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> |
7b679138 | 14 | #include <linux/blktrace_api.h> |
1da177e4 LT |
15 | |
16 | /* | |
17 | * tunables | |
18 | */ | |
fe094d98 JA |
19 | /* max queue in one round of service */ |
20 | static const int cfq_quantum = 4; | |
64100099 | 21 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
fe094d98 JA |
22 | /* maximum backwards seek, in KiB */ |
23 | static const int cfq_back_max = 16 * 1024; | |
24 | /* penalty of a backwards seek */ | |
25 | static const int cfq_back_penalty = 2; | |
64100099 | 26 | static const int cfq_slice_sync = HZ / 10; |
3b18152c | 27 | static int cfq_slice_async = HZ / 25; |
64100099 | 28 | static const int cfq_slice_async_rq = 2; |
caaa5f9f | 29 | static int cfq_slice_idle = HZ / 125; |
22e2c507 | 30 | |
d9e7620e | 31 | /* |
0871714e | 32 | * offset from end of service tree |
d9e7620e | 33 | */ |
0871714e | 34 | #define CFQ_IDLE_DELAY (HZ / 5) |
d9e7620e JA |
35 | |
36 | /* | |
37 | * below this threshold, we consider thinktime immediate | |
38 | */ | |
39 | #define CFQ_MIN_TT (2) | |
40 | ||
22e2c507 | 41 | #define CFQ_SLICE_SCALE (5) |
45333d5a | 42 | #define CFQ_HW_QUEUE_MIN (5) |
22e2c507 | 43 | |
fe094d98 JA |
44 | #define RQ_CIC(rq) \ |
45 | ((struct cfq_io_context *) (rq)->elevator_private) | |
7b679138 | 46 | #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2) |
1da177e4 | 47 | |
e18b890b CL |
48 | static struct kmem_cache *cfq_pool; |
49 | static struct kmem_cache *cfq_ioc_pool; | |
1da177e4 | 50 | |
4050cf16 | 51 | static DEFINE_PER_CPU(unsigned long, ioc_count); |
334e94de | 52 | static struct completion *ioc_gone; |
9a11b4ed | 53 | static DEFINE_SPINLOCK(ioc_gone_lock); |
334e94de | 54 | |
22e2c507 JA |
55 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR |
56 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
22e2c507 JA |
57 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) |
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 | ||
6118b70b JA |
73 | /* |
74 | * Per process-grouping structure | |
75 | */ | |
76 | struct cfq_queue { | |
77 | /* reference count */ | |
78 | atomic_t ref; | |
79 | /* various state flags, see below */ | |
80 | unsigned int flags; | |
81 | /* parent cfq_data */ | |
82 | struct cfq_data *cfqd; | |
83 | /* service_tree member */ | |
84 | struct rb_node rb_node; | |
85 | /* service_tree key */ | |
86 | unsigned long rb_key; | |
87 | /* prio tree member */ | |
88 | struct rb_node p_node; | |
89 | /* prio tree root we belong to, if any */ | |
90 | struct rb_root *p_root; | |
91 | /* sorted list of pending requests */ | |
92 | struct rb_root sort_list; | |
93 | /* if fifo isn't expired, next request to serve */ | |
94 | struct request *next_rq; | |
95 | /* requests queued in sort_list */ | |
96 | int queued[2]; | |
97 | /* currently allocated requests */ | |
98 | int allocated[2]; | |
99 | /* fifo list of requests in sort_list */ | |
100 | struct list_head fifo; | |
101 | ||
102 | unsigned long slice_end; | |
103 | long slice_resid; | |
104 | unsigned int slice_dispatch; | |
105 | ||
106 | /* pending metadata requests */ | |
107 | int meta_pending; | |
108 | /* number of requests that are on the dispatch list or inside driver */ | |
109 | int dispatched; | |
110 | ||
111 | /* io prio of this group */ | |
112 | unsigned short ioprio, org_ioprio; | |
113 | unsigned short ioprio_class, org_ioprio_class; | |
114 | ||
115 | pid_t pid; | |
116 | }; | |
117 | ||
22e2c507 JA |
118 | /* |
119 | * Per block device queue structure | |
120 | */ | |
1da177e4 | 121 | struct cfq_data { |
165125e1 | 122 | struct request_queue *queue; |
22e2c507 JA |
123 | |
124 | /* | |
125 | * rr list of queues with requests and the count of them | |
126 | */ | |
cc09e299 | 127 | struct cfq_rb_root service_tree; |
a36e71f9 JA |
128 | |
129 | /* | |
130 | * Each priority tree is sorted by next_request position. These | |
131 | * trees are used when determining if two or more queues are | |
132 | * interleaving requests (see cfq_close_cooperator). | |
133 | */ | |
134 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
135 | ||
22e2c507 | 136 | unsigned int busy_queues; |
3a9a3f6c DS |
137 | /* |
138 | * Used to track any pending rt requests so we can pre-empt current | |
139 | * non-RT cfqq in service when this value is non-zero. | |
140 | */ | |
141 | unsigned int busy_rt_queues; | |
22e2c507 | 142 | |
22e2c507 | 143 | int rq_in_driver; |
3ed9a296 | 144 | int sync_flight; |
45333d5a AC |
145 | |
146 | /* | |
147 | * queue-depth detection | |
148 | */ | |
149 | int rq_queued; | |
25776e35 | 150 | int hw_tag; |
45333d5a AC |
151 | int hw_tag_samples; |
152 | int rq_in_driver_peak; | |
1da177e4 | 153 | |
22e2c507 JA |
154 | /* |
155 | * idle window management | |
156 | */ | |
157 | struct timer_list idle_slice_timer; | |
158 | struct work_struct unplug_work; | |
1da177e4 | 159 | |
22e2c507 JA |
160 | struct cfq_queue *active_queue; |
161 | struct cfq_io_context *active_cic; | |
22e2c507 | 162 | |
c2dea2d1 VT |
163 | /* |
164 | * async queue for each priority case | |
165 | */ | |
166 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
167 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 168 | |
6d048f53 | 169 | sector_t last_position; |
1da177e4 | 170 | |
1da177e4 LT |
171 | /* |
172 | * tunables, see top of file | |
173 | */ | |
174 | unsigned int cfq_quantum; | |
22e2c507 | 175 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
176 | unsigned int cfq_back_penalty; |
177 | unsigned int cfq_back_max; | |
22e2c507 JA |
178 | unsigned int cfq_slice[2]; |
179 | unsigned int cfq_slice_async_rq; | |
180 | unsigned int cfq_slice_idle; | |
d9ff4187 AV |
181 | |
182 | struct list_head cic_list; | |
1da177e4 | 183 | |
6118b70b JA |
184 | /* |
185 | * Fallback dummy cfqq for extreme OOM conditions | |
186 | */ | |
187 | struct cfq_queue oom_cfqq; | |
1da177e4 LT |
188 | }; |
189 | ||
3b18152c | 190 | enum cfqq_state_flags { |
b0b8d749 JA |
191 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
192 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 193 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 JA |
194 | CFQ_CFQQ_FLAG_must_alloc, /* must be allowed rq alloc */ |
195 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ | |
b0b8d749 JA |
196 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
197 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
198 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 199 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 200 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
a36e71f9 | 201 | CFQ_CFQQ_FLAG_coop, /* has done a coop jump of the queue */ |
3b18152c JA |
202 | }; |
203 | ||
204 | #define CFQ_CFQQ_FNS(name) \ | |
205 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
206 | { \ | |
fe094d98 | 207 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
208 | } \ |
209 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
210 | { \ | |
fe094d98 | 211 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
212 | } \ |
213 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
214 | { \ | |
fe094d98 | 215 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
216 | } |
217 | ||
218 | CFQ_CFQQ_FNS(on_rr); | |
219 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 220 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c JA |
221 | CFQ_CFQQ_FNS(must_alloc); |
222 | CFQ_CFQQ_FNS(must_alloc_slice); | |
3b18152c JA |
223 | CFQ_CFQQ_FNS(fifo_expire); |
224 | CFQ_CFQQ_FNS(idle_window); | |
225 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 226 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 227 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 228 | CFQ_CFQQ_FNS(coop); |
3b18152c JA |
229 | #undef CFQ_CFQQ_FNS |
230 | ||
7b679138 JA |
231 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
232 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
233 | #define cfq_log(cfqd, fmt, args...) \ | |
234 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
235 | ||
165125e1 | 236 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
91fac317 | 237 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, int, |
fd0928df | 238 | struct io_context *, gfp_t); |
4ac845a2 | 239 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
240 | struct io_context *); |
241 | ||
242 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, | |
243 | int is_sync) | |
244 | { | |
245 | return cic->cfqq[!!is_sync]; | |
246 | } | |
247 | ||
248 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
249 | struct cfq_queue *cfqq, int is_sync) | |
250 | { | |
251 | cic->cfqq[!!is_sync] = cfqq; | |
252 | } | |
253 | ||
254 | /* | |
255 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
256 | * set (in which case it could also be direct WRITE). | |
257 | */ | |
258 | static inline int cfq_bio_sync(struct bio *bio) | |
259 | { | |
260 | if (bio_data_dir(bio) == READ || bio_sync(bio)) | |
261 | return 1; | |
262 | ||
263 | return 0; | |
264 | } | |
1da177e4 | 265 | |
99f95e52 AM |
266 | /* |
267 | * scheduler run of queue, if there are requests pending and no one in the | |
268 | * driver that will restart queueing | |
269 | */ | |
270 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) | |
271 | { | |
7b679138 JA |
272 | if (cfqd->busy_queues) { |
273 | cfq_log(cfqd, "schedule dispatch"); | |
18887ad9 | 274 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 275 | } |
99f95e52 AM |
276 | } |
277 | ||
165125e1 | 278 | static int cfq_queue_empty(struct request_queue *q) |
99f95e52 AM |
279 | { |
280 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
281 | ||
b4878f24 | 282 | return !cfqd->busy_queues; |
99f95e52 AM |
283 | } |
284 | ||
44f7c160 JA |
285 | /* |
286 | * Scale schedule slice based on io priority. Use the sync time slice only | |
287 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
288 | * io only, should not get full sync slice length. | |
289 | */ | |
d9e7620e JA |
290 | static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync, |
291 | unsigned short prio) | |
44f7c160 | 292 | { |
d9e7620e | 293 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 294 | |
d9e7620e JA |
295 | WARN_ON(prio >= IOPRIO_BE_NR); |
296 | ||
297 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
298 | } | |
44f7c160 | 299 | |
d9e7620e JA |
300 | static inline int |
301 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
302 | { | |
303 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
304 | } |
305 | ||
306 | static inline void | |
307 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
308 | { | |
309 | cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies; | |
7b679138 | 310 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
311 | } |
312 | ||
313 | /* | |
314 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
315 | * isn't valid until the first request from the dispatch is activated | |
316 | * and the slice time set. | |
317 | */ | |
318 | static inline int cfq_slice_used(struct cfq_queue *cfqq) | |
319 | { | |
320 | if (cfq_cfqq_slice_new(cfqq)) | |
321 | return 0; | |
322 | if (time_before(jiffies, cfqq->slice_end)) | |
323 | return 0; | |
324 | ||
325 | return 1; | |
326 | } | |
327 | ||
1da177e4 | 328 | /* |
5e705374 | 329 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 330 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 331 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 332 | */ |
5e705374 JA |
333 | static struct request * |
334 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2) | |
1da177e4 LT |
335 | { |
336 | sector_t last, s1, s2, d1 = 0, d2 = 0; | |
1da177e4 | 337 | unsigned long back_max; |
e8a99053 AM |
338 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
339 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
340 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 341 | |
5e705374 JA |
342 | if (rq1 == NULL || rq1 == rq2) |
343 | return rq2; | |
344 | if (rq2 == NULL) | |
345 | return rq1; | |
9c2c38a1 | 346 | |
5e705374 JA |
347 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
348 | return rq1; | |
349 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
350 | return rq2; | |
374f84ac JA |
351 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
352 | return rq1; | |
353 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
354 | return rq2; | |
1da177e4 | 355 | |
83096ebf TH |
356 | s1 = blk_rq_pos(rq1); |
357 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 358 | |
6d048f53 | 359 | last = cfqd->last_position; |
1da177e4 | 360 | |
1da177e4 LT |
361 | /* |
362 | * by definition, 1KiB is 2 sectors | |
363 | */ | |
364 | back_max = cfqd->cfq_back_max * 2; | |
365 | ||
366 | /* | |
367 | * Strict one way elevator _except_ in the case where we allow | |
368 | * short backward seeks which are biased as twice the cost of a | |
369 | * similar forward seek. | |
370 | */ | |
371 | if (s1 >= last) | |
372 | d1 = s1 - last; | |
373 | else if (s1 + back_max >= last) | |
374 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
375 | else | |
e8a99053 | 376 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
377 | |
378 | if (s2 >= last) | |
379 | d2 = s2 - last; | |
380 | else if (s2 + back_max >= last) | |
381 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
382 | else | |
e8a99053 | 383 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
384 | |
385 | /* Found required data */ | |
e8a99053 AM |
386 | |
387 | /* | |
388 | * By doing switch() on the bit mask "wrap" we avoid having to | |
389 | * check two variables for all permutations: --> faster! | |
390 | */ | |
391 | switch (wrap) { | |
5e705374 | 392 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 393 | if (d1 < d2) |
5e705374 | 394 | return rq1; |
e8a99053 | 395 | else if (d2 < d1) |
5e705374 | 396 | return rq2; |
e8a99053 AM |
397 | else { |
398 | if (s1 >= s2) | |
5e705374 | 399 | return rq1; |
e8a99053 | 400 | else |
5e705374 | 401 | return rq2; |
e8a99053 | 402 | } |
1da177e4 | 403 | |
e8a99053 | 404 | case CFQ_RQ2_WRAP: |
5e705374 | 405 | return rq1; |
e8a99053 | 406 | case CFQ_RQ1_WRAP: |
5e705374 JA |
407 | return rq2; |
408 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
409 | default: |
410 | /* | |
411 | * Since both rqs are wrapped, | |
412 | * start with the one that's further behind head | |
413 | * (--> only *one* back seek required), | |
414 | * since back seek takes more time than forward. | |
415 | */ | |
416 | if (s1 <= s2) | |
5e705374 | 417 | return rq1; |
1da177e4 | 418 | else |
5e705374 | 419 | return rq2; |
1da177e4 LT |
420 | } |
421 | } | |
422 | ||
498d3aa2 JA |
423 | /* |
424 | * The below is leftmost cache rbtree addon | |
425 | */ | |
0871714e | 426 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 JA |
427 | { |
428 | if (!root->left) | |
429 | root->left = rb_first(&root->rb); | |
430 | ||
0871714e JA |
431 | if (root->left) |
432 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
433 | ||
434 | return NULL; | |
cc09e299 JA |
435 | } |
436 | ||
a36e71f9 JA |
437 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
438 | { | |
439 | rb_erase(n, root); | |
440 | RB_CLEAR_NODE(n); | |
441 | } | |
442 | ||
cc09e299 JA |
443 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
444 | { | |
445 | if (root->left == n) | |
446 | root->left = NULL; | |
a36e71f9 | 447 | rb_erase_init(n, &root->rb); |
cc09e299 JA |
448 | } |
449 | ||
1da177e4 LT |
450 | /* |
451 | * would be nice to take fifo expire time into account as well | |
452 | */ | |
5e705374 JA |
453 | static struct request * |
454 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
455 | struct request *last) | |
1da177e4 | 456 | { |
21183b07 JA |
457 | struct rb_node *rbnext = rb_next(&last->rb_node); |
458 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 459 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 460 | |
21183b07 | 461 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
462 | |
463 | if (rbprev) | |
5e705374 | 464 | prev = rb_entry_rq(rbprev); |
1da177e4 | 465 | |
21183b07 | 466 | if (rbnext) |
5e705374 | 467 | next = rb_entry_rq(rbnext); |
21183b07 JA |
468 | else { |
469 | rbnext = rb_first(&cfqq->sort_list); | |
470 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 471 | next = rb_entry_rq(rbnext); |
21183b07 | 472 | } |
1da177e4 | 473 | |
21183b07 | 474 | return cfq_choose_req(cfqd, next, prev); |
1da177e4 LT |
475 | } |
476 | ||
d9e7620e JA |
477 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
478 | struct cfq_queue *cfqq) | |
1da177e4 | 479 | { |
d9e7620e JA |
480 | /* |
481 | * just an approximation, should be ok. | |
482 | */ | |
67e6b49e JA |
483 | return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
484 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); | |
d9e7620e JA |
485 | } |
486 | ||
498d3aa2 JA |
487 | /* |
488 | * The cfqd->service_tree holds all pending cfq_queue's that have | |
489 | * requests waiting to be processed. It is sorted in the order that | |
490 | * we will service the queues. | |
491 | */ | |
a36e71f9 JA |
492 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
493 | int add_front) | |
d9e7620e | 494 | { |
0871714e JA |
495 | struct rb_node **p, *parent; |
496 | struct cfq_queue *__cfqq; | |
d9e7620e | 497 | unsigned long rb_key; |
498d3aa2 | 498 | int left; |
d9e7620e | 499 | |
0871714e JA |
500 | if (cfq_class_idle(cfqq)) { |
501 | rb_key = CFQ_IDLE_DELAY; | |
502 | parent = rb_last(&cfqd->service_tree.rb); | |
503 | if (parent && parent != &cfqq->rb_node) { | |
504 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
505 | rb_key += __cfqq->rb_key; | |
506 | } else | |
507 | rb_key += jiffies; | |
508 | } else if (!add_front) { | |
edd75ffd JA |
509 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
510 | rb_key += cfqq->slice_resid; | |
511 | cfqq->slice_resid = 0; | |
512 | } else | |
513 | rb_key = 0; | |
1da177e4 | 514 | |
d9e7620e | 515 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
99f9628a | 516 | /* |
d9e7620e | 517 | * same position, nothing more to do |
99f9628a | 518 | */ |
d9e7620e JA |
519 | if (rb_key == cfqq->rb_key) |
520 | return; | |
1da177e4 | 521 | |
cc09e299 | 522 | cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree); |
1da177e4 | 523 | } |
d9e7620e | 524 | |
498d3aa2 | 525 | left = 1; |
0871714e JA |
526 | parent = NULL; |
527 | p = &cfqd->service_tree.rb.rb_node; | |
d9e7620e | 528 | while (*p) { |
67060e37 | 529 | struct rb_node **n; |
cc09e299 | 530 | |
d9e7620e JA |
531 | parent = *p; |
532 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
533 | ||
0c534e0a JA |
534 | /* |
535 | * sort RT queues first, we always want to give | |
67060e37 JA |
536 | * preference to them. IDLE queues goes to the back. |
537 | * after that, sort on the next service time. | |
0c534e0a JA |
538 | */ |
539 | if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq)) | |
67060e37 | 540 | n = &(*p)->rb_left; |
0c534e0a | 541 | else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq)) |
67060e37 JA |
542 | n = &(*p)->rb_right; |
543 | else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq)) | |
544 | n = &(*p)->rb_left; | |
545 | else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq)) | |
546 | n = &(*p)->rb_right; | |
0c534e0a | 547 | else if (rb_key < __cfqq->rb_key) |
67060e37 JA |
548 | n = &(*p)->rb_left; |
549 | else | |
550 | n = &(*p)->rb_right; | |
551 | ||
552 | if (n == &(*p)->rb_right) | |
cc09e299 | 553 | left = 0; |
67060e37 JA |
554 | |
555 | p = n; | |
d9e7620e JA |
556 | } |
557 | ||
cc09e299 JA |
558 | if (left) |
559 | cfqd->service_tree.left = &cfqq->rb_node; | |
560 | ||
d9e7620e JA |
561 | cfqq->rb_key = rb_key; |
562 | rb_link_node(&cfqq->rb_node, parent, p); | |
cc09e299 | 563 | rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb); |
1da177e4 LT |
564 | } |
565 | ||
a36e71f9 | 566 | static struct cfq_queue * |
f2d1f0ae JA |
567 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
568 | sector_t sector, struct rb_node **ret_parent, | |
569 | struct rb_node ***rb_link) | |
a36e71f9 | 570 | { |
a36e71f9 JA |
571 | struct rb_node **p, *parent; |
572 | struct cfq_queue *cfqq = NULL; | |
573 | ||
574 | parent = NULL; | |
575 | p = &root->rb_node; | |
576 | while (*p) { | |
577 | struct rb_node **n; | |
578 | ||
579 | parent = *p; | |
580 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
581 | ||
582 | /* | |
583 | * Sort strictly based on sector. Smallest to the left, | |
584 | * largest to the right. | |
585 | */ | |
2e46e8b2 | 586 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 587 | n = &(*p)->rb_right; |
2e46e8b2 | 588 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
589 | n = &(*p)->rb_left; |
590 | else | |
591 | break; | |
592 | p = n; | |
3ac6c9f8 | 593 | cfqq = NULL; |
a36e71f9 JA |
594 | } |
595 | ||
596 | *ret_parent = parent; | |
597 | if (rb_link) | |
598 | *rb_link = p; | |
3ac6c9f8 | 599 | return cfqq; |
a36e71f9 JA |
600 | } |
601 | ||
602 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
603 | { | |
a36e71f9 JA |
604 | struct rb_node **p, *parent; |
605 | struct cfq_queue *__cfqq; | |
606 | ||
f2d1f0ae JA |
607 | if (cfqq->p_root) { |
608 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
609 | cfqq->p_root = NULL; | |
610 | } | |
a36e71f9 JA |
611 | |
612 | if (cfq_class_idle(cfqq)) | |
613 | return; | |
614 | if (!cfqq->next_rq) | |
615 | return; | |
616 | ||
f2d1f0ae | 617 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
618 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
619 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
620 | if (!__cfqq) { |
621 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
622 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
623 | } else | |
624 | cfqq->p_root = NULL; | |
a36e71f9 JA |
625 | } |
626 | ||
498d3aa2 JA |
627 | /* |
628 | * Update cfqq's position in the service tree. | |
629 | */ | |
edd75ffd | 630 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 631 | { |
6d048f53 JA |
632 | /* |
633 | * Resorting requires the cfqq to be on the RR list already. | |
634 | */ | |
a36e71f9 | 635 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 636 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
637 | cfq_prio_tree_add(cfqd, cfqq); |
638 | } | |
6d048f53 JA |
639 | } |
640 | ||
1da177e4 LT |
641 | /* |
642 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 643 | * the pending list according to last request service |
1da177e4 | 644 | */ |
febffd61 | 645 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 646 | { |
7b679138 | 647 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
648 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
649 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 | 650 | cfqd->busy_queues++; |
3a9a3f6c DS |
651 | if (cfq_class_rt(cfqq)) |
652 | cfqd->busy_rt_queues++; | |
1da177e4 | 653 | |
edd75ffd | 654 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
655 | } |
656 | ||
498d3aa2 JA |
657 | /* |
658 | * Called when the cfqq no longer has requests pending, remove it from | |
659 | * the service tree. | |
660 | */ | |
febffd61 | 661 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 662 | { |
7b679138 | 663 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
664 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
665 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 666 | |
cc09e299 JA |
667 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) |
668 | cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree); | |
f2d1f0ae JA |
669 | if (cfqq->p_root) { |
670 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
671 | cfqq->p_root = NULL; | |
672 | } | |
d9e7620e | 673 | |
1da177e4 LT |
674 | BUG_ON(!cfqd->busy_queues); |
675 | cfqd->busy_queues--; | |
3a9a3f6c DS |
676 | if (cfq_class_rt(cfqq)) |
677 | cfqd->busy_rt_queues--; | |
1da177e4 LT |
678 | } |
679 | ||
680 | /* | |
681 | * rb tree support functions | |
682 | */ | |
febffd61 | 683 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 684 | { |
5e705374 | 685 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 686 | struct cfq_data *cfqd = cfqq->cfqd; |
5e705374 | 687 | const int sync = rq_is_sync(rq); |
1da177e4 | 688 | |
b4878f24 JA |
689 | BUG_ON(!cfqq->queued[sync]); |
690 | cfqq->queued[sync]--; | |
1da177e4 | 691 | |
5e705374 | 692 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 693 | |
dd67d051 | 694 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
b4878f24 | 695 | cfq_del_cfqq_rr(cfqd, cfqq); |
1da177e4 LT |
696 | } |
697 | ||
5e705374 | 698 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 699 | { |
5e705374 | 700 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 701 | struct cfq_data *cfqd = cfqq->cfqd; |
a36e71f9 | 702 | struct request *__alias, *prev; |
1da177e4 | 703 | |
5380a101 | 704 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
705 | |
706 | /* | |
707 | * looks a little odd, but the first insert might return an alias. | |
708 | * if that happens, put the alias on the dispatch list | |
709 | */ | |
21183b07 | 710 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 711 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
712 | |
713 | if (!cfq_cfqq_on_rr(cfqq)) | |
714 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
715 | |
716 | /* | |
717 | * check if this request is a better next-serve candidate | |
718 | */ | |
a36e71f9 | 719 | prev = cfqq->next_rq; |
5044eed4 | 720 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq); |
a36e71f9 JA |
721 | |
722 | /* | |
723 | * adjust priority tree position, if ->next_rq changes | |
724 | */ | |
725 | if (prev != cfqq->next_rq) | |
726 | cfq_prio_tree_add(cfqd, cfqq); | |
727 | ||
5044eed4 | 728 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
729 | } |
730 | ||
febffd61 | 731 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 732 | { |
5380a101 JA |
733 | elv_rb_del(&cfqq->sort_list, rq); |
734 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 735 | cfq_add_rq_rb(rq); |
1da177e4 LT |
736 | } |
737 | ||
206dc69b JA |
738 | static struct request * |
739 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 740 | { |
206dc69b | 741 | struct task_struct *tsk = current; |
91fac317 | 742 | struct cfq_io_context *cic; |
206dc69b | 743 | struct cfq_queue *cfqq; |
1da177e4 | 744 | |
4ac845a2 | 745 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
746 | if (!cic) |
747 | return NULL; | |
748 | ||
749 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
750 | if (cfqq) { |
751 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
752 | ||
21183b07 | 753 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 754 | } |
1da177e4 | 755 | |
1da177e4 LT |
756 | return NULL; |
757 | } | |
758 | ||
165125e1 | 759 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 760 | { |
22e2c507 | 761 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 762 | |
b4878f24 | 763 | cfqd->rq_in_driver++; |
7b679138 JA |
764 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
765 | cfqd->rq_in_driver); | |
25776e35 | 766 | |
5b93629b | 767 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
768 | } |
769 | ||
165125e1 | 770 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 771 | { |
b4878f24 JA |
772 | struct cfq_data *cfqd = q->elevator->elevator_data; |
773 | ||
774 | WARN_ON(!cfqd->rq_in_driver); | |
775 | cfqd->rq_in_driver--; | |
7b679138 JA |
776 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
777 | cfqd->rq_in_driver); | |
1da177e4 LT |
778 | } |
779 | ||
b4878f24 | 780 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 781 | { |
5e705374 | 782 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 783 | |
5e705374 JA |
784 | if (cfqq->next_rq == rq) |
785 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 786 | |
b4878f24 | 787 | list_del_init(&rq->queuelist); |
5e705374 | 788 | cfq_del_rq_rb(rq); |
374f84ac | 789 | |
45333d5a | 790 | cfqq->cfqd->rq_queued--; |
374f84ac JA |
791 | if (rq_is_meta(rq)) { |
792 | WARN_ON(!cfqq->meta_pending); | |
793 | cfqq->meta_pending--; | |
794 | } | |
1da177e4 LT |
795 | } |
796 | ||
165125e1 JA |
797 | static int cfq_merge(struct request_queue *q, struct request **req, |
798 | struct bio *bio) | |
1da177e4 LT |
799 | { |
800 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
801 | struct request *__rq; | |
1da177e4 | 802 | |
206dc69b | 803 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 804 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
805 | *req = __rq; |
806 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
807 | } |
808 | ||
809 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
810 | } |
811 | ||
165125e1 | 812 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 813 | int type) |
1da177e4 | 814 | { |
21183b07 | 815 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 816 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 817 | |
5e705374 | 818 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 819 | } |
1da177e4 LT |
820 | } |
821 | ||
822 | static void | |
165125e1 | 823 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
824 | struct request *next) |
825 | { | |
22e2c507 JA |
826 | /* |
827 | * reposition in fifo if next is older than rq | |
828 | */ | |
829 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
830 | time_before(next->start_time, rq->start_time)) | |
831 | list_move(&rq->queuelist, &next->queuelist); | |
832 | ||
b4878f24 | 833 | cfq_remove_request(next); |
22e2c507 JA |
834 | } |
835 | ||
165125e1 | 836 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
837 | struct bio *bio) |
838 | { | |
839 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 840 | struct cfq_io_context *cic; |
da775265 | 841 | struct cfq_queue *cfqq; |
da775265 JA |
842 | |
843 | /* | |
ec8acb69 | 844 | * Disallow merge of a sync bio into an async request. |
da775265 | 845 | */ |
91fac317 | 846 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
da775265 JA |
847 | return 0; |
848 | ||
849 | /* | |
719d3402 JA |
850 | * Lookup the cfqq that this bio will be queued with. Allow |
851 | * merge only if rq is queued there. | |
da775265 | 852 | */ |
4ac845a2 | 853 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 VT |
854 | if (!cic) |
855 | return 0; | |
719d3402 | 856 | |
91fac317 | 857 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
719d3402 JA |
858 | if (cfqq == RQ_CFQQ(rq)) |
859 | return 1; | |
da775265 | 860 | |
ec8acb69 | 861 | return 0; |
da775265 JA |
862 | } |
863 | ||
febffd61 JA |
864 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
865 | struct cfq_queue *cfqq) | |
22e2c507 JA |
866 | { |
867 | if (cfqq) { | |
7b679138 | 868 | cfq_log_cfqq(cfqd, cfqq, "set_active"); |
22e2c507 | 869 | cfqq->slice_end = 0; |
2f5cb738 JA |
870 | cfqq->slice_dispatch = 0; |
871 | ||
2f5cb738 | 872 | cfq_clear_cfqq_wait_request(cfqq); |
b029195d | 873 | cfq_clear_cfqq_must_dispatch(cfqq); |
3b18152c JA |
874 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
875 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 876 | cfq_mark_cfqq_slice_new(cfqq); |
2f5cb738 JA |
877 | |
878 | del_timer(&cfqd->idle_slice_timer); | |
22e2c507 JA |
879 | } |
880 | ||
881 | cfqd->active_queue = cfqq; | |
882 | } | |
883 | ||
7b14e3b5 JA |
884 | /* |
885 | * current cfqq expired its slice (or was too idle), select new one | |
886 | */ | |
887 | static void | |
888 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
6084cdda | 889 | int timed_out) |
7b14e3b5 | 890 | { |
7b679138 JA |
891 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
892 | ||
7b14e3b5 JA |
893 | if (cfq_cfqq_wait_request(cfqq)) |
894 | del_timer(&cfqd->idle_slice_timer); | |
895 | ||
7b14e3b5 JA |
896 | cfq_clear_cfqq_wait_request(cfqq); |
897 | ||
898 | /* | |
6084cdda | 899 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 900 | */ |
7b679138 | 901 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) { |
c5b680f3 | 902 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b679138 JA |
903 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
904 | } | |
7b14e3b5 | 905 | |
edd75ffd | 906 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
907 | |
908 | if (cfqq == cfqd->active_queue) | |
909 | cfqd->active_queue = NULL; | |
910 | ||
911 | if (cfqd->active_cic) { | |
912 | put_io_context(cfqd->active_cic->ioc); | |
913 | cfqd->active_cic = NULL; | |
914 | } | |
7b14e3b5 JA |
915 | } |
916 | ||
6084cdda | 917 | static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out) |
7b14e3b5 JA |
918 | { |
919 | struct cfq_queue *cfqq = cfqd->active_queue; | |
920 | ||
921 | if (cfqq) | |
6084cdda | 922 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
923 | } |
924 | ||
498d3aa2 JA |
925 | /* |
926 | * Get next queue for service. Unless we have a queue preemption, | |
927 | * we'll simply select the first cfqq in the service tree. | |
928 | */ | |
6d048f53 | 929 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 930 | { |
edd75ffd JA |
931 | if (RB_EMPTY_ROOT(&cfqd->service_tree.rb)) |
932 | return NULL; | |
d9e7620e | 933 | |
0871714e | 934 | return cfq_rb_first(&cfqd->service_tree); |
6d048f53 JA |
935 | } |
936 | ||
498d3aa2 JA |
937 | /* |
938 | * Get and set a new active queue for service. | |
939 | */ | |
a36e71f9 JA |
940 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
941 | struct cfq_queue *cfqq) | |
6d048f53 | 942 | { |
a36e71f9 JA |
943 | if (!cfqq) { |
944 | cfqq = cfq_get_next_queue(cfqd); | |
945 | if (cfqq) | |
946 | cfq_clear_cfqq_coop(cfqq); | |
947 | } | |
6d048f53 | 948 | |
22e2c507 | 949 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 950 | return cfqq; |
22e2c507 JA |
951 | } |
952 | ||
d9e7620e JA |
953 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
954 | struct request *rq) | |
955 | { | |
83096ebf TH |
956 | if (blk_rq_pos(rq) >= cfqd->last_position) |
957 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 958 | else |
83096ebf | 959 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
960 | } |
961 | ||
04dc6e71 JM |
962 | #define CIC_SEEK_THR 8 * 1024 |
963 | #define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR) | |
964 | ||
6d048f53 JA |
965 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq) |
966 | { | |
967 | struct cfq_io_context *cic = cfqd->active_cic; | |
04dc6e71 | 968 | sector_t sdist = cic->seek_mean; |
6d048f53 JA |
969 | |
970 | if (!sample_valid(cic->seek_samples)) | |
04dc6e71 | 971 | sdist = CIC_SEEK_THR; |
6d048f53 | 972 | |
04dc6e71 | 973 | return cfq_dist_from_last(cfqd, rq) <= sdist; |
6d048f53 JA |
974 | } |
975 | ||
a36e71f9 JA |
976 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
977 | struct cfq_queue *cur_cfqq) | |
978 | { | |
f2d1f0ae | 979 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
980 | struct rb_node *parent, *node; |
981 | struct cfq_queue *__cfqq; | |
982 | sector_t sector = cfqd->last_position; | |
983 | ||
984 | if (RB_EMPTY_ROOT(root)) | |
985 | return NULL; | |
986 | ||
987 | /* | |
988 | * First, if we find a request starting at the end of the last | |
989 | * request, choose it. | |
990 | */ | |
f2d1f0ae | 991 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
992 | if (__cfqq) |
993 | return __cfqq; | |
994 | ||
995 | /* | |
996 | * If the exact sector wasn't found, the parent of the NULL leaf | |
997 | * will contain the closest sector. | |
998 | */ | |
999 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
1000 | if (cfq_rq_close(cfqd, __cfqq->next_rq)) | |
1001 | return __cfqq; | |
1002 | ||
2e46e8b2 | 1003 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1004 | node = rb_next(&__cfqq->p_node); |
1005 | else | |
1006 | node = rb_prev(&__cfqq->p_node); | |
1007 | if (!node) | |
1008 | return NULL; | |
1009 | ||
1010 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
1011 | if (cfq_rq_close(cfqd, __cfqq->next_rq)) | |
1012 | return __cfqq; | |
1013 | ||
1014 | return NULL; | |
1015 | } | |
1016 | ||
1017 | /* | |
1018 | * cfqd - obvious | |
1019 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1020 | * closely cooperating with itself. | |
1021 | * | |
1022 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1023 | * one request, and that cfqd->last_position reflects a position on the disk | |
1024 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1025 | * assumption. | |
1026 | */ | |
1027 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
1028 | struct cfq_queue *cur_cfqq, | |
1029 | int probe) | |
6d048f53 | 1030 | { |
a36e71f9 JA |
1031 | struct cfq_queue *cfqq; |
1032 | ||
1033 | /* | |
1034 | * A valid cfq_io_context is necessary to compare requests against | |
1035 | * the seek_mean of the current cfqq. | |
1036 | */ | |
1037 | if (!cfqd->active_cic) | |
1038 | return NULL; | |
1039 | ||
6d048f53 | 1040 | /* |
d9e7620e JA |
1041 | * We should notice if some of the queues are cooperating, eg |
1042 | * working closely on the same area of the disk. In that case, | |
1043 | * we can group them together and don't waste time idling. | |
6d048f53 | 1044 | */ |
a36e71f9 JA |
1045 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1046 | if (!cfqq) | |
1047 | return NULL; | |
1048 | ||
1049 | if (cfq_cfqq_coop(cfqq)) | |
1050 | return NULL; | |
1051 | ||
1052 | if (!probe) | |
1053 | cfq_mark_cfqq_coop(cfqq); | |
1054 | return cfqq; | |
6d048f53 JA |
1055 | } |
1056 | ||
6d048f53 | 1057 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1058 | { |
1792669c | 1059 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1060 | struct cfq_io_context *cic; |
7b14e3b5 JA |
1061 | unsigned long sl; |
1062 | ||
a68bbddb | 1063 | /* |
f7d7b7a7 JA |
1064 | * SSD device without seek penalty, disable idling. But only do so |
1065 | * for devices that support queuing, otherwise we still have a problem | |
1066 | * with sync vs async workloads. | |
a68bbddb | 1067 | */ |
f7d7b7a7 | 1068 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
1069 | return; |
1070 | ||
dd67d051 | 1071 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 1072 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
1073 | |
1074 | /* | |
1075 | * idle is disabled, either manually or by past process history | |
1076 | */ | |
6d048f53 JA |
1077 | if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq)) |
1078 | return; | |
1079 | ||
7b679138 JA |
1080 | /* |
1081 | * still requests with the driver, don't idle | |
1082 | */ | |
1083 | if (cfqd->rq_in_driver) | |
1084 | return; | |
1085 | ||
22e2c507 JA |
1086 | /* |
1087 | * task has exited, don't wait | |
1088 | */ | |
206dc69b | 1089 | cic = cfqd->active_cic; |
66dac98e | 1090 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
1091 | return; |
1092 | ||
3b18152c | 1093 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 1094 | |
206dc69b JA |
1095 | /* |
1096 | * we don't want to idle for seeks, but we do want to allow | |
1097 | * fair distribution of slice time for a process doing back-to-back | |
1098 | * seeks. so allow a little bit of time for him to submit a new rq | |
1099 | */ | |
6d048f53 | 1100 | sl = cfqd->cfq_slice_idle; |
caaa5f9f | 1101 | if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic)) |
d9e7620e | 1102 | sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT)); |
206dc69b | 1103 | |
7b14e3b5 | 1104 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
9481ffdc | 1105 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); |
1da177e4 LT |
1106 | } |
1107 | ||
498d3aa2 JA |
1108 | /* |
1109 | * Move request from internal lists to the request queue dispatch list. | |
1110 | */ | |
165125e1 | 1111 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 1112 | { |
3ed9a296 | 1113 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1114 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 1115 | |
7b679138 JA |
1116 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
1117 | ||
5380a101 | 1118 | cfq_remove_request(rq); |
6d048f53 | 1119 | cfqq->dispatched++; |
5380a101 | 1120 | elv_dispatch_sort(q, rq); |
3ed9a296 JA |
1121 | |
1122 | if (cfq_cfqq_sync(cfqq)) | |
1123 | cfqd->sync_flight++; | |
1da177e4 LT |
1124 | } |
1125 | ||
1126 | /* | |
1127 | * return expired entry, or NULL to just start from scratch in rbtree | |
1128 | */ | |
febffd61 | 1129 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 LT |
1130 | { |
1131 | struct cfq_data *cfqd = cfqq->cfqd; | |
22e2c507 | 1132 | struct request *rq; |
89850f7e | 1133 | int fifo; |
1da177e4 | 1134 | |
3b18152c | 1135 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 1136 | return NULL; |
cb887411 JA |
1137 | |
1138 | cfq_mark_cfqq_fifo_expire(cfqq); | |
1139 | ||
89850f7e JA |
1140 | if (list_empty(&cfqq->fifo)) |
1141 | return NULL; | |
1da177e4 | 1142 | |
6d048f53 | 1143 | fifo = cfq_cfqq_sync(cfqq); |
89850f7e | 1144 | rq = rq_entry_fifo(cfqq->fifo.next); |
1da177e4 | 1145 | |
6d048f53 | 1146 | if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) |
7b679138 | 1147 | rq = NULL; |
1da177e4 | 1148 | |
7b679138 | 1149 | cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 1150 | return rq; |
1da177e4 LT |
1151 | } |
1152 | ||
22e2c507 JA |
1153 | static inline int |
1154 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1155 | { | |
1156 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 1157 | |
22e2c507 | 1158 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 1159 | |
22e2c507 | 1160 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
1161 | } |
1162 | ||
22e2c507 | 1163 | /* |
498d3aa2 JA |
1164 | * Select a queue for service. If we have a current active queue, |
1165 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 1166 | */ |
1b5ed5e1 | 1167 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 1168 | { |
a36e71f9 | 1169 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 1170 | |
22e2c507 JA |
1171 | cfqq = cfqd->active_queue; |
1172 | if (!cfqq) | |
1173 | goto new_queue; | |
1da177e4 | 1174 | |
22e2c507 | 1175 | /* |
6d048f53 | 1176 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 1177 | */ |
b029195d | 1178 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) |
3b18152c | 1179 | goto expire; |
1da177e4 | 1180 | |
3a9a3f6c DS |
1181 | /* |
1182 | * If we have a RT cfqq waiting, then we pre-empt the current non-rt | |
1183 | * cfqq. | |
1184 | */ | |
1185 | if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) { | |
1186 | /* | |
1187 | * We simulate this as cfqq timed out so that it gets to bank | |
1188 | * the remaining of its time slice. | |
1189 | */ | |
1190 | cfq_log_cfqq(cfqd, cfqq, "preempt"); | |
1191 | cfq_slice_expired(cfqd, 1); | |
1192 | goto new_queue; | |
1193 | } | |
1194 | ||
22e2c507 | 1195 | /* |
6d048f53 JA |
1196 | * The active queue has requests and isn't expired, allow it to |
1197 | * dispatch. | |
22e2c507 | 1198 | */ |
dd67d051 | 1199 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 1200 | goto keep_queue; |
6d048f53 | 1201 | |
a36e71f9 JA |
1202 | /* |
1203 | * If another queue has a request waiting within our mean seek | |
1204 | * distance, let it run. The expire code will check for close | |
1205 | * cooperators and put the close queue at the front of the service | |
1206 | * tree. | |
1207 | */ | |
1208 | new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0); | |
1209 | if (new_cfqq) | |
1210 | goto expire; | |
1211 | ||
6d048f53 JA |
1212 | /* |
1213 | * No requests pending. If the active queue still has requests in | |
1214 | * flight or is idling for a new request, allow either of these | |
1215 | * conditions to happen (or time out) before selecting a new queue. | |
1216 | */ | |
cc197479 JA |
1217 | if (timer_pending(&cfqd->idle_slice_timer) || |
1218 | (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) { | |
caaa5f9f JA |
1219 | cfqq = NULL; |
1220 | goto keep_queue; | |
22e2c507 JA |
1221 | } |
1222 | ||
3b18152c | 1223 | expire: |
6084cdda | 1224 | cfq_slice_expired(cfqd, 0); |
3b18152c | 1225 | new_queue: |
a36e71f9 | 1226 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 1227 | keep_queue: |
3b18152c | 1228 | return cfqq; |
22e2c507 JA |
1229 | } |
1230 | ||
febffd61 | 1231 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
1232 | { |
1233 | int dispatched = 0; | |
1234 | ||
1235 | while (cfqq->next_rq) { | |
1236 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
1237 | dispatched++; | |
1238 | } | |
1239 | ||
1240 | BUG_ON(!list_empty(&cfqq->fifo)); | |
1241 | return dispatched; | |
1242 | } | |
1243 | ||
498d3aa2 JA |
1244 | /* |
1245 | * Drain our current requests. Used for barriers and when switching | |
1246 | * io schedulers on-the-fly. | |
1247 | */ | |
d9e7620e | 1248 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 1249 | { |
0871714e | 1250 | struct cfq_queue *cfqq; |
d9e7620e | 1251 | int dispatched = 0; |
1b5ed5e1 | 1252 | |
0871714e | 1253 | while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL) |
d9e7620e | 1254 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); |
1b5ed5e1 | 1255 | |
6084cdda | 1256 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
1257 | |
1258 | BUG_ON(cfqd->busy_queues); | |
1259 | ||
6923715a | 1260 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
1261 | return dispatched; |
1262 | } | |
1263 | ||
2f5cb738 JA |
1264 | /* |
1265 | * Dispatch a request from cfqq, moving them to the request queue | |
1266 | * dispatch list. | |
1267 | */ | |
1268 | static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1269 | { | |
1270 | struct request *rq; | |
1271 | ||
1272 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
1273 | ||
1274 | /* | |
1275 | * follow expired path, else get first next available | |
1276 | */ | |
1277 | rq = cfq_check_fifo(cfqq); | |
1278 | if (!rq) | |
1279 | rq = cfqq->next_rq; | |
1280 | ||
1281 | /* | |
1282 | * insert request into driver dispatch list | |
1283 | */ | |
1284 | cfq_dispatch_insert(cfqd->queue, rq); | |
1285 | ||
1286 | if (!cfqd->active_cic) { | |
1287 | struct cfq_io_context *cic = RQ_CIC(rq); | |
1288 | ||
d9c7d394 | 1289 | atomic_long_inc(&cic->ioc->refcount); |
2f5cb738 JA |
1290 | cfqd->active_cic = cic; |
1291 | } | |
1292 | } | |
1293 | ||
1294 | /* | |
1295 | * Find the cfqq that we need to service and move a request from that to the | |
1296 | * dispatch list | |
1297 | */ | |
165125e1 | 1298 | static int cfq_dispatch_requests(struct request_queue *q, int force) |
22e2c507 JA |
1299 | { |
1300 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
6d048f53 | 1301 | struct cfq_queue *cfqq; |
2f5cb738 | 1302 | unsigned int max_dispatch; |
22e2c507 JA |
1303 | |
1304 | if (!cfqd->busy_queues) | |
1305 | return 0; | |
1306 | ||
1b5ed5e1 TH |
1307 | if (unlikely(force)) |
1308 | return cfq_forced_dispatch(cfqd); | |
1309 | ||
2f5cb738 JA |
1310 | cfqq = cfq_select_queue(cfqd); |
1311 | if (!cfqq) | |
1312 | return 0; | |
1313 | ||
1314 | /* | |
1315 | * If this is an async queue and we have sync IO in flight, let it wait | |
1316 | */ | |
1317 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) | |
1318 | return 0; | |
1319 | ||
1320 | max_dispatch = cfqd->cfq_quantum; | |
1321 | if (cfq_class_idle(cfqq)) | |
1322 | max_dispatch = 1; | |
b4878f24 | 1323 | |
2f5cb738 JA |
1324 | /* |
1325 | * Does this cfqq already have too much IO in flight? | |
1326 | */ | |
1327 | if (cfqq->dispatched >= max_dispatch) { | |
1328 | /* | |
1329 | * idle queue must always only have a single IO in flight | |
1330 | */ | |
3ed9a296 | 1331 | if (cfq_class_idle(cfqq)) |
2f5cb738 | 1332 | return 0; |
3ed9a296 | 1333 | |
2f5cb738 JA |
1334 | /* |
1335 | * We have other queues, don't allow more IO from this one | |
1336 | */ | |
1337 | if (cfqd->busy_queues > 1) | |
1338 | return 0; | |
9ede209e | 1339 | |
2f5cb738 JA |
1340 | /* |
1341 | * we are the only queue, allow up to 4 times of 'quantum' | |
1342 | */ | |
1343 | if (cfqq->dispatched >= 4 * max_dispatch) | |
1344 | return 0; | |
1345 | } | |
3ed9a296 | 1346 | |
2f5cb738 JA |
1347 | /* |
1348 | * Dispatch a request from this cfqq | |
1349 | */ | |
1350 | cfq_dispatch_request(cfqd, cfqq); | |
1351 | cfqq->slice_dispatch++; | |
b029195d | 1352 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 1353 | |
2f5cb738 JA |
1354 | /* |
1355 | * expire an async queue immediately if it has used up its slice. idle | |
1356 | * queue always expire after 1 dispatch round. | |
1357 | */ | |
1358 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
1359 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
1360 | cfq_class_idle(cfqq))) { | |
1361 | cfqq->slice_end = jiffies + 1; | |
1362 | cfq_slice_expired(cfqd, 0); | |
1da177e4 LT |
1363 | } |
1364 | ||
2f5cb738 JA |
1365 | cfq_log(cfqd, "dispatched a request"); |
1366 | return 1; | |
1da177e4 LT |
1367 | } |
1368 | ||
1da177e4 | 1369 | /* |
5e705374 JA |
1370 | * task holds one reference to the queue, dropped when task exits. each rq |
1371 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 LT |
1372 | * |
1373 | * queue lock must be held here. | |
1374 | */ | |
1375 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
1376 | { | |
22e2c507 JA |
1377 | struct cfq_data *cfqd = cfqq->cfqd; |
1378 | ||
1379 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
1380 | |
1381 | if (!atomic_dec_and_test(&cfqq->ref)) | |
1382 | return; | |
1383 | ||
7b679138 | 1384 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 1385 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 1386 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
3b18152c | 1387 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 | 1388 | |
28f95cbc | 1389 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 1390 | __cfq_slice_expired(cfqd, cfqq, 0); |
28f95cbc JA |
1391 | cfq_schedule_dispatch(cfqd); |
1392 | } | |
22e2c507 | 1393 | |
1da177e4 LT |
1394 | kmem_cache_free(cfq_pool, cfqq); |
1395 | } | |
1396 | ||
d6de8be7 JA |
1397 | /* |
1398 | * Must always be called with the rcu_read_lock() held | |
1399 | */ | |
07416d29 JA |
1400 | static void |
1401 | __call_for_each_cic(struct io_context *ioc, | |
1402 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1403 | { | |
1404 | struct cfq_io_context *cic; | |
1405 | struct hlist_node *n; | |
1406 | ||
1407 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) | |
1408 | func(ioc, cic); | |
1409 | } | |
1410 | ||
4ac845a2 | 1411 | /* |
34e6bbf2 | 1412 | * Call func for each cic attached to this ioc. |
4ac845a2 | 1413 | */ |
34e6bbf2 | 1414 | static void |
4ac845a2 JA |
1415 | call_for_each_cic(struct io_context *ioc, |
1416 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1da177e4 | 1417 | { |
4ac845a2 | 1418 | rcu_read_lock(); |
07416d29 | 1419 | __call_for_each_cic(ioc, func); |
4ac845a2 | 1420 | rcu_read_unlock(); |
34e6bbf2 FC |
1421 | } |
1422 | ||
1423 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
1424 | { | |
1425 | struct cfq_io_context *cic; | |
1426 | ||
1427 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
1428 | ||
1429 | kmem_cache_free(cfq_ioc_pool, cic); | |
1430 | elv_ioc_count_dec(ioc_count); | |
1431 | ||
9a11b4ed JA |
1432 | if (ioc_gone) { |
1433 | /* | |
1434 | * CFQ scheduler is exiting, grab exit lock and check | |
1435 | * the pending io context count. If it hits zero, | |
1436 | * complete ioc_gone and set it back to NULL | |
1437 | */ | |
1438 | spin_lock(&ioc_gone_lock); | |
1439 | if (ioc_gone && !elv_ioc_count_read(ioc_count)) { | |
1440 | complete(ioc_gone); | |
1441 | ioc_gone = NULL; | |
1442 | } | |
1443 | spin_unlock(&ioc_gone_lock); | |
1444 | } | |
34e6bbf2 | 1445 | } |
4ac845a2 | 1446 | |
34e6bbf2 FC |
1447 | static void cfq_cic_free(struct cfq_io_context *cic) |
1448 | { | |
1449 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
1450 | } |
1451 | ||
1452 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
1453 | { | |
1454 | unsigned long flags; | |
1455 | ||
1456 | BUG_ON(!cic->dead_key); | |
1457 | ||
1458 | spin_lock_irqsave(&ioc->lock, flags); | |
1459 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
ffc4e759 | 1460 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
1461 | spin_unlock_irqrestore(&ioc->lock, flags); |
1462 | ||
34e6bbf2 | 1463 | cfq_cic_free(cic); |
4ac845a2 JA |
1464 | } |
1465 | ||
d6de8be7 JA |
1466 | /* |
1467 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
1468 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
1469 | * and ->trim() which is called with the task lock held | |
1470 | */ | |
4ac845a2 JA |
1471 | static void cfq_free_io_context(struct io_context *ioc) |
1472 | { | |
4ac845a2 | 1473 | /* |
34e6bbf2 FC |
1474 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
1475 | * so no more cic's are allowed to be linked into this ioc. So it | |
1476 | * should be ok to iterate over the known list, we will see all cic's | |
1477 | * since no new ones are added. | |
4ac845a2 | 1478 | */ |
07416d29 | 1479 | __call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
1480 | } |
1481 | ||
89850f7e | 1482 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1483 | { |
28f95cbc | 1484 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 1485 | __cfq_slice_expired(cfqd, cfqq, 0); |
28f95cbc JA |
1486 | cfq_schedule_dispatch(cfqd); |
1487 | } | |
22e2c507 | 1488 | |
89850f7e JA |
1489 | cfq_put_queue(cfqq); |
1490 | } | |
22e2c507 | 1491 | |
89850f7e JA |
1492 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
1493 | struct cfq_io_context *cic) | |
1494 | { | |
4faa3c81 FC |
1495 | struct io_context *ioc = cic->ioc; |
1496 | ||
fc46379d | 1497 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
1498 | |
1499 | /* | |
1500 | * Make sure key == NULL is seen for dead queues | |
1501 | */ | |
fc46379d | 1502 | smp_wmb(); |
4ac845a2 | 1503 | cic->dead_key = (unsigned long) cic->key; |
fc46379d JA |
1504 | cic->key = NULL; |
1505 | ||
4faa3c81 FC |
1506 | if (ioc->ioc_data == cic) |
1507 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
1508 | ||
ff6657c6 JA |
1509 | if (cic->cfqq[BLK_RW_ASYNC]) { |
1510 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
1511 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
1512 | } |
1513 | ||
ff6657c6 JA |
1514 | if (cic->cfqq[BLK_RW_SYNC]) { |
1515 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
1516 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 1517 | } |
89850f7e JA |
1518 | } |
1519 | ||
4ac845a2 JA |
1520 | static void cfq_exit_single_io_context(struct io_context *ioc, |
1521 | struct cfq_io_context *cic) | |
89850f7e JA |
1522 | { |
1523 | struct cfq_data *cfqd = cic->key; | |
1524 | ||
89850f7e | 1525 | if (cfqd) { |
165125e1 | 1526 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 1527 | unsigned long flags; |
89850f7e | 1528 | |
4ac845a2 | 1529 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
1530 | |
1531 | /* | |
1532 | * Ensure we get a fresh copy of the ->key to prevent | |
1533 | * race between exiting task and queue | |
1534 | */ | |
1535 | smp_read_barrier_depends(); | |
1536 | if (cic->key) | |
1537 | __cfq_exit_single_io_context(cfqd, cic); | |
1538 | ||
4ac845a2 | 1539 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 1540 | } |
1da177e4 LT |
1541 | } |
1542 | ||
498d3aa2 JA |
1543 | /* |
1544 | * The process that ioc belongs to has exited, we need to clean up | |
1545 | * and put the internal structures we have that belongs to that process. | |
1546 | */ | |
e2d74ac0 | 1547 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 1548 | { |
4ac845a2 | 1549 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
1550 | } |
1551 | ||
22e2c507 | 1552 | static struct cfq_io_context * |
8267e268 | 1553 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1554 | { |
b5deef90 | 1555 | struct cfq_io_context *cic; |
1da177e4 | 1556 | |
94f6030c CL |
1557 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
1558 | cfqd->queue->node); | |
1da177e4 | 1559 | if (cic) { |
22e2c507 | 1560 | cic->last_end_request = jiffies; |
553698f9 | 1561 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 1562 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
1563 | cic->dtor = cfq_free_io_context; |
1564 | cic->exit = cfq_exit_io_context; | |
4050cf16 | 1565 | elv_ioc_count_inc(ioc_count); |
1da177e4 LT |
1566 | } |
1567 | ||
1568 | return cic; | |
1569 | } | |
1570 | ||
fd0928df | 1571 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
1572 | { |
1573 | struct task_struct *tsk = current; | |
1574 | int ioprio_class; | |
1575 | ||
3b18152c | 1576 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
1577 | return; |
1578 | ||
fd0928df | 1579 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 1580 | switch (ioprio_class) { |
fe094d98 JA |
1581 | default: |
1582 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
1583 | case IOPRIO_CLASS_NONE: | |
1584 | /* | |
6d63c275 | 1585 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
1586 | */ |
1587 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 1588 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
1589 | break; |
1590 | case IOPRIO_CLASS_RT: | |
1591 | cfqq->ioprio = task_ioprio(ioc); | |
1592 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
1593 | break; | |
1594 | case IOPRIO_CLASS_BE: | |
1595 | cfqq->ioprio = task_ioprio(ioc); | |
1596 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
1597 | break; | |
1598 | case IOPRIO_CLASS_IDLE: | |
1599 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
1600 | cfqq->ioprio = 7; | |
1601 | cfq_clear_cfqq_idle_window(cfqq); | |
1602 | break; | |
22e2c507 JA |
1603 | } |
1604 | ||
1605 | /* | |
1606 | * keep track of original prio settings in case we have to temporarily | |
1607 | * elevate the priority of this queue | |
1608 | */ | |
1609 | cfqq->org_ioprio = cfqq->ioprio; | |
1610 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 1611 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
1612 | } |
1613 | ||
febffd61 | 1614 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 1615 | { |
478a82b0 AV |
1616 | struct cfq_data *cfqd = cic->key; |
1617 | struct cfq_queue *cfqq; | |
c1b707d2 | 1618 | unsigned long flags; |
35e6077c | 1619 | |
caaa5f9f JA |
1620 | if (unlikely(!cfqd)) |
1621 | return; | |
1622 | ||
c1b707d2 | 1623 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 1624 | |
ff6657c6 | 1625 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
1626 | if (cfqq) { |
1627 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
1628 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
1629 | GFP_ATOMIC); | |
caaa5f9f | 1630 | if (new_cfqq) { |
ff6657c6 | 1631 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
1632 | cfq_put_queue(cfqq); |
1633 | } | |
22e2c507 | 1634 | } |
caaa5f9f | 1635 | |
ff6657c6 | 1636 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
1637 | if (cfqq) |
1638 | cfq_mark_cfqq_prio_changed(cfqq); | |
1639 | ||
c1b707d2 | 1640 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
1641 | } |
1642 | ||
fc46379d | 1643 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 1644 | { |
4ac845a2 | 1645 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 1646 | ioc->ioprio_changed = 0; |
22e2c507 JA |
1647 | } |
1648 | ||
d5036d77 JA |
1649 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
1650 | pid_t pid, int is_sync) | |
1651 | { | |
1652 | RB_CLEAR_NODE(&cfqq->rb_node); | |
1653 | RB_CLEAR_NODE(&cfqq->p_node); | |
1654 | INIT_LIST_HEAD(&cfqq->fifo); | |
1655 | ||
1656 | atomic_set(&cfqq->ref, 0); | |
1657 | cfqq->cfqd = cfqd; | |
1658 | ||
1659 | cfq_mark_cfqq_prio_changed(cfqq); | |
1660 | ||
1661 | if (is_sync) { | |
1662 | if (!cfq_class_idle(cfqq)) | |
1663 | cfq_mark_cfqq_idle_window(cfqq); | |
1664 | cfq_mark_cfqq_sync(cfqq); | |
1665 | } | |
1666 | cfqq->pid = pid; | |
1667 | } | |
1668 | ||
22e2c507 | 1669 | static struct cfq_queue * |
15c31be4 | 1670 | cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync, |
fd0928df | 1671 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 1672 | { |
22e2c507 | 1673 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 1674 | struct cfq_io_context *cic; |
22e2c507 JA |
1675 | |
1676 | retry: | |
4ac845a2 | 1677 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
1678 | /* cic always exists here */ |
1679 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 1680 | |
6118b70b JA |
1681 | /* |
1682 | * Always try a new alloc if we fell back to the OOM cfqq | |
1683 | * originally, since it should just be a temporary situation. | |
1684 | */ | |
1685 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
1686 | cfqq = NULL; | |
22e2c507 JA |
1687 | if (new_cfqq) { |
1688 | cfqq = new_cfqq; | |
1689 | new_cfqq = NULL; | |
1690 | } else if (gfp_mask & __GFP_WAIT) { | |
1691 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 1692 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 1693 | gfp_mask | __GFP_ZERO, |
94f6030c | 1694 | cfqd->queue->node); |
22e2c507 | 1695 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
1696 | if (new_cfqq) |
1697 | goto retry; | |
22e2c507 | 1698 | } else { |
94f6030c CL |
1699 | cfqq = kmem_cache_alloc_node(cfq_pool, |
1700 | gfp_mask | __GFP_ZERO, | |
1701 | cfqd->queue->node); | |
22e2c507 JA |
1702 | } |
1703 | ||
6118b70b JA |
1704 | if (cfqq) { |
1705 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
1706 | cfq_init_prio_data(cfqq, ioc); | |
1707 | cfq_log_cfqq(cfqd, cfqq, "alloced"); | |
1708 | } else | |
1709 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
1710 | } |
1711 | ||
1712 | if (new_cfqq) | |
1713 | kmem_cache_free(cfq_pool, new_cfqq); | |
1714 | ||
22e2c507 JA |
1715 | return cfqq; |
1716 | } | |
1717 | ||
c2dea2d1 VT |
1718 | static struct cfq_queue ** |
1719 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
1720 | { | |
fe094d98 | 1721 | switch (ioprio_class) { |
c2dea2d1 VT |
1722 | case IOPRIO_CLASS_RT: |
1723 | return &cfqd->async_cfqq[0][ioprio]; | |
1724 | case IOPRIO_CLASS_BE: | |
1725 | return &cfqd->async_cfqq[1][ioprio]; | |
1726 | case IOPRIO_CLASS_IDLE: | |
1727 | return &cfqd->async_idle_cfqq; | |
1728 | default: | |
1729 | BUG(); | |
1730 | } | |
1731 | } | |
1732 | ||
15c31be4 | 1733 | static struct cfq_queue * |
fd0928df | 1734 | cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc, |
15c31be4 JA |
1735 | gfp_t gfp_mask) |
1736 | { | |
fd0928df JA |
1737 | const int ioprio = task_ioprio(ioc); |
1738 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 1739 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
1740 | struct cfq_queue *cfqq = NULL; |
1741 | ||
c2dea2d1 VT |
1742 | if (!is_sync) { |
1743 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
1744 | cfqq = *async_cfqq; | |
1745 | } | |
1746 | ||
6118b70b | 1747 | if (!cfqq) |
fd0928df | 1748 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
1749 | |
1750 | /* | |
1751 | * pin the queue now that it's allocated, scheduler exit will prune it | |
1752 | */ | |
c2dea2d1 | 1753 | if (!is_sync && !(*async_cfqq)) { |
15c31be4 | 1754 | atomic_inc(&cfqq->ref); |
c2dea2d1 | 1755 | *async_cfqq = cfqq; |
15c31be4 JA |
1756 | } |
1757 | ||
1758 | atomic_inc(&cfqq->ref); | |
1759 | return cfqq; | |
1760 | } | |
1761 | ||
498d3aa2 JA |
1762 | /* |
1763 | * We drop cfq io contexts lazily, so we may find a dead one. | |
1764 | */ | |
dbecf3ab | 1765 | static void |
4ac845a2 JA |
1766 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
1767 | struct cfq_io_context *cic) | |
dbecf3ab | 1768 | { |
4ac845a2 JA |
1769 | unsigned long flags; |
1770 | ||
fc46379d | 1771 | WARN_ON(!list_empty(&cic->queue_list)); |
597bc485 | 1772 | |
4ac845a2 JA |
1773 | spin_lock_irqsave(&ioc->lock, flags); |
1774 | ||
4faa3c81 | 1775 | BUG_ON(ioc->ioc_data == cic); |
597bc485 | 1776 | |
4ac845a2 | 1777 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); |
ffc4e759 | 1778 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
1779 | spin_unlock_irqrestore(&ioc->lock, flags); |
1780 | ||
1781 | cfq_cic_free(cic); | |
dbecf3ab OH |
1782 | } |
1783 | ||
e2d74ac0 | 1784 | static struct cfq_io_context * |
4ac845a2 | 1785 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 1786 | { |
e2d74ac0 | 1787 | struct cfq_io_context *cic; |
d6de8be7 | 1788 | unsigned long flags; |
4ac845a2 | 1789 | void *k; |
e2d74ac0 | 1790 | |
91fac317 VT |
1791 | if (unlikely(!ioc)) |
1792 | return NULL; | |
1793 | ||
d6de8be7 JA |
1794 | rcu_read_lock(); |
1795 | ||
597bc485 JA |
1796 | /* |
1797 | * we maintain a last-hit cache, to avoid browsing over the tree | |
1798 | */ | |
4ac845a2 | 1799 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
1800 | if (cic && cic->key == cfqd) { |
1801 | rcu_read_unlock(); | |
597bc485 | 1802 | return cic; |
d6de8be7 | 1803 | } |
597bc485 | 1804 | |
4ac845a2 | 1805 | do { |
4ac845a2 JA |
1806 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); |
1807 | rcu_read_unlock(); | |
1808 | if (!cic) | |
1809 | break; | |
be3b0753 OH |
1810 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
1811 | k = cic->key; | |
1812 | if (unlikely(!k)) { | |
4ac845a2 | 1813 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 1814 | rcu_read_lock(); |
4ac845a2 | 1815 | continue; |
dbecf3ab | 1816 | } |
e2d74ac0 | 1817 | |
d6de8be7 | 1818 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 1819 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 1820 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
1821 | break; |
1822 | } while (1); | |
e2d74ac0 | 1823 | |
4ac845a2 | 1824 | return cic; |
e2d74ac0 JA |
1825 | } |
1826 | ||
4ac845a2 JA |
1827 | /* |
1828 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
1829 | * the process specific cfq io context when entered from the block layer. | |
1830 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
1831 | */ | |
febffd61 JA |
1832 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
1833 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 1834 | { |
0261d688 | 1835 | unsigned long flags; |
4ac845a2 | 1836 | int ret; |
e2d74ac0 | 1837 | |
4ac845a2 JA |
1838 | ret = radix_tree_preload(gfp_mask); |
1839 | if (!ret) { | |
1840 | cic->ioc = ioc; | |
1841 | cic->key = cfqd; | |
e2d74ac0 | 1842 | |
4ac845a2 JA |
1843 | spin_lock_irqsave(&ioc->lock, flags); |
1844 | ret = radix_tree_insert(&ioc->radix_root, | |
1845 | (unsigned long) cfqd, cic); | |
ffc4e759 JA |
1846 | if (!ret) |
1847 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 1848 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 1849 | |
4ac845a2 JA |
1850 | radix_tree_preload_end(); |
1851 | ||
1852 | if (!ret) { | |
1853 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
1854 | list_add(&cic->queue_list, &cfqd->cic_list); | |
1855 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
1856 | } | |
e2d74ac0 JA |
1857 | } |
1858 | ||
4ac845a2 JA |
1859 | if (ret) |
1860 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 1861 | |
4ac845a2 | 1862 | return ret; |
e2d74ac0 JA |
1863 | } |
1864 | ||
1da177e4 LT |
1865 | /* |
1866 | * Setup general io context and cfq io context. There can be several cfq | |
1867 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 1868 | * than one device managed by cfq. |
1da177e4 LT |
1869 | */ |
1870 | static struct cfq_io_context * | |
e2d74ac0 | 1871 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1872 | { |
22e2c507 | 1873 | struct io_context *ioc = NULL; |
1da177e4 | 1874 | struct cfq_io_context *cic; |
1da177e4 | 1875 | |
22e2c507 | 1876 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 1877 | |
b5deef90 | 1878 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
1879 | if (!ioc) |
1880 | return NULL; | |
1881 | ||
4ac845a2 | 1882 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
1883 | if (cic) |
1884 | goto out; | |
1da177e4 | 1885 | |
e2d74ac0 JA |
1886 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
1887 | if (cic == NULL) | |
1888 | goto err; | |
1da177e4 | 1889 | |
4ac845a2 JA |
1890 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
1891 | goto err_free; | |
1892 | ||
1da177e4 | 1893 | out: |
fc46379d JA |
1894 | smp_read_barrier_depends(); |
1895 | if (unlikely(ioc->ioprio_changed)) | |
1896 | cfq_ioc_set_ioprio(ioc); | |
1897 | ||
1da177e4 | 1898 | return cic; |
4ac845a2 JA |
1899 | err_free: |
1900 | cfq_cic_free(cic); | |
1da177e4 LT |
1901 | err: |
1902 | put_io_context(ioc); | |
1903 | return NULL; | |
1904 | } | |
1905 | ||
22e2c507 JA |
1906 | static void |
1907 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 1908 | { |
aaf1228d JA |
1909 | unsigned long elapsed = jiffies - cic->last_end_request; |
1910 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 1911 | |
22e2c507 JA |
1912 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
1913 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
1914 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
1915 | } | |
1da177e4 | 1916 | |
206dc69b | 1917 | static void |
6d048f53 JA |
1918 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic, |
1919 | struct request *rq) | |
206dc69b JA |
1920 | { |
1921 | sector_t sdist; | |
1922 | u64 total; | |
1923 | ||
4d00aa47 JM |
1924 | if (!cic->last_request_pos) |
1925 | sdist = 0; | |
83096ebf TH |
1926 | else if (cic->last_request_pos < blk_rq_pos(rq)) |
1927 | sdist = blk_rq_pos(rq) - cic->last_request_pos; | |
206dc69b | 1928 | else |
83096ebf | 1929 | sdist = cic->last_request_pos - blk_rq_pos(rq); |
206dc69b JA |
1930 | |
1931 | /* | |
1932 | * Don't allow the seek distance to get too large from the | |
1933 | * odd fragment, pagein, etc | |
1934 | */ | |
1935 | if (cic->seek_samples <= 60) /* second&third seek */ | |
1936 | sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024); | |
1937 | else | |
1938 | sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64); | |
1939 | ||
1940 | cic->seek_samples = (7*cic->seek_samples + 256) / 8; | |
1941 | cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8; | |
1942 | total = cic->seek_total + (cic->seek_samples/2); | |
1943 | do_div(total, cic->seek_samples); | |
1944 | cic->seek_mean = (sector_t)total; | |
1945 | } | |
1da177e4 | 1946 | |
22e2c507 JA |
1947 | /* |
1948 | * Disable idle window if the process thinks too long or seeks so much that | |
1949 | * it doesn't matter | |
1950 | */ | |
1951 | static void | |
1952 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1953 | struct cfq_io_context *cic) | |
1954 | { | |
7b679138 | 1955 | int old_idle, enable_idle; |
1be92f2f | 1956 | |
0871714e JA |
1957 | /* |
1958 | * Don't idle for async or idle io prio class | |
1959 | */ | |
1960 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
1961 | return; |
1962 | ||
c265a7f4 | 1963 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 1964 | |
66dac98e | 1965 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || |
caaa5f9f | 1966 | (cfqd->hw_tag && CIC_SEEKY(cic))) |
22e2c507 JA |
1967 | enable_idle = 0; |
1968 | else if (sample_valid(cic->ttime_samples)) { | |
1969 | if (cic->ttime_mean > cfqd->cfq_slice_idle) | |
1970 | enable_idle = 0; | |
1971 | else | |
1972 | enable_idle = 1; | |
1da177e4 LT |
1973 | } |
1974 | ||
7b679138 JA |
1975 | if (old_idle != enable_idle) { |
1976 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
1977 | if (enable_idle) | |
1978 | cfq_mark_cfqq_idle_window(cfqq); | |
1979 | else | |
1980 | cfq_clear_cfqq_idle_window(cfqq); | |
1981 | } | |
22e2c507 | 1982 | } |
1da177e4 | 1983 | |
22e2c507 JA |
1984 | /* |
1985 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
1986 | * no or if we aren't sure, a 1 will cause a preempt. | |
1987 | */ | |
1988 | static int | |
1989 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, | |
5e705374 | 1990 | struct request *rq) |
22e2c507 | 1991 | { |
6d048f53 | 1992 | struct cfq_queue *cfqq; |
22e2c507 | 1993 | |
6d048f53 JA |
1994 | cfqq = cfqd->active_queue; |
1995 | if (!cfqq) | |
22e2c507 JA |
1996 | return 0; |
1997 | ||
6d048f53 JA |
1998 | if (cfq_slice_used(cfqq)) |
1999 | return 1; | |
2000 | ||
2001 | if (cfq_class_idle(new_cfqq)) | |
caaa5f9f | 2002 | return 0; |
22e2c507 JA |
2003 | |
2004 | if (cfq_class_idle(cfqq)) | |
2005 | return 1; | |
1e3335de | 2006 | |
374f84ac JA |
2007 | /* |
2008 | * if the new request is sync, but the currently running queue is | |
2009 | * not, let the sync request have priority. | |
2010 | */ | |
5e705374 | 2011 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
22e2c507 | 2012 | return 1; |
1e3335de | 2013 | |
374f84ac JA |
2014 | /* |
2015 | * So both queues are sync. Let the new request get disk time if | |
2016 | * it's a metadata request and the current queue is doing regular IO. | |
2017 | */ | |
2018 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
2019 | return 1; | |
22e2c507 | 2020 | |
3a9a3f6c DS |
2021 | /* |
2022 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
2023 | */ | |
2024 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
2025 | return 1; | |
2026 | ||
1e3335de JA |
2027 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
2028 | return 0; | |
2029 | ||
2030 | /* | |
2031 | * if this request is as-good as one we would expect from the | |
2032 | * current cfqq, let it preempt | |
2033 | */ | |
6d048f53 | 2034 | if (cfq_rq_close(cfqd, rq)) |
1e3335de JA |
2035 | return 1; |
2036 | ||
22e2c507 JA |
2037 | return 0; |
2038 | } | |
2039 | ||
2040 | /* | |
2041 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
2042 | * let it have half of its nominal slice. | |
2043 | */ | |
2044 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2045 | { | |
7b679138 | 2046 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
6084cdda | 2047 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 2048 | |
bf572256 JA |
2049 | /* |
2050 | * Put the new queue at the front of the of the current list, | |
2051 | * so we know that it will be selected next. | |
2052 | */ | |
2053 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
2054 | |
2055 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 2056 | |
44f7c160 JA |
2057 | cfqq->slice_end = 0; |
2058 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
2059 | } |
2060 | ||
22e2c507 | 2061 | /* |
5e705374 | 2062 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
2063 | * something we should do about it |
2064 | */ | |
2065 | static void | |
5e705374 JA |
2066 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
2067 | struct request *rq) | |
22e2c507 | 2068 | { |
5e705374 | 2069 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 2070 | |
45333d5a | 2071 | cfqd->rq_queued++; |
374f84ac JA |
2072 | if (rq_is_meta(rq)) |
2073 | cfqq->meta_pending++; | |
2074 | ||
9c2c38a1 | 2075 | cfq_update_io_thinktime(cfqd, cic); |
6d048f53 | 2076 | cfq_update_io_seektime(cfqd, cic, rq); |
9c2c38a1 JA |
2077 | cfq_update_idle_window(cfqd, cfqq, cic); |
2078 | ||
83096ebf | 2079 | cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
2080 | |
2081 | if (cfqq == cfqd->active_queue) { | |
2082 | /* | |
b029195d JA |
2083 | * Remember that we saw a request from this process, but |
2084 | * don't start queuing just yet. Otherwise we risk seeing lots | |
2085 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
2086 | * and merging. If the request is already larger than a single |
2087 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
2088 | * merging is already done. Ditto for a busy system that |
2089 | * has other work pending, don't risk delaying until the | |
2090 | * idle timer unplug to continue working. | |
22e2c507 | 2091 | */ |
d6ceb25e | 2092 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
2093 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
2094 | cfqd->busy_queues > 1) { | |
d6ceb25e | 2095 | del_timer(&cfqd->idle_slice_timer); |
a7f55792 | 2096 | __blk_run_queue(cfqd->queue); |
d6ceb25e | 2097 | } |
b029195d | 2098 | cfq_mark_cfqq_must_dispatch(cfqq); |
d6ceb25e | 2099 | } |
5e705374 | 2100 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
2101 | /* |
2102 | * not the active queue - expire current slice if it is | |
2103 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
2104 | * has some old slice time left and is of higher priority or |
2105 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
2106 | */ |
2107 | cfq_preempt_queue(cfqd, cfqq); | |
a7f55792 | 2108 | __blk_run_queue(cfqd->queue); |
22e2c507 | 2109 | } |
1da177e4 LT |
2110 | } |
2111 | ||
165125e1 | 2112 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2113 | { |
b4878f24 | 2114 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 2115 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 2116 | |
7b679138 | 2117 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 2118 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 2119 | |
5e705374 | 2120 | cfq_add_rq_rb(rq); |
1da177e4 | 2121 | |
22e2c507 JA |
2122 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
2123 | ||
5e705374 | 2124 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
2125 | } |
2126 | ||
45333d5a AC |
2127 | /* |
2128 | * Update hw_tag based on peak queue depth over 50 samples under | |
2129 | * sufficient load. | |
2130 | */ | |
2131 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
2132 | { | |
2133 | if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak) | |
2134 | cfqd->rq_in_driver_peak = cfqd->rq_in_driver; | |
2135 | ||
2136 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
2137 | cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN) | |
2138 | return; | |
2139 | ||
2140 | if (cfqd->hw_tag_samples++ < 50) | |
2141 | return; | |
2142 | ||
2143 | if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN) | |
2144 | cfqd->hw_tag = 1; | |
2145 | else | |
2146 | cfqd->hw_tag = 0; | |
2147 | ||
2148 | cfqd->hw_tag_samples = 0; | |
2149 | cfqd->rq_in_driver_peak = 0; | |
2150 | } | |
2151 | ||
165125e1 | 2152 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2153 | { |
5e705374 | 2154 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 2155 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 2156 | const int sync = rq_is_sync(rq); |
b4878f24 | 2157 | unsigned long now; |
1da177e4 | 2158 | |
b4878f24 | 2159 | now = jiffies; |
7b679138 | 2160 | cfq_log_cfqq(cfqd, cfqq, "complete"); |
1da177e4 | 2161 | |
45333d5a AC |
2162 | cfq_update_hw_tag(cfqd); |
2163 | ||
b4878f24 | 2164 | WARN_ON(!cfqd->rq_in_driver); |
6d048f53 | 2165 | WARN_ON(!cfqq->dispatched); |
b4878f24 | 2166 | cfqd->rq_in_driver--; |
6d048f53 | 2167 | cfqq->dispatched--; |
1da177e4 | 2168 | |
3ed9a296 JA |
2169 | if (cfq_cfqq_sync(cfqq)) |
2170 | cfqd->sync_flight--; | |
2171 | ||
caaa5f9f | 2172 | if (sync) |
5e705374 | 2173 | RQ_CIC(rq)->last_end_request = now; |
caaa5f9f JA |
2174 | |
2175 | /* | |
2176 | * If this is the active queue, check if it needs to be expired, | |
2177 | * or if we want to idle in case it has no pending requests. | |
2178 | */ | |
2179 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
2180 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
2181 | ||
44f7c160 JA |
2182 | if (cfq_cfqq_slice_new(cfqq)) { |
2183 | cfq_set_prio_slice(cfqd, cfqq); | |
2184 | cfq_clear_cfqq_slice_new(cfqq); | |
2185 | } | |
a36e71f9 JA |
2186 | /* |
2187 | * If there are no requests waiting in this queue, and | |
2188 | * there are other queues ready to issue requests, AND | |
2189 | * those other queues are issuing requests within our | |
2190 | * mean seek distance, give them a chance to run instead | |
2191 | * of idling. | |
2192 | */ | |
0871714e | 2193 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
6084cdda | 2194 | cfq_slice_expired(cfqd, 1); |
a36e71f9 JA |
2195 | else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) && |
2196 | sync && !rq_noidle(rq)) | |
6d048f53 | 2197 | cfq_arm_slice_timer(cfqd); |
caaa5f9f | 2198 | } |
6d048f53 JA |
2199 | |
2200 | if (!cfqd->rq_in_driver) | |
2201 | cfq_schedule_dispatch(cfqd); | |
1da177e4 LT |
2202 | } |
2203 | ||
22e2c507 JA |
2204 | /* |
2205 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
2206 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
2207 | */ | |
2208 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 2209 | { |
22e2c507 JA |
2210 | if (has_fs_excl()) { |
2211 | /* | |
2212 | * boost idle prio on transactions that would lock out other | |
2213 | * users of the filesystem | |
2214 | */ | |
2215 | if (cfq_class_idle(cfqq)) | |
2216 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2217 | if (cfqq->ioprio > IOPRIO_NORM) | |
2218 | cfqq->ioprio = IOPRIO_NORM; | |
2219 | } else { | |
2220 | /* | |
2221 | * check if we need to unboost the queue | |
2222 | */ | |
2223 | if (cfqq->ioprio_class != cfqq->org_ioprio_class) | |
2224 | cfqq->ioprio_class = cfqq->org_ioprio_class; | |
2225 | if (cfqq->ioprio != cfqq->org_ioprio) | |
2226 | cfqq->ioprio = cfqq->org_ioprio; | |
2227 | } | |
22e2c507 | 2228 | } |
1da177e4 | 2229 | |
89850f7e | 2230 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 2231 | { |
3b18152c | 2232 | if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) && |
99f95e52 | 2233 | !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 2234 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 2235 | return ELV_MQUEUE_MUST; |
3b18152c | 2236 | } |
1da177e4 | 2237 | |
22e2c507 | 2238 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
2239 | } |
2240 | ||
165125e1 | 2241 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
2242 | { |
2243 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2244 | struct task_struct *tsk = current; | |
91fac317 | 2245 | struct cfq_io_context *cic; |
22e2c507 JA |
2246 | struct cfq_queue *cfqq; |
2247 | ||
2248 | /* | |
2249 | * don't force setup of a queue from here, as a call to may_queue | |
2250 | * does not necessarily imply that a request actually will be queued. | |
2251 | * so just lookup a possibly existing queue, or return 'may queue' | |
2252 | * if that fails | |
2253 | */ | |
4ac845a2 | 2254 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
2255 | if (!cic) |
2256 | return ELV_MQUEUE_MAY; | |
2257 | ||
b0b78f81 | 2258 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 2259 | if (cfqq) { |
fd0928df | 2260 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
2261 | cfq_prio_boost(cfqq); |
2262 | ||
89850f7e | 2263 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
2264 | } |
2265 | ||
2266 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
2267 | } |
2268 | ||
1da177e4 LT |
2269 | /* |
2270 | * queue lock held here | |
2271 | */ | |
bb37b94c | 2272 | static void cfq_put_request(struct request *rq) |
1da177e4 | 2273 | { |
5e705374 | 2274 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 2275 | |
5e705374 | 2276 | if (cfqq) { |
22e2c507 | 2277 | const int rw = rq_data_dir(rq); |
1da177e4 | 2278 | |
22e2c507 JA |
2279 | BUG_ON(!cfqq->allocated[rw]); |
2280 | cfqq->allocated[rw]--; | |
1da177e4 | 2281 | |
5e705374 | 2282 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 2283 | |
1da177e4 | 2284 | rq->elevator_private = NULL; |
5e705374 | 2285 | rq->elevator_private2 = NULL; |
1da177e4 | 2286 | |
1da177e4 LT |
2287 | cfq_put_queue(cfqq); |
2288 | } | |
2289 | } | |
2290 | ||
2291 | /* | |
22e2c507 | 2292 | * Allocate cfq data structures associated with this request. |
1da177e4 | 2293 | */ |
22e2c507 | 2294 | static int |
165125e1 | 2295 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
2296 | { |
2297 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2298 | struct cfq_io_context *cic; | |
2299 | const int rw = rq_data_dir(rq); | |
7749a8d4 | 2300 | const int is_sync = rq_is_sync(rq); |
22e2c507 | 2301 | struct cfq_queue *cfqq; |
1da177e4 LT |
2302 | unsigned long flags; |
2303 | ||
2304 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
2305 | ||
e2d74ac0 | 2306 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 2307 | |
1da177e4 LT |
2308 | spin_lock_irqsave(q->queue_lock, flags); |
2309 | ||
22e2c507 JA |
2310 | if (!cic) |
2311 | goto queue_fail; | |
2312 | ||
91fac317 VT |
2313 | cfqq = cic_to_cfqq(cic, is_sync); |
2314 | if (!cfqq) { | |
fd0928df | 2315 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 | 2316 | |
22e2c507 JA |
2317 | if (!cfqq) |
2318 | goto queue_fail; | |
1da177e4 | 2319 | |
91fac317 VT |
2320 | cic_set_cfqq(cic, cfqq, is_sync); |
2321 | } | |
1da177e4 LT |
2322 | |
2323 | cfqq->allocated[rw]++; | |
3b18152c | 2324 | cfq_clear_cfqq_must_alloc(cfqq); |
22e2c507 | 2325 | atomic_inc(&cfqq->ref); |
1da177e4 | 2326 | |
5e705374 | 2327 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 2328 | |
5e705374 JA |
2329 | rq->elevator_private = cic; |
2330 | rq->elevator_private2 = cfqq; | |
2331 | return 0; | |
1da177e4 | 2332 | |
22e2c507 JA |
2333 | queue_fail: |
2334 | if (cic) | |
2335 | put_io_context(cic->ioc); | |
89850f7e | 2336 | |
3b18152c | 2337 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 2338 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 2339 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
2340 | return 1; |
2341 | } | |
2342 | ||
65f27f38 | 2343 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 2344 | { |
65f27f38 DH |
2345 | struct cfq_data *cfqd = |
2346 | container_of(work, struct cfq_data, unplug_work); | |
165125e1 | 2347 | struct request_queue *q = cfqd->queue; |
22e2c507 | 2348 | |
40bb54d1 | 2349 | spin_lock_irq(q->queue_lock); |
a7f55792 | 2350 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 2351 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
2352 | } |
2353 | ||
2354 | /* | |
2355 | * Timer running if the active_queue is currently idling inside its time slice | |
2356 | */ | |
2357 | static void cfq_idle_slice_timer(unsigned long data) | |
2358 | { | |
2359 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
2360 | struct cfq_queue *cfqq; | |
2361 | unsigned long flags; | |
3c6bd2f8 | 2362 | int timed_out = 1; |
22e2c507 | 2363 | |
7b679138 JA |
2364 | cfq_log(cfqd, "idle timer fired"); |
2365 | ||
22e2c507 JA |
2366 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
2367 | ||
fe094d98 JA |
2368 | cfqq = cfqd->active_queue; |
2369 | if (cfqq) { | |
3c6bd2f8 JA |
2370 | timed_out = 0; |
2371 | ||
b029195d JA |
2372 | /* |
2373 | * We saw a request before the queue expired, let it through | |
2374 | */ | |
2375 | if (cfq_cfqq_must_dispatch(cfqq)) | |
2376 | goto out_kick; | |
2377 | ||
22e2c507 JA |
2378 | /* |
2379 | * expired | |
2380 | */ | |
44f7c160 | 2381 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
2382 | goto expire; |
2383 | ||
2384 | /* | |
2385 | * only expire and reinvoke request handler, if there are | |
2386 | * other queues with pending requests | |
2387 | */ | |
caaa5f9f | 2388 | if (!cfqd->busy_queues) |
22e2c507 | 2389 | goto out_cont; |
22e2c507 JA |
2390 | |
2391 | /* | |
2392 | * not expired and it has a request pending, let it dispatch | |
2393 | */ | |
75e50984 | 2394 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 2395 | goto out_kick; |
22e2c507 JA |
2396 | } |
2397 | expire: | |
6084cdda | 2398 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 2399 | out_kick: |
3b18152c | 2400 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
2401 | out_cont: |
2402 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2403 | } | |
2404 | ||
3b18152c JA |
2405 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
2406 | { | |
2407 | del_timer_sync(&cfqd->idle_slice_timer); | |
64d01dc9 | 2408 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 2409 | } |
22e2c507 | 2410 | |
c2dea2d1 VT |
2411 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
2412 | { | |
2413 | int i; | |
2414 | ||
2415 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
2416 | if (cfqd->async_cfqq[0][i]) | |
2417 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
2418 | if (cfqd->async_cfqq[1][i]) | |
2419 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 2420 | } |
2389d1ef ON |
2421 | |
2422 | if (cfqd->async_idle_cfqq) | |
2423 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
2424 | } |
2425 | ||
b374d18a | 2426 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 2427 | { |
22e2c507 | 2428 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 2429 | struct request_queue *q = cfqd->queue; |
22e2c507 | 2430 | |
3b18152c | 2431 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 2432 | |
d9ff4187 | 2433 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 2434 | |
d9ff4187 | 2435 | if (cfqd->active_queue) |
6084cdda | 2436 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
2437 | |
2438 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
2439 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
2440 | struct cfq_io_context, | |
2441 | queue_list); | |
89850f7e JA |
2442 | |
2443 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 2444 | } |
e2d74ac0 | 2445 | |
c2dea2d1 | 2446 | cfq_put_async_queues(cfqd); |
15c31be4 | 2447 | |
d9ff4187 | 2448 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
2449 | |
2450 | cfq_shutdown_timer_wq(cfqd); | |
2451 | ||
a90d742e | 2452 | kfree(cfqd); |
1da177e4 LT |
2453 | } |
2454 | ||
165125e1 | 2455 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
2456 | { |
2457 | struct cfq_data *cfqd; | |
26a2ac00 | 2458 | int i; |
1da177e4 | 2459 | |
94f6030c | 2460 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 2461 | if (!cfqd) |
bc1c1169 | 2462 | return NULL; |
1da177e4 | 2463 | |
cc09e299 | 2464 | cfqd->service_tree = CFQ_RB_ROOT; |
26a2ac00 JA |
2465 | |
2466 | /* | |
2467 | * Not strictly needed (since RB_ROOT just clears the node and we | |
2468 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
2469 | * to add magic to the rb code | |
2470 | */ | |
2471 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
2472 | cfqd->prio_trees[i] = RB_ROOT; | |
2473 | ||
6118b70b JA |
2474 | /* |
2475 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
2476 | * Grab a permanent reference to it, so that the normal code flow | |
2477 | * will not attempt to free it. | |
2478 | */ | |
2479 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
2480 | atomic_inc(&cfqd->oom_cfqq.ref); | |
2481 | ||
d9ff4187 | 2482 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 2483 | |
1da177e4 | 2484 | cfqd->queue = q; |
1da177e4 | 2485 | |
22e2c507 JA |
2486 | init_timer(&cfqd->idle_slice_timer); |
2487 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
2488 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
2489 | ||
65f27f38 | 2490 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 2491 | |
1da177e4 | 2492 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
2493 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
2494 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
2495 | cfqd->cfq_back_max = cfq_back_max; |
2496 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
2497 | cfqd->cfq_slice[0] = cfq_slice_async; |
2498 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
2499 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
2500 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
45333d5a | 2501 | cfqd->hw_tag = 1; |
3b18152c | 2502 | |
bc1c1169 | 2503 | return cfqd; |
1da177e4 LT |
2504 | } |
2505 | ||
2506 | static void cfq_slab_kill(void) | |
2507 | { | |
d6de8be7 JA |
2508 | /* |
2509 | * Caller already ensured that pending RCU callbacks are completed, | |
2510 | * so we should have no busy allocations at this point. | |
2511 | */ | |
1da177e4 LT |
2512 | if (cfq_pool) |
2513 | kmem_cache_destroy(cfq_pool); | |
2514 | if (cfq_ioc_pool) | |
2515 | kmem_cache_destroy(cfq_ioc_pool); | |
2516 | } | |
2517 | ||
2518 | static int __init cfq_slab_setup(void) | |
2519 | { | |
0a31bd5f | 2520 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
2521 | if (!cfq_pool) |
2522 | goto fail; | |
2523 | ||
34e6bbf2 | 2524 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
2525 | if (!cfq_ioc_pool) |
2526 | goto fail; | |
2527 | ||
2528 | return 0; | |
2529 | fail: | |
2530 | cfq_slab_kill(); | |
2531 | return -ENOMEM; | |
2532 | } | |
2533 | ||
1da177e4 LT |
2534 | /* |
2535 | * sysfs parts below --> | |
2536 | */ | |
1da177e4 LT |
2537 | static ssize_t |
2538 | cfq_var_show(unsigned int var, char *page) | |
2539 | { | |
2540 | return sprintf(page, "%d\n", var); | |
2541 | } | |
2542 | ||
2543 | static ssize_t | |
2544 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
2545 | { | |
2546 | char *p = (char *) page; | |
2547 | ||
2548 | *var = simple_strtoul(p, &p, 10); | |
2549 | return count; | |
2550 | } | |
2551 | ||
1da177e4 | 2552 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 2553 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 2554 | { \ |
3d1ab40f | 2555 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2556 | unsigned int __data = __VAR; \ |
2557 | if (__CONV) \ | |
2558 | __data = jiffies_to_msecs(__data); \ | |
2559 | return cfq_var_show(__data, (page)); \ | |
2560 | } | |
2561 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
2562 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
2563 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
2564 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
2565 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
2566 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
2567 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
2568 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
2569 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
1da177e4 LT |
2570 | #undef SHOW_FUNCTION |
2571 | ||
2572 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 2573 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 2574 | { \ |
3d1ab40f | 2575 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2576 | unsigned int __data; \ |
2577 | int ret = cfq_var_store(&__data, (page), count); \ | |
2578 | if (__data < (MIN)) \ | |
2579 | __data = (MIN); \ | |
2580 | else if (__data > (MAX)) \ | |
2581 | __data = (MAX); \ | |
2582 | if (__CONV) \ | |
2583 | *(__PTR) = msecs_to_jiffies(__data); \ | |
2584 | else \ | |
2585 | *(__PTR) = __data; \ | |
2586 | return ret; \ | |
2587 | } | |
2588 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
2589 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
2590 | UINT_MAX, 1); | |
2591 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
2592 | UINT_MAX, 1); | |
e572ec7e | 2593 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
2594 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
2595 | UINT_MAX, 0); | |
22e2c507 JA |
2596 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
2597 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
2598 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
2599 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
2600 | UINT_MAX, 0); | |
1da177e4 LT |
2601 | #undef STORE_FUNCTION |
2602 | ||
e572ec7e AV |
2603 | #define CFQ_ATTR(name) \ |
2604 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
2605 | ||
2606 | static struct elv_fs_entry cfq_attrs[] = { | |
2607 | CFQ_ATTR(quantum), | |
e572ec7e AV |
2608 | CFQ_ATTR(fifo_expire_sync), |
2609 | CFQ_ATTR(fifo_expire_async), | |
2610 | CFQ_ATTR(back_seek_max), | |
2611 | CFQ_ATTR(back_seek_penalty), | |
2612 | CFQ_ATTR(slice_sync), | |
2613 | CFQ_ATTR(slice_async), | |
2614 | CFQ_ATTR(slice_async_rq), | |
2615 | CFQ_ATTR(slice_idle), | |
e572ec7e | 2616 | __ATTR_NULL |
1da177e4 LT |
2617 | }; |
2618 | ||
1da177e4 LT |
2619 | static struct elevator_type iosched_cfq = { |
2620 | .ops = { | |
2621 | .elevator_merge_fn = cfq_merge, | |
2622 | .elevator_merged_fn = cfq_merged_request, | |
2623 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 2624 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 2625 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 2626 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 2627 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
2628 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
2629 | .elevator_queue_empty_fn = cfq_queue_empty, | |
2630 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
2631 | .elevator_former_req_fn = elv_rb_former_request, |
2632 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
2633 | .elevator_set_req_fn = cfq_set_request, |
2634 | .elevator_put_req_fn = cfq_put_request, | |
2635 | .elevator_may_queue_fn = cfq_may_queue, | |
2636 | .elevator_init_fn = cfq_init_queue, | |
2637 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 2638 | .trim = cfq_free_io_context, |
1da177e4 | 2639 | }, |
3d1ab40f | 2640 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
2641 | .elevator_name = "cfq", |
2642 | .elevator_owner = THIS_MODULE, | |
2643 | }; | |
2644 | ||
2645 | static int __init cfq_init(void) | |
2646 | { | |
22e2c507 JA |
2647 | /* |
2648 | * could be 0 on HZ < 1000 setups | |
2649 | */ | |
2650 | if (!cfq_slice_async) | |
2651 | cfq_slice_async = 1; | |
2652 | if (!cfq_slice_idle) | |
2653 | cfq_slice_idle = 1; | |
2654 | ||
1da177e4 LT |
2655 | if (cfq_slab_setup()) |
2656 | return -ENOMEM; | |
2657 | ||
2fdd82bd | 2658 | elv_register(&iosched_cfq); |
1da177e4 | 2659 | |
2fdd82bd | 2660 | return 0; |
1da177e4 LT |
2661 | } |
2662 | ||
2663 | static void __exit cfq_exit(void) | |
2664 | { | |
6e9a4738 | 2665 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 2666 | elv_unregister(&iosched_cfq); |
334e94de | 2667 | ioc_gone = &all_gone; |
fba82272 OH |
2668 | /* ioc_gone's update must be visible before reading ioc_count */ |
2669 | smp_wmb(); | |
d6de8be7 JA |
2670 | |
2671 | /* | |
2672 | * this also protects us from entering cfq_slab_kill() with | |
2673 | * pending RCU callbacks | |
2674 | */ | |
4050cf16 | 2675 | if (elv_ioc_count_read(ioc_count)) |
9a11b4ed | 2676 | wait_for_completion(&all_gone); |
83521d3e | 2677 | cfq_slab_kill(); |
1da177e4 LT |
2678 | } |
2679 | ||
2680 | module_init(cfq_init); | |
2681 | module_exit(cfq_exit); | |
2682 | ||
2683 | MODULE_AUTHOR("Jens Axboe"); | |
2684 | MODULE_LICENSE("GPL"); | |
2685 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |