blob: aa1e9535e3588472803ea079acc293238d60b803 [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 */
Shaohua Li4b27e1b2009-11-03 20:25:02 +0100199 CFQ_CFQQ_FLAG_coop_preempt, /* coop preempt */
Jens Axboe3b181522005-06-27 10:56:24 +0200200};
201
202#define CFQ_CFQQ_FNS(name) \
203static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
204{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100205 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200206} \
207static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
208{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100209 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200210} \
211static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
212{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100213 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200214}
215
216CFQ_CFQQ_FNS(on_rr);
217CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200218CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200219CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200220CFQ_CFQQ_FNS(fifo_expire);
221CFQ_CFQQ_FNS(idle_window);
222CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100223CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200224CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200225CFQ_CFQQ_FNS(coop);
Shaohua Li4b27e1b2009-11-03 20:25:02 +0100226CFQ_CFQQ_FNS(coop_preempt);
Jens Axboe3b181522005-06-27 10:56:24 +0200227#undef CFQ_CFQQ_FNS
228
Jens Axboe7b679132008-05-30 12:23:07 +0200229#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
230 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
231#define cfq_log(cfqd, fmt, args...) \
232 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
233
Jens Axboe165125e2007-07-24 09:28:11 +0200234static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200235static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100236 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100237static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200238 struct io_context *);
239
Jens Axboe5ad531d2009-07-03 12:57:48 +0200240static inline int rq_in_driver(struct cfq_data *cfqd)
241{
242 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
243}
244
Vasily Tarasov91fac312007-04-25 12:29:51 +0200245static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200246 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200247{
Jens Axboea6151c32009-10-07 20:02:57 +0200248 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200249}
250
251static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200252 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200253{
Jens Axboea6151c32009-10-07 20:02:57 +0200254 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200255}
256
257/*
258 * We regard a request as SYNC, if it's either a read or has the SYNC bit
259 * set (in which case it could also be direct WRITE).
260 */
Jens Axboea6151c32009-10-07 20:02:57 +0200261static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200262{
Jens Axboea6151c32009-10-07 20:02:57 +0200263 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200264}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700267 * scheduler run of queue, if there are requests pending and no one in the
268 * driver that will restart queueing
269 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200270static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700271{
Jens Axboe7b679132008-05-30 12:23:07 +0200272 if (cfqd->busy_queues) {
273 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200274 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200275 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700276}
277
Jens Axboe165125e2007-07-24 09:28:11 +0200278static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700279{
280 struct cfq_data *cfqd = q->elevator->elevator_data;
281
Jens Axboeb4878f22005-10-20 16:42:29 +0200282 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700283}
284
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100286 * Scale schedule slice based on io priority. Use the sync time slice only
287 * if a queue is marked sync and has sync io queued. A sync queue with async
288 * io only, should not get full sync slice length.
289 */
Jens Axboea6151c32009-10-07 20:02:57 +0200290static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200291 unsigned short prio)
292{
293 const int base_slice = cfqd->cfq_slice[sync];
294
295 WARN_ON(prio >= IOPRIO_BE_NR);
296
297 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
298}
299
Jens Axboe44f7c162007-01-19 11:51:58 +1100300static inline int
301cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
302{
Jens Axboed9e76202007-04-20 14:27:50 +0200303 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100304}
305
306static inline void
307cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
308{
309 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200310 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100311}
312
313/*
314 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
315 * isn't valid until the first request from the dispatch is activated
316 * and the slice time set.
317 */
Jens Axboea6151c32009-10-07 20:02:57 +0200318static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100319{
320 if (cfq_cfqq_slice_new(cfqq))
321 return 0;
322 if (time_before(jiffies, cfqq->slice_end))
323 return 0;
324
325 return 1;
326}
327
328/*
Jens Axboe5e705372006-07-13 12:39:25 +0200329 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200331 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332 */
Jens Axboe5e705372006-07-13 12:39:25 +0200333static struct request *
334cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335{
336 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200338#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
339#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
340 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341
Jens Axboe5e705372006-07-13 12:39:25 +0200342 if (rq1 == NULL || rq1 == rq2)
343 return rq2;
344 if (rq2 == NULL)
345 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200346
Jens Axboe5e705372006-07-13 12:39:25 +0200347 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
348 return rq1;
349 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
350 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200351 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
352 return rq1;
353 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
354 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355
Tejun Heo83096eb2009-05-07 22:24:39 +0900356 s1 = blk_rq_pos(rq1);
357 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358
Jens Axboe6d048f52007-04-25 12:44:27 +0200359 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 /*
362 * by definition, 1KiB is 2 sectors
363 */
364 back_max = cfqd->cfq_back_max * 2;
365
366 /*
367 * Strict one way elevator _except_ in the case where we allow
368 * short backward seeks which are biased as twice the cost of a
369 * similar forward seek.
370 */
371 if (s1 >= last)
372 d1 = s1 - last;
373 else if (s1 + back_max >= last)
374 d1 = (last - s1) * cfqd->cfq_back_penalty;
375 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200376 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377
378 if (s2 >= last)
379 d2 = s2 - last;
380 else if (s2 + back_max >= last)
381 d2 = (last - s2) * cfqd->cfq_back_penalty;
382 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200383 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384
385 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386
Andreas Mohre8a99052006-03-28 08:59:49 +0200387 /*
388 * By doing switch() on the bit mask "wrap" we avoid having to
389 * check two variables for all permutations: --> faster!
390 */
391 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200392 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200393 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200394 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200395 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200396 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200397 else {
398 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200399 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200400 else
Jens Axboe5e705372006-07-13 12:39:25 +0200401 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200402 }
403
404 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200405 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200406 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200407 return rq2;
408 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200409 default:
410 /*
411 * Since both rqs are wrapped,
412 * start with the one that's further behind head
413 * (--> only *one* back seek required),
414 * since back seek takes more time than forward.
415 */
416 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200417 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 else
Jens Axboe5e705372006-07-13 12:39:25 +0200419 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 }
421}
422
Jens Axboe498d3aa22007-04-26 12:54:48 +0200423/*
424 * The below is leftmost cache rbtree addon
425 */
Jens Axboe08717142008-01-28 11:38:15 +0100426static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200427{
428 if (!root->left)
429 root->left = rb_first(&root->rb);
430
Jens Axboe08717142008-01-28 11:38:15 +0100431 if (root->left)
432 return rb_entry(root->left, struct cfq_queue, rb_node);
433
434 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200435}
436
Jens Axboea36e71f2009-04-15 12:15:11 +0200437static void rb_erase_init(struct rb_node *n, struct rb_root *root)
438{
439 rb_erase(n, root);
440 RB_CLEAR_NODE(n);
441}
442
Jens Axboecc09e292007-04-26 12:53:50 +0200443static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
444{
445 if (root->left == n)
446 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200447 rb_erase_init(n, &root->rb);
Jens Axboecc09e292007-04-26 12:53:50 +0200448}
449
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450/*
451 * would be nice to take fifo expire time into account as well
452 */
Jens Axboe5e705372006-07-13 12:39:25 +0200453static struct request *
454cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
455 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456{
Jens Axboe21183b02006-07-13 12:33:14 +0200457 struct rb_node *rbnext = rb_next(&last->rb_node);
458 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200459 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460
Jens Axboe21183b02006-07-13 12:33:14 +0200461 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462
463 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200464 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200465
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200467 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200468 else {
469 rbnext = rb_first(&cfqq->sort_list);
470 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200471 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200472 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473
Jens Axboe21183b02006-07-13 12:33:14 +0200474 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475}
476
Jens Axboed9e76202007-04-20 14:27:50 +0200477static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
478 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479{
Jens Axboed9e76202007-04-20 14:27:50 +0200480 /*
481 * just an approximation, should be ok.
482 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200483 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
484 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200485}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486
Jens Axboe498d3aa22007-04-26 12:54:48 +0200487/*
488 * The cfqd->service_tree holds all pending cfq_queue's that have
489 * requests waiting to be processed. It is sorted in the order that
490 * we will service the queues.
491 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200492static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +0200493 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200494{
Jens Axboe08717142008-01-28 11:38:15 +0100495 struct rb_node **p, *parent;
496 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +0200497 unsigned long rb_key;
Jens Axboe498d3aa22007-04-26 12:54:48 +0200498 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200499
Jens Axboe08717142008-01-28 11:38:15 +0100500 if (cfq_class_idle(cfqq)) {
501 rb_key = CFQ_IDLE_DELAY;
502 parent = rb_last(&cfqd->service_tree.rb);
503 if (parent && parent != &cfqq->rb_node) {
504 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
505 rb_key += __cfqq->rb_key;
506 } else
507 rb_key += jiffies;
508 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +0200509 /*
510 * Get our rb key offset. Subtract any residual slice
511 * value carried from last service. A negative resid
512 * count indicates slice overrun, and this should position
513 * the next service time further away in the tree.
514 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200515 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +0200516 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +0200517 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +0200518 } else {
519 rb_key = -HZ;
520 __cfqq = cfq_rb_first(&cfqd->service_tree);
521 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
522 }
Jens Axboed9e76202007-04-20 14:27:50 +0200523
524 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100525 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200526 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100527 */
Jens Axboed9e76202007-04-20 14:27:50 +0200528 if (rb_key == cfqq->rb_key)
529 return;
Jens Axboe53b037442006-07-28 09:48:51 +0200530
Jens Axboecc09e292007-04-26 12:53:50 +0200531 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200532 }
Jens Axboed9e76202007-04-20 14:27:50 +0200533
Jens Axboe498d3aa22007-04-26 12:54:48 +0200534 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +0100535 parent = NULL;
536 p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200537 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +0200538 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200539
Jens Axboed9e76202007-04-20 14:27:50 +0200540 parent = *p;
541 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
542
Jens Axboe0c534e02007-04-18 20:01:57 +0200543 /*
544 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200545 * preference to them. IDLE queues goes to the back.
546 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200547 */
548 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200549 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200550 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200551 n = &(*p)->rb_right;
552 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
553 n = &(*p)->rb_left;
554 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
555 n = &(*p)->rb_right;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +0200556 else if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +0200557 n = &(*p)->rb_left;
558 else
559 n = &(*p)->rb_right;
560
561 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200562 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200563
564 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200565 }
566
Jens Axboecc09e292007-04-26 12:53:50 +0200567 if (left)
568 cfqd->service_tree.left = &cfqq->rb_node;
569
Jens Axboed9e76202007-04-20 14:27:50 +0200570 cfqq->rb_key = rb_key;
571 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200572 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573}
574
Jens Axboea36e71f2009-04-15 12:15:11 +0200575static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200576cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
577 sector_t sector, struct rb_node **ret_parent,
578 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +0200579{
Jens Axboea36e71f2009-04-15 12:15:11 +0200580 struct rb_node **p, *parent;
581 struct cfq_queue *cfqq = NULL;
582
583 parent = NULL;
584 p = &root->rb_node;
585 while (*p) {
586 struct rb_node **n;
587
588 parent = *p;
589 cfqq = rb_entry(parent, struct cfq_queue, p_node);
590
591 /*
592 * Sort strictly based on sector. Smallest to the left,
593 * largest to the right.
594 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900595 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +0200596 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900597 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +0200598 n = &(*p)->rb_left;
599 else
600 break;
601 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200602 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200603 }
604
605 *ret_parent = parent;
606 if (rb_link)
607 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200608 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +0200609}
610
611static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
612{
Jens Axboea36e71f2009-04-15 12:15:11 +0200613 struct rb_node **p, *parent;
614 struct cfq_queue *__cfqq;
615
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200616 if (cfqq->p_root) {
617 rb_erase(&cfqq->p_node, cfqq->p_root);
618 cfqq->p_root = NULL;
619 }
Jens Axboea36e71f2009-04-15 12:15:11 +0200620
621 if (cfq_class_idle(cfqq))
622 return;
623 if (!cfqq->next_rq)
624 return;
625
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200626 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900627 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
628 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +0200629 if (!__cfqq) {
630 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200631 rb_insert_color(&cfqq->p_node, cfqq->p_root);
632 } else
633 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200634}
635
Jens Axboe498d3aa22007-04-26 12:54:48 +0200636/*
637 * Update cfqq's position in the service tree.
638 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200639static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200640{
Jens Axboe6d048f52007-04-25 12:44:27 +0200641 /*
642 * Resorting requires the cfqq to be on the RR list already.
643 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200644 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200645 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +0200646 cfq_prio_tree_add(cfqd, cfqq);
647 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200648}
649
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650/*
651 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200652 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 */
Jens Axboefebffd62008-01-28 13:19:43 +0100654static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655{
Jens Axboe7b679132008-05-30 12:23:07 +0200656 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200657 BUG_ON(cfq_cfqq_on_rr(cfqq));
658 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 cfqd->busy_queues++;
660
Jens Axboeedd75ff2007-04-19 12:03:34 +0200661 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662}
663
Jens Axboe498d3aa22007-04-26 12:54:48 +0200664/*
665 * Called when the cfqq no longer has requests pending, remove it from
666 * the service tree.
667 */
Jens Axboefebffd62008-01-28 13:19:43 +0100668static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669{
Jens Axboe7b679132008-05-30 12:23:07 +0200670 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200671 BUG_ON(!cfq_cfqq_on_rr(cfqq));
672 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673
Jens Axboecc09e292007-04-26 12:53:50 +0200674 if (!RB_EMPTY_NODE(&cfqq->rb_node))
675 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200676 if (cfqq->p_root) {
677 rb_erase(&cfqq->p_node, cfqq->p_root);
678 cfqq->p_root = NULL;
679 }
Jens Axboed9e76202007-04-20 14:27:50 +0200680
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 BUG_ON(!cfqd->busy_queues);
682 cfqd->busy_queues--;
683}
684
685/*
686 * rb tree support functions
687 */
Jens Axboefebffd62008-01-28 13:19:43 +0100688static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689{
Jens Axboe5e705372006-07-13 12:39:25 +0200690 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200691 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200692 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693
Jens Axboeb4878f22005-10-20 16:42:29 +0200694 BUG_ON(!cfqq->queued[sync]);
695 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696
Jens Axboe5e705372006-07-13 12:39:25 +0200697 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698
Jens Axboedd67d052006-06-21 09:36:18 +0200699 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200700 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701}
702
Jens Axboe5e705372006-07-13 12:39:25 +0200703static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704{
Jens Axboe5e705372006-07-13 12:39:25 +0200705 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +0200707 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708
Jens Axboe5380a102006-07-13 12:37:56 +0200709 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710
711 /*
712 * looks a little odd, but the first insert might return an alias.
713 * if that happens, put the alias on the dispatch list
714 */
Jens Axboe21183b02006-07-13 12:33:14 +0200715 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200716 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100717
718 if (!cfq_cfqq_on_rr(cfqq))
719 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200720
721 /*
722 * check if this request is a better next-serve candidate
723 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200724 prev = cfqq->next_rq;
Jens Axboe5044eed2007-04-25 11:53:48 +0200725 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
Jens Axboea36e71f2009-04-15 12:15:11 +0200726
727 /*
728 * adjust priority tree position, if ->next_rq changes
729 */
730 if (prev != cfqq->next_rq)
731 cfq_prio_tree_add(cfqd, cfqq);
732
Jens Axboe5044eed2007-04-25 11:53:48 +0200733 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734}
735
Jens Axboefebffd62008-01-28 13:19:43 +0100736static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737{
Jens Axboe5380a102006-07-13 12:37:56 +0200738 elv_rb_del(&cfqq->sort_list, rq);
739 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200740 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741}
742
Jens Axboe206dc692006-03-28 13:03:44 +0200743static struct request *
744cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745{
Jens Axboe206dc692006-03-28 13:03:44 +0200746 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200747 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200748 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749
Jens Axboe4ac845a2008-01-24 08:44:49 +0100750 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200751 if (!cic)
752 return NULL;
753
754 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200755 if (cfqq) {
756 sector_t sector = bio->bi_sector + bio_sectors(bio);
757
Jens Axboe21183b02006-07-13 12:33:14 +0200758 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200759 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 return NULL;
762}
763
Jens Axboe165125e2007-07-24 09:28:11 +0200764static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200765{
766 struct cfq_data *cfqd = q->elevator->elevator_data;
767
Jens Axboe5ad531d2009-07-03 12:57:48 +0200768 cfqd->rq_in_driver[rq_is_sync(rq)]++;
Jens Axboe7b679132008-05-30 12:23:07 +0200769 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Jens Axboe5ad531d2009-07-03 12:57:48 +0200770 rq_in_driver(cfqd));
Jens Axboe25776e32006-06-01 10:12:26 +0200771
Tejun Heo5b936292009-05-07 22:24:38 +0900772 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200773}
774
Jens Axboe165125e2007-07-24 09:28:11 +0200775static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776{
Jens Axboe22e2c502005-06-27 10:55:12 +0200777 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5ad531d2009-07-03 12:57:48 +0200778 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779
Jens Axboe5ad531d2009-07-03 12:57:48 +0200780 WARN_ON(!cfqd->rq_in_driver[sync]);
781 cfqd->rq_in_driver[sync]--;
Jens Axboe7b679132008-05-30 12:23:07 +0200782 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Jens Axboe5ad531d2009-07-03 12:57:48 +0200783 rq_in_driver(cfqd));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784}
785
Jens Axboeb4878f22005-10-20 16:42:29 +0200786static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787{
Jens Axboe5e705372006-07-13 12:39:25 +0200788 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200789
Jens Axboe5e705372006-07-13 12:39:25 +0200790 if (cfqq->next_rq == rq)
791 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792
Jens Axboeb4878f22005-10-20 16:42:29 +0200793 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200794 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200795
Aaron Carroll45333d52008-08-26 15:52:36 +0200796 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +0200797 if (rq_is_meta(rq)) {
798 WARN_ON(!cfqq->meta_pending);
799 cfqq->meta_pending--;
800 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801}
802
Jens Axboe165125e2007-07-24 09:28:11 +0200803static int cfq_merge(struct request_queue *q, struct request **req,
804 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805{
806 struct cfq_data *cfqd = q->elevator->elevator_data;
807 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808
Jens Axboe206dc692006-03-28 13:03:44 +0200809 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200810 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200811 *req = __rq;
812 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 }
814
815 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816}
817
Jens Axboe165125e2007-07-24 09:28:11 +0200818static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200819 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820{
Jens Axboe21183b02006-07-13 12:33:14 +0200821 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200822 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823
Jens Axboe5e705372006-07-13 12:39:25 +0200824 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826}
827
828static void
Jens Axboe165125e2007-07-24 09:28:11 +0200829cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 struct request *next)
831{
Jens Axboe22e2c502005-06-27 10:55:12 +0200832 /*
833 * reposition in fifo if next is older than rq
834 */
835 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +0200836 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +0200837 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +0200838 rq_set_fifo_time(rq, rq_fifo_time(next));
839 }
Jens Axboe22e2c502005-06-27 10:55:12 +0200840
Jens Axboeb4878f22005-10-20 16:42:29 +0200841 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200842}
843
Jens Axboe165125e2007-07-24 09:28:11 +0200844static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100845 struct bio *bio)
846{
847 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200848 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100849 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100850
851 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100852 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100853 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200854 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +0200855 return false;
Jens Axboeda775262006-12-20 11:04:12 +0100856
857 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100858 * Lookup the cfqq that this bio will be queued with. Allow
859 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100860 */
Jens Axboe4ac845a2008-01-24 08:44:49 +0100861 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200862 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +0200863 return false;
Jens Axboe719d3402006-12-22 09:38:53 +0100864
Vasily Tarasov91fac312007-04-25 12:29:51 +0200865 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +0200866 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +0100867}
868
Jens Axboefebffd62008-01-28 13:19:43 +0100869static void __cfq_set_active_queue(struct cfq_data *cfqd,
870 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +0200871{
872 if (cfqq) {
Jens Axboe7b679132008-05-30 12:23:07 +0200873 cfq_log_cfqq(cfqd, cfqq, "set_active");
Jens Axboe22e2c502005-06-27 10:55:12 +0200874 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +0200875 cfqq->slice_dispatch = 0;
876
Jens Axboe2f5cb732009-04-07 08:51:19 +0200877 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +0200878 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200879 cfq_clear_cfqq_must_alloc_slice(cfqq);
880 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100881 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +0200882
883 del_timer(&cfqd->idle_slice_timer);
Jens Axboe22e2c502005-06-27 10:55:12 +0200884 }
885
886 cfqd->active_queue = cfqq;
887}
888
889/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100890 * current cfqq expired its slice (or was too idle), select new one
891 */
892static void
893__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +0200894 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100895{
Jens Axboe7b679132008-05-30 12:23:07 +0200896 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
897
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100898 if (cfq_cfqq_wait_request(cfqq))
899 del_timer(&cfqd->idle_slice_timer);
900
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100901 cfq_clear_cfqq_wait_request(cfqq);
902
903 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200904 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100905 */
Jens Axboe7b679132008-05-30 12:23:07 +0200906 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +1100907 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200908 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
909 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100910
Jens Axboeedd75ff2007-04-19 12:03:34 +0200911 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100912
913 if (cfqq == cfqd->active_queue)
914 cfqd->active_queue = NULL;
915
916 if (cfqd->active_cic) {
917 put_io_context(cfqd->active_cic->ioc);
918 cfqd->active_cic = NULL;
919 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100920}
921
Jens Axboea6151c32009-10-07 20:02:57 +0200922static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100923{
924 struct cfq_queue *cfqq = cfqd->active_queue;
925
926 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200927 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100928}
929
Jens Axboe498d3aa22007-04-26 12:54:48 +0200930/*
931 * Get next queue for service. Unless we have a queue preemption,
932 * we'll simply select the first cfqq in the service tree.
933 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200934static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200935{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200936 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
937 return NULL;
938
Jens Axboe08717142008-01-28 11:38:15 +0100939 return cfq_rb_first(&cfqd->service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +0200940}
941
Jens Axboe498d3aa22007-04-26 12:54:48 +0200942/*
943 * Get and set a new active queue for service.
944 */
Jens Axboea36e71f2009-04-15 12:15:11 +0200945static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
946 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200947{
Jens Axboea36e71f2009-04-15 12:15:11 +0200948 if (!cfqq) {
949 cfqq = cfq_get_next_queue(cfqd);
Shaohua Li4b27e1b2009-11-03 20:25:02 +0100950 if (cfqq && !cfq_cfqq_coop_preempt(cfqq))
Jens Axboea36e71f2009-04-15 12:15:11 +0200951 cfq_clear_cfqq_coop(cfqq);
952 }
Jens Axboe6d048f52007-04-25 12:44:27 +0200953
Shaohua Li4b27e1b2009-11-03 20:25:02 +0100954 if (cfqq)
955 cfq_clear_cfqq_coop_preempt(cfqq);
956
Jens Axboe22e2c502005-06-27 10:55:12 +0200957 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200958 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200959}
960
Jens Axboed9e76202007-04-20 14:27:50 +0200961static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
962 struct request *rq)
963{
Tejun Heo83096eb2009-05-07 22:24:39 +0900964 if (blk_rq_pos(rq) >= cfqd->last_position)
965 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +0200966 else
Tejun Heo83096eb2009-05-07 22:24:39 +0900967 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +0200968}
969
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200970#define CIC_SEEK_THR 8 * 1024
971#define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
972
Jens Axboe6d048f52007-04-25 12:44:27 +0200973static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
974{
975 struct cfq_io_context *cic = cfqd->active_cic;
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200976 sector_t sdist = cic->seek_mean;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200977
Jens Axboe6d048f52007-04-25 12:44:27 +0200978 if (!sample_valid(cic->seek_samples))
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200979 sdist = CIC_SEEK_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +0200980
Jeff Moyer04dc6e72009-04-21 07:31:56 +0200981 return cfq_dist_from_last(cfqd, rq) <= sdist;
Jens Axboe6d048f52007-04-25 12:44:27 +0200982}
983
Jens Axboea36e71f2009-04-15 12:15:11 +0200984static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
985 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200986{
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200987 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +0200988 struct rb_node *parent, *node;
989 struct cfq_queue *__cfqq;
990 sector_t sector = cfqd->last_position;
991
992 if (RB_EMPTY_ROOT(root))
993 return NULL;
994
995 /*
996 * First, if we find a request starting at the end of the last
997 * request, choose it.
998 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +0200999 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001000 if (__cfqq)
1001 return __cfqq;
1002
1003 /*
1004 * If the exact sector wasn't found, the parent of the NULL leaf
1005 * will contain the closest sector.
1006 */
1007 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1008 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1009 return __cfqq;
1010
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001011 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001012 node = rb_next(&__cfqq->p_node);
1013 else
1014 node = rb_prev(&__cfqq->p_node);
1015 if (!node)
1016 return NULL;
1017
1018 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1019 if (cfq_rq_close(cfqd, __cfqq->next_rq))
1020 return __cfqq;
1021
1022 return NULL;
1023}
1024
1025/*
1026 * cfqd - obvious
1027 * cur_cfqq - passed in so that we don't decide that the current queue is
1028 * closely cooperating with itself.
1029 *
1030 * So, basically we're assuming that that cur_cfqq has dispatched at least
1031 * one request, and that cfqd->last_position reflects a position on the disk
1032 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1033 * assumption.
1034 */
1035static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1036 struct cfq_queue *cur_cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001037 bool probe)
Jens Axboea36e71f2009-04-15 12:15:11 +02001038{
1039 struct cfq_queue *cfqq;
1040
1041 /*
1042 * A valid cfq_io_context is necessary to compare requests against
1043 * the seek_mean of the current cfqq.
1044 */
1045 if (!cfqd->active_cic)
1046 return NULL;
1047
Jens Axboe6d048f52007-04-25 12:44:27 +02001048 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001049 * We should notice if some of the queues are cooperating, eg
1050 * working closely on the same area of the disk. In that case,
1051 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001052 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001053 cfqq = cfqq_close(cfqd, cur_cfqq);
1054 if (!cfqq)
1055 return NULL;
1056
1057 if (cfq_cfqq_coop(cfqq))
1058 return NULL;
1059
1060 if (!probe)
1061 cfq_mark_cfqq_coop(cfqq);
1062 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001063}
1064
Jens Axboe6d048f52007-04-25 12:44:27 +02001065static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001066{
Jens Axboe17926692007-01-19 11:59:30 +11001067 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001068 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001069 unsigned long sl;
1070
Jens Axboea68bbdd2008-09-24 13:03:33 +02001071 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001072 * SSD device without seek penalty, disable idling. But only do so
1073 * for devices that support queuing, otherwise we still have a problem
1074 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001075 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001076 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001077 return;
1078
Jens Axboedd67d052006-06-21 09:36:18 +02001079 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001080 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001081
1082 /*
1083 * idle is disabled, either manually or by past process history
1084 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001085 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1086 return;
1087
Jens Axboe22e2c502005-06-27 10:55:12 +02001088 /*
Jens Axboe7b679132008-05-30 12:23:07 +02001089 * still requests with the driver, don't idle
1090 */
Jens Axboe5ad531d2009-07-03 12:57:48 +02001091 if (rq_in_driver(cfqd))
Jens Axboe7b679132008-05-30 12:23:07 +02001092 return;
1093
1094 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001095 * task has exited, don't wait
1096 */
Jens Axboe206dc692006-03-28 13:03:44 +02001097 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001098 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001099 return;
1100
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001101 /*
1102 * If our average think time is larger than the remaining time
1103 * slice, then don't idle. This avoids overrunning the allotted
1104 * time slice.
1105 */
1106 if (sample_valid(cic->ttime_samples) &&
1107 (cfqq->slice_end - jiffies < cic->ttime_mean))
1108 return;
1109
Jens Axboe3b181522005-06-27 10:56:24 +02001110 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001111
Jens Axboe206dc692006-03-28 13:03:44 +02001112 /*
1113 * we don't want to idle for seeks, but we do want to allow
1114 * fair distribution of slice time for a process doing back-to-back
1115 * seeks. so allow a little bit of time for him to submit a new rq
1116 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001117 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001118 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +02001119 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +02001120
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001121 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe9481ffd2009-04-15 12:14:13 +02001122 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123}
1124
Jens Axboe498d3aa22007-04-26 12:54:48 +02001125/*
1126 * Move request from internal lists to the request queue dispatch list.
1127 */
Jens Axboe165125e2007-07-24 09:28:11 +02001128static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001130 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001131 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001132
Jens Axboe7b679132008-05-30 12:23:07 +02001133 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1134
Jeff Moyer06d21882009-09-11 17:08:59 +02001135 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001136 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001137 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001138 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001139
1140 if (cfq_cfqq_sync(cfqq))
1141 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142}
1143
1144/*
1145 * return expired entry, or NULL to just start from scratch in rbtree
1146 */
Jens Axboefebffd62008-01-28 13:19:43 +01001147static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148{
Jens Axboe30996f42009-10-05 11:03:39 +02001149 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150
Jens Axboe3b181522005-06-27 10:56:24 +02001151 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001153
1154 cfq_mark_cfqq_fifo_expire(cfqq);
1155
Jens Axboe89850f72006-07-22 16:48:31 +02001156 if (list_empty(&cfqq->fifo))
1157 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158
Jens Axboe89850f72006-07-22 16:48:31 +02001159 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02001160 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02001161 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162
Jens Axboe30996f42009-10-05 11:03:39 +02001163 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001164 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165}
1166
Jens Axboe22e2c502005-06-27 10:55:12 +02001167static inline int
1168cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1169{
1170 const int base_rq = cfqd->cfq_slice_async_rq;
1171
1172 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1173
1174 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1175}
1176
1177/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02001178 * Select a queue for service. If we have a current active queue,
1179 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02001180 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001181static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001182{
Jens Axboea36e71f2009-04-15 12:15:11 +02001183 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001184
1185 cfqq = cfqd->active_queue;
1186 if (!cfqq)
1187 goto new_queue;
1188
1189 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001190 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02001191 */
Jens Axboeb0291952009-04-07 11:38:31 +02001192 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001193 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +02001194
1195 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001196 * The active queue has requests and isn't expired, allow it to
1197 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02001198 */
Jens Axboedd67d052006-06-21 09:36:18 +02001199 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001200 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02001201
1202 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02001203 * If another queue has a request waiting within our mean seek
1204 * distance, let it run. The expire code will check for close
1205 * cooperators and put the close queue at the front of the service
1206 * tree.
1207 */
1208 new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1209 if (new_cfqq)
1210 goto expire;
1211
1212 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001213 * No requests pending. If the active queue still has requests in
1214 * flight or is idling for a new request, allow either of these
1215 * conditions to happen (or time out) before selecting a new queue.
1216 */
Jens Axboecc197472007-04-20 20:45:39 +02001217 if (timer_pending(&cfqd->idle_slice_timer) ||
1218 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001219 cfqq = NULL;
1220 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001221 }
1222
Jens Axboe3b181522005-06-27 10:56:24 +02001223expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001224 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001225new_queue:
Jens Axboea36e71f2009-04-15 12:15:11 +02001226 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001227keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001228 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001229}
1230
Jens Axboefebffd62008-01-28 13:19:43 +01001231static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02001232{
1233 int dispatched = 0;
1234
1235 while (cfqq->next_rq) {
1236 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1237 dispatched++;
1238 }
1239
1240 BUG_ON(!list_empty(&cfqq->fifo));
1241 return dispatched;
1242}
1243
Jens Axboe498d3aa22007-04-26 12:54:48 +02001244/*
1245 * Drain our current requests. Used for barriers and when switching
1246 * io schedulers on-the-fly.
1247 */
Jens Axboed9e76202007-04-20 14:27:50 +02001248static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001249{
Jens Axboe08717142008-01-28 11:38:15 +01001250 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001251 int dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001252
Jens Axboe08717142008-01-28 11:38:15 +01001253 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
Jens Axboed9e76202007-04-20 14:27:50 +02001254 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001255
Jens Axboe6084cdd2007-04-23 08:25:00 +02001256 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001257
1258 BUG_ON(cfqd->busy_queues);
1259
Jeff Moyer69237152009-06-12 15:29:30 +02001260 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001261 return dispatched;
1262}
1263
Jens Axboe0b182d62009-10-06 20:49:37 +02001264static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02001265{
Jens Axboe2f5cb732009-04-07 08:51:19 +02001266 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267
Jens Axboe2f5cb732009-04-07 08:51:19 +02001268 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02001269 * Drain async requests before we start sync IO
1270 */
1271 if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02001272 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02001273
1274 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02001275 * If this is an async queue and we have sync IO in flight, let it wait
1276 */
1277 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02001278 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001279
1280 max_dispatch = cfqd->cfq_quantum;
1281 if (cfq_class_idle(cfqq))
1282 max_dispatch = 1;
1283
1284 /*
1285 * Does this cfqq already have too much IO in flight?
1286 */
1287 if (cfqq->dispatched >= max_dispatch) {
1288 /*
1289 * idle queue must always only have a single IO in flight
1290 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02001291 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02001292 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001293
Jens Axboe2f5cb732009-04-07 08:51:19 +02001294 /*
1295 * We have other queues, don't allow more IO from this one
1296 */
1297 if (cfqd->busy_queues > 1)
Jens Axboe0b182d62009-10-06 20:49:37 +02001298 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11001299
Jens Axboe2f5cb732009-04-07 08:51:19 +02001300 /*
Jens Axboe8e296752009-10-03 16:26:03 +02001301 * Sole queue user, allow bigger slice
Vivek Goyal365722b2009-10-03 15:21:27 +02001302 */
Jens Axboe8e296752009-10-03 16:26:03 +02001303 max_dispatch *= 4;
1304 }
1305
1306 /*
1307 * Async queues must wait a bit before being allowed dispatch.
1308 * We also ramp up the dispatch depth gradually for async IO,
1309 * based on the last sync IO we serviced
1310 */
Jens Axboe963b72f2009-10-03 19:42:18 +02001311 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Jens Axboe8e296752009-10-03 16:26:03 +02001312 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1313 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02001314
Jens Axboe61f0c1d2009-10-03 19:46:03 +02001315 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02001316 if (!depth && !cfqq->dispatched)
1317 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02001318 if (depth < max_dispatch)
1319 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 }
1321
Jens Axboe0b182d62009-10-06 20:49:37 +02001322 /*
1323 * If we're below the current max, allow a dispatch
1324 */
1325 return cfqq->dispatched < max_dispatch;
1326}
1327
1328/*
1329 * Dispatch a request from cfqq, moving them to the request queue
1330 * dispatch list.
1331 */
1332static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1333{
1334 struct request *rq;
1335
1336 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1337
1338 if (!cfq_may_dispatch(cfqd, cfqq))
1339 return false;
1340
1341 /*
1342 * follow expired path, else get first next available
1343 */
1344 rq = cfq_check_fifo(cfqq);
1345 if (!rq)
1346 rq = cfqq->next_rq;
1347
1348 /*
1349 * insert request into driver dispatch list
1350 */
1351 cfq_dispatch_insert(cfqd->queue, rq);
1352
1353 if (!cfqd->active_cic) {
1354 struct cfq_io_context *cic = RQ_CIC(rq);
1355
1356 atomic_long_inc(&cic->ioc->refcount);
1357 cfqd->active_cic = cic;
1358 }
1359
1360 return true;
1361}
1362
1363/*
1364 * Find the cfqq that we need to service and move a request from that to the
1365 * dispatch list
1366 */
1367static int cfq_dispatch_requests(struct request_queue *q, int force)
1368{
1369 struct cfq_data *cfqd = q->elevator->elevator_data;
1370 struct cfq_queue *cfqq;
1371
1372 if (!cfqd->busy_queues)
1373 return 0;
1374
1375 if (unlikely(force))
1376 return cfq_forced_dispatch(cfqd);
1377
1378 cfqq = cfq_select_queue(cfqd);
1379 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02001380 return 0;
1381
Jens Axboe2f5cb732009-04-07 08:51:19 +02001382 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02001383 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02001384 */
Jens Axboe0b182d62009-10-06 20:49:37 +02001385 if (!cfq_dispatch_request(cfqd, cfqq))
1386 return 0;
1387
Jens Axboe2f5cb732009-04-07 08:51:19 +02001388 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02001389 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001390
1391 /*
1392 * expire an async queue immediately if it has used up its slice. idle
1393 * queue always expire after 1 dispatch round.
1394 */
1395 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1396 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1397 cfq_class_idle(cfqq))) {
1398 cfqq->slice_end = jiffies + 1;
1399 cfq_slice_expired(cfqd, 0);
1400 }
1401
Shan Weib217a902009-09-01 10:06:42 +02001402 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02001403 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404}
1405
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406/*
Jens Axboe5e705372006-07-13 12:39:25 +02001407 * task holds one reference to the queue, dropped when task exits. each rq
1408 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 *
1410 * queue lock must be held here.
1411 */
1412static void cfq_put_queue(struct cfq_queue *cfqq)
1413{
Jens Axboe22e2c502005-06-27 10:55:12 +02001414 struct cfq_data *cfqd = cfqq->cfqd;
1415
1416 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417
1418 if (!atomic_dec_and_test(&cfqq->ref))
1419 return;
1420
Jens Axboe7b679132008-05-30 12:23:07 +02001421 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001423 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001424 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001426 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001427 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02001428 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001429 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001430
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 kmem_cache_free(cfq_pool, cfqq);
1432}
1433
Jens Axboed6de8be2008-05-28 14:46:59 +02001434/*
1435 * Must always be called with the rcu_read_lock() held
1436 */
Jens Axboe07416d22008-05-07 09:17:12 +02001437static void
1438__call_for_each_cic(struct io_context *ioc,
1439 void (*func)(struct io_context *, struct cfq_io_context *))
1440{
1441 struct cfq_io_context *cic;
1442 struct hlist_node *n;
1443
1444 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1445 func(ioc, cic);
1446}
1447
Jens Axboe4ac845a2008-01-24 08:44:49 +01001448/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001449 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001450 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001451static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001452call_for_each_cic(struct io_context *ioc,
1453 void (*func)(struct io_context *, struct cfq_io_context *))
1454{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001455 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02001456 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001457 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001458}
Jens Axboe4ac845a2008-01-24 08:44:49 +01001459
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001460static void cfq_cic_free_rcu(struct rcu_head *head)
1461{
1462 struct cfq_io_context *cic;
1463
1464 cic = container_of(head, struct cfq_io_context, rcu_head);
1465
1466 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09001467 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001468
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001469 if (ioc_gone) {
1470 /*
1471 * CFQ scheduler is exiting, grab exit lock and check
1472 * the pending io context count. If it hits zero,
1473 * complete ioc_gone and set it back to NULL
1474 */
1475 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09001476 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001477 complete(ioc_gone);
1478 ioc_gone = NULL;
1479 }
1480 spin_unlock(&ioc_gone_lock);
1481 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001482}
1483
1484static void cfq_cic_free(struct cfq_io_context *cic)
1485{
1486 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001487}
1488
1489static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1490{
1491 unsigned long flags;
1492
1493 BUG_ON(!cic->dead_key);
1494
1495 spin_lock_irqsave(&ioc->lock, flags);
1496 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01001497 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001498 spin_unlock_irqrestore(&ioc->lock, flags);
1499
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001500 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001501}
1502
Jens Axboed6de8be2008-05-28 14:46:59 +02001503/*
1504 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1505 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1506 * and ->trim() which is called with the task lock held
1507 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001508static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001510 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001511 * ioc->refcount is zero here, or we are called from elv_unregister(),
1512 * so no more cic's are allowed to be linked into this ioc. So it
1513 * should be ok to iterate over the known list, we will see all cic's
1514 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001515 */
Jens Axboe07416d22008-05-07 09:17:12 +02001516 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517}
1518
Jens Axboe89850f72006-07-22 16:48:31 +02001519static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1520{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001521 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001522 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02001523 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001524 }
Jens Axboe89850f72006-07-22 16:48:31 +02001525
1526 cfq_put_queue(cfqq);
1527}
1528
1529static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1530 struct cfq_io_context *cic)
1531{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001532 struct io_context *ioc = cic->ioc;
1533
Jens Axboefc463792006-08-29 09:05:44 +02001534 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001535
1536 /*
1537 * Make sure key == NULL is seen for dead queues
1538 */
Jens Axboefc463792006-08-29 09:05:44 +02001539 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001540 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02001541 cic->key = NULL;
1542
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001543 if (ioc->ioc_data == cic)
1544 rcu_assign_pointer(ioc->ioc_data, NULL);
1545
Jens Axboeff6657c2009-04-08 10:58:57 +02001546 if (cic->cfqq[BLK_RW_ASYNC]) {
1547 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1548 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001549 }
1550
Jens Axboeff6657c2009-04-08 10:58:57 +02001551 if (cic->cfqq[BLK_RW_SYNC]) {
1552 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1553 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02001554 }
Jens Axboe89850f72006-07-22 16:48:31 +02001555}
1556
Jens Axboe4ac845a2008-01-24 08:44:49 +01001557static void cfq_exit_single_io_context(struct io_context *ioc,
1558 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001559{
Al Viro478a82b2006-03-18 13:25:24 -05001560 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001561
Jens Axboe89850f72006-07-22 16:48:31 +02001562 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001563 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001564 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02001565
Jens Axboe4ac845a2008-01-24 08:44:49 +01001566 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01001567
1568 /*
1569 * Ensure we get a fresh copy of the ->key to prevent
1570 * race between exiting task and queue
1571 */
1572 smp_read_barrier_depends();
1573 if (cic->key)
1574 __cfq_exit_single_io_context(cfqd, cic);
1575
Jens Axboe4ac845a2008-01-24 08:44:49 +01001576 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05001577 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001578}
1579
Jens Axboe498d3aa22007-04-26 12:54:48 +02001580/*
1581 * The process that ioc belongs to has exited, we need to clean up
1582 * and put the internal structures we have that belongs to that process.
1583 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001584static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001586 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587}
1588
Jens Axboe22e2c502005-06-27 10:55:12 +02001589static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001590cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591{
Jens Axboeb5deef92006-07-19 23:39:40 +02001592 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593
Christoph Lameter94f60302007-07-17 04:03:29 -07001594 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1595 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001597 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001598 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01001599 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001600 cic->dtor = cfq_free_io_context;
1601 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09001602 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 }
1604
1605 return cic;
1606}
1607
Jens Axboefd0928d2008-01-24 08:52:45 +01001608static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02001609{
1610 struct task_struct *tsk = current;
1611 int ioprio_class;
1612
Jens Axboe3b181522005-06-27 10:56:24 +02001613 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001614 return;
1615
Jens Axboefd0928d2008-01-24 08:52:45 +01001616 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001617 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01001618 default:
1619 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1620 case IOPRIO_CLASS_NONE:
1621 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02001622 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01001623 */
1624 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02001625 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01001626 break;
1627 case IOPRIO_CLASS_RT:
1628 cfqq->ioprio = task_ioprio(ioc);
1629 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1630 break;
1631 case IOPRIO_CLASS_BE:
1632 cfqq->ioprio = task_ioprio(ioc);
1633 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1634 break;
1635 case IOPRIO_CLASS_IDLE:
1636 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1637 cfqq->ioprio = 7;
1638 cfq_clear_cfqq_idle_window(cfqq);
1639 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02001640 }
1641
1642 /*
1643 * keep track of original prio settings in case we have to temporarily
1644 * elevate the priority of this queue
1645 */
1646 cfqq->org_ioprio = cfqq->ioprio;
1647 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001648 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001649}
1650
Jens Axboefebffd62008-01-28 13:19:43 +01001651static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001652{
Al Viro478a82b2006-03-18 13:25:24 -05001653 struct cfq_data *cfqd = cic->key;
1654 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001655 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001656
Jens Axboecaaa5f92006-06-16 11:23:00 +02001657 if (unlikely(!cfqd))
1658 return;
1659
Jens Axboec1b707d2006-10-30 19:54:23 +01001660 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001661
Jens Axboeff6657c2009-04-08 10:58:57 +02001662 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001663 if (cfqq) {
1664 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02001665 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1666 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001667 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02001668 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001669 cfq_put_queue(cfqq);
1670 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001671 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001672
Jens Axboeff6657c2009-04-08 10:58:57 +02001673 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02001674 if (cfqq)
1675 cfq_mark_cfqq_prio_changed(cfqq);
1676
Jens Axboec1b707d2006-10-30 19:54:23 +01001677 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001678}
1679
Jens Axboefc463792006-08-29 09:05:44 +02001680static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001682 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02001683 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684}
1685
Jens Axboed5036d72009-06-26 10:44:34 +02001686static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001687 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02001688{
1689 RB_CLEAR_NODE(&cfqq->rb_node);
1690 RB_CLEAR_NODE(&cfqq->p_node);
1691 INIT_LIST_HEAD(&cfqq->fifo);
1692
1693 atomic_set(&cfqq->ref, 0);
1694 cfqq->cfqd = cfqd;
1695
1696 cfq_mark_cfqq_prio_changed(cfqq);
1697
1698 if (is_sync) {
1699 if (!cfq_class_idle(cfqq))
1700 cfq_mark_cfqq_idle_window(cfqq);
1701 cfq_mark_cfqq_sync(cfqq);
1702 }
1703 cfqq->pid = pid;
1704}
1705
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02001707cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01001708 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001711 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712
1713retry:
Jens Axboe4ac845a2008-01-24 08:44:49 +01001714 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001715 /* cic always exists here */
1716 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717
Jens Axboe6118b702009-06-30 09:34:12 +02001718 /*
1719 * Always try a new alloc if we fell back to the OOM cfqq
1720 * originally, since it should just be a temporary situation.
1721 */
1722 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1723 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 if (new_cfqq) {
1725 cfqq = new_cfqq;
1726 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001727 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001729 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02001730 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07001731 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02001733 if (new_cfqq)
1734 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001735 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001736 cfqq = kmem_cache_alloc_node(cfq_pool,
1737 gfp_mask | __GFP_ZERO,
1738 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001739 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740
Jens Axboe6118b702009-06-30 09:34:12 +02001741 if (cfqq) {
1742 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1743 cfq_init_prio_data(cfqq, ioc);
1744 cfq_log_cfqq(cfqd, cfqq, "alloced");
1745 } else
1746 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 }
1748
1749 if (new_cfqq)
1750 kmem_cache_free(cfq_pool, new_cfqq);
1751
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 return cfqq;
1753}
1754
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001755static struct cfq_queue **
1756cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1757{
Jens Axboefe094d92008-01-31 13:08:54 +01001758 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001759 case IOPRIO_CLASS_RT:
1760 return &cfqd->async_cfqq[0][ioprio];
1761 case IOPRIO_CLASS_BE:
1762 return &cfqd->async_cfqq[1][ioprio];
1763 case IOPRIO_CLASS_IDLE:
1764 return &cfqd->async_idle_cfqq;
1765 default:
1766 BUG();
1767 }
1768}
1769
Jens Axboe15c31be2007-07-10 13:43:25 +02001770static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02001771cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02001772 gfp_t gfp_mask)
1773{
Jens Axboefd0928d2008-01-24 08:52:45 +01001774 const int ioprio = task_ioprio(ioc);
1775 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001776 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001777 struct cfq_queue *cfqq = NULL;
1778
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001779 if (!is_sync) {
1780 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1781 cfqq = *async_cfqq;
1782 }
1783
Jens Axboe6118b702009-06-30 09:34:12 +02001784 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01001785 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02001786
1787 /*
1788 * pin the queue now that it's allocated, scheduler exit will prune it
1789 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001790 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001791 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001792 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001793 }
1794
1795 atomic_inc(&cfqq->ref);
1796 return cfqq;
1797}
1798
Jens Axboe498d3aa22007-04-26 12:54:48 +02001799/*
1800 * We drop cfq io contexts lazily, so we may find a dead one.
1801 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001802static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001803cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1804 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001805{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001806 unsigned long flags;
1807
Jens Axboefc463792006-08-29 09:05:44 +02001808 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001809
Jens Axboe4ac845a2008-01-24 08:44:49 +01001810 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02001811
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001812 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001813
1814 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01001815 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001816 spin_unlock_irqrestore(&ioc->lock, flags);
1817
1818 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001819}
1820
Jens Axboee2d74ac2006-03-28 08:59:01 +02001821static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01001822cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001823{
Jens Axboee2d74ac2006-03-28 08:59:01 +02001824 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001825 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001826 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001827
Vasily Tarasov91fac312007-04-25 12:29:51 +02001828 if (unlikely(!ioc))
1829 return NULL;
1830
Jens Axboed6de8be2008-05-28 14:46:59 +02001831 rcu_read_lock();
1832
Jens Axboe597bc482007-04-24 21:23:53 +02001833 /*
1834 * we maintain a last-hit cache, to avoid browsing over the tree
1835 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001836 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02001837 if (cic && cic->key == cfqd) {
1838 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02001839 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001840 }
Jens Axboe597bc482007-04-24 21:23:53 +02001841
Jens Axboe4ac845a2008-01-24 08:44:49 +01001842 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001843 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1844 rcu_read_unlock();
1845 if (!cic)
1846 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001847 /* ->key must be copied to avoid race with cfq_exit_queue() */
1848 k = cic->key;
1849 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001850 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001851 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001852 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001853 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001854
Jens Axboed6de8be2008-05-28 14:46:59 +02001855 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001856 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001857 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001858 break;
1859 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001860
Jens Axboe4ac845a2008-01-24 08:44:49 +01001861 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001862}
1863
Jens Axboe4ac845a2008-01-24 08:44:49 +01001864/*
1865 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1866 * the process specific cfq io context when entered from the block layer.
1867 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1868 */
Jens Axboefebffd62008-01-28 13:19:43 +01001869static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1870 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001871{
Jens Axboe0261d682006-10-30 19:07:48 +01001872 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001873 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001874
Jens Axboe4ac845a2008-01-24 08:44:49 +01001875 ret = radix_tree_preload(gfp_mask);
1876 if (!ret) {
1877 cic->ioc = ioc;
1878 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001879
Jens Axboe4ac845a2008-01-24 08:44:49 +01001880 spin_lock_irqsave(&ioc->lock, flags);
1881 ret = radix_tree_insert(&ioc->radix_root,
1882 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01001883 if (!ret)
1884 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001885 spin_unlock_irqrestore(&ioc->lock, flags);
1886
1887 radix_tree_preload_end();
1888
1889 if (!ret) {
1890 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1891 list_add(&cic->queue_list, &cfqd->cic_list);
1892 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001893 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001894 }
1895
Jens Axboe4ac845a2008-01-24 08:44:49 +01001896 if (ret)
1897 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02001898
Jens Axboe4ac845a2008-01-24 08:44:49 +01001899 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001900}
1901
Jens Axboe22e2c502005-06-27 10:55:12 +02001902/*
1903 * Setup general io context and cfq io context. There can be several cfq
1904 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001905 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001906 */
1907static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001908cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909{
Jens Axboe22e2c502005-06-27 10:55:12 +02001910 struct io_context *ioc = NULL;
1911 struct cfq_io_context *cic;
1912
1913 might_sleep_if(gfp_mask & __GFP_WAIT);
1914
Jens Axboeb5deef92006-07-19 23:39:40 +02001915 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001916 if (!ioc)
1917 return NULL;
1918
Jens Axboe4ac845a2008-01-24 08:44:49 +01001919 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001920 if (cic)
1921 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001922
Jens Axboee2d74ac2006-03-28 08:59:01 +02001923 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1924 if (cic == NULL)
1925 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001926
Jens Axboe4ac845a2008-01-24 08:44:49 +01001927 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1928 goto err_free;
1929
Jens Axboe22e2c502005-06-27 10:55:12 +02001930out:
Jens Axboefc463792006-08-29 09:05:44 +02001931 smp_read_barrier_depends();
1932 if (unlikely(ioc->ioprio_changed))
1933 cfq_ioc_set_ioprio(ioc);
1934
Jens Axboe22e2c502005-06-27 10:55:12 +02001935 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001936err_free:
1937 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001938err:
1939 put_io_context(ioc);
1940 return NULL;
1941}
1942
1943static void
1944cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1945{
Jens Axboeaaf12282007-01-19 11:30:16 +11001946 unsigned long elapsed = jiffies - cic->last_end_request;
1947 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001948
1949 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1950 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1951 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1952}
1953
Jens Axboe206dc692006-03-28 13:03:44 +02001954static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001955cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1956 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001957{
1958 sector_t sdist;
1959 u64 total;
1960
Jeff Moyer4d00aa42009-04-21 07:25:04 +02001961 if (!cic->last_request_pos)
1962 sdist = 0;
Tejun Heo83096eb2009-05-07 22:24:39 +09001963 else if (cic->last_request_pos < blk_rq_pos(rq))
1964 sdist = blk_rq_pos(rq) - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001965 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001966 sdist = cic->last_request_pos - blk_rq_pos(rq);
Jens Axboe206dc692006-03-28 13:03:44 +02001967
1968 /*
1969 * Don't allow the seek distance to get too large from the
1970 * odd fragment, pagein, etc
1971 */
1972 if (cic->seek_samples <= 60) /* second&third seek */
1973 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1974 else
1975 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1976
1977 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1978 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1979 total = cic->seek_total + (cic->seek_samples/2);
1980 do_div(total, cic->seek_samples);
1981 cic->seek_mean = (sector_t)total;
1982}
Jens Axboe22e2c502005-06-27 10:55:12 +02001983
1984/*
1985 * Disable idle window if the process thinks too long or seeks so much that
1986 * it doesn't matter
1987 */
1988static void
1989cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1990 struct cfq_io_context *cic)
1991{
Jens Axboe7b679132008-05-30 12:23:07 +02001992 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02001993
Jens Axboe08717142008-01-28 11:38:15 +01001994 /*
1995 * Don't idle for async or idle io prio class
1996 */
1997 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02001998 return;
1999
Jens Axboec265a7f2008-06-26 13:49:33 +02002000 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002001
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01002002 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Jens Axboe963b72f2009-10-03 19:42:18 +02002003 (!cfqd->cfq_latency && cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02002004 enable_idle = 0;
2005 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccoloec60e4f2009-10-07 19:51:54 +02002006 unsigned int slice_idle = cfqd->cfq_slice_idle;
2007 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
2008 slice_idle = msecs_to_jiffies(CFQ_MIN_TT);
2009 if (cic->ttime_mean > slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02002010 enable_idle = 0;
2011 else
2012 enable_idle = 1;
2013 }
2014
Jens Axboe7b679132008-05-30 12:23:07 +02002015 if (old_idle != enable_idle) {
2016 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2017 if (enable_idle)
2018 cfq_mark_cfqq_idle_window(cfqq);
2019 else
2020 cfq_clear_cfqq_idle_window(cfqq);
2021 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002022}
2023
Jens Axboe22e2c502005-06-27 10:55:12 +02002024/*
2025 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2026 * no or if we aren't sure, a 1 will cause a preempt.
2027 */
Jens Axboea6151c32009-10-07 20:02:57 +02002028static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02002029cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02002030 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002031{
Jens Axboe6d048f52007-04-25 12:44:27 +02002032 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002033
Jens Axboe6d048f52007-04-25 12:44:27 +02002034 cfqq = cfqd->active_queue;
2035 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02002036 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02002037
Jens Axboe6d048f52007-04-25 12:44:27 +02002038 if (cfq_slice_used(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002039 return true;
Jens Axboe6d048f52007-04-25 12:44:27 +02002040
2041 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002042 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02002043
2044 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002045 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002046
Jens Axboe22e2c502005-06-27 10:55:12 +02002047 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02002048 * if the new request is sync, but the currently running queue is
2049 * not, let the sync request have priority.
2050 */
Jens Axboe5e705372006-07-13 12:39:25 +02002051 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002052 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002053
Jens Axboe374f84a2006-07-23 01:42:19 +02002054 /*
2055 * So both queues are sync. Let the new request get disk time if
2056 * it's a metadata request and the current queue is doing regular IO.
2057 */
2058 if (rq_is_meta(rq) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01002059 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02002060
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002061 /*
2062 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2063 */
2064 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002065 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002066
Jens Axboe1e3335d2007-02-14 19:59:49 +01002067 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02002068 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01002069
2070 /*
2071 * if this request is as-good as one we would expect from the
2072 * current cfqq, let it preempt
2073 */
Shaohua Li4b27e1b2009-11-03 20:25:02 +01002074 if (cfq_rq_close(cfqd, rq) && (!cfq_cfqq_coop(new_cfqq) ||
2075 cfqd->busy_queues == 1)) {
2076 /*
2077 * Mark new queue coop_preempt, so its coop flag will not be
2078 * cleared when new queue gets scheduled at the very first time
2079 */
2080 cfq_mark_cfqq_coop_preempt(new_cfqq);
2081 cfq_mark_cfqq_coop(new_cfqq);
Jens Axboea6151c32009-10-07 20:02:57 +02002082 return true;
Shaohua Li4b27e1b2009-11-03 20:25:02 +01002083 }
Jens Axboe1e3335d2007-02-14 19:59:49 +01002084
Jens Axboea6151c32009-10-07 20:02:57 +02002085 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02002086}
2087
2088/*
2089 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2090 * let it have half of its nominal slice.
2091 */
2092static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2093{
Jens Axboe7b679132008-05-30 12:23:07 +02002094 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02002095 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02002096
Jens Axboebf572252006-07-19 20:29:12 +02002097 /*
2098 * Put the new queue at the front of the of the current list,
2099 * so we know that it will be selected next.
2100 */
2101 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02002102
2103 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02002104
Jens Axboe44f7c162007-01-19 11:51:58 +11002105 cfqq->slice_end = 0;
2106 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002107}
2108
2109/*
Jens Axboe5e705372006-07-13 12:39:25 +02002110 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02002111 * something we should do about it
2112 */
2113static void
Jens Axboe5e705372006-07-13 12:39:25 +02002114cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2115 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002116{
Jens Axboe5e705372006-07-13 12:39:25 +02002117 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02002118
Aaron Carroll45333d52008-08-26 15:52:36 +02002119 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02002120 if (rq_is_meta(rq))
2121 cfqq->meta_pending++;
2122
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002123 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02002124 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02002125 cfq_update_idle_window(cfqd, cfqq, cic);
2126
Tejun Heo83096eb2009-05-07 22:24:39 +09002127 cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002128
2129 if (cfqq == cfqd->active_queue) {
2130 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002131 * Remember that we saw a request from this process, but
2132 * don't start queuing just yet. Otherwise we risk seeing lots
2133 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02002134 * and merging. If the request is already larger than a single
2135 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02002136 * merging is already done. Ditto for a busy system that
2137 * has other work pending, don't risk delaying until the
2138 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02002139 */
Jens Axboed6ceb252009-04-14 14:18:16 +02002140 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02002141 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2142 cfqd->busy_queues > 1) {
Jens Axboed6ceb252009-04-14 14:18:16 +02002143 del_timer(&cfqd->idle_slice_timer);
Tejun Heoa7f55792009-04-23 11:05:17 +09002144 __blk_run_queue(cfqd->queue);
Jens Axboed6ceb252009-04-14 14:18:16 +02002145 }
Jens Axboeb0291952009-04-07 11:38:31 +02002146 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboed6ceb252009-04-14 14:18:16 +02002147 }
Jens Axboe5e705372006-07-13 12:39:25 +02002148 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002149 /*
2150 * not the active queue - expire current slice if it is
2151 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01002152 * has some old slice time left and is of higher priority or
2153 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02002154 */
2155 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09002156 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002157 }
2158}
2159
Jens Axboe165125e2007-07-24 09:28:11 +02002160static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002161{
Jens Axboeb4878f22005-10-20 16:42:29 +02002162 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002163 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002164
Jens Axboe7b679132008-05-30 12:23:07 +02002165 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01002166 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167
Jens Axboe5e705372006-07-13 12:39:25 +02002168 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169
Jens Axboe30996f42009-10-05 11:03:39 +02002170 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02002171 list_add_tail(&rq->queuelist, &cfqq->fifo);
2172
Jens Axboe5e705372006-07-13 12:39:25 +02002173 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174}
2175
Aaron Carroll45333d52008-08-26 15:52:36 +02002176/*
2177 * Update hw_tag based on peak queue depth over 50 samples under
2178 * sufficient load.
2179 */
2180static void cfq_update_hw_tag(struct cfq_data *cfqd)
2181{
Jens Axboe5ad531d2009-07-03 12:57:48 +02002182 if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2183 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
Aaron Carroll45333d52008-08-26 15:52:36 +02002184
2185 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Jens Axboe5ad531d2009-07-03 12:57:48 +02002186 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02002187 return;
2188
2189 if (cfqd->hw_tag_samples++ < 50)
2190 return;
2191
2192 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2193 cfqd->hw_tag = 1;
2194 else
2195 cfqd->hw_tag = 0;
2196
2197 cfqd->hw_tag_samples = 0;
2198 cfqd->rq_in_driver_peak = 0;
2199}
2200
Jens Axboe165125e2007-07-24 09:28:11 +02002201static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202{
Jens Axboe5e705372006-07-13 12:39:25 +02002203 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002204 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02002205 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002206 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207
Jens Axboeb4878f22005-10-20 16:42:29 +02002208 now = jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02002209 cfq_log_cfqq(cfqd, cfqq, "complete");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210
Aaron Carroll45333d52008-08-26 15:52:36 +02002211 cfq_update_hw_tag(cfqd);
2212
Jens Axboe5ad531d2009-07-03 12:57:48 +02002213 WARN_ON(!cfqd->rq_in_driver[sync]);
Jens Axboe6d048f52007-04-25 12:44:27 +02002214 WARN_ON(!cfqq->dispatched);
Jens Axboe5ad531d2009-07-03 12:57:48 +02002215 cfqd->rq_in_driver[sync]--;
Jens Axboe6d048f52007-04-25 12:44:27 +02002216 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217
Jens Axboe3ed9a292007-04-23 08:33:33 +02002218 if (cfq_cfqq_sync(cfqq))
2219 cfqd->sync_flight--;
2220
Vivek Goyal365722b2009-10-03 15:21:27 +02002221 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02002222 RQ_CIC(rq)->last_end_request = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02002223 cfqd->last_end_sync_rq = now;
2224 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002225
2226 /*
2227 * If this is the active queue, check if it needs to be expired,
2228 * or if we want to idle in case it has no pending requests.
2229 */
2230 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02002231 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2232
Jens Axboe44f7c162007-01-19 11:51:58 +11002233 if (cfq_cfqq_slice_new(cfqq)) {
2234 cfq_set_prio_slice(cfqd, cfqq);
2235 cfq_clear_cfqq_slice_new(cfqq);
2236 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002237 /*
2238 * If there are no requests waiting in this queue, and
2239 * there are other queues ready to issue requests, AND
2240 * those other queues are issuing requests within our
2241 * mean seek distance, give them a chance to run instead
2242 * of idling.
2243 */
Jens Axboe08717142008-01-28 11:38:15 +01002244 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02002245 cfq_slice_expired(cfqd, 1);
Jens Axboea36e71f2009-04-15 12:15:11 +02002246 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2247 sync && !rq_noidle(rq))
Jens Axboe6d048f52007-04-25 12:44:27 +02002248 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002249 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002250
Jens Axboe5ad531d2009-07-03 12:57:48 +02002251 if (!rq_in_driver(cfqd))
Jens Axboe23e018a2009-10-05 08:52:35 +02002252 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253}
2254
Jens Axboe22e2c502005-06-27 10:55:12 +02002255/*
2256 * we temporarily boost lower priority queues if they are holding fs exclusive
2257 * resources. they are boosted to normal prio (CLASS_BE/4)
2258 */
2259static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260{
Jens Axboe22e2c502005-06-27 10:55:12 +02002261 if (has_fs_excl()) {
2262 /*
2263 * boost idle prio on transactions that would lock out other
2264 * users of the filesystem
2265 */
2266 if (cfq_class_idle(cfqq))
2267 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2268 if (cfqq->ioprio > IOPRIO_NORM)
2269 cfqq->ioprio = IOPRIO_NORM;
2270 } else {
2271 /*
2272 * check if we need to unboost the queue
2273 */
2274 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2275 cfqq->ioprio_class = cfqq->org_ioprio_class;
2276 if (cfqq->ioprio != cfqq->org_ioprio)
2277 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002279}
2280
Jens Axboe89850f72006-07-22 16:48:31 +02002281static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002282{
Jens Axboe1b379d82009-08-11 08:26:11 +02002283 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002284 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002285 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02002286 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002287
2288 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02002289}
2290
Jens Axboe165125e2007-07-24 09:28:11 +02002291static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02002292{
2293 struct cfq_data *cfqd = q->elevator->elevator_data;
2294 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002295 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02002296 struct cfq_queue *cfqq;
2297
2298 /*
2299 * don't force setup of a queue from here, as a call to may_queue
2300 * does not necessarily imply that a request actually will be queued.
2301 * so just lookup a possibly existing queue, or return 'may queue'
2302 * if that fails
2303 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002304 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002305 if (!cic)
2306 return ELV_MQUEUE_MAY;
2307
Jens Axboeb0b78f82009-04-08 10:56:08 +02002308 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02002309 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002310 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02002311 cfq_prio_boost(cfqq);
2312
Jens Axboe89850f72006-07-22 16:48:31 +02002313 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002314 }
2315
2316 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317}
2318
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319/*
2320 * queue lock held here
2321 */
Jens Axboebb37b942006-12-01 10:42:33 +01002322static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323{
Jens Axboe5e705372006-07-13 12:39:25 +02002324 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325
Jens Axboe5e705372006-07-13 12:39:25 +02002326 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002327 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328
Jens Axboe22e2c502005-06-27 10:55:12 +02002329 BUG_ON(!cfqq->allocated[rw]);
2330 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331
Jens Axboe5e705372006-07-13 12:39:25 +02002332 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02002335 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 cfq_put_queue(cfqq);
2338 }
2339}
2340
2341/*
Jens Axboe22e2c502005-06-27 10:55:12 +02002342 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343 */
Jens Axboe22e2c502005-06-27 10:55:12 +02002344static int
Jens Axboe165125e2007-07-24 09:28:11 +02002345cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346{
2347 struct cfq_data *cfqd = q->elevator->elevator_data;
2348 struct cfq_io_context *cic;
2349 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02002350 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002351 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 unsigned long flags;
2353
2354 might_sleep_if(gfp_mask & __GFP_WAIT);
2355
Jens Axboee2d74ac2006-03-28 08:59:01 +02002356 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02002357
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 spin_lock_irqsave(q->queue_lock, flags);
2359
Jens Axboe22e2c502005-06-27 10:55:12 +02002360 if (!cic)
2361 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362
Vasily Tarasov91fac312007-04-25 12:29:51 +02002363 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02002364 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002365 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002366 cic_set_cfqq(cic, cfqq, is_sync);
2367 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368
2369 cfqq->allocated[rw]++;
Jens Axboe22e2c502005-06-27 10:55:12 +02002370 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02002371
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 spin_unlock_irqrestore(q->queue_lock, flags);
2373
Jens Axboe5e705372006-07-13 12:39:25 +02002374 rq->elevator_private = cic;
2375 rq->elevator_private2 = cfqq;
2376 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02002377
Jens Axboe22e2c502005-06-27 10:55:12 +02002378queue_fail:
2379 if (cic)
2380 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02002381
Jens Axboe23e018a2009-10-05 08:52:35 +02002382 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02002384 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385 return 1;
2386}
2387
David Howells65f27f32006-11-22 14:55:48 +00002388static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02002389{
David Howells65f27f32006-11-22 14:55:48 +00002390 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02002391 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02002392 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002393
Jens Axboe40bb54d2009-04-15 12:11:10 +02002394 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09002395 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02002396 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02002397}
2398
2399/*
2400 * Timer running if the active_queue is currently idling inside its time slice
2401 */
2402static void cfq_idle_slice_timer(unsigned long data)
2403{
2404 struct cfq_data *cfqd = (struct cfq_data *) data;
2405 struct cfq_queue *cfqq;
2406 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002407 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002408
Jens Axboe7b679132008-05-30 12:23:07 +02002409 cfq_log(cfqd, "idle timer fired");
2410
Jens Axboe22e2c502005-06-27 10:55:12 +02002411 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2412
Jens Axboefe094d92008-01-31 13:08:54 +01002413 cfqq = cfqd->active_queue;
2414 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002415 timed_out = 0;
2416
Jens Axboe22e2c502005-06-27 10:55:12 +02002417 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002418 * We saw a request before the queue expired, let it through
2419 */
2420 if (cfq_cfqq_must_dispatch(cfqq))
2421 goto out_kick;
2422
2423 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002424 * expired
2425 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002426 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002427 goto expire;
2428
2429 /*
2430 * only expire and reinvoke request handler, if there are
2431 * other queues with pending requests
2432 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002433 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002434 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002435
2436 /*
2437 * not expired and it has a request pending, let it dispatch
2438 */
Jens Axboe75e50982009-04-07 08:56:14 +02002439 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002440 goto out_kick;
Jens Axboe22e2c502005-06-27 10:55:12 +02002441 }
2442expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002443 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002444out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02002445 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002446out_cont:
2447 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2448}
2449
Jens Axboe3b181522005-06-27 10:56:24 +02002450static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2451{
2452 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02002453 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002454}
Jens Axboe22e2c502005-06-27 10:55:12 +02002455
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002456static void cfq_put_async_queues(struct cfq_data *cfqd)
2457{
2458 int i;
2459
2460 for (i = 0; i < IOPRIO_BE_NR; i++) {
2461 if (cfqd->async_cfqq[0][i])
2462 cfq_put_queue(cfqd->async_cfqq[0][i]);
2463 if (cfqd->async_cfqq[1][i])
2464 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002465 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002466
2467 if (cfqd->async_idle_cfqq)
2468 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002469}
2470
Jens Axboeb374d182008-10-31 10:05:07 +01002471static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472{
Jens Axboe22e2c502005-06-27 10:55:12 +02002473 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002474 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002475
Jens Axboe3b181522005-06-27 10:56:24 +02002476 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002477
Al Virod9ff4182006-03-18 13:51:22 -05002478 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002479
Al Virod9ff4182006-03-18 13:51:22 -05002480 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002481 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002482
2483 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002484 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2485 struct cfq_io_context,
2486 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002487
2488 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002489 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002490
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002491 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002492
Al Virod9ff4182006-03-18 13:51:22 -05002493 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002494
2495 cfq_shutdown_timer_wq(cfqd);
2496
Al Viroa90d7422006-03-18 12:05:37 -05002497 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498}
2499
Jens Axboe165125e2007-07-24 09:28:11 +02002500static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501{
2502 struct cfq_data *cfqd;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002503 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504
Christoph Lameter94f60302007-07-17 04:03:29 -07002505 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002507 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508
Jens Axboecc09e292007-04-26 12:53:50 +02002509 cfqd->service_tree = CFQ_RB_ROOT;
Jens Axboe26a2ac02009-04-23 12:13:27 +02002510
2511 /*
2512 * Not strictly needed (since RB_ROOT just clears the node and we
2513 * zeroed cfqd on alloc), but better be safe in case someone decides
2514 * to add magic to the rb code
2515 */
2516 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2517 cfqd->prio_trees[i] = RB_ROOT;
2518
Jens Axboe6118b702009-06-30 09:34:12 +02002519 /*
2520 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2521 * Grab a permanent reference to it, so that the normal code flow
2522 * will not attempt to free it.
2523 */
2524 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2525 atomic_inc(&cfqd->oom_cfqq.ref);
2526
Al Virod9ff4182006-03-18 13:51:22 -05002527 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530
Jens Axboe22e2c502005-06-27 10:55:12 +02002531 init_timer(&cfqd->idle_slice_timer);
2532 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2533 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2534
Jens Axboe23e018a2009-10-05 08:52:35 +02002535 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002536
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002538 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2539 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 cfqd->cfq_back_max = cfq_back_max;
2541 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002542 cfqd->cfq_slice[0] = cfq_slice_async;
2543 cfqd->cfq_slice[1] = cfq_slice_sync;
2544 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2545 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02002546 cfqd->cfq_latency = 1;
Aaron Carroll45333d52008-08-26 15:52:36 +02002547 cfqd->hw_tag = 1;
Vivek Goyal365722b2009-10-03 15:21:27 +02002548 cfqd->last_end_sync_rq = jiffies;
Jens Axboebc1c1162006-06-08 08:49:06 +02002549 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550}
2551
2552static void cfq_slab_kill(void)
2553{
Jens Axboed6de8be2008-05-28 14:46:59 +02002554 /*
2555 * Caller already ensured that pending RCU callbacks are completed,
2556 * so we should have no busy allocations at this point.
2557 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558 if (cfq_pool)
2559 kmem_cache_destroy(cfq_pool);
2560 if (cfq_ioc_pool)
2561 kmem_cache_destroy(cfq_ioc_pool);
2562}
2563
2564static int __init cfq_slab_setup(void)
2565{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002566 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 if (!cfq_pool)
2568 goto fail;
2569
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002570 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571 if (!cfq_ioc_pool)
2572 goto fail;
2573
2574 return 0;
2575fail:
2576 cfq_slab_kill();
2577 return -ENOMEM;
2578}
2579
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580/*
2581 * sysfs parts below -->
2582 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583static ssize_t
2584cfq_var_show(unsigned int var, char *page)
2585{
2586 return sprintf(page, "%d\n", var);
2587}
2588
2589static ssize_t
2590cfq_var_store(unsigned int *var, const char *page, size_t count)
2591{
2592 char *p = (char *) page;
2593
2594 *var = simple_strtoul(p, &p, 10);
2595 return count;
2596}
2597
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002599static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002601 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602 unsigned int __data = __VAR; \
2603 if (__CONV) \
2604 __data = jiffies_to_msecs(__data); \
2605 return cfq_var_show(__data, (page)); \
2606}
2607SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002608SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2609SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002610SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2611SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002612SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2613SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2614SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2615SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02002616SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617#undef SHOW_FUNCTION
2618
2619#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002620static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002622 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623 unsigned int __data; \
2624 int ret = cfq_var_store(&__data, (page), count); \
2625 if (__data < (MIN)) \
2626 __data = (MIN); \
2627 else if (__data > (MAX)) \
2628 __data = (MAX); \
2629 if (__CONV) \
2630 *(__PTR) = msecs_to_jiffies(__data); \
2631 else \
2632 *(__PTR) = __data; \
2633 return ret; \
2634}
2635STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002636STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2637 UINT_MAX, 1);
2638STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2639 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002640STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002641STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2642 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002643STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2644STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2645STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01002646STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2647 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02002648STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649#undef STORE_FUNCTION
2650
Al Viroe572ec72006-03-18 22:27:18 -05002651#define CFQ_ATTR(name) \
2652 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002653
Al Viroe572ec72006-03-18 22:27:18 -05002654static struct elv_fs_entry cfq_attrs[] = {
2655 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002656 CFQ_ATTR(fifo_expire_sync),
2657 CFQ_ATTR(fifo_expire_async),
2658 CFQ_ATTR(back_seek_max),
2659 CFQ_ATTR(back_seek_penalty),
2660 CFQ_ATTR(slice_sync),
2661 CFQ_ATTR(slice_async),
2662 CFQ_ATTR(slice_async_rq),
2663 CFQ_ATTR(slice_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02002664 CFQ_ATTR(low_latency),
Al Viroe572ec72006-03-18 22:27:18 -05002665 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666};
2667
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668static struct elevator_type iosched_cfq = {
2669 .ops = {
2670 .elevator_merge_fn = cfq_merge,
2671 .elevator_merged_fn = cfq_merged_request,
2672 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002673 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002674 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002676 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677 .elevator_deactivate_req_fn = cfq_deactivate_request,
2678 .elevator_queue_empty_fn = cfq_queue_empty,
2679 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002680 .elevator_former_req_fn = elv_rb_former_request,
2681 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682 .elevator_set_req_fn = cfq_set_request,
2683 .elevator_put_req_fn = cfq_put_request,
2684 .elevator_may_queue_fn = cfq_may_queue,
2685 .elevator_init_fn = cfq_init_queue,
2686 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002687 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002688 },
Al Viro3d1ab402006-03-18 18:35:43 -05002689 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690 .elevator_name = "cfq",
2691 .elevator_owner = THIS_MODULE,
2692};
2693
2694static int __init cfq_init(void)
2695{
Jens Axboe22e2c502005-06-27 10:55:12 +02002696 /*
2697 * could be 0 on HZ < 1000 setups
2698 */
2699 if (!cfq_slice_async)
2700 cfq_slice_async = 1;
2701 if (!cfq_slice_idle)
2702 cfq_slice_idle = 1;
2703
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 if (cfq_slab_setup())
2705 return -ENOMEM;
2706
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002707 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002709 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710}
2711
2712static void __exit cfq_exit(void)
2713{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002714 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002716 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002717 /* ioc_gone's update must be visible before reading ioc_count */
2718 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02002719
2720 /*
2721 * this also protects us from entering cfq_slab_kill() with
2722 * pending RCU callbacks
2723 */
Tejun Heo245b2e72009-06-24 15:13:48 +09002724 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002725 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08002726 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727}
2728
2729module_init(cfq_init);
2730module_exit(cfq_exit);
2731
2732MODULE_AUTHOR("Jens Axboe");
2733MODULE_LICENSE("GPL");
2734MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");