blob: e8525fa72823a0c88adfd386e84e7b3e5db33af7 [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;
87
Jens Axboe22e2c502005-06-27 10:55:12 +020088 int rq_in_driver;
Jens Axboe3ed9a292007-04-23 08:33:33 +020089 int sync_flight;
Aaron Carroll45333d52008-08-26 15:52:36 +020090
91 /*
92 * queue-depth detection
93 */
94 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +020095 int hw_tag;
Aaron Carroll45333d52008-08-26 15:52:36 +020096 int hw_tag_samples;
97 int rq_in_driver_peak;
Jens Axboe22e2c502005-06-27 10:55:12 +020098
99 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200100 * idle window management
101 */
102 struct timer_list idle_slice_timer;
103 struct work_struct unplug_work;
104
105 struct cfq_queue *active_queue;
106 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200107
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200108 /*
109 * async queue for each priority case
110 */
111 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
112 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200113
Jens Axboe6d048f52007-04-25 12:44:27 +0200114 sector_t last_position;
Jens Axboe22e2c502005-06-27 10:55:12 +0200115 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117 /*
118 * tunables, see top of file
119 */
120 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200121 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 unsigned int cfq_back_penalty;
123 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200124 unsigned int cfq_slice[2];
125 unsigned int cfq_slice_async_rq;
126 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500127
128 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129};
130
Jens Axboe22e2c502005-06-27 10:55:12 +0200131/*
132 * Per process-grouping structure
133 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134struct cfq_queue {
135 /* reference count */
136 atomic_t ref;
Richard Kennedybe754d22008-05-23 06:52:00 +0200137 /* various state flags, see below */
138 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 /* parent cfq_data */
140 struct cfq_data *cfqd;
Jens Axboed9e76202007-04-20 14:27:50 +0200141 /* service_tree member */
142 struct rb_node rb_node;
143 /* service_tree key */
144 unsigned long rb_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145 /* sorted list of pending requests */
146 struct rb_root sort_list;
147 /* if fifo isn't expired, next request to serve */
Jens Axboe5e705372006-07-13 12:39:25 +0200148 struct request *next_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 /* requests queued in sort_list */
150 int queued[2];
151 /* currently allocated requests */
152 int allocated[2];
153 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200154 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
Jens Axboe22e2c502005-06-27 10:55:12 +0200156 unsigned long slice_end;
Jens Axboec5b680f2007-01-19 11:56:49 +1100157 long slice_resid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158
Richard Kennedybe754d22008-05-23 06:52:00 +0200159 /* pending metadata requests */
160 int meta_pending;
Jens Axboe6d048f52007-04-25 12:44:27 +0200161 /* number of requests that are on the dispatch list or inside driver */
162 int dispatched;
Jens Axboe22e2c502005-06-27 10:55:12 +0200163
164 /* io prio of this group */
165 unsigned short ioprio, org_ioprio;
166 unsigned short ioprio_class, org_ioprio_class;
167
Jens Axboe7b679132008-05-30 12:23:07 +0200168 pid_t pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169};
170
Jens Axboe3b181522005-06-27 10:56:24 +0200171enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100172 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
173 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
174 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
175 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
176 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
177 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
178 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
179 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
180 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
Jens Axboe44f7c162007-01-19 11:51:58 +1100181 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200182 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200183};
184
185#define CFQ_CFQQ_FNS(name) \
186static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
187{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100188 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200189} \
190static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
191{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100192 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200193} \
194static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
195{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100196 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200197}
198
199CFQ_CFQQ_FNS(on_rr);
200CFQ_CFQQ_FNS(wait_request);
201CFQ_CFQQ_FNS(must_alloc);
202CFQ_CFQQ_FNS(must_alloc_slice);
203CFQ_CFQQ_FNS(must_dispatch);
204CFQ_CFQQ_FNS(fifo_expire);
205CFQ_CFQQ_FNS(idle_window);
206CFQ_CFQQ_FNS(prio_changed);
Jens Axboe53b03742006-07-28 09:48:51 +0200207CFQ_CFQQ_FNS(queue_new);
Jens Axboe44f7c162007-01-19 11:51:58 +1100208CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200209CFQ_CFQQ_FNS(sync);
Jens Axboe3b181522005-06-27 10:56:24 +0200210#undef CFQ_CFQQ_FNS
211
Jens Axboe7b679132008-05-30 12:23:07 +0200212#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
213 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
214#define cfq_log(cfqd, fmt, args...) \
215 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
216
Jens Axboe165125e2007-07-24 09:28:11 +0200217static void cfq_dispatch_insert(struct request_queue *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200218static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
Jens Axboefd0928d2008-01-24 08:52:45 +0100219 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100220static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200221 struct io_context *);
222
223static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
224 int is_sync)
225{
226 return cic->cfqq[!!is_sync];
227}
228
229static inline void cic_set_cfqq(struct cfq_io_context *cic,
230 struct cfq_queue *cfqq, int is_sync)
231{
232 cic->cfqq[!!is_sync] = cfqq;
233}
234
235/*
236 * We regard a request as SYNC, if it's either a read or has the SYNC bit
237 * set (in which case it could also be direct WRITE).
238 */
239static inline int cfq_bio_sync(struct bio *bio)
240{
241 if (bio_data_dir(bio) == READ || bio_sync(bio))
242 return 1;
243
244 return 0;
245}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700248 * scheduler run of queue, if there are requests pending and no one in the
249 * driver that will restart queueing
250 */
251static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
252{
Jens Axboe7b679132008-05-30 12:23:07 +0200253 if (cfqd->busy_queues) {
254 cfq_log(cfqd, "schedule dispatch");
Jens Axboe18887ad2008-07-28 13:08:45 +0200255 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200256 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700257}
258
Jens Axboe165125e2007-07-24 09:28:11 +0200259static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700260{
261 struct cfq_data *cfqd = q->elevator->elevator_data;
262
Jens Axboeb4878f22005-10-20 16:42:29 +0200263 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700264}
265
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100267 * Scale schedule slice based on io priority. Use the sync time slice only
268 * if a queue is marked sync and has sync io queued. A sync queue with async
269 * io only, should not get full sync slice length.
270 */
Jens Axboed9e76202007-04-20 14:27:50 +0200271static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
272 unsigned short prio)
273{
274 const int base_slice = cfqd->cfq_slice[sync];
275
276 WARN_ON(prio >= IOPRIO_BE_NR);
277
278 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
279}
280
Jens Axboe44f7c162007-01-19 11:51:58 +1100281static inline int
282cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
283{
Jens Axboed9e76202007-04-20 14:27:50 +0200284 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100285}
286
287static inline void
288cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
289{
290 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200291 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100292}
293
294/*
295 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
296 * isn't valid until the first request from the dispatch is activated
297 * and the slice time set.
298 */
299static inline int cfq_slice_used(struct cfq_queue *cfqq)
300{
301 if (cfq_cfqq_slice_new(cfqq))
302 return 0;
303 if (time_before(jiffies, cfqq->slice_end))
304 return 0;
305
306 return 1;
307}
308
309/*
Jens Axboe5e705372006-07-13 12:39:25 +0200310 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200312 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 */
Jens Axboe5e705372006-07-13 12:39:25 +0200314static struct request *
315cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316{
317 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200319#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
320#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
321 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322
Jens Axboe5e705372006-07-13 12:39:25 +0200323 if (rq1 == NULL || rq1 == rq2)
324 return rq2;
325 if (rq2 == NULL)
326 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200327
Jens Axboe5e705372006-07-13 12:39:25 +0200328 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
329 return rq1;
330 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
331 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200332 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
333 return rq1;
334 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
335 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336
Jens Axboe5e705372006-07-13 12:39:25 +0200337 s1 = rq1->sector;
338 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339
Jens Axboe6d048f52007-04-25 12:44:27 +0200340 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 /*
343 * by definition, 1KiB is 2 sectors
344 */
345 back_max = cfqd->cfq_back_max * 2;
346
347 /*
348 * Strict one way elevator _except_ in the case where we allow
349 * short backward seeks which are biased as twice the cost of a
350 * similar forward seek.
351 */
352 if (s1 >= last)
353 d1 = s1 - last;
354 else if (s1 + back_max >= last)
355 d1 = (last - s1) * cfqd->cfq_back_penalty;
356 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200357 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358
359 if (s2 >= last)
360 d2 = s2 - last;
361 else if (s2 + back_max >= last)
362 d2 = (last - s2) * cfqd->cfq_back_penalty;
363 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200364 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365
366 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367
Andreas Mohre8a99052006-03-28 08:59:49 +0200368 /*
369 * By doing switch() on the bit mask "wrap" we avoid having to
370 * check two variables for all permutations: --> faster!
371 */
372 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200373 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200374 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200375 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200376 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200377 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200378 else {
379 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200380 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200381 else
Jens Axboe5e705372006-07-13 12:39:25 +0200382 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200383 }
384
385 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200386 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200387 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200388 return rq2;
389 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200390 default:
391 /*
392 * Since both rqs are wrapped,
393 * start with the one that's further behind head
394 * (--> only *one* back seek required),
395 * since back seek takes more time than forward.
396 */
397 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200398 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399 else
Jens Axboe5e705372006-07-13 12:39:25 +0200400 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 }
402}
403
Jens Axboe498d3aa2007-04-26 12:54:48 +0200404/*
405 * The below is leftmost cache rbtree addon
406 */
Jens Axboe08717142008-01-28 11:38:15 +0100407static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200408{
409 if (!root->left)
410 root->left = rb_first(&root->rb);
411
Jens Axboe08717142008-01-28 11:38:15 +0100412 if (root->left)
413 return rb_entry(root->left, struct cfq_queue, rb_node);
414
415 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200416}
417
418static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
419{
420 if (root->left == n)
421 root->left = NULL;
422
423 rb_erase(n, &root->rb);
424 RB_CLEAR_NODE(n);
425}
426
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427/*
428 * would be nice to take fifo expire time into account as well
429 */
Jens Axboe5e705372006-07-13 12:39:25 +0200430static struct request *
431cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
432 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433{
Jens Axboe21183b02006-07-13 12:33:14 +0200434 struct rb_node *rbnext = rb_next(&last->rb_node);
435 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200436 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437
Jens Axboe21183b02006-07-13 12:33:14 +0200438 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439
440 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200441 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200442
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200444 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200445 else {
446 rbnext = rb_first(&cfqq->sort_list);
447 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200448 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200449 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
Jens Axboe21183b02006-07-13 12:33:14 +0200451 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452}
453
Jens Axboed9e76202007-04-20 14:27:50 +0200454static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
455 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456{
Jens Axboed9e76202007-04-20 14:27:50 +0200457 /*
458 * just an approximation, should be ok.
459 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200460 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
461 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200462}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463
Jens Axboe498d3aa2007-04-26 12:54:48 +0200464/*
465 * The cfqd->service_tree holds all pending cfq_queue's that have
466 * requests waiting to be processed. It is sorted in the order that
467 * we will service the queues.
468 */
Jens Axboed9e76202007-04-20 14:27:50 +0200469static void cfq_service_tree_add(struct cfq_data *cfqd,
Jens Axboeedd75ff2007-04-19 12:03:34 +0200470 struct cfq_queue *cfqq, int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200471{
Jens Axboe08717142008-01-28 11:38:15 +0100472 struct rb_node **p, *parent;
473 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +0200474 unsigned long rb_key;
Jens Axboe498d3aa2007-04-26 12:54:48 +0200475 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200476
Jens Axboe08717142008-01-28 11:38:15 +0100477 if (cfq_class_idle(cfqq)) {
478 rb_key = CFQ_IDLE_DELAY;
479 parent = rb_last(&cfqd->service_tree.rb);
480 if (parent && parent != &cfqq->rb_node) {
481 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
482 rb_key += __cfqq->rb_key;
483 } else
484 rb_key += jiffies;
485 } else if (!add_front) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200486 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
487 rb_key += cfqq->slice_resid;
488 cfqq->slice_resid = 0;
489 } else
490 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200491
492 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100493 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200494 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100495 */
Jens Axboed9e76202007-04-20 14:27:50 +0200496 if (rb_key == cfqq->rb_key)
497 return;
Jens Axboe53b03742006-07-28 09:48:51 +0200498
Jens Axboecc09e292007-04-26 12:53:50 +0200499 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200500 }
Jens Axboed9e76202007-04-20 14:27:50 +0200501
Jens Axboe498d3aa2007-04-26 12:54:48 +0200502 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +0100503 parent = NULL;
504 p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200505 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +0200506 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200507
Jens Axboed9e76202007-04-20 14:27:50 +0200508 parent = *p;
509 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
510
Jens Axboe0c534e02007-04-18 20:01:57 +0200511 /*
512 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200513 * preference to them. IDLE queues goes to the back.
514 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200515 */
516 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200517 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200518 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200519 n = &(*p)->rb_right;
520 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
521 n = &(*p)->rb_left;
522 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
523 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200524 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200525 n = &(*p)->rb_left;
526 else
527 n = &(*p)->rb_right;
528
529 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200530 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200531
532 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200533 }
534
Jens Axboecc09e292007-04-26 12:53:50 +0200535 if (left)
536 cfqd->service_tree.left = &cfqq->rb_node;
537
Jens Axboed9e76202007-04-20 14:27:50 +0200538 cfqq->rb_key = rb_key;
539 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200540 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541}
542
Jens Axboe498d3aa2007-04-26 12:54:48 +0200543/*
544 * Update cfqq's position in the service tree.
545 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200546static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200547{
Jens Axboe6d048f52007-04-25 12:44:27 +0200548 /*
549 * Resorting requires the cfqq to be on the RR list already.
550 */
Jens Axboe498d3aa2007-04-26 12:54:48 +0200551 if (cfq_cfqq_on_rr(cfqq))
Jens Axboeedd75ff2007-04-19 12:03:34 +0200552 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboe6d048f52007-04-25 12:44:27 +0200553}
554
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555/*
556 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200557 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558 */
Jens Axboefebffd62008-01-28 13:19:43 +0100559static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560{
Jens Axboe7b679132008-05-30 12:23:07 +0200561 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200562 BUG_ON(cfq_cfqq_on_rr(cfqq));
563 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 cfqd->busy_queues++;
565
Jens Axboeedd75ff2007-04-19 12:03:34 +0200566 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567}
568
Jens Axboe498d3aa2007-04-26 12:54:48 +0200569/*
570 * Called when the cfqq no longer has requests pending, remove it from
571 * the service tree.
572 */
Jens Axboefebffd62008-01-28 13:19:43 +0100573static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574{
Jens Axboe7b679132008-05-30 12:23:07 +0200575 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200576 BUG_ON(!cfq_cfqq_on_rr(cfqq));
577 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578
Jens Axboecc09e292007-04-26 12:53:50 +0200579 if (!RB_EMPTY_NODE(&cfqq->rb_node))
580 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboed9e76202007-04-20 14:27:50 +0200581
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 BUG_ON(!cfqd->busy_queues);
583 cfqd->busy_queues--;
584}
585
586/*
587 * rb tree support functions
588 */
Jens Axboefebffd62008-01-28 13:19:43 +0100589static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590{
Jens Axboe5e705372006-07-13 12:39:25 +0200591 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200592 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200593 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594
Jens Axboeb4878f22005-10-20 16:42:29 +0200595 BUG_ON(!cfqq->queued[sync]);
596 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597
Jens Axboe5e705372006-07-13 12:39:25 +0200598 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599
Jens Axboedd67d052006-06-21 09:36:18 +0200600 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200601 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602}
603
Jens Axboe5e705372006-07-13 12:39:25 +0200604static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605{
Jens Axboe5e705372006-07-13 12:39:25 +0200606 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe21183b02006-07-13 12:33:14 +0200608 struct request *__alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609
Jens Axboe5380a102006-07-13 12:37:56 +0200610 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611
612 /*
613 * looks a little odd, but the first insert might return an alias.
614 * if that happens, put the alias on the dispatch list
615 */
Jens Axboe21183b02006-07-13 12:33:14 +0200616 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200617 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100618
619 if (!cfq_cfqq_on_rr(cfqq))
620 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200621
622 /*
623 * check if this request is a better next-serve candidate
624 */
625 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
626 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627}
628
Jens Axboefebffd62008-01-28 13:19:43 +0100629static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630{
Jens Axboe5380a102006-07-13 12:37:56 +0200631 elv_rb_del(&cfqq->sort_list, rq);
632 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200633 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634}
635
Jens Axboe206dc692006-03-28 13:03:44 +0200636static struct request *
637cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638{
Jens Axboe206dc692006-03-28 13:03:44 +0200639 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200640 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200641 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642
Jens Axboe4ac845a2008-01-24 08:44:49 +0100643 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200644 if (!cic)
645 return NULL;
646
647 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200648 if (cfqq) {
649 sector_t sector = bio->bi_sector + bio_sectors(bio);
650
Jens Axboe21183b02006-07-13 12:33:14 +0200651 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200652 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 return NULL;
655}
656
Jens Axboe165125e2007-07-24 09:28:11 +0200657static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200658{
659 struct cfq_data *cfqd = q->elevator->elevator_data;
660
661 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +0200662 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
663 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +0200664
Jens Axboe6d048f52007-04-25 12:44:27 +0200665 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
Jens Axboeb4878f22005-10-20 16:42:29 +0200666}
667
Jens Axboe165125e2007-07-24 09:28:11 +0200668static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669{
Jens Axboe22e2c502005-06-27 10:55:12 +0200670 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671
Jens Axboeb4878f22005-10-20 16:42:29 +0200672 WARN_ON(!cfqd->rq_in_driver);
673 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +0200674 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
675 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676}
677
Jens Axboeb4878f22005-10-20 16:42:29 +0200678static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679{
Jens Axboe5e705372006-07-13 12:39:25 +0200680 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200681
Jens Axboe5e705372006-07-13 12:39:25 +0200682 if (cfqq->next_rq == rq)
683 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684
Jens Axboeb4878f22005-10-20 16:42:29 +0200685 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200686 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200687
Aaron Carroll45333d52008-08-26 15:52:36 +0200688 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +0200689 if (rq_is_meta(rq)) {
690 WARN_ON(!cfqq->meta_pending);
691 cfqq->meta_pending--;
692 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693}
694
Jens Axboe165125e2007-07-24 09:28:11 +0200695static int cfq_merge(struct request_queue *q, struct request **req,
696 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697{
698 struct cfq_data *cfqd = q->elevator->elevator_data;
699 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700
Jens Axboe206dc692006-03-28 13:03:44 +0200701 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200702 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200703 *req = __rq;
704 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 }
706
707 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708}
709
Jens Axboe165125e2007-07-24 09:28:11 +0200710static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200711 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712{
Jens Axboe21183b02006-07-13 12:33:14 +0200713 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200714 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715
Jens Axboe5e705372006-07-13 12:39:25 +0200716 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718}
719
720static void
Jens Axboe165125e2007-07-24 09:28:11 +0200721cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722 struct request *next)
723{
Jens Axboe22e2c502005-06-27 10:55:12 +0200724 /*
725 * reposition in fifo if next is older than rq
726 */
727 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
728 time_before(next->start_time, rq->start_time))
729 list_move(&rq->queuelist, &next->queuelist);
730
Jens Axboeb4878f22005-10-20 16:42:29 +0200731 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200732}
733
Jens Axboe165125e2007-07-24 09:28:11 +0200734static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100735 struct bio *bio)
736{
737 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200738 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100739 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100740
741 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100742 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100743 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200744 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100745 return 0;
746
747 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100748 * Lookup the cfqq that this bio will be queued with. Allow
749 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100750 */
Jens Axboe4ac845a2008-01-24 08:44:49 +0100751 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200752 if (!cic)
753 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100754
Vasily Tarasov91fac312007-04-25 12:29:51 +0200755 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100756 if (cfqq == RQ_CFQQ(rq))
757 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100758
Jens Axboeec8acb62007-01-02 18:32:11 +0100759 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100760}
761
Jens Axboefebffd62008-01-28 13:19:43 +0100762static void __cfq_set_active_queue(struct cfq_data *cfqd,
763 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +0200764{
765 if (cfqq) {
Jens Axboe7b679132008-05-30 12:23:07 +0200766 cfq_log_cfqq(cfqd, cfqq, "set_active");
Jens Axboe22e2c502005-06-27 10:55:12 +0200767 cfqq->slice_end = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200768 cfq_clear_cfqq_must_alloc_slice(cfqq);
769 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100770 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe1afba042007-04-17 12:47:55 +0200771 cfq_clear_cfqq_queue_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200772 }
773
774 cfqd->active_queue = cfqq;
775}
776
777/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100778 * current cfqq expired its slice (or was too idle), select new one
779 */
780static void
781__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200782 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100783{
Jens Axboe7b679132008-05-30 12:23:07 +0200784 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
785
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100786 if (cfq_cfqq_wait_request(cfqq))
787 del_timer(&cfqd->idle_slice_timer);
788
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100789 cfq_clear_cfqq_must_dispatch(cfqq);
790 cfq_clear_cfqq_wait_request(cfqq);
791
792 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200793 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100794 */
Jens Axboe7b679132008-05-30 12:23:07 +0200795 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +1100796 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200797 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
798 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100799
Jens Axboeedd75ff2007-04-19 12:03:34 +0200800 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100801
802 if (cfqq == cfqd->active_queue)
803 cfqd->active_queue = NULL;
804
805 if (cfqd->active_cic) {
806 put_io_context(cfqd->active_cic->ioc);
807 cfqd->active_cic = NULL;
808 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100809}
810
Jens Axboe6084cdd2007-04-23 08:25:00 +0200811static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100812{
813 struct cfq_queue *cfqq = cfqd->active_queue;
814
815 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200816 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100817}
818
Jens Axboe498d3aa2007-04-26 12:54:48 +0200819/*
820 * Get next queue for service. Unless we have a queue preemption,
821 * we'll simply select the first cfqq in the service tree.
822 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200823static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200824{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200825 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
826 return NULL;
827
Jens Axboe08717142008-01-28 11:38:15 +0100828 return cfq_rb_first(&cfqd->service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +0200829}
830
Jens Axboe498d3aa2007-04-26 12:54:48 +0200831/*
832 * Get and set a new active queue for service.
833 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200834static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
835{
836 struct cfq_queue *cfqq;
837
Jens Axboed9e76202007-04-20 14:27:50 +0200838 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200839 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200840 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200841}
842
Jens Axboed9e76202007-04-20 14:27:50 +0200843static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
844 struct request *rq)
845{
846 if (rq->sector >= cfqd->last_position)
847 return rq->sector - cfqd->last_position;
848 else
849 return cfqd->last_position - rq->sector;
850}
851
Jens Axboe6d048f52007-04-25 12:44:27 +0200852static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
853{
854 struct cfq_io_context *cic = cfqd->active_cic;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200855
Jens Axboe6d048f52007-04-25 12:44:27 +0200856 if (!sample_valid(cic->seek_samples))
857 return 0;
858
859 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
860}
861
Jens Axboed9e76202007-04-20 14:27:50 +0200862static int cfq_close_cooperator(struct cfq_data *cfq_data,
863 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200864{
Jens Axboe6d048f52007-04-25 12:44:27 +0200865 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200866 * We should notice if some of the queues are cooperating, eg
867 * working closely on the same area of the disk. In that case,
868 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +0200869 */
Jens Axboed9e76202007-04-20 14:27:50 +0200870 return 0;
Jens Axboe6d048f52007-04-25 12:44:27 +0200871}
872
873#define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
874
875static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200876{
Jens Axboe17926692007-01-19 11:59:30 +1100877 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +0200878 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100879 unsigned long sl;
880
Jens Axboea68bbdd2008-09-24 13:03:33 +0200881 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +0200882 * SSD device without seek penalty, disable idling. But only do so
883 * for devices that support queuing, otherwise we still have a problem
884 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +0200885 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +0200886 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +0200887 return;
888
Jens Axboedd67d052006-06-21 09:36:18 +0200889 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +0200890 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +0200891
892 /*
893 * idle is disabled, either manually or by past process history
894 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200895 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
896 return;
897
Jens Axboe22e2c502005-06-27 10:55:12 +0200898 /*
Jens Axboe7b679132008-05-30 12:23:07 +0200899 * still requests with the driver, don't idle
900 */
901 if (cfqd->rq_in_driver)
902 return;
903
904 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200905 * task has exited, don't wait
906 */
Jens Axboe206dc692006-03-28 13:03:44 +0200907 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +0100908 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +0200909 return;
910
911 /*
912 * See if this prio level has a good candidate
913 */
Jens Axboe1afba042007-04-17 12:47:55 +0200914 if (cfq_close_cooperator(cfqd, cfqq) &&
915 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
Jens Axboe6d048f52007-04-25 12:44:27 +0200916 return;
Jens Axboe22e2c502005-06-27 10:55:12 +0200917
Jens Axboe3b181522005-06-27 10:56:24 +0200918 cfq_mark_cfqq_must_dispatch(cfqq);
919 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200920
Jens Axboe206dc692006-03-28 13:03:44 +0200921 /*
922 * we don't want to idle for seeks, but we do want to allow
923 * fair distribution of slice time for a process doing back-to-back
924 * seeks. so allow a little bit of time for him to submit a new rq
925 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200926 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200927 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +0200928 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +0200929
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100930 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe7b679132008-05-30 12:23:07 +0200931 cfq_log(cfqd, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932}
933
Jens Axboe498d3aa2007-04-26 12:54:48 +0200934/*
935 * Move request from internal lists to the request queue dispatch list.
936 */
Jens Axboe165125e2007-07-24 09:28:11 +0200937static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938{
Jens Axboe3ed9a292007-04-23 08:33:33 +0200939 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200940 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200941
Jens Axboe7b679132008-05-30 12:23:07 +0200942 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
943
Jens Axboe5380a102006-07-13 12:37:56 +0200944 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200945 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +0200946 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +0200947
948 if (cfq_cfqq_sync(cfqq))
949 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950}
951
952/*
953 * return expired entry, or NULL to just start from scratch in rbtree
954 */
Jens Axboefebffd62008-01-28 13:19:43 +0100955static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956{
957 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200958 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200959 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960
Jens Axboe3b181522005-06-27 10:56:24 +0200961 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +1100963
964 cfq_mark_cfqq_fifo_expire(cfqq);
965
Jens Axboe89850f72006-07-22 16:48:31 +0200966 if (list_empty(&cfqq->fifo))
967 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968
Jens Axboe6d048f52007-04-25 12:44:27 +0200969 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +0200970 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971
Jens Axboe6d048f52007-04-25 12:44:27 +0200972 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
Jens Axboe7b679132008-05-30 12:23:07 +0200973 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974
Jens Axboe7b679132008-05-30 12:23:07 +0200975 cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200976 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977}
978
Jens Axboe22e2c502005-06-27 10:55:12 +0200979static inline int
980cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
981{
982 const int base_rq = cfqd->cfq_slice_async_rq;
983
984 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
985
986 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
987}
988
989/*
Jens Axboe498d3aa2007-04-26 12:54:48 +0200990 * Select a queue for service. If we have a current active queue,
991 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +0200992 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +0100993static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200994{
Jens Axboe22e2c502005-06-27 10:55:12 +0200995 struct cfq_queue *cfqq;
996
997 cfqq = cfqd->active_queue;
998 if (!cfqq)
999 goto new_queue;
1000
1001 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001002 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02001003 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001004 if (cfq_slice_used(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001005 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +02001006
1007 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001008 * The active queue has requests and isn't expired, allow it to
1009 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02001010 */
Jens Axboedd67d052006-06-21 09:36:18 +02001011 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001012 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02001013
1014 /*
1015 * No requests pending. If the active queue still has requests in
1016 * flight or is idling for a new request, allow either of these
1017 * conditions to happen (or time out) before selecting a new queue.
1018 */
Jens Axboecc197472007-04-20 20:45:39 +02001019 if (timer_pending(&cfqd->idle_slice_timer) ||
1020 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001021 cfqq = NULL;
1022 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001023 }
1024
Jens Axboe3b181522005-06-27 10:56:24 +02001025expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001026 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001027new_queue:
1028 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001029keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001030 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001031}
1032
Jens Axboe498d3aa2007-04-26 12:54:48 +02001033/*
1034 * Dispatch some requests from cfqq, moving them to the request queue
1035 * dispatch list.
1036 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001037static int
1038__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1039 int max_dispatch)
1040{
1041 int dispatched = 0;
1042
Jens Axboedd67d052006-06-21 09:36:18 +02001043 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001044
1045 do {
Jens Axboe5e705372006-07-13 12:39:25 +02001046 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001047
1048 /*
1049 * follow expired path, else get first next available
1050 */
Jens Axboefe094d92008-01-31 13:08:54 +01001051 rq = cfq_check_fifo(cfqq);
1052 if (rq == NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001053 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001054
1055 /*
1056 * finally, insert request into driver dispatch list
1057 */
Jens Axboe5e705372006-07-13 12:39:25 +02001058 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001059
Jens Axboe22e2c502005-06-27 10:55:12 +02001060 dispatched++;
1061
1062 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +02001063 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1064 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001065 }
1066
Jens Axboedd67d052006-06-21 09:36:18 +02001067 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001068 break;
1069
1070 } while (dispatched < max_dispatch);
1071
1072 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001073 * expire an async queue immediately if it has used up its slice. idle
1074 * queue always expire after 1 dispatch round.
1075 */
Jens Axboea9938002007-04-20 08:55:52 +02001076 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
Jens Axboe20e493a2007-04-23 08:26:36 +02001077 dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboea9938002007-04-20 08:55:52 +02001078 cfq_class_idle(cfqq))) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001079 cfqq->slice_end = jiffies + 1;
Jens Axboe6084cdd2007-04-23 08:25:00 +02001080 cfq_slice_expired(cfqd, 0);
Jens Axboe44f7c162007-01-19 11:51:58 +11001081 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001082
1083 return dispatched;
1084}
1085
Jens Axboefebffd62008-01-28 13:19:43 +01001086static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02001087{
1088 int dispatched = 0;
1089
1090 while (cfqq->next_rq) {
1091 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1092 dispatched++;
1093 }
1094
1095 BUG_ON(!list_empty(&cfqq->fifo));
1096 return dispatched;
1097}
1098
Jens Axboe498d3aa2007-04-26 12:54:48 +02001099/*
1100 * Drain our current requests. Used for barriers and when switching
1101 * io schedulers on-the-fly.
1102 */
Jens Axboed9e76202007-04-20 14:27:50 +02001103static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001104{
Jens Axboe08717142008-01-28 11:38:15 +01001105 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001106 int dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001107
Jens Axboe08717142008-01-28 11:38:15 +01001108 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
Jens Axboed9e76202007-04-20 14:27:50 +02001109 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001110
Jens Axboe6084cdd2007-04-23 08:25:00 +02001111 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001112
1113 BUG_ON(cfqd->busy_queues);
1114
Jens Axboe7b679132008-05-30 12:23:07 +02001115 cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001116 return dispatched;
1117}
1118
Jens Axboe165125e2007-07-24 09:28:11 +02001119static int cfq_dispatch_requests(struct request_queue *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120{
1121 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001122 struct cfq_queue *cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001123 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124
Jens Axboe22e2c502005-06-27 10:55:12 +02001125 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 return 0;
1127
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001128 if (unlikely(force))
1129 return cfq_forced_dispatch(cfqd);
1130
Jens Axboecaaa5f92006-06-16 11:23:00 +02001131 dispatched = 0;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001132 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001133 int max_dispatch;
1134
Jens Axboe3ed9a292007-04-23 08:33:33 +02001135 max_dispatch = cfqd->cfq_quantum;
1136 if (cfq_class_idle(cfqq))
1137 max_dispatch = 1;
1138
Jens Axboe30e0dc22008-10-20 15:44:28 +02001139 if (cfqq->dispatched >= max_dispatch && cfqd->busy_queues > 1)
1140 break;
Jens Axboe9ede2092007-01-19 12:11:44 +11001141
Jens Axboe3ed9a292007-04-23 08:33:33 +02001142 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1143 break;
1144
Jens Axboe3b181522005-06-27 10:56:24 +02001145 cfq_clear_cfqq_must_dispatch(cfqq);
1146 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001147 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148
Jens Axboecaaa5f92006-06-16 11:23:00 +02001149 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 }
1151
Jens Axboe7b679132008-05-30 12:23:07 +02001152 cfq_log(cfqd, "dispatched=%d", dispatched);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001153 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154}
1155
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156/*
Jens Axboe5e705372006-07-13 12:39:25 +02001157 * task holds one reference to the queue, dropped when task exits. each rq
1158 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 *
1160 * queue lock must be held here.
1161 */
1162static void cfq_put_queue(struct cfq_queue *cfqq)
1163{
Jens Axboe22e2c502005-06-27 10:55:12 +02001164 struct cfq_data *cfqd = cfqq->cfqd;
1165
1166 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167
1168 if (!atomic_dec_and_test(&cfqq->ref))
1169 return;
1170
Jens Axboe7b679132008-05-30 12:23:07 +02001171 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001173 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001174 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001176 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001177 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001178 cfq_schedule_dispatch(cfqd);
1179 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001180
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 kmem_cache_free(cfq_pool, cfqq);
1182}
1183
Jens Axboed6de8be2008-05-28 14:46:59 +02001184/*
1185 * Must always be called with the rcu_read_lock() held
1186 */
Jens Axboe07416d22008-05-07 09:17:12 +02001187static void
1188__call_for_each_cic(struct io_context *ioc,
1189 void (*func)(struct io_context *, struct cfq_io_context *))
1190{
1191 struct cfq_io_context *cic;
1192 struct hlist_node *n;
1193
1194 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1195 func(ioc, cic);
1196}
1197
Jens Axboe4ac845a2008-01-24 08:44:49 +01001198/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001199 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001200 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001201static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001202call_for_each_cic(struct io_context *ioc,
1203 void (*func)(struct io_context *, struct cfq_io_context *))
1204{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001205 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02001206 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001207 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001208}
Jens Axboe4ac845a2008-01-24 08:44:49 +01001209
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001210static void cfq_cic_free_rcu(struct rcu_head *head)
1211{
1212 struct cfq_io_context *cic;
1213
1214 cic = container_of(head, struct cfq_io_context, rcu_head);
1215
1216 kmem_cache_free(cfq_ioc_pool, cic);
1217 elv_ioc_count_dec(ioc_count);
1218
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001219 if (ioc_gone) {
1220 /*
1221 * CFQ scheduler is exiting, grab exit lock and check
1222 * the pending io context count. If it hits zero,
1223 * complete ioc_gone and set it back to NULL
1224 */
1225 spin_lock(&ioc_gone_lock);
1226 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1227 complete(ioc_gone);
1228 ioc_gone = NULL;
1229 }
1230 spin_unlock(&ioc_gone_lock);
1231 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001232}
1233
1234static void cfq_cic_free(struct cfq_io_context *cic)
1235{
1236 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001237}
1238
1239static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1240{
1241 unsigned long flags;
1242
1243 BUG_ON(!cic->dead_key);
1244
1245 spin_lock_irqsave(&ioc->lock, flags);
1246 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01001247 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001248 spin_unlock_irqrestore(&ioc->lock, flags);
1249
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001250 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001251}
1252
Jens Axboed6de8be2008-05-28 14:46:59 +02001253/*
1254 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1255 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1256 * and ->trim() which is called with the task lock held
1257 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001258static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001260 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001261 * ioc->refcount is zero here, or we are called from elv_unregister(),
1262 * so no more cic's are allowed to be linked into this ioc. So it
1263 * should be ok to iterate over the known list, we will see all cic's
1264 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001265 */
Jens Axboe07416d22008-05-07 09:17:12 +02001266 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267}
1268
Jens Axboe89850f72006-07-22 16:48:31 +02001269static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1270{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001271 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001272 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001273 cfq_schedule_dispatch(cfqd);
1274 }
Jens Axboe89850f72006-07-22 16:48:31 +02001275
1276 cfq_put_queue(cfqq);
1277}
1278
1279static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1280 struct cfq_io_context *cic)
1281{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001282 struct io_context *ioc = cic->ioc;
1283
Jens Axboefc463792006-08-29 09:05:44 +02001284 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001285
1286 /*
1287 * Make sure key == NULL is seen for dead queues
1288 */
Jens Axboefc463792006-08-29 09:05:44 +02001289 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001290 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02001291 cic->key = NULL;
1292
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001293 if (ioc->ioc_data == cic)
1294 rcu_assign_pointer(ioc->ioc_data, NULL);
1295
Jens Axboe89850f72006-07-22 16:48:31 +02001296 if (cic->cfqq[ASYNC]) {
1297 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1298 cic->cfqq[ASYNC] = NULL;
1299 }
1300
1301 if (cic->cfqq[SYNC]) {
1302 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1303 cic->cfqq[SYNC] = NULL;
1304 }
Jens Axboe89850f72006-07-22 16:48:31 +02001305}
1306
Jens Axboe4ac845a2008-01-24 08:44:49 +01001307static void cfq_exit_single_io_context(struct io_context *ioc,
1308 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001309{
Al Viro478a82b2006-03-18 13:25:24 -05001310 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001311
Jens Axboe89850f72006-07-22 16:48:31 +02001312 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001313 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001314 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02001315
Jens Axboe4ac845a2008-01-24 08:44:49 +01001316 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01001317
1318 /*
1319 * Ensure we get a fresh copy of the ->key to prevent
1320 * race between exiting task and queue
1321 */
1322 smp_read_barrier_depends();
1323 if (cic->key)
1324 __cfq_exit_single_io_context(cfqd, cic);
1325
Jens Axboe4ac845a2008-01-24 08:44:49 +01001326 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05001327 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001328}
1329
Jens Axboe498d3aa2007-04-26 12:54:48 +02001330/*
1331 * The process that ioc belongs to has exited, we need to clean up
1332 * and put the internal structures we have that belongs to that process.
1333 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001334static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001336 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337}
1338
Jens Axboe22e2c502005-06-27 10:55:12 +02001339static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001340cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341{
Jens Axboeb5deef92006-07-19 23:39:40 +02001342 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343
Christoph Lameter94f60302007-07-17 04:03:29 -07001344 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1345 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001347 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001348 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01001349 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001350 cic->dtor = cfq_free_io_context;
1351 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001352 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 }
1354
1355 return cic;
1356}
1357
Jens Axboefd0928d2008-01-24 08:52:45 +01001358static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02001359{
1360 struct task_struct *tsk = current;
1361 int ioprio_class;
1362
Jens Axboe3b181522005-06-27 10:56:24 +02001363 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001364 return;
1365
Jens Axboefd0928d2008-01-24 08:52:45 +01001366 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001367 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01001368 default:
1369 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1370 case IOPRIO_CLASS_NONE:
1371 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02001372 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01001373 */
1374 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02001375 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01001376 break;
1377 case IOPRIO_CLASS_RT:
1378 cfqq->ioprio = task_ioprio(ioc);
1379 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1380 break;
1381 case IOPRIO_CLASS_BE:
1382 cfqq->ioprio = task_ioprio(ioc);
1383 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1384 break;
1385 case IOPRIO_CLASS_IDLE:
1386 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1387 cfqq->ioprio = 7;
1388 cfq_clear_cfqq_idle_window(cfqq);
1389 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02001390 }
1391
1392 /*
1393 * keep track of original prio settings in case we have to temporarily
1394 * elevate the priority of this queue
1395 */
1396 cfqq->org_ioprio = cfqq->ioprio;
1397 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001398 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001399}
1400
Jens Axboefebffd62008-01-28 13:19:43 +01001401static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001402{
Al Viro478a82b2006-03-18 13:25:24 -05001403 struct cfq_data *cfqd = cic->key;
1404 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001405 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001406
Jens Axboecaaa5f92006-06-16 11:23:00 +02001407 if (unlikely(!cfqd))
1408 return;
1409
Jens Axboec1b707d2006-10-30 19:54:23 +01001410 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001411
1412 cfqq = cic->cfqq[ASYNC];
1413 if (cfqq) {
1414 struct cfq_queue *new_cfqq;
Jens Axboefd0928d2008-01-24 08:52:45 +01001415 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001416 if (new_cfqq) {
1417 cic->cfqq[ASYNC] = new_cfqq;
1418 cfq_put_queue(cfqq);
1419 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001420 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001421
1422 cfqq = cic->cfqq[SYNC];
1423 if (cfqq)
1424 cfq_mark_cfqq_prio_changed(cfqq);
1425
Jens Axboec1b707d2006-10-30 19:54:23 +01001426 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001427}
1428
Jens Axboefc463792006-08-29 09:05:44 +02001429static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001431 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02001432 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433}
1434
1435static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001436cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01001437 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001440 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441
1442retry:
Jens Axboe4ac845a2008-01-24 08:44:49 +01001443 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001444 /* cic always exists here */
1445 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446
1447 if (!cfqq) {
1448 if (new_cfqq) {
1449 cfqq = new_cfqq;
1450 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001451 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001452 /*
1453 * Inform the allocator of the fact that we will
1454 * just repeat this allocation if it fails, to allow
1455 * the allocator to do whatever it needs to attempt to
1456 * free memory.
1457 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001459 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1460 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1461 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462 spin_lock_irq(cfqd->queue->queue_lock);
1463 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001464 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001465 cfqq = kmem_cache_alloc_node(cfq_pool,
1466 gfp_mask | __GFP_ZERO,
1467 cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001468 if (!cfqq)
1469 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001470 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471
Jens Axboed9e76202007-04-20 14:27:50 +02001472 RB_CLEAR_NODE(&cfqq->rb_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001473 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 atomic_set(&cfqq->ref, 0);
1476 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001477
Jens Axboe3b181522005-06-27 10:56:24 +02001478 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b03742006-07-28 09:48:51 +02001479 cfq_mark_cfqq_queue_new(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001480
Jens Axboefd0928d2008-01-24 08:52:45 +01001481 cfq_init_prio_data(cfqq, ioc);
Jens Axboe08717142008-01-28 11:38:15 +01001482
1483 if (is_sync) {
1484 if (!cfq_class_idle(cfqq))
1485 cfq_mark_cfqq_idle_window(cfqq);
1486 cfq_mark_cfqq_sync(cfqq);
1487 }
Jens Axboe7b679132008-05-30 12:23:07 +02001488 cfqq->pid = current->pid;
1489 cfq_log_cfqq(cfqd, cfqq, "alloced");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 }
1491
1492 if (new_cfqq)
1493 kmem_cache_free(cfq_pool, new_cfqq);
1494
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495out:
1496 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1497 return cfqq;
1498}
1499
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001500static struct cfq_queue **
1501cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1502{
Jens Axboefe094d92008-01-31 13:08:54 +01001503 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001504 case IOPRIO_CLASS_RT:
1505 return &cfqd->async_cfqq[0][ioprio];
1506 case IOPRIO_CLASS_BE:
1507 return &cfqd->async_cfqq[1][ioprio];
1508 case IOPRIO_CLASS_IDLE:
1509 return &cfqd->async_idle_cfqq;
1510 default:
1511 BUG();
1512 }
1513}
1514
Jens Axboe15c31be2007-07-10 13:43:25 +02001515static struct cfq_queue *
Jens Axboefd0928d2008-01-24 08:52:45 +01001516cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02001517 gfp_t gfp_mask)
1518{
Jens Axboefd0928d2008-01-24 08:52:45 +01001519 const int ioprio = task_ioprio(ioc);
1520 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001521 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001522 struct cfq_queue *cfqq = NULL;
1523
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001524 if (!is_sync) {
1525 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1526 cfqq = *async_cfqq;
1527 }
1528
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001529 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01001530 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001531 if (!cfqq)
1532 return NULL;
1533 }
Jens Axboe15c31be2007-07-10 13:43:25 +02001534
1535 /*
1536 * pin the queue now that it's allocated, scheduler exit will prune it
1537 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001538 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001539 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001540 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001541 }
1542
1543 atomic_inc(&cfqq->ref);
1544 return cfqq;
1545}
1546
Jens Axboe498d3aa2007-04-26 12:54:48 +02001547/*
1548 * We drop cfq io contexts lazily, so we may find a dead one.
1549 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001550static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001551cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1552 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001553{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001554 unsigned long flags;
1555
Jens Axboefc463792006-08-29 09:05:44 +02001556 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001557
Jens Axboe4ac845a2008-01-24 08:44:49 +01001558 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02001559
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001560 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001561
1562 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01001563 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001564 spin_unlock_irqrestore(&ioc->lock, flags);
1565
1566 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001567}
1568
Jens Axboee2d74ac2006-03-28 08:59:01 +02001569static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01001570cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001571{
Jens Axboee2d74ac2006-03-28 08:59:01 +02001572 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001573 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001574 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001575
Vasily Tarasov91fac312007-04-25 12:29:51 +02001576 if (unlikely(!ioc))
1577 return NULL;
1578
Jens Axboed6de8be2008-05-28 14:46:59 +02001579 rcu_read_lock();
1580
Jens Axboe597bc482007-04-24 21:23:53 +02001581 /*
1582 * we maintain a last-hit cache, to avoid browsing over the tree
1583 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001584 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02001585 if (cic && cic->key == cfqd) {
1586 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02001587 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001588 }
Jens Axboe597bc482007-04-24 21:23:53 +02001589
Jens Axboe4ac845a2008-01-24 08:44:49 +01001590 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001591 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1592 rcu_read_unlock();
1593 if (!cic)
1594 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001595 /* ->key must be copied to avoid race with cfq_exit_queue() */
1596 k = cic->key;
1597 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001598 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001599 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001600 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001601 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001602
Jens Axboed6de8be2008-05-28 14:46:59 +02001603 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001604 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001605 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001606 break;
1607 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001608
Jens Axboe4ac845a2008-01-24 08:44:49 +01001609 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001610}
1611
Jens Axboe4ac845a2008-01-24 08:44:49 +01001612/*
1613 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1614 * the process specific cfq io context when entered from the block layer.
1615 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1616 */
Jens Axboefebffd62008-01-28 13:19:43 +01001617static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1618 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001619{
Jens Axboe0261d682006-10-30 19:07:48 +01001620 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001621 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001622
Jens Axboe4ac845a2008-01-24 08:44:49 +01001623 ret = radix_tree_preload(gfp_mask);
1624 if (!ret) {
1625 cic->ioc = ioc;
1626 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001627
Jens Axboe4ac845a2008-01-24 08:44:49 +01001628 spin_lock_irqsave(&ioc->lock, flags);
1629 ret = radix_tree_insert(&ioc->radix_root,
1630 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01001631 if (!ret)
1632 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001633 spin_unlock_irqrestore(&ioc->lock, flags);
1634
1635 radix_tree_preload_end();
1636
1637 if (!ret) {
1638 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1639 list_add(&cic->queue_list, &cfqd->cic_list);
1640 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001641 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001642 }
1643
Jens Axboe4ac845a2008-01-24 08:44:49 +01001644 if (ret)
1645 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02001646
Jens Axboe4ac845a2008-01-24 08:44:49 +01001647 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001648}
1649
Jens Axboe22e2c502005-06-27 10:55:12 +02001650/*
1651 * Setup general io context and cfq io context. There can be several cfq
1652 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001653 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001654 */
1655static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001656cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657{
Jens Axboe22e2c502005-06-27 10:55:12 +02001658 struct io_context *ioc = NULL;
1659 struct cfq_io_context *cic;
1660
1661 might_sleep_if(gfp_mask & __GFP_WAIT);
1662
Jens Axboeb5deef92006-07-19 23:39:40 +02001663 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001664 if (!ioc)
1665 return NULL;
1666
Jens Axboe4ac845a2008-01-24 08:44:49 +01001667 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001668 if (cic)
1669 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001670
Jens Axboee2d74ac2006-03-28 08:59:01 +02001671 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1672 if (cic == NULL)
1673 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001674
Jens Axboe4ac845a2008-01-24 08:44:49 +01001675 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1676 goto err_free;
1677
Jens Axboe22e2c502005-06-27 10:55:12 +02001678out:
Jens Axboefc463792006-08-29 09:05:44 +02001679 smp_read_barrier_depends();
1680 if (unlikely(ioc->ioprio_changed))
1681 cfq_ioc_set_ioprio(ioc);
1682
Jens Axboe22e2c502005-06-27 10:55:12 +02001683 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001684err_free:
1685 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001686err:
1687 put_io_context(ioc);
1688 return NULL;
1689}
1690
1691static void
1692cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1693{
Jens Axboeaaf12282007-01-19 11:30:16 +11001694 unsigned long elapsed = jiffies - cic->last_end_request;
1695 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001696
1697 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1698 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1699 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1700}
1701
Jens Axboe206dc692006-03-28 13:03:44 +02001702static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001703cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1704 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001705{
1706 sector_t sdist;
1707 u64 total;
1708
Jens Axboe5e705372006-07-13 12:39:25 +02001709 if (cic->last_request_pos < rq->sector)
1710 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001711 else
Jens Axboe5e705372006-07-13 12:39:25 +02001712 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001713
1714 /*
1715 * Don't allow the seek distance to get too large from the
1716 * odd fragment, pagein, etc
1717 */
1718 if (cic->seek_samples <= 60) /* second&third seek */
1719 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1720 else
1721 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1722
1723 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1724 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1725 total = cic->seek_total + (cic->seek_samples/2);
1726 do_div(total, cic->seek_samples);
1727 cic->seek_mean = (sector_t)total;
1728}
Jens Axboe22e2c502005-06-27 10:55:12 +02001729
1730/*
1731 * Disable idle window if the process thinks too long or seeks so much that
1732 * it doesn't matter
1733 */
1734static void
1735cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1736 struct cfq_io_context *cic)
1737{
Jens Axboe7b679132008-05-30 12:23:07 +02001738 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02001739
Jens Axboe08717142008-01-28 11:38:15 +01001740 /*
1741 * Don't idle for async or idle io prio class
1742 */
1743 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02001744 return;
1745
Jens Axboec265a7f2008-06-26 13:49:33 +02001746 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001747
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001748 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Jens Axboecaaa5f92006-06-16 11:23:00 +02001749 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001750 enable_idle = 0;
1751 else if (sample_valid(cic->ttime_samples)) {
1752 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1753 enable_idle = 0;
1754 else
1755 enable_idle = 1;
1756 }
1757
Jens Axboe7b679132008-05-30 12:23:07 +02001758 if (old_idle != enable_idle) {
1759 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1760 if (enable_idle)
1761 cfq_mark_cfqq_idle_window(cfqq);
1762 else
1763 cfq_clear_cfqq_idle_window(cfqq);
1764 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001765}
1766
Jens Axboe22e2c502005-06-27 10:55:12 +02001767/*
1768 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1769 * no or if we aren't sure, a 1 will cause a preempt.
1770 */
1771static int
1772cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001773 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001774{
Jens Axboe6d048f52007-04-25 12:44:27 +02001775 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001776
Jens Axboe6d048f52007-04-25 12:44:27 +02001777 cfqq = cfqd->active_queue;
1778 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001779 return 0;
1780
Jens Axboe6d048f52007-04-25 12:44:27 +02001781 if (cfq_slice_used(cfqq))
1782 return 1;
1783
1784 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001785 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001786
1787 if (cfq_class_idle(cfqq))
1788 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001789
Jens Axboe22e2c502005-06-27 10:55:12 +02001790 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02001791 * if the new request is sync, but the currently running queue is
1792 * not, let the sync request have priority.
1793 */
Jens Axboe5e705372006-07-13 12:39:25 +02001794 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001795 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001796
Jens Axboe374f84a2006-07-23 01:42:19 +02001797 /*
1798 * So both queues are sync. Let the new request get disk time if
1799 * it's a metadata request and the current queue is doing regular IO.
1800 */
1801 if (rq_is_meta(rq) && !cfqq->meta_pending)
1802 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001803
Jens Axboe1e3335d2007-02-14 19:59:49 +01001804 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1805 return 0;
1806
1807 /*
1808 * if this request is as-good as one we would expect from the
1809 * current cfqq, let it preempt
1810 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001811 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01001812 return 1;
1813
Jens Axboe22e2c502005-06-27 10:55:12 +02001814 return 0;
1815}
1816
1817/*
1818 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1819 * let it have half of its nominal slice.
1820 */
1821static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1822{
Jens Axboe7b679132008-05-30 12:23:07 +02001823 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02001824 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001825
Jens Axboebf572252006-07-19 20:29:12 +02001826 /*
1827 * Put the new queue at the front of the of the current list,
1828 * so we know that it will be selected next.
1829 */
1830 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02001831
1832 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02001833
Jens Axboe44f7c162007-01-19 11:51:58 +11001834 cfqq->slice_end = 0;
1835 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001836}
1837
1838/*
Jens Axboe5e705372006-07-13 12:39:25 +02001839 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001840 * something we should do about it
1841 */
1842static void
Jens Axboe5e705372006-07-13 12:39:25 +02001843cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1844 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001845{
Jens Axboe5e705372006-07-13 12:39:25 +02001846 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001847
Aaron Carroll45333d52008-08-26 15:52:36 +02001848 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02001849 if (rq_is_meta(rq))
1850 cfqq->meta_pending++;
1851
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001852 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02001853 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001854 cfq_update_idle_window(cfqd, cfqq, cic);
1855
Jens Axboe5e705372006-07-13 12:39:25 +02001856 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001857
1858 if (cfqq == cfqd->active_queue) {
1859 /*
1860 * if we are waiting for a request for this queue, let it rip
1861 * immediately and flag that we must not expire this queue
1862 * just now
1863 */
Jens Axboe3b181522005-06-27 10:56:24 +02001864 if (cfq_cfqq_wait_request(cfqq)) {
1865 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001866 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001867 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001868 }
Jens Axboe5e705372006-07-13 12:39:25 +02001869 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001870 /*
1871 * not the active queue - expire current slice if it is
1872 * idle and has expired it's mean thinktime or this new queue
1873 * has some old slice time left and is of higher priority
1874 */
1875 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001876 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001877 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001878 }
1879}
1880
Jens Axboe165125e2007-07-24 09:28:11 +02001881static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001882{
Jens Axboeb4878f22005-10-20 16:42:29 +02001883 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001884 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001885
Jens Axboe7b679132008-05-30 12:23:07 +02001886 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01001887 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888
Jens Axboe5e705372006-07-13 12:39:25 +02001889 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890
Jens Axboe22e2c502005-06-27 10:55:12 +02001891 list_add_tail(&rq->queuelist, &cfqq->fifo);
1892
Jens Axboe5e705372006-07-13 12:39:25 +02001893 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894}
1895
Aaron Carroll45333d52008-08-26 15:52:36 +02001896/*
1897 * Update hw_tag based on peak queue depth over 50 samples under
1898 * sufficient load.
1899 */
1900static void cfq_update_hw_tag(struct cfq_data *cfqd)
1901{
1902 if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
1903 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
1904
1905 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
1906 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
1907 return;
1908
1909 if (cfqd->hw_tag_samples++ < 50)
1910 return;
1911
1912 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
1913 cfqd->hw_tag = 1;
1914 else
1915 cfqd->hw_tag = 0;
1916
1917 cfqd->hw_tag_samples = 0;
1918 cfqd->rq_in_driver_peak = 0;
1919}
1920
Jens Axboe165125e2007-07-24 09:28:11 +02001921static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922{
Jens Axboe5e705372006-07-13 12:39:25 +02001923 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001924 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001925 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001926 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927
Jens Axboeb4878f22005-10-20 16:42:29 +02001928 now = jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001929 cfq_log_cfqq(cfqd, cfqq, "complete");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930
Aaron Carroll45333d52008-08-26 15:52:36 +02001931 cfq_update_hw_tag(cfqd);
1932
Jens Axboeb4878f22005-10-20 16:42:29 +02001933 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02001934 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02001935 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02001936 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937
Jens Axboe3ed9a292007-04-23 08:33:33 +02001938 if (cfq_cfqq_sync(cfqq))
1939 cfqd->sync_flight--;
1940
Jens Axboeb4878f22005-10-20 16:42:29 +02001941 if (!cfq_class_idle(cfqq))
1942 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001943
Jens Axboecaaa5f92006-06-16 11:23:00 +02001944 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001945 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001946
1947 /*
1948 * If this is the active queue, check if it needs to be expired,
1949 * or if we want to idle in case it has no pending requests.
1950 */
1951 if (cfqd->active_queue == cfqq) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001952 if (cfq_cfqq_slice_new(cfqq)) {
1953 cfq_set_prio_slice(cfqd, cfqq);
1954 cfq_clear_cfqq_slice_new(cfqq);
1955 }
Jens Axboe08717142008-01-28 11:38:15 +01001956 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02001957 cfq_slice_expired(cfqd, 1);
Jens Axboe6d048f52007-04-25 12:44:27 +02001958 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1959 cfq_arm_slice_timer(cfqd);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001960 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001961
1962 if (!cfqd->rq_in_driver)
1963 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964}
1965
Jens Axboe22e2c502005-06-27 10:55:12 +02001966/*
1967 * we temporarily boost lower priority queues if they are holding fs exclusive
1968 * resources. they are boosted to normal prio (CLASS_BE/4)
1969 */
1970static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971{
Jens Axboe22e2c502005-06-27 10:55:12 +02001972 if (has_fs_excl()) {
1973 /*
1974 * boost idle prio on transactions that would lock out other
1975 * users of the filesystem
1976 */
1977 if (cfq_class_idle(cfqq))
1978 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1979 if (cfqq->ioprio > IOPRIO_NORM)
1980 cfqq->ioprio = IOPRIO_NORM;
1981 } else {
1982 /*
1983 * check if we need to unboost the queue
1984 */
1985 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1986 cfqq->ioprio_class = cfqq->org_ioprio_class;
1987 if (cfqq->ioprio != cfqq->org_ioprio)
1988 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001990}
1991
Jens Axboe89850f72006-07-22 16:48:31 +02001992static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001993{
Jens Axboe3b181522005-06-27 10:56:24 +02001994 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07001995 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02001996 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001997 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02001998 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001999
2000 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02002001}
2002
Jens Axboe165125e2007-07-24 09:28:11 +02002003static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02002004{
2005 struct cfq_data *cfqd = q->elevator->elevator_data;
2006 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002007 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02002008 struct cfq_queue *cfqq;
2009
2010 /*
2011 * don't force setup of a queue from here, as a call to may_queue
2012 * does not necessarily imply that a request actually will be queued.
2013 * so just lookup a possibly existing queue, or return 'may queue'
2014 * if that fails
2015 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002016 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002017 if (!cic)
2018 return ELV_MQUEUE_MAY;
2019
2020 cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02002021 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002022 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02002023 cfq_prio_boost(cfqq);
2024
Jens Axboe89850f72006-07-22 16:48:31 +02002025 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002026 }
2027
2028 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029}
2030
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031/*
2032 * queue lock held here
2033 */
Jens Axboebb37b942006-12-01 10:42:33 +01002034static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035{
Jens Axboe5e705372006-07-13 12:39:25 +02002036 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037
Jens Axboe5e705372006-07-13 12:39:25 +02002038 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002039 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040
Jens Axboe22e2c502005-06-27 10:55:12 +02002041 BUG_ON(!cfqq->allocated[rw]);
2042 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043
Jens Axboe5e705372006-07-13 12:39:25 +02002044 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02002047 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049 cfq_put_queue(cfqq);
2050 }
2051}
2052
2053/*
Jens Axboe22e2c502005-06-27 10:55:12 +02002054 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055 */
Jens Axboe22e2c502005-06-27 10:55:12 +02002056static int
Jens Axboe165125e2007-07-24 09:28:11 +02002057cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058{
2059 struct cfq_data *cfqd = q->elevator->elevator_data;
2060 struct cfq_io_context *cic;
2061 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01002062 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002063 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064 unsigned long flags;
2065
2066 might_sleep_if(gfp_mask & __GFP_WAIT);
2067
Jens Axboee2d74ac2006-03-28 08:59:01 +02002068 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02002069
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070 spin_lock_irqsave(q->queue_lock, flags);
2071
Jens Axboe22e2c502005-06-27 10:55:12 +02002072 if (!cic)
2073 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074
Vasily Tarasov91fac312007-04-25 12:29:51 +02002075 cfqq = cic_to_cfqq(cic, is_sync);
2076 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002077 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002078
Jens Axboe22e2c502005-06-27 10:55:12 +02002079 if (!cfqq)
2080 goto queue_fail;
2081
Vasily Tarasov91fac312007-04-25 12:29:51 +02002082 cic_set_cfqq(cic, cfqq, is_sync);
2083 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084
2085 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02002086 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002087 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02002088
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089 spin_unlock_irqrestore(q->queue_lock, flags);
2090
Jens Axboe5e705372006-07-13 12:39:25 +02002091 rq->elevator_private = cic;
2092 rq->elevator_private2 = cfqq;
2093 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02002094
Jens Axboe22e2c502005-06-27 10:55:12 +02002095queue_fail:
2096 if (cic)
2097 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02002098
Jens Axboe3b181522005-06-27 10:56:24 +02002099 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02002101 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102 return 1;
2103}
2104
David Howells65f27f32006-11-22 14:55:48 +00002105static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02002106{
David Howells65f27f32006-11-22 14:55:48 +00002107 struct cfq_data *cfqd =
2108 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02002109 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002110 unsigned long flags;
2111
2112 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02002113 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02002114 spin_unlock_irqrestore(q->queue_lock, flags);
2115}
2116
2117/*
2118 * Timer running if the active_queue is currently idling inside its time slice
2119 */
2120static void cfq_idle_slice_timer(unsigned long data)
2121{
2122 struct cfq_data *cfqd = (struct cfq_data *) data;
2123 struct cfq_queue *cfqq;
2124 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002125 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002126
Jens Axboe7b679132008-05-30 12:23:07 +02002127 cfq_log(cfqd, "idle timer fired");
2128
Jens Axboe22e2c502005-06-27 10:55:12 +02002129 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2130
Jens Axboefe094d92008-01-31 13:08:54 +01002131 cfqq = cfqd->active_queue;
2132 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002133 timed_out = 0;
2134
Jens Axboe22e2c502005-06-27 10:55:12 +02002135 /*
2136 * expired
2137 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002138 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002139 goto expire;
2140
2141 /*
2142 * only expire and reinvoke request handler, if there are
2143 * other queues with pending requests
2144 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002145 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002146 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002147
2148 /*
2149 * not expired and it has a request pending, let it dispatch
2150 */
Jens Axboedd67d052006-06-21 09:36:18 +02002151 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002152 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002153 goto out_kick;
2154 }
2155 }
2156expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002157 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002158out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002159 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002160out_cont:
2161 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2162}
2163
Jens Axboe3b181522005-06-27 10:56:24 +02002164static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2165{
2166 del_timer_sync(&cfqd->idle_slice_timer);
Cheng Renquan64d01dc2008-12-03 12:41:39 +01002167 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002168}
Jens Axboe22e2c502005-06-27 10:55:12 +02002169
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002170static void cfq_put_async_queues(struct cfq_data *cfqd)
2171{
2172 int i;
2173
2174 for (i = 0; i < IOPRIO_BE_NR; i++) {
2175 if (cfqd->async_cfqq[0][i])
2176 cfq_put_queue(cfqd->async_cfqq[0][i]);
2177 if (cfqd->async_cfqq[1][i])
2178 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002179 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002180
2181 if (cfqd->async_idle_cfqq)
2182 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002183}
2184
Jens Axboeb374d182008-10-31 10:05:07 +01002185static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186{
Jens Axboe22e2c502005-06-27 10:55:12 +02002187 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002188 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002189
Jens Axboe3b181522005-06-27 10:56:24 +02002190 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002191
Al Virod9ff4182006-03-18 13:51:22 -05002192 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002193
Al Virod9ff4182006-03-18 13:51:22 -05002194 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002195 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002196
2197 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002198 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2199 struct cfq_io_context,
2200 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002201
2202 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002203 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002204
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002205 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002206
Al Virod9ff4182006-03-18 13:51:22 -05002207 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002208
2209 cfq_shutdown_timer_wq(cfqd);
2210
Al Viroa90d7422006-03-18 12:05:37 -05002211 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212}
2213
Jens Axboe165125e2007-07-24 09:28:11 +02002214static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215{
2216 struct cfq_data *cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217
Christoph Lameter94f60302007-07-17 04:03:29 -07002218 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002220 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221
Jens Axboecc09e292007-04-26 12:53:50 +02002222 cfqd->service_tree = CFQ_RB_ROOT;
Al Virod9ff4182006-03-18 13:51:22 -05002223 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226
Jens Axboe22e2c502005-06-27 10:55:12 +02002227 init_timer(&cfqd->idle_slice_timer);
2228 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2229 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2230
David Howells65f27f32006-11-22 14:55:48 +00002231 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002232
Oleg Nesterovb70c8642007-11-07 09:46:13 +01002233 cfqd->last_end_request = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002235 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2236 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 cfqd->cfq_back_max = cfq_back_max;
2238 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002239 cfqd->cfq_slice[0] = cfq_slice_async;
2240 cfqd->cfq_slice[1] = cfq_slice_sync;
2241 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2242 cfqd->cfq_slice_idle = cfq_slice_idle;
Aaron Carroll45333d52008-08-26 15:52:36 +02002243 cfqd->hw_tag = 1;
Jens Axboe3b181522005-06-27 10:56:24 +02002244
Jens Axboebc1c1162006-06-08 08:49:06 +02002245 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246}
2247
2248static void cfq_slab_kill(void)
2249{
Jens Axboed6de8be2008-05-28 14:46:59 +02002250 /*
2251 * Caller already ensured that pending RCU callbacks are completed,
2252 * so we should have no busy allocations at this point.
2253 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254 if (cfq_pool)
2255 kmem_cache_destroy(cfq_pool);
2256 if (cfq_ioc_pool)
2257 kmem_cache_destroy(cfq_ioc_pool);
2258}
2259
2260static int __init cfq_slab_setup(void)
2261{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002262 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 if (!cfq_pool)
2264 goto fail;
2265
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002266 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 if (!cfq_ioc_pool)
2268 goto fail;
2269
2270 return 0;
2271fail:
2272 cfq_slab_kill();
2273 return -ENOMEM;
2274}
2275
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276/*
2277 * sysfs parts below -->
2278 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279static ssize_t
2280cfq_var_show(unsigned int var, char *page)
2281{
2282 return sprintf(page, "%d\n", var);
2283}
2284
2285static ssize_t
2286cfq_var_store(unsigned int *var, const char *page, size_t count)
2287{
2288 char *p = (char *) page;
2289
2290 *var = simple_strtoul(p, &p, 10);
2291 return count;
2292}
2293
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002295static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002297 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 unsigned int __data = __VAR; \
2299 if (__CONV) \
2300 __data = jiffies_to_msecs(__data); \
2301 return cfq_var_show(__data, (page)); \
2302}
2303SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002304SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2305SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002306SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2307SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002308SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2309SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2310SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2311SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312#undef SHOW_FUNCTION
2313
2314#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002315static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002317 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 unsigned int __data; \
2319 int ret = cfq_var_store(&__data, (page), count); \
2320 if (__data < (MIN)) \
2321 __data = (MIN); \
2322 else if (__data > (MAX)) \
2323 __data = (MAX); \
2324 if (__CONV) \
2325 *(__PTR) = msecs_to_jiffies(__data); \
2326 else \
2327 *(__PTR) = __data; \
2328 return ret; \
2329}
2330STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002331STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2332 UINT_MAX, 1);
2333STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2334 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002335STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002336STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2337 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002338STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2339STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2340STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01002341STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2342 UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343#undef STORE_FUNCTION
2344
Al Viroe572ec72006-03-18 22:27:18 -05002345#define CFQ_ATTR(name) \
2346 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002347
Al Viroe572ec72006-03-18 22:27:18 -05002348static struct elv_fs_entry cfq_attrs[] = {
2349 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002350 CFQ_ATTR(fifo_expire_sync),
2351 CFQ_ATTR(fifo_expire_async),
2352 CFQ_ATTR(back_seek_max),
2353 CFQ_ATTR(back_seek_penalty),
2354 CFQ_ATTR(slice_sync),
2355 CFQ_ATTR(slice_async),
2356 CFQ_ATTR(slice_async_rq),
2357 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002358 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359};
2360
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361static struct elevator_type iosched_cfq = {
2362 .ops = {
2363 .elevator_merge_fn = cfq_merge,
2364 .elevator_merged_fn = cfq_merged_request,
2365 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002366 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002367 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002369 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370 .elevator_deactivate_req_fn = cfq_deactivate_request,
2371 .elevator_queue_empty_fn = cfq_queue_empty,
2372 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002373 .elevator_former_req_fn = elv_rb_former_request,
2374 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375 .elevator_set_req_fn = cfq_set_request,
2376 .elevator_put_req_fn = cfq_put_request,
2377 .elevator_may_queue_fn = cfq_may_queue,
2378 .elevator_init_fn = cfq_init_queue,
2379 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002380 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 },
Al Viro3d1ab402006-03-18 18:35:43 -05002382 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 .elevator_name = "cfq",
2384 .elevator_owner = THIS_MODULE,
2385};
2386
2387static int __init cfq_init(void)
2388{
Jens Axboe22e2c502005-06-27 10:55:12 +02002389 /*
2390 * could be 0 on HZ < 1000 setups
2391 */
2392 if (!cfq_slice_async)
2393 cfq_slice_async = 1;
2394 if (!cfq_slice_idle)
2395 cfq_slice_idle = 1;
2396
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397 if (cfq_slab_setup())
2398 return -ENOMEM;
2399
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002400 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002402 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403}
2404
2405static void __exit cfq_exit(void)
2406{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002407 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002409 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002410 /* ioc_gone's update must be visible before reading ioc_count */
2411 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02002412
2413 /*
2414 * this also protects us from entering cfq_slab_kill() with
2415 * pending RCU callbacks
2416 */
Jens Axboe4050cf12006-07-19 05:07:12 +02002417 if (elv_ioc_count_read(ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002418 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08002419 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420}
2421
2422module_init(cfq_init);
2423module_exit(cfq_exit);
2424
2425MODULE_AUTHOR("Jens Axboe");
2426MODULE_LICENSE("GPL");
2427MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");