blob: 833ec18eaa63e2a20b847797c99511e41314b474 [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;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 /*
127 * tunables, see top of file
128 */
129 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200130 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 unsigned int cfq_back_penalty;
132 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200133 unsigned int cfq_slice[2];
134 unsigned int cfq_slice_async_rq;
135 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500136
137 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138};
139
Jens Axboe22e2c502005-06-27 10:55:12 +0200140/*
141 * Per process-grouping structure
142 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143struct cfq_queue {
144 /* reference count */
145 atomic_t ref;
Richard Kennedybe754d22008-05-23 06:52:00 +0200146 /* various state flags, see below */
147 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148 /* parent cfq_data */
149 struct cfq_data *cfqd;
Jens Axboed9e76202007-04-20 14:27:50 +0200150 /* service_tree member */
151 struct rb_node rb_node;
152 /* service_tree key */
153 unsigned long rb_key;
Jens Axboea36e71f2009-04-15 12:15:11 +0200154 /* prio tree member */
155 struct rb_node p_node;
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200156 /* prio tree root we belong to, if any */
157 struct rb_root *p_root;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158 /* sorted list of pending requests */
159 struct rb_root sort_list;
160 /* if fifo isn't expired, next request to serve */
Jens Axboe5e705372006-07-13 12:39:25 +0200161 struct request *next_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162 /* requests queued in sort_list */
163 int queued[2];
164 /* currently allocated requests */
165 int allocated[2];
166 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200167 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168
Jens Axboe22e2c502005-06-27 10:55:12 +0200169 unsigned long slice_end;
Jens Axboec5b680f2007-01-19 11:56:49 +1100170 long slice_resid;
Jens Axboe2f5cb732009-04-07 08:51:19 +0200171 unsigned int slice_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172
Richard Kennedybe754d22008-05-23 06:52:00 +0200173 /* pending metadata requests */
174 int meta_pending;
Jens Axboe6d048f52007-04-25 12:44:27 +0200175 /* number of requests that are on the dispatch list or inside driver */
176 int dispatched;
Jens Axboe22e2c502005-06-27 10:55:12 +0200177
178 /* io prio of this group */
179 unsigned short ioprio, org_ioprio;
180 unsigned short ioprio_class, org_ioprio_class;
181
Jens Axboe7b679132008-05-30 12:23:07 +0200182 pid_t pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183};
184
Jens Axboe3b181522005-06-27 10:56:24 +0200185enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100186 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
187 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200188 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100189 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
190 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100191 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
192 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
193 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100194 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200195 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboea36e71f2009-04-15 12:15:11 +0200196 CFQ_CFQQ_FLAG_coop, /* has done a coop jump of the queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200197};
198
199#define CFQ_CFQQ_FNS(name) \
200static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
201{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100202 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200203} \
204static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
205{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100206 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200207} \
208static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
209{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100210 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200211}
212
213CFQ_CFQQ_FNS(on_rr);
214CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200215CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200216CFQ_CFQQ_FNS(must_alloc);
217CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200218CFQ_CFQQ_FNS(fifo_expire);
219CFQ_CFQQ_FNS(idle_window);
220CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100221CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200222CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200223CFQ_CFQQ_FNS(coop);
Jens Axboe3b181522005-06-27 10:56:24 +0200224#undef CFQ_CFQQ_FNS
225
Jens Axboe7b679132008-05-30 12:23:07 +0200226#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
227 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
228#define cfq_log(cfqd, fmt, args...) \
229 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
230
Jens Axboe165125e2007-07-24 09:28:11 +0200231static void cfq_dispatch_insert(struct request_queue *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200232static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
Jens Axboefd0928d2008-01-24 08:52:45 +0100233 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100234static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200235 struct io_context *);
236
237static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
238 int is_sync)
239{
240 return cic->cfqq[!!is_sync];
241}
242
243static inline void cic_set_cfqq(struct cfq_io_context *cic,
244 struct cfq_queue *cfqq, int is_sync)
245{
246 cic->cfqq[!!is_sync] = cfqq;
247}
248
249/*
250 * We regard a request as SYNC, if it's either a read or has the SYNC bit
251 * set (in which case it could also be direct WRITE).
252 */
253static inline int cfq_bio_sync(struct bio *bio)
254{
255 if (bio_data_dir(bio) == READ || bio_sync(bio))
256 return 1;
257
258 return 0;
259}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700262 * scheduler run of queue, if there are requests pending and no one in the
263 * driver that will restart queueing
264 */
265static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
266{
Jens Axboe7b679132008-05-30 12:23:07 +0200267 if (cfqd->busy_queues) {
268 cfq_log(cfqd, "schedule dispatch");
Jens Axboe18887ad2008-07-28 13:08:45 +0200269 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200270 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700271}
272
Jens Axboe165125e2007-07-24 09:28:11 +0200273static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700274{
275 struct cfq_data *cfqd = q->elevator->elevator_data;
276
Jens Axboeb4878f22005-10-20 16:42:29 +0200277 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700278}
279
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100281 * Scale schedule slice based on io priority. Use the sync time slice only
282 * if a queue is marked sync and has sync io queued. A sync queue with async
283 * io only, should not get full sync slice length.
284 */
Jens Axboed9e76202007-04-20 14:27:50 +0200285static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
286 unsigned short prio)
287{
288 const int base_slice = cfqd->cfq_slice[sync];
289
290 WARN_ON(prio >= IOPRIO_BE_NR);
291
292 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
293}
294
Jens Axboe44f7c162007-01-19 11:51:58 +1100295static inline int
296cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
297{
Jens Axboed9e76202007-04-20 14:27:50 +0200298 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100299}
300
301static inline void
302cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
303{
304 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200305 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100306}
307
308/*
309 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
310 * isn't valid until the first request from the dispatch is activated
311 * and the slice time set.
312 */
313static inline int cfq_slice_used(struct cfq_queue *cfqq)
314{
315 if (cfq_cfqq_slice_new(cfqq))
316 return 0;
317 if (time_before(jiffies, cfqq->slice_end))
318 return 0;
319
320 return 1;
321}
322
323/*
Jens Axboe5e705372006-07-13 12:39:25 +0200324 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200326 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327 */
Jens Axboe5e705372006-07-13 12:39:25 +0200328static struct request *
329cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330{
331 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200333#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
334#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
335 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336
Jens Axboe5e705372006-07-13 12:39:25 +0200337 if (rq1 == NULL || rq1 == rq2)
338 return rq2;
339 if (rq2 == NULL)
340 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200341
Jens Axboe5e705372006-07-13 12:39:25 +0200342 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
343 return rq1;
344 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
345 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200346 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
347 return rq1;
348 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
349 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350
Tejun Heo83096eb2009-05-07 22:24:39 +0900351 s1 = blk_rq_pos(rq1);
352 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353
Jens Axboe6d048f52007-04-25 12:44:27 +0200354 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356 /*
357 * by definition, 1KiB is 2 sectors
358 */
359 back_max = cfqd->cfq_back_max * 2;
360
361 /*
362 * Strict one way elevator _except_ in the case where we allow
363 * short backward seeks which are biased as twice the cost of a
364 * similar forward seek.
365 */
366 if (s1 >= last)
367 d1 = s1 - last;
368 else if (s1 + back_max >= last)
369 d1 = (last - s1) * cfqd->cfq_back_penalty;
370 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200371 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372
373 if (s2 >= last)
374 d2 = s2 - last;
375 else if (s2 + back_max >= last)
376 d2 = (last - s2) * cfqd->cfq_back_penalty;
377 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200378 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379
380 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381
Andreas Mohre8a99052006-03-28 08:59:49 +0200382 /*
383 * By doing switch() on the bit mask "wrap" we avoid having to
384 * check two variables for all permutations: --> faster!
385 */
386 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200387 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200388 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200389 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200390 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200391 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200392 else {
393 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200394 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200395 else
Jens Axboe5e705372006-07-13 12:39:25 +0200396 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200397 }
398
399 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200400 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200401 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200402 return rq2;
403 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200404 default:
405 /*
406 * Since both rqs are wrapped,
407 * start with the one that's further behind head
408 * (--> only *one* back seek required),
409 * since back seek takes more time than forward.
410 */
411 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200412 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 else
Jens Axboe5e705372006-07-13 12:39:25 +0200414 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 }
416}
417
Jens Axboe498d3aa22007-04-26 12:54:48 +0200418/*
419 * The below is leftmost cache rbtree addon
420 */
Jens Axboe08717142008-01-28 11:38:15 +0100421static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200422{
423 if (!root->left)
424 root->left = rb_first(&root->rb);
425
Jens Axboe08717142008-01-28 11:38:15 +0100426 if (root->left)
427 return rb_entry(root->left, struct cfq_queue, rb_node);
428
429 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200430}
431
Jens Axboea36e71f2009-04-15 12:15:11 +0200432static void rb_erase_init(struct rb_node *n, struct rb_root *root)
433{
434 rb_erase(n, root);
435 RB_CLEAR_NODE(n);
436}
437
Jens Axboecc09e292007-04-26 12:53:50 +0200438static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
439{
440 if (root->left == n)
441 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200442 rb_erase_init(n, &root->rb);
Jens Axboecc09e292007-04-26 12:53:50 +0200443}
444
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445/*
446 * would be nice to take fifo expire time into account as well
447 */
Jens Axboe5e705372006-07-13 12:39:25 +0200448static struct request *
449cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
450 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451{
Jens Axboe21183b02006-07-13 12:33:14 +0200452 struct rb_node *rbnext = rb_next(&last->rb_node);
453 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200454 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
Jens Axboe21183b02006-07-13 12:33:14 +0200456 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457
458 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200459 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200460
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200462 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200463 else {
464 rbnext = rb_first(&cfqq->sort_list);
465 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200466 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200467 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468
Jens Axboe21183b02006-07-13 12:33:14 +0200469 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470}
471
Jens Axboed9e76202007-04-20 14:27:50 +0200472static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
473 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474{
Jens Axboed9e76202007-04-20 14:27:50 +0200475 /*
476 * just an approximation, should be ok.
477 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200478 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
479 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200480}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481
Jens Axboe498d3aa22007-04-26 12:54:48 +0200482/*
483 * The cfqd->service_tree holds all pending cfq_queue's that have
484 * requests waiting to be processed. It is sorted in the order that
485 * we will service the queues.
486 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200487static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
488 int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200489{
Jens Axboe08717142008-01-28 11:38:15 +0100490 struct rb_node **p, *parent;
491 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +0200492 unsigned long rb_key;
Jens Axboe498d3aa22007-04-26 12:54:48 +0200493 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200494
Jens Axboe08717142008-01-28 11:38:15 +0100495 if (cfq_class_idle(cfqq)) {
496 rb_key = CFQ_IDLE_DELAY;
497 parent = rb_last(&cfqd->service_tree.rb);
498 if (parent && parent != &cfqq->rb_node) {
499 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
500 rb_key += __cfqq->rb_key;
501 } else
502 rb_key += jiffies;
503 } else if (!add_front) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200504 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
505 rb_key += cfqq->slice_resid;
506 cfqq->slice_resid = 0;
507 } else
508 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200509
510 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100511 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200512 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100513 */
Jens Axboed9e76202007-04-20 14:27:50 +0200514 if (rb_key == cfqq->rb_key)
515 return;
Jens Axboe53b037442006-07-28 09:48:51 +0200516
Jens Axboecc09e292007-04-26 12:53:50 +0200517 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200518 }
Jens Axboed9e76202007-04-20 14:27:50 +0200519
Jens Axboe498d3aa22007-04-26 12:54:48 +0200520 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +0100521 parent = NULL;
522 p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200523 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +0200524 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200525
Jens Axboed9e76202007-04-20 14:27:50 +0200526 parent = *p;
527 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
528
Jens Axboe0c534e02007-04-18 20:01:57 +0200529 /*
530 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200531 * preference to them. IDLE queues goes to the back.
532 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200533 */
534 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200535 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200536 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200537 n = &(*p)->rb_right;
538 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
539 n = &(*p)->rb_left;
540 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
541 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200542 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200543 n = &(*p)->rb_left;
544 else
545 n = &(*p)->rb_right;
546
547 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200548 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200549
550 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200551 }
552
Jens Axboecc09e292007-04-26 12:53:50 +0200553 if (left)
554 cfqd->service_tree.left = &cfqq->rb_node;
555
Jens Axboed9e76202007-04-20 14:27:50 +0200556 cfqq->rb_key = rb_key;
557 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200558 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559}
560
Jens Axboea36e71f2009-04-15 12:15:11 +0200561static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200562cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
563 sector_t sector, struct rb_node **ret_parent,
564 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +0200565{
Jens Axboea36e71f2009-04-15 12:15:11 +0200566 struct rb_node **p, *parent;
567 struct cfq_queue *cfqq = NULL;
568
569 parent = NULL;
570 p = &root->rb_node;
571 while (*p) {
572 struct rb_node **n;
573
574 parent = *p;
575 cfqq = rb_entry(parent, struct cfq_queue, p_node);
576
577 /*
578 * Sort strictly based on sector. Smallest to the left,
579 * largest to the right.
580 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900581 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +0200582 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900583 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +0200584 n = &(*p)->rb_left;
585 else
586 break;
587 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200588 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200589 }
590
591 *ret_parent = parent;
592 if (rb_link)
593 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200594 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +0200595}
596
597static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
598{
Jens Axboea36e71f2009-04-15 12:15:11 +0200599 struct rb_node **p, *parent;
600 struct cfq_queue *__cfqq;
601
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200602 if (cfqq->p_root) {
603 rb_erase(&cfqq->p_node, cfqq->p_root);
604 cfqq->p_root = NULL;
605 }
Jens Axboea36e71f2009-04-15 12:15:11 +0200606
607 if (cfq_class_idle(cfqq))
608 return;
609 if (!cfqq->next_rq)
610 return;
611
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200612 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900613 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
614 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200615 if (!__cfqq) {
616 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200617 rb_insert_color(&cfqq->p_node, cfqq->p_root);
618 } else
619 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200620}
621
Jens Axboe498d3aa22007-04-26 12:54:48 +0200622/*
623 * Update cfqq's position in the service tree.
624 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200625static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200626{
Jens Axboe6d048f52007-04-25 12:44:27 +0200627 /*
628 * Resorting requires the cfqq to be on the RR list already.
629 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200630 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200631 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +0200632 cfq_prio_tree_add(cfqd, cfqq);
633 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200634}
635
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636/*
637 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200638 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 */
Jens Axboefebffd62008-01-28 13:19:43 +0100640static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641{
Jens Axboe7b679132008-05-30 12:23:07 +0200642 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200643 BUG_ON(cfq_cfqq_on_rr(cfqq));
644 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 cfqd->busy_queues++;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +0100646 if (cfq_class_rt(cfqq))
647 cfqd->busy_rt_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648
Jens Axboeedd75ff2007-04-19 12:03:34 +0200649 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650}
651
Jens Axboe498d3aa22007-04-26 12:54:48 +0200652/*
653 * Called when the cfqq no longer has requests pending, remove it from
654 * the service tree.
655 */
Jens Axboefebffd62008-01-28 13:19:43 +0100656static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657{
Jens Axboe7b679132008-05-30 12:23:07 +0200658 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200659 BUG_ON(!cfq_cfqq_on_rr(cfqq));
660 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661
Jens Axboecc09e292007-04-26 12:53:50 +0200662 if (!RB_EMPTY_NODE(&cfqq->rb_node))
663 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200664 if (cfqq->p_root) {
665 rb_erase(&cfqq->p_node, cfqq->p_root);
666 cfqq->p_root = NULL;
667 }
Jens Axboed9e76202007-04-20 14:27:50 +0200668
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 BUG_ON(!cfqd->busy_queues);
670 cfqd->busy_queues--;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +0100671 if (cfq_class_rt(cfqq))
672 cfqd->busy_rt_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673}
674
675/*
676 * rb tree support functions
677 */
Jens Axboefebffd62008-01-28 13:19:43 +0100678static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679{
Jens Axboe5e705372006-07-13 12:39:25 +0200680 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200681 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200682 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
Jens Axboeb4878f22005-10-20 16:42:29 +0200684 BUG_ON(!cfqq->queued[sync]);
685 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686
Jens Axboe5e705372006-07-13 12:39:25 +0200687 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688
Jens Axboedd67d052006-06-21 09:36:18 +0200689 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200690 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691}
692
Jens Axboe5e705372006-07-13 12:39:25 +0200693static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694{
Jens Axboe5e705372006-07-13 12:39:25 +0200695 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +0200697 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698
Jens Axboe5380a102006-07-13 12:37:56 +0200699 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700
701 /*
702 * looks a little odd, but the first insert might return an alias.
703 * if that happens, put the alias on the dispatch list
704 */
Jens Axboe21183b02006-07-13 12:33:14 +0200705 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200706 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100707
708 if (!cfq_cfqq_on_rr(cfqq))
709 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200710
711 /*
712 * check if this request is a better next-serve candidate
713 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200714 prev = cfqq->next_rq;
Jens Axboe5044eed2007-04-25 11:53:48 +0200715 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
Jens Axboea36e71f2009-04-15 12:15:11 +0200716
717 /*
718 * adjust priority tree position, if ->next_rq changes
719 */
720 if (prev != cfqq->next_rq)
721 cfq_prio_tree_add(cfqd, cfqq);
722
Jens Axboe5044eed2007-04-25 11:53:48 +0200723 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724}
725
Jens Axboefebffd62008-01-28 13:19:43 +0100726static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727{
Jens Axboe5380a102006-07-13 12:37:56 +0200728 elv_rb_del(&cfqq->sort_list, rq);
729 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200730 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731}
732
Jens Axboe206dc692006-03-28 13:03:44 +0200733static struct request *
734cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735{
Jens Axboe206dc692006-03-28 13:03:44 +0200736 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200737 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200738 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739
Jens Axboe4ac845a2008-01-24 08:44:49 +0100740 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200741 if (!cic)
742 return NULL;
743
744 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200745 if (cfqq) {
746 sector_t sector = bio->bi_sector + bio_sectors(bio);
747
Jens Axboe21183b02006-07-13 12:33:14 +0200748 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200749 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 return NULL;
752}
753
Jens Axboe165125e2007-07-24 09:28:11 +0200754static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200755{
756 struct cfq_data *cfqd = q->elevator->elevator_data;
757
758 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +0200759 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
760 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +0200761
Tejun Heo5b936292009-05-07 22:24:38 +0900762 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200763}
764
Jens Axboe165125e2007-07-24 09:28:11 +0200765static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766{
Jens Axboe22e2c502005-06-27 10:55:12 +0200767 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768
Jens Axboeb4878f22005-10-20 16:42:29 +0200769 WARN_ON(!cfqd->rq_in_driver);
770 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +0200771 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
772 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773}
774
Jens Axboeb4878f22005-10-20 16:42:29 +0200775static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776{
Jens Axboe5e705372006-07-13 12:39:25 +0200777 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200778
Jens Axboe5e705372006-07-13 12:39:25 +0200779 if (cfqq->next_rq == rq)
780 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781
Jens Axboeb4878f22005-10-20 16:42:29 +0200782 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200783 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200784
Aaron Carroll45333d52008-08-26 15:52:36 +0200785 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +0200786 if (rq_is_meta(rq)) {
787 WARN_ON(!cfqq->meta_pending);
788 cfqq->meta_pending--;
789 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790}
791
Jens Axboe165125e2007-07-24 09:28:11 +0200792static int cfq_merge(struct request_queue *q, struct request **req,
793 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794{
795 struct cfq_data *cfqd = q->elevator->elevator_data;
796 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797
Jens Axboe206dc692006-03-28 13:03:44 +0200798 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200799 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200800 *req = __rq;
801 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 }
803
804 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805}
806
Jens Axboe165125e2007-07-24 09:28:11 +0200807static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200808 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809{
Jens Axboe21183b02006-07-13 12:33:14 +0200810 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200811 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812
Jens Axboe5e705372006-07-13 12:39:25 +0200813 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815}
816
817static void
Jens Axboe165125e2007-07-24 09:28:11 +0200818cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 struct request *next)
820{
Jens Axboe22e2c502005-06-27 10:55:12 +0200821 /*
822 * reposition in fifo if next is older than rq
823 */
824 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
825 time_before(next->start_time, rq->start_time))
826 list_move(&rq->queuelist, &next->queuelist);
827
Jens Axboeb4878f22005-10-20 16:42:29 +0200828 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200829}
830
Jens Axboe165125e2007-07-24 09:28:11 +0200831static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100832 struct bio *bio)
833{
834 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200835 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100836 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100837
838 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100839 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100840 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200841 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100842 return 0;
843
844 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100845 * Lookup the cfqq that this bio will be queued with. Allow
846 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100847 */
Jens Axboe4ac845a2008-01-24 08:44:49 +0100848 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200849 if (!cic)
850 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100851
Vasily Tarasov91fac312007-04-25 12:29:51 +0200852 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100853 if (cfqq == RQ_CFQQ(rq))
854 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100855
Jens Axboeec8acb62007-01-02 18:32:11 +0100856 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100857}
858
Jens Axboefebffd62008-01-28 13:19:43 +0100859static void __cfq_set_active_queue(struct cfq_data *cfqd,
860 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +0200861{
862 if (cfqq) {
Jens Axboe7b679132008-05-30 12:23:07 +0200863 cfq_log_cfqq(cfqd, cfqq, "set_active");
Jens Axboe22e2c502005-06-27 10:55:12 +0200864 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +0200865 cfqq->slice_dispatch = 0;
866
Jens Axboe2f5cb732009-04-07 08:51:19 +0200867 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +0200868 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200869 cfq_clear_cfqq_must_alloc_slice(cfqq);
870 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100871 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +0200872
873 del_timer(&cfqd->idle_slice_timer);
Jens Axboe22e2c502005-06-27 10:55:12 +0200874 }
875
876 cfqd->active_queue = cfqq;
877}
878
879/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100880 * current cfqq expired its slice (or was too idle), select new one
881 */
882static void
883__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200884 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100885{
Jens Axboe7b679132008-05-30 12:23:07 +0200886 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
887
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100888 if (cfq_cfqq_wait_request(cfqq))
889 del_timer(&cfqd->idle_slice_timer);
890
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100891 cfq_clear_cfqq_wait_request(cfqq);
892
893 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200894 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100895 */
Jens Axboe7b679132008-05-30 12:23:07 +0200896 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +1100897 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200898 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
899 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100900
Jens Axboeedd75ff2007-04-19 12:03:34 +0200901 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100902
903 if (cfqq == cfqd->active_queue)
904 cfqd->active_queue = NULL;
905
906 if (cfqd->active_cic) {
907 put_io_context(cfqd->active_cic->ioc);
908 cfqd->active_cic = NULL;
909 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100910}
911
Jens Axboe6084cdd2007-04-23 08:25:00 +0200912static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100913{
914 struct cfq_queue *cfqq = cfqd->active_queue;
915
916 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200917 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100918}
919
Jens Axboe498d3aa22007-04-26 12:54:48 +0200920/*
921 * Get next queue for service. Unless we have a queue preemption,
922 * we'll simply select the first cfqq in the service tree.
923 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200924static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200925{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200926 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
927 return NULL;
928
Jens Axboe08717142008-01-28 11:38:15 +0100929 return cfq_rb_first(&cfqd->service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +0200930}
931
Jens Axboe498d3aa22007-04-26 12:54:48 +0200932/*
933 * Get and set a new active queue for service.
934 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200935static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
936 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200937{
Jens Axboea36e71f2009-04-15 12:15:11 +0200938 if (!cfqq) {
939 cfqq = cfq_get_next_queue(cfqd);
940 if (cfqq)
941 cfq_clear_cfqq_coop(cfqq);
942 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200943
Jens Axboe22e2c502005-06-27 10:55:12 +0200944 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200945 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200946}
947
Jens Axboed9e76202007-04-20 14:27:50 +0200948static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
949 struct request *rq)
950{
Tejun Heo83096eb2009-05-07 22:24:39 +0900951 if (blk_rq_pos(rq) >= cfqd->last_position)
952 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +0200953 else
Tejun Heo83096eb2009-05-07 22:24:39 +0900954 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +0200955}
956
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200957#define CIC_SEEK_THR 8 * 1024
958#define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
959
Jens Axboe6d048f52007-04-25 12:44:27 +0200960static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
961{
962 struct cfq_io_context *cic = cfqd->active_cic;
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200963 sector_t sdist = cic->seek_mean;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200964
Jens Axboe6d048f52007-04-25 12:44:27 +0200965 if (!sample_valid(cic->seek_samples))
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200966 sdist = CIC_SEEK_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +0200967
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200968 return cfq_dist_from_last(cfqd, rq) <= sdist;
Jens Axboe6d048f52007-04-25 12:44:27 +0200969}
970
Jens Axboea36e71f2009-04-15 12:15:11 +0200971static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
972 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200973{
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200974 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +0200975 struct rb_node *parent, *node;
976 struct cfq_queue *__cfqq;
977 sector_t sector = cfqd->last_position;
978
979 if (RB_EMPTY_ROOT(root))
980 return NULL;
981
982 /*
983 * First, if we find a request starting at the end of the last
984 * request, choose it.
985 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200986 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +0200987 if (__cfqq)
988 return __cfqq;
989
990 /*
991 * If the exact sector wasn't found, the parent of the NULL leaf
992 * will contain the closest sector.
993 */
994 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
995 if (cfq_rq_close(cfqd, __cfqq->next_rq))
996 return __cfqq;
997
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900998 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +0200999 node = rb_next(&__cfqq->p_node);
1000 else
1001 node = rb_prev(&__cfqq->p_node);
1002 if (!node)
1003 return NULL;
1004
1005 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1006 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1007 return __cfqq;
1008
1009 return NULL;
1010}
1011
1012/*
1013 * cfqd - obvious
1014 * cur_cfqq - passed in so that we don't decide that the current queue is
1015 * closely cooperating with itself.
1016 *
1017 * So, basically we're assuming that that cur_cfqq has dispatched at least
1018 * one request, and that cfqd->last_position reflects a position on the disk
1019 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1020 * assumption.
1021 */
1022static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1023 struct cfq_queue *cur_cfqq,
1024 int probe)
1025{
1026 struct cfq_queue *cfqq;
1027
1028 /*
1029 * A valid cfq_io_context is necessary to compare requests against
1030 * the seek_mean of the current cfqq.
1031 */
1032 if (!cfqd->active_cic)
1033 return NULL;
1034
Jens Axboe6d048f52007-04-25 12:44:27 +02001035 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001036 * We should notice if some of the queues are cooperating, eg
1037 * working closely on the same area of the disk. In that case,
1038 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001039 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001040 cfqq = cfqq_close(cfqd, cur_cfqq);
1041 if (!cfqq)
1042 return NULL;
1043
1044 if (cfq_cfqq_coop(cfqq))
1045 return NULL;
1046
1047 if (!probe)
1048 cfq_mark_cfqq_coop(cfqq);
1049 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001050}
1051
Jens Axboe6d048f52007-04-25 12:44:27 +02001052static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001053{
Jens Axboe17926692007-01-19 11:59:30 +11001054 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001055 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001056 unsigned long sl;
1057
Jens Axboea68bbdd2008-09-24 13:03:33 +02001058 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001059 * SSD device without seek penalty, disable idling. But only do so
1060 * for devices that support queuing, otherwise we still have a problem
1061 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001062 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001063 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001064 return;
1065
Jens Axboedd67d052006-06-21 09:36:18 +02001066 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001067 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001068
1069 /*
1070 * idle is disabled, either manually or by past process history
1071 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001072 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1073 return;
1074
Jens Axboe22e2c502005-06-27 10:55:12 +02001075 /*
Jens Axboe7b679132008-05-30 12:23:07 +02001076 * still requests with the driver, don't idle
1077 */
1078 if (cfqd->rq_in_driver)
1079 return;
1080
1081 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001082 * task has exited, don't wait
1083 */
Jens Axboe206dc692006-03-28 13:03:44 +02001084 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001085 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001086 return;
1087
Jens Axboe3b181522005-06-27 10:56:24 +02001088 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001089
Jens Axboe206dc692006-03-28 13:03:44 +02001090 /*
1091 * we don't want to idle for seeks, but we do want to allow
1092 * fair distribution of slice time for a process doing back-to-back
1093 * seeks. so allow a little bit of time for him to submit a new rq
1094 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001095 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001096 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +02001097 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +02001098
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001099 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe9481ffd2009-04-15 12:14:13 +02001100 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101}
1102
Jens Axboe498d3aa22007-04-26 12:54:48 +02001103/*
1104 * Move request from internal lists to the request queue dispatch list.
1105 */
Jens Axboe165125e2007-07-24 09:28:11 +02001106static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001108 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001109 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001110
Jens Axboe7b679132008-05-30 12:23:07 +02001111 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1112
Jens Axboe5380a102006-07-13 12:37:56 +02001113 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001114 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001115 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001116
1117 if (cfq_cfqq_sync(cfqq))
1118 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119}
1120
1121/*
1122 * return expired entry, or NULL to just start from scratch in rbtree
1123 */
Jens Axboefebffd62008-01-28 13:19:43 +01001124static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125{
1126 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +02001127 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +02001128 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129
Jens Axboe3b181522005-06-27 10:56:24 +02001130 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001132
1133 cfq_mark_cfqq_fifo_expire(cfqq);
1134
Jens Axboe89850f72006-07-22 16:48:31 +02001135 if (list_empty(&cfqq->fifo))
1136 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137
Jens Axboe6d048f52007-04-25 12:44:27 +02001138 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +02001139 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140
Jens Axboe6d048f52007-04-25 12:44:27 +02001141 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
Jens Axboe7b679132008-05-30 12:23:07 +02001142 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143
Jens Axboe7b679132008-05-30 12:23:07 +02001144 cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001145 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146}
1147
Jens Axboe22e2c502005-06-27 10:55:12 +02001148static inline int
1149cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1150{
1151 const int base_rq = cfqd->cfq_slice_async_rq;
1152
1153 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1154
1155 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1156}
1157
1158/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02001159 * Select a queue for service. If we have a current active queue,
1160 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02001161 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001162static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001163{
Jens Axboea36e71f2009-04-15 12:15:11 +02001164 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001165
1166 cfqq = cfqd->active_queue;
1167 if (!cfqq)
1168 goto new_queue;
1169
1170 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001171 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02001172 */
Jens Axboeb0291952009-04-07 11:38:31 +02001173 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001174 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +02001175
1176 /*
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01001177 * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1178 * cfqq.
1179 */
1180 if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1181 /*
1182 * We simulate this as cfqq timed out so that it gets to bank
1183 * the remaining of its time slice.
1184 */
1185 cfq_log_cfqq(cfqd, cfqq, "preempt");
1186 cfq_slice_expired(cfqd, 1);
1187 goto new_queue;
1188 }
1189
1190 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001191 * The active queue has requests and isn't expired, allow it to
1192 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02001193 */
Jens Axboedd67d052006-06-21 09:36:18 +02001194 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001195 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02001196
1197 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02001198 * If another queue has a request waiting within our mean seek
1199 * distance, let it run. The expire code will check for close
1200 * cooperators and put the close queue at the front of the service
1201 * tree.
1202 */
1203 new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1204 if (new_cfqq)
1205 goto expire;
1206
1207 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001208 * No requests pending. If the active queue still has requests in
1209 * flight or is idling for a new request, allow either of these
1210 * conditions to happen (or time out) before selecting a new queue.
1211 */
Jens Axboecc197472007-04-20 20:45:39 +02001212 if (timer_pending(&cfqd->idle_slice_timer) ||
1213 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001214 cfqq = NULL;
1215 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001216 }
1217
Jens Axboe3b181522005-06-27 10:56:24 +02001218expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001219 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001220new_queue:
Jens Axboea36e71f2009-04-15 12:15:11 +02001221 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001222keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001223 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001224}
1225
Jens Axboefebffd62008-01-28 13:19:43 +01001226static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02001227{
1228 int dispatched = 0;
1229
1230 while (cfqq->next_rq) {
1231 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1232 dispatched++;
1233 }
1234
1235 BUG_ON(!list_empty(&cfqq->fifo));
1236 return dispatched;
1237}
1238
Jens Axboe498d3aa22007-04-26 12:54:48 +02001239/*
1240 * Drain our current requests. Used for barriers and when switching
1241 * io schedulers on-the-fly.
1242 */
Jens Axboed9e76202007-04-20 14:27:50 +02001243static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001244{
Jens Axboe08717142008-01-28 11:38:15 +01001245 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001246 int dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001247
Jens Axboe08717142008-01-28 11:38:15 +01001248 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
Jens Axboed9e76202007-04-20 14:27:50 +02001249 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001250
Jens Axboe6084cdd2007-04-23 08:25:00 +02001251 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001252
1253 BUG_ON(cfqd->busy_queues);
1254
Jeff Moyer69237152009-06-12 15:29:30 +02001255 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001256 return dispatched;
1257}
1258
Jens Axboe2f5cb732009-04-07 08:51:19 +02001259/*
1260 * Dispatch a request from cfqq, moving them to the request queue
1261 * dispatch list.
1262 */
1263static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1264{
1265 struct request *rq;
1266
1267 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1268
1269 /*
1270 * follow expired path, else get first next available
1271 */
1272 rq = cfq_check_fifo(cfqq);
1273 if (!rq)
1274 rq = cfqq->next_rq;
1275
1276 /*
1277 * insert request into driver dispatch list
1278 */
1279 cfq_dispatch_insert(cfqd->queue, rq);
1280
1281 if (!cfqd->active_cic) {
1282 struct cfq_io_context *cic = RQ_CIC(rq);
1283
Nikanth Karthikesand9c7d392009-06-10 12:57:06 -07001284 atomic_long_inc(&cic->ioc->refcount);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001285 cfqd->active_cic = cic;
1286 }
1287}
1288
1289/*
1290 * Find the cfqq that we need to service and move a request from that to the
1291 * dispatch list
1292 */
Jens Axboe165125e2007-07-24 09:28:11 +02001293static int cfq_dispatch_requests(struct request_queue *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294{
1295 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001296 struct cfq_queue *cfqq;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001297 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298
Jens Axboe22e2c502005-06-27 10:55:12 +02001299 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 return 0;
1301
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001302 if (unlikely(force))
1303 return cfq_forced_dispatch(cfqd);
1304
Jens Axboe2f5cb732009-04-07 08:51:19 +02001305 cfqq = cfq_select_queue(cfqd);
1306 if (!cfqq)
1307 return 0;
Jens Axboeb4878f22005-10-20 16:42:29 +02001308
Jens Axboe2f5cb732009-04-07 08:51:19 +02001309 /*
1310 * If this is an async queue and we have sync IO in flight, let it wait
1311 */
1312 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1313 return 0;
1314
1315 max_dispatch = cfqd->cfq_quantum;
1316 if (cfq_class_idle(cfqq))
1317 max_dispatch = 1;
1318
1319 /*
1320 * Does this cfqq already have too much IO in flight?
1321 */
1322 if (cfqq->dispatched >= max_dispatch) {
1323 /*
1324 * idle queue must always only have a single IO in flight
1325 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02001326 if (cfq_class_idle(cfqq))
Jens Axboe2f5cb732009-04-07 08:51:19 +02001327 return 0;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001328
Jens Axboe2f5cb732009-04-07 08:51:19 +02001329 /*
1330 * We have other queues, don't allow more IO from this one
1331 */
1332 if (cfqd->busy_queues > 1)
1333 return 0;
Jens Axboe9ede2092007-01-19 12:11:44 +11001334
Jens Axboe2f5cb732009-04-07 08:51:19 +02001335 /*
1336 * we are the only queue, allow up to 4 times of 'quantum'
1337 */
1338 if (cfqq->dispatched >= 4 * max_dispatch)
1339 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 }
1341
Jens Axboe2f5cb732009-04-07 08:51:19 +02001342 /*
1343 * Dispatch a request from this cfqq
1344 */
1345 cfq_dispatch_request(cfqd, cfqq);
1346 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02001347 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001348
1349 /*
1350 * expire an async queue immediately if it has used up its slice. idle
1351 * queue always expire after 1 dispatch round.
1352 */
1353 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1354 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1355 cfq_class_idle(cfqq))) {
1356 cfqq->slice_end = jiffies + 1;
1357 cfq_slice_expired(cfqd, 0);
1358 }
1359
1360 cfq_log(cfqd, "dispatched a request");
1361 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362}
1363
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364/*
Jens Axboe5e705372006-07-13 12:39:25 +02001365 * task holds one reference to the queue, dropped when task exits. each rq
1366 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 *
1368 * queue lock must be held here.
1369 */
1370static void cfq_put_queue(struct cfq_queue *cfqq)
1371{
Jens Axboe22e2c502005-06-27 10:55:12 +02001372 struct cfq_data *cfqd = cfqq->cfqd;
1373
1374 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375
1376 if (!atomic_dec_and_test(&cfqq->ref))
1377 return;
1378
Jens Axboe7b679132008-05-30 12:23:07 +02001379 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001381 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001382 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001384 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001385 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001386 cfq_schedule_dispatch(cfqd);
1387 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001388
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 kmem_cache_free(cfq_pool, cfqq);
1390}
1391
Jens Axboed6de8be2008-05-28 14:46:59 +02001392/*
1393 * Must always be called with the rcu_read_lock() held
1394 */
Jens Axboe07416d22008-05-07 09:17:12 +02001395static void
1396__call_for_each_cic(struct io_context *ioc,
1397 void (*func)(struct io_context *, struct cfq_io_context *))
1398{
1399 struct cfq_io_context *cic;
1400 struct hlist_node *n;
1401
1402 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1403 func(ioc, cic);
1404}
1405
Jens Axboe4ac845a2008-01-24 08:44:49 +01001406/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001407 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001408 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001409static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001410call_for_each_cic(struct io_context *ioc,
1411 void (*func)(struct io_context *, struct cfq_io_context *))
1412{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001413 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02001414 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001415 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001416}
Jens Axboe4ac845a2008-01-24 08:44:49 +01001417
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001418static void cfq_cic_free_rcu(struct rcu_head *head)
1419{
1420 struct cfq_io_context *cic;
1421
1422 cic = container_of(head, struct cfq_io_context, rcu_head);
1423
1424 kmem_cache_free(cfq_ioc_pool, cic);
1425 elv_ioc_count_dec(ioc_count);
1426
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001427 if (ioc_gone) {
1428 /*
1429 * CFQ scheduler is exiting, grab exit lock and check
1430 * the pending io context count. If it hits zero,
1431 * complete ioc_gone and set it back to NULL
1432 */
1433 spin_lock(&ioc_gone_lock);
1434 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1435 complete(ioc_gone);
1436 ioc_gone = NULL;
1437 }
1438 spin_unlock(&ioc_gone_lock);
1439 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001440}
1441
1442static void cfq_cic_free(struct cfq_io_context *cic)
1443{
1444 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001445}
1446
1447static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1448{
1449 unsigned long flags;
1450
1451 BUG_ON(!cic->dead_key);
1452
1453 spin_lock_irqsave(&ioc->lock, flags);
1454 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01001455 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001456 spin_unlock_irqrestore(&ioc->lock, flags);
1457
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001458 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001459}
1460
Jens Axboed6de8be2008-05-28 14:46:59 +02001461/*
1462 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1463 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1464 * and ->trim() which is called with the task lock held
1465 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001466static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001468 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001469 * ioc->refcount is zero here, or we are called from elv_unregister(),
1470 * so no more cic's are allowed to be linked into this ioc. So it
1471 * should be ok to iterate over the known list, we will see all cic's
1472 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001473 */
Jens Axboe07416d22008-05-07 09:17:12 +02001474 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475}
1476
Jens Axboe89850f72006-07-22 16:48:31 +02001477static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1478{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001479 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001480 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001481 cfq_schedule_dispatch(cfqd);
1482 }
Jens Axboe89850f72006-07-22 16:48:31 +02001483
1484 cfq_put_queue(cfqq);
1485}
1486
1487static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1488 struct cfq_io_context *cic)
1489{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001490 struct io_context *ioc = cic->ioc;
1491
Jens Axboefc463792006-08-29 09:05:44 +02001492 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001493
1494 /*
1495 * Make sure key == NULL is seen for dead queues
1496 */
Jens Axboefc463792006-08-29 09:05:44 +02001497 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001498 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02001499 cic->key = NULL;
1500
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001501 if (ioc->ioc_data == cic)
1502 rcu_assign_pointer(ioc->ioc_data, NULL);
1503
Jens Axboeff6657c2009-04-08 10:58:57 +02001504 if (cic->cfqq[BLK_RW_ASYNC]) {
1505 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1506 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001507 }
1508
Jens Axboeff6657c2009-04-08 10:58:57 +02001509 if (cic->cfqq[BLK_RW_SYNC]) {
1510 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1511 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001512 }
Jens Axboe89850f72006-07-22 16:48:31 +02001513}
1514
Jens Axboe4ac845a2008-01-24 08:44:49 +01001515static void cfq_exit_single_io_context(struct io_context *ioc,
1516 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001517{
Al Viro478a82b2006-03-18 13:25:24 -05001518 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001519
Jens Axboe89850f72006-07-22 16:48:31 +02001520 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001521 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001522 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02001523
Jens Axboe4ac845a2008-01-24 08:44:49 +01001524 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01001525
1526 /*
1527 * Ensure we get a fresh copy of the ->key to prevent
1528 * race between exiting task and queue
1529 */
1530 smp_read_barrier_depends();
1531 if (cic->key)
1532 __cfq_exit_single_io_context(cfqd, cic);
1533
Jens Axboe4ac845a2008-01-24 08:44:49 +01001534 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05001535 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001536}
1537
Jens Axboe498d3aa22007-04-26 12:54:48 +02001538/*
1539 * The process that ioc belongs to has exited, we need to clean up
1540 * and put the internal structures we have that belongs to that process.
1541 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001542static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001544 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545}
1546
Jens Axboe22e2c502005-06-27 10:55:12 +02001547static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001548cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549{
Jens Axboeb5deef92006-07-19 23:39:40 +02001550 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551
Christoph Lameter94f60302007-07-17 04:03:29 -07001552 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1553 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001555 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001556 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01001557 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001558 cic->dtor = cfq_free_io_context;
1559 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001560 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 }
1562
1563 return cic;
1564}
1565
Jens Axboefd0928d2008-01-24 08:52:45 +01001566static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02001567{
1568 struct task_struct *tsk = current;
1569 int ioprio_class;
1570
Jens Axboe3b181522005-06-27 10:56:24 +02001571 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001572 return;
1573
Jens Axboefd0928d2008-01-24 08:52:45 +01001574 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001575 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01001576 default:
1577 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1578 case IOPRIO_CLASS_NONE:
1579 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02001580 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01001581 */
1582 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02001583 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01001584 break;
1585 case IOPRIO_CLASS_RT:
1586 cfqq->ioprio = task_ioprio(ioc);
1587 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1588 break;
1589 case IOPRIO_CLASS_BE:
1590 cfqq->ioprio = task_ioprio(ioc);
1591 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1592 break;
1593 case IOPRIO_CLASS_IDLE:
1594 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1595 cfqq->ioprio = 7;
1596 cfq_clear_cfqq_idle_window(cfqq);
1597 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02001598 }
1599
1600 /*
1601 * keep track of original prio settings in case we have to temporarily
1602 * elevate the priority of this queue
1603 */
1604 cfqq->org_ioprio = cfqq->ioprio;
1605 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001606 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001607}
1608
Jens Axboefebffd62008-01-28 13:19:43 +01001609static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001610{
Al Viro478a82b2006-03-18 13:25:24 -05001611 struct cfq_data *cfqd = cic->key;
1612 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001613 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001614
Jens Axboecaaa5f92006-06-16 11:23:00 +02001615 if (unlikely(!cfqd))
1616 return;
1617
Jens Axboec1b707d2006-10-30 19:54:23 +01001618 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001619
Jens Axboeff6657c2009-04-08 10:58:57 +02001620 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001621 if (cfqq) {
1622 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02001623 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1624 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001625 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02001626 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001627 cfq_put_queue(cfqq);
1628 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001629 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001630
Jens Axboeff6657c2009-04-08 10:58:57 +02001631 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001632 if (cfqq)
1633 cfq_mark_cfqq_prio_changed(cfqq);
1634
Jens Axboec1b707d2006-10-30 19:54:23 +01001635 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001636}
1637
Jens Axboefc463792006-08-29 09:05:44 +02001638static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001640 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02001641 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642}
1643
1644static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001645cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01001646 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001649 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650
1651retry:
Jens Axboe4ac845a2008-01-24 08:44:49 +01001652 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001653 /* cic always exists here */
1654 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655
1656 if (!cfqq) {
1657 if (new_cfqq) {
1658 cfqq = new_cfqq;
1659 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001660 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001661 /*
1662 * Inform the allocator of the fact that we will
1663 * just repeat this allocation if it fails, to allow
1664 * the allocator to do whatever it needs to attempt to
1665 * free memory.
1666 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001668 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1669 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1670 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671 spin_lock_irq(cfqd->queue->queue_lock);
1672 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001673 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001674 cfqq = kmem_cache_alloc_node(cfq_pool,
1675 gfp_mask | __GFP_ZERO,
1676 cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001677 if (!cfqq)
1678 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001679 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680
Jens Axboed9e76202007-04-20 14:27:50 +02001681 RB_CLEAR_NODE(&cfqq->rb_node);
Jens Axboea36e71f2009-04-15 12:15:11 +02001682 RB_CLEAR_NODE(&cfqq->p_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001683 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 atomic_set(&cfqq->ref, 0);
1686 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001687
Jens Axboe3b181522005-06-27 10:56:24 +02001688 cfq_mark_cfqq_prio_changed(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001689
Jens Axboefd0928d2008-01-24 08:52:45 +01001690 cfq_init_prio_data(cfqq, ioc);
Jens Axboe08717142008-01-28 11:38:15 +01001691
1692 if (is_sync) {
1693 if (!cfq_class_idle(cfqq))
1694 cfq_mark_cfqq_idle_window(cfqq);
1695 cfq_mark_cfqq_sync(cfqq);
1696 }
Jens Axboe7b679132008-05-30 12:23:07 +02001697 cfqq->pid = current->pid;
1698 cfq_log_cfqq(cfqd, cfqq, "alloced");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 }
1700
1701 if (new_cfqq)
1702 kmem_cache_free(cfq_pool, new_cfqq);
1703
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704out:
1705 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1706 return cfqq;
1707}
1708
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001709static struct cfq_queue **
1710cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1711{
Jens Axboefe094d92008-01-31 13:08:54 +01001712 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001713 case IOPRIO_CLASS_RT:
1714 return &cfqd->async_cfqq[0][ioprio];
1715 case IOPRIO_CLASS_BE:
1716 return &cfqd->async_cfqq[1][ioprio];
1717 case IOPRIO_CLASS_IDLE:
1718 return &cfqd->async_idle_cfqq;
1719 default:
1720 BUG();
1721 }
1722}
1723
Jens Axboe15c31be2007-07-10 13:43:25 +02001724static struct cfq_queue *
Jens Axboefd0928d2008-01-24 08:52:45 +01001725cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02001726 gfp_t gfp_mask)
1727{
Jens Axboefd0928d2008-01-24 08:52:45 +01001728 const int ioprio = task_ioprio(ioc);
1729 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001730 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001731 struct cfq_queue *cfqq = NULL;
1732
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001733 if (!is_sync) {
1734 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1735 cfqq = *async_cfqq;
1736 }
1737
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001738 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01001739 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001740 if (!cfqq)
1741 return NULL;
1742 }
Jens Axboe15c31be2007-07-10 13:43:25 +02001743
1744 /*
1745 * pin the queue now that it's allocated, scheduler exit will prune it
1746 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001747 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001748 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001749 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001750 }
1751
1752 atomic_inc(&cfqq->ref);
1753 return cfqq;
1754}
1755
Jens Axboe498d3aa22007-04-26 12:54:48 +02001756/*
1757 * We drop cfq io contexts lazily, so we may find a dead one.
1758 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001759static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001760cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1761 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001762{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001763 unsigned long flags;
1764
Jens Axboefc463792006-08-29 09:05:44 +02001765 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001766
Jens Axboe4ac845a2008-01-24 08:44:49 +01001767 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02001768
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001769 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001770
1771 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01001772 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001773 spin_unlock_irqrestore(&ioc->lock, flags);
1774
1775 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001776}
1777
Jens Axboee2d74ac2006-03-28 08:59:01 +02001778static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01001779cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001780{
Jens Axboee2d74ac2006-03-28 08:59:01 +02001781 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001782 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001783 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001784
Vasily Tarasov91fac312007-04-25 12:29:51 +02001785 if (unlikely(!ioc))
1786 return NULL;
1787
Jens Axboed6de8be2008-05-28 14:46:59 +02001788 rcu_read_lock();
1789
Jens Axboe597bc482007-04-24 21:23:53 +02001790 /*
1791 * we maintain a last-hit cache, to avoid browsing over the tree
1792 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001793 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02001794 if (cic && cic->key == cfqd) {
1795 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02001796 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001797 }
Jens Axboe597bc482007-04-24 21:23:53 +02001798
Jens Axboe4ac845a2008-01-24 08:44:49 +01001799 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001800 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1801 rcu_read_unlock();
1802 if (!cic)
1803 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001804 /* ->key must be copied to avoid race with cfq_exit_queue() */
1805 k = cic->key;
1806 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001807 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001808 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001809 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001810 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001811
Jens Axboed6de8be2008-05-28 14:46:59 +02001812 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001813 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001814 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001815 break;
1816 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001817
Jens Axboe4ac845a2008-01-24 08:44:49 +01001818 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001819}
1820
Jens Axboe4ac845a2008-01-24 08:44:49 +01001821/*
1822 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1823 * the process specific cfq io context when entered from the block layer.
1824 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1825 */
Jens Axboefebffd62008-01-28 13:19:43 +01001826static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1827 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001828{
Jens Axboe0261d682006-10-30 19:07:48 +01001829 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001830 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001831
Jens Axboe4ac845a2008-01-24 08:44:49 +01001832 ret = radix_tree_preload(gfp_mask);
1833 if (!ret) {
1834 cic->ioc = ioc;
1835 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001836
Jens Axboe4ac845a2008-01-24 08:44:49 +01001837 spin_lock_irqsave(&ioc->lock, flags);
1838 ret = radix_tree_insert(&ioc->radix_root,
1839 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01001840 if (!ret)
1841 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001842 spin_unlock_irqrestore(&ioc->lock, flags);
1843
1844 radix_tree_preload_end();
1845
1846 if (!ret) {
1847 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1848 list_add(&cic->queue_list, &cfqd->cic_list);
1849 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001850 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001851 }
1852
Jens Axboe4ac845a2008-01-24 08:44:49 +01001853 if (ret)
1854 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02001855
Jens Axboe4ac845a2008-01-24 08:44:49 +01001856 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001857}
1858
Jens Axboe22e2c502005-06-27 10:55:12 +02001859/*
1860 * Setup general io context and cfq io context. There can be several cfq
1861 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001862 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001863 */
1864static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001865cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866{
Jens Axboe22e2c502005-06-27 10:55:12 +02001867 struct io_context *ioc = NULL;
1868 struct cfq_io_context *cic;
1869
1870 might_sleep_if(gfp_mask & __GFP_WAIT);
1871
Jens Axboeb5deef92006-07-19 23:39:40 +02001872 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001873 if (!ioc)
1874 return NULL;
1875
Jens Axboe4ac845a2008-01-24 08:44:49 +01001876 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001877 if (cic)
1878 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001879
Jens Axboee2d74ac2006-03-28 08:59:01 +02001880 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1881 if (cic == NULL)
1882 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001883
Jens Axboe4ac845a2008-01-24 08:44:49 +01001884 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1885 goto err_free;
1886
Jens Axboe22e2c502005-06-27 10:55:12 +02001887out:
Jens Axboefc463792006-08-29 09:05:44 +02001888 smp_read_barrier_depends();
1889 if (unlikely(ioc->ioprio_changed))
1890 cfq_ioc_set_ioprio(ioc);
1891
Jens Axboe22e2c502005-06-27 10:55:12 +02001892 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001893err_free:
1894 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001895err:
1896 put_io_context(ioc);
1897 return NULL;
1898}
1899
1900static void
1901cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1902{
Jens Axboeaaf12282007-01-19 11:30:16 +11001903 unsigned long elapsed = jiffies - cic->last_end_request;
1904 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001905
1906 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1907 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1908 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1909}
1910
Jens Axboe206dc692006-03-28 13:03:44 +02001911static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001912cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1913 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001914{
1915 sector_t sdist;
1916 u64 total;
1917
Jeff Moyer4d00aa42009-04-21 07:25:04 +02001918 if (!cic->last_request_pos)
1919 sdist = 0;
Tejun Heo83096eb2009-05-07 22:24:39 +09001920 else if (cic->last_request_pos < blk_rq_pos(rq))
1921 sdist = blk_rq_pos(rq) - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001922 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001923 sdist = cic->last_request_pos - blk_rq_pos(rq);
Jens Axboe206dc692006-03-28 13:03:44 +02001924
1925 /*
1926 * Don't allow the seek distance to get too large from the
1927 * odd fragment, pagein, etc
1928 */
1929 if (cic->seek_samples <= 60) /* second&third seek */
1930 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1931 else
1932 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1933
1934 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1935 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1936 total = cic->seek_total + (cic->seek_samples/2);
1937 do_div(total, cic->seek_samples);
1938 cic->seek_mean = (sector_t)total;
1939}
Jens Axboe22e2c502005-06-27 10:55:12 +02001940
1941/*
1942 * Disable idle window if the process thinks too long or seeks so much that
1943 * it doesn't matter
1944 */
1945static void
1946cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1947 struct cfq_io_context *cic)
1948{
Jens Axboe7b679132008-05-30 12:23:07 +02001949 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02001950
Jens Axboe08717142008-01-28 11:38:15 +01001951 /*
1952 * Don't idle for async or idle io prio class
1953 */
1954 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02001955 return;
1956
Jens Axboec265a7f2008-06-26 13:49:33 +02001957 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001958
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001959 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Jens Axboecaaa5f92006-06-16 11:23:00 +02001960 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001961 enable_idle = 0;
1962 else if (sample_valid(cic->ttime_samples)) {
1963 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1964 enable_idle = 0;
1965 else
1966 enable_idle = 1;
1967 }
1968
Jens Axboe7b679132008-05-30 12:23:07 +02001969 if (old_idle != enable_idle) {
1970 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1971 if (enable_idle)
1972 cfq_mark_cfqq_idle_window(cfqq);
1973 else
1974 cfq_clear_cfqq_idle_window(cfqq);
1975 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001976}
1977
Jens Axboe22e2c502005-06-27 10:55:12 +02001978/*
1979 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1980 * no or if we aren't sure, a 1 will cause a preempt.
1981 */
1982static int
1983cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001984 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001985{
Jens Axboe6d048f52007-04-25 12:44:27 +02001986 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001987
Jens Axboe6d048f52007-04-25 12:44:27 +02001988 cfqq = cfqd->active_queue;
1989 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001990 return 0;
1991
Jens Axboe6d048f52007-04-25 12:44:27 +02001992 if (cfq_slice_used(cfqq))
1993 return 1;
1994
1995 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001996 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001997
1998 if (cfq_class_idle(cfqq))
1999 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002000
Jens Axboe22e2c502005-06-27 10:55:12 +02002001 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02002002 * if the new request is sync, but the currently running queue is
2003 * not, let the sync request have priority.
2004 */
Jens Axboe5e705372006-07-13 12:39:25 +02002005 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002006 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002007
Jens Axboe374f84a2006-07-23 01:42:19 +02002008 /*
2009 * So both queues are sync. Let the new request get disk time if
2010 * it's a metadata request and the current queue is doing regular IO.
2011 */
2012 if (rq_is_meta(rq) && !cfqq->meta_pending)
2013 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002014
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002015 /*
2016 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2017 */
2018 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2019 return 1;
2020
Jens Axboe1e3335d2007-02-14 19:59:49 +01002021 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2022 return 0;
2023
2024 /*
2025 * if this request is as-good as one we would expect from the
2026 * current cfqq, let it preempt
2027 */
Jens Axboe6d048f52007-04-25 12:44:27 +02002028 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01002029 return 1;
2030
Jens Axboe22e2c502005-06-27 10:55:12 +02002031 return 0;
2032}
2033
2034/*
2035 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2036 * let it have half of its nominal slice.
2037 */
2038static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2039{
Jens Axboe7b679132008-05-30 12:23:07 +02002040 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02002041 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02002042
Jens Axboebf572252006-07-19 20:29:12 +02002043 /*
2044 * Put the new queue at the front of the of the current list,
2045 * so we know that it will be selected next.
2046 */
2047 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02002048
2049 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02002050
Jens Axboe44f7c162007-01-19 11:51:58 +11002051 cfqq->slice_end = 0;
2052 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002053}
2054
2055/*
Jens Axboe5e705372006-07-13 12:39:25 +02002056 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02002057 * something we should do about it
2058 */
2059static void
Jens Axboe5e705372006-07-13 12:39:25 +02002060cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2061 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002062{
Jens Axboe5e705372006-07-13 12:39:25 +02002063 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02002064
Aaron Carroll45333d52008-08-26 15:52:36 +02002065 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02002066 if (rq_is_meta(rq))
2067 cfqq->meta_pending++;
2068
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002069 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02002070 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002071 cfq_update_idle_window(cfqd, cfqq, cic);
2072
Tejun Heo83096eb2009-05-07 22:24:39 +09002073 cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002074
2075 if (cfqq == cfqd->active_queue) {
2076 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002077 * Remember that we saw a request from this process, but
2078 * don't start queuing just yet. Otherwise we risk seeing lots
2079 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02002080 * and merging. If the request is already larger than a single
2081 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02002082 * merging is already done. Ditto for a busy system that
2083 * has other work pending, don't risk delaying until the
2084 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02002085 */
Jens Axboed6ceb252009-04-14 14:18:16 +02002086 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02002087 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2088 cfqd->busy_queues > 1) {
Jens Axboed6ceb252009-04-14 14:18:16 +02002089 del_timer(&cfqd->idle_slice_timer);
Tejun Heoa7f55792009-04-23 11:05:17 +09002090 __blk_run_queue(cfqd->queue);
Jens Axboed6ceb252009-04-14 14:18:16 +02002091 }
Jens Axboeb0291952009-04-07 11:38:31 +02002092 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboed6ceb252009-04-14 14:18:16 +02002093 }
Jens Axboe5e705372006-07-13 12:39:25 +02002094 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002095 /*
2096 * not the active queue - expire current slice if it is
2097 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002098 * has some old slice time left and is of higher priority or
2099 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02002100 */
2101 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09002102 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002103 }
2104}
2105
Jens Axboe165125e2007-07-24 09:28:11 +02002106static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002107{
Jens Axboeb4878f22005-10-20 16:42:29 +02002108 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002109 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002110
Jens Axboe7b679132008-05-30 12:23:07 +02002111 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01002112 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113
Jens Axboe5e705372006-07-13 12:39:25 +02002114 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115
Jens Axboe22e2c502005-06-27 10:55:12 +02002116 list_add_tail(&rq->queuelist, &cfqq->fifo);
2117
Jens Axboe5e705372006-07-13 12:39:25 +02002118 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119}
2120
Aaron Carroll45333d52008-08-26 15:52:36 +02002121/*
2122 * Update hw_tag based on peak queue depth over 50 samples under
2123 * sufficient load.
2124 */
2125static void cfq_update_hw_tag(struct cfq_data *cfqd)
2126{
2127 if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
2128 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
2129
2130 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2131 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
2132 return;
2133
2134 if (cfqd->hw_tag_samples++ < 50)
2135 return;
2136
2137 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2138 cfqd->hw_tag = 1;
2139 else
2140 cfqd->hw_tag = 0;
2141
2142 cfqd->hw_tag_samples = 0;
2143 cfqd->rq_in_driver_peak = 0;
2144}
2145
Jens Axboe165125e2007-07-24 09:28:11 +02002146static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147{
Jens Axboe5e705372006-07-13 12:39:25 +02002148 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002149 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02002150 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002151 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152
Jens Axboeb4878f22005-10-20 16:42:29 +02002153 now = jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02002154 cfq_log_cfqq(cfqd, cfqq, "complete");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155
Aaron Carroll45333d52008-08-26 15:52:36 +02002156 cfq_update_hw_tag(cfqd);
2157
Jens Axboeb4878f22005-10-20 16:42:29 +02002158 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02002159 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02002160 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02002161 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162
Jens Axboe3ed9a292007-04-23 08:33:33 +02002163 if (cfq_cfqq_sync(cfqq))
2164 cfqd->sync_flight--;
2165
Jens Axboecaaa5f92006-06-16 11:23:00 +02002166 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02002167 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002168
2169 /*
2170 * If this is the active queue, check if it needs to be expired,
2171 * or if we want to idle in case it has no pending requests.
2172 */
2173 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02002174 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2175
Jens Axboe44f7c162007-01-19 11:51:58 +11002176 if (cfq_cfqq_slice_new(cfqq)) {
2177 cfq_set_prio_slice(cfqd, cfqq);
2178 cfq_clear_cfqq_slice_new(cfqq);
2179 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002180 /*
2181 * If there are no requests waiting in this queue, and
2182 * there are other queues ready to issue requests, AND
2183 * those other queues are issuing requests within our
2184 * mean seek distance, give them a chance to run instead
2185 * of idling.
2186 */
Jens Axboe08717142008-01-28 11:38:15 +01002187 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02002188 cfq_slice_expired(cfqd, 1);
Jens Axboea36e71f2009-04-15 12:15:11 +02002189 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2190 sync && !rq_noidle(rq))
Jens Axboe6d048f52007-04-25 12:44:27 +02002191 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002192 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002193
2194 if (!cfqd->rq_in_driver)
2195 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196}
2197
Jens Axboe22e2c502005-06-27 10:55:12 +02002198/*
2199 * we temporarily boost lower priority queues if they are holding fs exclusive
2200 * resources. they are boosted to normal prio (CLASS_BE/4)
2201 */
2202static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203{
Jens Axboe22e2c502005-06-27 10:55:12 +02002204 if (has_fs_excl()) {
2205 /*
2206 * boost idle prio on transactions that would lock out other
2207 * users of the filesystem
2208 */
2209 if (cfq_class_idle(cfqq))
2210 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2211 if (cfqq->ioprio > IOPRIO_NORM)
2212 cfqq->ioprio = IOPRIO_NORM;
2213 } else {
2214 /*
2215 * check if we need to unboost the queue
2216 */
2217 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2218 cfqq->ioprio_class = cfqq->org_ioprio_class;
2219 if (cfqq->ioprio != cfqq->org_ioprio)
2220 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002222}
2223
Jens Axboe89850f72006-07-22 16:48:31 +02002224static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002225{
Jens Axboe3b181522005-06-27 10:56:24 +02002226 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07002227 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002228 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002229 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02002230 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002231
2232 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02002233}
2234
Jens Axboe165125e2007-07-24 09:28:11 +02002235static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02002236{
2237 struct cfq_data *cfqd = q->elevator->elevator_data;
2238 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002239 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02002240 struct cfq_queue *cfqq;
2241
2242 /*
2243 * don't force setup of a queue from here, as a call to may_queue
2244 * does not necessarily imply that a request actually will be queued.
2245 * so just lookup a possibly existing queue, or return 'may queue'
2246 * if that fails
2247 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002248 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002249 if (!cic)
2250 return ELV_MQUEUE_MAY;
2251
Jens Axboeb0b78f82009-04-08 10:56:08 +02002252 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02002253 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002254 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02002255 cfq_prio_boost(cfqq);
2256
Jens Axboe89850f72006-07-22 16:48:31 +02002257 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002258 }
2259
2260 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261}
2262
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263/*
2264 * queue lock held here
2265 */
Jens Axboebb37b942006-12-01 10:42:33 +01002266static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267{
Jens Axboe5e705372006-07-13 12:39:25 +02002268 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269
Jens Axboe5e705372006-07-13 12:39:25 +02002270 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002271 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272
Jens Axboe22e2c502005-06-27 10:55:12 +02002273 BUG_ON(!cfqq->allocated[rw]);
2274 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275
Jens Axboe5e705372006-07-13 12:39:25 +02002276 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02002279 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 cfq_put_queue(cfqq);
2282 }
2283}
2284
2285/*
Jens Axboe22e2c502005-06-27 10:55:12 +02002286 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 */
Jens Axboe22e2c502005-06-27 10:55:12 +02002288static int
Jens Axboe165125e2007-07-24 09:28:11 +02002289cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290{
2291 struct cfq_data *cfqd = q->elevator->elevator_data;
2292 struct cfq_io_context *cic;
2293 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01002294 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002295 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 unsigned long flags;
2297
2298 might_sleep_if(gfp_mask & __GFP_WAIT);
2299
Jens Axboee2d74ac2006-03-28 08:59:01 +02002300 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02002301
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302 spin_lock_irqsave(q->queue_lock, flags);
2303
Jens Axboe22e2c502005-06-27 10:55:12 +02002304 if (!cic)
2305 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306
Vasily Tarasov91fac312007-04-25 12:29:51 +02002307 cfqq = cic_to_cfqq(cic, is_sync);
2308 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002309 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002310
Jens Axboe22e2c502005-06-27 10:55:12 +02002311 if (!cfqq)
2312 goto queue_fail;
2313
Vasily Tarasov91fac312007-04-25 12:29:51 +02002314 cic_set_cfqq(cic, cfqq, is_sync);
2315 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316
2317 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02002318 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002319 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02002320
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 spin_unlock_irqrestore(q->queue_lock, flags);
2322
Jens Axboe5e705372006-07-13 12:39:25 +02002323 rq->elevator_private = cic;
2324 rq->elevator_private2 = cfqq;
2325 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02002326
Jens Axboe22e2c502005-06-27 10:55:12 +02002327queue_fail:
2328 if (cic)
2329 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02002330
Jens Axboe3b181522005-06-27 10:56:24 +02002331 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02002333 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 return 1;
2335}
2336
David Howells65f27f32006-11-22 14:55:48 +00002337static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02002338{
David Howells65f27f32006-11-22 14:55:48 +00002339 struct cfq_data *cfqd =
2340 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02002341 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002342
Jens Axboe40bb54d2009-04-15 12:11:10 +02002343 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09002344 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02002345 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02002346}
2347
2348/*
2349 * Timer running if the active_queue is currently idling inside its time slice
2350 */
2351static void cfq_idle_slice_timer(unsigned long data)
2352{
2353 struct cfq_data *cfqd = (struct cfq_data *) data;
2354 struct cfq_queue *cfqq;
2355 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002356 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002357
Jens Axboe7b679132008-05-30 12:23:07 +02002358 cfq_log(cfqd, "idle timer fired");
2359
Jens Axboe22e2c502005-06-27 10:55:12 +02002360 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2361
Jens Axboefe094d92008-01-31 13:08:54 +01002362 cfqq = cfqd->active_queue;
2363 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002364 timed_out = 0;
2365
Jens Axboe22e2c502005-06-27 10:55:12 +02002366 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002367 * We saw a request before the queue expired, let it through
2368 */
2369 if (cfq_cfqq_must_dispatch(cfqq))
2370 goto out_kick;
2371
2372 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002373 * expired
2374 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002375 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002376 goto expire;
2377
2378 /*
2379 * only expire and reinvoke request handler, if there are
2380 * other queues with pending requests
2381 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002382 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002383 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002384
2385 /*
2386 * not expired and it has a request pending, let it dispatch
2387 */
Jens Axboe75e50982009-04-07 08:56:14 +02002388 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002389 goto out_kick;
Jens Axboe22e2c502005-06-27 10:55:12 +02002390 }
2391expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002392 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002393out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002394 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002395out_cont:
2396 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2397}
2398
Jens Axboe3b181522005-06-27 10:56:24 +02002399static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2400{
2401 del_timer_sync(&cfqd->idle_slice_timer);
Cheng Renquan64d01dc2008-12-03 12:41:39 +01002402 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002403}
Jens Axboe22e2c502005-06-27 10:55:12 +02002404
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002405static void cfq_put_async_queues(struct cfq_data *cfqd)
2406{
2407 int i;
2408
2409 for (i = 0; i < IOPRIO_BE_NR; i++) {
2410 if (cfqd->async_cfqq[0][i])
2411 cfq_put_queue(cfqd->async_cfqq[0][i]);
2412 if (cfqd->async_cfqq[1][i])
2413 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002414 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002415
2416 if (cfqd->async_idle_cfqq)
2417 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002418}
2419
Jens Axboeb374d182008-10-31 10:05:07 +01002420static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421{
Jens Axboe22e2c502005-06-27 10:55:12 +02002422 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002423 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002424
Jens Axboe3b181522005-06-27 10:56:24 +02002425 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002426
Al Virod9ff4182006-03-18 13:51:22 -05002427 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002428
Al Virod9ff4182006-03-18 13:51:22 -05002429 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002430 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002431
2432 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002433 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2434 struct cfq_io_context,
2435 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002436
2437 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002438 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002439
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002440 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002441
Al Virod9ff4182006-03-18 13:51:22 -05002442 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002443
2444 cfq_shutdown_timer_wq(cfqd);
2445
Al Viroa90d7422006-03-18 12:05:37 -05002446 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447}
2448
Jens Axboe165125e2007-07-24 09:28:11 +02002449static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450{
2451 struct cfq_data *cfqd;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002452 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453
Christoph Lameter94f60302007-07-17 04:03:29 -07002454 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002456 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457
Jens Axboecc09e292007-04-26 12:53:50 +02002458 cfqd->service_tree = CFQ_RB_ROOT;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002459
2460 /*
2461 * Not strictly needed (since RB_ROOT just clears the node and we
2462 * zeroed cfqd on alloc), but better be safe in case someone decides
2463 * to add magic to the rb code
2464 */
2465 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2466 cfqd->prio_trees[i] = RB_ROOT;
2467
Al Virod9ff4182006-03-18 13:51:22 -05002468 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471
Jens Axboe22e2c502005-06-27 10:55:12 +02002472 init_timer(&cfqd->idle_slice_timer);
2473 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2474 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2475
David Howells65f27f32006-11-22 14:55:48 +00002476 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002477
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002479 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2480 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 cfqd->cfq_back_max = cfq_back_max;
2482 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002483 cfqd->cfq_slice[0] = cfq_slice_async;
2484 cfqd->cfq_slice[1] = cfq_slice_sync;
2485 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2486 cfqd->cfq_slice_idle = cfq_slice_idle;
Aaron Carroll45333d52008-08-26 15:52:36 +02002487 cfqd->hw_tag = 1;
Jens Axboe3b181522005-06-27 10:56:24 +02002488
Jens Axboebc1c1162006-06-08 08:49:06 +02002489 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490}
2491
2492static void cfq_slab_kill(void)
2493{
Jens Axboed6de8be2008-05-28 14:46:59 +02002494 /*
2495 * Caller already ensured that pending RCU callbacks are completed,
2496 * so we should have no busy allocations at this point.
2497 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498 if (cfq_pool)
2499 kmem_cache_destroy(cfq_pool);
2500 if (cfq_ioc_pool)
2501 kmem_cache_destroy(cfq_ioc_pool);
2502}
2503
2504static int __init cfq_slab_setup(void)
2505{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002506 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 if (!cfq_pool)
2508 goto fail;
2509
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002510 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 if (!cfq_ioc_pool)
2512 goto fail;
2513
2514 return 0;
2515fail:
2516 cfq_slab_kill();
2517 return -ENOMEM;
2518}
2519
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520/*
2521 * sysfs parts below -->
2522 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523static ssize_t
2524cfq_var_show(unsigned int var, char *page)
2525{
2526 return sprintf(page, "%d\n", var);
2527}
2528
2529static ssize_t
2530cfq_var_store(unsigned int *var, const char *page, size_t count)
2531{
2532 char *p = (char *) page;
2533
2534 *var = simple_strtoul(p, &p, 10);
2535 return count;
2536}
2537
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002539static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002541 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542 unsigned int __data = __VAR; \
2543 if (__CONV) \
2544 __data = jiffies_to_msecs(__data); \
2545 return cfq_var_show(__data, (page)); \
2546}
2547SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002548SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2549SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002550SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2551SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002552SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2553SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2554SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2555SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556#undef SHOW_FUNCTION
2557
2558#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002559static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002561 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562 unsigned int __data; \
2563 int ret = cfq_var_store(&__data, (page), count); \
2564 if (__data < (MIN)) \
2565 __data = (MIN); \
2566 else if (__data > (MAX)) \
2567 __data = (MAX); \
2568 if (__CONV) \
2569 *(__PTR) = msecs_to_jiffies(__data); \
2570 else \
2571 *(__PTR) = __data; \
2572 return ret; \
2573}
2574STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002575STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2576 UINT_MAX, 1);
2577STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2578 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002579STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002580STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2581 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002582STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2583STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2584STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01002585STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2586 UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587#undef STORE_FUNCTION
2588
Al Viroe572ec72006-03-18 22:27:18 -05002589#define CFQ_ATTR(name) \
2590 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002591
Al Viroe572ec72006-03-18 22:27:18 -05002592static struct elv_fs_entry cfq_attrs[] = {
2593 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002594 CFQ_ATTR(fifo_expire_sync),
2595 CFQ_ATTR(fifo_expire_async),
2596 CFQ_ATTR(back_seek_max),
2597 CFQ_ATTR(back_seek_penalty),
2598 CFQ_ATTR(slice_sync),
2599 CFQ_ATTR(slice_async),
2600 CFQ_ATTR(slice_async_rq),
2601 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002602 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603};
2604
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605static struct elevator_type iosched_cfq = {
2606 .ops = {
2607 .elevator_merge_fn = cfq_merge,
2608 .elevator_merged_fn = cfq_merged_request,
2609 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002610 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002611 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002613 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 .elevator_deactivate_req_fn = cfq_deactivate_request,
2615 .elevator_queue_empty_fn = cfq_queue_empty,
2616 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002617 .elevator_former_req_fn = elv_rb_former_request,
2618 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 .elevator_set_req_fn = cfq_set_request,
2620 .elevator_put_req_fn = cfq_put_request,
2621 .elevator_may_queue_fn = cfq_may_queue,
2622 .elevator_init_fn = cfq_init_queue,
2623 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002624 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625 },
Al Viro3d1ab402006-03-18 18:35:43 -05002626 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 .elevator_name = "cfq",
2628 .elevator_owner = THIS_MODULE,
2629};
2630
2631static int __init cfq_init(void)
2632{
Jens Axboe22e2c502005-06-27 10:55:12 +02002633 /*
2634 * could be 0 on HZ < 1000 setups
2635 */
2636 if (!cfq_slice_async)
2637 cfq_slice_async = 1;
2638 if (!cfq_slice_idle)
2639 cfq_slice_idle = 1;
2640
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 if (cfq_slab_setup())
2642 return -ENOMEM;
2643
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002644 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002646 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647}
2648
2649static void __exit cfq_exit(void)
2650{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002651 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002653 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002654 /* ioc_gone's update must be visible before reading ioc_count */
2655 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02002656
2657 /*
2658 * this also protects us from entering cfq_slab_kill() with
2659 * pending RCU callbacks
2660 */
Jens Axboe4050cf12006-07-19 05:07:12 +02002661 if (elv_ioc_count_read(ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002662 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08002663 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664}
2665
2666module_init(cfq_init);
2667module_exit(cfq_exit);
2668
2669MODULE_AUTHOR("Jens Axboe");
2670MODULE_LICENSE("GPL");
2671MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");