blob: 9c4b679908f41cf034cdddc79a9bff8bcb1f4e85 [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
Tejun Heo245b2e72009-06-24 15:13:48 +090051static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050052static struct completion *ioc_gone;
Jens Axboe9a11b4e2008-05-29 09:32:08 +020053static DEFINE_SPINLOCK(ioc_gone_lock);
Al Viro334e94d2006-03-18 15:05:53 -050054
Jens Axboe22e2c502005-06-27 10:55:12 +020055#define CFQ_PRIO_LISTS IOPRIO_BE_NR
56#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020057#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
58
Jens Axboe206dc692006-03-28 13:03:44 +020059#define sample_valid(samples) ((samples) > 80)
60
Jens Axboe22e2c502005-06-27 10:55:12 +020061/*
Jens Axboecc09e292007-04-26 12:53:50 +020062 * Most of our rbtree usage is for sorting with min extraction, so
63 * if we cache the leftmost node we don't have to walk down the tree
64 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
65 * move this into the elevator for the rq sorting as well.
66 */
67struct cfq_rb_root {
68 struct rb_root rb;
69 struct rb_node *left;
70};
71#define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
72
73/*
Jens Axboe6118b702009-06-30 09:34:12 +020074 * Per process-grouping structure
75 */
76struct cfq_queue {
77 /* reference count */
78 atomic_t ref;
79 /* various state flags, see below */
80 unsigned int flags;
81 /* parent cfq_data */
82 struct cfq_data *cfqd;
83 /* service_tree member */
84 struct rb_node rb_node;
85 /* service_tree key */
86 unsigned long rb_key;
87 /* prio tree member */
88 struct rb_node p_node;
89 /* prio tree root we belong to, if any */
90 struct rb_root *p_root;
91 /* sorted list of pending requests */
92 struct rb_root sort_list;
93 /* if fifo isn't expired, next request to serve */
94 struct request *next_rq;
95 /* requests queued in sort_list */
96 int queued[2];
97 /* currently allocated requests */
98 int allocated[2];
99 /* fifo list of requests in sort_list */
100 struct list_head fifo;
101
102 unsigned long slice_end;
103 long slice_resid;
104 unsigned int slice_dispatch;
105
106 /* pending metadata requests */
107 int meta_pending;
108 /* number of requests that are on the dispatch list or inside driver */
109 int dispatched;
110
111 /* io prio of this group */
112 unsigned short ioprio, org_ioprio;
113 unsigned short ioprio_class, org_ioprio_class;
114
115 pid_t pid;
116};
117
118/*
Jens Axboe22e2c502005-06-27 10:55:12 +0200119 * Per block device queue structure
120 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121struct cfq_data {
Jens Axboe165125e2007-07-24 09:28:11 +0200122 struct request_queue *queue;
Jens Axboe22e2c502005-06-27 10:55:12 +0200123
124 /*
125 * rr list of queues with requests and the count of them
126 */
Jens Axboecc09e292007-04-26 12:53:50 +0200127 struct cfq_rb_root service_tree;
Jens Axboea36e71f2009-04-15 12:15:11 +0200128
129 /*
130 * Each priority tree is sorted by next_request position. These
131 * trees are used when determining if two or more queues are
132 * interleaving requests (see cfq_close_cooperator).
133 */
134 struct rb_root prio_trees[CFQ_PRIO_LISTS];
135
Jens Axboe22e2c502005-06-27 10:55:12 +0200136 unsigned int busy_queues;
137
Jens Axboe5ad531d2009-07-03 12:57:48 +0200138 int rq_in_driver[2];
Jens Axboe3ed9a292007-04-23 08:33:33 +0200139 int sync_flight;
Aaron Carroll45333d52008-08-26 15:52:36 +0200140
141 /*
142 * queue-depth detection
143 */
144 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200145 int hw_tag;
Aaron Carroll45333d52008-08-26 15:52:36 +0200146 int hw_tag_samples;
147 int rq_in_driver_peak;
Jens Axboe22e2c502005-06-27 10:55:12 +0200148
149 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200150 * idle window management
151 */
152 struct timer_list idle_slice_timer;
Jens Axboe8e296752009-10-03 16:26:03 +0200153 struct delayed_work unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200154
155 struct cfq_queue *active_queue;
156 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200157
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200158 /*
159 * async queue for each priority case
160 */
161 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
162 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200163
Jens Axboe6d048f52007-04-25 12:44:27 +0200164 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 /*
167 * tunables, see top of file
168 */
169 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200170 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 unsigned int cfq_back_penalty;
172 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200173 unsigned int cfq_slice[2];
174 unsigned int cfq_slice_async_rq;
175 unsigned int cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200176 unsigned int cfq_latency;
Al Virod9ff4182006-03-18 13:51:22 -0500177
178 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179
Jens Axboe6118b702009-06-30 09:34:12 +0200180 /*
181 * Fallback dummy cfqq for extreme OOM conditions
182 */
183 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200184
185 unsigned long last_end_sync_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186};
187
Jens Axboe3b181522005-06-27 10:56:24 +0200188enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100189 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
190 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200191 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100192 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100193 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
194 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
195 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100196 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200197 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboea36e71f2009-04-15 12:15:11 +0200198 CFQ_CFQQ_FLAG_coop, /* has done a coop jump of the queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200199};
200
201#define CFQ_CFQQ_FNS(name) \
202static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
203{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100204 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200205} \
206static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
207{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100208 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200209} \
210static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
211{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100212 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200213}
214
215CFQ_CFQQ_FNS(on_rr);
216CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200217CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200218CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200219CFQ_CFQQ_FNS(fifo_expire);
220CFQ_CFQQ_FNS(idle_window);
221CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100222CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200223CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200224CFQ_CFQQ_FNS(coop);
Jens Axboe3b181522005-06-27 10:56:24 +0200225#undef CFQ_CFQQ_FNS
226
Jens Axboe7b679132008-05-30 12:23:07 +0200227#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
228 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
229#define cfq_log(cfqd, fmt, args...) \
230 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
231
Jens Axboe165125e2007-07-24 09:28:11 +0200232static void cfq_dispatch_insert(struct request_queue *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200233static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
Jens Axboefd0928d2008-01-24 08:52:45 +0100234 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100235static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200236 struct io_context *);
237
Jens Axboe5ad531d2009-07-03 12:57:48 +0200238static inline int rq_in_driver(struct cfq_data *cfqd)
239{
240 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
241}
242
Vasily Tarasov91fac312007-04-25 12:29:51 +0200243static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
244 int is_sync)
245{
246 return cic->cfqq[!!is_sync];
247}
248
249static inline void cic_set_cfqq(struct cfq_io_context *cic,
250 struct cfq_queue *cfqq, int is_sync)
251{
252 cic->cfqq[!!is_sync] = cfqq;
253}
254
255/*
256 * We regard a request as SYNC, if it's either a read or has the SYNC bit
257 * set (in which case it could also be direct WRITE).
258 */
259static inline int cfq_bio_sync(struct bio *bio)
260{
Jens Axboe1f98a132009-09-11 14:32:04 +0200261 if (bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO))
Vasily Tarasov91fac312007-04-25 12:29:51 +0200262 return 1;
263
264 return 0;
265}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700268 * scheduler run of queue, if there are requests pending and no one in the
269 * driver that will restart queueing
270 */
Jens Axboe8e296752009-10-03 16:26:03 +0200271static inline void cfq_schedule_dispatch(struct cfq_data *cfqd,
272 unsigned long delay)
Andrew Morton99f95e52005-06-27 20:14:05 -0700273{
Jens Axboe7b679132008-05-30 12:23:07 +0200274 if (cfqd->busy_queues) {
275 cfq_log(cfqd, "schedule dispatch");
Jens Axboe8e296752009-10-03 16:26:03 +0200276 kblockd_schedule_delayed_work(cfqd->queue, &cfqd->unplug_work,
277 delay);
Jens Axboe7b679132008-05-30 12:23:07 +0200278 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700279}
280
Jens Axboe165125e2007-07-24 09:28:11 +0200281static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700282{
283 struct cfq_data *cfqd = q->elevator->elevator_data;
284
Jens Axboeb4878f22005-10-20 16:42:29 +0200285 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700286}
287
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100289 * Scale schedule slice based on io priority. Use the sync time slice only
290 * if a queue is marked sync and has sync io queued. A sync queue with async
291 * io only, should not get full sync slice length.
292 */
Jens Axboed9e76202007-04-20 14:27:50 +0200293static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
294 unsigned short prio)
295{
296 const int base_slice = cfqd->cfq_slice[sync];
297
298 WARN_ON(prio >= IOPRIO_BE_NR);
299
300 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
301}
302
Jens Axboe44f7c162007-01-19 11:51:58 +1100303static inline int
304cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
305{
Jens Axboed9e76202007-04-20 14:27:50 +0200306 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100307}
308
309static inline void
310cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
311{
312 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200313 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100314}
315
316/*
317 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
318 * isn't valid until the first request from the dispatch is activated
319 * and the slice time set.
320 */
321static inline int cfq_slice_used(struct cfq_queue *cfqq)
322{
323 if (cfq_cfqq_slice_new(cfqq))
324 return 0;
325 if (time_before(jiffies, cfqq->slice_end))
326 return 0;
327
328 return 1;
329}
330
331/*
Jens Axboe5e705372006-07-13 12:39:25 +0200332 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200334 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 */
Jens Axboe5e705372006-07-13 12:39:25 +0200336static struct request *
337cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338{
339 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200341#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
342#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
343 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344
Jens Axboe5e705372006-07-13 12:39:25 +0200345 if (rq1 == NULL || rq1 == rq2)
346 return rq2;
347 if (rq2 == NULL)
348 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200349
Jens Axboe5e705372006-07-13 12:39:25 +0200350 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
351 return rq1;
352 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
353 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200354 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
355 return rq1;
356 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
357 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358
Tejun Heo83096eb2009-05-07 22:24:39 +0900359 s1 = blk_rq_pos(rq1);
360 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361
Jens Axboe6d048f52007-04-25 12:44:27 +0200362 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 /*
365 * by definition, 1KiB is 2 sectors
366 */
367 back_max = cfqd->cfq_back_max * 2;
368
369 /*
370 * Strict one way elevator _except_ in the case where we allow
371 * short backward seeks which are biased as twice the cost of a
372 * similar forward seek.
373 */
374 if (s1 >= last)
375 d1 = s1 - last;
376 else if (s1 + back_max >= last)
377 d1 = (last - s1) * cfqd->cfq_back_penalty;
378 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200379 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380
381 if (s2 >= last)
382 d2 = s2 - last;
383 else if (s2 + back_max >= last)
384 d2 = (last - s2) * cfqd->cfq_back_penalty;
385 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200386 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387
388 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
Andreas Mohre8a99052006-03-28 08:59:49 +0200390 /*
391 * By doing switch() on the bit mask "wrap" we avoid having to
392 * check two variables for all permutations: --> faster!
393 */
394 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200395 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200396 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200397 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200398 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200399 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200400 else {
401 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200402 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200403 else
Jens Axboe5e705372006-07-13 12:39:25 +0200404 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200405 }
406
407 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200408 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200409 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200410 return rq2;
411 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200412 default:
413 /*
414 * Since both rqs are wrapped,
415 * start with the one that's further behind head
416 * (--> only *one* back seek required),
417 * since back seek takes more time than forward.
418 */
419 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200420 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 else
Jens Axboe5e705372006-07-13 12:39:25 +0200422 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 }
424}
425
Jens Axboe498d3aa2007-04-26 12:54:48 +0200426/*
427 * The below is leftmost cache rbtree addon
428 */
Jens Axboe08717142008-01-28 11:38:15 +0100429static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200430{
431 if (!root->left)
432 root->left = rb_first(&root->rb);
433
Jens Axboe08717142008-01-28 11:38:15 +0100434 if (root->left)
435 return rb_entry(root->left, struct cfq_queue, rb_node);
436
437 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200438}
439
Jens Axboea36e71f2009-04-15 12:15:11 +0200440static void rb_erase_init(struct rb_node *n, struct rb_root *root)
441{
442 rb_erase(n, root);
443 RB_CLEAR_NODE(n);
444}
445
Jens Axboecc09e292007-04-26 12:53:50 +0200446static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
447{
448 if (root->left == n)
449 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200450 rb_erase_init(n, &root->rb);
Jens Axboecc09e292007-04-26 12:53:50 +0200451}
452
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453/*
454 * would be nice to take fifo expire time into account as well
455 */
Jens Axboe5e705372006-07-13 12:39:25 +0200456static struct request *
457cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
458 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459{
Jens Axboe21183b02006-07-13 12:33:14 +0200460 struct rb_node *rbnext = rb_next(&last->rb_node);
461 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200462 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463
Jens Axboe21183b02006-07-13 12:33:14 +0200464 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465
466 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200467 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200468
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200470 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200471 else {
472 rbnext = rb_first(&cfqq->sort_list);
473 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200474 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200475 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476
Jens Axboe21183b02006-07-13 12:33:14 +0200477 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478}
479
Jens Axboed9e76202007-04-20 14:27:50 +0200480static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
481 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482{
Jens Axboed9e76202007-04-20 14:27:50 +0200483 /*
484 * just an approximation, should be ok.
485 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200486 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
487 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200488}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489
Jens Axboe498d3aa2007-04-26 12:54:48 +0200490/*
491 * The cfqd->service_tree holds all pending cfq_queue's that have
492 * requests waiting to be processed. It is sorted in the order that
493 * we will service the queues.
494 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200495static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
496 int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200497{
Jens Axboe08717142008-01-28 11:38:15 +0100498 struct rb_node **p, *parent;
499 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +0200500 unsigned long rb_key;
Jens Axboe498d3aa2007-04-26 12:54:48 +0200501 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200502
Jens Axboe08717142008-01-28 11:38:15 +0100503 if (cfq_class_idle(cfqq)) {
504 rb_key = CFQ_IDLE_DELAY;
505 parent = rb_last(&cfqd->service_tree.rb);
506 if (parent && parent != &cfqq->rb_node) {
507 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
508 rb_key += __cfqq->rb_key;
509 } else
510 rb_key += jiffies;
511 } else if (!add_front) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200512 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
513 rb_key += cfqq->slice_resid;
514 cfqq->slice_resid = 0;
515 } else
516 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200517
518 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100519 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200520 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100521 */
Jens Axboed9e76202007-04-20 14:27:50 +0200522 if (rb_key == cfqq->rb_key)
523 return;
Jens Axboe53b03742006-07-28 09:48:51 +0200524
Jens Axboecc09e292007-04-26 12:53:50 +0200525 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200526 }
Jens Axboed9e76202007-04-20 14:27:50 +0200527
Jens Axboe498d3aa2007-04-26 12:54:48 +0200528 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +0100529 parent = NULL;
530 p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200531 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +0200532 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200533
Jens Axboed9e76202007-04-20 14:27:50 +0200534 parent = *p;
535 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
536
Jens Axboe0c534e02007-04-18 20:01:57 +0200537 /*
538 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200539 * preference to them. IDLE queues goes to the back.
540 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200541 */
542 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200543 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200544 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200545 n = &(*p)->rb_right;
546 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
547 n = &(*p)->rb_left;
548 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
549 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200550 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200551 n = &(*p)->rb_left;
552 else
553 n = &(*p)->rb_right;
554
555 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200556 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200557
558 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200559 }
560
Jens Axboecc09e292007-04-26 12:53:50 +0200561 if (left)
562 cfqd->service_tree.left = &cfqq->rb_node;
563
Jens Axboed9e76202007-04-20 14:27:50 +0200564 cfqq->rb_key = rb_key;
565 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200566 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567}
568
Jens Axboea36e71f2009-04-15 12:15:11 +0200569static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200570cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
571 sector_t sector, struct rb_node **ret_parent,
572 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +0200573{
Jens Axboea36e71f2009-04-15 12:15:11 +0200574 struct rb_node **p, *parent;
575 struct cfq_queue *cfqq = NULL;
576
577 parent = NULL;
578 p = &root->rb_node;
579 while (*p) {
580 struct rb_node **n;
581
582 parent = *p;
583 cfqq = rb_entry(parent, struct cfq_queue, p_node);
584
585 /*
586 * Sort strictly based on sector. Smallest to the left,
587 * largest to the right.
588 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900589 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +0200590 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900591 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +0200592 n = &(*p)->rb_left;
593 else
594 break;
595 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200596 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200597 }
598
599 *ret_parent = parent;
600 if (rb_link)
601 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200602 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +0200603}
604
605static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
606{
Jens Axboea36e71f2009-04-15 12:15:11 +0200607 struct rb_node **p, *parent;
608 struct cfq_queue *__cfqq;
609
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200610 if (cfqq->p_root) {
611 rb_erase(&cfqq->p_node, cfqq->p_root);
612 cfqq->p_root = NULL;
613 }
Jens Axboea36e71f2009-04-15 12:15:11 +0200614
615 if (cfq_class_idle(cfqq))
616 return;
617 if (!cfqq->next_rq)
618 return;
619
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200620 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900621 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
622 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200623 if (!__cfqq) {
624 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200625 rb_insert_color(&cfqq->p_node, cfqq->p_root);
626 } else
627 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200628}
629
Jens Axboe498d3aa2007-04-26 12:54:48 +0200630/*
631 * Update cfqq's position in the service tree.
632 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200633static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200634{
Jens Axboe6d048f52007-04-25 12:44:27 +0200635 /*
636 * Resorting requires the cfqq to be on the RR list already.
637 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200638 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200639 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +0200640 cfq_prio_tree_add(cfqd, cfqq);
641 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200642}
643
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644/*
645 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200646 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 */
Jens Axboefebffd62008-01-28 13:19:43 +0100648static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649{
Jens Axboe7b679132008-05-30 12:23:07 +0200650 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200651 BUG_ON(cfq_cfqq_on_rr(cfqq));
652 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 cfqd->busy_queues++;
654
Jens Axboeedd75ff2007-04-19 12:03:34 +0200655 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656}
657
Jens Axboe498d3aa2007-04-26 12:54:48 +0200658/*
659 * Called when the cfqq no longer has requests pending, remove it from
660 * the service tree.
661 */
Jens Axboefebffd62008-01-28 13:19:43 +0100662static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663{
Jens Axboe7b679132008-05-30 12:23:07 +0200664 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200665 BUG_ON(!cfq_cfqq_on_rr(cfqq));
666 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667
Jens Axboecc09e292007-04-26 12:53:50 +0200668 if (!RB_EMPTY_NODE(&cfqq->rb_node))
669 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200670 if (cfqq->p_root) {
671 rb_erase(&cfqq->p_node, cfqq->p_root);
672 cfqq->p_root = NULL;
673 }
Jens Axboed9e76202007-04-20 14:27:50 +0200674
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 BUG_ON(!cfqd->busy_queues);
676 cfqd->busy_queues--;
677}
678
679/*
680 * rb tree support functions
681 */
Jens Axboefebffd62008-01-28 13:19:43 +0100682static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683{
Jens Axboe5e705372006-07-13 12:39:25 +0200684 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200685 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200686 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687
Jens Axboeb4878f22005-10-20 16:42:29 +0200688 BUG_ON(!cfqq->queued[sync]);
689 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690
Jens Axboe5e705372006-07-13 12:39:25 +0200691 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692
Jens Axboedd67d052006-06-21 09:36:18 +0200693 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200694 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695}
696
Jens Axboe5e705372006-07-13 12:39:25 +0200697static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698{
Jens Axboe5e705372006-07-13 12:39:25 +0200699 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +0200701 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702
Jens Axboe5380a102006-07-13 12:37:56 +0200703 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704
705 /*
706 * looks a little odd, but the first insert might return an alias.
707 * if that happens, put the alias on the dispatch list
708 */
Jens Axboe21183b02006-07-13 12:33:14 +0200709 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200710 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100711
712 if (!cfq_cfqq_on_rr(cfqq))
713 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200714
715 /*
716 * check if this request is a better next-serve candidate
717 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200718 prev = cfqq->next_rq;
Jens Axboe5044eed2007-04-25 11:53:48 +0200719 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
Jens Axboea36e71f2009-04-15 12:15:11 +0200720
721 /*
722 * adjust priority tree position, if ->next_rq changes
723 */
724 if (prev != cfqq->next_rq)
725 cfq_prio_tree_add(cfqd, cfqq);
726
Jens Axboe5044eed2007-04-25 11:53:48 +0200727 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728}
729
Jens Axboefebffd62008-01-28 13:19:43 +0100730static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731{
Jens Axboe5380a102006-07-13 12:37:56 +0200732 elv_rb_del(&cfqq->sort_list, rq);
733 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200734 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735}
736
Jens Axboe206dc692006-03-28 13:03:44 +0200737static struct request *
738cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739{
Jens Axboe206dc692006-03-28 13:03:44 +0200740 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200741 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200742 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743
Jens Axboe4ac845a2008-01-24 08:44:49 +0100744 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200745 if (!cic)
746 return NULL;
747
748 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200749 if (cfqq) {
750 sector_t sector = bio->bi_sector + bio_sectors(bio);
751
Jens Axboe21183b02006-07-13 12:33:14 +0200752 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200753 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 return NULL;
756}
757
Jens Axboe165125e2007-07-24 09:28:11 +0200758static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200759{
760 struct cfq_data *cfqd = q->elevator->elevator_data;
761
Jens Axboe5ad531d2009-07-03 12:57:48 +0200762 cfqd->rq_in_driver[rq_is_sync(rq)]++;
Jens Axboe7b679132008-05-30 12:23:07 +0200763 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Jens Axboe5ad531d2009-07-03 12:57:48 +0200764 rq_in_driver(cfqd));
Jens Axboe25776e32006-06-01 10:12:26 +0200765
Tejun Heo5b936292009-05-07 22:24:38 +0900766 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200767}
768
Jens Axboe165125e2007-07-24 09:28:11 +0200769static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770{
Jens Axboe22e2c502005-06-27 10:55:12 +0200771 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5ad531d2009-07-03 12:57:48 +0200772 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773
Jens Axboe5ad531d2009-07-03 12:57:48 +0200774 WARN_ON(!cfqd->rq_in_driver[sync]);
775 cfqd->rq_in_driver[sync]--;
Jens Axboe7b679132008-05-30 12:23:07 +0200776 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Jens Axboe5ad531d2009-07-03 12:57:48 +0200777 rq_in_driver(cfqd));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778}
779
Jens Axboeb4878f22005-10-20 16:42:29 +0200780static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781{
Jens Axboe5e705372006-07-13 12:39:25 +0200782 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200783
Jens Axboe5e705372006-07-13 12:39:25 +0200784 if (cfqq->next_rq == rq)
785 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786
Jens Axboeb4878f22005-10-20 16:42:29 +0200787 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200788 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200789
Aaron Carroll45333d52008-08-26 15:52:36 +0200790 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +0200791 if (rq_is_meta(rq)) {
792 WARN_ON(!cfqq->meta_pending);
793 cfqq->meta_pending--;
794 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795}
796
Jens Axboe165125e2007-07-24 09:28:11 +0200797static int cfq_merge(struct request_queue *q, struct request **req,
798 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799{
800 struct cfq_data *cfqd = q->elevator->elevator_data;
801 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802
Jens Axboe206dc692006-03-28 13:03:44 +0200803 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200804 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200805 *req = __rq;
806 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 }
808
809 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810}
811
Jens Axboe165125e2007-07-24 09:28:11 +0200812static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200813 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814{
Jens Axboe21183b02006-07-13 12:33:14 +0200815 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200816 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817
Jens Axboe5e705372006-07-13 12:39:25 +0200818 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820}
821
822static void
Jens Axboe165125e2007-07-24 09:28:11 +0200823cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 struct request *next)
825{
Jens Axboe22e2c502005-06-27 10:55:12 +0200826 /*
827 * reposition in fifo if next is older than rq
828 */
829 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
830 time_before(next->start_time, rq->start_time))
831 list_move(&rq->queuelist, &next->queuelist);
832
Jens Axboeb4878f22005-10-20 16:42:29 +0200833 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200834}
835
Jens Axboe165125e2007-07-24 09:28:11 +0200836static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100837 struct bio *bio)
838{
839 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200840 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100841 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100842
843 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100844 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100845 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200846 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100847 return 0;
848
849 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100850 * Lookup the cfqq that this bio will be queued with. Allow
851 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100852 */
Jens Axboe4ac845a2008-01-24 08:44:49 +0100853 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200854 if (!cic)
855 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100856
Vasily Tarasov91fac312007-04-25 12:29:51 +0200857 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100858 if (cfqq == RQ_CFQQ(rq))
859 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100860
Jens Axboeec8acb62007-01-02 18:32:11 +0100861 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100862}
863
Jens Axboefebffd62008-01-28 13:19:43 +0100864static void __cfq_set_active_queue(struct cfq_data *cfqd,
865 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +0200866{
867 if (cfqq) {
Jens Axboe7b679132008-05-30 12:23:07 +0200868 cfq_log_cfqq(cfqd, cfqq, "set_active");
Jens Axboe22e2c502005-06-27 10:55:12 +0200869 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +0200870 cfqq->slice_dispatch = 0;
871
Jens Axboe2f5cb732009-04-07 08:51:19 +0200872 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +0200873 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200874 cfq_clear_cfqq_must_alloc_slice(cfqq);
875 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100876 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +0200877
878 del_timer(&cfqd->idle_slice_timer);
Jens Axboe22e2c502005-06-27 10:55:12 +0200879 }
880
881 cfqd->active_queue = cfqq;
882}
883
884/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100885 * current cfqq expired its slice (or was too idle), select new one
886 */
887static void
888__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200889 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100890{
Jens Axboe7b679132008-05-30 12:23:07 +0200891 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
892
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100893 if (cfq_cfqq_wait_request(cfqq))
894 del_timer(&cfqd->idle_slice_timer);
895
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100896 cfq_clear_cfqq_wait_request(cfqq);
897
898 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200899 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100900 */
Jens Axboe7b679132008-05-30 12:23:07 +0200901 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +1100902 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200903 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
904 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100905
Jens Axboeedd75ff2007-04-19 12:03:34 +0200906 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100907
908 if (cfqq == cfqd->active_queue)
909 cfqd->active_queue = NULL;
910
911 if (cfqd->active_cic) {
912 put_io_context(cfqd->active_cic->ioc);
913 cfqd->active_cic = NULL;
914 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100915}
916
Jens Axboe6084cdd2007-04-23 08:25:00 +0200917static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100918{
919 struct cfq_queue *cfqq = cfqd->active_queue;
920
921 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200922 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100923}
924
Jens Axboe498d3aa2007-04-26 12:54:48 +0200925/*
926 * Get next queue for service. Unless we have a queue preemption,
927 * we'll simply select the first cfqq in the service tree.
928 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200929static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200930{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200931 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
932 return NULL;
933
Jens Axboe08717142008-01-28 11:38:15 +0100934 return cfq_rb_first(&cfqd->service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +0200935}
936
Jens Axboe498d3aa2007-04-26 12:54:48 +0200937/*
938 * Get and set a new active queue for service.
939 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200940static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
941 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200942{
Jens Axboea36e71f2009-04-15 12:15:11 +0200943 if (!cfqq) {
944 cfqq = cfq_get_next_queue(cfqd);
945 if (cfqq)
946 cfq_clear_cfqq_coop(cfqq);
947 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200948
Jens Axboe22e2c502005-06-27 10:55:12 +0200949 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200950 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200951}
952
Jens Axboed9e76202007-04-20 14:27:50 +0200953static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
954 struct request *rq)
955{
Tejun Heo83096eb2009-05-07 22:24:39 +0900956 if (blk_rq_pos(rq) >= cfqd->last_position)
957 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +0200958 else
Tejun Heo83096eb2009-05-07 22:24:39 +0900959 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +0200960}
961
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200962#define CIC_SEEK_THR 8 * 1024
963#define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
964
Jens Axboe6d048f52007-04-25 12:44:27 +0200965static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
966{
967 struct cfq_io_context *cic = cfqd->active_cic;
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200968 sector_t sdist = cic->seek_mean;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200969
Jens Axboe6d048f52007-04-25 12:44:27 +0200970 if (!sample_valid(cic->seek_samples))
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200971 sdist = CIC_SEEK_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +0200972
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200973 return cfq_dist_from_last(cfqd, rq) <= sdist;
Jens Axboe6d048f52007-04-25 12:44:27 +0200974}
975
Jens Axboea36e71f2009-04-15 12:15:11 +0200976static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
977 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200978{
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200979 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +0200980 struct rb_node *parent, *node;
981 struct cfq_queue *__cfqq;
982 sector_t sector = cfqd->last_position;
983
984 if (RB_EMPTY_ROOT(root))
985 return NULL;
986
987 /*
988 * First, if we find a request starting at the end of the last
989 * request, choose it.
990 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200991 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +0200992 if (__cfqq)
993 return __cfqq;
994
995 /*
996 * If the exact sector wasn't found, the parent of the NULL leaf
997 * will contain the closest sector.
998 */
999 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1000 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1001 return __cfqq;
1002
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001003 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001004 node = rb_next(&__cfqq->p_node);
1005 else
1006 node = rb_prev(&__cfqq->p_node);
1007 if (!node)
1008 return NULL;
1009
1010 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1011 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1012 return __cfqq;
1013
1014 return NULL;
1015}
1016
1017/*
1018 * cfqd - obvious
1019 * cur_cfqq - passed in so that we don't decide that the current queue is
1020 * closely cooperating with itself.
1021 *
1022 * So, basically we're assuming that that cur_cfqq has dispatched at least
1023 * one request, and that cfqd->last_position reflects a position on the disk
1024 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1025 * assumption.
1026 */
1027static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1028 struct cfq_queue *cur_cfqq,
1029 int probe)
1030{
1031 struct cfq_queue *cfqq;
1032
1033 /*
1034 * A valid cfq_io_context is necessary to compare requests against
1035 * the seek_mean of the current cfqq.
1036 */
1037 if (!cfqd->active_cic)
1038 return NULL;
1039
Jens Axboe6d048f52007-04-25 12:44:27 +02001040 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001041 * We should notice if some of the queues are cooperating, eg
1042 * working closely on the same area of the disk. In that case,
1043 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001044 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001045 cfqq = cfqq_close(cfqd, cur_cfqq);
1046 if (!cfqq)
1047 return NULL;
1048
1049 if (cfq_cfqq_coop(cfqq))
1050 return NULL;
1051
1052 if (!probe)
1053 cfq_mark_cfqq_coop(cfqq);
1054 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001055}
1056
Jens Axboe6d048f52007-04-25 12:44:27 +02001057static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001058{
Jens Axboe17926692007-01-19 11:59:30 +11001059 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001060 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001061 unsigned long sl;
1062
Jens Axboea68bbdd2008-09-24 13:03:33 +02001063 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001064 * SSD device without seek penalty, disable idling. But only do so
1065 * for devices that support queuing, otherwise we still have a problem
1066 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001067 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001068 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001069 return;
1070
Jens Axboedd67d052006-06-21 09:36:18 +02001071 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001072 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001073
1074 /*
1075 * idle is disabled, either manually or by past process history
1076 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001077 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1078 return;
1079
Jens Axboe22e2c502005-06-27 10:55:12 +02001080 /*
Jens Axboe7b679132008-05-30 12:23:07 +02001081 * still requests with the driver, don't idle
1082 */
Jens Axboe5ad531d2009-07-03 12:57:48 +02001083 if (rq_in_driver(cfqd))
Jens Axboe7b679132008-05-30 12:23:07 +02001084 return;
1085
1086 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001087 * task has exited, don't wait
1088 */
Jens Axboe206dc692006-03-28 13:03:44 +02001089 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001090 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001091 return;
1092
Jens Axboe3b181522005-06-27 10:56:24 +02001093 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001094
Jens Axboe206dc692006-03-28 13:03:44 +02001095 /*
1096 * we don't want to idle for seeks, but we do want to allow
1097 * fair distribution of slice time for a process doing back-to-back
1098 * seeks. so allow a little bit of time for him to submit a new rq
1099 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001100 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001101 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +02001102 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +02001103
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001104 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe9481ffd2009-04-15 12:14:13 +02001105 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106}
1107
Jens Axboe498d3aa2007-04-26 12:54:48 +02001108/*
1109 * Move request from internal lists to the request queue dispatch list.
1110 */
Jens Axboe165125e2007-07-24 09:28:11 +02001111static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001113 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001114 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001115
Jens Axboe7b679132008-05-30 12:23:07 +02001116 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1117
Jeff Moyer06d21882009-09-11 17:08:59 +02001118 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001119 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001120 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001121 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001122
1123 if (cfq_cfqq_sync(cfqq))
1124 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125}
1126
1127/*
1128 * return expired entry, or NULL to just start from scratch in rbtree
1129 */
Jens Axboefebffd62008-01-28 13:19:43 +01001130static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131{
1132 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +02001133 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +02001134 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135
Jens Axboe3b181522005-06-27 10:56:24 +02001136 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001138
1139 cfq_mark_cfqq_fifo_expire(cfqq);
1140
Jens Axboe89850f72006-07-22 16:48:31 +02001141 if (list_empty(&cfqq->fifo))
1142 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143
Jens Axboe6d048f52007-04-25 12:44:27 +02001144 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +02001145 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146
Jens Axboe6d048f52007-04-25 12:44:27 +02001147 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
Jens Axboe7b679132008-05-30 12:23:07 +02001148 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149
Jens Axboe7b679132008-05-30 12:23:07 +02001150 cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001151 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152}
1153
Jens Axboe22e2c502005-06-27 10:55:12 +02001154static inline int
1155cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1156{
1157 const int base_rq = cfqd->cfq_slice_async_rq;
1158
1159 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1160
1161 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1162}
1163
1164/*
Jens Axboe498d3aa2007-04-26 12:54:48 +02001165 * Select a queue for service. If we have a current active queue,
1166 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02001167 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001168static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001169{
Jens Axboea36e71f2009-04-15 12:15:11 +02001170 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001171
1172 cfqq = cfqd->active_queue;
1173 if (!cfqq)
1174 goto new_queue;
1175
1176 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001177 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02001178 */
Jens Axboeb0291952009-04-07 11:38:31 +02001179 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001180 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +02001181
1182 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001183 * The active queue has requests and isn't expired, allow it to
1184 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02001185 */
Jens Axboedd67d052006-06-21 09:36:18 +02001186 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001187 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02001188
1189 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02001190 * If another queue has a request waiting within our mean seek
1191 * distance, let it run. The expire code will check for close
1192 * cooperators and put the close queue at the front of the service
1193 * tree.
1194 */
1195 new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1196 if (new_cfqq)
1197 goto expire;
1198
1199 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001200 * No requests pending. If the active queue still has requests in
1201 * flight or is idling for a new request, allow either of these
1202 * conditions to happen (or time out) before selecting a new queue.
1203 */
Jens Axboecc197472007-04-20 20:45:39 +02001204 if (timer_pending(&cfqd->idle_slice_timer) ||
1205 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001206 cfqq = NULL;
1207 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001208 }
1209
Jens Axboe3b181522005-06-27 10:56:24 +02001210expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001211 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001212new_queue:
Jens Axboea36e71f2009-04-15 12:15:11 +02001213 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001214keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001215 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001216}
1217
Jens Axboefebffd62008-01-28 13:19:43 +01001218static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02001219{
1220 int dispatched = 0;
1221
1222 while (cfqq->next_rq) {
1223 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1224 dispatched++;
1225 }
1226
1227 BUG_ON(!list_empty(&cfqq->fifo));
1228 return dispatched;
1229}
1230
Jens Axboe498d3aa2007-04-26 12:54:48 +02001231/*
1232 * Drain our current requests. Used for barriers and when switching
1233 * io schedulers on-the-fly.
1234 */
Jens Axboed9e76202007-04-20 14:27:50 +02001235static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001236{
Jens Axboe08717142008-01-28 11:38:15 +01001237 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001238 int dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001239
Jens Axboe08717142008-01-28 11:38:15 +01001240 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
Jens Axboed9e76202007-04-20 14:27:50 +02001241 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001242
Jens Axboe6084cdd2007-04-23 08:25:00 +02001243 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001244
1245 BUG_ON(cfqd->busy_queues);
1246
Jeff Moyer69237152009-06-12 15:29:30 +02001247 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001248 return dispatched;
1249}
1250
Jens Axboe2f5cb732009-04-07 08:51:19 +02001251/*
1252 * Dispatch a request from cfqq, moving them to the request queue
1253 * dispatch list.
1254 */
1255static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1256{
1257 struct request *rq;
1258
1259 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1260
1261 /*
1262 * follow expired path, else get first next available
1263 */
1264 rq = cfq_check_fifo(cfqq);
1265 if (!rq)
1266 rq = cfqq->next_rq;
1267
1268 /*
1269 * insert request into driver dispatch list
1270 */
1271 cfq_dispatch_insert(cfqd->queue, rq);
1272
1273 if (!cfqd->active_cic) {
1274 struct cfq_io_context *cic = RQ_CIC(rq);
1275
Nikanth Karthikesand9c7d392009-06-10 12:57:06 -07001276 atomic_long_inc(&cic->ioc->refcount);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001277 cfqd->active_cic = cic;
1278 }
1279}
1280
1281/*
1282 * Find the cfqq that we need to service and move a request from that to the
1283 * dispatch list
1284 */
Jens Axboe165125e2007-07-24 09:28:11 +02001285static int cfq_dispatch_requests(struct request_queue *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286{
1287 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001288 struct cfq_queue *cfqq;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001289 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290
Jens Axboe22e2c502005-06-27 10:55:12 +02001291 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 return 0;
1293
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001294 if (unlikely(force))
1295 return cfq_forced_dispatch(cfqd);
1296
Jens Axboe2f5cb732009-04-07 08:51:19 +02001297 cfqq = cfq_select_queue(cfqd);
1298 if (!cfqq)
1299 return 0;
Jens Axboeb4878f22005-10-20 16:42:29 +02001300
Jens Axboe2f5cb732009-04-07 08:51:19 +02001301 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02001302 * Drain async requests before we start sync IO
1303 */
1304 if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1305 return 0;
1306
1307 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02001308 * If this is an async queue and we have sync IO in flight, let it wait
1309 */
1310 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1311 return 0;
1312
1313 max_dispatch = cfqd->cfq_quantum;
1314 if (cfq_class_idle(cfqq))
1315 max_dispatch = 1;
1316
1317 /*
1318 * Does this cfqq already have too much IO in flight?
1319 */
1320 if (cfqq->dispatched >= max_dispatch) {
1321 /*
1322 * idle queue must always only have a single IO in flight
1323 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02001324 if (cfq_class_idle(cfqq))
Jens Axboe2f5cb732009-04-07 08:51:19 +02001325 return 0;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001326
Jens Axboe2f5cb732009-04-07 08:51:19 +02001327 /*
1328 * We have other queues, don't allow more IO from this one
1329 */
1330 if (cfqd->busy_queues > 1)
1331 return 0;
Jens Axboe9ede2092007-01-19 12:11:44 +11001332
Jens Axboe2f5cb732009-04-07 08:51:19 +02001333 /*
Jens Axboe8e296752009-10-03 16:26:03 +02001334 * Sole queue user, allow bigger slice
Vivek Goyal365722b2009-10-03 15:21:27 +02001335 */
Jens Axboe8e296752009-10-03 16:26:03 +02001336 max_dispatch *= 4;
1337 }
1338
1339 /*
1340 * Async queues must wait a bit before being allowed dispatch.
1341 * We also ramp up the dispatch depth gradually for async IO,
1342 * based on the last sync IO we serviced
1343 */
Jens Axboe963b72f2009-10-03 19:42:18 +02001344 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Jens Axboe8e296752009-10-03 16:26:03 +02001345 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1346 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02001347
Jens Axboe61f0c1d2009-10-03 19:46:03 +02001348 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02001349 if (!depth && !cfqq->dispatched)
1350 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02001351 if (depth < max_dispatch)
1352 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 }
1354
Jens Axboe8e296752009-10-03 16:26:03 +02001355 if (cfqq->dispatched >= max_dispatch)
1356 return 0;
1357
Jens Axboe2f5cb732009-04-07 08:51:19 +02001358 /*
1359 * Dispatch a request from this cfqq
1360 */
1361 cfq_dispatch_request(cfqd, cfqq);
1362 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02001363 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001364
1365 /*
1366 * expire an async queue immediately if it has used up its slice. idle
1367 * queue always expire after 1 dispatch round.
1368 */
1369 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1370 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1371 cfq_class_idle(cfqq))) {
1372 cfqq->slice_end = jiffies + 1;
1373 cfq_slice_expired(cfqd, 0);
1374 }
1375
Shan Weib217a902009-09-01 10:06:42 +02001376 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02001377 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378}
1379
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380/*
Jens Axboe5e705372006-07-13 12:39:25 +02001381 * task holds one reference to the queue, dropped when task exits. each rq
1382 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 *
1384 * queue lock must be held here.
1385 */
1386static void cfq_put_queue(struct cfq_queue *cfqq)
1387{
Jens Axboe22e2c502005-06-27 10:55:12 +02001388 struct cfq_data *cfqd = cfqq->cfqd;
1389
1390 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391
1392 if (!atomic_dec_and_test(&cfqq->ref))
1393 return;
1394
Jens Axboe7b679132008-05-30 12:23:07 +02001395 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001397 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001398 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001400 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001401 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe8e296752009-10-03 16:26:03 +02001402 cfq_schedule_dispatch(cfqd, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001403 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001404
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405 kmem_cache_free(cfq_pool, cfqq);
1406}
1407
Jens Axboed6de8be2008-05-28 14:46:59 +02001408/*
1409 * Must always be called with the rcu_read_lock() held
1410 */
Jens Axboe07416d22008-05-07 09:17:12 +02001411static void
1412__call_for_each_cic(struct io_context *ioc,
1413 void (*func)(struct io_context *, struct cfq_io_context *))
1414{
1415 struct cfq_io_context *cic;
1416 struct hlist_node *n;
1417
1418 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1419 func(ioc, cic);
1420}
1421
Jens Axboe4ac845a2008-01-24 08:44:49 +01001422/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001423 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001424 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001425static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001426call_for_each_cic(struct io_context *ioc,
1427 void (*func)(struct io_context *, struct cfq_io_context *))
1428{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001429 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02001430 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001431 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001432}
Jens Axboe4ac845a2008-01-24 08:44:49 +01001433
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001434static void cfq_cic_free_rcu(struct rcu_head *head)
1435{
1436 struct cfq_io_context *cic;
1437
1438 cic = container_of(head, struct cfq_io_context, rcu_head);
1439
1440 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09001441 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001442
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001443 if (ioc_gone) {
1444 /*
1445 * CFQ scheduler is exiting, grab exit lock and check
1446 * the pending io context count. If it hits zero,
1447 * complete ioc_gone and set it back to NULL
1448 */
1449 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09001450 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001451 complete(ioc_gone);
1452 ioc_gone = NULL;
1453 }
1454 spin_unlock(&ioc_gone_lock);
1455 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001456}
1457
1458static void cfq_cic_free(struct cfq_io_context *cic)
1459{
1460 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001461}
1462
1463static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1464{
1465 unsigned long flags;
1466
1467 BUG_ON(!cic->dead_key);
1468
1469 spin_lock_irqsave(&ioc->lock, flags);
1470 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01001471 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001472 spin_unlock_irqrestore(&ioc->lock, flags);
1473
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001474 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001475}
1476
Jens Axboed6de8be2008-05-28 14:46:59 +02001477/*
1478 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1479 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1480 * and ->trim() which is called with the task lock held
1481 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001482static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001484 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001485 * ioc->refcount is zero here, or we are called from elv_unregister(),
1486 * so no more cic's are allowed to be linked into this ioc. So it
1487 * should be ok to iterate over the known list, we will see all cic's
1488 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001489 */
Jens Axboe07416d22008-05-07 09:17:12 +02001490 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491}
1492
Jens Axboe89850f72006-07-22 16:48:31 +02001493static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1494{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001495 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001496 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe8e296752009-10-03 16:26:03 +02001497 cfq_schedule_dispatch(cfqd, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001498 }
Jens Axboe89850f72006-07-22 16:48:31 +02001499
1500 cfq_put_queue(cfqq);
1501}
1502
1503static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1504 struct cfq_io_context *cic)
1505{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001506 struct io_context *ioc = cic->ioc;
1507
Jens Axboefc463792006-08-29 09:05:44 +02001508 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001509
1510 /*
1511 * Make sure key == NULL is seen for dead queues
1512 */
Jens Axboefc463792006-08-29 09:05:44 +02001513 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001514 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02001515 cic->key = NULL;
1516
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001517 if (ioc->ioc_data == cic)
1518 rcu_assign_pointer(ioc->ioc_data, NULL);
1519
Jens Axboeff6657c2009-04-08 10:58:57 +02001520 if (cic->cfqq[BLK_RW_ASYNC]) {
1521 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1522 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001523 }
1524
Jens Axboeff6657c2009-04-08 10:58:57 +02001525 if (cic->cfqq[BLK_RW_SYNC]) {
1526 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1527 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001528 }
Jens Axboe89850f72006-07-22 16:48:31 +02001529}
1530
Jens Axboe4ac845a2008-01-24 08:44:49 +01001531static void cfq_exit_single_io_context(struct io_context *ioc,
1532 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001533{
Al Viro478a82b2006-03-18 13:25:24 -05001534 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001535
Jens Axboe89850f72006-07-22 16:48:31 +02001536 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001537 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001538 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02001539
Jens Axboe4ac845a2008-01-24 08:44:49 +01001540 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01001541
1542 /*
1543 * Ensure we get a fresh copy of the ->key to prevent
1544 * race between exiting task and queue
1545 */
1546 smp_read_barrier_depends();
1547 if (cic->key)
1548 __cfq_exit_single_io_context(cfqd, cic);
1549
Jens Axboe4ac845a2008-01-24 08:44:49 +01001550 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05001551 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001552}
1553
Jens Axboe498d3aa2007-04-26 12:54:48 +02001554/*
1555 * The process that ioc belongs to has exited, we need to clean up
1556 * and put the internal structures we have that belongs to that process.
1557 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001558static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001560 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561}
1562
Jens Axboe22e2c502005-06-27 10:55:12 +02001563static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001564cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565{
Jens Axboeb5deef92006-07-19 23:39:40 +02001566 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567
Christoph Lameter94f60302007-07-17 04:03:29 -07001568 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1569 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001571 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001572 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01001573 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001574 cic->dtor = cfq_free_io_context;
1575 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09001576 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 }
1578
1579 return cic;
1580}
1581
Jens Axboefd0928d2008-01-24 08:52:45 +01001582static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02001583{
1584 struct task_struct *tsk = current;
1585 int ioprio_class;
1586
Jens Axboe3b181522005-06-27 10:56:24 +02001587 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001588 return;
1589
Jens Axboefd0928d2008-01-24 08:52:45 +01001590 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001591 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01001592 default:
1593 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1594 case IOPRIO_CLASS_NONE:
1595 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02001596 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01001597 */
1598 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02001599 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01001600 break;
1601 case IOPRIO_CLASS_RT:
1602 cfqq->ioprio = task_ioprio(ioc);
1603 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1604 break;
1605 case IOPRIO_CLASS_BE:
1606 cfqq->ioprio = task_ioprio(ioc);
1607 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1608 break;
1609 case IOPRIO_CLASS_IDLE:
1610 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1611 cfqq->ioprio = 7;
1612 cfq_clear_cfqq_idle_window(cfqq);
1613 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02001614 }
1615
1616 /*
1617 * keep track of original prio settings in case we have to temporarily
1618 * elevate the priority of this queue
1619 */
1620 cfqq->org_ioprio = cfqq->ioprio;
1621 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001622 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001623}
1624
Jens Axboefebffd62008-01-28 13:19:43 +01001625static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001626{
Al Viro478a82b2006-03-18 13:25:24 -05001627 struct cfq_data *cfqd = cic->key;
1628 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001629 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001630
Jens Axboecaaa5f92006-06-16 11:23:00 +02001631 if (unlikely(!cfqd))
1632 return;
1633
Jens Axboec1b707d2006-10-30 19:54:23 +01001634 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001635
Jens Axboeff6657c2009-04-08 10:58:57 +02001636 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001637 if (cfqq) {
1638 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02001639 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1640 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001641 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02001642 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001643 cfq_put_queue(cfqq);
1644 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001645 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001646
Jens Axboeff6657c2009-04-08 10:58:57 +02001647 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001648 if (cfqq)
1649 cfq_mark_cfqq_prio_changed(cfqq);
1650
Jens Axboec1b707d2006-10-30 19:54:23 +01001651 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001652}
1653
Jens Axboefc463792006-08-29 09:05:44 +02001654static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001656 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02001657 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658}
1659
Jens Axboed5036d72009-06-26 10:44:34 +02001660static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1661 pid_t pid, int is_sync)
1662{
1663 RB_CLEAR_NODE(&cfqq->rb_node);
1664 RB_CLEAR_NODE(&cfqq->p_node);
1665 INIT_LIST_HEAD(&cfqq->fifo);
1666
1667 atomic_set(&cfqq->ref, 0);
1668 cfqq->cfqd = cfqd;
1669
1670 cfq_mark_cfqq_prio_changed(cfqq);
1671
1672 if (is_sync) {
1673 if (!cfq_class_idle(cfqq))
1674 cfq_mark_cfqq_idle_window(cfqq);
1675 cfq_mark_cfqq_sync(cfqq);
1676 }
1677 cfqq->pid = pid;
1678}
1679
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001681cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01001682 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001685 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
1687retry:
Jens Axboe4ac845a2008-01-24 08:44:49 +01001688 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001689 /* cic always exists here */
1690 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691
Jens Axboe6118b702009-06-30 09:34:12 +02001692 /*
1693 * Always try a new alloc if we fell back to the OOM cfqq
1694 * originally, since it should just be a temporary situation.
1695 */
1696 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1697 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698 if (new_cfqq) {
1699 cfqq = new_cfqq;
1700 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001701 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001703 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02001704 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07001705 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02001707 if (new_cfqq)
1708 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001709 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001710 cfqq = kmem_cache_alloc_node(cfq_pool,
1711 gfp_mask | __GFP_ZERO,
1712 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001713 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714
Jens Axboe6118b702009-06-30 09:34:12 +02001715 if (cfqq) {
1716 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1717 cfq_init_prio_data(cfqq, ioc);
1718 cfq_log_cfqq(cfqd, cfqq, "alloced");
1719 } else
1720 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721 }
1722
1723 if (new_cfqq)
1724 kmem_cache_free(cfq_pool, new_cfqq);
1725
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 return cfqq;
1727}
1728
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001729static struct cfq_queue **
1730cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1731{
Jens Axboefe094d92008-01-31 13:08:54 +01001732 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001733 case IOPRIO_CLASS_RT:
1734 return &cfqd->async_cfqq[0][ioprio];
1735 case IOPRIO_CLASS_BE:
1736 return &cfqd->async_cfqq[1][ioprio];
1737 case IOPRIO_CLASS_IDLE:
1738 return &cfqd->async_idle_cfqq;
1739 default:
1740 BUG();
1741 }
1742}
1743
Jens Axboe15c31be2007-07-10 13:43:25 +02001744static struct cfq_queue *
Jens Axboefd0928d2008-01-24 08:52:45 +01001745cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02001746 gfp_t gfp_mask)
1747{
Jens Axboefd0928d2008-01-24 08:52:45 +01001748 const int ioprio = task_ioprio(ioc);
1749 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001750 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001751 struct cfq_queue *cfqq = NULL;
1752
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001753 if (!is_sync) {
1754 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1755 cfqq = *async_cfqq;
1756 }
1757
Jens Axboe6118b702009-06-30 09:34:12 +02001758 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01001759 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02001760
1761 /*
1762 * pin the queue now that it's allocated, scheduler exit will prune it
1763 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001764 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001765 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001766 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001767 }
1768
1769 atomic_inc(&cfqq->ref);
1770 return cfqq;
1771}
1772
Jens Axboe498d3aa2007-04-26 12:54:48 +02001773/*
1774 * We drop cfq io contexts lazily, so we may find a dead one.
1775 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001776static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001777cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1778 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001779{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001780 unsigned long flags;
1781
Jens Axboefc463792006-08-29 09:05:44 +02001782 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001783
Jens Axboe4ac845a2008-01-24 08:44:49 +01001784 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02001785
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001786 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001787
1788 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01001789 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001790 spin_unlock_irqrestore(&ioc->lock, flags);
1791
1792 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001793}
1794
Jens Axboee2d74ac2006-03-28 08:59:01 +02001795static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01001796cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001797{
Jens Axboee2d74ac2006-03-28 08:59:01 +02001798 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001799 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001800 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001801
Vasily Tarasov91fac312007-04-25 12:29:51 +02001802 if (unlikely(!ioc))
1803 return NULL;
1804
Jens Axboed6de8be2008-05-28 14:46:59 +02001805 rcu_read_lock();
1806
Jens Axboe597bc482007-04-24 21:23:53 +02001807 /*
1808 * we maintain a last-hit cache, to avoid browsing over the tree
1809 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001810 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02001811 if (cic && cic->key == cfqd) {
1812 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02001813 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001814 }
Jens Axboe597bc482007-04-24 21:23:53 +02001815
Jens Axboe4ac845a2008-01-24 08:44:49 +01001816 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001817 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1818 rcu_read_unlock();
1819 if (!cic)
1820 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001821 /* ->key must be copied to avoid race with cfq_exit_queue() */
1822 k = cic->key;
1823 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001824 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001825 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001826 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001827 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001828
Jens Axboed6de8be2008-05-28 14:46:59 +02001829 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001830 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001831 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001832 break;
1833 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001834
Jens Axboe4ac845a2008-01-24 08:44:49 +01001835 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001836}
1837
Jens Axboe4ac845a2008-01-24 08:44:49 +01001838/*
1839 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1840 * the process specific cfq io context when entered from the block layer.
1841 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1842 */
Jens Axboefebffd62008-01-28 13:19:43 +01001843static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1844 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001845{
Jens Axboe0261d682006-10-30 19:07:48 +01001846 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001847 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001848
Jens Axboe4ac845a2008-01-24 08:44:49 +01001849 ret = radix_tree_preload(gfp_mask);
1850 if (!ret) {
1851 cic->ioc = ioc;
1852 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001853
Jens Axboe4ac845a2008-01-24 08:44:49 +01001854 spin_lock_irqsave(&ioc->lock, flags);
1855 ret = radix_tree_insert(&ioc->radix_root,
1856 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01001857 if (!ret)
1858 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001859 spin_unlock_irqrestore(&ioc->lock, flags);
1860
1861 radix_tree_preload_end();
1862
1863 if (!ret) {
1864 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1865 list_add(&cic->queue_list, &cfqd->cic_list);
1866 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001867 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001868 }
1869
Jens Axboe4ac845a2008-01-24 08:44:49 +01001870 if (ret)
1871 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02001872
Jens Axboe4ac845a2008-01-24 08:44:49 +01001873 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001874}
1875
Jens Axboe22e2c502005-06-27 10:55:12 +02001876/*
1877 * Setup general io context and cfq io context. There can be several cfq
1878 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001879 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001880 */
1881static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001882cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883{
Jens Axboe22e2c502005-06-27 10:55:12 +02001884 struct io_context *ioc = NULL;
1885 struct cfq_io_context *cic;
1886
1887 might_sleep_if(gfp_mask & __GFP_WAIT);
1888
Jens Axboeb5deef92006-07-19 23:39:40 +02001889 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001890 if (!ioc)
1891 return NULL;
1892
Jens Axboe4ac845a2008-01-24 08:44:49 +01001893 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001894 if (cic)
1895 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001896
Jens Axboee2d74ac2006-03-28 08:59:01 +02001897 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1898 if (cic == NULL)
1899 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001900
Jens Axboe4ac845a2008-01-24 08:44:49 +01001901 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1902 goto err_free;
1903
Jens Axboe22e2c502005-06-27 10:55:12 +02001904out:
Jens Axboefc463792006-08-29 09:05:44 +02001905 smp_read_barrier_depends();
1906 if (unlikely(ioc->ioprio_changed))
1907 cfq_ioc_set_ioprio(ioc);
1908
Jens Axboe22e2c502005-06-27 10:55:12 +02001909 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001910err_free:
1911 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001912err:
1913 put_io_context(ioc);
1914 return NULL;
1915}
1916
1917static void
1918cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1919{
Jens Axboeaaf12282007-01-19 11:30:16 +11001920 unsigned long elapsed = jiffies - cic->last_end_request;
1921 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001922
1923 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1924 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1925 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1926}
1927
Jens Axboe206dc692006-03-28 13:03:44 +02001928static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001929cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1930 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001931{
1932 sector_t sdist;
1933 u64 total;
1934
Jeff Moyer4d00aa42009-04-21 07:25:04 +02001935 if (!cic->last_request_pos)
1936 sdist = 0;
Tejun Heo83096eb2009-05-07 22:24:39 +09001937 else if (cic->last_request_pos < blk_rq_pos(rq))
1938 sdist = blk_rq_pos(rq) - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001939 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001940 sdist = cic->last_request_pos - blk_rq_pos(rq);
Jens Axboe206dc692006-03-28 13:03:44 +02001941
1942 /*
1943 * Don't allow the seek distance to get too large from the
1944 * odd fragment, pagein, etc
1945 */
1946 if (cic->seek_samples <= 60) /* second&third seek */
1947 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1948 else
1949 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1950
1951 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1952 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1953 total = cic->seek_total + (cic->seek_samples/2);
1954 do_div(total, cic->seek_samples);
1955 cic->seek_mean = (sector_t)total;
1956}
Jens Axboe22e2c502005-06-27 10:55:12 +02001957
1958/*
1959 * Disable idle window if the process thinks too long or seeks so much that
1960 * it doesn't matter
1961 */
1962static void
1963cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1964 struct cfq_io_context *cic)
1965{
Jens Axboe7b679132008-05-30 12:23:07 +02001966 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02001967
Jens Axboe08717142008-01-28 11:38:15 +01001968 /*
1969 * Don't idle for async or idle io prio class
1970 */
1971 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02001972 return;
1973
Jens Axboec265a7f2008-06-26 13:49:33 +02001974 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001975
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001976 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Jens Axboe963b72f2009-10-03 19:42:18 +02001977 (!cfqd->cfq_latency && cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001978 enable_idle = 0;
1979 else if (sample_valid(cic->ttime_samples)) {
1980 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1981 enable_idle = 0;
1982 else
1983 enable_idle = 1;
1984 }
1985
Jens Axboe7b679132008-05-30 12:23:07 +02001986 if (old_idle != enable_idle) {
1987 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1988 if (enable_idle)
1989 cfq_mark_cfqq_idle_window(cfqq);
1990 else
1991 cfq_clear_cfqq_idle_window(cfqq);
1992 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001993}
1994
Jens Axboe22e2c502005-06-27 10:55:12 +02001995/*
1996 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1997 * no or if we aren't sure, a 1 will cause a preempt.
1998 */
1999static int
2000cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02002001 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002002{
Jens Axboe6d048f52007-04-25 12:44:27 +02002003 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002004
Jens Axboe6d048f52007-04-25 12:44:27 +02002005 cfqq = cfqd->active_queue;
2006 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002007 return 0;
2008
Jens Axboe6d048f52007-04-25 12:44:27 +02002009 if (cfq_slice_used(cfqq))
2010 return 1;
2011
2012 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02002013 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02002014
2015 if (cfq_class_idle(cfqq))
2016 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002017
Jens Axboe22e2c502005-06-27 10:55:12 +02002018 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02002019 * if the new request is sync, but the currently running queue is
2020 * not, let the sync request have priority.
2021 */
Jens Axboe5e705372006-07-13 12:39:25 +02002022 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002023 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002024
Jens Axboe374f84a2006-07-23 01:42:19 +02002025 /*
2026 * So both queues are sync. Let the new request get disk time if
2027 * it's a metadata request and the current queue is doing regular IO.
2028 */
2029 if (rq_is_meta(rq) && !cfqq->meta_pending)
2030 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002031
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002032 /*
2033 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2034 */
2035 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2036 return 1;
2037
Jens Axboe1e3335d2007-02-14 19:59:49 +01002038 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2039 return 0;
2040
2041 /*
2042 * if this request is as-good as one we would expect from the
2043 * current cfqq, let it preempt
2044 */
Jens Axboe6d048f52007-04-25 12:44:27 +02002045 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01002046 return 1;
2047
Jens Axboe22e2c502005-06-27 10:55:12 +02002048 return 0;
2049}
2050
2051/*
2052 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2053 * let it have half of its nominal slice.
2054 */
2055static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2056{
Jens Axboe7b679132008-05-30 12:23:07 +02002057 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02002058 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02002059
Jens Axboebf572252006-07-19 20:29:12 +02002060 /*
2061 * Put the new queue at the front of the of the current list,
2062 * so we know that it will be selected next.
2063 */
2064 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02002065
2066 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02002067
Jens Axboe44f7c162007-01-19 11:51:58 +11002068 cfqq->slice_end = 0;
2069 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002070}
2071
2072/*
Jens Axboe5e705372006-07-13 12:39:25 +02002073 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02002074 * something we should do about it
2075 */
2076static void
Jens Axboe5e705372006-07-13 12:39:25 +02002077cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2078 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002079{
Jens Axboe5e705372006-07-13 12:39:25 +02002080 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02002081
Aaron Carroll45333d52008-08-26 15:52:36 +02002082 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02002083 if (rq_is_meta(rq))
2084 cfqq->meta_pending++;
2085
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002086 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02002087 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002088 cfq_update_idle_window(cfqd, cfqq, cic);
2089
Tejun Heo83096eb2009-05-07 22:24:39 +09002090 cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002091
2092 if (cfqq == cfqd->active_queue) {
2093 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002094 * Remember that we saw a request from this process, but
2095 * don't start queuing just yet. Otherwise we risk seeing lots
2096 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02002097 * and merging. If the request is already larger than a single
2098 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02002099 * merging is already done. Ditto for a busy system that
2100 * has other work pending, don't risk delaying until the
2101 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02002102 */
Jens Axboed6ceb252009-04-14 14:18:16 +02002103 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02002104 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2105 cfqd->busy_queues > 1) {
Jens Axboed6ceb252009-04-14 14:18:16 +02002106 del_timer(&cfqd->idle_slice_timer);
Tejun Heoa7f55792009-04-23 11:05:17 +09002107 __blk_run_queue(cfqd->queue);
Jens Axboed6ceb252009-04-14 14:18:16 +02002108 }
Jens Axboeb0291952009-04-07 11:38:31 +02002109 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboed6ceb252009-04-14 14:18:16 +02002110 }
Jens Axboe5e705372006-07-13 12:39:25 +02002111 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002112 /*
2113 * not the active queue - expire current slice if it is
2114 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002115 * has some old slice time left and is of higher priority or
2116 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02002117 */
2118 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09002119 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002120 }
2121}
2122
Jens Axboe165125e2007-07-24 09:28:11 +02002123static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002124{
Jens Axboeb4878f22005-10-20 16:42:29 +02002125 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002126 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002127
Jens Axboe7b679132008-05-30 12:23:07 +02002128 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01002129 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130
Jens Axboe5e705372006-07-13 12:39:25 +02002131 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132
Jens Axboe22e2c502005-06-27 10:55:12 +02002133 list_add_tail(&rq->queuelist, &cfqq->fifo);
2134
Jens Axboe5e705372006-07-13 12:39:25 +02002135 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136}
2137
Aaron Carroll45333d52008-08-26 15:52:36 +02002138/*
2139 * Update hw_tag based on peak queue depth over 50 samples under
2140 * sufficient load.
2141 */
2142static void cfq_update_hw_tag(struct cfq_data *cfqd)
2143{
Jens Axboe5ad531d2009-07-03 12:57:48 +02002144 if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2145 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
Aaron Carroll45333d52008-08-26 15:52:36 +02002146
2147 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Jens Axboe5ad531d2009-07-03 12:57:48 +02002148 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02002149 return;
2150
2151 if (cfqd->hw_tag_samples++ < 50)
2152 return;
2153
2154 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2155 cfqd->hw_tag = 1;
2156 else
2157 cfqd->hw_tag = 0;
2158
2159 cfqd->hw_tag_samples = 0;
2160 cfqd->rq_in_driver_peak = 0;
2161}
2162
Jens Axboe165125e2007-07-24 09:28:11 +02002163static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164{
Jens Axboe5e705372006-07-13 12:39:25 +02002165 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002166 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02002167 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002168 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169
Jens Axboeb4878f22005-10-20 16:42:29 +02002170 now = jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02002171 cfq_log_cfqq(cfqd, cfqq, "complete");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172
Aaron Carroll45333d52008-08-26 15:52:36 +02002173 cfq_update_hw_tag(cfqd);
2174
Jens Axboe5ad531d2009-07-03 12:57:48 +02002175 WARN_ON(!cfqd->rq_in_driver[sync]);
Jens Axboe6d048f52007-04-25 12:44:27 +02002176 WARN_ON(!cfqq->dispatched);
Jens Axboe5ad531d2009-07-03 12:57:48 +02002177 cfqd->rq_in_driver[sync]--;
Jens Axboe6d048f52007-04-25 12:44:27 +02002178 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179
Jens Axboe3ed9a292007-04-23 08:33:33 +02002180 if (cfq_cfqq_sync(cfqq))
2181 cfqd->sync_flight--;
2182
Vivek Goyal365722b2009-10-03 15:21:27 +02002183 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02002184 RQ_CIC(rq)->last_end_request = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02002185 cfqd->last_end_sync_rq = now;
2186 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002187
2188 /*
2189 * If this is the active queue, check if it needs to be expired,
2190 * or if we want to idle in case it has no pending requests.
2191 */
2192 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02002193 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2194
Jens Axboe44f7c162007-01-19 11:51:58 +11002195 if (cfq_cfqq_slice_new(cfqq)) {
2196 cfq_set_prio_slice(cfqd, cfqq);
2197 cfq_clear_cfqq_slice_new(cfqq);
2198 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002199 /*
2200 * If there are no requests waiting in this queue, and
2201 * there are other queues ready to issue requests, AND
2202 * those other queues are issuing requests within our
2203 * mean seek distance, give them a chance to run instead
2204 * of idling.
2205 */
Jens Axboe08717142008-01-28 11:38:15 +01002206 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02002207 cfq_slice_expired(cfqd, 1);
Jens Axboea36e71f2009-04-15 12:15:11 +02002208 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2209 sync && !rq_noidle(rq))
Jens Axboe6d048f52007-04-25 12:44:27 +02002210 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002211 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002212
Jens Axboe5ad531d2009-07-03 12:57:48 +02002213 if (!rq_in_driver(cfqd))
Jens Axboe8e296752009-10-03 16:26:03 +02002214 cfq_schedule_dispatch(cfqd, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215}
2216
Jens Axboe22e2c502005-06-27 10:55:12 +02002217/*
2218 * we temporarily boost lower priority queues if they are holding fs exclusive
2219 * resources. they are boosted to normal prio (CLASS_BE/4)
2220 */
2221static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222{
Jens Axboe22e2c502005-06-27 10:55:12 +02002223 if (has_fs_excl()) {
2224 /*
2225 * boost idle prio on transactions that would lock out other
2226 * users of the filesystem
2227 */
2228 if (cfq_class_idle(cfqq))
2229 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2230 if (cfqq->ioprio > IOPRIO_NORM)
2231 cfqq->ioprio = IOPRIO_NORM;
2232 } else {
2233 /*
2234 * check if we need to unboost the queue
2235 */
2236 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2237 cfqq->ioprio_class = cfqq->org_ioprio_class;
2238 if (cfqq->ioprio != cfqq->org_ioprio)
2239 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002241}
2242
Jens Axboe89850f72006-07-22 16:48:31 +02002243static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002244{
Jens Axboe1b379d82009-08-11 08:26:11 +02002245 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002246 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002247 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02002248 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002249
2250 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02002251}
2252
Jens Axboe165125e2007-07-24 09:28:11 +02002253static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02002254{
2255 struct cfq_data *cfqd = q->elevator->elevator_data;
2256 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002257 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02002258 struct cfq_queue *cfqq;
2259
2260 /*
2261 * don't force setup of a queue from here, as a call to may_queue
2262 * does not necessarily imply that a request actually will be queued.
2263 * so just lookup a possibly existing queue, or return 'may queue'
2264 * if that fails
2265 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002266 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002267 if (!cic)
2268 return ELV_MQUEUE_MAY;
2269
Jens Axboeb0b78f82009-04-08 10:56:08 +02002270 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02002271 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002272 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02002273 cfq_prio_boost(cfqq);
2274
Jens Axboe89850f72006-07-22 16:48:31 +02002275 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002276 }
2277
2278 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279}
2280
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281/*
2282 * queue lock held here
2283 */
Jens Axboebb37b942006-12-01 10:42:33 +01002284static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285{
Jens Axboe5e705372006-07-13 12:39:25 +02002286 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287
Jens Axboe5e705372006-07-13 12:39:25 +02002288 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002289 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290
Jens Axboe22e2c502005-06-27 10:55:12 +02002291 BUG_ON(!cfqq->allocated[rw]);
2292 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293
Jens Axboe5e705372006-07-13 12:39:25 +02002294 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02002297 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 cfq_put_queue(cfqq);
2300 }
2301}
2302
2303/*
Jens Axboe22e2c502005-06-27 10:55:12 +02002304 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 */
Jens Axboe22e2c502005-06-27 10:55:12 +02002306static int
Jens Axboe165125e2007-07-24 09:28:11 +02002307cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308{
2309 struct cfq_data *cfqd = q->elevator->elevator_data;
2310 struct cfq_io_context *cic;
2311 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01002312 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002313 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314 unsigned long flags;
2315
2316 might_sleep_if(gfp_mask & __GFP_WAIT);
2317
Jens Axboee2d74ac2006-03-28 08:59:01 +02002318 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02002319
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 spin_lock_irqsave(q->queue_lock, flags);
2321
Jens Axboe22e2c502005-06-27 10:55:12 +02002322 if (!cic)
2323 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324
Vasily Tarasov91fac312007-04-25 12:29:51 +02002325 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02002326 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002327 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002328 cic_set_cfqq(cic, cfqq, is_sync);
2329 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330
2331 cfqq->allocated[rw]++;
Jens Axboe22e2c502005-06-27 10:55:12 +02002332 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02002333
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 spin_unlock_irqrestore(q->queue_lock, flags);
2335
Jens Axboe5e705372006-07-13 12:39:25 +02002336 rq->elevator_private = cic;
2337 rq->elevator_private2 = cfqq;
2338 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02002339
Jens Axboe22e2c502005-06-27 10:55:12 +02002340queue_fail:
2341 if (cic)
2342 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02002343
Jens Axboe8e296752009-10-03 16:26:03 +02002344 cfq_schedule_dispatch(cfqd, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02002346 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 return 1;
2348}
2349
David Howells65f27f32006-11-22 14:55:48 +00002350static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02002351{
David Howells65f27f32006-11-22 14:55:48 +00002352 struct cfq_data *cfqd =
Jens Axboe8e296752009-10-03 16:26:03 +02002353 container_of(work, struct cfq_data, unplug_work.work);
Jens Axboe165125e2007-07-24 09:28:11 +02002354 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002355
Jens Axboe40bb54d2009-04-15 12:11:10 +02002356 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09002357 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02002358 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02002359}
2360
2361/*
2362 * Timer running if the active_queue is currently idling inside its time slice
2363 */
2364static void cfq_idle_slice_timer(unsigned long data)
2365{
2366 struct cfq_data *cfqd = (struct cfq_data *) data;
2367 struct cfq_queue *cfqq;
2368 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002369 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002370
Jens Axboe7b679132008-05-30 12:23:07 +02002371 cfq_log(cfqd, "idle timer fired");
2372
Jens Axboe22e2c502005-06-27 10:55:12 +02002373 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2374
Jens Axboefe094d92008-01-31 13:08:54 +01002375 cfqq = cfqd->active_queue;
2376 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002377 timed_out = 0;
2378
Jens Axboe22e2c502005-06-27 10:55:12 +02002379 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002380 * We saw a request before the queue expired, let it through
2381 */
2382 if (cfq_cfqq_must_dispatch(cfqq))
2383 goto out_kick;
2384
2385 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002386 * expired
2387 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002388 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002389 goto expire;
2390
2391 /*
2392 * only expire and reinvoke request handler, if there are
2393 * other queues with pending requests
2394 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002395 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002396 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002397
2398 /*
2399 * not expired and it has a request pending, let it dispatch
2400 */
Jens Axboe75e50982009-04-07 08:56:14 +02002401 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002402 goto out_kick;
Jens Axboe22e2c502005-06-27 10:55:12 +02002403 }
2404expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002405 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002406out_kick:
Jens Axboe8e296752009-10-03 16:26:03 +02002407 cfq_schedule_dispatch(cfqd, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002408out_cont:
2409 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2410}
2411
Jens Axboe3b181522005-06-27 10:56:24 +02002412static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2413{
2414 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe8e296752009-10-03 16:26:03 +02002415 cancel_delayed_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002416}
Jens Axboe22e2c502005-06-27 10:55:12 +02002417
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002418static void cfq_put_async_queues(struct cfq_data *cfqd)
2419{
2420 int i;
2421
2422 for (i = 0; i < IOPRIO_BE_NR; i++) {
2423 if (cfqd->async_cfqq[0][i])
2424 cfq_put_queue(cfqd->async_cfqq[0][i]);
2425 if (cfqd->async_cfqq[1][i])
2426 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002427 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002428
2429 if (cfqd->async_idle_cfqq)
2430 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002431}
2432
Jens Axboeb374d182008-10-31 10:05:07 +01002433static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434{
Jens Axboe22e2c502005-06-27 10:55:12 +02002435 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002436 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002437
Jens Axboe3b181522005-06-27 10:56:24 +02002438 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002439
Al Virod9ff4182006-03-18 13:51:22 -05002440 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002441
Al Virod9ff4182006-03-18 13:51:22 -05002442 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002443 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002444
2445 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002446 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2447 struct cfq_io_context,
2448 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002449
2450 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002451 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002452
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002453 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002454
Al Virod9ff4182006-03-18 13:51:22 -05002455 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002456
2457 cfq_shutdown_timer_wq(cfqd);
2458
Al Viroa90d7422006-03-18 12:05:37 -05002459 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460}
2461
Jens Axboe165125e2007-07-24 09:28:11 +02002462static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463{
2464 struct cfq_data *cfqd;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002465 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466
Christoph Lameter94f60302007-07-17 04:03:29 -07002467 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002469 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470
Jens Axboecc09e292007-04-26 12:53:50 +02002471 cfqd->service_tree = CFQ_RB_ROOT;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002472
2473 /*
2474 * Not strictly needed (since RB_ROOT just clears the node and we
2475 * zeroed cfqd on alloc), but better be safe in case someone decides
2476 * to add magic to the rb code
2477 */
2478 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2479 cfqd->prio_trees[i] = RB_ROOT;
2480
Jens Axboe6118b702009-06-30 09:34:12 +02002481 /*
2482 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2483 * Grab a permanent reference to it, so that the normal code flow
2484 * will not attempt to free it.
2485 */
2486 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2487 atomic_inc(&cfqd->oom_cfqq.ref);
2488
Al Virod9ff4182006-03-18 13:51:22 -05002489 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492
Jens Axboe22e2c502005-06-27 10:55:12 +02002493 init_timer(&cfqd->idle_slice_timer);
2494 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2495 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2496
Jens Axboe8e296752009-10-03 16:26:03 +02002497 INIT_DELAYED_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002498
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002500 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2501 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 cfqd->cfq_back_max = cfq_back_max;
2503 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002504 cfqd->cfq_slice[0] = cfq_slice_async;
2505 cfqd->cfq_slice[1] = cfq_slice_sync;
2506 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2507 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02002508 cfqd->cfq_latency = 1;
Aaron Carroll45333d52008-08-26 15:52:36 +02002509 cfqd->hw_tag = 1;
Vivek Goyal365722b2009-10-03 15:21:27 +02002510 cfqd->last_end_sync_rq = jiffies;
Jens Axboebc1c1162006-06-08 08:49:06 +02002511 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512}
2513
2514static void cfq_slab_kill(void)
2515{
Jens Axboed6de8be2008-05-28 14:46:59 +02002516 /*
2517 * Caller already ensured that pending RCU callbacks are completed,
2518 * so we should have no busy allocations at this point.
2519 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520 if (cfq_pool)
2521 kmem_cache_destroy(cfq_pool);
2522 if (cfq_ioc_pool)
2523 kmem_cache_destroy(cfq_ioc_pool);
2524}
2525
2526static int __init cfq_slab_setup(void)
2527{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002528 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 if (!cfq_pool)
2530 goto fail;
2531
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002532 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533 if (!cfq_ioc_pool)
2534 goto fail;
2535
2536 return 0;
2537fail:
2538 cfq_slab_kill();
2539 return -ENOMEM;
2540}
2541
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542/*
2543 * sysfs parts below -->
2544 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545static ssize_t
2546cfq_var_show(unsigned int var, char *page)
2547{
2548 return sprintf(page, "%d\n", var);
2549}
2550
2551static ssize_t
2552cfq_var_store(unsigned int *var, const char *page, size_t count)
2553{
2554 char *p = (char *) page;
2555
2556 *var = simple_strtoul(p, &p, 10);
2557 return count;
2558}
2559
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002561static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002563 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 unsigned int __data = __VAR; \
2565 if (__CONV) \
2566 __data = jiffies_to_msecs(__data); \
2567 return cfq_var_show(__data, (page)); \
2568}
2569SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002570SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2571SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002572SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2573SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002574SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2575SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2576SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2577SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02002578SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579#undef SHOW_FUNCTION
2580
2581#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002582static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002584 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585 unsigned int __data; \
2586 int ret = cfq_var_store(&__data, (page), count); \
2587 if (__data < (MIN)) \
2588 __data = (MIN); \
2589 else if (__data > (MAX)) \
2590 __data = (MAX); \
2591 if (__CONV) \
2592 *(__PTR) = msecs_to_jiffies(__data); \
2593 else \
2594 *(__PTR) = __data; \
2595 return ret; \
2596}
2597STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002598STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2599 UINT_MAX, 1);
2600STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2601 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002602STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002603STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2604 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002605STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2606STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2607STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01002608STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2609 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02002610STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611#undef STORE_FUNCTION
2612
Al Viroe572ec72006-03-18 22:27:18 -05002613#define CFQ_ATTR(name) \
2614 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002615
Al Viroe572ec72006-03-18 22:27:18 -05002616static struct elv_fs_entry cfq_attrs[] = {
2617 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002618 CFQ_ATTR(fifo_expire_sync),
2619 CFQ_ATTR(fifo_expire_async),
2620 CFQ_ATTR(back_seek_max),
2621 CFQ_ATTR(back_seek_penalty),
2622 CFQ_ATTR(slice_sync),
2623 CFQ_ATTR(slice_async),
2624 CFQ_ATTR(slice_async_rq),
2625 CFQ_ATTR(slice_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02002626 CFQ_ATTR(low_latency),
Al Viroe572ec72006-03-18 22:27:18 -05002627 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628};
2629
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630static struct elevator_type iosched_cfq = {
2631 .ops = {
2632 .elevator_merge_fn = cfq_merge,
2633 .elevator_merged_fn = cfq_merged_request,
2634 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002635 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002636 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002638 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639 .elevator_deactivate_req_fn = cfq_deactivate_request,
2640 .elevator_queue_empty_fn = cfq_queue_empty,
2641 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002642 .elevator_former_req_fn = elv_rb_former_request,
2643 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 .elevator_set_req_fn = cfq_set_request,
2645 .elevator_put_req_fn = cfq_put_request,
2646 .elevator_may_queue_fn = cfq_may_queue,
2647 .elevator_init_fn = cfq_init_queue,
2648 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002649 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650 },
Al Viro3d1ab402006-03-18 18:35:43 -05002651 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652 .elevator_name = "cfq",
2653 .elevator_owner = THIS_MODULE,
2654};
2655
2656static int __init cfq_init(void)
2657{
Jens Axboe22e2c502005-06-27 10:55:12 +02002658 /*
2659 * could be 0 on HZ < 1000 setups
2660 */
2661 if (!cfq_slice_async)
2662 cfq_slice_async = 1;
2663 if (!cfq_slice_idle)
2664 cfq_slice_idle = 1;
2665
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666 if (cfq_slab_setup())
2667 return -ENOMEM;
2668
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002669 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002671 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672}
2673
2674static void __exit cfq_exit(void)
2675{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002676 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002678 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002679 /* ioc_gone's update must be visible before reading ioc_count */
2680 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02002681
2682 /*
2683 * this also protects us from entering cfq_slab_kill() with
2684 * pending RCU callbacks
2685 */
Tejun Heo245b2e72009-06-24 15:13:48 +09002686 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002687 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08002688 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689}
2690
2691module_init(cfq_init);
2692module_exit(cfq_exit);
2693
2694MODULE_AUTHOR("Jens Axboe");
2695MODULE_LICENSE("GPL");
2696MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");