26 12:54:48 +0200385/*
386 * The below is leftmost cache rbtree addon
387 */
Jens Axboe08717142008-01-28 11:38:15 +0100388static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200389{
390 if (!root->left)
391 root->left = rb_first(&root->rb);
392
Jens Axboe08717142008-01-28 11:38:15 +0100393 if (root->left)
394 return rb_entry(root->left, struct cfq_queue, rb_node);
395
396 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200397}
398
399static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
400{
401 if (root->left == n)
402 root->left = NULL;
403
404 rb_erase(n, &root->rb);
405 RB_CLEAR_NODE(n);
406}
407
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408/*
409 * would be nice to take fifo expire time into account as well
410 */
Jens Axboe5e705372006-07-13 12:39:25 +0200411static struct request *
412cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
413 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414{
Jens Axboe21183b02006-07-13 12:33:14 +0200415 struct rb_node *rbnext = rb_next(&last->rb_node);
416 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200417 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418
Jens Axboe21183b02006-07-13 12:33:14 +0200419 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420
421 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200422 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200423
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200425 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200426 else {
427 rbnext = rb_first(&cfqq->sort_list);
428 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200429 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200430 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431
Jens Axboe21183b02006-07-13 12:33:14 +0200432 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433}
434
Jens Axboed9e76202007-04-20 14:27:50 +0200435static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
436 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437{
Jens Axboed9e76202007-04-20 14:27:50 +0200438 /*
439 * just an approximation, should be ok.
440 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200441 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
442 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200443}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444
Jens Axboe498d3aa22007-04-26 12:54:48 +0200445/*
446 * The cfqd->service_tree holds all pending cfq_queue's that have
447 * requests waiting to be processed. It is sorted in the order that
448 * we will service the queues.
449 */
Jens Axboed9e76202007-04-20 14:27:50 +0200450static void cfq_service_tree_add(struct cfq_data *cfqd,
Jens Axboeedd75ff2007-04-19 12:03:34 +0200451 struct cfq_queue *cfqq, int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200452{
Jens Axboe08717142008-01-28 11:38:15 +0100453 struct rb_node **p, *parent;
454 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +0200455 unsigned long rb_key;
Jens Axboe498d3aa22007-04-26 12:54:48 +0200456 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200457
Jens Axboe08717142008-01-28 11:38:15 +0100458 if (cfq_class_idle(cfqq)) {
459 rb_key = CFQ_IDLE_DELAY;
460 parent = rb_last(&cfqd->service_tree.rb);
461 if (parent && parent != &cfqq->rb_node) {
462 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
463 rb_key += __cfqq->rb_key;
464 } else
465 rb_key += jiffies;
466 } else if (!add_front) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200467 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
468 rb_key += cfqq->slice_resid;
469 cfqq->slice_resid = 0;
470 } else
471 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200472
473 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100474 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200475 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100476 */
Jens Axboed9e76202007-04-20 14:27:50 +0200477 if (rb_key == cfqq->rb_key)
478 return;
Jens Axboe53b037442006-07-28 09:48:51 +0200479
Jens Axboecc09e292007-04-26 12:53:50 +0200480 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200481 }
Jens Axboed9e76202007-04-20 14:27:50 +0200482
Jens Axboe498d3aa22007-04-26 12:54:48 +0200483 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +0100484 parent = NULL;
485 p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200486 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +0200487 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200488
Jens Axboed9e76202007-04-20 14:27:50 +0200489 parent = *p;
490 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
491
Jens Axboe0c534e02007-04-18 20:01:57 +0200492 /*
493 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200494 * preference to them. IDLE queues goes to the back.
495 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200496 */
497 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200498 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200499 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200500 n = &(*p)->rb_right;
501 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
502 n = &(*p)->rb_left;
503 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
504 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200505 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200506 n = &(*p)->rb_left;
507 else
508 n = &(*p)->rb_right;
509
510 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200511 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200512
513 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200514 }
515
Jens Axboecc09e292007-04-26 12:53:50 +0200516 if (left)
517 cfqd->service_tree.left = &cfqq->rb_node;
518
Jens Axboed9e76202007-04-20 14:27:50 +0200519 cfqq->rb_key = rb_key;
520 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200521 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522}
523
Jens Axboe498d3aa22007-04-26 12:54:48 +0200524/*
525 * Update cfqq's position in the service tree.
526 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200527static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200528{
Jens Axboe6d048f52007-04-25 12:44:27 +0200529 /*
530 * Resorting requires the cfqq to be on the RR list already.
531 */
Jens Axboe498d3aa22007-04-26 12:54:48 +0200532 if (cfq_cfqq_on_rr(cfqq))
Jens Axboeedd75ff2007-04-19 12:03:34 +0200533 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboe6d048f52007-04-25 12:44:27 +0200534}
535
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536/*
537 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200538 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539 */
Jens Axboefebffd62008-01-28 13:19:43 +0100540static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541{
Jens Axboe3b181522005-06-27 10:56:24 +0200542 BUG_ON(cfq_cfqq_on_rr(cfqq));
543 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 cfqd->busy_queues++;
545
Jens Axboeedd75ff2007-04-19 12:03:34 +0200546 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547}
548
Jens Axboe498d3aa22007-04-26 12:54:48 +0200549/*
550 * Called when the cfqq no longer has requests pending, remove it from
551 * the service tree.
552 */
Jens Axboefebffd62008-01-28 13:19:43 +0100553static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554{
Jens Axboe3b181522005-06-27 10:56:24 +0200555 BUG_ON(!cfq_cfqq_on_rr(cfqq));
556 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557
Jens Axboecc09e292007-04-26 12:53:50 +0200558 if (!RB_EMPTY_NODE(&cfqq->rb_node))
559 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboed9e76202007-04-20 14:27:50 +0200560
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 BUG_ON(!cfqd->busy_queues);
562 cfqd->busy_queues--;
563}
564
565/*
566 * rb tree support functions
567 */
Jens Axboefebffd62008-01-28 13:19:43 +0100568static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569{
Jens Axboe5e705372006-07-13 12:39:25 +0200570 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200571 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200572 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573
Jens Axboeb4878f22005-10-20 16:42:29 +0200574 BUG_ON(!cfqq->queued[sync]);
575 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576
Jens Axboe5e705372006-07-13 12:39:25 +0200577 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578
Jens Axboedd67d052006-06-21 09:36:18 +0200579 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200580 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581}
582
Jens Axboe5e705372006-07-13 12:39:25 +0200583static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584{
Jens Axboe5e705372006-07-13 12:39:25 +0200585 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe21183b02006-07-13 12:33:14 +0200587 struct request *__alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588
Jens Axboe5380a102006-07-13 12:37:56 +0200589 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590
591 /*
592 * looks a little odd, but the first insert might return an alias.
593 * if that happens, put the alias on the dispatch list
594 */
Jens Axboe21183b02006-07-13 12:33:14 +0200595 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200596 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100597
598 if (!cfq_cfqq_on_rr(cfqq))
599 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200600
601 /*
602 * check if this request is a better next-serve candidate
603 */
604 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
605 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606}
607
Jens Axboefebffd62008-01-28 13:19:43 +0100608static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609{
Jens Axboe5380a102006-07-13 12:37:56 +0200610 elv_rb_del(&cfqq->sort_list, rq);
611 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200612 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613}
614
Jens Axboe206dc692006-03-28 13:03:44 +0200615static struct request *
616cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617{
Jens Axboe206dc692006-03-28 13:03:44 +0200618 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200619 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200620 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621
Jens Axboe4ac845a2008-01-24 08:44:49 +0100622 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200623 if (!cic)
624 return NULL;
625
626 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200627 if (cfqq) {
628 sector_t sector = bio->bi_sector + bio_sectors(bio);
629
Jens Axboe21183b02006-07-13 12:33:14 +0200630 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200631 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633 return NULL;
634}
635
Jens Axboe165125e2007-07-24 09:28:11 +0200636static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200637{
638 struct cfq_data *cfqd = q->elevator->elevator_data;
639
640 cfqd->rq_in_driver++;
Jens Axboe25776e32006-06-01 10:12:26 +0200641
642 /*
643 * If the depth is larger 1, it really could be queueing. But lets
644 * make the mark a little higher - idling could still be good for
645 * low queueing, and a low queueing number could also just indicate
646 * a SCSI mid layer like behaviour where limit+1 is often seen.
647 */
648 if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
649 cfqd->hw_tag = 1;
Jens Axboe6d048f52007-04-25 12:44:27 +0200650
651 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
Jens Axboeb4878f22005-10-20 16:42:29 +0200652}
653
Jens Axboe165125e2007-07-24 09:28:11 +0200654static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655{
Jens Axboe22e2c502005-06-27 10:55:12 +0200656 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657
Jens Axboeb4878f22005-10-20 16:42:29 +0200658 WARN_ON(!cfqd->rq_in_driver);
659 cfqd->rq_in_driver--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660}
661
Jens Axboeb4878f22005-10-20 16:42:29 +0200662static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663{
Jens Axboe5e705372006-07-13 12:39:25 +0200664 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200665
Jens Axboe5e705372006-07-13 12:39:25 +0200666 if (cfqq->next_rq == rq)
667 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668
Jens Axboeb4878f22005-10-20 16:42:29 +0200669 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200670 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200671
672 if (rq_is_meta(rq)) {
673 WARN_ON(!cfqq->meta_pending);
674 cfqq->meta_pending--;
675 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676}
677
Jens Axboe165125e2007-07-24 09:28:11 +0200678static int cfq_merge(struct request_queue *q, struct request **req,
679 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680{
681 struct cfq_data *cfqd = q->elevator->elevator_data;
682 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
Jens Axboe206dc692006-03-28 13:03:44 +0200684 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200685 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200686 *req = __rq;
687 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 }
689
690 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691}
692
Jens Axboe165125e2007-07-24 09:28:11 +0200693static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200694 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695{
Jens Axboe21183b02006-07-13 12:33:14 +0200696 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200697 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698
Jens Axboe5e705372006-07-13 12:39:25 +0200699 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701}
702
703static void
Jens Axboe165125e2007-07-24 09:28:11 +0200704cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 struct request *next)
706{
Jens Axboe22e2c502005-06-27 10:55:12 +0200707 /*
708 * reposition in fifo if next is older than rq
709 */
710 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
711 time_before(next->start_time, rq->start_time))
712 list_move(&rq->queuelist, &next->queuelist);
713
Jens Axboeb4878f22005-10-20 16:42:29 +0200714 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200715}
716
Jens Axboe165125e2007-07-24 09:28:11 +0200717static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100718 struct bio *bio)
719{
720 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200721 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100722 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100723
724 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100725 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100726 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200727 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100728 return 0;
729
730 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100731 * Lookup the cfqq that this bio will be queued with. Allow
732 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100733 */
Jens Axboe4ac845a2008-01-24 08:44:49 +0100734 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200735 if (!cic)
736 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100737
Vasily Tarasov91fac312007-04-25 12:29:51 +0200738 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100739 if (cfqq == RQ_CFQQ(rq))
740 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100741
Jens Axboeec8acb62007-01-02 18:32:11 +0100742 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100743}
744
Jens Axboefebffd62008-01-28 13:19:43 +0100745static void __cfq_set_active_queue(struct cfq_data *cfqd,
746 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +0200747{
748 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +0200749 cfqq->slice_end = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200750 cfq_clear_cfqq_must_alloc_slice(cfqq);
751 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100752 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe1afba042007-04-17 12:47:55 +0200753 cfq_clear_cfqq_queue_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200754 }
755
756 cfqd->active_queue = cfqq;
757}
758
759/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100760 * current cfqq expired its slice (or was too idle), select new one
761 */
762static void
763__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200764 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100765{
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100766 if (cfq_cfqq_wait_request(cfqq))
767 del_timer(&cfqd->idle_slice_timer);
768
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100769 cfq_clear_cfqq_must_dispatch(cfqq);
770 cfq_clear_cfqq_wait_request(cfqq);
771
772 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200773 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100774 */
Jens Axboe3c6bd2f2007-01-19 12:06:33 +1100775 if (timed_out && !cfq_cfqq_slice_new(cfqq))
Jens Axboec5b680f2007-01-19 11:56:49 +1100776 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100777
Jens Axboeedd75ff2007-04-19 12:03:34 +0200778 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100779
780 if (cfqq == cfqd->active_queue)
781 cfqd->active_queue = NULL;
782
783 if (cfqd->active_cic) {
784 put_io_context(cfqd->active_cic->ioc);
785 cfqd->active_cic = NULL;
786 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100787}
788
Jens Axboe6084cdd2007-04-23 08:25:00 +0200789static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100790{
791 struct cfq_queue *cfqq = cfqd->active_queue;
792
793 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200794 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100795}
796
Jens Axboe498d3aa22007-04-26 12:54:48 +0200797/*
798 * Get next queue for service. Unless we have a queue preemption,
799 * we'll simply select the first cfqq in the service tree.
800 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200801static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200802{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200803 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
804 return NULL;
805
Jens Axboe08717142008-01-28 11:38:15 +0100806 return cfq_rb_first(&cfqd->service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +0200807}
808
Jens Axboe498d3aa22007-04-26 12:54:48 +0200809/*
810 * Get and set a new active queue for service.
811 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200812static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
813{
814 struct cfq_queue *cfqq;
815
Jens Axboed9e76202007-04-20 14:27:50 +0200816 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200817 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200818 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200819}
820
Jens Axboed9e76202007-04-20 14:27:50 +0200821static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
822 struct request *rq)
823{
824 if (rq->sector >= cfqd->last_position)
825 return rq->sector - cfqd->last_position;
826 else
827 return cfqd->last_position - rq->sector;
828}
829
Jens Axboe6d048f52007-04-25 12:44:27 +0200830static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
831{
832 struct cfq_io_context *cic = cfqd->active_cic;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200833
Jens Axboe6d048f52007-04-25 12:44:27 +0200834 if (!sample_valid(cic->seek_samples))
835 return 0;
836
837 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
838}
839
Jens Axboed9e76202007-04-20 14:27:50 +0200840static int cfq_close_cooperator(struct cfq_data *cfq_data,
841 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200842{
Jens Axboe6d048f52007-04-25 12:44:27 +0200843 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200844 * We should notice if some of the queues are cooperating, eg
845 * working closely on the same area of the disk. In that case,
846 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +0200847 */
Jens Axboed9e76202007-04-20 14:27:50 +0200848 return 0;
Jens Axboe6d048f52007-04-25 12:44:27 +0200849}
850
851#define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
852
853static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200854{
Jens Axboe17926692007-01-19 11:59:30 +1100855 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +0200856 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100857 unsigned long sl;
858
Jens Axboedd67d052006-06-21 09:36:18 +0200859 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +0200860 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +0200861
862 /*
863 * idle is disabled, either manually or by past process history
864 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200865 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
866 return;
867
Jens Axboe22e2c502005-06-27 10:55:12 +0200868 /*
869 * task has exited, don't wait
870 */
Jens Axboe206dc692006-03-28 13:03:44 +0200871 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +0100872 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +0200873 return;
874
875 /*
876 * See if this prio level has a good candidate
877 */
Jens Axboe1afba042007-04-17 12:47:55 +0200878 if (cfq_close_cooperator(cfqd, cfqq) &&
879 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
Jens Axboe6d048f52007-04-25 12:44:27 +0200880 return;
Jens Axboe22e2c502005-06-27 10:55:12 +0200881
Jens Axboe3b181522005-06-27 10:56:24 +0200882 cfq_mark_cfqq_must_dispatch(cfqq);
883 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200884
Jens Axboe206dc692006-03-28 13:03:44 +0200885 /*
886 * we don't want to idle for seeks, but we do want to allow
887 * fair distribution of slice time for a process doing back-to-back
888 * seeks. so allow a little bit of time for him to submit a new rq
889 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200890 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200891 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +0200892 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +0200893
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100894 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895}
896
Jens Axboe498d3aa22007-04-26 12:54:48 +0200897/*
898 * Move request from internal lists to the request queue dispatch list.
899 */
Jens Axboe165125e2007-07-24 09:28:11 +0200900static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901{
Jens Axboe3ed9a292007-04-23 08:33:33 +0200902 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200903 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200904
Jens Axboe5380a102006-07-13 12:37:56 +0200905 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200906 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +0200907 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +0200908
909 if (cfq_cfqq_sync(cfqq))
910 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911}
912
913/*
914 * return expired entry, or NULL to just start from scratch in rbtree
915 */
Jens Axboefebffd62008-01-28 13:19:43 +0100916static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917{
918 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200919 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200920 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921
Jens Axboe3b181522005-06-27 10:56:24 +0200922 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +1100924
925 cfq_mark_cfqq_fifo_expire(cfqq);
926
Jens Axboe89850f72006-07-22 16:48:31 +0200927 if (list_empty(&cfqq->fifo))
928 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929
Jens Axboe6d048f52007-04-25 12:44:27 +0200930 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +0200931 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932
Jens Axboe6d048f52007-04-25 12:44:27 +0200933 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
934 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935
Jens Axboe6d048f52007-04-25 12:44:27 +0200936 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937}
938
Jens Axboe22e2c502005-06-27 10:55:12 +0200939static inline int
940cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
941{
942 const int base_rq = cfqd->cfq_slice_async_rq;
943
944 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
945
946 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
947}
948
949/*
Jens Axboe498d3aa22007-04-26 12:54:48 +0200950 * Select a queue for service. If we have a current active queue,
951 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +0200952 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100953static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200954{
Jens Axboe22e2c502005-06-27 10:55:12 +0200955 struct cfq_queue *cfqq;
956
957 cfqq = cfqd->active_queue;
958 if (!cfqq)
959 goto new_queue;
960
961 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200962 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +0200963 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200964 if (cfq_slice_used(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +0200965 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +0200966
967 /*
Jens Axboe6d048f52007-04-25 12:44:27 +0200968 * The active queue has requests and isn't expired, allow it to
969 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +0200970 */
Jens Axboedd67d052006-06-21 09:36:18 +0200971 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +0200972 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +0200973
974 /*
975 * No requests pending. If the active queue still has requests in
976 * flight or is idling for a new request, allow either of these
977 * conditions to happen (or time out) before selecting a new queue.
978 */
Jens Axboecc197472007-04-20 20:45:39 +0200979 if (timer_pending(&cfqd->idle_slice_timer) ||
980 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +0200981 cfqq = NULL;
982 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +0200983 }
984
Jens Axboe3b181522005-06-27 10:56:24 +0200985expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +0200986 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +0200987new_queue:
988 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200989keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +0200990 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200991}
992
Jens Axboe498d3aa22007-04-26 12:54:48 +0200993/*
994 * Dispatch some requests from cfqq, moving them to the request queue
995 * dispatch list.
996 */
Jens Axboe22e2c502005-06-27 10:55:12 +0200997static int
998__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
999 int max_dispatch)
1000{
1001 int dispatched = 0;
1002
Jens Axboedd67d052006-06-21 09:36:18 +02001003 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001004
1005 do {
Jens Axboe5e705372006-07-13 12:39:25 +02001006 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001007
1008 /*
1009 * follow expired path, else get first next available
1010 */
Jens Axboefe094d92008-01-31 13:08:54 +01001011 rq = cfq_check_fifo(cfqq);
1012 if (rq == NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001013 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001014
1015 /*
1016 * finally, insert request into driver dispatch list
1017 */
Jens Axboe5e705372006-07-13 12:39:25 +02001018 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001019
Jens Axboe22e2c502005-06-27 10:55:12 +02001020 dispatched++;
1021
1022 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +02001023 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1024 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001025 }
1026
Jens Axboedd67d052006-06-21 09:36:18 +02001027 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001028 break;
1029
1030 } while (dispatched < max_dispatch);
1031
1032 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001033 * expire an async queue immediately if it has used up its slice. idle
1034 * queue always expire after 1 dispatch round.
1035 */
Jens Axboea9938002007-04-20 08:55:52 +02001036 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
Jens Axboe20e493a2007-04-23 08:26:36 +02001037 dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboea9938002007-04-20 08:55:52 +02001038 cfq_class_idle(cfqq))) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001039 cfqq->slice_end = jiffies + 1;
Jens Axboe6084cdd2007-04-23 08:25:00 +02001040 cfq_slice_expired(cfqd, 0);
Jens Axboe44f7c162007-01-19 11:51:58 +11001041 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001042
1043 return dispatched;
1044}
1045
Jens Axboefebffd62008-01-28 13:19:43 +01001046static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02001047{
1048 int dispatched = 0;
1049
1050 while (cfqq->next_rq) {
1051 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1052 dispatched++;
1053 }
1054
1055 BUG_ON(!list_empty(&cfqq->fifo));
1056 return dispatched;
1057}
1058
Jens Axboe498d3aa22007-04-26 12:54:48 +02001059/*
1060 * Drain our current requests. Used for barriers and when switching
1061 * io schedulers on-the-fly.
1062 */
Jens Axboed9e76202007-04-20 14:27:50 +02001063static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001064{
Jens Axboe08717142008-01-28 11:38:15 +01001065 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001066 int dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001067
Jens Axboe08717142008-01-28 11:38:15 +01001068 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
Jens Axboed9e76202007-04-20 14:27:50 +02001069 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001070
Jens Axboe6084cdd2007-04-23 08:25:00 +02001071 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001072
1073 BUG_ON(cfqd->busy_queues);
1074
1075 return dispatched;
1076}
1077
Jens Axboe165125e2007-07-24 09:28:11 +02001078static int cfq_dispatch_requests(struct request_queue *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079{
1080 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001081 struct cfq_queue *cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001082 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083
Jens Axboe22e2c502005-06-27 10:55:12 +02001084 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085 return 0;
1086
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001087 if (unlikely(force))
1088 return cfq_forced_dispatch(cfqd);
1089
Jens Axboecaaa5f92006-06-16 11:23:00 +02001090 dispatched = 0;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001091 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001092 int max_dispatch;
1093
Jens Axboe3ed9a292007-04-23 08:33:33 +02001094 max_dispatch = cfqd->cfq_quantum;
1095 if (cfq_class_idle(cfqq))
1096 max_dispatch = 1;
1097
1098 if (cfqq->dispatched >= max_dispatch) {
1099 if (cfqd->busy_queues > 1)
Jens Axboe6d048f52007-04-25 12:44:27 +02001100 break;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001101 if (cfqq->dispatched >= 4 * max_dispatch)
Jens Axboea9938002007-04-20 08:55:52 +02001102 break;
1103 }
Jens Axboe9ede2092007-01-19 12:11:44 +11001104
Jens Axboe3ed9a292007-04-23 08:33:33 +02001105 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1106 break;
1107
Jens Axboe3b181522005-06-27 10:56:24 +02001108 cfq_clear_cfqq_must_dispatch(cfqq);
1109 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001110 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111
Jens Axboecaaa5f92006-06-16 11:23:00 +02001112 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 }
1114
Jens Axboecaaa5f92006-06-16 11:23:00 +02001115 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116}
1117
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118/*
Jens Axboe5e705372006-07-13 12:39:25 +02001119 * task holds one reference to the queue, dropped when task exits. each rq
1120 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 *
1122 * queue lock must be held here.
1123 */
1124static void cfq_put_queue(struct cfq_queue *cfqq)
1125{
Jens Axboe22e2c502005-06-27 10:55:12 +02001126 struct cfq_data *cfqd = cfqq->cfqd;
1127
1128 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129
1130 if (!atomic_dec_and_test(&cfqq->ref))
1131 return;
1132
1133 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001134 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001135 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001137 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001138 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001139 cfq_schedule_dispatch(cfqd);
1140 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001141
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 kmem_cache_free(cfq_pool, cfqq);
1143}
1144
Jens Axboed6de8be2008-05-28 14:46:59 +02001145/*
1146 * Must always be called with the rcu_read_lock() held
1147 */
Jens Axboe07416d22008-05-07 09:17:12 +02001148static void
1149__call_for_each_cic(struct io_context *ioc,
1150 void (*func)(struct io_context *, struct cfq_io_context *))
1151{
1152 struct cfq_io_context *cic;
1153 struct hlist_node *n;
1154
1155 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1156 func(ioc, cic);
1157}
1158
Jens Axboe4ac845a2008-01-24 08:44:49 +01001159/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001160 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001161 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001162static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001163call_for_each_cic(struct io_context *ioc,
1164 void (*func)(struct io_context *, struct cfq_io_context *))
1165{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001166 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02001167 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001168 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001169}
Jens Axboe4ac845a2008-01-24 08:44:49 +01001170
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001171static void cfq_cic_free_rcu(struct rcu_head *head)
1172{
1173 struct cfq_io_context *cic;
1174
1175 cic = container_of(head, struct cfq_io_context, rcu_head);
1176
1177 kmem_cache_free(cfq_ioc_pool, cic);
1178 elv_ioc_count_dec(ioc_count);
1179
1180 if (ioc_gone && !elv_ioc_count_read(ioc_count))
1181 complete(ioc_gone);
1182}
1183
1184static void cfq_cic_free(struct cfq_io_context *cic)
1185{
1186 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001187}
1188
1189static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1190{
1191 unsigned long flags;
1192
1193 BUG_ON(!cic->dead_key);
1194
1195 spin_lock_irqsave(&ioc->lock, flags);
1196 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01001197 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001198 spin_unlock_irqrestore(&ioc->lock, flags);
1199
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001200 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001201}
1202
Jens Axboed6de8be2008-05-28 14:46:59 +02001203/*
1204 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1205 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1206 * and ->trim() which is called with the task lock held
1207 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001208static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001210 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001211 * ioc->refcount is zero here, or we are called from elv_unregister(),
1212 * so no more cic's are allowed to be linked into this ioc. So it
1213 * should be ok to iterate over the known list, we will see all cic's
1214 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001215 */
Jens Axboe07416d22008-05-07 09:17:12 +02001216 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217}
1218
Jens Axboe89850f72006-07-22 16:48:31 +02001219static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1220{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001221 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001222 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001223 cfq_schedule_dispatch(cfqd);
1224 }
Jens Axboe89850f72006-07-22 16:48:31 +02001225
1226 cfq_put_queue(cfqq);
1227}
1228
1229static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1230 struct cfq_io_context *cic)
1231{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001232 struct io_context *ioc = cic->ioc;
1233
Jens Axboefc463792006-08-29 09:05:44 +02001234 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001235
1236 /*
1237 * Make sure key == NULL is seen for dead queues
1238 */
Jens Axboefc463792006-08-29 09:05:44 +02001239 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001240 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02001241 cic->key = NULL;
1242
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001243 if (ioc->ioc_data == cic)
1244 rcu_assign_pointer(ioc->ioc_data, NULL);
1245
Jens Axboe89850f72006-07-22 16:48:31 +02001246 if (cic->cfqq[ASYNC]) {
1247 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1248 cic->cfqq[ASYNC] = NULL;
1249 }
1250
1251 if (cic->cfqq[SYNC]) {
1252 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1253 cic->cfqq[SYNC] = NULL;
1254 }
Jens Axboe89850f72006-07-22 16:48:31 +02001255}
1256
Jens Axboe4ac845a2008-01-24 08:44:49 +01001257static void cfq_exit_single_io_context(struct io_context *ioc,
1258 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001259{
Al Viro478a82b2006-03-18 13:25:24 -05001260 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001261
Jens Axboe89850f72006-07-22 16:48:31 +02001262 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001263 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001264 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02001265
Jens Axboe4ac845a2008-01-24 08:44:49 +01001266 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe89850f72006-07-22 16:48:31 +02001267 __cfq_exit_single_io_context(cfqd, cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001268 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05001269 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001270}
1271
Jens Axboe498d3aa22007-04-26 12:54:48 +02001272/*
1273 * The process that ioc belongs to has exited, we need to clean up
1274 * and put the internal structures we have that belongs to that process.
1275 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001276static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001278 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279}
1280
Jens Axboe22e2c502005-06-27 10:55:12 +02001281static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001282cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283{
Jens Axboeb5deef92006-07-19 23:39:40 +02001284 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285
Christoph Lameter94f60302007-07-17 04:03:29 -07001286 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1287 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001289 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001290 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01001291 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001292 cic->dtor = cfq_free_io_context;
1293 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001294 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001295 }
1296
1297 return cic;
1298}
1299
Jens Axboefd0928d2008-01-24 08:52:45 +01001300static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02001301{
1302 struct task_struct *tsk = current;
1303 int ioprio_class;
1304
Jens Axboe3b181522005-06-27 10:56:24 +02001305 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001306 return;
1307
Jens Axboefd0928d2008-01-24 08:52:45 +01001308 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001309 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01001310 default:
1311 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1312 case IOPRIO_CLASS_NONE:
1313 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02001314 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01001315 */
1316 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02001317 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01001318 break;
1319 case IOPRIO_CLASS_RT:
1320 cfqq->ioprio = task_ioprio(ioc);
1321 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1322 break;
1323 case IOPRIO_CLASS_BE:
1324 cfqq->ioprio = task_ioprio(ioc);
1325 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1326 break;
1327 case IOPRIO_CLASS_IDLE:
1328 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1329 cfqq->ioprio = 7;
1330 cfq_clear_cfqq_idle_window(cfqq);
1331 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02001332 }
1333
1334 /*
1335 * keep track of original prio settings in case we have to temporarily
1336 * elevate the priority of this queue
1337 */
1338 cfqq->org_ioprio = cfqq->ioprio;
1339 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001340 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001341}
1342
Jens Axboefebffd62008-01-28 13:19:43 +01001343static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001344{
Al Viro478a82b2006-03-18 13:25:24 -05001345 struct cfq_data *cfqd = cic->key;
1346 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001347 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001348
Jens Axboecaaa5f92006-06-16 11:23:00 +02001349 if (unlikely(!cfqd))
1350 return;
1351
Jens Axboec1b707d2006-10-30 19:54:23 +01001352 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001353
1354 cfqq = cic->cfqq[ASYNC];
1355 if (cfqq) {
1356 struct cfq_queue *new_cfqq;
Jens Axboefd0928d2008-01-24 08:52:45 +01001357 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001358 if (new_cfqq) {
1359 cic->cfqq[ASYNC] = new_cfqq;
1360 cfq_put_queue(cfqq);
1361 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001362 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001363
1364 cfqq = cic->cfqq[SYNC];
1365 if (cfqq)
1366 cfq_mark_cfqq_prio_changed(cfqq);
1367
Jens Axboec1b707d2006-10-30 19:54:23 +01001368 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001369}
1370
Jens Axboefc463792006-08-29 09:05:44 +02001371static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001373 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02001374 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375}
1376
1377static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001378cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01001379 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001382 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383
1384retry:
Jens Axboe4ac845a2008-01-24 08:44:49 +01001385 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001386 /* cic always exists here */
1387 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388
1389 if (!cfqq) {
1390 if (new_cfqq) {
1391 cfqq = new_cfqq;
1392 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001393 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001394 /*
1395 * Inform the allocator of the fact that we will
1396 * just repeat this allocation if it fails, to allow
1397 * the allocator to do whatever it needs to attempt to
1398 * free memory.
1399 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001401 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1402 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1403 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 spin_lock_irq(cfqd->queue->queue_lock);
1405 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001406 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001407 cfqq = kmem_cache_alloc_node(cfq_pool,
1408 gfp_mask | __GFP_ZERO,
1409 cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001410 if (!cfqq)
1411 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001412 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413
Jens Axboed9e76202007-04-20 14:27:50 +02001414 RB_CLEAR_NODE(&cfqq->rb_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001415 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417 atomic_set(&cfqq->ref, 0);
1418 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001419
Jens Axboe3b181522005-06-27 10:56:24 +02001420 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b037442006-07-28 09:48:51 +02001421 cfq_mark_cfqq_queue_new(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001422
Jens Axboefd0928d2008-01-24 08:52:45 +01001423 cfq_init_prio_data(cfqq, ioc);
Jens Axboe08717142008-01-28 11:38:15 +01001424
1425 if (is_sync) {
1426 if (!cfq_class_idle(cfqq))
1427 cfq_mark_cfqq_idle_window(cfqq);
1428 cfq_mark_cfqq_sync(cfqq);
1429 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 }
1431
1432 if (new_cfqq)
1433 kmem_cache_free(cfq_pool, new_cfqq);
1434
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435out:
1436 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1437 return cfqq;
1438}
1439
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001440static struct cfq_queue **
1441cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1442{
Jens Axboefe094d92008-01-31 13:08:54 +01001443 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001444 case IOPRIO_CLASS_RT:
1445 return &cfqd->async_cfqq[0][ioprio];
1446 case IOPRIO_CLASS_BE:
1447 return &cfqd->async_cfqq[1][ioprio];
1448 case IOPRIO_CLASS_IDLE:
1449 return &cfqd->async_idle_cfqq;
1450 default:
1451 BUG();
1452 }
1453}
1454
Jens Axboe15c31be2007-07-10 13:43:25 +02001455static struct cfq_queue *
Jens Axboefd0928d2008-01-24 08:52:45 +01001456cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02001457 gfp_t gfp_mask)
1458{
Jens Axboefd0928d2008-01-24 08:52:45 +01001459 const int ioprio = task_ioprio(ioc);
1460 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001461 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001462 struct cfq_queue *cfqq = NULL;
1463
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001464 if (!is_sync) {
1465 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1466 cfqq = *async_cfqq;
1467 }
1468
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001469 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01001470 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001471 if (!cfqq)
1472 return NULL;
1473 }
Jens Axboe15c31be2007-07-10 13:43:25 +02001474
1475 /*
1476 * pin the queue now that it's allocated, scheduler exit will prune it
1477 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001478 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001479 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001480 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001481 }
1482
1483 atomic_inc(&cfqq->ref);
1484 return cfqq;
1485}
1486
Jens Axboe498d3aa22007-04-26 12:54:48 +02001487/*
1488 * We drop cfq io contexts lazily, so we may find a dead one.
1489 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001490static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001491cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1492 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001493{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001494 unsigned long flags;
1495
Jens Axboefc463792006-08-29 09:05:44 +02001496 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001497
Jens Axboe4ac845a2008-01-24 08:44:49 +01001498 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02001499
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001500 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001501
1502 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01001503 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001504 spin_unlock_irqrestore(&ioc->lock, flags);
1505
1506 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001507}
1508
Jens Axboee2d74ac2006-03-28 08:59:01 +02001509static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01001510cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001511{
Jens Axboee2d74ac2006-03-28 08:59:01 +02001512 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001513 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001514 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001515
Vasily Tarasov91fac312007-04-25 12:29:51 +02001516 if (unlikely(!ioc))
1517 return NULL;
1518
Jens Axboed6de8be2008-05-28 14:46:59 +02001519 rcu_read_lock();
1520
Jens Axboe597bc482007-04-24 21:23:53 +02001521 /*
1522 * we maintain a last-hit cache, to avoid browsing over the tree
1523 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001524 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02001525 if (cic && cic->key == cfqd) {
1526 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02001527 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001528 }
Jens Axboe597bc482007-04-24 21:23:53 +02001529
Jens Axboe4ac845a2008-01-24 08:44:49 +01001530 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001531 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1532 rcu_read_unlock();
1533 if (!cic)
1534 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001535 /* ->key must be copied to avoid race with cfq_exit_queue() */
1536 k = cic->key;
1537 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001538 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001539 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001540 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001541 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001542
Jens Axboed6de8be2008-05-28 14:46:59 +02001543 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001544 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001545 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001546 break;
1547 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001548
Jens Axboe4ac845a2008-01-24 08:44:49 +01001549 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001550}
1551
Jens Axboe4ac845a2008-01-24 08:44:49 +01001552/*
1553 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1554 * the process specific cfq io context when entered from the block layer.
1555 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1556 */
Jens Axboefebffd62008-01-28 13:19:43 +01001557static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1558 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001559{
Jens Axboe0261d682006-10-30 19:07:48 +01001560 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001561 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001562
Jens Axboe4ac845a2008-01-24 08:44:49 +01001563 ret = radix_tree_preload(gfp_mask);
1564 if (!ret) {
1565 cic->ioc = ioc;
1566 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001567
Jens Axboe4ac845a2008-01-24 08:44:49 +01001568 spin_lock_irqsave(&ioc->lock, flags);
1569 ret = radix_tree_insert(&ioc->radix_root,
1570 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01001571 if (!ret)
1572 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001573 spin_unlock_irqrestore(&ioc->lock, flags);
1574
1575 radix_tree_preload_end();
1576
1577 if (!ret) {
1578 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1579 list_add(&cic->queue_list, &cfqd->cic_list);
1580 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001581 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001582 }
1583
Jens Axboe4ac845a2008-01-24 08:44:49 +01001584 if (ret)
1585 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02001586
Jens Axboe4ac845a2008-01-24 08:44:49 +01001587 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001588}
1589
Jens Axboe22e2c502005-06-27 10:55:12 +02001590/*
1591 * Setup general io context and cfq io context. There can be several cfq
1592 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001593 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001594 */
1595static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001596cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597{
Jens Axboe22e2c502005-06-27 10:55:12 +02001598 struct io_context *ioc = NULL;
1599 struct cfq_io_context *cic;
1600
1601 might_sleep_if(gfp_mask & __GFP_WAIT);
1602
Jens Axboeb5deef92006-07-19 23:39:40 +02001603 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001604 if (!ioc)
1605 return NULL;
1606
Jens Axboe4ac845a2008-01-24 08:44:49 +01001607 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001608 if (cic)
1609 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001610
Jens Axboee2d74ac2006-03-28 08:59:01 +02001611 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1612 if (cic == NULL)
1613 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001614
Jens Axboe4ac845a2008-01-24 08:44:49 +01001615 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1616 goto err_free;
1617
Jens Axboe22e2c502005-06-27 10:55:12 +02001618out:
Jens Axboefc463792006-08-29 09:05:44 +02001619 smp_read_barrier_depends();
1620 if (unlikely(ioc->ioprio_changed))
1621 cfq_ioc_set_ioprio(ioc);
1622
Jens Axboe22e2c502005-06-27 10:55:12 +02001623 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001624err_free:
1625 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001626err:
1627 put_io_context(ioc);
1628 return NULL;
1629}
1630
1631static void
1632cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1633{
Jens Axboeaaf12282007-01-19 11:30:16 +11001634 unsigned long elapsed = jiffies - cic->last_end_request;
1635 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001636
1637 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1638 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1639 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1640}
1641
Jens Axboe206dc692006-03-28 13:03:44 +02001642static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001643cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1644 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001645{
1646 sector_t sdist;
1647 u64 total;
1648
Jens Axboe5e705372006-07-13 12:39:25 +02001649 if (cic->last_request_pos < rq->sector)
1650 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001651 else
Jens Axboe5e705372006-07-13 12:39:25 +02001652 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001653
1654 /*
1655 * Don't allow the seek distance to get too large from the
1656 * odd fragment, pagein, etc
1657 */
1658 if (cic->seek_samples <= 60) /* second&third seek */
1659 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1660 else
1661 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1662
1663 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1664 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1665 total = cic->seek_total + (cic->seek_samples/2);
1666 do_div(total, cic->seek_samples);
1667 cic->seek_mean = (sector_t)total;
1668}
Jens Axboe22e2c502005-06-27 10:55:12 +02001669
1670/*
1671 * Disable idle window if the process thinks too long or seeks so much that
1672 * it doesn't matter
1673 */
1674static void
1675cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1676 struct cfq_io_context *cic)
1677{
Jens Axboe1be92f22007-04-19 14:32:26 +02001678 int enable_idle;
1679
Jens Axboe08717142008-01-28 11:38:15 +01001680 /*
1681 * Don't idle for async or idle io prio class
1682 */
1683 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02001684 return;
1685
1686 enable_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001687
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001688 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Jens Axboecaaa5f92006-06-16 11:23:00 +02001689 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001690 enable_idle = 0;
1691 else if (sample_valid(cic->ttime_samples)) {
1692 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1693 enable_idle = 0;
1694 else
1695 enable_idle = 1;
1696 }
1697
Jens Axboe3b181522005-06-27 10:56:24 +02001698 if (enable_idle)
1699 cfq_mark_cfqq_idle_window(cfqq);
1700 else
1701 cfq_clear_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001702}
1703
Jens Axboe22e2c502005-06-27 10:55:12 +02001704/*
1705 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1706 * no or if we aren't sure, a 1 will cause a preempt.
1707 */
1708static int
1709cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001710 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001711{
Jens Axboe6d048f52007-04-25 12:44:27 +02001712 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001713
Jens Axboe6d048f52007-04-25 12:44:27 +02001714 cfqq = cfqd->active_queue;
1715 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001716 return 0;
1717
Jens Axboe6d048f52007-04-25 12:44:27 +02001718 if (cfq_slice_used(cfqq))
1719 return 1;
1720
1721 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001722 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001723
1724 if (cfq_class_idle(cfqq))
1725 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001726
Jens Axboe22e2c502005-06-27 10:55:12 +02001727 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02001728 * if the new request is sync, but the currently running queue is
1729 * not, let the sync request have priority.
1730 */
Jens Axboe5e705372006-07-13 12:39:25 +02001731 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001732 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001733
Jens Axboe374f84a2006-07-23 01:42:19 +02001734 /*
1735 * So both queues are sync. Let the new request get disk time if
1736 * it's a metadata request and the current queue is doing regular IO.
1737 */
1738 if (rq_is_meta(rq) && !cfqq->meta_pending)
1739 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001740
Jens Axboe1e3335d2007-02-14 19:59:49 +01001741 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1742 return 0;
1743
1744 /*
1745 * if this request is as-good as one we would expect from the
1746 * current cfqq, let it preempt
1747 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001748 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01001749 return 1;
1750
Jens Axboe22e2c502005-06-27 10:55:12 +02001751 return 0;
1752}
1753
1754/*
1755 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1756 * let it have half of its nominal slice.
1757 */
1758static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1759{
Jens Axboe6084cdd2007-04-23 08:25:00 +02001760 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001761
Jens Axboebf572252006-07-19 20:29:12 +02001762 /*
1763 * Put the new queue at the front of the of the current list,
1764 * so we know that it will be selected next.
1765 */
1766 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02001767
1768 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02001769
Jens Axboe44f7c162007-01-19 11:51:58 +11001770 cfqq->slice_end = 0;
1771 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001772}
1773
1774/*
Jens Axboe5e705372006-07-13 12:39:25 +02001775 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001776 * something we should do about it
1777 */
1778static void
Jens Axboe5e705372006-07-13 12:39:25 +02001779cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1780 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001781{
Jens Axboe5e705372006-07-13 12:39:25 +02001782 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001783
Jens Axboe374f84a2006-07-23 01:42:19 +02001784 if (rq_is_meta(rq))
1785 cfqq->meta_pending++;
1786
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001787 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02001788 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001789 cfq_update_idle_window(cfqd, cfqq, cic);
1790
Jens Axboe5e705372006-07-13 12:39:25 +02001791 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001792
1793 if (cfqq == cfqd->active_queue) {
1794 /*
1795 * if we are waiting for a request for this queue, let it rip
1796 * immediately and flag that we must not expire this queue
1797 * just now
1798 */
Jens Axboe3b181522005-06-27 10:56:24 +02001799 if (cfq_cfqq_wait_request(cfqq)) {
1800 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001801 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001802 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001803 }
Jens Axboe5e705372006-07-13 12:39:25 +02001804 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001805 /*
1806 * not the active queue - expire current slice if it is
1807 * idle and has expired it's mean thinktime or this new queue
1808 * has some old slice time left and is of higher priority
1809 */
1810 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001811 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001812 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001813 }
1814}
1815
Jens Axboe165125e2007-07-24 09:28:11 +02001816static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001817{
Jens Axboeb4878f22005-10-20 16:42:29 +02001818 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001819 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001820
Jens Axboefd0928d2008-01-24 08:52:45 +01001821 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822
Jens Axboe5e705372006-07-13 12:39:25 +02001823 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824
Jens Axboe22e2c502005-06-27 10:55:12 +02001825 list_add_tail(&rq->queuelist, &cfqq->fifo);
1826
Jens Axboe5e705372006-07-13 12:39:25 +02001827 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828}
1829
Jens Axboe165125e2007-07-24 09:28:11 +02001830static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831{
Jens Axboe5e705372006-07-13 12:39:25 +02001832 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001833 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001834 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001835 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836
Jens Axboeb4878f22005-10-20 16:42:29 +02001837 now = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838
Jens Axboeb4878f22005-10-20 16:42:29 +02001839 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02001840 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02001841 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02001842 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843
Jens Axboe3ed9a292007-04-23 08:33:33 +02001844 if (cfq_cfqq_sync(cfqq))
1845 cfqd->sync_flight--;
1846
Jens Axboeb4878f22005-10-20 16:42:29 +02001847 if (!cfq_class_idle(cfqq))
1848 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001849
Jens Axboecaaa5f92006-06-16 11:23:00 +02001850 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001851 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001852
1853 /*
1854 * If this is the active queue, check if it needs to be expired,
1855 * or if we want to idle in case it has no pending requests.
1856 */
1857 if (cfqd->active_queue == cfqq) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001858 if (cfq_cfqq_slice_new(cfqq)) {
1859 cfq_set_prio_slice(cfqd, cfqq);
1860 cfq_clear_cfqq_slice_new(cfqq);
1861 }
Jens Axboe08717142008-01-28 11:38:15 +01001862 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02001863 cfq_slice_expired(cfqd, 1);
Jens Axboe6d048f52007-04-25 12:44:27 +02001864 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1865 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001866 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001867
1868 if (!cfqd->rq_in_driver)
1869 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870}
1871
Jens Axboe22e2c502005-06-27 10:55:12 +02001872/*
1873 * we temporarily boost lower priority queues if they are holding fs exclusive
1874 * resources. they are boosted to normal prio (CLASS_BE/4)
1875 */
1876static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877{
Jens Axboe22e2c502005-06-27 10:55:12 +02001878 if (has_fs_excl()) {
1879 /*
1880 * boost idle prio on transactions that would lock out other
1881 * users of the filesystem
1882 */
1883 if (cfq_class_idle(cfqq))
1884 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1885 if (cfqq->ioprio > IOPRIO_NORM)
1886 cfqq->ioprio = IOPRIO_NORM;
1887 } else {
1888 /*
1889 * check if we need to unboost the queue
1890 */
1891 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1892 cfqq->ioprio_class = cfqq->org_ioprio_class;
1893 if (cfqq->ioprio != cfqq->org_ioprio)
1894 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001896}
1897
Jens Axboe89850f72006-07-22 16:48:31 +02001898static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001899{
Jens Axboe3b181522005-06-27 10:56:24 +02001900 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001901 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001902 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001903 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001904 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001905
1906 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02001907}
1908
Jens Axboe165125e2007-07-24 09:28:11 +02001909static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02001910{
1911 struct cfq_data *cfqd = q->elevator->elevator_data;
1912 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001913 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02001914 struct cfq_queue *cfqq;
1915
1916 /*
1917 * don't force setup of a queue from here, as a call to may_queue
1918 * does not necessarily imply that a request actually will be queued.
1919 * so just lookup a possibly existing queue, or return 'may queue'
1920 * if that fails
1921 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001922 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001923 if (!cic)
1924 return ELV_MQUEUE_MAY;
1925
1926 cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02001927 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01001928 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02001929 cfq_prio_boost(cfqq);
1930
Jens Axboe89850f72006-07-22 16:48:31 +02001931 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001932 }
1933
1934 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935}
1936
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937/*
1938 * queue lock held here
1939 */
Jens Axboebb37b942006-12-01 10:42:33 +01001940static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941{
Jens Axboe5e705372006-07-13 12:39:25 +02001942 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943
Jens Axboe5e705372006-07-13 12:39:25 +02001944 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001945 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946
Jens Axboe22e2c502005-06-27 10:55:12 +02001947 BUG_ON(!cfqq->allocated[rw]);
1948 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949
Jens Axboe5e705372006-07-13 12:39:25 +02001950 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02001953 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955 cfq_put_queue(cfqq);
1956 }
1957}
1958
1959/*
Jens Axboe22e2c502005-06-27 10:55:12 +02001960 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001962static int
Jens Axboe165125e2007-07-24 09:28:11 +02001963cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964{
1965 struct cfq_data *cfqd = q->elevator->elevator_data;
1966 struct cfq_io_context *cic;
1967 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01001968 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001969 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970 unsigned long flags;
1971
1972 might_sleep_if(gfp_mask & __GFP_WAIT);
1973
Jens Axboee2d74ac2006-03-28 08:59:01 +02001974 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02001975
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 spin_lock_irqsave(q->queue_lock, flags);
1977
Jens Axboe22e2c502005-06-27 10:55:12 +02001978 if (!cic)
1979 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980
Vasily Tarasov91fac312007-04-25 12:29:51 +02001981 cfqq = cic_to_cfqq(cic, is_sync);
1982 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01001983 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001984
Jens Axboe22e2c502005-06-27 10:55:12 +02001985 if (!cfqq)
1986 goto queue_fail;
1987
Vasily Tarasov91fac312007-04-25 12:29:51 +02001988 cic_set_cfqq(cic, cfqq, is_sync);
1989 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990
1991 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02001992 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001993 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02001994
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995 spin_unlock_irqrestore(q->queue_lock, flags);
1996
Jens Axboe5e705372006-07-13 12:39:25 +02001997 rq->elevator_private = cic;
1998 rq->elevator_private2 = cfqq;
1999 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02002000
Jens Axboe22e2c502005-06-27 10:55:12 +02002001queue_fail:
2002 if (cic)
2003 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02002004
Jens Axboe3b181522005-06-27 10:56:24 +02002005 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 spin_unlock_irqrestore(q->queue_lock, flags);
2007 return 1;
2008}
2009
David Howells65f27f32006-11-22 14:55:48 +00002010static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02002011{
David Howells65f27f32006-11-22 14:55:48 +00002012 struct cfq_data *cfqd =
2013 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02002014 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002015 unsigned long flags;
2016
2017 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02002018 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02002019 spin_unlock_irqrestore(q->queue_lock, flags);
2020}
2021
2022/*
2023 * Timer running if the active_queue is currently idling inside its time slice
2024 */
2025static void cfq_idle_slice_timer(unsigned long data)
2026{
2027 struct cfq_data *cfqd = (struct cfq_data *) data;
2028 struct cfq_queue *cfqq;
2029 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002030 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002031
2032 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2033
Jens Axboefe094d92008-01-31 13:08:54 +01002034 cfqq = cfqd->active_queue;
2035 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002036 timed_out = 0;
2037
Jens Axboe22e2c502005-06-27 10:55:12 +02002038 /*
2039 * expired
2040 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002041 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002042 goto expire;
2043
2044 /*
2045 * only expire and reinvoke request handler, if there are
2046 * other queues with pending requests
2047 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002048 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002049 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002050
2051 /*
2052 * not expired and it has a request pending, let it dispatch
2053 */
Jens Axboedd67d052006-06-21 09:36:18 +02002054 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002055 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002056 goto out_kick;
2057 }
2058 }
2059expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002060 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002061out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002062 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002063out_cont:
2064 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2065}
2066
Jens Axboe3b181522005-06-27 10:56:24 +02002067static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2068{
2069 del_timer_sync(&cfqd->idle_slice_timer);
Oleg Nesterov43108642007-10-23 15:08:18 +02002070 kblockd_flush_work(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002071}
Jens Axboe22e2c502005-06-27 10:55:12 +02002072
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002073static void cfq_put_async_queues(struct cfq_data *cfqd)
2074{
2075 int i;
2076
2077 for (i = 0; i < IOPRIO_BE_NR; i++) {
2078 if (cfqd->async_cfqq[0][i])
2079 cfq_put_queue(cfqd->async_cfqq[0][i]);
2080 if (cfqd->async_cfqq[1][i])
2081 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002082 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002083
2084 if (cfqd->async_idle_cfqq)
2085 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002086}
2087
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088static void cfq_exit_queue(elevator_t *e)
2089{
Jens Axboe22e2c502005-06-27 10:55:12 +02002090 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002091 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002092
Jens Axboe3b181522005-06-27 10:56:24 +02002093 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002094
Al Virod9ff4182006-03-18 13:51:22 -05002095 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002096
Al Virod9ff4182006-03-18 13:51:22 -05002097 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002098 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002099
2100 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002101 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2102 struct cfq_io_context,
2103 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002104
2105 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002106 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002107
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002108 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002109
Al Virod9ff4182006-03-18 13:51:22 -05002110 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002111
2112 cfq_shutdown_timer_wq(cfqd);
2113
Al Viroa90d7422006-03-18 12:05:37 -05002114 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115}
2116
Jens Axboe165125e2007-07-24 09:28:11 +02002117static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118{
2119 struct cfq_data *cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120
Christoph Lameter94f60302007-07-17 04:03:29 -07002121 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002123 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124
Jens Axboecc09e292007-04-26 12:53:50 +02002125 cfqd->service_tree = CFQ_RB_ROOT;
Al Virod9ff4182006-03-18 13:51:22 -05002126 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129
Jens Axboe22e2c502005-06-27 10:55:12 +02002130 init_timer(&cfqd->idle_slice_timer);
2131 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2132 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2133
David Howells65f27f32006-11-22 14:55:48 +00002134 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002135
Oleg Nesterovb70c8642007-11-07 09:46:13 +01002136 cfqd->last_end_request = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002138 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2139 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140 cfqd->cfq_back_max = cfq_back_max;
2141 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002142 cfqd->cfq_slice[0] = cfq_slice_async;
2143 cfqd->cfq_slice[1] = cfq_slice_sync;
2144 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2145 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe3b181522005-06-27 10:56:24 +02002146
Jens Axboebc1c1162006-06-08 08:49:06 +02002147 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148}
2149
2150static void cfq_slab_kill(void)
2151{
Jens Axboed6de8be2008-05-28 14:46:59 +02002152 /*
2153 * Caller already ensured that pending RCU callbacks are completed,
2154 * so we should have no busy allocations at this point.
2155 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 if (cfq_pool)
2157 kmem_cache_destroy(cfq_pool);
2158 if (cfq_ioc_pool)
2159 kmem_cache_destroy(cfq_ioc_pool);
2160}
2161
2162static int __init cfq_slab_setup(void)
2163{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002164 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165 if (!cfq_pool)
2166 goto fail;
2167
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002168 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169 if (!cfq_ioc_pool)
2170 goto fail;
2171
2172 return 0;
2173fail:
2174 cfq_slab_kill();
2175 return -ENOMEM;
2176}
2177
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178/*
2179 * sysfs parts below -->
2180 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181static ssize_t
2182cfq_var_show(unsigned int var, char *page)
2183{
2184 return sprintf(page, "%d\n", var);
2185}
2186
2187static ssize_t
2188cfq_var_store(unsigned int *var, const char *page, size_t count)
2189{
2190 char *p = (char *) page;
2191
2192 *var = simple_strtoul(p, &p, 10);
2193 return count;
2194}
2195
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002197static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002199 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 unsigned int __data = __VAR; \
2201 if (__CONV) \
2202 __data = jiffies_to_msecs(__data); \
2203 return cfq_var_show(__data, (page)); \
2204}
2205SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002206SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2207SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002208SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2209SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002210SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2211SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2212SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2213SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214#undef SHOW_FUNCTION
2215
2216#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Al Viro3d1ab402006-03-18 18:35:43 -05002217static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002219 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 unsigned int __data; \
2221 int ret = cfq_var_store(&__data, (page), count); \
2222 if (__data < (MIN)) \
2223 __data = (MIN); \
2224 else if (__data > (MAX)) \
2225 __data = (MAX); \
2226 if (__CONV) \
2227 *(__PTR) = msecs_to_jiffies(__data); \
2228 else \
2229 *(__PTR) = __data; \
2230 return ret; \
2231}
2232STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002233STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2234 UINT_MAX, 1);
2235STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2236 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002237STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002238STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2239 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002240STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2241STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2242STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01002243STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2244 UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245#undef STORE_FUNCTION
2246
Al Viroe572ec72006-03-18 22:27:18 -05002247#define CFQ_ATTR(name) \
2248 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002249
Al Viroe572ec72006-03-18 22:27:18 -05002250static struct elv_fs_entry cfq_attrs[] = {
2251 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002252 CFQ_ATTR(fifo_expire_sync),
2253 CFQ_ATTR(fifo_expire_async),
2254 CFQ_ATTR(back_seek_max),
2255 CFQ_ATTR(back_seek_penalty),
2256 CFQ_ATTR(slice_sync),
2257 CFQ_ATTR(slice_async),
2258 CFQ_ATTR(slice_async_rq),
2259 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002260 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261};
2262
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263static struct elevator_type iosched_cfq = {
2264 .ops = {
2265 .elevator_merge_fn = cfq_merge,
2266 .elevator_merged_fn = cfq_merged_request,
2267 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002268 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002269 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002271 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 .elevator_deactivate_req_fn = cfq_deactivate_request,
2273 .elevator_queue_empty_fn = cfq_queue_empty,
2274 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002275 .elevator_former_req_fn = elv_rb_former_request,
2276 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 .elevator_set_req_fn = cfq_set_request,
2278 .elevator_put_req_fn = cfq_put_request,
2279 .elevator_may_queue_fn = cfq_may_queue,
2280 .elevator_init_fn = cfq_init_queue,
2281 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002282 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 },
Al Viro3d1ab402006-03-18 18:35:43 -05002284 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 .elevator_name = "cfq",
2286 .elevator_owner = THIS_MODULE,
2287};
2288
2289static int __init cfq_init(void)
2290{
Jens Axboe22e2c502005-06-27 10:55:12 +02002291 /*
2292 * could be 0 on HZ < 1000 setups
2293 */
2294 if (!cfq_slice_async)
2295 cfq_slice_async = 1;
2296 if (!cfq_slice_idle)
2297 cfq_slice_idle = 1;
2298
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 if (cfq_slab_setup())
2300 return -ENOMEM;
2301
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002302 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002304 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305}
2306
2307static void __exit cfq_exit(void)
2308{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002309 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002311 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002312 /* ioc_gone's update must be visible before reading ioc_count */
2313 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02002314
2315 /*
2316 * this also protects us from entering cfq_slab_kill() with
2317 * pending RCU callbacks
2318 */
Jens Axboe4050cf12006-07-19 05:07:12 +02002319 if (elv_ioc_count_read(ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002320 wait_for_completion(ioc_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08002321 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322}
2323
2324module_init(cfq_init);
2325module_exit(cfq_exit);
2326
2327MODULE_AUTHOR("Jens Axboe");
2328MODULE_LICENSE("GPL");
2329MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");