blob: a55a9bd75bd1baf616a3a1b7118acaeee328759f [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 Axboe22e2c502005-06-27 10:55:12 +020074 * Per block device queue structure
75 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070076struct cfq_data {
Jens Axboe165125e2007-07-24 09:28:11 +020077 struct request_queue *queue;
Jens Axboe22e2c502005-06-27 10:55:12 +020078
79 /*
80 * rr list of queues with requests and the count of them
81 */
Jens Axboecc09e292007-04-26 12:53:50 +020082 struct cfq_rb_root service_tree;
Jens Axboea36e71f2009-04-15 12:15:11 +020083
84 /*
85 * Each priority tree is sorted by next_request position. These
86 * trees are used when determining if two or more queues are
87 * interleaving requests (see cfq_close_cooperator).
88 */
89 struct rb_root prio_trees[CFQ_PRIO_LISTS];
90
Jens Axboe22e2c502005-06-27 10:55:12 +020091 unsigned int busy_queues;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +010092 /*
93 * Used to track any pending rt requests so we can pre-empt current
94 * non-RT cfqq in service when this value is non-zero.
95 */
96 unsigned int busy_rt_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +020097
Jens Axboe22e2c502005-06-27 10:55:12 +020098 int rq_in_driver;
Jens Axboe3ed9a292007-04-23 08:33:33 +020099 int sync_flight;
Aaron Carroll45333d52008-08-26 15:52:36 +0200100
101 /*
102 * queue-depth detection
103 */
104 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200105 int hw_tag;
Aaron Carroll45333d52008-08-26 15:52:36 +0200106 int hw_tag_samples;
107 int rq_in_driver_peak;
Jens Axboe22e2c502005-06-27 10:55:12 +0200108
109 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200110 * idle window management
111 */
112 struct timer_list idle_slice_timer;
113 struct work_struct unplug_work;
114
115 struct cfq_queue *active_queue;
116 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200117
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200118 /*
119 * async queue for each priority case
120 */
121 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
122 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200123
Jens Axboe6d048f52007-04-25 12:44:27 +0200124 sector_t last_position;
Jens Axboe22e2c502005-06-27 10:55:12 +0200125 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127 /*
128 * tunables, see top of file
129 */
130 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200131 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 unsigned int cfq_back_penalty;
133 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200134 unsigned int cfq_slice[2];
135 unsigned int cfq_slice_async_rq;
136 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500137
138 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139};
140
Jens Axboe22e2c502005-06-27 10:55:12 +0200141/*
142 * Per process-grouping structure
143 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144struct cfq_queue {
145 /* reference count */
146 atomic_t ref;
Richard Kennedybe754d22008-05-23 06:52:00 +0200147 /* various state flags, see below */
148 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 /* parent cfq_data */
150 struct cfq_data *cfqd;
Jens Axboed9e76202007-04-20 14:27:50 +0200151 /* service_tree member */
152 struct rb_node rb_node;
153 /* service_tree key */
154 unsigned long rb_key;
Jens Axboea36e71f2009-04-15 12:15:11 +0200155 /* prio tree member */
156 struct rb_node p_node;
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200157 /* prio tree root we belong to, if any */
158 struct rb_root *p_root;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 /* sorted list of pending requests */
160 struct rb_root sort_list;
161 /* if fifo isn't expired, next request to serve */
Jens Axboe5e705372006-07-13 12:39:25 +0200162 struct request *next_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163 /* requests queued in sort_list */
164 int queued[2];
165 /* currently allocated requests */
166 int allocated[2];
167 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200168 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169
Jens Axboe22e2c502005-06-27 10:55:12 +0200170 unsigned long slice_end;
Jens Axboec5b680f2007-01-19 11:56:49 +1100171 long slice_resid;
Jens Axboe2f5cb732009-04-07 08:51:19 +0200172 unsigned int slice_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173
Richard Kennedybe754d22008-05-23 06:52:00 +0200174 /* pending metadata requests */
175 int meta_pending;
Jens Axboe6d048f52007-04-25 12:44:27 +0200176 /* number of requests that are on the dispatch list or inside driver */
177 int dispatched;
Jens Axboe22e2c502005-06-27 10:55:12 +0200178
179 /* io prio of this group */
180 unsigned short ioprio, org_ioprio;
181 unsigned short ioprio_class, org_ioprio_class;
182
Jens Axboe7b679132008-05-30 12:23:07 +0200183 pid_t pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184};
185
Jens Axboe3b181522005-06-27 10:56:24 +0200186enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100187 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
188 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200189 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100190 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
191 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100192 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
193 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
194 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100195 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200196 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboea36e71f2009-04-15 12:15:11 +0200197 CFQ_CFQQ_FLAG_coop, /* has done a coop jump of the queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200198};
199
200#define CFQ_CFQQ_FNS(name) \
201static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
202{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100203 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200204} \
205static inline void cfq_clear_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 int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
210{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100211 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200212}
213
214CFQ_CFQQ_FNS(on_rr);
215CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200216CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200217CFQ_CFQQ_FNS(must_alloc);
218CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200219CFQ_CFQQ_FNS(fifo_expire);
220CFQ_CFQQ_FNS(idle_window);
221CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100222CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200223CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200224CFQ_CFQQ_FNS(coop);
Jens Axboe3b181522005-06-27 10:56:24 +0200225#undef CFQ_CFQQ_FNS
226
Jens Axboe7b679132008-05-30 12:23:07 +0200227#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
228 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
229#define cfq_log(cfqd, fmt, args...) \
230 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
231
Jens Axboe165125e2007-07-24 09:28:11 +0200232static void cfq_dispatch_insert(struct request_queue *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200233static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
Jens Axboefd0928d2008-01-24 08:52:45 +0100234 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100235static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200236 struct io_context *);
237
238static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
239 int is_sync)
240{
241 return cic->cfqq[!!is_sync];
242}
243
244static inline void cic_set_cfqq(struct cfq_io_context *cic,
245 struct cfq_queue *cfqq, int is_sync)
246{
247 cic->cfqq[!!is_sync] = cfqq;
248}
249
250/*
251 * We regard a request as SYNC, if it's either a read or has the SYNC bit
252 * set (in which case it could also be direct WRITE).
253 */
254static inline int cfq_bio_sync(struct bio *bio)
255{
256 if (bio_data_dir(bio) == READ || bio_sync(bio))
257 return 1;
258
259 return 0;
260}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700263 * scheduler run of queue, if there are requests pending and no one in the
264 * driver that will restart queueing
265 */
266static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
267{
Jens Axboe7b679132008-05-30 12:23:07 +0200268 if (cfqd->busy_queues) {
269 cfq_log(cfqd, "schedule dispatch");
Jens Axboe18887ad2008-07-28 13:08:45 +0200270 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200271 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700272}
273
Jens Axboe165125e2007-07-24 09:28:11 +0200274static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700275{
276 struct cfq_data *cfqd = q->elevator->elevator_data;
277
Jens Axboeb4878f22005-10-20 16:42:29 +0200278 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700279}
280
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100282 * Scale schedule slice based on io priority. Use the sync time slice only
283 * if a queue is marked sync and has sync io queued. A sync queue with async
284 * io only, should not get full sync slice length.
285 */
Jens Axboed9e76202007-04-20 14:27:50 +0200286static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
287 unsigned short prio)
288{
289 const int base_slice = cfqd->cfq_slice[sync];
290
291 WARN_ON(prio >= IOPRIO_BE_NR);
292
293 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
294}
295
Jens Axboe44f7c162007-01-19 11:51:58 +1100296static inline int
297cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
298{
Jens Axboed9e76202007-04-20 14:27:50 +0200299 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100300}
301
302static inline void
303cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
304{
305 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200306 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100307}
308
309/*
310 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
311 * isn't valid until the first request from the dispatch is activated
312 * and the slice time set.
313 */
314static inline int cfq_slice_used(struct cfq_queue *cfqq)
315{
316 if (cfq_cfqq_slice_new(cfqq))
317 return 0;
318 if (time_before(jiffies, cfqq->slice_end))
319 return 0;
320
321 return 1;
322}
323
324/*
Jens Axboe5e705372006-07-13 12:39:25 +0200325 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200327 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328 */
Jens Axboe5e705372006-07-13 12:39:25 +0200329static struct request *
330cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331{
332 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200334#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
335#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
336 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337
Jens Axboe5e705372006-07-13 12:39:25 +0200338 if (rq1 == NULL || rq1 == rq2)
339 return rq2;
340 if (rq2 == NULL)
341 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200342
Jens Axboe5e705372006-07-13 12:39:25 +0200343 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
344 return rq1;
345 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
346 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200347 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
348 return rq1;
349 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
350 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351
Jens Axboe5e705372006-07-13 12:39:25 +0200352 s1 = rq1->sector;
353 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354
Jens Axboe6d048f52007-04-25 12:44:27 +0200355 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357 /*
358 * by definition, 1KiB is 2 sectors
359 */
360 back_max = cfqd->cfq_back_max * 2;
361
362 /*
363 * Strict one way elevator _except_ in the case where we allow
364 * short backward seeks which are biased as twice the cost of a
365 * similar forward seek.
366 */
367 if (s1 >= last)
368 d1 = s1 - last;
369 else if (s1 + back_max >= last)
370 d1 = (last - s1) * cfqd->cfq_back_penalty;
371 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200372 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
374 if (s2 >= last)
375 d2 = s2 - last;
376 else if (s2 + back_max >= last)
377 d2 = (last - s2) * cfqd->cfq_back_penalty;
378 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200379 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380
381 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382
Andreas Mohre8a99052006-03-28 08:59:49 +0200383 /*
384 * By doing switch() on the bit mask "wrap" we avoid having to
385 * check two variables for all permutations: --> faster!
386 */
387 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200388 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200389 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200390 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200391 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200392 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200393 else {
394 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200395 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200396 else
Jens Axboe5e705372006-07-13 12:39:25 +0200397 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200398 }
399
400 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200401 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200402 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200403 return rq2;
404 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200405 default:
406 /*
407 * Since both rqs are wrapped,
408 * start with the one that's further behind head
409 * (--> only *one* back seek required),
410 * since back seek takes more time than forward.
411 */
412 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200413 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 else
Jens Axboe5e705372006-07-13 12:39:25 +0200415 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 }
417}
418
Jens Axboe498d3aa2007-04-26 12:54:48 +0200419/*
420 * The below is leftmost cache rbtree addon
421 */
Jens Axboe08717142008-01-28 11:38:15 +0100422static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200423{
424 if (!root->left)
425 root->left = rb_first(&root->rb);
426
Jens Axboe08717142008-01-28 11:38:15 +0100427 if (root->left)
428 return rb_entry(root->left, struct cfq_queue, rb_node);
429
430 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200431}
432
Jens Axboea36e71f2009-04-15 12:15:11 +0200433static void rb_erase_init(struct rb_node *n, struct rb_root *root)
434{
435 rb_erase(n, root);
436 RB_CLEAR_NODE(n);
437}
438
Jens Axboecc09e292007-04-26 12:53:50 +0200439static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
440{
441 if (root->left == n)
442 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200443 rb_erase_init(n, &root->rb);
Jens Axboecc09e292007-04-26 12:53:50 +0200444}
445
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446/*
447 * would be nice to take fifo expire time into account as well
448 */
Jens Axboe5e705372006-07-13 12:39:25 +0200449static struct request *
450cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
451 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452{
Jens Axboe21183b02006-07-13 12:33:14 +0200453 struct rb_node *rbnext = rb_next(&last->rb_node);
454 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200455 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456
Jens Axboe21183b02006-07-13 12:33:14 +0200457 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458
459 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200460 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200461
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200463 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200464 else {
465 rbnext = rb_first(&cfqq->sort_list);
466 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200467 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200468 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469
Jens Axboe21183b02006-07-13 12:33:14 +0200470 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471}
472
Jens Axboed9e76202007-04-20 14:27:50 +0200473static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
474 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475{
Jens Axboed9e76202007-04-20 14:27:50 +0200476 /*
477 * just an approximation, should be ok.
478 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200479 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
480 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200481}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482
Jens Axboe498d3aa2007-04-26 12:54:48 +0200483/*
484 * The cfqd->service_tree holds all pending cfq_queue's that have
485 * requests waiting to be processed. It is sorted in the order that
486 * we will service the queues.
487 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200488static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
489 int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200490{
Jens Axboe08717142008-01-28 11:38:15 +0100491 struct rb_node **p, *parent;
492 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +0200493 unsigned long rb_key;
Jens Axboe498d3aa2007-04-26 12:54:48 +0200494 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200495
Jens Axboe08717142008-01-28 11:38:15 +0100496 if (cfq_class_idle(cfqq)) {
497 rb_key = CFQ_IDLE_DELAY;
498 parent = rb_last(&cfqd->service_tree.rb);
499 if (parent && parent != &cfqq->rb_node) {
500 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
501 rb_key += __cfqq->rb_key;
502 } else
503 rb_key += jiffies;
504 } else if (!add_front) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200505 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
506 rb_key += cfqq->slice_resid;
507 cfqq->slice_resid = 0;
508 } else
509 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200510
511 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100512 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200513 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100514 */
Jens Axboed9e76202007-04-20 14:27:50 +0200515 if (rb_key == cfqq->rb_key)
516 return;
Jens Axboe53b03742006-07-28 09:48:51 +0200517
Jens Axboecc09e292007-04-26 12:53:50 +0200518 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200519 }
Jens Axboed9e76202007-04-20 14:27:50 +0200520
Jens Axboe498d3aa2007-04-26 12:54:48 +0200521 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +0100522 parent = NULL;
523 p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200524 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +0200525 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200526
Jens Axboed9e76202007-04-20 14:27:50 +0200527 parent = *p;
528 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
529
Jens Axboe0c534e02007-04-18 20:01:57 +0200530 /*
531 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200532 * preference to them. IDLE queues goes to the back.
533 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200534 */
535 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200536 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200537 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200538 n = &(*p)->rb_right;
539 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
540 n = &(*p)->rb_left;
541 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
542 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200543 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200544 n = &(*p)->rb_left;
545 else
546 n = &(*p)->rb_right;
547
548 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200549 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200550
551 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200552 }
553
Jens Axboecc09e292007-04-26 12:53:50 +0200554 if (left)
555 cfqd->service_tree.left = &cfqq->rb_node;
556
Jens Axboed9e76202007-04-20 14:27:50 +0200557 cfqq->rb_key = rb_key;
558 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200559 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560}
561
Jens Axboea36e71f2009-04-15 12:15:11 +0200562static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200563cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
564 sector_t sector, struct rb_node **ret_parent,
565 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +0200566{
Jens Axboea36e71f2009-04-15 12:15:11 +0200567 struct rb_node **p, *parent;
568 struct cfq_queue *cfqq = NULL;
569
570 parent = NULL;
571 p = &root->rb_node;
572 while (*p) {
573 struct rb_node **n;
574
575 parent = *p;
576 cfqq = rb_entry(parent, struct cfq_queue, p_node);
577
578 /*
579 * Sort strictly based on sector. Smallest to the left,
580 * largest to the right.
581 */
582 if (sector > cfqq->next_rq->sector)
583 n = &(*p)->rb_right;
584 else if (sector < cfqq->next_rq->sector)
585 n = &(*p)->rb_left;
586 else
587 break;
588 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200589 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200590 }
591
592 *ret_parent = parent;
593 if (rb_link)
594 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200595 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +0200596}
597
598static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
599{
Jens Axboea36e71f2009-04-15 12:15:11 +0200600 struct rb_node **p, *parent;
601 struct cfq_queue *__cfqq;
602
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200603 if (cfqq->p_root) {
604 rb_erase(&cfqq->p_node, cfqq->p_root);
605 cfqq->p_root = NULL;
606 }
Jens Axboea36e71f2009-04-15 12:15:11 +0200607
608 if (cfq_class_idle(cfqq))
609 return;
610 if (!cfqq->next_rq)
611 return;
612
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200613 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
614 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, cfqq->next_rq->sector,
Jens Axboea36e71f2009-04-15 12:15:11 +0200615 &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200616 if (!__cfqq) {
617 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200618 rb_insert_color(&cfqq->p_node, cfqq->p_root);
619 } else
620 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200621}
622
Jens Axboe498d3aa2007-04-26 12:54:48 +0200623/*
624 * Update cfqq's position in the service tree.
625 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200626static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200627{
Jens Axboe6d048f52007-04-25 12:44:27 +0200628 /*
629 * Resorting requires the cfqq to be on the RR list already.
630 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200631 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200632 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +0200633 cfq_prio_tree_add(cfqd, cfqq);
634 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200635}
636
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637/*
638 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200639 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640 */
Jens Axboefebffd62008-01-28 13:19:43 +0100641static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642{
Jens Axboe7b679132008-05-30 12:23:07 +0200643 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200644 BUG_ON(cfq_cfqq_on_rr(cfqq));
645 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646 cfqd->busy_queues++;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +0100647 if (cfq_class_rt(cfqq))
648 cfqd->busy_rt_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649
Jens Axboeedd75ff2007-04-19 12:03:34 +0200650 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651}
652
Jens Axboe498d3aa2007-04-26 12:54:48 +0200653/*
654 * Called when the cfqq no longer has requests pending, remove it from
655 * the service tree.
656 */
Jens Axboefebffd62008-01-28 13:19:43 +0100657static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658{
Jens Axboe7b679132008-05-30 12:23:07 +0200659 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200660 BUG_ON(!cfq_cfqq_on_rr(cfqq));
661 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
Jens Axboecc09e292007-04-26 12:53:50 +0200663 if (!RB_EMPTY_NODE(&cfqq->rb_node))
664 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200665 if (cfqq->p_root) {
666 rb_erase(&cfqq->p_node, cfqq->p_root);
667 cfqq->p_root = NULL;
668 }
Jens Axboed9e76202007-04-20 14:27:50 +0200669
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670 BUG_ON(!cfqd->busy_queues);
671 cfqd->busy_queues--;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +0100672 if (cfq_class_rt(cfqq))
673 cfqd->busy_rt_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674}
675
676/*
677 * rb tree support functions
678 */
Jens Axboefebffd62008-01-28 13:19:43 +0100679static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680{
Jens Axboe5e705372006-07-13 12:39:25 +0200681 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200682 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200683 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684
Jens Axboeb4878f22005-10-20 16:42:29 +0200685 BUG_ON(!cfqq->queued[sync]);
686 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687
Jens Axboe5e705372006-07-13 12:39:25 +0200688 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689
Jens Axboedd67d052006-06-21 09:36:18 +0200690 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200691 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692}
693
Jens Axboe5e705372006-07-13 12:39:25 +0200694static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695{
Jens Axboe5e705372006-07-13 12:39:25 +0200696 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +0200698 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699
Jens Axboe5380a102006-07-13 12:37:56 +0200700 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701
702 /*
703 * looks a little odd, but the first insert might return an alias.
704 * if that happens, put the alias on the dispatch list
705 */
Jens Axboe21183b02006-07-13 12:33:14 +0200706 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200707 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100708
709 if (!cfq_cfqq_on_rr(cfqq))
710 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200711
712 /*
713 * check if this request is a better next-serve candidate
714 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200715 prev = cfqq->next_rq;
Jens Axboe5044eed2007-04-25 11:53:48 +0200716 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
Jens Axboea36e71f2009-04-15 12:15:11 +0200717
718 /*
719 * adjust priority tree position, if ->next_rq changes
720 */
721 if (prev != cfqq->next_rq)
722 cfq_prio_tree_add(cfqd, cfqq);
723
Jens Axboe5044eed2007-04-25 11:53:48 +0200724 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725}
726
Jens Axboefebffd62008-01-28 13:19:43 +0100727static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728{
Jens Axboe5380a102006-07-13 12:37:56 +0200729 elv_rb_del(&cfqq->sort_list, rq);
730 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200731 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732}
733
Jens Axboe206dc692006-03-28 13:03:44 +0200734static struct request *
735cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736{
Jens Axboe206dc692006-03-28 13:03:44 +0200737 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200738 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200739 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740
Jens Axboe4ac845a2008-01-24 08:44:49 +0100741 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200742 if (!cic)
743 return NULL;
744
745 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200746 if (cfqq) {
747 sector_t sector = bio->bi_sector + bio_sectors(bio);
748
Jens Axboe21183b02006-07-13 12:33:14 +0200749 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200750 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 return NULL;
753}
754
Jens Axboe165125e2007-07-24 09:28:11 +0200755static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200756{
757 struct cfq_data *cfqd = q->elevator->elevator_data;
758
759 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +0200760 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
761 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +0200762
Jens Axboe6d048f52007-04-25 12:44:27 +0200763 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
Jens Axboeb4878f22005-10-20 16:42:29 +0200764}
765
Jens Axboe165125e2007-07-24 09:28:11 +0200766static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767{
Jens Axboe22e2c502005-06-27 10:55:12 +0200768 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769
Jens Axboeb4878f22005-10-20 16:42:29 +0200770 WARN_ON(!cfqd->rq_in_driver);
771 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +0200772 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
773 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774}
775
Jens Axboeb4878f22005-10-20 16:42:29 +0200776static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777{
Jens Axboe5e705372006-07-13 12:39:25 +0200778 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200779
Jens Axboe5e705372006-07-13 12:39:25 +0200780 if (cfqq->next_rq == rq)
781 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782
Jens Axboeb4878f22005-10-20 16:42:29 +0200783 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200784 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200785
Aaron Carroll45333d52008-08-26 15:52:36 +0200786 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +0200787 if (rq_is_meta(rq)) {
788 WARN_ON(!cfqq->meta_pending);
789 cfqq->meta_pending--;
790 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791}
792
Jens Axboe165125e2007-07-24 09:28:11 +0200793static int cfq_merge(struct request_queue *q, struct request **req,
794 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795{
796 struct cfq_data *cfqd = q->elevator->elevator_data;
797 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798
Jens Axboe206dc692006-03-28 13:03:44 +0200799 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200800 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200801 *req = __rq;
802 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 }
804
805 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806}
807
Jens Axboe165125e2007-07-24 09:28:11 +0200808static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200809 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810{
Jens Axboe21183b02006-07-13 12:33:14 +0200811 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200812 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813
Jens Axboe5e705372006-07-13 12:39:25 +0200814 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816}
817
818static void
Jens Axboe165125e2007-07-24 09:28:11 +0200819cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 struct request *next)
821{
Jens Axboe22e2c502005-06-27 10:55:12 +0200822 /*
823 * reposition in fifo if next is older than rq
824 */
825 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
826 time_before(next->start_time, rq->start_time))
827 list_move(&rq->queuelist, &next->queuelist);
828
Jens Axboeb4878f22005-10-20 16:42:29 +0200829 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200830}
831
Jens Axboe165125e2007-07-24 09:28:11 +0200832static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100833 struct bio *bio)
834{
835 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200836 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100837 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100838
839 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100840 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100841 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200842 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100843 return 0;
844
845 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100846 * Lookup the cfqq that this bio will be queued with. Allow
847 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100848 */
Jens Axboe4ac845a2008-01-24 08:44:49 +0100849 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200850 if (!cic)
851 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100852
Vasily Tarasov91fac312007-04-25 12:29:51 +0200853 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100854 if (cfqq == RQ_CFQQ(rq))
855 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100856
Jens Axboeec8acb62007-01-02 18:32:11 +0100857 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100858}
859
Jens Axboefebffd62008-01-28 13:19:43 +0100860static void __cfq_set_active_queue(struct cfq_data *cfqd,
861 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +0200862{
863 if (cfqq) {
Jens Axboe7b679132008-05-30 12:23:07 +0200864 cfq_log_cfqq(cfqd, cfqq, "set_active");
Jens Axboe22e2c502005-06-27 10:55:12 +0200865 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +0200866 cfqq->slice_dispatch = 0;
867
Jens Axboe2f5cb732009-04-07 08:51:19 +0200868 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +0200869 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200870 cfq_clear_cfqq_must_alloc_slice(cfqq);
871 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100872 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +0200873
874 del_timer(&cfqd->idle_slice_timer);
Jens Axboe22e2c502005-06-27 10:55:12 +0200875 }
876
877 cfqd->active_queue = cfqq;
878}
879
880/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100881 * current cfqq expired its slice (or was too idle), select new one
882 */
883static void
884__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200885 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100886{
Jens Axboe7b679132008-05-30 12:23:07 +0200887 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
888
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100889 if (cfq_cfqq_wait_request(cfqq))
890 del_timer(&cfqd->idle_slice_timer);
891
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100892 cfq_clear_cfqq_wait_request(cfqq);
893
894 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200895 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100896 */
Jens Axboe7b679132008-05-30 12:23:07 +0200897 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +1100898 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200899 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
900 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100901
Jens Axboeedd75ff2007-04-19 12:03:34 +0200902 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100903
904 if (cfqq == cfqd->active_queue)
905 cfqd->active_queue = NULL;
906
907 if (cfqd->active_cic) {
908 put_io_context(cfqd->active_cic->ioc);
909 cfqd->active_cic = NULL;
910 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100911}
912
Jens Axboe6084cdd2007-04-23 08:25:00 +0200913static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100914{
915 struct cfq_queue *cfqq = cfqd->active_queue;
916
917 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200918 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100919}
920
Jens Axboe498d3aa2007-04-26 12:54:48 +0200921/*
922 * Get next queue for service. Unless we have a queue preemption,
923 * we'll simply select the first cfqq in the service tree.
924 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200925static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200926{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200927 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
928 return NULL;
929
Jens Axboe08717142008-01-28 11:38:15 +0100930 return cfq_rb_first(&cfqd->service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +0200931}
932
Jens Axboe498d3aa2007-04-26 12:54:48 +0200933/*
934 * Get and set a new active queue for service.
935 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200936static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
937 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200938{
Jens Axboea36e71f2009-04-15 12:15:11 +0200939 if (!cfqq) {
940 cfqq = cfq_get_next_queue(cfqd);
941 if (cfqq)
942 cfq_clear_cfqq_coop(cfqq);
943 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200944
Jens Axboe22e2c502005-06-27 10:55:12 +0200945 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200946 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200947}
948
Jens Axboed9e76202007-04-20 14:27:50 +0200949static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
950 struct request *rq)
951{
952 if (rq->sector >= cfqd->last_position)
953 return rq->sector - cfqd->last_position;
954 else
955 return cfqd->last_position - rq->sector;
956}
957
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200958#define CIC_SEEK_THR 8 * 1024
959#define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
960
Jens Axboe6d048f52007-04-25 12:44:27 +0200961static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
962{
963 struct cfq_io_context *cic = cfqd->active_cic;
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200964 sector_t sdist = cic->seek_mean;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200965
Jens Axboe6d048f52007-04-25 12:44:27 +0200966 if (!sample_valid(cic->seek_samples))
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200967 sdist = CIC_SEEK_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +0200968
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200969 return cfq_dist_from_last(cfqd, rq) <= sdist;
Jens Axboe6d048f52007-04-25 12:44:27 +0200970}
971
Jens Axboea36e71f2009-04-15 12:15:11 +0200972static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
973 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200974{
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200975 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +0200976 struct rb_node *parent, *node;
977 struct cfq_queue *__cfqq;
978 sector_t sector = cfqd->last_position;
979
980 if (RB_EMPTY_ROOT(root))
981 return NULL;
982
983 /*
984 * First, if we find a request starting at the end of the last
985 * request, choose it.
986 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200987 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +0200988 if (__cfqq)
989 return __cfqq;
990
991 /*
992 * If the exact sector wasn't found, the parent of the NULL leaf
993 * will contain the closest sector.
994 */
995 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
996 if (cfq_rq_close(cfqd, __cfqq->next_rq))
997 return __cfqq;
998
999 if (__cfqq->next_rq->sector < sector)
1000 node = rb_next(&__cfqq->p_node);
1001 else
1002 node = rb_prev(&__cfqq->p_node);
1003 if (!node)
1004 return NULL;
1005
1006 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1007 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1008 return __cfqq;
1009
1010 return NULL;
1011}
1012
1013/*
1014 * cfqd - obvious
1015 * cur_cfqq - passed in so that we don't decide that the current queue is
1016 * closely cooperating with itself.
1017 *
1018 * So, basically we're assuming that that cur_cfqq has dispatched at least
1019 * one request, and that cfqd->last_position reflects a position on the disk
1020 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1021 * assumption.
1022 */
1023static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1024 struct cfq_queue *cur_cfqq,
1025 int probe)
1026{
1027 struct cfq_queue *cfqq;
1028
1029 /*
1030 * A valid cfq_io_context is necessary to compare requests against
1031 * the seek_mean of the current cfqq.
1032 */
1033 if (!cfqd->active_cic)
1034 return NULL;
1035
Jens Axboe6d048f52007-04-25 12:44:27 +02001036 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001037 * We should notice if some of the queues are cooperating, eg
1038 * working closely on the same area of the disk. In that case,
1039 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001040 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001041 cfqq = cfqq_close(cfqd, cur_cfqq);
1042 if (!cfqq)
1043 return NULL;
1044
1045 if (cfq_cfqq_coop(cfqq))
1046 return NULL;
1047
1048 if (!probe)
1049 cfq_mark_cfqq_coop(cfqq);
1050 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001051}
1052
Jens Axboe6d048f52007-04-25 12:44:27 +02001053static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001054{
Jens Axboe17926692007-01-19 11:59:30 +11001055 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001056 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001057 unsigned long sl;
1058
Jens Axboea68bbdd2008-09-24 13:03:33 +02001059 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001060 * SSD device without seek penalty, disable idling. But only do so
1061 * for devices that support queuing, otherwise we still have a problem
1062 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001063 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001064 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001065 return;
1066
Jens Axboedd67d052006-06-21 09:36:18 +02001067 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001068 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001069
1070 /*
1071 * idle is disabled, either manually or by past process history
1072 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001073 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1074 return;
1075
Jens Axboe22e2c502005-06-27 10:55:12 +02001076 /*
Jens Axboe7b679132008-05-30 12:23:07 +02001077 * still requests with the driver, don't idle
1078 */
1079 if (cfqd->rq_in_driver)
1080 return;
1081
1082 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001083 * task has exited, don't wait
1084 */
Jens Axboe206dc692006-03-28 13:03:44 +02001085 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001086 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001087 return;
1088
Jens Axboe3b181522005-06-27 10:56:24 +02001089 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001090
Jens Axboe206dc692006-03-28 13:03:44 +02001091 /*
1092 * we don't want to idle for seeks, but we do want to allow
1093 * fair distribution of slice time for a process doing back-to-back
1094 * seeks. so allow a little bit of time for him to submit a new rq
1095 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001096 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001097 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +02001098 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +02001099
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001100 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe9481ffd2009-04-15 12:14:13 +02001101 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102}
1103
Jens Axboe498d3aa2007-04-26 12:54:48 +02001104/*
1105 * Move request from internal lists to the request queue dispatch list.
1106 */
Jens Axboe165125e2007-07-24 09:28:11 +02001107static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001109 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001110 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001111
Jens Axboe7b679132008-05-30 12:23:07 +02001112 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1113
Jens Axboe5380a102006-07-13 12:37:56 +02001114 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001115 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001116 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001117
1118 if (cfq_cfqq_sync(cfqq))
1119 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120}
1121
1122/*
1123 * return expired entry, or NULL to just start from scratch in rbtree
1124 */
Jens Axboefebffd62008-01-28 13:19:43 +01001125static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126{
1127 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +02001128 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +02001129 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
Jens Axboe3b181522005-06-27 10:56:24 +02001131 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001133
1134 cfq_mark_cfqq_fifo_expire(cfqq);
1135
Jens Axboe89850f72006-07-22 16:48:31 +02001136 if (list_empty(&cfqq->fifo))
1137 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138
Jens Axboe6d048f52007-04-25 12:44:27 +02001139 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +02001140 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141
Jens Axboe6d048f52007-04-25 12:44:27 +02001142 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
Jens Axboe7b679132008-05-30 12:23:07 +02001143 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144
Jens Axboe7b679132008-05-30 12:23:07 +02001145 cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001146 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147}
1148
Jens Axboe22e2c502005-06-27 10:55:12 +02001149static inline int
1150cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1151{
1152 const int base_rq = cfqd->cfq_slice_async_rq;
1153
1154 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1155
1156 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1157}
1158
1159/*
Jens Axboe498d3aa2007-04-26 12:54:48 +02001160 * Select a queue for service. If we have a current active queue,
1161 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02001162 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001163static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001164{
Jens Axboea36e71f2009-04-15 12:15:11 +02001165 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001166
1167 cfqq = cfqd->active_queue;
1168 if (!cfqq)
1169 goto new_queue;
1170
1171 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001172 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02001173 */
Jens Axboeb0291952009-04-07 11:38:31 +02001174 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001175 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +02001176
1177 /*
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01001178 * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1179 * cfqq.
1180 */
1181 if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1182 /*
1183 * We simulate this as cfqq timed out so that it gets to bank
1184 * the remaining of its time slice.
1185 */
1186 cfq_log_cfqq(cfqd, cfqq, "preempt");
1187 cfq_slice_expired(cfqd, 1);
1188 goto new_queue;
1189 }
1190
1191 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001192 * The active queue has requests and isn't expired, allow it to
1193 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02001194 */
Jens Axboedd67d052006-06-21 09:36:18 +02001195 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001196 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02001197
1198 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02001199 * If another queue has a request waiting within our mean seek
1200 * distance, let it run. The expire code will check for close
1201 * cooperators and put the close queue at the front of the service
1202 * tree.
1203 */
1204 new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1205 if (new_cfqq)
1206 goto expire;
1207
1208 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001209 * No requests pending. If the active queue still has requests in
1210 * flight or is idling for a new request, allow either of these
1211 * conditions to happen (or time out) before selecting a new queue.
1212 */
Jens Axboecc197472007-04-20 20:45:39 +02001213 if (timer_pending(&cfqd->idle_slice_timer) ||
1214 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001215 cfqq = NULL;
1216 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001217 }
1218
Jens Axboe3b181522005-06-27 10:56:24 +02001219expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001220 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001221new_queue:
Jens Axboea36e71f2009-04-15 12:15:11 +02001222 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001223keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001224 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001225}
1226
Jens Axboefebffd62008-01-28 13:19:43 +01001227static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02001228{
1229 int dispatched = 0;
1230
1231 while (cfqq->next_rq) {
1232 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1233 dispatched++;
1234 }
1235
1236 BUG_ON(!list_empty(&cfqq->fifo));
1237 return dispatched;
1238}
1239
Jens Axboe498d3aa2007-04-26 12:54:48 +02001240/*
1241 * Drain our current requests. Used for barriers and when switching
1242 * io schedulers on-the-fly.
1243 */
Jens Axboed9e76202007-04-20 14:27:50 +02001244static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001245{
Jens Axboe08717142008-01-28 11:38:15 +01001246 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001247 int dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001248
Jens Axboe08717142008-01-28 11:38:15 +01001249 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
Jens Axboed9e76202007-04-20 14:27:50 +02001250 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001251
Jens Axboe6084cdd2007-04-23 08:25:00 +02001252 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001253
1254 BUG_ON(cfqd->busy_queues);
1255
Jens Axboe7b679132008-05-30 12:23:07 +02001256 cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001257 return dispatched;
1258}
1259
Jens Axboe2f5cb732009-04-07 08:51:19 +02001260/*
1261 * Dispatch a request from cfqq, moving them to the request queue
1262 * dispatch list.
1263 */
1264static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1265{
1266 struct request *rq;
1267
1268 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1269
1270 /*
1271 * follow expired path, else get first next available
1272 */
1273 rq = cfq_check_fifo(cfqq);
1274 if (!rq)
1275 rq = cfqq->next_rq;
1276
1277 /*
1278 * insert request into driver dispatch list
1279 */
1280 cfq_dispatch_insert(cfqd->queue, rq);
1281
1282 if (!cfqd->active_cic) {
1283 struct cfq_io_context *cic = RQ_CIC(rq);
1284
1285 atomic_inc(&cic->ioc->refcount);
1286 cfqd->active_cic = cic;
1287 }
1288}
1289
1290/*
1291 * Find the cfqq that we need to service and move a request from that to the
1292 * dispatch list
1293 */
Jens Axboe165125e2007-07-24 09:28:11 +02001294static int cfq_dispatch_requests(struct request_queue *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001295{
1296 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001297 struct cfq_queue *cfqq;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001298 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299
Jens Axboe22e2c502005-06-27 10:55:12 +02001300 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 return 0;
1302
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001303 if (unlikely(force))
1304 return cfq_forced_dispatch(cfqd);
1305
Jens Axboe2f5cb732009-04-07 08:51:19 +02001306 cfqq = cfq_select_queue(cfqd);
1307 if (!cfqq)
1308 return 0;
Jens Axboeb4878f22005-10-20 16:42:29 +02001309
Jens Axboe2f5cb732009-04-07 08:51:19 +02001310 /*
1311 * If this is an async queue and we have sync IO in flight, let it wait
1312 */
1313 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1314 return 0;
1315
1316 max_dispatch = cfqd->cfq_quantum;
1317 if (cfq_class_idle(cfqq))
1318 max_dispatch = 1;
1319
1320 /*
1321 * Does this cfqq already have too much IO in flight?
1322 */
1323 if (cfqq->dispatched >= max_dispatch) {
1324 /*
1325 * idle queue must always only have a single IO in flight
1326 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02001327 if (cfq_class_idle(cfqq))
Jens Axboe2f5cb732009-04-07 08:51:19 +02001328 return 0;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001329
Jens Axboe2f5cb732009-04-07 08:51:19 +02001330 /*
1331 * We have other queues, don't allow more IO from this one
1332 */
1333 if (cfqd->busy_queues > 1)
1334 return 0;
Jens Axboe9ede2092007-01-19 12:11:44 +11001335
Jens Axboe2f5cb732009-04-07 08:51:19 +02001336 /*
1337 * we are the only queue, allow up to 4 times of 'quantum'
1338 */
1339 if (cfqq->dispatched >= 4 * max_dispatch)
1340 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341 }
1342
Jens Axboe2f5cb732009-04-07 08:51:19 +02001343 /*
1344 * Dispatch a request from this cfqq
1345 */
1346 cfq_dispatch_request(cfqd, cfqq);
1347 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02001348 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001349
1350 /*
1351 * expire an async queue immediately if it has used up its slice. idle
1352 * queue always expire after 1 dispatch round.
1353 */
1354 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1355 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1356 cfq_class_idle(cfqq))) {
1357 cfqq->slice_end = jiffies + 1;
1358 cfq_slice_expired(cfqd, 0);
1359 }
1360
1361 cfq_log(cfqd, "dispatched a request");
1362 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363}
1364
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365/*
Jens Axboe5e705372006-07-13 12:39:25 +02001366 * task holds one reference to the queue, dropped when task exits. each rq
1367 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 *
1369 * queue lock must be held here.
1370 */
1371static void cfq_put_queue(struct cfq_queue *cfqq)
1372{
Jens Axboe22e2c502005-06-27 10:55:12 +02001373 struct cfq_data *cfqd = cfqq->cfqd;
1374
1375 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376
1377 if (!atomic_dec_and_test(&cfqq->ref))
1378 return;
1379
Jens Axboe7b679132008-05-30 12:23:07 +02001380 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001382 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001383 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001385 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001386 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001387 cfq_schedule_dispatch(cfqd);
1388 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001389
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 kmem_cache_free(cfq_pool, cfqq);
1391}
1392
Jens Axboed6de8be2008-05-28 14:46:59 +02001393/*
1394 * Must always be called with the rcu_read_lock() held
1395 */
Jens Axboe07416d22008-05-07 09:17:12 +02001396static void
1397__call_for_each_cic(struct io_context *ioc,
1398 void (*func)(struct io_context *, struct cfq_io_context *))
1399{
1400 struct cfq_io_context *cic;
1401 struct hlist_node *n;
1402
1403 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1404 func(ioc, cic);
1405}
1406
Jens Axboe4ac845a2008-01-24 08:44:49 +01001407/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001408 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001409 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001410static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001411call_for_each_cic(struct io_context *ioc,
1412 void (*func)(struct io_context *, struct cfq_io_context *))
1413{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001414 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02001415 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001416 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001417}
Jens Axboe4ac845a2008-01-24 08:44:49 +01001418
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001419static void cfq_cic_free_rcu(struct rcu_head *head)
1420{
1421 struct cfq_io_context *cic;
1422
1423 cic = container_of(head, struct cfq_io_context, rcu_head);
1424
1425 kmem_cache_free(cfq_ioc_pool, cic);
1426 elv_ioc_count_dec(ioc_count);
1427
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001428 if (ioc_gone) {
1429 /*
1430 * CFQ scheduler is exiting, grab exit lock and check
1431 * the pending io context count. If it hits zero,
1432 * complete ioc_gone and set it back to NULL
1433 */
1434 spin_lock(&ioc_gone_lock);
1435 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1436 complete(ioc_gone);
1437 ioc_gone = NULL;
1438 }
1439 spin_unlock(&ioc_gone_lock);
1440 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001441}
1442
1443static void cfq_cic_free(struct cfq_io_context *cic)
1444{
1445 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001446}
1447
1448static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1449{
1450 unsigned long flags;
1451
1452 BUG_ON(!cic->dead_key);
1453
1454 spin_lock_irqsave(&ioc->lock, flags);
1455 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01001456 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001457 spin_unlock_irqrestore(&ioc->lock, flags);
1458
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001459 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001460}
1461
Jens Axboed6de8be2008-05-28 14:46:59 +02001462/*
1463 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1464 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1465 * and ->trim() which is called with the task lock held
1466 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001467static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001469 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001470 * ioc->refcount is zero here, or we are called from elv_unregister(),
1471 * so no more cic's are allowed to be linked into this ioc. So it
1472 * should be ok to iterate over the known list, we will see all cic's
1473 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001474 */
Jens Axboe07416d22008-05-07 09:17:12 +02001475 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476}
1477
Jens Axboe89850f72006-07-22 16:48:31 +02001478static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1479{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001480 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001481 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001482 cfq_schedule_dispatch(cfqd);
1483 }
Jens Axboe89850f72006-07-22 16:48:31 +02001484
1485 cfq_put_queue(cfqq);
1486}
1487
1488static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1489 struct cfq_io_context *cic)
1490{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001491 struct io_context *ioc = cic->ioc;
1492
Jens Axboefc463792006-08-29 09:05:44 +02001493 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001494
1495 /*
1496 * Make sure key == NULL is seen for dead queues
1497 */
Jens Axboefc463792006-08-29 09:05:44 +02001498 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001499 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02001500 cic->key = NULL;
1501
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001502 if (ioc->ioc_data == cic)
1503 rcu_assign_pointer(ioc->ioc_data, NULL);
1504
Jens Axboeff6657c2009-04-08 10:58:57 +02001505 if (cic->cfqq[BLK_RW_ASYNC]) {
1506 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1507 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001508 }
1509
Jens Axboeff6657c2009-04-08 10:58:57 +02001510 if (cic->cfqq[BLK_RW_SYNC]) {
1511 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1512 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001513 }
Jens Axboe89850f72006-07-22 16:48:31 +02001514}
1515
Jens Axboe4ac845a2008-01-24 08:44:49 +01001516static void cfq_exit_single_io_context(struct io_context *ioc,
1517 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001518{
Al Viro478a82b2006-03-18 13:25:24 -05001519 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001520
Jens Axboe89850f72006-07-22 16:48:31 +02001521 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001522 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001523 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02001524
Jens Axboe4ac845a2008-01-24 08:44:49 +01001525 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01001526
1527 /*
1528 * Ensure we get a fresh copy of the ->key to prevent
1529 * race between exiting task and queue
1530 */
1531 smp_read_barrier_depends();
1532 if (cic->key)
1533 __cfq_exit_single_io_context(cfqd, cic);
1534
Jens Axboe4ac845a2008-01-24 08:44:49 +01001535 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05001536 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001537}
1538
Jens Axboe498d3aa2007-04-26 12:54:48 +02001539/*
1540 * The process that ioc belongs to has exited, we need to clean up
1541 * and put the internal structures we have that belongs to that process.
1542 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001543static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001545 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546}
1547
Jens Axboe22e2c502005-06-27 10:55:12 +02001548static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001549cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550{
Jens Axboeb5deef92006-07-19 23:39:40 +02001551 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552
Christoph Lameter94f60302007-07-17 04:03:29 -07001553 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1554 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001556 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001557 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01001558 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001559 cic->dtor = cfq_free_io_context;
1560 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001561 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 }
1563
1564 return cic;
1565}
1566
Jens Axboefd0928d2008-01-24 08:52:45 +01001567static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02001568{
1569 struct task_struct *tsk = current;
1570 int ioprio_class;
1571
Jens Axboe3b181522005-06-27 10:56:24 +02001572 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001573 return;
1574
Jens Axboefd0928d2008-01-24 08:52:45 +01001575 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001576 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01001577 default:
1578 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1579 case IOPRIO_CLASS_NONE:
1580 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02001581 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01001582 */
1583 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02001584 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01001585 break;
1586 case IOPRIO_CLASS_RT:
1587 cfqq->ioprio = task_ioprio(ioc);
1588 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1589 break;
1590 case IOPRIO_CLASS_BE:
1591 cfqq->ioprio = task_ioprio(ioc);
1592 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1593 break;
1594 case IOPRIO_CLASS_IDLE:
1595 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1596 cfqq->ioprio = 7;
1597 cfq_clear_cfqq_idle_window(cfqq);
1598 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02001599 }
1600
1601 /*
1602 * keep track of original prio settings in case we have to temporarily
1603 * elevate the priority of this queue
1604 */
1605 cfqq->org_ioprio = cfqq->ioprio;
1606 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001607 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001608}
1609
Jens Axboefebffd62008-01-28 13:19:43 +01001610static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001611{
Al Viro478a82b2006-03-18 13:25:24 -05001612 struct cfq_data *cfqd = cic->key;
1613 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001614 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001615
Jens Axboecaaa5f92006-06-16 11:23:00 +02001616 if (unlikely(!cfqd))
1617 return;
1618
Jens Axboec1b707d2006-10-30 19:54:23 +01001619 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001620
Jens Axboeff6657c2009-04-08 10:58:57 +02001621 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001622 if (cfqq) {
1623 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02001624 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1625 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001626 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02001627 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001628 cfq_put_queue(cfqq);
1629 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001630 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001631
Jens Axboeff6657c2009-04-08 10:58:57 +02001632 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001633 if (cfqq)
1634 cfq_mark_cfqq_prio_changed(cfqq);
1635
Jens Axboec1b707d2006-10-30 19:54:23 +01001636 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001637}
1638
Jens Axboefc463792006-08-29 09:05:44 +02001639static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001641 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02001642 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643}
1644
1645static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001646cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01001647 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001650 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651
1652retry:
Jens Axboe4ac845a2008-01-24 08:44:49 +01001653 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001654 /* cic always exists here */
1655 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656
1657 if (!cfqq) {
1658 if (new_cfqq) {
1659 cfqq = new_cfqq;
1660 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001661 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001662 /*
1663 * Inform the allocator of the fact that we will
1664 * just repeat this allocation if it fails, to allow
1665 * the allocator to do whatever it needs to attempt to
1666 * free memory.
1667 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001669 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1670 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1671 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 spin_lock_irq(cfqd->queue->queue_lock);
1673 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001674 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001675 cfqq = kmem_cache_alloc_node(cfq_pool,
1676 gfp_mask | __GFP_ZERO,
1677 cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001678 if (!cfqq)
1679 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001680 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681
Jens Axboed9e76202007-04-20 14:27:50 +02001682 RB_CLEAR_NODE(&cfqq->rb_node);
Jens Axboea36e71f2009-04-15 12:15:11 +02001683 RB_CLEAR_NODE(&cfqq->p_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001684 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 atomic_set(&cfqq->ref, 0);
1687 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001688
Jens Axboe3b181522005-06-27 10:56:24 +02001689 cfq_mark_cfqq_prio_changed(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001690
Jens Axboefd0928d2008-01-24 08:52:45 +01001691 cfq_init_prio_data(cfqq, ioc);
Jens Axboe08717142008-01-28 11:38:15 +01001692
1693 if (is_sync) {
1694 if (!cfq_class_idle(cfqq))
1695 cfq_mark_cfqq_idle_window(cfqq);
1696 cfq_mark_cfqq_sync(cfqq);
1697 }
Jens Axboe7b679132008-05-30 12:23:07 +02001698 cfqq->pid = current->pid;
1699 cfq_log_cfqq(cfqd, cfqq, "alloced");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 }
1701
1702 if (new_cfqq)
1703 kmem_cache_free(cfq_pool, new_cfqq);
1704
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705out:
1706 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1707 return cfqq;
1708}
1709
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001710static struct cfq_queue **
1711cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1712{
Jens Axboefe094d92008-01-31 13:08:54 +01001713 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001714 case IOPRIO_CLASS_RT:
1715 return &cfqd->async_cfqq[0][ioprio];
1716 case IOPRIO_CLASS_BE:
1717 return &cfqd->async_cfqq[1][ioprio];
1718 case IOPRIO_CLASS_IDLE:
1719 return &cfqd->async_idle_cfqq;
1720 default:
1721 BUG();
1722 }
1723}
1724
Jens Axboe15c31be2007-07-10 13:43:25 +02001725static struct cfq_queue *
Jens Axboefd0928d2008-01-24 08:52:45 +01001726cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02001727 gfp_t gfp_mask)
1728{
Jens Axboefd0928d2008-01-24 08:52:45 +01001729 const int ioprio = task_ioprio(ioc);
1730 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001731 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001732 struct cfq_queue *cfqq = NULL;
1733
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001734 if (!is_sync) {
1735 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1736 cfqq = *async_cfqq;
1737 }
1738
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001739 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01001740 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001741 if (!cfqq)
1742 return NULL;
1743 }
Jens Axboe15c31be2007-07-10 13:43:25 +02001744
1745 /*
1746 * pin the queue now that it's allocated, scheduler exit will prune it
1747 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001748 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001749 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001750 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001751 }
1752
1753 atomic_inc(&cfqq->ref);
1754 return cfqq;
1755}
1756
Jens Axboe498d3aa2007-04-26 12:54:48 +02001757/*
1758 * We drop cfq io contexts lazily, so we may find a dead one.
1759 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001760static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001761cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1762 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001763{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001764 unsigned long flags;
1765
Jens Axboefc463792006-08-29 09:05:44 +02001766 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001767
Jens Axboe4ac845a2008-01-24 08:44:49 +01001768 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02001769
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001770 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001771
1772 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01001773 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001774 spin_unlock_irqrestore(&ioc->lock, flags);
1775
1776 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001777}
1778
Jens Axboee2d74ac2006-03-28 08:59:01 +02001779static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01001780cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001781{
Jens Axboee2d74ac2006-03-28 08:59:01 +02001782 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001783 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001784 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001785
Vasily Tarasov91fac312007-04-25 12:29:51 +02001786 if (unlikely(!ioc))
1787 return NULL;
1788
Jens Axboed6de8be2008-05-28 14:46:59 +02001789 rcu_read_lock();
1790
Jens Axboe597bc482007-04-24 21:23:53 +02001791 /*
1792 * we maintain a last-hit cache, to avoid browsing over the tree
1793 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001794 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02001795 if (cic && cic->key == cfqd) {
1796 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02001797 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001798 }
Jens Axboe597bc482007-04-24 21:23:53 +02001799
Jens Axboe4ac845a2008-01-24 08:44:49 +01001800 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001801 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1802 rcu_read_unlock();
1803 if (!cic)
1804 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001805 /* ->key must be copied to avoid race with cfq_exit_queue() */
1806 k = cic->key;
1807 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001808 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001809 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001810 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001811 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001812
Jens Axboed6de8be2008-05-28 14:46:59 +02001813 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001814 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001815 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001816 break;
1817 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001818
Jens Axboe4ac845a2008-01-24 08:44:49 +01001819 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001820}
1821
Jens Axboe4ac845a2008-01-24 08:44:49 +01001822/*
1823 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1824 * the process specific cfq io context when entered from the block layer.
1825 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1826 */
Jens Axboefebffd62008-01-28 13:19:43 +01001827static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1828 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001829{
Jens Axboe0261d682006-10-30 19:07:48 +01001830 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001831 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001832
Jens Axboe4ac845a2008-01-24 08:44:49 +01001833 ret = radix_tree_preload(gfp_mask);
1834 if (!ret) {
1835 cic->ioc = ioc;
1836 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001837
Jens Axboe4ac845a2008-01-24 08:44:49 +01001838 spin_lock_irqsave(&ioc->lock, flags);
1839 ret = radix_tree_insert(&ioc->radix_root,
1840 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01001841 if (!ret)
1842 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001843 spin_unlock_irqrestore(&ioc->lock, flags);
1844
1845 radix_tree_preload_end();
1846
1847 if (!ret) {
1848 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1849 list_add(&cic->queue_list, &cfqd->cic_list);
1850 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001851 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001852 }
1853
Jens Axboe4ac845a2008-01-24 08:44:49 +01001854 if (ret)
1855 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02001856
Jens Axboe4ac845a2008-01-24 08:44:49 +01001857 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001858}
1859
Jens Axboe22e2c502005-06-27 10:55:12 +02001860/*
1861 * Setup general io context and cfq io context. There can be several cfq
1862 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001863 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001864 */
1865static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001866cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867{
Jens Axboe22e2c502005-06-27 10:55:12 +02001868 struct io_context *ioc = NULL;
1869 struct cfq_io_context *cic;
1870
1871 might_sleep_if(gfp_mask & __GFP_WAIT);
1872
Jens Axboeb5deef92006-07-19 23:39:40 +02001873 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001874 if (!ioc)
1875 return NULL;
1876
Jens Axboe4ac845a2008-01-24 08:44:49 +01001877 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001878 if (cic)
1879 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001880
Jens Axboee2d74ac2006-03-28 08:59:01 +02001881 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1882 if (cic == NULL)
1883 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001884
Jens Axboe4ac845a2008-01-24 08:44:49 +01001885 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1886 goto err_free;
1887
Jens Axboe22e2c502005-06-27 10:55:12 +02001888out:
Jens Axboefc463792006-08-29 09:05:44 +02001889 smp_read_barrier_depends();
1890 if (unlikely(ioc->ioprio_changed))
1891 cfq_ioc_set_ioprio(ioc);
1892
Jens Axboe22e2c502005-06-27 10:55:12 +02001893 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001894err_free:
1895 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001896err:
1897 put_io_context(ioc);
1898 return NULL;
1899}
1900
1901static void
1902cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1903{
Jens Axboeaaf12282007-01-19 11:30:16 +11001904 unsigned long elapsed = jiffies - cic->last_end_request;
1905 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001906
1907 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1908 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1909 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1910}
1911
Jens Axboe206dc692006-03-28 13:03:44 +02001912static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001913cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1914 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001915{
1916 sector_t sdist;
1917 u64 total;
1918
Jeff Moyer4d00aa42009-04-21 07:25:04 +02001919 if (!cic->last_request_pos)
1920 sdist = 0;
1921 else if (cic->last_request_pos < rq->sector)
Jens Axboe5e705372006-07-13 12:39:25 +02001922 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001923 else
Jens Axboe5e705372006-07-13 12:39:25 +02001924 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001925
1926 /*
1927 * Don't allow the seek distance to get too large from the
1928 * odd fragment, pagein, etc
1929 */
1930 if (cic->seek_samples <= 60) /* second&third seek */
1931 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1932 else
1933 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1934
1935 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1936 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1937 total = cic->seek_total + (cic->seek_samples/2);
1938 do_div(total, cic->seek_samples);
1939 cic->seek_mean = (sector_t)total;
1940}
Jens Axboe22e2c502005-06-27 10:55:12 +02001941
1942/*
1943 * Disable idle window if the process thinks too long or seeks so much that
1944 * it doesn't matter
1945 */
1946static void
1947cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1948 struct cfq_io_context *cic)
1949{
Jens Axboe7b679132008-05-30 12:23:07 +02001950 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02001951
Jens Axboe08717142008-01-28 11:38:15 +01001952 /*
1953 * Don't idle for async or idle io prio class
1954 */
1955 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02001956 return;
1957
Jens Axboec265a7f2008-06-26 13:49:33 +02001958 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001959
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001960 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Jens Axboecaaa5f92006-06-16 11:23:00 +02001961 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001962 enable_idle = 0;
1963 else if (sample_valid(cic->ttime_samples)) {
1964 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1965 enable_idle = 0;
1966 else
1967 enable_idle = 1;
1968 }
1969
Jens Axboe7b679132008-05-30 12:23:07 +02001970 if (old_idle != enable_idle) {
1971 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1972 if (enable_idle)
1973 cfq_mark_cfqq_idle_window(cfqq);
1974 else
1975 cfq_clear_cfqq_idle_window(cfqq);
1976 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001977}
1978
Jens Axboe22e2c502005-06-27 10:55:12 +02001979/*
1980 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1981 * no or if we aren't sure, a 1 will cause a preempt.
1982 */
1983static int
1984cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001985 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001986{
Jens Axboe6d048f52007-04-25 12:44:27 +02001987 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001988
Jens Axboe6d048f52007-04-25 12:44:27 +02001989 cfqq = cfqd->active_queue;
1990 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001991 return 0;
1992
Jens Axboe6d048f52007-04-25 12:44:27 +02001993 if (cfq_slice_used(cfqq))
1994 return 1;
1995
1996 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001997 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001998
1999 if (cfq_class_idle(cfqq))
2000 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002001
Jens Axboe22e2c502005-06-27 10:55:12 +02002002 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02002003 * if the new request is sync, but the currently running queue is
2004 * not, let the sync request have priority.
2005 */
Jens Axboe5e705372006-07-13 12:39:25 +02002006 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002007 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002008
Jens Axboe374f84a2006-07-23 01:42:19 +02002009 /*
2010 * So both queues are sync. Let the new request get disk time if
2011 * it's a metadata request and the current queue is doing regular IO.
2012 */
2013 if (rq_is_meta(rq) && !cfqq->meta_pending)
2014 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002015
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002016 /*
2017 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2018 */
2019 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2020 return 1;
2021
Jens Axboe1e3335d2007-02-14 19:59:49 +01002022 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2023 return 0;
2024
2025 /*
2026 * if this request is as-good as one we would expect from the
2027 * current cfqq, let it preempt
2028 */
Jens Axboe6d048f52007-04-25 12:44:27 +02002029 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01002030 return 1;
2031
Jens Axboe22e2c502005-06-27 10:55:12 +02002032 return 0;
2033}
2034
2035/*
2036 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2037 * let it have half of its nominal slice.
2038 */
2039static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2040{
Jens Axboe7b679132008-05-30 12:23:07 +02002041 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02002042 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02002043
Jens Axboebf572252006-07-19 20:29:12 +02002044 /*
2045 * Put the new queue at the front of the of the current list,
2046 * so we know that it will be selected next.
2047 */
2048 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02002049
2050 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02002051
Jens Axboe44f7c162007-01-19 11:51:58 +11002052 cfqq->slice_end = 0;
2053 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002054}
2055
2056/*
Jens Axboe5e705372006-07-13 12:39:25 +02002057 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02002058 * something we should do about it
2059 */
2060static void
Jens Axboe5e705372006-07-13 12:39:25 +02002061cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2062 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002063{
Jens Axboe5e705372006-07-13 12:39:25 +02002064 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02002065
Aaron Carroll45333d52008-08-26 15:52:36 +02002066 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02002067 if (rq_is_meta(rq))
2068 cfqq->meta_pending++;
2069
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002070 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02002071 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002072 cfq_update_idle_window(cfqd, cfqq, cic);
2073
Jens Axboe5e705372006-07-13 12:39:25 +02002074 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02002075
2076 if (cfqq == cfqd->active_queue) {
2077 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002078 * Remember that we saw a request from this process, but
2079 * don't start queuing just yet. Otherwise we risk seeing lots
2080 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02002081 * and merging. If the request is already larger than a single
2082 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02002083 * merging is already done. Ditto for a busy system that
2084 * has other work pending, don't risk delaying until the
2085 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02002086 */
Jens Axboed6ceb252009-04-14 14:18:16 +02002087 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02002088 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2089 cfqd->busy_queues > 1) {
Jens Axboed6ceb252009-04-14 14:18:16 +02002090 del_timer(&cfqd->idle_slice_timer);
2091 blk_start_queueing(cfqd->queue);
2092 }
Jens Axboeb0291952009-04-07 11:38:31 +02002093 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboed6ceb252009-04-14 14:18:16 +02002094 }
Jens Axboe5e705372006-07-13 12:39:25 +02002095 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002096 /*
2097 * not the active queue - expire current slice if it is
2098 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002099 * has some old slice time left and is of higher priority or
2100 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02002101 */
2102 cfq_preempt_queue(cfqd, cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02002103 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002104 }
2105}
2106
Jens Axboe165125e2007-07-24 09:28:11 +02002107static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002108{
Jens Axboeb4878f22005-10-20 16:42:29 +02002109 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002110 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002111
Jens Axboe7b679132008-05-30 12:23:07 +02002112 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01002113 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114
Jens Axboe5e705372006-07-13 12:39:25 +02002115 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116
Jens Axboe22e2c502005-06-27 10:55:12 +02002117 list_add_tail(&rq->queuelist, &cfqq->fifo);
2118
Jens Axboe5e705372006-07-13 12:39:25 +02002119 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120}
2121
Aaron Carroll45333d52008-08-26 15:52:36 +02002122/*
2123 * Update hw_tag based on peak queue depth over 50 samples under
2124 * sufficient load.
2125 */
2126static void cfq_update_hw_tag(struct cfq_data *cfqd)
2127{
2128 if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
2129 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
2130
2131 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2132 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
2133 return;
2134
2135 if (cfqd->hw_tag_samples++ < 50)
2136 return;
2137
2138 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2139 cfqd->hw_tag = 1;
2140 else
2141 cfqd->hw_tag = 0;
2142
2143 cfqd->hw_tag_samples = 0;
2144 cfqd->rq_in_driver_peak = 0;
2145}
2146
Jens Axboe165125e2007-07-24 09:28:11 +02002147static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148{
Jens Axboe5e705372006-07-13 12:39:25 +02002149 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002150 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02002151 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002152 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153
Jens Axboeb4878f22005-10-20 16:42:29 +02002154 now = jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02002155 cfq_log_cfqq(cfqd, cfqq, "complete");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156
Aaron Carroll45333d52008-08-26 15:52:36 +02002157 cfq_update_hw_tag(cfqd);
2158
Jens Axboeb4878f22005-10-20 16:42:29 +02002159 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02002160 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02002161 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02002162 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163
Jens Axboe3ed9a292007-04-23 08:33:33 +02002164 if (cfq_cfqq_sync(cfqq))
2165 cfqd->sync_flight--;
2166
Jens Axboeb4878f22005-10-20 16:42:29 +02002167 if (!cfq_class_idle(cfqq))
2168 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02002169
Jens Axboecaaa5f92006-06-16 11:23:00 +02002170 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02002171 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002172
2173 /*
2174 * If this is the active queue, check if it needs to be expired,
2175 * or if we want to idle in case it has no pending requests.
2176 */
2177 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02002178 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2179
Jens Axboe44f7c162007-01-19 11:51:58 +11002180 if (cfq_cfqq_slice_new(cfqq)) {
2181 cfq_set_prio_slice(cfqd, cfqq);
2182 cfq_clear_cfqq_slice_new(cfqq);
2183 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002184 /*
2185 * If there are no requests waiting in this queue, and
2186 * there are other queues ready to issue requests, AND
2187 * those other queues are issuing requests within our
2188 * mean seek distance, give them a chance to run instead
2189 * of idling.
2190 */
Jens Axboe08717142008-01-28 11:38:15 +01002191 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02002192 cfq_slice_expired(cfqd, 1);
Jens Axboea36e71f2009-04-15 12:15:11 +02002193 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2194 sync && !rq_noidle(rq))
Jens Axboe6d048f52007-04-25 12:44:27 +02002195 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002196 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002197
2198 if (!cfqd->rq_in_driver)
2199 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200}
2201
Jens Axboe22e2c502005-06-27 10:55:12 +02002202/*
2203 * we temporarily boost lower priority queues if they are holding fs exclusive
2204 * resources. they are boosted to normal prio (CLASS_BE/4)
2205 */
2206static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207{
Jens Axboe22e2c502005-06-27 10:55:12 +02002208 if (has_fs_excl()) {
2209 /*
2210 * boost idle prio on transactions that would lock out other
2211 * users of the filesystem
2212 */
2213 if (cfq_class_idle(cfqq))
2214 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2215 if (cfqq->ioprio > IOPRIO_NORM)
2216 cfqq->ioprio = IOPRIO_NORM;
2217 } else {
2218 /*
2219 * check if we need to unboost the queue
2220 */
2221 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2222 cfqq->ioprio_class = cfqq->org_ioprio_class;
2223 if (cfqq->ioprio != cfqq->org_ioprio)
2224 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002226}
2227
Jens Axboe89850f72006-07-22 16:48:31 +02002228static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002229{
Jens Axboe3b181522005-06-27 10:56:24 +02002230 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07002231 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002232 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002233 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02002234 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002235
2236 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02002237}
2238
Jens Axboe165125e2007-07-24 09:28:11 +02002239static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02002240{
2241 struct cfq_data *cfqd = q->elevator->elevator_data;
2242 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002243 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02002244 struct cfq_queue *cfqq;
2245
2246 /*
2247 * don't force setup of a queue from here, as a call to may_queue
2248 * does not necessarily imply that a request actually will be queued.
2249 * so just lookup a possibly existing queue, or return 'may queue'
2250 * if that fails
2251 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002252 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002253 if (!cic)
2254 return ELV_MQUEUE_MAY;
2255
Jens Axboeb0b78f82009-04-08 10:56:08 +02002256 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02002257 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002258 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02002259 cfq_prio_boost(cfqq);
2260
Jens Axboe89850f72006-07-22 16:48:31 +02002261 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002262 }
2263
2264 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265}
2266
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267/*
2268 * queue lock held here
2269 */
Jens Axboebb37b942006-12-01 10:42:33 +01002270static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271{
Jens Axboe5e705372006-07-13 12:39:25 +02002272 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273
Jens Axboe5e705372006-07-13 12:39:25 +02002274 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002275 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276
Jens Axboe22e2c502005-06-27 10:55:12 +02002277 BUG_ON(!cfqq->allocated[rw]);
2278 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279
Jens Axboe5e705372006-07-13 12:39:25 +02002280 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02002283 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 cfq_put_queue(cfqq);
2286 }
2287}
2288
2289/*
Jens Axboe22e2c502005-06-27 10:55:12 +02002290 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291 */
Jens Axboe22e2c502005-06-27 10:55:12 +02002292static int
Jens Axboe165125e2007-07-24 09:28:11 +02002293cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294{
2295 struct cfq_data *cfqd = q->elevator->elevator_data;
2296 struct cfq_io_context *cic;
2297 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01002298 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002299 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 unsigned long flags;
2301
2302 might_sleep_if(gfp_mask & __GFP_WAIT);
2303
Jens Axboee2d74ac2006-03-28 08:59:01 +02002304 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02002305
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306 spin_lock_irqsave(q->queue_lock, flags);
2307
Jens Axboe22e2c502005-06-27 10:55:12 +02002308 if (!cic)
2309 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310
Vasily Tarasov91fac312007-04-25 12:29:51 +02002311 cfqq = cic_to_cfqq(cic, is_sync);
2312 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002313 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002314
Jens Axboe22e2c502005-06-27 10:55:12 +02002315 if (!cfqq)
2316 goto queue_fail;
2317
Vasily Tarasov91fac312007-04-25 12:29:51 +02002318 cic_set_cfqq(cic, cfqq, is_sync);
2319 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320
2321 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02002322 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002323 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02002324
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 spin_unlock_irqrestore(q->queue_lock, flags);
2326
Jens Axboe5e705372006-07-13 12:39:25 +02002327 rq->elevator_private = cic;
2328 rq->elevator_private2 = cfqq;
2329 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02002330
Jens Axboe22e2c502005-06-27 10:55:12 +02002331queue_fail:
2332 if (cic)
2333 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02002334
Jens Axboe3b181522005-06-27 10:56:24 +02002335 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02002337 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 return 1;
2339}
2340
David Howells65f27f32006-11-22 14:55:48 +00002341static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02002342{
David Howells65f27f32006-11-22 14:55:48 +00002343 struct cfq_data *cfqd =
2344 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02002345 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002346
Jens Axboe40bb54d2009-04-15 12:11:10 +02002347 spin_lock_irq(q->queue_lock);
Jens Axboedc72ef42006-07-20 14:54:05 +02002348 blk_start_queueing(q);
Jens Axboe40bb54d2009-04-15 12:11:10 +02002349 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02002350}
2351
2352/*
2353 * Timer running if the active_queue is currently idling inside its time slice
2354 */
2355static void cfq_idle_slice_timer(unsigned long data)
2356{
2357 struct cfq_data *cfqd = (struct cfq_data *) data;
2358 struct cfq_queue *cfqq;
2359 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002360 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002361
Jens Axboe7b679132008-05-30 12:23:07 +02002362 cfq_log(cfqd, "idle timer fired");
2363
Jens Axboe22e2c502005-06-27 10:55:12 +02002364 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2365
Jens Axboefe094d92008-01-31 13:08:54 +01002366 cfqq = cfqd->active_queue;
2367 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002368 timed_out = 0;
2369
Jens Axboe22e2c502005-06-27 10:55:12 +02002370 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002371 * We saw a request before the queue expired, let it through
2372 */
2373 if (cfq_cfqq_must_dispatch(cfqq))
2374 goto out_kick;
2375
2376 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002377 * expired
2378 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002379 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002380 goto expire;
2381
2382 /*
2383 * only expire and reinvoke request handler, if there are
2384 * other queues with pending requests
2385 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002386 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002387 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002388
2389 /*
2390 * not expired and it has a request pending, let it dispatch
2391 */
Jens Axboe75e50982009-04-07 08:56:14 +02002392 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002393 goto out_kick;
Jens Axboe22e2c502005-06-27 10:55:12 +02002394 }
2395expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002396 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002397out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002398 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002399out_cont:
2400 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2401}
2402
Jens Axboe3b181522005-06-27 10:56:24 +02002403static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2404{
2405 del_timer_sync(&cfqd->idle_slice_timer);
Cheng Renquan64d01dc2008-12-03 12:41:39 +01002406 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002407}
Jens Axboe22e2c502005-06-27 10:55:12 +02002408
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002409static void cfq_put_async_queues(struct cfq_data *cfqd)
2410{
2411 int i;
2412
2413 for (i = 0; i < IOPRIO_BE_NR; i++) {
2414 if (cfqd->async_cfqq[0][i])
2415 cfq_put_queue(cfqd->async_cfqq[0][i]);
2416 if (cfqd->async_cfqq[1][i])
2417 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002418 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002419
2420 if (cfqd->async_idle_cfqq)
2421 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002422}
2423
Jens Axboeb374d182008-10-31 10:05:07 +01002424static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425{
Jens Axboe22e2c502005-06-27 10:55:12 +02002426 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002427 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002428
Jens Axboe3b181522005-06-27 10:56:24 +02002429 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002430
Al Virod9ff4182006-03-18 13:51:22 -05002431 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002432
Al Virod9ff4182006-03-18 13:51:22 -05002433 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002434 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002435
2436 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002437 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2438 struct cfq_io_context,
2439 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002440
2441 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002442 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002443
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002444 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002445
Al Virod9ff4182006-03-18 13:51:22 -05002446 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002447
2448 cfq_shutdown_timer_wq(cfqd);
2449
Al Viroa90d7422006-03-18 12:05:37 -05002450 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451}
2452
Jens Axboe165125e2007-07-24 09:28:11 +02002453static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454{
2455 struct cfq_data *cfqd;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002456 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457
Christoph Lameter94f60302007-07-17 04:03:29 -07002458 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002460 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461
Jens Axboecc09e292007-04-26 12:53:50 +02002462 cfqd->service_tree = CFQ_RB_ROOT;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002463
2464 /*
2465 * Not strictly needed (since RB_ROOT just clears the node and we
2466 * zeroed cfqd on alloc), but better be safe in case someone decides
2467 * to add magic to the rb code
2468 */
2469 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2470 cfqd->prio_trees[i] = RB_ROOT;
2471
Al Virod9ff4182006-03-18 13:51:22 -05002472 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475
Jens Axboe22e2c502005-06-27 10:55:12 +02002476 init_timer(&cfqd->idle_slice_timer);
2477 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2478 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2479
David Howells65f27f32006-11-22 14:55:48 +00002480 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002481
Oleg Nesterovb70c8642007-11-07 09:46:13 +01002482 cfqd->last_end_request = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002484 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2485 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486 cfqd->cfq_back_max = cfq_back_max;
2487 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002488 cfqd->cfq_slice[0] = cfq_slice_async;
2489 cfqd->cfq_slice[1] = cfq_slice_sync;
2490 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2491 cfqd->cfq_slice_idle = cfq_slice_idle;
Aaron Carroll45333d52008-08-26 15:52:36 +02002492 cfqd->hw_tag = 1;
Jens Axboe3b181522005-06-27 10:56:24 +02002493
Jens Axboebc1c1162006-06-08 08:49:06 +02002494 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495}
2496
2497static void cfq_slab_kill(void)
2498{
Jens Axboed6de8be2008-05-28 14:46:59 +02002499 /*
2500 * Caller already ensured that pending RCU callbacks are completed,
2501 * so we should have no busy allocations at this point.
2502 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503 if (cfq_pool)
2504 kmem_cache_destroy(cfq_pool);
2505 if (cfq_ioc_pool)
2506 kmem_cache_destroy(cfq_ioc_pool);
2507}
2508
2509static int __init cfq_slab_setup(void)
2510{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002511 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512 if (!cfq_pool)
2513 goto fail;
2514
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002515 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 if (!cfq_ioc_pool)
2517 goto fail;
2518
2519 return 0;
2520fail:
2521 cfq_slab_kill();
2522 return -ENOMEM;
2523}
2524
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525/*
2526 * sysfs parts below -->
2527 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528static ssize_t
2529cfq_var_show(unsigned int var, char *page)
2530{
2531 return sprintf(page, "%d\n", var);
2532}
2533
2534static ssize_t
2535cfq_var_store(unsigned int *var, const char *page, size_t count)
2536{
2537 char *p = (char *) page;
2538
2539 *var = simple_strtoul(p, &p, 10);
2540 return count;
2541}
2542
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002544static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002546 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 unsigned int __data = __VAR; \
2548 if (__CONV) \
2549 __data = jiffies_to_msecs(__data); \
2550 return cfq_var_show(__data, (page)); \
2551}
2552SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002553SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2554SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002555SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2556SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002557SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2558SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2559SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2560SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561#undef SHOW_FUNCTION
2562
2563#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002564static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002566 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 unsigned int __data; \
2568 int ret = cfq_var_store(&__data, (page), count); \
2569 if (__data < (MIN)) \
2570 __data = (MIN); \
2571 else if (__data > (MAX)) \
2572 __data = (MAX); \
2573 if (__CONV) \
2574 *(__PTR) = msecs_to_jiffies(__data); \
2575 else \
2576 *(__PTR) = __data; \
2577 return ret; \
2578}
2579STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002580STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2581 UINT_MAX, 1);
2582STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2583 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002584STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002585STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2586 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002587STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2588STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2589STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01002590STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2591 UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592#undef STORE_FUNCTION
2593
Al Viroe572ec72006-03-18 22:27:18 -05002594#define CFQ_ATTR(name) \
2595 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002596
Al Viroe572ec72006-03-18 22:27:18 -05002597static struct elv_fs_entry cfq_attrs[] = {
2598 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002599 CFQ_ATTR(fifo_expire_sync),
2600 CFQ_ATTR(fifo_expire_async),
2601 CFQ_ATTR(back_seek_max),
2602 CFQ_ATTR(back_seek_penalty),
2603 CFQ_ATTR(slice_sync),
2604 CFQ_ATTR(slice_async),
2605 CFQ_ATTR(slice_async_rq),
2606 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002607 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608};
2609
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610static struct elevator_type iosched_cfq = {
2611 .ops = {
2612 .elevator_merge_fn = cfq_merge,
2613 .elevator_merged_fn = cfq_merged_request,
2614 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002615 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002616 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002618 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 .elevator_deactivate_req_fn = cfq_deactivate_request,
2620 .elevator_queue_empty_fn = cfq_queue_empty,
2621 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002622 .elevator_former_req_fn = elv_rb_former_request,
2623 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624 .elevator_set_req_fn = cfq_set_request,
2625 .elevator_put_req_fn = cfq_put_request,
2626 .elevator_may_queue_fn = cfq_may_queue,
2627 .elevator_init_fn = cfq_init_queue,
2628 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002629 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630 },
Al Viro3d1ab402006-03-18 18:35:43 -05002631 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 .elevator_name = "cfq",
2633 .elevator_owner = THIS_MODULE,
2634};
2635
2636static int __init cfq_init(void)
2637{
Jens Axboe22e2c502005-06-27 10:55:12 +02002638 /*
2639 * could be 0 on HZ < 1000 setups
2640 */
2641 if (!cfq_slice_async)
2642 cfq_slice_async = 1;
2643 if (!cfq_slice_idle)
2644 cfq_slice_idle = 1;
2645
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646 if (cfq_slab_setup())
2647 return -ENOMEM;
2648
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002649 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002651 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652}
2653
2654static void __exit cfq_exit(void)
2655{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002656 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002658 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002659 /* ioc_gone's update must be visible before reading ioc_count */
2660 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02002661
2662 /*
2663 * this also protects us from entering cfq_slab_kill() with
2664 * pending RCU callbacks
2665 */
Jens Axboe4050cf12006-07-19 05:07:12 +02002666 if (elv_ioc_count_read(ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002667 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08002668 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669}
2670
2671module_init(cfq_init);
2672module_exit(cfq_exit);
2673
2674MODULE_AUTHOR("Jens Axboe");
2675MODULE_LICENSE("GPL");
2676MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");