blob: 9e809345f71abd718cecc27f6fd4ea068ed4775c [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * CFQ, or complete fairness queueing, disk scheduler.
3 *
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6 *
Jens Axboe0fe23472006-09-04 15:41:16 +02007 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/module.h>
Al Viro1cc9be62006-03-18 12:29:52 -050010#include <linux/blkdev.h>
11#include <linux/elevator.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070012#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020013#include <linux/ioprio.h>
Jens Axboe7b679132008-05-30 12:23:07 +020014#include <linux/blktrace_api.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070015
16/*
17 * tunables
18 */
Jens Axboefe094d92008-01-31 13:08:54 +010019/* max queue in one round of service */
20static const int cfq_quantum = 4;
Arjan van de Ven64100092006-01-06 09:46:02 +010021static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010022/* maximum backwards seek, in KiB */
23static const int cfq_back_max = 16 * 1024;
24/* penalty of a backwards seek */
25static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010026static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020027static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010028static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020029static int cfq_slice_idle = HZ / 125;
Jens Axboe22e2c502005-06-27 10:55:12 +020030
Jens Axboed9e76202007-04-20 14:27:50 +020031/*
Jens Axboe08717142008-01-28 11:38:15 +010032 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020033 */
Jens Axboe08717142008-01-28 11:38:15 +010034#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020035
36/*
37 * below this threshold, we consider thinktime immediate
38 */
39#define CFQ_MIN_TT (2)
40
Jens Axboe22e2c502005-06-27 10:55:12 +020041#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020042#define CFQ_HW_QUEUE_MIN (5)
Jens Axboe22e2c502005-06-27 10:55:12 +020043
Jens Axboefe094d92008-01-31 13:08:54 +010044#define RQ_CIC(rq) \
45 ((struct cfq_io_context *) (rq)->elevator_private)
Jens Axboe7b679132008-05-30 12:23:07 +020046#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
Linus Torvalds1da177e2005-04-16 15:20:36 -070047
Christoph Lametere18b8902006-12-06 20:33:20 -080048static struct kmem_cache *cfq_pool;
49static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070050
Jens Axboe4050cf12006-07-19 05:07:12 +020051static DEFINE_PER_CPU(unsigned long, ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050052static struct completion *ioc_gone;
Jens Axboe9a11b4e2008-05-29 09:32:08 +020053static DEFINE_SPINLOCK(ioc_gone_lock);
Al Viro334e94d2006-03-18 15:05:53 -050054
Jens Axboe22e2c502005-06-27 10:55:12 +020055#define CFQ_PRIO_LISTS IOPRIO_BE_NR
56#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020057#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
58
Jens Axboe3b181522005-06-27 10:56:24 +020059#define ASYNC (0)
60#define SYNC (1)
61
Jens Axboe206dc692006-03-28 13:03:44 +020062#define sample_valid(samples) ((samples) > 80)
63
Jens Axboe22e2c502005-06-27 10:55:12 +020064/*
Jens Axboecc09e292007-04-26 12:53:50 +020065 * Most of our rbtree usage is for sorting with min extraction, so
66 * if we cache the leftmost node we don't have to walk down the tree
67 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
68 * move this into the elevator for the rq sorting as well.
69 */
70struct cfq_rb_root {
71 struct rb_root rb;
72 struct rb_node *left;
73};
74#define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
75
76/*
Jens Axboe22e2c502005-06-27 10:55:12 +020077 * Per block device queue structure
78 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070079struct cfq_data {
Jens Axboe165125e2007-07-24 09:28:11 +020080 struct request_queue *queue;
Jens Axboe22e2c502005-06-27 10:55:12 +020081
82 /*
83 * rr list of queues with requests and the count of them
84 */
Jens Axboecc09e292007-04-26 12:53:50 +020085 struct cfq_rb_root service_tree;
Jens Axboe22e2c502005-06-27 10:55:12 +020086 unsigned int busy_queues;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +010087 /*
88 * Used to track any pending rt requests so we can pre-empt current
89 * non-RT cfqq in service when this value is non-zero.
90 */
91 unsigned int busy_rt_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +020092
Jens Axboe22e2c502005-06-27 10:55:12 +020093 int rq_in_driver;
Jens Axboe3ed9a292007-04-23 08:33:33 +020094 int sync_flight;
Aaron Carroll45333d52008-08-26 15:52:36 +020095
96 /*
97 * queue-depth detection
98 */
99 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200100 int hw_tag;
Aaron Carroll45333d52008-08-26 15:52:36 +0200101 int hw_tag_samples;
102 int rq_in_driver_peak;
Jens Axboe22e2c502005-06-27 10:55:12 +0200103
104 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200105 * idle window management
106 */
107 struct timer_list idle_slice_timer;
108 struct work_struct unplug_work;
109
110 struct cfq_queue *active_queue;
111 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200112
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200113 /*
114 * async queue for each priority case
115 */
116 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
117 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200118
Jens Axboe6d048f52007-04-25 12:44:27 +0200119 sector_t last_position;
Jens Axboe22e2c502005-06-27 10:55:12 +0200120 unsigned long last_end_request;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 /*
123 * tunables, see top of file
124 */
125 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200126 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127 unsigned int cfq_back_penalty;
128 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200129 unsigned int cfq_slice[2];
130 unsigned int cfq_slice_async_rq;
131 unsigned int cfq_slice_idle;
Al Virod9ff4182006-03-18 13:51:22 -0500132
133 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134};
135
Jens Axboe22e2c502005-06-27 10:55:12 +0200136/*
137 * Per process-grouping structure
138 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139struct cfq_queue {
140 /* reference count */
141 atomic_t ref;
Richard Kennedybe754d22008-05-23 06:52:00 +0200142 /* various state flags, see below */
143 unsigned int flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 /* parent cfq_data */
145 struct cfq_data *cfqd;
Jens Axboed9e76202007-04-20 14:27:50 +0200146 /* service_tree member */
147 struct rb_node rb_node;
148 /* service_tree key */
149 unsigned long rb_key;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 /* sorted list of pending requests */
151 struct rb_root sort_list;
152 /* if fifo isn't expired, next request to serve */
Jens Axboe5e705372006-07-13 12:39:25 +0200153 struct request *next_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 /* requests queued in sort_list */
155 int queued[2];
156 /* currently allocated requests */
157 int allocated[2];
158 /* fifo list of requests in sort_list */
Jens Axboe22e2c502005-06-27 10:55:12 +0200159 struct list_head fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160
Jens Axboe22e2c502005-06-27 10:55:12 +0200161 unsigned long slice_end;
Jens Axboec5b680f2007-01-19 11:56:49 +1100162 long slice_resid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163
Richard Kennedybe754d22008-05-23 06:52:00 +0200164 /* pending metadata requests */
165 int meta_pending;
Jens Axboe6d048f52007-04-25 12:44:27 +0200166 /* number of requests that are on the dispatch list or inside driver */
167 int dispatched;
Jens Axboe22e2c502005-06-27 10:55:12 +0200168
169 /* io prio of this group */
170 unsigned short ioprio, org_ioprio;
171 unsigned short ioprio_class, org_ioprio_class;
172
Jens Axboe7b679132008-05-30 12:23:07 +0200173 pid_t pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174};
175
Jens Axboe3b181522005-06-27 10:56:24 +0200176enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100177 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
178 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
179 CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */
180 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
181 CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */
182 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
183 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
184 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
185 CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */
Jens Axboe44f7c162007-01-19 11:51:58 +1100186 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200187 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200188};
189
190#define CFQ_CFQQ_FNS(name) \
191static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
192{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100193 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200194} \
195static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
196{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100197 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200198} \
199static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
200{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100201 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200202}
203
204CFQ_CFQQ_FNS(on_rr);
205CFQ_CFQQ_FNS(wait_request);
206CFQ_CFQQ_FNS(must_alloc);
207CFQ_CFQQ_FNS(must_alloc_slice);
208CFQ_CFQQ_FNS(must_dispatch);
209CFQ_CFQQ_FNS(fifo_expire);
210CFQ_CFQQ_FNS(idle_window);
211CFQ_CFQQ_FNS(prio_changed);
Jens Axboe53b03742006-07-28 09:48:51 +0200212CFQ_CFQQ_FNS(queue_new);
Jens Axboe44f7c162007-01-19 11:51:58 +1100213CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200214CFQ_CFQQ_FNS(sync);
Jens Axboe3b181522005-06-27 10:56:24 +0200215#undef CFQ_CFQQ_FNS
216
Jens Axboe7b679132008-05-30 12:23:07 +0200217#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
218 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
219#define cfq_log(cfqd, fmt, args...) \
220 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
221
Jens Axboe165125e2007-07-24 09:28:11 +0200222static void cfq_dispatch_insert(struct request_queue *, struct request *);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200223static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
Jens Axboefd0928d2008-01-24 08:52:45 +0100224 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100225static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200226 struct io_context *);
227
228static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
229 int is_sync)
230{
231 return cic->cfqq[!!is_sync];
232}
233
234static inline void cic_set_cfqq(struct cfq_io_context *cic,
235 struct cfq_queue *cfqq, int is_sync)
236{
237 cic->cfqq[!!is_sync] = cfqq;
238}
239
240/*
241 * We regard a request as SYNC, if it's either a read or has the SYNC bit
242 * set (in which case it could also be direct WRITE).
243 */
244static inline int cfq_bio_sync(struct bio *bio)
245{
246 if (bio_data_dir(bio) == READ || bio_sync(bio))
247 return 1;
248
249 return 0;
250}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700253 * scheduler run of queue, if there are requests pending and no one in the
254 * driver that will restart queueing
255 */
256static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
257{
Jens Axboe7b679132008-05-30 12:23:07 +0200258 if (cfqd->busy_queues) {
259 cfq_log(cfqd, "schedule dispatch");
Jens Axboe18887ad2008-07-28 13:08:45 +0200260 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200261 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700262}
263
Jens Axboe165125e2007-07-24 09:28:11 +0200264static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700265{
266 struct cfq_data *cfqd = q->elevator->elevator_data;
267
Jens Axboeb4878f22005-10-20 16:42:29 +0200268 return !cfqd->busy_queues;
Andrew Morton99f95e52005-06-27 20:14:05 -0700269}
270
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100272 * Scale schedule slice based on io priority. Use the sync time slice only
273 * if a queue is marked sync and has sync io queued. A sync queue with async
274 * io only, should not get full sync slice length.
275 */
Jens Axboed9e76202007-04-20 14:27:50 +0200276static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
277 unsigned short prio)
278{
279 const int base_slice = cfqd->cfq_slice[sync];
280
281 WARN_ON(prio >= IOPRIO_BE_NR);
282
283 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
284}
285
Jens Axboe44f7c162007-01-19 11:51:58 +1100286static inline int
287cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
288{
Jens Axboed9e76202007-04-20 14:27:50 +0200289 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100290}
291
292static inline void
293cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
294{
295 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200296 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100297}
298
299/*
300 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
301 * isn't valid until the first request from the dispatch is activated
302 * and the slice time set.
303 */
304static inline int cfq_slice_used(struct cfq_queue *cfqq)
305{
306 if (cfq_cfqq_slice_new(cfqq))
307 return 0;
308 if (time_before(jiffies, cfqq->slice_end))
309 return 0;
310
311 return 1;
312}
313
314/*
Jens Axboe5e705372006-07-13 12:39:25 +0200315 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200317 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 */
Jens Axboe5e705372006-07-13 12:39:25 +0200319static struct request *
320cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321{
322 sector_t last, s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200324#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
325#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
326 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327
Jens Axboe5e705372006-07-13 12:39:25 +0200328 if (rq1 == NULL || rq1 == rq2)
329 return rq2;
330 if (rq2 == NULL)
331 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200332
Jens Axboe5e705372006-07-13 12:39:25 +0200333 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
334 return rq1;
335 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
336 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200337 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
338 return rq1;
339 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
340 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341
Jens Axboe5e705372006-07-13 12:39:25 +0200342 s1 = rq1->sector;
343 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344
Jens Axboe6d048f52007-04-25 12:44:27 +0200345 last = cfqd->last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 /*
348 * by definition, 1KiB is 2 sectors
349 */
350 back_max = cfqd->cfq_back_max * 2;
351
352 /*
353 * Strict one way elevator _except_ in the case where we allow
354 * short backward seeks which are biased as twice the cost of a
355 * similar forward seek.
356 */
357 if (s1 >= last)
358 d1 = s1 - last;
359 else if (s1 + back_max >= last)
360 d1 = (last - s1) * cfqd->cfq_back_penalty;
361 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200362 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363
364 if (s2 >= last)
365 d2 = s2 - last;
366 else if (s2 + back_max >= last)
367 d2 = (last - s2) * cfqd->cfq_back_penalty;
368 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200369 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370
371 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372
Andreas Mohre8a99052006-03-28 08:59:49 +0200373 /*
374 * By doing switch() on the bit mask "wrap" we avoid having to
375 * check two variables for all permutations: --> faster!
376 */
377 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200378 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200379 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200380 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200381 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200382 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200383 else {
384 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200385 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200386 else
Jens Axboe5e705372006-07-13 12:39:25 +0200387 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200388 }
389
390 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200391 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200392 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200393 return rq2;
394 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200395 default:
396 /*
397 * Since both rqs are wrapped,
398 * start with the one that's further behind head
399 * (--> only *one* back seek required),
400 * since back seek takes more time than forward.
401 */
402 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200403 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404 else
Jens Axboe5e705372006-07-13 12:39:25 +0200405 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 }
407}
408
Jens Axboe498d3aa2007-04-26 12:54:48 +0200409/*
410 * The below is leftmost cache rbtree addon
411 */
Jens Axboe08717142008-01-28 11:38:15 +0100412static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200413{
414 if (!root->left)
415 root->left = rb_first(&root->rb);
416
Jens Axboe08717142008-01-28 11:38:15 +0100417 if (root->left)
418 return rb_entry(root->left, struct cfq_queue, rb_node);
419
420 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200421}
422
423static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
424{
425 if (root->left == n)
426 root->left = NULL;
427
428 rb_erase(n, &root->rb);
429 RB_CLEAR_NODE(n);
430}
431
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432/*
433 * would be nice to take fifo expire time into account as well
434 */
Jens Axboe5e705372006-07-13 12:39:25 +0200435static struct request *
436cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
437 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438{
Jens Axboe21183b02006-07-13 12:33:14 +0200439 struct rb_node *rbnext = rb_next(&last->rb_node);
440 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200441 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442
Jens Axboe21183b02006-07-13 12:33:14 +0200443 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444
445 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200446 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200447
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200449 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200450 else {
451 rbnext = rb_first(&cfqq->sort_list);
452 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200453 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200454 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
Jens Axboe21183b02006-07-13 12:33:14 +0200456 return cfq_choose_req(cfqd, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457}
458
Jens Axboed9e76202007-04-20 14:27:50 +0200459static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
460 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461{
Jens Axboed9e76202007-04-20 14:27:50 +0200462 /*
463 * just an approximation, should be ok.
464 */
Jens Axboe67e6b492007-04-20 14:18:00 +0200465 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
466 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200467}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468
Jens Axboe498d3aa2007-04-26 12:54:48 +0200469/*
470 * The cfqd->service_tree holds all pending cfq_queue's that have
471 * requests waiting to be processed. It is sorted in the order that
472 * we will service the queues.
473 */
Jens Axboed9e76202007-04-20 14:27:50 +0200474static void cfq_service_tree_add(struct cfq_data *cfqd,
Jens Axboeedd75ff2007-04-19 12:03:34 +0200475 struct cfq_queue *cfqq, int add_front)
Jens Axboed9e76202007-04-20 14:27:50 +0200476{
Jens Axboe08717142008-01-28 11:38:15 +0100477 struct rb_node **p, *parent;
478 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +0200479 unsigned long rb_key;
Jens Axboe498d3aa2007-04-26 12:54:48 +0200480 int left;
Jens Axboed9e76202007-04-20 14:27:50 +0200481
Jens Axboe08717142008-01-28 11:38:15 +0100482 if (cfq_class_idle(cfqq)) {
483 rb_key = CFQ_IDLE_DELAY;
484 parent = rb_last(&cfqd->service_tree.rb);
485 if (parent && parent != &cfqq->rb_node) {
486 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
487 rb_key += __cfqq->rb_key;
488 } else
489 rb_key += jiffies;
490 } else if (!add_front) {
Jens Axboeedd75ff2007-04-19 12:03:34 +0200491 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
492 rb_key += cfqq->slice_resid;
493 cfqq->slice_resid = 0;
494 } else
495 rb_key = 0;
Jens Axboed9e76202007-04-20 14:27:50 +0200496
497 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Jens Axboe99f96282007-02-05 11:56:25 +0100498 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200499 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +0100500 */
Jens Axboed9e76202007-04-20 14:27:50 +0200501 if (rb_key == cfqq->rb_key)
502 return;
Jens Axboe53b03742006-07-28 09:48:51 +0200503
Jens Axboecc09e292007-04-26 12:53:50 +0200504 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboe22e2c502005-06-27 10:55:12 +0200505 }
Jens Axboed9e76202007-04-20 14:27:50 +0200506
Jens Axboe498d3aa2007-04-26 12:54:48 +0200507 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +0100508 parent = NULL;
509 p = &cfqd->service_tree.rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +0200510 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +0200511 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +0200512
Jens Axboed9e76202007-04-20 14:27:50 +0200513 parent = *p;
514 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
515
Jens Axboe0c534e02007-04-18 20:01:57 +0200516 /*
517 * sort RT queues first, we always want to give
Jens Axboe67060e32007-04-18 20:13:32 +0200518 * preference to them. IDLE queues goes to the back.
519 * after that, sort on the next service time.
Jens Axboe0c534e02007-04-18 20:01:57 +0200520 */
521 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200522 n = &(*p)->rb_left;
Jens Axboe0c534e02007-04-18 20:01:57 +0200523 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
Jens Axboe67060e32007-04-18 20:13:32 +0200524 n = &(*p)->rb_right;
525 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
526 n = &(*p)->rb_left;
527 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
528 n = &(*p)->rb_right;
Jens Axboe0c534e02007-04-18 20:01:57 +0200529 else if (rb_key < __cfqq->rb_key)
Jens Axboe67060e32007-04-18 20:13:32 +0200530 n = &(*p)->rb_left;
531 else
532 n = &(*p)->rb_right;
533
534 if (n == &(*p)->rb_right)
Jens Axboecc09e292007-04-26 12:53:50 +0200535 left = 0;
Jens Axboe67060e32007-04-18 20:13:32 +0200536
537 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +0200538 }
539
Jens Axboecc09e292007-04-26 12:53:50 +0200540 if (left)
541 cfqd->service_tree.left = &cfqq->rb_node;
542
Jens Axboed9e76202007-04-20 14:27:50 +0200543 cfqq->rb_key = rb_key;
544 rb_link_node(&cfqq->rb_node, parent, p);
Jens Axboecc09e292007-04-26 12:53:50 +0200545 rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546}
547
Jens Axboe498d3aa2007-04-26 12:54:48 +0200548/*
549 * Update cfqq's position in the service tree.
550 */
Jens Axboeedd75ff2007-04-19 12:03:34 +0200551static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200552{
Jens Axboe6d048f52007-04-25 12:44:27 +0200553 /*
554 * Resorting requires the cfqq to be on the RR list already.
555 */
Jens Axboe498d3aa2007-04-26 12:54:48 +0200556 if (cfq_cfqq_on_rr(cfqq))
Jens Axboeedd75ff2007-04-19 12:03:34 +0200557 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboe6d048f52007-04-25 12:44:27 +0200558}
559
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560/*
561 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +0200562 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 */
Jens Axboefebffd62008-01-28 13:19:43 +0100564static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565{
Jens Axboe7b679132008-05-30 12:23:07 +0200566 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200567 BUG_ON(cfq_cfqq_on_rr(cfqq));
568 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 cfqd->busy_queues++;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +0100570 if (cfq_class_rt(cfqq))
571 cfqd->busy_rt_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572
Jens Axboeedd75ff2007-04-19 12:03:34 +0200573 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574}
575
Jens Axboe498d3aa2007-04-26 12:54:48 +0200576/*
577 * Called when the cfqq no longer has requests pending, remove it from
578 * the service tree.
579 */
Jens Axboefebffd62008-01-28 13:19:43 +0100580static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581{
Jens Axboe7b679132008-05-30 12:23:07 +0200582 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +0200583 BUG_ON(!cfq_cfqq_on_rr(cfqq));
584 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585
Jens Axboecc09e292007-04-26 12:53:50 +0200586 if (!RB_EMPTY_NODE(&cfqq->rb_node))
587 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
Jens Axboed9e76202007-04-20 14:27:50 +0200588
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589 BUG_ON(!cfqd->busy_queues);
590 cfqd->busy_queues--;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +0100591 if (cfq_class_rt(cfqq))
592 cfqd->busy_rt_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593}
594
595/*
596 * rb tree support functions
597 */
Jens Axboefebffd62008-01-28 13:19:43 +0100598static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599{
Jens Axboe5e705372006-07-13 12:39:25 +0200600 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +0200601 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5e705372006-07-13 12:39:25 +0200602 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603
Jens Axboeb4878f22005-10-20 16:42:29 +0200604 BUG_ON(!cfqq->queued[sync]);
605 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606
Jens Axboe5e705372006-07-13 12:39:25 +0200607 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608
Jens Axboedd67d052006-06-21 09:36:18 +0200609 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboeb4878f22005-10-20 16:42:29 +0200610 cfq_del_cfqq_rr(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611}
612
Jens Axboe5e705372006-07-13 12:39:25 +0200613static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614{
Jens Axboe5e705372006-07-13 12:39:25 +0200615 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe21183b02006-07-13 12:33:14 +0200617 struct request *__alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618
Jens Axboe5380a102006-07-13 12:37:56 +0200619 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620
621 /*
622 * looks a little odd, but the first insert might return an alias.
623 * if that happens, put the alias on the dispatch list
624 */
Jens Axboe21183b02006-07-13 12:33:14 +0200625 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +0200626 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +0100627
628 if (!cfq_cfqq_on_rr(cfqq))
629 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +0200630
631 /*
632 * check if this request is a better next-serve candidate
633 */
634 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
635 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636}
637
Jens Axboefebffd62008-01-28 13:19:43 +0100638static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639{
Jens Axboe5380a102006-07-13 12:37:56 +0200640 elv_rb_del(&cfqq->sort_list, rq);
641 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +0200642 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643}
644
Jens Axboe206dc692006-03-28 13:03:44 +0200645static struct request *
646cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647{
Jens Axboe206dc692006-03-28 13:03:44 +0200648 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200649 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +0200650 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651
Jens Axboe4ac845a2008-01-24 08:44:49 +0100652 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200653 if (!cic)
654 return NULL;
655
656 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +0200657 if (cfqq) {
658 sector_t sector = bio->bi_sector + bio_sectors(bio);
659
Jens Axboe21183b02006-07-13 12:33:14 +0200660 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +0200661 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 return NULL;
664}
665
Jens Axboe165125e2007-07-24 09:28:11 +0200666static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200667{
668 struct cfq_data *cfqd = q->elevator->elevator_data;
669
670 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +0200671 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
672 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +0200673
Jens Axboe6d048f52007-04-25 12:44:27 +0200674 cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
Jens Axboeb4878f22005-10-20 16:42:29 +0200675}
676
Jens Axboe165125e2007-07-24 09:28:11 +0200677static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678{
Jens Axboe22e2c502005-06-27 10:55:12 +0200679 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680
Jens Axboeb4878f22005-10-20 16:42:29 +0200681 WARN_ON(!cfqd->rq_in_driver);
682 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +0200683 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
684 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685}
686
Jens Axboeb4878f22005-10-20 16:42:29 +0200687static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688{
Jens Axboe5e705372006-07-13 12:39:25 +0200689 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +0200690
Jens Axboe5e705372006-07-13 12:39:25 +0200691 if (cfqq->next_rq == rq)
692 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693
Jens Axboeb4878f22005-10-20 16:42:29 +0200694 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +0200695 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +0200696
Aaron Carroll45333d52008-08-26 15:52:36 +0200697 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +0200698 if (rq_is_meta(rq)) {
699 WARN_ON(!cfqq->meta_pending);
700 cfqq->meta_pending--;
701 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702}
703
Jens Axboe165125e2007-07-24 09:28:11 +0200704static int cfq_merge(struct request_queue *q, struct request **req,
705 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706{
707 struct cfq_data *cfqd = q->elevator->elevator_data;
708 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709
Jens Axboe206dc692006-03-28 13:03:44 +0200710 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +0200711 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +0200712 *req = __rq;
713 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 }
715
716 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717}
718
Jens Axboe165125e2007-07-24 09:28:11 +0200719static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +0200720 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721{
Jens Axboe21183b02006-07-13 12:33:14 +0200722 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +0200723 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724
Jens Axboe5e705372006-07-13 12:39:25 +0200725 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727}
728
729static void
Jens Axboe165125e2007-07-24 09:28:11 +0200730cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731 struct request *next)
732{
Jens Axboe22e2c502005-06-27 10:55:12 +0200733 /*
734 * reposition in fifo if next is older than rq
735 */
736 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
737 time_before(next->start_time, rq->start_time))
738 list_move(&rq->queuelist, &next->queuelist);
739
Jens Axboeb4878f22005-10-20 16:42:29 +0200740 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +0200741}
742
Jens Axboe165125e2007-07-24 09:28:11 +0200743static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +0100744 struct bio *bio)
745{
746 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200747 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +0100748 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +0100749
750 /*
Jens Axboeec8acb62007-01-02 18:32:11 +0100751 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +0100752 */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200753 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboeda775262006-12-20 11:04:12 +0100754 return 0;
755
756 /*
Jens Axboe719d3402006-12-22 09:38:53 +0100757 * Lookup the cfqq that this bio will be queued with. Allow
758 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +0100759 */
Jens Axboe4ac845a2008-01-24 08:44:49 +0100760 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200761 if (!cic)
762 return 0;
Jens Axboe719d3402006-12-22 09:38:53 +0100763
Vasily Tarasov91fac312007-04-25 12:29:51 +0200764 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe719d3402006-12-22 09:38:53 +0100765 if (cfqq == RQ_CFQQ(rq))
766 return 1;
Jens Axboeda775262006-12-20 11:04:12 +0100767
Jens Axboeec8acb62007-01-02 18:32:11 +0100768 return 0;
Jens Axboeda775262006-12-20 11:04:12 +0100769}
770
Jens Axboefebffd62008-01-28 13:19:43 +0100771static void __cfq_set_active_queue(struct cfq_data *cfqd,
772 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +0200773{
774 if (cfqq) {
Jens Axboe7b679132008-05-30 12:23:07 +0200775 cfq_log_cfqq(cfqd, cfqq, "set_active");
Jens Axboe22e2c502005-06-27 10:55:12 +0200776 cfqq->slice_end = 0;
Jens Axboe3b181522005-06-27 10:56:24 +0200777 cfq_clear_cfqq_must_alloc_slice(cfqq);
778 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +1100779 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe1afba042007-04-17 12:47:55 +0200780 cfq_clear_cfqq_queue_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200781 }
782
783 cfqd->active_queue = cfqq;
784}
785
786/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100787 * current cfqq expired its slice (or was too idle), select new one
788 */
789static void
790__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6084cdd2007-04-23 08:25:00 +0200791 int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100792{
Jens Axboe7b679132008-05-30 12:23:07 +0200793 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
794
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100795 if (cfq_cfqq_wait_request(cfqq))
796 del_timer(&cfqd->idle_slice_timer);
797
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100798 cfq_clear_cfqq_must_dispatch(cfqq);
799 cfq_clear_cfqq_wait_request(cfqq);
800
801 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +0200802 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100803 */
Jens Axboe7b679132008-05-30 12:23:07 +0200804 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +1100805 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +0200806 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
807 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100808
Jens Axboeedd75ff2007-04-19 12:03:34 +0200809 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100810
811 if (cfqq == cfqd->active_queue)
812 cfqd->active_queue = NULL;
813
814 if (cfqd->active_cic) {
815 put_io_context(cfqd->active_cic->ioc);
816 cfqd->active_cic = NULL;
817 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100818}
819
Jens Axboe6084cdd2007-04-23 08:25:00 +0200820static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100821{
822 struct cfq_queue *cfqq = cfqd->active_queue;
823
824 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +0200825 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100826}
827
Jens Axboe498d3aa2007-04-26 12:54:48 +0200828/*
829 * Get next queue for service. Unless we have a queue preemption,
830 * we'll simply select the first cfqq in the service tree.
831 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200832static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200833{
Jens Axboeedd75ff2007-04-19 12:03:34 +0200834 if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
835 return NULL;
836
Jens Axboe08717142008-01-28 11:38:15 +0100837 return cfq_rb_first(&cfqd->service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +0200838}
839
Jens Axboe498d3aa2007-04-26 12:54:48 +0200840/*
841 * Get and set a new active queue for service.
842 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200843static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
844{
845 struct cfq_queue *cfqq;
846
Jens Axboed9e76202007-04-20 14:27:50 +0200847 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +0200848 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +0200849 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +0200850}
851
Jens Axboed9e76202007-04-20 14:27:50 +0200852static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
853 struct request *rq)
854{
855 if (rq->sector >= cfqd->last_position)
856 return rq->sector - cfqd->last_position;
857 else
858 return cfqd->last_position - rq->sector;
859}
860
Jens Axboe6d048f52007-04-25 12:44:27 +0200861static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
862{
863 struct cfq_io_context *cic = cfqd->active_cic;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200864
Jens Axboe6d048f52007-04-25 12:44:27 +0200865 if (!sample_valid(cic->seek_samples))
866 return 0;
867
868 return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
869}
870
Jens Axboed9e76202007-04-20 14:27:50 +0200871static int cfq_close_cooperator(struct cfq_data *cfq_data,
872 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +0200873{
Jens Axboe6d048f52007-04-25 12:44:27 +0200874 /*
Jens Axboed9e76202007-04-20 14:27:50 +0200875 * We should notice if some of the queues are cooperating, eg
876 * working closely on the same area of the disk. In that case,
877 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +0200878 */
Jens Axboed9e76202007-04-20 14:27:50 +0200879 return 0;
Jens Axboe6d048f52007-04-25 12:44:27 +0200880}
881
882#define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
883
884static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +0200885{
Jens Axboe17926692007-01-19 11:59:30 +1100886 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +0200887 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100888 unsigned long sl;
889
Jens Axboea68bbdd2008-09-24 13:03:33 +0200890 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +0200891 * SSD device without seek penalty, disable idling. But only do so
892 * for devices that support queuing, otherwise we still have a problem
893 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +0200894 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +0200895 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +0200896 return;
897
Jens Axboedd67d052006-06-21 09:36:18 +0200898 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +0200899 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +0200900
901 /*
902 * idle is disabled, either manually or by past process history
903 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200904 if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
905 return;
906
Jens Axboe22e2c502005-06-27 10:55:12 +0200907 /*
Jens Axboe7b679132008-05-30 12:23:07 +0200908 * still requests with the driver, don't idle
909 */
910 if (cfqd->rq_in_driver)
911 return;
912
913 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200914 * task has exited, don't wait
915 */
Jens Axboe206dc692006-03-28 13:03:44 +0200916 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +0100917 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +0200918 return;
919
920 /*
921 * See if this prio level has a good candidate
922 */
Jens Axboe1afba042007-04-17 12:47:55 +0200923 if (cfq_close_cooperator(cfqd, cfqq) &&
924 (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
Jens Axboe6d048f52007-04-25 12:44:27 +0200925 return;
Jens Axboe22e2c502005-06-27 10:55:12 +0200926
Jens Axboe3b181522005-06-27 10:56:24 +0200927 cfq_mark_cfqq_must_dispatch(cfqq);
928 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200929
Jens Axboe206dc692006-03-28 13:03:44 +0200930 /*
931 * we don't want to idle for seeks, but we do want to allow
932 * fair distribution of slice time for a process doing back-to-back
933 * seeks. so allow a little bit of time for him to submit a new rq
934 */
Jens Axboe6d048f52007-04-25 12:44:27 +0200935 sl = cfqd->cfq_slice_idle;
Jens Axboecaaa5f92006-06-16 11:23:00 +0200936 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
Jens Axboed9e76202007-04-20 14:27:50 +0200937 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
Jens Axboe206dc692006-03-28 13:03:44 +0200938
Jens Axboe7b14e3b2006-02-28 09:35:11 +0100939 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe7b679132008-05-30 12:23:07 +0200940 cfq_log(cfqd, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941}
942
Jens Axboe498d3aa2007-04-26 12:54:48 +0200943/*
944 * Move request from internal lists to the request queue dispatch list.
945 */
Jens Axboe165125e2007-07-24 09:28:11 +0200946static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947{
Jens Axboe3ed9a292007-04-23 08:33:33 +0200948 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +0200949 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +0200950
Jens Axboe7b679132008-05-30 12:23:07 +0200951 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
952
Jens Axboe5380a102006-07-13 12:37:56 +0200953 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200954 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +0200955 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +0200956
957 if (cfq_cfqq_sync(cfqq))
958 cfqd->sync_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959}
960
961/*
962 * return expired entry, or NULL to just start from scratch in rbtree
963 */
Jens Axboefebffd62008-01-28 13:19:43 +0100964static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965{
966 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe22e2c502005-06-27 10:55:12 +0200967 struct request *rq;
Jens Axboe89850f72006-07-22 16:48:31 +0200968 int fifo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969
Jens Axboe3b181522005-06-27 10:56:24 +0200970 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +1100972
973 cfq_mark_cfqq_fifo_expire(cfqq);
974
Jens Axboe89850f72006-07-22 16:48:31 +0200975 if (list_empty(&cfqq->fifo))
976 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977
Jens Axboe6d048f52007-04-25 12:44:27 +0200978 fifo = cfq_cfqq_sync(cfqq);
Jens Axboe89850f72006-07-22 16:48:31 +0200979 rq = rq_entry_fifo(cfqq->fifo.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980
Jens Axboe6d048f52007-04-25 12:44:27 +0200981 if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
Jens Axboe7b679132008-05-30 12:23:07 +0200982 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983
Jens Axboe7b679132008-05-30 12:23:07 +0200984 cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +0200985 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986}
987
Jens Axboe22e2c502005-06-27 10:55:12 +0200988static inline int
989cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
990{
991 const int base_rq = cfqd->cfq_slice_async_rq;
992
993 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
994
995 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
996}
997
998/*
Jens Axboe498d3aa2007-04-26 12:54:48 +0200999 * Select a queue for service. If we have a current active queue,
1000 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02001001 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001002static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001003{
Jens Axboe22e2c502005-06-27 10:55:12 +02001004 struct cfq_queue *cfqq;
1005
1006 cfqq = cfqd->active_queue;
1007 if (!cfqq)
1008 goto new_queue;
1009
1010 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001011 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02001012 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001013 if (cfq_slice_used(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02001014 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +02001015
1016 /*
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01001017 * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1018 * cfqq.
1019 */
1020 if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1021 /*
1022 * We simulate this as cfqq timed out so that it gets to bank
1023 * the remaining of its time slice.
1024 */
1025 cfq_log_cfqq(cfqd, cfqq, "preempt");
1026 cfq_slice_expired(cfqd, 1);
1027 goto new_queue;
1028 }
1029
1030 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02001031 * The active queue has requests and isn't expired, allow it to
1032 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02001033 */
Jens Axboedd67d052006-06-21 09:36:18 +02001034 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001035 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02001036
1037 /*
1038 * No requests pending. If the active queue still has requests in
1039 * flight or is idling for a new request, allow either of these
1040 * conditions to happen (or time out) before selecting a new queue.
1041 */
Jens Axboecc197472007-04-20 20:45:39 +02001042 if (timer_pending(&cfqd->idle_slice_timer) ||
1043 (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02001044 cfqq = NULL;
1045 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02001046 }
1047
Jens Axboe3b181522005-06-27 10:56:24 +02001048expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02001049 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02001050new_queue:
1051 cfqq = cfq_set_active_queue(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02001052keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02001053 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001054}
1055
Jens Axboe498d3aa2007-04-26 12:54:48 +02001056/*
1057 * Dispatch some requests from cfqq, moving them to the request queue
1058 * dispatch list.
1059 */
Jens Axboe22e2c502005-06-27 10:55:12 +02001060static int
1061__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1062 int max_dispatch)
1063{
1064 int dispatched = 0;
1065
Jens Axboedd67d052006-06-21 09:36:18 +02001066 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001067
1068 do {
Jens Axboe5e705372006-07-13 12:39:25 +02001069 struct request *rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001070
1071 /*
1072 * follow expired path, else get first next available
1073 */
Jens Axboefe094d92008-01-31 13:08:54 +01001074 rq = cfq_check_fifo(cfqq);
1075 if (rq == NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001076 rq = cfqq->next_rq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001077
1078 /*
1079 * finally, insert request into driver dispatch list
1080 */
Jens Axboe5e705372006-07-13 12:39:25 +02001081 cfq_dispatch_insert(cfqd->queue, rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001082
Jens Axboe22e2c502005-06-27 10:55:12 +02001083 dispatched++;
1084
1085 if (!cfqd->active_cic) {
Jens Axboe5e705372006-07-13 12:39:25 +02001086 atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1087 cfqd->active_cic = RQ_CIC(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001088 }
1089
Jens Axboedd67d052006-06-21 09:36:18 +02001090 if (RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02001091 break;
1092
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01001093 /*
1094 * If there is a non-empty RT cfqq waiting for current
1095 * cfqq's timeslice to complete, pre-empt this cfqq
1096 */
1097 if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues)
1098 break;
1099
Jens Axboe22e2c502005-06-27 10:55:12 +02001100 } while (dispatched < max_dispatch);
1101
1102 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001103 * expire an async queue immediately if it has used up its slice. idle
1104 * queue always expire after 1 dispatch round.
1105 */
Jens Axboea9938002007-04-20 08:55:52 +02001106 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
Jens Axboe20e493a2007-04-23 08:26:36 +02001107 dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
Jens Axboea9938002007-04-20 08:55:52 +02001108 cfq_class_idle(cfqq))) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001109 cfqq->slice_end = jiffies + 1;
Jens Axboe6084cdd2007-04-23 08:25:00 +02001110 cfq_slice_expired(cfqd, 0);
Jens Axboe44f7c162007-01-19 11:51:58 +11001111 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001112
1113 return dispatched;
1114}
1115
Jens Axboefebffd62008-01-28 13:19:43 +01001116static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02001117{
1118 int dispatched = 0;
1119
1120 while (cfqq->next_rq) {
1121 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1122 dispatched++;
1123 }
1124
1125 BUG_ON(!list_empty(&cfqq->fifo));
1126 return dispatched;
1127}
1128
Jens Axboe498d3aa2007-04-26 12:54:48 +02001129/*
1130 * Drain our current requests. Used for barriers and when switching
1131 * io schedulers on-the-fly.
1132 */
Jens Axboed9e76202007-04-20 14:27:50 +02001133static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001134{
Jens Axboe08717142008-01-28 11:38:15 +01001135 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001136 int dispatched = 0;
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001137
Jens Axboe08717142008-01-28 11:38:15 +01001138 while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
Jens Axboed9e76202007-04-20 14:27:50 +02001139 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001140
Jens Axboe6084cdd2007-04-23 08:25:00 +02001141 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001142
1143 BUG_ON(cfqd->busy_queues);
1144
Jens Axboe7b679132008-05-30 12:23:07 +02001145 cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001146 return dispatched;
1147}
1148
Jens Axboe165125e2007-07-24 09:28:11 +02001149static int cfq_dispatch_requests(struct request_queue *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150{
1151 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe6d048f52007-04-25 12:44:27 +02001152 struct cfq_queue *cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001153 int dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154
Jens Axboe22e2c502005-06-27 10:55:12 +02001155 if (!cfqd->busy_queues)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 return 0;
1157
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01001158 if (unlikely(force))
1159 return cfq_forced_dispatch(cfqd);
1160
Jens Axboecaaa5f92006-06-16 11:23:00 +02001161 dispatched = 0;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001162 while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
Jens Axboeb4878f22005-10-20 16:42:29 +02001163 int max_dispatch;
1164
Jens Axboe3ed9a292007-04-23 08:33:33 +02001165 max_dispatch = cfqd->cfq_quantum;
1166 if (cfq_class_idle(cfqq))
1167 max_dispatch = 1;
1168
Jens Axboe30e0dc22008-10-20 15:44:28 +02001169 if (cfqq->dispatched >= max_dispatch && cfqd->busy_queues > 1)
1170 break;
Jens Axboe9ede2092007-01-19 12:11:44 +11001171
Jens Axboe3ed9a292007-04-23 08:33:33 +02001172 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1173 break;
1174
Jens Axboe3b181522005-06-27 10:56:24 +02001175 cfq_clear_cfqq_must_dispatch(cfqq);
1176 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001177 del_timer(&cfqd->idle_slice_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178
Jens Axboecaaa5f92006-06-16 11:23:00 +02001179 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 }
1181
Jens Axboe7b679132008-05-30 12:23:07 +02001182 cfq_log(cfqd, "dispatched=%d", dispatched);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001183 return dispatched;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184}
1185
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186/*
Jens Axboe5e705372006-07-13 12:39:25 +02001187 * task holds one reference to the queue, dropped when task exits. each rq
1188 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 *
1190 * queue lock must be held here.
1191 */
1192static void cfq_put_queue(struct cfq_queue *cfqq)
1193{
Jens Axboe22e2c502005-06-27 10:55:12 +02001194 struct cfq_data *cfqd = cfqq->cfqd;
1195
1196 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197
1198 if (!atomic_dec_and_test(&cfqq->ref))
1199 return;
1200
Jens Axboe7b679132008-05-30 12:23:07 +02001201 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02001203 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Jens Axboe3b181522005-06-27 10:56:24 +02001204 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001206 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001207 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001208 cfq_schedule_dispatch(cfqd);
1209 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001210
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 kmem_cache_free(cfq_pool, cfqq);
1212}
1213
Jens Axboed6de8be2008-05-28 14:46:59 +02001214/*
1215 * Must always be called with the rcu_read_lock() held
1216 */
Jens Axboe07416d22008-05-07 09:17:12 +02001217static void
1218__call_for_each_cic(struct io_context *ioc,
1219 void (*func)(struct io_context *, struct cfq_io_context *))
1220{
1221 struct cfq_io_context *cic;
1222 struct hlist_node *n;
1223
1224 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1225 func(ioc, cic);
1226}
1227
Jens Axboe4ac845a2008-01-24 08:44:49 +01001228/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001229 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001230 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001231static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001232call_for_each_cic(struct io_context *ioc,
1233 void (*func)(struct io_context *, struct cfq_io_context *))
1234{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001235 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02001236 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001237 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001238}
Jens Axboe4ac845a2008-01-24 08:44:49 +01001239
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001240static void cfq_cic_free_rcu(struct rcu_head *head)
1241{
1242 struct cfq_io_context *cic;
1243
1244 cic = container_of(head, struct cfq_io_context, rcu_head);
1245
1246 kmem_cache_free(cfq_ioc_pool, cic);
1247 elv_ioc_count_dec(ioc_count);
1248
Jens Axboe9a11b4e2008-05-29 09:32:08 +02001249 if (ioc_gone) {
1250 /*
1251 * CFQ scheduler is exiting, grab exit lock and check
1252 * the pending io context count. If it hits zero,
1253 * complete ioc_gone and set it back to NULL
1254 */
1255 spin_lock(&ioc_gone_lock);
1256 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1257 complete(ioc_gone);
1258 ioc_gone = NULL;
1259 }
1260 spin_unlock(&ioc_gone_lock);
1261 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001262}
1263
1264static void cfq_cic_free(struct cfq_io_context *cic)
1265{
1266 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001267}
1268
1269static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1270{
1271 unsigned long flags;
1272
1273 BUG_ON(!cic->dead_key);
1274
1275 spin_lock_irqsave(&ioc->lock, flags);
1276 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01001277 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001278 spin_unlock_irqrestore(&ioc->lock, flags);
1279
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001280 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001281}
1282
Jens Axboed6de8be2008-05-28 14:46:59 +02001283/*
1284 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1285 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1286 * and ->trim() which is called with the task lock held
1287 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001288static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001290 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02001291 * ioc->refcount is zero here, or we are called from elv_unregister(),
1292 * so no more cic's are allowed to be linked into this ioc. So it
1293 * should be ok to iterate over the known list, we will see all cic's
1294 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01001295 */
Jens Axboe07416d22008-05-07 09:17:12 +02001296 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297}
1298
Jens Axboe89850f72006-07-22 16:48:31 +02001299static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1300{
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001301 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02001302 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11001303 cfq_schedule_dispatch(cfqd);
1304 }
Jens Axboe89850f72006-07-22 16:48:31 +02001305
1306 cfq_put_queue(cfqq);
1307}
1308
1309static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1310 struct cfq_io_context *cic)
1311{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001312 struct io_context *ioc = cic->ioc;
1313
Jens Axboefc463792006-08-29 09:05:44 +02001314 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001315
1316 /*
1317 * Make sure key == NULL is seen for dead queues
1318 */
Jens Axboefc463792006-08-29 09:05:44 +02001319 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001320 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02001321 cic->key = NULL;
1322
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001323 if (ioc->ioc_data == cic)
1324 rcu_assign_pointer(ioc->ioc_data, NULL);
1325
Jens Axboe89850f72006-07-22 16:48:31 +02001326 if (cic->cfqq[ASYNC]) {
1327 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1328 cic->cfqq[ASYNC] = NULL;
1329 }
1330
1331 if (cic->cfqq[SYNC]) {
1332 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1333 cic->cfqq[SYNC] = NULL;
1334 }
Jens Axboe89850f72006-07-22 16:48:31 +02001335}
1336
Jens Axboe4ac845a2008-01-24 08:44:49 +01001337static void cfq_exit_single_io_context(struct io_context *ioc,
1338 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001339{
Al Viro478a82b2006-03-18 13:25:24 -05001340 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02001341
Jens Axboe89850f72006-07-22 16:48:31 +02001342 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02001343 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001344 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02001345
Jens Axboe4ac845a2008-01-24 08:44:49 +01001346 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01001347
1348 /*
1349 * Ensure we get a fresh copy of the ->key to prevent
1350 * race between exiting task and queue
1351 */
1352 smp_read_barrier_depends();
1353 if (cic->key)
1354 __cfq_exit_single_io_context(cfqd, cic);
1355
Jens Axboe4ac845a2008-01-24 08:44:49 +01001356 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05001357 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001358}
1359
Jens Axboe498d3aa2007-04-26 12:54:48 +02001360/*
1361 * The process that ioc belongs to has exited, we need to clean up
1362 * and put the internal structures we have that belongs to that process.
1363 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02001364static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001366 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367}
1368
Jens Axboe22e2c502005-06-27 10:55:12 +02001369static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04001370cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371{
Jens Axboeb5deef92006-07-19 23:39:40 +02001372 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373
Christoph Lameter94f60302007-07-17 04:03:29 -07001374 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1375 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001377 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02001378 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01001379 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02001380 cic->dtor = cfq_free_io_context;
1381 cic->exit = cfq_exit_io_context;
Jens Axboe4050cf12006-07-19 05:07:12 +02001382 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 }
1384
1385 return cic;
1386}
1387
Jens Axboefd0928d2008-01-24 08:52:45 +01001388static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02001389{
1390 struct task_struct *tsk = current;
1391 int ioprio_class;
1392
Jens Axboe3b181522005-06-27 10:56:24 +02001393 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001394 return;
1395
Jens Axboefd0928d2008-01-24 08:52:45 +01001396 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001397 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01001398 default:
1399 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1400 case IOPRIO_CLASS_NONE:
1401 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02001402 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01001403 */
1404 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02001405 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01001406 break;
1407 case IOPRIO_CLASS_RT:
1408 cfqq->ioprio = task_ioprio(ioc);
1409 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1410 break;
1411 case IOPRIO_CLASS_BE:
1412 cfqq->ioprio = task_ioprio(ioc);
1413 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1414 break;
1415 case IOPRIO_CLASS_IDLE:
1416 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1417 cfqq->ioprio = 7;
1418 cfq_clear_cfqq_idle_window(cfqq);
1419 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02001420 }
1421
1422 /*
1423 * keep track of original prio settings in case we have to temporarily
1424 * elevate the priority of this queue
1425 */
1426 cfqq->org_ioprio = cfqq->ioprio;
1427 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02001428 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001429}
1430
Jens Axboefebffd62008-01-28 13:19:43 +01001431static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02001432{
Al Viro478a82b2006-03-18 13:25:24 -05001433 struct cfq_data *cfqd = cic->key;
1434 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01001435 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02001436
Jens Axboecaaa5f92006-06-16 11:23:00 +02001437 if (unlikely(!cfqd))
1438 return;
1439
Jens Axboec1b707d2006-10-30 19:54:23 +01001440 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001441
1442 cfqq = cic->cfqq[ASYNC];
1443 if (cfqq) {
1444 struct cfq_queue *new_cfqq;
Jens Axboefd0928d2008-01-24 08:52:45 +01001445 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02001446 if (new_cfqq) {
1447 cic->cfqq[ASYNC] = new_cfqq;
1448 cfq_put_queue(cfqq);
1449 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001450 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001451
1452 cfqq = cic->cfqq[SYNC];
1453 if (cfqq)
1454 cfq_mark_cfqq_prio_changed(cfqq);
1455
Jens Axboec1b707d2006-10-30 19:54:23 +01001456 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02001457}
1458
Jens Axboefc463792006-08-29 09:05:44 +02001459static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001461 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02001462 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463}
1464
1465static struct cfq_queue *
Jens Axboe15c31be2007-07-10 13:43:25 +02001466cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01001467 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001470 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471
1472retry:
Jens Axboe4ac845a2008-01-24 08:44:49 +01001473 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001474 /* cic always exists here */
1475 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476
1477 if (!cfqq) {
1478 if (new_cfqq) {
1479 cfqq = new_cfqq;
1480 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001481 } else if (gfp_mask & __GFP_WAIT) {
Jens Axboe89850f72006-07-22 16:48:31 +02001482 /*
1483 * Inform the allocator of the fact that we will
1484 * just repeat this allocation if it fails, to allow
1485 * the allocator to do whatever it needs to attempt to
1486 * free memory.
1487 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07001489 new_cfqq = kmem_cache_alloc_node(cfq_pool,
1490 gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1491 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 spin_lock_irq(cfqd->queue->queue_lock);
1493 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02001494 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07001495 cfqq = kmem_cache_alloc_node(cfq_pool,
1496 gfp_mask | __GFP_ZERO,
1497 cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001498 if (!cfqq)
1499 goto out;
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02001500 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501
Jens Axboed9e76202007-04-20 14:27:50 +02001502 RB_CLEAR_NODE(&cfqq->rb_node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001503 INIT_LIST_HEAD(&cfqq->fifo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 atomic_set(&cfqq->ref, 0);
1506 cfqq->cfqd = cfqd;
Jens Axboec5b680f2007-01-19 11:56:49 +11001507
Jens Axboe3b181522005-06-27 10:56:24 +02001508 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe53b03742006-07-28 09:48:51 +02001509 cfq_mark_cfqq_queue_new(cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001510
Jens Axboefd0928d2008-01-24 08:52:45 +01001511 cfq_init_prio_data(cfqq, ioc);
Jens Axboe08717142008-01-28 11:38:15 +01001512
1513 if (is_sync) {
1514 if (!cfq_class_idle(cfqq))
1515 cfq_mark_cfqq_idle_window(cfqq);
1516 cfq_mark_cfqq_sync(cfqq);
1517 }
Jens Axboe7b679132008-05-30 12:23:07 +02001518 cfqq->pid = current->pid;
1519 cfq_log_cfqq(cfqd, cfqq, "alloced");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520 }
1521
1522 if (new_cfqq)
1523 kmem_cache_free(cfq_pool, new_cfqq);
1524
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525out:
1526 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1527 return cfqq;
1528}
1529
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001530static struct cfq_queue **
1531cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1532{
Jens Axboefe094d92008-01-31 13:08:54 +01001533 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001534 case IOPRIO_CLASS_RT:
1535 return &cfqd->async_cfqq[0][ioprio];
1536 case IOPRIO_CLASS_BE:
1537 return &cfqd->async_cfqq[1][ioprio];
1538 case IOPRIO_CLASS_IDLE:
1539 return &cfqd->async_idle_cfqq;
1540 default:
1541 BUG();
1542 }
1543}
1544
Jens Axboe15c31be2007-07-10 13:43:25 +02001545static struct cfq_queue *
Jens Axboefd0928d2008-01-24 08:52:45 +01001546cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02001547 gfp_t gfp_mask)
1548{
Jens Axboefd0928d2008-01-24 08:52:45 +01001549 const int ioprio = task_ioprio(ioc);
1550 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001551 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02001552 struct cfq_queue *cfqq = NULL;
1553
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001554 if (!is_sync) {
1555 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1556 cfqq = *async_cfqq;
1557 }
1558
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001559 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01001560 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Oleg Nesterov0a0836a2007-10-23 15:08:21 +02001561 if (!cfqq)
1562 return NULL;
1563 }
Jens Axboe15c31be2007-07-10 13:43:25 +02001564
1565 /*
1566 * pin the queue now that it's allocated, scheduler exit will prune it
1567 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001568 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02001569 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02001570 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02001571 }
1572
1573 atomic_inc(&cfqq->ref);
1574 return cfqq;
1575}
1576
Jens Axboe498d3aa2007-04-26 12:54:48 +02001577/*
1578 * We drop cfq io contexts lazily, so we may find a dead one.
1579 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001580static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01001581cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1582 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001583{
Jens Axboe4ac845a2008-01-24 08:44:49 +01001584 unsigned long flags;
1585
Jens Axboefc463792006-08-29 09:05:44 +02001586 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02001587
Jens Axboe4ac845a2008-01-24 08:44:49 +01001588 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02001589
Fabio Checconi4faa3c82008-04-10 08:28:01 +02001590 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001591
1592 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01001593 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001594 spin_unlock_irqrestore(&ioc->lock, flags);
1595
1596 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001597}
1598
Jens Axboee2d74ac2006-03-28 08:59:01 +02001599static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01001600cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001601{
Jens Axboee2d74ac2006-03-28 08:59:01 +02001602 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001603 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001604 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001605
Vasily Tarasov91fac312007-04-25 12:29:51 +02001606 if (unlikely(!ioc))
1607 return NULL;
1608
Jens Axboed6de8be2008-05-28 14:46:59 +02001609 rcu_read_lock();
1610
Jens Axboe597bc482007-04-24 21:23:53 +02001611 /*
1612 * we maintain a last-hit cache, to avoid browsing over the tree
1613 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001614 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02001615 if (cic && cic->key == cfqd) {
1616 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02001617 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02001618 }
Jens Axboe597bc482007-04-24 21:23:53 +02001619
Jens Axboe4ac845a2008-01-24 08:44:49 +01001620 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001621 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1622 rcu_read_unlock();
1623 if (!cic)
1624 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02001625 /* ->key must be copied to avoid race with cfq_exit_queue() */
1626 k = cic->key;
1627 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01001628 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001629 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01001630 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001631 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001632
Jens Axboed6de8be2008-05-28 14:46:59 +02001633 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001634 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02001635 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001636 break;
1637 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001638
Jens Axboe4ac845a2008-01-24 08:44:49 +01001639 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001640}
1641
Jens Axboe4ac845a2008-01-24 08:44:49 +01001642/*
1643 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1644 * the process specific cfq io context when entered from the block layer.
1645 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1646 */
Jens Axboefebffd62008-01-28 13:19:43 +01001647static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1648 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02001649{
Jens Axboe0261d682006-10-30 19:07:48 +01001650 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001651 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001652
Jens Axboe4ac845a2008-01-24 08:44:49 +01001653 ret = radix_tree_preload(gfp_mask);
1654 if (!ret) {
1655 cic->ioc = ioc;
1656 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001657
Jens Axboe4ac845a2008-01-24 08:44:49 +01001658 spin_lock_irqsave(&ioc->lock, flags);
1659 ret = radix_tree_insert(&ioc->radix_root,
1660 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01001661 if (!ret)
1662 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01001663 spin_unlock_irqrestore(&ioc->lock, flags);
1664
1665 radix_tree_preload_end();
1666
1667 if (!ret) {
1668 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1669 list_add(&cic->queue_list, &cfqd->cic_list);
1670 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02001671 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02001672 }
1673
Jens Axboe4ac845a2008-01-24 08:44:49 +01001674 if (ret)
1675 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02001676
Jens Axboe4ac845a2008-01-24 08:44:49 +01001677 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02001678}
1679
Jens Axboe22e2c502005-06-27 10:55:12 +02001680/*
1681 * Setup general io context and cfq io context. There can be several cfq
1682 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02001683 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02001684 */
1685static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02001686cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687{
Jens Axboe22e2c502005-06-27 10:55:12 +02001688 struct io_context *ioc = NULL;
1689 struct cfq_io_context *cic;
1690
1691 might_sleep_if(gfp_mask & __GFP_WAIT);
1692
Jens Axboeb5deef92006-07-19 23:39:40 +02001693 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02001694 if (!ioc)
1695 return NULL;
1696
Jens Axboe4ac845a2008-01-24 08:44:49 +01001697 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02001698 if (cic)
1699 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02001700
Jens Axboee2d74ac2006-03-28 08:59:01 +02001701 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1702 if (cic == NULL)
1703 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02001704
Jens Axboe4ac845a2008-01-24 08:44:49 +01001705 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1706 goto err_free;
1707
Jens Axboe22e2c502005-06-27 10:55:12 +02001708out:
Jens Axboefc463792006-08-29 09:05:44 +02001709 smp_read_barrier_depends();
1710 if (unlikely(ioc->ioprio_changed))
1711 cfq_ioc_set_ioprio(ioc);
1712
Jens Axboe22e2c502005-06-27 10:55:12 +02001713 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01001714err_free:
1715 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02001716err:
1717 put_io_context(ioc);
1718 return NULL;
1719}
1720
1721static void
1722cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1723{
Jens Axboeaaf12282007-01-19 11:30:16 +11001724 unsigned long elapsed = jiffies - cic->last_end_request;
1725 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02001726
1727 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1728 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1729 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1730}
1731
Jens Axboe206dc692006-03-28 13:03:44 +02001732static void
Jens Axboe6d048f52007-04-25 12:44:27 +02001733cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1734 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02001735{
1736 sector_t sdist;
1737 u64 total;
1738
Jens Axboe5e705372006-07-13 12:39:25 +02001739 if (cic->last_request_pos < rq->sector)
1740 sdist = rq->sector - cic->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02001741 else
Jens Axboe5e705372006-07-13 12:39:25 +02001742 sdist = cic->last_request_pos - rq->sector;
Jens Axboe206dc692006-03-28 13:03:44 +02001743
1744 /*
1745 * Don't allow the seek distance to get too large from the
1746 * odd fragment, pagein, etc
1747 */
1748 if (cic->seek_samples <= 60) /* second&third seek */
1749 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1750 else
1751 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1752
1753 cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1754 cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1755 total = cic->seek_total + (cic->seek_samples/2);
1756 do_div(total, cic->seek_samples);
1757 cic->seek_mean = (sector_t)total;
1758}
Jens Axboe22e2c502005-06-27 10:55:12 +02001759
1760/*
1761 * Disable idle window if the process thinks too long or seeks so much that
1762 * it doesn't matter
1763 */
1764static void
1765cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1766 struct cfq_io_context *cic)
1767{
Jens Axboe7b679132008-05-30 12:23:07 +02001768 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02001769
Jens Axboe08717142008-01-28 11:38:15 +01001770 /*
1771 * Don't idle for async or idle io prio class
1772 */
1773 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02001774 return;
1775
Jens Axboec265a7f2008-06-26 13:49:33 +02001776 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001777
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001778 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Jens Axboecaaa5f92006-06-16 11:23:00 +02001779 (cfqd->hw_tag && CIC_SEEKY(cic)))
Jens Axboe22e2c502005-06-27 10:55:12 +02001780 enable_idle = 0;
1781 else if (sample_valid(cic->ttime_samples)) {
1782 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1783 enable_idle = 0;
1784 else
1785 enable_idle = 1;
1786 }
1787
Jens Axboe7b679132008-05-30 12:23:07 +02001788 if (old_idle != enable_idle) {
1789 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1790 if (enable_idle)
1791 cfq_mark_cfqq_idle_window(cfqq);
1792 else
1793 cfq_clear_cfqq_idle_window(cfqq);
1794 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001795}
1796
Jens Axboe22e2c502005-06-27 10:55:12 +02001797/*
1798 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1799 * no or if we aren't sure, a 1 will cause a preempt.
1800 */
1801static int
1802cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02001803 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001804{
Jens Axboe6d048f52007-04-25 12:44:27 +02001805 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001806
Jens Axboe6d048f52007-04-25 12:44:27 +02001807 cfqq = cfqd->active_queue;
1808 if (!cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001809 return 0;
1810
Jens Axboe6d048f52007-04-25 12:44:27 +02001811 if (cfq_slice_used(cfqq))
1812 return 1;
1813
1814 if (cfq_class_idle(new_cfqq))
Jens Axboecaaa5f92006-06-16 11:23:00 +02001815 return 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001816
1817 if (cfq_class_idle(cfqq))
1818 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001819
Jens Axboe22e2c502005-06-27 10:55:12 +02001820 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02001821 * if the new request is sync, but the currently running queue is
1822 * not, let the sync request have priority.
1823 */
Jens Axboe5e705372006-07-13 12:39:25 +02001824 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02001825 return 1;
Jens Axboe1e3335d2007-02-14 19:59:49 +01001826
Jens Axboe374f84a2006-07-23 01:42:19 +02001827 /*
1828 * So both queues are sync. Let the new request get disk time if
1829 * it's a metadata request and the current queue is doing regular IO.
1830 */
1831 if (rq_is_meta(rq) && !cfqq->meta_pending)
1832 return 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02001833
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01001834 /*
1835 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
1836 */
1837 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
1838 return 1;
1839
Jens Axboe1e3335d2007-02-14 19:59:49 +01001840 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1841 return 0;
1842
1843 /*
1844 * if this request is as-good as one we would expect from the
1845 * current cfqq, let it preempt
1846 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001847 if (cfq_rq_close(cfqd, rq))
Jens Axboe1e3335d2007-02-14 19:59:49 +01001848 return 1;
1849
Jens Axboe22e2c502005-06-27 10:55:12 +02001850 return 0;
1851}
1852
1853/*
1854 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1855 * let it have half of its nominal slice.
1856 */
1857static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1858{
Jens Axboe7b679132008-05-30 12:23:07 +02001859 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02001860 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02001861
Jens Axboebf572252006-07-19 20:29:12 +02001862 /*
1863 * Put the new queue at the front of the of the current list,
1864 * so we know that it will be selected next.
1865 */
1866 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02001867
1868 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02001869
Jens Axboe44f7c162007-01-19 11:51:58 +11001870 cfqq->slice_end = 0;
1871 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001872}
1873
1874/*
Jens Axboe5e705372006-07-13 12:39:25 +02001875 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02001876 * something we should do about it
1877 */
1878static void
Jens Axboe5e705372006-07-13 12:39:25 +02001879cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1880 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001881{
Jens Axboe5e705372006-07-13 12:39:25 +02001882 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02001883
Aaron Carroll45333d52008-08-26 15:52:36 +02001884 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02001885 if (rq_is_meta(rq))
1886 cfqq->meta_pending++;
1887
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001888 cfq_update_io_thinktime(cfqd, cic);
Jens Axboe6d048f52007-04-25 12:44:27 +02001889 cfq_update_io_seektime(cfqd, cic, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001890 cfq_update_idle_window(cfqd, cfqq, cic);
1891
Jens Axboe5e705372006-07-13 12:39:25 +02001892 cic->last_request_pos = rq->sector + rq->nr_sectors;
Jens Axboe22e2c502005-06-27 10:55:12 +02001893
1894 if (cfqq == cfqd->active_queue) {
1895 /*
1896 * if we are waiting for a request for this queue, let it rip
1897 * immediately and flag that we must not expire this queue
1898 * just now
1899 */
Jens Axboe3b181522005-06-27 10:56:24 +02001900 if (cfq_cfqq_wait_request(cfqq)) {
1901 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001902 del_timer(&cfqd->idle_slice_timer);
Jens Axboedc72ef42006-07-20 14:54:05 +02001903 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001904 }
Jens Axboe5e705372006-07-13 12:39:25 +02001905 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001906 /*
1907 * not the active queue - expire current slice if it is
1908 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01001909 * has some old slice time left and is of higher priority or
1910 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02001911 */
1912 cfq_preempt_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001913 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboedc72ef42006-07-20 14:54:05 +02001914 blk_start_queueing(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02001915 }
1916}
1917
Jens Axboe165125e2007-07-24 09:28:11 +02001918static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001919{
Jens Axboeb4878f22005-10-20 16:42:29 +02001920 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001921 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001922
Jens Axboe7b679132008-05-30 12:23:07 +02001923 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01001924 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925
Jens Axboe5e705372006-07-13 12:39:25 +02001926 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927
Jens Axboe22e2c502005-06-27 10:55:12 +02001928 list_add_tail(&rq->queuelist, &cfqq->fifo);
1929
Jens Axboe5e705372006-07-13 12:39:25 +02001930 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931}
1932
Aaron Carroll45333d52008-08-26 15:52:36 +02001933/*
1934 * Update hw_tag based on peak queue depth over 50 samples under
1935 * sufficient load.
1936 */
1937static void cfq_update_hw_tag(struct cfq_data *cfqd)
1938{
1939 if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
1940 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
1941
1942 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
1943 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
1944 return;
1945
1946 if (cfqd->hw_tag_samples++ < 50)
1947 return;
1948
1949 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
1950 cfqd->hw_tag = 1;
1951 else
1952 cfqd->hw_tag = 0;
1953
1954 cfqd->hw_tag_samples = 0;
1955 cfqd->rq_in_driver_peak = 0;
1956}
1957
Jens Axboe165125e2007-07-24 09:28:11 +02001958static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959{
Jens Axboe5e705372006-07-13 12:39:25 +02001960 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001961 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02001962 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001963 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964
Jens Axboeb4878f22005-10-20 16:42:29 +02001965 now = jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001966 cfq_log_cfqq(cfqd, cfqq, "complete");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967
Aaron Carroll45333d52008-08-26 15:52:36 +02001968 cfq_update_hw_tag(cfqd);
1969
Jens Axboeb4878f22005-10-20 16:42:29 +02001970 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02001971 WARN_ON(!cfqq->dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02001972 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02001973 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974
Jens Axboe3ed9a292007-04-23 08:33:33 +02001975 if (cfq_cfqq_sync(cfqq))
1976 cfqd->sync_flight--;
1977
Jens Axboeb4878f22005-10-20 16:42:29 +02001978 if (!cfq_class_idle(cfqq))
1979 cfqd->last_end_request = now;
Jens Axboe3b181522005-06-27 10:56:24 +02001980
Jens Axboecaaa5f92006-06-16 11:23:00 +02001981 if (sync)
Jens Axboe5e705372006-07-13 12:39:25 +02001982 RQ_CIC(rq)->last_end_request = now;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001983
1984 /*
1985 * If this is the active queue, check if it needs to be expired,
1986 * or if we want to idle in case it has no pending requests.
1987 */
1988 if (cfqd->active_queue == cfqq) {
Jens Axboe44f7c162007-01-19 11:51:58 +11001989 if (cfq_cfqq_slice_new(cfqq)) {
1990 cfq_set_prio_slice(cfqd, cfqq);
1991 cfq_clear_cfqq_slice_new(cfqq);
1992 }
Jens Axboe08717142008-01-28 11:38:15 +01001993 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02001994 cfq_slice_expired(cfqd, 1);
Jens Axboeaeb6faf2009-04-06 14:48:07 +02001995 else if (sync && !rq_noidle(rq) &&
1996 RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe6d048f52007-04-25 12:44:27 +02001997 cfq_arm_slice_timer(cfqd);
Jens Axboeaeb6faf2009-04-06 14:48:07 +02001998 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02001999 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002000
2001 if (!cfqd->rq_in_driver)
2002 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003}
2004
Jens Axboe22e2c502005-06-27 10:55:12 +02002005/*
2006 * we temporarily boost lower priority queues if they are holding fs exclusive
2007 * resources. they are boosted to normal prio (CLASS_BE/4)
2008 */
2009static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010{
Jens Axboe22e2c502005-06-27 10:55:12 +02002011 if (has_fs_excl()) {
2012 /*
2013 * boost idle prio on transactions that would lock out other
2014 * users of the filesystem
2015 */
2016 if (cfq_class_idle(cfqq))
2017 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2018 if (cfqq->ioprio > IOPRIO_NORM)
2019 cfqq->ioprio = IOPRIO_NORM;
2020 } else {
2021 /*
2022 * check if we need to unboost the queue
2023 */
2024 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2025 cfqq->ioprio_class = cfqq->org_ioprio_class;
2026 if (cfqq->ioprio != cfqq->org_ioprio)
2027 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002029}
2030
Jens Axboe89850f72006-07-22 16:48:31 +02002031static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002032{
Jens Axboe3b181522005-06-27 10:56:24 +02002033 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
Andrew Morton99f95e52005-06-27 20:14:05 -07002034 !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002035 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002036 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02002037 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002038
2039 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02002040}
2041
Jens Axboe165125e2007-07-24 09:28:11 +02002042static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02002043{
2044 struct cfq_data *cfqd = q->elevator->elevator_data;
2045 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002046 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02002047 struct cfq_queue *cfqq;
2048
2049 /*
2050 * don't force setup of a queue from here, as a call to may_queue
2051 * does not necessarily imply that a request actually will be queued.
2052 * so just lookup a possibly existing queue, or return 'may queue'
2053 * if that fails
2054 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002055 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002056 if (!cic)
2057 return ELV_MQUEUE_MAY;
2058
2059 cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
Jens Axboe22e2c502005-06-27 10:55:12 +02002060 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002061 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02002062 cfq_prio_boost(cfqq);
2063
Jens Axboe89850f72006-07-22 16:48:31 +02002064 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002065 }
2066
2067 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068}
2069
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070/*
2071 * queue lock held here
2072 */
Jens Axboebb37b942006-12-01 10:42:33 +01002073static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074{
Jens Axboe5e705372006-07-13 12:39:25 +02002075 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076
Jens Axboe5e705372006-07-13 12:39:25 +02002077 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002078 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079
Jens Axboe22e2c502005-06-27 10:55:12 +02002080 BUG_ON(!cfqq->allocated[rw]);
2081 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082
Jens Axboe5e705372006-07-13 12:39:25 +02002083 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02002086 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088 cfq_put_queue(cfqq);
2089 }
2090}
2091
2092/*
Jens Axboe22e2c502005-06-27 10:55:12 +02002093 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 */
Jens Axboe22e2c502005-06-27 10:55:12 +02002095static int
Jens Axboe165125e2007-07-24 09:28:11 +02002096cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097{
2098 struct cfq_data *cfqd = q->elevator->elevator_data;
2099 struct cfq_io_context *cic;
2100 const int rw = rq_data_dir(rq);
Jens Axboe7749a8d2006-12-13 13:02:26 +01002101 const int is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002102 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 unsigned long flags;
2104
2105 might_sleep_if(gfp_mask & __GFP_WAIT);
2106
Jens Axboee2d74ac2006-03-28 08:59:01 +02002107 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02002108
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109 spin_lock_irqsave(q->queue_lock, flags);
2110
Jens Axboe22e2c502005-06-27 10:55:12 +02002111 if (!cic)
2112 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113
Vasily Tarasov91fac312007-04-25 12:29:51 +02002114 cfqq = cic_to_cfqq(cic, is_sync);
2115 if (!cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01002116 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002117
Jens Axboe22e2c502005-06-27 10:55:12 +02002118 if (!cfqq)
2119 goto queue_fail;
2120
Vasily Tarasov91fac312007-04-25 12:29:51 +02002121 cic_set_cfqq(cic, cfqq, is_sync);
2122 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123
2124 cfqq->allocated[rw]++;
Jens Axboe3b181522005-06-27 10:56:24 +02002125 cfq_clear_cfqq_must_alloc(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002126 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02002127
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 spin_unlock_irqrestore(q->queue_lock, flags);
2129
Jens Axboe5e705372006-07-13 12:39:25 +02002130 rq->elevator_private = cic;
2131 rq->elevator_private2 = cfqq;
2132 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02002133
Jens Axboe22e2c502005-06-27 10:55:12 +02002134queue_fail:
2135 if (cic)
2136 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02002137
Jens Axboe3b181522005-06-27 10:56:24 +02002138 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02002140 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 return 1;
2142}
2143
David Howells65f27f32006-11-22 14:55:48 +00002144static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02002145{
David Howells65f27f32006-11-22 14:55:48 +00002146 struct cfq_data *cfqd =
2147 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02002148 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002149 unsigned long flags;
2150
2151 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02002152 blk_start_queueing(q);
Jens Axboe22e2c502005-06-27 10:55:12 +02002153 spin_unlock_irqrestore(q->queue_lock, flags);
2154}
2155
2156/*
2157 * Timer running if the active_queue is currently idling inside its time slice
2158 */
2159static void cfq_idle_slice_timer(unsigned long data)
2160{
2161 struct cfq_data *cfqd = (struct cfq_data *) data;
2162 struct cfq_queue *cfqq;
2163 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002164 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02002165
Jens Axboe7b679132008-05-30 12:23:07 +02002166 cfq_log(cfqd, "idle timer fired");
2167
Jens Axboe22e2c502005-06-27 10:55:12 +02002168 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2169
Jens Axboefe094d92008-01-31 13:08:54 +01002170 cfqq = cfqd->active_queue;
2171 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11002172 timed_out = 0;
2173
Jens Axboe22e2c502005-06-27 10:55:12 +02002174 /*
2175 * expired
2176 */
Jens Axboe44f7c162007-01-19 11:51:58 +11002177 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002178 goto expire;
2179
2180 /*
2181 * only expire and reinvoke request handler, if there are
2182 * other queues with pending requests
2183 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02002184 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02002185 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02002186
2187 /*
2188 * not expired and it has a request pending, let it dispatch
2189 */
Jens Axboedd67d052006-06-21 09:36:18 +02002190 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
Jens Axboe3b181522005-06-27 10:56:24 +02002191 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002192 goto out_kick;
2193 }
2194 }
2195expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002196 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02002197out_kick:
Jens Axboe3b181522005-06-27 10:56:24 +02002198 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02002199out_cont:
2200 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2201}
2202
Jens Axboe3b181522005-06-27 10:56:24 +02002203static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2204{
2205 del_timer_sync(&cfqd->idle_slice_timer);
Cheng Renquan64d01dc2008-12-03 12:41:39 +01002206 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02002207}
Jens Axboe22e2c502005-06-27 10:55:12 +02002208
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002209static void cfq_put_async_queues(struct cfq_data *cfqd)
2210{
2211 int i;
2212
2213 for (i = 0; i < IOPRIO_BE_NR; i++) {
2214 if (cfqd->async_cfqq[0][i])
2215 cfq_put_queue(cfqd->async_cfqq[0][i]);
2216 if (cfqd->async_cfqq[1][i])
2217 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002218 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01002219
2220 if (cfqd->async_idle_cfqq)
2221 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002222}
2223
Jens Axboeb374d182008-10-31 10:05:07 +01002224static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225{
Jens Axboe22e2c502005-06-27 10:55:12 +02002226 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02002227 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002228
Jens Axboe3b181522005-06-27 10:56:24 +02002229 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002230
Al Virod9ff4182006-03-18 13:51:22 -05002231 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002232
Al Virod9ff4182006-03-18 13:51:22 -05002233 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02002234 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002235
2236 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05002237 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2238 struct cfq_io_context,
2239 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02002240
2241 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05002242 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002243
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002244 cfq_put_async_queues(cfqd);
Jens Axboe15c31be2007-07-10 13:43:25 +02002245
Al Virod9ff4182006-03-18 13:51:22 -05002246 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05002247
2248 cfq_shutdown_timer_wq(cfqd);
2249
Al Viroa90d7422006-03-18 12:05:37 -05002250 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251}
2252
Jens Axboe165125e2007-07-24 09:28:11 +02002253static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254{
2255 struct cfq_data *cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256
Christoph Lameter94f60302007-07-17 04:03:29 -07002257 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02002259 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260
Jens Axboecc09e292007-04-26 12:53:50 +02002261 cfqd->service_tree = CFQ_RB_ROOT;
Al Virod9ff4182006-03-18 13:51:22 -05002262 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265
Jens Axboe22e2c502005-06-27 10:55:12 +02002266 init_timer(&cfqd->idle_slice_timer);
2267 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2268 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2269
David Howells65f27f32006-11-22 14:55:48 +00002270 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02002271
Oleg Nesterovb70c8642007-11-07 09:46:13 +01002272 cfqd->last_end_request = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02002274 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2275 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 cfqd->cfq_back_max = cfq_back_max;
2277 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02002278 cfqd->cfq_slice[0] = cfq_slice_async;
2279 cfqd->cfq_slice[1] = cfq_slice_sync;
2280 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2281 cfqd->cfq_slice_idle = cfq_slice_idle;
Aaron Carroll45333d52008-08-26 15:52:36 +02002282 cfqd->hw_tag = 1;
Jens Axboe3b181522005-06-27 10:56:24 +02002283
Jens Axboebc1c1162006-06-08 08:49:06 +02002284 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285}
2286
2287static void cfq_slab_kill(void)
2288{
Jens Axboed6de8be2008-05-28 14:46:59 +02002289 /*
2290 * Caller already ensured that pending RCU callbacks are completed,
2291 * so we should have no busy allocations at this point.
2292 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 if (cfq_pool)
2294 kmem_cache_destroy(cfq_pool);
2295 if (cfq_ioc_pool)
2296 kmem_cache_destroy(cfq_ioc_pool);
2297}
2298
2299static int __init cfq_slab_setup(void)
2300{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07002301 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302 if (!cfq_pool)
2303 goto fail;
2304
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002305 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306 if (!cfq_ioc_pool)
2307 goto fail;
2308
2309 return 0;
2310fail:
2311 cfq_slab_kill();
2312 return -ENOMEM;
2313}
2314
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315/*
2316 * sysfs parts below -->
2317 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318static ssize_t
2319cfq_var_show(unsigned int var, char *page)
2320{
2321 return sprintf(page, "%d\n", var);
2322}
2323
2324static ssize_t
2325cfq_var_store(unsigned int *var, const char *page, size_t count)
2326{
2327 char *p = (char *) page;
2328
2329 *var = simple_strtoul(p, &p, 10);
2330 return count;
2331}
2332
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002334static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002336 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 unsigned int __data = __VAR; \
2338 if (__CONV) \
2339 __data = jiffies_to_msecs(__data); \
2340 return cfq_var_show(__data, (page)); \
2341}
2342SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002343SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2344SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05002345SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2346SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002347SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2348SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2349SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2350SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351#undef SHOW_FUNCTION
2352
2353#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01002354static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355{ \
Al Viro3d1ab402006-03-18 18:35:43 -05002356 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 unsigned int __data; \
2358 int ret = cfq_var_store(&__data, (page), count); \
2359 if (__data < (MIN)) \
2360 __data = (MIN); \
2361 else if (__data > (MAX)) \
2362 __data = (MAX); \
2363 if (__CONV) \
2364 *(__PTR) = msecs_to_jiffies(__data); \
2365 else \
2366 *(__PTR) = __data; \
2367 return ret; \
2368}
2369STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002370STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2371 UINT_MAX, 1);
2372STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2373 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05002374STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01002375STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2376 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02002377STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2378STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2379STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01002380STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2381 UINT_MAX, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382#undef STORE_FUNCTION
2383
Al Viroe572ec72006-03-18 22:27:18 -05002384#define CFQ_ATTR(name) \
2385 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02002386
Al Viroe572ec72006-03-18 22:27:18 -05002387static struct elv_fs_entry cfq_attrs[] = {
2388 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05002389 CFQ_ATTR(fifo_expire_sync),
2390 CFQ_ATTR(fifo_expire_async),
2391 CFQ_ATTR(back_seek_max),
2392 CFQ_ATTR(back_seek_penalty),
2393 CFQ_ATTR(slice_sync),
2394 CFQ_ATTR(slice_async),
2395 CFQ_ATTR(slice_async_rq),
2396 CFQ_ATTR(slice_idle),
Al Viroe572ec72006-03-18 22:27:18 -05002397 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398};
2399
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400static struct elevator_type iosched_cfq = {
2401 .ops = {
2402 .elevator_merge_fn = cfq_merge,
2403 .elevator_merged_fn = cfq_merged_request,
2404 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01002405 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02002406 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02002408 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409 .elevator_deactivate_req_fn = cfq_deactivate_request,
2410 .elevator_queue_empty_fn = cfq_queue_empty,
2411 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02002412 .elevator_former_req_fn = elv_rb_former_request,
2413 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414 .elevator_set_req_fn = cfq_set_request,
2415 .elevator_put_req_fn = cfq_put_request,
2416 .elevator_may_queue_fn = cfq_may_queue,
2417 .elevator_init_fn = cfq_init_queue,
2418 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02002419 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420 },
Al Viro3d1ab402006-03-18 18:35:43 -05002421 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422 .elevator_name = "cfq",
2423 .elevator_owner = THIS_MODULE,
2424};
2425
2426static int __init cfq_init(void)
2427{
Jens Axboe22e2c502005-06-27 10:55:12 +02002428 /*
2429 * could be 0 on HZ < 1000 setups
2430 */
2431 if (!cfq_slice_async)
2432 cfq_slice_async = 1;
2433 if (!cfq_slice_idle)
2434 cfq_slice_idle = 1;
2435
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436 if (cfq_slab_setup())
2437 return -ENOMEM;
2438
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002439 elv_register(&iosched_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01002441 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442}
2443
2444static void __exit cfq_exit(void)
2445{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07002446 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05002448 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02002449 /* ioc_gone's update must be visible before reading ioc_count */
2450 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02002451
2452 /*
2453 * this also protects us from entering cfq_slab_kill() with
2454 * pending RCU callbacks
2455 */
Jens Axboe4050cf12006-07-19 05:07:12 +02002456 if (elv_ioc_count_read(ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002457 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08002458 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459}
2460
2461module_init(cfq_init);
2462module_exit(cfq_exit);
2463
2464MODULE_AUTHOR("Jens Axboe");
2465MODULE_LICENSE("GPL");
2466MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");