blob: fd7080ed793599127647baafbed009d8e11ef1af [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>
Jens Axboe7b679132008-05-30 12:23:07 +020014#include <linux/blktrace_api.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070015
16/*
17 * tunables
18 */
Jens Axboefe094d92008-01-31 13:08:54 +010019/* max queue in one round of service */
20static const int cfq_quantum = 4;
Arjan van de Ven64100092006-01-06 09:46:02 +010021static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010022/* maximum backwards seek, in KiB */
23static const int cfq_back_max = 16 * 1024;
24/* penalty of a backwards seek */
25static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010026static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020027static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010028static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020029static int cfq_slice_idle = HZ / 125;
Jens Axboe22e2c502005-06-27 10:55:12 +020030
Jens Axboed9e76202007-04-20 14:27:50 +020031/*
Jens Axboe08717142008-01-28 11:38:15 +010032 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020033 */
Jens Axboe08717142008-01-28 11:38:15 +010034#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020035
36/*
37 * below this threshold, we consider thinktime immediate
38 */
39#define CFQ_MIN_TT (2)
40
Jens Axboe22e2c502005-06-27 10:55:12 +020041#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020042#define CFQ_HW_QUEUE_MIN (5)
Jens Axboe22e2c502005-06-27 10:55:12 +020043
Jens Axboefe094d92008-01-31 13:08:54 +010044#define RQ_CIC(rq) \
45 ((struct cfq_io_context *) (rq)->elevator_private)
Jens Axboe7b679132008-05-30 12:23:07 +020046#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
Linus Torvalds1da177e2005-04-16 15:20:36 -070047
Christoph Lametere18b8902006-12-06 20:33:20 -080048static struct kmem_cache *cfq_pool;
49static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070050
Jens Axboe4050cf12006-07-19 05:07:12 +020051static DEFINE_PER_CPU(unsigned long, ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050052static struct completion *ioc_gone;
Jens Axboe9a11b4e2008-05-29 09:32:08 +020053static DEFINE_SPINLOCK(ioc_gone_lock);
Al Viro334e94d2006-03-18 15:05:53 -050054
Jens Axboe22e2c502005-06-27 10:55:12 +020055#define CFQ_PRIO_LISTS IOPRIO_BE_NR
56#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020057#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
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 Axboe6118b702009-06-30 09:34:12 +020074 * Per process-grouping structure
75 */
76struct cfq_queue {
77 /* reference count */
78 atomic_t ref;
79 /* various state flags, see below */
80 unsigned int flags;
81 /* parent cfq_data */
82 struct cfq_data *cfqd;
83 /* service_tree member */
84 struct rb_node rb_node;
85 /* service_tree key */
86 unsigned long rb_key;
87 /* prio tree member */
88 struct rb_node p_node;
89 /* prio tree root we belong to, if any */
90 struct rb_root *p_root;
91 /* sorted list of pending requests */
92 struct rb_root sort_list;
93 /* if fifo isn't expired, next request to serve */
94 struct request *next_rq;
95 /* requests queued in sort_list */
96 int queued[2];
97 /* currently allocated requests */
98 int allocated[2];
99 /* fifo list of requests in sort_list */
100 struct list_head fifo;
101
102 unsigned long slice_end;
103 long slice_resid;
104 unsigned int slice_dispatch;
105
106 /* pending metadata requests */
107 int meta_pending;
108 /* number of requests that are on the dispatch list or inside driver */
109 int dispatched;
110
111 /* io prio of this group */
112 unsigned short ioprio, org_ioprio;
113 unsigned short ioprio_class, org_ioprio_class;
114
115 pid_t pid;
116};
117
118/*
Jens Axboe22e2c502005-06-27 10:55:12 +0200119 * Per block device queue structure
120 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121struct cfq_data {
Jens Axboe165125e2007-07-24 09:28:11 +0200122 struct request_queue *queue;
Jens Axboe22e2c502005-06-27 10:55:12 +0200123
124 /*
125 * rr list of queues with requests and the count of them
126 */
Jens Axboecc09e292007-04-26 12:53:50 +0200127 struct cfq_rb_root service_tree;
Jens Axboea36e71f2009-04-15 12:15:11 +0200128
129 /*
130 * Each priority tree is sorted by next_request position. These
131 * trees are used when determining if two or more queues are
132 * interleaving requests (see cfq_close_cooperator).
133 */
134 struct rb_root prio_trees[CFQ_PRIO_LISTS];
135
Jens Axboe22e2c502005-06-27 10:55:12 +0200136 unsigned int busy_queues;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +0100137 /*
138 * Used to track any pending rt requests so we can pre-empt current
139 * non-RT cfqq in service when this value is non-zero.
140 */
141 unsigned int busy_rt_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +0200142
Jens Axboe22e2c502005-06-27 10:55:12 +0200143 int rq_in_driver;
Jens Axboe3ed9a292007-04-23 08:33:33 +0200144 int sync_flight;
Aaron Carroll45333d52008-08-26 15:52:36 +0200145
146 /*
147 * queue-depth detection
148 */
149 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200150 int hw_tag;
Aaron Carroll45333d52008-08-26 15:52:36 +0200151 int hw_tag_samples;
152 int rq_in_driver_peak;
Jens Axboe22e2c502005-06-27 10:55:12 +0200153
154 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200155 * idle window management
156 */
157 struct timer_list idle_slice_timer;
158 struct work_struct unplug_work;
159
160 struct cfq_queue *active_queue;
161 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200162
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200163 /*
164 * async queue for each priority case
165 */
166 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
167 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200168
Jens Axboe6d048f52007-04-25 12:44:27 +0200169 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 /*
172 * tunables, see top of file
173 */
174 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200175 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 unsigned int cfq_back_penalty;
177 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200178 unsigned int cfq_slice[2];
179 unsigned int cfq_slice_async_rq;
180 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500181
182 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183
Jens Axboe6118b702009-06-30 09:34:12 +0200184 /*
185 * Fallback dummy cfqq for extreme OOM conditions
186 */
187 struct cfq_queue oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188};
189
Jens Axboe3b181522005-06-27 10:56:24 +0200190enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100191 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
192 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200193 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100194 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
195 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100196 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
197 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
198 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100199 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200200 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboea36e71f2009-04-15 12:15:11 +0200201 CFQ_CFQQ_FLAG_coop, /* has done a coop jump of the queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200202};
203
204#define CFQ_CFQQ_FNS(name) \
205static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
206{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100207 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200208} \
209static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
210{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100211 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200212} \
213static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
214{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100215 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200216}
217
218CFQ_CFQQ_FNS(on_rr);
219CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200220CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200221CFQ_CFQQ_FNS(must_alloc);
222CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200223CFQ_CFQQ_FNS(fifo_expire);
224CFQ_CFQQ_FNS(idle_window);
225CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100226CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200227CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200228CFQ_CFQQ_FNS(coop);
Jens Axboe3b181522005-06-27 10:56:24 +0200229#undef CFQ_CFQQ_FNS
230
Jens Axboe7b679132008-05-30 12:23:07 +0200231#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
232 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
233#define cfq_log(cfqd, fmt, args...) \
234 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
235
Jens Axboe165125e2007-07-24 09:28:11 +0200236static void cfq_dispatch_insert(struct request_queue *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200237static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
Jens Axboefd0928d2008-01-24 08:52:45 +0100238 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100239static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200240 struct io_context *);
241
242static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
243 int is_sync)
244{
245 return cic->cfqq[!!is_sync];
246}
247
248static inline void cic_set_cfqq(struct cfq_io_context *cic,
249 struct cfq_queue *cfqq, int is_sync)
250{
251 cic->cfqq[!!is_sync] = cfqq;
252}
253
254/*
255 * We regard a request as SYNC, if it's either a read or has the SYNC bit
256 * set (in which case it could also be direct WRITE).
257 */
258static inline int cfq_bio_sync(struct bio *bio)
259{
260 if (bio_data_dir(bio) == READ || bio_sync(bio))
261 return 1;
262
263 return 0;
264}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700267 * scheduler run of queue, if there are requests pending and no one in the
268 * driver that will restart queueing
269 */
270static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
271{
Jens Axboe7b679132008-05-30 12:23:07 +0200272 if (cfqd->busy_queues) {
273 cfq_log(cfqd, "schedule dispatch");
Jens Axboe18887ad2008-07-28 13:08:45 +0200274 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200275 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700276}
277
Jens Axboe165125e2007-07-24 09:28:11 +0200278static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700279{
280 struct cfq_data *cfqd = q->elevator->elevator_data;
281
Jens Axboeb4878f22005-10-20 16:42:29 +0200282 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700283}
284
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100286 * Scale schedule slice based on io priority. Use the sync time slice only
287 * if a queue is marked sync and has sync io queued. A sync queue with async
288 * io only, should not get full sync slice length.
289 */
Jens Axboed9e76202007-04-20 14:27:50 +0200290static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
291 unsigned short prio)
292{
293 const int base_slice = cfqd->cfq_slice[sync];
294
295 WARN_ON(prio >= IOPRIO_BE_NR);
296
297 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
298}
299
Jens Axboe44f7c162007-01-19 11:51:58 +1100300static inline int
301cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
302{
Jens Axboed9e76202007-04-20 14:27:50 +0200303 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100304}
305
306static inline void
307cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
308{
309 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200310 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100311}
312
313/*
314 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
315 * isn't valid until the first request from the dispatch is activated
316 * and the slice time set.
317 */
318static inline int cfq_slice_used(struct cfq_queue *cfqq)
319{
320 if (cfq_cfqq_slice_new(cfqq))
321 return 0;
322 if (time_before(jiffies, cfqq->slice_end))
323 return 0;
324
325 return 1;
326}
327
328/*
Jens Axboe5e705372006-07-13 12:39:25 +0200329 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200331 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332 */
Jens Axboe5e705372006-07-13 12:39:25 +0200333static struct request *
334cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335{
336 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200338#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
339#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
340 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341
Jens Axboe5e705372006-07-13 12:39:25 +0200342 if (rq1 == NULL || rq1 == rq2)
343 return rq2;
344 if (rq2 == NULL)
345 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200346
Jens Axboe5e705372006-07-13 12:39:25 +0200347 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
348 return rq1;
349 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
350 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200351 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
352 return rq1;
353 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
354 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355
Tejun Heo83096eb2009-05-07 22:24:39 +0900356 s1 = blk_rq_pos(rq1);
357 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358
Jens Axboe6d048f52007-04-25 12:44:27 +0200359 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 /*
362 * by definition, 1KiB is 2 sectors
363 */
364 back_max = cfqd->cfq_back_max * 2;
365
366 /*
367 * Strict one way elevator _except_ in the case where we allow
368 * short backward seeks which are biased as twice the cost of a
369 * similar forward seek.
370 */
371 if (s1 >= last)
372 d1 = s1 - last;
373 else if (s1 + back_max >= last)
374 d1 = (last - s1) * cfqd->cfq_back_penalty;
375 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200376 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377
378 if (s2 >= last)
379 d2 = s2 - last;
380 else if (s2 + back_max >= last)
381 d2 = (last - s2) * cfqd->cfq_back_penalty;
382 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200383 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384
385 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386
Andreas Mohre8a99052006-03-28 08:59:49 +0200387 /*
388 * By doing switch() on the bit mask "wrap" we avoid having to
389 * check two variables for all permutations: --> faster!
390 */
391 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200392 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200393 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200394 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200395 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200396 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200397 else {
398 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200399 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200400 else
Jens Axboe5e705372006-07-13 12:39:25 +0200401 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200402 }
403
404 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200405 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200406 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200407 return rq2;
408 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200409 default:
410 /*
411 * Since both rqs are wrapped,
412 * start with the one that's further behind head
413 * (--> only *one* back seek required),
414 * since back seek takes more time than forward.
415 */
416 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200417 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 else
Jens Axboe5e705372006-07-13 12:39:25 +0200419 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 }
421}
422
Jens Axboe498d3aa22007-04-26 12:54:48 +0200423/*
424 * The below is leftmost cache rbtree addon
425 */
Jens Axboe08717142008-01-28 11:38:15 +0100426static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200427{
428 if (!root->left)
429 root->left = rb_first(&root->rb);
430
Jens Axboe08717142008-01-28 11:38:15 +0100431 if (root->left)
432 return rb_entry(root->left, struct cfq_queue, rb_node);
433
434 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200435}
436
Jens Axboea36e71f2009-04-15 12:15:11 +0200437static void rb_erase_init(struct rb_node *n, struct rb_root *root)
438{
439 rb_erase(n, root);
440 RB_CLEAR_NODE(n);
441}
442
Jens Axboecc09e292007-04-26 12:53:50 +0200443static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
444{
445 if (root->left == n)
446 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200447 rb_erase_init(n, &root->rb);
Jens Axboecc09e292007-04-26 12:53:50 +0200448}
449
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450/*
451 * would be nice to take fifo expire time into account as well
452 */
Jens Axboe5e705372006-07-13 12:39:25 +0200453static struct request *
454cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
455 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456{
Jens Axboe21183b02006-07-13 12:33:14 +0200457 struct rb_node *rbnext = rb_next(&last->rb_node);
458 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200459 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460
Jens Axboe21183b02006-07-13 12:33:14 +0200461 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462
463 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200464 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200465
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200467 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200468 else {
469 rbnext = rb_first(&cfqq->sort_list);
470 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200471 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200472 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473
Jens Axboe21183b02006-07-13 12:33:14 +0200474 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475}
476
Jens Axboed9e76202007-04-20 14:27:50 +0200477static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
478 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479{
Jens Axboed9e76202007-04-20 14:27:50 +0200480 /*
481 * just an approximation, should be ok.
482 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200483 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
484 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200485}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486
Jens Axboe498d3aa22007-04-26 12:54:48 +0200487/*
488 * The cfqd->service_tree holds all pending cfq_queue's that have
489 * requests waiting to be processed. It is sorted in the order that
490 * we will service the queues.
491 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200492static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
493 int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200494{
Jens Axboe08717142008-01-28 11:38:15 +0100495 struct rb_node **p, *parent;
496 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +0200497 unsigned long rb_key;
Jens Axboe498d3aa22007-04-26 12:54:48 +0200498 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200499
Jens Axboe08717142008-01-28 11:38:15 +0100500 if (cfq_class_idle(cfqq)) {
501 rb_key = CFQ_IDLE_DELAY;
502 parent = rb_last(&cfqd->service_tree.rb);
503 if (parent && parent != &cfqq->rb_node) {
504 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
505 rb_key += __cfqq->rb_key;
506 } else
507 rb_key += jiffies;
508 } else if (!add_front) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200509 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
510 rb_key += cfqq->slice_resid;
511 cfqq->slice_resid = 0;
512 } else
513 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200514
515 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100516 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200517 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100518 */
Jens Axboed9e76202007-04-20 14:27:50 +0200519 if (rb_key == cfqq->rb_key)
520 return;
Jens Axboe53b037442006-07-28 09:48:51 +0200521
Jens Axboecc09e292007-04-26 12:53:50 +0200522 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200523 }
Jens Axboed9e76202007-04-20 14:27:50 +0200524
Jens Axboe498d3aa22007-04-26 12:54:48 +0200525 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +0100526 parent = NULL;
527 p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200528 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +0200529 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200530
Jens Axboed9e76202007-04-20 14:27:50 +0200531 parent = *p;
532 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
533
Jens Axboe0c534e02007-04-18 20:01:57 +0200534 /*
535 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200536 * preference to them. IDLE queues goes to the back.
537 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200538 */
539 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200540 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200541 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200542 n = &(*p)->rb_right;
543 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
544 n = &(*p)->rb_left;
545 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
546 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200547 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200548 n = &(*p)->rb_left;
549 else
550 n = &(*p)->rb_right;
551
552 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200553 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200554
555 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200556 }
557
Jens Axboecc09e292007-04-26 12:53:50 +0200558 if (left)
559 cfqd->service_tree.left = &cfqq->rb_node;
560
Jens Axboed9e76202007-04-20 14:27:50 +0200561 cfqq->rb_key = rb_key;
562 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200563 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564}
565
Jens Axboea36e71f2009-04-15 12:15:11 +0200566static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200567cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
568 sector_t sector, struct rb_node **ret_parent,
569 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +0200570{
Jens Axboea36e71f2009-04-15 12:15:11 +0200571 struct rb_node **p, *parent;
572 struct cfq_queue *cfqq = NULL;
573
574 parent = NULL;
575 p = &root->rb_node;
576 while (*p) {
577 struct rb_node **n;
578
579 parent = *p;
580 cfqq = rb_entry(parent, struct cfq_queue, p_node);
581
582 /*
583 * Sort strictly based on sector. Smallest to the left,
584 * largest to the right.
585 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900586 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +0200587 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900588 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +0200589 n = &(*p)->rb_left;
590 else
591 break;
592 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200593 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200594 }
595
596 *ret_parent = parent;
597 if (rb_link)
598 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200599 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +0200600}
601
602static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
603{
Jens Axboea36e71f2009-04-15 12:15:11 +0200604 struct rb_node **p, *parent;
605 struct cfq_queue *__cfqq;
606
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200607 if (cfqq->p_root) {
608 rb_erase(&cfqq->p_node, cfqq->p_root);
609 cfqq->p_root = NULL;
610 }
Jens Axboea36e71f2009-04-15 12:15:11 +0200611
612 if (cfq_class_idle(cfqq))
613 return;
614 if (!cfqq->next_rq)
615 return;
616
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200617 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900618 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
619 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200620 if (!__cfqq) {
621 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200622 rb_insert_color(&cfqq->p_node, cfqq->p_root);
623 } else
624 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200625}
626
Jens Axboe498d3aa22007-04-26 12:54:48 +0200627/*
628 * Update cfqq's position in the service tree.
629 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200630static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200631{
Jens Axboe6d048f52007-04-25 12:44:27 +0200632 /*
633 * Resorting requires the cfqq to be on the RR list already.
634 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200635 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200636 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +0200637 cfq_prio_tree_add(cfqd, cfqq);
638 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200639}
640
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641/*
642 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200643 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 */
Jens Axboefebffd62008-01-28 13:19:43 +0100645static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646{
Jens Axboe7b679132008-05-30 12:23:07 +0200647 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200648 BUG_ON(cfq_cfqq_on_rr(cfqq));
649 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650 cfqd->busy_queues++;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +0100651 if (cfq_class_rt(cfqq))
652 cfqd->busy_rt_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653
Jens Axboeedd75ff2007-04-19 12:03:34 +0200654 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655}
656
Jens Axboe498d3aa22007-04-26 12:54:48 +0200657/*
658 * Called when the cfqq no longer has requests pending, remove it from
659 * the service tree.
660 */
Jens Axboefebffd62008-01-28 13:19:43 +0100661static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662{
Jens Axboe7b679132008-05-30 12:23:07 +0200663 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200664 BUG_ON(!cfq_cfqq_on_rr(cfqq));
665 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666
Jens Axboecc09e292007-04-26 12:53:50 +0200667 if (!RB_EMPTY_NODE(&cfqq->rb_node))
668 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200669 if (cfqq->p_root) {
670 rb_erase(&cfqq->p_node, cfqq->p_root);
671 cfqq->p_root = NULL;
672 }
Jens Axboed9e76202007-04-20 14:27:50 +0200673
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 BUG_ON(!cfqd->busy_queues);
675 cfqd->busy_queues--;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +0100676 if (cfq_class_rt(cfqq))
677 cfqd->busy_rt_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678}
679
680/*
681 * rb tree support functions
682 */
Jens Axboefebffd62008-01-28 13:19:43 +0100683static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684{
Jens Axboe5e705372006-07-13 12:39:25 +0200685 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200686 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200687 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688
Jens Axboeb4878f22005-10-20 16:42:29 +0200689 BUG_ON(!cfqq->queued[sync]);
690 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691
Jens Axboe5e705372006-07-13 12:39:25 +0200692 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693
Jens Axboedd67d052006-06-21 09:36:18 +0200694 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200695 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696}
697
Jens Axboe5e705372006-07-13 12:39:25 +0200698static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699{
Jens Axboe5e705372006-07-13 12:39:25 +0200700 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +0200702 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703
Jens Axboe5380a102006-07-13 12:37:56 +0200704 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705
706 /*
707 * looks a little odd, but the first insert might return an alias.
708 * if that happens, put the alias on the dispatch list
709 */
Jens Axboe21183b02006-07-13 12:33:14 +0200710 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200711 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100712
713 if (!cfq_cfqq_on_rr(cfqq))
714 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200715
716 /*
717 * check if this request is a better next-serve candidate
718 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200719 prev = cfqq->next_rq;
Jens Axboe5044eed2007-04-25 11:53:48 +0200720 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
Jens Axboea36e71f2009-04-15 12:15:11 +0200721
722 /*
723 * adjust priority tree position, if ->next_rq changes
724 */
725 if (prev != cfqq->next_rq)
726 cfq_prio_tree_add(cfqd, cfqq);
727
Jens Axboe5044eed2007-04-25 11:53:48 +0200728 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729}
730
Jens Axboefebffd62008-01-28 13:19:43 +0100731static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732{
Jens Axboe5380a102006-07-13 12:37:56 +0200733 elv_rb_del(&cfqq->sort_list, rq);
734 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200735 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736}
737
Jens Axboe206dc692006-03-28 13:03:44 +0200738static struct request *
739cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740{
Jens Axboe206dc692006-03-28 13:03:44 +0200741 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200742 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200743 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744
Jens Axboe4ac845a2008-01-24 08:44:49 +0100745 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200746 if (!cic)
747 return NULL;
748
749 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200750 if (cfqq) {
751 sector_t sector = bio->bi_sector + bio_sectors(bio);
752
Jens Axboe21183b02006-07-13 12:33:14 +0200753 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200754 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 return NULL;
757}
758
Jens Axboe165125e2007-07-24 09:28:11 +0200759static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200760{
761 struct cfq_data *cfqd = q->elevator->elevator_data;
762
763 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +0200764 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
765 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +0200766
Tejun Heo5b936292009-05-07 22:24:38 +0900767 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200768}
769
Jens Axboe165125e2007-07-24 09:28:11 +0200770static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771{
Jens Axboe22e2c502005-06-27 10:55:12 +0200772 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773
Jens Axboeb4878f22005-10-20 16:42:29 +0200774 WARN_ON(!cfqd->rq_in_driver);
775 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +0200776 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
777 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778}
779
Jens Axboeb4878f22005-10-20 16:42:29 +0200780static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781{
Jens Axboe5e705372006-07-13 12:39:25 +0200782 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200783
Jens Axboe5e705372006-07-13 12:39:25 +0200784 if (cfqq->next_rq == rq)
785 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786
Jens Axboeb4878f22005-10-20 16:42:29 +0200787 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200788 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200789
Aaron Carroll45333d52008-08-26 15:52:36 +0200790 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +0200791 if (rq_is_meta(rq)) {
792 WARN_ON(!cfqq->meta_pending);
793 cfqq->meta_pending--;
794 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795}
796
Jens Axboe165125e2007-07-24 09:28:11 +0200797static int cfq_merge(struct request_queue *q, struct request **req,
798 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799{
800 struct cfq_data *cfqd = q->elevator->elevator_data;
801 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802
Jens Axboe206dc692006-03-28 13:03:44 +0200803 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200804 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200805 *req = __rq;
806 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 }
808
809 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810}
811
Jens Axboe165125e2007-07-24 09:28:11 +0200812static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200813 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814{
Jens Axboe21183b02006-07-13 12:33:14 +0200815 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200816 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817
Jens Axboe5e705372006-07-13 12:39:25 +0200818 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820}
821
822static void
Jens Axboe165125e2007-07-24 09:28:11 +0200823cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 struct request *next)
825{
Jens Axboe22e2c502005-06-27 10:55:12 +0200826 /*
827 * reposition in fifo if next is older than rq
828 */
829 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
830 time_before(next->start_time, rq->start_time))
831 list_move(&rq->queuelist, &next->queuelist);
832
Jens Axboeb4878f22005-10-20 16:42:29 +0200833 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200834}
835
Jens Axboe165125e2007-07-24 09:28:11 +0200836static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100837 struct bio *bio)
838{
839 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200840 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100841 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100842
843 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100844 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100845 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200846 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100847 return 0;
848
849 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100850 * Lookup the cfqq that this bio will be queued with. Allow
851 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100852 */
Jens Axboe4ac845a2008-01-24 08:44:49 +0100853 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200854 if (!cic)
855 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100856
Vasily Tarasov91fac312007-04-25 12:29:51 +0200857 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100858 if (cfqq == RQ_CFQQ(rq))
859 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100860
Jens Axboeec8acb62007-01-02 18:32:11 +0100861 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100862}
863
Jens Axboefebffd62008-01-28 13:19:43 +0100864static void __cfq_set_active_queue(struct cfq_data *cfqd,
865 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +0200866{
867 if (cfqq) {
Jens Axboe7b679132008-05-30 12:23:07 +0200868 cfq_log_cfqq(cfqd, cfqq, "set_active");
Jens Axboe22e2c502005-06-27 10:55:12 +0200869 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +0200870 cfqq->slice_dispatch = 0;
871
Jens Axboe2f5cb732009-04-07 08:51:19 +0200872 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +0200873 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200874 cfq_clear_cfqq_must_alloc_slice(cfqq);
875 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100876 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +0200877
878 del_timer(&cfqd->idle_slice_timer);
Jens Axboe22e2c502005-06-27 10:55:12 +0200879 }
880
881 cfqd->active_queue = cfqq;
882}
883
884/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100885 * current cfqq expired its slice (or was too idle), select new one
886 */
887static void
888__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200889 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100890{
Jens Axboe7b679132008-05-30 12:23:07 +0200891 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
892
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100893 if (cfq_cfqq_wait_request(cfqq))
894 del_timer(&cfqd->idle_slice_timer);
895
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100896 cfq_clear_cfqq_wait_request(cfqq);
897
898 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200899 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100900 */
Jens Axboe7b679132008-05-30 12:23:07 +0200901 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +1100902 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200903 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
904 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100905
Jens Axboeedd75ff2007-04-19 12:03:34 +0200906 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100907
908 if (cfqq == cfqd->active_queue)
909 cfqd->active_queue = NULL;
910
911 if (cfqd->active_cic) {
912 put_io_context(cfqd->active_cic->ioc);
913 cfqd->active_cic = NULL;
914 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100915}
916
Jens Axboe6084cdd2007-04-23 08:25:00 +0200917static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100918{
919 struct cfq_queue *cfqq = cfqd->active_queue;
920
921 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200922 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100923}
924
Jens Axboe498d3aa22007-04-26 12:54:48 +0200925/*
926 * Get next queue for service. Unless we have a queue preemption,
927 * we'll simply select the first cfqq in the service tree.
928 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200929static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200930{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200931 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
932 return NULL;
933
Jens Axboe08717142008-01-28 11:38:15 +0100934 return cfq_rb_first(&cfqd->service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +0200935}
936
Jens Axboe498d3aa22007-04-26 12:54:48 +0200937/*
938 * Get and set a new active queue for service.
939 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200940static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
941 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200942{
Jens Axboea36e71f2009-04-15 12:15:11 +0200943 if (!cfqq) {
944 cfqq = cfq_get_next_queue(cfqd);
945 if (cfqq)
946 cfq_clear_cfqq_coop(cfqq);
947 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200948
Jens Axboe22e2c502005-06-27 10:55:12 +0200949 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200950 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200951}
952
Jens Axboed9e76202007-04-20 14:27:50 +0200953static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
954 struct request *rq)
955{
Tejun Heo83096eb2009-05-07 22:24:39 +0900956 if (blk_rq_pos(rq) >= cfqd->last_position)
957 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +0200958 else
Tejun Heo83096eb2009-05-07 22:24:39 +0900959 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +0200960}
961
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200962#define CIC_SEEK_THR 8 * 1024
963#define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
964
Jens Axboe6d048f52007-04-25 12:44:27 +0200965static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
966{
967 struct cfq_io_context *cic = cfqd->active_cic;
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200968 sector_t sdist = cic->seek_mean;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200969
Jens Axboe6d048f52007-04-25 12:44:27 +0200970 if (!sample_valid(cic->seek_samples))
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200971 sdist = CIC_SEEK_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +0200972
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200973 return cfq_dist_from_last(cfqd, rq) <= sdist;
Jens Axboe6d048f52007-04-25 12:44:27 +0200974}
975
Jens Axboea36e71f2009-04-15 12:15:11 +0200976static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
977 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200978{
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200979 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +0200980 struct rb_node *parent, *node;
981 struct cfq_queue *__cfqq;
982 sector_t sector = cfqd->last_position;
983
984 if (RB_EMPTY_ROOT(root))
985 return NULL;
986
987 /*
988 * First, if we find a request starting at the end of the last
989 * request, choose it.
990 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200991 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +0200992 if (__cfqq)
993 return __cfqq;
994
995 /*
996 * If the exact sector wasn't found, the parent of the NULL leaf
997 * will contain the closest sector.
998 */
999 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1000 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1001 return __cfqq;
1002
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001003 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001004 node = rb_next(&__cfqq->p_node);
1005 else
1006 node = rb_prev(&__cfqq->p_node);
1007 if (!node)
1008 return NULL;
1009
1010 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1011 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1012 return __cfqq;
1013
1014 return NULL;
1015}
1016
1017/*
1018 * cfqd - obvious
1019 * cur_cfqq - passed in so that we don't decide that the current queue is
1020 * closely cooperating with itself.
1021 *
1022 * So, basically we're assuming that that cur_cfqq has dispatched at least
1023 * one request, and that cfqd->last_position reflects a position on the disk
1024 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1025 * assumption.
1026 */
1027static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1028 struct cfq_queue *cur_cfqq,
1029 int probe)
1030{
1031 struct cfq_queue *cfqq;
1032
1033 /*
1034 * A valid cfq_io_context is necessary to compare requests against
1035 * the seek_mean of the current cfqq.
1036 */
1037 if (!cfqd->active_cic)
1038 return NULL;
1039
Jens Axboe6d048f52007-04-25 12:44:27 +02001040 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001041 * We should notice if some of the queues are cooperating, eg
1042 * working closely on the same area of the disk. In that case,
1043 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001044 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001045 cfqq = cfqq_close(cfqd, cur_cfqq);
1046 if (!cfqq)
1047 return NULL;
1048
1049 if (cfq_cfqq_coop(cfqq))
1050 return NULL;
1051
1052 if (!probe)
1053 cfq_mark_cfqq_coop(cfqq);
1054 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001055}
1056
Jens Axboe6d048f52007-04-25 12:44:27 +02001057static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001058{
Jens Axboe17926692007-01-19 11:59:30 +11001059 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001060 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001061 unsigned long sl;
1062
Jens Axboea68bbdd2008-09-24 13:03:33 +02001063 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001064 * SSD device without seek penalty, disable idling. But only do so
1065 * for devices that support queuing, otherwise we still have a problem
1066 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001067 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001068 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001069 return;
1070
Jens Axboedd67d052006-06-21 09:36:18 +02001071 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001072 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001073
1074 /*
1075 * idle is disabled, either manually or by past process history
1076 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001077 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1078 return;
1079
Jens Axboe22e2c502005-06-27 10:55:12 +02001080 /*
Jens Axboe7b679132008-05-30 12:23:07 +02001081 * still requests with the driver, don't idle
1082 */
1083 if (cfqd->rq_in_driver)
1084 return;
1085
1086 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001087 * task has exited, don't wait
1088 */
Jens Axboe206dc692006-03-28 13:03:44 +02001089 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001090 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001091 return;
1092
Jens Axboe3b181522005-06-27 10:56:24 +02001093 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001094
Jens Axboe206dc692006-03-28 13:03:44 +02001095 /*
1096 * we don't want to idle for seeks, but we do want to allow
1097 * fair distribution of slice time for a process doing back-to-back
1098 * seeks. so allow a little bit of time for him to submit a new rq
1099 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001100 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001101 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +02001102 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +02001103
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001104 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe9481ffd2009-04-15 12:14:13 +02001105 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106}
1107
Jens Axboe498d3aa22007-04-26 12:54:48 +02001108/*
1109 * Move request from internal lists to the request queue dispatch list.
1110 */
Jens Axboe165125e2007-07-24 09:28:11 +02001111static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001113 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001114 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001115
Jens Axboe7b679132008-05-30 12:23:07 +02001116 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1117
Jens Axboe5380a102006-07-13 12:37:56 +02001118 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001119 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001120 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001121
1122 if (cfq_cfqq_sync(cfqq))
1123 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124}
1125
1126/*
1127 * return expired entry, or NULL to just start from scratch in rbtree
1128 */
Jens Axboefebffd62008-01-28 13:19:43 +01001129static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130{
1131 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +02001132 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +02001133 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134
Jens Axboe3b181522005-06-27 10:56:24 +02001135 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001137
1138 cfq_mark_cfqq_fifo_expire(cfqq);
1139
Jens Axboe89850f72006-07-22 16:48:31 +02001140 if (list_empty(&cfqq->fifo))
1141 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142
Jens Axboe6d048f52007-04-25 12:44:27 +02001143 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +02001144 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145
Jens Axboe6d048f52007-04-25 12:44:27 +02001146 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
Jens Axboe7b679132008-05-30 12:23:07 +02001147 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148
Jens Axboe7b679132008-05-30 12:23:07 +02001149 cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001150 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151}
1152
Jens Axboe22e2c502005-06-27 10:55:12 +02001153static inline int
1154cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1155{
1156 const int base_rq = cfqd->cfq_slice_async_rq;
1157
1158 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1159
1160 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1161}
1162
1163/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02001164 * Select a queue for service. If we have a current active queue,
1165 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02001166 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001167static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001168{
Jens Axboea36e71f2009-04-15 12:15:11 +02001169 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001170
1171 cfqq = cfqd->active_queue;
1172 if (!cfqq)
1173 goto new_queue;
1174
1175 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001176 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02001177 */
Jens Axboeb0291952009-04-07 11:38:31 +02001178 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001179 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +02001180
1181 /*
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01001182 * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1183 * cfqq.
1184 */
1185 if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1186 /*
1187 * We simulate this as cfqq timed out so that it gets to bank
1188 * the remaining of its time slice.
1189 */
1190 cfq_log_cfqq(cfqd, cfqq, "preempt");
1191 cfq_slice_expired(cfqd, 1);
1192 goto new_queue;
1193 }
1194
1195 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001196 * The active queue has requests and isn't expired, allow it to
1197 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02001198 */
Jens Axboedd67d052006-06-21 09:36:18 +02001199 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001200 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02001201
1202 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02001203 * If another queue has a request waiting within our mean seek
1204 * distance, let it run. The expire code will check for close
1205 * cooperators and put the close queue at the front of the service
1206 * tree.
1207 */
1208 new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1209 if (new_cfqq)
1210 goto expire;
1211
1212 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001213 * No requests pending. If the active queue still has requests in
1214 * flight or is idling for a new request, allow either of these
1215 * conditions to happen (or time out) before selecting a new queue.
1216 */
Jens Axboecc197472007-04-20 20:45:39 +02001217 if (timer_pending(&cfqd->idle_slice_timer) ||
1218 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001219 cfqq = NULL;
1220 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001221 }
1222
Jens Axboe3b181522005-06-27 10:56:24 +02001223expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001224 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001225new_queue:
Jens Axboea36e71f2009-04-15 12:15:11 +02001226 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001227keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001228 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001229}
1230
Jens Axboefebffd62008-01-28 13:19:43 +01001231static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02001232{
1233 int dispatched = 0;
1234
1235 while (cfqq->next_rq) {
1236 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1237 dispatched++;
1238 }
1239
1240 BUG_ON(!list_empty(&cfqq->fifo));
1241 return dispatched;
1242}
1243
Jens Axboe498d3aa22007-04-26 12:54:48 +02001244/*
1245 * Drain our current requests. Used for barriers and when switching
1246 * io schedulers on-the-fly.
1247 */
Jens Axboed9e76202007-04-20 14:27:50 +02001248static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001249{
Jens Axboe08717142008-01-28 11:38:15 +01001250 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001251 int dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001252
Jens Axboe08717142008-01-28 11:38:15 +01001253 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
Jens Axboed9e76202007-04-20 14:27:50 +02001254 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001255
Jens Axboe6084cdd2007-04-23 08:25:00 +02001256 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001257
1258 BUG_ON(cfqd->busy_queues);
1259
Jeff Moyer69237152009-06-12 15:29:30 +02001260 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001261 return dispatched;
1262}
1263
Jens Axboe2f5cb732009-04-07 08:51:19 +02001264/*
1265 * Dispatch a request from cfqq, moving them to the request queue
1266 * dispatch list.
1267 */
1268static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1269{
1270 struct request *rq;
1271
1272 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1273
1274 /*
1275 * follow expired path, else get first next available
1276 */
1277 rq = cfq_check_fifo(cfqq);
1278 if (!rq)
1279 rq = cfqq->next_rq;
1280
1281 /*
1282 * insert request into driver dispatch list
1283 */
1284 cfq_dispatch_insert(cfqd->queue, rq);
1285
1286 if (!cfqd->active_cic) {
1287 struct cfq_io_context *cic = RQ_CIC(rq);
1288
Nikanth Karthikesand9c7d392009-06-10 12:57:06 -07001289 atomic_long_inc(&cic->ioc->refcount);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001290 cfqd->active_cic = cic;
1291 }
1292}
1293
1294/*
1295 * Find the cfqq that we need to service and move a request from that to the
1296 * dispatch list
1297 */
Jens Axboe165125e2007-07-24 09:28:11 +02001298static int cfq_dispatch_requests(struct request_queue *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299{
1300 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001301 struct cfq_queue *cfqq;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001302 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303
Jens Axboe22e2c502005-06-27 10:55:12 +02001304 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 return 0;
1306
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001307 if (unlikely(force))
1308 return cfq_forced_dispatch(cfqd);
1309
Jens Axboe2f5cb732009-04-07 08:51:19 +02001310 cfqq = cfq_select_queue(cfqd);
1311 if (!cfqq)
1312 return 0;
Jens Axboeb4878f22005-10-20 16:42:29 +02001313
Jens Axboe2f5cb732009-04-07 08:51:19 +02001314 /*
1315 * If this is an async queue and we have sync IO in flight, let it wait
1316 */
1317 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1318 return 0;
1319
1320 max_dispatch = cfqd->cfq_quantum;
1321 if (cfq_class_idle(cfqq))
1322 max_dispatch = 1;
1323
1324 /*
1325 * Does this cfqq already have too much IO in flight?
1326 */
1327 if (cfqq->dispatched >= max_dispatch) {
1328 /*
1329 * idle queue must always only have a single IO in flight
1330 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02001331 if (cfq_class_idle(cfqq))
Jens Axboe2f5cb732009-04-07 08:51:19 +02001332 return 0;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001333
Jens Axboe2f5cb732009-04-07 08:51:19 +02001334 /*
1335 * We have other queues, don't allow more IO from this one
1336 */
1337 if (cfqd->busy_queues > 1)
1338 return 0;
Jens Axboe9ede2092007-01-19 12:11:44 +11001339
Jens Axboe2f5cb732009-04-07 08:51:19 +02001340 /*
1341 * we are the only queue, allow up to 4 times of 'quantum'
1342 */
1343 if (cfqq->dispatched >= 4 * max_dispatch)
1344 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 }
1346
Jens Axboe2f5cb732009-04-07 08:51:19 +02001347 /*
1348 * Dispatch a request from this cfqq
1349 */
1350 cfq_dispatch_request(cfqd, cfqq);
1351 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02001352 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001353
1354 /*
1355 * expire an async queue immediately if it has used up its slice. idle
1356 * queue always expire after 1 dispatch round.
1357 */
1358 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1359 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1360 cfq_class_idle(cfqq))) {
1361 cfqq->slice_end = jiffies + 1;
1362 cfq_slice_expired(cfqd, 0);
1363 }
1364
1365 cfq_log(cfqd, "dispatched a request");
1366 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367}
1368
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369/*
Jens Axboe5e705372006-07-13 12:39:25 +02001370 * task holds one reference to the queue, dropped when task exits. each rq
1371 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 *
1373 * queue lock must be held here.
1374 */
1375static void cfq_put_queue(struct cfq_queue *cfqq)
1376{
Jens Axboe22e2c502005-06-27 10:55:12 +02001377 struct cfq_data *cfqd = cfqq->cfqd;
1378
1379 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380
1381 if (!atomic_dec_and_test(&cfqq->ref))
1382 return;
1383
Jens Axboe7b679132008-05-30 12:23:07 +02001384 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001386 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001387 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001389 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001390 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001391 cfq_schedule_dispatch(cfqd);
1392 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001393
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 kmem_cache_free(cfq_pool, cfqq);
1395}
1396
Jens Axboed6de8be2008-05-28 14:46:59 +02001397/*
1398 * Must always be called with the rcu_read_lock() held
1399 */
Jens Axboe07416d22008-05-07 09:17:12 +02001400static void
1401__call_for_each_cic(struct io_context *ioc,
1402 void (*func)(struct io_context *, struct cfq_io_context *))
1403{
1404 struct cfq_io_context *cic;
1405 struct hlist_node *n;
1406
1407 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1408 func(ioc, cic);
1409}
1410
Jens Axboe4ac845a2008-01-24 08:44:49 +01001411/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001412 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001413 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001414static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001415call_for_each_cic(struct io_context *ioc,
1416 void (*func)(struct io_context *, struct cfq_io_context *))
1417{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001418 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02001419 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001420 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001421}
Jens Axboe4ac845a2008-01-24 08:44:49 +01001422
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001423static void cfq_cic_free_rcu(struct rcu_head *head)
1424{
1425 struct cfq_io_context *cic;
1426
1427 cic = container_of(head, struct cfq_io_context, rcu_head);
1428
1429 kmem_cache_free(cfq_ioc_pool, cic);
1430 elv_ioc_count_dec(ioc_count);
1431
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001432 if (ioc_gone) {
1433 /*
1434 * CFQ scheduler is exiting, grab exit lock and check
1435 * the pending io context count. If it hits zero,
1436 * complete ioc_gone and set it back to NULL
1437 */
1438 spin_lock(&ioc_gone_lock);
1439 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1440 complete(ioc_gone);
1441 ioc_gone = NULL;
1442 }
1443 spin_unlock(&ioc_gone_lock);
1444 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001445}
1446
1447static void cfq_cic_free(struct cfq_io_context *cic)
1448{
1449 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001450}
1451
1452static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1453{
1454 unsigned long flags;
1455
1456 BUG_ON(!cic->dead_key);
1457
1458 spin_lock_irqsave(&ioc->lock, flags);
1459 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01001460 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001461 spin_unlock_irqrestore(&ioc->lock, flags);
1462
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001463 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001464}
1465
Jens Axboed6de8be2008-05-28 14:46:59 +02001466/*
1467 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1468 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1469 * and ->trim() which is called with the task lock held
1470 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001471static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001473 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001474 * ioc->refcount is zero here, or we are called from elv_unregister(),
1475 * so no more cic's are allowed to be linked into this ioc. So it
1476 * should be ok to iterate over the known list, we will see all cic's
1477 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001478 */
Jens Axboe07416d22008-05-07 09:17:12 +02001479 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480}
1481
Jens Axboe89850f72006-07-22 16:48:31 +02001482static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1483{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001484 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001485 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001486 cfq_schedule_dispatch(cfqd);
1487 }
Jens Axboe89850f72006-07-22 16:48:31 +02001488
1489 cfq_put_queue(cfqq);
1490}
1491
1492static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1493 struct cfq_io_context *cic)
1494{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001495 struct io_context *ioc = cic->ioc;
1496
Jens Axboefc463792006-08-29 09:05:44 +02001497 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001498
1499 /*
1500 * Make sure key == NULL is seen for dead queues
1501 */
Jens Axboefc463792006-08-29 09:05:44 +02001502 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001503 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02001504 cic->key = NULL;
1505
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001506 if (ioc->ioc_data == cic)
1507 rcu_assign_pointer(ioc->ioc_data, NULL);
1508
Jens Axboeff6657c2009-04-08 10:58:57 +02001509 if (cic->cfqq[BLK_RW_ASYNC]) {
1510 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1511 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001512 }
1513
Jens Axboeff6657c2009-04-08 10:58:57 +02001514 if (cic->cfqq[BLK_RW_SYNC]) {
1515 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1516 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001517 }
Jens Axboe89850f72006-07-22 16:48:31 +02001518}
1519
Jens Axboe4ac845a2008-01-24 08:44:49 +01001520static void cfq_exit_single_io_context(struct io_context *ioc,
1521 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001522{
Al Viro478a82b2006-03-18 13:25:24 -05001523 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001524
Jens Axboe89850f72006-07-22 16:48:31 +02001525 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001526 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001527 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02001528
Jens Axboe4ac845a2008-01-24 08:44:49 +01001529 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01001530
1531 /*
1532 * Ensure we get a fresh copy of the ->key to prevent
1533 * race between exiting task and queue
1534 */
1535 smp_read_barrier_depends();
1536 if (cic->key)
1537 __cfq_exit_single_io_context(cfqd, cic);
1538
Jens Axboe4ac845a2008-01-24 08:44:49 +01001539 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05001540 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001541}
1542
Jens Axboe498d3aa22007-04-26 12:54:48 +02001543/*
1544 * The process that ioc belongs to has exited, we need to clean up
1545 * and put the internal structures we have that belongs to that process.
1546 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001547static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001549 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550}
1551
Jens Axboe22e2c502005-06-27 10:55:12 +02001552static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001553cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554{
Jens Axboeb5deef92006-07-19 23:39:40 +02001555 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556
Christoph Lameter94f60302007-07-17 04:03:29 -07001557 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1558 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001560 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001561 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01001562 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001563 cic->dtor = cfq_free_io_context;
1564 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001565 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566 }
1567
1568 return cic;
1569}
1570
Jens Axboefd0928d2008-01-24 08:52:45 +01001571static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02001572{
1573 struct task_struct *tsk = current;
1574 int ioprio_class;
1575
Jens Axboe3b181522005-06-27 10:56:24 +02001576 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001577 return;
1578
Jens Axboefd0928d2008-01-24 08:52:45 +01001579 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001580 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01001581 default:
1582 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1583 case IOPRIO_CLASS_NONE:
1584 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02001585 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01001586 */
1587 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02001588 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01001589 break;
1590 case IOPRIO_CLASS_RT:
1591 cfqq->ioprio = task_ioprio(ioc);
1592 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1593 break;
1594 case IOPRIO_CLASS_BE:
1595 cfqq->ioprio = task_ioprio(ioc);
1596 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1597 break;
1598 case IOPRIO_CLASS_IDLE:
1599 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1600 cfqq->ioprio = 7;
1601 cfq_clear_cfqq_idle_window(cfqq);
1602 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02001603 }
1604
1605 /*
1606 * keep track of original prio settings in case we have to temporarily
1607 * elevate the priority of this queue
1608 */
1609 cfqq->org_ioprio = cfqq->ioprio;
1610 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001611 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001612}
1613
Jens Axboefebffd62008-01-28 13:19:43 +01001614static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001615{
Al Viro478a82b2006-03-18 13:25:24 -05001616 struct cfq_data *cfqd = cic->key;
1617 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001618 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001619
Jens Axboecaaa5f92006-06-16 11:23:00 +02001620 if (unlikely(!cfqd))
1621 return;
1622
Jens Axboec1b707d2006-10-30 19:54:23 +01001623 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001624
Jens Axboeff6657c2009-04-08 10:58:57 +02001625 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001626 if (cfqq) {
1627 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02001628 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1629 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001630 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02001631 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001632 cfq_put_queue(cfqq);
1633 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001634 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001635
Jens Axboeff6657c2009-04-08 10:58:57 +02001636 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001637 if (cfqq)
1638 cfq_mark_cfqq_prio_changed(cfqq);
1639
Jens Axboec1b707d2006-10-30 19:54:23 +01001640 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001641}
1642
Jens Axboefc463792006-08-29 09:05:44 +02001643static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001645 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02001646 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647}
1648
Jens Axboed5036d72009-06-26 10:44:34 +02001649static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1650 pid_t pid, int is_sync)
1651{
1652 RB_CLEAR_NODE(&cfqq->rb_node);
1653 RB_CLEAR_NODE(&cfqq->p_node);
1654 INIT_LIST_HEAD(&cfqq->fifo);
1655
1656 atomic_set(&cfqq->ref, 0);
1657 cfqq->cfqd = cfqd;
1658
1659 cfq_mark_cfqq_prio_changed(cfqq);
1660
1661 if (is_sync) {
1662 if (!cfq_class_idle(cfqq))
1663 cfq_mark_cfqq_idle_window(cfqq);
1664 cfq_mark_cfqq_sync(cfqq);
1665 }
1666 cfqq->pid = pid;
1667}
1668
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001670cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01001671 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001674 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675
1676retry:
Jens Axboe4ac845a2008-01-24 08:44:49 +01001677 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001678 /* cic always exists here */
1679 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680
Jens Axboe6118b702009-06-30 09:34:12 +02001681 /*
1682 * Always try a new alloc if we fell back to the OOM cfqq
1683 * originally, since it should just be a temporary situation.
1684 */
1685 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1686 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 if (new_cfqq) {
1688 cfqq = new_cfqq;
1689 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001690 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001692 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02001693 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07001694 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02001696 if (new_cfqq)
1697 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001698 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001699 cfqq = kmem_cache_alloc_node(cfq_pool,
1700 gfp_mask | __GFP_ZERO,
1701 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001702 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703
Jens Axboe6118b702009-06-30 09:34:12 +02001704 if (cfqq) {
1705 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1706 cfq_init_prio_data(cfqq, ioc);
1707 cfq_log_cfqq(cfqd, cfqq, "alloced");
1708 } else
1709 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710 }
1711
1712 if (new_cfqq)
1713 kmem_cache_free(cfq_pool, new_cfqq);
1714
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 return cfqq;
1716}
1717
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001718static struct cfq_queue **
1719cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1720{
Jens Axboefe094d92008-01-31 13:08:54 +01001721 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001722 case IOPRIO_CLASS_RT:
1723 return &cfqd->async_cfqq[0][ioprio];
1724 case IOPRIO_CLASS_BE:
1725 return &cfqd->async_cfqq[1][ioprio];
1726 case IOPRIO_CLASS_IDLE:
1727 return &cfqd->async_idle_cfqq;
1728 default:
1729 BUG();
1730 }
1731}
1732
Jens Axboe15c31be2007-07-10 13:43:25 +02001733static struct cfq_queue *
Jens Axboefd0928d2008-01-24 08:52:45 +01001734cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02001735 gfp_t gfp_mask)
1736{
Jens Axboefd0928d2008-01-24 08:52:45 +01001737 const int ioprio = task_ioprio(ioc);
1738 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001739 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001740 struct cfq_queue *cfqq = NULL;
1741
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001742 if (!is_sync) {
1743 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1744 cfqq = *async_cfqq;
1745 }
1746
Jens Axboe6118b702009-06-30 09:34:12 +02001747 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01001748 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02001749
1750 /*
1751 * pin the queue now that it's allocated, scheduler exit will prune it
1752 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001753 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001754 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001755 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001756 }
1757
1758 atomic_inc(&cfqq->ref);
1759 return cfqq;
1760}
1761
Jens Axboe498d3aa22007-04-26 12:54:48 +02001762/*
1763 * We drop cfq io contexts lazily, so we may find a dead one.
1764 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001765static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001766cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1767 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001768{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001769 unsigned long flags;
1770
Jens Axboefc463792006-08-29 09:05:44 +02001771 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001772
Jens Axboe4ac845a2008-01-24 08:44:49 +01001773 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02001774
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001775 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001776
1777 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01001778 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001779 spin_unlock_irqrestore(&ioc->lock, flags);
1780
1781 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001782}
1783
Jens Axboee2d74ac2006-03-28 08:59:01 +02001784static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01001785cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001786{
Jens Axboee2d74ac2006-03-28 08:59:01 +02001787 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001788 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001789 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001790
Vasily Tarasov91fac312007-04-25 12:29:51 +02001791 if (unlikely(!ioc))
1792 return NULL;
1793
Jens Axboed6de8be2008-05-28 14:46:59 +02001794 rcu_read_lock();
1795
Jens Axboe597bc482007-04-24 21:23:53 +02001796 /*
1797 * we maintain a last-hit cache, to avoid browsing over the tree
1798 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001799 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02001800 if (cic && cic->key == cfqd) {
1801 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02001802 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001803 }
Jens Axboe597bc482007-04-24 21:23:53 +02001804
Jens Axboe4ac845a2008-01-24 08:44:49 +01001805 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001806 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1807 rcu_read_unlock();
1808 if (!cic)
1809 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001810 /* ->key must be copied to avoid race with cfq_exit_queue() */
1811 k = cic->key;
1812 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001813 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001814 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001815 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001816 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001817
Jens Axboed6de8be2008-05-28 14:46:59 +02001818 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001819 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001820 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001821 break;
1822 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001823
Jens Axboe4ac845a2008-01-24 08:44:49 +01001824 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001825}
1826
Jens Axboe4ac845a2008-01-24 08:44:49 +01001827/*
1828 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1829 * the process specific cfq io context when entered from the block layer.
1830 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1831 */
Jens Axboefebffd62008-01-28 13:19:43 +01001832static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1833 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001834{
Jens Axboe0261d682006-10-30 19:07:48 +01001835 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001836 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001837
Jens Axboe4ac845a2008-01-24 08:44:49 +01001838 ret = radix_tree_preload(gfp_mask);
1839 if (!ret) {
1840 cic->ioc = ioc;
1841 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001842
Jens Axboe4ac845a2008-01-24 08:44:49 +01001843 spin_lock_irqsave(&ioc->lock, flags);
1844 ret = radix_tree_insert(&ioc->radix_root,
1845 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01001846 if (!ret)
1847 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001848 spin_unlock_irqrestore(&ioc->lock, flags);
1849
1850 radix_tree_preload_end();
1851
1852 if (!ret) {
1853 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1854 list_add(&cic->queue_list, &cfqd->cic_list);
1855 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001856 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001857 }
1858
Jens Axboe4ac845a2008-01-24 08:44:49 +01001859 if (ret)
1860 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02001861
Jens Axboe4ac845a2008-01-24 08:44:49 +01001862 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001863}
1864
Jens Axboe22e2c502005-06-27 10:55:12 +02001865/*
1866 * Setup general io context and cfq io context. There can be several cfq
1867 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001868 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001869 */
1870static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001871cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872{
Jens Axboe22e2c502005-06-27 10:55:12 +02001873 struct io_context *ioc = NULL;
1874 struct cfq_io_context *cic;
1875
1876 might_sleep_if(gfp_mask & __GFP_WAIT);
1877
Jens Axboeb5deef92006-07-19 23:39:40 +02001878 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001879 if (!ioc)
1880 return NULL;
1881
Jens Axboe4ac845a2008-01-24 08:44:49 +01001882 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001883 if (cic)
1884 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001885
Jens Axboee2d74ac2006-03-28 08:59:01 +02001886 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1887 if (cic == NULL)
1888 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001889
Jens Axboe4ac845a2008-01-24 08:44:49 +01001890 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1891 goto err_free;
1892
Jens Axboe22e2c502005-06-27 10:55:12 +02001893out:
Jens Axboefc463792006-08-29 09:05:44 +02001894 smp_read_barrier_depends();
1895 if (unlikely(ioc->ioprio_changed))
1896 cfq_ioc_set_ioprio(ioc);
1897
Jens Axboe22e2c502005-06-27 10:55:12 +02001898 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001899err_free:
1900 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001901err:
1902 put_io_context(ioc);
1903 return NULL;
1904}
1905
1906static void
1907cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1908{
Jens Axboeaaf12282007-01-19 11:30:16 +11001909 unsigned long elapsed = jiffies - cic->last_end_request;
1910 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001911
1912 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1913 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1914 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1915}
1916
Jens Axboe206dc692006-03-28 13:03:44 +02001917static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001918cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1919 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001920{
1921 sector_t sdist;
1922 u64 total;
1923
Jeff Moyer4d00aa42009-04-21 07:25:04 +02001924 if (!cic->last_request_pos)
1925 sdist = 0;
Tejun Heo83096eb2009-05-07 22:24:39 +09001926 else if (cic->last_request_pos < blk_rq_pos(rq))
1927 sdist = blk_rq_pos(rq) - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001928 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001929 sdist = cic->last_request_pos - blk_rq_pos(rq);
Jens Axboe206dc692006-03-28 13:03:44 +02001930
1931 /*
1932 * Don't allow the seek distance to get too large from the
1933 * odd fragment, pagein, etc
1934 */
1935 if (cic->seek_samples <= 60) /* second&third seek */
1936 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1937 else
1938 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1939
1940 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1941 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1942 total = cic->seek_total + (cic->seek_samples/2);
1943 do_div(total, cic->seek_samples);
1944 cic->seek_mean = (sector_t)total;
1945}
Jens Axboe22e2c502005-06-27 10:55:12 +02001946
1947/*
1948 * Disable idle window if the process thinks too long or seeks so much that
1949 * it doesn't matter
1950 */
1951static void
1952cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1953 struct cfq_io_context *cic)
1954{
Jens Axboe7b679132008-05-30 12:23:07 +02001955 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02001956
Jens Axboe08717142008-01-28 11:38:15 +01001957 /*
1958 * Don't idle for async or idle io prio class
1959 */
1960 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02001961 return;
1962
Jens Axboec265a7f2008-06-26 13:49:33 +02001963 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001964
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001965 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Jens Axboecaaa5f92006-06-16 11:23:00 +02001966 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001967 enable_idle = 0;
1968 else if (sample_valid(cic->ttime_samples)) {
1969 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1970 enable_idle = 0;
1971 else
1972 enable_idle = 1;
1973 }
1974
Jens Axboe7b679132008-05-30 12:23:07 +02001975 if (old_idle != enable_idle) {
1976 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1977 if (enable_idle)
1978 cfq_mark_cfqq_idle_window(cfqq);
1979 else
1980 cfq_clear_cfqq_idle_window(cfqq);
1981 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001982}
1983
Jens Axboe22e2c502005-06-27 10:55:12 +02001984/*
1985 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1986 * no or if we aren't sure, a 1 will cause a preempt.
1987 */
1988static int
1989cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001990 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001991{
Jens Axboe6d048f52007-04-25 12:44:27 +02001992 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001993
Jens Axboe6d048f52007-04-25 12:44:27 +02001994 cfqq = cfqd->active_queue;
1995 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001996 return 0;
1997
Jens Axboe6d048f52007-04-25 12:44:27 +02001998 if (cfq_slice_used(cfqq))
1999 return 1;
2000
2001 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02002002 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02002003
2004 if (cfq_class_idle(cfqq))
2005 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002006
Jens Axboe22e2c502005-06-27 10:55:12 +02002007 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02002008 * if the new request is sync, but the currently running queue is
2009 * not, let the sync request have priority.
2010 */
Jens Axboe5e705372006-07-13 12:39:25 +02002011 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002012 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002013
Jens Axboe374f84a2006-07-23 01:42:19 +02002014 /*
2015 * So both queues are sync. Let the new request get disk time if
2016 * it's a metadata request and the current queue is doing regular IO.
2017 */
2018 if (rq_is_meta(rq) && !cfqq->meta_pending)
2019 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002020
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002021 /*
2022 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2023 */
2024 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2025 return 1;
2026
Jens Axboe1e3335d2007-02-14 19:59:49 +01002027 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2028 return 0;
2029
2030 /*
2031 * if this request is as-good as one we would expect from the
2032 * current cfqq, let it preempt
2033 */
Jens Axboe6d048f52007-04-25 12:44:27 +02002034 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01002035 return 1;
2036
Jens Axboe22e2c502005-06-27 10:55:12 +02002037 return 0;
2038}
2039
2040/*
2041 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2042 * let it have half of its nominal slice.
2043 */
2044static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2045{
Jens Axboe7b679132008-05-30 12:23:07 +02002046 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02002047 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02002048
Jens Axboebf572252006-07-19 20:29:12 +02002049 /*
2050 * Put the new queue at the front of the of the current list,
2051 * so we know that it will be selected next.
2052 */
2053 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02002054
2055 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02002056
Jens Axboe44f7c162007-01-19 11:51:58 +11002057 cfqq->slice_end = 0;
2058 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002059}
2060
2061/*
Jens Axboe5e705372006-07-13 12:39:25 +02002062 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02002063 * something we should do about it
2064 */
2065static void
Jens Axboe5e705372006-07-13 12:39:25 +02002066cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2067 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002068{
Jens Axboe5e705372006-07-13 12:39:25 +02002069 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02002070
Aaron Carroll45333d52008-08-26 15:52:36 +02002071 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02002072 if (rq_is_meta(rq))
2073 cfqq->meta_pending++;
2074
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002075 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02002076 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002077 cfq_update_idle_window(cfqd, cfqq, cic);
2078
Tejun Heo83096eb2009-05-07 22:24:39 +09002079 cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002080
2081 if (cfqq == cfqd->active_queue) {
2082 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002083 * Remember that we saw a request from this process, but
2084 * don't start queuing just yet. Otherwise we risk seeing lots
2085 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02002086 * and merging. If the request is already larger than a single
2087 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02002088 * merging is already done. Ditto for a busy system that
2089 * has other work pending, don't risk delaying until the
2090 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02002091 */
Jens Axboed6ceb252009-04-14 14:18:16 +02002092 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02002093 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2094 cfqd->busy_queues > 1) {
Jens Axboed6ceb252009-04-14 14:18:16 +02002095 del_timer(&cfqd->idle_slice_timer);
Tejun Heoa7f55792009-04-23 11:05:17 +09002096 __blk_run_queue(cfqd->queue);
Jens Axboed6ceb252009-04-14 14:18:16 +02002097 }
Jens Axboeb0291952009-04-07 11:38:31 +02002098 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboed6ceb252009-04-14 14:18:16 +02002099 }
Jens Axboe5e705372006-07-13 12:39:25 +02002100 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002101 /*
2102 * not the active queue - expire current slice if it is
2103 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002104 * has some old slice time left and is of higher priority or
2105 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02002106 */
2107 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09002108 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002109 }
2110}
2111
Jens Axboe165125e2007-07-24 09:28:11 +02002112static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002113{
Jens Axboeb4878f22005-10-20 16:42:29 +02002114 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002115 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002116
Jens Axboe7b679132008-05-30 12:23:07 +02002117 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01002118 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119
Jens Axboe5e705372006-07-13 12:39:25 +02002120 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121
Jens Axboe22e2c502005-06-27 10:55:12 +02002122 list_add_tail(&rq->queuelist, &cfqq->fifo);
2123
Jens Axboe5e705372006-07-13 12:39:25 +02002124 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125}
2126
Aaron Carroll45333d52008-08-26 15:52:36 +02002127/*
2128 * Update hw_tag based on peak queue depth over 50 samples under
2129 * sufficient load.
2130 */
2131static void cfq_update_hw_tag(struct cfq_data *cfqd)
2132{
2133 if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
2134 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
2135
2136 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2137 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
2138 return;
2139
2140 if (cfqd->hw_tag_samples++ < 50)
2141 return;
2142
2143 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2144 cfqd->hw_tag = 1;
2145 else
2146 cfqd->hw_tag = 0;
2147
2148 cfqd->hw_tag_samples = 0;
2149 cfqd->rq_in_driver_peak = 0;
2150}
2151
Jens Axboe165125e2007-07-24 09:28:11 +02002152static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153{
Jens Axboe5e705372006-07-13 12:39:25 +02002154 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002155 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02002156 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002157 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158
Jens Axboeb4878f22005-10-20 16:42:29 +02002159 now = jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02002160 cfq_log_cfqq(cfqd, cfqq, "complete");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161
Aaron Carroll45333d52008-08-26 15:52:36 +02002162 cfq_update_hw_tag(cfqd);
2163
Jens Axboeb4878f22005-10-20 16:42:29 +02002164 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02002165 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02002166 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02002167 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168
Jens Axboe3ed9a292007-04-23 08:33:33 +02002169 if (cfq_cfqq_sync(cfqq))
2170 cfqd->sync_flight--;
2171
Jens Axboecaaa5f92006-06-16 11:23:00 +02002172 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02002173 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002174
2175 /*
2176 * If this is the active queue, check if it needs to be expired,
2177 * or if we want to idle in case it has no pending requests.
2178 */
2179 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02002180 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2181
Jens Axboe44f7c162007-01-19 11:51:58 +11002182 if (cfq_cfqq_slice_new(cfqq)) {
2183 cfq_set_prio_slice(cfqd, cfqq);
2184 cfq_clear_cfqq_slice_new(cfqq);
2185 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002186 /*
2187 * If there are no requests waiting in this queue, and
2188 * there are other queues ready to issue requests, AND
2189 * those other queues are issuing requests within our
2190 * mean seek distance, give them a chance to run instead
2191 * of idling.
2192 */
Jens Axboe08717142008-01-28 11:38:15 +01002193 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02002194 cfq_slice_expired(cfqd, 1);
Jens Axboea36e71f2009-04-15 12:15:11 +02002195 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2196 sync && !rq_noidle(rq))
Jens Axboe6d048f52007-04-25 12:44:27 +02002197 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002198 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002199
2200 if (!cfqd->rq_in_driver)
2201 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202}
2203
Jens Axboe22e2c502005-06-27 10:55:12 +02002204/*
2205 * we temporarily boost lower priority queues if they are holding fs exclusive
2206 * resources. they are boosted to normal prio (CLASS_BE/4)
2207 */
2208static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209{
Jens Axboe22e2c502005-06-27 10:55:12 +02002210 if (has_fs_excl()) {
2211 /*
2212 * boost idle prio on transactions that would lock out other
2213 * users of the filesystem
2214 */
2215 if (cfq_class_idle(cfqq))
2216 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2217 if (cfqq->ioprio > IOPRIO_NORM)
2218 cfqq->ioprio = IOPRIO_NORM;
2219 } else {
2220 /*
2221 * check if we need to unboost the queue
2222 */
2223 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2224 cfqq->ioprio_class = cfqq->org_ioprio_class;
2225 if (cfqq->ioprio != cfqq->org_ioprio)
2226 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002228}
2229
Jens Axboe89850f72006-07-22 16:48:31 +02002230static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002231{
Jens Axboe3b181522005-06-27 10:56:24 +02002232 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07002233 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002234 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002235 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02002236 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002237
2238 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02002239}
2240
Jens Axboe165125e2007-07-24 09:28:11 +02002241static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02002242{
2243 struct cfq_data *cfqd = q->elevator->elevator_data;
2244 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002245 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02002246 struct cfq_queue *cfqq;
2247
2248 /*
2249 * don't force setup of a queue from here, as a call to may_queue
2250 * does not necessarily imply that a request actually will be queued.
2251 * so just lookup a possibly existing queue, or return 'may queue'
2252 * if that fails
2253 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002254 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002255 if (!cic)
2256 return ELV_MQUEUE_MAY;
2257
Jens Axboeb0b78f82009-04-08 10:56:08 +02002258 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02002259 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002260 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02002261 cfq_prio_boost(cfqq);
2262
Jens Axboe89850f72006-07-22 16:48:31 +02002263 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002264 }
2265
2266 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267}
2268
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269/*
2270 * queue lock held here
2271 */
Jens Axboebb37b942006-12-01 10:42:33 +01002272static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273{
Jens Axboe5e705372006-07-13 12:39:25 +02002274 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275
Jens Axboe5e705372006-07-13 12:39:25 +02002276 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002277 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278
Jens Axboe22e2c502005-06-27 10:55:12 +02002279 BUG_ON(!cfqq->allocated[rw]);
2280 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281
Jens Axboe5e705372006-07-13 12:39:25 +02002282 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02002285 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 cfq_put_queue(cfqq);
2288 }
2289}
2290
2291/*
Jens Axboe22e2c502005-06-27 10:55:12 +02002292 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 */
Jens Axboe22e2c502005-06-27 10:55:12 +02002294static int
Jens Axboe165125e2007-07-24 09:28:11 +02002295cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296{
2297 struct cfq_data *cfqd = q->elevator->elevator_data;
2298 struct cfq_io_context *cic;
2299 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01002300 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002301 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302 unsigned long flags;
2303
2304 might_sleep_if(gfp_mask & __GFP_WAIT);
2305
Jens Axboee2d74ac2006-03-28 08:59:01 +02002306 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02002307
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 spin_lock_irqsave(q->queue_lock, flags);
2309
Jens Axboe22e2c502005-06-27 10:55:12 +02002310 if (!cic)
2311 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312
Vasily Tarasov91fac312007-04-25 12:29:51 +02002313 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02002314 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002315 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002316 cic_set_cfqq(cic, cfqq, is_sync);
2317 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318
2319 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02002320 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002321 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02002322
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 spin_unlock_irqrestore(q->queue_lock, flags);
2324
Jens Axboe5e705372006-07-13 12:39:25 +02002325 rq->elevator_private = cic;
2326 rq->elevator_private2 = cfqq;
2327 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02002328
Jens Axboe22e2c502005-06-27 10:55:12 +02002329queue_fail:
2330 if (cic)
2331 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02002332
Jens Axboe3b181522005-06-27 10:56:24 +02002333 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02002335 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 return 1;
2337}
2338
David Howells65f27f32006-11-22 14:55:48 +00002339static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02002340{
David Howells65f27f32006-11-22 14:55:48 +00002341 struct cfq_data *cfqd =
2342 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02002343 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002344
Jens Axboe40bb54d2009-04-15 12:11:10 +02002345 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09002346 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02002347 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02002348}
2349
2350/*
2351 * Timer running if the active_queue is currently idling inside its time slice
2352 */
2353static void cfq_idle_slice_timer(unsigned long data)
2354{
2355 struct cfq_data *cfqd = (struct cfq_data *) data;
2356 struct cfq_queue *cfqq;
2357 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002358 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002359
Jens Axboe7b679132008-05-30 12:23:07 +02002360 cfq_log(cfqd, "idle timer fired");
2361
Jens Axboe22e2c502005-06-27 10:55:12 +02002362 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2363
Jens Axboefe094d92008-01-31 13:08:54 +01002364 cfqq = cfqd->active_queue;
2365 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002366 timed_out = 0;
2367
Jens Axboe22e2c502005-06-27 10:55:12 +02002368 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002369 * We saw a request before the queue expired, let it through
2370 */
2371 if (cfq_cfqq_must_dispatch(cfqq))
2372 goto out_kick;
2373
2374 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002375 * expired
2376 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002377 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002378 goto expire;
2379
2380 /*
2381 * only expire and reinvoke request handler, if there are
2382 * other queues with pending requests
2383 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002384 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002385 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002386
2387 /*
2388 * not expired and it has a request pending, let it dispatch
2389 */
Jens Axboe75e50982009-04-07 08:56:14 +02002390 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002391 goto out_kick;
Jens Axboe22e2c502005-06-27 10:55:12 +02002392 }
2393expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002394 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002395out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002396 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002397out_cont:
2398 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2399}
2400
Jens Axboe3b181522005-06-27 10:56:24 +02002401static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2402{
2403 del_timer_sync(&cfqd->idle_slice_timer);
Cheng Renquan64d01dc2008-12-03 12:41:39 +01002404 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002405}
Jens Axboe22e2c502005-06-27 10:55:12 +02002406
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002407static void cfq_put_async_queues(struct cfq_data *cfqd)
2408{
2409 int i;
2410
2411 for (i = 0; i < IOPRIO_BE_NR; i++) {
2412 if (cfqd->async_cfqq[0][i])
2413 cfq_put_queue(cfqd->async_cfqq[0][i]);
2414 if (cfqd->async_cfqq[1][i])
2415 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002416 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002417
2418 if (cfqd->async_idle_cfqq)
2419 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002420}
2421
Jens Axboeb374d182008-10-31 10:05:07 +01002422static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423{
Jens Axboe22e2c502005-06-27 10:55:12 +02002424 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002425 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002426
Jens Axboe3b181522005-06-27 10:56:24 +02002427 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002428
Al Virod9ff4182006-03-18 13:51:22 -05002429 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002430
Al Virod9ff4182006-03-18 13:51:22 -05002431 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002432 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002433
2434 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002435 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2436 struct cfq_io_context,
2437 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002438
2439 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002440 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002441
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002442 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002443
Al Virod9ff4182006-03-18 13:51:22 -05002444 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002445
2446 cfq_shutdown_timer_wq(cfqd);
2447
Al Viroa90d7422006-03-18 12:05:37 -05002448 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449}
2450
Jens Axboe165125e2007-07-24 09:28:11 +02002451static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452{
2453 struct cfq_data *cfqd;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002454 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455
Christoph Lameter94f60302007-07-17 04:03:29 -07002456 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002458 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459
Jens Axboecc09e292007-04-26 12:53:50 +02002460 cfqd->service_tree = CFQ_RB_ROOT;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002461
2462 /*
2463 * Not strictly needed (since RB_ROOT just clears the node and we
2464 * zeroed cfqd on alloc), but better be safe in case someone decides
2465 * to add magic to the rb code
2466 */
2467 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2468 cfqd->prio_trees[i] = RB_ROOT;
2469
Jens Axboe6118b702009-06-30 09:34:12 +02002470 /*
2471 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2472 * Grab a permanent reference to it, so that the normal code flow
2473 * will not attempt to free it.
2474 */
2475 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2476 atomic_inc(&cfqd->oom_cfqq.ref);
2477
Al Virod9ff4182006-03-18 13:51:22 -05002478 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481
Jens Axboe22e2c502005-06-27 10:55:12 +02002482 init_timer(&cfqd->idle_slice_timer);
2483 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2484 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2485
David Howells65f27f32006-11-22 14:55:48 +00002486 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002487
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002489 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2490 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 cfqd->cfq_back_max = cfq_back_max;
2492 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002493 cfqd->cfq_slice[0] = cfq_slice_async;
2494 cfqd->cfq_slice[1] = cfq_slice_sync;
2495 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2496 cfqd->cfq_slice_idle = cfq_slice_idle;
Aaron Carroll45333d52008-08-26 15:52:36 +02002497 cfqd->hw_tag = 1;
Jens Axboe3b181522005-06-27 10:56:24 +02002498
Jens Axboebc1c1162006-06-08 08:49:06 +02002499 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500}
2501
2502static void cfq_slab_kill(void)
2503{
Jens Axboed6de8be2008-05-28 14:46:59 +02002504 /*
2505 * Caller already ensured that pending RCU callbacks are completed,
2506 * so we should have no busy allocations at this point.
2507 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508 if (cfq_pool)
2509 kmem_cache_destroy(cfq_pool);
2510 if (cfq_ioc_pool)
2511 kmem_cache_destroy(cfq_ioc_pool);
2512}
2513
2514static int __init cfq_slab_setup(void)
2515{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002516 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517 if (!cfq_pool)
2518 goto fail;
2519
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002520 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521 if (!cfq_ioc_pool)
2522 goto fail;
2523
2524 return 0;
2525fail:
2526 cfq_slab_kill();
2527 return -ENOMEM;
2528}
2529
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530/*
2531 * sysfs parts below -->
2532 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533static ssize_t
2534cfq_var_show(unsigned int var, char *page)
2535{
2536 return sprintf(page, "%d\n", var);
2537}
2538
2539static ssize_t
2540cfq_var_store(unsigned int *var, const char *page, size_t count)
2541{
2542 char *p = (char *) page;
2543
2544 *var = simple_strtoul(p, &p, 10);
2545 return count;
2546}
2547
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002549static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002551 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 unsigned int __data = __VAR; \
2553 if (__CONV) \
2554 __data = jiffies_to_msecs(__data); \
2555 return cfq_var_show(__data, (page)); \
2556}
2557SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002558SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2559SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002560SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2561SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002562SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2563SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2564SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2565SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566#undef SHOW_FUNCTION
2567
2568#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002569static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002571 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 unsigned int __data; \
2573 int ret = cfq_var_store(&__data, (page), count); \
2574 if (__data < (MIN)) \
2575 __data = (MIN); \
2576 else if (__data > (MAX)) \
2577 __data = (MAX); \
2578 if (__CONV) \
2579 *(__PTR) = msecs_to_jiffies(__data); \
2580 else \
2581 *(__PTR) = __data; \
2582 return ret; \
2583}
2584STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002585STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2586 UINT_MAX, 1);
2587STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2588 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002589STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002590STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2591 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002592STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2593STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2594STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01002595STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2596 UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597#undef STORE_FUNCTION
2598
Al Viroe572ec72006-03-18 22:27:18 -05002599#define CFQ_ATTR(name) \
2600 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002601
Al Viroe572ec72006-03-18 22:27:18 -05002602static struct elv_fs_entry cfq_attrs[] = {
2603 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002604 CFQ_ATTR(fifo_expire_sync),
2605 CFQ_ATTR(fifo_expire_async),
2606 CFQ_ATTR(back_seek_max),
2607 CFQ_ATTR(back_seek_penalty),
2608 CFQ_ATTR(slice_sync),
2609 CFQ_ATTR(slice_async),
2610 CFQ_ATTR(slice_async_rq),
2611 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002612 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613};
2614
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615static struct elevator_type iosched_cfq = {
2616 .ops = {
2617 .elevator_merge_fn = cfq_merge,
2618 .elevator_merged_fn = cfq_merged_request,
2619 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002620 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002621 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002623 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624 .elevator_deactivate_req_fn = cfq_deactivate_request,
2625 .elevator_queue_empty_fn = cfq_queue_empty,
2626 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002627 .elevator_former_req_fn = elv_rb_former_request,
2628 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 .elevator_set_req_fn = cfq_set_request,
2630 .elevator_put_req_fn = cfq_put_request,
2631 .elevator_may_queue_fn = cfq_may_queue,
2632 .elevator_init_fn = cfq_init_queue,
2633 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002634 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635 },
Al Viro3d1ab402006-03-18 18:35:43 -05002636 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 .elevator_name = "cfq",
2638 .elevator_owner = THIS_MODULE,
2639};
2640
2641static int __init cfq_init(void)
2642{
Jens Axboe22e2c502005-06-27 10:55:12 +02002643 /*
2644 * could be 0 on HZ < 1000 setups
2645 */
2646 if (!cfq_slice_async)
2647 cfq_slice_async = 1;
2648 if (!cfq_slice_idle)
2649 cfq_slice_idle = 1;
2650
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651 if (cfq_slab_setup())
2652 return -ENOMEM;
2653
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002654 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002656 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657}
2658
2659static void __exit cfq_exit(void)
2660{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002661 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002663 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002664 /* ioc_gone's update must be visible before reading ioc_count */
2665 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02002666
2667 /*
2668 * this also protects us from entering cfq_slab_kill() with
2669 * pending RCU callbacks
2670 */
Jens Axboe4050cf12006-07-19 05:07:12 +02002671 if (elv_ioc_count_read(ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002672 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08002673 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674}
2675
2676module_init(cfq_init);
2677module_exit(cfq_exit);
2678
2679MODULE_AUTHOR("Jens Axboe");
2680MODULE_LICENSE("GPL");
2681MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");