]>
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> | |
ad5ebd2f | 12 | #include <linux/jiffies.h> |
1da177e4 | 13 | #include <linux/rbtree.h> |
22e2c507 | 14 | #include <linux/ioprio.h> |
7b679138 | 15 | #include <linux/blktrace_api.h> |
1da177e4 LT |
16 | |
17 | /* | |
18 | * tunables | |
19 | */ | |
fe094d98 JA |
20 | /* max queue in one round of service */ |
21 | static const int cfq_quantum = 4; | |
64100099 | 22 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
fe094d98 JA |
23 | /* maximum backwards seek, in KiB */ |
24 | static const int cfq_back_max = 16 * 1024; | |
25 | /* penalty of a backwards seek */ | |
26 | static const int cfq_back_penalty = 2; | |
64100099 | 27 | static const int cfq_slice_sync = HZ / 10; |
3b18152c | 28 | static int cfq_slice_async = HZ / 25; |
64100099 | 29 | static const int cfq_slice_async_rq = 2; |
caaa5f9f | 30 | static int cfq_slice_idle = HZ / 125; |
5db5d642 CZ |
31 | static const int cfq_target_latency = HZ * 3/10; /* 300 ms */ |
32 | static const int cfq_hist_divisor = 4; | |
22e2c507 | 33 | |
d9e7620e | 34 | /* |
0871714e | 35 | * offset from end of service tree |
d9e7620e | 36 | */ |
0871714e | 37 | #define CFQ_IDLE_DELAY (HZ / 5) |
d9e7620e JA |
38 | |
39 | /* | |
40 | * below this threshold, we consider thinktime immediate | |
41 | */ | |
42 | #define CFQ_MIN_TT (2) | |
43 | ||
e6c5bc73 JM |
44 | /* |
45 | * Allow merged cfqqs to perform this amount of seeky I/O before | |
46 | * deciding to break the queues up again. | |
47 | */ | |
48 | #define CFQQ_COOP_TOUT (HZ) | |
49 | ||
22e2c507 | 50 | #define CFQ_SLICE_SCALE (5) |
45333d5a | 51 | #define CFQ_HW_QUEUE_MIN (5) |
22e2c507 | 52 | |
fe094d98 JA |
53 | #define RQ_CIC(rq) \ |
54 | ((struct cfq_io_context *) (rq)->elevator_private) | |
7b679138 | 55 | #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2) |
1da177e4 | 56 | |
e18b890b CL |
57 | static struct kmem_cache *cfq_pool; |
58 | static struct kmem_cache *cfq_ioc_pool; | |
1da177e4 | 59 | |
245b2e70 | 60 | static DEFINE_PER_CPU(unsigned long, cfq_ioc_count); |
334e94de | 61 | static struct completion *ioc_gone; |
9a11b4ed | 62 | static DEFINE_SPINLOCK(ioc_gone_lock); |
334e94de | 63 | |
22e2c507 JA |
64 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR |
65 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
22e2c507 JA |
66 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) |
67 | ||
206dc69b | 68 | #define sample_valid(samples) ((samples) > 80) |
1fa8f6d6 | 69 | #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node) |
206dc69b | 70 | |
cc09e299 JA |
71 | /* |
72 | * Most of our rbtree usage is for sorting with min extraction, so | |
73 | * if we cache the leftmost node we don't have to walk down the tree | |
74 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
75 | * move this into the elevator for the rq sorting as well. | |
76 | */ | |
77 | struct cfq_rb_root { | |
78 | struct rb_root rb; | |
79 | struct rb_node *left; | |
aa6f6a3d | 80 | unsigned count; |
1fa8f6d6 | 81 | u64 min_vdisktime; |
cc09e299 | 82 | }; |
1fa8f6d6 | 83 | #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, } |
cc09e299 | 84 | |
6118b70b JA |
85 | /* |
86 | * Per process-grouping structure | |
87 | */ | |
88 | struct cfq_queue { | |
89 | /* reference count */ | |
90 | atomic_t ref; | |
91 | /* various state flags, see below */ | |
92 | unsigned int flags; | |
93 | /* parent cfq_data */ | |
94 | struct cfq_data *cfqd; | |
95 | /* service_tree member */ | |
96 | struct rb_node rb_node; | |
97 | /* service_tree key */ | |
98 | unsigned long rb_key; | |
99 | /* prio tree member */ | |
100 | struct rb_node p_node; | |
101 | /* prio tree root we belong to, if any */ | |
102 | struct rb_root *p_root; | |
103 | /* sorted list of pending requests */ | |
104 | struct rb_root sort_list; | |
105 | /* if fifo isn't expired, next request to serve */ | |
106 | struct request *next_rq; | |
107 | /* requests queued in sort_list */ | |
108 | int queued[2]; | |
109 | /* currently allocated requests */ | |
110 | int allocated[2]; | |
111 | /* fifo list of requests in sort_list */ | |
112 | struct list_head fifo; | |
113 | ||
114 | unsigned long slice_end; | |
115 | long slice_resid; | |
116 | unsigned int slice_dispatch; | |
117 | ||
118 | /* pending metadata requests */ | |
119 | int meta_pending; | |
120 | /* number of requests that are on the dispatch list or inside driver */ | |
121 | int dispatched; | |
122 | ||
123 | /* io prio of this group */ | |
124 | unsigned short ioprio, org_ioprio; | |
125 | unsigned short ioprio_class, org_ioprio_class; | |
126 | ||
b2c18e1e JM |
127 | unsigned int seek_samples; |
128 | u64 seek_total; | |
129 | sector_t seek_mean; | |
130 | sector_t last_request_pos; | |
e6c5bc73 | 131 | unsigned long seeky_start; |
b2c18e1e | 132 | |
6118b70b | 133 | pid_t pid; |
df5fe3e8 | 134 | |
aa6f6a3d | 135 | struct cfq_rb_root *service_tree; |
df5fe3e8 | 136 | struct cfq_queue *new_cfqq; |
cdb16e8f | 137 | struct cfq_group *cfqg; |
6118b70b JA |
138 | }; |
139 | ||
c0324a02 | 140 | /* |
718eee05 | 141 | * First index in the service_trees. |
c0324a02 CZ |
142 | * IDLE is handled separately, so it has negative index |
143 | */ | |
144 | enum wl_prio_t { | |
c0324a02 | 145 | BE_WORKLOAD = 0, |
615f0259 VG |
146 | RT_WORKLOAD = 1, |
147 | IDLE_WORKLOAD = 2, | |
c0324a02 CZ |
148 | }; |
149 | ||
718eee05 CZ |
150 | /* |
151 | * Second index in the service_trees. | |
152 | */ | |
153 | enum wl_type_t { | |
154 | ASYNC_WORKLOAD = 0, | |
155 | SYNC_NOIDLE_WORKLOAD = 1, | |
156 | SYNC_WORKLOAD = 2 | |
157 | }; | |
158 | ||
cdb16e8f VG |
159 | /* This is per cgroup per device grouping structure */ |
160 | struct cfq_group { | |
1fa8f6d6 VG |
161 | /* group service_tree member */ |
162 | struct rb_node rb_node; | |
163 | ||
164 | /* group service_tree key */ | |
165 | u64 vdisktime; | |
166 | bool on_st; | |
167 | ||
168 | /* number of cfqq currently on this group */ | |
169 | int nr_cfqq; | |
170 | ||
cdb16e8f VG |
171 | /* |
172 | * rr lists of queues with requests, onle rr for each priority class. | |
173 | * Counts are embedded in the cfq_rb_root | |
174 | */ | |
175 | struct cfq_rb_root service_trees[2][3]; | |
176 | struct cfq_rb_root service_tree_idle; | |
177 | }; | |
718eee05 | 178 | |
22e2c507 JA |
179 | /* |
180 | * Per block device queue structure | |
181 | */ | |
1da177e4 | 182 | struct cfq_data { |
165125e1 | 183 | struct request_queue *queue; |
1fa8f6d6 VG |
184 | /* Root service tree for cfq_groups */ |
185 | struct cfq_rb_root grp_service_tree; | |
cdb16e8f | 186 | struct cfq_group root_group; |
22e2c507 | 187 | |
c0324a02 CZ |
188 | /* |
189 | * The priority currently being served | |
22e2c507 | 190 | */ |
c0324a02 | 191 | enum wl_prio_t serving_prio; |
718eee05 CZ |
192 | enum wl_type_t serving_type; |
193 | unsigned long workload_expires; | |
cdb16e8f | 194 | struct cfq_group *serving_group; |
8e550632 | 195 | bool noidle_tree_requires_idle; |
a36e71f9 JA |
196 | |
197 | /* | |
198 | * Each priority tree is sorted by next_request position. These | |
199 | * trees are used when determining if two or more queues are | |
200 | * interleaving requests (see cfq_close_cooperator). | |
201 | */ | |
202 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
203 | ||
22e2c507 | 204 | unsigned int busy_queues; |
5db5d642 | 205 | unsigned int busy_queues_avg[2]; |
22e2c507 | 206 | |
5ad531db | 207 | int rq_in_driver[2]; |
3ed9a296 | 208 | int sync_flight; |
45333d5a AC |
209 | |
210 | /* | |
211 | * queue-depth detection | |
212 | */ | |
213 | int rq_queued; | |
25776e35 | 214 | int hw_tag; |
e459dd08 CZ |
215 | /* |
216 | * hw_tag can be | |
217 | * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) | |
218 | * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) | |
219 | * 0 => no NCQ | |
220 | */ | |
221 | int hw_tag_est_depth; | |
222 | unsigned int hw_tag_samples; | |
1da177e4 | 223 | |
22e2c507 JA |
224 | /* |
225 | * idle window management | |
226 | */ | |
227 | struct timer_list idle_slice_timer; | |
23e018a1 | 228 | struct work_struct unplug_work; |
1da177e4 | 229 | |
22e2c507 JA |
230 | struct cfq_queue *active_queue; |
231 | struct cfq_io_context *active_cic; | |
22e2c507 | 232 | |
c2dea2d1 VT |
233 | /* |
234 | * async queue for each priority case | |
235 | */ | |
236 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
237 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 238 | |
6d048f53 | 239 | sector_t last_position; |
1da177e4 | 240 | |
1da177e4 LT |
241 | /* |
242 | * tunables, see top of file | |
243 | */ | |
244 | unsigned int cfq_quantum; | |
22e2c507 | 245 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
246 | unsigned int cfq_back_penalty; |
247 | unsigned int cfq_back_max; | |
22e2c507 JA |
248 | unsigned int cfq_slice[2]; |
249 | unsigned int cfq_slice_async_rq; | |
250 | unsigned int cfq_slice_idle; | |
963b72fc | 251 | unsigned int cfq_latency; |
d9ff4187 AV |
252 | |
253 | struct list_head cic_list; | |
1da177e4 | 254 | |
6118b70b JA |
255 | /* |
256 | * Fallback dummy cfqq for extreme OOM conditions | |
257 | */ | |
258 | struct cfq_queue oom_cfqq; | |
365722bb VG |
259 | |
260 | unsigned long last_end_sync_rq; | |
1da177e4 LT |
261 | }; |
262 | ||
cdb16e8f VG |
263 | static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, |
264 | enum wl_prio_t prio, | |
718eee05 | 265 | enum wl_type_t type, |
c0324a02 CZ |
266 | struct cfq_data *cfqd) |
267 | { | |
1fa8f6d6 VG |
268 | if (!cfqg) |
269 | return NULL; | |
270 | ||
c0324a02 | 271 | if (prio == IDLE_WORKLOAD) |
cdb16e8f | 272 | return &cfqg->service_tree_idle; |
c0324a02 | 273 | |
cdb16e8f | 274 | return &cfqg->service_trees[prio][type]; |
c0324a02 CZ |
275 | } |
276 | ||
3b18152c | 277 | enum cfqq_state_flags { |
b0b8d749 JA |
278 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
279 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 280 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 | 281 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ |
b0b8d749 JA |
282 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
283 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
284 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 285 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 286 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
b3b6d040 | 287 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ |
76280aff | 288 | CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ |
3b18152c JA |
289 | }; |
290 | ||
291 | #define CFQ_CFQQ_FNS(name) \ | |
292 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
293 | { \ | |
fe094d98 | 294 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
295 | } \ |
296 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
297 | { \ | |
fe094d98 | 298 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
299 | } \ |
300 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
301 | { \ | |
fe094d98 | 302 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
303 | } |
304 | ||
305 | CFQ_CFQQ_FNS(on_rr); | |
306 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 307 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c | 308 | CFQ_CFQQ_FNS(must_alloc_slice); |
3b18152c JA |
309 | CFQ_CFQQ_FNS(fifo_expire); |
310 | CFQ_CFQQ_FNS(idle_window); | |
311 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 312 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 313 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 314 | CFQ_CFQQ_FNS(coop); |
76280aff | 315 | CFQ_CFQQ_FNS(deep); |
3b18152c JA |
316 | #undef CFQ_CFQQ_FNS |
317 | ||
7b679138 JA |
318 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
319 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
320 | #define cfq_log(cfqd, fmt, args...) \ | |
321 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
322 | ||
615f0259 VG |
323 | /* Traverses through cfq group service trees */ |
324 | #define for_each_cfqg_st(cfqg, i, j, st) \ | |
325 | for (i = 0; i <= IDLE_WORKLOAD; i++) \ | |
326 | for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ | |
327 | : &cfqg->service_tree_idle; \ | |
328 | (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ | |
329 | (i == IDLE_WORKLOAD && j == 0); \ | |
330 | j++, st = i < IDLE_WORKLOAD ? \ | |
331 | &cfqg->service_trees[i][j]: NULL) \ | |
332 | ||
333 | ||
c0324a02 CZ |
334 | static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) |
335 | { | |
336 | if (cfq_class_idle(cfqq)) | |
337 | return IDLE_WORKLOAD; | |
338 | if (cfq_class_rt(cfqq)) | |
339 | return RT_WORKLOAD; | |
340 | return BE_WORKLOAD; | |
341 | } | |
342 | ||
718eee05 CZ |
343 | |
344 | static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) | |
345 | { | |
346 | if (!cfq_cfqq_sync(cfqq)) | |
347 | return ASYNC_WORKLOAD; | |
348 | if (!cfq_cfqq_idle_window(cfqq)) | |
349 | return SYNC_NOIDLE_WORKLOAD; | |
350 | return SYNC_WORKLOAD; | |
351 | } | |
352 | ||
c0324a02 CZ |
353 | static inline int cfq_busy_queues_wl(enum wl_prio_t wl, struct cfq_data *cfqd) |
354 | { | |
cdb16e8f VG |
355 | struct cfq_group *cfqg = &cfqd->root_group; |
356 | ||
c0324a02 | 357 | if (wl == IDLE_WORKLOAD) |
cdb16e8f | 358 | return cfqg->service_tree_idle.count; |
c0324a02 | 359 | |
cdb16e8f VG |
360 | return cfqg->service_trees[wl][ASYNC_WORKLOAD].count |
361 | + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count | |
362 | + cfqg->service_trees[wl][SYNC_WORKLOAD].count; | |
c0324a02 CZ |
363 | } |
364 | ||
165125e1 | 365 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
a6151c3a | 366 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, |
fd0928df | 367 | struct io_context *, gfp_t); |
4ac845a2 | 368 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
369 | struct io_context *); |
370 | ||
5ad531db JA |
371 | static inline int rq_in_driver(struct cfq_data *cfqd) |
372 | { | |
373 | return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1]; | |
374 | } | |
375 | ||
91fac317 | 376 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, |
a6151c3a | 377 | bool is_sync) |
91fac317 | 378 | { |
a6151c3a | 379 | return cic->cfqq[is_sync]; |
91fac317 VT |
380 | } |
381 | ||
382 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 383 | struct cfq_queue *cfqq, bool is_sync) |
91fac317 | 384 | { |
a6151c3a | 385 | cic->cfqq[is_sync] = cfqq; |
91fac317 VT |
386 | } |
387 | ||
388 | /* | |
389 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
390 | * set (in which case it could also be direct WRITE). | |
391 | */ | |
a6151c3a | 392 | static inline bool cfq_bio_sync(struct bio *bio) |
91fac317 | 393 | { |
a6151c3a | 394 | return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO); |
91fac317 | 395 | } |
1da177e4 | 396 | |
99f95e52 AM |
397 | /* |
398 | * scheduler run of queue, if there are requests pending and no one in the | |
399 | * driver that will restart queueing | |
400 | */ | |
23e018a1 | 401 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) |
99f95e52 | 402 | { |
7b679138 JA |
403 | if (cfqd->busy_queues) { |
404 | cfq_log(cfqd, "schedule dispatch"); | |
23e018a1 | 405 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 406 | } |
99f95e52 AM |
407 | } |
408 | ||
165125e1 | 409 | static int cfq_queue_empty(struct request_queue *q) |
99f95e52 AM |
410 | { |
411 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
412 | ||
f04a6424 | 413 | return !cfqd->rq_queued; |
99f95e52 AM |
414 | } |
415 | ||
44f7c160 JA |
416 | /* |
417 | * Scale schedule slice based on io priority. Use the sync time slice only | |
418 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
419 | * io only, should not get full sync slice length. | |
420 | */ | |
a6151c3a | 421 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, |
d9e7620e | 422 | unsigned short prio) |
44f7c160 | 423 | { |
d9e7620e | 424 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 425 | |
d9e7620e JA |
426 | WARN_ON(prio >= IOPRIO_BE_NR); |
427 | ||
428 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
429 | } | |
44f7c160 | 430 | |
d9e7620e JA |
431 | static inline int |
432 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
433 | { | |
434 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
435 | } |
436 | ||
5db5d642 CZ |
437 | /* |
438 | * get averaged number of queues of RT/BE priority. | |
439 | * average is updated, with a formula that gives more weight to higher numbers, | |
440 | * to quickly follows sudden increases and decrease slowly | |
441 | */ | |
442 | ||
5869619c JA |
443 | static inline unsigned cfq_get_avg_queues(struct cfq_data *cfqd, bool rt) |
444 | { | |
5db5d642 CZ |
445 | unsigned min_q, max_q; |
446 | unsigned mult = cfq_hist_divisor - 1; | |
447 | unsigned round = cfq_hist_divisor / 2; | |
c0324a02 | 448 | unsigned busy = cfq_busy_queues_wl(rt, cfqd); |
5db5d642 CZ |
449 | |
450 | min_q = min(cfqd->busy_queues_avg[rt], busy); | |
451 | max_q = max(cfqd->busy_queues_avg[rt], busy); | |
452 | cfqd->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
453 | cfq_hist_divisor; | |
454 | return cfqd->busy_queues_avg[rt]; | |
455 | } | |
456 | ||
44f7c160 JA |
457 | static inline void |
458 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
459 | { | |
5db5d642 CZ |
460 | unsigned slice = cfq_prio_to_slice(cfqd, cfqq); |
461 | if (cfqd->cfq_latency) { | |
462 | /* interested queues (we consider only the ones with the same | |
463 | * priority class) */ | |
464 | unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq)); | |
465 | unsigned sync_slice = cfqd->cfq_slice[1]; | |
466 | unsigned expect_latency = sync_slice * iq; | |
467 | if (expect_latency > cfq_target_latency) { | |
468 | unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; | |
469 | /* scale low_slice according to IO priority | |
470 | * and sync vs async */ | |
471 | unsigned low_slice = | |
472 | min(slice, base_low_slice * slice / sync_slice); | |
473 | /* the adapted slice value is scaled to fit all iqs | |
474 | * into the target latency */ | |
475 | slice = max(slice * cfq_target_latency / expect_latency, | |
476 | low_slice); | |
477 | } | |
478 | } | |
479 | cfqq->slice_end = jiffies + slice; | |
7b679138 | 480 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
481 | } |
482 | ||
483 | /* | |
484 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
485 | * isn't valid until the first request from the dispatch is activated | |
486 | * and the slice time set. | |
487 | */ | |
a6151c3a | 488 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) |
44f7c160 JA |
489 | { |
490 | if (cfq_cfqq_slice_new(cfqq)) | |
491 | return 0; | |
492 | if (time_before(jiffies, cfqq->slice_end)) | |
493 | return 0; | |
494 | ||
495 | return 1; | |
496 | } | |
497 | ||
1da177e4 | 498 | /* |
5e705374 | 499 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 500 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 501 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 502 | */ |
5e705374 | 503 | static struct request * |
cf7c25cf | 504 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) |
1da177e4 | 505 | { |
cf7c25cf | 506 | sector_t s1, s2, d1 = 0, d2 = 0; |
1da177e4 | 507 | unsigned long back_max; |
e8a99053 AM |
508 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
509 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
510 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 511 | |
5e705374 JA |
512 | if (rq1 == NULL || rq1 == rq2) |
513 | return rq2; | |
514 | if (rq2 == NULL) | |
515 | return rq1; | |
9c2c38a1 | 516 | |
5e705374 JA |
517 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
518 | return rq1; | |
519 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
520 | return rq2; | |
374f84ac JA |
521 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
522 | return rq1; | |
523 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
524 | return rq2; | |
1da177e4 | 525 | |
83096ebf TH |
526 | s1 = blk_rq_pos(rq1); |
527 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 528 | |
1da177e4 LT |
529 | /* |
530 | * by definition, 1KiB is 2 sectors | |
531 | */ | |
532 | back_max = cfqd->cfq_back_max * 2; | |
533 | ||
534 | /* | |
535 | * Strict one way elevator _except_ in the case where we allow | |
536 | * short backward seeks which are biased as twice the cost of a | |
537 | * similar forward seek. | |
538 | */ | |
539 | if (s1 >= last) | |
540 | d1 = s1 - last; | |
541 | else if (s1 + back_max >= last) | |
542 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
543 | else | |
e8a99053 | 544 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
545 | |
546 | if (s2 >= last) | |
547 | d2 = s2 - last; | |
548 | else if (s2 + back_max >= last) | |
549 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
550 | else | |
e8a99053 | 551 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
552 | |
553 | /* Found required data */ | |
e8a99053 AM |
554 | |
555 | /* | |
556 | * By doing switch() on the bit mask "wrap" we avoid having to | |
557 | * check two variables for all permutations: --> faster! | |
558 | */ | |
559 | switch (wrap) { | |
5e705374 | 560 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 561 | if (d1 < d2) |
5e705374 | 562 | return rq1; |
e8a99053 | 563 | else if (d2 < d1) |
5e705374 | 564 | return rq2; |
e8a99053 AM |
565 | else { |
566 | if (s1 >= s2) | |
5e705374 | 567 | return rq1; |
e8a99053 | 568 | else |
5e705374 | 569 | return rq2; |
e8a99053 | 570 | } |
1da177e4 | 571 | |
e8a99053 | 572 | case CFQ_RQ2_WRAP: |
5e705374 | 573 | return rq1; |
e8a99053 | 574 | case CFQ_RQ1_WRAP: |
5e705374 JA |
575 | return rq2; |
576 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
577 | default: |
578 | /* | |
579 | * Since both rqs are wrapped, | |
580 | * start with the one that's further behind head | |
581 | * (--> only *one* back seek required), | |
582 | * since back seek takes more time than forward. | |
583 | */ | |
584 | if (s1 <= s2) | |
5e705374 | 585 | return rq1; |
1da177e4 | 586 | else |
5e705374 | 587 | return rq2; |
1da177e4 LT |
588 | } |
589 | } | |
590 | ||
498d3aa2 JA |
591 | /* |
592 | * The below is leftmost cache rbtree addon | |
593 | */ | |
0871714e | 594 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 | 595 | { |
615f0259 VG |
596 | /* Service tree is empty */ |
597 | if (!root->count) | |
598 | return NULL; | |
599 | ||
cc09e299 JA |
600 | if (!root->left) |
601 | root->left = rb_first(&root->rb); | |
602 | ||
0871714e JA |
603 | if (root->left) |
604 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
605 | ||
606 | return NULL; | |
cc09e299 JA |
607 | } |
608 | ||
1fa8f6d6 VG |
609 | static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) |
610 | { | |
611 | if (!root->left) | |
612 | root->left = rb_first(&root->rb); | |
613 | ||
614 | if (root->left) | |
615 | return rb_entry_cfqg(root->left); | |
616 | ||
617 | return NULL; | |
618 | } | |
619 | ||
a36e71f9 JA |
620 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
621 | { | |
622 | rb_erase(n, root); | |
623 | RB_CLEAR_NODE(n); | |
624 | } | |
625 | ||
cc09e299 JA |
626 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
627 | { | |
628 | if (root->left == n) | |
629 | root->left = NULL; | |
a36e71f9 | 630 | rb_erase_init(n, &root->rb); |
aa6f6a3d | 631 | --root->count; |
cc09e299 JA |
632 | } |
633 | ||
1da177e4 LT |
634 | /* |
635 | * would be nice to take fifo expire time into account as well | |
636 | */ | |
5e705374 JA |
637 | static struct request * |
638 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
639 | struct request *last) | |
1da177e4 | 640 | { |
21183b07 JA |
641 | struct rb_node *rbnext = rb_next(&last->rb_node); |
642 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 643 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 644 | |
21183b07 | 645 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
646 | |
647 | if (rbprev) | |
5e705374 | 648 | prev = rb_entry_rq(rbprev); |
1da177e4 | 649 | |
21183b07 | 650 | if (rbnext) |
5e705374 | 651 | next = rb_entry_rq(rbnext); |
21183b07 JA |
652 | else { |
653 | rbnext = rb_first(&cfqq->sort_list); | |
654 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 655 | next = rb_entry_rq(rbnext); |
21183b07 | 656 | } |
1da177e4 | 657 | |
cf7c25cf | 658 | return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); |
1da177e4 LT |
659 | } |
660 | ||
d9e7620e JA |
661 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
662 | struct cfq_queue *cfqq) | |
1da177e4 | 663 | { |
d9e7620e JA |
664 | /* |
665 | * just an approximation, should be ok. | |
666 | */ | |
cdb16e8f | 667 | return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
464191c6 | 668 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); |
d9e7620e JA |
669 | } |
670 | ||
1fa8f6d6 VG |
671 | static inline s64 |
672 | cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
673 | { | |
674 | return cfqg->vdisktime - st->min_vdisktime; | |
675 | } | |
676 | ||
677 | static void | |
678 | __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
679 | { | |
680 | struct rb_node **node = &st->rb.rb_node; | |
681 | struct rb_node *parent = NULL; | |
682 | struct cfq_group *__cfqg; | |
683 | s64 key = cfqg_key(st, cfqg); | |
684 | int left = 1; | |
685 | ||
686 | while (*node != NULL) { | |
687 | parent = *node; | |
688 | __cfqg = rb_entry_cfqg(parent); | |
689 | ||
690 | if (key < cfqg_key(st, __cfqg)) | |
691 | node = &parent->rb_left; | |
692 | else { | |
693 | node = &parent->rb_right; | |
694 | left = 0; | |
695 | } | |
696 | } | |
697 | ||
698 | if (left) | |
699 | st->left = &cfqg->rb_node; | |
700 | ||
701 | rb_link_node(&cfqg->rb_node, parent, node); | |
702 | rb_insert_color(&cfqg->rb_node, &st->rb); | |
703 | } | |
704 | ||
705 | static void | |
706 | cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
707 | { | |
708 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
709 | struct cfq_group *__cfqg; | |
710 | struct rb_node *n; | |
711 | ||
712 | cfqg->nr_cfqq++; | |
713 | if (cfqg->on_st) | |
714 | return; | |
715 | ||
716 | /* | |
717 | * Currently put the group at the end. Later implement something | |
718 | * so that groups get lesser vtime based on their weights, so that | |
719 | * if group does not loose all if it was not continously backlogged. | |
720 | */ | |
721 | n = rb_last(&st->rb); | |
722 | if (n) { | |
723 | __cfqg = rb_entry_cfqg(n); | |
724 | cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; | |
725 | } else | |
726 | cfqg->vdisktime = st->min_vdisktime; | |
727 | ||
728 | __cfq_group_service_tree_add(st, cfqg); | |
729 | cfqg->on_st = true; | |
730 | } | |
731 | ||
732 | static void | |
733 | cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
734 | { | |
735 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
736 | ||
737 | BUG_ON(cfqg->nr_cfqq < 1); | |
738 | cfqg->nr_cfqq--; | |
739 | /* If there are other cfq queues under this group, don't delete it */ | |
740 | if (cfqg->nr_cfqq) | |
741 | return; | |
742 | ||
743 | cfqg->on_st = false; | |
744 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) | |
745 | cfq_rb_erase(&cfqg->rb_node, st); | |
746 | } | |
747 | ||
498d3aa2 | 748 | /* |
c0324a02 | 749 | * The cfqd->service_trees holds all pending cfq_queue's that have |
498d3aa2 JA |
750 | * requests waiting to be processed. It is sorted in the order that |
751 | * we will service the queues. | |
752 | */ | |
a36e71f9 | 753 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 754 | bool add_front) |
d9e7620e | 755 | { |
0871714e JA |
756 | struct rb_node **p, *parent; |
757 | struct cfq_queue *__cfqq; | |
d9e7620e | 758 | unsigned long rb_key; |
c0324a02 | 759 | struct cfq_rb_root *service_tree; |
498d3aa2 | 760 | int left; |
d9e7620e | 761 | |
cdb16e8f VG |
762 | service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), |
763 | cfqq_type(cfqq), cfqd); | |
0871714e JA |
764 | if (cfq_class_idle(cfqq)) { |
765 | rb_key = CFQ_IDLE_DELAY; | |
aa6f6a3d | 766 | parent = rb_last(&service_tree->rb); |
0871714e JA |
767 | if (parent && parent != &cfqq->rb_node) { |
768 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
769 | rb_key += __cfqq->rb_key; | |
770 | } else | |
771 | rb_key += jiffies; | |
772 | } else if (!add_front) { | |
b9c8946b JA |
773 | /* |
774 | * Get our rb key offset. Subtract any residual slice | |
775 | * value carried from last service. A negative resid | |
776 | * count indicates slice overrun, and this should position | |
777 | * the next service time further away in the tree. | |
778 | */ | |
edd75ffd | 779 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
b9c8946b | 780 | rb_key -= cfqq->slice_resid; |
edd75ffd | 781 | cfqq->slice_resid = 0; |
48e025e6 CZ |
782 | } else { |
783 | rb_key = -HZ; | |
aa6f6a3d | 784 | __cfqq = cfq_rb_first(service_tree); |
48e025e6 CZ |
785 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; |
786 | } | |
1da177e4 | 787 | |
d9e7620e | 788 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
99f9628a | 789 | /* |
d9e7620e | 790 | * same position, nothing more to do |
99f9628a | 791 | */ |
c0324a02 CZ |
792 | if (rb_key == cfqq->rb_key && |
793 | cfqq->service_tree == service_tree) | |
d9e7620e | 794 | return; |
1da177e4 | 795 | |
aa6f6a3d CZ |
796 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); |
797 | cfqq->service_tree = NULL; | |
1da177e4 | 798 | } |
d9e7620e | 799 | |
498d3aa2 | 800 | left = 1; |
0871714e | 801 | parent = NULL; |
aa6f6a3d CZ |
802 | cfqq->service_tree = service_tree; |
803 | p = &service_tree->rb.rb_node; | |
d9e7620e | 804 | while (*p) { |
67060e37 | 805 | struct rb_node **n; |
cc09e299 | 806 | |
d9e7620e JA |
807 | parent = *p; |
808 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
809 | ||
0c534e0a | 810 | /* |
c0324a02 | 811 | * sort by key, that represents service time. |
0c534e0a | 812 | */ |
c0324a02 | 813 | if (time_before(rb_key, __cfqq->rb_key)) |
67060e37 | 814 | n = &(*p)->rb_left; |
c0324a02 | 815 | else { |
67060e37 | 816 | n = &(*p)->rb_right; |
cc09e299 | 817 | left = 0; |
c0324a02 | 818 | } |
67060e37 JA |
819 | |
820 | p = n; | |
d9e7620e JA |
821 | } |
822 | ||
cc09e299 | 823 | if (left) |
aa6f6a3d | 824 | service_tree->left = &cfqq->rb_node; |
cc09e299 | 825 | |
d9e7620e JA |
826 | cfqq->rb_key = rb_key; |
827 | rb_link_node(&cfqq->rb_node, parent, p); | |
aa6f6a3d CZ |
828 | rb_insert_color(&cfqq->rb_node, &service_tree->rb); |
829 | service_tree->count++; | |
1fa8f6d6 | 830 | cfq_group_service_tree_add(cfqd, cfqq->cfqg); |
1da177e4 LT |
831 | } |
832 | ||
a36e71f9 | 833 | static struct cfq_queue * |
f2d1f0ae JA |
834 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
835 | sector_t sector, struct rb_node **ret_parent, | |
836 | struct rb_node ***rb_link) | |
a36e71f9 | 837 | { |
a36e71f9 JA |
838 | struct rb_node **p, *parent; |
839 | struct cfq_queue *cfqq = NULL; | |
840 | ||
841 | parent = NULL; | |
842 | p = &root->rb_node; | |
843 | while (*p) { | |
844 | struct rb_node **n; | |
845 | ||
846 | parent = *p; | |
847 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
848 | ||
849 | /* | |
850 | * Sort strictly based on sector. Smallest to the left, | |
851 | * largest to the right. | |
852 | */ | |
2e46e8b2 | 853 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 854 | n = &(*p)->rb_right; |
2e46e8b2 | 855 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
856 | n = &(*p)->rb_left; |
857 | else | |
858 | break; | |
859 | p = n; | |
3ac6c9f8 | 860 | cfqq = NULL; |
a36e71f9 JA |
861 | } |
862 | ||
863 | *ret_parent = parent; | |
864 | if (rb_link) | |
865 | *rb_link = p; | |
3ac6c9f8 | 866 | return cfqq; |
a36e71f9 JA |
867 | } |
868 | ||
869 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
870 | { | |
a36e71f9 JA |
871 | struct rb_node **p, *parent; |
872 | struct cfq_queue *__cfqq; | |
873 | ||
f2d1f0ae JA |
874 | if (cfqq->p_root) { |
875 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
876 | cfqq->p_root = NULL; | |
877 | } | |
a36e71f9 JA |
878 | |
879 | if (cfq_class_idle(cfqq)) | |
880 | return; | |
881 | if (!cfqq->next_rq) | |
882 | return; | |
883 | ||
f2d1f0ae | 884 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
885 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
886 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
887 | if (!__cfqq) { |
888 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
889 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
890 | } else | |
891 | cfqq->p_root = NULL; | |
a36e71f9 JA |
892 | } |
893 | ||
498d3aa2 JA |
894 | /* |
895 | * Update cfqq's position in the service tree. | |
896 | */ | |
edd75ffd | 897 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 898 | { |
6d048f53 JA |
899 | /* |
900 | * Resorting requires the cfqq to be on the RR list already. | |
901 | */ | |
a36e71f9 | 902 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 903 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
904 | cfq_prio_tree_add(cfqd, cfqq); |
905 | } | |
6d048f53 JA |
906 | } |
907 | ||
1da177e4 LT |
908 | /* |
909 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 910 | * the pending list according to last request service |
1da177e4 | 911 | */ |
febffd61 | 912 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 913 | { |
7b679138 | 914 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
915 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
916 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 LT |
917 | cfqd->busy_queues++; |
918 | ||
edd75ffd | 919 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
920 | } |
921 | ||
498d3aa2 JA |
922 | /* |
923 | * Called when the cfqq no longer has requests pending, remove it from | |
924 | * the service tree. | |
925 | */ | |
febffd61 | 926 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 927 | { |
7b679138 | 928 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
929 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
930 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 931 | |
aa6f6a3d CZ |
932 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
933 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
934 | cfqq->service_tree = NULL; | |
935 | } | |
f2d1f0ae JA |
936 | if (cfqq->p_root) { |
937 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
938 | cfqq->p_root = NULL; | |
939 | } | |
d9e7620e | 940 | |
1fa8f6d6 | 941 | cfq_group_service_tree_del(cfqd, cfqq->cfqg); |
1da177e4 LT |
942 | BUG_ON(!cfqd->busy_queues); |
943 | cfqd->busy_queues--; | |
944 | } | |
945 | ||
946 | /* | |
947 | * rb tree support functions | |
948 | */ | |
febffd61 | 949 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 950 | { |
5e705374 | 951 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
5e705374 | 952 | const int sync = rq_is_sync(rq); |
1da177e4 | 953 | |
b4878f24 JA |
954 | BUG_ON(!cfqq->queued[sync]); |
955 | cfqq->queued[sync]--; | |
1da177e4 | 956 | |
5e705374 | 957 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 958 | |
f04a6424 VG |
959 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { |
960 | /* | |
961 | * Queue will be deleted from service tree when we actually | |
962 | * expire it later. Right now just remove it from prio tree | |
963 | * as it is empty. | |
964 | */ | |
965 | if (cfqq->p_root) { | |
966 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
967 | cfqq->p_root = NULL; | |
968 | } | |
969 | } | |
1da177e4 LT |
970 | } |
971 | ||
5e705374 | 972 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 973 | { |
5e705374 | 974 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 975 | struct cfq_data *cfqd = cfqq->cfqd; |
a36e71f9 | 976 | struct request *__alias, *prev; |
1da177e4 | 977 | |
5380a101 | 978 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
979 | |
980 | /* | |
981 | * looks a little odd, but the first insert might return an alias. | |
982 | * if that happens, put the alias on the dispatch list | |
983 | */ | |
21183b07 | 984 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 985 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
986 | |
987 | if (!cfq_cfqq_on_rr(cfqq)) | |
988 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
989 | |
990 | /* | |
991 | * check if this request is a better next-serve candidate | |
992 | */ | |
a36e71f9 | 993 | prev = cfqq->next_rq; |
cf7c25cf | 994 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); |
a36e71f9 JA |
995 | |
996 | /* | |
997 | * adjust priority tree position, if ->next_rq changes | |
998 | */ | |
999 | if (prev != cfqq->next_rq) | |
1000 | cfq_prio_tree_add(cfqd, cfqq); | |
1001 | ||
5044eed4 | 1002 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
1003 | } |
1004 | ||
febffd61 | 1005 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 1006 | { |
5380a101 JA |
1007 | elv_rb_del(&cfqq->sort_list, rq); |
1008 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 1009 | cfq_add_rq_rb(rq); |
1da177e4 LT |
1010 | } |
1011 | ||
206dc69b JA |
1012 | static struct request * |
1013 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 1014 | { |
206dc69b | 1015 | struct task_struct *tsk = current; |
91fac317 | 1016 | struct cfq_io_context *cic; |
206dc69b | 1017 | struct cfq_queue *cfqq; |
1da177e4 | 1018 | |
4ac845a2 | 1019 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
1020 | if (!cic) |
1021 | return NULL; | |
1022 | ||
1023 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
1024 | if (cfqq) { |
1025 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
1026 | ||
21183b07 | 1027 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 1028 | } |
1da177e4 | 1029 | |
1da177e4 LT |
1030 | return NULL; |
1031 | } | |
1032 | ||
165125e1 | 1033 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1034 | { |
22e2c507 | 1035 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 1036 | |
5ad531db | 1037 | cfqd->rq_in_driver[rq_is_sync(rq)]++; |
7b679138 | 1038 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
5ad531db | 1039 | rq_in_driver(cfqd)); |
25776e35 | 1040 | |
5b93629b | 1041 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
1042 | } |
1043 | ||
165125e1 | 1044 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1045 | { |
b4878f24 | 1046 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5ad531db | 1047 | const int sync = rq_is_sync(rq); |
b4878f24 | 1048 | |
5ad531db JA |
1049 | WARN_ON(!cfqd->rq_in_driver[sync]); |
1050 | cfqd->rq_in_driver[sync]--; | |
7b679138 | 1051 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
5ad531db | 1052 | rq_in_driver(cfqd)); |
1da177e4 LT |
1053 | } |
1054 | ||
b4878f24 | 1055 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 1056 | { |
5e705374 | 1057 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 1058 | |
5e705374 JA |
1059 | if (cfqq->next_rq == rq) |
1060 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 1061 | |
b4878f24 | 1062 | list_del_init(&rq->queuelist); |
5e705374 | 1063 | cfq_del_rq_rb(rq); |
374f84ac | 1064 | |
45333d5a | 1065 | cfqq->cfqd->rq_queued--; |
374f84ac JA |
1066 | if (rq_is_meta(rq)) { |
1067 | WARN_ON(!cfqq->meta_pending); | |
1068 | cfqq->meta_pending--; | |
1069 | } | |
1da177e4 LT |
1070 | } |
1071 | ||
165125e1 JA |
1072 | static int cfq_merge(struct request_queue *q, struct request **req, |
1073 | struct bio *bio) | |
1da177e4 LT |
1074 | { |
1075 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1076 | struct request *__rq; | |
1da177e4 | 1077 | |
206dc69b | 1078 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 1079 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
1080 | *req = __rq; |
1081 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1082 | } |
1083 | ||
1084 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1085 | } |
1086 | ||
165125e1 | 1087 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 1088 | int type) |
1da177e4 | 1089 | { |
21183b07 | 1090 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 1091 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 1092 | |
5e705374 | 1093 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 1094 | } |
1da177e4 LT |
1095 | } |
1096 | ||
1097 | static void | |
165125e1 | 1098 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
1099 | struct request *next) |
1100 | { | |
cf7c25cf | 1101 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 JA |
1102 | /* |
1103 | * reposition in fifo if next is older than rq | |
1104 | */ | |
1105 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
30996f40 | 1106 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
22e2c507 | 1107 | list_move(&rq->queuelist, &next->queuelist); |
30996f40 JA |
1108 | rq_set_fifo_time(rq, rq_fifo_time(next)); |
1109 | } | |
22e2c507 | 1110 | |
cf7c25cf CZ |
1111 | if (cfqq->next_rq == next) |
1112 | cfqq->next_rq = rq; | |
b4878f24 | 1113 | cfq_remove_request(next); |
22e2c507 JA |
1114 | } |
1115 | ||
165125e1 | 1116 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
1117 | struct bio *bio) |
1118 | { | |
1119 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 1120 | struct cfq_io_context *cic; |
da775265 | 1121 | struct cfq_queue *cfqq; |
da775265 JA |
1122 | |
1123 | /* | |
ec8acb69 | 1124 | * Disallow merge of a sync bio into an async request. |
da775265 | 1125 | */ |
91fac317 | 1126 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
a6151c3a | 1127 | return false; |
da775265 JA |
1128 | |
1129 | /* | |
719d3402 JA |
1130 | * Lookup the cfqq that this bio will be queued with. Allow |
1131 | * merge only if rq is queued there. | |
da775265 | 1132 | */ |
4ac845a2 | 1133 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 | 1134 | if (!cic) |
a6151c3a | 1135 | return false; |
719d3402 | 1136 | |
91fac317 | 1137 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
a6151c3a | 1138 | return cfqq == RQ_CFQQ(rq); |
da775265 JA |
1139 | } |
1140 | ||
febffd61 JA |
1141 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
1142 | struct cfq_queue *cfqq) | |
22e2c507 JA |
1143 | { |
1144 | if (cfqq) { | |
7b679138 | 1145 | cfq_log_cfqq(cfqd, cfqq, "set_active"); |
22e2c507 | 1146 | cfqq->slice_end = 0; |
2f5cb738 JA |
1147 | cfqq->slice_dispatch = 0; |
1148 | ||
2f5cb738 | 1149 | cfq_clear_cfqq_wait_request(cfqq); |
b029195d | 1150 | cfq_clear_cfqq_must_dispatch(cfqq); |
3b18152c JA |
1151 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
1152 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 1153 | cfq_mark_cfqq_slice_new(cfqq); |
2f5cb738 JA |
1154 | |
1155 | del_timer(&cfqd->idle_slice_timer); | |
22e2c507 JA |
1156 | } |
1157 | ||
1158 | cfqd->active_queue = cfqq; | |
1159 | } | |
1160 | ||
7b14e3b5 JA |
1161 | /* |
1162 | * current cfqq expired its slice (or was too idle), select new one | |
1163 | */ | |
1164 | static void | |
1165 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
a6151c3a | 1166 | bool timed_out) |
7b14e3b5 | 1167 | { |
7b679138 JA |
1168 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
1169 | ||
7b14e3b5 JA |
1170 | if (cfq_cfqq_wait_request(cfqq)) |
1171 | del_timer(&cfqd->idle_slice_timer); | |
1172 | ||
7b14e3b5 JA |
1173 | cfq_clear_cfqq_wait_request(cfqq); |
1174 | ||
1175 | /* | |
6084cdda | 1176 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 1177 | */ |
7b679138 | 1178 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) { |
c5b680f3 | 1179 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b679138 JA |
1180 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
1181 | } | |
7b14e3b5 | 1182 | |
f04a6424 VG |
1183 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
1184 | cfq_del_cfqq_rr(cfqd, cfqq); | |
1185 | ||
edd75ffd | 1186 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
1187 | |
1188 | if (cfqq == cfqd->active_queue) | |
1189 | cfqd->active_queue = NULL; | |
1190 | ||
1191 | if (cfqd->active_cic) { | |
1192 | put_io_context(cfqd->active_cic->ioc); | |
1193 | cfqd->active_cic = NULL; | |
1194 | } | |
7b14e3b5 JA |
1195 | } |
1196 | ||
a6151c3a | 1197 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) |
7b14e3b5 JA |
1198 | { |
1199 | struct cfq_queue *cfqq = cfqd->active_queue; | |
1200 | ||
1201 | if (cfqq) | |
6084cdda | 1202 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
1203 | } |
1204 | ||
498d3aa2 JA |
1205 | /* |
1206 | * Get next queue for service. Unless we have a queue preemption, | |
1207 | * we'll simply select the first cfqq in the service tree. | |
1208 | */ | |
6d048f53 | 1209 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 1210 | { |
c0324a02 | 1211 | struct cfq_rb_root *service_tree = |
cdb16e8f VG |
1212 | service_tree_for(cfqd->serving_group, cfqd->serving_prio, |
1213 | cfqd->serving_type, cfqd); | |
d9e7620e | 1214 | |
f04a6424 VG |
1215 | if (!cfqd->rq_queued) |
1216 | return NULL; | |
1217 | ||
1fa8f6d6 VG |
1218 | /* There is nothing to dispatch */ |
1219 | if (!service_tree) | |
1220 | return NULL; | |
c0324a02 CZ |
1221 | if (RB_EMPTY_ROOT(&service_tree->rb)) |
1222 | return NULL; | |
1223 | return cfq_rb_first(service_tree); | |
6d048f53 JA |
1224 | } |
1225 | ||
f04a6424 VG |
1226 | static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) |
1227 | { | |
1228 | struct cfq_group *cfqg = &cfqd->root_group; | |
1229 | struct cfq_queue *cfqq; | |
1230 | int i, j; | |
1231 | struct cfq_rb_root *st; | |
1232 | ||
1233 | if (!cfqd->rq_queued) | |
1234 | return NULL; | |
1235 | ||
1236 | for_each_cfqg_st(cfqg, i, j, st) | |
1237 | if ((cfqq = cfq_rb_first(st)) != NULL) | |
1238 | return cfqq; | |
1239 | return NULL; | |
1240 | } | |
1241 | ||
498d3aa2 JA |
1242 | /* |
1243 | * Get and set a new active queue for service. | |
1244 | */ | |
a36e71f9 JA |
1245 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
1246 | struct cfq_queue *cfqq) | |
6d048f53 | 1247 | { |
e00ef799 | 1248 | if (!cfqq) |
a36e71f9 | 1249 | cfqq = cfq_get_next_queue(cfqd); |
6d048f53 | 1250 | |
22e2c507 | 1251 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 1252 | return cfqq; |
22e2c507 JA |
1253 | } |
1254 | ||
d9e7620e JA |
1255 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
1256 | struct request *rq) | |
1257 | { | |
83096ebf TH |
1258 | if (blk_rq_pos(rq) >= cfqd->last_position) |
1259 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 1260 | else |
83096ebf | 1261 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
1262 | } |
1263 | ||
b2c18e1e JM |
1264 | #define CFQQ_SEEK_THR 8 * 1024 |
1265 | #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR) | |
04dc6e71 | 1266 | |
b2c18e1e JM |
1267 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
1268 | struct request *rq) | |
6d048f53 | 1269 | { |
b2c18e1e | 1270 | sector_t sdist = cfqq->seek_mean; |
6d048f53 | 1271 | |
b2c18e1e JM |
1272 | if (!sample_valid(cfqq->seek_samples)) |
1273 | sdist = CFQQ_SEEK_THR; | |
6d048f53 | 1274 | |
04dc6e71 | 1275 | return cfq_dist_from_last(cfqd, rq) <= sdist; |
6d048f53 JA |
1276 | } |
1277 | ||
a36e71f9 JA |
1278 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
1279 | struct cfq_queue *cur_cfqq) | |
1280 | { | |
f2d1f0ae | 1281 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
1282 | struct rb_node *parent, *node; |
1283 | struct cfq_queue *__cfqq; | |
1284 | sector_t sector = cfqd->last_position; | |
1285 | ||
1286 | if (RB_EMPTY_ROOT(root)) | |
1287 | return NULL; | |
1288 | ||
1289 | /* | |
1290 | * First, if we find a request starting at the end of the last | |
1291 | * request, choose it. | |
1292 | */ | |
f2d1f0ae | 1293 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
1294 | if (__cfqq) |
1295 | return __cfqq; | |
1296 | ||
1297 | /* | |
1298 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1299 | * will contain the closest sector. | |
1300 | */ | |
1301 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
b2c18e1e | 1302 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1303 | return __cfqq; |
1304 | ||
2e46e8b2 | 1305 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1306 | node = rb_next(&__cfqq->p_node); |
1307 | else | |
1308 | node = rb_prev(&__cfqq->p_node); | |
1309 | if (!node) | |
1310 | return NULL; | |
1311 | ||
1312 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
b2c18e1e | 1313 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1314 | return __cfqq; |
1315 | ||
1316 | return NULL; | |
1317 | } | |
1318 | ||
1319 | /* | |
1320 | * cfqd - obvious | |
1321 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1322 | * closely cooperating with itself. | |
1323 | * | |
1324 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1325 | * one request, and that cfqd->last_position reflects a position on the disk | |
1326 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1327 | * assumption. | |
1328 | */ | |
1329 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
b3b6d040 | 1330 | struct cfq_queue *cur_cfqq) |
6d048f53 | 1331 | { |
a36e71f9 JA |
1332 | struct cfq_queue *cfqq; |
1333 | ||
e6c5bc73 JM |
1334 | if (!cfq_cfqq_sync(cur_cfqq)) |
1335 | return NULL; | |
1336 | if (CFQQ_SEEKY(cur_cfqq)) | |
1337 | return NULL; | |
1338 | ||
6d048f53 | 1339 | /* |
d9e7620e JA |
1340 | * We should notice if some of the queues are cooperating, eg |
1341 | * working closely on the same area of the disk. In that case, | |
1342 | * we can group them together and don't waste time idling. | |
6d048f53 | 1343 | */ |
a36e71f9 JA |
1344 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1345 | if (!cfqq) | |
1346 | return NULL; | |
1347 | ||
df5fe3e8 JM |
1348 | /* |
1349 | * It only makes sense to merge sync queues. | |
1350 | */ | |
1351 | if (!cfq_cfqq_sync(cfqq)) | |
1352 | return NULL; | |
e6c5bc73 JM |
1353 | if (CFQQ_SEEKY(cfqq)) |
1354 | return NULL; | |
df5fe3e8 | 1355 | |
c0324a02 CZ |
1356 | /* |
1357 | * Do not merge queues of different priority classes | |
1358 | */ | |
1359 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
1360 | return NULL; | |
1361 | ||
a36e71f9 | 1362 | return cfqq; |
6d048f53 JA |
1363 | } |
1364 | ||
a6d44e98 CZ |
1365 | /* |
1366 | * Determine whether we should enforce idle window for this queue. | |
1367 | */ | |
1368 | ||
1369 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1370 | { | |
1371 | enum wl_prio_t prio = cfqq_prio(cfqq); | |
718eee05 | 1372 | struct cfq_rb_root *service_tree = cfqq->service_tree; |
a6d44e98 | 1373 | |
f04a6424 VG |
1374 | BUG_ON(!service_tree); |
1375 | BUG_ON(!service_tree->count); | |
1376 | ||
a6d44e98 CZ |
1377 | /* We never do for idle class queues. */ |
1378 | if (prio == IDLE_WORKLOAD) | |
1379 | return false; | |
1380 | ||
1381 | /* We do for queues that were marked with idle window flag. */ | |
1382 | if (cfq_cfqq_idle_window(cfqq)) | |
1383 | return true; | |
1384 | ||
1385 | /* | |
1386 | * Otherwise, we do only if they are the last ones | |
1387 | * in their service tree. | |
1388 | */ | |
f04a6424 | 1389 | return service_tree->count == 1; |
a6d44e98 CZ |
1390 | } |
1391 | ||
6d048f53 | 1392 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1393 | { |
1792669c | 1394 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1395 | struct cfq_io_context *cic; |
7b14e3b5 JA |
1396 | unsigned long sl; |
1397 | ||
a68bbddb | 1398 | /* |
f7d7b7a7 JA |
1399 | * SSD device without seek penalty, disable idling. But only do so |
1400 | * for devices that support queuing, otherwise we still have a problem | |
1401 | * with sync vs async workloads. | |
a68bbddb | 1402 | */ |
f7d7b7a7 | 1403 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
1404 | return; |
1405 | ||
dd67d051 | 1406 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 1407 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
1408 | |
1409 | /* | |
1410 | * idle is disabled, either manually or by past process history | |
1411 | */ | |
a6d44e98 | 1412 | if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq)) |
6d048f53 JA |
1413 | return; |
1414 | ||
7b679138 | 1415 | /* |
8e550632 | 1416 | * still active requests from this queue, don't idle |
7b679138 | 1417 | */ |
8e550632 | 1418 | if (cfqq->dispatched) |
7b679138 JA |
1419 | return; |
1420 | ||
22e2c507 JA |
1421 | /* |
1422 | * task has exited, don't wait | |
1423 | */ | |
206dc69b | 1424 | cic = cfqd->active_cic; |
66dac98e | 1425 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
1426 | return; |
1427 | ||
355b659c CZ |
1428 | /* |
1429 | * If our average think time is larger than the remaining time | |
1430 | * slice, then don't idle. This avoids overrunning the allotted | |
1431 | * time slice. | |
1432 | */ | |
1433 | if (sample_valid(cic->ttime_samples) && | |
1434 | (cfqq->slice_end - jiffies < cic->ttime_mean)) | |
1435 | return; | |
1436 | ||
3b18152c | 1437 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 1438 | |
6d048f53 | 1439 | sl = cfqd->cfq_slice_idle; |
206dc69b | 1440 | |
7b14e3b5 | 1441 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
9481ffdc | 1442 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); |
1da177e4 LT |
1443 | } |
1444 | ||
498d3aa2 JA |
1445 | /* |
1446 | * Move request from internal lists to the request queue dispatch list. | |
1447 | */ | |
165125e1 | 1448 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 1449 | { |
3ed9a296 | 1450 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1451 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 1452 | |
7b679138 JA |
1453 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
1454 | ||
06d21886 | 1455 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); |
5380a101 | 1456 | cfq_remove_request(rq); |
6d048f53 | 1457 | cfqq->dispatched++; |
5380a101 | 1458 | elv_dispatch_sort(q, rq); |
3ed9a296 JA |
1459 | |
1460 | if (cfq_cfqq_sync(cfqq)) | |
1461 | cfqd->sync_flight++; | |
1da177e4 LT |
1462 | } |
1463 | ||
1464 | /* | |
1465 | * return expired entry, or NULL to just start from scratch in rbtree | |
1466 | */ | |
febffd61 | 1467 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 | 1468 | { |
30996f40 | 1469 | struct request *rq = NULL; |
1da177e4 | 1470 | |
3b18152c | 1471 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 1472 | return NULL; |
cb887411 JA |
1473 | |
1474 | cfq_mark_cfqq_fifo_expire(cfqq); | |
1475 | ||
89850f7e JA |
1476 | if (list_empty(&cfqq->fifo)) |
1477 | return NULL; | |
1da177e4 | 1478 | |
89850f7e | 1479 | rq = rq_entry_fifo(cfqq->fifo.next); |
30996f40 | 1480 | if (time_before(jiffies, rq_fifo_time(rq))) |
7b679138 | 1481 | rq = NULL; |
1da177e4 | 1482 | |
30996f40 | 1483 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 1484 | return rq; |
1da177e4 LT |
1485 | } |
1486 | ||
22e2c507 JA |
1487 | static inline int |
1488 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1489 | { | |
1490 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 1491 | |
22e2c507 | 1492 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 1493 | |
22e2c507 | 1494 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
1495 | } |
1496 | ||
df5fe3e8 JM |
1497 | /* |
1498 | * Must be called with the queue_lock held. | |
1499 | */ | |
1500 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
1501 | { | |
1502 | int process_refs, io_refs; | |
1503 | ||
1504 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
1505 | process_refs = atomic_read(&cfqq->ref) - io_refs; | |
1506 | BUG_ON(process_refs < 0); | |
1507 | return process_refs; | |
1508 | } | |
1509 | ||
1510 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
1511 | { | |
e6c5bc73 | 1512 | int process_refs, new_process_refs; |
df5fe3e8 JM |
1513 | struct cfq_queue *__cfqq; |
1514 | ||
1515 | /* Avoid a circular list and skip interim queue merges */ | |
1516 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
1517 | if (__cfqq == cfqq) | |
1518 | return; | |
1519 | new_cfqq = __cfqq; | |
1520 | } | |
1521 | ||
1522 | process_refs = cfqq_process_refs(cfqq); | |
1523 | /* | |
1524 | * If the process for the cfqq has gone away, there is no | |
1525 | * sense in merging the queues. | |
1526 | */ | |
1527 | if (process_refs == 0) | |
1528 | return; | |
1529 | ||
e6c5bc73 JM |
1530 | /* |
1531 | * Merge in the direction of the lesser amount of work. | |
1532 | */ | |
1533 | new_process_refs = cfqq_process_refs(new_cfqq); | |
1534 | if (new_process_refs >= process_refs) { | |
1535 | cfqq->new_cfqq = new_cfqq; | |
1536 | atomic_add(process_refs, &new_cfqq->ref); | |
1537 | } else { | |
1538 | new_cfqq->new_cfqq = cfqq; | |
1539 | atomic_add(new_process_refs, &cfqq->ref); | |
1540 | } | |
df5fe3e8 JM |
1541 | } |
1542 | ||
cdb16e8f VG |
1543 | static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, |
1544 | struct cfq_group *cfqg, enum wl_prio_t prio, | |
1545 | bool prio_changed) | |
718eee05 CZ |
1546 | { |
1547 | struct cfq_queue *queue; | |
1548 | int i; | |
1549 | bool key_valid = false; | |
1550 | unsigned long lowest_key = 0; | |
1551 | enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; | |
1552 | ||
1553 | if (prio_changed) { | |
1554 | /* | |
1555 | * When priorities switched, we prefer starting | |
1556 | * from SYNC_NOIDLE (first choice), or just SYNC | |
1557 | * over ASYNC | |
1558 | */ | |
cdb16e8f | 1559 | if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) |
718eee05 CZ |
1560 | return cur_best; |
1561 | cur_best = SYNC_WORKLOAD; | |
cdb16e8f | 1562 | if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) |
718eee05 CZ |
1563 | return cur_best; |
1564 | ||
1565 | return ASYNC_WORKLOAD; | |
1566 | } | |
1567 | ||
1568 | for (i = 0; i < 3; ++i) { | |
1569 | /* otherwise, select the one with lowest rb_key */ | |
cdb16e8f | 1570 | queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd)); |
718eee05 CZ |
1571 | if (queue && |
1572 | (!key_valid || time_before(queue->rb_key, lowest_key))) { | |
1573 | lowest_key = queue->rb_key; | |
1574 | cur_best = i; | |
1575 | key_valid = true; | |
1576 | } | |
1577 | } | |
1578 | ||
1579 | return cur_best; | |
1580 | } | |
1581 | ||
cdb16e8f | 1582 | static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) |
718eee05 CZ |
1583 | { |
1584 | enum wl_prio_t previous_prio = cfqd->serving_prio; | |
1585 | bool prio_changed; | |
1586 | unsigned slice; | |
1587 | unsigned count; | |
cdb16e8f | 1588 | struct cfq_rb_root *st; |
718eee05 | 1589 | |
1fa8f6d6 VG |
1590 | if (!cfqg) { |
1591 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1592 | cfqd->workload_expires = jiffies + 1; | |
1593 | return; | |
1594 | } | |
1595 | ||
718eee05 CZ |
1596 | /* Choose next priority. RT > BE > IDLE */ |
1597 | if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd)) | |
1598 | cfqd->serving_prio = RT_WORKLOAD; | |
1599 | else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd)) | |
1600 | cfqd->serving_prio = BE_WORKLOAD; | |
1601 | else { | |
1602 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1603 | cfqd->workload_expires = jiffies + 1; | |
1604 | return; | |
1605 | } | |
1606 | ||
1607 | /* | |
1608 | * For RT and BE, we have to choose also the type | |
1609 | * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload | |
1610 | * expiration time | |
1611 | */ | |
1612 | prio_changed = (cfqd->serving_prio != previous_prio); | |
cdb16e8f VG |
1613 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, |
1614 | cfqd); | |
1615 | count = st->count; | |
718eee05 CZ |
1616 | |
1617 | /* | |
1618 | * If priority didn't change, check workload expiration, | |
1619 | * and that we still have other queues ready | |
1620 | */ | |
1621 | if (!prio_changed && count && | |
1622 | !time_after(jiffies, cfqd->workload_expires)) | |
1623 | return; | |
1624 | ||
1625 | /* otherwise select new workload type */ | |
1626 | cfqd->serving_type = | |
cdb16e8f VG |
1627 | cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed); |
1628 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, | |
1629 | cfqd); | |
1630 | count = st->count; | |
718eee05 CZ |
1631 | |
1632 | /* | |
1633 | * the workload slice is computed as a fraction of target latency | |
1634 | * proportional to the number of queues in that workload, over | |
1635 | * all the queues in the same priority class | |
1636 | */ | |
1637 | slice = cfq_target_latency * count / | |
1638 | max_t(unsigned, cfqd->busy_queues_avg[cfqd->serving_prio], | |
1639 | cfq_busy_queues_wl(cfqd->serving_prio, cfqd)); | |
1640 | ||
1641 | if (cfqd->serving_type == ASYNC_WORKLOAD) | |
1642 | /* async workload slice is scaled down according to | |
1643 | * the sync/async slice ratio. */ | |
1644 | slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; | |
1645 | else | |
1646 | /* sync workload slice is at least 2 * cfq_slice_idle */ | |
1647 | slice = max(slice, 2 * cfqd->cfq_slice_idle); | |
1648 | ||
1649 | slice = max_t(unsigned, slice, CFQ_MIN_TT); | |
1650 | cfqd->workload_expires = jiffies + slice; | |
8e550632 | 1651 | cfqd->noidle_tree_requires_idle = false; |
718eee05 CZ |
1652 | } |
1653 | ||
1fa8f6d6 VG |
1654 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) |
1655 | { | |
1656 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
1657 | ||
1658 | if (RB_EMPTY_ROOT(&st->rb)) | |
1659 | return NULL; | |
1660 | return cfq_rb_first_group(st); | |
1661 | } | |
1662 | ||
cdb16e8f VG |
1663 | static void cfq_choose_cfqg(struct cfq_data *cfqd) |
1664 | { | |
1fa8f6d6 VG |
1665 | struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); |
1666 | ||
1667 | cfqd->serving_group = cfqg; | |
1668 | choose_service_tree(cfqd, cfqg); | |
cdb16e8f VG |
1669 | } |
1670 | ||
22e2c507 | 1671 | /* |
498d3aa2 JA |
1672 | * Select a queue for service. If we have a current active queue, |
1673 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 1674 | */ |
1b5ed5e1 | 1675 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 1676 | { |
a36e71f9 | 1677 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 1678 | |
22e2c507 JA |
1679 | cfqq = cfqd->active_queue; |
1680 | if (!cfqq) | |
1681 | goto new_queue; | |
1da177e4 | 1682 | |
f04a6424 VG |
1683 | if (!cfqd->rq_queued) |
1684 | return NULL; | |
22e2c507 | 1685 | /* |
6d048f53 | 1686 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 1687 | */ |
b029195d | 1688 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) |
3b18152c | 1689 | goto expire; |
1da177e4 | 1690 | |
22e2c507 | 1691 | /* |
6d048f53 JA |
1692 | * The active queue has requests and isn't expired, allow it to |
1693 | * dispatch. | |
22e2c507 | 1694 | */ |
dd67d051 | 1695 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 1696 | goto keep_queue; |
6d048f53 | 1697 | |
a36e71f9 JA |
1698 | /* |
1699 | * If another queue has a request waiting within our mean seek | |
1700 | * distance, let it run. The expire code will check for close | |
1701 | * cooperators and put the close queue at the front of the service | |
df5fe3e8 | 1702 | * tree. If possible, merge the expiring queue with the new cfqq. |
a36e71f9 | 1703 | */ |
b3b6d040 | 1704 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); |
df5fe3e8 JM |
1705 | if (new_cfqq) { |
1706 | if (!cfqq->new_cfqq) | |
1707 | cfq_setup_merge(cfqq, new_cfqq); | |
a36e71f9 | 1708 | goto expire; |
df5fe3e8 | 1709 | } |
a36e71f9 | 1710 | |
6d048f53 JA |
1711 | /* |
1712 | * No requests pending. If the active queue still has requests in | |
1713 | * flight or is idling for a new request, allow either of these | |
1714 | * conditions to happen (or time out) before selecting a new queue. | |
1715 | */ | |
cc197479 | 1716 | if (timer_pending(&cfqd->idle_slice_timer) || |
a6d44e98 | 1717 | (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) { |
caaa5f9f JA |
1718 | cfqq = NULL; |
1719 | goto keep_queue; | |
22e2c507 JA |
1720 | } |
1721 | ||
3b18152c | 1722 | expire: |
6084cdda | 1723 | cfq_slice_expired(cfqd, 0); |
3b18152c | 1724 | new_queue: |
718eee05 CZ |
1725 | /* |
1726 | * Current queue expired. Check if we have to switch to a new | |
1727 | * service tree | |
1728 | */ | |
1729 | if (!new_cfqq) | |
cdb16e8f | 1730 | cfq_choose_cfqg(cfqd); |
718eee05 | 1731 | |
a36e71f9 | 1732 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 1733 | keep_queue: |
3b18152c | 1734 | return cfqq; |
22e2c507 JA |
1735 | } |
1736 | ||
febffd61 | 1737 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
1738 | { |
1739 | int dispatched = 0; | |
1740 | ||
1741 | while (cfqq->next_rq) { | |
1742 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
1743 | dispatched++; | |
1744 | } | |
1745 | ||
1746 | BUG_ON(!list_empty(&cfqq->fifo)); | |
f04a6424 VG |
1747 | |
1748 | /* By default cfqq is not expired if it is empty. Do it explicitly */ | |
1749 | __cfq_slice_expired(cfqq->cfqd, cfqq, 0); | |
d9e7620e JA |
1750 | return dispatched; |
1751 | } | |
1752 | ||
498d3aa2 JA |
1753 | /* |
1754 | * Drain our current requests. Used for barriers and when switching | |
1755 | * io schedulers on-the-fly. | |
1756 | */ | |
d9e7620e | 1757 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 1758 | { |
0871714e | 1759 | struct cfq_queue *cfqq; |
d9e7620e | 1760 | int dispatched = 0; |
cdb16e8f | 1761 | |
f04a6424 VG |
1762 | while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) |
1763 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); | |
1b5ed5e1 | 1764 | |
6084cdda | 1765 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
1766 | BUG_ON(cfqd->busy_queues); |
1767 | ||
6923715a | 1768 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
1769 | return dispatched; |
1770 | } | |
1771 | ||
0b182d61 | 1772 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
2f5cb738 | 1773 | { |
2f5cb738 | 1774 | unsigned int max_dispatch; |
22e2c507 | 1775 | |
5ad531db JA |
1776 | /* |
1777 | * Drain async requests before we start sync IO | |
1778 | */ | |
a6d44e98 | 1779 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC]) |
0b182d61 | 1780 | return false; |
5ad531db | 1781 | |
2f5cb738 JA |
1782 | /* |
1783 | * If this is an async queue and we have sync IO in flight, let it wait | |
1784 | */ | |
1785 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) | |
0b182d61 | 1786 | return false; |
2f5cb738 JA |
1787 | |
1788 | max_dispatch = cfqd->cfq_quantum; | |
1789 | if (cfq_class_idle(cfqq)) | |
1790 | max_dispatch = 1; | |
b4878f24 | 1791 | |
2f5cb738 JA |
1792 | /* |
1793 | * Does this cfqq already have too much IO in flight? | |
1794 | */ | |
1795 | if (cfqq->dispatched >= max_dispatch) { | |
1796 | /* | |
1797 | * idle queue must always only have a single IO in flight | |
1798 | */ | |
3ed9a296 | 1799 | if (cfq_class_idle(cfqq)) |
0b182d61 | 1800 | return false; |
3ed9a296 | 1801 | |
2f5cb738 JA |
1802 | /* |
1803 | * We have other queues, don't allow more IO from this one | |
1804 | */ | |
1805 | if (cfqd->busy_queues > 1) | |
0b182d61 | 1806 | return false; |
9ede209e | 1807 | |
365722bb | 1808 | /* |
474b18cc | 1809 | * Sole queue user, no limit |
365722bb | 1810 | */ |
474b18cc | 1811 | max_dispatch = -1; |
8e296755 JA |
1812 | } |
1813 | ||
1814 | /* | |
1815 | * Async queues must wait a bit before being allowed dispatch. | |
1816 | * We also ramp up the dispatch depth gradually for async IO, | |
1817 | * based on the last sync IO we serviced | |
1818 | */ | |
963b72fc | 1819 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { |
8e296755 JA |
1820 | unsigned long last_sync = jiffies - cfqd->last_end_sync_rq; |
1821 | unsigned int depth; | |
365722bb | 1822 | |
61f0c1dc | 1823 | depth = last_sync / cfqd->cfq_slice[1]; |
e00c54c3 JA |
1824 | if (!depth && !cfqq->dispatched) |
1825 | depth = 1; | |
8e296755 JA |
1826 | if (depth < max_dispatch) |
1827 | max_dispatch = depth; | |
2f5cb738 | 1828 | } |
3ed9a296 | 1829 | |
0b182d61 JA |
1830 | /* |
1831 | * If we're below the current max, allow a dispatch | |
1832 | */ | |
1833 | return cfqq->dispatched < max_dispatch; | |
1834 | } | |
1835 | ||
1836 | /* | |
1837 | * Dispatch a request from cfqq, moving them to the request queue | |
1838 | * dispatch list. | |
1839 | */ | |
1840 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1841 | { | |
1842 | struct request *rq; | |
1843 | ||
1844 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
1845 | ||
1846 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
1847 | return false; | |
1848 | ||
1849 | /* | |
1850 | * follow expired path, else get first next available | |
1851 | */ | |
1852 | rq = cfq_check_fifo(cfqq); | |
1853 | if (!rq) | |
1854 | rq = cfqq->next_rq; | |
1855 | ||
1856 | /* | |
1857 | * insert request into driver dispatch list | |
1858 | */ | |
1859 | cfq_dispatch_insert(cfqd->queue, rq); | |
1860 | ||
1861 | if (!cfqd->active_cic) { | |
1862 | struct cfq_io_context *cic = RQ_CIC(rq); | |
1863 | ||
1864 | atomic_long_inc(&cic->ioc->refcount); | |
1865 | cfqd->active_cic = cic; | |
1866 | } | |
1867 | ||
1868 | return true; | |
1869 | } | |
1870 | ||
1871 | /* | |
1872 | * Find the cfqq that we need to service and move a request from that to the | |
1873 | * dispatch list | |
1874 | */ | |
1875 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
1876 | { | |
1877 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1878 | struct cfq_queue *cfqq; | |
1879 | ||
1880 | if (!cfqd->busy_queues) | |
1881 | return 0; | |
1882 | ||
1883 | if (unlikely(force)) | |
1884 | return cfq_forced_dispatch(cfqd); | |
1885 | ||
1886 | cfqq = cfq_select_queue(cfqd); | |
1887 | if (!cfqq) | |
8e296755 JA |
1888 | return 0; |
1889 | ||
2f5cb738 | 1890 | /* |
0b182d61 | 1891 | * Dispatch a request from this cfqq, if it is allowed |
2f5cb738 | 1892 | */ |
0b182d61 JA |
1893 | if (!cfq_dispatch_request(cfqd, cfqq)) |
1894 | return 0; | |
1895 | ||
2f5cb738 | 1896 | cfqq->slice_dispatch++; |
b029195d | 1897 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 1898 | |
2f5cb738 JA |
1899 | /* |
1900 | * expire an async queue immediately if it has used up its slice. idle | |
1901 | * queue always expire after 1 dispatch round. | |
1902 | */ | |
1903 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
1904 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
1905 | cfq_class_idle(cfqq))) { | |
1906 | cfqq->slice_end = jiffies + 1; | |
1907 | cfq_slice_expired(cfqd, 0); | |
1da177e4 LT |
1908 | } |
1909 | ||
b217a903 | 1910 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); |
2f5cb738 | 1911 | return 1; |
1da177e4 LT |
1912 | } |
1913 | ||
1da177e4 | 1914 | /* |
5e705374 JA |
1915 | * task holds one reference to the queue, dropped when task exits. each rq |
1916 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 LT |
1917 | * |
1918 | * queue lock must be held here. | |
1919 | */ | |
1920 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
1921 | { | |
22e2c507 JA |
1922 | struct cfq_data *cfqd = cfqq->cfqd; |
1923 | ||
1924 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
1925 | |
1926 | if (!atomic_dec_and_test(&cfqq->ref)) | |
1927 | return; | |
1928 | ||
7b679138 | 1929 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 1930 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 1931 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
1da177e4 | 1932 | |
28f95cbc | 1933 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 1934 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 1935 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 1936 | } |
22e2c507 | 1937 | |
f04a6424 | 1938 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 LT |
1939 | kmem_cache_free(cfq_pool, cfqq); |
1940 | } | |
1941 | ||
d6de8be7 JA |
1942 | /* |
1943 | * Must always be called with the rcu_read_lock() held | |
1944 | */ | |
07416d29 JA |
1945 | static void |
1946 | __call_for_each_cic(struct io_context *ioc, | |
1947 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1948 | { | |
1949 | struct cfq_io_context *cic; | |
1950 | struct hlist_node *n; | |
1951 | ||
1952 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) | |
1953 | func(ioc, cic); | |
1954 | } | |
1955 | ||
4ac845a2 | 1956 | /* |
34e6bbf2 | 1957 | * Call func for each cic attached to this ioc. |
4ac845a2 | 1958 | */ |
34e6bbf2 | 1959 | static void |
4ac845a2 JA |
1960 | call_for_each_cic(struct io_context *ioc, |
1961 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1da177e4 | 1962 | { |
4ac845a2 | 1963 | rcu_read_lock(); |
07416d29 | 1964 | __call_for_each_cic(ioc, func); |
4ac845a2 | 1965 | rcu_read_unlock(); |
34e6bbf2 FC |
1966 | } |
1967 | ||
1968 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
1969 | { | |
1970 | struct cfq_io_context *cic; | |
1971 | ||
1972 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
1973 | ||
1974 | kmem_cache_free(cfq_ioc_pool, cic); | |
245b2e70 | 1975 | elv_ioc_count_dec(cfq_ioc_count); |
34e6bbf2 | 1976 | |
9a11b4ed JA |
1977 | if (ioc_gone) { |
1978 | /* | |
1979 | * CFQ scheduler is exiting, grab exit lock and check | |
1980 | * the pending io context count. If it hits zero, | |
1981 | * complete ioc_gone and set it back to NULL | |
1982 | */ | |
1983 | spin_lock(&ioc_gone_lock); | |
245b2e70 | 1984 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { |
9a11b4ed JA |
1985 | complete(ioc_gone); |
1986 | ioc_gone = NULL; | |
1987 | } | |
1988 | spin_unlock(&ioc_gone_lock); | |
1989 | } | |
34e6bbf2 | 1990 | } |
4ac845a2 | 1991 | |
34e6bbf2 FC |
1992 | static void cfq_cic_free(struct cfq_io_context *cic) |
1993 | { | |
1994 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
1995 | } |
1996 | ||
1997 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
1998 | { | |
1999 | unsigned long flags; | |
2000 | ||
2001 | BUG_ON(!cic->dead_key); | |
2002 | ||
2003 | spin_lock_irqsave(&ioc->lock, flags); | |
2004 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
ffc4e759 | 2005 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2006 | spin_unlock_irqrestore(&ioc->lock, flags); |
2007 | ||
34e6bbf2 | 2008 | cfq_cic_free(cic); |
4ac845a2 JA |
2009 | } |
2010 | ||
d6de8be7 JA |
2011 | /* |
2012 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
2013 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
2014 | * and ->trim() which is called with the task lock held | |
2015 | */ | |
4ac845a2 JA |
2016 | static void cfq_free_io_context(struct io_context *ioc) |
2017 | { | |
4ac845a2 | 2018 | /* |
34e6bbf2 FC |
2019 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
2020 | * so no more cic's are allowed to be linked into this ioc. So it | |
2021 | * should be ok to iterate over the known list, we will see all cic's | |
2022 | * since no new ones are added. | |
4ac845a2 | 2023 | */ |
07416d29 | 2024 | __call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
2025 | } |
2026 | ||
89850f7e | 2027 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 2028 | { |
df5fe3e8 JM |
2029 | struct cfq_queue *__cfqq, *next; |
2030 | ||
28f95cbc | 2031 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 2032 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 2033 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 2034 | } |
22e2c507 | 2035 | |
df5fe3e8 JM |
2036 | /* |
2037 | * If this queue was scheduled to merge with another queue, be | |
2038 | * sure to drop the reference taken on that queue (and others in | |
2039 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
2040 | */ | |
2041 | __cfqq = cfqq->new_cfqq; | |
2042 | while (__cfqq) { | |
2043 | if (__cfqq == cfqq) { | |
2044 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
2045 | break; | |
2046 | } | |
2047 | next = __cfqq->new_cfqq; | |
2048 | cfq_put_queue(__cfqq); | |
2049 | __cfqq = next; | |
2050 | } | |
2051 | ||
89850f7e JA |
2052 | cfq_put_queue(cfqq); |
2053 | } | |
22e2c507 | 2054 | |
89850f7e JA |
2055 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
2056 | struct cfq_io_context *cic) | |
2057 | { | |
4faa3c81 FC |
2058 | struct io_context *ioc = cic->ioc; |
2059 | ||
fc46379d | 2060 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
2061 | |
2062 | /* | |
2063 | * Make sure key == NULL is seen for dead queues | |
2064 | */ | |
fc46379d | 2065 | smp_wmb(); |
4ac845a2 | 2066 | cic->dead_key = (unsigned long) cic->key; |
fc46379d JA |
2067 | cic->key = NULL; |
2068 | ||
4faa3c81 FC |
2069 | if (ioc->ioc_data == cic) |
2070 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
2071 | ||
ff6657c6 JA |
2072 | if (cic->cfqq[BLK_RW_ASYNC]) { |
2073 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
2074 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
2075 | } |
2076 | ||
ff6657c6 JA |
2077 | if (cic->cfqq[BLK_RW_SYNC]) { |
2078 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
2079 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 2080 | } |
89850f7e JA |
2081 | } |
2082 | ||
4ac845a2 JA |
2083 | static void cfq_exit_single_io_context(struct io_context *ioc, |
2084 | struct cfq_io_context *cic) | |
89850f7e JA |
2085 | { |
2086 | struct cfq_data *cfqd = cic->key; | |
2087 | ||
89850f7e | 2088 | if (cfqd) { |
165125e1 | 2089 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 2090 | unsigned long flags; |
89850f7e | 2091 | |
4ac845a2 | 2092 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
2093 | |
2094 | /* | |
2095 | * Ensure we get a fresh copy of the ->key to prevent | |
2096 | * race between exiting task and queue | |
2097 | */ | |
2098 | smp_read_barrier_depends(); | |
2099 | if (cic->key) | |
2100 | __cfq_exit_single_io_context(cfqd, cic); | |
2101 | ||
4ac845a2 | 2102 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 2103 | } |
1da177e4 LT |
2104 | } |
2105 | ||
498d3aa2 JA |
2106 | /* |
2107 | * The process that ioc belongs to has exited, we need to clean up | |
2108 | * and put the internal structures we have that belongs to that process. | |
2109 | */ | |
e2d74ac0 | 2110 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 2111 | { |
4ac845a2 | 2112 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
2113 | } |
2114 | ||
22e2c507 | 2115 | static struct cfq_io_context * |
8267e268 | 2116 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2117 | { |
b5deef90 | 2118 | struct cfq_io_context *cic; |
1da177e4 | 2119 | |
94f6030c CL |
2120 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
2121 | cfqd->queue->node); | |
1da177e4 | 2122 | if (cic) { |
22e2c507 | 2123 | cic->last_end_request = jiffies; |
553698f9 | 2124 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 2125 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
2126 | cic->dtor = cfq_free_io_context; |
2127 | cic->exit = cfq_exit_io_context; | |
245b2e70 | 2128 | elv_ioc_count_inc(cfq_ioc_count); |
1da177e4 LT |
2129 | } |
2130 | ||
2131 | return cic; | |
2132 | } | |
2133 | ||
fd0928df | 2134 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
2135 | { |
2136 | struct task_struct *tsk = current; | |
2137 | int ioprio_class; | |
2138 | ||
3b18152c | 2139 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
2140 | return; |
2141 | ||
fd0928df | 2142 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 2143 | switch (ioprio_class) { |
fe094d98 JA |
2144 | default: |
2145 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
2146 | case IOPRIO_CLASS_NONE: | |
2147 | /* | |
6d63c275 | 2148 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
2149 | */ |
2150 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 2151 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
2152 | break; |
2153 | case IOPRIO_CLASS_RT: | |
2154 | cfqq->ioprio = task_ioprio(ioc); | |
2155 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
2156 | break; | |
2157 | case IOPRIO_CLASS_BE: | |
2158 | cfqq->ioprio = task_ioprio(ioc); | |
2159 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2160 | break; | |
2161 | case IOPRIO_CLASS_IDLE: | |
2162 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
2163 | cfqq->ioprio = 7; | |
2164 | cfq_clear_cfqq_idle_window(cfqq); | |
2165 | break; | |
22e2c507 JA |
2166 | } |
2167 | ||
2168 | /* | |
2169 | * keep track of original prio settings in case we have to temporarily | |
2170 | * elevate the priority of this queue | |
2171 | */ | |
2172 | cfqq->org_ioprio = cfqq->ioprio; | |
2173 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 2174 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
2175 | } |
2176 | ||
febffd61 | 2177 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 2178 | { |
478a82b0 AV |
2179 | struct cfq_data *cfqd = cic->key; |
2180 | struct cfq_queue *cfqq; | |
c1b707d2 | 2181 | unsigned long flags; |
35e6077c | 2182 | |
caaa5f9f JA |
2183 | if (unlikely(!cfqd)) |
2184 | return; | |
2185 | ||
c1b707d2 | 2186 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 2187 | |
ff6657c6 | 2188 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
2189 | if (cfqq) { |
2190 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
2191 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
2192 | GFP_ATOMIC); | |
caaa5f9f | 2193 | if (new_cfqq) { |
ff6657c6 | 2194 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
2195 | cfq_put_queue(cfqq); |
2196 | } | |
22e2c507 | 2197 | } |
caaa5f9f | 2198 | |
ff6657c6 | 2199 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
2200 | if (cfqq) |
2201 | cfq_mark_cfqq_prio_changed(cfqq); | |
2202 | ||
c1b707d2 | 2203 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
2204 | } |
2205 | ||
fc46379d | 2206 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 2207 | { |
4ac845a2 | 2208 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 2209 | ioc->ioprio_changed = 0; |
22e2c507 JA |
2210 | } |
2211 | ||
d5036d77 | 2212 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 2213 | pid_t pid, bool is_sync) |
d5036d77 JA |
2214 | { |
2215 | RB_CLEAR_NODE(&cfqq->rb_node); | |
2216 | RB_CLEAR_NODE(&cfqq->p_node); | |
2217 | INIT_LIST_HEAD(&cfqq->fifo); | |
2218 | ||
2219 | atomic_set(&cfqq->ref, 0); | |
2220 | cfqq->cfqd = cfqd; | |
2221 | ||
2222 | cfq_mark_cfqq_prio_changed(cfqq); | |
2223 | ||
2224 | if (is_sync) { | |
2225 | if (!cfq_class_idle(cfqq)) | |
2226 | cfq_mark_cfqq_idle_window(cfqq); | |
2227 | cfq_mark_cfqq_sync(cfqq); | |
2228 | } | |
2229 | cfqq->pid = pid; | |
2230 | } | |
2231 | ||
cdb16e8f VG |
2232 | static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) |
2233 | { | |
2234 | cfqq->cfqg = cfqg; | |
2235 | } | |
2236 | ||
2237 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create) | |
2238 | { | |
2239 | return &cfqd->root_group; | |
2240 | } | |
2241 | ||
22e2c507 | 2242 | static struct cfq_queue * |
a6151c3a | 2243 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, |
fd0928df | 2244 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 2245 | { |
22e2c507 | 2246 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 2247 | struct cfq_io_context *cic; |
cdb16e8f | 2248 | struct cfq_group *cfqg; |
22e2c507 JA |
2249 | |
2250 | retry: | |
cdb16e8f | 2251 | cfqg = cfq_get_cfqg(cfqd, 1); |
4ac845a2 | 2252 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
2253 | /* cic always exists here */ |
2254 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 2255 | |
6118b70b JA |
2256 | /* |
2257 | * Always try a new alloc if we fell back to the OOM cfqq | |
2258 | * originally, since it should just be a temporary situation. | |
2259 | */ | |
2260 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
2261 | cfqq = NULL; | |
22e2c507 JA |
2262 | if (new_cfqq) { |
2263 | cfqq = new_cfqq; | |
2264 | new_cfqq = NULL; | |
2265 | } else if (gfp_mask & __GFP_WAIT) { | |
2266 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 2267 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 2268 | gfp_mask | __GFP_ZERO, |
94f6030c | 2269 | cfqd->queue->node); |
22e2c507 | 2270 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
2271 | if (new_cfqq) |
2272 | goto retry; | |
22e2c507 | 2273 | } else { |
94f6030c CL |
2274 | cfqq = kmem_cache_alloc_node(cfq_pool, |
2275 | gfp_mask | __GFP_ZERO, | |
2276 | cfqd->queue->node); | |
22e2c507 JA |
2277 | } |
2278 | ||
6118b70b JA |
2279 | if (cfqq) { |
2280 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
2281 | cfq_init_prio_data(cfqq, ioc); | |
cdb16e8f | 2282 | cfq_link_cfqq_cfqg(cfqq, cfqg); |
6118b70b JA |
2283 | cfq_log_cfqq(cfqd, cfqq, "alloced"); |
2284 | } else | |
2285 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
2286 | } |
2287 | ||
2288 | if (new_cfqq) | |
2289 | kmem_cache_free(cfq_pool, new_cfqq); | |
2290 | ||
22e2c507 JA |
2291 | return cfqq; |
2292 | } | |
2293 | ||
c2dea2d1 VT |
2294 | static struct cfq_queue ** |
2295 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
2296 | { | |
fe094d98 | 2297 | switch (ioprio_class) { |
c2dea2d1 VT |
2298 | case IOPRIO_CLASS_RT: |
2299 | return &cfqd->async_cfqq[0][ioprio]; | |
2300 | case IOPRIO_CLASS_BE: | |
2301 | return &cfqd->async_cfqq[1][ioprio]; | |
2302 | case IOPRIO_CLASS_IDLE: | |
2303 | return &cfqd->async_idle_cfqq; | |
2304 | default: | |
2305 | BUG(); | |
2306 | } | |
2307 | } | |
2308 | ||
15c31be4 | 2309 | static struct cfq_queue * |
a6151c3a | 2310 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, |
15c31be4 JA |
2311 | gfp_t gfp_mask) |
2312 | { | |
fd0928df JA |
2313 | const int ioprio = task_ioprio(ioc); |
2314 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 2315 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
2316 | struct cfq_queue *cfqq = NULL; |
2317 | ||
c2dea2d1 VT |
2318 | if (!is_sync) { |
2319 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
2320 | cfqq = *async_cfqq; | |
2321 | } | |
2322 | ||
6118b70b | 2323 | if (!cfqq) |
fd0928df | 2324 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
2325 | |
2326 | /* | |
2327 | * pin the queue now that it's allocated, scheduler exit will prune it | |
2328 | */ | |
c2dea2d1 | 2329 | if (!is_sync && !(*async_cfqq)) { |
15c31be4 | 2330 | atomic_inc(&cfqq->ref); |
c2dea2d1 | 2331 | *async_cfqq = cfqq; |
15c31be4 JA |
2332 | } |
2333 | ||
2334 | atomic_inc(&cfqq->ref); | |
2335 | return cfqq; | |
2336 | } | |
2337 | ||
498d3aa2 JA |
2338 | /* |
2339 | * We drop cfq io contexts lazily, so we may find a dead one. | |
2340 | */ | |
dbecf3ab | 2341 | static void |
4ac845a2 JA |
2342 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
2343 | struct cfq_io_context *cic) | |
dbecf3ab | 2344 | { |
4ac845a2 JA |
2345 | unsigned long flags; |
2346 | ||
fc46379d | 2347 | WARN_ON(!list_empty(&cic->queue_list)); |
597bc485 | 2348 | |
4ac845a2 JA |
2349 | spin_lock_irqsave(&ioc->lock, flags); |
2350 | ||
4faa3c81 | 2351 | BUG_ON(ioc->ioc_data == cic); |
597bc485 | 2352 | |
4ac845a2 | 2353 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); |
ffc4e759 | 2354 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2355 | spin_unlock_irqrestore(&ioc->lock, flags); |
2356 | ||
2357 | cfq_cic_free(cic); | |
dbecf3ab OH |
2358 | } |
2359 | ||
e2d74ac0 | 2360 | static struct cfq_io_context * |
4ac845a2 | 2361 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 2362 | { |
e2d74ac0 | 2363 | struct cfq_io_context *cic; |
d6de8be7 | 2364 | unsigned long flags; |
4ac845a2 | 2365 | void *k; |
e2d74ac0 | 2366 | |
91fac317 VT |
2367 | if (unlikely(!ioc)) |
2368 | return NULL; | |
2369 | ||
d6de8be7 JA |
2370 | rcu_read_lock(); |
2371 | ||
597bc485 JA |
2372 | /* |
2373 | * we maintain a last-hit cache, to avoid browsing over the tree | |
2374 | */ | |
4ac845a2 | 2375 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
2376 | if (cic && cic->key == cfqd) { |
2377 | rcu_read_unlock(); | |
597bc485 | 2378 | return cic; |
d6de8be7 | 2379 | } |
597bc485 | 2380 | |
4ac845a2 | 2381 | do { |
4ac845a2 JA |
2382 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); |
2383 | rcu_read_unlock(); | |
2384 | if (!cic) | |
2385 | break; | |
be3b0753 OH |
2386 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
2387 | k = cic->key; | |
2388 | if (unlikely(!k)) { | |
4ac845a2 | 2389 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 2390 | rcu_read_lock(); |
4ac845a2 | 2391 | continue; |
dbecf3ab | 2392 | } |
e2d74ac0 | 2393 | |
d6de8be7 | 2394 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 2395 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 2396 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
2397 | break; |
2398 | } while (1); | |
e2d74ac0 | 2399 | |
4ac845a2 | 2400 | return cic; |
e2d74ac0 JA |
2401 | } |
2402 | ||
4ac845a2 JA |
2403 | /* |
2404 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
2405 | * the process specific cfq io context when entered from the block layer. | |
2406 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
2407 | */ | |
febffd61 JA |
2408 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
2409 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 2410 | { |
0261d688 | 2411 | unsigned long flags; |
4ac845a2 | 2412 | int ret; |
e2d74ac0 | 2413 | |
4ac845a2 JA |
2414 | ret = radix_tree_preload(gfp_mask); |
2415 | if (!ret) { | |
2416 | cic->ioc = ioc; | |
2417 | cic->key = cfqd; | |
e2d74ac0 | 2418 | |
4ac845a2 JA |
2419 | spin_lock_irqsave(&ioc->lock, flags); |
2420 | ret = radix_tree_insert(&ioc->radix_root, | |
2421 | (unsigned long) cfqd, cic); | |
ffc4e759 JA |
2422 | if (!ret) |
2423 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 2424 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 2425 | |
4ac845a2 JA |
2426 | radix_tree_preload_end(); |
2427 | ||
2428 | if (!ret) { | |
2429 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
2430 | list_add(&cic->queue_list, &cfqd->cic_list); | |
2431 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2432 | } | |
e2d74ac0 JA |
2433 | } |
2434 | ||
4ac845a2 JA |
2435 | if (ret) |
2436 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 2437 | |
4ac845a2 | 2438 | return ret; |
e2d74ac0 JA |
2439 | } |
2440 | ||
1da177e4 LT |
2441 | /* |
2442 | * Setup general io context and cfq io context. There can be several cfq | |
2443 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 2444 | * than one device managed by cfq. |
1da177e4 LT |
2445 | */ |
2446 | static struct cfq_io_context * | |
e2d74ac0 | 2447 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2448 | { |
22e2c507 | 2449 | struct io_context *ioc = NULL; |
1da177e4 | 2450 | struct cfq_io_context *cic; |
1da177e4 | 2451 | |
22e2c507 | 2452 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 2453 | |
b5deef90 | 2454 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
2455 | if (!ioc) |
2456 | return NULL; | |
2457 | ||
4ac845a2 | 2458 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
2459 | if (cic) |
2460 | goto out; | |
1da177e4 | 2461 | |
e2d74ac0 JA |
2462 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
2463 | if (cic == NULL) | |
2464 | goto err; | |
1da177e4 | 2465 | |
4ac845a2 JA |
2466 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
2467 | goto err_free; | |
2468 | ||
1da177e4 | 2469 | out: |
fc46379d JA |
2470 | smp_read_barrier_depends(); |
2471 | if (unlikely(ioc->ioprio_changed)) | |
2472 | cfq_ioc_set_ioprio(ioc); | |
2473 | ||
1da177e4 | 2474 | return cic; |
4ac845a2 JA |
2475 | err_free: |
2476 | cfq_cic_free(cic); | |
1da177e4 LT |
2477 | err: |
2478 | put_io_context(ioc); | |
2479 | return NULL; | |
2480 | } | |
2481 | ||
22e2c507 JA |
2482 | static void |
2483 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 2484 | { |
aaf1228d JA |
2485 | unsigned long elapsed = jiffies - cic->last_end_request; |
2486 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 2487 | |
22e2c507 JA |
2488 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
2489 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
2490 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
2491 | } | |
1da177e4 | 2492 | |
206dc69b | 2493 | static void |
b2c18e1e | 2494 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
6d048f53 | 2495 | struct request *rq) |
206dc69b JA |
2496 | { |
2497 | sector_t sdist; | |
2498 | u64 total; | |
2499 | ||
b2c18e1e | 2500 | if (!cfqq->last_request_pos) |
4d00aa47 | 2501 | sdist = 0; |
b2c18e1e JM |
2502 | else if (cfqq->last_request_pos < blk_rq_pos(rq)) |
2503 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
206dc69b | 2504 | else |
b2c18e1e | 2505 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); |
206dc69b JA |
2506 | |
2507 | /* | |
2508 | * Don't allow the seek distance to get too large from the | |
2509 | * odd fragment, pagein, etc | |
2510 | */ | |
b2c18e1e JM |
2511 | if (cfqq->seek_samples <= 60) /* second&third seek */ |
2512 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024); | |
206dc69b | 2513 | else |
b2c18e1e | 2514 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64); |
206dc69b | 2515 | |
b2c18e1e JM |
2516 | cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8; |
2517 | cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8; | |
2518 | total = cfqq->seek_total + (cfqq->seek_samples/2); | |
2519 | do_div(total, cfqq->seek_samples); | |
2520 | cfqq->seek_mean = (sector_t)total; | |
e6c5bc73 JM |
2521 | |
2522 | /* | |
2523 | * If this cfqq is shared between multiple processes, check to | |
2524 | * make sure that those processes are still issuing I/Os within | |
2525 | * the mean seek distance. If not, it may be time to break the | |
2526 | * queues apart again. | |
2527 | */ | |
2528 | if (cfq_cfqq_coop(cfqq)) { | |
2529 | if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start) | |
2530 | cfqq->seeky_start = jiffies; | |
2531 | else if (!CFQQ_SEEKY(cfqq)) | |
2532 | cfqq->seeky_start = 0; | |
2533 | } | |
206dc69b | 2534 | } |
1da177e4 | 2535 | |
22e2c507 JA |
2536 | /* |
2537 | * Disable idle window if the process thinks too long or seeks so much that | |
2538 | * it doesn't matter | |
2539 | */ | |
2540 | static void | |
2541 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2542 | struct cfq_io_context *cic) | |
2543 | { | |
7b679138 | 2544 | int old_idle, enable_idle; |
1be92f2f | 2545 | |
0871714e JA |
2546 | /* |
2547 | * Don't idle for async or idle io prio class | |
2548 | */ | |
2549 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
2550 | return; |
2551 | ||
c265a7f4 | 2552 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 2553 | |
76280aff CZ |
2554 | if (cfqq->queued[0] + cfqq->queued[1] >= 4) |
2555 | cfq_mark_cfqq_deep(cfqq); | |
2556 | ||
66dac98e | 2557 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || |
76280aff CZ |
2558 | (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples) |
2559 | && CFQQ_SEEKY(cfqq))) | |
22e2c507 JA |
2560 | enable_idle = 0; |
2561 | else if (sample_valid(cic->ttime_samples)) { | |
718eee05 | 2562 | if (cic->ttime_mean > cfqd->cfq_slice_idle) |
22e2c507 JA |
2563 | enable_idle = 0; |
2564 | else | |
2565 | enable_idle = 1; | |
1da177e4 LT |
2566 | } |
2567 | ||
7b679138 JA |
2568 | if (old_idle != enable_idle) { |
2569 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
2570 | if (enable_idle) | |
2571 | cfq_mark_cfqq_idle_window(cfqq); | |
2572 | else | |
2573 | cfq_clear_cfqq_idle_window(cfqq); | |
2574 | } | |
22e2c507 | 2575 | } |
1da177e4 | 2576 | |
22e2c507 JA |
2577 | /* |
2578 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
2579 | * no or if we aren't sure, a 1 will cause a preempt. | |
2580 | */ | |
a6151c3a | 2581 | static bool |
22e2c507 | 2582 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, |
5e705374 | 2583 | struct request *rq) |
22e2c507 | 2584 | { |
6d048f53 | 2585 | struct cfq_queue *cfqq; |
22e2c507 | 2586 | |
6d048f53 JA |
2587 | cfqq = cfqd->active_queue; |
2588 | if (!cfqq) | |
a6151c3a | 2589 | return false; |
22e2c507 | 2590 | |
6d048f53 | 2591 | if (cfq_slice_used(cfqq)) |
a6151c3a | 2592 | return true; |
6d048f53 JA |
2593 | |
2594 | if (cfq_class_idle(new_cfqq)) | |
a6151c3a | 2595 | return false; |
22e2c507 JA |
2596 | |
2597 | if (cfq_class_idle(cfqq)) | |
a6151c3a | 2598 | return true; |
1e3335de | 2599 | |
f04a6424 | 2600 | /* Allow preemption only if we are idling on sync-noidle tree */ |
e4a22919 CZ |
2601 | if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && |
2602 | cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && | |
f04a6424 VG |
2603 | new_cfqq->service_tree->count == 2 && |
2604 | RB_EMPTY_ROOT(&cfqq->sort_list)) | |
718eee05 CZ |
2605 | return true; |
2606 | ||
374f84ac JA |
2607 | /* |
2608 | * if the new request is sync, but the currently running queue is | |
2609 | * not, let the sync request have priority. | |
2610 | */ | |
5e705374 | 2611 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
a6151c3a | 2612 | return true; |
1e3335de | 2613 | |
374f84ac JA |
2614 | /* |
2615 | * So both queues are sync. Let the new request get disk time if | |
2616 | * it's a metadata request and the current queue is doing regular IO. | |
2617 | */ | |
2618 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
e6ec4fe2 | 2619 | return true; |
22e2c507 | 2620 | |
3a9a3f6c DS |
2621 | /* |
2622 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
2623 | */ | |
2624 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
a6151c3a | 2625 | return true; |
3a9a3f6c | 2626 | |
1e3335de | 2627 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
a6151c3a | 2628 | return false; |
1e3335de JA |
2629 | |
2630 | /* | |
2631 | * if this request is as-good as one we would expect from the | |
2632 | * current cfqq, let it preempt | |
2633 | */ | |
e00ef799 | 2634 | if (cfq_rq_close(cfqd, cfqq, rq)) |
a6151c3a | 2635 | return true; |
1e3335de | 2636 | |
a6151c3a | 2637 | return false; |
22e2c507 JA |
2638 | } |
2639 | ||
2640 | /* | |
2641 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
2642 | * let it have half of its nominal slice. | |
2643 | */ | |
2644 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2645 | { | |
7b679138 | 2646 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
6084cdda | 2647 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 2648 | |
bf572256 JA |
2649 | /* |
2650 | * Put the new queue at the front of the of the current list, | |
2651 | * so we know that it will be selected next. | |
2652 | */ | |
2653 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
2654 | |
2655 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 2656 | |
44f7c160 JA |
2657 | cfqq->slice_end = 0; |
2658 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
2659 | } |
2660 | ||
22e2c507 | 2661 | /* |
5e705374 | 2662 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
2663 | * something we should do about it |
2664 | */ | |
2665 | static void | |
5e705374 JA |
2666 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
2667 | struct request *rq) | |
22e2c507 | 2668 | { |
5e705374 | 2669 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 2670 | |
45333d5a | 2671 | cfqd->rq_queued++; |
374f84ac JA |
2672 | if (rq_is_meta(rq)) |
2673 | cfqq->meta_pending++; | |
2674 | ||
9c2c38a1 | 2675 | cfq_update_io_thinktime(cfqd, cic); |
b2c18e1e | 2676 | cfq_update_io_seektime(cfqd, cfqq, rq); |
9c2c38a1 JA |
2677 | cfq_update_idle_window(cfqd, cfqq, cic); |
2678 | ||
b2c18e1e | 2679 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
2680 | |
2681 | if (cfqq == cfqd->active_queue) { | |
2682 | /* | |
b029195d JA |
2683 | * Remember that we saw a request from this process, but |
2684 | * don't start queuing just yet. Otherwise we risk seeing lots | |
2685 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
2686 | * and merging. If the request is already larger than a single |
2687 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
2688 | * merging is already done. Ditto for a busy system that |
2689 | * has other work pending, don't risk delaying until the | |
2690 | * idle timer unplug to continue working. | |
22e2c507 | 2691 | */ |
d6ceb25e | 2692 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
2693 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
2694 | cfqd->busy_queues > 1) { | |
d6ceb25e | 2695 | del_timer(&cfqd->idle_slice_timer); |
bf791937 VG |
2696 | __blk_run_queue(cfqd->queue); |
2697 | } else | |
2698 | cfq_mark_cfqq_must_dispatch(cfqq); | |
d6ceb25e | 2699 | } |
5e705374 | 2700 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
2701 | /* |
2702 | * not the active queue - expire current slice if it is | |
2703 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
2704 | * has some old slice time left and is of higher priority or |
2705 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
2706 | */ |
2707 | cfq_preempt_queue(cfqd, cfqq); | |
a7f55792 | 2708 | __blk_run_queue(cfqd->queue); |
22e2c507 | 2709 | } |
1da177e4 LT |
2710 | } |
2711 | ||
165125e1 | 2712 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2713 | { |
b4878f24 | 2714 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 2715 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 2716 | |
7b679138 | 2717 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 2718 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 2719 | |
30996f40 | 2720 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); |
22e2c507 | 2721 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
aa6f6a3d | 2722 | cfq_add_rq_rb(rq); |
22e2c507 | 2723 | |
5e705374 | 2724 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
2725 | } |
2726 | ||
45333d5a AC |
2727 | /* |
2728 | * Update hw_tag based on peak queue depth over 50 samples under | |
2729 | * sufficient load. | |
2730 | */ | |
2731 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
2732 | { | |
1a1238a7 SL |
2733 | struct cfq_queue *cfqq = cfqd->active_queue; |
2734 | ||
e459dd08 CZ |
2735 | if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth) |
2736 | cfqd->hw_tag_est_depth = rq_in_driver(cfqd); | |
2737 | ||
2738 | if (cfqd->hw_tag == 1) | |
2739 | return; | |
45333d5a AC |
2740 | |
2741 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
5ad531db | 2742 | rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
2743 | return; |
2744 | ||
1a1238a7 SL |
2745 | /* |
2746 | * If active queue hasn't enough requests and can idle, cfq might not | |
2747 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
2748 | * case | |
2749 | */ | |
2750 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
2751 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
2752 | CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN) | |
2753 | return; | |
2754 | ||
45333d5a AC |
2755 | if (cfqd->hw_tag_samples++ < 50) |
2756 | return; | |
2757 | ||
e459dd08 | 2758 | if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
2759 | cfqd->hw_tag = 1; |
2760 | else | |
2761 | cfqd->hw_tag = 0; | |
45333d5a AC |
2762 | } |
2763 | ||
165125e1 | 2764 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2765 | { |
5e705374 | 2766 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 2767 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 2768 | const int sync = rq_is_sync(rq); |
b4878f24 | 2769 | unsigned long now; |
1da177e4 | 2770 | |
b4878f24 | 2771 | now = jiffies; |
7b679138 | 2772 | cfq_log_cfqq(cfqd, cfqq, "complete"); |
1da177e4 | 2773 | |
45333d5a AC |
2774 | cfq_update_hw_tag(cfqd); |
2775 | ||
5ad531db | 2776 | WARN_ON(!cfqd->rq_in_driver[sync]); |
6d048f53 | 2777 | WARN_ON(!cfqq->dispatched); |
5ad531db | 2778 | cfqd->rq_in_driver[sync]--; |
6d048f53 | 2779 | cfqq->dispatched--; |
1da177e4 | 2780 | |
3ed9a296 JA |
2781 | if (cfq_cfqq_sync(cfqq)) |
2782 | cfqd->sync_flight--; | |
2783 | ||
365722bb | 2784 | if (sync) { |
5e705374 | 2785 | RQ_CIC(rq)->last_end_request = now; |
365722bb VG |
2786 | cfqd->last_end_sync_rq = now; |
2787 | } | |
caaa5f9f JA |
2788 | |
2789 | /* | |
2790 | * If this is the active queue, check if it needs to be expired, | |
2791 | * or if we want to idle in case it has no pending requests. | |
2792 | */ | |
2793 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
2794 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
2795 | ||
44f7c160 JA |
2796 | if (cfq_cfqq_slice_new(cfqq)) { |
2797 | cfq_set_prio_slice(cfqd, cfqq); | |
2798 | cfq_clear_cfqq_slice_new(cfqq); | |
2799 | } | |
a36e71f9 | 2800 | /* |
8e550632 CZ |
2801 | * Idling is not enabled on: |
2802 | * - expired queues | |
2803 | * - idle-priority queues | |
2804 | * - async queues | |
2805 | * - queues with still some requests queued | |
2806 | * - when there is a close cooperator | |
a36e71f9 | 2807 | */ |
0871714e | 2808 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
6084cdda | 2809 | cfq_slice_expired(cfqd, 1); |
8e550632 CZ |
2810 | else if (sync && cfqq_empty && |
2811 | !cfq_close_cooperator(cfqd, cfqq)) { | |
2812 | cfqd->noidle_tree_requires_idle |= !rq_noidle(rq); | |
2813 | /* | |
2814 | * Idling is enabled for SYNC_WORKLOAD. | |
2815 | * SYNC_NOIDLE_WORKLOAD idles at the end of the tree | |
2816 | * only if we processed at least one !rq_noidle request | |
2817 | */ | |
2818 | if (cfqd->serving_type == SYNC_WORKLOAD | |
2819 | || cfqd->noidle_tree_requires_idle) | |
2820 | cfq_arm_slice_timer(cfqd); | |
2821 | } | |
caaa5f9f | 2822 | } |
6d048f53 | 2823 | |
5ad531db | 2824 | if (!rq_in_driver(cfqd)) |
23e018a1 | 2825 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
2826 | } |
2827 | ||
22e2c507 JA |
2828 | /* |
2829 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
2830 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
2831 | */ | |
2832 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 2833 | { |
22e2c507 JA |
2834 | if (has_fs_excl()) { |
2835 | /* | |
2836 | * boost idle prio on transactions that would lock out other | |
2837 | * users of the filesystem | |
2838 | */ | |
2839 | if (cfq_class_idle(cfqq)) | |
2840 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2841 | if (cfqq->ioprio > IOPRIO_NORM) | |
2842 | cfqq->ioprio = IOPRIO_NORM; | |
2843 | } else { | |
2844 | /* | |
dddb7451 | 2845 | * unboost the queue (if needed) |
22e2c507 | 2846 | */ |
dddb7451 CZ |
2847 | cfqq->ioprio_class = cfqq->org_ioprio_class; |
2848 | cfqq->ioprio = cfqq->org_ioprio; | |
22e2c507 | 2849 | } |
22e2c507 | 2850 | } |
1da177e4 | 2851 | |
89850f7e | 2852 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 2853 | { |
1b379d8d | 2854 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 2855 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 2856 | return ELV_MQUEUE_MUST; |
3b18152c | 2857 | } |
1da177e4 | 2858 | |
22e2c507 | 2859 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
2860 | } |
2861 | ||
165125e1 | 2862 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
2863 | { |
2864 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2865 | struct task_struct *tsk = current; | |
91fac317 | 2866 | struct cfq_io_context *cic; |
22e2c507 JA |
2867 | struct cfq_queue *cfqq; |
2868 | ||
2869 | /* | |
2870 | * don't force setup of a queue from here, as a call to may_queue | |
2871 | * does not necessarily imply that a request actually will be queued. | |
2872 | * so just lookup a possibly existing queue, or return 'may queue' | |
2873 | * if that fails | |
2874 | */ | |
4ac845a2 | 2875 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
2876 | if (!cic) |
2877 | return ELV_MQUEUE_MAY; | |
2878 | ||
b0b78f81 | 2879 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 2880 | if (cfqq) { |
fd0928df | 2881 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
2882 | cfq_prio_boost(cfqq); |
2883 | ||
89850f7e | 2884 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
2885 | } |
2886 | ||
2887 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
2888 | } |
2889 | ||
1da177e4 LT |
2890 | /* |
2891 | * queue lock held here | |
2892 | */ | |
bb37b94c | 2893 | static void cfq_put_request(struct request *rq) |
1da177e4 | 2894 | { |
5e705374 | 2895 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 2896 | |
5e705374 | 2897 | if (cfqq) { |
22e2c507 | 2898 | const int rw = rq_data_dir(rq); |
1da177e4 | 2899 | |
22e2c507 JA |
2900 | BUG_ON(!cfqq->allocated[rw]); |
2901 | cfqq->allocated[rw]--; | |
1da177e4 | 2902 | |
5e705374 | 2903 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 2904 | |
1da177e4 | 2905 | rq->elevator_private = NULL; |
5e705374 | 2906 | rq->elevator_private2 = NULL; |
1da177e4 | 2907 | |
1da177e4 LT |
2908 | cfq_put_queue(cfqq); |
2909 | } | |
2910 | } | |
2911 | ||
df5fe3e8 JM |
2912 | static struct cfq_queue * |
2913 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, | |
2914 | struct cfq_queue *cfqq) | |
2915 | { | |
2916 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
2917 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
b3b6d040 | 2918 | cfq_mark_cfqq_coop(cfqq->new_cfqq); |
df5fe3e8 JM |
2919 | cfq_put_queue(cfqq); |
2920 | return cic_to_cfqq(cic, 1); | |
2921 | } | |
2922 | ||
e6c5bc73 JM |
2923 | static int should_split_cfqq(struct cfq_queue *cfqq) |
2924 | { | |
2925 | if (cfqq->seeky_start && | |
2926 | time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT)) | |
2927 | return 1; | |
2928 | return 0; | |
2929 | } | |
2930 | ||
2931 | /* | |
2932 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
2933 | * was the last process referring to said cfqq. | |
2934 | */ | |
2935 | static struct cfq_queue * | |
2936 | split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) | |
2937 | { | |
2938 | if (cfqq_process_refs(cfqq) == 1) { | |
2939 | cfqq->seeky_start = 0; | |
2940 | cfqq->pid = current->pid; | |
2941 | cfq_clear_cfqq_coop(cfqq); | |
2942 | return cfqq; | |
2943 | } | |
2944 | ||
2945 | cic_set_cfqq(cic, NULL, 1); | |
2946 | cfq_put_queue(cfqq); | |
2947 | return NULL; | |
2948 | } | |
1da177e4 | 2949 | /* |
22e2c507 | 2950 | * Allocate cfq data structures associated with this request. |
1da177e4 | 2951 | */ |
22e2c507 | 2952 | static int |
165125e1 | 2953 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
2954 | { |
2955 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2956 | struct cfq_io_context *cic; | |
2957 | const int rw = rq_data_dir(rq); | |
a6151c3a | 2958 | const bool is_sync = rq_is_sync(rq); |
22e2c507 | 2959 | struct cfq_queue *cfqq; |
1da177e4 LT |
2960 | unsigned long flags; |
2961 | ||
2962 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
2963 | ||
e2d74ac0 | 2964 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 2965 | |
1da177e4 LT |
2966 | spin_lock_irqsave(q->queue_lock, flags); |
2967 | ||
22e2c507 JA |
2968 | if (!cic) |
2969 | goto queue_fail; | |
2970 | ||
e6c5bc73 | 2971 | new_queue: |
91fac317 | 2972 | cfqq = cic_to_cfqq(cic, is_sync); |
32f2e807 | 2973 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { |
fd0928df | 2974 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 | 2975 | cic_set_cfqq(cic, cfqq, is_sync); |
df5fe3e8 | 2976 | } else { |
e6c5bc73 JM |
2977 | /* |
2978 | * If the queue was seeky for too long, break it apart. | |
2979 | */ | |
2980 | if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) { | |
2981 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); | |
2982 | cfqq = split_cfqq(cic, cfqq); | |
2983 | if (!cfqq) | |
2984 | goto new_queue; | |
2985 | } | |
2986 | ||
df5fe3e8 JM |
2987 | /* |
2988 | * Check to see if this queue is scheduled to merge with | |
2989 | * another, closely cooperating queue. The merging of | |
2990 | * queues happens here as it must be done in process context. | |
2991 | * The reference on new_cfqq was taken in merge_cfqqs. | |
2992 | */ | |
2993 | if (cfqq->new_cfqq) | |
2994 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
91fac317 | 2995 | } |
1da177e4 LT |
2996 | |
2997 | cfqq->allocated[rw]++; | |
22e2c507 | 2998 | atomic_inc(&cfqq->ref); |
1da177e4 | 2999 | |
5e705374 | 3000 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 3001 | |
5e705374 JA |
3002 | rq->elevator_private = cic; |
3003 | rq->elevator_private2 = cfqq; | |
3004 | return 0; | |
1da177e4 | 3005 | |
22e2c507 JA |
3006 | queue_fail: |
3007 | if (cic) | |
3008 | put_io_context(cic->ioc); | |
89850f7e | 3009 | |
23e018a1 | 3010 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 3011 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 3012 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
3013 | return 1; |
3014 | } | |
3015 | ||
65f27f38 | 3016 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 3017 | { |
65f27f38 | 3018 | struct cfq_data *cfqd = |
23e018a1 | 3019 | container_of(work, struct cfq_data, unplug_work); |
165125e1 | 3020 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3021 | |
40bb54d1 | 3022 | spin_lock_irq(q->queue_lock); |
a7f55792 | 3023 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 3024 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
3025 | } |
3026 | ||
3027 | /* | |
3028 | * Timer running if the active_queue is currently idling inside its time slice | |
3029 | */ | |
3030 | static void cfq_idle_slice_timer(unsigned long data) | |
3031 | { | |
3032 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
3033 | struct cfq_queue *cfqq; | |
3034 | unsigned long flags; | |
3c6bd2f8 | 3035 | int timed_out = 1; |
22e2c507 | 3036 | |
7b679138 JA |
3037 | cfq_log(cfqd, "idle timer fired"); |
3038 | ||
22e2c507 JA |
3039 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
3040 | ||
fe094d98 JA |
3041 | cfqq = cfqd->active_queue; |
3042 | if (cfqq) { | |
3c6bd2f8 JA |
3043 | timed_out = 0; |
3044 | ||
b029195d JA |
3045 | /* |
3046 | * We saw a request before the queue expired, let it through | |
3047 | */ | |
3048 | if (cfq_cfqq_must_dispatch(cfqq)) | |
3049 | goto out_kick; | |
3050 | ||
22e2c507 JA |
3051 | /* |
3052 | * expired | |
3053 | */ | |
44f7c160 | 3054 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
3055 | goto expire; |
3056 | ||
3057 | /* | |
3058 | * only expire and reinvoke request handler, if there are | |
3059 | * other queues with pending requests | |
3060 | */ | |
caaa5f9f | 3061 | if (!cfqd->busy_queues) |
22e2c507 | 3062 | goto out_cont; |
22e2c507 JA |
3063 | |
3064 | /* | |
3065 | * not expired and it has a request pending, let it dispatch | |
3066 | */ | |
75e50984 | 3067 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 3068 | goto out_kick; |
76280aff CZ |
3069 | |
3070 | /* | |
3071 | * Queue depth flag is reset only when the idle didn't succeed | |
3072 | */ | |
3073 | cfq_clear_cfqq_deep(cfqq); | |
22e2c507 JA |
3074 | } |
3075 | expire: | |
6084cdda | 3076 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 3077 | out_kick: |
23e018a1 | 3078 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
3079 | out_cont: |
3080 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
3081 | } | |
3082 | ||
3b18152c JA |
3083 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
3084 | { | |
3085 | del_timer_sync(&cfqd->idle_slice_timer); | |
23e018a1 | 3086 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 3087 | } |
22e2c507 | 3088 | |
c2dea2d1 VT |
3089 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
3090 | { | |
3091 | int i; | |
3092 | ||
3093 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
3094 | if (cfqd->async_cfqq[0][i]) | |
3095 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
3096 | if (cfqd->async_cfqq[1][i]) | |
3097 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 3098 | } |
2389d1ef ON |
3099 | |
3100 | if (cfqd->async_idle_cfqq) | |
3101 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
3102 | } |
3103 | ||
b374d18a | 3104 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 3105 | { |
22e2c507 | 3106 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 3107 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3108 | |
3b18152c | 3109 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 3110 | |
d9ff4187 | 3111 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 3112 | |
d9ff4187 | 3113 | if (cfqd->active_queue) |
6084cdda | 3114 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
3115 | |
3116 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
3117 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
3118 | struct cfq_io_context, | |
3119 | queue_list); | |
89850f7e JA |
3120 | |
3121 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 3122 | } |
e2d74ac0 | 3123 | |
c2dea2d1 | 3124 | cfq_put_async_queues(cfqd); |
15c31be4 | 3125 | |
d9ff4187 | 3126 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
3127 | |
3128 | cfq_shutdown_timer_wq(cfqd); | |
3129 | ||
a90d742e | 3130 | kfree(cfqd); |
1da177e4 LT |
3131 | } |
3132 | ||
165125e1 | 3133 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
3134 | { |
3135 | struct cfq_data *cfqd; | |
718eee05 | 3136 | int i, j; |
cdb16e8f | 3137 | struct cfq_group *cfqg; |
615f0259 | 3138 | struct cfq_rb_root *st; |
1da177e4 | 3139 | |
94f6030c | 3140 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 3141 | if (!cfqd) |
bc1c1169 | 3142 | return NULL; |
1da177e4 | 3143 | |
1fa8f6d6 VG |
3144 | /* Init root service tree */ |
3145 | cfqd->grp_service_tree = CFQ_RB_ROOT; | |
3146 | ||
cdb16e8f VG |
3147 | /* Init root group */ |
3148 | cfqg = &cfqd->root_group; | |
615f0259 VG |
3149 | for_each_cfqg_st(cfqg, i, j, st) |
3150 | *st = CFQ_RB_ROOT; | |
1fa8f6d6 | 3151 | RB_CLEAR_NODE(&cfqg->rb_node); |
26a2ac00 JA |
3152 | |
3153 | /* | |
3154 | * Not strictly needed (since RB_ROOT just clears the node and we | |
3155 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
3156 | * to add magic to the rb code | |
3157 | */ | |
3158 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
3159 | cfqd->prio_trees[i] = RB_ROOT; | |
3160 | ||
6118b70b JA |
3161 | /* |
3162 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
3163 | * Grab a permanent reference to it, so that the normal code flow | |
3164 | * will not attempt to free it. | |
3165 | */ | |
3166 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
3167 | atomic_inc(&cfqd->oom_cfqq.ref); | |
cdb16e8f | 3168 | cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group); |
6118b70b | 3169 | |
d9ff4187 | 3170 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 3171 | |
1da177e4 | 3172 | cfqd->queue = q; |
1da177e4 | 3173 | |
22e2c507 JA |
3174 | init_timer(&cfqd->idle_slice_timer); |
3175 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
3176 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
3177 | ||
23e018a1 | 3178 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 3179 | |
1da177e4 | 3180 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
3181 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
3182 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
3183 | cfqd->cfq_back_max = cfq_back_max; |
3184 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
3185 | cfqd->cfq_slice[0] = cfq_slice_async; |
3186 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
3187 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
3188 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
963b72fc | 3189 | cfqd->cfq_latency = 1; |
e459dd08 | 3190 | cfqd->hw_tag = -1; |
365722bb | 3191 | cfqd->last_end_sync_rq = jiffies; |
bc1c1169 | 3192 | return cfqd; |
1da177e4 LT |
3193 | } |
3194 | ||
3195 | static void cfq_slab_kill(void) | |
3196 | { | |
d6de8be7 JA |
3197 | /* |
3198 | * Caller already ensured that pending RCU callbacks are completed, | |
3199 | * so we should have no busy allocations at this point. | |
3200 | */ | |
1da177e4 LT |
3201 | if (cfq_pool) |
3202 | kmem_cache_destroy(cfq_pool); | |
3203 | if (cfq_ioc_pool) | |
3204 | kmem_cache_destroy(cfq_ioc_pool); | |
3205 | } | |
3206 | ||
3207 | static int __init cfq_slab_setup(void) | |
3208 | { | |
0a31bd5f | 3209 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
3210 | if (!cfq_pool) |
3211 | goto fail; | |
3212 | ||
34e6bbf2 | 3213 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
3214 | if (!cfq_ioc_pool) |
3215 | goto fail; | |
3216 | ||
3217 | return 0; | |
3218 | fail: | |
3219 | cfq_slab_kill(); | |
3220 | return -ENOMEM; | |
3221 | } | |
3222 | ||
1da177e4 LT |
3223 | /* |
3224 | * sysfs parts below --> | |
3225 | */ | |
1da177e4 LT |
3226 | static ssize_t |
3227 | cfq_var_show(unsigned int var, char *page) | |
3228 | { | |
3229 | return sprintf(page, "%d\n", var); | |
3230 | } | |
3231 | ||
3232 | static ssize_t | |
3233 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
3234 | { | |
3235 | char *p = (char *) page; | |
3236 | ||
3237 | *var = simple_strtoul(p, &p, 10); | |
3238 | return count; | |
3239 | } | |
3240 | ||
1da177e4 | 3241 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 3242 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 3243 | { \ |
3d1ab40f | 3244 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3245 | unsigned int __data = __VAR; \ |
3246 | if (__CONV) \ | |
3247 | __data = jiffies_to_msecs(__data); \ | |
3248 | return cfq_var_show(__data, (page)); \ | |
3249 | } | |
3250 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
3251 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
3252 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
3253 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
3254 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
3255 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
3256 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
3257 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
3258 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
963b72fc | 3259 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); |
1da177e4 LT |
3260 | #undef SHOW_FUNCTION |
3261 | ||
3262 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 3263 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 3264 | { \ |
3d1ab40f | 3265 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3266 | unsigned int __data; \ |
3267 | int ret = cfq_var_store(&__data, (page), count); \ | |
3268 | if (__data < (MIN)) \ | |
3269 | __data = (MIN); \ | |
3270 | else if (__data > (MAX)) \ | |
3271 | __data = (MAX); \ | |
3272 | if (__CONV) \ | |
3273 | *(__PTR) = msecs_to_jiffies(__data); \ | |
3274 | else \ | |
3275 | *(__PTR) = __data; \ | |
3276 | return ret; \ | |
3277 | } | |
3278 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
3279 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
3280 | UINT_MAX, 1); | |
3281 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
3282 | UINT_MAX, 1); | |
e572ec7e | 3283 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
3284 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
3285 | UINT_MAX, 0); | |
22e2c507 JA |
3286 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
3287 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
3288 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
3289 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
3290 | UINT_MAX, 0); | |
963b72fc | 3291 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); |
1da177e4 LT |
3292 | #undef STORE_FUNCTION |
3293 | ||
e572ec7e AV |
3294 | #define CFQ_ATTR(name) \ |
3295 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
3296 | ||
3297 | static struct elv_fs_entry cfq_attrs[] = { | |
3298 | CFQ_ATTR(quantum), | |
e572ec7e AV |
3299 | CFQ_ATTR(fifo_expire_sync), |
3300 | CFQ_ATTR(fifo_expire_async), | |
3301 | CFQ_ATTR(back_seek_max), | |
3302 | CFQ_ATTR(back_seek_penalty), | |
3303 | CFQ_ATTR(slice_sync), | |
3304 | CFQ_ATTR(slice_async), | |
3305 | CFQ_ATTR(slice_async_rq), | |
3306 | CFQ_ATTR(slice_idle), | |
963b72fc | 3307 | CFQ_ATTR(low_latency), |
e572ec7e | 3308 | __ATTR_NULL |
1da177e4 LT |
3309 | }; |
3310 | ||
1da177e4 LT |
3311 | static struct elevator_type iosched_cfq = { |
3312 | .ops = { | |
3313 | .elevator_merge_fn = cfq_merge, | |
3314 | .elevator_merged_fn = cfq_merged_request, | |
3315 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 3316 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 3317 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 3318 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 3319 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
3320 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
3321 | .elevator_queue_empty_fn = cfq_queue_empty, | |
3322 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
3323 | .elevator_former_req_fn = elv_rb_former_request, |
3324 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
3325 | .elevator_set_req_fn = cfq_set_request, |
3326 | .elevator_put_req_fn = cfq_put_request, | |
3327 | .elevator_may_queue_fn = cfq_may_queue, | |
3328 | .elevator_init_fn = cfq_init_queue, | |
3329 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 3330 | .trim = cfq_free_io_context, |
1da177e4 | 3331 | }, |
3d1ab40f | 3332 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
3333 | .elevator_name = "cfq", |
3334 | .elevator_owner = THIS_MODULE, | |
3335 | }; | |
3336 | ||
3337 | static int __init cfq_init(void) | |
3338 | { | |
22e2c507 JA |
3339 | /* |
3340 | * could be 0 on HZ < 1000 setups | |
3341 | */ | |
3342 | if (!cfq_slice_async) | |
3343 | cfq_slice_async = 1; | |
3344 | if (!cfq_slice_idle) | |
3345 | cfq_slice_idle = 1; | |
3346 | ||
1da177e4 LT |
3347 | if (cfq_slab_setup()) |
3348 | return -ENOMEM; | |
3349 | ||
2fdd82bd | 3350 | elv_register(&iosched_cfq); |
1da177e4 | 3351 | |
2fdd82bd | 3352 | return 0; |
1da177e4 LT |
3353 | } |
3354 | ||
3355 | static void __exit cfq_exit(void) | |
3356 | { | |
6e9a4738 | 3357 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 3358 | elv_unregister(&iosched_cfq); |
334e94de | 3359 | ioc_gone = &all_gone; |
fba82272 OH |
3360 | /* ioc_gone's update must be visible before reading ioc_count */ |
3361 | smp_wmb(); | |
d6de8be7 JA |
3362 | |
3363 | /* | |
3364 | * this also protects us from entering cfq_slab_kill() with | |
3365 | * pending RCU callbacks | |
3366 | */ | |
245b2e70 | 3367 | if (elv_ioc_count_read(cfq_ioc_count)) |
9a11b4ed | 3368 | wait_for_completion(&all_gone); |
83521d3e | 3369 | cfq_slab_kill(); |
1da177e4 LT |
3370 | } |
3371 | ||
3372 | module_init(cfq_init); | |
3373 | module_exit(cfq_exit); | |
3374 | ||
3375 | MODULE_AUTHOR("Jens Axboe"); | |
3376 | MODULE_LICENSE("GPL"); | |
3377 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |