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 | |
245b2e70 | 51 | static DEFINE_PER_CPU(unsigned long, cfq_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 JA |
136 | unsigned int busy_queues; |
137 | ||
5ad531db | 138 | int rq_in_driver[2]; |
3ed9a296 | 139 | int sync_flight; |
45333d5a AC |
140 | |
141 | /* | |
142 | * queue-depth detection | |
143 | */ | |
144 | int rq_queued; | |
25776e35 | 145 | int hw_tag; |
45333d5a AC |
146 | int hw_tag_samples; |
147 | int rq_in_driver_peak; | |
1da177e4 | 148 | |
22e2c507 JA |
149 | /* |
150 | * idle window management | |
151 | */ | |
152 | struct timer_list idle_slice_timer; | |
23e018a1 | 153 | struct work_struct unplug_work; |
1da177e4 | 154 | |
22e2c507 JA |
155 | struct cfq_queue *active_queue; |
156 | struct cfq_io_context *active_cic; | |
22e2c507 | 157 | |
c2dea2d1 VT |
158 | /* |
159 | * async queue for each priority case | |
160 | */ | |
161 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
162 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 163 | |
6d048f53 | 164 | sector_t last_position; |
1da177e4 | 165 | |
1da177e4 LT |
166 | /* |
167 | * tunables, see top of file | |
168 | */ | |
169 | unsigned int cfq_quantum; | |
22e2c507 | 170 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
171 | unsigned int cfq_back_penalty; |
172 | unsigned int cfq_back_max; | |
22e2c507 JA |
173 | unsigned int cfq_slice[2]; |
174 | unsigned int cfq_slice_async_rq; | |
175 | unsigned int cfq_slice_idle; | |
963b72fc | 176 | unsigned int cfq_latency; |
d9ff4187 AV |
177 | |
178 | struct list_head cic_list; | |
1da177e4 | 179 | |
6118b70b JA |
180 | /* |
181 | * Fallback dummy cfqq for extreme OOM conditions | |
182 | */ | |
183 | struct cfq_queue oom_cfqq; | |
365722bb VG |
184 | |
185 | unsigned long last_end_sync_rq; | |
1da177e4 LT |
186 | }; |
187 | ||
3b18152c | 188 | enum cfqq_state_flags { |
b0b8d749 JA |
189 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
190 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 191 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 | 192 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ |
b0b8d749 JA |
193 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
194 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
195 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 196 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 197 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
a36e71f9 | 198 | CFQ_CFQQ_FLAG_coop, /* has done a coop jump of the queue */ |
3b18152c JA |
199 | }; |
200 | ||
201 | #define CFQ_CFQQ_FNS(name) \ | |
202 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
203 | { \ | |
fe094d98 | 204 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
205 | } \ |
206 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
207 | { \ | |
fe094d98 | 208 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
209 | } \ |
210 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
211 | { \ | |
fe094d98 | 212 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
213 | } |
214 | ||
215 | CFQ_CFQQ_FNS(on_rr); | |
216 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 217 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c | 218 | CFQ_CFQQ_FNS(must_alloc_slice); |
3b18152c JA |
219 | CFQ_CFQQ_FNS(fifo_expire); |
220 | CFQ_CFQQ_FNS(idle_window); | |
221 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 222 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 223 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 224 | CFQ_CFQQ_FNS(coop); |
3b18152c JA |
225 | #undef CFQ_CFQQ_FNS |
226 | ||
7b679138 JA |
227 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
228 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
229 | #define cfq_log(cfqd, fmt, args...) \ | |
230 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
231 | ||
165125e1 | 232 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
91fac317 | 233 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, int, |
fd0928df | 234 | struct io_context *, gfp_t); |
4ac845a2 | 235 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
236 | struct io_context *); |
237 | ||
5ad531db JA |
238 | static inline int rq_in_driver(struct cfq_data *cfqd) |
239 | { | |
240 | return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1]; | |
241 | } | |
242 | ||
91fac317 VT |
243 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, |
244 | int is_sync) | |
245 | { | |
246 | return cic->cfqq[!!is_sync]; | |
247 | } | |
248 | ||
249 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
250 | struct cfq_queue *cfqq, int is_sync) | |
251 | { | |
252 | cic->cfqq[!!is_sync] = cfqq; | |
253 | } | |
254 | ||
255 | /* | |
256 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
257 | * set (in which case it could also be direct WRITE). | |
258 | */ | |
259 | static inline int cfq_bio_sync(struct bio *bio) | |
260 | { | |
1f98a13f | 261 | if (bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO)) |
91fac317 VT |
262 | return 1; |
263 | ||
264 | return 0; | |
265 | } | |
1da177e4 | 266 | |
99f95e52 AM |
267 | /* |
268 | * scheduler run of queue, if there are requests pending and no one in the | |
269 | * driver that will restart queueing | |
270 | */ | |
23e018a1 | 271 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) |
99f95e52 | 272 | { |
7b679138 JA |
273 | if (cfqd->busy_queues) { |
274 | cfq_log(cfqd, "schedule dispatch"); | |
23e018a1 | 275 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 276 | } |
99f95e52 AM |
277 | } |
278 | ||
165125e1 | 279 | static int cfq_queue_empty(struct request_queue *q) |
99f95e52 AM |
280 | { |
281 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
282 | ||
b4878f24 | 283 | return !cfqd->busy_queues; |
99f95e52 AM |
284 | } |
285 | ||
44f7c160 JA |
286 | /* |
287 | * Scale schedule slice based on io priority. Use the sync time slice only | |
288 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
289 | * io only, should not get full sync slice length. | |
290 | */ | |
d9e7620e JA |
291 | static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync, |
292 | unsigned short prio) | |
44f7c160 | 293 | { |
d9e7620e | 294 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 295 | |
d9e7620e JA |
296 | WARN_ON(prio >= IOPRIO_BE_NR); |
297 | ||
298 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
299 | } | |
44f7c160 | 300 | |
d9e7620e JA |
301 | static inline int |
302 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
303 | { | |
304 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
305 | } |
306 | ||
307 | static inline void | |
308 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
309 | { | |
310 | cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies; | |
7b679138 | 311 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
312 | } |
313 | ||
314 | /* | |
315 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
316 | * isn't valid until the first request from the dispatch is activated | |
317 | * and the slice time set. | |
318 | */ | |
319 | static inline int cfq_slice_used(struct cfq_queue *cfqq) | |
320 | { | |
321 | if (cfq_cfqq_slice_new(cfqq)) | |
322 | return 0; | |
323 | if (time_before(jiffies, cfqq->slice_end)) | |
324 | return 0; | |
325 | ||
326 | return 1; | |
327 | } | |
328 | ||
1da177e4 | 329 | /* |
5e705374 | 330 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 331 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 332 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 333 | */ |
5e705374 JA |
334 | static struct request * |
335 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2) | |
1da177e4 LT |
336 | { |
337 | sector_t last, s1, s2, d1 = 0, d2 = 0; | |
1da177e4 | 338 | unsigned long back_max; |
e8a99053 AM |
339 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
340 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
341 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 342 | |
5e705374 JA |
343 | if (rq1 == NULL || rq1 == rq2) |
344 | return rq2; | |
345 | if (rq2 == NULL) | |
346 | return rq1; | |
9c2c38a1 | 347 | |
5e705374 JA |
348 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
349 | return rq1; | |
350 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
351 | return rq2; | |
374f84ac JA |
352 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
353 | return rq1; | |
354 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
355 | return rq2; | |
1da177e4 | 356 | |
83096ebf TH |
357 | s1 = blk_rq_pos(rq1); |
358 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 359 | |
6d048f53 | 360 | last = cfqd->last_position; |
1da177e4 | 361 | |
1da177e4 LT |
362 | /* |
363 | * by definition, 1KiB is 2 sectors | |
364 | */ | |
365 | back_max = cfqd->cfq_back_max * 2; | |
366 | ||
367 | /* | |
368 | * Strict one way elevator _except_ in the case where we allow | |
369 | * short backward seeks which are biased as twice the cost of a | |
370 | * similar forward seek. | |
371 | */ | |
372 | if (s1 >= last) | |
373 | d1 = s1 - last; | |
374 | else if (s1 + back_max >= last) | |
375 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
376 | else | |
e8a99053 | 377 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
378 | |
379 | if (s2 >= last) | |
380 | d2 = s2 - last; | |
381 | else if (s2 + back_max >= last) | |
382 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
383 | else | |
e8a99053 | 384 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
385 | |
386 | /* Found required data */ | |
e8a99053 AM |
387 | |
388 | /* | |
389 | * By doing switch() on the bit mask "wrap" we avoid having to | |
390 | * check two variables for all permutations: --> faster! | |
391 | */ | |
392 | switch (wrap) { | |
5e705374 | 393 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 394 | if (d1 < d2) |
5e705374 | 395 | return rq1; |
e8a99053 | 396 | else if (d2 < d1) |
5e705374 | 397 | return rq2; |
e8a99053 AM |
398 | else { |
399 | if (s1 >= s2) | |
5e705374 | 400 | return rq1; |
e8a99053 | 401 | else |
5e705374 | 402 | return rq2; |
e8a99053 | 403 | } |
1da177e4 | 404 | |
e8a99053 | 405 | case CFQ_RQ2_WRAP: |
5e705374 | 406 | return rq1; |
e8a99053 | 407 | case CFQ_RQ1_WRAP: |
5e705374 JA |
408 | return rq2; |
409 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
410 | default: |
411 | /* | |
412 | * Since both rqs are wrapped, | |
413 | * start with the one that's further behind head | |
414 | * (--> only *one* back seek required), | |
415 | * since back seek takes more time than forward. | |
416 | */ | |
417 | if (s1 <= s2) | |
5e705374 | 418 | return rq1; |
1da177e4 | 419 | else |
5e705374 | 420 | return rq2; |
1da177e4 LT |
421 | } |
422 | } | |
423 | ||
498d3aa2 JA |
424 | /* |
425 | * The below is leftmost cache rbtree addon | |
426 | */ | |
0871714e | 427 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 JA |
428 | { |
429 | if (!root->left) | |
430 | root->left = rb_first(&root->rb); | |
431 | ||
0871714e JA |
432 | if (root->left) |
433 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
434 | ||
435 | return NULL; | |
cc09e299 JA |
436 | } |
437 | ||
a36e71f9 JA |
438 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
439 | { | |
440 | rb_erase(n, root); | |
441 | RB_CLEAR_NODE(n); | |
442 | } | |
443 | ||
cc09e299 JA |
444 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
445 | { | |
446 | if (root->left == n) | |
447 | root->left = NULL; | |
a36e71f9 | 448 | rb_erase_init(n, &root->rb); |
cc09e299 JA |
449 | } |
450 | ||
1da177e4 LT |
451 | /* |
452 | * would be nice to take fifo expire time into account as well | |
453 | */ | |
5e705374 JA |
454 | static struct request * |
455 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
456 | struct request *last) | |
1da177e4 | 457 | { |
21183b07 JA |
458 | struct rb_node *rbnext = rb_next(&last->rb_node); |
459 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 460 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 461 | |
21183b07 | 462 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
463 | |
464 | if (rbprev) | |
5e705374 | 465 | prev = rb_entry_rq(rbprev); |
1da177e4 | 466 | |
21183b07 | 467 | if (rbnext) |
5e705374 | 468 | next = rb_entry_rq(rbnext); |
21183b07 JA |
469 | else { |
470 | rbnext = rb_first(&cfqq->sort_list); | |
471 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 472 | next = rb_entry_rq(rbnext); |
21183b07 | 473 | } |
1da177e4 | 474 | |
21183b07 | 475 | return cfq_choose_req(cfqd, next, prev); |
1da177e4 LT |
476 | } |
477 | ||
d9e7620e JA |
478 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
479 | struct cfq_queue *cfqq) | |
1da177e4 | 480 | { |
d9e7620e JA |
481 | /* |
482 | * just an approximation, should be ok. | |
483 | */ | |
67e6b49e JA |
484 | return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
485 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); | |
d9e7620e JA |
486 | } |
487 | ||
498d3aa2 JA |
488 | /* |
489 | * The cfqd->service_tree holds all pending cfq_queue's that have | |
490 | * requests waiting to be processed. It is sorted in the order that | |
491 | * we will service the queues. | |
492 | */ | |
a36e71f9 JA |
493 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
494 | int add_front) | |
d9e7620e | 495 | { |
0871714e JA |
496 | struct rb_node **p, *parent; |
497 | struct cfq_queue *__cfqq; | |
d9e7620e | 498 | unsigned long rb_key; |
498d3aa2 | 499 | int left; |
d9e7620e | 500 | |
0871714e JA |
501 | if (cfq_class_idle(cfqq)) { |
502 | rb_key = CFQ_IDLE_DELAY; | |
503 | parent = rb_last(&cfqd->service_tree.rb); | |
504 | if (parent && parent != &cfqq->rb_node) { | |
505 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
506 | rb_key += __cfqq->rb_key; | |
507 | } else | |
508 | rb_key += jiffies; | |
509 | } else if (!add_front) { | |
b9c8946b JA |
510 | /* |
511 | * Get our rb key offset. Subtract any residual slice | |
512 | * value carried from last service. A negative resid | |
513 | * count indicates slice overrun, and this should position | |
514 | * the next service time further away in the tree. | |
515 | */ | |
edd75ffd | 516 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
b9c8946b | 517 | rb_key -= cfqq->slice_resid; |
edd75ffd | 518 | cfqq->slice_resid = 0; |
48e025e6 CZ |
519 | } else { |
520 | rb_key = -HZ; | |
521 | __cfqq = cfq_rb_first(&cfqd->service_tree); | |
522 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; | |
523 | } | |
1da177e4 | 524 | |
d9e7620e | 525 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
99f9628a | 526 | /* |
d9e7620e | 527 | * same position, nothing more to do |
99f9628a | 528 | */ |
d9e7620e JA |
529 | if (rb_key == cfqq->rb_key) |
530 | return; | |
1da177e4 | 531 | |
cc09e299 | 532 | cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree); |
1da177e4 | 533 | } |
d9e7620e | 534 | |
498d3aa2 | 535 | left = 1; |
0871714e JA |
536 | parent = NULL; |
537 | p = &cfqd->service_tree.rb.rb_node; | |
d9e7620e | 538 | while (*p) { |
67060e37 | 539 | struct rb_node **n; |
cc09e299 | 540 | |
d9e7620e JA |
541 | parent = *p; |
542 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
543 | ||
0c534e0a JA |
544 | /* |
545 | * sort RT queues first, we always want to give | |
67060e37 JA |
546 | * preference to them. IDLE queues goes to the back. |
547 | * after that, sort on the next service time. | |
0c534e0a JA |
548 | */ |
549 | if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq)) | |
67060e37 | 550 | n = &(*p)->rb_left; |
0c534e0a | 551 | else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq)) |
67060e37 JA |
552 | n = &(*p)->rb_right; |
553 | else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq)) | |
554 | n = &(*p)->rb_left; | |
555 | else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq)) | |
556 | n = &(*p)->rb_right; | |
48e025e6 | 557 | else if (time_before(rb_key, __cfqq->rb_key)) |
67060e37 JA |
558 | n = &(*p)->rb_left; |
559 | else | |
560 | n = &(*p)->rb_right; | |
561 | ||
562 | if (n == &(*p)->rb_right) | |
cc09e299 | 563 | left = 0; |
67060e37 JA |
564 | |
565 | p = n; | |
d9e7620e JA |
566 | } |
567 | ||
cc09e299 JA |
568 | if (left) |
569 | cfqd->service_tree.left = &cfqq->rb_node; | |
570 | ||
d9e7620e JA |
571 | cfqq->rb_key = rb_key; |
572 | rb_link_node(&cfqq->rb_node, parent, p); | |
cc09e299 | 573 | rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb); |
1da177e4 LT |
574 | } |
575 | ||
a36e71f9 | 576 | static struct cfq_queue * |
f2d1f0ae JA |
577 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
578 | sector_t sector, struct rb_node **ret_parent, | |
579 | struct rb_node ***rb_link) | |
a36e71f9 | 580 | { |
a36e71f9 JA |
581 | struct rb_node **p, *parent; |
582 | struct cfq_queue *cfqq = NULL; | |
583 | ||
584 | parent = NULL; | |
585 | p = &root->rb_node; | |
586 | while (*p) { | |
587 | struct rb_node **n; | |
588 | ||
589 | parent = *p; | |
590 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
591 | ||
592 | /* | |
593 | * Sort strictly based on sector. Smallest to the left, | |
594 | * largest to the right. | |
595 | */ | |
2e46e8b2 | 596 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 597 | n = &(*p)->rb_right; |
2e46e8b2 | 598 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
599 | n = &(*p)->rb_left; |
600 | else | |
601 | break; | |
602 | p = n; | |
3ac6c9f8 | 603 | cfqq = NULL; |
a36e71f9 JA |
604 | } |
605 | ||
606 | *ret_parent = parent; | |
607 | if (rb_link) | |
608 | *rb_link = p; | |
3ac6c9f8 | 609 | return cfqq; |
a36e71f9 JA |
610 | } |
611 | ||
612 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
613 | { | |
a36e71f9 JA |
614 | struct rb_node **p, *parent; |
615 | struct cfq_queue *__cfqq; | |
616 | ||
f2d1f0ae JA |
617 | if (cfqq->p_root) { |
618 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
619 | cfqq->p_root = NULL; | |
620 | } | |
a36e71f9 JA |
621 | |
622 | if (cfq_class_idle(cfqq)) | |
623 | return; | |
624 | if (!cfqq->next_rq) | |
625 | return; | |
626 | ||
f2d1f0ae | 627 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
628 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
629 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
630 | if (!__cfqq) { |
631 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
632 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
633 | } else | |
634 | cfqq->p_root = NULL; | |
a36e71f9 JA |
635 | } |
636 | ||
498d3aa2 JA |
637 | /* |
638 | * Update cfqq's position in the service tree. | |
639 | */ | |
edd75ffd | 640 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 641 | { |
6d048f53 JA |
642 | /* |
643 | * Resorting requires the cfqq to be on the RR list already. | |
644 | */ | |
a36e71f9 | 645 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 646 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
647 | cfq_prio_tree_add(cfqd, cfqq); |
648 | } | |
6d048f53 JA |
649 | } |
650 | ||
1da177e4 LT |
651 | /* |
652 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 653 | * the pending list according to last request service |
1da177e4 | 654 | */ |
febffd61 | 655 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 656 | { |
7b679138 | 657 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
658 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
659 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 LT |
660 | cfqd->busy_queues++; |
661 | ||
edd75ffd | 662 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
663 | } |
664 | ||
498d3aa2 JA |
665 | /* |
666 | * Called when the cfqq no longer has requests pending, remove it from | |
667 | * the service tree. | |
668 | */ | |
febffd61 | 669 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 670 | { |
7b679138 | 671 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
672 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
673 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 674 | |
cc09e299 JA |
675 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) |
676 | cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree); | |
f2d1f0ae JA |
677 | if (cfqq->p_root) { |
678 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
679 | cfqq->p_root = NULL; | |
680 | } | |
d9e7620e | 681 | |
1da177e4 LT |
682 | BUG_ON(!cfqd->busy_queues); |
683 | cfqd->busy_queues--; | |
684 | } | |
685 | ||
686 | /* | |
687 | * rb tree support functions | |
688 | */ | |
febffd61 | 689 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 690 | { |
5e705374 | 691 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 692 | struct cfq_data *cfqd = cfqq->cfqd; |
5e705374 | 693 | const int sync = rq_is_sync(rq); |
1da177e4 | 694 | |
b4878f24 JA |
695 | BUG_ON(!cfqq->queued[sync]); |
696 | cfqq->queued[sync]--; | |
1da177e4 | 697 | |
5e705374 | 698 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 699 | |
dd67d051 | 700 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
b4878f24 | 701 | cfq_del_cfqq_rr(cfqd, cfqq); |
1da177e4 LT |
702 | } |
703 | ||
5e705374 | 704 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 705 | { |
5e705374 | 706 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 707 | struct cfq_data *cfqd = cfqq->cfqd; |
a36e71f9 | 708 | struct request *__alias, *prev; |
1da177e4 | 709 | |
5380a101 | 710 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
711 | |
712 | /* | |
713 | * looks a little odd, but the first insert might return an alias. | |
714 | * if that happens, put the alias on the dispatch list | |
715 | */ | |
21183b07 | 716 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 717 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
718 | |
719 | if (!cfq_cfqq_on_rr(cfqq)) | |
720 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
721 | |
722 | /* | |
723 | * check if this request is a better next-serve candidate | |
724 | */ | |
a36e71f9 | 725 | prev = cfqq->next_rq; |
5044eed4 | 726 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq); |
a36e71f9 JA |
727 | |
728 | /* | |
729 | * adjust priority tree position, if ->next_rq changes | |
730 | */ | |
731 | if (prev != cfqq->next_rq) | |
732 | cfq_prio_tree_add(cfqd, cfqq); | |
733 | ||
5044eed4 | 734 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
735 | } |
736 | ||
febffd61 | 737 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 738 | { |
5380a101 JA |
739 | elv_rb_del(&cfqq->sort_list, rq); |
740 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 741 | cfq_add_rq_rb(rq); |
1da177e4 LT |
742 | } |
743 | ||
206dc69b JA |
744 | static struct request * |
745 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 746 | { |
206dc69b | 747 | struct task_struct *tsk = current; |
91fac317 | 748 | struct cfq_io_context *cic; |
206dc69b | 749 | struct cfq_queue *cfqq; |
1da177e4 | 750 | |
4ac845a2 | 751 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
752 | if (!cic) |
753 | return NULL; | |
754 | ||
755 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
756 | if (cfqq) { |
757 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
758 | ||
21183b07 | 759 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 760 | } |
1da177e4 | 761 | |
1da177e4 LT |
762 | return NULL; |
763 | } | |
764 | ||
165125e1 | 765 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 766 | { |
22e2c507 | 767 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 768 | |
5ad531db | 769 | cfqd->rq_in_driver[rq_is_sync(rq)]++; |
7b679138 | 770 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
5ad531db | 771 | rq_in_driver(cfqd)); |
25776e35 | 772 | |
5b93629b | 773 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
774 | } |
775 | ||
165125e1 | 776 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 777 | { |
b4878f24 | 778 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5ad531db | 779 | const int sync = rq_is_sync(rq); |
b4878f24 | 780 | |
5ad531db JA |
781 | WARN_ON(!cfqd->rq_in_driver[sync]); |
782 | cfqd->rq_in_driver[sync]--; | |
7b679138 | 783 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
5ad531db | 784 | rq_in_driver(cfqd)); |
1da177e4 LT |
785 | } |
786 | ||
b4878f24 | 787 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 788 | { |
5e705374 | 789 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 790 | |
5e705374 JA |
791 | if (cfqq->next_rq == rq) |
792 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 793 | |
b4878f24 | 794 | list_del_init(&rq->queuelist); |
5e705374 | 795 | cfq_del_rq_rb(rq); |
374f84ac | 796 | |
45333d5a | 797 | cfqq->cfqd->rq_queued--; |
374f84ac JA |
798 | if (rq_is_meta(rq)) { |
799 | WARN_ON(!cfqq->meta_pending); | |
800 | cfqq->meta_pending--; | |
801 | } | |
1da177e4 LT |
802 | } |
803 | ||
165125e1 JA |
804 | static int cfq_merge(struct request_queue *q, struct request **req, |
805 | struct bio *bio) | |
1da177e4 LT |
806 | { |
807 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
808 | struct request *__rq; | |
1da177e4 | 809 | |
206dc69b | 810 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 811 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
812 | *req = __rq; |
813 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
814 | } |
815 | ||
816 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
817 | } |
818 | ||
165125e1 | 819 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 820 | int type) |
1da177e4 | 821 | { |
21183b07 | 822 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 823 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 824 | |
5e705374 | 825 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 826 | } |
1da177e4 LT |
827 | } |
828 | ||
829 | static void | |
165125e1 | 830 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
831 | struct request *next) |
832 | { | |
22e2c507 JA |
833 | /* |
834 | * reposition in fifo if next is older than rq | |
835 | */ | |
836 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
30996f40 | 837 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
22e2c507 | 838 | list_move(&rq->queuelist, &next->queuelist); |
30996f40 JA |
839 | rq_set_fifo_time(rq, rq_fifo_time(next)); |
840 | } | |
22e2c507 | 841 | |
b4878f24 | 842 | cfq_remove_request(next); |
22e2c507 JA |
843 | } |
844 | ||
165125e1 | 845 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
846 | struct bio *bio) |
847 | { | |
848 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 849 | struct cfq_io_context *cic; |
da775265 | 850 | struct cfq_queue *cfqq; |
da775265 JA |
851 | |
852 | /* | |
ec8acb69 | 853 | * Disallow merge of a sync bio into an async request. |
da775265 | 854 | */ |
91fac317 | 855 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
da775265 JA |
856 | return 0; |
857 | ||
858 | /* | |
719d3402 JA |
859 | * Lookup the cfqq that this bio will be queued with. Allow |
860 | * merge only if rq is queued there. | |
da775265 | 861 | */ |
4ac845a2 | 862 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 VT |
863 | if (!cic) |
864 | return 0; | |
719d3402 | 865 | |
91fac317 | 866 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
719d3402 JA |
867 | if (cfqq == RQ_CFQQ(rq)) |
868 | return 1; | |
da775265 | 869 | |
ec8acb69 | 870 | return 0; |
da775265 JA |
871 | } |
872 | ||
febffd61 JA |
873 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
874 | struct cfq_queue *cfqq) | |
22e2c507 JA |
875 | { |
876 | if (cfqq) { | |
7b679138 | 877 | cfq_log_cfqq(cfqd, cfqq, "set_active"); |
22e2c507 | 878 | cfqq->slice_end = 0; |
2f5cb738 JA |
879 | cfqq->slice_dispatch = 0; |
880 | ||
2f5cb738 | 881 | cfq_clear_cfqq_wait_request(cfqq); |
b029195d | 882 | cfq_clear_cfqq_must_dispatch(cfqq); |
3b18152c JA |
883 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
884 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 885 | cfq_mark_cfqq_slice_new(cfqq); |
2f5cb738 JA |
886 | |
887 | del_timer(&cfqd->idle_slice_timer); | |
22e2c507 JA |
888 | } |
889 | ||
890 | cfqd->active_queue = cfqq; | |
891 | } | |
892 | ||
7b14e3b5 JA |
893 | /* |
894 | * current cfqq expired its slice (or was too idle), select new one | |
895 | */ | |
896 | static void | |
897 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
6084cdda | 898 | int timed_out) |
7b14e3b5 | 899 | { |
7b679138 JA |
900 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
901 | ||
7b14e3b5 JA |
902 | if (cfq_cfqq_wait_request(cfqq)) |
903 | del_timer(&cfqd->idle_slice_timer); | |
904 | ||
7b14e3b5 JA |
905 | cfq_clear_cfqq_wait_request(cfqq); |
906 | ||
907 | /* | |
6084cdda | 908 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 909 | */ |
7b679138 | 910 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) { |
c5b680f3 | 911 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b679138 JA |
912 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
913 | } | |
7b14e3b5 | 914 | |
edd75ffd | 915 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
916 | |
917 | if (cfqq == cfqd->active_queue) | |
918 | cfqd->active_queue = NULL; | |
919 | ||
920 | if (cfqd->active_cic) { | |
921 | put_io_context(cfqd->active_cic->ioc); | |
922 | cfqd->active_cic = NULL; | |
923 | } | |
7b14e3b5 JA |
924 | } |
925 | ||
6084cdda | 926 | static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out) |
7b14e3b5 JA |
927 | { |
928 | struct cfq_queue *cfqq = cfqd->active_queue; | |
929 | ||
930 | if (cfqq) | |
6084cdda | 931 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
932 | } |
933 | ||
498d3aa2 JA |
934 | /* |
935 | * Get next queue for service. Unless we have a queue preemption, | |
936 | * we'll simply select the first cfqq in the service tree. | |
937 | */ | |
6d048f53 | 938 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 939 | { |
edd75ffd JA |
940 | if (RB_EMPTY_ROOT(&cfqd->service_tree.rb)) |
941 | return NULL; | |
d9e7620e | 942 | |
0871714e | 943 | return cfq_rb_first(&cfqd->service_tree); |
6d048f53 JA |
944 | } |
945 | ||
498d3aa2 JA |
946 | /* |
947 | * Get and set a new active queue for service. | |
948 | */ | |
a36e71f9 JA |
949 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
950 | struct cfq_queue *cfqq) | |
6d048f53 | 951 | { |
a36e71f9 JA |
952 | if (!cfqq) { |
953 | cfqq = cfq_get_next_queue(cfqd); | |
954 | if (cfqq) | |
955 | cfq_clear_cfqq_coop(cfqq); | |
956 | } | |
6d048f53 | 957 | |
22e2c507 | 958 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 959 | return cfqq; |
22e2c507 JA |
960 | } |
961 | ||
d9e7620e JA |
962 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
963 | struct request *rq) | |
964 | { | |
83096ebf TH |
965 | if (blk_rq_pos(rq) >= cfqd->last_position) |
966 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 967 | else |
83096ebf | 968 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
969 | } |
970 | ||
04dc6e71 JM |
971 | #define CIC_SEEK_THR 8 * 1024 |
972 | #define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR) | |
973 | ||
6d048f53 JA |
974 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq) |
975 | { | |
976 | struct cfq_io_context *cic = cfqd->active_cic; | |
04dc6e71 | 977 | sector_t sdist = cic->seek_mean; |
6d048f53 JA |
978 | |
979 | if (!sample_valid(cic->seek_samples)) | |
04dc6e71 | 980 | sdist = CIC_SEEK_THR; |
6d048f53 | 981 | |
04dc6e71 | 982 | return cfq_dist_from_last(cfqd, rq) <= sdist; |
6d048f53 JA |
983 | } |
984 | ||
a36e71f9 JA |
985 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
986 | struct cfq_queue *cur_cfqq) | |
987 | { | |
f2d1f0ae | 988 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
989 | struct rb_node *parent, *node; |
990 | struct cfq_queue *__cfqq; | |
991 | sector_t sector = cfqd->last_position; | |
992 | ||
993 | if (RB_EMPTY_ROOT(root)) | |
994 | return NULL; | |
995 | ||
996 | /* | |
997 | * First, if we find a request starting at the end of the last | |
998 | * request, choose it. | |
999 | */ | |
f2d1f0ae | 1000 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
1001 | if (__cfqq) |
1002 | return __cfqq; | |
1003 | ||
1004 | /* | |
1005 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1006 | * will contain the closest sector. | |
1007 | */ | |
1008 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
1009 | if (cfq_rq_close(cfqd, __cfqq->next_rq)) | |
1010 | return __cfqq; | |
1011 | ||
2e46e8b2 | 1012 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1013 | node = rb_next(&__cfqq->p_node); |
1014 | else | |
1015 | node = rb_prev(&__cfqq->p_node); | |
1016 | if (!node) | |
1017 | return NULL; | |
1018 | ||
1019 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
1020 | if (cfq_rq_close(cfqd, __cfqq->next_rq)) | |
1021 | return __cfqq; | |
1022 | ||
1023 | return NULL; | |
1024 | } | |
1025 | ||
1026 | /* | |
1027 | * cfqd - obvious | |
1028 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1029 | * closely cooperating with itself. | |
1030 | * | |
1031 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1032 | * one request, and that cfqd->last_position reflects a position on the disk | |
1033 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1034 | * assumption. | |
1035 | */ | |
1036 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
1037 | struct cfq_queue *cur_cfqq, | |
1038 | int probe) | |
6d048f53 | 1039 | { |
a36e71f9 JA |
1040 | struct cfq_queue *cfqq; |
1041 | ||
1042 | /* | |
1043 | * A valid cfq_io_context is necessary to compare requests against | |
1044 | * the seek_mean of the current cfqq. | |
1045 | */ | |
1046 | if (!cfqd->active_cic) | |
1047 | return NULL; | |
1048 | ||
6d048f53 | 1049 | /* |
d9e7620e JA |
1050 | * We should notice if some of the queues are cooperating, eg |
1051 | * working closely on the same area of the disk. In that case, | |
1052 | * we can group them together and don't waste time idling. | |
6d048f53 | 1053 | */ |
a36e71f9 JA |
1054 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1055 | if (!cfqq) | |
1056 | return NULL; | |
1057 | ||
1058 | if (cfq_cfqq_coop(cfqq)) | |
1059 | return NULL; | |
1060 | ||
1061 | if (!probe) | |
1062 | cfq_mark_cfqq_coop(cfqq); | |
1063 | return cfqq; | |
6d048f53 JA |
1064 | } |
1065 | ||
6d048f53 | 1066 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1067 | { |
1792669c | 1068 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1069 | struct cfq_io_context *cic; |
7b14e3b5 JA |
1070 | unsigned long sl; |
1071 | ||
a68bbddb | 1072 | /* |
f7d7b7a7 JA |
1073 | * SSD device without seek penalty, disable idling. But only do so |
1074 | * for devices that support queuing, otherwise we still have a problem | |
1075 | * with sync vs async workloads. | |
a68bbddb | 1076 | */ |
f7d7b7a7 | 1077 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
1078 | return; |
1079 | ||
dd67d051 | 1080 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 1081 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
1082 | |
1083 | /* | |
1084 | * idle is disabled, either manually or by past process history | |
1085 | */ | |
6d048f53 JA |
1086 | if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq)) |
1087 | return; | |
1088 | ||
7b679138 JA |
1089 | /* |
1090 | * still requests with the driver, don't idle | |
1091 | */ | |
5ad531db | 1092 | if (rq_in_driver(cfqd)) |
7b679138 JA |
1093 | return; |
1094 | ||
22e2c507 JA |
1095 | /* |
1096 | * task has exited, don't wait | |
1097 | */ | |
206dc69b | 1098 | cic = cfqd->active_cic; |
66dac98e | 1099 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
1100 | return; |
1101 | ||
3b18152c | 1102 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 1103 | |
206dc69b JA |
1104 | /* |
1105 | * we don't want to idle for seeks, but we do want to allow | |
1106 | * fair distribution of slice time for a process doing back-to-back | |
1107 | * seeks. so allow a little bit of time for him to submit a new rq | |
1108 | */ | |
6d048f53 | 1109 | sl = cfqd->cfq_slice_idle; |
caaa5f9f | 1110 | if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic)) |
d9e7620e | 1111 | sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT)); |
206dc69b | 1112 | |
7b14e3b5 | 1113 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
9481ffdc | 1114 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); |
1da177e4 LT |
1115 | } |
1116 | ||
498d3aa2 JA |
1117 | /* |
1118 | * Move request from internal lists to the request queue dispatch list. | |
1119 | */ | |
165125e1 | 1120 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 1121 | { |
3ed9a296 | 1122 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1123 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 1124 | |
7b679138 JA |
1125 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
1126 | ||
06d21886 | 1127 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); |
5380a101 | 1128 | cfq_remove_request(rq); |
6d048f53 | 1129 | cfqq->dispatched++; |
5380a101 | 1130 | elv_dispatch_sort(q, rq); |
3ed9a296 JA |
1131 | |
1132 | if (cfq_cfqq_sync(cfqq)) | |
1133 | cfqd->sync_flight++; | |
1da177e4 LT |
1134 | } |
1135 | ||
1136 | /* | |
1137 | * return expired entry, or NULL to just start from scratch in rbtree | |
1138 | */ | |
febffd61 | 1139 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 | 1140 | { |
30996f40 | 1141 | struct request *rq = NULL; |
1da177e4 | 1142 | |
3b18152c | 1143 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 1144 | return NULL; |
cb887411 JA |
1145 | |
1146 | cfq_mark_cfqq_fifo_expire(cfqq); | |
1147 | ||
89850f7e JA |
1148 | if (list_empty(&cfqq->fifo)) |
1149 | return NULL; | |
1da177e4 | 1150 | |
89850f7e | 1151 | rq = rq_entry_fifo(cfqq->fifo.next); |
30996f40 | 1152 | if (time_before(jiffies, rq_fifo_time(rq))) |
7b679138 | 1153 | rq = NULL; |
1da177e4 | 1154 | |
30996f40 | 1155 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 1156 | return rq; |
1da177e4 LT |
1157 | } |
1158 | ||
22e2c507 JA |
1159 | static inline int |
1160 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1161 | { | |
1162 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 1163 | |
22e2c507 | 1164 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 1165 | |
22e2c507 | 1166 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
1167 | } |
1168 | ||
22e2c507 | 1169 | /* |
498d3aa2 JA |
1170 | * Select a queue for service. If we have a current active queue, |
1171 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 1172 | */ |
1b5ed5e1 | 1173 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 1174 | { |
a36e71f9 | 1175 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 1176 | |
22e2c507 JA |
1177 | cfqq = cfqd->active_queue; |
1178 | if (!cfqq) | |
1179 | goto new_queue; | |
1da177e4 | 1180 | |
22e2c507 | 1181 | /* |
6d048f53 | 1182 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 1183 | */ |
b029195d | 1184 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) |
3b18152c | 1185 | goto expire; |
1da177e4 | 1186 | |
22e2c507 | 1187 | /* |
6d048f53 JA |
1188 | * The active queue has requests and isn't expired, allow it to |
1189 | * dispatch. | |
22e2c507 | 1190 | */ |
dd67d051 | 1191 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 1192 | goto keep_queue; |
6d048f53 | 1193 | |
a36e71f9 JA |
1194 | /* |
1195 | * If another queue has a request waiting within our mean seek | |
1196 | * distance, let it run. The expire code will check for close | |
1197 | * cooperators and put the close queue at the front of the service | |
1198 | * tree. | |
1199 | */ | |
1200 | new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0); | |
1201 | if (new_cfqq) | |
1202 | goto expire; | |
1203 | ||
6d048f53 JA |
1204 | /* |
1205 | * No requests pending. If the active queue still has requests in | |
1206 | * flight or is idling for a new request, allow either of these | |
1207 | * conditions to happen (or time out) before selecting a new queue. | |
1208 | */ | |
cc197479 JA |
1209 | if (timer_pending(&cfqd->idle_slice_timer) || |
1210 | (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) { | |
caaa5f9f JA |
1211 | cfqq = NULL; |
1212 | goto keep_queue; | |
22e2c507 JA |
1213 | } |
1214 | ||
3b18152c | 1215 | expire: |
6084cdda | 1216 | cfq_slice_expired(cfqd, 0); |
3b18152c | 1217 | new_queue: |
a36e71f9 | 1218 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 1219 | keep_queue: |
3b18152c | 1220 | return cfqq; |
22e2c507 JA |
1221 | } |
1222 | ||
febffd61 | 1223 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
1224 | { |
1225 | int dispatched = 0; | |
1226 | ||
1227 | while (cfqq->next_rq) { | |
1228 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
1229 | dispatched++; | |
1230 | } | |
1231 | ||
1232 | BUG_ON(!list_empty(&cfqq->fifo)); | |
1233 | return dispatched; | |
1234 | } | |
1235 | ||
498d3aa2 JA |
1236 | /* |
1237 | * Drain our current requests. Used for barriers and when switching | |
1238 | * io schedulers on-the-fly. | |
1239 | */ | |
d9e7620e | 1240 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 1241 | { |
0871714e | 1242 | struct cfq_queue *cfqq; |
d9e7620e | 1243 | int dispatched = 0; |
1b5ed5e1 | 1244 | |
0871714e | 1245 | while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL) |
d9e7620e | 1246 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); |
1b5ed5e1 | 1247 | |
6084cdda | 1248 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
1249 | |
1250 | BUG_ON(cfqd->busy_queues); | |
1251 | ||
6923715a | 1252 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
1253 | return dispatched; |
1254 | } | |
1255 | ||
0b182d61 | 1256 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
2f5cb738 | 1257 | { |
2f5cb738 | 1258 | unsigned int max_dispatch; |
22e2c507 | 1259 | |
5ad531db JA |
1260 | /* |
1261 | * Drain async requests before we start sync IO | |
1262 | */ | |
1263 | if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC]) | |
0b182d61 | 1264 | return false; |
5ad531db | 1265 | |
2f5cb738 JA |
1266 | /* |
1267 | * If this is an async queue and we have sync IO in flight, let it wait | |
1268 | */ | |
1269 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) | |
0b182d61 | 1270 | return false; |
2f5cb738 JA |
1271 | |
1272 | max_dispatch = cfqd->cfq_quantum; | |
1273 | if (cfq_class_idle(cfqq)) | |
1274 | max_dispatch = 1; | |
b4878f24 | 1275 | |
2f5cb738 JA |
1276 | /* |
1277 | * Does this cfqq already have too much IO in flight? | |
1278 | */ | |
1279 | if (cfqq->dispatched >= max_dispatch) { | |
1280 | /* | |
1281 | * idle queue must always only have a single IO in flight | |
1282 | */ | |
3ed9a296 | 1283 | if (cfq_class_idle(cfqq)) |
0b182d61 | 1284 | return false; |
3ed9a296 | 1285 | |
2f5cb738 JA |
1286 | /* |
1287 | * We have other queues, don't allow more IO from this one | |
1288 | */ | |
1289 | if (cfqd->busy_queues > 1) | |
0b182d61 | 1290 | return false; |
9ede209e | 1291 | |
365722bb | 1292 | /* |
8e296755 | 1293 | * Sole queue user, allow bigger slice |
365722bb | 1294 | */ |
8e296755 JA |
1295 | max_dispatch *= 4; |
1296 | } | |
1297 | ||
1298 | /* | |
1299 | * Async queues must wait a bit before being allowed dispatch. | |
1300 | * We also ramp up the dispatch depth gradually for async IO, | |
1301 | * based on the last sync IO we serviced | |
1302 | */ | |
963b72fc | 1303 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { |
8e296755 JA |
1304 | unsigned long last_sync = jiffies - cfqd->last_end_sync_rq; |
1305 | unsigned int depth; | |
365722bb | 1306 | |
61f0c1dc | 1307 | depth = last_sync / cfqd->cfq_slice[1]; |
e00c54c3 JA |
1308 | if (!depth && !cfqq->dispatched) |
1309 | depth = 1; | |
8e296755 JA |
1310 | if (depth < max_dispatch) |
1311 | max_dispatch = depth; | |
2f5cb738 | 1312 | } |
3ed9a296 | 1313 | |
0b182d61 JA |
1314 | /* |
1315 | * If we're below the current max, allow a dispatch | |
1316 | */ | |
1317 | return cfqq->dispatched < max_dispatch; | |
1318 | } | |
1319 | ||
1320 | /* | |
1321 | * Dispatch a request from cfqq, moving them to the request queue | |
1322 | * dispatch list. | |
1323 | */ | |
1324 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1325 | { | |
1326 | struct request *rq; | |
1327 | ||
1328 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
1329 | ||
1330 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
1331 | return false; | |
1332 | ||
1333 | /* | |
1334 | * follow expired path, else get first next available | |
1335 | */ | |
1336 | rq = cfq_check_fifo(cfqq); | |
1337 | if (!rq) | |
1338 | rq = cfqq->next_rq; | |
1339 | ||
1340 | /* | |
1341 | * insert request into driver dispatch list | |
1342 | */ | |
1343 | cfq_dispatch_insert(cfqd->queue, rq); | |
1344 | ||
1345 | if (!cfqd->active_cic) { | |
1346 | struct cfq_io_context *cic = RQ_CIC(rq); | |
1347 | ||
1348 | atomic_long_inc(&cic->ioc->refcount); | |
1349 | cfqd->active_cic = cic; | |
1350 | } | |
1351 | ||
1352 | return true; | |
1353 | } | |
1354 | ||
1355 | /* | |
1356 | * Find the cfqq that we need to service and move a request from that to the | |
1357 | * dispatch list | |
1358 | */ | |
1359 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
1360 | { | |
1361 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1362 | struct cfq_queue *cfqq; | |
1363 | ||
1364 | if (!cfqd->busy_queues) | |
1365 | return 0; | |
1366 | ||
1367 | if (unlikely(force)) | |
1368 | return cfq_forced_dispatch(cfqd); | |
1369 | ||
1370 | cfqq = cfq_select_queue(cfqd); | |
1371 | if (!cfqq) | |
8e296755 JA |
1372 | return 0; |
1373 | ||
2f5cb738 | 1374 | /* |
0b182d61 | 1375 | * Dispatch a request from this cfqq, if it is allowed |
2f5cb738 | 1376 | */ |
0b182d61 JA |
1377 | if (!cfq_dispatch_request(cfqd, cfqq)) |
1378 | return 0; | |
1379 | ||
2f5cb738 | 1380 | cfqq->slice_dispatch++; |
b029195d | 1381 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 1382 | |
2f5cb738 JA |
1383 | /* |
1384 | * expire an async queue immediately if it has used up its slice. idle | |
1385 | * queue always expire after 1 dispatch round. | |
1386 | */ | |
1387 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
1388 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
1389 | cfq_class_idle(cfqq))) { | |
1390 | cfqq->slice_end = jiffies + 1; | |
1391 | cfq_slice_expired(cfqd, 0); | |
1da177e4 LT |
1392 | } |
1393 | ||
b217a903 | 1394 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); |
2f5cb738 | 1395 | return 1; |
1da177e4 LT |
1396 | } |
1397 | ||
1da177e4 | 1398 | /* |
5e705374 JA |
1399 | * task holds one reference to the queue, dropped when task exits. each rq |
1400 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 LT |
1401 | * |
1402 | * queue lock must be held here. | |
1403 | */ | |
1404 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
1405 | { | |
22e2c507 JA |
1406 | struct cfq_data *cfqd = cfqq->cfqd; |
1407 | ||
1408 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
1409 | |
1410 | if (!atomic_dec_and_test(&cfqq->ref)) | |
1411 | return; | |
1412 | ||
7b679138 | 1413 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 1414 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 1415 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
3b18152c | 1416 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 | 1417 | |
28f95cbc | 1418 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 1419 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 1420 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 1421 | } |
22e2c507 | 1422 | |
1da177e4 LT |
1423 | kmem_cache_free(cfq_pool, cfqq); |
1424 | } | |
1425 | ||
d6de8be7 JA |
1426 | /* |
1427 | * Must always be called with the rcu_read_lock() held | |
1428 | */ | |
07416d29 JA |
1429 | static void |
1430 | __call_for_each_cic(struct io_context *ioc, | |
1431 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1432 | { | |
1433 | struct cfq_io_context *cic; | |
1434 | struct hlist_node *n; | |
1435 | ||
1436 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) | |
1437 | func(ioc, cic); | |
1438 | } | |
1439 | ||
4ac845a2 | 1440 | /* |
34e6bbf2 | 1441 | * Call func for each cic attached to this ioc. |
4ac845a2 | 1442 | */ |
34e6bbf2 | 1443 | static void |
4ac845a2 JA |
1444 | call_for_each_cic(struct io_context *ioc, |
1445 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1da177e4 | 1446 | { |
4ac845a2 | 1447 | rcu_read_lock(); |
07416d29 | 1448 | __call_for_each_cic(ioc, func); |
4ac845a2 | 1449 | rcu_read_unlock(); |
34e6bbf2 FC |
1450 | } |
1451 | ||
1452 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
1453 | { | |
1454 | struct cfq_io_context *cic; | |
1455 | ||
1456 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
1457 | ||
1458 | kmem_cache_free(cfq_ioc_pool, cic); | |
245b2e70 | 1459 | elv_ioc_count_dec(cfq_ioc_count); |
34e6bbf2 | 1460 | |
9a11b4ed JA |
1461 | if (ioc_gone) { |
1462 | /* | |
1463 | * CFQ scheduler is exiting, grab exit lock and check | |
1464 | * the pending io context count. If it hits zero, | |
1465 | * complete ioc_gone and set it back to NULL | |
1466 | */ | |
1467 | spin_lock(&ioc_gone_lock); | |
245b2e70 | 1468 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { |
9a11b4ed JA |
1469 | complete(ioc_gone); |
1470 | ioc_gone = NULL; | |
1471 | } | |
1472 | spin_unlock(&ioc_gone_lock); | |
1473 | } | |
34e6bbf2 | 1474 | } |
4ac845a2 | 1475 | |
34e6bbf2 FC |
1476 | static void cfq_cic_free(struct cfq_io_context *cic) |
1477 | { | |
1478 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
1479 | } |
1480 | ||
1481 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
1482 | { | |
1483 | unsigned long flags; | |
1484 | ||
1485 | BUG_ON(!cic->dead_key); | |
1486 | ||
1487 | spin_lock_irqsave(&ioc->lock, flags); | |
1488 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
ffc4e759 | 1489 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
1490 | spin_unlock_irqrestore(&ioc->lock, flags); |
1491 | ||
34e6bbf2 | 1492 | cfq_cic_free(cic); |
4ac845a2 JA |
1493 | } |
1494 | ||
d6de8be7 JA |
1495 | /* |
1496 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
1497 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
1498 | * and ->trim() which is called with the task lock held | |
1499 | */ | |
4ac845a2 JA |
1500 | static void cfq_free_io_context(struct io_context *ioc) |
1501 | { | |
4ac845a2 | 1502 | /* |
34e6bbf2 FC |
1503 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
1504 | * so no more cic's are allowed to be linked into this ioc. So it | |
1505 | * should be ok to iterate over the known list, we will see all cic's | |
1506 | * since no new ones are added. | |
4ac845a2 | 1507 | */ |
07416d29 | 1508 | __call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
1509 | } |
1510 | ||
89850f7e | 1511 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1512 | { |
28f95cbc | 1513 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 1514 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 1515 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 1516 | } |
22e2c507 | 1517 | |
89850f7e JA |
1518 | cfq_put_queue(cfqq); |
1519 | } | |
22e2c507 | 1520 | |
89850f7e JA |
1521 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
1522 | struct cfq_io_context *cic) | |
1523 | { | |
4faa3c81 FC |
1524 | struct io_context *ioc = cic->ioc; |
1525 | ||
fc46379d | 1526 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
1527 | |
1528 | /* | |
1529 | * Make sure key == NULL is seen for dead queues | |
1530 | */ | |
fc46379d | 1531 | smp_wmb(); |
4ac845a2 | 1532 | cic->dead_key = (unsigned long) cic->key; |
fc46379d JA |
1533 | cic->key = NULL; |
1534 | ||
4faa3c81 FC |
1535 | if (ioc->ioc_data == cic) |
1536 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
1537 | ||
ff6657c6 JA |
1538 | if (cic->cfqq[BLK_RW_ASYNC]) { |
1539 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
1540 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
1541 | } |
1542 | ||
ff6657c6 JA |
1543 | if (cic->cfqq[BLK_RW_SYNC]) { |
1544 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
1545 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 1546 | } |
89850f7e JA |
1547 | } |
1548 | ||
4ac845a2 JA |
1549 | static void cfq_exit_single_io_context(struct io_context *ioc, |
1550 | struct cfq_io_context *cic) | |
89850f7e JA |
1551 | { |
1552 | struct cfq_data *cfqd = cic->key; | |
1553 | ||
89850f7e | 1554 | if (cfqd) { |
165125e1 | 1555 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 1556 | unsigned long flags; |
89850f7e | 1557 | |
4ac845a2 | 1558 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
1559 | |
1560 | /* | |
1561 | * Ensure we get a fresh copy of the ->key to prevent | |
1562 | * race between exiting task and queue | |
1563 | */ | |
1564 | smp_read_barrier_depends(); | |
1565 | if (cic->key) | |
1566 | __cfq_exit_single_io_context(cfqd, cic); | |
1567 | ||
4ac845a2 | 1568 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 1569 | } |
1da177e4 LT |
1570 | } |
1571 | ||
498d3aa2 JA |
1572 | /* |
1573 | * The process that ioc belongs to has exited, we need to clean up | |
1574 | * and put the internal structures we have that belongs to that process. | |
1575 | */ | |
e2d74ac0 | 1576 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 1577 | { |
4ac845a2 | 1578 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
1579 | } |
1580 | ||
22e2c507 | 1581 | static struct cfq_io_context * |
8267e268 | 1582 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1583 | { |
b5deef90 | 1584 | struct cfq_io_context *cic; |
1da177e4 | 1585 | |
94f6030c CL |
1586 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
1587 | cfqd->queue->node); | |
1da177e4 | 1588 | if (cic) { |
22e2c507 | 1589 | cic->last_end_request = jiffies; |
553698f9 | 1590 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 1591 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
1592 | cic->dtor = cfq_free_io_context; |
1593 | cic->exit = cfq_exit_io_context; | |
245b2e70 | 1594 | elv_ioc_count_inc(cfq_ioc_count); |
1da177e4 LT |
1595 | } |
1596 | ||
1597 | return cic; | |
1598 | } | |
1599 | ||
fd0928df | 1600 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
1601 | { |
1602 | struct task_struct *tsk = current; | |
1603 | int ioprio_class; | |
1604 | ||
3b18152c | 1605 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
1606 | return; |
1607 | ||
fd0928df | 1608 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 1609 | switch (ioprio_class) { |
fe094d98 JA |
1610 | default: |
1611 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
1612 | case IOPRIO_CLASS_NONE: | |
1613 | /* | |
6d63c275 | 1614 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
1615 | */ |
1616 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 1617 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
1618 | break; |
1619 | case IOPRIO_CLASS_RT: | |
1620 | cfqq->ioprio = task_ioprio(ioc); | |
1621 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
1622 | break; | |
1623 | case IOPRIO_CLASS_BE: | |
1624 | cfqq->ioprio = task_ioprio(ioc); | |
1625 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
1626 | break; | |
1627 | case IOPRIO_CLASS_IDLE: | |
1628 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
1629 | cfqq->ioprio = 7; | |
1630 | cfq_clear_cfqq_idle_window(cfqq); | |
1631 | break; | |
22e2c507 JA |
1632 | } |
1633 | ||
1634 | /* | |
1635 | * keep track of original prio settings in case we have to temporarily | |
1636 | * elevate the priority of this queue | |
1637 | */ | |
1638 | cfqq->org_ioprio = cfqq->ioprio; | |
1639 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 1640 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
1641 | } |
1642 | ||
febffd61 | 1643 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 1644 | { |
478a82b0 AV |
1645 | struct cfq_data *cfqd = cic->key; |
1646 | struct cfq_queue *cfqq; | |
c1b707d2 | 1647 | unsigned long flags; |
35e6077c | 1648 | |
caaa5f9f JA |
1649 | if (unlikely(!cfqd)) |
1650 | return; | |
1651 | ||
c1b707d2 | 1652 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 1653 | |
ff6657c6 | 1654 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
1655 | if (cfqq) { |
1656 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
1657 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
1658 | GFP_ATOMIC); | |
caaa5f9f | 1659 | if (new_cfqq) { |
ff6657c6 | 1660 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
1661 | cfq_put_queue(cfqq); |
1662 | } | |
22e2c507 | 1663 | } |
caaa5f9f | 1664 | |
ff6657c6 | 1665 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
1666 | if (cfqq) |
1667 | cfq_mark_cfqq_prio_changed(cfqq); | |
1668 | ||
c1b707d2 | 1669 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
1670 | } |
1671 | ||
fc46379d | 1672 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 1673 | { |
4ac845a2 | 1674 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 1675 | ioc->ioprio_changed = 0; |
22e2c507 JA |
1676 | } |
1677 | ||
d5036d77 JA |
1678 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
1679 | pid_t pid, int is_sync) | |
1680 | { | |
1681 | RB_CLEAR_NODE(&cfqq->rb_node); | |
1682 | RB_CLEAR_NODE(&cfqq->p_node); | |
1683 | INIT_LIST_HEAD(&cfqq->fifo); | |
1684 | ||
1685 | atomic_set(&cfqq->ref, 0); | |
1686 | cfqq->cfqd = cfqd; | |
1687 | ||
1688 | cfq_mark_cfqq_prio_changed(cfqq); | |
1689 | ||
1690 | if (is_sync) { | |
1691 | if (!cfq_class_idle(cfqq)) | |
1692 | cfq_mark_cfqq_idle_window(cfqq); | |
1693 | cfq_mark_cfqq_sync(cfqq); | |
1694 | } | |
1695 | cfqq->pid = pid; | |
1696 | } | |
1697 | ||
22e2c507 | 1698 | static struct cfq_queue * |
15c31be4 | 1699 | cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync, |
fd0928df | 1700 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 1701 | { |
22e2c507 | 1702 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 1703 | struct cfq_io_context *cic; |
22e2c507 JA |
1704 | |
1705 | retry: | |
4ac845a2 | 1706 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
1707 | /* cic always exists here */ |
1708 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 1709 | |
6118b70b JA |
1710 | /* |
1711 | * Always try a new alloc if we fell back to the OOM cfqq | |
1712 | * originally, since it should just be a temporary situation. | |
1713 | */ | |
1714 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
1715 | cfqq = NULL; | |
22e2c507 JA |
1716 | if (new_cfqq) { |
1717 | cfqq = new_cfqq; | |
1718 | new_cfqq = NULL; | |
1719 | } else if (gfp_mask & __GFP_WAIT) { | |
1720 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 1721 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 1722 | gfp_mask | __GFP_ZERO, |
94f6030c | 1723 | cfqd->queue->node); |
22e2c507 | 1724 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
1725 | if (new_cfqq) |
1726 | goto retry; | |
22e2c507 | 1727 | } else { |
94f6030c CL |
1728 | cfqq = kmem_cache_alloc_node(cfq_pool, |
1729 | gfp_mask | __GFP_ZERO, | |
1730 | cfqd->queue->node); | |
22e2c507 JA |
1731 | } |
1732 | ||
6118b70b JA |
1733 | if (cfqq) { |
1734 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
1735 | cfq_init_prio_data(cfqq, ioc); | |
1736 | cfq_log_cfqq(cfqd, cfqq, "alloced"); | |
1737 | } else | |
1738 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
1739 | } |
1740 | ||
1741 | if (new_cfqq) | |
1742 | kmem_cache_free(cfq_pool, new_cfqq); | |
1743 | ||
22e2c507 JA |
1744 | return cfqq; |
1745 | } | |
1746 | ||
c2dea2d1 VT |
1747 | static struct cfq_queue ** |
1748 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
1749 | { | |
fe094d98 | 1750 | switch (ioprio_class) { |
c2dea2d1 VT |
1751 | case IOPRIO_CLASS_RT: |
1752 | return &cfqd->async_cfqq[0][ioprio]; | |
1753 | case IOPRIO_CLASS_BE: | |
1754 | return &cfqd->async_cfqq[1][ioprio]; | |
1755 | case IOPRIO_CLASS_IDLE: | |
1756 | return &cfqd->async_idle_cfqq; | |
1757 | default: | |
1758 | BUG(); | |
1759 | } | |
1760 | } | |
1761 | ||
15c31be4 | 1762 | static struct cfq_queue * |
fd0928df | 1763 | cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc, |
15c31be4 JA |
1764 | gfp_t gfp_mask) |
1765 | { | |
fd0928df JA |
1766 | const int ioprio = task_ioprio(ioc); |
1767 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 1768 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
1769 | struct cfq_queue *cfqq = NULL; |
1770 | ||
c2dea2d1 VT |
1771 | if (!is_sync) { |
1772 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
1773 | cfqq = *async_cfqq; | |
1774 | } | |
1775 | ||
6118b70b | 1776 | if (!cfqq) |
fd0928df | 1777 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
1778 | |
1779 | /* | |
1780 | * pin the queue now that it's allocated, scheduler exit will prune it | |
1781 | */ | |
c2dea2d1 | 1782 | if (!is_sync && !(*async_cfqq)) { |
15c31be4 | 1783 | atomic_inc(&cfqq->ref); |
c2dea2d1 | 1784 | *async_cfqq = cfqq; |
15c31be4 JA |
1785 | } |
1786 | ||
1787 | atomic_inc(&cfqq->ref); | |
1788 | return cfqq; | |
1789 | } | |
1790 | ||
498d3aa2 JA |
1791 | /* |
1792 | * We drop cfq io contexts lazily, so we may find a dead one. | |
1793 | */ | |
dbecf3ab | 1794 | static void |
4ac845a2 JA |
1795 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
1796 | struct cfq_io_context *cic) | |
dbecf3ab | 1797 | { |
4ac845a2 JA |
1798 | unsigned long flags; |
1799 | ||
fc46379d | 1800 | WARN_ON(!list_empty(&cic->queue_list)); |
597bc485 | 1801 | |
4ac845a2 JA |
1802 | spin_lock_irqsave(&ioc->lock, flags); |
1803 | ||
4faa3c81 | 1804 | BUG_ON(ioc->ioc_data == cic); |
597bc485 | 1805 | |
4ac845a2 | 1806 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); |
ffc4e759 | 1807 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
1808 | spin_unlock_irqrestore(&ioc->lock, flags); |
1809 | ||
1810 | cfq_cic_free(cic); | |
dbecf3ab OH |
1811 | } |
1812 | ||
e2d74ac0 | 1813 | static struct cfq_io_context * |
4ac845a2 | 1814 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 1815 | { |
e2d74ac0 | 1816 | struct cfq_io_context *cic; |
d6de8be7 | 1817 | unsigned long flags; |
4ac845a2 | 1818 | void *k; |
e2d74ac0 | 1819 | |
91fac317 VT |
1820 | if (unlikely(!ioc)) |
1821 | return NULL; | |
1822 | ||
d6de8be7 JA |
1823 | rcu_read_lock(); |
1824 | ||
597bc485 JA |
1825 | /* |
1826 | * we maintain a last-hit cache, to avoid browsing over the tree | |
1827 | */ | |
4ac845a2 | 1828 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
1829 | if (cic && cic->key == cfqd) { |
1830 | rcu_read_unlock(); | |
597bc485 | 1831 | return cic; |
d6de8be7 | 1832 | } |
597bc485 | 1833 | |
4ac845a2 | 1834 | do { |
4ac845a2 JA |
1835 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); |
1836 | rcu_read_unlock(); | |
1837 | if (!cic) | |
1838 | break; | |
be3b0753 OH |
1839 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
1840 | k = cic->key; | |
1841 | if (unlikely(!k)) { | |
4ac845a2 | 1842 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 1843 | rcu_read_lock(); |
4ac845a2 | 1844 | continue; |
dbecf3ab | 1845 | } |
e2d74ac0 | 1846 | |
d6de8be7 | 1847 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 1848 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 1849 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
1850 | break; |
1851 | } while (1); | |
e2d74ac0 | 1852 | |
4ac845a2 | 1853 | return cic; |
e2d74ac0 JA |
1854 | } |
1855 | ||
4ac845a2 JA |
1856 | /* |
1857 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
1858 | * the process specific cfq io context when entered from the block layer. | |
1859 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
1860 | */ | |
febffd61 JA |
1861 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
1862 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 1863 | { |
0261d688 | 1864 | unsigned long flags; |
4ac845a2 | 1865 | int ret; |
e2d74ac0 | 1866 | |
4ac845a2 JA |
1867 | ret = radix_tree_preload(gfp_mask); |
1868 | if (!ret) { | |
1869 | cic->ioc = ioc; | |
1870 | cic->key = cfqd; | |
e2d74ac0 | 1871 | |
4ac845a2 JA |
1872 | spin_lock_irqsave(&ioc->lock, flags); |
1873 | ret = radix_tree_insert(&ioc->radix_root, | |
1874 | (unsigned long) cfqd, cic); | |
ffc4e759 JA |
1875 | if (!ret) |
1876 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 1877 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 1878 | |
4ac845a2 JA |
1879 | radix_tree_preload_end(); |
1880 | ||
1881 | if (!ret) { | |
1882 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
1883 | list_add(&cic->queue_list, &cfqd->cic_list); | |
1884 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
1885 | } | |
e2d74ac0 JA |
1886 | } |
1887 | ||
4ac845a2 JA |
1888 | if (ret) |
1889 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 1890 | |
4ac845a2 | 1891 | return ret; |
e2d74ac0 JA |
1892 | } |
1893 | ||
1da177e4 LT |
1894 | /* |
1895 | * Setup general io context and cfq io context. There can be several cfq | |
1896 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 1897 | * than one device managed by cfq. |
1da177e4 LT |
1898 | */ |
1899 | static struct cfq_io_context * | |
e2d74ac0 | 1900 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1901 | { |
22e2c507 | 1902 | struct io_context *ioc = NULL; |
1da177e4 | 1903 | struct cfq_io_context *cic; |
1da177e4 | 1904 | |
22e2c507 | 1905 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 1906 | |
b5deef90 | 1907 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
1908 | if (!ioc) |
1909 | return NULL; | |
1910 | ||
4ac845a2 | 1911 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
1912 | if (cic) |
1913 | goto out; | |
1da177e4 | 1914 | |
e2d74ac0 JA |
1915 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
1916 | if (cic == NULL) | |
1917 | goto err; | |
1da177e4 | 1918 | |
4ac845a2 JA |
1919 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
1920 | goto err_free; | |
1921 | ||
1da177e4 | 1922 | out: |
fc46379d JA |
1923 | smp_read_barrier_depends(); |
1924 | if (unlikely(ioc->ioprio_changed)) | |
1925 | cfq_ioc_set_ioprio(ioc); | |
1926 | ||
1da177e4 | 1927 | return cic; |
4ac845a2 JA |
1928 | err_free: |
1929 | cfq_cic_free(cic); | |
1da177e4 LT |
1930 | err: |
1931 | put_io_context(ioc); | |
1932 | return NULL; | |
1933 | } | |
1934 | ||
22e2c507 JA |
1935 | static void |
1936 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 1937 | { |
aaf1228d JA |
1938 | unsigned long elapsed = jiffies - cic->last_end_request; |
1939 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 1940 | |
22e2c507 JA |
1941 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
1942 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
1943 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
1944 | } | |
1da177e4 | 1945 | |
206dc69b | 1946 | static void |
6d048f53 JA |
1947 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic, |
1948 | struct request *rq) | |
206dc69b JA |
1949 | { |
1950 | sector_t sdist; | |
1951 | u64 total; | |
1952 | ||
4d00aa47 JM |
1953 | if (!cic->last_request_pos) |
1954 | sdist = 0; | |
83096ebf TH |
1955 | else if (cic->last_request_pos < blk_rq_pos(rq)) |
1956 | sdist = blk_rq_pos(rq) - cic->last_request_pos; | |
206dc69b | 1957 | else |
83096ebf | 1958 | sdist = cic->last_request_pos - blk_rq_pos(rq); |
206dc69b JA |
1959 | |
1960 | /* | |
1961 | * Don't allow the seek distance to get too large from the | |
1962 | * odd fragment, pagein, etc | |
1963 | */ | |
1964 | if (cic->seek_samples <= 60) /* second&third seek */ | |
1965 | sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024); | |
1966 | else | |
1967 | sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64); | |
1968 | ||
1969 | cic->seek_samples = (7*cic->seek_samples + 256) / 8; | |
1970 | cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8; | |
1971 | total = cic->seek_total + (cic->seek_samples/2); | |
1972 | do_div(total, cic->seek_samples); | |
1973 | cic->seek_mean = (sector_t)total; | |
1974 | } | |
1da177e4 | 1975 | |
22e2c507 JA |
1976 | /* |
1977 | * Disable idle window if the process thinks too long or seeks so much that | |
1978 | * it doesn't matter | |
1979 | */ | |
1980 | static void | |
1981 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1982 | struct cfq_io_context *cic) | |
1983 | { | |
7b679138 | 1984 | int old_idle, enable_idle; |
1be92f2f | 1985 | |
0871714e JA |
1986 | /* |
1987 | * Don't idle for async or idle io prio class | |
1988 | */ | |
1989 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
1990 | return; |
1991 | ||
c265a7f4 | 1992 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 1993 | |
66dac98e | 1994 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || |
963b72fc | 1995 | (!cfqd->cfq_latency && cfqd->hw_tag && CIC_SEEKY(cic))) |
22e2c507 JA |
1996 | enable_idle = 0; |
1997 | else if (sample_valid(cic->ttime_samples)) { | |
ec60e4f6 CZ |
1998 | unsigned int slice_idle = cfqd->cfq_slice_idle; |
1999 | if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic)) | |
2000 | slice_idle = msecs_to_jiffies(CFQ_MIN_TT); | |
2001 | if (cic->ttime_mean > slice_idle) | |
22e2c507 JA |
2002 | enable_idle = 0; |
2003 | else | |
2004 | enable_idle = 1; | |
1da177e4 LT |
2005 | } |
2006 | ||
7b679138 JA |
2007 | if (old_idle != enable_idle) { |
2008 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
2009 | if (enable_idle) | |
2010 | cfq_mark_cfqq_idle_window(cfqq); | |
2011 | else | |
2012 | cfq_clear_cfqq_idle_window(cfqq); | |
2013 | } | |
22e2c507 | 2014 | } |
1da177e4 | 2015 | |
22e2c507 JA |
2016 | /* |
2017 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
2018 | * no or if we aren't sure, a 1 will cause a preempt. | |
2019 | */ | |
2020 | static int | |
2021 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, | |
5e705374 | 2022 | struct request *rq) |
22e2c507 | 2023 | { |
6d048f53 | 2024 | struct cfq_queue *cfqq; |
22e2c507 | 2025 | |
6d048f53 JA |
2026 | cfqq = cfqd->active_queue; |
2027 | if (!cfqq) | |
22e2c507 JA |
2028 | return 0; |
2029 | ||
6d048f53 JA |
2030 | if (cfq_slice_used(cfqq)) |
2031 | return 1; | |
2032 | ||
2033 | if (cfq_class_idle(new_cfqq)) | |
caaa5f9f | 2034 | return 0; |
22e2c507 JA |
2035 | |
2036 | if (cfq_class_idle(cfqq)) | |
2037 | return 1; | |
1e3335de | 2038 | |
374f84ac JA |
2039 | /* |
2040 | * if the new request is sync, but the currently running queue is | |
2041 | * not, let the sync request have priority. | |
2042 | */ | |
5e705374 | 2043 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
22e2c507 | 2044 | return 1; |
1e3335de | 2045 | |
374f84ac JA |
2046 | /* |
2047 | * So both queues are sync. Let the new request get disk time if | |
2048 | * it's a metadata request and the current queue is doing regular IO. | |
2049 | */ | |
2050 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
2051 | return 1; | |
22e2c507 | 2052 | |
3a9a3f6c DS |
2053 | /* |
2054 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
2055 | */ | |
2056 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
2057 | return 1; | |
2058 | ||
1e3335de JA |
2059 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
2060 | return 0; | |
2061 | ||
2062 | /* | |
2063 | * if this request is as-good as one we would expect from the | |
2064 | * current cfqq, let it preempt | |
2065 | */ | |
6d048f53 | 2066 | if (cfq_rq_close(cfqd, rq)) |
1e3335de JA |
2067 | return 1; |
2068 | ||
22e2c507 JA |
2069 | return 0; |
2070 | } | |
2071 | ||
2072 | /* | |
2073 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
2074 | * let it have half of its nominal slice. | |
2075 | */ | |
2076 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2077 | { | |
7b679138 | 2078 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
6084cdda | 2079 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 2080 | |
bf572256 JA |
2081 | /* |
2082 | * Put the new queue at the front of the of the current list, | |
2083 | * so we know that it will be selected next. | |
2084 | */ | |
2085 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
2086 | |
2087 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 2088 | |
44f7c160 JA |
2089 | cfqq->slice_end = 0; |
2090 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
2091 | } |
2092 | ||
22e2c507 | 2093 | /* |
5e705374 | 2094 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
2095 | * something we should do about it |
2096 | */ | |
2097 | static void | |
5e705374 JA |
2098 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
2099 | struct request *rq) | |
22e2c507 | 2100 | { |
5e705374 | 2101 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 2102 | |
45333d5a | 2103 | cfqd->rq_queued++; |
374f84ac JA |
2104 | if (rq_is_meta(rq)) |
2105 | cfqq->meta_pending++; | |
2106 | ||
9c2c38a1 | 2107 | cfq_update_io_thinktime(cfqd, cic); |
6d048f53 | 2108 | cfq_update_io_seektime(cfqd, cic, rq); |
9c2c38a1 JA |
2109 | cfq_update_idle_window(cfqd, cfqq, cic); |
2110 | ||
83096ebf | 2111 | cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
2112 | |
2113 | if (cfqq == cfqd->active_queue) { | |
2114 | /* | |
b029195d JA |
2115 | * Remember that we saw a request from this process, but |
2116 | * don't start queuing just yet. Otherwise we risk seeing lots | |
2117 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
2118 | * and merging. If the request is already larger than a single |
2119 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
2120 | * merging is already done. Ditto for a busy system that |
2121 | * has other work pending, don't risk delaying until the | |
2122 | * idle timer unplug to continue working. | |
22e2c507 | 2123 | */ |
d6ceb25e | 2124 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
2125 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
2126 | cfqd->busy_queues > 1) { | |
d6ceb25e | 2127 | del_timer(&cfqd->idle_slice_timer); |
a7f55792 | 2128 | __blk_run_queue(cfqd->queue); |
d6ceb25e | 2129 | } |
b029195d | 2130 | cfq_mark_cfqq_must_dispatch(cfqq); |
d6ceb25e | 2131 | } |
5e705374 | 2132 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
2133 | /* |
2134 | * not the active queue - expire current slice if it is | |
2135 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
2136 | * has some old slice time left and is of higher priority or |
2137 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
2138 | */ |
2139 | cfq_preempt_queue(cfqd, cfqq); | |
a7f55792 | 2140 | __blk_run_queue(cfqd->queue); |
22e2c507 | 2141 | } |
1da177e4 LT |
2142 | } |
2143 | ||
165125e1 | 2144 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2145 | { |
b4878f24 | 2146 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 2147 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 2148 | |
7b679138 | 2149 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 2150 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 2151 | |
5e705374 | 2152 | cfq_add_rq_rb(rq); |
1da177e4 | 2153 | |
30996f40 | 2154 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); |
22e2c507 JA |
2155 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
2156 | ||
5e705374 | 2157 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
2158 | } |
2159 | ||
45333d5a AC |
2160 | /* |
2161 | * Update hw_tag based on peak queue depth over 50 samples under | |
2162 | * sufficient load. | |
2163 | */ | |
2164 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
2165 | { | |
5ad531db JA |
2166 | if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak) |
2167 | cfqd->rq_in_driver_peak = rq_in_driver(cfqd); | |
45333d5a AC |
2168 | |
2169 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
5ad531db | 2170 | rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
2171 | return; |
2172 | ||
2173 | if (cfqd->hw_tag_samples++ < 50) | |
2174 | return; | |
2175 | ||
2176 | if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN) | |
2177 | cfqd->hw_tag = 1; | |
2178 | else | |
2179 | cfqd->hw_tag = 0; | |
2180 | ||
2181 | cfqd->hw_tag_samples = 0; | |
2182 | cfqd->rq_in_driver_peak = 0; | |
2183 | } | |
2184 | ||
165125e1 | 2185 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2186 | { |
5e705374 | 2187 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 2188 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 2189 | const int sync = rq_is_sync(rq); |
b4878f24 | 2190 | unsigned long now; |
1da177e4 | 2191 | |
b4878f24 | 2192 | now = jiffies; |
7b679138 | 2193 | cfq_log_cfqq(cfqd, cfqq, "complete"); |
1da177e4 | 2194 | |
45333d5a AC |
2195 | cfq_update_hw_tag(cfqd); |
2196 | ||
5ad531db | 2197 | WARN_ON(!cfqd->rq_in_driver[sync]); |
6d048f53 | 2198 | WARN_ON(!cfqq->dispatched); |
5ad531db | 2199 | cfqd->rq_in_driver[sync]--; |
6d048f53 | 2200 | cfqq->dispatched--; |
1da177e4 | 2201 | |
3ed9a296 JA |
2202 | if (cfq_cfqq_sync(cfqq)) |
2203 | cfqd->sync_flight--; | |
2204 | ||
365722bb | 2205 | if (sync) { |
5e705374 | 2206 | RQ_CIC(rq)->last_end_request = now; |
365722bb VG |
2207 | cfqd->last_end_sync_rq = now; |
2208 | } | |
caaa5f9f JA |
2209 | |
2210 | /* | |
2211 | * If this is the active queue, check if it needs to be expired, | |
2212 | * or if we want to idle in case it has no pending requests. | |
2213 | */ | |
2214 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
2215 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
2216 | ||
44f7c160 JA |
2217 | if (cfq_cfqq_slice_new(cfqq)) { |
2218 | cfq_set_prio_slice(cfqd, cfqq); | |
2219 | cfq_clear_cfqq_slice_new(cfqq); | |
2220 | } | |
a36e71f9 JA |
2221 | /* |
2222 | * If there are no requests waiting in this queue, and | |
2223 | * there are other queues ready to issue requests, AND | |
2224 | * those other queues are issuing requests within our | |
2225 | * mean seek distance, give them a chance to run instead | |
2226 | * of idling. | |
2227 | */ | |
0871714e | 2228 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
6084cdda | 2229 | cfq_slice_expired(cfqd, 1); |
a36e71f9 JA |
2230 | else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) && |
2231 | sync && !rq_noidle(rq)) | |
6d048f53 | 2232 | cfq_arm_slice_timer(cfqd); |
caaa5f9f | 2233 | } |
6d048f53 | 2234 | |
5ad531db | 2235 | if (!rq_in_driver(cfqd)) |
23e018a1 | 2236 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
2237 | } |
2238 | ||
22e2c507 JA |
2239 | /* |
2240 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
2241 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
2242 | */ | |
2243 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 2244 | { |
22e2c507 JA |
2245 | if (has_fs_excl()) { |
2246 | /* | |
2247 | * boost idle prio on transactions that would lock out other | |
2248 | * users of the filesystem | |
2249 | */ | |
2250 | if (cfq_class_idle(cfqq)) | |
2251 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2252 | if (cfqq->ioprio > IOPRIO_NORM) | |
2253 | cfqq->ioprio = IOPRIO_NORM; | |
2254 | } else { | |
2255 | /* | |
2256 | * check if we need to unboost the queue | |
2257 | */ | |
2258 | if (cfqq->ioprio_class != cfqq->org_ioprio_class) | |
2259 | cfqq->ioprio_class = cfqq->org_ioprio_class; | |
2260 | if (cfqq->ioprio != cfqq->org_ioprio) | |
2261 | cfqq->ioprio = cfqq->org_ioprio; | |
2262 | } | |
22e2c507 | 2263 | } |
1da177e4 | 2264 | |
89850f7e | 2265 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 2266 | { |
1b379d8d | 2267 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 2268 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 2269 | return ELV_MQUEUE_MUST; |
3b18152c | 2270 | } |
1da177e4 | 2271 | |
22e2c507 | 2272 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
2273 | } |
2274 | ||
165125e1 | 2275 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
2276 | { |
2277 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2278 | struct task_struct *tsk = current; | |
91fac317 | 2279 | struct cfq_io_context *cic; |
22e2c507 JA |
2280 | struct cfq_queue *cfqq; |
2281 | ||
2282 | /* | |
2283 | * don't force setup of a queue from here, as a call to may_queue | |
2284 | * does not necessarily imply that a request actually will be queued. | |
2285 | * so just lookup a possibly existing queue, or return 'may queue' | |
2286 | * if that fails | |
2287 | */ | |
4ac845a2 | 2288 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
2289 | if (!cic) |
2290 | return ELV_MQUEUE_MAY; | |
2291 | ||
b0b78f81 | 2292 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 2293 | if (cfqq) { |
fd0928df | 2294 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
2295 | cfq_prio_boost(cfqq); |
2296 | ||
89850f7e | 2297 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
2298 | } |
2299 | ||
2300 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
2301 | } |
2302 | ||
1da177e4 LT |
2303 | /* |
2304 | * queue lock held here | |
2305 | */ | |
bb37b94c | 2306 | static void cfq_put_request(struct request *rq) |
1da177e4 | 2307 | { |
5e705374 | 2308 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 2309 | |
5e705374 | 2310 | if (cfqq) { |
22e2c507 | 2311 | const int rw = rq_data_dir(rq); |
1da177e4 | 2312 | |
22e2c507 JA |
2313 | BUG_ON(!cfqq->allocated[rw]); |
2314 | cfqq->allocated[rw]--; | |
1da177e4 | 2315 | |
5e705374 | 2316 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 2317 | |
1da177e4 | 2318 | rq->elevator_private = NULL; |
5e705374 | 2319 | rq->elevator_private2 = NULL; |
1da177e4 | 2320 | |
1da177e4 LT |
2321 | cfq_put_queue(cfqq); |
2322 | } | |
2323 | } | |
2324 | ||
2325 | /* | |
22e2c507 | 2326 | * Allocate cfq data structures associated with this request. |
1da177e4 | 2327 | */ |
22e2c507 | 2328 | static int |
165125e1 | 2329 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
2330 | { |
2331 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2332 | struct cfq_io_context *cic; | |
2333 | const int rw = rq_data_dir(rq); | |
7749a8d4 | 2334 | const int is_sync = rq_is_sync(rq); |
22e2c507 | 2335 | struct cfq_queue *cfqq; |
1da177e4 LT |
2336 | unsigned long flags; |
2337 | ||
2338 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
2339 | ||
e2d74ac0 | 2340 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 2341 | |
1da177e4 LT |
2342 | spin_lock_irqsave(q->queue_lock, flags); |
2343 | ||
22e2c507 JA |
2344 | if (!cic) |
2345 | goto queue_fail; | |
2346 | ||
91fac317 | 2347 | cfqq = cic_to_cfqq(cic, is_sync); |
32f2e807 | 2348 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { |
fd0928df | 2349 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 VT |
2350 | cic_set_cfqq(cic, cfqq, is_sync); |
2351 | } | |
1da177e4 LT |
2352 | |
2353 | cfqq->allocated[rw]++; | |
22e2c507 | 2354 | atomic_inc(&cfqq->ref); |
1da177e4 | 2355 | |
5e705374 | 2356 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 2357 | |
5e705374 JA |
2358 | rq->elevator_private = cic; |
2359 | rq->elevator_private2 = cfqq; | |
2360 | return 0; | |
1da177e4 | 2361 | |
22e2c507 JA |
2362 | queue_fail: |
2363 | if (cic) | |
2364 | put_io_context(cic->ioc); | |
89850f7e | 2365 | |
23e018a1 | 2366 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 2367 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 2368 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
2369 | return 1; |
2370 | } | |
2371 | ||
65f27f38 | 2372 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 2373 | { |
65f27f38 | 2374 | struct cfq_data *cfqd = |
23e018a1 | 2375 | container_of(work, struct cfq_data, unplug_work); |
165125e1 | 2376 | struct request_queue *q = cfqd->queue; |
22e2c507 | 2377 | |
40bb54d1 | 2378 | spin_lock_irq(q->queue_lock); |
a7f55792 | 2379 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 2380 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
2381 | } |
2382 | ||
2383 | /* | |
2384 | * Timer running if the active_queue is currently idling inside its time slice | |
2385 | */ | |
2386 | static void cfq_idle_slice_timer(unsigned long data) | |
2387 | { | |
2388 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
2389 | struct cfq_queue *cfqq; | |
2390 | unsigned long flags; | |
3c6bd2f8 | 2391 | int timed_out = 1; |
22e2c507 | 2392 | |
7b679138 JA |
2393 | cfq_log(cfqd, "idle timer fired"); |
2394 | ||
22e2c507 JA |
2395 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
2396 | ||
fe094d98 JA |
2397 | cfqq = cfqd->active_queue; |
2398 | if (cfqq) { | |
3c6bd2f8 JA |
2399 | timed_out = 0; |
2400 | ||
b029195d JA |
2401 | /* |
2402 | * We saw a request before the queue expired, let it through | |
2403 | */ | |
2404 | if (cfq_cfqq_must_dispatch(cfqq)) | |
2405 | goto out_kick; | |
2406 | ||
22e2c507 JA |
2407 | /* |
2408 | * expired | |
2409 | */ | |
44f7c160 | 2410 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
2411 | goto expire; |
2412 | ||
2413 | /* | |
2414 | * only expire and reinvoke request handler, if there are | |
2415 | * other queues with pending requests | |
2416 | */ | |
caaa5f9f | 2417 | if (!cfqd->busy_queues) |
22e2c507 | 2418 | goto out_cont; |
22e2c507 JA |
2419 | |
2420 | /* | |
2421 | * not expired and it has a request pending, let it dispatch | |
2422 | */ | |
75e50984 | 2423 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 2424 | goto out_kick; |
22e2c507 JA |
2425 | } |
2426 | expire: | |
6084cdda | 2427 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 2428 | out_kick: |
23e018a1 | 2429 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
2430 | out_cont: |
2431 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2432 | } | |
2433 | ||
3b18152c JA |
2434 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
2435 | { | |
2436 | del_timer_sync(&cfqd->idle_slice_timer); | |
23e018a1 | 2437 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 2438 | } |
22e2c507 | 2439 | |
c2dea2d1 VT |
2440 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
2441 | { | |
2442 | int i; | |
2443 | ||
2444 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
2445 | if (cfqd->async_cfqq[0][i]) | |
2446 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
2447 | if (cfqd->async_cfqq[1][i]) | |
2448 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 2449 | } |
2389d1ef ON |
2450 | |
2451 | if (cfqd->async_idle_cfqq) | |
2452 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
2453 | } |
2454 | ||
b374d18a | 2455 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 2456 | { |
22e2c507 | 2457 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 2458 | struct request_queue *q = cfqd->queue; |
22e2c507 | 2459 | |
3b18152c | 2460 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 2461 | |
d9ff4187 | 2462 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 2463 | |
d9ff4187 | 2464 | if (cfqd->active_queue) |
6084cdda | 2465 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
2466 | |
2467 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
2468 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
2469 | struct cfq_io_context, | |
2470 | queue_list); | |
89850f7e JA |
2471 | |
2472 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 2473 | } |
e2d74ac0 | 2474 | |
c2dea2d1 | 2475 | cfq_put_async_queues(cfqd); |
15c31be4 | 2476 | |
d9ff4187 | 2477 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
2478 | |
2479 | cfq_shutdown_timer_wq(cfqd); | |
2480 | ||
a90d742e | 2481 | kfree(cfqd); |
1da177e4 LT |
2482 | } |
2483 | ||
165125e1 | 2484 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
2485 | { |
2486 | struct cfq_data *cfqd; | |
26a2ac00 | 2487 | int i; |
1da177e4 | 2488 | |
94f6030c | 2489 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 2490 | if (!cfqd) |
bc1c1169 | 2491 | return NULL; |
1da177e4 | 2492 | |
cc09e299 | 2493 | cfqd->service_tree = CFQ_RB_ROOT; |
26a2ac00 JA |
2494 | |
2495 | /* | |
2496 | * Not strictly needed (since RB_ROOT just clears the node and we | |
2497 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
2498 | * to add magic to the rb code | |
2499 | */ | |
2500 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
2501 | cfqd->prio_trees[i] = RB_ROOT; | |
2502 | ||
6118b70b JA |
2503 | /* |
2504 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
2505 | * Grab a permanent reference to it, so that the normal code flow | |
2506 | * will not attempt to free it. | |
2507 | */ | |
2508 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
2509 | atomic_inc(&cfqd->oom_cfqq.ref); | |
2510 | ||
d9ff4187 | 2511 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 2512 | |
1da177e4 | 2513 | cfqd->queue = q; |
1da177e4 | 2514 | |
22e2c507 JA |
2515 | init_timer(&cfqd->idle_slice_timer); |
2516 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
2517 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
2518 | ||
23e018a1 | 2519 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 2520 | |
1da177e4 | 2521 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
2522 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
2523 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
2524 | cfqd->cfq_back_max = cfq_back_max; |
2525 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
2526 | cfqd->cfq_slice[0] = cfq_slice_async; |
2527 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
2528 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
2529 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
963b72fc | 2530 | cfqd->cfq_latency = 1; |
45333d5a | 2531 | cfqd->hw_tag = 1; |
365722bb | 2532 | cfqd->last_end_sync_rq = jiffies; |
bc1c1169 | 2533 | return cfqd; |
1da177e4 LT |
2534 | } |
2535 | ||
2536 | static void cfq_slab_kill(void) | |
2537 | { | |
d6de8be7 JA |
2538 | /* |
2539 | * Caller already ensured that pending RCU callbacks are completed, | |
2540 | * so we should have no busy allocations at this point. | |
2541 | */ | |
1da177e4 LT |
2542 | if (cfq_pool) |
2543 | kmem_cache_destroy(cfq_pool); | |
2544 | if (cfq_ioc_pool) | |
2545 | kmem_cache_destroy(cfq_ioc_pool); | |
2546 | } | |
2547 | ||
2548 | static int __init cfq_slab_setup(void) | |
2549 | { | |
0a31bd5f | 2550 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
2551 | if (!cfq_pool) |
2552 | goto fail; | |
2553 | ||
34e6bbf2 | 2554 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
2555 | if (!cfq_ioc_pool) |
2556 | goto fail; | |
2557 | ||
2558 | return 0; | |
2559 | fail: | |
2560 | cfq_slab_kill(); | |
2561 | return -ENOMEM; | |
2562 | } | |
2563 | ||
1da177e4 LT |
2564 | /* |
2565 | * sysfs parts below --> | |
2566 | */ | |
1da177e4 LT |
2567 | static ssize_t |
2568 | cfq_var_show(unsigned int var, char *page) | |
2569 | { | |
2570 | return sprintf(page, "%d\n", var); | |
2571 | } | |
2572 | ||
2573 | static ssize_t | |
2574 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
2575 | { | |
2576 | char *p = (char *) page; | |
2577 | ||
2578 | *var = simple_strtoul(p, &p, 10); | |
2579 | return count; | |
2580 | } | |
2581 | ||
1da177e4 | 2582 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 2583 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 2584 | { \ |
3d1ab40f | 2585 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2586 | unsigned int __data = __VAR; \ |
2587 | if (__CONV) \ | |
2588 | __data = jiffies_to_msecs(__data); \ | |
2589 | return cfq_var_show(__data, (page)); \ | |
2590 | } | |
2591 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
2592 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
2593 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
2594 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
2595 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
2596 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
2597 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
2598 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
2599 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
963b72fc | 2600 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); |
1da177e4 LT |
2601 | #undef SHOW_FUNCTION |
2602 | ||
2603 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 2604 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 2605 | { \ |
3d1ab40f | 2606 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2607 | unsigned int __data; \ |
2608 | int ret = cfq_var_store(&__data, (page), count); \ | |
2609 | if (__data < (MIN)) \ | |
2610 | __data = (MIN); \ | |
2611 | else if (__data > (MAX)) \ | |
2612 | __data = (MAX); \ | |
2613 | if (__CONV) \ | |
2614 | *(__PTR) = msecs_to_jiffies(__data); \ | |
2615 | else \ | |
2616 | *(__PTR) = __data; \ | |
2617 | return ret; \ | |
2618 | } | |
2619 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
2620 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
2621 | UINT_MAX, 1); | |
2622 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
2623 | UINT_MAX, 1); | |
e572ec7e | 2624 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
2625 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
2626 | UINT_MAX, 0); | |
22e2c507 JA |
2627 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
2628 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
2629 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
2630 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
2631 | UINT_MAX, 0); | |
963b72fc | 2632 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); |
1da177e4 LT |
2633 | #undef STORE_FUNCTION |
2634 | ||
e572ec7e AV |
2635 | #define CFQ_ATTR(name) \ |
2636 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
2637 | ||
2638 | static struct elv_fs_entry cfq_attrs[] = { | |
2639 | CFQ_ATTR(quantum), | |
e572ec7e AV |
2640 | CFQ_ATTR(fifo_expire_sync), |
2641 | CFQ_ATTR(fifo_expire_async), | |
2642 | CFQ_ATTR(back_seek_max), | |
2643 | CFQ_ATTR(back_seek_penalty), | |
2644 | CFQ_ATTR(slice_sync), | |
2645 | CFQ_ATTR(slice_async), | |
2646 | CFQ_ATTR(slice_async_rq), | |
2647 | CFQ_ATTR(slice_idle), | |
963b72fc | 2648 | CFQ_ATTR(low_latency), |
e572ec7e | 2649 | __ATTR_NULL |
1da177e4 LT |
2650 | }; |
2651 | ||
1da177e4 LT |
2652 | static struct elevator_type iosched_cfq = { |
2653 | .ops = { | |
2654 | .elevator_merge_fn = cfq_merge, | |
2655 | .elevator_merged_fn = cfq_merged_request, | |
2656 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 2657 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 2658 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 2659 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 2660 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
2661 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
2662 | .elevator_queue_empty_fn = cfq_queue_empty, | |
2663 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
2664 | .elevator_former_req_fn = elv_rb_former_request, |
2665 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
2666 | .elevator_set_req_fn = cfq_set_request, |
2667 | .elevator_put_req_fn = cfq_put_request, | |
2668 | .elevator_may_queue_fn = cfq_may_queue, | |
2669 | .elevator_init_fn = cfq_init_queue, | |
2670 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 2671 | .trim = cfq_free_io_context, |
1da177e4 | 2672 | }, |
3d1ab40f | 2673 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
2674 | .elevator_name = "cfq", |
2675 | .elevator_owner = THIS_MODULE, | |
2676 | }; | |
2677 | ||
2678 | static int __init cfq_init(void) | |
2679 | { | |
22e2c507 JA |
2680 | /* |
2681 | * could be 0 on HZ < 1000 setups | |
2682 | */ | |
2683 | if (!cfq_slice_async) | |
2684 | cfq_slice_async = 1; | |
2685 | if (!cfq_slice_idle) | |
2686 | cfq_slice_idle = 1; | |
2687 | ||
1da177e4 LT |
2688 | if (cfq_slab_setup()) |
2689 | return -ENOMEM; | |
2690 | ||
2fdd82bd | 2691 | elv_register(&iosched_cfq); |
1da177e4 | 2692 | |
2fdd82bd | 2693 | return 0; |
1da177e4 LT |
2694 | } |
2695 | ||
2696 | static void __exit cfq_exit(void) | |
2697 | { | |
6e9a4738 | 2698 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 2699 | elv_unregister(&iosched_cfq); |
334e94de | 2700 | ioc_gone = &all_gone; |
fba82272 OH |
2701 | /* ioc_gone's update must be visible before reading ioc_count */ |
2702 | smp_wmb(); | |
d6de8be7 JA |
2703 | |
2704 | /* | |
2705 | * this also protects us from entering cfq_slab_kill() with | |
2706 | * pending RCU callbacks | |
2707 | */ | |
245b2e70 | 2708 | if (elv_ioc_count_read(cfq_ioc_count)) |
9a11b4ed | 2709 | wait_for_completion(&all_gone); |
83521d3e | 2710 | cfq_slab_kill(); |
1da177e4 LT |
2711 | } |
2712 | ||
2713 | module_init(cfq_init); | |
2714 | module_exit(cfq_exit); | |
2715 | ||
2716 | MODULE_AUTHOR("Jens Axboe"); | |
2717 | MODULE_LICENSE("GPL"); | |
2718 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |