Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
1da177e4 LT |
2 | * 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 | * | |
0fe23479 | 7 | * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk> |
1da177e4 | 8 | */ |
1da177e4 | 9 | #include <linux/module.h> |
1cc9be68 AV |
10 | #include <linux/blkdev.h> |
11 | #include <linux/elevator.h> | |
1da177e4 LT |
12 | #include <linux/hash.h> |
13 | #include <linux/rbtree.h> | |
22e2c507 | 14 | #include <linux/ioprio.h> |
1da177e4 LT |
15 | |
16 | /* | |
17 | * tunables | |
18 | */ | |
64100099 | 19 | static const int cfq_quantum = 4; /* max queue in one round of service */ |
64100099 AV |
20 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
21 | static const int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */ | |
22 | static const int cfq_back_penalty = 2; /* penalty of a backwards seek */ | |
1da177e4 | 23 | |
64100099 | 24 | static const int cfq_slice_sync = HZ / 10; |
3b18152c | 25 | static int cfq_slice_async = HZ / 25; |
64100099 | 26 | static const int cfq_slice_async_rq = 2; |
caaa5f9f | 27 | static int cfq_slice_idle = HZ / 125; |
22e2c507 | 28 | |
d9e7620e JA |
29 | /* |
30 | * grace period before allowing idle class to get disk access | |
31 | */ | |
22e2c507 | 32 | #define CFQ_IDLE_GRACE (HZ / 10) |
d9e7620e JA |
33 | |
34 | /* | |
35 | * below this threshold, we consider thinktime immediate | |
36 | */ | |
37 | #define CFQ_MIN_TT (2) | |
38 | ||
22e2c507 JA |
39 | #define CFQ_SLICE_SCALE (5) |
40 | ||
41 | #define CFQ_KEY_ASYNC (0) | |
22e2c507 | 42 | |
1da177e4 LT |
43 | /* |
44 | * for the hash of cfqq inside the cfqd | |
45 | */ | |
46 | #define CFQ_QHASH_SHIFT 6 | |
47 | #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT) | |
1da177e4 | 48 | |
5e705374 JA |
49 | #define RQ_CIC(rq) ((struct cfq_io_context*)(rq)->elevator_private) |
50 | #define RQ_CFQQ(rq) ((rq)->elevator_private2) | |
1da177e4 | 51 | |
e18b890b CL |
52 | static struct kmem_cache *cfq_pool; |
53 | static struct kmem_cache *cfq_ioc_pool; | |
1da177e4 | 54 | |
4050cf16 | 55 | static DEFINE_PER_CPU(unsigned long, ioc_count); |
334e94de AV |
56 | static struct completion *ioc_gone; |
57 | ||
22e2c507 JA |
58 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR |
59 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
22e2c507 JA |
60 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) |
61 | ||
3b18152c JA |
62 | #define ASYNC (0) |
63 | #define SYNC (1) | |
64 | ||
6d048f53 | 65 | #define cfq_cfqq_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC) |
22e2c507 | 66 | |
206dc69b JA |
67 | #define sample_valid(samples) ((samples) > 80) |
68 | ||
cc09e299 JA |
69 | /* |
70 | * Most of our rbtree usage is for sorting with min extraction, so | |
71 | * if we cache the leftmost node we don't have to walk down the tree | |
72 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
73 | * move this into the elevator for the rq sorting as well. | |
74 | */ | |
75 | struct cfq_rb_root { | |
76 | struct rb_root rb; | |
77 | struct rb_node *left; | |
78 | }; | |
79 | #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, } | |
80 | ||
22e2c507 JA |
81 | /* |
82 | * Per block device queue structure | |
83 | */ | |
1da177e4 | 84 | struct cfq_data { |
22e2c507 JA |
85 | request_queue_t *queue; |
86 | ||
87 | /* | |
88 | * rr list of queues with requests and the count of them | |
89 | */ | |
cc09e299 | 90 | struct cfq_rb_root service_tree; |
22e2c507 JA |
91 | unsigned int busy_queues; |
92 | ||
22e2c507 JA |
93 | /* |
94 | * cfqq lookup hash | |
95 | */ | |
1da177e4 | 96 | struct hlist_head *cfq_hash; |
1da177e4 | 97 | |
22e2c507 | 98 | int rq_in_driver; |
3ed9a296 | 99 | int sync_flight; |
25776e35 | 100 | int hw_tag; |
1da177e4 | 101 | |
22e2c507 JA |
102 | /* |
103 | * idle window management | |
104 | */ | |
105 | struct timer_list idle_slice_timer; | |
106 | struct work_struct unplug_work; | |
1da177e4 | 107 | |
22e2c507 JA |
108 | struct cfq_queue *active_queue; |
109 | struct cfq_io_context *active_cic; | |
22e2c507 JA |
110 | |
111 | struct timer_list idle_class_timer; | |
1da177e4 | 112 | |
6d048f53 | 113 | sector_t last_position; |
22e2c507 | 114 | unsigned long last_end_request; |
1da177e4 | 115 | |
1da177e4 LT |
116 | /* |
117 | * tunables, see top of file | |
118 | */ | |
119 | unsigned int cfq_quantum; | |
22e2c507 | 120 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
121 | unsigned int cfq_back_penalty; |
122 | unsigned int cfq_back_max; | |
22e2c507 JA |
123 | unsigned int cfq_slice[2]; |
124 | unsigned int cfq_slice_async_rq; | |
125 | unsigned int cfq_slice_idle; | |
d9ff4187 AV |
126 | |
127 | struct list_head cic_list; | |
6d048f53 JA |
128 | |
129 | sector_t new_seek_mean; | |
130 | u64 new_seek_total; | |
1da177e4 LT |
131 | }; |
132 | ||
22e2c507 JA |
133 | /* |
134 | * Per process-grouping structure | |
135 | */ | |
1da177e4 LT |
136 | struct cfq_queue { |
137 | /* reference count */ | |
138 | atomic_t ref; | |
139 | /* parent cfq_data */ | |
140 | struct cfq_data *cfqd; | |
22e2c507 | 141 | /* cfqq lookup hash */ |
1da177e4 LT |
142 | struct hlist_node cfq_hash; |
143 | /* hash key */ | |
22e2c507 | 144 | unsigned int key; |
d9e7620e JA |
145 | /* service_tree member */ |
146 | struct rb_node rb_node; | |
147 | /* service_tree key */ | |
148 | unsigned long rb_key; | |
1da177e4 LT |
149 | /* sorted list of pending requests */ |
150 | struct rb_root sort_list; | |
151 | /* if fifo isn't expired, next request to serve */ | |
5e705374 | 152 | struct request *next_rq; |
1da177e4 LT |
153 | /* requests queued in sort_list */ |
154 | int queued[2]; | |
155 | /* currently allocated requests */ | |
156 | int allocated[2]; | |
374f84ac JA |
157 | /* pending metadata requests */ |
158 | int meta_pending; | |
1da177e4 | 159 | /* fifo list of requests in sort_list */ |
22e2c507 | 160 | struct list_head fifo; |
1da177e4 | 161 | |
22e2c507 | 162 | unsigned long slice_end; |
c5b680f3 | 163 | long slice_resid; |
1da177e4 | 164 | |
6d048f53 JA |
165 | /* number of requests that are on the dispatch list or inside driver */ |
166 | int dispatched; | |
22e2c507 JA |
167 | |
168 | /* io prio of this group */ | |
169 | unsigned short ioprio, org_ioprio; | |
170 | unsigned short ioprio_class, org_ioprio_class; | |
171 | ||
3b18152c JA |
172 | /* various state flags, see below */ |
173 | unsigned int flags; | |
6d048f53 JA |
174 | |
175 | sector_t last_request_pos; | |
1da177e4 LT |
176 | }; |
177 | ||
3b18152c | 178 | enum cfqq_state_flags { |
b0b8d749 JA |
179 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
180 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
181 | CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */ | |
182 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ | |
183 | CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */ | |
184 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ | |
185 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
186 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
187 | CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */ | |
44f7c160 | 188 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
3b18152c JA |
189 | }; |
190 | ||
191 | #define CFQ_CFQQ_FNS(name) \ | |
192 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
193 | { \ | |
194 | cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ | |
195 | } \ | |
196 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
197 | { \ | |
198 | cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ | |
199 | } \ | |
200 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
201 | { \ | |
202 | return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ | |
203 | } | |
204 | ||
205 | CFQ_CFQQ_FNS(on_rr); | |
206 | CFQ_CFQQ_FNS(wait_request); | |
207 | CFQ_CFQQ_FNS(must_alloc); | |
208 | CFQ_CFQQ_FNS(must_alloc_slice); | |
209 | CFQ_CFQQ_FNS(must_dispatch); | |
210 | CFQ_CFQQ_FNS(fifo_expire); | |
211 | CFQ_CFQQ_FNS(idle_window); | |
212 | CFQ_CFQQ_FNS(prio_changed); | |
53b03744 | 213 | CFQ_CFQQ_FNS(queue_new); |
44f7c160 | 214 | CFQ_CFQQ_FNS(slice_new); |
3b18152c JA |
215 | #undef CFQ_CFQQ_FNS |
216 | ||
3b18152c | 217 | static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short); |
5e705374 | 218 | static void cfq_dispatch_insert(request_queue_t *, struct request *); |
498d3aa2 | 219 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, unsigned int, struct task_struct *, gfp_t); |
1da177e4 | 220 | |
99f95e52 AM |
221 | /* |
222 | * scheduler run of queue, if there are requests pending and no one in the | |
223 | * driver that will restart queueing | |
224 | */ | |
225 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) | |
226 | { | |
7b14e3b5 | 227 | if (cfqd->busy_queues) |
99f95e52 AM |
228 | kblockd_schedule_work(&cfqd->unplug_work); |
229 | } | |
230 | ||
231 | static int cfq_queue_empty(request_queue_t *q) | |
232 | { | |
233 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
234 | ||
b4878f24 | 235 | return !cfqd->busy_queues; |
99f95e52 AM |
236 | } |
237 | ||
7749a8d4 | 238 | static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync) |
206dc69b | 239 | { |
7749a8d4 JA |
240 | /* |
241 | * Use the per-process queue, for read requests and syncronous writes | |
242 | */ | |
243 | if (!(rw & REQ_RW) || is_sync) | |
206dc69b JA |
244 | return task->pid; |
245 | ||
246 | return CFQ_KEY_ASYNC; | |
247 | } | |
248 | ||
44f7c160 JA |
249 | /* |
250 | * Scale schedule slice based on io priority. Use the sync time slice only | |
251 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
252 | * io only, should not get full sync slice length. | |
253 | */ | |
d9e7620e JA |
254 | static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync, |
255 | unsigned short prio) | |
44f7c160 | 256 | { |
d9e7620e | 257 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 258 | |
d9e7620e JA |
259 | WARN_ON(prio >= IOPRIO_BE_NR); |
260 | ||
261 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
262 | } | |
44f7c160 | 263 | |
d9e7620e JA |
264 | static inline int |
265 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
266 | { | |
267 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
268 | } |
269 | ||
270 | static inline void | |
271 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
272 | { | |
273 | cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies; | |
274 | } | |
275 | ||
276 | /* | |
277 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
278 | * isn't valid until the first request from the dispatch is activated | |
279 | * and the slice time set. | |
280 | */ | |
281 | static inline int cfq_slice_used(struct cfq_queue *cfqq) | |
282 | { | |
283 | if (cfq_cfqq_slice_new(cfqq)) | |
284 | return 0; | |
285 | if (time_before(jiffies, cfqq->slice_end)) | |
286 | return 0; | |
287 | ||
288 | return 1; | |
289 | } | |
290 | ||
1da177e4 | 291 | /* |
5e705374 | 292 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 293 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 294 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 295 | */ |
5e705374 JA |
296 | static struct request * |
297 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2) | |
1da177e4 LT |
298 | { |
299 | sector_t last, s1, s2, d1 = 0, d2 = 0; | |
1da177e4 | 300 | unsigned long back_max; |
e8a99053 AM |
301 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
302 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
303 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 304 | |
5e705374 JA |
305 | if (rq1 == NULL || rq1 == rq2) |
306 | return rq2; | |
307 | if (rq2 == NULL) | |
308 | return rq1; | |
9c2c38a1 | 309 | |
5e705374 JA |
310 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
311 | return rq1; | |
312 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
313 | return rq2; | |
374f84ac JA |
314 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
315 | return rq1; | |
316 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
317 | return rq2; | |
1da177e4 | 318 | |
5e705374 JA |
319 | s1 = rq1->sector; |
320 | s2 = rq2->sector; | |
1da177e4 | 321 | |
6d048f53 | 322 | last = cfqd->last_position; |
1da177e4 | 323 | |
1da177e4 LT |
324 | /* |
325 | * by definition, 1KiB is 2 sectors | |
326 | */ | |
327 | back_max = cfqd->cfq_back_max * 2; | |
328 | ||
329 | /* | |
330 | * Strict one way elevator _except_ in the case where we allow | |
331 | * short backward seeks which are biased as twice the cost of a | |
332 | * similar forward seek. | |
333 | */ | |
334 | if (s1 >= last) | |
335 | d1 = s1 - last; | |
336 | else if (s1 + back_max >= last) | |
337 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
338 | else | |
e8a99053 | 339 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
340 | |
341 | if (s2 >= last) | |
342 | d2 = s2 - last; | |
343 | else if (s2 + back_max >= last) | |
344 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
345 | else | |
e8a99053 | 346 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
347 | |
348 | /* Found required data */ | |
e8a99053 AM |
349 | |
350 | /* | |
351 | * By doing switch() on the bit mask "wrap" we avoid having to | |
352 | * check two variables for all permutations: --> faster! | |
353 | */ | |
354 | switch (wrap) { | |
5e705374 | 355 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 356 | if (d1 < d2) |
5e705374 | 357 | return rq1; |
e8a99053 | 358 | else if (d2 < d1) |
5e705374 | 359 | return rq2; |
e8a99053 AM |
360 | else { |
361 | if (s1 >= s2) | |
5e705374 | 362 | return rq1; |
e8a99053 | 363 | else |
5e705374 | 364 | return rq2; |
e8a99053 | 365 | } |
1da177e4 | 366 | |
e8a99053 | 367 | case CFQ_RQ2_WRAP: |
5e705374 | 368 | return rq1; |
e8a99053 | 369 | case CFQ_RQ1_WRAP: |
5e705374 JA |
370 | return rq2; |
371 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
372 | default: |
373 | /* | |
374 | * Since both rqs are wrapped, | |
375 | * start with the one that's further behind head | |
376 | * (--> only *one* back seek required), | |
377 | * since back seek takes more time than forward. | |
378 | */ | |
379 | if (s1 <= s2) | |
5e705374 | 380 | return rq1; |
1da177e4 | 381 | else |
5e705374 | 382 | return rq2; |
1da177e4 LT |
383 | } |
384 | } | |
385 | ||
498d3aa2 JA |
386 | /* |
387 | * The below is leftmost cache rbtree addon | |
388 | */ | |
cc09e299 JA |
389 | static struct rb_node *cfq_rb_first(struct cfq_rb_root *root) |
390 | { | |
391 | if (!root->left) | |
392 | root->left = rb_first(&root->rb); | |
393 | ||
394 | return root->left; | |
395 | } | |
396 | ||
397 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) | |
398 | { | |
399 | if (root->left == n) | |
400 | root->left = NULL; | |
401 | ||
402 | rb_erase(n, &root->rb); | |
403 | RB_CLEAR_NODE(n); | |
404 | } | |
405 | ||
1da177e4 LT |
406 | /* |
407 | * would be nice to take fifo expire time into account as well | |
408 | */ | |
5e705374 JA |
409 | static struct request * |
410 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
411 | struct request *last) | |
1da177e4 | 412 | { |
21183b07 JA |
413 | struct rb_node *rbnext = rb_next(&last->rb_node); |
414 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 415 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 416 | |
21183b07 | 417 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
418 | |
419 | if (rbprev) | |
5e705374 | 420 | prev = rb_entry_rq(rbprev); |
1da177e4 | 421 | |
21183b07 | 422 | if (rbnext) |
5e705374 | 423 | next = rb_entry_rq(rbnext); |
21183b07 JA |
424 | else { |
425 | rbnext = rb_first(&cfqq->sort_list); | |
426 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 427 | next = rb_entry_rq(rbnext); |
21183b07 | 428 | } |
1da177e4 | 429 | |
21183b07 | 430 | return cfq_choose_req(cfqd, next, prev); |
1da177e4 LT |
431 | } |
432 | ||
d9e7620e JA |
433 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
434 | struct cfq_queue *cfqq) | |
1da177e4 | 435 | { |
d9e7620e JA |
436 | /* |
437 | * just an approximation, should be ok. | |
438 | */ | |
67e6b49e JA |
439 | return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
440 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); | |
d9e7620e JA |
441 | } |
442 | ||
498d3aa2 JA |
443 | /* |
444 | * The cfqd->service_tree holds all pending cfq_queue's that have | |
445 | * requests waiting to be processed. It is sorted in the order that | |
446 | * we will service the queues. | |
447 | */ | |
d9e7620e | 448 | static void cfq_service_tree_add(struct cfq_data *cfqd, |
edd75ffd | 449 | struct cfq_queue *cfqq, int add_front) |
d9e7620e | 450 | { |
cc09e299 | 451 | struct rb_node **p = &cfqd->service_tree.rb.rb_node; |
d9e7620e | 452 | struct rb_node *parent = NULL; |
d9e7620e | 453 | unsigned long rb_key; |
498d3aa2 | 454 | int left; |
d9e7620e | 455 | |
edd75ffd JA |
456 | if (!add_front) { |
457 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; | |
458 | rb_key += cfqq->slice_resid; | |
459 | cfqq->slice_resid = 0; | |
460 | } else | |
461 | rb_key = 0; | |
1da177e4 | 462 | |
d9e7620e | 463 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
99f9628a | 464 | /* |
d9e7620e | 465 | * same position, nothing more to do |
99f9628a | 466 | */ |
d9e7620e JA |
467 | if (rb_key == cfqq->rb_key) |
468 | return; | |
1da177e4 | 469 | |
cc09e299 | 470 | cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree); |
1da177e4 | 471 | } |
d9e7620e | 472 | |
498d3aa2 | 473 | left = 1; |
d9e7620e | 474 | while (*p) { |
cc09e299 | 475 | struct cfq_queue *__cfqq; |
67060e37 | 476 | struct rb_node **n; |
cc09e299 | 477 | |
d9e7620e JA |
478 | parent = *p; |
479 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
480 | ||
0c534e0a JA |
481 | /* |
482 | * sort RT queues first, we always want to give | |
67060e37 JA |
483 | * preference to them. IDLE queues goes to the back. |
484 | * after that, sort on the next service time. | |
0c534e0a JA |
485 | */ |
486 | if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq)) | |
67060e37 | 487 | n = &(*p)->rb_left; |
0c534e0a | 488 | else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq)) |
67060e37 JA |
489 | n = &(*p)->rb_right; |
490 | else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq)) | |
491 | n = &(*p)->rb_left; | |
492 | else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq)) | |
493 | n = &(*p)->rb_right; | |
0c534e0a | 494 | else if (rb_key < __cfqq->rb_key) |
67060e37 JA |
495 | n = &(*p)->rb_left; |
496 | else | |
497 | n = &(*p)->rb_right; | |
498 | ||
499 | if (n == &(*p)->rb_right) | |
cc09e299 | 500 | left = 0; |
67060e37 JA |
501 | |
502 | p = n; | |
d9e7620e JA |
503 | } |
504 | ||
cc09e299 JA |
505 | if (left) |
506 | cfqd->service_tree.left = &cfqq->rb_node; | |
507 | ||
d9e7620e JA |
508 | cfqq->rb_key = rb_key; |
509 | rb_link_node(&cfqq->rb_node, parent, p); | |
cc09e299 | 510 | rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb); |
1da177e4 LT |
511 | } |
512 | ||
498d3aa2 JA |
513 | /* |
514 | * Update cfqq's position in the service tree. | |
515 | */ | |
edd75ffd | 516 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 517 | { |
6d048f53 JA |
518 | /* |
519 | * Resorting requires the cfqq to be on the RR list already. | |
520 | */ | |
498d3aa2 | 521 | if (cfq_cfqq_on_rr(cfqq)) |
edd75ffd | 522 | cfq_service_tree_add(cfqd, cfqq, 0); |
6d048f53 JA |
523 | } |
524 | ||
1da177e4 LT |
525 | /* |
526 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 527 | * the pending list according to last request service |
1da177e4 LT |
528 | */ |
529 | static inline void | |
b4878f24 | 530 | cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 531 | { |
3b18152c JA |
532 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
533 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 LT |
534 | cfqd->busy_queues++; |
535 | ||
edd75ffd | 536 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
537 | } |
538 | ||
498d3aa2 JA |
539 | /* |
540 | * Called when the cfqq no longer has requests pending, remove it from | |
541 | * the service tree. | |
542 | */ | |
1da177e4 LT |
543 | static inline void |
544 | cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
545 | { | |
3b18152c JA |
546 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
547 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 548 | |
cc09e299 JA |
549 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) |
550 | cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree); | |
d9e7620e | 551 | |
1da177e4 LT |
552 | BUG_ON(!cfqd->busy_queues); |
553 | cfqd->busy_queues--; | |
554 | } | |
555 | ||
556 | /* | |
557 | * rb tree support functions | |
558 | */ | |
5e705374 | 559 | static inline void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 560 | { |
5e705374 | 561 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 562 | struct cfq_data *cfqd = cfqq->cfqd; |
5e705374 | 563 | const int sync = rq_is_sync(rq); |
1da177e4 | 564 | |
b4878f24 JA |
565 | BUG_ON(!cfqq->queued[sync]); |
566 | cfqq->queued[sync]--; | |
1da177e4 | 567 | |
5e705374 | 568 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 569 | |
dd67d051 | 570 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
b4878f24 | 571 | cfq_del_cfqq_rr(cfqd, cfqq); |
1da177e4 LT |
572 | } |
573 | ||
5e705374 | 574 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 575 | { |
5e705374 | 576 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 577 | struct cfq_data *cfqd = cfqq->cfqd; |
21183b07 | 578 | struct request *__alias; |
1da177e4 | 579 | |
5380a101 | 580 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
581 | |
582 | /* | |
583 | * looks a little odd, but the first insert might return an alias. | |
584 | * if that happens, put the alias on the dispatch list | |
585 | */ | |
21183b07 | 586 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 587 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
588 | |
589 | if (!cfq_cfqq_on_rr(cfqq)) | |
590 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
591 | |
592 | /* | |
593 | * check if this request is a better next-serve candidate | |
594 | */ | |
595 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq); | |
596 | BUG_ON(!cfqq->next_rq); | |
1da177e4 LT |
597 | } |
598 | ||
599 | static inline void | |
5e705374 | 600 | cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 601 | { |
5380a101 JA |
602 | elv_rb_del(&cfqq->sort_list, rq); |
603 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 604 | cfq_add_rq_rb(rq); |
1da177e4 LT |
605 | } |
606 | ||
206dc69b JA |
607 | static struct request * |
608 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 609 | { |
206dc69b | 610 | struct task_struct *tsk = current; |
7749a8d4 | 611 | pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio)); |
206dc69b | 612 | struct cfq_queue *cfqq; |
1da177e4 | 613 | |
206dc69b | 614 | cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio); |
89850f7e JA |
615 | if (cfqq) { |
616 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
617 | ||
21183b07 | 618 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 619 | } |
1da177e4 | 620 | |
1da177e4 LT |
621 | return NULL; |
622 | } | |
623 | ||
b4878f24 | 624 | static void cfq_activate_request(request_queue_t *q, struct request *rq) |
1da177e4 | 625 | { |
22e2c507 | 626 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 627 | |
b4878f24 | 628 | cfqd->rq_in_driver++; |
25776e35 JA |
629 | |
630 | /* | |
631 | * If the depth is larger 1, it really could be queueing. But lets | |
632 | * make the mark a little higher - idling could still be good for | |
633 | * low queueing, and a low queueing number could also just indicate | |
634 | * a SCSI mid layer like behaviour where limit+1 is often seen. | |
635 | */ | |
636 | if (!cfqd->hw_tag && cfqd->rq_in_driver > 4) | |
637 | cfqd->hw_tag = 1; | |
6d048f53 JA |
638 | |
639 | cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors; | |
1da177e4 LT |
640 | } |
641 | ||
b4878f24 | 642 | static void cfq_deactivate_request(request_queue_t *q, struct request *rq) |
1da177e4 | 643 | { |
b4878f24 JA |
644 | struct cfq_data *cfqd = q->elevator->elevator_data; |
645 | ||
646 | WARN_ON(!cfqd->rq_in_driver); | |
647 | cfqd->rq_in_driver--; | |
1da177e4 LT |
648 | } |
649 | ||
b4878f24 | 650 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 651 | { |
5e705374 | 652 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 653 | |
5e705374 JA |
654 | if (cfqq->next_rq == rq) |
655 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 656 | |
b4878f24 | 657 | list_del_init(&rq->queuelist); |
5e705374 | 658 | cfq_del_rq_rb(rq); |
374f84ac JA |
659 | |
660 | if (rq_is_meta(rq)) { | |
661 | WARN_ON(!cfqq->meta_pending); | |
662 | cfqq->meta_pending--; | |
663 | } | |
1da177e4 LT |
664 | } |
665 | ||
498d3aa2 | 666 | static int cfq_merge(request_queue_t *q, struct request **req, struct bio *bio) |
1da177e4 LT |
667 | { |
668 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
669 | struct request *__rq; | |
1da177e4 | 670 | |
206dc69b | 671 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 672 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
673 | *req = __rq; |
674 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
675 | } |
676 | ||
677 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
678 | } |
679 | ||
21183b07 JA |
680 | static void cfq_merged_request(request_queue_t *q, struct request *req, |
681 | int type) | |
1da177e4 | 682 | { |
21183b07 | 683 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 684 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 685 | |
5e705374 | 686 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 687 | } |
1da177e4 LT |
688 | } |
689 | ||
690 | static void | |
691 | cfq_merged_requests(request_queue_t *q, struct request *rq, | |
692 | struct request *next) | |
693 | { | |
22e2c507 JA |
694 | /* |
695 | * reposition in fifo if next is older than rq | |
696 | */ | |
697 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
698 | time_before(next->start_time, rq->start_time)) | |
699 | list_move(&rq->queuelist, &next->queuelist); | |
700 | ||
b4878f24 | 701 | cfq_remove_request(next); |
22e2c507 JA |
702 | } |
703 | ||
da775265 JA |
704 | static int cfq_allow_merge(request_queue_t *q, struct request *rq, |
705 | struct bio *bio) | |
706 | { | |
707 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
708 | const int rw = bio_data_dir(bio); | |
709 | struct cfq_queue *cfqq; | |
710 | pid_t key; | |
711 | ||
712 | /* | |
ec8acb69 | 713 | * Disallow merge of a sync bio into an async request. |
da775265 | 714 | */ |
ec8acb69 | 715 | if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq)) |
da775265 JA |
716 | return 0; |
717 | ||
718 | /* | |
719d3402 JA |
719 | * Lookup the cfqq that this bio will be queued with. Allow |
720 | * merge only if rq is queued there. | |
da775265 | 721 | */ |
719d3402 | 722 | key = cfq_queue_pid(current, rw, bio_sync(bio)); |
da775265 | 723 | cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio); |
719d3402 JA |
724 | |
725 | if (cfqq == RQ_CFQQ(rq)) | |
726 | return 1; | |
da775265 | 727 | |
ec8acb69 | 728 | return 0; |
da775265 JA |
729 | } |
730 | ||
22e2c507 JA |
731 | static inline void |
732 | __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
733 | { | |
734 | if (cfqq) { | |
735 | /* | |
736 | * stop potential idle class queues waiting service | |
737 | */ | |
738 | del_timer(&cfqd->idle_class_timer); | |
739 | ||
22e2c507 | 740 | cfqq->slice_end = 0; |
3b18152c JA |
741 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
742 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 743 | cfq_mark_cfqq_slice_new(cfqq); |
1afba045 | 744 | cfq_clear_cfqq_queue_new(cfqq); |
22e2c507 JA |
745 | } |
746 | ||
747 | cfqd->active_queue = cfqq; | |
748 | } | |
749 | ||
7b14e3b5 JA |
750 | /* |
751 | * current cfqq expired its slice (or was too idle), select new one | |
752 | */ | |
753 | static void | |
754 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
6084cdda | 755 | int timed_out) |
7b14e3b5 | 756 | { |
7b14e3b5 JA |
757 | if (cfq_cfqq_wait_request(cfqq)) |
758 | del_timer(&cfqd->idle_slice_timer); | |
759 | ||
7b14e3b5 JA |
760 | cfq_clear_cfqq_must_dispatch(cfqq); |
761 | cfq_clear_cfqq_wait_request(cfqq); | |
762 | ||
763 | /* | |
6084cdda | 764 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 765 | */ |
3c6bd2f8 | 766 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) |
c5b680f3 | 767 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b14e3b5 | 768 | |
edd75ffd | 769 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
770 | |
771 | if (cfqq == cfqd->active_queue) | |
772 | cfqd->active_queue = NULL; | |
773 | ||
774 | if (cfqd->active_cic) { | |
775 | put_io_context(cfqd->active_cic->ioc); | |
776 | cfqd->active_cic = NULL; | |
777 | } | |
7b14e3b5 JA |
778 | } |
779 | ||
6084cdda | 780 | static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out) |
7b14e3b5 JA |
781 | { |
782 | struct cfq_queue *cfqq = cfqd->active_queue; | |
783 | ||
784 | if (cfqq) | |
6084cdda | 785 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
786 | } |
787 | ||
498d3aa2 JA |
788 | /* |
789 | * Get next queue for service. Unless we have a queue preemption, | |
790 | * we'll simply select the first cfqq in the service tree. | |
791 | */ | |
6d048f53 | 792 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 793 | { |
edd75ffd JA |
794 | struct cfq_queue *cfqq; |
795 | struct rb_node *n; | |
22e2c507 | 796 | |
edd75ffd JA |
797 | if (RB_EMPTY_ROOT(&cfqd->service_tree.rb)) |
798 | return NULL; | |
d9e7620e | 799 | |
edd75ffd JA |
800 | n = cfq_rb_first(&cfqd->service_tree); |
801 | cfqq = rb_entry(n, struct cfq_queue, rb_node); | |
498d3aa2 | 802 | |
edd75ffd JA |
803 | if (cfq_class_idle(cfqq)) { |
804 | unsigned long end; | |
805 | ||
806 | /* | |
807 | * if we have idle queues and no rt or be queues had | |
808 | * pending requests, either allow immediate service if | |
809 | * the grace period has passed or arm the idle grace | |
810 | * timer | |
811 | */ | |
812 | end = cfqd->last_end_request + CFQ_IDLE_GRACE; | |
813 | if (time_before(jiffies, end)) { | |
814 | mod_timer(&cfqd->idle_class_timer, end); | |
815 | cfqq = NULL; | |
67060e37 | 816 | } |
22e2c507 JA |
817 | } |
818 | ||
6d048f53 JA |
819 | return cfqq; |
820 | } | |
821 | ||
498d3aa2 JA |
822 | /* |
823 | * Get and set a new active queue for service. | |
824 | */ | |
6d048f53 JA |
825 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd) |
826 | { | |
827 | struct cfq_queue *cfqq; | |
828 | ||
d9e7620e | 829 | cfqq = cfq_get_next_queue(cfqd); |
22e2c507 | 830 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 831 | return cfqq; |
22e2c507 JA |
832 | } |
833 | ||
d9e7620e JA |
834 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
835 | struct request *rq) | |
836 | { | |
837 | if (rq->sector >= cfqd->last_position) | |
838 | return rq->sector - cfqd->last_position; | |
839 | else | |
840 | return cfqd->last_position - rq->sector; | |
841 | } | |
842 | ||
6d048f53 JA |
843 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq) |
844 | { | |
845 | struct cfq_io_context *cic = cfqd->active_cic; | |
846 | ||
847 | if (!sample_valid(cic->seek_samples)) | |
848 | return 0; | |
849 | ||
850 | return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean; | |
851 | } | |
852 | ||
d9e7620e JA |
853 | static int cfq_close_cooperator(struct cfq_data *cfq_data, |
854 | struct cfq_queue *cfqq) | |
6d048f53 | 855 | { |
6d048f53 | 856 | /* |
d9e7620e JA |
857 | * We should notice if some of the queues are cooperating, eg |
858 | * working closely on the same area of the disk. In that case, | |
859 | * we can group them together and don't waste time idling. | |
6d048f53 | 860 | */ |
d9e7620e | 861 | return 0; |
6d048f53 JA |
862 | } |
863 | ||
864 | #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024)) | |
caaa5f9f | 865 | |
6d048f53 | 866 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 867 | { |
1792669c | 868 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 869 | struct cfq_io_context *cic; |
7b14e3b5 JA |
870 | unsigned long sl; |
871 | ||
dd67d051 | 872 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 873 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
874 | |
875 | /* | |
876 | * idle is disabled, either manually or by past process history | |
877 | */ | |
6d048f53 JA |
878 | if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq)) |
879 | return; | |
880 | ||
22e2c507 JA |
881 | /* |
882 | * task has exited, don't wait | |
883 | */ | |
206dc69b JA |
884 | cic = cfqd->active_cic; |
885 | if (!cic || !cic->ioc->task) | |
6d048f53 JA |
886 | return; |
887 | ||
888 | /* | |
889 | * See if this prio level has a good candidate | |
890 | */ | |
1afba045 JA |
891 | if (cfq_close_cooperator(cfqd, cfqq) && |
892 | (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2)) | |
6d048f53 | 893 | return; |
22e2c507 | 894 | |
3b18152c JA |
895 | cfq_mark_cfqq_must_dispatch(cfqq); |
896 | cfq_mark_cfqq_wait_request(cfqq); | |
22e2c507 | 897 | |
206dc69b JA |
898 | /* |
899 | * we don't want to idle for seeks, but we do want to allow | |
900 | * fair distribution of slice time for a process doing back-to-back | |
901 | * seeks. so allow a little bit of time for him to submit a new rq | |
902 | */ | |
6d048f53 | 903 | sl = cfqd->cfq_slice_idle; |
caaa5f9f | 904 | if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic)) |
d9e7620e | 905 | sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT)); |
206dc69b | 906 | |
7b14e3b5 | 907 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
1da177e4 LT |
908 | } |
909 | ||
498d3aa2 JA |
910 | /* |
911 | * Move request from internal lists to the request queue dispatch list. | |
912 | */ | |
5e705374 | 913 | static void cfq_dispatch_insert(request_queue_t *q, struct request *rq) |
1da177e4 | 914 | { |
3ed9a296 | 915 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 916 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 917 | |
5380a101 | 918 | cfq_remove_request(rq); |
6d048f53 | 919 | cfqq->dispatched++; |
5380a101 | 920 | elv_dispatch_sort(q, rq); |
3ed9a296 JA |
921 | |
922 | if (cfq_cfqq_sync(cfqq)) | |
923 | cfqd->sync_flight++; | |
1da177e4 LT |
924 | } |
925 | ||
926 | /* | |
927 | * return expired entry, or NULL to just start from scratch in rbtree | |
928 | */ | |
5e705374 | 929 | static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 LT |
930 | { |
931 | struct cfq_data *cfqd = cfqq->cfqd; | |
22e2c507 | 932 | struct request *rq; |
89850f7e | 933 | int fifo; |
1da177e4 | 934 | |
3b18152c | 935 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 936 | return NULL; |
cb887411 JA |
937 | |
938 | cfq_mark_cfqq_fifo_expire(cfqq); | |
939 | ||
89850f7e JA |
940 | if (list_empty(&cfqq->fifo)) |
941 | return NULL; | |
1da177e4 | 942 | |
6d048f53 | 943 | fifo = cfq_cfqq_sync(cfqq); |
89850f7e | 944 | rq = rq_entry_fifo(cfqq->fifo.next); |
1da177e4 | 945 | |
6d048f53 JA |
946 | if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) |
947 | return NULL; | |
1da177e4 | 948 | |
6d048f53 | 949 | return rq; |
1da177e4 LT |
950 | } |
951 | ||
22e2c507 JA |
952 | static inline int |
953 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
954 | { | |
955 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 956 | |
22e2c507 | 957 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 958 | |
22e2c507 | 959 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
960 | } |
961 | ||
22e2c507 | 962 | /* |
498d3aa2 JA |
963 | * Select a queue for service. If we have a current active queue, |
964 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 965 | */ |
1b5ed5e1 | 966 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 967 | { |
1da177e4 | 968 | struct cfq_queue *cfqq; |
1da177e4 | 969 | |
22e2c507 JA |
970 | cfqq = cfqd->active_queue; |
971 | if (!cfqq) | |
972 | goto new_queue; | |
1da177e4 | 973 | |
22e2c507 | 974 | /* |
6d048f53 | 975 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 976 | */ |
6d048f53 | 977 | if (cfq_slice_used(cfqq)) |
3b18152c | 978 | goto expire; |
1da177e4 | 979 | |
22e2c507 | 980 | /* |
6d048f53 JA |
981 | * The active queue has requests and isn't expired, allow it to |
982 | * dispatch. | |
22e2c507 | 983 | */ |
dd67d051 | 984 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 985 | goto keep_queue; |
6d048f53 JA |
986 | |
987 | /* | |
988 | * No requests pending. If the active queue still has requests in | |
989 | * flight or is idling for a new request, allow either of these | |
990 | * conditions to happen (or time out) before selecting a new queue. | |
991 | */ | |
cc197479 JA |
992 | if (timer_pending(&cfqd->idle_slice_timer) || |
993 | (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) { | |
caaa5f9f JA |
994 | cfqq = NULL; |
995 | goto keep_queue; | |
22e2c507 JA |
996 | } |
997 | ||
3b18152c | 998 | expire: |
6084cdda | 999 | cfq_slice_expired(cfqd, 0); |
3b18152c JA |
1000 | new_queue: |
1001 | cfqq = cfq_set_active_queue(cfqd); | |
22e2c507 | 1002 | keep_queue: |
3b18152c | 1003 | return cfqq; |
22e2c507 JA |
1004 | } |
1005 | ||
498d3aa2 JA |
1006 | /* |
1007 | * Dispatch some requests from cfqq, moving them to the request queue | |
1008 | * dispatch list. | |
1009 | */ | |
22e2c507 JA |
1010 | static int |
1011 | __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1012 | int max_dispatch) | |
1013 | { | |
1014 | int dispatched = 0; | |
1015 | ||
dd67d051 | 1016 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); |
22e2c507 JA |
1017 | |
1018 | do { | |
5e705374 | 1019 | struct request *rq; |
1da177e4 LT |
1020 | |
1021 | /* | |
22e2c507 | 1022 | * follow expired path, else get first next available |
1da177e4 | 1023 | */ |
5e705374 JA |
1024 | if ((rq = cfq_check_fifo(cfqq)) == NULL) |
1025 | rq = cfqq->next_rq; | |
22e2c507 JA |
1026 | |
1027 | /* | |
1028 | * finally, insert request into driver dispatch list | |
1029 | */ | |
5e705374 | 1030 | cfq_dispatch_insert(cfqd->queue, rq); |
1da177e4 | 1031 | |
22e2c507 | 1032 | dispatched++; |
1da177e4 | 1033 | |
22e2c507 | 1034 | if (!cfqd->active_cic) { |
5e705374 JA |
1035 | atomic_inc(&RQ_CIC(rq)->ioc->refcount); |
1036 | cfqd->active_cic = RQ_CIC(rq); | |
22e2c507 | 1037 | } |
1da177e4 | 1038 | |
dd67d051 | 1039 | if (RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 JA |
1040 | break; |
1041 | ||
1042 | } while (dispatched < max_dispatch); | |
1043 | ||
22e2c507 JA |
1044 | /* |
1045 | * expire an async queue immediately if it has used up its slice. idle | |
1046 | * queue always expire after 1 dispatch round. | |
1047 | */ | |
a9938006 | 1048 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && |
20e493a8 | 1049 | dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) || |
a9938006 | 1050 | cfq_class_idle(cfqq))) { |
44f7c160 | 1051 | cfqq->slice_end = jiffies + 1; |
6084cdda | 1052 | cfq_slice_expired(cfqd, 0); |
44f7c160 | 1053 | } |
22e2c507 JA |
1054 | |
1055 | return dispatched; | |
1056 | } | |
1057 | ||
d9e7620e JA |
1058 | static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
1059 | { | |
1060 | int dispatched = 0; | |
1061 | ||
1062 | while (cfqq->next_rq) { | |
1063 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
1064 | dispatched++; | |
1065 | } | |
1066 | ||
1067 | BUG_ON(!list_empty(&cfqq->fifo)); | |
1068 | return dispatched; | |
1069 | } | |
1070 | ||
498d3aa2 JA |
1071 | /* |
1072 | * Drain our current requests. Used for barriers and when switching | |
1073 | * io schedulers on-the-fly. | |
1074 | */ | |
d9e7620e | 1075 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 1076 | { |
d9e7620e JA |
1077 | int dispatched = 0; |
1078 | struct rb_node *n; | |
1079 | ||
cc09e299 | 1080 | while ((n = cfq_rb_first(&cfqd->service_tree)) != NULL) { |
d9e7620e | 1081 | struct cfq_queue *cfqq = rb_entry(n, struct cfq_queue, rb_node); |
1b5ed5e1 | 1082 | |
d9e7620e JA |
1083 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); |
1084 | } | |
1b5ed5e1 | 1085 | |
6084cdda | 1086 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
1087 | |
1088 | BUG_ON(cfqd->busy_queues); | |
1089 | ||
1090 | return dispatched; | |
1091 | } | |
1092 | ||
d9e7620e | 1093 | static int cfq_dispatch_requests(request_queue_t *q, int force) |
22e2c507 JA |
1094 | { |
1095 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
6d048f53 | 1096 | struct cfq_queue *cfqq; |
caaa5f9f | 1097 | int dispatched; |
22e2c507 JA |
1098 | |
1099 | if (!cfqd->busy_queues) | |
1100 | return 0; | |
1101 | ||
1b5ed5e1 TH |
1102 | if (unlikely(force)) |
1103 | return cfq_forced_dispatch(cfqd); | |
1104 | ||
caaa5f9f | 1105 | dispatched = 0; |
caaa5f9f | 1106 | while ((cfqq = cfq_select_queue(cfqd)) != NULL) { |
b4878f24 JA |
1107 | int max_dispatch; |
1108 | ||
3ed9a296 JA |
1109 | max_dispatch = cfqd->cfq_quantum; |
1110 | if (cfq_class_idle(cfqq)) | |
1111 | max_dispatch = 1; | |
1112 | ||
1113 | if (cfqq->dispatched >= max_dispatch) { | |
1114 | if (cfqd->busy_queues > 1) | |
6d048f53 | 1115 | break; |
3ed9a296 | 1116 | if (cfqq->dispatched >= 4 * max_dispatch) |
a9938006 JA |
1117 | break; |
1118 | } | |
9ede209e | 1119 | |
3ed9a296 JA |
1120 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) |
1121 | break; | |
1122 | ||
3b18152c JA |
1123 | cfq_clear_cfqq_must_dispatch(cfqq); |
1124 | cfq_clear_cfqq_wait_request(cfqq); | |
22e2c507 JA |
1125 | del_timer(&cfqd->idle_slice_timer); |
1126 | ||
caaa5f9f | 1127 | dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch); |
1da177e4 LT |
1128 | } |
1129 | ||
caaa5f9f | 1130 | return dispatched; |
1da177e4 LT |
1131 | } |
1132 | ||
1da177e4 | 1133 | /* |
5e705374 JA |
1134 | * task holds one reference to the queue, dropped when task exits. each rq |
1135 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 LT |
1136 | * |
1137 | * queue lock must be held here. | |
1138 | */ | |
1139 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
1140 | { | |
22e2c507 JA |
1141 | struct cfq_data *cfqd = cfqq->cfqd; |
1142 | ||
1143 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
1144 | |
1145 | if (!atomic_dec_and_test(&cfqq->ref)) | |
1146 | return; | |
1147 | ||
1148 | BUG_ON(rb_first(&cfqq->sort_list)); | |
22e2c507 | 1149 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
3b18152c | 1150 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 | 1151 | |
28f95cbc | 1152 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 1153 | __cfq_slice_expired(cfqd, cfqq, 0); |
28f95cbc JA |
1154 | cfq_schedule_dispatch(cfqd); |
1155 | } | |
22e2c507 | 1156 | |
1da177e4 LT |
1157 | /* |
1158 | * it's on the empty list and still hashed | |
1159 | */ | |
1da177e4 LT |
1160 | hlist_del(&cfqq->cfq_hash); |
1161 | kmem_cache_free(cfq_pool, cfqq); | |
1162 | } | |
1163 | ||
1ea25ecb | 1164 | static struct cfq_queue * |
3b18152c JA |
1165 | __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio, |
1166 | const int hashval) | |
1da177e4 LT |
1167 | { |
1168 | struct hlist_head *hash_list = &cfqd->cfq_hash[hashval]; | |
206dc69b JA |
1169 | struct hlist_node *entry; |
1170 | struct cfq_queue *__cfqq; | |
1da177e4 | 1171 | |
206dc69b | 1172 | hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) { |
b0a6916b | 1173 | const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio); |
1da177e4 | 1174 | |
206dc69b | 1175 | if (__cfqq->key == key && (__p == prio || !prio)) |
1da177e4 LT |
1176 | return __cfqq; |
1177 | } | |
1178 | ||
1179 | return NULL; | |
1180 | } | |
1181 | ||
1182 | static struct cfq_queue * | |
3b18152c | 1183 | cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio) |
1da177e4 | 1184 | { |
3b18152c | 1185 | return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT)); |
1da177e4 LT |
1186 | } |
1187 | ||
e2d74ac0 | 1188 | static void cfq_free_io_context(struct io_context *ioc) |
1da177e4 | 1189 | { |
22e2c507 | 1190 | struct cfq_io_context *__cic; |
e2d74ac0 JA |
1191 | struct rb_node *n; |
1192 | int freed = 0; | |
1da177e4 | 1193 | |
e2d74ac0 JA |
1194 | while ((n = rb_first(&ioc->cic_root)) != NULL) { |
1195 | __cic = rb_entry(n, struct cfq_io_context, rb_node); | |
1196 | rb_erase(&__cic->rb_node, &ioc->cic_root); | |
22e2c507 | 1197 | kmem_cache_free(cfq_ioc_pool, __cic); |
334e94de | 1198 | freed++; |
1da177e4 LT |
1199 | } |
1200 | ||
4050cf16 JA |
1201 | elv_ioc_count_mod(ioc_count, -freed); |
1202 | ||
1203 | if (ioc_gone && !elv_ioc_count_read(ioc_count)) | |
334e94de | 1204 | complete(ioc_gone); |
1da177e4 LT |
1205 | } |
1206 | ||
89850f7e | 1207 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1208 | { |
28f95cbc | 1209 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 1210 | __cfq_slice_expired(cfqd, cfqq, 0); |
28f95cbc JA |
1211 | cfq_schedule_dispatch(cfqd); |
1212 | } | |
22e2c507 | 1213 | |
89850f7e JA |
1214 | cfq_put_queue(cfqq); |
1215 | } | |
22e2c507 | 1216 | |
89850f7e JA |
1217 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
1218 | struct cfq_io_context *cic) | |
1219 | { | |
fc46379d JA |
1220 | list_del_init(&cic->queue_list); |
1221 | smp_wmb(); | |
1222 | cic->key = NULL; | |
1223 | ||
12a05732 | 1224 | if (cic->cfqq[ASYNC]) { |
89850f7e | 1225 | cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]); |
12a05732 AV |
1226 | cic->cfqq[ASYNC] = NULL; |
1227 | } | |
1228 | ||
1229 | if (cic->cfqq[SYNC]) { | |
89850f7e | 1230 | cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]); |
12a05732 AV |
1231 | cic->cfqq[SYNC] = NULL; |
1232 | } | |
89850f7e JA |
1233 | } |
1234 | ||
89850f7e JA |
1235 | static void cfq_exit_single_io_context(struct cfq_io_context *cic) |
1236 | { | |
1237 | struct cfq_data *cfqd = cic->key; | |
1238 | ||
89850f7e JA |
1239 | if (cfqd) { |
1240 | request_queue_t *q = cfqd->queue; | |
1241 | ||
fc46379d | 1242 | spin_lock_irq(q->queue_lock); |
89850f7e | 1243 | __cfq_exit_single_io_context(cfqd, cic); |
fc46379d | 1244 | spin_unlock_irq(q->queue_lock); |
89850f7e | 1245 | } |
1da177e4 LT |
1246 | } |
1247 | ||
498d3aa2 JA |
1248 | /* |
1249 | * The process that ioc belongs to has exited, we need to clean up | |
1250 | * and put the internal structures we have that belongs to that process. | |
1251 | */ | |
e2d74ac0 | 1252 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 1253 | { |
22e2c507 | 1254 | struct cfq_io_context *__cic; |
e2d74ac0 | 1255 | struct rb_node *n; |
22e2c507 | 1256 | |
1da177e4 LT |
1257 | /* |
1258 | * put the reference this task is holding to the various queues | |
1259 | */ | |
e2d74ac0 JA |
1260 | |
1261 | n = rb_first(&ioc->cic_root); | |
1262 | while (n != NULL) { | |
1263 | __cic = rb_entry(n, struct cfq_io_context, rb_node); | |
1264 | ||
22e2c507 | 1265 | cfq_exit_single_io_context(__cic); |
e2d74ac0 | 1266 | n = rb_next(n); |
1da177e4 | 1267 | } |
1da177e4 LT |
1268 | } |
1269 | ||
22e2c507 | 1270 | static struct cfq_io_context * |
8267e268 | 1271 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1272 | { |
b5deef90 | 1273 | struct cfq_io_context *cic; |
1da177e4 | 1274 | |
b5deef90 | 1275 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node); |
1da177e4 | 1276 | if (cic) { |
553698f9 | 1277 | memset(cic, 0, sizeof(*cic)); |
22e2c507 | 1278 | cic->last_end_request = jiffies; |
553698f9 | 1279 | INIT_LIST_HEAD(&cic->queue_list); |
22e2c507 JA |
1280 | cic->dtor = cfq_free_io_context; |
1281 | cic->exit = cfq_exit_io_context; | |
4050cf16 | 1282 | elv_ioc_count_inc(ioc_count); |
1da177e4 LT |
1283 | } |
1284 | ||
1285 | return cic; | |
1286 | } | |
1287 | ||
22e2c507 JA |
1288 | static void cfq_init_prio_data(struct cfq_queue *cfqq) |
1289 | { | |
1290 | struct task_struct *tsk = current; | |
1291 | int ioprio_class; | |
1292 | ||
3b18152c | 1293 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
1294 | return; |
1295 | ||
1296 | ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio); | |
1297 | switch (ioprio_class) { | |
1298 | default: | |
1299 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
1300 | case IOPRIO_CLASS_NONE: | |
1301 | /* | |
1302 | * no prio set, place us in the middle of the BE classes | |
1303 | */ | |
1304 | cfqq->ioprio = task_nice_ioprio(tsk); | |
1305 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
1306 | break; | |
1307 | case IOPRIO_CLASS_RT: | |
1308 | cfqq->ioprio = task_ioprio(tsk); | |
1309 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
1310 | break; | |
1311 | case IOPRIO_CLASS_BE: | |
1312 | cfqq->ioprio = task_ioprio(tsk); | |
1313 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
1314 | break; | |
1315 | case IOPRIO_CLASS_IDLE: | |
1316 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
1317 | cfqq->ioprio = 7; | |
3b18152c | 1318 | cfq_clear_cfqq_idle_window(cfqq); |
22e2c507 JA |
1319 | break; |
1320 | } | |
1321 | ||
1322 | /* | |
1323 | * keep track of original prio settings in case we have to temporarily | |
1324 | * elevate the priority of this queue | |
1325 | */ | |
1326 | cfqq->org_ioprio = cfqq->ioprio; | |
1327 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 1328 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
1329 | } |
1330 | ||
478a82b0 | 1331 | static inline void changed_ioprio(struct cfq_io_context *cic) |
22e2c507 | 1332 | { |
478a82b0 AV |
1333 | struct cfq_data *cfqd = cic->key; |
1334 | struct cfq_queue *cfqq; | |
c1b707d2 | 1335 | unsigned long flags; |
35e6077c | 1336 | |
caaa5f9f JA |
1337 | if (unlikely(!cfqd)) |
1338 | return; | |
1339 | ||
c1b707d2 | 1340 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f JA |
1341 | |
1342 | cfqq = cic->cfqq[ASYNC]; | |
1343 | if (cfqq) { | |
1344 | struct cfq_queue *new_cfqq; | |
1345 | new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task, | |
1346 | GFP_ATOMIC); | |
1347 | if (new_cfqq) { | |
1348 | cic->cfqq[ASYNC] = new_cfqq; | |
1349 | cfq_put_queue(cfqq); | |
1350 | } | |
22e2c507 | 1351 | } |
caaa5f9f JA |
1352 | |
1353 | cfqq = cic->cfqq[SYNC]; | |
1354 | if (cfqq) | |
1355 | cfq_mark_cfqq_prio_changed(cfqq); | |
1356 | ||
c1b707d2 | 1357 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
1358 | } |
1359 | ||
fc46379d | 1360 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 1361 | { |
a6a0763a | 1362 | struct cfq_io_context *cic; |
e2d74ac0 | 1363 | struct rb_node *n; |
a6a0763a | 1364 | |
fc46379d | 1365 | ioc->ioprio_changed = 0; |
a6a0763a | 1366 | |
e2d74ac0 JA |
1367 | n = rb_first(&ioc->cic_root); |
1368 | while (n != NULL) { | |
1369 | cic = rb_entry(n, struct cfq_io_context, rb_node); | |
3793c65c | 1370 | |
478a82b0 | 1371 | changed_ioprio(cic); |
e2d74ac0 JA |
1372 | n = rb_next(n); |
1373 | } | |
22e2c507 JA |
1374 | } |
1375 | ||
1376 | static struct cfq_queue * | |
6f325a13 | 1377 | cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, |
8267e268 | 1378 | gfp_t gfp_mask) |
22e2c507 JA |
1379 | { |
1380 | const int hashval = hash_long(key, CFQ_QHASH_SHIFT); | |
1381 | struct cfq_queue *cfqq, *new_cfqq = NULL; | |
6f325a13 | 1382 | unsigned short ioprio; |
22e2c507 JA |
1383 | |
1384 | retry: | |
6f325a13 | 1385 | ioprio = tsk->ioprio; |
3b18152c | 1386 | cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval); |
22e2c507 JA |
1387 | |
1388 | if (!cfqq) { | |
1389 | if (new_cfqq) { | |
1390 | cfqq = new_cfqq; | |
1391 | new_cfqq = NULL; | |
1392 | } else if (gfp_mask & __GFP_WAIT) { | |
89850f7e JA |
1393 | /* |
1394 | * Inform the allocator of the fact that we will | |
1395 | * just repeat this allocation if it fails, to allow | |
1396 | * the allocator to do whatever it needs to attempt to | |
1397 | * free memory. | |
1398 | */ | |
22e2c507 | 1399 | spin_unlock_irq(cfqd->queue->queue_lock); |
b5deef90 | 1400 | new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node); |
22e2c507 JA |
1401 | spin_lock_irq(cfqd->queue->queue_lock); |
1402 | goto retry; | |
1403 | } else { | |
b5deef90 | 1404 | cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node); |
22e2c507 JA |
1405 | if (!cfqq) |
1406 | goto out; | |
1407 | } | |
1408 | ||
1409 | memset(cfqq, 0, sizeof(*cfqq)); | |
1410 | ||
1411 | INIT_HLIST_NODE(&cfqq->cfq_hash); | |
d9e7620e | 1412 | RB_CLEAR_NODE(&cfqq->rb_node); |
22e2c507 JA |
1413 | INIT_LIST_HEAD(&cfqq->fifo); |
1414 | ||
1415 | cfqq->key = key; | |
1416 | hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]); | |
1417 | atomic_set(&cfqq->ref, 0); | |
1418 | cfqq->cfqd = cfqd; | |
c5b680f3 | 1419 | |
a9938006 JA |
1420 | if (key != CFQ_KEY_ASYNC) |
1421 | cfq_mark_cfqq_idle_window(cfqq); | |
1422 | ||
3b18152c | 1423 | cfq_mark_cfqq_prio_changed(cfqq); |
53b03744 | 1424 | cfq_mark_cfqq_queue_new(cfqq); |
3b18152c | 1425 | cfq_init_prio_data(cfqq); |
22e2c507 JA |
1426 | } |
1427 | ||
1428 | if (new_cfqq) | |
1429 | kmem_cache_free(cfq_pool, new_cfqq); | |
1430 | ||
1431 | atomic_inc(&cfqq->ref); | |
1432 | out: | |
1433 | WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq); | |
1434 | return cfqq; | |
1435 | } | |
1436 | ||
498d3aa2 JA |
1437 | /* |
1438 | * We drop cfq io contexts lazily, so we may find a dead one. | |
1439 | */ | |
dbecf3ab OH |
1440 | static void |
1441 | cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic) | |
1442 | { | |
fc46379d | 1443 | WARN_ON(!list_empty(&cic->queue_list)); |
dbecf3ab | 1444 | rb_erase(&cic->rb_node, &ioc->cic_root); |
dbecf3ab | 1445 | kmem_cache_free(cfq_ioc_pool, cic); |
4050cf16 | 1446 | elv_ioc_count_dec(ioc_count); |
dbecf3ab OH |
1447 | } |
1448 | ||
e2d74ac0 JA |
1449 | static struct cfq_io_context * |
1450 | cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc) | |
1451 | { | |
dbecf3ab | 1452 | struct rb_node *n; |
e2d74ac0 | 1453 | struct cfq_io_context *cic; |
be3b0753 | 1454 | void *k, *key = cfqd; |
e2d74ac0 | 1455 | |
dbecf3ab OH |
1456 | restart: |
1457 | n = ioc->cic_root.rb_node; | |
e2d74ac0 JA |
1458 | while (n) { |
1459 | cic = rb_entry(n, struct cfq_io_context, rb_node); | |
be3b0753 OH |
1460 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
1461 | k = cic->key; | |
1462 | if (unlikely(!k)) { | |
dbecf3ab OH |
1463 | cfq_drop_dead_cic(ioc, cic); |
1464 | goto restart; | |
1465 | } | |
e2d74ac0 | 1466 | |
be3b0753 | 1467 | if (key < k) |
e2d74ac0 | 1468 | n = n->rb_left; |
be3b0753 | 1469 | else if (key > k) |
e2d74ac0 JA |
1470 | n = n->rb_right; |
1471 | else | |
1472 | return cic; | |
1473 | } | |
1474 | ||
1475 | return NULL; | |
1476 | } | |
1477 | ||
1478 | static inline void | |
1479 | cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, | |
1480 | struct cfq_io_context *cic) | |
1481 | { | |
dbecf3ab OH |
1482 | struct rb_node **p; |
1483 | struct rb_node *parent; | |
e2d74ac0 | 1484 | struct cfq_io_context *__cic; |
0261d688 | 1485 | unsigned long flags; |
be3b0753 | 1486 | void *k; |
e2d74ac0 | 1487 | |
e2d74ac0 JA |
1488 | cic->ioc = ioc; |
1489 | cic->key = cfqd; | |
1490 | ||
dbecf3ab OH |
1491 | restart: |
1492 | parent = NULL; | |
1493 | p = &ioc->cic_root.rb_node; | |
e2d74ac0 JA |
1494 | while (*p) { |
1495 | parent = *p; | |
1496 | __cic = rb_entry(parent, struct cfq_io_context, rb_node); | |
be3b0753 OH |
1497 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
1498 | k = __cic->key; | |
1499 | if (unlikely(!k)) { | |
be33c3a6 | 1500 | cfq_drop_dead_cic(ioc, __cic); |
dbecf3ab OH |
1501 | goto restart; |
1502 | } | |
e2d74ac0 | 1503 | |
be3b0753 | 1504 | if (cic->key < k) |
e2d74ac0 | 1505 | p = &(*p)->rb_left; |
be3b0753 | 1506 | else if (cic->key > k) |
e2d74ac0 JA |
1507 | p = &(*p)->rb_right; |
1508 | else | |
1509 | BUG(); | |
1510 | } | |
1511 | ||
1512 | rb_link_node(&cic->rb_node, parent, p); | |
1513 | rb_insert_color(&cic->rb_node, &ioc->cic_root); | |
fc46379d | 1514 | |
0261d688 | 1515 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
e2d74ac0 | 1516 | list_add(&cic->queue_list, &cfqd->cic_list); |
0261d688 | 1517 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
e2d74ac0 JA |
1518 | } |
1519 | ||
1da177e4 LT |
1520 | /* |
1521 | * Setup general io context and cfq io context. There can be several cfq | |
1522 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 1523 | * than one device managed by cfq. |
1da177e4 LT |
1524 | */ |
1525 | static struct cfq_io_context * | |
e2d74ac0 | 1526 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1527 | { |
22e2c507 | 1528 | struct io_context *ioc = NULL; |
1da177e4 | 1529 | struct cfq_io_context *cic; |
1da177e4 | 1530 | |
22e2c507 | 1531 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 1532 | |
b5deef90 | 1533 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
1534 | if (!ioc) |
1535 | return NULL; | |
1536 | ||
e2d74ac0 JA |
1537 | cic = cfq_cic_rb_lookup(cfqd, ioc); |
1538 | if (cic) | |
1539 | goto out; | |
1da177e4 | 1540 | |
e2d74ac0 JA |
1541 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
1542 | if (cic == NULL) | |
1543 | goto err; | |
1da177e4 | 1544 | |
e2d74ac0 | 1545 | cfq_cic_link(cfqd, ioc, cic); |
1da177e4 | 1546 | out: |
fc46379d JA |
1547 | smp_read_barrier_depends(); |
1548 | if (unlikely(ioc->ioprio_changed)) | |
1549 | cfq_ioc_set_ioprio(ioc); | |
1550 | ||
1da177e4 LT |
1551 | return cic; |
1552 | err: | |
1553 | put_io_context(ioc); | |
1554 | return NULL; | |
1555 | } | |
1556 | ||
22e2c507 JA |
1557 | static void |
1558 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 1559 | { |
aaf1228d JA |
1560 | unsigned long elapsed = jiffies - cic->last_end_request; |
1561 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 1562 | |
22e2c507 JA |
1563 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
1564 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
1565 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
1566 | } | |
1da177e4 | 1567 | |
206dc69b | 1568 | static void |
6d048f53 JA |
1569 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic, |
1570 | struct request *rq) | |
206dc69b JA |
1571 | { |
1572 | sector_t sdist; | |
1573 | u64 total; | |
1574 | ||
5e705374 JA |
1575 | if (cic->last_request_pos < rq->sector) |
1576 | sdist = rq->sector - cic->last_request_pos; | |
206dc69b | 1577 | else |
5e705374 | 1578 | sdist = cic->last_request_pos - rq->sector; |
206dc69b | 1579 | |
6d048f53 JA |
1580 | if (!cic->seek_samples) { |
1581 | cfqd->new_seek_total = (7*cic->seek_total + (u64)256*sdist) / 8; | |
1582 | cfqd->new_seek_mean = cfqd->new_seek_total / 256; | |
1583 | } | |
1584 | ||
206dc69b JA |
1585 | /* |
1586 | * Don't allow the seek distance to get too large from the | |
1587 | * odd fragment, pagein, etc | |
1588 | */ | |
1589 | if (cic->seek_samples <= 60) /* second&third seek */ | |
1590 | sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024); | |
1591 | else | |
1592 | sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64); | |
1593 | ||
1594 | cic->seek_samples = (7*cic->seek_samples + 256) / 8; | |
1595 | cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8; | |
1596 | total = cic->seek_total + (cic->seek_samples/2); | |
1597 | do_div(total, cic->seek_samples); | |
1598 | cic->seek_mean = (sector_t)total; | |
1599 | } | |
1da177e4 | 1600 | |
22e2c507 JA |
1601 | /* |
1602 | * Disable idle window if the process thinks too long or seeks so much that | |
1603 | * it doesn't matter | |
1604 | */ | |
1605 | static void | |
1606 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1607 | struct cfq_io_context *cic) | |
1608 | { | |
1be92f2f JA |
1609 | int enable_idle; |
1610 | ||
1611 | if (!cfq_cfqq_sync(cfqq)) | |
1612 | return; | |
1613 | ||
1614 | enable_idle = cfq_cfqq_idle_window(cfqq); | |
1da177e4 | 1615 | |
caaa5f9f JA |
1616 | if (!cic->ioc->task || !cfqd->cfq_slice_idle || |
1617 | (cfqd->hw_tag && CIC_SEEKY(cic))) | |
22e2c507 JA |
1618 | enable_idle = 0; |
1619 | else if (sample_valid(cic->ttime_samples)) { | |
1620 | if (cic->ttime_mean > cfqd->cfq_slice_idle) | |
1621 | enable_idle = 0; | |
1622 | else | |
1623 | enable_idle = 1; | |
1da177e4 LT |
1624 | } |
1625 | ||
3b18152c JA |
1626 | if (enable_idle) |
1627 | cfq_mark_cfqq_idle_window(cfqq); | |
1628 | else | |
1629 | cfq_clear_cfqq_idle_window(cfqq); | |
22e2c507 | 1630 | } |
1da177e4 | 1631 | |
22e2c507 JA |
1632 | /* |
1633 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
1634 | * no or if we aren't sure, a 1 will cause a preempt. | |
1635 | */ | |
1636 | static int | |
1637 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, | |
5e705374 | 1638 | struct request *rq) |
22e2c507 | 1639 | { |
6d048f53 | 1640 | struct cfq_queue *cfqq; |
22e2c507 | 1641 | |
6d048f53 JA |
1642 | cfqq = cfqd->active_queue; |
1643 | if (!cfqq) | |
22e2c507 JA |
1644 | return 0; |
1645 | ||
6d048f53 JA |
1646 | if (cfq_slice_used(cfqq)) |
1647 | return 1; | |
1648 | ||
1649 | if (cfq_class_idle(new_cfqq)) | |
caaa5f9f | 1650 | return 0; |
22e2c507 JA |
1651 | |
1652 | if (cfq_class_idle(cfqq)) | |
1653 | return 1; | |
1e3335de | 1654 | |
374f84ac JA |
1655 | /* |
1656 | * if the new request is sync, but the currently running queue is | |
1657 | * not, let the sync request have priority. | |
1658 | */ | |
5e705374 | 1659 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
22e2c507 | 1660 | return 1; |
1e3335de | 1661 | |
374f84ac JA |
1662 | /* |
1663 | * So both queues are sync. Let the new request get disk time if | |
1664 | * it's a metadata request and the current queue is doing regular IO. | |
1665 | */ | |
1666 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
1667 | return 1; | |
22e2c507 | 1668 | |
1e3335de JA |
1669 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
1670 | return 0; | |
1671 | ||
1672 | /* | |
1673 | * if this request is as-good as one we would expect from the | |
1674 | * current cfqq, let it preempt | |
1675 | */ | |
6d048f53 | 1676 | if (cfq_rq_close(cfqd, rq)) |
1e3335de JA |
1677 | return 1; |
1678 | ||
22e2c507 JA |
1679 | return 0; |
1680 | } | |
1681 | ||
1682 | /* | |
1683 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
1684 | * let it have half of its nominal slice. | |
1685 | */ | |
1686 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1687 | { | |
6084cdda | 1688 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 1689 | |
bf572256 JA |
1690 | /* |
1691 | * Put the new queue at the front of the of the current list, | |
1692 | * so we know that it will be selected next. | |
1693 | */ | |
1694 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
1695 | |
1696 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 1697 | |
44f7c160 JA |
1698 | cfqq->slice_end = 0; |
1699 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
1700 | } |
1701 | ||
22e2c507 | 1702 | /* |
5e705374 | 1703 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
1704 | * something we should do about it |
1705 | */ | |
1706 | static void | |
5e705374 JA |
1707 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
1708 | struct request *rq) | |
22e2c507 | 1709 | { |
5e705374 | 1710 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 1711 | |
374f84ac JA |
1712 | if (rq_is_meta(rq)) |
1713 | cfqq->meta_pending++; | |
1714 | ||
9c2c38a1 | 1715 | cfq_update_io_thinktime(cfqd, cic); |
6d048f53 | 1716 | cfq_update_io_seektime(cfqd, cic, rq); |
9c2c38a1 JA |
1717 | cfq_update_idle_window(cfqd, cfqq, cic); |
1718 | ||
5e705374 | 1719 | cic->last_request_pos = rq->sector + rq->nr_sectors; |
6d048f53 | 1720 | cfqq->last_request_pos = cic->last_request_pos; |
22e2c507 JA |
1721 | |
1722 | if (cfqq == cfqd->active_queue) { | |
1723 | /* | |
1724 | * if we are waiting for a request for this queue, let it rip | |
1725 | * immediately and flag that we must not expire this queue | |
1726 | * just now | |
1727 | */ | |
3b18152c JA |
1728 | if (cfq_cfqq_wait_request(cfqq)) { |
1729 | cfq_mark_cfqq_must_dispatch(cfqq); | |
22e2c507 | 1730 | del_timer(&cfqd->idle_slice_timer); |
dc72ef4a | 1731 | blk_start_queueing(cfqd->queue); |
22e2c507 | 1732 | } |
5e705374 | 1733 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
1734 | /* |
1735 | * not the active queue - expire current slice if it is | |
1736 | * idle and has expired it's mean thinktime or this new queue | |
1737 | * has some old slice time left and is of higher priority | |
1738 | */ | |
1739 | cfq_preempt_queue(cfqd, cfqq); | |
3b18152c | 1740 | cfq_mark_cfqq_must_dispatch(cfqq); |
dc72ef4a | 1741 | blk_start_queueing(cfqd->queue); |
22e2c507 | 1742 | } |
1da177e4 LT |
1743 | } |
1744 | ||
b4878f24 | 1745 | static void cfq_insert_request(request_queue_t *q, struct request *rq) |
1da177e4 | 1746 | { |
b4878f24 | 1747 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1748 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 JA |
1749 | |
1750 | cfq_init_prio_data(cfqq); | |
1da177e4 | 1751 | |
5e705374 | 1752 | cfq_add_rq_rb(rq); |
1da177e4 | 1753 | |
22e2c507 JA |
1754 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
1755 | ||
5e705374 | 1756 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
1757 | } |
1758 | ||
1da177e4 LT |
1759 | static void cfq_completed_request(request_queue_t *q, struct request *rq) |
1760 | { | |
5e705374 | 1761 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 1762 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 1763 | const int sync = rq_is_sync(rq); |
b4878f24 | 1764 | unsigned long now; |
1da177e4 | 1765 | |
b4878f24 | 1766 | now = jiffies; |
1da177e4 | 1767 | |
b4878f24 | 1768 | WARN_ON(!cfqd->rq_in_driver); |
6d048f53 | 1769 | WARN_ON(!cfqq->dispatched); |
b4878f24 | 1770 | cfqd->rq_in_driver--; |
6d048f53 | 1771 | cfqq->dispatched--; |
1da177e4 | 1772 | |
3ed9a296 JA |
1773 | if (cfq_cfqq_sync(cfqq)) |
1774 | cfqd->sync_flight--; | |
1775 | ||
b4878f24 JA |
1776 | if (!cfq_class_idle(cfqq)) |
1777 | cfqd->last_end_request = now; | |
3b18152c | 1778 | |
caaa5f9f | 1779 | if (sync) |
5e705374 | 1780 | RQ_CIC(rq)->last_end_request = now; |
caaa5f9f JA |
1781 | |
1782 | /* | |
1783 | * If this is the active queue, check if it needs to be expired, | |
1784 | * or if we want to idle in case it has no pending requests. | |
1785 | */ | |
1786 | if (cfqd->active_queue == cfqq) { | |
44f7c160 JA |
1787 | if (cfq_cfqq_slice_new(cfqq)) { |
1788 | cfq_set_prio_slice(cfqd, cfqq); | |
1789 | cfq_clear_cfqq_slice_new(cfqq); | |
1790 | } | |
1791 | if (cfq_slice_used(cfqq)) | |
6084cdda | 1792 | cfq_slice_expired(cfqd, 1); |
6d048f53 JA |
1793 | else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) |
1794 | cfq_arm_slice_timer(cfqd); | |
caaa5f9f | 1795 | } |
6d048f53 JA |
1796 | |
1797 | if (!cfqd->rq_in_driver) | |
1798 | cfq_schedule_dispatch(cfqd); | |
1da177e4 LT |
1799 | } |
1800 | ||
22e2c507 JA |
1801 | /* |
1802 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
1803 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
1804 | */ | |
1805 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 1806 | { |
22e2c507 JA |
1807 | if (has_fs_excl()) { |
1808 | /* | |
1809 | * boost idle prio on transactions that would lock out other | |
1810 | * users of the filesystem | |
1811 | */ | |
1812 | if (cfq_class_idle(cfqq)) | |
1813 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
1814 | if (cfqq->ioprio > IOPRIO_NORM) | |
1815 | cfqq->ioprio = IOPRIO_NORM; | |
1816 | } else { | |
1817 | /* | |
1818 | * check if we need to unboost the queue | |
1819 | */ | |
1820 | if (cfqq->ioprio_class != cfqq->org_ioprio_class) | |
1821 | cfqq->ioprio_class = cfqq->org_ioprio_class; | |
1822 | if (cfqq->ioprio != cfqq->org_ioprio) | |
1823 | cfqq->ioprio = cfqq->org_ioprio; | |
1824 | } | |
22e2c507 | 1825 | } |
1da177e4 | 1826 | |
89850f7e | 1827 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 1828 | { |
3b18152c | 1829 | if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) && |
99f95e52 | 1830 | !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 1831 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 1832 | return ELV_MQUEUE_MUST; |
3b18152c | 1833 | } |
1da177e4 | 1834 | |
22e2c507 | 1835 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
1836 | } |
1837 | ||
cb78b285 | 1838 | static int cfq_may_queue(request_queue_t *q, int rw) |
22e2c507 JA |
1839 | { |
1840 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1841 | struct task_struct *tsk = current; | |
1842 | struct cfq_queue *cfqq; | |
7749a8d4 JA |
1843 | unsigned int key; |
1844 | ||
1845 | key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC); | |
22e2c507 JA |
1846 | |
1847 | /* | |
1848 | * don't force setup of a queue from here, as a call to may_queue | |
1849 | * does not necessarily imply that a request actually will be queued. | |
1850 | * so just lookup a possibly existing queue, or return 'may queue' | |
1851 | * if that fails | |
1852 | */ | |
7749a8d4 | 1853 | cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio); |
22e2c507 JA |
1854 | if (cfqq) { |
1855 | cfq_init_prio_data(cfqq); | |
1856 | cfq_prio_boost(cfqq); | |
1857 | ||
89850f7e | 1858 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
1859 | } |
1860 | ||
1861 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
1862 | } |
1863 | ||
1da177e4 LT |
1864 | /* |
1865 | * queue lock held here | |
1866 | */ | |
bb37b94c | 1867 | static void cfq_put_request(struct request *rq) |
1da177e4 | 1868 | { |
5e705374 | 1869 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 1870 | |
5e705374 | 1871 | if (cfqq) { |
22e2c507 | 1872 | const int rw = rq_data_dir(rq); |
1da177e4 | 1873 | |
22e2c507 JA |
1874 | BUG_ON(!cfqq->allocated[rw]); |
1875 | cfqq->allocated[rw]--; | |
1da177e4 | 1876 | |
5e705374 | 1877 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 1878 | |
1da177e4 | 1879 | rq->elevator_private = NULL; |
5e705374 | 1880 | rq->elevator_private2 = NULL; |
1da177e4 | 1881 | |
1da177e4 LT |
1882 | cfq_put_queue(cfqq); |
1883 | } | |
1884 | } | |
1885 | ||
1886 | /* | |
22e2c507 | 1887 | * Allocate cfq data structures associated with this request. |
1da177e4 | 1888 | */ |
22e2c507 | 1889 | static int |
cb78b285 | 1890 | cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
1891 | { |
1892 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3b18152c | 1893 | struct task_struct *tsk = current; |
1da177e4 LT |
1894 | struct cfq_io_context *cic; |
1895 | const int rw = rq_data_dir(rq); | |
7749a8d4 JA |
1896 | const int is_sync = rq_is_sync(rq); |
1897 | pid_t key = cfq_queue_pid(tsk, rw, is_sync); | |
22e2c507 | 1898 | struct cfq_queue *cfqq; |
1da177e4 LT |
1899 | unsigned long flags; |
1900 | ||
1901 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
1902 | ||
e2d74ac0 | 1903 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 1904 | |
1da177e4 LT |
1905 | spin_lock_irqsave(q->queue_lock, flags); |
1906 | ||
22e2c507 JA |
1907 | if (!cic) |
1908 | goto queue_fail; | |
1909 | ||
12a05732 | 1910 | if (!cic->cfqq[is_sync]) { |
6f325a13 | 1911 | cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask); |
22e2c507 JA |
1912 | if (!cfqq) |
1913 | goto queue_fail; | |
1da177e4 | 1914 | |
12a05732 | 1915 | cic->cfqq[is_sync] = cfqq; |
22e2c507 | 1916 | } else |
12a05732 | 1917 | cfqq = cic->cfqq[is_sync]; |
1da177e4 LT |
1918 | |
1919 | cfqq->allocated[rw]++; | |
3b18152c | 1920 | cfq_clear_cfqq_must_alloc(cfqq); |
22e2c507 | 1921 | atomic_inc(&cfqq->ref); |
1da177e4 | 1922 | |
5e705374 | 1923 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 1924 | |
5e705374 JA |
1925 | rq->elevator_private = cic; |
1926 | rq->elevator_private2 = cfqq; | |
1927 | return 0; | |
1da177e4 | 1928 | |
22e2c507 JA |
1929 | queue_fail: |
1930 | if (cic) | |
1931 | put_io_context(cic->ioc); | |
89850f7e | 1932 | |
3b18152c | 1933 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
1934 | spin_unlock_irqrestore(q->queue_lock, flags); |
1935 | return 1; | |
1936 | } | |
1937 | ||
65f27f38 | 1938 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 1939 | { |
65f27f38 DH |
1940 | struct cfq_data *cfqd = |
1941 | container_of(work, struct cfq_data, unplug_work); | |
1942 | request_queue_t *q = cfqd->queue; | |
22e2c507 JA |
1943 | unsigned long flags; |
1944 | ||
1945 | spin_lock_irqsave(q->queue_lock, flags); | |
dc72ef4a | 1946 | blk_start_queueing(q); |
22e2c507 JA |
1947 | spin_unlock_irqrestore(q->queue_lock, flags); |
1948 | } | |
1949 | ||
1950 | /* | |
1951 | * Timer running if the active_queue is currently idling inside its time slice | |
1952 | */ | |
1953 | static void cfq_idle_slice_timer(unsigned long data) | |
1954 | { | |
1955 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
1956 | struct cfq_queue *cfqq; | |
1957 | unsigned long flags; | |
3c6bd2f8 | 1958 | int timed_out = 1; |
22e2c507 JA |
1959 | |
1960 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
1961 | ||
1962 | if ((cfqq = cfqd->active_queue) != NULL) { | |
3c6bd2f8 JA |
1963 | timed_out = 0; |
1964 | ||
22e2c507 JA |
1965 | /* |
1966 | * expired | |
1967 | */ | |
44f7c160 | 1968 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
1969 | goto expire; |
1970 | ||
1971 | /* | |
1972 | * only expire and reinvoke request handler, if there are | |
1973 | * other queues with pending requests | |
1974 | */ | |
caaa5f9f | 1975 | if (!cfqd->busy_queues) |
22e2c507 | 1976 | goto out_cont; |
22e2c507 JA |
1977 | |
1978 | /* | |
1979 | * not expired and it has a request pending, let it dispatch | |
1980 | */ | |
dd67d051 | 1981 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) { |
3b18152c | 1982 | cfq_mark_cfqq_must_dispatch(cfqq); |
22e2c507 JA |
1983 | goto out_kick; |
1984 | } | |
1985 | } | |
1986 | expire: | |
6084cdda | 1987 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 1988 | out_kick: |
3b18152c | 1989 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
1990 | out_cont: |
1991 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
1992 | } | |
1993 | ||
1994 | /* | |
1995 | * Timer running if an idle class queue is waiting for service | |
1996 | */ | |
1997 | static void cfq_idle_class_timer(unsigned long data) | |
1998 | { | |
1999 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
2000 | unsigned long flags, end; | |
2001 | ||
2002 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
2003 | ||
2004 | /* | |
2005 | * race with a non-idle queue, reset timer | |
2006 | */ | |
2007 | end = cfqd->last_end_request + CFQ_IDLE_GRACE; | |
ae818a38 JA |
2008 | if (!time_after_eq(jiffies, end)) |
2009 | mod_timer(&cfqd->idle_class_timer, end); | |
2010 | else | |
3b18152c | 2011 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
2012 | |
2013 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2014 | } | |
2015 | ||
3b18152c JA |
2016 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
2017 | { | |
2018 | del_timer_sync(&cfqd->idle_slice_timer); | |
2019 | del_timer_sync(&cfqd->idle_class_timer); | |
2020 | blk_sync_queue(cfqd->queue); | |
2021 | } | |
22e2c507 | 2022 | |
1da177e4 LT |
2023 | static void cfq_exit_queue(elevator_t *e) |
2024 | { | |
22e2c507 | 2025 | struct cfq_data *cfqd = e->elevator_data; |
d9ff4187 | 2026 | request_queue_t *q = cfqd->queue; |
22e2c507 | 2027 | |
3b18152c | 2028 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 2029 | |
d9ff4187 | 2030 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 2031 | |
d9ff4187 | 2032 | if (cfqd->active_queue) |
6084cdda | 2033 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
2034 | |
2035 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
2036 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
2037 | struct cfq_io_context, | |
2038 | queue_list); | |
89850f7e JA |
2039 | |
2040 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 2041 | } |
e2d74ac0 | 2042 | |
d9ff4187 | 2043 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
2044 | |
2045 | cfq_shutdown_timer_wq(cfqd); | |
2046 | ||
a90d742e AV |
2047 | kfree(cfqd->cfq_hash); |
2048 | kfree(cfqd); | |
1da177e4 LT |
2049 | } |
2050 | ||
bb37b94c | 2051 | static void *cfq_init_queue(request_queue_t *q) |
1da177e4 LT |
2052 | { |
2053 | struct cfq_data *cfqd; | |
2054 | int i; | |
2055 | ||
b5deef90 | 2056 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node); |
1da177e4 | 2057 | if (!cfqd) |
bc1c1169 | 2058 | return NULL; |
1da177e4 LT |
2059 | |
2060 | memset(cfqd, 0, sizeof(*cfqd)); | |
22e2c507 | 2061 | |
cc09e299 | 2062 | cfqd->service_tree = CFQ_RB_ROOT; |
d9ff4187 | 2063 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 2064 | |
b5deef90 | 2065 | cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node); |
1da177e4 | 2066 | if (!cfqd->cfq_hash) |
5e705374 | 2067 | goto out_free; |
1da177e4 | 2068 | |
1da177e4 LT |
2069 | for (i = 0; i < CFQ_QHASH_ENTRIES; i++) |
2070 | INIT_HLIST_HEAD(&cfqd->cfq_hash[i]); | |
2071 | ||
1da177e4 | 2072 | cfqd->queue = q; |
1da177e4 | 2073 | |
22e2c507 JA |
2074 | init_timer(&cfqd->idle_slice_timer); |
2075 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
2076 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
2077 | ||
2078 | init_timer(&cfqd->idle_class_timer); | |
2079 | cfqd->idle_class_timer.function = cfq_idle_class_timer; | |
2080 | cfqd->idle_class_timer.data = (unsigned long) cfqd; | |
2081 | ||
65f27f38 | 2082 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 2083 | |
1da177e4 | 2084 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
2085 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
2086 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
2087 | cfqd->cfq_back_max = cfq_back_max; |
2088 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
2089 | cfqd->cfq_slice[0] = cfq_slice_async; |
2090 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
2091 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
2092 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
3b18152c | 2093 | |
bc1c1169 | 2094 | return cfqd; |
5e705374 | 2095 | out_free: |
1da177e4 | 2096 | kfree(cfqd); |
bc1c1169 | 2097 | return NULL; |
1da177e4 LT |
2098 | } |
2099 | ||
2100 | static void cfq_slab_kill(void) | |
2101 | { | |
1da177e4 LT |
2102 | if (cfq_pool) |
2103 | kmem_cache_destroy(cfq_pool); | |
2104 | if (cfq_ioc_pool) | |
2105 | kmem_cache_destroy(cfq_ioc_pool); | |
2106 | } | |
2107 | ||
2108 | static int __init cfq_slab_setup(void) | |
2109 | { | |
1da177e4 LT |
2110 | cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0, |
2111 | NULL, NULL); | |
2112 | if (!cfq_pool) | |
2113 | goto fail; | |
2114 | ||
2115 | cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool", | |
2116 | sizeof(struct cfq_io_context), 0, 0, NULL, NULL); | |
2117 | if (!cfq_ioc_pool) | |
2118 | goto fail; | |
2119 | ||
2120 | return 0; | |
2121 | fail: | |
2122 | cfq_slab_kill(); | |
2123 | return -ENOMEM; | |
2124 | } | |
2125 | ||
1da177e4 LT |
2126 | /* |
2127 | * sysfs parts below --> | |
2128 | */ | |
1da177e4 LT |
2129 | static ssize_t |
2130 | cfq_var_show(unsigned int var, char *page) | |
2131 | { | |
2132 | return sprintf(page, "%d\n", var); | |
2133 | } | |
2134 | ||
2135 | static ssize_t | |
2136 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
2137 | { | |
2138 | char *p = (char *) page; | |
2139 | ||
2140 | *var = simple_strtoul(p, &p, 10); | |
2141 | return count; | |
2142 | } | |
2143 | ||
1da177e4 | 2144 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
3d1ab40f | 2145 | static ssize_t __FUNC(elevator_t *e, char *page) \ |
1da177e4 | 2146 | { \ |
3d1ab40f | 2147 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2148 | unsigned int __data = __VAR; \ |
2149 | if (__CONV) \ | |
2150 | __data = jiffies_to_msecs(__data); \ | |
2151 | return cfq_var_show(__data, (page)); \ | |
2152 | } | |
2153 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
2154 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
2155 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
2156 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
2157 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
2158 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
2159 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
2160 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
2161 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
1da177e4 LT |
2162 | #undef SHOW_FUNCTION |
2163 | ||
2164 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
3d1ab40f | 2165 | static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \ |
1da177e4 | 2166 | { \ |
3d1ab40f | 2167 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2168 | unsigned int __data; \ |
2169 | int ret = cfq_var_store(&__data, (page), count); \ | |
2170 | if (__data < (MIN)) \ | |
2171 | __data = (MIN); \ | |
2172 | else if (__data > (MAX)) \ | |
2173 | __data = (MAX); \ | |
2174 | if (__CONV) \ | |
2175 | *(__PTR) = msecs_to_jiffies(__data); \ | |
2176 | else \ | |
2177 | *(__PTR) = __data; \ | |
2178 | return ret; \ | |
2179 | } | |
2180 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
22e2c507 JA |
2181 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1); |
2182 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1); | |
e572ec7e AV |
2183 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
2184 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0); | |
22e2c507 JA |
2185 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
2186 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
2187 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
2188 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0); | |
1da177e4 LT |
2189 | #undef STORE_FUNCTION |
2190 | ||
e572ec7e AV |
2191 | #define CFQ_ATTR(name) \ |
2192 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
2193 | ||
2194 | static struct elv_fs_entry cfq_attrs[] = { | |
2195 | CFQ_ATTR(quantum), | |
e572ec7e AV |
2196 | CFQ_ATTR(fifo_expire_sync), |
2197 | CFQ_ATTR(fifo_expire_async), | |
2198 | CFQ_ATTR(back_seek_max), | |
2199 | CFQ_ATTR(back_seek_penalty), | |
2200 | CFQ_ATTR(slice_sync), | |
2201 | CFQ_ATTR(slice_async), | |
2202 | CFQ_ATTR(slice_async_rq), | |
2203 | CFQ_ATTR(slice_idle), | |
e572ec7e | 2204 | __ATTR_NULL |
1da177e4 LT |
2205 | }; |
2206 | ||
1da177e4 LT |
2207 | static struct elevator_type iosched_cfq = { |
2208 | .ops = { | |
2209 | .elevator_merge_fn = cfq_merge, | |
2210 | .elevator_merged_fn = cfq_merged_request, | |
2211 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 2212 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 2213 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 2214 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 2215 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
2216 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
2217 | .elevator_queue_empty_fn = cfq_queue_empty, | |
2218 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
2219 | .elevator_former_req_fn = elv_rb_former_request, |
2220 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
2221 | .elevator_set_req_fn = cfq_set_request, |
2222 | .elevator_put_req_fn = cfq_put_request, | |
2223 | .elevator_may_queue_fn = cfq_may_queue, | |
2224 | .elevator_init_fn = cfq_init_queue, | |
2225 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 2226 | .trim = cfq_free_io_context, |
1da177e4 | 2227 | }, |
3d1ab40f | 2228 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
2229 | .elevator_name = "cfq", |
2230 | .elevator_owner = THIS_MODULE, | |
2231 | }; | |
2232 | ||
2233 | static int __init cfq_init(void) | |
2234 | { | |
2235 | int ret; | |
2236 | ||
22e2c507 JA |
2237 | /* |
2238 | * could be 0 on HZ < 1000 setups | |
2239 | */ | |
2240 | if (!cfq_slice_async) | |
2241 | cfq_slice_async = 1; | |
2242 | if (!cfq_slice_idle) | |
2243 | cfq_slice_idle = 1; | |
2244 | ||
1da177e4 LT |
2245 | if (cfq_slab_setup()) |
2246 | return -ENOMEM; | |
2247 | ||
2248 | ret = elv_register(&iosched_cfq); | |
22e2c507 JA |
2249 | if (ret) |
2250 | cfq_slab_kill(); | |
1da177e4 | 2251 | |
1da177e4 LT |
2252 | return ret; |
2253 | } | |
2254 | ||
2255 | static void __exit cfq_exit(void) | |
2256 | { | |
6e9a4738 | 2257 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 2258 | elv_unregister(&iosched_cfq); |
334e94de | 2259 | ioc_gone = &all_gone; |
fba82272 OH |
2260 | /* ioc_gone's update must be visible before reading ioc_count */ |
2261 | smp_wmb(); | |
4050cf16 | 2262 | if (elv_ioc_count_read(ioc_count)) |
fba82272 | 2263 | wait_for_completion(ioc_gone); |
334e94de | 2264 | synchronize_rcu(); |
83521d3e | 2265 | cfq_slab_kill(); |
1da177e4 LT |
2266 | } |
2267 | ||
2268 | module_init(cfq_init); | |
2269 | module_exit(cfq_exit); | |
2270 | ||
2271 | MODULE_AUTHOR("Jens Axboe"); | |
2272 | MODULE_LICENSE("GPL"); | |
2273 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |