blob: 5802e322b7ad31cb7c0e2ea9c3b39ada3b5baa65 [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 Axboe23e018a2009-10-05 08:52:35 +0200153 struct work_struct 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 Axboeb0b8d7492007-01-19 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 Axboeb0b8d7492007-01-19 11:35:30 +1100192 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 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 *);
Jens Axboea6151c32009-10-07 20:02:57 +0200233static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
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,
Jens Axboea6151c32009-10-07 20:02:57 +0200244 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200245{
Jens Axboea6151c32009-10-07 20:02:57 +0200246 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200247}
248
249static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200250 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200251{
Jens Axboea6151c32009-10-07 20:02:57 +0200252 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200253}
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 */
Jens Axboea6151c32009-10-07 20:02:57 +0200259static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200260{
Jens Axboea6151c32009-10-07 20:02:57 +0200261 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200262}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700265 * scheduler run of queue, if there are requests pending and no one in the
266 * driver that will restart queueing
267 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200268static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700269{
Jens Axboe7b679132008-05-30 12:23:07 +0200270 if (cfqd->busy_queues) {
271 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200272 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200273 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700274}
275
Jens Axboe165125e2007-07-24 09:28:11 +0200276static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700277{
278 struct cfq_data *cfqd = q->elevator->elevator_data;
279
Jens Axboeb4878f22005-10-20 16:42:29 +0200280 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700281}
282
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100284 * Scale schedule slice based on io priority. Use the sync time slice only
285 * if a queue is marked sync and has sync io queued. A sync queue with async
286 * io only, should not get full sync slice length.
287 */
Jens Axboea6151c32009-10-07 20:02:57 +0200288static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200289 unsigned short prio)
290{
291 const int base_slice = cfqd->cfq_slice[sync];
292
293 WARN_ON(prio >= IOPRIO_BE_NR);
294
295 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
296}
297
Jens Axboe44f7c162007-01-19 11:51:58 +1100298static inline int
299cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
300{
Jens Axboed9e76202007-04-20 14:27:50 +0200301 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100302}
303
304static inline void
305cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
306{
307 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200308 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100309}
310
311/*
312 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
313 * isn't valid until the first request from the dispatch is activated
314 * and the slice time set.
315 */
Jens Axboea6151c32009-10-07 20:02:57 +0200316static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100317{
318 if (cfq_cfqq_slice_new(cfqq))
319 return 0;
320 if (time_before(jiffies, cfqq->slice_end))
321 return 0;
322
323 return 1;
324}
325
326/*
Jens Axboe5e705372006-07-13 12:39:25 +0200327 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200329 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330 */
Jens Axboe5e705372006-07-13 12:39:25 +0200331static struct request *
332cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333{
334 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200336#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
337#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
338 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339
Jens Axboe5e705372006-07-13 12:39:25 +0200340 if (rq1 == NULL || rq1 == rq2)
341 return rq2;
342 if (rq2 == NULL)
343 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200344
Jens Axboe5e705372006-07-13 12:39:25 +0200345 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
346 return rq1;
347 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
348 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200349 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
350 return rq1;
351 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
352 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353
Tejun Heo83096eb2009-05-07 22:24:39 +0900354 s1 = blk_rq_pos(rq1);
355 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356
Jens Axboe6d048f52007-04-25 12:44:27 +0200357 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 /*
360 * by definition, 1KiB is 2 sectors
361 */
362 back_max = cfqd->cfq_back_max * 2;
363
364 /*
365 * Strict one way elevator _except_ in the case where we allow
366 * short backward seeks which are biased as twice the cost of a
367 * similar forward seek.
368 */
369 if (s1 >= last)
370 d1 = s1 - last;
371 else if (s1 + back_max >= last)
372 d1 = (last - s1) * cfqd->cfq_back_penalty;
373 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200374 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375
376 if (s2 >= last)
377 d2 = s2 - last;
378 else if (s2 + back_max >= last)
379 d2 = (last - s2) * cfqd->cfq_back_penalty;
380 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200381 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382
383 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384
Andreas Mohre8a99052006-03-28 08:59:49 +0200385 /*
386 * By doing switch() on the bit mask "wrap" we avoid having to
387 * check two variables for all permutations: --> faster!
388 */
389 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200390 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200391 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200392 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200393 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200394 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200395 else {
396 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200397 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200398 else
Jens Axboe5e705372006-07-13 12:39:25 +0200399 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200400 }
401
402 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200403 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200404 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200405 return rq2;
406 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200407 default:
408 /*
409 * Since both rqs are wrapped,
410 * start with the one that's further behind head
411 * (--> only *one* back seek required),
412 * since back seek takes more time than forward.
413 */
414 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200415 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 else
Jens Axboe5e705372006-07-13 12:39:25 +0200417 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 }
419}
420
Jens Axboe498d3aa22007-04-26 12:54:48 +0200421/*
422 * The below is leftmost cache rbtree addon
423 */
Jens Axboe08717142008-01-28 11:38:15 +0100424static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200425{
426 if (!root->left)
427 root->left = rb_first(&root->rb);
428
Jens Axboe08717142008-01-28 11:38:15 +0100429 if (root->left)
430 return rb_entry(root->left, struct cfq_queue, rb_node);
431
432 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200433}
434
Jens Axboea36e71f2009-04-15 12:15:11 +0200435static void rb_erase_init(struct rb_node *n, struct rb_root *root)
436{
437 rb_erase(n, root);
438 RB_CLEAR_NODE(n);
439}
440
Jens Axboecc09e292007-04-26 12:53:50 +0200441static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
442{
443 if (root->left == n)
444 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200445 rb_erase_init(n, &root->rb);
Jens Axboecc09e292007-04-26 12:53:50 +0200446}
447
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448/*
449 * would be nice to take fifo expire time into account as well
450 */
Jens Axboe5e705372006-07-13 12:39:25 +0200451static struct request *
452cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
453 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454{
Jens Axboe21183b02006-07-13 12:33:14 +0200455 struct rb_node *rbnext = rb_next(&last->rb_node);
456 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200457 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458
Jens Axboe21183b02006-07-13 12:33:14 +0200459 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460
461 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200462 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200463
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200465 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200466 else {
467 rbnext = rb_first(&cfqq->sort_list);
468 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200469 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200470 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471
Jens Axboe21183b02006-07-13 12:33:14 +0200472 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473}
474
Jens Axboed9e76202007-04-20 14:27:50 +0200475static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
476 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477{
Jens Axboed9e76202007-04-20 14:27:50 +0200478 /*
479 * just an approximation, should be ok.
480 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200481 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
482 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200483}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484
Jens Axboe498d3aa22007-04-26 12:54:48 +0200485/*
486 * The cfqd->service_tree holds all pending cfq_queue's that have
487 * requests waiting to be processed. It is sorted in the order that
488 * we will service the queues.
489 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200490static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +0200491 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200492{
Jens Axboe08717142008-01-28 11:38:15 +0100493 struct rb_node **p, *parent;
494 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +0200495 unsigned long rb_key;
Jens Axboe498d3aa22007-04-26 12:54:48 +0200496 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200497
Jens Axboe08717142008-01-28 11:38:15 +0100498 if (cfq_class_idle(cfqq)) {
499 rb_key = CFQ_IDLE_DELAY;
500 parent = rb_last(&cfqd->service_tree.rb);
501 if (parent && parent != &cfqq->rb_node) {
502 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
503 rb_key += __cfqq->rb_key;
504 } else
505 rb_key += jiffies;
506 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +0200507 /*
508 * Get our rb key offset. Subtract any residual slice
509 * value carried from last service. A negative resid
510 * count indicates slice overrun, and this should position
511 * the next service time further away in the tree.
512 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200513 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +0200514 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +0200515 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +0200516 } else {
517 rb_key = -HZ;
518 __cfqq = cfq_rb_first(&cfqd->service_tree);
519 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
520 }
Jens Axboed9e76202007-04-20 14:27:50 +0200521
522 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100523 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200524 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100525 */
Jens Axboed9e76202007-04-20 14:27:50 +0200526 if (rb_key == cfqq->rb_key)
527 return;
Jens Axboe53b037442006-07-28 09:48:51 +0200528
Jens Axboecc09e292007-04-26 12:53:50 +0200529 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200530 }
Jens Axboed9e76202007-04-20 14:27:50 +0200531
Jens Axboe498d3aa22007-04-26 12:54:48 +0200532 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +0100533 parent = NULL;
534 p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200535 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +0200536 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200537
Jens Axboed9e76202007-04-20 14:27:50 +0200538 parent = *p;
539 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
540
Jens Axboe0c534e02007-04-18 20:01:57 +0200541 /*
542 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200543 * preference to them. IDLE queues goes to the back.
544 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200545 */
546 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200547 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200548 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200549 n = &(*p)->rb_right;
550 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
551 n = &(*p)->rb_left;
552 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
553 n = &(*p)->rb_right;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +0200554 else if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +0200555 n = &(*p)->rb_left;
556 else
557 n = &(*p)->rb_right;
558
559 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200560 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200561
562 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200563 }
564
Jens Axboecc09e292007-04-26 12:53:50 +0200565 if (left)
566 cfqd->service_tree.left = &cfqq->rb_node;
567
Jens Axboed9e76202007-04-20 14:27:50 +0200568 cfqq->rb_key = rb_key;
569 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200570 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571}
572
Jens Axboea36e71f2009-04-15 12:15:11 +0200573static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200574cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
575 sector_t sector, struct rb_node **ret_parent,
576 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +0200577{
Jens Axboea36e71f2009-04-15 12:15:11 +0200578 struct rb_node **p, *parent;
579 struct cfq_queue *cfqq = NULL;
580
581 parent = NULL;
582 p = &root->rb_node;
583 while (*p) {
584 struct rb_node **n;
585
586 parent = *p;
587 cfqq = rb_entry(parent, struct cfq_queue, p_node);
588
589 /*
590 * Sort strictly based on sector. Smallest to the left,
591 * largest to the right.
592 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900593 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +0200594 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900595 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +0200596 n = &(*p)->rb_left;
597 else
598 break;
599 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200600 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200601 }
602
603 *ret_parent = parent;
604 if (rb_link)
605 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200606 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +0200607}
608
609static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
610{
Jens Axboea36e71f2009-04-15 12:15:11 +0200611 struct rb_node **p, *parent;
612 struct cfq_queue *__cfqq;
613
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200614 if (cfqq->p_root) {
615 rb_erase(&cfqq->p_node, cfqq->p_root);
616 cfqq->p_root = NULL;
617 }
Jens Axboea36e71f2009-04-15 12:15:11 +0200618
619 if (cfq_class_idle(cfqq))
620 return;
621 if (!cfqq->next_rq)
622 return;
623
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200624 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900625 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
626 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200627 if (!__cfqq) {
628 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200629 rb_insert_color(&cfqq->p_node, cfqq->p_root);
630 } else
631 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200632}
633
Jens Axboe498d3aa22007-04-26 12:54:48 +0200634/*
635 * Update cfqq's position in the service tree.
636 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200637static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200638{
Jens Axboe6d048f52007-04-25 12:44:27 +0200639 /*
640 * Resorting requires the cfqq to be on the RR list already.
641 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200642 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200643 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +0200644 cfq_prio_tree_add(cfqd, cfqq);
645 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200646}
647
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648/*
649 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200650 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 */
Jens Axboefebffd62008-01-28 13:19:43 +0100652static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653{
Jens Axboe7b679132008-05-30 12:23:07 +0200654 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200655 BUG_ON(cfq_cfqq_on_rr(cfqq));
656 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 cfqd->busy_queues++;
658
Jens Axboeedd75ff2007-04-19 12:03:34 +0200659 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660}
661
Jens Axboe498d3aa22007-04-26 12:54:48 +0200662/*
663 * Called when the cfqq no longer has requests pending, remove it from
664 * the service tree.
665 */
Jens Axboefebffd62008-01-28 13:19:43 +0100666static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667{
Jens Axboe7b679132008-05-30 12:23:07 +0200668 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200669 BUG_ON(!cfq_cfqq_on_rr(cfqq));
670 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671
Jens Axboecc09e292007-04-26 12:53:50 +0200672 if (!RB_EMPTY_NODE(&cfqq->rb_node))
673 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200674 if (cfqq->p_root) {
675 rb_erase(&cfqq->p_node, cfqq->p_root);
676 cfqq->p_root = NULL;
677 }
Jens Axboed9e76202007-04-20 14:27:50 +0200678
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 BUG_ON(!cfqd->busy_queues);
680 cfqd->busy_queues--;
681}
682
683/*
684 * rb tree support functions
685 */
Jens Axboefebffd62008-01-28 13:19:43 +0100686static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687{
Jens Axboe5e705372006-07-13 12:39:25 +0200688 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200689 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200690 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691
Jens Axboeb4878f22005-10-20 16:42:29 +0200692 BUG_ON(!cfqq->queued[sync]);
693 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694
Jens Axboe5e705372006-07-13 12:39:25 +0200695 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696
Jens Axboedd67d052006-06-21 09:36:18 +0200697 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200698 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699}
700
Jens Axboe5e705372006-07-13 12:39:25 +0200701static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702{
Jens Axboe5e705372006-07-13 12:39:25 +0200703 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +0200705 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706
Jens Axboe5380a102006-07-13 12:37:56 +0200707 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708
709 /*
710 * looks a little odd, but the first insert might return an alias.
711 * if that happens, put the alias on the dispatch list
712 */
Jens Axboe21183b02006-07-13 12:33:14 +0200713 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200714 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100715
716 if (!cfq_cfqq_on_rr(cfqq))
717 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200718
719 /*
720 * check if this request is a better next-serve candidate
721 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200722 prev = cfqq->next_rq;
Jens Axboe5044eed2007-04-25 11:53:48 +0200723 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
Jens Axboea36e71f2009-04-15 12:15:11 +0200724
725 /*
726 * adjust priority tree position, if ->next_rq changes
727 */
728 if (prev != cfqq->next_rq)
729 cfq_prio_tree_add(cfqd, cfqq);
730
Jens Axboe5044eed2007-04-25 11:53:48 +0200731 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732}
733
Jens Axboefebffd62008-01-28 13:19:43 +0100734static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735{
Jens Axboe5380a102006-07-13 12:37:56 +0200736 elv_rb_del(&cfqq->sort_list, rq);
737 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200738 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739}
740
Jens Axboe206dc692006-03-28 13:03:44 +0200741static struct request *
742cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743{
Jens Axboe206dc692006-03-28 13:03:44 +0200744 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200745 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200746 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747
Jens Axboe4ac845a2008-01-24 08:44:49 +0100748 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200749 if (!cic)
750 return NULL;
751
752 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200753 if (cfqq) {
754 sector_t sector = bio->bi_sector + bio_sectors(bio);
755
Jens Axboe21183b02006-07-13 12:33:14 +0200756 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200757 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 return NULL;
760}
761
Jens Axboe165125e2007-07-24 09:28:11 +0200762static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200763{
764 struct cfq_data *cfqd = q->elevator->elevator_data;
765
Jens Axboe5ad531d2009-07-03 12:57:48 +0200766 cfqd->rq_in_driver[rq_is_sync(rq)]++;
Jens Axboe7b679132008-05-30 12:23:07 +0200767 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Jens Axboe5ad531d2009-07-03 12:57:48 +0200768 rq_in_driver(cfqd));
Jens Axboe25776e32006-06-01 10:12:26 +0200769
Tejun Heo5b936292009-05-07 22:24:38 +0900770 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200771}
772
Jens Axboe165125e2007-07-24 09:28:11 +0200773static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774{
Jens Axboe22e2c502005-06-27 10:55:12 +0200775 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5ad531d2009-07-03 12:57:48 +0200776 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777
Jens Axboe5ad531d2009-07-03 12:57:48 +0200778 WARN_ON(!cfqd->rq_in_driver[sync]);
779 cfqd->rq_in_driver[sync]--;
Jens Axboe7b679132008-05-30 12:23:07 +0200780 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Jens Axboe5ad531d2009-07-03 12:57:48 +0200781 rq_in_driver(cfqd));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782}
783
Jens Axboeb4878f22005-10-20 16:42:29 +0200784static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785{
Jens Axboe5e705372006-07-13 12:39:25 +0200786 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200787
Jens Axboe5e705372006-07-13 12:39:25 +0200788 if (cfqq->next_rq == rq)
789 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790
Jens Axboeb4878f22005-10-20 16:42:29 +0200791 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200792 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200793
Aaron Carroll45333d52008-08-26 15:52:36 +0200794 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +0200795 if (rq_is_meta(rq)) {
796 WARN_ON(!cfqq->meta_pending);
797 cfqq->meta_pending--;
798 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799}
800
Jens Axboe165125e2007-07-24 09:28:11 +0200801static int cfq_merge(struct request_queue *q, struct request **req,
802 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803{
804 struct cfq_data *cfqd = q->elevator->elevator_data;
805 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806
Jens Axboe206dc692006-03-28 13:03:44 +0200807 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200808 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200809 *req = __rq;
810 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 }
812
813 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814}
815
Jens Axboe165125e2007-07-24 09:28:11 +0200816static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200817 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818{
Jens Axboe21183b02006-07-13 12:33:14 +0200819 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200820 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821
Jens Axboe5e705372006-07-13 12:39:25 +0200822 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824}
825
826static void
Jens Axboe165125e2007-07-24 09:28:11 +0200827cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 struct request *next)
829{
Jens Axboe22e2c502005-06-27 10:55:12 +0200830 /*
831 * reposition in fifo if next is older than rq
832 */
833 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +0200834 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +0200835 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +0200836 rq_set_fifo_time(rq, rq_fifo_time(next));
837 }
Jens Axboe22e2c502005-06-27 10:55:12 +0200838
Jens Axboeb4878f22005-10-20 16:42:29 +0200839 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200840}
841
Jens Axboe165125e2007-07-24 09:28:11 +0200842static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100843 struct bio *bio)
844{
845 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200846 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100847 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100848
849 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100850 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100851 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200852 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +0200853 return false;
Jens Axboeda775262006-12-20 11:04:12 +0100854
855 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100856 * Lookup the cfqq that this bio will be queued with. Allow
857 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100858 */
Jens Axboe4ac845a2008-01-24 08:44:49 +0100859 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200860 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +0200861 return false;
Jens Axboe719d3402006-12-22 09:38:53 +0100862
Vasily Tarasov91fac312007-04-25 12:29:51 +0200863 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +0200864 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +0100865}
866
Jens Axboefebffd62008-01-28 13:19:43 +0100867static void __cfq_set_active_queue(struct cfq_data *cfqd,
868 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +0200869{
870 if (cfqq) {
Jens Axboe7b679132008-05-30 12:23:07 +0200871 cfq_log_cfqq(cfqd, cfqq, "set_active");
Jens Axboe22e2c502005-06-27 10:55:12 +0200872 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +0200873 cfqq->slice_dispatch = 0;
874
Jens Axboe2f5cb732009-04-07 08:51:19 +0200875 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +0200876 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200877 cfq_clear_cfqq_must_alloc_slice(cfqq);
878 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100879 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +0200880
881 del_timer(&cfqd->idle_slice_timer);
Jens Axboe22e2c502005-06-27 10:55:12 +0200882 }
883
884 cfqd->active_queue = cfqq;
885}
886
887/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100888 * current cfqq expired its slice (or was too idle), select new one
889 */
890static void
891__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +0200892 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100893{
Jens Axboe7b679132008-05-30 12:23:07 +0200894 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
895
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100896 if (cfq_cfqq_wait_request(cfqq))
897 del_timer(&cfqd->idle_slice_timer);
898
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100899 cfq_clear_cfqq_wait_request(cfqq);
900
901 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200902 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100903 */
Jens Axboe7b679132008-05-30 12:23:07 +0200904 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +1100905 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200906 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
907 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100908
Jens Axboeedd75ff2007-04-19 12:03:34 +0200909 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100910
911 if (cfqq == cfqd->active_queue)
912 cfqd->active_queue = NULL;
913
914 if (cfqd->active_cic) {
915 put_io_context(cfqd->active_cic->ioc);
916 cfqd->active_cic = NULL;
917 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100918}
919
Jens Axboea6151c32009-10-07 20:02:57 +0200920static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100921{
922 struct cfq_queue *cfqq = cfqd->active_queue;
923
924 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200925 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100926}
927
Jens Axboe498d3aa22007-04-26 12:54:48 +0200928/*
929 * Get next queue for service. Unless we have a queue preemption,
930 * we'll simply select the first cfqq in the service tree.
931 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200932static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200933{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200934 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
935 return NULL;
936
Jens Axboe08717142008-01-28 11:38:15 +0100937 return cfq_rb_first(&cfqd->service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +0200938}
939
Jens Axboe498d3aa22007-04-26 12:54:48 +0200940/*
941 * Get and set a new active queue for service.
942 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200943static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
944 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200945{
Jens Axboea36e71f2009-04-15 12:15:11 +0200946 if (!cfqq) {
947 cfqq = cfq_get_next_queue(cfqd);
948 if (cfqq)
949 cfq_clear_cfqq_coop(cfqq);
950 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200951
Jens Axboe22e2c502005-06-27 10:55:12 +0200952 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200953 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200954}
955
Jens Axboed9e76202007-04-20 14:27:50 +0200956static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
957 struct request *rq)
958{
Tejun Heo83096eb2009-05-07 22:24:39 +0900959 if (blk_rq_pos(rq) >= cfqd->last_position)
960 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +0200961 else
Tejun Heo83096eb2009-05-07 22:24:39 +0900962 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +0200963}
964
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200965#define CIC_SEEK_THR 8 * 1024
966#define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
967
Jens Axboe6d048f52007-04-25 12:44:27 +0200968static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
969{
970 struct cfq_io_context *cic = cfqd->active_cic;
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200971 sector_t sdist = cic->seek_mean;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200972
Jens Axboe6d048f52007-04-25 12:44:27 +0200973 if (!sample_valid(cic->seek_samples))
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200974 sdist = CIC_SEEK_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +0200975
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200976 return cfq_dist_from_last(cfqd, rq) <= sdist;
Jens Axboe6d048f52007-04-25 12:44:27 +0200977}
978
Jens Axboea36e71f2009-04-15 12:15:11 +0200979static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
980 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200981{
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200982 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +0200983 struct rb_node *parent, *node;
984 struct cfq_queue *__cfqq;
985 sector_t sector = cfqd->last_position;
986
987 if (RB_EMPTY_ROOT(root))
988 return NULL;
989
990 /*
991 * First, if we find a request starting at the end of the last
992 * request, choose it.
993 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200994 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +0200995 if (__cfqq)
996 return __cfqq;
997
998 /*
999 * If the exact sector wasn't found, the parent of the NULL leaf
1000 * will contain the closest sector.
1001 */
1002 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1003 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1004 return __cfqq;
1005
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001006 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001007 node = rb_next(&__cfqq->p_node);
1008 else
1009 node = rb_prev(&__cfqq->p_node);
1010 if (!node)
1011 return NULL;
1012
1013 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1014 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1015 return __cfqq;
1016
1017 return NULL;
1018}
1019
1020/*
1021 * cfqd - obvious
1022 * cur_cfqq - passed in so that we don't decide that the current queue is
1023 * closely cooperating with itself.
1024 *
1025 * So, basically we're assuming that that cur_cfqq has dispatched at least
1026 * one request, and that cfqd->last_position reflects a position on the disk
1027 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1028 * assumption.
1029 */
1030static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1031 struct cfq_queue *cur_cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001032 bool probe)
Jens Axboea36e71f2009-04-15 12:15:11 +02001033{
1034 struct cfq_queue *cfqq;
1035
1036 /*
1037 * A valid cfq_io_context is necessary to compare requests against
1038 * the seek_mean of the current cfqq.
1039 */
1040 if (!cfqd->active_cic)
1041 return NULL;
1042
Jens Axboe6d048f52007-04-25 12:44:27 +02001043 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001044 * We should notice if some of the queues are cooperating, eg
1045 * working closely on the same area of the disk. In that case,
1046 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001047 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001048 cfqq = cfqq_close(cfqd, cur_cfqq);
1049 if (!cfqq)
1050 return NULL;
1051
1052 if (cfq_cfqq_coop(cfqq))
1053 return NULL;
1054
1055 if (!probe)
1056 cfq_mark_cfqq_coop(cfqq);
1057 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001058}
1059
Jens Axboe6d048f52007-04-25 12:44:27 +02001060static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001061{
Jens Axboe17926692007-01-19 11:59:30 +11001062 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001063 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001064 unsigned long sl;
1065
Jens Axboea68bbdd2008-09-24 13:03:33 +02001066 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001067 * SSD device without seek penalty, disable idling. But only do so
1068 * for devices that support queuing, otherwise we still have a problem
1069 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001070 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001071 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001072 return;
1073
Jens Axboedd67d052006-06-21 09:36:18 +02001074 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001075 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001076
1077 /*
1078 * idle is disabled, either manually or by past process history
1079 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001080 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1081 return;
1082
Jens Axboe22e2c502005-06-27 10:55:12 +02001083 /*
Jens Axboe7b679132008-05-30 12:23:07 +02001084 * still requests with the driver, don't idle
1085 */
Jens Axboe5ad531d2009-07-03 12:57:48 +02001086 if (rq_in_driver(cfqd))
Jens Axboe7b679132008-05-30 12:23:07 +02001087 return;
1088
1089 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001090 * task has exited, don't wait
1091 */
Jens Axboe206dc692006-03-28 13:03:44 +02001092 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001093 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001094 return;
1095
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001096 /*
1097 * If our average think time is larger than the remaining time
1098 * slice, then don't idle. This avoids overrunning the allotted
1099 * time slice.
1100 */
1101 if (sample_valid(cic->ttime_samples) &&
1102 (cfqq->slice_end - jiffies < cic->ttime_mean))
1103 return;
1104
Jens Axboe3b181522005-06-27 10:56:24 +02001105 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001106
Jens Axboe206dc692006-03-28 13:03:44 +02001107 /*
1108 * we don't want to idle for seeks, but we do want to allow
1109 * fair distribution of slice time for a process doing back-to-back
1110 * seeks. so allow a little bit of time for him to submit a new rq
1111 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001112 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001113 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +02001114 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +02001115
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001116 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe9481ffd2009-04-15 12:14:13 +02001117 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118}
1119
Jens Axboe498d3aa22007-04-26 12:54:48 +02001120/*
1121 * Move request from internal lists to the request queue dispatch list.
1122 */
Jens Axboe165125e2007-07-24 09:28:11 +02001123static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001125 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001126 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001127
Jens Axboe7b679132008-05-30 12:23:07 +02001128 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1129
Jeff Moyer06d21882009-09-11 17:08:59 +02001130 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001131 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001132 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001133 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001134
1135 if (cfq_cfqq_sync(cfqq))
1136 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137}
1138
1139/*
1140 * return expired entry, or NULL to just start from scratch in rbtree
1141 */
Jens Axboefebffd62008-01-28 13:19:43 +01001142static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143{
Jens Axboe30996f42009-10-05 11:03:39 +02001144 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145
Jens Axboe3b181522005-06-27 10:56:24 +02001146 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001148
1149 cfq_mark_cfqq_fifo_expire(cfqq);
1150
Jens Axboe89850f72006-07-22 16:48:31 +02001151 if (list_empty(&cfqq->fifo))
1152 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153
Jens Axboe89850f72006-07-22 16:48:31 +02001154 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02001155 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02001156 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157
Jens Axboe30996f42009-10-05 11:03:39 +02001158 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001159 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160}
1161
Jens Axboe22e2c502005-06-27 10:55:12 +02001162static inline int
1163cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1164{
1165 const int base_rq = cfqd->cfq_slice_async_rq;
1166
1167 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1168
1169 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1170}
1171
1172/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02001173 * Select a queue for service. If we have a current active queue,
1174 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02001175 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001176static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001177{
Jens Axboea36e71f2009-04-15 12:15:11 +02001178 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001179
1180 cfqq = cfqd->active_queue;
1181 if (!cfqq)
1182 goto new_queue;
1183
1184 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001185 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02001186 */
Jens Axboeb0291952009-04-07 11:38:31 +02001187 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001188 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +02001189
1190 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001191 * The active queue has requests and isn't expired, allow it to
1192 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02001193 */
Jens Axboedd67d052006-06-21 09:36:18 +02001194 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001195 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02001196
1197 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02001198 * If another queue has a request waiting within our mean seek
1199 * distance, let it run. The expire code will check for close
1200 * cooperators and put the close queue at the front of the service
1201 * tree.
1202 */
1203 new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1204 if (new_cfqq)
1205 goto expire;
1206
1207 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001208 * No requests pending. If the active queue still has requests in
1209 * flight or is idling for a new request, allow either of these
1210 * conditions to happen (or time out) before selecting a new queue.
1211 */
Jens Axboecc197472007-04-20 20:45:39 +02001212 if (timer_pending(&cfqd->idle_slice_timer) ||
1213 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001214 cfqq = NULL;
1215 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001216 }
1217
Jens Axboe3b181522005-06-27 10:56:24 +02001218expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001219 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001220new_queue:
Jens Axboea36e71f2009-04-15 12:15:11 +02001221 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001222keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001223 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001224}
1225
Jens Axboefebffd62008-01-28 13:19:43 +01001226static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02001227{
1228 int dispatched = 0;
1229
1230 while (cfqq->next_rq) {
1231 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1232 dispatched++;
1233 }
1234
1235 BUG_ON(!list_empty(&cfqq->fifo));
1236 return dispatched;
1237}
1238
Jens Axboe498d3aa22007-04-26 12:54:48 +02001239/*
1240 * Drain our current requests. Used for barriers and when switching
1241 * io schedulers on-the-fly.
1242 */
Jens Axboed9e76202007-04-20 14:27:50 +02001243static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001244{
Jens Axboe08717142008-01-28 11:38:15 +01001245 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001246 int dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001247
Jens Axboe08717142008-01-28 11:38:15 +01001248 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
Jens Axboed9e76202007-04-20 14:27:50 +02001249 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001250
Jens Axboe6084cdd2007-04-23 08:25:00 +02001251 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001252
1253 BUG_ON(cfqd->busy_queues);
1254
Jeff Moyer69237152009-06-12 15:29:30 +02001255 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001256 return dispatched;
1257}
1258
Jens Axboe0b182d62009-10-06 20:49:37 +02001259static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02001260{
Jens Axboe2f5cb732009-04-07 08:51:19 +02001261 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262
Jens Axboe2f5cb732009-04-07 08:51:19 +02001263 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02001264 * Drain async requests before we start sync IO
1265 */
1266 if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02001267 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02001268
1269 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02001270 * If this is an async queue and we have sync IO in flight, let it wait
1271 */
1272 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02001273 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001274
1275 max_dispatch = cfqd->cfq_quantum;
1276 if (cfq_class_idle(cfqq))
1277 max_dispatch = 1;
1278
1279 /*
1280 * Does this cfqq already have too much IO in flight?
1281 */
1282 if (cfqq->dispatched >= max_dispatch) {
1283 /*
1284 * idle queue must always only have a single IO in flight
1285 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02001286 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02001287 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001288
Jens Axboe2f5cb732009-04-07 08:51:19 +02001289 /*
1290 * We have other queues, don't allow more IO from this one
1291 */
1292 if (cfqd->busy_queues > 1)
Jens Axboe0b182d62009-10-06 20:49:37 +02001293 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11001294
Jens Axboe2f5cb732009-04-07 08:51:19 +02001295 /*
Jens Axboe8e296752009-10-03 16:26:03 +02001296 * Sole queue user, allow bigger slice
Vivek Goyal365722b2009-10-03 15:21:27 +02001297 */
Jens Axboe8e296752009-10-03 16:26:03 +02001298 max_dispatch *= 4;
1299 }
1300
1301 /*
1302 * Async queues must wait a bit before being allowed dispatch.
1303 * We also ramp up the dispatch depth gradually for async IO,
1304 * based on the last sync IO we serviced
1305 */
Jens Axboe963b72f2009-10-03 19:42:18 +02001306 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Jens Axboe8e296752009-10-03 16:26:03 +02001307 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1308 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02001309
Jens Axboe61f0c1d2009-10-03 19:46:03 +02001310 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02001311 if (!depth && !cfqq->dispatched)
1312 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02001313 if (depth < max_dispatch)
1314 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 }
1316
Jens Axboe0b182d62009-10-06 20:49:37 +02001317 /*
1318 * If we're below the current max, allow a dispatch
1319 */
1320 return cfqq->dispatched < max_dispatch;
1321}
1322
1323/*
1324 * Dispatch a request from cfqq, moving them to the request queue
1325 * dispatch list.
1326 */
1327static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1328{
1329 struct request *rq;
1330
1331 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1332
1333 if (!cfq_may_dispatch(cfqd, cfqq))
1334 return false;
1335
1336 /*
1337 * follow expired path, else get first next available
1338 */
1339 rq = cfq_check_fifo(cfqq);
1340 if (!rq)
1341 rq = cfqq->next_rq;
1342
1343 /*
1344 * insert request into driver dispatch list
1345 */
1346 cfq_dispatch_insert(cfqd->queue, rq);
1347
1348 if (!cfqd->active_cic) {
1349 struct cfq_io_context *cic = RQ_CIC(rq);
1350
1351 atomic_long_inc(&cic->ioc->refcount);
1352 cfqd->active_cic = cic;
1353 }
1354
1355 return true;
1356}
1357
1358/*
1359 * Find the cfqq that we need to service and move a request from that to the
1360 * dispatch list
1361 */
1362static int cfq_dispatch_requests(struct request_queue *q, int force)
1363{
1364 struct cfq_data *cfqd = q->elevator->elevator_data;
1365 struct cfq_queue *cfqq;
1366
1367 if (!cfqd->busy_queues)
1368 return 0;
1369
1370 if (unlikely(force))
1371 return cfq_forced_dispatch(cfqd);
1372
1373 cfqq = cfq_select_queue(cfqd);
1374 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02001375 return 0;
1376
Jens Axboe2f5cb732009-04-07 08:51:19 +02001377 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02001378 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02001379 */
Jens Axboe0b182d62009-10-06 20:49:37 +02001380 if (!cfq_dispatch_request(cfqd, cfqq))
1381 return 0;
1382
Jens Axboe2f5cb732009-04-07 08:51:19 +02001383 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02001384 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001385
1386 /*
1387 * expire an async queue immediately if it has used up its slice. idle
1388 * queue always expire after 1 dispatch round.
1389 */
1390 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1391 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1392 cfq_class_idle(cfqq))) {
1393 cfqq->slice_end = jiffies + 1;
1394 cfq_slice_expired(cfqd, 0);
1395 }
1396
Shan Weib217a902009-09-01 10:06:42 +02001397 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02001398 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399}
1400
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401/*
Jens Axboe5e705372006-07-13 12:39:25 +02001402 * task holds one reference to the queue, dropped when task exits. each rq
1403 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 *
1405 * queue lock must be held here.
1406 */
1407static void cfq_put_queue(struct cfq_queue *cfqq)
1408{
Jens Axboe22e2c502005-06-27 10:55:12 +02001409 struct cfq_data *cfqd = cfqq->cfqd;
1410
1411 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412
1413 if (!atomic_dec_and_test(&cfqq->ref))
1414 return;
1415
Jens Axboe7b679132008-05-30 12:23:07 +02001416 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001418 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001419 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001421 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001422 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02001423 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001424 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001425
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 kmem_cache_free(cfq_pool, cfqq);
1427}
1428
Jens Axboed6de8be2008-05-28 14:46:59 +02001429/*
1430 * Must always be called with the rcu_read_lock() held
1431 */
Jens Axboe07416d22008-05-07 09:17:12 +02001432static void
1433__call_for_each_cic(struct io_context *ioc,
1434 void (*func)(struct io_context *, struct cfq_io_context *))
1435{
1436 struct cfq_io_context *cic;
1437 struct hlist_node *n;
1438
1439 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1440 func(ioc, cic);
1441}
1442
Jens Axboe4ac845a2008-01-24 08:44:49 +01001443/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001444 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001445 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001446static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001447call_for_each_cic(struct io_context *ioc,
1448 void (*func)(struct io_context *, struct cfq_io_context *))
1449{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001450 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02001451 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001452 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001453}
Jens Axboe4ac845a2008-01-24 08:44:49 +01001454
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001455static void cfq_cic_free_rcu(struct rcu_head *head)
1456{
1457 struct cfq_io_context *cic;
1458
1459 cic = container_of(head, struct cfq_io_context, rcu_head);
1460
1461 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09001462 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001463
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001464 if (ioc_gone) {
1465 /*
1466 * CFQ scheduler is exiting, grab exit lock and check
1467 * the pending io context count. If it hits zero,
1468 * complete ioc_gone and set it back to NULL
1469 */
1470 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09001471 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001472 complete(ioc_gone);
1473 ioc_gone = NULL;
1474 }
1475 spin_unlock(&ioc_gone_lock);
1476 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001477}
1478
1479static void cfq_cic_free(struct cfq_io_context *cic)
1480{
1481 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001482}
1483
1484static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1485{
1486 unsigned long flags;
1487
1488 BUG_ON(!cic->dead_key);
1489
1490 spin_lock_irqsave(&ioc->lock, flags);
1491 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01001492 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001493 spin_unlock_irqrestore(&ioc->lock, flags);
1494
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001495 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001496}
1497
Jens Axboed6de8be2008-05-28 14:46:59 +02001498/*
1499 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1500 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1501 * and ->trim() which is called with the task lock held
1502 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001503static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001505 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001506 * ioc->refcount is zero here, or we are called from elv_unregister(),
1507 * so no more cic's are allowed to be linked into this ioc. So it
1508 * should be ok to iterate over the known list, we will see all cic's
1509 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001510 */
Jens Axboe07416d22008-05-07 09:17:12 +02001511 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512}
1513
Jens Axboe89850f72006-07-22 16:48:31 +02001514static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1515{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001516 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001517 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02001518 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001519 }
Jens Axboe89850f72006-07-22 16:48:31 +02001520
1521 cfq_put_queue(cfqq);
1522}
1523
1524static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1525 struct cfq_io_context *cic)
1526{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001527 struct io_context *ioc = cic->ioc;
1528
Jens Axboefc463792006-08-29 09:05:44 +02001529 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001530
1531 /*
1532 * Make sure key == NULL is seen for dead queues
1533 */
Jens Axboefc463792006-08-29 09:05:44 +02001534 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001535 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02001536 cic->key = NULL;
1537
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001538 if (ioc->ioc_data == cic)
1539 rcu_assign_pointer(ioc->ioc_data, NULL);
1540
Jens Axboeff6657c2009-04-08 10:58:57 +02001541 if (cic->cfqq[BLK_RW_ASYNC]) {
1542 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1543 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001544 }
1545
Jens Axboeff6657c2009-04-08 10:58:57 +02001546 if (cic->cfqq[BLK_RW_SYNC]) {
1547 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1548 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001549 }
Jens Axboe89850f72006-07-22 16:48:31 +02001550}
1551
Jens Axboe4ac845a2008-01-24 08:44:49 +01001552static void cfq_exit_single_io_context(struct io_context *ioc,
1553 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001554{
Al Viro478a82b2006-03-18 13:25:24 -05001555 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001556
Jens Axboe89850f72006-07-22 16:48:31 +02001557 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001558 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001559 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02001560
Jens Axboe4ac845a2008-01-24 08:44:49 +01001561 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01001562
1563 /*
1564 * Ensure we get a fresh copy of the ->key to prevent
1565 * race between exiting task and queue
1566 */
1567 smp_read_barrier_depends();
1568 if (cic->key)
1569 __cfq_exit_single_io_context(cfqd, cic);
1570
Jens Axboe4ac845a2008-01-24 08:44:49 +01001571 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05001572 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001573}
1574
Jens Axboe498d3aa22007-04-26 12:54:48 +02001575/*
1576 * The process that ioc belongs to has exited, we need to clean up
1577 * and put the internal structures we have that belongs to that process.
1578 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001579static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001581 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582}
1583
Jens Axboe22e2c502005-06-27 10:55:12 +02001584static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001585cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586{
Jens Axboeb5deef92006-07-19 23:39:40 +02001587 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588
Christoph Lameter94f60302007-07-17 04:03:29 -07001589 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1590 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001592 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001593 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01001594 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001595 cic->dtor = cfq_free_io_context;
1596 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09001597 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598 }
1599
1600 return cic;
1601}
1602
Jens Axboefd0928d2008-01-24 08:52:45 +01001603static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02001604{
1605 struct task_struct *tsk = current;
1606 int ioprio_class;
1607
Jens Axboe3b181522005-06-27 10:56:24 +02001608 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001609 return;
1610
Jens Axboefd0928d2008-01-24 08:52:45 +01001611 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001612 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01001613 default:
1614 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1615 case IOPRIO_CLASS_NONE:
1616 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02001617 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01001618 */
1619 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02001620 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01001621 break;
1622 case IOPRIO_CLASS_RT:
1623 cfqq->ioprio = task_ioprio(ioc);
1624 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1625 break;
1626 case IOPRIO_CLASS_BE:
1627 cfqq->ioprio = task_ioprio(ioc);
1628 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1629 break;
1630 case IOPRIO_CLASS_IDLE:
1631 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1632 cfqq->ioprio = 7;
1633 cfq_clear_cfqq_idle_window(cfqq);
1634 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02001635 }
1636
1637 /*
1638 * keep track of original prio settings in case we have to temporarily
1639 * elevate the priority of this queue
1640 */
1641 cfqq->org_ioprio = cfqq->ioprio;
1642 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001643 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001644}
1645
Jens Axboefebffd62008-01-28 13:19:43 +01001646static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001647{
Al Viro478a82b2006-03-18 13:25:24 -05001648 struct cfq_data *cfqd = cic->key;
1649 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001650 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001651
Jens Axboecaaa5f92006-06-16 11:23:00 +02001652 if (unlikely(!cfqd))
1653 return;
1654
Jens Axboec1b707d2006-10-30 19:54:23 +01001655 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001656
Jens Axboeff6657c2009-04-08 10:58:57 +02001657 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001658 if (cfqq) {
1659 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02001660 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1661 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001662 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02001663 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001664 cfq_put_queue(cfqq);
1665 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001666 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001667
Jens Axboeff6657c2009-04-08 10:58:57 +02001668 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001669 if (cfqq)
1670 cfq_mark_cfqq_prio_changed(cfqq);
1671
Jens Axboec1b707d2006-10-30 19:54:23 +01001672 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001673}
1674
Jens Axboefc463792006-08-29 09:05:44 +02001675static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001677 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02001678 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679}
1680
Jens Axboed5036d72009-06-26 10:44:34 +02001681static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001682 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02001683{
1684 RB_CLEAR_NODE(&cfqq->rb_node);
1685 RB_CLEAR_NODE(&cfqq->p_node);
1686 INIT_LIST_HEAD(&cfqq->fifo);
1687
1688 atomic_set(&cfqq->ref, 0);
1689 cfqq->cfqd = cfqd;
1690
1691 cfq_mark_cfqq_prio_changed(cfqq);
1692
1693 if (is_sync) {
1694 if (!cfq_class_idle(cfqq))
1695 cfq_mark_cfqq_idle_window(cfqq);
1696 cfq_mark_cfqq_sync(cfqq);
1697 }
1698 cfqq->pid = pid;
1699}
1700
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02001702cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01001703 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001706 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707
1708retry:
Jens Axboe4ac845a2008-01-24 08:44:49 +01001709 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001710 /* cic always exists here */
1711 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712
Jens Axboe6118b702009-06-30 09:34:12 +02001713 /*
1714 * Always try a new alloc if we fell back to the OOM cfqq
1715 * originally, since it should just be a temporary situation.
1716 */
1717 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1718 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 if (new_cfqq) {
1720 cfqq = new_cfqq;
1721 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001722 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001724 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02001725 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07001726 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02001728 if (new_cfqq)
1729 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001730 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001731 cfqq = kmem_cache_alloc_node(cfq_pool,
1732 gfp_mask | __GFP_ZERO,
1733 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001734 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735
Jens Axboe6118b702009-06-30 09:34:12 +02001736 if (cfqq) {
1737 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1738 cfq_init_prio_data(cfqq, ioc);
1739 cfq_log_cfqq(cfqd, cfqq, "alloced");
1740 } else
1741 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 }
1743
1744 if (new_cfqq)
1745 kmem_cache_free(cfq_pool, new_cfqq);
1746
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 return cfqq;
1748}
1749
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001750static struct cfq_queue **
1751cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1752{
Jens Axboefe094d92008-01-31 13:08:54 +01001753 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001754 case IOPRIO_CLASS_RT:
1755 return &cfqd->async_cfqq[0][ioprio];
1756 case IOPRIO_CLASS_BE:
1757 return &cfqd->async_cfqq[1][ioprio];
1758 case IOPRIO_CLASS_IDLE:
1759 return &cfqd->async_idle_cfqq;
1760 default:
1761 BUG();
1762 }
1763}
1764
Jens Axboe15c31be2007-07-10 13:43:25 +02001765static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02001766cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02001767 gfp_t gfp_mask)
1768{
Jens Axboefd0928d2008-01-24 08:52:45 +01001769 const int ioprio = task_ioprio(ioc);
1770 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001771 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001772 struct cfq_queue *cfqq = NULL;
1773
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001774 if (!is_sync) {
1775 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1776 cfqq = *async_cfqq;
1777 }
1778
Jens Axboe6118b702009-06-30 09:34:12 +02001779 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01001780 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02001781
1782 /*
1783 * pin the queue now that it's allocated, scheduler exit will prune it
1784 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001785 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001786 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001787 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001788 }
1789
1790 atomic_inc(&cfqq->ref);
1791 return cfqq;
1792}
1793
Jens Axboe498d3aa22007-04-26 12:54:48 +02001794/*
1795 * We drop cfq io contexts lazily, so we may find a dead one.
1796 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001797static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001798cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1799 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001800{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001801 unsigned long flags;
1802
Jens Axboefc463792006-08-29 09:05:44 +02001803 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001804
Jens Axboe4ac845a2008-01-24 08:44:49 +01001805 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02001806
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001807 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001808
1809 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01001810 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001811 spin_unlock_irqrestore(&ioc->lock, flags);
1812
1813 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001814}
1815
Jens Axboee2d74ac2006-03-28 08:59:01 +02001816static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01001817cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001818{
Jens Axboee2d74ac2006-03-28 08:59:01 +02001819 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001820 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001821 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001822
Vasily Tarasov91fac312007-04-25 12:29:51 +02001823 if (unlikely(!ioc))
1824 return NULL;
1825
Jens Axboed6de8be2008-05-28 14:46:59 +02001826 rcu_read_lock();
1827
Jens Axboe597bc482007-04-24 21:23:53 +02001828 /*
1829 * we maintain a last-hit cache, to avoid browsing over the tree
1830 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001831 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02001832 if (cic && cic->key == cfqd) {
1833 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02001834 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001835 }
Jens Axboe597bc482007-04-24 21:23:53 +02001836
Jens Axboe4ac845a2008-01-24 08:44:49 +01001837 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001838 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1839 rcu_read_unlock();
1840 if (!cic)
1841 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001842 /* ->key must be copied to avoid race with cfq_exit_queue() */
1843 k = cic->key;
1844 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001845 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001846 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001847 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001848 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001849
Jens Axboed6de8be2008-05-28 14:46:59 +02001850 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001851 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001852 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001853 break;
1854 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001855
Jens Axboe4ac845a2008-01-24 08:44:49 +01001856 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001857}
1858
Jens Axboe4ac845a2008-01-24 08:44:49 +01001859/*
1860 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1861 * the process specific cfq io context when entered from the block layer.
1862 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1863 */
Jens Axboefebffd62008-01-28 13:19:43 +01001864static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1865 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001866{
Jens Axboe0261d682006-10-30 19:07:48 +01001867 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001868 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001869
Jens Axboe4ac845a2008-01-24 08:44:49 +01001870 ret = radix_tree_preload(gfp_mask);
1871 if (!ret) {
1872 cic->ioc = ioc;
1873 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001874
Jens Axboe4ac845a2008-01-24 08:44:49 +01001875 spin_lock_irqsave(&ioc->lock, flags);
1876 ret = radix_tree_insert(&ioc->radix_root,
1877 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01001878 if (!ret)
1879 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001880 spin_unlock_irqrestore(&ioc->lock, flags);
1881
1882 radix_tree_preload_end();
1883
1884 if (!ret) {
1885 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1886 list_add(&cic->queue_list, &cfqd->cic_list);
1887 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001888 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001889 }
1890
Jens Axboe4ac845a2008-01-24 08:44:49 +01001891 if (ret)
1892 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02001893
Jens Axboe4ac845a2008-01-24 08:44:49 +01001894 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001895}
1896
Jens Axboe22e2c502005-06-27 10:55:12 +02001897/*
1898 * Setup general io context and cfq io context. There can be several cfq
1899 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001900 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001901 */
1902static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001903cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904{
Jens Axboe22e2c502005-06-27 10:55:12 +02001905 struct io_context *ioc = NULL;
1906 struct cfq_io_context *cic;
1907
1908 might_sleep_if(gfp_mask & __GFP_WAIT);
1909
Jens Axboeb5deef92006-07-19 23:39:40 +02001910 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001911 if (!ioc)
1912 return NULL;
1913
Jens Axboe4ac845a2008-01-24 08:44:49 +01001914 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001915 if (cic)
1916 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001917
Jens Axboee2d74ac2006-03-28 08:59:01 +02001918 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1919 if (cic == NULL)
1920 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001921
Jens Axboe4ac845a2008-01-24 08:44:49 +01001922 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1923 goto err_free;
1924
Jens Axboe22e2c502005-06-27 10:55:12 +02001925out:
Jens Axboefc463792006-08-29 09:05:44 +02001926 smp_read_barrier_depends();
1927 if (unlikely(ioc->ioprio_changed))
1928 cfq_ioc_set_ioprio(ioc);
1929
Jens Axboe22e2c502005-06-27 10:55:12 +02001930 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001931err_free:
1932 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001933err:
1934 put_io_context(ioc);
1935 return NULL;
1936}
1937
1938static void
1939cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1940{
Jens Axboeaaf12282007-01-19 11:30:16 +11001941 unsigned long elapsed = jiffies - cic->last_end_request;
1942 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001943
1944 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1945 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1946 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1947}
1948
Jens Axboe206dc692006-03-28 13:03:44 +02001949static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001950cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1951 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001952{
1953 sector_t sdist;
1954 u64 total;
1955
Jeff Moyer4d00aa42009-04-21 07:25:04 +02001956 if (!cic->last_request_pos)
1957 sdist = 0;
Tejun Heo83096eb2009-05-07 22:24:39 +09001958 else if (cic->last_request_pos < blk_rq_pos(rq))
1959 sdist = blk_rq_pos(rq) - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001960 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001961 sdist = cic->last_request_pos - blk_rq_pos(rq);
Jens Axboe206dc692006-03-28 13:03:44 +02001962
1963 /*
1964 * Don't allow the seek distance to get too large from the
1965 * odd fragment, pagein, etc
1966 */
1967 if (cic->seek_samples <= 60) /* second&third seek */
1968 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1969 else
1970 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1971
1972 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1973 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1974 total = cic->seek_total + (cic->seek_samples/2);
1975 do_div(total, cic->seek_samples);
1976 cic->seek_mean = (sector_t)total;
1977}
Jens Axboe22e2c502005-06-27 10:55:12 +02001978
1979/*
1980 * Disable idle window if the process thinks too long or seeks so much that
1981 * it doesn't matter
1982 */
1983static void
1984cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1985 struct cfq_io_context *cic)
1986{
Jens Axboe7b679132008-05-30 12:23:07 +02001987 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02001988
Jens Axboe08717142008-01-28 11:38:15 +01001989 /*
1990 * Don't idle for async or idle io prio class
1991 */
1992 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02001993 return;
1994
Jens Axboec265a7f2008-06-26 13:49:33 +02001995 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001996
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001997 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Jens Axboe963b72f2009-10-03 19:42:18 +02001998 (!cfqd->cfq_latency && cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001999 enable_idle = 0;
2000 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccoloec60e4f2009-10-07 19:51:54 +02002001 unsigned int slice_idle = cfqd->cfq_slice_idle;
2002 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
2003 slice_idle = msecs_to_jiffies(CFQ_MIN_TT);
2004 if (cic->ttime_mean > slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02002005 enable_idle = 0;
2006 else
2007 enable_idle = 1;
2008 }
2009
Jens Axboe7b679132008-05-30 12:23:07 +02002010 if (old_idle != enable_idle) {
2011 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2012 if (enable_idle)
2013 cfq_mark_cfqq_idle_window(cfqq);
2014 else
2015 cfq_clear_cfqq_idle_window(cfqq);
2016 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002017}
2018
Jens Axboe22e2c502005-06-27 10:55:12 +02002019/*
2020 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2021 * no or if we aren't sure, a 1 will cause a preempt.
2022 */
Jens Axboea6151c32009-10-07 20:02:57 +02002023static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02002024cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02002025 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002026{
Jens Axboe6d048f52007-04-25 12:44:27 +02002027 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002028
Jens Axboe6d048f52007-04-25 12:44:27 +02002029 cfqq = cfqd->active_queue;
2030 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02002031 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02002032
Jens Axboe6d048f52007-04-25 12:44:27 +02002033 if (cfq_slice_used(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002034 return true;
Jens Axboe6d048f52007-04-25 12:44:27 +02002035
2036 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002037 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02002038
2039 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002040 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002041
Jens Axboe22e2c502005-06-27 10:55:12 +02002042 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02002043 * if the new request is sync, but the currently running queue is
2044 * not, let the sync request have priority.
2045 */
Jens Axboe5e705372006-07-13 12:39:25 +02002046 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002047 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002048
Jens Axboe374f84a2006-07-23 01:42:19 +02002049 /*
2050 * So both queues are sync. Let the new request get disk time if
2051 * it's a metadata request and the current queue is doing regular IO.
2052 */
2053 if (rq_is_meta(rq) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01002054 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02002055
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002056 /*
2057 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2058 */
2059 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002060 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002061
Jens Axboe1e3335d2007-02-14 19:59:49 +01002062 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002063 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002064
2065 /*
2066 * if this request is as-good as one we would expect from the
2067 * current cfqq, let it preempt
2068 */
Jens Axboe6d048f52007-04-25 12:44:27 +02002069 if (cfq_rq_close(cfqd, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02002070 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002071
Jens Axboea6151c32009-10-07 20:02:57 +02002072 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02002073}
2074
2075/*
2076 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2077 * let it have half of its nominal slice.
2078 */
2079static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2080{
Jens Axboe7b679132008-05-30 12:23:07 +02002081 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02002082 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02002083
Jens Axboebf572252006-07-19 20:29:12 +02002084 /*
2085 * Put the new queue at the front of the of the current list,
2086 * so we know that it will be selected next.
2087 */
2088 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02002089
2090 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02002091
Jens Axboe44f7c162007-01-19 11:51:58 +11002092 cfqq->slice_end = 0;
2093 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002094}
2095
2096/*
Jens Axboe5e705372006-07-13 12:39:25 +02002097 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02002098 * something we should do about it
2099 */
2100static void
Jens Axboe5e705372006-07-13 12:39:25 +02002101cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2102 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002103{
Jens Axboe5e705372006-07-13 12:39:25 +02002104 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02002105
Aaron Carroll45333d52008-08-26 15:52:36 +02002106 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02002107 if (rq_is_meta(rq))
2108 cfqq->meta_pending++;
2109
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002110 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02002111 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002112 cfq_update_idle_window(cfqd, cfqq, cic);
2113
Tejun Heo83096eb2009-05-07 22:24:39 +09002114 cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002115
2116 if (cfqq == cfqd->active_queue) {
2117 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002118 * Remember that we saw a request from this process, but
2119 * don't start queuing just yet. Otherwise we risk seeing lots
2120 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02002121 * and merging. If the request is already larger than a single
2122 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02002123 * merging is already done. Ditto for a busy system that
2124 * has other work pending, don't risk delaying until the
2125 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02002126 */
Jens Axboed6ceb252009-04-14 14:18:16 +02002127 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02002128 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2129 cfqd->busy_queues > 1) {
Jens Axboed6ceb252009-04-14 14:18:16 +02002130 del_timer(&cfqd->idle_slice_timer);
Tejun Heoa7f55792009-04-23 11:05:17 +09002131 __blk_run_queue(cfqd->queue);
Jens Axboed6ceb252009-04-14 14:18:16 +02002132 }
Jens Axboeb0291952009-04-07 11:38:31 +02002133 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboed6ceb252009-04-14 14:18:16 +02002134 }
Jens Axboe5e705372006-07-13 12:39:25 +02002135 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002136 /*
2137 * not the active queue - expire current slice if it is
2138 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002139 * has some old slice time left and is of higher priority or
2140 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02002141 */
2142 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09002143 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002144 }
2145}
2146
Jens Axboe165125e2007-07-24 09:28:11 +02002147static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002148{
Jens Axboeb4878f22005-10-20 16:42:29 +02002149 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002150 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002151
Jens Axboe7b679132008-05-30 12:23:07 +02002152 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01002153 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154
Jens Axboe5e705372006-07-13 12:39:25 +02002155 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156
Jens Axboe30996f42009-10-05 11:03:39 +02002157 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02002158 list_add_tail(&rq->queuelist, &cfqq->fifo);
2159
Jens Axboe5e705372006-07-13 12:39:25 +02002160 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161}
2162
Aaron Carroll45333d52008-08-26 15:52:36 +02002163/*
2164 * Update hw_tag based on peak queue depth over 50 samples under
2165 * sufficient load.
2166 */
2167static void cfq_update_hw_tag(struct cfq_data *cfqd)
2168{
Jens Axboe5ad531d2009-07-03 12:57:48 +02002169 if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2170 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
Aaron Carroll45333d52008-08-26 15:52:36 +02002171
2172 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Jens Axboe5ad531d2009-07-03 12:57:48 +02002173 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02002174 return;
2175
2176 if (cfqd->hw_tag_samples++ < 50)
2177 return;
2178
2179 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2180 cfqd->hw_tag = 1;
2181 else
2182 cfqd->hw_tag = 0;
2183
2184 cfqd->hw_tag_samples = 0;
2185 cfqd->rq_in_driver_peak = 0;
2186}
2187
Jens Axboe165125e2007-07-24 09:28:11 +02002188static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189{
Jens Axboe5e705372006-07-13 12:39:25 +02002190 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002191 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02002192 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002193 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194
Jens Axboeb4878f22005-10-20 16:42:29 +02002195 now = jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02002196 cfq_log_cfqq(cfqd, cfqq, "complete");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197
Aaron Carroll45333d52008-08-26 15:52:36 +02002198 cfq_update_hw_tag(cfqd);
2199
Jens Axboe5ad531d2009-07-03 12:57:48 +02002200 WARN_ON(!cfqd->rq_in_driver[sync]);
Jens Axboe6d048f52007-04-25 12:44:27 +02002201 WARN_ON(!cfqq->dispatched);
Jens Axboe5ad531d2009-07-03 12:57:48 +02002202 cfqd->rq_in_driver[sync]--;
Jens Axboe6d048f52007-04-25 12:44:27 +02002203 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204
Jens Axboe3ed9a292007-04-23 08:33:33 +02002205 if (cfq_cfqq_sync(cfqq))
2206 cfqd->sync_flight--;
2207
Vivek Goyal365722b2009-10-03 15:21:27 +02002208 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02002209 RQ_CIC(rq)->last_end_request = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02002210 cfqd->last_end_sync_rq = now;
2211 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002212
2213 /*
2214 * If this is the active queue, check if it needs to be expired,
2215 * or if we want to idle in case it has no pending requests.
2216 */
2217 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02002218 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2219
Jens Axboe44f7c162007-01-19 11:51:58 +11002220 if (cfq_cfqq_slice_new(cfqq)) {
2221 cfq_set_prio_slice(cfqd, cfqq);
2222 cfq_clear_cfqq_slice_new(cfqq);
2223 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002224 /*
2225 * If there are no requests waiting in this queue, and
2226 * there are other queues ready to issue requests, AND
2227 * those other queues are issuing requests within our
2228 * mean seek distance, give them a chance to run instead
2229 * of idling.
2230 */
Jens Axboe08717142008-01-28 11:38:15 +01002231 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02002232 cfq_slice_expired(cfqd, 1);
Jens Axboea36e71f2009-04-15 12:15:11 +02002233 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2234 sync && !rq_noidle(rq))
Jens Axboe6d048f52007-04-25 12:44:27 +02002235 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002236 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002237
Jens Axboe5ad531d2009-07-03 12:57:48 +02002238 if (!rq_in_driver(cfqd))
Jens Axboe23e018a2009-10-05 08:52:35 +02002239 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240}
2241
Jens Axboe22e2c502005-06-27 10:55:12 +02002242/*
2243 * we temporarily boost lower priority queues if they are holding fs exclusive
2244 * resources. they are boosted to normal prio (CLASS_BE/4)
2245 */
2246static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247{
Jens Axboe22e2c502005-06-27 10:55:12 +02002248 if (has_fs_excl()) {
2249 /*
2250 * boost idle prio on transactions that would lock out other
2251 * users of the filesystem
2252 */
2253 if (cfq_class_idle(cfqq))
2254 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2255 if (cfqq->ioprio > IOPRIO_NORM)
2256 cfqq->ioprio = IOPRIO_NORM;
2257 } else {
2258 /*
2259 * check if we need to unboost the queue
2260 */
2261 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2262 cfqq->ioprio_class = cfqq->org_ioprio_class;
2263 if (cfqq->ioprio != cfqq->org_ioprio)
2264 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002266}
2267
Jens Axboe89850f72006-07-22 16:48:31 +02002268static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002269{
Jens Axboe1b379d82009-08-11 08:26:11 +02002270 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002271 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002272 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02002273 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002274
2275 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02002276}
2277
Jens Axboe165125e2007-07-24 09:28:11 +02002278static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02002279{
2280 struct cfq_data *cfqd = q->elevator->elevator_data;
2281 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002282 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02002283 struct cfq_queue *cfqq;
2284
2285 /*
2286 * don't force setup of a queue from here, as a call to may_queue
2287 * does not necessarily imply that a request actually will be queued.
2288 * so just lookup a possibly existing queue, or return 'may queue'
2289 * if that fails
2290 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002291 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002292 if (!cic)
2293 return ELV_MQUEUE_MAY;
2294
Jens Axboeb0b78f82009-04-08 10:56:08 +02002295 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02002296 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002297 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02002298 cfq_prio_boost(cfqq);
2299
Jens Axboe89850f72006-07-22 16:48:31 +02002300 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002301 }
2302
2303 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304}
2305
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306/*
2307 * queue lock held here
2308 */
Jens Axboebb37b942006-12-01 10:42:33 +01002309static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310{
Jens Axboe5e705372006-07-13 12:39:25 +02002311 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312
Jens Axboe5e705372006-07-13 12:39:25 +02002313 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002314 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315
Jens Axboe22e2c502005-06-27 10:55:12 +02002316 BUG_ON(!cfqq->allocated[rw]);
2317 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318
Jens Axboe5e705372006-07-13 12:39:25 +02002319 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02002322 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 cfq_put_queue(cfqq);
2325 }
2326}
2327
2328/*
Jens Axboe22e2c502005-06-27 10:55:12 +02002329 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 */
Jens Axboe22e2c502005-06-27 10:55:12 +02002331static int
Jens Axboe165125e2007-07-24 09:28:11 +02002332cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333{
2334 struct cfq_data *cfqd = q->elevator->elevator_data;
2335 struct cfq_io_context *cic;
2336 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02002337 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002338 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 unsigned long flags;
2340
2341 might_sleep_if(gfp_mask & __GFP_WAIT);
2342
Jens Axboee2d74ac2006-03-28 08:59:01 +02002343 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02002344
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 spin_lock_irqsave(q->queue_lock, flags);
2346
Jens Axboe22e2c502005-06-27 10:55:12 +02002347 if (!cic)
2348 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349
Vasily Tarasov91fac312007-04-25 12:29:51 +02002350 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02002351 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002352 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002353 cic_set_cfqq(cic, cfqq, is_sync);
2354 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355
2356 cfqq->allocated[rw]++;
Jens Axboe22e2c502005-06-27 10:55:12 +02002357 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02002358
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 spin_unlock_irqrestore(q->queue_lock, flags);
2360
Jens Axboe5e705372006-07-13 12:39:25 +02002361 rq->elevator_private = cic;
2362 rq->elevator_private2 = cfqq;
2363 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02002364
Jens Axboe22e2c502005-06-27 10:55:12 +02002365queue_fail:
2366 if (cic)
2367 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02002368
Jens Axboe23e018a2009-10-05 08:52:35 +02002369 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02002371 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 return 1;
2373}
2374
David Howells65f27f32006-11-22 14:55:48 +00002375static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02002376{
David Howells65f27f32006-11-22 14:55:48 +00002377 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02002378 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02002379 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002380
Jens Axboe40bb54d2009-04-15 12:11:10 +02002381 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09002382 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02002383 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02002384}
2385
2386/*
2387 * Timer running if the active_queue is currently idling inside its time slice
2388 */
2389static void cfq_idle_slice_timer(unsigned long data)
2390{
2391 struct cfq_data *cfqd = (struct cfq_data *) data;
2392 struct cfq_queue *cfqq;
2393 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002394 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002395
Jens Axboe7b679132008-05-30 12:23:07 +02002396 cfq_log(cfqd, "idle timer fired");
2397
Jens Axboe22e2c502005-06-27 10:55:12 +02002398 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2399
Jens Axboefe094d92008-01-31 13:08:54 +01002400 cfqq = cfqd->active_queue;
2401 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002402 timed_out = 0;
2403
Jens Axboe22e2c502005-06-27 10:55:12 +02002404 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002405 * We saw a request before the queue expired, let it through
2406 */
2407 if (cfq_cfqq_must_dispatch(cfqq))
2408 goto out_kick;
2409
2410 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002411 * expired
2412 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002413 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002414 goto expire;
2415
2416 /*
2417 * only expire and reinvoke request handler, if there are
2418 * other queues with pending requests
2419 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002420 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002421 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002422
2423 /*
2424 * not expired and it has a request pending, let it dispatch
2425 */
Jens Axboe75e50982009-04-07 08:56:14 +02002426 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002427 goto out_kick;
Jens Axboe22e2c502005-06-27 10:55:12 +02002428 }
2429expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002430 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002431out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02002432 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002433out_cont:
2434 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2435}
2436
Jens Axboe3b181522005-06-27 10:56:24 +02002437static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2438{
2439 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02002440 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002441}
Jens Axboe22e2c502005-06-27 10:55:12 +02002442
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002443static void cfq_put_async_queues(struct cfq_data *cfqd)
2444{
2445 int i;
2446
2447 for (i = 0; i < IOPRIO_BE_NR; i++) {
2448 if (cfqd->async_cfqq[0][i])
2449 cfq_put_queue(cfqd->async_cfqq[0][i]);
2450 if (cfqd->async_cfqq[1][i])
2451 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002452 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002453
2454 if (cfqd->async_idle_cfqq)
2455 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002456}
2457
Jens Axboeb374d182008-10-31 10:05:07 +01002458static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459{
Jens Axboe22e2c502005-06-27 10:55:12 +02002460 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002461 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002462
Jens Axboe3b181522005-06-27 10:56:24 +02002463 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002464
Al Virod9ff4182006-03-18 13:51:22 -05002465 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002466
Al Virod9ff4182006-03-18 13:51:22 -05002467 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002468 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002469
2470 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002471 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2472 struct cfq_io_context,
2473 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002474
2475 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002476 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002477
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002478 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002479
Al Virod9ff4182006-03-18 13:51:22 -05002480 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002481
2482 cfq_shutdown_timer_wq(cfqd);
2483
Al Viroa90d7422006-03-18 12:05:37 -05002484 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485}
2486
Jens Axboe165125e2007-07-24 09:28:11 +02002487static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488{
2489 struct cfq_data *cfqd;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002490 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491
Christoph Lameter94f60302007-07-17 04:03:29 -07002492 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002494 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495
Jens Axboecc09e292007-04-26 12:53:50 +02002496 cfqd->service_tree = CFQ_RB_ROOT;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002497
2498 /*
2499 * Not strictly needed (since RB_ROOT just clears the node and we
2500 * zeroed cfqd on alloc), but better be safe in case someone decides
2501 * to add magic to the rb code
2502 */
2503 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2504 cfqd->prio_trees[i] = RB_ROOT;
2505
Jens Axboe6118b702009-06-30 09:34:12 +02002506 /*
2507 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2508 * Grab a permanent reference to it, so that the normal code flow
2509 * will not attempt to free it.
2510 */
2511 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2512 atomic_inc(&cfqd->oom_cfqq.ref);
2513
Al Virod9ff4182006-03-18 13:51:22 -05002514 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517
Jens Axboe22e2c502005-06-27 10:55:12 +02002518 init_timer(&cfqd->idle_slice_timer);
2519 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2520 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2521
Jens Axboe23e018a2009-10-05 08:52:35 +02002522 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002523
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002525 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2526 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 cfqd->cfq_back_max = cfq_back_max;
2528 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002529 cfqd->cfq_slice[0] = cfq_slice_async;
2530 cfqd->cfq_slice[1] = cfq_slice_sync;
2531 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2532 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02002533 cfqd->cfq_latency = 1;
Aaron Carroll45333d52008-08-26 15:52:36 +02002534 cfqd->hw_tag = 1;
Vivek Goyal365722b2009-10-03 15:21:27 +02002535 cfqd->last_end_sync_rq = jiffies;
Jens Axboebc1c1162006-06-08 08:49:06 +02002536 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537}
2538
2539static void cfq_slab_kill(void)
2540{
Jens Axboed6de8be2008-05-28 14:46:59 +02002541 /*
2542 * Caller already ensured that pending RCU callbacks are completed,
2543 * so we should have no busy allocations at this point.
2544 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545 if (cfq_pool)
2546 kmem_cache_destroy(cfq_pool);
2547 if (cfq_ioc_pool)
2548 kmem_cache_destroy(cfq_ioc_pool);
2549}
2550
2551static int __init cfq_slab_setup(void)
2552{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002553 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 if (!cfq_pool)
2555 goto fail;
2556
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002557 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558 if (!cfq_ioc_pool)
2559 goto fail;
2560
2561 return 0;
2562fail:
2563 cfq_slab_kill();
2564 return -ENOMEM;
2565}
2566
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567/*
2568 * sysfs parts below -->
2569 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570static ssize_t
2571cfq_var_show(unsigned int var, char *page)
2572{
2573 return sprintf(page, "%d\n", var);
2574}
2575
2576static ssize_t
2577cfq_var_store(unsigned int *var, const char *page, size_t count)
2578{
2579 char *p = (char *) page;
2580
2581 *var = simple_strtoul(p, &p, 10);
2582 return count;
2583}
2584
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002586static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002588 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589 unsigned int __data = __VAR; \
2590 if (__CONV) \
2591 __data = jiffies_to_msecs(__data); \
2592 return cfq_var_show(__data, (page)); \
2593}
2594SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002595SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2596SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002597SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2598SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002599SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2600SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2601SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2602SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02002603SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604#undef SHOW_FUNCTION
2605
2606#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002607static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002609 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610 unsigned int __data; \
2611 int ret = cfq_var_store(&__data, (page), count); \
2612 if (__data < (MIN)) \
2613 __data = (MIN); \
2614 else if (__data > (MAX)) \
2615 __data = (MAX); \
2616 if (__CONV) \
2617 *(__PTR) = msecs_to_jiffies(__data); \
2618 else \
2619 *(__PTR) = __data; \
2620 return ret; \
2621}
2622STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002623STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2624 UINT_MAX, 1);
2625STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2626 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002627STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002628STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2629 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002630STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2631STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2632STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01002633STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2634 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02002635STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636#undef STORE_FUNCTION
2637
Al Viroe572ec72006-03-18 22:27:18 -05002638#define CFQ_ATTR(name) \
2639 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002640
Al Viroe572ec72006-03-18 22:27:18 -05002641static struct elv_fs_entry cfq_attrs[] = {
2642 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002643 CFQ_ATTR(fifo_expire_sync),
2644 CFQ_ATTR(fifo_expire_async),
2645 CFQ_ATTR(back_seek_max),
2646 CFQ_ATTR(back_seek_penalty),
2647 CFQ_ATTR(slice_sync),
2648 CFQ_ATTR(slice_async),
2649 CFQ_ATTR(slice_async_rq),
2650 CFQ_ATTR(slice_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02002651 CFQ_ATTR(low_latency),
Al Viroe572ec72006-03-18 22:27:18 -05002652 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653};
2654
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655static struct elevator_type iosched_cfq = {
2656 .ops = {
2657 .elevator_merge_fn = cfq_merge,
2658 .elevator_merged_fn = cfq_merged_request,
2659 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002660 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002661 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002663 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 .elevator_deactivate_req_fn = cfq_deactivate_request,
2665 .elevator_queue_empty_fn = cfq_queue_empty,
2666 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002667 .elevator_former_req_fn = elv_rb_former_request,
2668 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669 .elevator_set_req_fn = cfq_set_request,
2670 .elevator_put_req_fn = cfq_put_request,
2671 .elevator_may_queue_fn = cfq_may_queue,
2672 .elevator_init_fn = cfq_init_queue,
2673 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002674 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 },
Al Viro3d1ab402006-03-18 18:35:43 -05002676 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677 .elevator_name = "cfq",
2678 .elevator_owner = THIS_MODULE,
2679};
2680
2681static int __init cfq_init(void)
2682{
Jens Axboe22e2c502005-06-27 10:55:12 +02002683 /*
2684 * could be 0 on HZ < 1000 setups
2685 */
2686 if (!cfq_slice_async)
2687 cfq_slice_async = 1;
2688 if (!cfq_slice_idle)
2689 cfq_slice_idle = 1;
2690
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691 if (cfq_slab_setup())
2692 return -ENOMEM;
2693
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002694 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002695
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002696 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697}
2698
2699static void __exit cfq_exit(void)
2700{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002701 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002703 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002704 /* ioc_gone's update must be visible before reading ioc_count */
2705 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02002706
2707 /*
2708 * this also protects us from entering cfq_slab_kill() with
2709 * pending RCU callbacks
2710 */
Tejun Heo245b2e72009-06-24 15:13:48 +09002711 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002712 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08002713 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714}
2715
2716module_init(cfq_init);
2717module_exit(cfq_exit);
2718
2719MODULE_AUTHOR("Jens Axboe");
2720MODULE_LICENSE("GPL");
2721MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");