blob: a4809de6fea656a4ac6091f6c61c38f536a4f766 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * CFQ, or complete fairness queueing, disk scheduler.
3 *
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6 *
Jens Axboe0fe23472006-09-04 15:41:16 +02007 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/module.h>
Al Viro1cc9be62006-03-18 12:29:52 -050010#include <linux/blkdev.h>
11#include <linux/elevator.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070012#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020013#include <linux/ioprio.h>
Jens Axboe7b679132008-05-30 12:23:07 +020014#include <linux/blktrace_api.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070015
16/*
17 * tunables
18 */
Jens Axboefe094d92008-01-31 13:08:54 +010019/* max queue in one round of service */
20static const int cfq_quantum = 4;
Arjan van de Ven64100092006-01-06 09:46:02 +010021static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010022/* maximum backwards seek, in KiB */
23static const int cfq_back_max = 16 * 1024;
24/* penalty of a backwards seek */
25static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010026static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020027static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010028static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020029static int cfq_slice_idle = HZ / 125;
Jens Axboe22e2c502005-06-27 10:55:12 +020030
Jens Axboed9e76202007-04-20 14:27:50 +020031/*
Jens Axboe08717142008-01-28 11:38:15 +010032 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020033 */
Jens Axboe08717142008-01-28 11:38:15 +010034#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020035
36/*
37 * below this threshold, we consider thinktime immediate
38 */
39#define CFQ_MIN_TT (2)
40
Jens Axboe22e2c502005-06-27 10:55:12 +020041#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020042#define CFQ_HW_QUEUE_MIN (5)
Jens Axboe22e2c502005-06-27 10:55:12 +020043
Jens Axboefe094d92008-01-31 13:08:54 +010044#define RQ_CIC(rq) \
45 ((struct cfq_io_context *) (rq)->elevator_private)
Jens Axboe7b679132008-05-30 12:23:07 +020046#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
Linus Torvalds1da177e2005-04-16 15:20:36 -070047
Christoph Lametere18b8902006-12-06 20:33:20 -080048static struct kmem_cache *cfq_pool;
49static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070050
Jens Axboe4050cf12006-07-19 05:07:12 +020051static DEFINE_PER_CPU(unsigned long, ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050052static struct completion *ioc_gone;
Jens Axboe9a11b4e2008-05-29 09:32:08 +020053static DEFINE_SPINLOCK(ioc_gone_lock);
Al Viro334e94d2006-03-18 15:05:53 -050054
Jens Axboe22e2c502005-06-27 10:55:12 +020055#define CFQ_PRIO_LISTS IOPRIO_BE_NR
56#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020057#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
58
Jens Axboe3b181522005-06-27 10:56:24 +020059#define ASYNC (0)
60#define SYNC (1)
61
Jens Axboe206dc692006-03-28 13:03:44 +020062#define sample_valid(samples) ((samples) > 80)
63
Jens Axboe22e2c502005-06-27 10:55:12 +020064/*
Jens Axboecc09e292007-04-26 12:53:50 +020065 * Most of our rbtree usage is for sorting with min extraction, so
66 * if we cache the leftmost node we don't have to walk down the tree
67 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
68 * move this into the elevator for the rq sorting as well.
69 */
70struct cfq_rb_root {
71 struct rb_root rb;
72 struct rb_node *left;
73};
74#define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
75
76/*
Jens Axboe22e2c502005-06-27 10:55:12 +020077 * Per block device queue structure
78 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070079struct cfq_data {
Jens Axboe165125e2007-07-24 09:28:11 +020080 struct request_queue *queue;
Jens Axboe22e2c502005-06-27 10:55:12 +020081
82 /*
83 * rr list of queues with requests and the count of them
84 */
Jens Axboecc09e292007-04-26 12:53:50 +020085 struct cfq_rb_root service_tree;
Jens Axboe22e2c502005-06-27 10:55:12 +020086 unsigned int busy_queues;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +010087 /*
88 * Used to track any pending rt requests so we can pre-empt current
89 * non-RT cfqq in service when this value is non-zero.
90 */
91 unsigned int busy_rt_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +020092
Jens Axboe22e2c502005-06-27 10:55:12 +020093 int rq_in_driver;
Jens Axboe3ed9a292007-04-23 08:33:33 +020094 int sync_flight;
Aaron Carroll45333d52008-08-26 15:52:36 +020095
96 /*
97 * queue-depth detection
98 */
99 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200100 int hw_tag;
Aaron Carroll45333d52008-08-26 15:52:36 +0200101 int hw_tag_samples;
102 int rq_in_driver_peak;
Jens Axboe22e2c502005-06-27 10:55:12 +0200103
104 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200105 * idle window management
106 */
107 struct timer_list idle_slice_timer;
108 struct work_struct unplug_work;
109
110 struct cfq_queue *active_queue;
111 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200112
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200113 /*
114 * async queue for each priority case
115 */
116 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
117 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200118
Jens Axboe6d048f52007-04-25 12:44:27 +0200119 sector_t last_position;
Jens Axboe22e2c502005-06-27 10:55:12 +0200120 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 /*
123 * tunables, see top of file
124 */
125 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200126 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127 unsigned int cfq_back_penalty;
128 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200129 unsigned int cfq_slice[2];
130 unsigned int cfq_slice_async_rq;
131 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500132
133 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134};
135
Jens Axboe22e2c502005-06-27 10:55:12 +0200136/*
137 * Per process-grouping structure
138 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139struct cfq_queue {
140 /* reference count */
141 atomic_t ref;
Richard Kennedybe754d22008-05-23 06:52:00 +0200142 /* various state flags, see below */
143 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 /* parent cfq_data */
145 struct cfq_data *cfqd;
Jens Axboed9e76202007-04-20 14:27:50 +0200146 /* service_tree member */
147 struct rb_node rb_node;
148 /* service_tree key */
149 unsigned long rb_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 /* sorted list of pending requests */
151 struct rb_root sort_list;
152 /* if fifo isn't expired, next request to serve */
Jens Axboe5e705372006-07-13 12:39:25 +0200153 struct request *next_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 /* requests queued in sort_list */
155 int queued[2];
156 /* currently allocated requests */
157 int allocated[2];
158 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200159 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160
Jens Axboe22e2c502005-06-27 10:55:12 +0200161 unsigned long slice_end;
Jens Axboec5b680f2007-01-19 11:56:49 +1100162 long slice_resid;
Jens Axboe2f5cb732009-04-07 08:51:19 +0200163 unsigned int slice_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164
Richard Kennedybe754d22008-05-23 06:52:00 +0200165 /* pending metadata requests */
166 int meta_pending;
Jens Axboe6d048f52007-04-25 12:44:27 +0200167 /* number of requests that are on the dispatch list or inside driver */
168 int dispatched;
Jens Axboe22e2c502005-06-27 10:55:12 +0200169
170 /* io prio of this group */
171 unsigned short ioprio, org_ioprio;
172 unsigned short ioprio_class, org_ioprio_class;
173
Jens Axboe7b679132008-05-30 12:23:07 +0200174 pid_t pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175};
176
Jens Axboe3b181522005-06-27 10:56:24 +0200177enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100178 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
179 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200180 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100181 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
182 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100183 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
184 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
185 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100186 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200187 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200188};
189
190#define CFQ_CFQQ_FNS(name) \
191static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
192{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100193 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200194} \
195static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
196{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100197 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200198} \
199static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
200{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100201 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200202}
203
204CFQ_CFQQ_FNS(on_rr);
205CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200206CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200207CFQ_CFQQ_FNS(must_alloc);
208CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200209CFQ_CFQQ_FNS(fifo_expire);
210CFQ_CFQQ_FNS(idle_window);
211CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100212CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200213CFQ_CFQQ_FNS(sync);
Jens Axboe3b181522005-06-27 10:56:24 +0200214#undef CFQ_CFQQ_FNS
215
Jens Axboe7b679132008-05-30 12:23:07 +0200216#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
217 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
218#define cfq_log(cfqd, fmt, args...) \
219 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
220
Jens Axboe165125e2007-07-24 09:28:11 +0200221static void cfq_dispatch_insert(struct request_queue *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200222static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
Jens Axboefd0928d2008-01-24 08:52:45 +0100223 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100224static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200225 struct io_context *);
226
227static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
228 int is_sync)
229{
230 return cic->cfqq[!!is_sync];
231}
232
233static inline void cic_set_cfqq(struct cfq_io_context *cic,
234 struct cfq_queue *cfqq, int is_sync)
235{
236 cic->cfqq[!!is_sync] = cfqq;
237}
238
239/*
240 * We regard a request as SYNC, if it's either a read or has the SYNC bit
241 * set (in which case it could also be direct WRITE).
242 */
243static inline int cfq_bio_sync(struct bio *bio)
244{
245 if (bio_data_dir(bio) == READ || bio_sync(bio))
246 return 1;
247
248 return 0;
249}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700252 * scheduler run of queue, if there are requests pending and no one in the
253 * driver that will restart queueing
254 */
255static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
256{
Jens Axboe7b679132008-05-30 12:23:07 +0200257 if (cfqd->busy_queues) {
258 cfq_log(cfqd, "schedule dispatch");
Jens Axboe18887ad2008-07-28 13:08:45 +0200259 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200260 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700261}
262
Jens Axboe165125e2007-07-24 09:28:11 +0200263static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700264{
265 struct cfq_data *cfqd = q->elevator->elevator_data;
266
Jens Axboeb4878f22005-10-20 16:42:29 +0200267 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700268}
269
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100271 * Scale schedule slice based on io priority. Use the sync time slice only
272 * if a queue is marked sync and has sync io queued. A sync queue with async
273 * io only, should not get full sync slice length.
274 */
Jens Axboed9e76202007-04-20 14:27:50 +0200275static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
276 unsigned short prio)
277{
278 const int base_slice = cfqd->cfq_slice[sync];
279
280 WARN_ON(prio >= IOPRIO_BE_NR);
281
282 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
283}
284
Jens Axboe44f7c162007-01-19 11:51:58 +1100285static inline int
286cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
287{
Jens Axboed9e76202007-04-20 14:27:50 +0200288 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100289}
290
291static inline void
292cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
293{
294 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200295 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100296}
297
298/*
299 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
300 * isn't valid until the first request from the dispatch is activated
301 * and the slice time set.
302 */
303static inline int cfq_slice_used(struct cfq_queue *cfqq)
304{
305 if (cfq_cfqq_slice_new(cfqq))
306 return 0;
307 if (time_before(jiffies, cfqq->slice_end))
308 return 0;
309
310 return 1;
311}
312
313/*
Jens Axboe5e705372006-07-13 12:39:25 +0200314 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200316 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317 */
Jens Axboe5e705372006-07-13 12:39:25 +0200318static struct request *
319cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320{
321 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200323#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
324#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
325 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326
Jens Axboe5e705372006-07-13 12:39:25 +0200327 if (rq1 == NULL || rq1 == rq2)
328 return rq2;
329 if (rq2 == NULL)
330 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200331
Jens Axboe5e705372006-07-13 12:39:25 +0200332 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
333 return rq1;
334 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
335 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200336 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
337 return rq1;
338 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
339 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340
Jens Axboe5e705372006-07-13 12:39:25 +0200341 s1 = rq1->sector;
342 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343
Jens Axboe6d048f52007-04-25 12:44:27 +0200344 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346 /*
347 * by definition, 1KiB is 2 sectors
348 */
349 back_max = cfqd->cfq_back_max * 2;
350
351 /*
352 * Strict one way elevator _except_ in the case where we allow
353 * short backward seeks which are biased as twice the cost of a
354 * similar forward seek.
355 */
356 if (s1 >= last)
357 d1 = s1 - last;
358 else if (s1 + back_max >= last)
359 d1 = (last - s1) * cfqd->cfq_back_penalty;
360 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200361 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362
363 if (s2 >= last)
364 d2 = s2 - last;
365 else if (s2 + back_max >= last)
366 d2 = (last - s2) * cfqd->cfq_back_penalty;
367 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200368 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369
370 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371
Andreas Mohre8a99052006-03-28 08:59:49 +0200372 /*
373 * By doing switch() on the bit mask "wrap" we avoid having to
374 * check two variables for all permutations: --> faster!
375 */
376 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200377 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200378 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200379 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200380 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200381 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200382 else {
383 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200384 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200385 else
Jens Axboe5e705372006-07-13 12:39:25 +0200386 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200387 }
388
389 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200390 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200391 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200392 return rq2;
393 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200394 default:
395 /*
396 * Since both rqs are wrapped,
397 * start with the one that's further behind head
398 * (--> only *one* back seek required),
399 * since back seek takes more time than forward.
400 */
401 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200402 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 else
Jens Axboe5e705372006-07-13 12:39:25 +0200404 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 }
406}
407
Jens Axboe498d3aa2007-04-26 12:54:48 +0200408/*
409 * The below is leftmost cache rbtree addon
410 */
Jens Axboe08717142008-01-28 11:38:15 +0100411static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200412{
413 if (!root->left)
414 root->left = rb_first(&root->rb);
415
Jens Axboe08717142008-01-28 11:38:15 +0100416 if (root->left)
417 return rb_entry(root->left, struct cfq_queue, rb_node);
418
419 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200420}
421
422static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
423{
424 if (root->left == n)
425 root->left = NULL;
426
427 rb_erase(n, &root->rb);
428 RB_CLEAR_NODE(n);
429}
430
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431/*
432 * would be nice to take fifo expire time into account as well
433 */
Jens Axboe5e705372006-07-13 12:39:25 +0200434static struct request *
435cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
436 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437{
Jens Axboe21183b02006-07-13 12:33:14 +0200438 struct rb_node *rbnext = rb_next(&last->rb_node);
439 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200440 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441
Jens Axboe21183b02006-07-13 12:33:14 +0200442 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
444 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200445 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200446
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200448 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200449 else {
450 rbnext = rb_first(&cfqq->sort_list);
451 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200452 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200453 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454
Jens Axboe21183b02006-07-13 12:33:14 +0200455 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456}
457
Jens Axboed9e76202007-04-20 14:27:50 +0200458static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
459 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460{
Jens Axboed9e76202007-04-20 14:27:50 +0200461 /*
462 * just an approximation, should be ok.
463 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200464 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
465 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200466}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467
Jens Axboe498d3aa2007-04-26 12:54:48 +0200468/*
469 * The cfqd->service_tree holds all pending cfq_queue's that have
470 * requests waiting to be processed. It is sorted in the order that
471 * we will service the queues.
472 */
Jens Axboed9e76202007-04-20 14:27:50 +0200473static void cfq_service_tree_add(struct cfq_data *cfqd,
Jens Axboeedd75ff2007-04-19 12:03:34 +0200474 struct cfq_queue *cfqq, int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200475{
Jens Axboe08717142008-01-28 11:38:15 +0100476 struct rb_node **p, *parent;
477 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +0200478 unsigned long rb_key;
Jens Axboe498d3aa2007-04-26 12:54:48 +0200479 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200480
Jens Axboe08717142008-01-28 11:38:15 +0100481 if (cfq_class_idle(cfqq)) {
482 rb_key = CFQ_IDLE_DELAY;
483 parent = rb_last(&cfqd->service_tree.rb);
484 if (parent && parent != &cfqq->rb_node) {
485 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
486 rb_key += __cfqq->rb_key;
487 } else
488 rb_key += jiffies;
489 } else if (!add_front) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200490 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
491 rb_key += cfqq->slice_resid;
492 cfqq->slice_resid = 0;
493 } else
494 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200495
496 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100497 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200498 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100499 */
Jens Axboed9e76202007-04-20 14:27:50 +0200500 if (rb_key == cfqq->rb_key)
501 return;
Jens Axboe53b03742006-07-28 09:48:51 +0200502
Jens Axboecc09e292007-04-26 12:53:50 +0200503 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200504 }
Jens Axboed9e76202007-04-20 14:27:50 +0200505
Jens Axboe498d3aa2007-04-26 12:54:48 +0200506 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +0100507 parent = NULL;
508 p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200509 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +0200510 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200511
Jens Axboed9e76202007-04-20 14:27:50 +0200512 parent = *p;
513 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
514
Jens Axboe0c534e02007-04-18 20:01:57 +0200515 /*
516 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200517 * preference to them. IDLE queues goes to the back.
518 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200519 */
520 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200521 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200522 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200523 n = &(*p)->rb_right;
524 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
525 n = &(*p)->rb_left;
526 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
527 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200528 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200529 n = &(*p)->rb_left;
530 else
531 n = &(*p)->rb_right;
532
533 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200534 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200535
536 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200537 }
538
Jens Axboecc09e292007-04-26 12:53:50 +0200539 if (left)
540 cfqd->service_tree.left = &cfqq->rb_node;
541
Jens Axboed9e76202007-04-20 14:27:50 +0200542 cfqq->rb_key = rb_key;
543 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200544 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545}
546
Jens Axboe498d3aa2007-04-26 12:54:48 +0200547/*
548 * Update cfqq's position in the service tree.
549 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200550static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200551{
Jens Axboe6d048f52007-04-25 12:44:27 +0200552 /*
553 * Resorting requires the cfqq to be on the RR list already.
554 */
Jens Axboe498d3aa2007-04-26 12:54:48 +0200555 if (cfq_cfqq_on_rr(cfqq))
Jens Axboeedd75ff2007-04-19 12:03:34 +0200556 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboe6d048f52007-04-25 12:44:27 +0200557}
558
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559/*
560 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200561 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 */
Jens Axboefebffd62008-01-28 13:19:43 +0100563static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564{
Jens Axboe7b679132008-05-30 12:23:07 +0200565 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200566 BUG_ON(cfq_cfqq_on_rr(cfqq));
567 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 cfqd->busy_queues++;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +0100569 if (cfq_class_rt(cfqq))
570 cfqd->busy_rt_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571
Jens Axboeedd75ff2007-04-19 12:03:34 +0200572 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573}
574
Jens Axboe498d3aa2007-04-26 12:54:48 +0200575/*
576 * Called when the cfqq no longer has requests pending, remove it from
577 * the service tree.
578 */
Jens Axboefebffd62008-01-28 13:19:43 +0100579static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580{
Jens Axboe7b679132008-05-30 12:23:07 +0200581 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200582 BUG_ON(!cfq_cfqq_on_rr(cfqq));
583 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584
Jens Axboecc09e292007-04-26 12:53:50 +0200585 if (!RB_EMPTY_NODE(&cfqq->rb_node))
586 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboed9e76202007-04-20 14:27:50 +0200587
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 BUG_ON(!cfqd->busy_queues);
589 cfqd->busy_queues--;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +0100590 if (cfq_class_rt(cfqq))
591 cfqd->busy_rt_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592}
593
594/*
595 * rb tree support functions
596 */
Jens Axboefebffd62008-01-28 13:19:43 +0100597static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598{
Jens Axboe5e705372006-07-13 12:39:25 +0200599 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200600 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200601 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602
Jens Axboeb4878f22005-10-20 16:42:29 +0200603 BUG_ON(!cfqq->queued[sync]);
604 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
Jens Axboe5e705372006-07-13 12:39:25 +0200606 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607
Jens Axboedd67d052006-06-21 09:36:18 +0200608 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200609 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610}
611
Jens Axboe5e705372006-07-13 12:39:25 +0200612static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613{
Jens Axboe5e705372006-07-13 12:39:25 +0200614 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe21183b02006-07-13 12:33:14 +0200616 struct request *__alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617
Jens Axboe5380a102006-07-13 12:37:56 +0200618 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619
620 /*
621 * looks a little odd, but the first insert might return an alias.
622 * if that happens, put the alias on the dispatch list
623 */
Jens Axboe21183b02006-07-13 12:33:14 +0200624 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200625 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100626
627 if (!cfq_cfqq_on_rr(cfqq))
628 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200629
630 /*
631 * check if this request is a better next-serve candidate
632 */
633 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
634 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635}
636
Jens Axboefebffd62008-01-28 13:19:43 +0100637static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638{
Jens Axboe5380a102006-07-13 12:37:56 +0200639 elv_rb_del(&cfqq->sort_list, rq);
640 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200641 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642}
643
Jens Axboe206dc692006-03-28 13:03:44 +0200644static struct request *
645cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646{
Jens Axboe206dc692006-03-28 13:03:44 +0200647 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200648 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200649 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650
Jens Axboe4ac845a2008-01-24 08:44:49 +0100651 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200652 if (!cic)
653 return NULL;
654
655 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200656 if (cfqq) {
657 sector_t sector = bio->bi_sector + bio_sectors(bio);
658
Jens Axboe21183b02006-07-13 12:33:14 +0200659 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200660 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662 return NULL;
663}
664
Jens Axboe165125e2007-07-24 09:28:11 +0200665static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200666{
667 struct cfq_data *cfqd = q->elevator->elevator_data;
668
669 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +0200670 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
671 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +0200672
Jens Axboe6d048f52007-04-25 12:44:27 +0200673 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
Jens Axboeb4878f22005-10-20 16:42:29 +0200674}
675
Jens Axboe165125e2007-07-24 09:28:11 +0200676static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677{
Jens Axboe22e2c502005-06-27 10:55:12 +0200678 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679
Jens Axboeb4878f22005-10-20 16:42:29 +0200680 WARN_ON(!cfqd->rq_in_driver);
681 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +0200682 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
683 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684}
685
Jens Axboeb4878f22005-10-20 16:42:29 +0200686static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687{
Jens Axboe5e705372006-07-13 12:39:25 +0200688 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200689
Jens Axboe5e705372006-07-13 12:39:25 +0200690 if (cfqq->next_rq == rq)
691 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692
Jens Axboeb4878f22005-10-20 16:42:29 +0200693 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200694 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200695
Aaron Carroll45333d52008-08-26 15:52:36 +0200696 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +0200697 if (rq_is_meta(rq)) {
698 WARN_ON(!cfqq->meta_pending);
699 cfqq->meta_pending--;
700 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701}
702
Jens Axboe165125e2007-07-24 09:28:11 +0200703static int cfq_merge(struct request_queue *q, struct request **req,
704 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705{
706 struct cfq_data *cfqd = q->elevator->elevator_data;
707 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708
Jens Axboe206dc692006-03-28 13:03:44 +0200709 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200710 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200711 *req = __rq;
712 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 }
714
715 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716}
717
Jens Axboe165125e2007-07-24 09:28:11 +0200718static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200719 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720{
Jens Axboe21183b02006-07-13 12:33:14 +0200721 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200722 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723
Jens Axboe5e705372006-07-13 12:39:25 +0200724 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726}
727
728static void
Jens Axboe165125e2007-07-24 09:28:11 +0200729cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730 struct request *next)
731{
Jens Axboe22e2c502005-06-27 10:55:12 +0200732 /*
733 * reposition in fifo if next is older than rq
734 */
735 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
736 time_before(next->start_time, rq->start_time))
737 list_move(&rq->queuelist, &next->queuelist);
738
Jens Axboeb4878f22005-10-20 16:42:29 +0200739 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200740}
741
Jens Axboe165125e2007-07-24 09:28:11 +0200742static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100743 struct bio *bio)
744{
745 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200746 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100747 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100748
749 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100750 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100751 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200752 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100753 return 0;
754
755 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100756 * Lookup the cfqq that this bio will be queued with. Allow
757 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100758 */
Jens Axboe4ac845a2008-01-24 08:44:49 +0100759 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200760 if (!cic)
761 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100762
Vasily Tarasov91fac312007-04-25 12:29:51 +0200763 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100764 if (cfqq == RQ_CFQQ(rq))
765 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100766
Jens Axboeec8acb62007-01-02 18:32:11 +0100767 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100768}
769
Jens Axboefebffd62008-01-28 13:19:43 +0100770static void __cfq_set_active_queue(struct cfq_data *cfqd,
771 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +0200772{
773 if (cfqq) {
Jens Axboe7b679132008-05-30 12:23:07 +0200774 cfq_log_cfqq(cfqd, cfqq, "set_active");
Jens Axboe22e2c502005-06-27 10:55:12 +0200775 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +0200776 cfqq->slice_dispatch = 0;
777
Jens Axboe2f5cb732009-04-07 08:51:19 +0200778 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +0200779 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200780 cfq_clear_cfqq_must_alloc_slice(cfqq);
781 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100782 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +0200783
784 del_timer(&cfqd->idle_slice_timer);
Jens Axboe22e2c502005-06-27 10:55:12 +0200785 }
786
787 cfqd->active_queue = cfqq;
788}
789
790/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100791 * current cfqq expired its slice (or was too idle), select new one
792 */
793static void
794__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200795 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100796{
Jens Axboe7b679132008-05-30 12:23:07 +0200797 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
798
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100799 if (cfq_cfqq_wait_request(cfqq))
800 del_timer(&cfqd->idle_slice_timer);
801
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100802 cfq_clear_cfqq_wait_request(cfqq);
803
804 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200805 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100806 */
Jens Axboe7b679132008-05-30 12:23:07 +0200807 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +1100808 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200809 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
810 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100811
Jens Axboeedd75ff2007-04-19 12:03:34 +0200812 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100813
814 if (cfqq == cfqd->active_queue)
815 cfqd->active_queue = NULL;
816
817 if (cfqd->active_cic) {
818 put_io_context(cfqd->active_cic->ioc);
819 cfqd->active_cic = NULL;
820 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100821}
822
Jens Axboe6084cdd2007-04-23 08:25:00 +0200823static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100824{
825 struct cfq_queue *cfqq = cfqd->active_queue;
826
827 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200828 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100829}
830
Jens Axboe498d3aa2007-04-26 12:54:48 +0200831/*
832 * Get next queue for service. Unless we have a queue preemption,
833 * we'll simply select the first cfqq in the service tree.
834 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200835static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200836{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200837 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
838 return NULL;
839
Jens Axboe08717142008-01-28 11:38:15 +0100840 return cfq_rb_first(&cfqd->service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +0200841}
842
Jens Axboe498d3aa2007-04-26 12:54:48 +0200843/*
844 * Get and set a new active queue for service.
845 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200846static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
847{
848 struct cfq_queue *cfqq;
849
Jens Axboed9e76202007-04-20 14:27:50 +0200850 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200851 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200852 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200853}
854
Jens Axboed9e76202007-04-20 14:27:50 +0200855static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
856 struct request *rq)
857{
858 if (rq->sector >= cfqd->last_position)
859 return rq->sector - cfqd->last_position;
860 else
861 return cfqd->last_position - rq->sector;
862}
863
Jens Axboe6d048f52007-04-25 12:44:27 +0200864static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
865{
866 struct cfq_io_context *cic = cfqd->active_cic;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200867
Jens Axboe6d048f52007-04-25 12:44:27 +0200868 if (!sample_valid(cic->seek_samples))
869 return 0;
870
871 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
872}
873
Jens Axboed9e76202007-04-20 14:27:50 +0200874static int cfq_close_cooperator(struct cfq_data *cfq_data,
875 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200876{
Jens Axboe6d048f52007-04-25 12:44:27 +0200877 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200878 * We should notice if some of the queues are cooperating, eg
879 * working closely on the same area of the disk. In that case,
880 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +0200881 */
Jens Axboed9e76202007-04-20 14:27:50 +0200882 return 0;
Jens Axboe6d048f52007-04-25 12:44:27 +0200883}
884
885#define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
886
887static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200888{
Jens Axboe17926692007-01-19 11:59:30 +1100889 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +0200890 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100891 unsigned long sl;
892
Jens Axboea68bbdd2008-09-24 13:03:33 +0200893 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +0200894 * SSD device without seek penalty, disable idling. But only do so
895 * for devices that support queuing, otherwise we still have a problem
896 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +0200897 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +0200898 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +0200899 return;
900
Jens Axboedd67d052006-06-21 09:36:18 +0200901 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +0200902 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +0200903
904 /*
905 * idle is disabled, either manually or by past process history
906 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200907 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
908 return;
909
Jens Axboe22e2c502005-06-27 10:55:12 +0200910 /*
Jens Axboe7b679132008-05-30 12:23:07 +0200911 * still requests with the driver, don't idle
912 */
913 if (cfqd->rq_in_driver)
914 return;
915
916 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200917 * task has exited, don't wait
918 */
Jens Axboe206dc692006-03-28 13:03:44 +0200919 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +0100920 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +0200921 return;
922
923 /*
924 * See if this prio level has a good candidate
925 */
Jens Axboe1afba042007-04-17 12:47:55 +0200926 if (cfq_close_cooperator(cfqd, cfqq) &&
927 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
Jens Axboe6d048f52007-04-25 12:44:27 +0200928 return;
Jens Axboe22e2c502005-06-27 10:55:12 +0200929
Jens Axboe3b181522005-06-27 10:56:24 +0200930 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200931
Jens Axboe206dc692006-03-28 13:03:44 +0200932 /*
933 * we don't want to idle for seeks, but we do want to allow
934 * fair distribution of slice time for a process doing back-to-back
935 * seeks. so allow a little bit of time for him to submit a new rq
936 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200937 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200938 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +0200939 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +0200940
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100941 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe7b679132008-05-30 12:23:07 +0200942 cfq_log(cfqd, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943}
944
Jens Axboe498d3aa2007-04-26 12:54:48 +0200945/*
946 * Move request from internal lists to the request queue dispatch list.
947 */
Jens Axboe165125e2007-07-24 09:28:11 +0200948static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949{
Jens Axboe3ed9a292007-04-23 08:33:33 +0200950 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200951 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200952
Jens Axboe7b679132008-05-30 12:23:07 +0200953 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
954
Jens Axboe5380a102006-07-13 12:37:56 +0200955 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200956 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +0200957 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +0200958
959 if (cfq_cfqq_sync(cfqq))
960 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961}
962
963/*
964 * return expired entry, or NULL to just start from scratch in rbtree
965 */
Jens Axboefebffd62008-01-28 13:19:43 +0100966static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967{
968 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200969 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200970 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971
Jens Axboe3b181522005-06-27 10:56:24 +0200972 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +1100974
975 cfq_mark_cfqq_fifo_expire(cfqq);
976
Jens Axboe89850f72006-07-22 16:48:31 +0200977 if (list_empty(&cfqq->fifo))
978 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979
Jens Axboe6d048f52007-04-25 12:44:27 +0200980 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +0200981 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982
Jens Axboe6d048f52007-04-25 12:44:27 +0200983 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
Jens Axboe7b679132008-05-30 12:23:07 +0200984 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985
Jens Axboe7b679132008-05-30 12:23:07 +0200986 cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200987 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988}
989
Jens Axboe22e2c502005-06-27 10:55:12 +0200990static inline int
991cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
992{
993 const int base_rq = cfqd->cfq_slice_async_rq;
994
995 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
996
997 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
998}
999
1000/*
Jens Axboe498d3aa2007-04-26 12:54:48 +02001001 * Select a queue for service. If we have a current active queue,
1002 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02001003 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001004static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001005{
Jens Axboe22e2c502005-06-27 10:55:12 +02001006 struct cfq_queue *cfqq;
1007
1008 cfqq = cfqd->active_queue;
1009 if (!cfqq)
1010 goto new_queue;
1011
1012 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001013 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02001014 */
Jens Axboeb0291952009-04-07 11:38:31 +02001015 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001016 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +02001017
1018 /*
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01001019 * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1020 * cfqq.
1021 */
1022 if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1023 /*
1024 * We simulate this as cfqq timed out so that it gets to bank
1025 * the remaining of its time slice.
1026 */
1027 cfq_log_cfqq(cfqd, cfqq, "preempt");
1028 cfq_slice_expired(cfqd, 1);
1029 goto new_queue;
1030 }
1031
1032 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001033 * The active queue has requests and isn't expired, allow it to
1034 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02001035 */
Jens Axboedd67d052006-06-21 09:36:18 +02001036 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001037 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02001038
1039 /*
1040 * No requests pending. If the active queue still has requests in
1041 * flight or is idling for a new request, allow either of these
1042 * conditions to happen (or time out) before selecting a new queue.
1043 */
Jens Axboecc197472007-04-20 20:45:39 +02001044 if (timer_pending(&cfqd->idle_slice_timer) ||
1045 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001046 cfqq = NULL;
1047 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001048 }
1049
Jens Axboe3b181522005-06-27 10:56:24 +02001050expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001051 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001052new_queue:
1053 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001054keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001055 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001056}
1057
Jens Axboefebffd62008-01-28 13:19:43 +01001058static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02001059{
1060 int dispatched = 0;
1061
1062 while (cfqq->next_rq) {
1063 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1064 dispatched++;
1065 }
1066
1067 BUG_ON(!list_empty(&cfqq->fifo));
1068 return dispatched;
1069}
1070
Jens Axboe498d3aa2007-04-26 12:54:48 +02001071/*
1072 * Drain our current requests. Used for barriers and when switching
1073 * io schedulers on-the-fly.
1074 */
Jens Axboed9e76202007-04-20 14:27:50 +02001075static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001076{
Jens Axboe08717142008-01-28 11:38:15 +01001077 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001078 int dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001079
Jens Axboe08717142008-01-28 11:38:15 +01001080 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
Jens Axboed9e76202007-04-20 14:27:50 +02001081 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001082
Jens Axboe6084cdd2007-04-23 08:25:00 +02001083 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001084
1085 BUG_ON(cfqd->busy_queues);
1086
Jens Axboe7b679132008-05-30 12:23:07 +02001087 cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001088 return dispatched;
1089}
1090
Jens Axboe2f5cb732009-04-07 08:51:19 +02001091/*
1092 * Dispatch a request from cfqq, moving them to the request queue
1093 * dispatch list.
1094 */
1095static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1096{
1097 struct request *rq;
1098
1099 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1100
1101 /*
1102 * follow expired path, else get first next available
1103 */
1104 rq = cfq_check_fifo(cfqq);
1105 if (!rq)
1106 rq = cfqq->next_rq;
1107
1108 /*
1109 * insert request into driver dispatch list
1110 */
1111 cfq_dispatch_insert(cfqd->queue, rq);
1112
1113 if (!cfqd->active_cic) {
1114 struct cfq_io_context *cic = RQ_CIC(rq);
1115
1116 atomic_inc(&cic->ioc->refcount);
1117 cfqd->active_cic = cic;
1118 }
1119}
1120
1121/*
1122 * Find the cfqq that we need to service and move a request from that to the
1123 * dispatch list
1124 */
Jens Axboe165125e2007-07-24 09:28:11 +02001125static int cfq_dispatch_requests(struct request_queue *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126{
1127 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001128 struct cfq_queue *cfqq;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001129 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
Jens Axboe22e2c502005-06-27 10:55:12 +02001131 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 return 0;
1133
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001134 if (unlikely(force))
1135 return cfq_forced_dispatch(cfqd);
1136
Jens Axboe2f5cb732009-04-07 08:51:19 +02001137 cfqq = cfq_select_queue(cfqd);
1138 if (!cfqq)
1139 return 0;
Jens Axboeb4878f22005-10-20 16:42:29 +02001140
Jens Axboe2f5cb732009-04-07 08:51:19 +02001141 /*
1142 * If this is an async queue and we have sync IO in flight, let it wait
1143 */
1144 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1145 return 0;
1146
1147 max_dispatch = cfqd->cfq_quantum;
1148 if (cfq_class_idle(cfqq))
1149 max_dispatch = 1;
1150
1151 /*
1152 * Does this cfqq already have too much IO in flight?
1153 */
1154 if (cfqq->dispatched >= max_dispatch) {
1155 /*
1156 * idle queue must always only have a single IO in flight
1157 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02001158 if (cfq_class_idle(cfqq))
Jens Axboe2f5cb732009-04-07 08:51:19 +02001159 return 0;
Jens Axboe3ed9a292007-04-23 08:33:33 +02001160
Jens Axboe2f5cb732009-04-07 08:51:19 +02001161 /*
1162 * We have other queues, don't allow more IO from this one
1163 */
1164 if (cfqd->busy_queues > 1)
1165 return 0;
Jens Axboe9ede2092007-01-19 12:11:44 +11001166
Jens Axboe2f5cb732009-04-07 08:51:19 +02001167 /*
1168 * we are the only queue, allow up to 4 times of 'quantum'
1169 */
1170 if (cfqq->dispatched >= 4 * max_dispatch)
1171 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 }
1173
Jens Axboe2f5cb732009-04-07 08:51:19 +02001174 /*
1175 * Dispatch a request from this cfqq
1176 */
1177 cfq_dispatch_request(cfqd, cfqq);
1178 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02001179 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001180
1181 /*
1182 * expire an async queue immediately if it has used up its slice. idle
1183 * queue always expire after 1 dispatch round.
1184 */
1185 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1186 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1187 cfq_class_idle(cfqq))) {
1188 cfqq->slice_end = jiffies + 1;
1189 cfq_slice_expired(cfqd, 0);
1190 }
1191
1192 cfq_log(cfqd, "dispatched a request");
1193 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194}
1195
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196/*
Jens Axboe5e705372006-07-13 12:39:25 +02001197 * task holds one reference to the queue, dropped when task exits. each rq
1198 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 *
1200 * queue lock must be held here.
1201 */
1202static void cfq_put_queue(struct cfq_queue *cfqq)
1203{
Jens Axboe22e2c502005-06-27 10:55:12 +02001204 struct cfq_data *cfqd = cfqq->cfqd;
1205
1206 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207
1208 if (!atomic_dec_and_test(&cfqq->ref))
1209 return;
1210
Jens Axboe7b679132008-05-30 12:23:07 +02001211 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001213 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001214 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001216 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001217 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001218 cfq_schedule_dispatch(cfqd);
1219 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001220
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 kmem_cache_free(cfq_pool, cfqq);
1222}
1223
Jens Axboed6de8be2008-05-28 14:46:59 +02001224/*
1225 * Must always be called with the rcu_read_lock() held
1226 */
Jens Axboe07416d22008-05-07 09:17:12 +02001227static void
1228__call_for_each_cic(struct io_context *ioc,
1229 void (*func)(struct io_context *, struct cfq_io_context *))
1230{
1231 struct cfq_io_context *cic;
1232 struct hlist_node *n;
1233
1234 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1235 func(ioc, cic);
1236}
1237
Jens Axboe4ac845a2008-01-24 08:44:49 +01001238/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001239 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001240 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001241static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001242call_for_each_cic(struct io_context *ioc,
1243 void (*func)(struct io_context *, struct cfq_io_context *))
1244{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001245 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02001246 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001247 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001248}
Jens Axboe4ac845a2008-01-24 08:44:49 +01001249
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001250static void cfq_cic_free_rcu(struct rcu_head *head)
1251{
1252 struct cfq_io_context *cic;
1253
1254 cic = container_of(head, struct cfq_io_context, rcu_head);
1255
1256 kmem_cache_free(cfq_ioc_pool, cic);
1257 elv_ioc_count_dec(ioc_count);
1258
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001259 if (ioc_gone) {
1260 /*
1261 * CFQ scheduler is exiting, grab exit lock and check
1262 * the pending io context count. If it hits zero,
1263 * complete ioc_gone and set it back to NULL
1264 */
1265 spin_lock(&ioc_gone_lock);
1266 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1267 complete(ioc_gone);
1268 ioc_gone = NULL;
1269 }
1270 spin_unlock(&ioc_gone_lock);
1271 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001272}
1273
1274static void cfq_cic_free(struct cfq_io_context *cic)
1275{
1276 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001277}
1278
1279static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1280{
1281 unsigned long flags;
1282
1283 BUG_ON(!cic->dead_key);
1284
1285 spin_lock_irqsave(&ioc->lock, flags);
1286 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01001287 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001288 spin_unlock_irqrestore(&ioc->lock, flags);
1289
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001290 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001291}
1292
Jens Axboed6de8be2008-05-28 14:46:59 +02001293/*
1294 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1295 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1296 * and ->trim() which is called with the task lock held
1297 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001298static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001300 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001301 * ioc->refcount is zero here, or we are called from elv_unregister(),
1302 * so no more cic's are allowed to be linked into this ioc. So it
1303 * should be ok to iterate over the known list, we will see all cic's
1304 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001305 */
Jens Axboe07416d22008-05-07 09:17:12 +02001306 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307}
1308
Jens Axboe89850f72006-07-22 16:48:31 +02001309static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1310{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001311 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001312 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001313 cfq_schedule_dispatch(cfqd);
1314 }
Jens Axboe89850f72006-07-22 16:48:31 +02001315
1316 cfq_put_queue(cfqq);
1317}
1318
1319static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1320 struct cfq_io_context *cic)
1321{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001322 struct io_context *ioc = cic->ioc;
1323
Jens Axboefc463792006-08-29 09:05:44 +02001324 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001325
1326 /*
1327 * Make sure key == NULL is seen for dead queues
1328 */
Jens Axboefc463792006-08-29 09:05:44 +02001329 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001330 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02001331 cic->key = NULL;
1332
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001333 if (ioc->ioc_data == cic)
1334 rcu_assign_pointer(ioc->ioc_data, NULL);
1335
Jens Axboe89850f72006-07-22 16:48:31 +02001336 if (cic->cfqq[ASYNC]) {
1337 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1338 cic->cfqq[ASYNC] = NULL;
1339 }
1340
1341 if (cic->cfqq[SYNC]) {
1342 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1343 cic->cfqq[SYNC] = NULL;
1344 }
Jens Axboe89850f72006-07-22 16:48:31 +02001345}
1346
Jens Axboe4ac845a2008-01-24 08:44:49 +01001347static void cfq_exit_single_io_context(struct io_context *ioc,
1348 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001349{
Al Viro478a82b2006-03-18 13:25:24 -05001350 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001351
Jens Axboe89850f72006-07-22 16:48:31 +02001352 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001353 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001354 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02001355
Jens Axboe4ac845a2008-01-24 08:44:49 +01001356 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01001357
1358 /*
1359 * Ensure we get a fresh copy of the ->key to prevent
1360 * race between exiting task and queue
1361 */
1362 smp_read_barrier_depends();
1363 if (cic->key)
1364 __cfq_exit_single_io_context(cfqd, cic);
1365
Jens Axboe4ac845a2008-01-24 08:44:49 +01001366 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05001367 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001368}
1369
Jens Axboe498d3aa2007-04-26 12:54:48 +02001370/*
1371 * The process that ioc belongs to has exited, we need to clean up
1372 * and put the internal structures we have that belongs to that process.
1373 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001374static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001376 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377}
1378
Jens Axboe22e2c502005-06-27 10:55:12 +02001379static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001380cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381{
Jens Axboeb5deef92006-07-19 23:39:40 +02001382 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383
Christoph Lameter94f60302007-07-17 04:03:29 -07001384 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1385 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001387 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001388 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01001389 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001390 cic->dtor = cfq_free_io_context;
1391 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001392 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393 }
1394
1395 return cic;
1396}
1397
Jens Axboefd0928d2008-01-24 08:52:45 +01001398static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02001399{
1400 struct task_struct *tsk = current;
1401 int ioprio_class;
1402
Jens Axboe3b181522005-06-27 10:56:24 +02001403 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001404 return;
1405
Jens Axboefd0928d2008-01-24 08:52:45 +01001406 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001407 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01001408 default:
1409 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1410 case IOPRIO_CLASS_NONE:
1411 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02001412 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01001413 */
1414 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02001415 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01001416 break;
1417 case IOPRIO_CLASS_RT:
1418 cfqq->ioprio = task_ioprio(ioc);
1419 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1420 break;
1421 case IOPRIO_CLASS_BE:
1422 cfqq->ioprio = task_ioprio(ioc);
1423 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1424 break;
1425 case IOPRIO_CLASS_IDLE:
1426 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1427 cfqq->ioprio = 7;
1428 cfq_clear_cfqq_idle_window(cfqq);
1429 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02001430 }
1431
1432 /*
1433 * keep track of original prio settings in case we have to temporarily
1434 * elevate the priority of this queue
1435 */
1436 cfqq->org_ioprio = cfqq->ioprio;
1437 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001438 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001439}
1440
Jens Axboefebffd62008-01-28 13:19:43 +01001441static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001442{
Al Viro478a82b2006-03-18 13:25:24 -05001443 struct cfq_data *cfqd = cic->key;
1444 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001445 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001446
Jens Axboecaaa5f92006-06-16 11:23:00 +02001447 if (unlikely(!cfqd))
1448 return;
1449
Jens Axboec1b707d2006-10-30 19:54:23 +01001450 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001451
1452 cfqq = cic->cfqq[ASYNC];
1453 if (cfqq) {
1454 struct cfq_queue *new_cfqq;
Jens Axboefd0928d2008-01-24 08:52:45 +01001455 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001456 if (new_cfqq) {
1457 cic->cfqq[ASYNC] = new_cfqq;
1458 cfq_put_queue(cfqq);
1459 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001460 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001461
1462 cfqq = cic->cfqq[SYNC];
1463 if (cfqq)
1464 cfq_mark_cfqq_prio_changed(cfqq);
1465
Jens Axboec1b707d2006-10-30 19:54:23 +01001466 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001467}
1468
Jens Axboefc463792006-08-29 09:05:44 +02001469static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001471 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02001472 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473}
1474
1475static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001476cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01001477 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001480 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481
1482retry:
Jens Axboe4ac845a2008-01-24 08:44:49 +01001483 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001484 /* cic always exists here */
1485 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486
1487 if (!cfqq) {
1488 if (new_cfqq) {
1489 cfqq = new_cfqq;
1490 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001491 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001492 /*
1493 * Inform the allocator of the fact that we will
1494 * just repeat this allocation if it fails, to allow
1495 * the allocator to do whatever it needs to attempt to
1496 * free memory.
1497 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001499 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1500 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1501 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 spin_lock_irq(cfqd->queue->queue_lock);
1503 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001504 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001505 cfqq = kmem_cache_alloc_node(cfq_pool,
1506 gfp_mask | __GFP_ZERO,
1507 cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001508 if (!cfqq)
1509 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001510 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511
Jens Axboed9e76202007-04-20 14:27:50 +02001512 RB_CLEAR_NODE(&cfqq->rb_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001513 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515 atomic_set(&cfqq->ref, 0);
1516 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001517
Jens Axboe3b181522005-06-27 10:56:24 +02001518 cfq_mark_cfqq_prio_changed(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001519
Jens Axboefd0928d2008-01-24 08:52:45 +01001520 cfq_init_prio_data(cfqq, ioc);
Jens Axboe08717142008-01-28 11:38:15 +01001521
1522 if (is_sync) {
1523 if (!cfq_class_idle(cfqq))
1524 cfq_mark_cfqq_idle_window(cfqq);
1525 cfq_mark_cfqq_sync(cfqq);
1526 }
Jens Axboe7b679132008-05-30 12:23:07 +02001527 cfqq->pid = current->pid;
1528 cfq_log_cfqq(cfqd, cfqq, "alloced");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 }
1530
1531 if (new_cfqq)
1532 kmem_cache_free(cfq_pool, new_cfqq);
1533
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534out:
1535 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1536 return cfqq;
1537}
1538
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001539static struct cfq_queue **
1540cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1541{
Jens Axboefe094d92008-01-31 13:08:54 +01001542 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001543 case IOPRIO_CLASS_RT:
1544 return &cfqd->async_cfqq[0][ioprio];
1545 case IOPRIO_CLASS_BE:
1546 return &cfqd->async_cfqq[1][ioprio];
1547 case IOPRIO_CLASS_IDLE:
1548 return &cfqd->async_idle_cfqq;
1549 default:
1550 BUG();
1551 }
1552}
1553
Jens Axboe15c31be2007-07-10 13:43:25 +02001554static struct cfq_queue *
Jens Axboefd0928d2008-01-24 08:52:45 +01001555cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02001556 gfp_t gfp_mask)
1557{
Jens Axboefd0928d2008-01-24 08:52:45 +01001558 const int ioprio = task_ioprio(ioc);
1559 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001560 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001561 struct cfq_queue *cfqq = NULL;
1562
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001563 if (!is_sync) {
1564 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1565 cfqq = *async_cfqq;
1566 }
1567
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001568 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01001569 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001570 if (!cfqq)
1571 return NULL;
1572 }
Jens Axboe15c31be2007-07-10 13:43:25 +02001573
1574 /*
1575 * pin the queue now that it's allocated, scheduler exit will prune it
1576 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001577 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001578 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001579 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001580 }
1581
1582 atomic_inc(&cfqq->ref);
1583 return cfqq;
1584}
1585
Jens Axboe498d3aa2007-04-26 12:54:48 +02001586/*
1587 * We drop cfq io contexts lazily, so we may find a dead one.
1588 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001589static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001590cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1591 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001592{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001593 unsigned long flags;
1594
Jens Axboefc463792006-08-29 09:05:44 +02001595 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001596
Jens Axboe4ac845a2008-01-24 08:44:49 +01001597 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02001598
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001599 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001600
1601 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01001602 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001603 spin_unlock_irqrestore(&ioc->lock, flags);
1604
1605 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001606}
1607
Jens Axboee2d74ac2006-03-28 08:59:01 +02001608static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01001609cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001610{
Jens Axboee2d74ac2006-03-28 08:59:01 +02001611 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001612 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001613 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001614
Vasily Tarasov91fac312007-04-25 12:29:51 +02001615 if (unlikely(!ioc))
1616 return NULL;
1617
Jens Axboed6de8be2008-05-28 14:46:59 +02001618 rcu_read_lock();
1619
Jens Axboe597bc482007-04-24 21:23:53 +02001620 /*
1621 * we maintain a last-hit cache, to avoid browsing over the tree
1622 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001623 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02001624 if (cic && cic->key == cfqd) {
1625 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02001626 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001627 }
Jens Axboe597bc482007-04-24 21:23:53 +02001628
Jens Axboe4ac845a2008-01-24 08:44:49 +01001629 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001630 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1631 rcu_read_unlock();
1632 if (!cic)
1633 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001634 /* ->key must be copied to avoid race with cfq_exit_queue() */
1635 k = cic->key;
1636 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001637 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001638 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001639 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001640 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001641
Jens Axboed6de8be2008-05-28 14:46:59 +02001642 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001643 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001644 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001645 break;
1646 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001647
Jens Axboe4ac845a2008-01-24 08:44:49 +01001648 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001649}
1650
Jens Axboe4ac845a2008-01-24 08:44:49 +01001651/*
1652 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1653 * the process specific cfq io context when entered from the block layer.
1654 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1655 */
Jens Axboefebffd62008-01-28 13:19:43 +01001656static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1657 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001658{
Jens Axboe0261d682006-10-30 19:07:48 +01001659 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001660 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001661
Jens Axboe4ac845a2008-01-24 08:44:49 +01001662 ret = radix_tree_preload(gfp_mask);
1663 if (!ret) {
1664 cic->ioc = ioc;
1665 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001666
Jens Axboe4ac845a2008-01-24 08:44:49 +01001667 spin_lock_irqsave(&ioc->lock, flags);
1668 ret = radix_tree_insert(&ioc->radix_root,
1669 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01001670 if (!ret)
1671 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001672 spin_unlock_irqrestore(&ioc->lock, flags);
1673
1674 radix_tree_preload_end();
1675
1676 if (!ret) {
1677 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1678 list_add(&cic->queue_list, &cfqd->cic_list);
1679 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001680 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001681 }
1682
Jens Axboe4ac845a2008-01-24 08:44:49 +01001683 if (ret)
1684 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02001685
Jens Axboe4ac845a2008-01-24 08:44:49 +01001686 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001687}
1688
Jens Axboe22e2c502005-06-27 10:55:12 +02001689/*
1690 * Setup general io context and cfq io context. There can be several cfq
1691 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001692 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001693 */
1694static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001695cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696{
Jens Axboe22e2c502005-06-27 10:55:12 +02001697 struct io_context *ioc = NULL;
1698 struct cfq_io_context *cic;
1699
1700 might_sleep_if(gfp_mask & __GFP_WAIT);
1701
Jens Axboeb5deef92006-07-19 23:39:40 +02001702 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001703 if (!ioc)
1704 return NULL;
1705
Jens Axboe4ac845a2008-01-24 08:44:49 +01001706 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001707 if (cic)
1708 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001709
Jens Axboee2d74ac2006-03-28 08:59:01 +02001710 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1711 if (cic == NULL)
1712 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001713
Jens Axboe4ac845a2008-01-24 08:44:49 +01001714 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1715 goto err_free;
1716
Jens Axboe22e2c502005-06-27 10:55:12 +02001717out:
Jens Axboefc463792006-08-29 09:05:44 +02001718 smp_read_barrier_depends();
1719 if (unlikely(ioc->ioprio_changed))
1720 cfq_ioc_set_ioprio(ioc);
1721
Jens Axboe22e2c502005-06-27 10:55:12 +02001722 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001723err_free:
1724 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001725err:
1726 put_io_context(ioc);
1727 return NULL;
1728}
1729
1730static void
1731cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1732{
Jens Axboeaaf12282007-01-19 11:30:16 +11001733 unsigned long elapsed = jiffies - cic->last_end_request;
1734 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001735
1736 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1737 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1738 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1739}
1740
Jens Axboe206dc692006-03-28 13:03:44 +02001741static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001742cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1743 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001744{
1745 sector_t sdist;
1746 u64 total;
1747
Jens Axboe5e705372006-07-13 12:39:25 +02001748 if (cic->last_request_pos < rq->sector)
1749 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001750 else
Jens Axboe5e705372006-07-13 12:39:25 +02001751 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001752
1753 /*
1754 * Don't allow the seek distance to get too large from the
1755 * odd fragment, pagein, etc
1756 */
1757 if (cic->seek_samples <= 60) /* second&third seek */
1758 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1759 else
1760 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1761
1762 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1763 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1764 total = cic->seek_total + (cic->seek_samples/2);
1765 do_div(total, cic->seek_samples);
1766 cic->seek_mean = (sector_t)total;
1767}
Jens Axboe22e2c502005-06-27 10:55:12 +02001768
1769/*
1770 * Disable idle window if the process thinks too long or seeks so much that
1771 * it doesn't matter
1772 */
1773static void
1774cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1775 struct cfq_io_context *cic)
1776{
Jens Axboe7b679132008-05-30 12:23:07 +02001777 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02001778
Jens Axboe08717142008-01-28 11:38:15 +01001779 /*
1780 * Don't idle for async or idle io prio class
1781 */
1782 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02001783 return;
1784
Jens Axboec265a7f2008-06-26 13:49:33 +02001785 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001786
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001787 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Jens Axboecaaa5f92006-06-16 11:23:00 +02001788 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001789 enable_idle = 0;
1790 else if (sample_valid(cic->ttime_samples)) {
1791 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1792 enable_idle = 0;
1793 else
1794 enable_idle = 1;
1795 }
1796
Jens Axboe7b679132008-05-30 12:23:07 +02001797 if (old_idle != enable_idle) {
1798 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1799 if (enable_idle)
1800 cfq_mark_cfqq_idle_window(cfqq);
1801 else
1802 cfq_clear_cfqq_idle_window(cfqq);
1803 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001804}
1805
Jens Axboe22e2c502005-06-27 10:55:12 +02001806/*
1807 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1808 * no or if we aren't sure, a 1 will cause a preempt.
1809 */
1810static int
1811cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001812 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001813{
Jens Axboe6d048f52007-04-25 12:44:27 +02001814 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001815
Jens Axboe6d048f52007-04-25 12:44:27 +02001816 cfqq = cfqd->active_queue;
1817 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001818 return 0;
1819
Jens Axboe6d048f52007-04-25 12:44:27 +02001820 if (cfq_slice_used(cfqq))
1821 return 1;
1822
1823 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001824 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001825
1826 if (cfq_class_idle(cfqq))
1827 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001828
Jens Axboe22e2c502005-06-27 10:55:12 +02001829 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02001830 * if the new request is sync, but the currently running queue is
1831 * not, let the sync request have priority.
1832 */
Jens Axboe5e705372006-07-13 12:39:25 +02001833 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001834 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001835
Jens Axboe374f84a2006-07-23 01:42:19 +02001836 /*
1837 * So both queues are sync. Let the new request get disk time if
1838 * it's a metadata request and the current queue is doing regular IO.
1839 */
1840 if (rq_is_meta(rq) && !cfqq->meta_pending)
1841 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001842
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01001843 /*
1844 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
1845 */
1846 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
1847 return 1;
1848
Jens Axboe1e3335d2007-02-14 19:59:49 +01001849 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1850 return 0;
1851
1852 /*
1853 * if this request is as-good as one we would expect from the
1854 * current cfqq, let it preempt
1855 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001856 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01001857 return 1;
1858
Jens Axboe22e2c502005-06-27 10:55:12 +02001859 return 0;
1860}
1861
1862/*
1863 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1864 * let it have half of its nominal slice.
1865 */
1866static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1867{
Jens Axboe7b679132008-05-30 12:23:07 +02001868 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02001869 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001870
Jens Axboebf572252006-07-19 20:29:12 +02001871 /*
1872 * Put the new queue at the front of the of the current list,
1873 * so we know that it will be selected next.
1874 */
1875 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02001876
1877 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02001878
Jens Axboe44f7c162007-01-19 11:51:58 +11001879 cfqq->slice_end = 0;
1880 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001881}
1882
1883/*
Jens Axboe5e705372006-07-13 12:39:25 +02001884 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001885 * something we should do about it
1886 */
1887static void
Jens Axboe5e705372006-07-13 12:39:25 +02001888cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1889 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001890{
Jens Axboe5e705372006-07-13 12:39:25 +02001891 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001892
Aaron Carroll45333d52008-08-26 15:52:36 +02001893 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02001894 if (rq_is_meta(rq))
1895 cfqq->meta_pending++;
1896
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001897 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02001898 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001899 cfq_update_idle_window(cfqd, cfqq, cic);
1900
Jens Axboe5e705372006-07-13 12:39:25 +02001901 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001902
1903 if (cfqq == cfqd->active_queue) {
1904 /*
Jens Axboeb0291952009-04-07 11:38:31 +02001905 * Remember that we saw a request from this process, but
1906 * don't start queuing just yet. Otherwise we risk seeing lots
1907 * of tiny requests, because we disrupt the normal plugging
1908 * and merging.
Jens Axboe22e2c502005-06-27 10:55:12 +02001909 */
Jens Axboeb0291952009-04-07 11:38:31 +02001910 if (cfq_cfqq_wait_request(cfqq))
1911 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe5e705372006-07-13 12:39:25 +02001912 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001913 /*
1914 * not the active queue - expire current slice if it is
1915 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01001916 * has some old slice time left and is of higher priority or
1917 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02001918 */
1919 cfq_preempt_queue(cfqd, cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001920 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001921 }
1922}
1923
Jens Axboe165125e2007-07-24 09:28:11 +02001924static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001925{
Jens Axboeb4878f22005-10-20 16:42:29 +02001926 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001927 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001928
Jens Axboe7b679132008-05-30 12:23:07 +02001929 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01001930 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931
Jens Axboe5e705372006-07-13 12:39:25 +02001932 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933
Jens Axboe22e2c502005-06-27 10:55:12 +02001934 list_add_tail(&rq->queuelist, &cfqq->fifo);
1935
Jens Axboe5e705372006-07-13 12:39:25 +02001936 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937}
1938
Aaron Carroll45333d52008-08-26 15:52:36 +02001939/*
1940 * Update hw_tag based on peak queue depth over 50 samples under
1941 * sufficient load.
1942 */
1943static void cfq_update_hw_tag(struct cfq_data *cfqd)
1944{
1945 if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
1946 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
1947
1948 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
1949 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
1950 return;
1951
1952 if (cfqd->hw_tag_samples++ < 50)
1953 return;
1954
1955 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
1956 cfqd->hw_tag = 1;
1957 else
1958 cfqd->hw_tag = 0;
1959
1960 cfqd->hw_tag_samples = 0;
1961 cfqd->rq_in_driver_peak = 0;
1962}
1963
Jens Axboe165125e2007-07-24 09:28:11 +02001964static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965{
Jens Axboe5e705372006-07-13 12:39:25 +02001966 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001967 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001968 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001969 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970
Jens Axboeb4878f22005-10-20 16:42:29 +02001971 now = jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001972 cfq_log_cfqq(cfqd, cfqq, "complete");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973
Aaron Carroll45333d52008-08-26 15:52:36 +02001974 cfq_update_hw_tag(cfqd);
1975
Jens Axboeb4878f22005-10-20 16:42:29 +02001976 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02001977 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02001978 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02001979 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980
Jens Axboe3ed9a292007-04-23 08:33:33 +02001981 if (cfq_cfqq_sync(cfqq))
1982 cfqd->sync_flight--;
1983
Jens Axboeb4878f22005-10-20 16:42:29 +02001984 if (!cfq_class_idle(cfqq))
1985 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001986
Jens Axboecaaa5f92006-06-16 11:23:00 +02001987 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001988 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001989
1990 /*
1991 * If this is the active queue, check if it needs to be expired,
1992 * or if we want to idle in case it has no pending requests.
1993 */
1994 if (cfqd->active_queue == cfqq) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001995 if (cfq_cfqq_slice_new(cfqq)) {
1996 cfq_set_prio_slice(cfqd, cfqq);
1997 cfq_clear_cfqq_slice_new(cfqq);
1998 }
Jens Axboe08717142008-01-28 11:38:15 +01001999 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02002000 cfq_slice_expired(cfqd, 1);
Jens Axboeaeb6faf2009-04-06 14:48:07 +02002001 else if (sync && !rq_noidle(rq) &&
2002 RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe6d048f52007-04-25 12:44:27 +02002003 cfq_arm_slice_timer(cfqd);
Jens Axboeaeb6faf2009-04-06 14:48:07 +02002004 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002005 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002006
2007 if (!cfqd->rq_in_driver)
2008 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009}
2010
Jens Axboe22e2c502005-06-27 10:55:12 +02002011/*
2012 * we temporarily boost lower priority queues if they are holding fs exclusive
2013 * resources. they are boosted to normal prio (CLASS_BE/4)
2014 */
2015static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016{
Jens Axboe22e2c502005-06-27 10:55:12 +02002017 if (has_fs_excl()) {
2018 /*
2019 * boost idle prio on transactions that would lock out other
2020 * users of the filesystem
2021 */
2022 if (cfq_class_idle(cfqq))
2023 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2024 if (cfqq->ioprio > IOPRIO_NORM)
2025 cfqq->ioprio = IOPRIO_NORM;
2026 } else {
2027 /*
2028 * check if we need to unboost the queue
2029 */
2030 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2031 cfqq->ioprio_class = cfqq->org_ioprio_class;
2032 if (cfqq->ioprio != cfqq->org_ioprio)
2033 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002035}
2036
Jens Axboe89850f72006-07-22 16:48:31 +02002037static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002038{
Jens Axboe3b181522005-06-27 10:56:24 +02002039 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07002040 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002041 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002042 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02002043 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002044
2045 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02002046}
2047
Jens Axboe165125e2007-07-24 09:28:11 +02002048static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02002049{
2050 struct cfq_data *cfqd = q->elevator->elevator_data;
2051 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002052 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02002053 struct cfq_queue *cfqq;
2054
2055 /*
2056 * don't force setup of a queue from here, as a call to may_queue
2057 * does not necessarily imply that a request actually will be queued.
2058 * so just lookup a possibly existing queue, or return 'may queue'
2059 * if that fails
2060 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002061 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002062 if (!cic)
2063 return ELV_MQUEUE_MAY;
2064
2065 cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02002066 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002067 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02002068 cfq_prio_boost(cfqq);
2069
Jens Axboe89850f72006-07-22 16:48:31 +02002070 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002071 }
2072
2073 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074}
2075
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076/*
2077 * queue lock held here
2078 */
Jens Axboebb37b942006-12-01 10:42:33 +01002079static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080{
Jens Axboe5e705372006-07-13 12:39:25 +02002081 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082
Jens Axboe5e705372006-07-13 12:39:25 +02002083 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002084 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085
Jens Axboe22e2c502005-06-27 10:55:12 +02002086 BUG_ON(!cfqq->allocated[rw]);
2087 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088
Jens Axboe5e705372006-07-13 12:39:25 +02002089 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02002092 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 cfq_put_queue(cfqq);
2095 }
2096}
2097
2098/*
Jens Axboe22e2c502005-06-27 10:55:12 +02002099 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100 */
Jens Axboe22e2c502005-06-27 10:55:12 +02002101static int
Jens Axboe165125e2007-07-24 09:28:11 +02002102cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103{
2104 struct cfq_data *cfqd = q->elevator->elevator_data;
2105 struct cfq_io_context *cic;
2106 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01002107 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002108 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109 unsigned long flags;
2110
2111 might_sleep_if(gfp_mask & __GFP_WAIT);
2112
Jens Axboee2d74ac2006-03-28 08:59:01 +02002113 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02002114
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 spin_lock_irqsave(q->queue_lock, flags);
2116
Jens Axboe22e2c502005-06-27 10:55:12 +02002117 if (!cic)
2118 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119
Vasily Tarasov91fac312007-04-25 12:29:51 +02002120 cfqq = cic_to_cfqq(cic, is_sync);
2121 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002122 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002123
Jens Axboe22e2c502005-06-27 10:55:12 +02002124 if (!cfqq)
2125 goto queue_fail;
2126
Vasily Tarasov91fac312007-04-25 12:29:51 +02002127 cic_set_cfqq(cic, cfqq, is_sync);
2128 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129
2130 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02002131 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002132 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02002133
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134 spin_unlock_irqrestore(q->queue_lock, flags);
2135
Jens Axboe5e705372006-07-13 12:39:25 +02002136 rq->elevator_private = cic;
2137 rq->elevator_private2 = cfqq;
2138 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02002139
Jens Axboe22e2c502005-06-27 10:55:12 +02002140queue_fail:
2141 if (cic)
2142 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02002143
Jens Axboe3b181522005-06-27 10:56:24 +02002144 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02002146 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 return 1;
2148}
2149
David Howells65f27f32006-11-22 14:55:48 +00002150static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02002151{
David Howells65f27f32006-11-22 14:55:48 +00002152 struct cfq_data *cfqd =
2153 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02002154 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002155 unsigned long flags;
2156
2157 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02002158 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02002159 spin_unlock_irqrestore(q->queue_lock, flags);
2160}
2161
2162/*
2163 * Timer running if the active_queue is currently idling inside its time slice
2164 */
2165static void cfq_idle_slice_timer(unsigned long data)
2166{
2167 struct cfq_data *cfqd = (struct cfq_data *) data;
2168 struct cfq_queue *cfqq;
2169 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002170 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002171
Jens Axboe7b679132008-05-30 12:23:07 +02002172 cfq_log(cfqd, "idle timer fired");
2173
Jens Axboe22e2c502005-06-27 10:55:12 +02002174 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2175
Jens Axboefe094d92008-01-31 13:08:54 +01002176 cfqq = cfqd->active_queue;
2177 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002178 timed_out = 0;
2179
Jens Axboe22e2c502005-06-27 10:55:12 +02002180 /*
Jens Axboeb0291952009-04-07 11:38:31 +02002181 * We saw a request before the queue expired, let it through
2182 */
2183 if (cfq_cfqq_must_dispatch(cfqq))
2184 goto out_kick;
2185
2186 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002187 * expired
2188 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002189 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002190 goto expire;
2191
2192 /*
2193 * only expire and reinvoke request handler, if there are
2194 * other queues with pending requests
2195 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002196 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002197 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002198
2199 /*
2200 * not expired and it has a request pending, let it dispatch
2201 */
Jens Axboe75e50982009-04-07 08:56:14 +02002202 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002203 goto out_kick;
Jens Axboe22e2c502005-06-27 10:55:12 +02002204 }
2205expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002206 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002207out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002208 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002209out_cont:
2210 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2211}
2212
Jens Axboe3b181522005-06-27 10:56:24 +02002213static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2214{
2215 del_timer_sync(&cfqd->idle_slice_timer);
Cheng Renquan64d01dc2008-12-03 12:41:39 +01002216 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002217}
Jens Axboe22e2c502005-06-27 10:55:12 +02002218
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002219static void cfq_put_async_queues(struct cfq_data *cfqd)
2220{
2221 int i;
2222
2223 for (i = 0; i < IOPRIO_BE_NR; i++) {
2224 if (cfqd->async_cfqq[0][i])
2225 cfq_put_queue(cfqd->async_cfqq[0][i]);
2226 if (cfqd->async_cfqq[1][i])
2227 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002228 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002229
2230 if (cfqd->async_idle_cfqq)
2231 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002232}
2233
Jens Axboeb374d182008-10-31 10:05:07 +01002234static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235{
Jens Axboe22e2c502005-06-27 10:55:12 +02002236 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002237 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002238
Jens Axboe3b181522005-06-27 10:56:24 +02002239 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002240
Al Virod9ff4182006-03-18 13:51:22 -05002241 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002242
Al Virod9ff4182006-03-18 13:51:22 -05002243 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002244 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002245
2246 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002247 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2248 struct cfq_io_context,
2249 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002250
2251 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002252 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002253
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002254 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002255
Al Virod9ff4182006-03-18 13:51:22 -05002256 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002257
2258 cfq_shutdown_timer_wq(cfqd);
2259
Al Viroa90d7422006-03-18 12:05:37 -05002260 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261}
2262
Jens Axboe165125e2007-07-24 09:28:11 +02002263static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264{
2265 struct cfq_data *cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266
Christoph Lameter94f60302007-07-17 04:03:29 -07002267 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002269 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270
Jens Axboecc09e292007-04-26 12:53:50 +02002271 cfqd->service_tree = CFQ_RB_ROOT;
Al Virod9ff4182006-03-18 13:51:22 -05002272 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275
Jens Axboe22e2c502005-06-27 10:55:12 +02002276 init_timer(&cfqd->idle_slice_timer);
2277 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2278 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2279
David Howells65f27f32006-11-22 14:55:48 +00002280 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002281
Oleg Nesterovb70c8642007-11-07 09:46:13 +01002282 cfqd->last_end_request = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002284 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2285 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 cfqd->cfq_back_max = cfq_back_max;
2287 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002288 cfqd->cfq_slice[0] = cfq_slice_async;
2289 cfqd->cfq_slice[1] = cfq_slice_sync;
2290 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2291 cfqd->cfq_slice_idle = cfq_slice_idle;
Aaron Carroll45333d52008-08-26 15:52:36 +02002292 cfqd->hw_tag = 1;
Jens Axboe3b181522005-06-27 10:56:24 +02002293
Jens Axboebc1c1162006-06-08 08:49:06 +02002294 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295}
2296
2297static void cfq_slab_kill(void)
2298{
Jens Axboed6de8be2008-05-28 14:46:59 +02002299 /*
2300 * Caller already ensured that pending RCU callbacks are completed,
2301 * so we should have no busy allocations at this point.
2302 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303 if (cfq_pool)
2304 kmem_cache_destroy(cfq_pool);
2305 if (cfq_ioc_pool)
2306 kmem_cache_destroy(cfq_ioc_pool);
2307}
2308
2309static int __init cfq_slab_setup(void)
2310{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002311 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 if (!cfq_pool)
2313 goto fail;
2314
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002315 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316 if (!cfq_ioc_pool)
2317 goto fail;
2318
2319 return 0;
2320fail:
2321 cfq_slab_kill();
2322 return -ENOMEM;
2323}
2324
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325/*
2326 * sysfs parts below -->
2327 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328static ssize_t
2329cfq_var_show(unsigned int var, char *page)
2330{
2331 return sprintf(page, "%d\n", var);
2332}
2333
2334static ssize_t
2335cfq_var_store(unsigned int *var, const char *page, size_t count)
2336{
2337 char *p = (char *) page;
2338
2339 *var = simple_strtoul(p, &p, 10);
2340 return count;
2341}
2342
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002344static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002346 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 unsigned int __data = __VAR; \
2348 if (__CONV) \
2349 __data = jiffies_to_msecs(__data); \
2350 return cfq_var_show(__data, (page)); \
2351}
2352SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002353SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2354SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002355SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2356SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002357SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2358SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2359SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2360SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361#undef SHOW_FUNCTION
2362
2363#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002364static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002366 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 unsigned int __data; \
2368 int ret = cfq_var_store(&__data, (page), count); \
2369 if (__data < (MIN)) \
2370 __data = (MIN); \
2371 else if (__data > (MAX)) \
2372 __data = (MAX); \
2373 if (__CONV) \
2374 *(__PTR) = msecs_to_jiffies(__data); \
2375 else \
2376 *(__PTR) = __data; \
2377 return ret; \
2378}
2379STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002380STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2381 UINT_MAX, 1);
2382STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2383 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002384STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002385STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2386 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002387STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2388STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2389STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01002390STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2391 UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392#undef STORE_FUNCTION
2393
Al Viroe572ec72006-03-18 22:27:18 -05002394#define CFQ_ATTR(name) \
2395 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002396
Al Viroe572ec72006-03-18 22:27:18 -05002397static struct elv_fs_entry cfq_attrs[] = {
2398 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002399 CFQ_ATTR(fifo_expire_sync),
2400 CFQ_ATTR(fifo_expire_async),
2401 CFQ_ATTR(back_seek_max),
2402 CFQ_ATTR(back_seek_penalty),
2403 CFQ_ATTR(slice_sync),
2404 CFQ_ATTR(slice_async),
2405 CFQ_ATTR(slice_async_rq),
2406 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002407 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408};
2409
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410static struct elevator_type iosched_cfq = {
2411 .ops = {
2412 .elevator_merge_fn = cfq_merge,
2413 .elevator_merged_fn = cfq_merged_request,
2414 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002415 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002416 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002418 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 .elevator_deactivate_req_fn = cfq_deactivate_request,
2420 .elevator_queue_empty_fn = cfq_queue_empty,
2421 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002422 .elevator_former_req_fn = elv_rb_former_request,
2423 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 .elevator_set_req_fn = cfq_set_request,
2425 .elevator_put_req_fn = cfq_put_request,
2426 .elevator_may_queue_fn = cfq_may_queue,
2427 .elevator_init_fn = cfq_init_queue,
2428 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002429 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430 },
Al Viro3d1ab402006-03-18 18:35:43 -05002431 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 .elevator_name = "cfq",
2433 .elevator_owner = THIS_MODULE,
2434};
2435
2436static int __init cfq_init(void)
2437{
Jens Axboe22e2c502005-06-27 10:55:12 +02002438 /*
2439 * could be 0 on HZ < 1000 setups
2440 */
2441 if (!cfq_slice_async)
2442 cfq_slice_async = 1;
2443 if (!cfq_slice_idle)
2444 cfq_slice_idle = 1;
2445
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 if (cfq_slab_setup())
2447 return -ENOMEM;
2448
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002449 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002451 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452}
2453
2454static void __exit cfq_exit(void)
2455{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002456 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002458 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002459 /* ioc_gone's update must be visible before reading ioc_count */
2460 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02002461
2462 /*
2463 * this also protects us from entering cfq_slab_kill() with
2464 * pending RCU callbacks
2465 */
Jens Axboe4050cf12006-07-19 05:07:12 +02002466 if (elv_ioc_count_read(ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002467 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08002468 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469}
2470
2471module_init(cfq_init);
2472module_exit(cfq_exit);
2473
2474MODULE_AUTHOR("Jens Axboe");
2475MODULE_LICENSE("GPL");
2476MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");