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 | 12 | #include <linux/rbtree.h> |
22e2c507 | 13 | #include <linux/ioprio.h> |
1da177e4 LT |
14 | |
15 | /* | |
16 | * tunables | |
17 | */ | |
fe094d98 JA |
18 | /* max queue in one round of service */ |
19 | static const int cfq_quantum = 4; | |
64100099 | 20 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
fe094d98 JA |
21 | /* maximum backwards seek, in KiB */ |
22 | static const int cfq_back_max = 16 * 1024; | |
23 | /* penalty of a backwards seek */ | |
24 | static const int cfq_back_penalty = 2; | |
64100099 | 25 | static const int cfq_slice_sync = HZ / 10; |
3b18152c | 26 | static int cfq_slice_async = HZ / 25; |
64100099 | 27 | static const int cfq_slice_async_rq = 2; |
caaa5f9f | 28 | static int cfq_slice_idle = HZ / 125; |
22e2c507 | 29 | |
d9e7620e | 30 | /* |
0871714e | 31 | * offset from end of service tree |
d9e7620e | 32 | */ |
0871714e | 33 | #define CFQ_IDLE_DELAY (HZ / 5) |
d9e7620e JA |
34 | |
35 | /* | |
36 | * below this threshold, we consider thinktime immediate | |
37 | */ | |
38 | #define CFQ_MIN_TT (2) | |
39 | ||
22e2c507 JA |
40 | #define CFQ_SLICE_SCALE (5) |
41 | ||
fe094d98 JA |
42 | #define RQ_CIC(rq) \ |
43 | ((struct cfq_io_context *) (rq)->elevator_private) | |
5e705374 | 44 | #define RQ_CFQQ(rq) ((rq)->elevator_private2) |
1da177e4 | 45 | |
e18b890b CL |
46 | static struct kmem_cache *cfq_pool; |
47 | static struct kmem_cache *cfq_ioc_pool; | |
1da177e4 | 48 | |
4050cf16 | 49 | static DEFINE_PER_CPU(unsigned long, ioc_count); |
334e94de AV |
50 | static struct completion *ioc_gone; |
51 | ||
22e2c507 JA |
52 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR |
53 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
22e2c507 JA |
54 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) |
55 | ||
3b18152c JA |
56 | #define ASYNC (0) |
57 | #define SYNC (1) | |
58 | ||
206dc69b JA |
59 | #define sample_valid(samples) ((samples) > 80) |
60 | ||
cc09e299 JA |
61 | /* |
62 | * Most of our rbtree usage is for sorting with min extraction, so | |
63 | * if we cache the leftmost node we don't have to walk down the tree | |
64 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
65 | * move this into the elevator for the rq sorting as well. | |
66 | */ | |
67 | struct cfq_rb_root { | |
68 | struct rb_root rb; | |
69 | struct rb_node *left; | |
70 | }; | |
71 | #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, } | |
72 | ||
22e2c507 JA |
73 | /* |
74 | * Per block device queue structure | |
75 | */ | |
1da177e4 | 76 | struct cfq_data { |
165125e1 | 77 | struct request_queue *queue; |
22e2c507 JA |
78 | |
79 | /* | |
80 | * rr list of queues with requests and the count of them | |
81 | */ | |
cc09e299 | 82 | struct cfq_rb_root service_tree; |
22e2c507 JA |
83 | unsigned int busy_queues; |
84 | ||
22e2c507 | 85 | int rq_in_driver; |
3ed9a296 | 86 | int sync_flight; |
25776e35 | 87 | int hw_tag; |
1da177e4 | 88 | |
22e2c507 JA |
89 | /* |
90 | * idle window management | |
91 | */ | |
92 | struct timer_list idle_slice_timer; | |
93 | struct work_struct unplug_work; | |
1da177e4 | 94 | |
22e2c507 JA |
95 | struct cfq_queue *active_queue; |
96 | struct cfq_io_context *active_cic; | |
22e2c507 | 97 | |
c2dea2d1 VT |
98 | /* |
99 | * async queue for each priority case | |
100 | */ | |
101 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
102 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 103 | |
6d048f53 | 104 | sector_t last_position; |
22e2c507 | 105 | unsigned long last_end_request; |
1da177e4 | 106 | |
1da177e4 LT |
107 | /* |
108 | * tunables, see top of file | |
109 | */ | |
110 | unsigned int cfq_quantum; | |
22e2c507 | 111 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
112 | unsigned int cfq_back_penalty; |
113 | unsigned int cfq_back_max; | |
22e2c507 JA |
114 | unsigned int cfq_slice[2]; |
115 | unsigned int cfq_slice_async_rq; | |
116 | unsigned int cfq_slice_idle; | |
d9ff4187 AV |
117 | |
118 | struct list_head cic_list; | |
1da177e4 LT |
119 | }; |
120 | ||
22e2c507 JA |
121 | /* |
122 | * Per process-grouping structure | |
123 | */ | |
1da177e4 LT |
124 | struct cfq_queue { |
125 | /* reference count */ | |
126 | atomic_t ref; | |
127 | /* parent cfq_data */ | |
128 | struct cfq_data *cfqd; | |
d9e7620e JA |
129 | /* service_tree member */ |
130 | struct rb_node rb_node; | |
131 | /* service_tree key */ | |
132 | unsigned long rb_key; | |
1da177e4 LT |
133 | /* sorted list of pending requests */ |
134 | struct rb_root sort_list; | |
135 | /* if fifo isn't expired, next request to serve */ | |
5e705374 | 136 | struct request *next_rq; |
1da177e4 LT |
137 | /* requests queued in sort_list */ |
138 | int queued[2]; | |
139 | /* currently allocated requests */ | |
140 | int allocated[2]; | |
374f84ac JA |
141 | /* pending metadata requests */ |
142 | int meta_pending; | |
1da177e4 | 143 | /* fifo list of requests in sort_list */ |
22e2c507 | 144 | struct list_head fifo; |
1da177e4 | 145 | |
22e2c507 | 146 | unsigned long slice_end; |
c5b680f3 | 147 | long slice_resid; |
1da177e4 | 148 | |
6d048f53 JA |
149 | /* number of requests that are on the dispatch list or inside driver */ |
150 | int dispatched; | |
22e2c507 JA |
151 | |
152 | /* io prio of this group */ | |
153 | unsigned short ioprio, org_ioprio; | |
154 | unsigned short ioprio_class, org_ioprio_class; | |
155 | ||
3b18152c JA |
156 | /* various state flags, see below */ |
157 | unsigned int flags; | |
1da177e4 LT |
158 | }; |
159 | ||
3b18152c | 160 | enum cfqq_state_flags { |
b0b8d749 JA |
161 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
162 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
163 | CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */ | |
164 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ | |
165 | CFQ_CFQQ_FLAG_must_dispatch, /* must dispatch, even if expired */ | |
166 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ | |
167 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
168 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
169 | CFQ_CFQQ_FLAG_queue_new, /* queue never been serviced */ | |
44f7c160 | 170 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 171 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
3b18152c JA |
172 | }; |
173 | ||
174 | #define CFQ_CFQQ_FNS(name) \ | |
175 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
176 | { \ | |
fe094d98 | 177 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
178 | } \ |
179 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
180 | { \ | |
fe094d98 | 181 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
182 | } \ |
183 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
184 | { \ | |
fe094d98 | 185 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
186 | } |
187 | ||
188 | CFQ_CFQQ_FNS(on_rr); | |
189 | CFQ_CFQQ_FNS(wait_request); | |
190 | CFQ_CFQQ_FNS(must_alloc); | |
191 | CFQ_CFQQ_FNS(must_alloc_slice); | |
192 | CFQ_CFQQ_FNS(must_dispatch); | |
193 | CFQ_CFQQ_FNS(fifo_expire); | |
194 | CFQ_CFQQ_FNS(idle_window); | |
195 | CFQ_CFQQ_FNS(prio_changed); | |
53b03744 | 196 | CFQ_CFQQ_FNS(queue_new); |
44f7c160 | 197 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 198 | CFQ_CFQQ_FNS(sync); |
3b18152c JA |
199 | #undef CFQ_CFQQ_FNS |
200 | ||
165125e1 | 201 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
91fac317 | 202 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, int, |
fd0928df | 203 | struct io_context *, gfp_t); |
4ac845a2 | 204 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
205 | struct io_context *); |
206 | ||
207 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, | |
208 | int is_sync) | |
209 | { | |
210 | return cic->cfqq[!!is_sync]; | |
211 | } | |
212 | ||
213 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
214 | struct cfq_queue *cfqq, int is_sync) | |
215 | { | |
216 | cic->cfqq[!!is_sync] = cfqq; | |
217 | } | |
218 | ||
219 | /* | |
220 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
221 | * set (in which case it could also be direct WRITE). | |
222 | */ | |
223 | static inline int cfq_bio_sync(struct bio *bio) | |
224 | { | |
225 | if (bio_data_dir(bio) == READ || bio_sync(bio)) | |
226 | return 1; | |
227 | ||
228 | return 0; | |
229 | } | |
1da177e4 | 230 | |
99f95e52 AM |
231 | /* |
232 | * scheduler run of queue, if there are requests pending and no one in the | |
233 | * driver that will restart queueing | |
234 | */ | |
235 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) | |
236 | { | |
7b14e3b5 | 237 | if (cfqd->busy_queues) |
99f95e52 AM |
238 | kblockd_schedule_work(&cfqd->unplug_work); |
239 | } | |
240 | ||
165125e1 | 241 | static int cfq_queue_empty(struct request_queue *q) |
99f95e52 AM |
242 | { |
243 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
244 | ||
b4878f24 | 245 | return !cfqd->busy_queues; |
99f95e52 AM |
246 | } |
247 | ||
44f7c160 JA |
248 | /* |
249 | * Scale schedule slice based on io priority. Use the sync time slice only | |
250 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
251 | * io only, should not get full sync slice length. | |
252 | */ | |
d9e7620e JA |
253 | static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync, |
254 | unsigned short prio) | |
44f7c160 | 255 | { |
d9e7620e | 256 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 257 | |
d9e7620e JA |
258 | WARN_ON(prio >= IOPRIO_BE_NR); |
259 | ||
260 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
261 | } | |
44f7c160 | 262 | |
d9e7620e JA |
263 | static inline int |
264 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
265 | { | |
266 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
267 | } |
268 | ||
269 | static inline void | |
270 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
271 | { | |
272 | cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies; | |
273 | } | |
274 | ||
275 | /* | |
276 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
277 | * isn't valid until the first request from the dispatch is activated | |
278 | * and the slice time set. | |
279 | */ | |
280 | static inline int cfq_slice_used(struct cfq_queue *cfqq) | |
281 | { | |
282 | if (cfq_cfqq_slice_new(cfqq)) | |
283 | return 0; | |
284 | if (time_before(jiffies, cfqq->slice_end)) | |
285 | return 0; | |
286 | ||
287 | return 1; | |
288 | } | |
289 | ||
1da177e4 | 290 | /* |
5e705374 | 291 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 292 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 293 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 294 | */ |
5e705374 JA |
295 | static struct request * |
296 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2) | |
1da177e4 LT |
297 | { |
298 | sector_t last, s1, s2, d1 = 0, d2 = 0; | |
1da177e4 | 299 | unsigned long back_max; |
e8a99053 AM |
300 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
301 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
302 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 303 | |
5e705374 JA |
304 | if (rq1 == NULL || rq1 == rq2) |
305 | return rq2; | |
306 | if (rq2 == NULL) | |
307 | return rq1; | |
9c2c38a1 | 308 | |
5e705374 JA |
309 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
310 | return rq1; | |
311 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
312 | return rq2; | |
374f84ac JA |
313 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
314 | return rq1; | |
315 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
316 | return rq2; | |
1da177e4 | 317 | |
5e705374 JA |
318 | s1 = rq1->sector; |
319 | s2 = rq2->sector; | |
1da177e4 | 320 | |
6d048f53 | 321 | last = cfqd->last_position; |
1da177e4 | 322 | |
1da177e4 LT |
323 | /* |
324 | * by definition, 1KiB is 2 sectors | |
325 | */ | |
326 | back_max = cfqd->cfq_back_max * 2; | |
327 | ||
328 | /* | |
329 | * Strict one way elevator _except_ in the case where we allow | |
330 | * short backward seeks which are biased as twice the cost of a | |
331 | * similar forward seek. | |
332 | */ | |
333 | if (s1 >= last) | |
334 | d1 = s1 - last; | |
335 | else if (s1 + back_max >= last) | |
336 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
337 | else | |
e8a99053 | 338 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
339 | |
340 | if (s2 >= last) | |
341 | d2 = s2 - last; | |
342 | else if (s2 + back_max >= last) | |
343 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
344 | else | |
e8a99053 | 345 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
346 | |
347 | /* Found required data */ | |
e8a99053 AM |
348 | |
349 | /* | |
350 | * By doing switch() on the bit mask "wrap" we avoid having to | |
351 | * check two variables for all permutations: --> faster! | |
352 | */ | |
353 | switch (wrap) { | |
5e705374 | 354 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 355 | if (d1 < d2) |
5e705374 | 356 | return rq1; |
e8a99053 | 357 | else if (d2 < d1) |
5e705374 | 358 | return rq2; |
e8a99053 AM |
359 | else { |
360 | if (s1 >= s2) | |
5e705374 | 361 | return rq1; |
e8a99053 | 362 | else |
5e705374 | 363 | return rq2; |
e8a99053 | 364 | } |
1da177e4 | 365 | |
e8a99053 | 366 | case CFQ_RQ2_WRAP: |
5e705374 | 367 | return rq1; |
e8a99053 | 368 | case CFQ_RQ1_WRAP: |
5e705374 JA |
369 | return rq2; |
370 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
371 | default: |
372 | /* | |
373 | * Since both rqs are wrapped, | |
374 | * start with the one that's further behind head | |
375 | * (--> only *one* back seek required), | |
376 | * since back seek takes more time than forward. | |
377 | */ | |
378 | if (s1 <= s2) | |
5e705374 | 379 | return rq1; |
1da177e4 | 380 | else |
5e705374 | 381 | return rq2; |
1da177e4 LT |
382 | } |
383 | } | |
384 | ||
498d3aa2 JA |
385 | /* |
386 | * The below is leftmost cache rbtree addon | |
387 | */ | |
0871714e | 388 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 JA |
389 | { |
390 | if (!root->left) | |
391 | root->left = rb_first(&root->rb); | |
392 | ||
0871714e JA |
393 | if (root->left) |
394 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
395 | ||
396 | return NULL; | |
cc09e299 JA |
397 | } |
398 | ||
399 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) | |
400 | { | |
401 | if (root->left == n) | |
402 | root->left = NULL; | |
403 | ||
404 | rb_erase(n, &root->rb); | |
405 | RB_CLEAR_NODE(n); | |
406 | } | |
407 | ||
1da177e4 LT |
408 | /* |
409 | * would be nice to take fifo expire time into account as well | |
410 | */ | |
5e705374 JA |
411 | static struct request * |
412 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
413 | struct request *last) | |
1da177e4 | 414 | { |
21183b07 JA |
415 | struct rb_node *rbnext = rb_next(&last->rb_node); |
416 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 417 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 418 | |
21183b07 | 419 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
420 | |
421 | if (rbprev) | |
5e705374 | 422 | prev = rb_entry_rq(rbprev); |
1da177e4 | 423 | |
21183b07 | 424 | if (rbnext) |
5e705374 | 425 | next = rb_entry_rq(rbnext); |
21183b07 JA |
426 | else { |
427 | rbnext = rb_first(&cfqq->sort_list); | |
428 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 429 | next = rb_entry_rq(rbnext); |
21183b07 | 430 | } |
1da177e4 | 431 | |
21183b07 | 432 | return cfq_choose_req(cfqd, next, prev); |
1da177e4 LT |
433 | } |
434 | ||
d9e7620e JA |
435 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
436 | struct cfq_queue *cfqq) | |
1da177e4 | 437 | { |
d9e7620e JA |
438 | /* |
439 | * just an approximation, should be ok. | |
440 | */ | |
67e6b49e JA |
441 | return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
442 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); | |
d9e7620e JA |
443 | } |
444 | ||
498d3aa2 JA |
445 | /* |
446 | * The cfqd->service_tree holds all pending cfq_queue's that have | |
447 | * requests waiting to be processed. It is sorted in the order that | |
448 | * we will service the queues. | |
449 | */ | |
d9e7620e | 450 | static void cfq_service_tree_add(struct cfq_data *cfqd, |
edd75ffd | 451 | struct cfq_queue *cfqq, int add_front) |
d9e7620e | 452 | { |
0871714e JA |
453 | struct rb_node **p, *parent; |
454 | struct cfq_queue *__cfqq; | |
d9e7620e | 455 | unsigned long rb_key; |
498d3aa2 | 456 | int left; |
d9e7620e | 457 | |
0871714e JA |
458 | if (cfq_class_idle(cfqq)) { |
459 | rb_key = CFQ_IDLE_DELAY; | |
460 | parent = rb_last(&cfqd->service_tree.rb); | |
461 | if (parent && parent != &cfqq->rb_node) { | |
462 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
463 | rb_key += __cfqq->rb_key; | |
464 | } else | |
465 | rb_key += jiffies; | |
466 | } else if (!add_front) { | |
edd75ffd JA |
467 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
468 | rb_key += cfqq->slice_resid; | |
469 | cfqq->slice_resid = 0; | |
470 | } else | |
471 | rb_key = 0; | |
1da177e4 | 472 | |
d9e7620e | 473 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
99f9628a | 474 | /* |
d9e7620e | 475 | * same position, nothing more to do |
99f9628a | 476 | */ |
d9e7620e JA |
477 | if (rb_key == cfqq->rb_key) |
478 | return; | |
1da177e4 | 479 | |
cc09e299 | 480 | cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree); |
1da177e4 | 481 | } |
d9e7620e | 482 | |
498d3aa2 | 483 | left = 1; |
0871714e JA |
484 | parent = NULL; |
485 | p = &cfqd->service_tree.rb.rb_node; | |
d9e7620e | 486 | while (*p) { |
67060e37 | 487 | struct rb_node **n; |
cc09e299 | 488 | |
d9e7620e JA |
489 | parent = *p; |
490 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
491 | ||
0c534e0a JA |
492 | /* |
493 | * sort RT queues first, we always want to give | |
67060e37 JA |
494 | * preference to them. IDLE queues goes to the back. |
495 | * after that, sort on the next service time. | |
0c534e0a JA |
496 | */ |
497 | if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq)) | |
67060e37 | 498 | n = &(*p)->rb_left; |
0c534e0a | 499 | else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq)) |
67060e37 JA |
500 | n = &(*p)->rb_right; |
501 | else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq)) | |
502 | n = &(*p)->rb_left; | |
503 | else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq)) | |
504 | n = &(*p)->rb_right; | |
0c534e0a | 505 | else if (rb_key < __cfqq->rb_key) |
67060e37 JA |
506 | n = &(*p)->rb_left; |
507 | else | |
508 | n = &(*p)->rb_right; | |
509 | ||
510 | if (n == &(*p)->rb_right) | |
cc09e299 | 511 | left = 0; |
67060e37 JA |
512 | |
513 | p = n; | |
d9e7620e JA |
514 | } |
515 | ||
cc09e299 JA |
516 | if (left) |
517 | cfqd->service_tree.left = &cfqq->rb_node; | |
518 | ||
d9e7620e JA |
519 | cfqq->rb_key = rb_key; |
520 | rb_link_node(&cfqq->rb_node, parent, p); | |
cc09e299 | 521 | rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb); |
1da177e4 LT |
522 | } |
523 | ||
498d3aa2 JA |
524 | /* |
525 | * Update cfqq's position in the service tree. | |
526 | */ | |
edd75ffd | 527 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 528 | { |
6d048f53 JA |
529 | /* |
530 | * Resorting requires the cfqq to be on the RR list already. | |
531 | */ | |
498d3aa2 | 532 | if (cfq_cfqq_on_rr(cfqq)) |
edd75ffd | 533 | cfq_service_tree_add(cfqd, cfqq, 0); |
6d048f53 JA |
534 | } |
535 | ||
1da177e4 LT |
536 | /* |
537 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 538 | * the pending list according to last request service |
1da177e4 | 539 | */ |
febffd61 | 540 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 541 | { |
3b18152c JA |
542 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
543 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 LT |
544 | cfqd->busy_queues++; |
545 | ||
edd75ffd | 546 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
547 | } |
548 | ||
498d3aa2 JA |
549 | /* |
550 | * Called when the cfqq no longer has requests pending, remove it from | |
551 | * the service tree. | |
552 | */ | |
febffd61 | 553 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 554 | { |
3b18152c JA |
555 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
556 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 557 | |
cc09e299 JA |
558 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) |
559 | cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree); | |
d9e7620e | 560 | |
1da177e4 LT |
561 | BUG_ON(!cfqd->busy_queues); |
562 | cfqd->busy_queues--; | |
563 | } | |
564 | ||
565 | /* | |
566 | * rb tree support functions | |
567 | */ | |
febffd61 | 568 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 569 | { |
5e705374 | 570 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 571 | struct cfq_data *cfqd = cfqq->cfqd; |
5e705374 | 572 | const int sync = rq_is_sync(rq); |
1da177e4 | 573 | |
b4878f24 JA |
574 | BUG_ON(!cfqq->queued[sync]); |
575 | cfqq->queued[sync]--; | |
1da177e4 | 576 | |
5e705374 | 577 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 578 | |
dd67d051 | 579 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
b4878f24 | 580 | cfq_del_cfqq_rr(cfqd, cfqq); |
1da177e4 LT |
581 | } |
582 | ||
5e705374 | 583 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 584 | { |
5e705374 | 585 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 586 | struct cfq_data *cfqd = cfqq->cfqd; |
21183b07 | 587 | struct request *__alias; |
1da177e4 | 588 | |
5380a101 | 589 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
590 | |
591 | /* | |
592 | * looks a little odd, but the first insert might return an alias. | |
593 | * if that happens, put the alias on the dispatch list | |
594 | */ | |
21183b07 | 595 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 596 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
597 | |
598 | if (!cfq_cfqq_on_rr(cfqq)) | |
599 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
600 | |
601 | /* | |
602 | * check if this request is a better next-serve candidate | |
603 | */ | |
604 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq); | |
605 | BUG_ON(!cfqq->next_rq); | |
1da177e4 LT |
606 | } |
607 | ||
febffd61 | 608 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 609 | { |
5380a101 JA |
610 | elv_rb_del(&cfqq->sort_list, rq); |
611 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 612 | cfq_add_rq_rb(rq); |
1da177e4 LT |
613 | } |
614 | ||
206dc69b JA |
615 | static struct request * |
616 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 617 | { |
206dc69b | 618 | struct task_struct *tsk = current; |
91fac317 | 619 | struct cfq_io_context *cic; |
206dc69b | 620 | struct cfq_queue *cfqq; |
1da177e4 | 621 | |
4ac845a2 | 622 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
623 | if (!cic) |
624 | return NULL; | |
625 | ||
626 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
627 | if (cfqq) { |
628 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
629 | ||
21183b07 | 630 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 631 | } |
1da177e4 | 632 | |
1da177e4 LT |
633 | return NULL; |
634 | } | |
635 | ||
165125e1 | 636 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 637 | { |
22e2c507 | 638 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 639 | |
b4878f24 | 640 | cfqd->rq_in_driver++; |
25776e35 JA |
641 | |
642 | /* | |
643 | * If the depth is larger 1, it really could be queueing. But lets | |
644 | * make the mark a little higher - idling could still be good for | |
645 | * low queueing, and a low queueing number could also just indicate | |
646 | * a SCSI mid layer like behaviour where limit+1 is often seen. | |
647 | */ | |
648 | if (!cfqd->hw_tag && cfqd->rq_in_driver > 4) | |
649 | cfqd->hw_tag = 1; | |
6d048f53 JA |
650 | |
651 | cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors; | |
1da177e4 LT |
652 | } |
653 | ||
165125e1 | 654 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 655 | { |
b4878f24 JA |
656 | struct cfq_data *cfqd = q->elevator->elevator_data; |
657 | ||
658 | WARN_ON(!cfqd->rq_in_driver); | |
659 | cfqd->rq_in_driver--; | |
1da177e4 LT |
660 | } |
661 | ||
b4878f24 | 662 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 663 | { |
5e705374 | 664 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 665 | |
5e705374 JA |
666 | if (cfqq->next_rq == rq) |
667 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 668 | |
b4878f24 | 669 | list_del_init(&rq->queuelist); |
5e705374 | 670 | cfq_del_rq_rb(rq); |
374f84ac JA |
671 | |
672 | if (rq_is_meta(rq)) { | |
673 | WARN_ON(!cfqq->meta_pending); | |
674 | cfqq->meta_pending--; | |
675 | } | |
1da177e4 LT |
676 | } |
677 | ||
165125e1 JA |
678 | static int cfq_merge(struct request_queue *q, struct request **req, |
679 | struct bio *bio) | |
1da177e4 LT |
680 | { |
681 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
682 | struct request *__rq; | |
1da177e4 | 683 | |
206dc69b | 684 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 685 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
686 | *req = __rq; |
687 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
688 | } |
689 | ||
690 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
691 | } |
692 | ||
165125e1 | 693 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 694 | int type) |
1da177e4 | 695 | { |
21183b07 | 696 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 697 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 698 | |
5e705374 | 699 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 700 | } |
1da177e4 LT |
701 | } |
702 | ||
703 | static void | |
165125e1 | 704 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
705 | struct request *next) |
706 | { | |
22e2c507 JA |
707 | /* |
708 | * reposition in fifo if next is older than rq | |
709 | */ | |
710 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
711 | time_before(next->start_time, rq->start_time)) | |
712 | list_move(&rq->queuelist, &next->queuelist); | |
713 | ||
b4878f24 | 714 | cfq_remove_request(next); |
22e2c507 JA |
715 | } |
716 | ||
165125e1 | 717 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
718 | struct bio *bio) |
719 | { | |
720 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 721 | struct cfq_io_context *cic; |
da775265 | 722 | struct cfq_queue *cfqq; |
da775265 JA |
723 | |
724 | /* | |
ec8acb69 | 725 | * Disallow merge of a sync bio into an async request. |
da775265 | 726 | */ |
91fac317 | 727 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
da775265 JA |
728 | return 0; |
729 | ||
730 | /* | |
719d3402 JA |
731 | * Lookup the cfqq that this bio will be queued with. Allow |
732 | * merge only if rq is queued there. | |
da775265 | 733 | */ |
4ac845a2 | 734 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 VT |
735 | if (!cic) |
736 | return 0; | |
719d3402 | 737 | |
91fac317 | 738 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
719d3402 JA |
739 | if (cfqq == RQ_CFQQ(rq)) |
740 | return 1; | |
da775265 | 741 | |
ec8acb69 | 742 | return 0; |
da775265 JA |
743 | } |
744 | ||
febffd61 JA |
745 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
746 | struct cfq_queue *cfqq) | |
22e2c507 JA |
747 | { |
748 | if (cfqq) { | |
22e2c507 | 749 | cfqq->slice_end = 0; |
3b18152c JA |
750 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
751 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 752 | cfq_mark_cfqq_slice_new(cfqq); |
1afba045 | 753 | cfq_clear_cfqq_queue_new(cfqq); |
22e2c507 JA |
754 | } |
755 | ||
756 | cfqd->active_queue = cfqq; | |
757 | } | |
758 | ||
7b14e3b5 JA |
759 | /* |
760 | * current cfqq expired its slice (or was too idle), select new one | |
761 | */ | |
762 | static void | |
763 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
6084cdda | 764 | int timed_out) |
7b14e3b5 | 765 | { |
7b14e3b5 JA |
766 | if (cfq_cfqq_wait_request(cfqq)) |
767 | del_timer(&cfqd->idle_slice_timer); | |
768 | ||
7b14e3b5 JA |
769 | cfq_clear_cfqq_must_dispatch(cfqq); |
770 | cfq_clear_cfqq_wait_request(cfqq); | |
771 | ||
772 | /* | |
6084cdda | 773 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 774 | */ |
3c6bd2f8 | 775 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) |
c5b680f3 | 776 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b14e3b5 | 777 | |
edd75ffd | 778 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
779 | |
780 | if (cfqq == cfqd->active_queue) | |
781 | cfqd->active_queue = NULL; | |
782 | ||
783 | if (cfqd->active_cic) { | |
784 | put_io_context(cfqd->active_cic->ioc); | |
785 | cfqd->active_cic = NULL; | |
786 | } | |
7b14e3b5 JA |
787 | } |
788 | ||
6084cdda | 789 | static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out) |
7b14e3b5 JA |
790 | { |
791 | struct cfq_queue *cfqq = cfqd->active_queue; | |
792 | ||
793 | if (cfqq) | |
6084cdda | 794 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
795 | } |
796 | ||
498d3aa2 JA |
797 | /* |
798 | * Get next queue for service. Unless we have a queue preemption, | |
799 | * we'll simply select the first cfqq in the service tree. | |
800 | */ | |
6d048f53 | 801 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 802 | { |
edd75ffd JA |
803 | if (RB_EMPTY_ROOT(&cfqd->service_tree.rb)) |
804 | return NULL; | |
d9e7620e | 805 | |
0871714e | 806 | return cfq_rb_first(&cfqd->service_tree); |
6d048f53 JA |
807 | } |
808 | ||
498d3aa2 JA |
809 | /* |
810 | * Get and set a new active queue for service. | |
811 | */ | |
6d048f53 JA |
812 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd) |
813 | { | |
814 | struct cfq_queue *cfqq; | |
815 | ||
d9e7620e | 816 | cfqq = cfq_get_next_queue(cfqd); |
22e2c507 | 817 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 818 | return cfqq; |
22e2c507 JA |
819 | } |
820 | ||
d9e7620e JA |
821 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
822 | struct request *rq) | |
823 | { | |
824 | if (rq->sector >= cfqd->last_position) | |
825 | return rq->sector - cfqd->last_position; | |
826 | else | |
827 | return cfqd->last_position - rq->sector; | |
828 | } | |
829 | ||
6d048f53 JA |
830 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq) |
831 | { | |
832 | struct cfq_io_context *cic = cfqd->active_cic; | |
833 | ||
834 | if (!sample_valid(cic->seek_samples)) | |
835 | return 0; | |
836 | ||
837 | return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean; | |
838 | } | |
839 | ||
d9e7620e JA |
840 | static int cfq_close_cooperator(struct cfq_data *cfq_data, |
841 | struct cfq_queue *cfqq) | |
6d048f53 | 842 | { |
6d048f53 | 843 | /* |
d9e7620e JA |
844 | * We should notice if some of the queues are cooperating, eg |
845 | * working closely on the same area of the disk. In that case, | |
846 | * we can group them together and don't waste time idling. | |
6d048f53 | 847 | */ |
d9e7620e | 848 | return 0; |
6d048f53 JA |
849 | } |
850 | ||
851 | #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024)) | |
caaa5f9f | 852 | |
6d048f53 | 853 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 854 | { |
1792669c | 855 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 856 | struct cfq_io_context *cic; |
7b14e3b5 JA |
857 | unsigned long sl; |
858 | ||
dd67d051 | 859 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 860 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
861 | |
862 | /* | |
863 | * idle is disabled, either manually or by past process history | |
864 | */ | |
6d048f53 JA |
865 | if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq)) |
866 | return; | |
867 | ||
22e2c507 JA |
868 | /* |
869 | * task has exited, don't wait | |
870 | */ | |
206dc69b | 871 | cic = cfqd->active_cic; |
66dac98e | 872 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
873 | return; |
874 | ||
875 | /* | |
876 | * See if this prio level has a good candidate | |
877 | */ | |
1afba045 JA |
878 | if (cfq_close_cooperator(cfqd, cfqq) && |
879 | (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2)) | |
6d048f53 | 880 | return; |
22e2c507 | 881 | |
3b18152c JA |
882 | cfq_mark_cfqq_must_dispatch(cfqq); |
883 | cfq_mark_cfqq_wait_request(cfqq); | |
22e2c507 | 884 | |
206dc69b JA |
885 | /* |
886 | * we don't want to idle for seeks, but we do want to allow | |
887 | * fair distribution of slice time for a process doing back-to-back | |
888 | * seeks. so allow a little bit of time for him to submit a new rq | |
889 | */ | |
6d048f53 | 890 | sl = cfqd->cfq_slice_idle; |
caaa5f9f | 891 | if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic)) |
d9e7620e | 892 | sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT)); |
206dc69b | 893 | |
7b14e3b5 | 894 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
1da177e4 LT |
895 | } |
896 | ||
498d3aa2 JA |
897 | /* |
898 | * Move request from internal lists to the request queue dispatch list. | |
899 | */ | |
165125e1 | 900 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 901 | { |
3ed9a296 | 902 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 903 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 904 | |
5380a101 | 905 | cfq_remove_request(rq); |
6d048f53 | 906 | cfqq->dispatched++; |
5380a101 | 907 | elv_dispatch_sort(q, rq); |
3ed9a296 JA |
908 | |
909 | if (cfq_cfqq_sync(cfqq)) | |
910 | cfqd->sync_flight++; | |
1da177e4 LT |
911 | } |
912 | ||
913 | /* | |
914 | * return expired entry, or NULL to just start from scratch in rbtree | |
915 | */ | |
febffd61 | 916 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 LT |
917 | { |
918 | struct cfq_data *cfqd = cfqq->cfqd; | |
22e2c507 | 919 | struct request *rq; |
89850f7e | 920 | int fifo; |
1da177e4 | 921 | |
3b18152c | 922 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 923 | return NULL; |
cb887411 JA |
924 | |
925 | cfq_mark_cfqq_fifo_expire(cfqq); | |
926 | ||
89850f7e JA |
927 | if (list_empty(&cfqq->fifo)) |
928 | return NULL; | |
1da177e4 | 929 | |
6d048f53 | 930 | fifo = cfq_cfqq_sync(cfqq); |
89850f7e | 931 | rq = rq_entry_fifo(cfqq->fifo.next); |
1da177e4 | 932 | |
6d048f53 JA |
933 | if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) |
934 | return NULL; | |
1da177e4 | 935 | |
6d048f53 | 936 | return rq; |
1da177e4 LT |
937 | } |
938 | ||
22e2c507 JA |
939 | static inline int |
940 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
941 | { | |
942 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 943 | |
22e2c507 | 944 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 945 | |
22e2c507 | 946 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
947 | } |
948 | ||
22e2c507 | 949 | /* |
498d3aa2 JA |
950 | * Select a queue for service. If we have a current active queue, |
951 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 952 | */ |
1b5ed5e1 | 953 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 954 | { |
1da177e4 | 955 | struct cfq_queue *cfqq; |
1da177e4 | 956 | |
22e2c507 JA |
957 | cfqq = cfqd->active_queue; |
958 | if (!cfqq) | |
959 | goto new_queue; | |
1da177e4 | 960 | |
22e2c507 | 961 | /* |
6d048f53 | 962 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 963 | */ |
6d048f53 | 964 | if (cfq_slice_used(cfqq)) |
3b18152c | 965 | goto expire; |
1da177e4 | 966 | |
22e2c507 | 967 | /* |
6d048f53 JA |
968 | * The active queue has requests and isn't expired, allow it to |
969 | * dispatch. | |
22e2c507 | 970 | */ |
dd67d051 | 971 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 972 | goto keep_queue; |
6d048f53 JA |
973 | |
974 | /* | |
975 | * No requests pending. If the active queue still has requests in | |
976 | * flight or is idling for a new request, allow either of these | |
977 | * conditions to happen (or time out) before selecting a new queue. | |
978 | */ | |
cc197479 JA |
979 | if (timer_pending(&cfqd->idle_slice_timer) || |
980 | (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) { | |
caaa5f9f JA |
981 | cfqq = NULL; |
982 | goto keep_queue; | |
22e2c507 JA |
983 | } |
984 | ||
3b18152c | 985 | expire: |
6084cdda | 986 | cfq_slice_expired(cfqd, 0); |
3b18152c JA |
987 | new_queue: |
988 | cfqq = cfq_set_active_queue(cfqd); | |
22e2c507 | 989 | keep_queue: |
3b18152c | 990 | return cfqq; |
22e2c507 JA |
991 | } |
992 | ||
498d3aa2 JA |
993 | /* |
994 | * Dispatch some requests from cfqq, moving them to the request queue | |
995 | * dispatch list. | |
996 | */ | |
22e2c507 JA |
997 | static int |
998 | __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
999 | int max_dispatch) | |
1000 | { | |
1001 | int dispatched = 0; | |
1002 | ||
dd67d051 | 1003 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); |
22e2c507 JA |
1004 | |
1005 | do { | |
5e705374 | 1006 | struct request *rq; |
1da177e4 LT |
1007 | |
1008 | /* | |
22e2c507 | 1009 | * follow expired path, else get first next available |
1da177e4 | 1010 | */ |
fe094d98 JA |
1011 | rq = cfq_check_fifo(cfqq); |
1012 | if (rq == NULL) | |
5e705374 | 1013 | rq = cfqq->next_rq; |
22e2c507 JA |
1014 | |
1015 | /* | |
1016 | * finally, insert request into driver dispatch list | |
1017 | */ | |
5e705374 | 1018 | cfq_dispatch_insert(cfqd->queue, rq); |
1da177e4 | 1019 | |
22e2c507 | 1020 | dispatched++; |
1da177e4 | 1021 | |
22e2c507 | 1022 | if (!cfqd->active_cic) { |
5e705374 JA |
1023 | atomic_inc(&RQ_CIC(rq)->ioc->refcount); |
1024 | cfqd->active_cic = RQ_CIC(rq); | |
22e2c507 | 1025 | } |
1da177e4 | 1026 | |
dd67d051 | 1027 | if (RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 JA |
1028 | break; |
1029 | ||
1030 | } while (dispatched < max_dispatch); | |
1031 | ||
22e2c507 JA |
1032 | /* |
1033 | * expire an async queue immediately if it has used up its slice. idle | |
1034 | * queue always expire after 1 dispatch round. | |
1035 | */ | |
a9938006 | 1036 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && |
20e493a8 | 1037 | dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) || |
a9938006 | 1038 | cfq_class_idle(cfqq))) { |
44f7c160 | 1039 | cfqq->slice_end = jiffies + 1; |
6084cdda | 1040 | cfq_slice_expired(cfqd, 0); |
44f7c160 | 1041 | } |
22e2c507 JA |
1042 | |
1043 | return dispatched; | |
1044 | } | |
1045 | ||
febffd61 | 1046 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
1047 | { |
1048 | int dispatched = 0; | |
1049 | ||
1050 | while (cfqq->next_rq) { | |
1051 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
1052 | dispatched++; | |
1053 | } | |
1054 | ||
1055 | BUG_ON(!list_empty(&cfqq->fifo)); | |
1056 | return dispatched; | |
1057 | } | |
1058 | ||
498d3aa2 JA |
1059 | /* |
1060 | * Drain our current requests. Used for barriers and when switching | |
1061 | * io schedulers on-the-fly. | |
1062 | */ | |
d9e7620e | 1063 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 1064 | { |
0871714e | 1065 | struct cfq_queue *cfqq; |
d9e7620e | 1066 | int dispatched = 0; |
1b5ed5e1 | 1067 | |
0871714e | 1068 | while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL) |
d9e7620e | 1069 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); |
1b5ed5e1 | 1070 | |
6084cdda | 1071 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
1072 | |
1073 | BUG_ON(cfqd->busy_queues); | |
1074 | ||
1075 | return dispatched; | |
1076 | } | |
1077 | ||
165125e1 | 1078 | static int cfq_dispatch_requests(struct request_queue *q, int force) |
22e2c507 JA |
1079 | { |
1080 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
6d048f53 | 1081 | struct cfq_queue *cfqq; |
caaa5f9f | 1082 | int dispatched; |
22e2c507 JA |
1083 | |
1084 | if (!cfqd->busy_queues) | |
1085 | return 0; | |
1086 | ||
1b5ed5e1 TH |
1087 | if (unlikely(force)) |
1088 | return cfq_forced_dispatch(cfqd); | |
1089 | ||
caaa5f9f | 1090 | dispatched = 0; |
caaa5f9f | 1091 | while ((cfqq = cfq_select_queue(cfqd)) != NULL) { |
b4878f24 JA |
1092 | int max_dispatch; |
1093 | ||
3ed9a296 JA |
1094 | max_dispatch = cfqd->cfq_quantum; |
1095 | if (cfq_class_idle(cfqq)) | |
1096 | max_dispatch = 1; | |
1097 | ||
1098 | if (cfqq->dispatched >= max_dispatch) { | |
1099 | if (cfqd->busy_queues > 1) | |
6d048f53 | 1100 | break; |
3ed9a296 | 1101 | if (cfqq->dispatched >= 4 * max_dispatch) |
a9938006 JA |
1102 | break; |
1103 | } | |
9ede209e | 1104 | |
3ed9a296 JA |
1105 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) |
1106 | break; | |
1107 | ||
3b18152c JA |
1108 | cfq_clear_cfqq_must_dispatch(cfqq); |
1109 | cfq_clear_cfqq_wait_request(cfqq); | |
22e2c507 JA |
1110 | del_timer(&cfqd->idle_slice_timer); |
1111 | ||
caaa5f9f | 1112 | dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch); |
1da177e4 LT |
1113 | } |
1114 | ||
caaa5f9f | 1115 | return dispatched; |
1da177e4 LT |
1116 | } |
1117 | ||
1da177e4 | 1118 | /* |
5e705374 JA |
1119 | * task holds one reference to the queue, dropped when task exits. each rq |
1120 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 LT |
1121 | * |
1122 | * queue lock must be held here. | |
1123 | */ | |
1124 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
1125 | { | |
22e2c507 JA |
1126 | struct cfq_data *cfqd = cfqq->cfqd; |
1127 | ||
1128 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
1129 | |
1130 | if (!atomic_dec_and_test(&cfqq->ref)) | |
1131 | return; | |
1132 | ||
1133 | BUG_ON(rb_first(&cfqq->sort_list)); | |
22e2c507 | 1134 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
3b18152c | 1135 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 | 1136 | |
28f95cbc | 1137 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 1138 | __cfq_slice_expired(cfqd, cfqq, 0); |
28f95cbc JA |
1139 | cfq_schedule_dispatch(cfqd); |
1140 | } | |
22e2c507 | 1141 | |
1da177e4 LT |
1142 | kmem_cache_free(cfq_pool, cfqq); |
1143 | } | |
1144 | ||
4ac845a2 JA |
1145 | /* |
1146 | * Call func for each cic attached to this ioc. Returns number of cic's seen. | |
1147 | */ | |
1148 | #define CIC_GANG_NR 16 | |
1149 | static unsigned int | |
1150 | call_for_each_cic(struct io_context *ioc, | |
1151 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1da177e4 | 1152 | { |
4ac845a2 JA |
1153 | struct cfq_io_context *cics[CIC_GANG_NR]; |
1154 | unsigned long index = 0; | |
1155 | unsigned int called = 0; | |
1156 | int nr; | |
1da177e4 | 1157 | |
4ac845a2 | 1158 | rcu_read_lock(); |
597bc485 | 1159 | |
4ac845a2 JA |
1160 | do { |
1161 | int i; | |
66dac98e | 1162 | |
4ac845a2 JA |
1163 | /* |
1164 | * Perhaps there's a better way - this just gang lookups from | |
1165 | * 0 to the end, restarting after each CIC_GANG_NR from the | |
1166 | * last key + 1. | |
1167 | */ | |
1168 | nr = radix_tree_gang_lookup(&ioc->radix_root, (void **) cics, | |
1169 | index, CIC_GANG_NR); | |
1170 | if (!nr) | |
1171 | break; | |
1172 | ||
1173 | called += nr; | |
1174 | index = 1 + (unsigned long) cics[nr - 1]->key; | |
1175 | ||
1176 | for (i = 0; i < nr; i++) | |
1177 | func(ioc, cics[i]); | |
1178 | } while (nr == CIC_GANG_NR); | |
1179 | ||
1180 | rcu_read_unlock(); | |
1181 | ||
1182 | return called; | |
1183 | } | |
1184 | ||
1185 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
1186 | { | |
1187 | unsigned long flags; | |
1188 | ||
1189 | BUG_ON(!cic->dead_key); | |
1190 | ||
1191 | spin_lock_irqsave(&ioc->lock, flags); | |
1192 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
1193 | spin_unlock_irqrestore(&ioc->lock, flags); | |
1194 | ||
1195 | kmem_cache_free(cfq_ioc_pool, cic); | |
1196 | } | |
1197 | ||
1198 | static void cfq_free_io_context(struct io_context *ioc) | |
1199 | { | |
1200 | int freed; | |
1201 | ||
1202 | /* | |
1203 | * ioc->refcount is zero here, so no more cic's are allowed to be | |
1204 | * linked into this ioc. So it should be ok to iterate over the known | |
1205 | * list, we will see all cic's since no new ones are added. | |
1206 | */ | |
1207 | freed = call_for_each_cic(ioc, cic_free_func); | |
1da177e4 | 1208 | |
4050cf16 JA |
1209 | elv_ioc_count_mod(ioc_count, -freed); |
1210 | ||
1211 | if (ioc_gone && !elv_ioc_count_read(ioc_count)) | |
334e94de | 1212 | complete(ioc_gone); |
1da177e4 LT |
1213 | } |
1214 | ||
89850f7e | 1215 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1216 | { |
28f95cbc | 1217 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 1218 | __cfq_slice_expired(cfqd, cfqq, 0); |
28f95cbc JA |
1219 | cfq_schedule_dispatch(cfqd); |
1220 | } | |
22e2c507 | 1221 | |
89850f7e JA |
1222 | cfq_put_queue(cfqq); |
1223 | } | |
22e2c507 | 1224 | |
89850f7e JA |
1225 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
1226 | struct cfq_io_context *cic) | |
1227 | { | |
fc46379d | 1228 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
1229 | |
1230 | /* | |
1231 | * Make sure key == NULL is seen for dead queues | |
1232 | */ | |
fc46379d | 1233 | smp_wmb(); |
4ac845a2 | 1234 | cic->dead_key = (unsigned long) cic->key; |
fc46379d JA |
1235 | cic->key = NULL; |
1236 | ||
12a05732 | 1237 | if (cic->cfqq[ASYNC]) { |
89850f7e | 1238 | cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]); |
12a05732 AV |
1239 | cic->cfqq[ASYNC] = NULL; |
1240 | } | |
1241 | ||
1242 | if (cic->cfqq[SYNC]) { | |
89850f7e | 1243 | cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]); |
12a05732 AV |
1244 | cic->cfqq[SYNC] = NULL; |
1245 | } | |
89850f7e JA |
1246 | } |
1247 | ||
4ac845a2 JA |
1248 | static void cfq_exit_single_io_context(struct io_context *ioc, |
1249 | struct cfq_io_context *cic) | |
89850f7e JA |
1250 | { |
1251 | struct cfq_data *cfqd = cic->key; | |
1252 | ||
89850f7e | 1253 | if (cfqd) { |
165125e1 | 1254 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 1255 | unsigned long flags; |
89850f7e | 1256 | |
4ac845a2 | 1257 | spin_lock_irqsave(q->queue_lock, flags); |
89850f7e | 1258 | __cfq_exit_single_io_context(cfqd, cic); |
4ac845a2 | 1259 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 1260 | } |
1da177e4 LT |
1261 | } |
1262 | ||
498d3aa2 JA |
1263 | /* |
1264 | * The process that ioc belongs to has exited, we need to clean up | |
1265 | * and put the internal structures we have that belongs to that process. | |
1266 | */ | |
e2d74ac0 | 1267 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 1268 | { |
4ac845a2 JA |
1269 | rcu_assign_pointer(ioc->ioc_data, NULL); |
1270 | call_for_each_cic(ioc, cfq_exit_single_io_context); | |
1da177e4 LT |
1271 | } |
1272 | ||
22e2c507 | 1273 | static struct cfq_io_context * |
8267e268 | 1274 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1275 | { |
b5deef90 | 1276 | struct cfq_io_context *cic; |
1da177e4 | 1277 | |
94f6030c CL |
1278 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
1279 | cfqd->queue->node); | |
1da177e4 | 1280 | if (cic) { |
22e2c507 | 1281 | cic->last_end_request = jiffies; |
553698f9 | 1282 | INIT_LIST_HEAD(&cic->queue_list); |
22e2c507 JA |
1283 | cic->dtor = cfq_free_io_context; |
1284 | cic->exit = cfq_exit_io_context; | |
4050cf16 | 1285 | elv_ioc_count_inc(ioc_count); |
1da177e4 LT |
1286 | } |
1287 | ||
1288 | return cic; | |
1289 | } | |
1290 | ||
fd0928df | 1291 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
1292 | { |
1293 | struct task_struct *tsk = current; | |
1294 | int ioprio_class; | |
1295 | ||
3b18152c | 1296 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
1297 | return; |
1298 | ||
fd0928df | 1299 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 1300 | switch (ioprio_class) { |
fe094d98 JA |
1301 | default: |
1302 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
1303 | case IOPRIO_CLASS_NONE: | |
1304 | /* | |
1305 | * no prio set, place us in the middle of the BE classes | |
1306 | */ | |
1307 | cfqq->ioprio = task_nice_ioprio(tsk); | |
1308 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
1309 | break; | |
1310 | case IOPRIO_CLASS_RT: | |
1311 | cfqq->ioprio = task_ioprio(ioc); | |
1312 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
1313 | break; | |
1314 | case IOPRIO_CLASS_BE: | |
1315 | cfqq->ioprio = task_ioprio(ioc); | |
1316 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
1317 | break; | |
1318 | case IOPRIO_CLASS_IDLE: | |
1319 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
1320 | cfqq->ioprio = 7; | |
1321 | cfq_clear_cfqq_idle_window(cfqq); | |
1322 | break; | |
22e2c507 JA |
1323 | } |
1324 | ||
1325 | /* | |
1326 | * keep track of original prio settings in case we have to temporarily | |
1327 | * elevate the priority of this queue | |
1328 | */ | |
1329 | cfqq->org_ioprio = cfqq->ioprio; | |
1330 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 1331 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
1332 | } |
1333 | ||
febffd61 | 1334 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 1335 | { |
478a82b0 AV |
1336 | struct cfq_data *cfqd = cic->key; |
1337 | struct cfq_queue *cfqq; | |
c1b707d2 | 1338 | unsigned long flags; |
35e6077c | 1339 | |
caaa5f9f JA |
1340 | if (unlikely(!cfqd)) |
1341 | return; | |
1342 | ||
c1b707d2 | 1343 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f JA |
1344 | |
1345 | cfqq = cic->cfqq[ASYNC]; | |
1346 | if (cfqq) { | |
1347 | struct cfq_queue *new_cfqq; | |
fd0928df | 1348 | new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC); |
caaa5f9f JA |
1349 | if (new_cfqq) { |
1350 | cic->cfqq[ASYNC] = new_cfqq; | |
1351 | cfq_put_queue(cfqq); | |
1352 | } | |
22e2c507 | 1353 | } |
caaa5f9f JA |
1354 | |
1355 | cfqq = cic->cfqq[SYNC]; | |
1356 | if (cfqq) | |
1357 | cfq_mark_cfqq_prio_changed(cfqq); | |
1358 | ||
c1b707d2 | 1359 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
1360 | } |
1361 | ||
fc46379d | 1362 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 1363 | { |
4ac845a2 | 1364 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 1365 | ioc->ioprio_changed = 0; |
22e2c507 JA |
1366 | } |
1367 | ||
1368 | static struct cfq_queue * | |
15c31be4 | 1369 | cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync, |
fd0928df | 1370 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 1371 | { |
22e2c507 | 1372 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 1373 | struct cfq_io_context *cic; |
22e2c507 JA |
1374 | |
1375 | retry: | |
4ac845a2 | 1376 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
1377 | /* cic always exists here */ |
1378 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 JA |
1379 | |
1380 | if (!cfqq) { | |
1381 | if (new_cfqq) { | |
1382 | cfqq = new_cfqq; | |
1383 | new_cfqq = NULL; | |
1384 | } else if (gfp_mask & __GFP_WAIT) { | |
89850f7e JA |
1385 | /* |
1386 | * Inform the allocator of the fact that we will | |
1387 | * just repeat this allocation if it fails, to allow | |
1388 | * the allocator to do whatever it needs to attempt to | |
1389 | * free memory. | |
1390 | */ | |
22e2c507 | 1391 | spin_unlock_irq(cfqd->queue->queue_lock); |
94f6030c CL |
1392 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
1393 | gfp_mask | __GFP_NOFAIL | __GFP_ZERO, | |
1394 | cfqd->queue->node); | |
22e2c507 JA |
1395 | spin_lock_irq(cfqd->queue->queue_lock); |
1396 | goto retry; | |
1397 | } else { | |
94f6030c CL |
1398 | cfqq = kmem_cache_alloc_node(cfq_pool, |
1399 | gfp_mask | __GFP_ZERO, | |
1400 | cfqd->queue->node); | |
22e2c507 JA |
1401 | if (!cfqq) |
1402 | goto out; | |
1403 | } | |
1404 | ||
d9e7620e | 1405 | RB_CLEAR_NODE(&cfqq->rb_node); |
22e2c507 JA |
1406 | INIT_LIST_HEAD(&cfqq->fifo); |
1407 | ||
22e2c507 JA |
1408 | atomic_set(&cfqq->ref, 0); |
1409 | cfqq->cfqd = cfqd; | |
c5b680f3 | 1410 | |
3b18152c | 1411 | cfq_mark_cfqq_prio_changed(cfqq); |
53b03744 | 1412 | cfq_mark_cfqq_queue_new(cfqq); |
91fac317 | 1413 | |
fd0928df | 1414 | cfq_init_prio_data(cfqq, ioc); |
0871714e JA |
1415 | |
1416 | if (is_sync) { | |
1417 | if (!cfq_class_idle(cfqq)) | |
1418 | cfq_mark_cfqq_idle_window(cfqq); | |
1419 | cfq_mark_cfqq_sync(cfqq); | |
1420 | } | |
22e2c507 JA |
1421 | } |
1422 | ||
1423 | if (new_cfqq) | |
1424 | kmem_cache_free(cfq_pool, new_cfqq); | |
1425 | ||
22e2c507 JA |
1426 | out: |
1427 | WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq); | |
1428 | return cfqq; | |
1429 | } | |
1430 | ||
c2dea2d1 VT |
1431 | static struct cfq_queue ** |
1432 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
1433 | { | |
fe094d98 | 1434 | switch (ioprio_class) { |
c2dea2d1 VT |
1435 | case IOPRIO_CLASS_RT: |
1436 | return &cfqd->async_cfqq[0][ioprio]; | |
1437 | case IOPRIO_CLASS_BE: | |
1438 | return &cfqd->async_cfqq[1][ioprio]; | |
1439 | case IOPRIO_CLASS_IDLE: | |
1440 | return &cfqd->async_idle_cfqq; | |
1441 | default: | |
1442 | BUG(); | |
1443 | } | |
1444 | } | |
1445 | ||
15c31be4 | 1446 | static struct cfq_queue * |
fd0928df | 1447 | cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc, |
15c31be4 JA |
1448 | gfp_t gfp_mask) |
1449 | { | |
fd0928df JA |
1450 | const int ioprio = task_ioprio(ioc); |
1451 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 1452 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
1453 | struct cfq_queue *cfqq = NULL; |
1454 | ||
c2dea2d1 VT |
1455 | if (!is_sync) { |
1456 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
1457 | cfqq = *async_cfqq; | |
1458 | } | |
1459 | ||
0a0836a0 | 1460 | if (!cfqq) { |
fd0928df | 1461 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
0a0836a0 ON |
1462 | if (!cfqq) |
1463 | return NULL; | |
1464 | } | |
15c31be4 JA |
1465 | |
1466 | /* | |
1467 | * pin the queue now that it's allocated, scheduler exit will prune it | |
1468 | */ | |
c2dea2d1 | 1469 | if (!is_sync && !(*async_cfqq)) { |
15c31be4 | 1470 | atomic_inc(&cfqq->ref); |
c2dea2d1 | 1471 | *async_cfqq = cfqq; |
15c31be4 JA |
1472 | } |
1473 | ||
1474 | atomic_inc(&cfqq->ref); | |
1475 | return cfqq; | |
1476 | } | |
1477 | ||
4ac845a2 JA |
1478 | static void cfq_cic_free(struct cfq_io_context *cic) |
1479 | { | |
1480 | kmem_cache_free(cfq_ioc_pool, cic); | |
1481 | elv_ioc_count_dec(ioc_count); | |
1482 | ||
1483 | if (ioc_gone && !elv_ioc_count_read(ioc_count)) | |
1484 | complete(ioc_gone); | |
1485 | } | |
1486 | ||
498d3aa2 JA |
1487 | /* |
1488 | * We drop cfq io contexts lazily, so we may find a dead one. | |
1489 | */ | |
dbecf3ab | 1490 | static void |
4ac845a2 JA |
1491 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
1492 | struct cfq_io_context *cic) | |
dbecf3ab | 1493 | { |
4ac845a2 JA |
1494 | unsigned long flags; |
1495 | ||
fc46379d | 1496 | WARN_ON(!list_empty(&cic->queue_list)); |
597bc485 | 1497 | |
4ac845a2 JA |
1498 | spin_lock_irqsave(&ioc->lock, flags); |
1499 | ||
597bc485 | 1500 | if (ioc->ioc_data == cic) |
4ac845a2 | 1501 | rcu_assign_pointer(ioc->ioc_data, NULL); |
597bc485 | 1502 | |
4ac845a2 JA |
1503 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); |
1504 | spin_unlock_irqrestore(&ioc->lock, flags); | |
1505 | ||
1506 | cfq_cic_free(cic); | |
dbecf3ab OH |
1507 | } |
1508 | ||
e2d74ac0 | 1509 | static struct cfq_io_context * |
4ac845a2 | 1510 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 1511 | { |
e2d74ac0 | 1512 | struct cfq_io_context *cic; |
4ac845a2 | 1513 | void *k; |
e2d74ac0 | 1514 | |
91fac317 VT |
1515 | if (unlikely(!ioc)) |
1516 | return NULL; | |
1517 | ||
597bc485 JA |
1518 | /* |
1519 | * we maintain a last-hit cache, to avoid browsing over the tree | |
1520 | */ | |
4ac845a2 | 1521 | cic = rcu_dereference(ioc->ioc_data); |
597bc485 JA |
1522 | if (cic && cic->key == cfqd) |
1523 | return cic; | |
1524 | ||
4ac845a2 JA |
1525 | do { |
1526 | rcu_read_lock(); | |
1527 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); | |
1528 | rcu_read_unlock(); | |
1529 | if (!cic) | |
1530 | break; | |
be3b0753 OH |
1531 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
1532 | k = cic->key; | |
1533 | if (unlikely(!k)) { | |
4ac845a2 JA |
1534 | cfq_drop_dead_cic(cfqd, ioc, cic); |
1535 | continue; | |
dbecf3ab | 1536 | } |
e2d74ac0 | 1537 | |
4ac845a2 JA |
1538 | rcu_assign_pointer(ioc->ioc_data, cic); |
1539 | break; | |
1540 | } while (1); | |
e2d74ac0 | 1541 | |
4ac845a2 | 1542 | return cic; |
e2d74ac0 JA |
1543 | } |
1544 | ||
4ac845a2 JA |
1545 | /* |
1546 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
1547 | * the process specific cfq io context when entered from the block layer. | |
1548 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
1549 | */ | |
febffd61 JA |
1550 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
1551 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 1552 | { |
0261d688 | 1553 | unsigned long flags; |
4ac845a2 | 1554 | int ret; |
e2d74ac0 | 1555 | |
4ac845a2 JA |
1556 | ret = radix_tree_preload(gfp_mask); |
1557 | if (!ret) { | |
1558 | cic->ioc = ioc; | |
1559 | cic->key = cfqd; | |
e2d74ac0 | 1560 | |
4ac845a2 JA |
1561 | spin_lock_irqsave(&ioc->lock, flags); |
1562 | ret = radix_tree_insert(&ioc->radix_root, | |
1563 | (unsigned long) cfqd, cic); | |
1564 | spin_unlock_irqrestore(&ioc->lock, flags); | |
e2d74ac0 | 1565 | |
4ac845a2 JA |
1566 | radix_tree_preload_end(); |
1567 | ||
1568 | if (!ret) { | |
1569 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
1570 | list_add(&cic->queue_list, &cfqd->cic_list); | |
1571 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
1572 | } | |
e2d74ac0 JA |
1573 | } |
1574 | ||
4ac845a2 JA |
1575 | if (ret) |
1576 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 1577 | |
4ac845a2 | 1578 | return ret; |
e2d74ac0 JA |
1579 | } |
1580 | ||
1da177e4 LT |
1581 | /* |
1582 | * Setup general io context and cfq io context. There can be several cfq | |
1583 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 1584 | * than one device managed by cfq. |
1da177e4 LT |
1585 | */ |
1586 | static struct cfq_io_context * | |
e2d74ac0 | 1587 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1588 | { |
22e2c507 | 1589 | struct io_context *ioc = NULL; |
1da177e4 | 1590 | struct cfq_io_context *cic; |
1da177e4 | 1591 | |
22e2c507 | 1592 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 1593 | |
b5deef90 | 1594 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
1595 | if (!ioc) |
1596 | return NULL; | |
1597 | ||
4ac845a2 | 1598 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
1599 | if (cic) |
1600 | goto out; | |
1da177e4 | 1601 | |
e2d74ac0 JA |
1602 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
1603 | if (cic == NULL) | |
1604 | goto err; | |
1da177e4 | 1605 | |
4ac845a2 JA |
1606 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
1607 | goto err_free; | |
1608 | ||
1da177e4 | 1609 | out: |
fc46379d JA |
1610 | smp_read_barrier_depends(); |
1611 | if (unlikely(ioc->ioprio_changed)) | |
1612 | cfq_ioc_set_ioprio(ioc); | |
1613 | ||
1da177e4 | 1614 | return cic; |
4ac845a2 JA |
1615 | err_free: |
1616 | cfq_cic_free(cic); | |
1da177e4 LT |
1617 | err: |
1618 | put_io_context(ioc); | |
1619 | return NULL; | |
1620 | } | |
1621 | ||
22e2c507 JA |
1622 | static void |
1623 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 1624 | { |
aaf1228d JA |
1625 | unsigned long elapsed = jiffies - cic->last_end_request; |
1626 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 1627 | |
22e2c507 JA |
1628 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
1629 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
1630 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
1631 | } | |
1da177e4 | 1632 | |
206dc69b | 1633 | static void |
6d048f53 JA |
1634 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic, |
1635 | struct request *rq) | |
206dc69b JA |
1636 | { |
1637 | sector_t sdist; | |
1638 | u64 total; | |
1639 | ||
5e705374 JA |
1640 | if (cic->last_request_pos < rq->sector) |
1641 | sdist = rq->sector - cic->last_request_pos; | |
206dc69b | 1642 | else |
5e705374 | 1643 | sdist = cic->last_request_pos - rq->sector; |
206dc69b JA |
1644 | |
1645 | /* | |
1646 | * Don't allow the seek distance to get too large from the | |
1647 | * odd fragment, pagein, etc | |
1648 | */ | |
1649 | if (cic->seek_samples <= 60) /* second&third seek */ | |
1650 | sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024); | |
1651 | else | |
1652 | sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64); | |
1653 | ||
1654 | cic->seek_samples = (7*cic->seek_samples + 256) / 8; | |
1655 | cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8; | |
1656 | total = cic->seek_total + (cic->seek_samples/2); | |
1657 | do_div(total, cic->seek_samples); | |
1658 | cic->seek_mean = (sector_t)total; | |
1659 | } | |
1da177e4 | 1660 | |
22e2c507 JA |
1661 | /* |
1662 | * Disable idle window if the process thinks too long or seeks so much that | |
1663 | * it doesn't matter | |
1664 | */ | |
1665 | static void | |
1666 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1667 | struct cfq_io_context *cic) | |
1668 | { | |
1be92f2f JA |
1669 | int enable_idle; |
1670 | ||
0871714e JA |
1671 | /* |
1672 | * Don't idle for async or idle io prio class | |
1673 | */ | |
1674 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
1675 | return; |
1676 | ||
1677 | enable_idle = cfq_cfqq_idle_window(cfqq); | |
1da177e4 | 1678 | |
66dac98e | 1679 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || |
caaa5f9f | 1680 | (cfqd->hw_tag && CIC_SEEKY(cic))) |
22e2c507 JA |
1681 | enable_idle = 0; |
1682 | else if (sample_valid(cic->ttime_samples)) { | |
1683 | if (cic->ttime_mean > cfqd->cfq_slice_idle) | |
1684 | enable_idle = 0; | |
1685 | else | |
1686 | enable_idle = 1; | |
1da177e4 LT |
1687 | } |
1688 | ||
3b18152c JA |
1689 | if (enable_idle) |
1690 | cfq_mark_cfqq_idle_window(cfqq); | |
1691 | else | |
1692 | cfq_clear_cfqq_idle_window(cfqq); | |
22e2c507 | 1693 | } |
1da177e4 | 1694 | |
22e2c507 JA |
1695 | /* |
1696 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
1697 | * no or if we aren't sure, a 1 will cause a preempt. | |
1698 | */ | |
1699 | static int | |
1700 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, | |
5e705374 | 1701 | struct request *rq) |
22e2c507 | 1702 | { |
6d048f53 | 1703 | struct cfq_queue *cfqq; |
22e2c507 | 1704 | |
6d048f53 JA |
1705 | cfqq = cfqd->active_queue; |
1706 | if (!cfqq) | |
22e2c507 JA |
1707 | return 0; |
1708 | ||
6d048f53 JA |
1709 | if (cfq_slice_used(cfqq)) |
1710 | return 1; | |
1711 | ||
1712 | if (cfq_class_idle(new_cfqq)) | |
caaa5f9f | 1713 | return 0; |
22e2c507 JA |
1714 | |
1715 | if (cfq_class_idle(cfqq)) | |
1716 | return 1; | |
1e3335de | 1717 | |
374f84ac JA |
1718 | /* |
1719 | * if the new request is sync, but the currently running queue is | |
1720 | * not, let the sync request have priority. | |
1721 | */ | |
5e705374 | 1722 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
22e2c507 | 1723 | return 1; |
1e3335de | 1724 | |
374f84ac JA |
1725 | /* |
1726 | * So both queues are sync. Let the new request get disk time if | |
1727 | * it's a metadata request and the current queue is doing regular IO. | |
1728 | */ | |
1729 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
1730 | return 1; | |
22e2c507 | 1731 | |
1e3335de JA |
1732 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
1733 | return 0; | |
1734 | ||
1735 | /* | |
1736 | * if this request is as-good as one we would expect from the | |
1737 | * current cfqq, let it preempt | |
1738 | */ | |
6d048f53 | 1739 | if (cfq_rq_close(cfqd, rq)) |
1e3335de JA |
1740 | return 1; |
1741 | ||
22e2c507 JA |
1742 | return 0; |
1743 | } | |
1744 | ||
1745 | /* | |
1746 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
1747 | * let it have half of its nominal slice. | |
1748 | */ | |
1749 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1750 | { | |
6084cdda | 1751 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 1752 | |
bf572256 JA |
1753 | /* |
1754 | * Put the new queue at the front of the of the current list, | |
1755 | * so we know that it will be selected next. | |
1756 | */ | |
1757 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
1758 | |
1759 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 1760 | |
44f7c160 JA |
1761 | cfqq->slice_end = 0; |
1762 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
1763 | } |
1764 | ||
22e2c507 | 1765 | /* |
5e705374 | 1766 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
1767 | * something we should do about it |
1768 | */ | |
1769 | static void | |
5e705374 JA |
1770 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
1771 | struct request *rq) | |
22e2c507 | 1772 | { |
5e705374 | 1773 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 1774 | |
374f84ac JA |
1775 | if (rq_is_meta(rq)) |
1776 | cfqq->meta_pending++; | |
1777 | ||
9c2c38a1 | 1778 | cfq_update_io_thinktime(cfqd, cic); |
6d048f53 | 1779 | cfq_update_io_seektime(cfqd, cic, rq); |
9c2c38a1 JA |
1780 | cfq_update_idle_window(cfqd, cfqq, cic); |
1781 | ||
5e705374 | 1782 | cic->last_request_pos = rq->sector + rq->nr_sectors; |
22e2c507 JA |
1783 | |
1784 | if (cfqq == cfqd->active_queue) { | |
1785 | /* | |
1786 | * if we are waiting for a request for this queue, let it rip | |
1787 | * immediately and flag that we must not expire this queue | |
1788 | * just now | |
1789 | */ | |
3b18152c JA |
1790 | if (cfq_cfqq_wait_request(cfqq)) { |
1791 | cfq_mark_cfqq_must_dispatch(cfqq); | |
22e2c507 | 1792 | del_timer(&cfqd->idle_slice_timer); |
dc72ef4a | 1793 | blk_start_queueing(cfqd->queue); |
22e2c507 | 1794 | } |
5e705374 | 1795 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
1796 | /* |
1797 | * not the active queue - expire current slice if it is | |
1798 | * idle and has expired it's mean thinktime or this new queue | |
1799 | * has some old slice time left and is of higher priority | |
1800 | */ | |
1801 | cfq_preempt_queue(cfqd, cfqq); | |
3b18152c | 1802 | cfq_mark_cfqq_must_dispatch(cfqq); |
dc72ef4a | 1803 | blk_start_queueing(cfqd->queue); |
22e2c507 | 1804 | } |
1da177e4 LT |
1805 | } |
1806 | ||
165125e1 | 1807 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1808 | { |
b4878f24 | 1809 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1810 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 1811 | |
fd0928df | 1812 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 1813 | |
5e705374 | 1814 | cfq_add_rq_rb(rq); |
1da177e4 | 1815 | |
22e2c507 JA |
1816 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
1817 | ||
5e705374 | 1818 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
1819 | } |
1820 | ||
165125e1 | 1821 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1822 | { |
5e705374 | 1823 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 1824 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 1825 | const int sync = rq_is_sync(rq); |
b4878f24 | 1826 | unsigned long now; |
1da177e4 | 1827 | |
b4878f24 | 1828 | now = jiffies; |
1da177e4 | 1829 | |
b4878f24 | 1830 | WARN_ON(!cfqd->rq_in_driver); |
6d048f53 | 1831 | WARN_ON(!cfqq->dispatched); |
b4878f24 | 1832 | cfqd->rq_in_driver--; |
6d048f53 | 1833 | cfqq->dispatched--; |
1da177e4 | 1834 | |
3ed9a296 JA |
1835 | if (cfq_cfqq_sync(cfqq)) |
1836 | cfqd->sync_flight--; | |
1837 | ||
b4878f24 JA |
1838 | if (!cfq_class_idle(cfqq)) |
1839 | cfqd->last_end_request = now; | |
3b18152c | 1840 | |
caaa5f9f | 1841 | if (sync) |
5e705374 | 1842 | RQ_CIC(rq)->last_end_request = now; |
caaa5f9f JA |
1843 | |
1844 | /* | |
1845 | * If this is the active queue, check if it needs to be expired, | |
1846 | * or if we want to idle in case it has no pending requests. | |
1847 | */ | |
1848 | if (cfqd->active_queue == cfqq) { | |
44f7c160 JA |
1849 | if (cfq_cfqq_slice_new(cfqq)) { |
1850 | cfq_set_prio_slice(cfqd, cfqq); | |
1851 | cfq_clear_cfqq_slice_new(cfqq); | |
1852 | } | |
0871714e | 1853 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
6084cdda | 1854 | cfq_slice_expired(cfqd, 1); |
6d048f53 JA |
1855 | else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) |
1856 | cfq_arm_slice_timer(cfqd); | |
caaa5f9f | 1857 | } |
6d048f53 JA |
1858 | |
1859 | if (!cfqd->rq_in_driver) | |
1860 | cfq_schedule_dispatch(cfqd); | |
1da177e4 LT |
1861 | } |
1862 | ||
22e2c507 JA |
1863 | /* |
1864 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
1865 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
1866 | */ | |
1867 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 1868 | { |
22e2c507 JA |
1869 | if (has_fs_excl()) { |
1870 | /* | |
1871 | * boost idle prio on transactions that would lock out other | |
1872 | * users of the filesystem | |
1873 | */ | |
1874 | if (cfq_class_idle(cfqq)) | |
1875 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
1876 | if (cfqq->ioprio > IOPRIO_NORM) | |
1877 | cfqq->ioprio = IOPRIO_NORM; | |
1878 | } else { | |
1879 | /* | |
1880 | * check if we need to unboost the queue | |
1881 | */ | |
1882 | if (cfqq->ioprio_class != cfqq->org_ioprio_class) | |
1883 | cfqq->ioprio_class = cfqq->org_ioprio_class; | |
1884 | if (cfqq->ioprio != cfqq->org_ioprio) | |
1885 | cfqq->ioprio = cfqq->org_ioprio; | |
1886 | } | |
22e2c507 | 1887 | } |
1da177e4 | 1888 | |
89850f7e | 1889 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 1890 | { |
3b18152c | 1891 | if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) && |
99f95e52 | 1892 | !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 1893 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 1894 | return ELV_MQUEUE_MUST; |
3b18152c | 1895 | } |
1da177e4 | 1896 | |
22e2c507 | 1897 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
1898 | } |
1899 | ||
165125e1 | 1900 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
1901 | { |
1902 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1903 | struct task_struct *tsk = current; | |
91fac317 | 1904 | struct cfq_io_context *cic; |
22e2c507 JA |
1905 | struct cfq_queue *cfqq; |
1906 | ||
1907 | /* | |
1908 | * don't force setup of a queue from here, as a call to may_queue | |
1909 | * does not necessarily imply that a request actually will be queued. | |
1910 | * so just lookup a possibly existing queue, or return 'may queue' | |
1911 | * if that fails | |
1912 | */ | |
4ac845a2 | 1913 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
1914 | if (!cic) |
1915 | return ELV_MQUEUE_MAY; | |
1916 | ||
1917 | cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC); | |
22e2c507 | 1918 | if (cfqq) { |
fd0928df | 1919 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
1920 | cfq_prio_boost(cfqq); |
1921 | ||
89850f7e | 1922 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
1923 | } |
1924 | ||
1925 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
1926 | } |
1927 | ||
1da177e4 LT |
1928 | /* |
1929 | * queue lock held here | |
1930 | */ | |
bb37b94c | 1931 | static void cfq_put_request(struct request *rq) |
1da177e4 | 1932 | { |
5e705374 | 1933 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 1934 | |
5e705374 | 1935 | if (cfqq) { |
22e2c507 | 1936 | const int rw = rq_data_dir(rq); |
1da177e4 | 1937 | |
22e2c507 JA |
1938 | BUG_ON(!cfqq->allocated[rw]); |
1939 | cfqq->allocated[rw]--; | |
1da177e4 | 1940 | |
5e705374 | 1941 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 1942 | |
1da177e4 | 1943 | rq->elevator_private = NULL; |
5e705374 | 1944 | rq->elevator_private2 = NULL; |
1da177e4 | 1945 | |
1da177e4 LT |
1946 | cfq_put_queue(cfqq); |
1947 | } | |
1948 | } | |
1949 | ||
1950 | /* | |
22e2c507 | 1951 | * Allocate cfq data structures associated with this request. |
1da177e4 | 1952 | */ |
22e2c507 | 1953 | static int |
165125e1 | 1954 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
1955 | { |
1956 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1957 | struct cfq_io_context *cic; | |
1958 | const int rw = rq_data_dir(rq); | |
7749a8d4 | 1959 | const int is_sync = rq_is_sync(rq); |
22e2c507 | 1960 | struct cfq_queue *cfqq; |
1da177e4 LT |
1961 | unsigned long flags; |
1962 | ||
1963 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
1964 | ||
e2d74ac0 | 1965 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 1966 | |
1da177e4 LT |
1967 | spin_lock_irqsave(q->queue_lock, flags); |
1968 | ||
22e2c507 JA |
1969 | if (!cic) |
1970 | goto queue_fail; | |
1971 | ||
91fac317 VT |
1972 | cfqq = cic_to_cfqq(cic, is_sync); |
1973 | if (!cfqq) { | |
fd0928df | 1974 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 | 1975 | |
22e2c507 JA |
1976 | if (!cfqq) |
1977 | goto queue_fail; | |
1da177e4 | 1978 | |
91fac317 VT |
1979 | cic_set_cfqq(cic, cfqq, is_sync); |
1980 | } | |
1da177e4 LT |
1981 | |
1982 | cfqq->allocated[rw]++; | |
3b18152c | 1983 | cfq_clear_cfqq_must_alloc(cfqq); |
22e2c507 | 1984 | atomic_inc(&cfqq->ref); |
1da177e4 | 1985 | |
5e705374 | 1986 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 1987 | |
5e705374 JA |
1988 | rq->elevator_private = cic; |
1989 | rq->elevator_private2 = cfqq; | |
1990 | return 0; | |
1da177e4 | 1991 | |
22e2c507 JA |
1992 | queue_fail: |
1993 | if (cic) | |
1994 | put_io_context(cic->ioc); | |
89850f7e | 1995 | |
3b18152c | 1996 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
1997 | spin_unlock_irqrestore(q->queue_lock, flags); |
1998 | return 1; | |
1999 | } | |
2000 | ||
65f27f38 | 2001 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 2002 | { |
65f27f38 DH |
2003 | struct cfq_data *cfqd = |
2004 | container_of(work, struct cfq_data, unplug_work); | |
165125e1 | 2005 | struct request_queue *q = cfqd->queue; |
22e2c507 JA |
2006 | unsigned long flags; |
2007 | ||
2008 | spin_lock_irqsave(q->queue_lock, flags); | |
dc72ef4a | 2009 | blk_start_queueing(q); |
22e2c507 JA |
2010 | spin_unlock_irqrestore(q->queue_lock, flags); |
2011 | } | |
2012 | ||
2013 | /* | |
2014 | * Timer running if the active_queue is currently idling inside its time slice | |
2015 | */ | |
2016 | static void cfq_idle_slice_timer(unsigned long data) | |
2017 | { | |
2018 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
2019 | struct cfq_queue *cfqq; | |
2020 | unsigned long flags; | |
3c6bd2f8 | 2021 | int timed_out = 1; |
22e2c507 JA |
2022 | |
2023 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
2024 | ||
fe094d98 JA |
2025 | cfqq = cfqd->active_queue; |
2026 | if (cfqq) { | |
3c6bd2f8 JA |
2027 | timed_out = 0; |
2028 | ||
22e2c507 JA |
2029 | /* |
2030 | * expired | |
2031 | */ | |
44f7c160 | 2032 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
2033 | goto expire; |
2034 | ||
2035 | /* | |
2036 | * only expire and reinvoke request handler, if there are | |
2037 | * other queues with pending requests | |
2038 | */ | |
caaa5f9f | 2039 | if (!cfqd->busy_queues) |
22e2c507 | 2040 | goto out_cont; |
22e2c507 JA |
2041 | |
2042 | /* | |
2043 | * not expired and it has a request pending, let it dispatch | |
2044 | */ | |
dd67d051 | 2045 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) { |
3b18152c | 2046 | cfq_mark_cfqq_must_dispatch(cfqq); |
22e2c507 JA |
2047 | goto out_kick; |
2048 | } | |
2049 | } | |
2050 | expire: | |
6084cdda | 2051 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 2052 | out_kick: |
3b18152c | 2053 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
2054 | out_cont: |
2055 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2056 | } | |
2057 | ||
3b18152c JA |
2058 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
2059 | { | |
2060 | del_timer_sync(&cfqd->idle_slice_timer); | |
4310864b | 2061 | kblockd_flush_work(&cfqd->unplug_work); |
3b18152c | 2062 | } |
22e2c507 | 2063 | |
c2dea2d1 VT |
2064 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
2065 | { | |
2066 | int i; | |
2067 | ||
2068 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
2069 | if (cfqd->async_cfqq[0][i]) | |
2070 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
2071 | if (cfqd->async_cfqq[1][i]) | |
2072 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 2073 | } |
2389d1ef ON |
2074 | |
2075 | if (cfqd->async_idle_cfqq) | |
2076 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
2077 | } |
2078 | ||
1da177e4 LT |
2079 | static void cfq_exit_queue(elevator_t *e) |
2080 | { | |
22e2c507 | 2081 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 2082 | struct request_queue *q = cfqd->queue; |
22e2c507 | 2083 | |
3b18152c | 2084 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 2085 | |
d9ff4187 | 2086 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 2087 | |
d9ff4187 | 2088 | if (cfqd->active_queue) |
6084cdda | 2089 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
2090 | |
2091 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
2092 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
2093 | struct cfq_io_context, | |
2094 | queue_list); | |
89850f7e JA |
2095 | |
2096 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 2097 | } |
e2d74ac0 | 2098 | |
c2dea2d1 | 2099 | cfq_put_async_queues(cfqd); |
15c31be4 | 2100 | |
d9ff4187 | 2101 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
2102 | |
2103 | cfq_shutdown_timer_wq(cfqd); | |
2104 | ||
a90d742e | 2105 | kfree(cfqd); |
1da177e4 LT |
2106 | } |
2107 | ||
165125e1 | 2108 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
2109 | { |
2110 | struct cfq_data *cfqd; | |
1da177e4 | 2111 | |
94f6030c | 2112 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 2113 | if (!cfqd) |
bc1c1169 | 2114 | return NULL; |
1da177e4 | 2115 | |
cc09e299 | 2116 | cfqd->service_tree = CFQ_RB_ROOT; |
d9ff4187 | 2117 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 2118 | |
1da177e4 | 2119 | cfqd->queue = q; |
1da177e4 | 2120 | |
22e2c507 JA |
2121 | init_timer(&cfqd->idle_slice_timer); |
2122 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
2123 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
2124 | ||
65f27f38 | 2125 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 2126 | |
b70c864d | 2127 | cfqd->last_end_request = jiffies; |
1da177e4 | 2128 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
2129 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
2130 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
2131 | cfqd->cfq_back_max = cfq_back_max; |
2132 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
2133 | cfqd->cfq_slice[0] = cfq_slice_async; |
2134 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
2135 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
2136 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
3b18152c | 2137 | |
bc1c1169 | 2138 | return cfqd; |
1da177e4 LT |
2139 | } |
2140 | ||
2141 | static void cfq_slab_kill(void) | |
2142 | { | |
1da177e4 LT |
2143 | if (cfq_pool) |
2144 | kmem_cache_destroy(cfq_pool); | |
2145 | if (cfq_ioc_pool) | |
2146 | kmem_cache_destroy(cfq_ioc_pool); | |
2147 | } | |
2148 | ||
2149 | static int __init cfq_slab_setup(void) | |
2150 | { | |
0a31bd5f | 2151 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
2152 | if (!cfq_pool) |
2153 | goto fail; | |
2154 | ||
4ac845a2 | 2155 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, SLAB_DESTROY_BY_RCU); |
1da177e4 LT |
2156 | if (!cfq_ioc_pool) |
2157 | goto fail; | |
2158 | ||
2159 | return 0; | |
2160 | fail: | |
2161 | cfq_slab_kill(); | |
2162 | return -ENOMEM; | |
2163 | } | |
2164 | ||
1da177e4 LT |
2165 | /* |
2166 | * sysfs parts below --> | |
2167 | */ | |
1da177e4 LT |
2168 | static ssize_t |
2169 | cfq_var_show(unsigned int var, char *page) | |
2170 | { | |
2171 | return sprintf(page, "%d\n", var); | |
2172 | } | |
2173 | ||
2174 | static ssize_t | |
2175 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
2176 | { | |
2177 | char *p = (char *) page; | |
2178 | ||
2179 | *var = simple_strtoul(p, &p, 10); | |
2180 | return count; | |
2181 | } | |
2182 | ||
1da177e4 | 2183 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
3d1ab40f | 2184 | static ssize_t __FUNC(elevator_t *e, char *page) \ |
1da177e4 | 2185 | { \ |
3d1ab40f | 2186 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2187 | unsigned int __data = __VAR; \ |
2188 | if (__CONV) \ | |
2189 | __data = jiffies_to_msecs(__data); \ | |
2190 | return cfq_var_show(__data, (page)); \ | |
2191 | } | |
2192 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
2193 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
2194 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
2195 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
2196 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
2197 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
2198 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
2199 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
2200 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
1da177e4 LT |
2201 | #undef SHOW_FUNCTION |
2202 | ||
2203 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
3d1ab40f | 2204 | static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \ |
1da177e4 | 2205 | { \ |
3d1ab40f | 2206 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2207 | unsigned int __data; \ |
2208 | int ret = cfq_var_store(&__data, (page), count); \ | |
2209 | if (__data < (MIN)) \ | |
2210 | __data = (MIN); \ | |
2211 | else if (__data > (MAX)) \ | |
2212 | __data = (MAX); \ | |
2213 | if (__CONV) \ | |
2214 | *(__PTR) = msecs_to_jiffies(__data); \ | |
2215 | else \ | |
2216 | *(__PTR) = __data; \ | |
2217 | return ret; \ | |
2218 | } | |
2219 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
2220 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
2221 | UINT_MAX, 1); | |
2222 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
2223 | UINT_MAX, 1); | |
e572ec7e | 2224 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
2225 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
2226 | UINT_MAX, 0); | |
22e2c507 JA |
2227 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
2228 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
2229 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
2230 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
2231 | UINT_MAX, 0); | |
1da177e4 LT |
2232 | #undef STORE_FUNCTION |
2233 | ||
e572ec7e AV |
2234 | #define CFQ_ATTR(name) \ |
2235 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
2236 | ||
2237 | static struct elv_fs_entry cfq_attrs[] = { | |
2238 | CFQ_ATTR(quantum), | |
e572ec7e AV |
2239 | CFQ_ATTR(fifo_expire_sync), |
2240 | CFQ_ATTR(fifo_expire_async), | |
2241 | CFQ_ATTR(back_seek_max), | |
2242 | CFQ_ATTR(back_seek_penalty), | |
2243 | CFQ_ATTR(slice_sync), | |
2244 | CFQ_ATTR(slice_async), | |
2245 | CFQ_ATTR(slice_async_rq), | |
2246 | CFQ_ATTR(slice_idle), | |
e572ec7e | 2247 | __ATTR_NULL |
1da177e4 LT |
2248 | }; |
2249 | ||
1da177e4 LT |
2250 | static struct elevator_type iosched_cfq = { |
2251 | .ops = { | |
2252 | .elevator_merge_fn = cfq_merge, | |
2253 | .elevator_merged_fn = cfq_merged_request, | |
2254 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 2255 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 2256 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 2257 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 2258 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
2259 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
2260 | .elevator_queue_empty_fn = cfq_queue_empty, | |
2261 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
2262 | .elevator_former_req_fn = elv_rb_former_request, |
2263 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
2264 | .elevator_set_req_fn = cfq_set_request, |
2265 | .elevator_put_req_fn = cfq_put_request, | |
2266 | .elevator_may_queue_fn = cfq_may_queue, | |
2267 | .elevator_init_fn = cfq_init_queue, | |
2268 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 2269 | .trim = cfq_free_io_context, |
1da177e4 | 2270 | }, |
3d1ab40f | 2271 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
2272 | .elevator_name = "cfq", |
2273 | .elevator_owner = THIS_MODULE, | |
2274 | }; | |
2275 | ||
2276 | static int __init cfq_init(void) | |
2277 | { | |
22e2c507 JA |
2278 | /* |
2279 | * could be 0 on HZ < 1000 setups | |
2280 | */ | |
2281 | if (!cfq_slice_async) | |
2282 | cfq_slice_async = 1; | |
2283 | if (!cfq_slice_idle) | |
2284 | cfq_slice_idle = 1; | |
2285 | ||
1da177e4 LT |
2286 | if (cfq_slab_setup()) |
2287 | return -ENOMEM; | |
2288 | ||
2fdd82bd | 2289 | elv_register(&iosched_cfq); |
1da177e4 | 2290 | |
2fdd82bd | 2291 | return 0; |
1da177e4 LT |
2292 | } |
2293 | ||
2294 | static void __exit cfq_exit(void) | |
2295 | { | |
6e9a4738 | 2296 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 2297 | elv_unregister(&iosched_cfq); |
334e94de | 2298 | ioc_gone = &all_gone; |
fba82272 OH |
2299 | /* ioc_gone's update must be visible before reading ioc_count */ |
2300 | smp_wmb(); | |
4050cf16 | 2301 | if (elv_ioc_count_read(ioc_count)) |
fba82272 | 2302 | wait_for_completion(ioc_gone); |
334e94de | 2303 | synchronize_rcu(); |
83521d3e | 2304 | cfq_slab_kill(); |
1da177e4 LT |
2305 | } |
2306 | ||
2307 | module_init(cfq_init); | |
2308 | module_exit(cfq_exit); | |
2309 | ||
2310 | MODULE_AUTHOR("Jens Axboe"); | |
2311 | MODULE_LICENSE("GPL"); | |
2312 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |