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 | ||
b2c18e1e JM |
115 | unsigned int seek_samples; |
116 | u64 seek_total; | |
117 | sector_t seek_mean; | |
118 | sector_t last_request_pos; | |
119 | ||
6118b70b JA |
120 | pid_t pid; |
121 | }; | |
122 | ||
22e2c507 JA |
123 | /* |
124 | * Per block device queue structure | |
125 | */ | |
1da177e4 | 126 | struct cfq_data { |
165125e1 | 127 | struct request_queue *queue; |
22e2c507 JA |
128 | |
129 | /* | |
130 | * rr list of queues with requests and the count of them | |
131 | */ | |
cc09e299 | 132 | struct cfq_rb_root service_tree; |
a36e71f9 JA |
133 | |
134 | /* | |
135 | * Each priority tree is sorted by next_request position. These | |
136 | * trees are used when determining if two or more queues are | |
137 | * interleaving requests (see cfq_close_cooperator). | |
138 | */ | |
139 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
140 | ||
22e2c507 JA |
141 | unsigned int busy_queues; |
142 | ||
5ad531db | 143 | int rq_in_driver[2]; |
3ed9a296 | 144 | int sync_flight; |
45333d5a AC |
145 | |
146 | /* | |
147 | * queue-depth detection | |
148 | */ | |
149 | int rq_queued; | |
25776e35 | 150 | int hw_tag; |
45333d5a AC |
151 | int hw_tag_samples; |
152 | int rq_in_driver_peak; | |
1da177e4 | 153 | |
22e2c507 JA |
154 | /* |
155 | * idle window management | |
156 | */ | |
157 | struct timer_list idle_slice_timer; | |
23e018a1 | 158 | struct work_struct unplug_work; |
1da177e4 | 159 | |
22e2c507 JA |
160 | struct cfq_queue *active_queue; |
161 | struct cfq_io_context *active_cic; | |
22e2c507 | 162 | |
c2dea2d1 VT |
163 | /* |
164 | * async queue for each priority case | |
165 | */ | |
166 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
167 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 168 | |
6d048f53 | 169 | sector_t last_position; |
1da177e4 | 170 | |
1da177e4 LT |
171 | /* |
172 | * tunables, see top of file | |
173 | */ | |
174 | unsigned int cfq_quantum; | |
22e2c507 | 175 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
176 | unsigned int cfq_back_penalty; |
177 | unsigned int cfq_back_max; | |
22e2c507 JA |
178 | unsigned int cfq_slice[2]; |
179 | unsigned int cfq_slice_async_rq; | |
180 | unsigned int cfq_slice_idle; | |
963b72fc | 181 | unsigned int cfq_latency; |
d9ff4187 AV |
182 | |
183 | struct list_head cic_list; | |
1da177e4 | 184 | |
6118b70b JA |
185 | /* |
186 | * Fallback dummy cfqq for extreme OOM conditions | |
187 | */ | |
188 | struct cfq_queue oom_cfqq; | |
365722bb VG |
189 | |
190 | unsigned long last_end_sync_rq; | |
1da177e4 LT |
191 | }; |
192 | ||
3b18152c | 193 | enum cfqq_state_flags { |
b0b8d749 JA |
194 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
195 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 196 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 | 197 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ |
b0b8d749 JA |
198 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
199 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
200 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 201 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 202 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
a36e71f9 | 203 | CFQ_CFQQ_FLAG_coop, /* has done a coop jump of the queue */ |
3b18152c JA |
204 | }; |
205 | ||
206 | #define CFQ_CFQQ_FNS(name) \ | |
207 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
208 | { \ | |
fe094d98 | 209 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
210 | } \ |
211 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
212 | { \ | |
fe094d98 | 213 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
214 | } \ |
215 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
216 | { \ | |
fe094d98 | 217 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
218 | } |
219 | ||
220 | CFQ_CFQQ_FNS(on_rr); | |
221 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 222 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c | 223 | CFQ_CFQQ_FNS(must_alloc_slice); |
3b18152c JA |
224 | CFQ_CFQQ_FNS(fifo_expire); |
225 | CFQ_CFQQ_FNS(idle_window); | |
226 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 227 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 228 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 229 | CFQ_CFQQ_FNS(coop); |
3b18152c JA |
230 | #undef CFQ_CFQQ_FNS |
231 | ||
7b679138 JA |
232 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
233 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
234 | #define cfq_log(cfqd, fmt, args...) \ | |
235 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
236 | ||
165125e1 | 237 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
a6151c3a | 238 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, |
fd0928df | 239 | struct io_context *, gfp_t); |
4ac845a2 | 240 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
241 | struct io_context *); |
242 | ||
5ad531db JA |
243 | static inline int rq_in_driver(struct cfq_data *cfqd) |
244 | { | |
245 | return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1]; | |
246 | } | |
247 | ||
91fac317 | 248 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, |
a6151c3a | 249 | bool is_sync) |
91fac317 | 250 | { |
a6151c3a | 251 | return cic->cfqq[is_sync]; |
91fac317 VT |
252 | } |
253 | ||
254 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 255 | struct cfq_queue *cfqq, bool is_sync) |
91fac317 | 256 | { |
a6151c3a | 257 | cic->cfqq[is_sync] = cfqq; |
91fac317 VT |
258 | } |
259 | ||
260 | /* | |
261 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
262 | * set (in which case it could also be direct WRITE). | |
263 | */ | |
a6151c3a | 264 | static inline bool cfq_bio_sync(struct bio *bio) |
91fac317 | 265 | { |
a6151c3a | 266 | return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO); |
91fac317 | 267 | } |
1da177e4 | 268 | |
99f95e52 AM |
269 | /* |
270 | * scheduler run of queue, if there are requests pending and no one in the | |
271 | * driver that will restart queueing | |
272 | */ | |
23e018a1 | 273 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) |
99f95e52 | 274 | { |
7b679138 JA |
275 | if (cfqd->busy_queues) { |
276 | cfq_log(cfqd, "schedule dispatch"); | |
23e018a1 | 277 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 278 | } |
99f95e52 AM |
279 | } |
280 | ||
165125e1 | 281 | static int cfq_queue_empty(struct request_queue *q) |
99f95e52 AM |
282 | { |
283 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
284 | ||
b4878f24 | 285 | return !cfqd->busy_queues; |
99f95e52 AM |
286 | } |
287 | ||
44f7c160 JA |
288 | /* |
289 | * Scale schedule slice based on io priority. Use the sync time slice only | |
290 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
291 | * io only, should not get full sync slice length. | |
292 | */ | |
a6151c3a | 293 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, |
d9e7620e | 294 | unsigned short prio) |
44f7c160 | 295 | { |
d9e7620e | 296 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 297 | |
d9e7620e JA |
298 | WARN_ON(prio >= IOPRIO_BE_NR); |
299 | ||
300 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
301 | } | |
44f7c160 | 302 | |
d9e7620e JA |
303 | static inline int |
304 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
305 | { | |
306 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
307 | } |
308 | ||
309 | static inline void | |
310 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
311 | { | |
312 | cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies; | |
7b679138 | 313 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
314 | } |
315 | ||
316 | /* | |
317 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
318 | * isn't valid until the first request from the dispatch is activated | |
319 | * and the slice time set. | |
320 | */ | |
a6151c3a | 321 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) |
44f7c160 JA |
322 | { |
323 | if (cfq_cfqq_slice_new(cfqq)) | |
324 | return 0; | |
325 | if (time_before(jiffies, cfqq->slice_end)) | |
326 | return 0; | |
327 | ||
328 | return 1; | |
329 | } | |
330 | ||
1da177e4 | 331 | /* |
5e705374 | 332 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 333 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 334 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 335 | */ |
5e705374 JA |
336 | static struct request * |
337 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2) | |
1da177e4 LT |
338 | { |
339 | sector_t last, s1, s2, d1 = 0, d2 = 0; | |
1da177e4 | 340 | unsigned long back_max; |
e8a99053 AM |
341 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
342 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
343 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 344 | |
5e705374 JA |
345 | if (rq1 == NULL || rq1 == rq2) |
346 | return rq2; | |
347 | if (rq2 == NULL) | |
348 | return rq1; | |
9c2c38a1 | 349 | |
5e705374 JA |
350 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
351 | return rq1; | |
352 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
353 | return rq2; | |
374f84ac JA |
354 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
355 | return rq1; | |
356 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
357 | return rq2; | |
1da177e4 | 358 | |
83096ebf TH |
359 | s1 = blk_rq_pos(rq1); |
360 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 361 | |
6d048f53 | 362 | last = cfqd->last_position; |
1da177e4 | 363 | |
1da177e4 LT |
364 | /* |
365 | * by definition, 1KiB is 2 sectors | |
366 | */ | |
367 | back_max = cfqd->cfq_back_max * 2; | |
368 | ||
369 | /* | |
370 | * Strict one way elevator _except_ in the case where we allow | |
371 | * short backward seeks which are biased as twice the cost of a | |
372 | * similar forward seek. | |
373 | */ | |
374 | if (s1 >= last) | |
375 | d1 = s1 - last; | |
376 | else if (s1 + back_max >= last) | |
377 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
378 | else | |
e8a99053 | 379 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
380 | |
381 | if (s2 >= last) | |
382 | d2 = s2 - last; | |
383 | else if (s2 + back_max >= last) | |
384 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
385 | else | |
e8a99053 | 386 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
387 | |
388 | /* Found required data */ | |
e8a99053 AM |
389 | |
390 | /* | |
391 | * By doing switch() on the bit mask "wrap" we avoid having to | |
392 | * check two variables for all permutations: --> faster! | |
393 | */ | |
394 | switch (wrap) { | |
5e705374 | 395 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 396 | if (d1 < d2) |
5e705374 | 397 | return rq1; |
e8a99053 | 398 | else if (d2 < d1) |
5e705374 | 399 | return rq2; |
e8a99053 AM |
400 | else { |
401 | if (s1 >= s2) | |
5e705374 | 402 | return rq1; |
e8a99053 | 403 | else |
5e705374 | 404 | return rq2; |
e8a99053 | 405 | } |
1da177e4 | 406 | |
e8a99053 | 407 | case CFQ_RQ2_WRAP: |
5e705374 | 408 | return rq1; |
e8a99053 | 409 | case CFQ_RQ1_WRAP: |
5e705374 JA |
410 | return rq2; |
411 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
412 | default: |
413 | /* | |
414 | * Since both rqs are wrapped, | |
415 | * start with the one that's further behind head | |
416 | * (--> only *one* back seek required), | |
417 | * since back seek takes more time than forward. | |
418 | */ | |
419 | if (s1 <= s2) | |
5e705374 | 420 | return rq1; |
1da177e4 | 421 | else |
5e705374 | 422 | return rq2; |
1da177e4 LT |
423 | } |
424 | } | |
425 | ||
498d3aa2 JA |
426 | /* |
427 | * The below is leftmost cache rbtree addon | |
428 | */ | |
0871714e | 429 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 JA |
430 | { |
431 | if (!root->left) | |
432 | root->left = rb_first(&root->rb); | |
433 | ||
0871714e JA |
434 | if (root->left) |
435 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
436 | ||
437 | return NULL; | |
cc09e299 JA |
438 | } |
439 | ||
a36e71f9 JA |
440 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
441 | { | |
442 | rb_erase(n, root); | |
443 | RB_CLEAR_NODE(n); | |
444 | } | |
445 | ||
cc09e299 JA |
446 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
447 | { | |
448 | if (root->left == n) | |
449 | root->left = NULL; | |
a36e71f9 | 450 | rb_erase_init(n, &root->rb); |
cc09e299 JA |
451 | } |
452 | ||
1da177e4 LT |
453 | /* |
454 | * would be nice to take fifo expire time into account as well | |
455 | */ | |
5e705374 JA |
456 | static struct request * |
457 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
458 | struct request *last) | |
1da177e4 | 459 | { |
21183b07 JA |
460 | struct rb_node *rbnext = rb_next(&last->rb_node); |
461 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 462 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 463 | |
21183b07 | 464 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
465 | |
466 | if (rbprev) | |
5e705374 | 467 | prev = rb_entry_rq(rbprev); |
1da177e4 | 468 | |
21183b07 | 469 | if (rbnext) |
5e705374 | 470 | next = rb_entry_rq(rbnext); |
21183b07 JA |
471 | else { |
472 | rbnext = rb_first(&cfqq->sort_list); | |
473 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 474 | next = rb_entry_rq(rbnext); |
21183b07 | 475 | } |
1da177e4 | 476 | |
21183b07 | 477 | return cfq_choose_req(cfqd, next, prev); |
1da177e4 LT |
478 | } |
479 | ||
d9e7620e JA |
480 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
481 | struct cfq_queue *cfqq) | |
1da177e4 | 482 | { |
d9e7620e JA |
483 | /* |
484 | * just an approximation, should be ok. | |
485 | */ | |
67e6b49e JA |
486 | return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
487 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); | |
d9e7620e JA |
488 | } |
489 | ||
498d3aa2 JA |
490 | /* |
491 | * The cfqd->service_tree holds all pending cfq_queue's that have | |
492 | * requests waiting to be processed. It is sorted in the order that | |
493 | * we will service the queues. | |
494 | */ | |
a36e71f9 | 495 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 496 | bool add_front) |
d9e7620e | 497 | { |
0871714e JA |
498 | struct rb_node **p, *parent; |
499 | struct cfq_queue *__cfqq; | |
d9e7620e | 500 | unsigned long rb_key; |
498d3aa2 | 501 | int left; |
d9e7620e | 502 | |
0871714e JA |
503 | if (cfq_class_idle(cfqq)) { |
504 | rb_key = CFQ_IDLE_DELAY; | |
505 | parent = rb_last(&cfqd->service_tree.rb); | |
506 | if (parent && parent != &cfqq->rb_node) { | |
507 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
508 | rb_key += __cfqq->rb_key; | |
509 | } else | |
510 | rb_key += jiffies; | |
511 | } else if (!add_front) { | |
b9c8946b JA |
512 | /* |
513 | * Get our rb key offset. Subtract any residual slice | |
514 | * value carried from last service. A negative resid | |
515 | * count indicates slice overrun, and this should position | |
516 | * the next service time further away in the tree. | |
517 | */ | |
edd75ffd | 518 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
b9c8946b | 519 | rb_key -= cfqq->slice_resid; |
edd75ffd | 520 | cfqq->slice_resid = 0; |
48e025e6 CZ |
521 | } else { |
522 | rb_key = -HZ; | |
523 | __cfqq = cfq_rb_first(&cfqd->service_tree); | |
524 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; | |
525 | } | |
1da177e4 | 526 | |
d9e7620e | 527 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
99f9628a | 528 | /* |
d9e7620e | 529 | * same position, nothing more to do |
99f9628a | 530 | */ |
d9e7620e JA |
531 | if (rb_key == cfqq->rb_key) |
532 | return; | |
1da177e4 | 533 | |
cc09e299 | 534 | cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree); |
1da177e4 | 535 | } |
d9e7620e | 536 | |
498d3aa2 | 537 | left = 1; |
0871714e JA |
538 | parent = NULL; |
539 | p = &cfqd->service_tree.rb.rb_node; | |
d9e7620e | 540 | while (*p) { |
67060e37 | 541 | struct rb_node **n; |
cc09e299 | 542 | |
d9e7620e JA |
543 | parent = *p; |
544 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
545 | ||
0c534e0a JA |
546 | /* |
547 | * sort RT queues first, we always want to give | |
67060e37 JA |
548 | * preference to them. IDLE queues goes to the back. |
549 | * after that, sort on the next service time. | |
0c534e0a JA |
550 | */ |
551 | if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq)) | |
67060e37 | 552 | n = &(*p)->rb_left; |
0c534e0a | 553 | else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq)) |
67060e37 JA |
554 | n = &(*p)->rb_right; |
555 | else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq)) | |
556 | n = &(*p)->rb_left; | |
557 | else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq)) | |
558 | n = &(*p)->rb_right; | |
48e025e6 | 559 | else if (time_before(rb_key, __cfqq->rb_key)) |
67060e37 JA |
560 | n = &(*p)->rb_left; |
561 | else | |
562 | n = &(*p)->rb_right; | |
563 | ||
564 | if (n == &(*p)->rb_right) | |
cc09e299 | 565 | left = 0; |
67060e37 JA |
566 | |
567 | p = n; | |
d9e7620e JA |
568 | } |
569 | ||
cc09e299 JA |
570 | if (left) |
571 | cfqd->service_tree.left = &cfqq->rb_node; | |
572 | ||
d9e7620e JA |
573 | cfqq->rb_key = rb_key; |
574 | rb_link_node(&cfqq->rb_node, parent, p); | |
cc09e299 | 575 | rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb); |
1da177e4 LT |
576 | } |
577 | ||
a36e71f9 | 578 | static struct cfq_queue * |
f2d1f0ae JA |
579 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
580 | sector_t sector, struct rb_node **ret_parent, | |
581 | struct rb_node ***rb_link) | |
a36e71f9 | 582 | { |
a36e71f9 JA |
583 | struct rb_node **p, *parent; |
584 | struct cfq_queue *cfqq = NULL; | |
585 | ||
586 | parent = NULL; | |
587 | p = &root->rb_node; | |
588 | while (*p) { | |
589 | struct rb_node **n; | |
590 | ||
591 | parent = *p; | |
592 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
593 | ||
594 | /* | |
595 | * Sort strictly based on sector. Smallest to the left, | |
596 | * largest to the right. | |
597 | */ | |
2e46e8b2 | 598 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 599 | n = &(*p)->rb_right; |
2e46e8b2 | 600 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
601 | n = &(*p)->rb_left; |
602 | else | |
603 | break; | |
604 | p = n; | |
3ac6c9f8 | 605 | cfqq = NULL; |
a36e71f9 JA |
606 | } |
607 | ||
608 | *ret_parent = parent; | |
609 | if (rb_link) | |
610 | *rb_link = p; | |
3ac6c9f8 | 611 | return cfqq; |
a36e71f9 JA |
612 | } |
613 | ||
614 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
615 | { | |
a36e71f9 JA |
616 | struct rb_node **p, *parent; |
617 | struct cfq_queue *__cfqq; | |
618 | ||
f2d1f0ae JA |
619 | if (cfqq->p_root) { |
620 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
621 | cfqq->p_root = NULL; | |
622 | } | |
a36e71f9 JA |
623 | |
624 | if (cfq_class_idle(cfqq)) | |
625 | return; | |
626 | if (!cfqq->next_rq) | |
627 | return; | |
628 | ||
f2d1f0ae | 629 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
630 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
631 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
632 | if (!__cfqq) { |
633 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
634 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
635 | } else | |
636 | cfqq->p_root = NULL; | |
a36e71f9 JA |
637 | } |
638 | ||
498d3aa2 JA |
639 | /* |
640 | * Update cfqq's position in the service tree. | |
641 | */ | |
edd75ffd | 642 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 643 | { |
6d048f53 JA |
644 | /* |
645 | * Resorting requires the cfqq to be on the RR list already. | |
646 | */ | |
a36e71f9 | 647 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 648 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
649 | cfq_prio_tree_add(cfqd, cfqq); |
650 | } | |
6d048f53 JA |
651 | } |
652 | ||
1da177e4 LT |
653 | /* |
654 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 655 | * the pending list according to last request service |
1da177e4 | 656 | */ |
febffd61 | 657 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 658 | { |
7b679138 | 659 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
660 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
661 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 LT |
662 | cfqd->busy_queues++; |
663 | ||
edd75ffd | 664 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
665 | } |
666 | ||
498d3aa2 JA |
667 | /* |
668 | * Called when the cfqq no longer has requests pending, remove it from | |
669 | * the service tree. | |
670 | */ | |
febffd61 | 671 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 672 | { |
7b679138 | 673 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
674 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
675 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 676 | |
cc09e299 JA |
677 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) |
678 | cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree); | |
f2d1f0ae JA |
679 | if (cfqq->p_root) { |
680 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
681 | cfqq->p_root = NULL; | |
682 | } | |
d9e7620e | 683 | |
1da177e4 LT |
684 | BUG_ON(!cfqd->busy_queues); |
685 | cfqd->busy_queues--; | |
686 | } | |
687 | ||
688 | /* | |
689 | * rb tree support functions | |
690 | */ | |
febffd61 | 691 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 692 | { |
5e705374 | 693 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 694 | struct cfq_data *cfqd = cfqq->cfqd; |
5e705374 | 695 | const int sync = rq_is_sync(rq); |
1da177e4 | 696 | |
b4878f24 JA |
697 | BUG_ON(!cfqq->queued[sync]); |
698 | cfqq->queued[sync]--; | |
1da177e4 | 699 | |
5e705374 | 700 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 701 | |
dd67d051 | 702 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
b4878f24 | 703 | cfq_del_cfqq_rr(cfqd, cfqq); |
1da177e4 LT |
704 | } |
705 | ||
5e705374 | 706 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 707 | { |
5e705374 | 708 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 709 | struct cfq_data *cfqd = cfqq->cfqd; |
a36e71f9 | 710 | struct request *__alias, *prev; |
1da177e4 | 711 | |
5380a101 | 712 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
713 | |
714 | /* | |
715 | * looks a little odd, but the first insert might return an alias. | |
716 | * if that happens, put the alias on the dispatch list | |
717 | */ | |
21183b07 | 718 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 719 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
720 | |
721 | if (!cfq_cfqq_on_rr(cfqq)) | |
722 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
723 | |
724 | /* | |
725 | * check if this request is a better next-serve candidate | |
726 | */ | |
a36e71f9 | 727 | prev = cfqq->next_rq; |
5044eed4 | 728 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq); |
a36e71f9 JA |
729 | |
730 | /* | |
731 | * adjust priority tree position, if ->next_rq changes | |
732 | */ | |
733 | if (prev != cfqq->next_rq) | |
734 | cfq_prio_tree_add(cfqd, cfqq); | |
735 | ||
5044eed4 | 736 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
737 | } |
738 | ||
febffd61 | 739 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 740 | { |
5380a101 JA |
741 | elv_rb_del(&cfqq->sort_list, rq); |
742 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 743 | cfq_add_rq_rb(rq); |
1da177e4 LT |
744 | } |
745 | ||
206dc69b JA |
746 | static struct request * |
747 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 748 | { |
206dc69b | 749 | struct task_struct *tsk = current; |
91fac317 | 750 | struct cfq_io_context *cic; |
206dc69b | 751 | struct cfq_queue *cfqq; |
1da177e4 | 752 | |
4ac845a2 | 753 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
754 | if (!cic) |
755 | return NULL; | |
756 | ||
757 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
758 | if (cfqq) { |
759 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
760 | ||
21183b07 | 761 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 762 | } |
1da177e4 | 763 | |
1da177e4 LT |
764 | return NULL; |
765 | } | |
766 | ||
165125e1 | 767 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 768 | { |
22e2c507 | 769 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 770 | |
5ad531db | 771 | cfqd->rq_in_driver[rq_is_sync(rq)]++; |
7b679138 | 772 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
5ad531db | 773 | rq_in_driver(cfqd)); |
25776e35 | 774 | |
5b93629b | 775 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
776 | } |
777 | ||
165125e1 | 778 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 779 | { |
b4878f24 | 780 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5ad531db | 781 | const int sync = rq_is_sync(rq); |
b4878f24 | 782 | |
5ad531db JA |
783 | WARN_ON(!cfqd->rq_in_driver[sync]); |
784 | cfqd->rq_in_driver[sync]--; | |
7b679138 | 785 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
5ad531db | 786 | rq_in_driver(cfqd)); |
1da177e4 LT |
787 | } |
788 | ||
b4878f24 | 789 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 790 | { |
5e705374 | 791 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 792 | |
5e705374 JA |
793 | if (cfqq->next_rq == rq) |
794 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 795 | |
b4878f24 | 796 | list_del_init(&rq->queuelist); |
5e705374 | 797 | cfq_del_rq_rb(rq); |
374f84ac | 798 | |
45333d5a | 799 | cfqq->cfqd->rq_queued--; |
374f84ac JA |
800 | if (rq_is_meta(rq)) { |
801 | WARN_ON(!cfqq->meta_pending); | |
802 | cfqq->meta_pending--; | |
803 | } | |
1da177e4 LT |
804 | } |
805 | ||
165125e1 JA |
806 | static int cfq_merge(struct request_queue *q, struct request **req, |
807 | struct bio *bio) | |
1da177e4 LT |
808 | { |
809 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
810 | struct request *__rq; | |
1da177e4 | 811 | |
206dc69b | 812 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 813 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
814 | *req = __rq; |
815 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
816 | } |
817 | ||
818 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
819 | } |
820 | ||
165125e1 | 821 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 822 | int type) |
1da177e4 | 823 | { |
21183b07 | 824 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 825 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 826 | |
5e705374 | 827 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 828 | } |
1da177e4 LT |
829 | } |
830 | ||
831 | static void | |
165125e1 | 832 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
833 | struct request *next) |
834 | { | |
22e2c507 JA |
835 | /* |
836 | * reposition in fifo if next is older than rq | |
837 | */ | |
838 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
30996f40 | 839 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
22e2c507 | 840 | list_move(&rq->queuelist, &next->queuelist); |
30996f40 JA |
841 | rq_set_fifo_time(rq, rq_fifo_time(next)); |
842 | } | |
22e2c507 | 843 | |
b4878f24 | 844 | cfq_remove_request(next); |
22e2c507 JA |
845 | } |
846 | ||
165125e1 | 847 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
848 | struct bio *bio) |
849 | { | |
850 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 851 | struct cfq_io_context *cic; |
da775265 | 852 | struct cfq_queue *cfqq; |
da775265 JA |
853 | |
854 | /* | |
ec8acb69 | 855 | * Disallow merge of a sync bio into an async request. |
da775265 | 856 | */ |
91fac317 | 857 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
a6151c3a | 858 | return false; |
da775265 JA |
859 | |
860 | /* | |
719d3402 JA |
861 | * Lookup the cfqq that this bio will be queued with. Allow |
862 | * merge only if rq is queued there. | |
da775265 | 863 | */ |
4ac845a2 | 864 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 | 865 | if (!cic) |
a6151c3a | 866 | return false; |
719d3402 | 867 | |
91fac317 | 868 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
a6151c3a | 869 | return cfqq == RQ_CFQQ(rq); |
da775265 JA |
870 | } |
871 | ||
febffd61 JA |
872 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
873 | struct cfq_queue *cfqq) | |
22e2c507 JA |
874 | { |
875 | if (cfqq) { | |
7b679138 | 876 | cfq_log_cfqq(cfqd, cfqq, "set_active"); |
22e2c507 | 877 | cfqq->slice_end = 0; |
2f5cb738 JA |
878 | cfqq->slice_dispatch = 0; |
879 | ||
2f5cb738 | 880 | cfq_clear_cfqq_wait_request(cfqq); |
b029195d | 881 | cfq_clear_cfqq_must_dispatch(cfqq); |
3b18152c JA |
882 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
883 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 884 | cfq_mark_cfqq_slice_new(cfqq); |
2f5cb738 JA |
885 | |
886 | del_timer(&cfqd->idle_slice_timer); | |
22e2c507 JA |
887 | } |
888 | ||
889 | cfqd->active_queue = cfqq; | |
890 | } | |
891 | ||
7b14e3b5 JA |
892 | /* |
893 | * current cfqq expired its slice (or was too idle), select new one | |
894 | */ | |
895 | static void | |
896 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
a6151c3a | 897 | bool timed_out) |
7b14e3b5 | 898 | { |
7b679138 JA |
899 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
900 | ||
7b14e3b5 JA |
901 | if (cfq_cfqq_wait_request(cfqq)) |
902 | del_timer(&cfqd->idle_slice_timer); | |
903 | ||
7b14e3b5 JA |
904 | cfq_clear_cfqq_wait_request(cfqq); |
905 | ||
906 | /* | |
6084cdda | 907 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 908 | */ |
7b679138 | 909 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) { |
c5b680f3 | 910 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b679138 JA |
911 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
912 | } | |
7b14e3b5 | 913 | |
edd75ffd | 914 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
915 | |
916 | if (cfqq == cfqd->active_queue) | |
917 | cfqd->active_queue = NULL; | |
918 | ||
919 | if (cfqd->active_cic) { | |
920 | put_io_context(cfqd->active_cic->ioc); | |
921 | cfqd->active_cic = NULL; | |
922 | } | |
7b14e3b5 JA |
923 | } |
924 | ||
a6151c3a | 925 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) |
7b14e3b5 JA |
926 | { |
927 | struct cfq_queue *cfqq = cfqd->active_queue; | |
928 | ||
929 | if (cfqq) | |
6084cdda | 930 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
931 | } |
932 | ||
498d3aa2 JA |
933 | /* |
934 | * Get next queue for service. Unless we have a queue preemption, | |
935 | * we'll simply select the first cfqq in the service tree. | |
936 | */ | |
6d048f53 | 937 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 938 | { |
edd75ffd JA |
939 | if (RB_EMPTY_ROOT(&cfqd->service_tree.rb)) |
940 | return NULL; | |
d9e7620e | 941 | |
0871714e | 942 | return cfq_rb_first(&cfqd->service_tree); |
6d048f53 JA |
943 | } |
944 | ||
498d3aa2 JA |
945 | /* |
946 | * Get and set a new active queue for service. | |
947 | */ | |
a36e71f9 JA |
948 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
949 | struct cfq_queue *cfqq) | |
6d048f53 | 950 | { |
a36e71f9 JA |
951 | if (!cfqq) { |
952 | cfqq = cfq_get_next_queue(cfqd); | |
953 | if (cfqq) | |
954 | cfq_clear_cfqq_coop(cfqq); | |
955 | } | |
6d048f53 | 956 | |
22e2c507 | 957 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 958 | return cfqq; |
22e2c507 JA |
959 | } |
960 | ||
d9e7620e JA |
961 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
962 | struct request *rq) | |
963 | { | |
83096ebf TH |
964 | if (blk_rq_pos(rq) >= cfqd->last_position) |
965 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 966 | else |
83096ebf | 967 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
968 | } |
969 | ||
b2c18e1e JM |
970 | #define CFQQ_SEEK_THR 8 * 1024 |
971 | #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR) | |
04dc6e71 | 972 | |
b2c18e1e JM |
973 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
974 | struct request *rq) | |
6d048f53 | 975 | { |
b2c18e1e | 976 | sector_t sdist = cfqq->seek_mean; |
6d048f53 | 977 | |
b2c18e1e JM |
978 | if (!sample_valid(cfqq->seek_samples)) |
979 | sdist = CFQQ_SEEK_THR; | |
6d048f53 | 980 | |
04dc6e71 | 981 | return cfq_dist_from_last(cfqd, rq) <= sdist; |
6d048f53 JA |
982 | } |
983 | ||
a36e71f9 JA |
984 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
985 | struct cfq_queue *cur_cfqq) | |
986 | { | |
f2d1f0ae | 987 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
988 | struct rb_node *parent, *node; |
989 | struct cfq_queue *__cfqq; | |
990 | sector_t sector = cfqd->last_position; | |
991 | ||
992 | if (RB_EMPTY_ROOT(root)) | |
993 | return NULL; | |
994 | ||
995 | /* | |
996 | * First, if we find a request starting at the end of the last | |
997 | * request, choose it. | |
998 | */ | |
f2d1f0ae | 999 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
1000 | if (__cfqq) |
1001 | return __cfqq; | |
1002 | ||
1003 | /* | |
1004 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1005 | * will contain the closest sector. | |
1006 | */ | |
1007 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
b2c18e1e | 1008 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1009 | return __cfqq; |
1010 | ||
2e46e8b2 | 1011 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1012 | node = rb_next(&__cfqq->p_node); |
1013 | else | |
1014 | node = rb_prev(&__cfqq->p_node); | |
1015 | if (!node) | |
1016 | return NULL; | |
1017 | ||
1018 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
b2c18e1e | 1019 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1020 | return __cfqq; |
1021 | ||
1022 | return NULL; | |
1023 | } | |
1024 | ||
1025 | /* | |
1026 | * cfqd - obvious | |
1027 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1028 | * closely cooperating with itself. | |
1029 | * | |
1030 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1031 | * one request, and that cfqd->last_position reflects a position on the disk | |
1032 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1033 | * assumption. | |
1034 | */ | |
1035 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
1036 | struct cfq_queue *cur_cfqq, | |
a6151c3a | 1037 | bool probe) |
6d048f53 | 1038 | { |
a36e71f9 JA |
1039 | struct cfq_queue *cfqq; |
1040 | ||
6d048f53 | 1041 | /* |
d9e7620e JA |
1042 | * We should notice if some of the queues are cooperating, eg |
1043 | * working closely on the same area of the disk. In that case, | |
1044 | * we can group them together and don't waste time idling. | |
6d048f53 | 1045 | */ |
a36e71f9 JA |
1046 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1047 | if (!cfqq) | |
1048 | return NULL; | |
1049 | ||
1050 | if (cfq_cfqq_coop(cfqq)) | |
1051 | return NULL; | |
1052 | ||
1053 | if (!probe) | |
1054 | cfq_mark_cfqq_coop(cfqq); | |
1055 | return cfqq; | |
6d048f53 JA |
1056 | } |
1057 | ||
6d048f53 | 1058 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1059 | { |
1792669c | 1060 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1061 | struct cfq_io_context *cic; |
7b14e3b5 JA |
1062 | unsigned long sl; |
1063 | ||
a68bbddb | 1064 | /* |
f7d7b7a7 JA |
1065 | * SSD device without seek penalty, disable idling. But only do so |
1066 | * for devices that support queuing, otherwise we still have a problem | |
1067 | * with sync vs async workloads. | |
a68bbddb | 1068 | */ |
f7d7b7a7 | 1069 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
1070 | return; |
1071 | ||
dd67d051 | 1072 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 1073 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
1074 | |
1075 | /* | |
1076 | * idle is disabled, either manually or by past process history | |
1077 | */ | |
6d048f53 JA |
1078 | if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq)) |
1079 | return; | |
1080 | ||
7b679138 JA |
1081 | /* |
1082 | * still requests with the driver, don't idle | |
1083 | */ | |
5ad531db | 1084 | if (rq_in_driver(cfqd)) |
7b679138 JA |
1085 | return; |
1086 | ||
22e2c507 JA |
1087 | /* |
1088 | * task has exited, don't wait | |
1089 | */ | |
206dc69b | 1090 | cic = cfqd->active_cic; |
66dac98e | 1091 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
1092 | return; |
1093 | ||
355b659c CZ |
1094 | /* |
1095 | * If our average think time is larger than the remaining time | |
1096 | * slice, then don't idle. This avoids overrunning the allotted | |
1097 | * time slice. | |
1098 | */ | |
1099 | if (sample_valid(cic->ttime_samples) && | |
1100 | (cfqq->slice_end - jiffies < cic->ttime_mean)) | |
1101 | return; | |
1102 | ||
3b18152c | 1103 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 1104 | |
206dc69b JA |
1105 | /* |
1106 | * we don't want to idle for seeks, but we do want to allow | |
1107 | * fair distribution of slice time for a process doing back-to-back | |
1108 | * seeks. so allow a little bit of time for him to submit a new rq | |
1109 | */ | |
6d048f53 | 1110 | sl = cfqd->cfq_slice_idle; |
b2c18e1e | 1111 | if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)) |
d9e7620e | 1112 | sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT)); |
206dc69b | 1113 | |
7b14e3b5 | 1114 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
9481ffdc | 1115 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); |
1da177e4 LT |
1116 | } |
1117 | ||
498d3aa2 JA |
1118 | /* |
1119 | * Move request from internal lists to the request queue dispatch list. | |
1120 | */ | |
165125e1 | 1121 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 1122 | { |
3ed9a296 | 1123 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1124 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 1125 | |
7b679138 JA |
1126 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
1127 | ||
06d21886 | 1128 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); |
5380a101 | 1129 | cfq_remove_request(rq); |
6d048f53 | 1130 | cfqq->dispatched++; |
5380a101 | 1131 | elv_dispatch_sort(q, rq); |
3ed9a296 JA |
1132 | |
1133 | if (cfq_cfqq_sync(cfqq)) | |
1134 | cfqd->sync_flight++; | |
1da177e4 LT |
1135 | } |
1136 | ||
1137 | /* | |
1138 | * return expired entry, or NULL to just start from scratch in rbtree | |
1139 | */ | |
febffd61 | 1140 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 | 1141 | { |
30996f40 | 1142 | struct request *rq = NULL; |
1da177e4 | 1143 | |
3b18152c | 1144 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 1145 | return NULL; |
cb887411 JA |
1146 | |
1147 | cfq_mark_cfqq_fifo_expire(cfqq); | |
1148 | ||
89850f7e JA |
1149 | if (list_empty(&cfqq->fifo)) |
1150 | return NULL; | |
1da177e4 | 1151 | |
89850f7e | 1152 | rq = rq_entry_fifo(cfqq->fifo.next); |
30996f40 | 1153 | if (time_before(jiffies, rq_fifo_time(rq))) |
7b679138 | 1154 | rq = NULL; |
1da177e4 | 1155 | |
30996f40 | 1156 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 1157 | return rq; |
1da177e4 LT |
1158 | } |
1159 | ||
22e2c507 JA |
1160 | static inline int |
1161 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1162 | { | |
1163 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 1164 | |
22e2c507 | 1165 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 1166 | |
22e2c507 | 1167 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
1168 | } |
1169 | ||
22e2c507 | 1170 | /* |
498d3aa2 JA |
1171 | * Select a queue for service. If we have a current active queue, |
1172 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 1173 | */ |
1b5ed5e1 | 1174 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 1175 | { |
a36e71f9 | 1176 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 1177 | |
22e2c507 JA |
1178 | cfqq = cfqd->active_queue; |
1179 | if (!cfqq) | |
1180 | goto new_queue; | |
1da177e4 | 1181 | |
22e2c507 | 1182 | /* |
6d048f53 | 1183 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 1184 | */ |
b029195d | 1185 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) |
3b18152c | 1186 | goto expire; |
1da177e4 | 1187 | |
22e2c507 | 1188 | /* |
6d048f53 JA |
1189 | * The active queue has requests and isn't expired, allow it to |
1190 | * dispatch. | |
22e2c507 | 1191 | */ |
dd67d051 | 1192 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 1193 | goto keep_queue; |
6d048f53 | 1194 | |
a36e71f9 JA |
1195 | /* |
1196 | * If another queue has a request waiting within our mean seek | |
1197 | * distance, let it run. The expire code will check for close | |
1198 | * cooperators and put the close queue at the front of the service | |
1199 | * tree. | |
1200 | */ | |
1201 | new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0); | |
1202 | if (new_cfqq) | |
1203 | goto expire; | |
1204 | ||
6d048f53 JA |
1205 | /* |
1206 | * No requests pending. If the active queue still has requests in | |
1207 | * flight or is idling for a new request, allow either of these | |
1208 | * conditions to happen (or time out) before selecting a new queue. | |
1209 | */ | |
cc197479 JA |
1210 | if (timer_pending(&cfqd->idle_slice_timer) || |
1211 | (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) { | |
caaa5f9f JA |
1212 | cfqq = NULL; |
1213 | goto keep_queue; | |
22e2c507 JA |
1214 | } |
1215 | ||
3b18152c | 1216 | expire: |
6084cdda | 1217 | cfq_slice_expired(cfqd, 0); |
3b18152c | 1218 | new_queue: |
a36e71f9 | 1219 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 1220 | keep_queue: |
3b18152c | 1221 | return cfqq; |
22e2c507 JA |
1222 | } |
1223 | ||
febffd61 | 1224 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
1225 | { |
1226 | int dispatched = 0; | |
1227 | ||
1228 | while (cfqq->next_rq) { | |
1229 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
1230 | dispatched++; | |
1231 | } | |
1232 | ||
1233 | BUG_ON(!list_empty(&cfqq->fifo)); | |
1234 | return dispatched; | |
1235 | } | |
1236 | ||
498d3aa2 JA |
1237 | /* |
1238 | * Drain our current requests. Used for barriers and when switching | |
1239 | * io schedulers on-the-fly. | |
1240 | */ | |
d9e7620e | 1241 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 1242 | { |
0871714e | 1243 | struct cfq_queue *cfqq; |
d9e7620e | 1244 | int dispatched = 0; |
1b5ed5e1 | 1245 | |
0871714e | 1246 | while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL) |
d9e7620e | 1247 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); |
1b5ed5e1 | 1248 | |
6084cdda | 1249 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
1250 | |
1251 | BUG_ON(cfqd->busy_queues); | |
1252 | ||
6923715a | 1253 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
1254 | return dispatched; |
1255 | } | |
1256 | ||
0b182d61 | 1257 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
2f5cb738 | 1258 | { |
2f5cb738 | 1259 | unsigned int max_dispatch; |
22e2c507 | 1260 | |
5ad531db JA |
1261 | /* |
1262 | * Drain async requests before we start sync IO | |
1263 | */ | |
1264 | if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC]) | |
0b182d61 | 1265 | return false; |
5ad531db | 1266 | |
2f5cb738 JA |
1267 | /* |
1268 | * If this is an async queue and we have sync IO in flight, let it wait | |
1269 | */ | |
1270 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) | |
0b182d61 | 1271 | return false; |
2f5cb738 JA |
1272 | |
1273 | max_dispatch = cfqd->cfq_quantum; | |
1274 | if (cfq_class_idle(cfqq)) | |
1275 | max_dispatch = 1; | |
b4878f24 | 1276 | |
2f5cb738 JA |
1277 | /* |
1278 | * Does this cfqq already have too much IO in flight? | |
1279 | */ | |
1280 | if (cfqq->dispatched >= max_dispatch) { | |
1281 | /* | |
1282 | * idle queue must always only have a single IO in flight | |
1283 | */ | |
3ed9a296 | 1284 | if (cfq_class_idle(cfqq)) |
0b182d61 | 1285 | return false; |
3ed9a296 | 1286 | |
2f5cb738 JA |
1287 | /* |
1288 | * We have other queues, don't allow more IO from this one | |
1289 | */ | |
1290 | if (cfqd->busy_queues > 1) | |
0b182d61 | 1291 | return false; |
9ede209e | 1292 | |
365722bb | 1293 | /* |
8e296755 | 1294 | * Sole queue user, allow bigger slice |
365722bb | 1295 | */ |
8e296755 JA |
1296 | max_dispatch *= 4; |
1297 | } | |
1298 | ||
1299 | /* | |
1300 | * Async queues must wait a bit before being allowed dispatch. | |
1301 | * We also ramp up the dispatch depth gradually for async IO, | |
1302 | * based on the last sync IO we serviced | |
1303 | */ | |
963b72fc | 1304 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { |
8e296755 JA |
1305 | unsigned long last_sync = jiffies - cfqd->last_end_sync_rq; |
1306 | unsigned int depth; | |
365722bb | 1307 | |
61f0c1dc | 1308 | depth = last_sync / cfqd->cfq_slice[1]; |
e00c54c3 JA |
1309 | if (!depth && !cfqq->dispatched) |
1310 | depth = 1; | |
8e296755 JA |
1311 | if (depth < max_dispatch) |
1312 | max_dispatch = depth; | |
2f5cb738 | 1313 | } |
3ed9a296 | 1314 | |
0b182d61 JA |
1315 | /* |
1316 | * If we're below the current max, allow a dispatch | |
1317 | */ | |
1318 | return cfqq->dispatched < max_dispatch; | |
1319 | } | |
1320 | ||
1321 | /* | |
1322 | * Dispatch a request from cfqq, moving them to the request queue | |
1323 | * dispatch list. | |
1324 | */ | |
1325 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1326 | { | |
1327 | struct request *rq; | |
1328 | ||
1329 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
1330 | ||
1331 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
1332 | return false; | |
1333 | ||
1334 | /* | |
1335 | * follow expired path, else get first next available | |
1336 | */ | |
1337 | rq = cfq_check_fifo(cfqq); | |
1338 | if (!rq) | |
1339 | rq = cfqq->next_rq; | |
1340 | ||
1341 | /* | |
1342 | * insert request into driver dispatch list | |
1343 | */ | |
1344 | cfq_dispatch_insert(cfqd->queue, rq); | |
1345 | ||
1346 | if (!cfqd->active_cic) { | |
1347 | struct cfq_io_context *cic = RQ_CIC(rq); | |
1348 | ||
1349 | atomic_long_inc(&cic->ioc->refcount); | |
1350 | cfqd->active_cic = cic; | |
1351 | } | |
1352 | ||
1353 | return true; | |
1354 | } | |
1355 | ||
1356 | /* | |
1357 | * Find the cfqq that we need to service and move a request from that to the | |
1358 | * dispatch list | |
1359 | */ | |
1360 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
1361 | { | |
1362 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1363 | struct cfq_queue *cfqq; | |
1364 | ||
1365 | if (!cfqd->busy_queues) | |
1366 | return 0; | |
1367 | ||
1368 | if (unlikely(force)) | |
1369 | return cfq_forced_dispatch(cfqd); | |
1370 | ||
1371 | cfqq = cfq_select_queue(cfqd); | |
1372 | if (!cfqq) | |
8e296755 JA |
1373 | return 0; |
1374 | ||
2f5cb738 | 1375 | /* |
0b182d61 | 1376 | * Dispatch a request from this cfqq, if it is allowed |
2f5cb738 | 1377 | */ |
0b182d61 JA |
1378 | if (!cfq_dispatch_request(cfqd, cfqq)) |
1379 | return 0; | |
1380 | ||
2f5cb738 | 1381 | cfqq->slice_dispatch++; |
b029195d | 1382 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 1383 | |
2f5cb738 JA |
1384 | /* |
1385 | * expire an async queue immediately if it has used up its slice. idle | |
1386 | * queue always expire after 1 dispatch round. | |
1387 | */ | |
1388 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
1389 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
1390 | cfq_class_idle(cfqq))) { | |
1391 | cfqq->slice_end = jiffies + 1; | |
1392 | cfq_slice_expired(cfqd, 0); | |
1da177e4 LT |
1393 | } |
1394 | ||
b217a903 | 1395 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); |
2f5cb738 | 1396 | return 1; |
1da177e4 LT |
1397 | } |
1398 | ||
1da177e4 | 1399 | /* |
5e705374 JA |
1400 | * task holds one reference to the queue, dropped when task exits. each rq |
1401 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 LT |
1402 | * |
1403 | * queue lock must be held here. | |
1404 | */ | |
1405 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
1406 | { | |
22e2c507 JA |
1407 | struct cfq_data *cfqd = cfqq->cfqd; |
1408 | ||
1409 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
1410 | |
1411 | if (!atomic_dec_and_test(&cfqq->ref)) | |
1412 | return; | |
1413 | ||
7b679138 | 1414 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 1415 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 1416 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
3b18152c | 1417 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 | 1418 | |
28f95cbc | 1419 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 1420 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 1421 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 1422 | } |
22e2c507 | 1423 | |
1da177e4 LT |
1424 | kmem_cache_free(cfq_pool, cfqq); |
1425 | } | |
1426 | ||
d6de8be7 JA |
1427 | /* |
1428 | * Must always be called with the rcu_read_lock() held | |
1429 | */ | |
07416d29 JA |
1430 | static void |
1431 | __call_for_each_cic(struct io_context *ioc, | |
1432 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1433 | { | |
1434 | struct cfq_io_context *cic; | |
1435 | struct hlist_node *n; | |
1436 | ||
1437 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) | |
1438 | func(ioc, cic); | |
1439 | } | |
1440 | ||
4ac845a2 | 1441 | /* |
34e6bbf2 | 1442 | * Call func for each cic attached to this ioc. |
4ac845a2 | 1443 | */ |
34e6bbf2 | 1444 | static void |
4ac845a2 JA |
1445 | call_for_each_cic(struct io_context *ioc, |
1446 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1da177e4 | 1447 | { |
4ac845a2 | 1448 | rcu_read_lock(); |
07416d29 | 1449 | __call_for_each_cic(ioc, func); |
4ac845a2 | 1450 | rcu_read_unlock(); |
34e6bbf2 FC |
1451 | } |
1452 | ||
1453 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
1454 | { | |
1455 | struct cfq_io_context *cic; | |
1456 | ||
1457 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
1458 | ||
1459 | kmem_cache_free(cfq_ioc_pool, cic); | |
245b2e70 | 1460 | elv_ioc_count_dec(cfq_ioc_count); |
34e6bbf2 | 1461 | |
9a11b4ed JA |
1462 | if (ioc_gone) { |
1463 | /* | |
1464 | * CFQ scheduler is exiting, grab exit lock and check | |
1465 | * the pending io context count. If it hits zero, | |
1466 | * complete ioc_gone and set it back to NULL | |
1467 | */ | |
1468 | spin_lock(&ioc_gone_lock); | |
245b2e70 | 1469 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { |
9a11b4ed JA |
1470 | complete(ioc_gone); |
1471 | ioc_gone = NULL; | |
1472 | } | |
1473 | spin_unlock(&ioc_gone_lock); | |
1474 | } | |
34e6bbf2 | 1475 | } |
4ac845a2 | 1476 | |
34e6bbf2 FC |
1477 | static void cfq_cic_free(struct cfq_io_context *cic) |
1478 | { | |
1479 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
1480 | } |
1481 | ||
1482 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
1483 | { | |
1484 | unsigned long flags; | |
1485 | ||
1486 | BUG_ON(!cic->dead_key); | |
1487 | ||
1488 | spin_lock_irqsave(&ioc->lock, flags); | |
1489 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
ffc4e759 | 1490 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
1491 | spin_unlock_irqrestore(&ioc->lock, flags); |
1492 | ||
34e6bbf2 | 1493 | cfq_cic_free(cic); |
4ac845a2 JA |
1494 | } |
1495 | ||
d6de8be7 JA |
1496 | /* |
1497 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
1498 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
1499 | * and ->trim() which is called with the task lock held | |
1500 | */ | |
4ac845a2 JA |
1501 | static void cfq_free_io_context(struct io_context *ioc) |
1502 | { | |
4ac845a2 | 1503 | /* |
34e6bbf2 FC |
1504 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
1505 | * so no more cic's are allowed to be linked into this ioc. So it | |
1506 | * should be ok to iterate over the known list, we will see all cic's | |
1507 | * since no new ones are added. | |
4ac845a2 | 1508 | */ |
07416d29 | 1509 | __call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
1510 | } |
1511 | ||
89850f7e | 1512 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1513 | { |
28f95cbc | 1514 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 1515 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 1516 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 1517 | } |
22e2c507 | 1518 | |
89850f7e JA |
1519 | cfq_put_queue(cfqq); |
1520 | } | |
22e2c507 | 1521 | |
89850f7e JA |
1522 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
1523 | struct cfq_io_context *cic) | |
1524 | { | |
4faa3c81 FC |
1525 | struct io_context *ioc = cic->ioc; |
1526 | ||
fc46379d | 1527 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
1528 | |
1529 | /* | |
1530 | * Make sure key == NULL is seen for dead queues | |
1531 | */ | |
fc46379d | 1532 | smp_wmb(); |
4ac845a2 | 1533 | cic->dead_key = (unsigned long) cic->key; |
fc46379d JA |
1534 | cic->key = NULL; |
1535 | ||
4faa3c81 FC |
1536 | if (ioc->ioc_data == cic) |
1537 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
1538 | ||
ff6657c6 JA |
1539 | if (cic->cfqq[BLK_RW_ASYNC]) { |
1540 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
1541 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
1542 | } |
1543 | ||
ff6657c6 JA |
1544 | if (cic->cfqq[BLK_RW_SYNC]) { |
1545 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
1546 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 1547 | } |
89850f7e JA |
1548 | } |
1549 | ||
4ac845a2 JA |
1550 | static void cfq_exit_single_io_context(struct io_context *ioc, |
1551 | struct cfq_io_context *cic) | |
89850f7e JA |
1552 | { |
1553 | struct cfq_data *cfqd = cic->key; | |
1554 | ||
89850f7e | 1555 | if (cfqd) { |
165125e1 | 1556 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 1557 | unsigned long flags; |
89850f7e | 1558 | |
4ac845a2 | 1559 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
1560 | |
1561 | /* | |
1562 | * Ensure we get a fresh copy of the ->key to prevent | |
1563 | * race between exiting task and queue | |
1564 | */ | |
1565 | smp_read_barrier_depends(); | |
1566 | if (cic->key) | |
1567 | __cfq_exit_single_io_context(cfqd, cic); | |
1568 | ||
4ac845a2 | 1569 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 1570 | } |
1da177e4 LT |
1571 | } |
1572 | ||
498d3aa2 JA |
1573 | /* |
1574 | * The process that ioc belongs to has exited, we need to clean up | |
1575 | * and put the internal structures we have that belongs to that process. | |
1576 | */ | |
e2d74ac0 | 1577 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 1578 | { |
4ac845a2 | 1579 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
1580 | } |
1581 | ||
22e2c507 | 1582 | static struct cfq_io_context * |
8267e268 | 1583 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1584 | { |
b5deef90 | 1585 | struct cfq_io_context *cic; |
1da177e4 | 1586 | |
94f6030c CL |
1587 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
1588 | cfqd->queue->node); | |
1da177e4 | 1589 | if (cic) { |
22e2c507 | 1590 | cic->last_end_request = jiffies; |
553698f9 | 1591 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 1592 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
1593 | cic->dtor = cfq_free_io_context; |
1594 | cic->exit = cfq_exit_io_context; | |
245b2e70 | 1595 | elv_ioc_count_inc(cfq_ioc_count); |
1da177e4 LT |
1596 | } |
1597 | ||
1598 | return cic; | |
1599 | } | |
1600 | ||
fd0928df | 1601 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
1602 | { |
1603 | struct task_struct *tsk = current; | |
1604 | int ioprio_class; | |
1605 | ||
3b18152c | 1606 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
1607 | return; |
1608 | ||
fd0928df | 1609 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 1610 | switch (ioprio_class) { |
fe094d98 JA |
1611 | default: |
1612 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
1613 | case IOPRIO_CLASS_NONE: | |
1614 | /* | |
6d63c275 | 1615 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
1616 | */ |
1617 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 1618 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
1619 | break; |
1620 | case IOPRIO_CLASS_RT: | |
1621 | cfqq->ioprio = task_ioprio(ioc); | |
1622 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
1623 | break; | |
1624 | case IOPRIO_CLASS_BE: | |
1625 | cfqq->ioprio = task_ioprio(ioc); | |
1626 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
1627 | break; | |
1628 | case IOPRIO_CLASS_IDLE: | |
1629 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
1630 | cfqq->ioprio = 7; | |
1631 | cfq_clear_cfqq_idle_window(cfqq); | |
1632 | break; | |
22e2c507 JA |
1633 | } |
1634 | ||
1635 | /* | |
1636 | * keep track of original prio settings in case we have to temporarily | |
1637 | * elevate the priority of this queue | |
1638 | */ | |
1639 | cfqq->org_ioprio = cfqq->ioprio; | |
1640 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 1641 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
1642 | } |
1643 | ||
febffd61 | 1644 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 1645 | { |
478a82b0 AV |
1646 | struct cfq_data *cfqd = cic->key; |
1647 | struct cfq_queue *cfqq; | |
c1b707d2 | 1648 | unsigned long flags; |
35e6077c | 1649 | |
caaa5f9f JA |
1650 | if (unlikely(!cfqd)) |
1651 | return; | |
1652 | ||
c1b707d2 | 1653 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 1654 | |
ff6657c6 | 1655 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
1656 | if (cfqq) { |
1657 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
1658 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
1659 | GFP_ATOMIC); | |
caaa5f9f | 1660 | if (new_cfqq) { |
ff6657c6 | 1661 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
1662 | cfq_put_queue(cfqq); |
1663 | } | |
22e2c507 | 1664 | } |
caaa5f9f | 1665 | |
ff6657c6 | 1666 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
1667 | if (cfqq) |
1668 | cfq_mark_cfqq_prio_changed(cfqq); | |
1669 | ||
c1b707d2 | 1670 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
1671 | } |
1672 | ||
fc46379d | 1673 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 1674 | { |
4ac845a2 | 1675 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 1676 | ioc->ioprio_changed = 0; |
22e2c507 JA |
1677 | } |
1678 | ||
d5036d77 | 1679 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 1680 | pid_t pid, bool is_sync) |
d5036d77 JA |
1681 | { |
1682 | RB_CLEAR_NODE(&cfqq->rb_node); | |
1683 | RB_CLEAR_NODE(&cfqq->p_node); | |
1684 | INIT_LIST_HEAD(&cfqq->fifo); | |
1685 | ||
1686 | atomic_set(&cfqq->ref, 0); | |
1687 | cfqq->cfqd = cfqd; | |
1688 | ||
1689 | cfq_mark_cfqq_prio_changed(cfqq); | |
1690 | ||
1691 | if (is_sync) { | |
1692 | if (!cfq_class_idle(cfqq)) | |
1693 | cfq_mark_cfqq_idle_window(cfqq); | |
1694 | cfq_mark_cfqq_sync(cfqq); | |
1695 | } | |
1696 | cfqq->pid = pid; | |
1697 | } | |
1698 | ||
22e2c507 | 1699 | static struct cfq_queue * |
a6151c3a | 1700 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, |
fd0928df | 1701 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 1702 | { |
22e2c507 | 1703 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 1704 | struct cfq_io_context *cic; |
22e2c507 JA |
1705 | |
1706 | retry: | |
4ac845a2 | 1707 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
1708 | /* cic always exists here */ |
1709 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 1710 | |
6118b70b JA |
1711 | /* |
1712 | * Always try a new alloc if we fell back to the OOM cfqq | |
1713 | * originally, since it should just be a temporary situation. | |
1714 | */ | |
1715 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
1716 | cfqq = NULL; | |
22e2c507 JA |
1717 | if (new_cfqq) { |
1718 | cfqq = new_cfqq; | |
1719 | new_cfqq = NULL; | |
1720 | } else if (gfp_mask & __GFP_WAIT) { | |
1721 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 1722 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 1723 | gfp_mask | __GFP_ZERO, |
94f6030c | 1724 | cfqd->queue->node); |
22e2c507 | 1725 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
1726 | if (new_cfqq) |
1727 | goto retry; | |
22e2c507 | 1728 | } else { |
94f6030c CL |
1729 | cfqq = kmem_cache_alloc_node(cfq_pool, |
1730 | gfp_mask | __GFP_ZERO, | |
1731 | cfqd->queue->node); | |
22e2c507 JA |
1732 | } |
1733 | ||
6118b70b JA |
1734 | if (cfqq) { |
1735 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
1736 | cfq_init_prio_data(cfqq, ioc); | |
1737 | cfq_log_cfqq(cfqd, cfqq, "alloced"); | |
1738 | } else | |
1739 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
1740 | } |
1741 | ||
1742 | if (new_cfqq) | |
1743 | kmem_cache_free(cfq_pool, new_cfqq); | |
1744 | ||
22e2c507 JA |
1745 | return cfqq; |
1746 | } | |
1747 | ||
c2dea2d1 VT |
1748 | static struct cfq_queue ** |
1749 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
1750 | { | |
fe094d98 | 1751 | switch (ioprio_class) { |
c2dea2d1 VT |
1752 | case IOPRIO_CLASS_RT: |
1753 | return &cfqd->async_cfqq[0][ioprio]; | |
1754 | case IOPRIO_CLASS_BE: | |
1755 | return &cfqd->async_cfqq[1][ioprio]; | |
1756 | case IOPRIO_CLASS_IDLE: | |
1757 | return &cfqd->async_idle_cfqq; | |
1758 | default: | |
1759 | BUG(); | |
1760 | } | |
1761 | } | |
1762 | ||
15c31be4 | 1763 | static struct cfq_queue * |
a6151c3a | 1764 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, |
15c31be4 JA |
1765 | gfp_t gfp_mask) |
1766 | { | |
fd0928df JA |
1767 | const int ioprio = task_ioprio(ioc); |
1768 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 1769 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
1770 | struct cfq_queue *cfqq = NULL; |
1771 | ||
c2dea2d1 VT |
1772 | if (!is_sync) { |
1773 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
1774 | cfqq = *async_cfqq; | |
1775 | } | |
1776 | ||
6118b70b | 1777 | if (!cfqq) |
fd0928df | 1778 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
1779 | |
1780 | /* | |
1781 | * pin the queue now that it's allocated, scheduler exit will prune it | |
1782 | */ | |
c2dea2d1 | 1783 | if (!is_sync && !(*async_cfqq)) { |
15c31be4 | 1784 | atomic_inc(&cfqq->ref); |
c2dea2d1 | 1785 | *async_cfqq = cfqq; |
15c31be4 JA |
1786 | } |
1787 | ||
1788 | atomic_inc(&cfqq->ref); | |
1789 | return cfqq; | |
1790 | } | |
1791 | ||
498d3aa2 JA |
1792 | /* |
1793 | * We drop cfq io contexts lazily, so we may find a dead one. | |
1794 | */ | |
dbecf3ab | 1795 | static void |
4ac845a2 JA |
1796 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
1797 | struct cfq_io_context *cic) | |
dbecf3ab | 1798 | { |
4ac845a2 JA |
1799 | unsigned long flags; |
1800 | ||
fc46379d | 1801 | WARN_ON(!list_empty(&cic->queue_list)); |
597bc485 | 1802 | |
4ac845a2 JA |
1803 | spin_lock_irqsave(&ioc->lock, flags); |
1804 | ||
4faa3c81 | 1805 | BUG_ON(ioc->ioc_data == cic); |
597bc485 | 1806 | |
4ac845a2 | 1807 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); |
ffc4e759 | 1808 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
1809 | spin_unlock_irqrestore(&ioc->lock, flags); |
1810 | ||
1811 | cfq_cic_free(cic); | |
dbecf3ab OH |
1812 | } |
1813 | ||
e2d74ac0 | 1814 | static struct cfq_io_context * |
4ac845a2 | 1815 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 1816 | { |
e2d74ac0 | 1817 | struct cfq_io_context *cic; |
d6de8be7 | 1818 | unsigned long flags; |
4ac845a2 | 1819 | void *k; |
e2d74ac0 | 1820 | |
91fac317 VT |
1821 | if (unlikely(!ioc)) |
1822 | return NULL; | |
1823 | ||
d6de8be7 JA |
1824 | rcu_read_lock(); |
1825 | ||
597bc485 JA |
1826 | /* |
1827 | * we maintain a last-hit cache, to avoid browsing over the tree | |
1828 | */ | |
4ac845a2 | 1829 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
1830 | if (cic && cic->key == cfqd) { |
1831 | rcu_read_unlock(); | |
597bc485 | 1832 | return cic; |
d6de8be7 | 1833 | } |
597bc485 | 1834 | |
4ac845a2 | 1835 | do { |
4ac845a2 JA |
1836 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); |
1837 | rcu_read_unlock(); | |
1838 | if (!cic) | |
1839 | break; | |
be3b0753 OH |
1840 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
1841 | k = cic->key; | |
1842 | if (unlikely(!k)) { | |
4ac845a2 | 1843 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 1844 | rcu_read_lock(); |
4ac845a2 | 1845 | continue; |
dbecf3ab | 1846 | } |
e2d74ac0 | 1847 | |
d6de8be7 | 1848 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 1849 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 1850 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
1851 | break; |
1852 | } while (1); | |
e2d74ac0 | 1853 | |
4ac845a2 | 1854 | return cic; |
e2d74ac0 JA |
1855 | } |
1856 | ||
4ac845a2 JA |
1857 | /* |
1858 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
1859 | * the process specific cfq io context when entered from the block layer. | |
1860 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
1861 | */ | |
febffd61 JA |
1862 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
1863 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 1864 | { |
0261d688 | 1865 | unsigned long flags; |
4ac845a2 | 1866 | int ret; |
e2d74ac0 | 1867 | |
4ac845a2 JA |
1868 | ret = radix_tree_preload(gfp_mask); |
1869 | if (!ret) { | |
1870 | cic->ioc = ioc; | |
1871 | cic->key = cfqd; | |
e2d74ac0 | 1872 | |
4ac845a2 JA |
1873 | spin_lock_irqsave(&ioc->lock, flags); |
1874 | ret = radix_tree_insert(&ioc->radix_root, | |
1875 | (unsigned long) cfqd, cic); | |
ffc4e759 JA |
1876 | if (!ret) |
1877 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 1878 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 1879 | |
4ac845a2 JA |
1880 | radix_tree_preload_end(); |
1881 | ||
1882 | if (!ret) { | |
1883 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
1884 | list_add(&cic->queue_list, &cfqd->cic_list); | |
1885 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
1886 | } | |
e2d74ac0 JA |
1887 | } |
1888 | ||
4ac845a2 JA |
1889 | if (ret) |
1890 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 1891 | |
4ac845a2 | 1892 | return ret; |
e2d74ac0 JA |
1893 | } |
1894 | ||
1da177e4 LT |
1895 | /* |
1896 | * Setup general io context and cfq io context. There can be several cfq | |
1897 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 1898 | * than one device managed by cfq. |
1da177e4 LT |
1899 | */ |
1900 | static struct cfq_io_context * | |
e2d74ac0 | 1901 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1902 | { |
22e2c507 | 1903 | struct io_context *ioc = NULL; |
1da177e4 | 1904 | struct cfq_io_context *cic; |
1da177e4 | 1905 | |
22e2c507 | 1906 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 1907 | |
b5deef90 | 1908 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
1909 | if (!ioc) |
1910 | return NULL; | |
1911 | ||
4ac845a2 | 1912 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
1913 | if (cic) |
1914 | goto out; | |
1da177e4 | 1915 | |
e2d74ac0 JA |
1916 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
1917 | if (cic == NULL) | |
1918 | goto err; | |
1da177e4 | 1919 | |
4ac845a2 JA |
1920 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
1921 | goto err_free; | |
1922 | ||
1da177e4 | 1923 | out: |
fc46379d JA |
1924 | smp_read_barrier_depends(); |
1925 | if (unlikely(ioc->ioprio_changed)) | |
1926 | cfq_ioc_set_ioprio(ioc); | |
1927 | ||
1da177e4 | 1928 | return cic; |
4ac845a2 JA |
1929 | err_free: |
1930 | cfq_cic_free(cic); | |
1da177e4 LT |
1931 | err: |
1932 | put_io_context(ioc); | |
1933 | return NULL; | |
1934 | } | |
1935 | ||
22e2c507 JA |
1936 | static void |
1937 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 1938 | { |
aaf1228d JA |
1939 | unsigned long elapsed = jiffies - cic->last_end_request; |
1940 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 1941 | |
22e2c507 JA |
1942 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
1943 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
1944 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
1945 | } | |
1da177e4 | 1946 | |
206dc69b | 1947 | static void |
b2c18e1e | 1948 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
6d048f53 | 1949 | struct request *rq) |
206dc69b JA |
1950 | { |
1951 | sector_t sdist; | |
1952 | u64 total; | |
1953 | ||
b2c18e1e | 1954 | if (!cfqq->last_request_pos) |
4d00aa47 | 1955 | sdist = 0; |
b2c18e1e JM |
1956 | else if (cfqq->last_request_pos < blk_rq_pos(rq)) |
1957 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
206dc69b | 1958 | else |
b2c18e1e | 1959 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); |
206dc69b JA |
1960 | |
1961 | /* | |
1962 | * Don't allow the seek distance to get too large from the | |
1963 | * odd fragment, pagein, etc | |
1964 | */ | |
b2c18e1e JM |
1965 | if (cfqq->seek_samples <= 60) /* second&third seek */ |
1966 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024); | |
206dc69b | 1967 | else |
b2c18e1e | 1968 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64); |
206dc69b | 1969 | |
b2c18e1e JM |
1970 | cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8; |
1971 | cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8; | |
1972 | total = cfqq->seek_total + (cfqq->seek_samples/2); | |
1973 | do_div(total, cfqq->seek_samples); | |
1974 | cfqq->seek_mean = (sector_t)total; | |
206dc69b | 1975 | } |
1da177e4 | 1976 | |
22e2c507 JA |
1977 | /* |
1978 | * Disable idle window if the process thinks too long or seeks so much that | |
1979 | * it doesn't matter | |
1980 | */ | |
1981 | static void | |
1982 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1983 | struct cfq_io_context *cic) | |
1984 | { | |
7b679138 | 1985 | int old_idle, enable_idle; |
1be92f2f | 1986 | |
0871714e JA |
1987 | /* |
1988 | * Don't idle for async or idle io prio class | |
1989 | */ | |
1990 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
1991 | return; |
1992 | ||
c265a7f4 | 1993 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 1994 | |
66dac98e | 1995 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || |
b2c18e1e | 1996 | (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq))) |
22e2c507 JA |
1997 | enable_idle = 0; |
1998 | else if (sample_valid(cic->ttime_samples)) { | |
ec60e4f6 | 1999 | unsigned int slice_idle = cfqd->cfq_slice_idle; |
b2c18e1e | 2000 | if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)) |
ec60e4f6 CZ |
2001 | slice_idle = msecs_to_jiffies(CFQ_MIN_TT); |
2002 | if (cic->ttime_mean > slice_idle) | |
22e2c507 JA |
2003 | enable_idle = 0; |
2004 | else | |
2005 | enable_idle = 1; | |
1da177e4 LT |
2006 | } |
2007 | ||
7b679138 JA |
2008 | if (old_idle != enable_idle) { |
2009 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
2010 | if (enable_idle) | |
2011 | cfq_mark_cfqq_idle_window(cfqq); | |
2012 | else | |
2013 | cfq_clear_cfqq_idle_window(cfqq); | |
2014 | } | |
22e2c507 | 2015 | } |
1da177e4 | 2016 | |
22e2c507 JA |
2017 | /* |
2018 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
2019 | * no or if we aren't sure, a 1 will cause a preempt. | |
2020 | */ | |
a6151c3a | 2021 | static bool |
22e2c507 | 2022 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, |
5e705374 | 2023 | struct request *rq) |
22e2c507 | 2024 | { |
6d048f53 | 2025 | struct cfq_queue *cfqq; |
22e2c507 | 2026 | |
6d048f53 JA |
2027 | cfqq = cfqd->active_queue; |
2028 | if (!cfqq) | |
a6151c3a | 2029 | return false; |
22e2c507 | 2030 | |
6d048f53 | 2031 | if (cfq_slice_used(cfqq)) |
a6151c3a | 2032 | return true; |
6d048f53 JA |
2033 | |
2034 | if (cfq_class_idle(new_cfqq)) | |
a6151c3a | 2035 | return false; |
22e2c507 JA |
2036 | |
2037 | if (cfq_class_idle(cfqq)) | |
a6151c3a | 2038 | return true; |
1e3335de | 2039 | |
374f84ac JA |
2040 | /* |
2041 | * if the new request is sync, but the currently running queue is | |
2042 | * not, let the sync request have priority. | |
2043 | */ | |
5e705374 | 2044 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
a6151c3a | 2045 | return true; |
1e3335de | 2046 | |
374f84ac JA |
2047 | /* |
2048 | * So both queues are sync. Let the new request get disk time if | |
2049 | * it's a metadata request and the current queue is doing regular IO. | |
2050 | */ | |
2051 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
a6151c3a | 2052 | return false; |
22e2c507 | 2053 | |
3a9a3f6c DS |
2054 | /* |
2055 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
2056 | */ | |
2057 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
a6151c3a | 2058 | return true; |
3a9a3f6c | 2059 | |
1e3335de | 2060 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
a6151c3a | 2061 | return false; |
1e3335de JA |
2062 | |
2063 | /* | |
2064 | * if this request is as-good as one we would expect from the | |
2065 | * current cfqq, let it preempt | |
2066 | */ | |
b2c18e1e | 2067 | if (cfq_rq_close(cfqd, cfqq, rq)) |
a6151c3a | 2068 | return true; |
1e3335de | 2069 | |
a6151c3a | 2070 | return false; |
22e2c507 JA |
2071 | } |
2072 | ||
2073 | /* | |
2074 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
2075 | * let it have half of its nominal slice. | |
2076 | */ | |
2077 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2078 | { | |
7b679138 | 2079 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
6084cdda | 2080 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 2081 | |
bf572256 JA |
2082 | /* |
2083 | * Put the new queue at the front of the of the current list, | |
2084 | * so we know that it will be selected next. | |
2085 | */ | |
2086 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
2087 | |
2088 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 2089 | |
44f7c160 JA |
2090 | cfqq->slice_end = 0; |
2091 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
2092 | } |
2093 | ||
22e2c507 | 2094 | /* |
5e705374 | 2095 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
2096 | * something we should do about it |
2097 | */ | |
2098 | static void | |
5e705374 JA |
2099 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
2100 | struct request *rq) | |
22e2c507 | 2101 | { |
5e705374 | 2102 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 2103 | |
45333d5a | 2104 | cfqd->rq_queued++; |
374f84ac JA |
2105 | if (rq_is_meta(rq)) |
2106 | cfqq->meta_pending++; | |
2107 | ||
9c2c38a1 | 2108 | cfq_update_io_thinktime(cfqd, cic); |
b2c18e1e | 2109 | cfq_update_io_seektime(cfqd, cfqq, rq); |
9c2c38a1 JA |
2110 | cfq_update_idle_window(cfqd, cfqq, cic); |
2111 | ||
b2c18e1e | 2112 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
2113 | |
2114 | if (cfqq == cfqd->active_queue) { | |
2115 | /* | |
b029195d JA |
2116 | * Remember that we saw a request from this process, but |
2117 | * don't start queuing just yet. Otherwise we risk seeing lots | |
2118 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
2119 | * and merging. If the request is already larger than a single |
2120 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
2121 | * merging is already done. Ditto for a busy system that |
2122 | * has other work pending, don't risk delaying until the | |
2123 | * idle timer unplug to continue working. | |
22e2c507 | 2124 | */ |
d6ceb25e | 2125 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
2126 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
2127 | cfqd->busy_queues > 1) { | |
d6ceb25e | 2128 | del_timer(&cfqd->idle_slice_timer); |
a7f55792 | 2129 | __blk_run_queue(cfqd->queue); |
d6ceb25e | 2130 | } |
b029195d | 2131 | cfq_mark_cfqq_must_dispatch(cfqq); |
d6ceb25e | 2132 | } |
5e705374 | 2133 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
2134 | /* |
2135 | * not the active queue - expire current slice if it is | |
2136 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
2137 | * has some old slice time left and is of higher priority or |
2138 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
2139 | */ |
2140 | cfq_preempt_queue(cfqd, cfqq); | |
a7f55792 | 2141 | __blk_run_queue(cfqd->queue); |
22e2c507 | 2142 | } |
1da177e4 LT |
2143 | } |
2144 | ||
165125e1 | 2145 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2146 | { |
b4878f24 | 2147 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 2148 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 2149 | |
7b679138 | 2150 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 2151 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 2152 | |
5e705374 | 2153 | cfq_add_rq_rb(rq); |
1da177e4 | 2154 | |
30996f40 | 2155 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); |
22e2c507 JA |
2156 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
2157 | ||
5e705374 | 2158 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
2159 | } |
2160 | ||
45333d5a AC |
2161 | /* |
2162 | * Update hw_tag based on peak queue depth over 50 samples under | |
2163 | * sufficient load. | |
2164 | */ | |
2165 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
2166 | { | |
5ad531db JA |
2167 | if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak) |
2168 | cfqd->rq_in_driver_peak = rq_in_driver(cfqd); | |
45333d5a AC |
2169 | |
2170 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
5ad531db | 2171 | rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
2172 | return; |
2173 | ||
2174 | if (cfqd->hw_tag_samples++ < 50) | |
2175 | return; | |
2176 | ||
2177 | if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN) | |
2178 | cfqd->hw_tag = 1; | |
2179 | else | |
2180 | cfqd->hw_tag = 0; | |
2181 | ||
2182 | cfqd->hw_tag_samples = 0; | |
2183 | cfqd->rq_in_driver_peak = 0; | |
2184 | } | |
2185 | ||
165125e1 | 2186 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2187 | { |
5e705374 | 2188 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 2189 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 2190 | const int sync = rq_is_sync(rq); |
b4878f24 | 2191 | unsigned long now; |
1da177e4 | 2192 | |
b4878f24 | 2193 | now = jiffies; |
7b679138 | 2194 | cfq_log_cfqq(cfqd, cfqq, "complete"); |
1da177e4 | 2195 | |
45333d5a AC |
2196 | cfq_update_hw_tag(cfqd); |
2197 | ||
5ad531db | 2198 | WARN_ON(!cfqd->rq_in_driver[sync]); |
6d048f53 | 2199 | WARN_ON(!cfqq->dispatched); |
5ad531db | 2200 | cfqd->rq_in_driver[sync]--; |
6d048f53 | 2201 | cfqq->dispatched--; |
1da177e4 | 2202 | |
3ed9a296 JA |
2203 | if (cfq_cfqq_sync(cfqq)) |
2204 | cfqd->sync_flight--; | |
2205 | ||
365722bb | 2206 | if (sync) { |
5e705374 | 2207 | RQ_CIC(rq)->last_end_request = now; |
365722bb VG |
2208 | cfqd->last_end_sync_rq = now; |
2209 | } | |
caaa5f9f JA |
2210 | |
2211 | /* | |
2212 | * If this is the active queue, check if it needs to be expired, | |
2213 | * or if we want to idle in case it has no pending requests. | |
2214 | */ | |
2215 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
2216 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
2217 | ||
44f7c160 JA |
2218 | if (cfq_cfqq_slice_new(cfqq)) { |
2219 | cfq_set_prio_slice(cfqd, cfqq); | |
2220 | cfq_clear_cfqq_slice_new(cfqq); | |
2221 | } | |
a36e71f9 JA |
2222 | /* |
2223 | * If there are no requests waiting in this queue, and | |
2224 | * there are other queues ready to issue requests, AND | |
2225 | * those other queues are issuing requests within our | |
2226 | * mean seek distance, give them a chance to run instead | |
2227 | * of idling. | |
2228 | */ | |
0871714e | 2229 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
6084cdda | 2230 | cfq_slice_expired(cfqd, 1); |
a36e71f9 JA |
2231 | else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) && |
2232 | sync && !rq_noidle(rq)) | |
6d048f53 | 2233 | cfq_arm_slice_timer(cfqd); |
caaa5f9f | 2234 | } |
6d048f53 | 2235 | |
5ad531db | 2236 | if (!rq_in_driver(cfqd)) |
23e018a1 | 2237 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
2238 | } |
2239 | ||
22e2c507 JA |
2240 | /* |
2241 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
2242 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
2243 | */ | |
2244 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 2245 | { |
22e2c507 JA |
2246 | if (has_fs_excl()) { |
2247 | /* | |
2248 | * boost idle prio on transactions that would lock out other | |
2249 | * users of the filesystem | |
2250 | */ | |
2251 | if (cfq_class_idle(cfqq)) | |
2252 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2253 | if (cfqq->ioprio > IOPRIO_NORM) | |
2254 | cfqq->ioprio = IOPRIO_NORM; | |
2255 | } else { | |
2256 | /* | |
2257 | * check if we need to unboost the queue | |
2258 | */ | |
2259 | if (cfqq->ioprio_class != cfqq->org_ioprio_class) | |
2260 | cfqq->ioprio_class = cfqq->org_ioprio_class; | |
2261 | if (cfqq->ioprio != cfqq->org_ioprio) | |
2262 | cfqq->ioprio = cfqq->org_ioprio; | |
2263 | } | |
22e2c507 | 2264 | } |
1da177e4 | 2265 | |
89850f7e | 2266 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 2267 | { |
1b379d8d | 2268 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 2269 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 2270 | return ELV_MQUEUE_MUST; |
3b18152c | 2271 | } |
1da177e4 | 2272 | |
22e2c507 | 2273 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
2274 | } |
2275 | ||
165125e1 | 2276 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
2277 | { |
2278 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2279 | struct task_struct *tsk = current; | |
91fac317 | 2280 | struct cfq_io_context *cic; |
22e2c507 JA |
2281 | struct cfq_queue *cfqq; |
2282 | ||
2283 | /* | |
2284 | * don't force setup of a queue from here, as a call to may_queue | |
2285 | * does not necessarily imply that a request actually will be queued. | |
2286 | * so just lookup a possibly existing queue, or return 'may queue' | |
2287 | * if that fails | |
2288 | */ | |
4ac845a2 | 2289 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
2290 | if (!cic) |
2291 | return ELV_MQUEUE_MAY; | |
2292 | ||
b0b78f81 | 2293 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 2294 | if (cfqq) { |
fd0928df | 2295 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
2296 | cfq_prio_boost(cfqq); |
2297 | ||
89850f7e | 2298 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
2299 | } |
2300 | ||
2301 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
2302 | } |
2303 | ||
1da177e4 LT |
2304 | /* |
2305 | * queue lock held here | |
2306 | */ | |
bb37b94c | 2307 | static void cfq_put_request(struct request *rq) |
1da177e4 | 2308 | { |
5e705374 | 2309 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 2310 | |
5e705374 | 2311 | if (cfqq) { |
22e2c507 | 2312 | const int rw = rq_data_dir(rq); |
1da177e4 | 2313 | |
22e2c507 JA |
2314 | BUG_ON(!cfqq->allocated[rw]); |
2315 | cfqq->allocated[rw]--; | |
1da177e4 | 2316 | |
5e705374 | 2317 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 2318 | |
1da177e4 | 2319 | rq->elevator_private = NULL; |
5e705374 | 2320 | rq->elevator_private2 = NULL; |
1da177e4 | 2321 | |
1da177e4 LT |
2322 | cfq_put_queue(cfqq); |
2323 | } | |
2324 | } | |
2325 | ||
2326 | /* | |
22e2c507 | 2327 | * Allocate cfq data structures associated with this request. |
1da177e4 | 2328 | */ |
22e2c507 | 2329 | static int |
165125e1 | 2330 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
2331 | { |
2332 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2333 | struct cfq_io_context *cic; | |
2334 | const int rw = rq_data_dir(rq); | |
a6151c3a | 2335 | const bool is_sync = rq_is_sync(rq); |
22e2c507 | 2336 | struct cfq_queue *cfqq; |
1da177e4 LT |
2337 | unsigned long flags; |
2338 | ||
2339 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
2340 | ||
e2d74ac0 | 2341 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 2342 | |
1da177e4 LT |
2343 | spin_lock_irqsave(q->queue_lock, flags); |
2344 | ||
22e2c507 JA |
2345 | if (!cic) |
2346 | goto queue_fail; | |
2347 | ||
91fac317 | 2348 | cfqq = cic_to_cfqq(cic, is_sync); |
32f2e807 | 2349 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { |
fd0928df | 2350 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 VT |
2351 | cic_set_cfqq(cic, cfqq, is_sync); |
2352 | } | |
1da177e4 LT |
2353 | |
2354 | cfqq->allocated[rw]++; | |
22e2c507 | 2355 | atomic_inc(&cfqq->ref); |
1da177e4 | 2356 | |
5e705374 | 2357 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 2358 | |
5e705374 JA |
2359 | rq->elevator_private = cic; |
2360 | rq->elevator_private2 = cfqq; | |
2361 | return 0; | |
1da177e4 | 2362 | |
22e2c507 JA |
2363 | queue_fail: |
2364 | if (cic) | |
2365 | put_io_context(cic->ioc); | |
89850f7e | 2366 | |
23e018a1 | 2367 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 2368 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 2369 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
2370 | return 1; |
2371 | } | |
2372 | ||
65f27f38 | 2373 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 2374 | { |
65f27f38 | 2375 | struct cfq_data *cfqd = |
23e018a1 | 2376 | container_of(work, struct cfq_data, unplug_work); |
165125e1 | 2377 | struct request_queue *q = cfqd->queue; |
22e2c507 | 2378 | |
40bb54d1 | 2379 | spin_lock_irq(q->queue_lock); |
a7f55792 | 2380 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 2381 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
2382 | } |
2383 | ||
2384 | /* | |
2385 | * Timer running if the active_queue is currently idling inside its time slice | |
2386 | */ | |
2387 | static void cfq_idle_slice_timer(unsigned long data) | |
2388 | { | |
2389 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
2390 | struct cfq_queue *cfqq; | |
2391 | unsigned long flags; | |
3c6bd2f8 | 2392 | int timed_out = 1; |
22e2c507 | 2393 | |
7b679138 JA |
2394 | cfq_log(cfqd, "idle timer fired"); |
2395 | ||
22e2c507 JA |
2396 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
2397 | ||
fe094d98 JA |
2398 | cfqq = cfqd->active_queue; |
2399 | if (cfqq) { | |
3c6bd2f8 JA |
2400 | timed_out = 0; |
2401 | ||
b029195d JA |
2402 | /* |
2403 | * We saw a request before the queue expired, let it through | |
2404 | */ | |
2405 | if (cfq_cfqq_must_dispatch(cfqq)) | |
2406 | goto out_kick; | |
2407 | ||
22e2c507 JA |
2408 | /* |
2409 | * expired | |
2410 | */ | |
44f7c160 | 2411 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
2412 | goto expire; |
2413 | ||
2414 | /* | |
2415 | * only expire and reinvoke request handler, if there are | |
2416 | * other queues with pending requests | |
2417 | */ | |
caaa5f9f | 2418 | if (!cfqd->busy_queues) |
22e2c507 | 2419 | goto out_cont; |
22e2c507 JA |
2420 | |
2421 | /* | |
2422 | * not expired and it has a request pending, let it dispatch | |
2423 | */ | |
75e50984 | 2424 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 2425 | goto out_kick; |
22e2c507 JA |
2426 | } |
2427 | expire: | |
6084cdda | 2428 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 2429 | out_kick: |
23e018a1 | 2430 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
2431 | out_cont: |
2432 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2433 | } | |
2434 | ||
3b18152c JA |
2435 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
2436 | { | |
2437 | del_timer_sync(&cfqd->idle_slice_timer); | |
23e018a1 | 2438 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 2439 | } |
22e2c507 | 2440 | |
c2dea2d1 VT |
2441 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
2442 | { | |
2443 | int i; | |
2444 | ||
2445 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
2446 | if (cfqd->async_cfqq[0][i]) | |
2447 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
2448 | if (cfqd->async_cfqq[1][i]) | |
2449 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 2450 | } |
2389d1ef ON |
2451 | |
2452 | if (cfqd->async_idle_cfqq) | |
2453 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
2454 | } |
2455 | ||
b374d18a | 2456 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 2457 | { |
22e2c507 | 2458 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 2459 | struct request_queue *q = cfqd->queue; |
22e2c507 | 2460 | |
3b18152c | 2461 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 2462 | |
d9ff4187 | 2463 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 2464 | |
d9ff4187 | 2465 | if (cfqd->active_queue) |
6084cdda | 2466 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
2467 | |
2468 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
2469 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
2470 | struct cfq_io_context, | |
2471 | queue_list); | |
89850f7e JA |
2472 | |
2473 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 2474 | } |
e2d74ac0 | 2475 | |
c2dea2d1 | 2476 | cfq_put_async_queues(cfqd); |
15c31be4 | 2477 | |
d9ff4187 | 2478 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
2479 | |
2480 | cfq_shutdown_timer_wq(cfqd); | |
2481 | ||
a90d742e | 2482 | kfree(cfqd); |
1da177e4 LT |
2483 | } |
2484 | ||
165125e1 | 2485 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
2486 | { |
2487 | struct cfq_data *cfqd; | |
26a2ac00 | 2488 | int i; |
1da177e4 | 2489 | |
94f6030c | 2490 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 2491 | if (!cfqd) |
bc1c1169 | 2492 | return NULL; |
1da177e4 | 2493 | |
cc09e299 | 2494 | cfqd->service_tree = CFQ_RB_ROOT; |
26a2ac00 JA |
2495 | |
2496 | /* | |
2497 | * Not strictly needed (since RB_ROOT just clears the node and we | |
2498 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
2499 | * to add magic to the rb code | |
2500 | */ | |
2501 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
2502 | cfqd->prio_trees[i] = RB_ROOT; | |
2503 | ||
6118b70b JA |
2504 | /* |
2505 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
2506 | * Grab a permanent reference to it, so that the normal code flow | |
2507 | * will not attempt to free it. | |
2508 | */ | |
2509 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
2510 | atomic_inc(&cfqd->oom_cfqq.ref); | |
2511 | ||
d9ff4187 | 2512 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 2513 | |
1da177e4 | 2514 | cfqd->queue = q; |
1da177e4 | 2515 | |
22e2c507 JA |
2516 | init_timer(&cfqd->idle_slice_timer); |
2517 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
2518 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
2519 | ||
23e018a1 | 2520 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 2521 | |
1da177e4 | 2522 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
2523 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
2524 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
2525 | cfqd->cfq_back_max = cfq_back_max; |
2526 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
2527 | cfqd->cfq_slice[0] = cfq_slice_async; |
2528 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
2529 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
2530 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
963b72fc | 2531 | cfqd->cfq_latency = 1; |
45333d5a | 2532 | cfqd->hw_tag = 1; |
365722bb | 2533 | cfqd->last_end_sync_rq = jiffies; |
bc1c1169 | 2534 | return cfqd; |
1da177e4 LT |
2535 | } |
2536 | ||
2537 | static void cfq_slab_kill(void) | |
2538 | { | |
d6de8be7 JA |
2539 | /* |
2540 | * Caller already ensured that pending RCU callbacks are completed, | |
2541 | * so we should have no busy allocations at this point. | |
2542 | */ | |
1da177e4 LT |
2543 | if (cfq_pool) |
2544 | kmem_cache_destroy(cfq_pool); | |
2545 | if (cfq_ioc_pool) | |
2546 | kmem_cache_destroy(cfq_ioc_pool); | |
2547 | } | |
2548 | ||
2549 | static int __init cfq_slab_setup(void) | |
2550 | { | |
0a31bd5f | 2551 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
2552 | if (!cfq_pool) |
2553 | goto fail; | |
2554 | ||
34e6bbf2 | 2555 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
2556 | if (!cfq_ioc_pool) |
2557 | goto fail; | |
2558 | ||
2559 | return 0; | |
2560 | fail: | |
2561 | cfq_slab_kill(); | |
2562 | return -ENOMEM; | |
2563 | } | |
2564 | ||
1da177e4 LT |
2565 | /* |
2566 | * sysfs parts below --> | |
2567 | */ | |
1da177e4 LT |
2568 | static ssize_t |
2569 | cfq_var_show(unsigned int var, char *page) | |
2570 | { | |
2571 | return sprintf(page, "%d\n", var); | |
2572 | } | |
2573 | ||
2574 | static ssize_t | |
2575 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
2576 | { | |
2577 | char *p = (char *) page; | |
2578 | ||
2579 | *var = simple_strtoul(p, &p, 10); | |
2580 | return count; | |
2581 | } | |
2582 | ||
1da177e4 | 2583 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 2584 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 2585 | { \ |
3d1ab40f | 2586 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2587 | unsigned int __data = __VAR; \ |
2588 | if (__CONV) \ | |
2589 | __data = jiffies_to_msecs(__data); \ | |
2590 | return cfq_var_show(__data, (page)); \ | |
2591 | } | |
2592 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
2593 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
2594 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
2595 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
2596 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
2597 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
2598 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
2599 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
2600 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
963b72fc | 2601 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); |
1da177e4 LT |
2602 | #undef SHOW_FUNCTION |
2603 | ||
2604 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 2605 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 2606 | { \ |
3d1ab40f | 2607 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2608 | unsigned int __data; \ |
2609 | int ret = cfq_var_store(&__data, (page), count); \ | |
2610 | if (__data < (MIN)) \ | |
2611 | __data = (MIN); \ | |
2612 | else if (__data > (MAX)) \ | |
2613 | __data = (MAX); \ | |
2614 | if (__CONV) \ | |
2615 | *(__PTR) = msecs_to_jiffies(__data); \ | |
2616 | else \ | |
2617 | *(__PTR) = __data; \ | |
2618 | return ret; \ | |
2619 | } | |
2620 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
2621 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
2622 | UINT_MAX, 1); | |
2623 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
2624 | UINT_MAX, 1); | |
e572ec7e | 2625 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
2626 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
2627 | UINT_MAX, 0); | |
22e2c507 JA |
2628 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
2629 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
2630 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
2631 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
2632 | UINT_MAX, 0); | |
963b72fc | 2633 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); |
1da177e4 LT |
2634 | #undef STORE_FUNCTION |
2635 | ||
e572ec7e AV |
2636 | #define CFQ_ATTR(name) \ |
2637 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
2638 | ||
2639 | static struct elv_fs_entry cfq_attrs[] = { | |
2640 | CFQ_ATTR(quantum), | |
e572ec7e AV |
2641 | CFQ_ATTR(fifo_expire_sync), |
2642 | CFQ_ATTR(fifo_expire_async), | |
2643 | CFQ_ATTR(back_seek_max), | |
2644 | CFQ_ATTR(back_seek_penalty), | |
2645 | CFQ_ATTR(slice_sync), | |
2646 | CFQ_ATTR(slice_async), | |
2647 | CFQ_ATTR(slice_async_rq), | |
2648 | CFQ_ATTR(slice_idle), | |
963b72fc | 2649 | CFQ_ATTR(low_latency), |
e572ec7e | 2650 | __ATTR_NULL |
1da177e4 LT |
2651 | }; |
2652 | ||
1da177e4 LT |
2653 | static struct elevator_type iosched_cfq = { |
2654 | .ops = { | |
2655 | .elevator_merge_fn = cfq_merge, | |
2656 | .elevator_merged_fn = cfq_merged_request, | |
2657 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 2658 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 2659 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 2660 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 2661 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
2662 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
2663 | .elevator_queue_empty_fn = cfq_queue_empty, | |
2664 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
2665 | .elevator_former_req_fn = elv_rb_former_request, |
2666 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
2667 | .elevator_set_req_fn = cfq_set_request, |
2668 | .elevator_put_req_fn = cfq_put_request, | |
2669 | .elevator_may_queue_fn = cfq_may_queue, | |
2670 | .elevator_init_fn = cfq_init_queue, | |
2671 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 2672 | .trim = cfq_free_io_context, |
1da177e4 | 2673 | }, |
3d1ab40f | 2674 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
2675 | .elevator_name = "cfq", |
2676 | .elevator_owner = THIS_MODULE, | |
2677 | }; | |
2678 | ||
2679 | static int __init cfq_init(void) | |
2680 | { | |
22e2c507 JA |
2681 | /* |
2682 | * could be 0 on HZ < 1000 setups | |
2683 | */ | |
2684 | if (!cfq_slice_async) | |
2685 | cfq_slice_async = 1; | |
2686 | if (!cfq_slice_idle) | |
2687 | cfq_slice_idle = 1; | |
2688 | ||
1da177e4 LT |
2689 | if (cfq_slab_setup()) |
2690 | return -ENOMEM; | |
2691 | ||
2fdd82bd | 2692 | elv_register(&iosched_cfq); |
1da177e4 | 2693 | |
2fdd82bd | 2694 | return 0; |
1da177e4 LT |
2695 | } |
2696 | ||
2697 | static void __exit cfq_exit(void) | |
2698 | { | |
6e9a4738 | 2699 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 2700 | elv_unregister(&iosched_cfq); |
334e94de | 2701 | ioc_gone = &all_gone; |
fba82272 OH |
2702 | /* ioc_gone's update must be visible before reading ioc_count */ |
2703 | smp_wmb(); | |
d6de8be7 JA |
2704 | |
2705 | /* | |
2706 | * this also protects us from entering cfq_slab_kill() with | |
2707 | * pending RCU callbacks | |
2708 | */ | |
245b2e70 | 2709 | if (elv_ioc_count_read(cfq_ioc_count)) |
9a11b4ed | 2710 | wait_for_completion(&all_gone); |
83521d3e | 2711 | cfq_slab_kill(); |
1da177e4 LT |
2712 | } |
2713 | ||
2714 | module_init(cfq_init); | |
2715 | module_exit(cfq_exit); | |
2716 | ||
2717 | MODULE_AUTHOR("Jens Axboe"); | |
2718 | MODULE_LICENSE("GPL"); | |
2719 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |