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> |
25bc6b07 | 16 | #include "blk-cgroup.h" |
1da177e4 LT |
17 | |
18 | /* | |
19 | * tunables | |
20 | */ | |
fe094d98 JA |
21 | /* max queue in one round of service */ |
22 | static const int cfq_quantum = 4; | |
64100099 | 23 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
fe094d98 JA |
24 | /* maximum backwards seek, in KiB */ |
25 | static const int cfq_back_max = 16 * 1024; | |
26 | /* penalty of a backwards seek */ | |
27 | static const int cfq_back_penalty = 2; | |
64100099 | 28 | static const int cfq_slice_sync = HZ / 10; |
3b18152c | 29 | static int cfq_slice_async = HZ / 25; |
64100099 | 30 | static const int cfq_slice_async_rq = 2; |
caaa5f9f | 31 | static int cfq_slice_idle = HZ / 125; |
5db5d642 CZ |
32 | static const int cfq_target_latency = HZ * 3/10; /* 300 ms */ |
33 | static const int cfq_hist_divisor = 4; | |
22e2c507 | 34 | |
d9e7620e | 35 | /* |
0871714e | 36 | * offset from end of service tree |
d9e7620e | 37 | */ |
0871714e | 38 | #define CFQ_IDLE_DELAY (HZ / 5) |
d9e7620e JA |
39 | |
40 | /* | |
41 | * below this threshold, we consider thinktime immediate | |
42 | */ | |
43 | #define CFQ_MIN_TT (2) | |
44 | ||
e6c5bc73 JM |
45 | /* |
46 | * Allow merged cfqqs to perform this amount of seeky I/O before | |
47 | * deciding to break the queues up again. | |
48 | */ | |
49 | #define CFQQ_COOP_TOUT (HZ) | |
50 | ||
22e2c507 | 51 | #define CFQ_SLICE_SCALE (5) |
45333d5a | 52 | #define CFQ_HW_QUEUE_MIN (5) |
25bc6b07 | 53 | #define CFQ_SERVICE_SHIFT 12 |
22e2c507 | 54 | |
fe094d98 JA |
55 | #define RQ_CIC(rq) \ |
56 | ((struct cfq_io_context *) (rq)->elevator_private) | |
7b679138 | 57 | #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2) |
1da177e4 | 58 | |
e18b890b CL |
59 | static struct kmem_cache *cfq_pool; |
60 | static struct kmem_cache *cfq_ioc_pool; | |
1da177e4 | 61 | |
245b2e70 | 62 | static DEFINE_PER_CPU(unsigned long, cfq_ioc_count); |
334e94de | 63 | static struct completion *ioc_gone; |
9a11b4ed | 64 | static DEFINE_SPINLOCK(ioc_gone_lock); |
334e94de | 65 | |
22e2c507 JA |
66 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR |
67 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
22e2c507 JA |
68 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) |
69 | ||
206dc69b | 70 | #define sample_valid(samples) ((samples) > 80) |
1fa8f6d6 | 71 | #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node) |
206dc69b | 72 | |
cc09e299 JA |
73 | /* |
74 | * Most of our rbtree usage is for sorting with min extraction, so | |
75 | * if we cache the leftmost node we don't have to walk down the tree | |
76 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
77 | * move this into the elevator for the rq sorting as well. | |
78 | */ | |
79 | struct cfq_rb_root { | |
80 | struct rb_root rb; | |
81 | struct rb_node *left; | |
aa6f6a3d | 82 | unsigned count; |
1fa8f6d6 | 83 | u64 min_vdisktime; |
25bc6b07 | 84 | struct rb_node *active; |
58ff82f3 | 85 | unsigned total_weight; |
cc09e299 | 86 | }; |
1fa8f6d6 | 87 | #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, } |
cc09e299 | 88 | |
6118b70b JA |
89 | /* |
90 | * Per process-grouping structure | |
91 | */ | |
92 | struct cfq_queue { | |
93 | /* reference count */ | |
94 | atomic_t ref; | |
95 | /* various state flags, see below */ | |
96 | unsigned int flags; | |
97 | /* parent cfq_data */ | |
98 | struct cfq_data *cfqd; | |
99 | /* service_tree member */ | |
100 | struct rb_node rb_node; | |
101 | /* service_tree key */ | |
102 | unsigned long rb_key; | |
103 | /* prio tree member */ | |
104 | struct rb_node p_node; | |
105 | /* prio tree root we belong to, if any */ | |
106 | struct rb_root *p_root; | |
107 | /* sorted list of pending requests */ | |
108 | struct rb_root sort_list; | |
109 | /* if fifo isn't expired, next request to serve */ | |
110 | struct request *next_rq; | |
111 | /* requests queued in sort_list */ | |
112 | int queued[2]; | |
113 | /* currently allocated requests */ | |
114 | int allocated[2]; | |
115 | /* fifo list of requests in sort_list */ | |
116 | struct list_head fifo; | |
117 | ||
118 | unsigned long slice_end; | |
119 | long slice_resid; | |
120 | unsigned int slice_dispatch; | |
121 | ||
122 | /* pending metadata requests */ | |
123 | int meta_pending; | |
124 | /* number of requests that are on the dispatch list or inside driver */ | |
125 | int dispatched; | |
126 | ||
127 | /* io prio of this group */ | |
128 | unsigned short ioprio, org_ioprio; | |
129 | unsigned short ioprio_class, org_ioprio_class; | |
130 | ||
b2c18e1e JM |
131 | unsigned int seek_samples; |
132 | u64 seek_total; | |
133 | sector_t seek_mean; | |
134 | sector_t last_request_pos; | |
e6c5bc73 | 135 | unsigned long seeky_start; |
b2c18e1e | 136 | |
6118b70b | 137 | pid_t pid; |
df5fe3e8 | 138 | |
aa6f6a3d | 139 | struct cfq_rb_root *service_tree; |
df5fe3e8 | 140 | struct cfq_queue *new_cfqq; |
cdb16e8f | 141 | struct cfq_group *cfqg; |
6118b70b JA |
142 | }; |
143 | ||
c0324a02 | 144 | /* |
718eee05 | 145 | * First index in the service_trees. |
c0324a02 CZ |
146 | * IDLE is handled separately, so it has negative index |
147 | */ | |
148 | enum wl_prio_t { | |
c0324a02 | 149 | BE_WORKLOAD = 0, |
615f0259 VG |
150 | RT_WORKLOAD = 1, |
151 | IDLE_WORKLOAD = 2, | |
c0324a02 CZ |
152 | }; |
153 | ||
718eee05 CZ |
154 | /* |
155 | * Second index in the service_trees. | |
156 | */ | |
157 | enum wl_type_t { | |
158 | ASYNC_WORKLOAD = 0, | |
159 | SYNC_NOIDLE_WORKLOAD = 1, | |
160 | SYNC_WORKLOAD = 2 | |
161 | }; | |
162 | ||
cdb16e8f VG |
163 | /* This is per cgroup per device grouping structure */ |
164 | struct cfq_group { | |
1fa8f6d6 VG |
165 | /* group service_tree member */ |
166 | struct rb_node rb_node; | |
167 | ||
168 | /* group service_tree key */ | |
169 | u64 vdisktime; | |
25bc6b07 | 170 | unsigned int weight; |
1fa8f6d6 VG |
171 | bool on_st; |
172 | ||
173 | /* number of cfqq currently on this group */ | |
174 | int nr_cfqq; | |
175 | ||
58ff82f3 VG |
176 | /* Per group busy queus average. Useful for workload slice calc. */ |
177 | unsigned int busy_queues_avg[2]; | |
cdb16e8f VG |
178 | /* |
179 | * rr lists of queues with requests, onle rr for each priority class. | |
180 | * Counts are embedded in the cfq_rb_root | |
181 | */ | |
182 | struct cfq_rb_root service_trees[2][3]; | |
183 | struct cfq_rb_root service_tree_idle; | |
184 | }; | |
718eee05 | 185 | |
22e2c507 JA |
186 | /* |
187 | * Per block device queue structure | |
188 | */ | |
1da177e4 | 189 | struct cfq_data { |
165125e1 | 190 | struct request_queue *queue; |
1fa8f6d6 VG |
191 | /* Root service tree for cfq_groups */ |
192 | struct cfq_rb_root grp_service_tree; | |
cdb16e8f | 193 | struct cfq_group root_group; |
58ff82f3 VG |
194 | /* Number of active cfq groups on group service tree */ |
195 | int nr_groups; | |
22e2c507 | 196 | |
c0324a02 CZ |
197 | /* |
198 | * The priority currently being served | |
22e2c507 | 199 | */ |
c0324a02 | 200 | enum wl_prio_t serving_prio; |
718eee05 CZ |
201 | enum wl_type_t serving_type; |
202 | unsigned long workload_expires; | |
cdb16e8f | 203 | struct cfq_group *serving_group; |
8e550632 | 204 | bool noidle_tree_requires_idle; |
a36e71f9 JA |
205 | |
206 | /* | |
207 | * Each priority tree is sorted by next_request position. These | |
208 | * trees are used when determining if two or more queues are | |
209 | * interleaving requests (see cfq_close_cooperator). | |
210 | */ | |
211 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
212 | ||
22e2c507 JA |
213 | unsigned int busy_queues; |
214 | ||
5ad531db | 215 | int rq_in_driver[2]; |
3ed9a296 | 216 | int sync_flight; |
45333d5a AC |
217 | |
218 | /* | |
219 | * queue-depth detection | |
220 | */ | |
221 | int rq_queued; | |
25776e35 | 222 | int hw_tag; |
e459dd08 CZ |
223 | /* |
224 | * hw_tag can be | |
225 | * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) | |
226 | * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) | |
227 | * 0 => no NCQ | |
228 | */ | |
229 | int hw_tag_est_depth; | |
230 | unsigned int hw_tag_samples; | |
1da177e4 | 231 | |
22e2c507 JA |
232 | /* |
233 | * idle window management | |
234 | */ | |
235 | struct timer_list idle_slice_timer; | |
23e018a1 | 236 | struct work_struct unplug_work; |
1da177e4 | 237 | |
22e2c507 JA |
238 | struct cfq_queue *active_queue; |
239 | struct cfq_io_context *active_cic; | |
22e2c507 | 240 | |
c2dea2d1 VT |
241 | /* |
242 | * async queue for each priority case | |
243 | */ | |
244 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
245 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 246 | |
6d048f53 | 247 | sector_t last_position; |
1da177e4 | 248 | |
1da177e4 LT |
249 | /* |
250 | * tunables, see top of file | |
251 | */ | |
252 | unsigned int cfq_quantum; | |
22e2c507 | 253 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
254 | unsigned int cfq_back_penalty; |
255 | unsigned int cfq_back_max; | |
22e2c507 JA |
256 | unsigned int cfq_slice[2]; |
257 | unsigned int cfq_slice_async_rq; | |
258 | unsigned int cfq_slice_idle; | |
963b72fc | 259 | unsigned int cfq_latency; |
d9ff4187 AV |
260 | |
261 | struct list_head cic_list; | |
1da177e4 | 262 | |
6118b70b JA |
263 | /* |
264 | * Fallback dummy cfqq for extreme OOM conditions | |
265 | */ | |
266 | struct cfq_queue oom_cfqq; | |
365722bb VG |
267 | |
268 | unsigned long last_end_sync_rq; | |
1da177e4 LT |
269 | }; |
270 | ||
cdb16e8f VG |
271 | static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, |
272 | enum wl_prio_t prio, | |
718eee05 | 273 | enum wl_type_t type, |
c0324a02 CZ |
274 | struct cfq_data *cfqd) |
275 | { | |
1fa8f6d6 VG |
276 | if (!cfqg) |
277 | return NULL; | |
278 | ||
c0324a02 | 279 | if (prio == IDLE_WORKLOAD) |
cdb16e8f | 280 | return &cfqg->service_tree_idle; |
c0324a02 | 281 | |
cdb16e8f | 282 | return &cfqg->service_trees[prio][type]; |
c0324a02 CZ |
283 | } |
284 | ||
3b18152c | 285 | enum cfqq_state_flags { |
b0b8d749 JA |
286 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
287 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 288 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 | 289 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ |
b0b8d749 JA |
290 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
291 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
292 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 293 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 294 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
b3b6d040 | 295 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ |
76280aff | 296 | CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ |
3b18152c JA |
297 | }; |
298 | ||
299 | #define CFQ_CFQQ_FNS(name) \ | |
300 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
301 | { \ | |
fe094d98 | 302 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
303 | } \ |
304 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
305 | { \ | |
fe094d98 | 306 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
307 | } \ |
308 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
309 | { \ | |
fe094d98 | 310 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
311 | } |
312 | ||
313 | CFQ_CFQQ_FNS(on_rr); | |
314 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 315 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c | 316 | CFQ_CFQQ_FNS(must_alloc_slice); |
3b18152c JA |
317 | CFQ_CFQQ_FNS(fifo_expire); |
318 | CFQ_CFQQ_FNS(idle_window); | |
319 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 320 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 321 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 322 | CFQ_CFQQ_FNS(coop); |
76280aff | 323 | CFQ_CFQQ_FNS(deep); |
3b18152c JA |
324 | #undef CFQ_CFQQ_FNS |
325 | ||
7b679138 JA |
326 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
327 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
328 | #define cfq_log(cfqd, fmt, args...) \ | |
329 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
330 | ||
615f0259 VG |
331 | /* Traverses through cfq group service trees */ |
332 | #define for_each_cfqg_st(cfqg, i, j, st) \ | |
333 | for (i = 0; i <= IDLE_WORKLOAD; i++) \ | |
334 | for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ | |
335 | : &cfqg->service_tree_idle; \ | |
336 | (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ | |
337 | (i == IDLE_WORKLOAD && j == 0); \ | |
338 | j++, st = i < IDLE_WORKLOAD ? \ | |
339 | &cfqg->service_trees[i][j]: NULL) \ | |
340 | ||
341 | ||
c0324a02 CZ |
342 | static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) |
343 | { | |
344 | if (cfq_class_idle(cfqq)) | |
345 | return IDLE_WORKLOAD; | |
346 | if (cfq_class_rt(cfqq)) | |
347 | return RT_WORKLOAD; | |
348 | return BE_WORKLOAD; | |
349 | } | |
350 | ||
718eee05 CZ |
351 | |
352 | static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) | |
353 | { | |
354 | if (!cfq_cfqq_sync(cfqq)) | |
355 | return ASYNC_WORKLOAD; | |
356 | if (!cfq_cfqq_idle_window(cfqq)) | |
357 | return SYNC_NOIDLE_WORKLOAD; | |
358 | return SYNC_WORKLOAD; | |
359 | } | |
360 | ||
58ff82f3 VG |
361 | static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl, |
362 | struct cfq_data *cfqd, | |
363 | struct cfq_group *cfqg) | |
c0324a02 CZ |
364 | { |
365 | if (wl == IDLE_WORKLOAD) | |
cdb16e8f | 366 | return cfqg->service_tree_idle.count; |
c0324a02 | 367 | |
cdb16e8f VG |
368 | return cfqg->service_trees[wl][ASYNC_WORKLOAD].count |
369 | + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count | |
370 | + cfqg->service_trees[wl][SYNC_WORKLOAD].count; | |
c0324a02 CZ |
371 | } |
372 | ||
165125e1 | 373 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
a6151c3a | 374 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, |
fd0928df | 375 | struct io_context *, gfp_t); |
4ac845a2 | 376 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
377 | struct io_context *); |
378 | ||
5ad531db JA |
379 | static inline int rq_in_driver(struct cfq_data *cfqd) |
380 | { | |
381 | return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1]; | |
382 | } | |
383 | ||
91fac317 | 384 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, |
a6151c3a | 385 | bool is_sync) |
91fac317 | 386 | { |
a6151c3a | 387 | return cic->cfqq[is_sync]; |
91fac317 VT |
388 | } |
389 | ||
390 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 391 | struct cfq_queue *cfqq, bool is_sync) |
91fac317 | 392 | { |
a6151c3a | 393 | cic->cfqq[is_sync] = cfqq; |
91fac317 VT |
394 | } |
395 | ||
396 | /* | |
397 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
398 | * set (in which case it could also be direct WRITE). | |
399 | */ | |
a6151c3a | 400 | static inline bool cfq_bio_sync(struct bio *bio) |
91fac317 | 401 | { |
a6151c3a | 402 | return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO); |
91fac317 | 403 | } |
1da177e4 | 404 | |
99f95e52 AM |
405 | /* |
406 | * scheduler run of queue, if there are requests pending and no one in the | |
407 | * driver that will restart queueing | |
408 | */ | |
23e018a1 | 409 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) |
99f95e52 | 410 | { |
7b679138 JA |
411 | if (cfqd->busy_queues) { |
412 | cfq_log(cfqd, "schedule dispatch"); | |
23e018a1 | 413 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 414 | } |
99f95e52 AM |
415 | } |
416 | ||
165125e1 | 417 | static int cfq_queue_empty(struct request_queue *q) |
99f95e52 AM |
418 | { |
419 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
420 | ||
f04a6424 | 421 | return !cfqd->rq_queued; |
99f95e52 AM |
422 | } |
423 | ||
44f7c160 JA |
424 | /* |
425 | * Scale schedule slice based on io priority. Use the sync time slice only | |
426 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
427 | * io only, should not get full sync slice length. | |
428 | */ | |
a6151c3a | 429 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, |
d9e7620e | 430 | unsigned short prio) |
44f7c160 | 431 | { |
d9e7620e | 432 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 433 | |
d9e7620e JA |
434 | WARN_ON(prio >= IOPRIO_BE_NR); |
435 | ||
436 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
437 | } | |
44f7c160 | 438 | |
d9e7620e JA |
439 | static inline int |
440 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
441 | { | |
442 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
443 | } |
444 | ||
25bc6b07 VG |
445 | static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg) |
446 | { | |
447 | u64 d = delta << CFQ_SERVICE_SHIFT; | |
448 | ||
449 | d = d * BLKIO_WEIGHT_DEFAULT; | |
450 | do_div(d, cfqg->weight); | |
451 | return d; | |
452 | } | |
453 | ||
454 | static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
455 | { | |
456 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
457 | if (delta > 0) | |
458 | min_vdisktime = vdisktime; | |
459 | ||
460 | return min_vdisktime; | |
461 | } | |
462 | ||
463 | static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
464 | { | |
465 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
466 | if (delta < 0) | |
467 | min_vdisktime = vdisktime; | |
468 | ||
469 | return min_vdisktime; | |
470 | } | |
471 | ||
472 | static void update_min_vdisktime(struct cfq_rb_root *st) | |
473 | { | |
474 | u64 vdisktime = st->min_vdisktime; | |
475 | struct cfq_group *cfqg; | |
476 | ||
477 | if (st->active) { | |
478 | cfqg = rb_entry_cfqg(st->active); | |
479 | vdisktime = cfqg->vdisktime; | |
480 | } | |
481 | ||
482 | if (st->left) { | |
483 | cfqg = rb_entry_cfqg(st->left); | |
484 | vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime); | |
485 | } | |
486 | ||
487 | st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime); | |
488 | } | |
489 | ||
5db5d642 CZ |
490 | /* |
491 | * get averaged number of queues of RT/BE priority. | |
492 | * average is updated, with a formula that gives more weight to higher numbers, | |
493 | * to quickly follows sudden increases and decrease slowly | |
494 | */ | |
495 | ||
58ff82f3 VG |
496 | static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, |
497 | struct cfq_group *cfqg, bool rt) | |
5869619c | 498 | { |
5db5d642 CZ |
499 | unsigned min_q, max_q; |
500 | unsigned mult = cfq_hist_divisor - 1; | |
501 | unsigned round = cfq_hist_divisor / 2; | |
58ff82f3 | 502 | unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); |
5db5d642 | 503 | |
58ff82f3 VG |
504 | min_q = min(cfqg->busy_queues_avg[rt], busy); |
505 | max_q = max(cfqg->busy_queues_avg[rt], busy); | |
506 | cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
5db5d642 | 507 | cfq_hist_divisor; |
58ff82f3 VG |
508 | return cfqg->busy_queues_avg[rt]; |
509 | } | |
510 | ||
511 | static inline unsigned | |
512 | cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
513 | { | |
514 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
515 | ||
516 | return cfq_target_latency * cfqg->weight / st->total_weight; | |
5db5d642 CZ |
517 | } |
518 | ||
44f7c160 JA |
519 | static inline void |
520 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
521 | { | |
5db5d642 CZ |
522 | unsigned slice = cfq_prio_to_slice(cfqd, cfqq); |
523 | if (cfqd->cfq_latency) { | |
58ff82f3 VG |
524 | /* |
525 | * interested queues (we consider only the ones with the same | |
526 | * priority class in the cfq group) | |
527 | */ | |
528 | unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, | |
529 | cfq_class_rt(cfqq)); | |
5db5d642 CZ |
530 | unsigned sync_slice = cfqd->cfq_slice[1]; |
531 | unsigned expect_latency = sync_slice * iq; | |
58ff82f3 VG |
532 | unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg); |
533 | ||
534 | if (expect_latency > group_slice) { | |
5db5d642 CZ |
535 | unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; |
536 | /* scale low_slice according to IO priority | |
537 | * and sync vs async */ | |
538 | unsigned low_slice = | |
539 | min(slice, base_low_slice * slice / sync_slice); | |
540 | /* the adapted slice value is scaled to fit all iqs | |
541 | * into the target latency */ | |
58ff82f3 | 542 | slice = max(slice * group_slice / expect_latency, |
5db5d642 CZ |
543 | low_slice); |
544 | } | |
545 | } | |
546 | cfqq->slice_end = jiffies + slice; | |
7b679138 | 547 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
548 | } |
549 | ||
550 | /* | |
551 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
552 | * isn't valid until the first request from the dispatch is activated | |
553 | * and the slice time set. | |
554 | */ | |
a6151c3a | 555 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) |
44f7c160 JA |
556 | { |
557 | if (cfq_cfqq_slice_new(cfqq)) | |
558 | return 0; | |
559 | if (time_before(jiffies, cfqq->slice_end)) | |
560 | return 0; | |
561 | ||
562 | return 1; | |
563 | } | |
564 | ||
1da177e4 | 565 | /* |
5e705374 | 566 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 567 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 568 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 569 | */ |
5e705374 | 570 | static struct request * |
cf7c25cf | 571 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) |
1da177e4 | 572 | { |
cf7c25cf | 573 | sector_t s1, s2, d1 = 0, d2 = 0; |
1da177e4 | 574 | unsigned long back_max; |
e8a99053 AM |
575 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
576 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
577 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 578 | |
5e705374 JA |
579 | if (rq1 == NULL || rq1 == rq2) |
580 | return rq2; | |
581 | if (rq2 == NULL) | |
582 | return rq1; | |
9c2c38a1 | 583 | |
5e705374 JA |
584 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
585 | return rq1; | |
586 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
587 | return rq2; | |
374f84ac JA |
588 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
589 | return rq1; | |
590 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
591 | return rq2; | |
1da177e4 | 592 | |
83096ebf TH |
593 | s1 = blk_rq_pos(rq1); |
594 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 595 | |
1da177e4 LT |
596 | /* |
597 | * by definition, 1KiB is 2 sectors | |
598 | */ | |
599 | back_max = cfqd->cfq_back_max * 2; | |
600 | ||
601 | /* | |
602 | * Strict one way elevator _except_ in the case where we allow | |
603 | * short backward seeks which are biased as twice the cost of a | |
604 | * similar forward seek. | |
605 | */ | |
606 | if (s1 >= last) | |
607 | d1 = s1 - last; | |
608 | else if (s1 + back_max >= last) | |
609 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
610 | else | |
e8a99053 | 611 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
612 | |
613 | if (s2 >= last) | |
614 | d2 = s2 - last; | |
615 | else if (s2 + back_max >= last) | |
616 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
617 | else | |
e8a99053 | 618 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
619 | |
620 | /* Found required data */ | |
e8a99053 AM |
621 | |
622 | /* | |
623 | * By doing switch() on the bit mask "wrap" we avoid having to | |
624 | * check two variables for all permutations: --> faster! | |
625 | */ | |
626 | switch (wrap) { | |
5e705374 | 627 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 628 | if (d1 < d2) |
5e705374 | 629 | return rq1; |
e8a99053 | 630 | else if (d2 < d1) |
5e705374 | 631 | return rq2; |
e8a99053 AM |
632 | else { |
633 | if (s1 >= s2) | |
5e705374 | 634 | return rq1; |
e8a99053 | 635 | else |
5e705374 | 636 | return rq2; |
e8a99053 | 637 | } |
1da177e4 | 638 | |
e8a99053 | 639 | case CFQ_RQ2_WRAP: |
5e705374 | 640 | return rq1; |
e8a99053 | 641 | case CFQ_RQ1_WRAP: |
5e705374 JA |
642 | return rq2; |
643 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
644 | default: |
645 | /* | |
646 | * Since both rqs are wrapped, | |
647 | * start with the one that's further behind head | |
648 | * (--> only *one* back seek required), | |
649 | * since back seek takes more time than forward. | |
650 | */ | |
651 | if (s1 <= s2) | |
5e705374 | 652 | return rq1; |
1da177e4 | 653 | else |
5e705374 | 654 | return rq2; |
1da177e4 LT |
655 | } |
656 | } | |
657 | ||
498d3aa2 JA |
658 | /* |
659 | * The below is leftmost cache rbtree addon | |
660 | */ | |
0871714e | 661 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 | 662 | { |
615f0259 VG |
663 | /* Service tree is empty */ |
664 | if (!root->count) | |
665 | return NULL; | |
666 | ||
cc09e299 JA |
667 | if (!root->left) |
668 | root->left = rb_first(&root->rb); | |
669 | ||
0871714e JA |
670 | if (root->left) |
671 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
672 | ||
673 | return NULL; | |
cc09e299 JA |
674 | } |
675 | ||
1fa8f6d6 VG |
676 | static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) |
677 | { | |
678 | if (!root->left) | |
679 | root->left = rb_first(&root->rb); | |
680 | ||
681 | if (root->left) | |
682 | return rb_entry_cfqg(root->left); | |
683 | ||
684 | return NULL; | |
685 | } | |
686 | ||
a36e71f9 JA |
687 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
688 | { | |
689 | rb_erase(n, root); | |
690 | RB_CLEAR_NODE(n); | |
691 | } | |
692 | ||
cc09e299 JA |
693 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
694 | { | |
695 | if (root->left == n) | |
696 | root->left = NULL; | |
a36e71f9 | 697 | rb_erase_init(n, &root->rb); |
aa6f6a3d | 698 | --root->count; |
cc09e299 JA |
699 | } |
700 | ||
1da177e4 LT |
701 | /* |
702 | * would be nice to take fifo expire time into account as well | |
703 | */ | |
5e705374 JA |
704 | static struct request * |
705 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
706 | struct request *last) | |
1da177e4 | 707 | { |
21183b07 JA |
708 | struct rb_node *rbnext = rb_next(&last->rb_node); |
709 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 710 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 711 | |
21183b07 | 712 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
713 | |
714 | if (rbprev) | |
5e705374 | 715 | prev = rb_entry_rq(rbprev); |
1da177e4 | 716 | |
21183b07 | 717 | if (rbnext) |
5e705374 | 718 | next = rb_entry_rq(rbnext); |
21183b07 JA |
719 | else { |
720 | rbnext = rb_first(&cfqq->sort_list); | |
721 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 722 | next = rb_entry_rq(rbnext); |
21183b07 | 723 | } |
1da177e4 | 724 | |
cf7c25cf | 725 | return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); |
1da177e4 LT |
726 | } |
727 | ||
d9e7620e JA |
728 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
729 | struct cfq_queue *cfqq) | |
1da177e4 | 730 | { |
d9e7620e JA |
731 | /* |
732 | * just an approximation, should be ok. | |
733 | */ | |
cdb16e8f | 734 | return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
464191c6 | 735 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); |
d9e7620e JA |
736 | } |
737 | ||
1fa8f6d6 VG |
738 | static inline s64 |
739 | cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
740 | { | |
741 | return cfqg->vdisktime - st->min_vdisktime; | |
742 | } | |
743 | ||
744 | static void | |
745 | __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
746 | { | |
747 | struct rb_node **node = &st->rb.rb_node; | |
748 | struct rb_node *parent = NULL; | |
749 | struct cfq_group *__cfqg; | |
750 | s64 key = cfqg_key(st, cfqg); | |
751 | int left = 1; | |
752 | ||
753 | while (*node != NULL) { | |
754 | parent = *node; | |
755 | __cfqg = rb_entry_cfqg(parent); | |
756 | ||
757 | if (key < cfqg_key(st, __cfqg)) | |
758 | node = &parent->rb_left; | |
759 | else { | |
760 | node = &parent->rb_right; | |
761 | left = 0; | |
762 | } | |
763 | } | |
764 | ||
765 | if (left) | |
766 | st->left = &cfqg->rb_node; | |
767 | ||
768 | rb_link_node(&cfqg->rb_node, parent, node); | |
769 | rb_insert_color(&cfqg->rb_node, &st->rb); | |
770 | } | |
771 | ||
772 | static void | |
773 | cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
774 | { | |
775 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
776 | struct cfq_group *__cfqg; | |
777 | struct rb_node *n; | |
778 | ||
779 | cfqg->nr_cfqq++; | |
780 | if (cfqg->on_st) | |
781 | return; | |
782 | ||
783 | /* | |
784 | * Currently put the group at the end. Later implement something | |
785 | * so that groups get lesser vtime based on their weights, so that | |
786 | * if group does not loose all if it was not continously backlogged. | |
787 | */ | |
788 | n = rb_last(&st->rb); | |
789 | if (n) { | |
790 | __cfqg = rb_entry_cfqg(n); | |
791 | cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; | |
792 | } else | |
793 | cfqg->vdisktime = st->min_vdisktime; | |
794 | ||
795 | __cfq_group_service_tree_add(st, cfqg); | |
796 | cfqg->on_st = true; | |
58ff82f3 VG |
797 | cfqd->nr_groups++; |
798 | st->total_weight += cfqg->weight; | |
1fa8f6d6 VG |
799 | } |
800 | ||
801 | static void | |
802 | cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
803 | { | |
804 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
805 | ||
25bc6b07 VG |
806 | if (st->active == &cfqg->rb_node) |
807 | st->active = NULL; | |
808 | ||
1fa8f6d6 VG |
809 | BUG_ON(cfqg->nr_cfqq < 1); |
810 | cfqg->nr_cfqq--; | |
25bc6b07 | 811 | |
1fa8f6d6 VG |
812 | /* If there are other cfq queues under this group, don't delete it */ |
813 | if (cfqg->nr_cfqq) | |
814 | return; | |
815 | ||
816 | cfqg->on_st = false; | |
58ff82f3 VG |
817 | cfqd->nr_groups--; |
818 | st->total_weight -= cfqg->weight; | |
1fa8f6d6 VG |
819 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) |
820 | cfq_rb_erase(&cfqg->rb_node, st); | |
821 | } | |
822 | ||
498d3aa2 | 823 | /* |
c0324a02 | 824 | * The cfqd->service_trees holds all pending cfq_queue's that have |
498d3aa2 JA |
825 | * requests waiting to be processed. It is sorted in the order that |
826 | * we will service the queues. | |
827 | */ | |
a36e71f9 | 828 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 829 | bool add_front) |
d9e7620e | 830 | { |
0871714e JA |
831 | struct rb_node **p, *parent; |
832 | struct cfq_queue *__cfqq; | |
d9e7620e | 833 | unsigned long rb_key; |
c0324a02 | 834 | struct cfq_rb_root *service_tree; |
498d3aa2 | 835 | int left; |
d9e7620e | 836 | |
cdb16e8f VG |
837 | service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), |
838 | cfqq_type(cfqq), cfqd); | |
0871714e JA |
839 | if (cfq_class_idle(cfqq)) { |
840 | rb_key = CFQ_IDLE_DELAY; | |
aa6f6a3d | 841 | parent = rb_last(&service_tree->rb); |
0871714e JA |
842 | if (parent && parent != &cfqq->rb_node) { |
843 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
844 | rb_key += __cfqq->rb_key; | |
845 | } else | |
846 | rb_key += jiffies; | |
847 | } else if (!add_front) { | |
b9c8946b JA |
848 | /* |
849 | * Get our rb key offset. Subtract any residual slice | |
850 | * value carried from last service. A negative resid | |
851 | * count indicates slice overrun, and this should position | |
852 | * the next service time further away in the tree. | |
853 | */ | |
edd75ffd | 854 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
b9c8946b | 855 | rb_key -= cfqq->slice_resid; |
edd75ffd | 856 | cfqq->slice_resid = 0; |
48e025e6 CZ |
857 | } else { |
858 | rb_key = -HZ; | |
aa6f6a3d | 859 | __cfqq = cfq_rb_first(service_tree); |
48e025e6 CZ |
860 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; |
861 | } | |
1da177e4 | 862 | |
d9e7620e | 863 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
99f9628a | 864 | /* |
d9e7620e | 865 | * same position, nothing more to do |
99f9628a | 866 | */ |
c0324a02 CZ |
867 | if (rb_key == cfqq->rb_key && |
868 | cfqq->service_tree == service_tree) | |
d9e7620e | 869 | return; |
1da177e4 | 870 | |
aa6f6a3d CZ |
871 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); |
872 | cfqq->service_tree = NULL; | |
1da177e4 | 873 | } |
d9e7620e | 874 | |
498d3aa2 | 875 | left = 1; |
0871714e | 876 | parent = NULL; |
aa6f6a3d CZ |
877 | cfqq->service_tree = service_tree; |
878 | p = &service_tree->rb.rb_node; | |
d9e7620e | 879 | while (*p) { |
67060e37 | 880 | struct rb_node **n; |
cc09e299 | 881 | |
d9e7620e JA |
882 | parent = *p; |
883 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
884 | ||
0c534e0a | 885 | /* |
c0324a02 | 886 | * sort by key, that represents service time. |
0c534e0a | 887 | */ |
c0324a02 | 888 | if (time_before(rb_key, __cfqq->rb_key)) |
67060e37 | 889 | n = &(*p)->rb_left; |
c0324a02 | 890 | else { |
67060e37 | 891 | n = &(*p)->rb_right; |
cc09e299 | 892 | left = 0; |
c0324a02 | 893 | } |
67060e37 JA |
894 | |
895 | p = n; | |
d9e7620e JA |
896 | } |
897 | ||
cc09e299 | 898 | if (left) |
aa6f6a3d | 899 | service_tree->left = &cfqq->rb_node; |
cc09e299 | 900 | |
d9e7620e JA |
901 | cfqq->rb_key = rb_key; |
902 | rb_link_node(&cfqq->rb_node, parent, p); | |
aa6f6a3d CZ |
903 | rb_insert_color(&cfqq->rb_node, &service_tree->rb); |
904 | service_tree->count++; | |
1fa8f6d6 | 905 | cfq_group_service_tree_add(cfqd, cfqq->cfqg); |
1da177e4 LT |
906 | } |
907 | ||
a36e71f9 | 908 | static struct cfq_queue * |
f2d1f0ae JA |
909 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
910 | sector_t sector, struct rb_node **ret_parent, | |
911 | struct rb_node ***rb_link) | |
a36e71f9 | 912 | { |
a36e71f9 JA |
913 | struct rb_node **p, *parent; |
914 | struct cfq_queue *cfqq = NULL; | |
915 | ||
916 | parent = NULL; | |
917 | p = &root->rb_node; | |
918 | while (*p) { | |
919 | struct rb_node **n; | |
920 | ||
921 | parent = *p; | |
922 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
923 | ||
924 | /* | |
925 | * Sort strictly based on sector. Smallest to the left, | |
926 | * largest to the right. | |
927 | */ | |
2e46e8b2 | 928 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 929 | n = &(*p)->rb_right; |
2e46e8b2 | 930 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
931 | n = &(*p)->rb_left; |
932 | else | |
933 | break; | |
934 | p = n; | |
3ac6c9f8 | 935 | cfqq = NULL; |
a36e71f9 JA |
936 | } |
937 | ||
938 | *ret_parent = parent; | |
939 | if (rb_link) | |
940 | *rb_link = p; | |
3ac6c9f8 | 941 | return cfqq; |
a36e71f9 JA |
942 | } |
943 | ||
944 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
945 | { | |
a36e71f9 JA |
946 | struct rb_node **p, *parent; |
947 | struct cfq_queue *__cfqq; | |
948 | ||
f2d1f0ae JA |
949 | if (cfqq->p_root) { |
950 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
951 | cfqq->p_root = NULL; | |
952 | } | |
a36e71f9 JA |
953 | |
954 | if (cfq_class_idle(cfqq)) | |
955 | return; | |
956 | if (!cfqq->next_rq) | |
957 | return; | |
958 | ||
f2d1f0ae | 959 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
960 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
961 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
962 | if (!__cfqq) { |
963 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
964 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
965 | } else | |
966 | cfqq->p_root = NULL; | |
a36e71f9 JA |
967 | } |
968 | ||
498d3aa2 JA |
969 | /* |
970 | * Update cfqq's position in the service tree. | |
971 | */ | |
edd75ffd | 972 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 973 | { |
6d048f53 JA |
974 | /* |
975 | * Resorting requires the cfqq to be on the RR list already. | |
976 | */ | |
a36e71f9 | 977 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 978 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
979 | cfq_prio_tree_add(cfqd, cfqq); |
980 | } | |
6d048f53 JA |
981 | } |
982 | ||
1da177e4 LT |
983 | /* |
984 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 985 | * the pending list according to last request service |
1da177e4 | 986 | */ |
febffd61 | 987 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 988 | { |
7b679138 | 989 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
990 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
991 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 LT |
992 | cfqd->busy_queues++; |
993 | ||
edd75ffd | 994 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
995 | } |
996 | ||
498d3aa2 JA |
997 | /* |
998 | * Called when the cfqq no longer has requests pending, remove it from | |
999 | * the service tree. | |
1000 | */ | |
febffd61 | 1001 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1002 | { |
7b679138 | 1003 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
1004 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
1005 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 1006 | |
aa6f6a3d CZ |
1007 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
1008 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
1009 | cfqq->service_tree = NULL; | |
1010 | } | |
f2d1f0ae JA |
1011 | if (cfqq->p_root) { |
1012 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1013 | cfqq->p_root = NULL; | |
1014 | } | |
d9e7620e | 1015 | |
1fa8f6d6 | 1016 | cfq_group_service_tree_del(cfqd, cfqq->cfqg); |
1da177e4 LT |
1017 | BUG_ON(!cfqd->busy_queues); |
1018 | cfqd->busy_queues--; | |
1019 | } | |
1020 | ||
1021 | /* | |
1022 | * rb tree support functions | |
1023 | */ | |
febffd61 | 1024 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 1025 | { |
5e705374 | 1026 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
5e705374 | 1027 | const int sync = rq_is_sync(rq); |
1da177e4 | 1028 | |
b4878f24 JA |
1029 | BUG_ON(!cfqq->queued[sync]); |
1030 | cfqq->queued[sync]--; | |
1da177e4 | 1031 | |
5e705374 | 1032 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 1033 | |
f04a6424 VG |
1034 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { |
1035 | /* | |
1036 | * Queue will be deleted from service tree when we actually | |
1037 | * expire it later. Right now just remove it from prio tree | |
1038 | * as it is empty. | |
1039 | */ | |
1040 | if (cfqq->p_root) { | |
1041 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1042 | cfqq->p_root = NULL; | |
1043 | } | |
1044 | } | |
1da177e4 LT |
1045 | } |
1046 | ||
5e705374 | 1047 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 1048 | { |
5e705374 | 1049 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 1050 | struct cfq_data *cfqd = cfqq->cfqd; |
a36e71f9 | 1051 | struct request *__alias, *prev; |
1da177e4 | 1052 | |
5380a101 | 1053 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
1054 | |
1055 | /* | |
1056 | * looks a little odd, but the first insert might return an alias. | |
1057 | * if that happens, put the alias on the dispatch list | |
1058 | */ | |
21183b07 | 1059 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 1060 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
1061 | |
1062 | if (!cfq_cfqq_on_rr(cfqq)) | |
1063 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
1064 | |
1065 | /* | |
1066 | * check if this request is a better next-serve candidate | |
1067 | */ | |
a36e71f9 | 1068 | prev = cfqq->next_rq; |
cf7c25cf | 1069 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); |
a36e71f9 JA |
1070 | |
1071 | /* | |
1072 | * adjust priority tree position, if ->next_rq changes | |
1073 | */ | |
1074 | if (prev != cfqq->next_rq) | |
1075 | cfq_prio_tree_add(cfqd, cfqq); | |
1076 | ||
5044eed4 | 1077 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
1078 | } |
1079 | ||
febffd61 | 1080 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 1081 | { |
5380a101 JA |
1082 | elv_rb_del(&cfqq->sort_list, rq); |
1083 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 1084 | cfq_add_rq_rb(rq); |
1da177e4 LT |
1085 | } |
1086 | ||
206dc69b JA |
1087 | static struct request * |
1088 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 1089 | { |
206dc69b | 1090 | struct task_struct *tsk = current; |
91fac317 | 1091 | struct cfq_io_context *cic; |
206dc69b | 1092 | struct cfq_queue *cfqq; |
1da177e4 | 1093 | |
4ac845a2 | 1094 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
1095 | if (!cic) |
1096 | return NULL; | |
1097 | ||
1098 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
1099 | if (cfqq) { |
1100 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
1101 | ||
21183b07 | 1102 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 1103 | } |
1da177e4 | 1104 | |
1da177e4 LT |
1105 | return NULL; |
1106 | } | |
1107 | ||
165125e1 | 1108 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1109 | { |
22e2c507 | 1110 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 1111 | |
5ad531db | 1112 | cfqd->rq_in_driver[rq_is_sync(rq)]++; |
7b679138 | 1113 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
5ad531db | 1114 | rq_in_driver(cfqd)); |
25776e35 | 1115 | |
5b93629b | 1116 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
1117 | } |
1118 | ||
165125e1 | 1119 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1120 | { |
b4878f24 | 1121 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5ad531db | 1122 | const int sync = rq_is_sync(rq); |
b4878f24 | 1123 | |
5ad531db JA |
1124 | WARN_ON(!cfqd->rq_in_driver[sync]); |
1125 | cfqd->rq_in_driver[sync]--; | |
7b679138 | 1126 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
5ad531db | 1127 | rq_in_driver(cfqd)); |
1da177e4 LT |
1128 | } |
1129 | ||
b4878f24 | 1130 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 1131 | { |
5e705374 | 1132 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 1133 | |
5e705374 JA |
1134 | if (cfqq->next_rq == rq) |
1135 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 1136 | |
b4878f24 | 1137 | list_del_init(&rq->queuelist); |
5e705374 | 1138 | cfq_del_rq_rb(rq); |
374f84ac | 1139 | |
45333d5a | 1140 | cfqq->cfqd->rq_queued--; |
374f84ac JA |
1141 | if (rq_is_meta(rq)) { |
1142 | WARN_ON(!cfqq->meta_pending); | |
1143 | cfqq->meta_pending--; | |
1144 | } | |
1da177e4 LT |
1145 | } |
1146 | ||
165125e1 JA |
1147 | static int cfq_merge(struct request_queue *q, struct request **req, |
1148 | struct bio *bio) | |
1da177e4 LT |
1149 | { |
1150 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1151 | struct request *__rq; | |
1da177e4 | 1152 | |
206dc69b | 1153 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 1154 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
1155 | *req = __rq; |
1156 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1157 | } |
1158 | ||
1159 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1160 | } |
1161 | ||
165125e1 | 1162 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 1163 | int type) |
1da177e4 | 1164 | { |
21183b07 | 1165 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 1166 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 1167 | |
5e705374 | 1168 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 1169 | } |
1da177e4 LT |
1170 | } |
1171 | ||
1172 | static void | |
165125e1 | 1173 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
1174 | struct request *next) |
1175 | { | |
cf7c25cf | 1176 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 JA |
1177 | /* |
1178 | * reposition in fifo if next is older than rq | |
1179 | */ | |
1180 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
30996f40 | 1181 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
22e2c507 | 1182 | list_move(&rq->queuelist, &next->queuelist); |
30996f40 JA |
1183 | rq_set_fifo_time(rq, rq_fifo_time(next)); |
1184 | } | |
22e2c507 | 1185 | |
cf7c25cf CZ |
1186 | if (cfqq->next_rq == next) |
1187 | cfqq->next_rq = rq; | |
b4878f24 | 1188 | cfq_remove_request(next); |
22e2c507 JA |
1189 | } |
1190 | ||
165125e1 | 1191 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
1192 | struct bio *bio) |
1193 | { | |
1194 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 1195 | struct cfq_io_context *cic; |
da775265 | 1196 | struct cfq_queue *cfqq; |
da775265 JA |
1197 | |
1198 | /* | |
ec8acb69 | 1199 | * Disallow merge of a sync bio into an async request. |
da775265 | 1200 | */ |
91fac317 | 1201 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
a6151c3a | 1202 | return false; |
da775265 JA |
1203 | |
1204 | /* | |
719d3402 JA |
1205 | * Lookup the cfqq that this bio will be queued with. Allow |
1206 | * merge only if rq is queued there. | |
da775265 | 1207 | */ |
4ac845a2 | 1208 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 | 1209 | if (!cic) |
a6151c3a | 1210 | return false; |
719d3402 | 1211 | |
91fac317 | 1212 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
a6151c3a | 1213 | return cfqq == RQ_CFQQ(rq); |
da775265 JA |
1214 | } |
1215 | ||
febffd61 JA |
1216 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
1217 | struct cfq_queue *cfqq) | |
22e2c507 JA |
1218 | { |
1219 | if (cfqq) { | |
7b679138 | 1220 | cfq_log_cfqq(cfqd, cfqq, "set_active"); |
22e2c507 | 1221 | cfqq->slice_end = 0; |
2f5cb738 JA |
1222 | cfqq->slice_dispatch = 0; |
1223 | ||
2f5cb738 | 1224 | cfq_clear_cfqq_wait_request(cfqq); |
b029195d | 1225 | cfq_clear_cfqq_must_dispatch(cfqq); |
3b18152c JA |
1226 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
1227 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 1228 | cfq_mark_cfqq_slice_new(cfqq); |
2f5cb738 JA |
1229 | |
1230 | del_timer(&cfqd->idle_slice_timer); | |
22e2c507 JA |
1231 | } |
1232 | ||
1233 | cfqd->active_queue = cfqq; | |
1234 | } | |
1235 | ||
7b14e3b5 JA |
1236 | /* |
1237 | * current cfqq expired its slice (or was too idle), select new one | |
1238 | */ | |
1239 | static void | |
1240 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
a6151c3a | 1241 | bool timed_out) |
7b14e3b5 | 1242 | { |
7b679138 JA |
1243 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
1244 | ||
7b14e3b5 JA |
1245 | if (cfq_cfqq_wait_request(cfqq)) |
1246 | del_timer(&cfqd->idle_slice_timer); | |
1247 | ||
7b14e3b5 JA |
1248 | cfq_clear_cfqq_wait_request(cfqq); |
1249 | ||
1250 | /* | |
6084cdda | 1251 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 1252 | */ |
7b679138 | 1253 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) { |
c5b680f3 | 1254 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b679138 JA |
1255 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
1256 | } | |
7b14e3b5 | 1257 | |
f04a6424 VG |
1258 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
1259 | cfq_del_cfqq_rr(cfqd, cfqq); | |
1260 | ||
edd75ffd | 1261 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
1262 | |
1263 | if (cfqq == cfqd->active_queue) | |
1264 | cfqd->active_queue = NULL; | |
1265 | ||
1266 | if (cfqd->active_cic) { | |
1267 | put_io_context(cfqd->active_cic->ioc); | |
1268 | cfqd->active_cic = NULL; | |
1269 | } | |
7b14e3b5 JA |
1270 | } |
1271 | ||
a6151c3a | 1272 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) |
7b14e3b5 JA |
1273 | { |
1274 | struct cfq_queue *cfqq = cfqd->active_queue; | |
1275 | ||
1276 | if (cfqq) | |
6084cdda | 1277 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
1278 | } |
1279 | ||
498d3aa2 JA |
1280 | /* |
1281 | * Get next queue for service. Unless we have a queue preemption, | |
1282 | * we'll simply select the first cfqq in the service tree. | |
1283 | */ | |
6d048f53 | 1284 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 1285 | { |
c0324a02 | 1286 | struct cfq_rb_root *service_tree = |
cdb16e8f VG |
1287 | service_tree_for(cfqd->serving_group, cfqd->serving_prio, |
1288 | cfqd->serving_type, cfqd); | |
d9e7620e | 1289 | |
f04a6424 VG |
1290 | if (!cfqd->rq_queued) |
1291 | return NULL; | |
1292 | ||
1fa8f6d6 VG |
1293 | /* There is nothing to dispatch */ |
1294 | if (!service_tree) | |
1295 | return NULL; | |
c0324a02 CZ |
1296 | if (RB_EMPTY_ROOT(&service_tree->rb)) |
1297 | return NULL; | |
1298 | return cfq_rb_first(service_tree); | |
6d048f53 JA |
1299 | } |
1300 | ||
f04a6424 VG |
1301 | static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) |
1302 | { | |
1303 | struct cfq_group *cfqg = &cfqd->root_group; | |
1304 | struct cfq_queue *cfqq; | |
1305 | int i, j; | |
1306 | struct cfq_rb_root *st; | |
1307 | ||
1308 | if (!cfqd->rq_queued) | |
1309 | return NULL; | |
1310 | ||
1311 | for_each_cfqg_st(cfqg, i, j, st) | |
1312 | if ((cfqq = cfq_rb_first(st)) != NULL) | |
1313 | return cfqq; | |
1314 | return NULL; | |
1315 | } | |
1316 | ||
498d3aa2 JA |
1317 | /* |
1318 | * Get and set a new active queue for service. | |
1319 | */ | |
a36e71f9 JA |
1320 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
1321 | struct cfq_queue *cfqq) | |
6d048f53 | 1322 | { |
e00ef799 | 1323 | if (!cfqq) |
a36e71f9 | 1324 | cfqq = cfq_get_next_queue(cfqd); |
6d048f53 | 1325 | |
22e2c507 | 1326 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 1327 | return cfqq; |
22e2c507 JA |
1328 | } |
1329 | ||
d9e7620e JA |
1330 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
1331 | struct request *rq) | |
1332 | { | |
83096ebf TH |
1333 | if (blk_rq_pos(rq) >= cfqd->last_position) |
1334 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 1335 | else |
83096ebf | 1336 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
1337 | } |
1338 | ||
b2c18e1e JM |
1339 | #define CFQQ_SEEK_THR 8 * 1024 |
1340 | #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR) | |
04dc6e71 | 1341 | |
b2c18e1e JM |
1342 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
1343 | struct request *rq) | |
6d048f53 | 1344 | { |
b2c18e1e | 1345 | sector_t sdist = cfqq->seek_mean; |
6d048f53 | 1346 | |
b2c18e1e JM |
1347 | if (!sample_valid(cfqq->seek_samples)) |
1348 | sdist = CFQQ_SEEK_THR; | |
6d048f53 | 1349 | |
04dc6e71 | 1350 | return cfq_dist_from_last(cfqd, rq) <= sdist; |
6d048f53 JA |
1351 | } |
1352 | ||
a36e71f9 JA |
1353 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
1354 | struct cfq_queue *cur_cfqq) | |
1355 | { | |
f2d1f0ae | 1356 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
1357 | struct rb_node *parent, *node; |
1358 | struct cfq_queue *__cfqq; | |
1359 | sector_t sector = cfqd->last_position; | |
1360 | ||
1361 | if (RB_EMPTY_ROOT(root)) | |
1362 | return NULL; | |
1363 | ||
1364 | /* | |
1365 | * First, if we find a request starting at the end of the last | |
1366 | * request, choose it. | |
1367 | */ | |
f2d1f0ae | 1368 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
1369 | if (__cfqq) |
1370 | return __cfqq; | |
1371 | ||
1372 | /* | |
1373 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1374 | * will contain the closest sector. | |
1375 | */ | |
1376 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
b2c18e1e | 1377 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1378 | return __cfqq; |
1379 | ||
2e46e8b2 | 1380 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1381 | node = rb_next(&__cfqq->p_node); |
1382 | else | |
1383 | node = rb_prev(&__cfqq->p_node); | |
1384 | if (!node) | |
1385 | return NULL; | |
1386 | ||
1387 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
b2c18e1e | 1388 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1389 | return __cfqq; |
1390 | ||
1391 | return NULL; | |
1392 | } | |
1393 | ||
1394 | /* | |
1395 | * cfqd - obvious | |
1396 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1397 | * closely cooperating with itself. | |
1398 | * | |
1399 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1400 | * one request, and that cfqd->last_position reflects a position on the disk | |
1401 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1402 | * assumption. | |
1403 | */ | |
1404 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
b3b6d040 | 1405 | struct cfq_queue *cur_cfqq) |
6d048f53 | 1406 | { |
a36e71f9 JA |
1407 | struct cfq_queue *cfqq; |
1408 | ||
e6c5bc73 JM |
1409 | if (!cfq_cfqq_sync(cur_cfqq)) |
1410 | return NULL; | |
1411 | if (CFQQ_SEEKY(cur_cfqq)) | |
1412 | return NULL; | |
1413 | ||
6d048f53 | 1414 | /* |
d9e7620e JA |
1415 | * We should notice if some of the queues are cooperating, eg |
1416 | * working closely on the same area of the disk. In that case, | |
1417 | * we can group them together and don't waste time idling. | |
6d048f53 | 1418 | */ |
a36e71f9 JA |
1419 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1420 | if (!cfqq) | |
1421 | return NULL; | |
1422 | ||
df5fe3e8 JM |
1423 | /* |
1424 | * It only makes sense to merge sync queues. | |
1425 | */ | |
1426 | if (!cfq_cfqq_sync(cfqq)) | |
1427 | return NULL; | |
e6c5bc73 JM |
1428 | if (CFQQ_SEEKY(cfqq)) |
1429 | return NULL; | |
df5fe3e8 | 1430 | |
c0324a02 CZ |
1431 | /* |
1432 | * Do not merge queues of different priority classes | |
1433 | */ | |
1434 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
1435 | return NULL; | |
1436 | ||
a36e71f9 | 1437 | return cfqq; |
6d048f53 JA |
1438 | } |
1439 | ||
a6d44e98 CZ |
1440 | /* |
1441 | * Determine whether we should enforce idle window for this queue. | |
1442 | */ | |
1443 | ||
1444 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1445 | { | |
1446 | enum wl_prio_t prio = cfqq_prio(cfqq); | |
718eee05 | 1447 | struct cfq_rb_root *service_tree = cfqq->service_tree; |
a6d44e98 | 1448 | |
f04a6424 VG |
1449 | BUG_ON(!service_tree); |
1450 | BUG_ON(!service_tree->count); | |
1451 | ||
a6d44e98 CZ |
1452 | /* We never do for idle class queues. */ |
1453 | if (prio == IDLE_WORKLOAD) | |
1454 | return false; | |
1455 | ||
1456 | /* We do for queues that were marked with idle window flag. */ | |
1457 | if (cfq_cfqq_idle_window(cfqq)) | |
1458 | return true; | |
1459 | ||
1460 | /* | |
1461 | * Otherwise, we do only if they are the last ones | |
1462 | * in their service tree. | |
1463 | */ | |
f04a6424 | 1464 | return service_tree->count == 1; |
a6d44e98 CZ |
1465 | } |
1466 | ||
6d048f53 | 1467 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1468 | { |
1792669c | 1469 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1470 | struct cfq_io_context *cic; |
7b14e3b5 JA |
1471 | unsigned long sl; |
1472 | ||
a68bbddb | 1473 | /* |
f7d7b7a7 JA |
1474 | * SSD device without seek penalty, disable idling. But only do so |
1475 | * for devices that support queuing, otherwise we still have a problem | |
1476 | * with sync vs async workloads. | |
a68bbddb | 1477 | */ |
f7d7b7a7 | 1478 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
1479 | return; |
1480 | ||
dd67d051 | 1481 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 1482 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
1483 | |
1484 | /* | |
1485 | * idle is disabled, either manually or by past process history | |
1486 | */ | |
a6d44e98 | 1487 | if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq)) |
6d048f53 JA |
1488 | return; |
1489 | ||
7b679138 | 1490 | /* |
8e550632 | 1491 | * still active requests from this queue, don't idle |
7b679138 | 1492 | */ |
8e550632 | 1493 | if (cfqq->dispatched) |
7b679138 JA |
1494 | return; |
1495 | ||
22e2c507 JA |
1496 | /* |
1497 | * task has exited, don't wait | |
1498 | */ | |
206dc69b | 1499 | cic = cfqd->active_cic; |
66dac98e | 1500 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
1501 | return; |
1502 | ||
355b659c CZ |
1503 | /* |
1504 | * If our average think time is larger than the remaining time | |
1505 | * slice, then don't idle. This avoids overrunning the allotted | |
1506 | * time slice. | |
1507 | */ | |
1508 | if (sample_valid(cic->ttime_samples) && | |
1509 | (cfqq->slice_end - jiffies < cic->ttime_mean)) | |
1510 | return; | |
1511 | ||
3b18152c | 1512 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 1513 | |
6d048f53 | 1514 | sl = cfqd->cfq_slice_idle; |
206dc69b | 1515 | |
7b14e3b5 | 1516 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
9481ffdc | 1517 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); |
1da177e4 LT |
1518 | } |
1519 | ||
498d3aa2 JA |
1520 | /* |
1521 | * Move request from internal lists to the request queue dispatch list. | |
1522 | */ | |
165125e1 | 1523 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 1524 | { |
3ed9a296 | 1525 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1526 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 1527 | |
7b679138 JA |
1528 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
1529 | ||
06d21886 | 1530 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); |
5380a101 | 1531 | cfq_remove_request(rq); |
6d048f53 | 1532 | cfqq->dispatched++; |
5380a101 | 1533 | elv_dispatch_sort(q, rq); |
3ed9a296 JA |
1534 | |
1535 | if (cfq_cfqq_sync(cfqq)) | |
1536 | cfqd->sync_flight++; | |
1da177e4 LT |
1537 | } |
1538 | ||
1539 | /* | |
1540 | * return expired entry, or NULL to just start from scratch in rbtree | |
1541 | */ | |
febffd61 | 1542 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 | 1543 | { |
30996f40 | 1544 | struct request *rq = NULL; |
1da177e4 | 1545 | |
3b18152c | 1546 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 1547 | return NULL; |
cb887411 JA |
1548 | |
1549 | cfq_mark_cfqq_fifo_expire(cfqq); | |
1550 | ||
89850f7e JA |
1551 | if (list_empty(&cfqq->fifo)) |
1552 | return NULL; | |
1da177e4 | 1553 | |
89850f7e | 1554 | rq = rq_entry_fifo(cfqq->fifo.next); |
30996f40 | 1555 | if (time_before(jiffies, rq_fifo_time(rq))) |
7b679138 | 1556 | rq = NULL; |
1da177e4 | 1557 | |
30996f40 | 1558 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 1559 | return rq; |
1da177e4 LT |
1560 | } |
1561 | ||
22e2c507 JA |
1562 | static inline int |
1563 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1564 | { | |
1565 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 1566 | |
22e2c507 | 1567 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 1568 | |
22e2c507 | 1569 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
1570 | } |
1571 | ||
df5fe3e8 JM |
1572 | /* |
1573 | * Must be called with the queue_lock held. | |
1574 | */ | |
1575 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
1576 | { | |
1577 | int process_refs, io_refs; | |
1578 | ||
1579 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
1580 | process_refs = atomic_read(&cfqq->ref) - io_refs; | |
1581 | BUG_ON(process_refs < 0); | |
1582 | return process_refs; | |
1583 | } | |
1584 | ||
1585 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
1586 | { | |
e6c5bc73 | 1587 | int process_refs, new_process_refs; |
df5fe3e8 JM |
1588 | struct cfq_queue *__cfqq; |
1589 | ||
1590 | /* Avoid a circular list and skip interim queue merges */ | |
1591 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
1592 | if (__cfqq == cfqq) | |
1593 | return; | |
1594 | new_cfqq = __cfqq; | |
1595 | } | |
1596 | ||
1597 | process_refs = cfqq_process_refs(cfqq); | |
1598 | /* | |
1599 | * If the process for the cfqq has gone away, there is no | |
1600 | * sense in merging the queues. | |
1601 | */ | |
1602 | if (process_refs == 0) | |
1603 | return; | |
1604 | ||
e6c5bc73 JM |
1605 | /* |
1606 | * Merge in the direction of the lesser amount of work. | |
1607 | */ | |
1608 | new_process_refs = cfqq_process_refs(new_cfqq); | |
1609 | if (new_process_refs >= process_refs) { | |
1610 | cfqq->new_cfqq = new_cfqq; | |
1611 | atomic_add(process_refs, &new_cfqq->ref); | |
1612 | } else { | |
1613 | new_cfqq->new_cfqq = cfqq; | |
1614 | atomic_add(new_process_refs, &cfqq->ref); | |
1615 | } | |
df5fe3e8 JM |
1616 | } |
1617 | ||
cdb16e8f VG |
1618 | static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, |
1619 | struct cfq_group *cfqg, enum wl_prio_t prio, | |
1620 | bool prio_changed) | |
718eee05 CZ |
1621 | { |
1622 | struct cfq_queue *queue; | |
1623 | int i; | |
1624 | bool key_valid = false; | |
1625 | unsigned long lowest_key = 0; | |
1626 | enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; | |
1627 | ||
1628 | if (prio_changed) { | |
1629 | /* | |
1630 | * When priorities switched, we prefer starting | |
1631 | * from SYNC_NOIDLE (first choice), or just SYNC | |
1632 | * over ASYNC | |
1633 | */ | |
cdb16e8f | 1634 | if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) |
718eee05 CZ |
1635 | return cur_best; |
1636 | cur_best = SYNC_WORKLOAD; | |
cdb16e8f | 1637 | if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) |
718eee05 CZ |
1638 | return cur_best; |
1639 | ||
1640 | return ASYNC_WORKLOAD; | |
1641 | } | |
1642 | ||
1643 | for (i = 0; i < 3; ++i) { | |
1644 | /* otherwise, select the one with lowest rb_key */ | |
cdb16e8f | 1645 | queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd)); |
718eee05 CZ |
1646 | if (queue && |
1647 | (!key_valid || time_before(queue->rb_key, lowest_key))) { | |
1648 | lowest_key = queue->rb_key; | |
1649 | cur_best = i; | |
1650 | key_valid = true; | |
1651 | } | |
1652 | } | |
1653 | ||
1654 | return cur_best; | |
1655 | } | |
1656 | ||
cdb16e8f | 1657 | static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) |
718eee05 CZ |
1658 | { |
1659 | enum wl_prio_t previous_prio = cfqd->serving_prio; | |
1660 | bool prio_changed; | |
1661 | unsigned slice; | |
1662 | unsigned count; | |
cdb16e8f | 1663 | struct cfq_rb_root *st; |
58ff82f3 | 1664 | unsigned group_slice; |
718eee05 | 1665 | |
1fa8f6d6 VG |
1666 | if (!cfqg) { |
1667 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1668 | cfqd->workload_expires = jiffies + 1; | |
1669 | return; | |
1670 | } | |
1671 | ||
718eee05 | 1672 | /* Choose next priority. RT > BE > IDLE */ |
58ff82f3 | 1673 | if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) |
718eee05 | 1674 | cfqd->serving_prio = RT_WORKLOAD; |
58ff82f3 | 1675 | else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) |
718eee05 CZ |
1676 | cfqd->serving_prio = BE_WORKLOAD; |
1677 | else { | |
1678 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1679 | cfqd->workload_expires = jiffies + 1; | |
1680 | return; | |
1681 | } | |
1682 | ||
1683 | /* | |
1684 | * For RT and BE, we have to choose also the type | |
1685 | * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload | |
1686 | * expiration time | |
1687 | */ | |
1688 | prio_changed = (cfqd->serving_prio != previous_prio); | |
cdb16e8f VG |
1689 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, |
1690 | cfqd); | |
1691 | count = st->count; | |
718eee05 CZ |
1692 | |
1693 | /* | |
1694 | * If priority didn't change, check workload expiration, | |
1695 | * and that we still have other queues ready | |
1696 | */ | |
1697 | if (!prio_changed && count && | |
1698 | !time_after(jiffies, cfqd->workload_expires)) | |
1699 | return; | |
1700 | ||
1701 | /* otherwise select new workload type */ | |
1702 | cfqd->serving_type = | |
cdb16e8f VG |
1703 | cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed); |
1704 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, | |
1705 | cfqd); | |
1706 | count = st->count; | |
718eee05 CZ |
1707 | |
1708 | /* | |
1709 | * the workload slice is computed as a fraction of target latency | |
1710 | * proportional to the number of queues in that workload, over | |
1711 | * all the queues in the same priority class | |
1712 | */ | |
58ff82f3 VG |
1713 | group_slice = cfq_group_slice(cfqd, cfqg); |
1714 | ||
1715 | slice = group_slice * count / | |
1716 | max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], | |
1717 | cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); | |
718eee05 CZ |
1718 | |
1719 | if (cfqd->serving_type == ASYNC_WORKLOAD) | |
1720 | /* async workload slice is scaled down according to | |
1721 | * the sync/async slice ratio. */ | |
1722 | slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; | |
1723 | else | |
1724 | /* sync workload slice is at least 2 * cfq_slice_idle */ | |
1725 | slice = max(slice, 2 * cfqd->cfq_slice_idle); | |
1726 | ||
1727 | slice = max_t(unsigned, slice, CFQ_MIN_TT); | |
1728 | cfqd->workload_expires = jiffies + slice; | |
8e550632 | 1729 | cfqd->noidle_tree_requires_idle = false; |
718eee05 CZ |
1730 | } |
1731 | ||
1fa8f6d6 VG |
1732 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) |
1733 | { | |
1734 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
25bc6b07 | 1735 | struct cfq_group *cfqg; |
1fa8f6d6 VG |
1736 | |
1737 | if (RB_EMPTY_ROOT(&st->rb)) | |
1738 | return NULL; | |
25bc6b07 VG |
1739 | cfqg = cfq_rb_first_group(st); |
1740 | st->active = &cfqg->rb_node; | |
1741 | update_min_vdisktime(st); | |
1742 | return cfqg; | |
1fa8f6d6 VG |
1743 | } |
1744 | ||
cdb16e8f VG |
1745 | static void cfq_choose_cfqg(struct cfq_data *cfqd) |
1746 | { | |
1fa8f6d6 VG |
1747 | struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); |
1748 | ||
1749 | cfqd->serving_group = cfqg; | |
1750 | choose_service_tree(cfqd, cfqg); | |
cdb16e8f VG |
1751 | } |
1752 | ||
22e2c507 | 1753 | /* |
498d3aa2 JA |
1754 | * Select a queue for service. If we have a current active queue, |
1755 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 1756 | */ |
1b5ed5e1 | 1757 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 1758 | { |
a36e71f9 | 1759 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 1760 | |
22e2c507 JA |
1761 | cfqq = cfqd->active_queue; |
1762 | if (!cfqq) | |
1763 | goto new_queue; | |
1da177e4 | 1764 | |
f04a6424 VG |
1765 | if (!cfqd->rq_queued) |
1766 | return NULL; | |
22e2c507 | 1767 | /* |
6d048f53 | 1768 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 1769 | */ |
b029195d | 1770 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) |
3b18152c | 1771 | goto expire; |
1da177e4 | 1772 | |
22e2c507 | 1773 | /* |
6d048f53 JA |
1774 | * The active queue has requests and isn't expired, allow it to |
1775 | * dispatch. | |
22e2c507 | 1776 | */ |
dd67d051 | 1777 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 1778 | goto keep_queue; |
6d048f53 | 1779 | |
a36e71f9 JA |
1780 | /* |
1781 | * If another queue has a request waiting within our mean seek | |
1782 | * distance, let it run. The expire code will check for close | |
1783 | * cooperators and put the close queue at the front of the service | |
df5fe3e8 | 1784 | * tree. If possible, merge the expiring queue with the new cfqq. |
a36e71f9 | 1785 | */ |
b3b6d040 | 1786 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); |
df5fe3e8 JM |
1787 | if (new_cfqq) { |
1788 | if (!cfqq->new_cfqq) | |
1789 | cfq_setup_merge(cfqq, new_cfqq); | |
a36e71f9 | 1790 | goto expire; |
df5fe3e8 | 1791 | } |
a36e71f9 | 1792 | |
6d048f53 JA |
1793 | /* |
1794 | * No requests pending. If the active queue still has requests in | |
1795 | * flight or is idling for a new request, allow either of these | |
1796 | * conditions to happen (or time out) before selecting a new queue. | |
1797 | */ | |
cc197479 | 1798 | if (timer_pending(&cfqd->idle_slice_timer) || |
a6d44e98 | 1799 | (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) { |
caaa5f9f JA |
1800 | cfqq = NULL; |
1801 | goto keep_queue; | |
22e2c507 JA |
1802 | } |
1803 | ||
3b18152c | 1804 | expire: |
6084cdda | 1805 | cfq_slice_expired(cfqd, 0); |
3b18152c | 1806 | new_queue: |
718eee05 CZ |
1807 | /* |
1808 | * Current queue expired. Check if we have to switch to a new | |
1809 | * service tree | |
1810 | */ | |
1811 | if (!new_cfqq) | |
cdb16e8f | 1812 | cfq_choose_cfqg(cfqd); |
718eee05 | 1813 | |
a36e71f9 | 1814 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 1815 | keep_queue: |
3b18152c | 1816 | return cfqq; |
22e2c507 JA |
1817 | } |
1818 | ||
febffd61 | 1819 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
1820 | { |
1821 | int dispatched = 0; | |
1822 | ||
1823 | while (cfqq->next_rq) { | |
1824 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
1825 | dispatched++; | |
1826 | } | |
1827 | ||
1828 | BUG_ON(!list_empty(&cfqq->fifo)); | |
f04a6424 VG |
1829 | |
1830 | /* By default cfqq is not expired if it is empty. Do it explicitly */ | |
1831 | __cfq_slice_expired(cfqq->cfqd, cfqq, 0); | |
d9e7620e JA |
1832 | return dispatched; |
1833 | } | |
1834 | ||
498d3aa2 JA |
1835 | /* |
1836 | * Drain our current requests. Used for barriers and when switching | |
1837 | * io schedulers on-the-fly. | |
1838 | */ | |
d9e7620e | 1839 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 1840 | { |
0871714e | 1841 | struct cfq_queue *cfqq; |
d9e7620e | 1842 | int dispatched = 0; |
cdb16e8f | 1843 | |
f04a6424 VG |
1844 | while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) |
1845 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); | |
1b5ed5e1 | 1846 | |
6084cdda | 1847 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
1848 | BUG_ON(cfqd->busy_queues); |
1849 | ||
6923715a | 1850 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
1851 | return dispatched; |
1852 | } | |
1853 | ||
0b182d61 | 1854 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
2f5cb738 | 1855 | { |
2f5cb738 | 1856 | unsigned int max_dispatch; |
22e2c507 | 1857 | |
5ad531db JA |
1858 | /* |
1859 | * Drain async requests before we start sync IO | |
1860 | */ | |
a6d44e98 | 1861 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC]) |
0b182d61 | 1862 | return false; |
5ad531db | 1863 | |
2f5cb738 JA |
1864 | /* |
1865 | * If this is an async queue and we have sync IO in flight, let it wait | |
1866 | */ | |
1867 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) | |
0b182d61 | 1868 | return false; |
2f5cb738 JA |
1869 | |
1870 | max_dispatch = cfqd->cfq_quantum; | |
1871 | if (cfq_class_idle(cfqq)) | |
1872 | max_dispatch = 1; | |
b4878f24 | 1873 | |
2f5cb738 JA |
1874 | /* |
1875 | * Does this cfqq already have too much IO in flight? | |
1876 | */ | |
1877 | if (cfqq->dispatched >= max_dispatch) { | |
1878 | /* | |
1879 | * idle queue must always only have a single IO in flight | |
1880 | */ | |
3ed9a296 | 1881 | if (cfq_class_idle(cfqq)) |
0b182d61 | 1882 | return false; |
3ed9a296 | 1883 | |
2f5cb738 JA |
1884 | /* |
1885 | * We have other queues, don't allow more IO from this one | |
1886 | */ | |
1887 | if (cfqd->busy_queues > 1) | |
0b182d61 | 1888 | return false; |
9ede209e | 1889 | |
365722bb | 1890 | /* |
474b18cc | 1891 | * Sole queue user, no limit |
365722bb | 1892 | */ |
474b18cc | 1893 | max_dispatch = -1; |
8e296755 JA |
1894 | } |
1895 | ||
1896 | /* | |
1897 | * Async queues must wait a bit before being allowed dispatch. | |
1898 | * We also ramp up the dispatch depth gradually for async IO, | |
1899 | * based on the last sync IO we serviced | |
1900 | */ | |
963b72fc | 1901 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { |
8e296755 JA |
1902 | unsigned long last_sync = jiffies - cfqd->last_end_sync_rq; |
1903 | unsigned int depth; | |
365722bb | 1904 | |
61f0c1dc | 1905 | depth = last_sync / cfqd->cfq_slice[1]; |
e00c54c3 JA |
1906 | if (!depth && !cfqq->dispatched) |
1907 | depth = 1; | |
8e296755 JA |
1908 | if (depth < max_dispatch) |
1909 | max_dispatch = depth; | |
2f5cb738 | 1910 | } |
3ed9a296 | 1911 | |
0b182d61 JA |
1912 | /* |
1913 | * If we're below the current max, allow a dispatch | |
1914 | */ | |
1915 | return cfqq->dispatched < max_dispatch; | |
1916 | } | |
1917 | ||
1918 | /* | |
1919 | * Dispatch a request from cfqq, moving them to the request queue | |
1920 | * dispatch list. | |
1921 | */ | |
1922 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1923 | { | |
1924 | struct request *rq; | |
1925 | ||
1926 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
1927 | ||
1928 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
1929 | return false; | |
1930 | ||
1931 | /* | |
1932 | * follow expired path, else get first next available | |
1933 | */ | |
1934 | rq = cfq_check_fifo(cfqq); | |
1935 | if (!rq) | |
1936 | rq = cfqq->next_rq; | |
1937 | ||
1938 | /* | |
1939 | * insert request into driver dispatch list | |
1940 | */ | |
1941 | cfq_dispatch_insert(cfqd->queue, rq); | |
1942 | ||
1943 | if (!cfqd->active_cic) { | |
1944 | struct cfq_io_context *cic = RQ_CIC(rq); | |
1945 | ||
1946 | atomic_long_inc(&cic->ioc->refcount); | |
1947 | cfqd->active_cic = cic; | |
1948 | } | |
1949 | ||
1950 | return true; | |
1951 | } | |
1952 | ||
1953 | /* | |
1954 | * Find the cfqq that we need to service and move a request from that to the | |
1955 | * dispatch list | |
1956 | */ | |
1957 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
1958 | { | |
1959 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1960 | struct cfq_queue *cfqq; | |
1961 | ||
1962 | if (!cfqd->busy_queues) | |
1963 | return 0; | |
1964 | ||
1965 | if (unlikely(force)) | |
1966 | return cfq_forced_dispatch(cfqd); | |
1967 | ||
1968 | cfqq = cfq_select_queue(cfqd); | |
1969 | if (!cfqq) | |
8e296755 JA |
1970 | return 0; |
1971 | ||
2f5cb738 | 1972 | /* |
0b182d61 | 1973 | * Dispatch a request from this cfqq, if it is allowed |
2f5cb738 | 1974 | */ |
0b182d61 JA |
1975 | if (!cfq_dispatch_request(cfqd, cfqq)) |
1976 | return 0; | |
1977 | ||
2f5cb738 | 1978 | cfqq->slice_dispatch++; |
b029195d | 1979 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 1980 | |
2f5cb738 JA |
1981 | /* |
1982 | * expire an async queue immediately if it has used up its slice. idle | |
1983 | * queue always expire after 1 dispatch round. | |
1984 | */ | |
1985 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
1986 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
1987 | cfq_class_idle(cfqq))) { | |
1988 | cfqq->slice_end = jiffies + 1; | |
1989 | cfq_slice_expired(cfqd, 0); | |
1da177e4 LT |
1990 | } |
1991 | ||
b217a903 | 1992 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); |
2f5cb738 | 1993 | return 1; |
1da177e4 LT |
1994 | } |
1995 | ||
1da177e4 | 1996 | /* |
5e705374 JA |
1997 | * task holds one reference to the queue, dropped when task exits. each rq |
1998 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 LT |
1999 | * |
2000 | * queue lock must be held here. | |
2001 | */ | |
2002 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
2003 | { | |
22e2c507 JA |
2004 | struct cfq_data *cfqd = cfqq->cfqd; |
2005 | ||
2006 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
2007 | |
2008 | if (!atomic_dec_and_test(&cfqq->ref)) | |
2009 | return; | |
2010 | ||
7b679138 | 2011 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 2012 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 2013 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
1da177e4 | 2014 | |
28f95cbc | 2015 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 2016 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 2017 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 2018 | } |
22e2c507 | 2019 | |
f04a6424 | 2020 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 LT |
2021 | kmem_cache_free(cfq_pool, cfqq); |
2022 | } | |
2023 | ||
d6de8be7 JA |
2024 | /* |
2025 | * Must always be called with the rcu_read_lock() held | |
2026 | */ | |
07416d29 JA |
2027 | static void |
2028 | __call_for_each_cic(struct io_context *ioc, | |
2029 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
2030 | { | |
2031 | struct cfq_io_context *cic; | |
2032 | struct hlist_node *n; | |
2033 | ||
2034 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) | |
2035 | func(ioc, cic); | |
2036 | } | |
2037 | ||
4ac845a2 | 2038 | /* |
34e6bbf2 | 2039 | * Call func for each cic attached to this ioc. |
4ac845a2 | 2040 | */ |
34e6bbf2 | 2041 | static void |
4ac845a2 JA |
2042 | call_for_each_cic(struct io_context *ioc, |
2043 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1da177e4 | 2044 | { |
4ac845a2 | 2045 | rcu_read_lock(); |
07416d29 | 2046 | __call_for_each_cic(ioc, func); |
4ac845a2 | 2047 | rcu_read_unlock(); |
34e6bbf2 FC |
2048 | } |
2049 | ||
2050 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
2051 | { | |
2052 | struct cfq_io_context *cic; | |
2053 | ||
2054 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
2055 | ||
2056 | kmem_cache_free(cfq_ioc_pool, cic); | |
245b2e70 | 2057 | elv_ioc_count_dec(cfq_ioc_count); |
34e6bbf2 | 2058 | |
9a11b4ed JA |
2059 | if (ioc_gone) { |
2060 | /* | |
2061 | * CFQ scheduler is exiting, grab exit lock and check | |
2062 | * the pending io context count. If it hits zero, | |
2063 | * complete ioc_gone and set it back to NULL | |
2064 | */ | |
2065 | spin_lock(&ioc_gone_lock); | |
245b2e70 | 2066 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { |
9a11b4ed JA |
2067 | complete(ioc_gone); |
2068 | ioc_gone = NULL; | |
2069 | } | |
2070 | spin_unlock(&ioc_gone_lock); | |
2071 | } | |
34e6bbf2 | 2072 | } |
4ac845a2 | 2073 | |
34e6bbf2 FC |
2074 | static void cfq_cic_free(struct cfq_io_context *cic) |
2075 | { | |
2076 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
2077 | } |
2078 | ||
2079 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
2080 | { | |
2081 | unsigned long flags; | |
2082 | ||
2083 | BUG_ON(!cic->dead_key); | |
2084 | ||
2085 | spin_lock_irqsave(&ioc->lock, flags); | |
2086 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
ffc4e759 | 2087 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2088 | spin_unlock_irqrestore(&ioc->lock, flags); |
2089 | ||
34e6bbf2 | 2090 | cfq_cic_free(cic); |
4ac845a2 JA |
2091 | } |
2092 | ||
d6de8be7 JA |
2093 | /* |
2094 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
2095 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
2096 | * and ->trim() which is called with the task lock held | |
2097 | */ | |
4ac845a2 JA |
2098 | static void cfq_free_io_context(struct io_context *ioc) |
2099 | { | |
4ac845a2 | 2100 | /* |
34e6bbf2 FC |
2101 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
2102 | * so no more cic's are allowed to be linked into this ioc. So it | |
2103 | * should be ok to iterate over the known list, we will see all cic's | |
2104 | * since no new ones are added. | |
4ac845a2 | 2105 | */ |
07416d29 | 2106 | __call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
2107 | } |
2108 | ||
89850f7e | 2109 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 2110 | { |
df5fe3e8 JM |
2111 | struct cfq_queue *__cfqq, *next; |
2112 | ||
28f95cbc | 2113 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 2114 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 2115 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 2116 | } |
22e2c507 | 2117 | |
df5fe3e8 JM |
2118 | /* |
2119 | * If this queue was scheduled to merge with another queue, be | |
2120 | * sure to drop the reference taken on that queue (and others in | |
2121 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
2122 | */ | |
2123 | __cfqq = cfqq->new_cfqq; | |
2124 | while (__cfqq) { | |
2125 | if (__cfqq == cfqq) { | |
2126 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
2127 | break; | |
2128 | } | |
2129 | next = __cfqq->new_cfqq; | |
2130 | cfq_put_queue(__cfqq); | |
2131 | __cfqq = next; | |
2132 | } | |
2133 | ||
89850f7e JA |
2134 | cfq_put_queue(cfqq); |
2135 | } | |
22e2c507 | 2136 | |
89850f7e JA |
2137 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
2138 | struct cfq_io_context *cic) | |
2139 | { | |
4faa3c81 FC |
2140 | struct io_context *ioc = cic->ioc; |
2141 | ||
fc46379d | 2142 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
2143 | |
2144 | /* | |
2145 | * Make sure key == NULL is seen for dead queues | |
2146 | */ | |
fc46379d | 2147 | smp_wmb(); |
4ac845a2 | 2148 | cic->dead_key = (unsigned long) cic->key; |
fc46379d JA |
2149 | cic->key = NULL; |
2150 | ||
4faa3c81 FC |
2151 | if (ioc->ioc_data == cic) |
2152 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
2153 | ||
ff6657c6 JA |
2154 | if (cic->cfqq[BLK_RW_ASYNC]) { |
2155 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
2156 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
2157 | } |
2158 | ||
ff6657c6 JA |
2159 | if (cic->cfqq[BLK_RW_SYNC]) { |
2160 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
2161 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 2162 | } |
89850f7e JA |
2163 | } |
2164 | ||
4ac845a2 JA |
2165 | static void cfq_exit_single_io_context(struct io_context *ioc, |
2166 | struct cfq_io_context *cic) | |
89850f7e JA |
2167 | { |
2168 | struct cfq_data *cfqd = cic->key; | |
2169 | ||
89850f7e | 2170 | if (cfqd) { |
165125e1 | 2171 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 2172 | unsigned long flags; |
89850f7e | 2173 | |
4ac845a2 | 2174 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
2175 | |
2176 | /* | |
2177 | * Ensure we get a fresh copy of the ->key to prevent | |
2178 | * race between exiting task and queue | |
2179 | */ | |
2180 | smp_read_barrier_depends(); | |
2181 | if (cic->key) | |
2182 | __cfq_exit_single_io_context(cfqd, cic); | |
2183 | ||
4ac845a2 | 2184 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 2185 | } |
1da177e4 LT |
2186 | } |
2187 | ||
498d3aa2 JA |
2188 | /* |
2189 | * The process that ioc belongs to has exited, we need to clean up | |
2190 | * and put the internal structures we have that belongs to that process. | |
2191 | */ | |
e2d74ac0 | 2192 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 2193 | { |
4ac845a2 | 2194 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
2195 | } |
2196 | ||
22e2c507 | 2197 | static struct cfq_io_context * |
8267e268 | 2198 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2199 | { |
b5deef90 | 2200 | struct cfq_io_context *cic; |
1da177e4 | 2201 | |
94f6030c CL |
2202 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
2203 | cfqd->queue->node); | |
1da177e4 | 2204 | if (cic) { |
22e2c507 | 2205 | cic->last_end_request = jiffies; |
553698f9 | 2206 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 2207 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
2208 | cic->dtor = cfq_free_io_context; |
2209 | cic->exit = cfq_exit_io_context; | |
245b2e70 | 2210 | elv_ioc_count_inc(cfq_ioc_count); |
1da177e4 LT |
2211 | } |
2212 | ||
2213 | return cic; | |
2214 | } | |
2215 | ||
fd0928df | 2216 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
2217 | { |
2218 | struct task_struct *tsk = current; | |
2219 | int ioprio_class; | |
2220 | ||
3b18152c | 2221 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
2222 | return; |
2223 | ||
fd0928df | 2224 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 2225 | switch (ioprio_class) { |
fe094d98 JA |
2226 | default: |
2227 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
2228 | case IOPRIO_CLASS_NONE: | |
2229 | /* | |
6d63c275 | 2230 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
2231 | */ |
2232 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 2233 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
2234 | break; |
2235 | case IOPRIO_CLASS_RT: | |
2236 | cfqq->ioprio = task_ioprio(ioc); | |
2237 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
2238 | break; | |
2239 | case IOPRIO_CLASS_BE: | |
2240 | cfqq->ioprio = task_ioprio(ioc); | |
2241 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2242 | break; | |
2243 | case IOPRIO_CLASS_IDLE: | |
2244 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
2245 | cfqq->ioprio = 7; | |
2246 | cfq_clear_cfqq_idle_window(cfqq); | |
2247 | break; | |
22e2c507 JA |
2248 | } |
2249 | ||
2250 | /* | |
2251 | * keep track of original prio settings in case we have to temporarily | |
2252 | * elevate the priority of this queue | |
2253 | */ | |
2254 | cfqq->org_ioprio = cfqq->ioprio; | |
2255 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 2256 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
2257 | } |
2258 | ||
febffd61 | 2259 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 2260 | { |
478a82b0 AV |
2261 | struct cfq_data *cfqd = cic->key; |
2262 | struct cfq_queue *cfqq; | |
c1b707d2 | 2263 | unsigned long flags; |
35e6077c | 2264 | |
caaa5f9f JA |
2265 | if (unlikely(!cfqd)) |
2266 | return; | |
2267 | ||
c1b707d2 | 2268 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 2269 | |
ff6657c6 | 2270 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
2271 | if (cfqq) { |
2272 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
2273 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
2274 | GFP_ATOMIC); | |
caaa5f9f | 2275 | if (new_cfqq) { |
ff6657c6 | 2276 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
2277 | cfq_put_queue(cfqq); |
2278 | } | |
22e2c507 | 2279 | } |
caaa5f9f | 2280 | |
ff6657c6 | 2281 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
2282 | if (cfqq) |
2283 | cfq_mark_cfqq_prio_changed(cfqq); | |
2284 | ||
c1b707d2 | 2285 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
2286 | } |
2287 | ||
fc46379d | 2288 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 2289 | { |
4ac845a2 | 2290 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 2291 | ioc->ioprio_changed = 0; |
22e2c507 JA |
2292 | } |
2293 | ||
d5036d77 | 2294 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 2295 | pid_t pid, bool is_sync) |
d5036d77 JA |
2296 | { |
2297 | RB_CLEAR_NODE(&cfqq->rb_node); | |
2298 | RB_CLEAR_NODE(&cfqq->p_node); | |
2299 | INIT_LIST_HEAD(&cfqq->fifo); | |
2300 | ||
2301 | atomic_set(&cfqq->ref, 0); | |
2302 | cfqq->cfqd = cfqd; | |
2303 | ||
2304 | cfq_mark_cfqq_prio_changed(cfqq); | |
2305 | ||
2306 | if (is_sync) { | |
2307 | if (!cfq_class_idle(cfqq)) | |
2308 | cfq_mark_cfqq_idle_window(cfqq); | |
2309 | cfq_mark_cfqq_sync(cfqq); | |
2310 | } | |
2311 | cfqq->pid = pid; | |
2312 | } | |
2313 | ||
cdb16e8f VG |
2314 | static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) |
2315 | { | |
2316 | cfqq->cfqg = cfqg; | |
2317 | } | |
2318 | ||
2319 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create) | |
2320 | { | |
2321 | return &cfqd->root_group; | |
2322 | } | |
2323 | ||
22e2c507 | 2324 | static struct cfq_queue * |
a6151c3a | 2325 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, |
fd0928df | 2326 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 2327 | { |
22e2c507 | 2328 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 2329 | struct cfq_io_context *cic; |
cdb16e8f | 2330 | struct cfq_group *cfqg; |
22e2c507 JA |
2331 | |
2332 | retry: | |
cdb16e8f | 2333 | cfqg = cfq_get_cfqg(cfqd, 1); |
4ac845a2 | 2334 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
2335 | /* cic always exists here */ |
2336 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 2337 | |
6118b70b JA |
2338 | /* |
2339 | * Always try a new alloc if we fell back to the OOM cfqq | |
2340 | * originally, since it should just be a temporary situation. | |
2341 | */ | |
2342 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
2343 | cfqq = NULL; | |
22e2c507 JA |
2344 | if (new_cfqq) { |
2345 | cfqq = new_cfqq; | |
2346 | new_cfqq = NULL; | |
2347 | } else if (gfp_mask & __GFP_WAIT) { | |
2348 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 2349 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 2350 | gfp_mask | __GFP_ZERO, |
94f6030c | 2351 | cfqd->queue->node); |
22e2c507 | 2352 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
2353 | if (new_cfqq) |
2354 | goto retry; | |
22e2c507 | 2355 | } else { |
94f6030c CL |
2356 | cfqq = kmem_cache_alloc_node(cfq_pool, |
2357 | gfp_mask | __GFP_ZERO, | |
2358 | cfqd->queue->node); | |
22e2c507 JA |
2359 | } |
2360 | ||
6118b70b JA |
2361 | if (cfqq) { |
2362 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
2363 | cfq_init_prio_data(cfqq, ioc); | |
cdb16e8f | 2364 | cfq_link_cfqq_cfqg(cfqq, cfqg); |
6118b70b JA |
2365 | cfq_log_cfqq(cfqd, cfqq, "alloced"); |
2366 | } else | |
2367 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
2368 | } |
2369 | ||
2370 | if (new_cfqq) | |
2371 | kmem_cache_free(cfq_pool, new_cfqq); | |
2372 | ||
22e2c507 JA |
2373 | return cfqq; |
2374 | } | |
2375 | ||
c2dea2d1 VT |
2376 | static struct cfq_queue ** |
2377 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
2378 | { | |
fe094d98 | 2379 | switch (ioprio_class) { |
c2dea2d1 VT |
2380 | case IOPRIO_CLASS_RT: |
2381 | return &cfqd->async_cfqq[0][ioprio]; | |
2382 | case IOPRIO_CLASS_BE: | |
2383 | return &cfqd->async_cfqq[1][ioprio]; | |
2384 | case IOPRIO_CLASS_IDLE: | |
2385 | return &cfqd->async_idle_cfqq; | |
2386 | default: | |
2387 | BUG(); | |
2388 | } | |
2389 | } | |
2390 | ||
15c31be4 | 2391 | static struct cfq_queue * |
a6151c3a | 2392 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, |
15c31be4 JA |
2393 | gfp_t gfp_mask) |
2394 | { | |
fd0928df JA |
2395 | const int ioprio = task_ioprio(ioc); |
2396 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 2397 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
2398 | struct cfq_queue *cfqq = NULL; |
2399 | ||
c2dea2d1 VT |
2400 | if (!is_sync) { |
2401 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
2402 | cfqq = *async_cfqq; | |
2403 | } | |
2404 | ||
6118b70b | 2405 | if (!cfqq) |
fd0928df | 2406 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
2407 | |
2408 | /* | |
2409 | * pin the queue now that it's allocated, scheduler exit will prune it | |
2410 | */ | |
c2dea2d1 | 2411 | if (!is_sync && !(*async_cfqq)) { |
15c31be4 | 2412 | atomic_inc(&cfqq->ref); |
c2dea2d1 | 2413 | *async_cfqq = cfqq; |
15c31be4 JA |
2414 | } |
2415 | ||
2416 | atomic_inc(&cfqq->ref); | |
2417 | return cfqq; | |
2418 | } | |
2419 | ||
498d3aa2 JA |
2420 | /* |
2421 | * We drop cfq io contexts lazily, so we may find a dead one. | |
2422 | */ | |
dbecf3ab | 2423 | static void |
4ac845a2 JA |
2424 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
2425 | struct cfq_io_context *cic) | |
dbecf3ab | 2426 | { |
4ac845a2 JA |
2427 | unsigned long flags; |
2428 | ||
fc46379d | 2429 | WARN_ON(!list_empty(&cic->queue_list)); |
597bc485 | 2430 | |
4ac845a2 JA |
2431 | spin_lock_irqsave(&ioc->lock, flags); |
2432 | ||
4faa3c81 | 2433 | BUG_ON(ioc->ioc_data == cic); |
597bc485 | 2434 | |
4ac845a2 | 2435 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); |
ffc4e759 | 2436 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2437 | spin_unlock_irqrestore(&ioc->lock, flags); |
2438 | ||
2439 | cfq_cic_free(cic); | |
dbecf3ab OH |
2440 | } |
2441 | ||
e2d74ac0 | 2442 | static struct cfq_io_context * |
4ac845a2 | 2443 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 2444 | { |
e2d74ac0 | 2445 | struct cfq_io_context *cic; |
d6de8be7 | 2446 | unsigned long flags; |
4ac845a2 | 2447 | void *k; |
e2d74ac0 | 2448 | |
91fac317 VT |
2449 | if (unlikely(!ioc)) |
2450 | return NULL; | |
2451 | ||
d6de8be7 JA |
2452 | rcu_read_lock(); |
2453 | ||
597bc485 JA |
2454 | /* |
2455 | * we maintain a last-hit cache, to avoid browsing over the tree | |
2456 | */ | |
4ac845a2 | 2457 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
2458 | if (cic && cic->key == cfqd) { |
2459 | rcu_read_unlock(); | |
597bc485 | 2460 | return cic; |
d6de8be7 | 2461 | } |
597bc485 | 2462 | |
4ac845a2 | 2463 | do { |
4ac845a2 JA |
2464 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); |
2465 | rcu_read_unlock(); | |
2466 | if (!cic) | |
2467 | break; | |
be3b0753 OH |
2468 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
2469 | k = cic->key; | |
2470 | if (unlikely(!k)) { | |
4ac845a2 | 2471 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 2472 | rcu_read_lock(); |
4ac845a2 | 2473 | continue; |
dbecf3ab | 2474 | } |
e2d74ac0 | 2475 | |
d6de8be7 | 2476 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 2477 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 2478 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
2479 | break; |
2480 | } while (1); | |
e2d74ac0 | 2481 | |
4ac845a2 | 2482 | return cic; |
e2d74ac0 JA |
2483 | } |
2484 | ||
4ac845a2 JA |
2485 | /* |
2486 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
2487 | * the process specific cfq io context when entered from the block layer. | |
2488 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
2489 | */ | |
febffd61 JA |
2490 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
2491 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 2492 | { |
0261d688 | 2493 | unsigned long flags; |
4ac845a2 | 2494 | int ret; |
e2d74ac0 | 2495 | |
4ac845a2 JA |
2496 | ret = radix_tree_preload(gfp_mask); |
2497 | if (!ret) { | |
2498 | cic->ioc = ioc; | |
2499 | cic->key = cfqd; | |
e2d74ac0 | 2500 | |
4ac845a2 JA |
2501 | spin_lock_irqsave(&ioc->lock, flags); |
2502 | ret = radix_tree_insert(&ioc->radix_root, | |
2503 | (unsigned long) cfqd, cic); | |
ffc4e759 JA |
2504 | if (!ret) |
2505 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 2506 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 2507 | |
4ac845a2 JA |
2508 | radix_tree_preload_end(); |
2509 | ||
2510 | if (!ret) { | |
2511 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
2512 | list_add(&cic->queue_list, &cfqd->cic_list); | |
2513 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2514 | } | |
e2d74ac0 JA |
2515 | } |
2516 | ||
4ac845a2 JA |
2517 | if (ret) |
2518 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 2519 | |
4ac845a2 | 2520 | return ret; |
e2d74ac0 JA |
2521 | } |
2522 | ||
1da177e4 LT |
2523 | /* |
2524 | * Setup general io context and cfq io context. There can be several cfq | |
2525 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 2526 | * than one device managed by cfq. |
1da177e4 LT |
2527 | */ |
2528 | static struct cfq_io_context * | |
e2d74ac0 | 2529 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2530 | { |
22e2c507 | 2531 | struct io_context *ioc = NULL; |
1da177e4 | 2532 | struct cfq_io_context *cic; |
1da177e4 | 2533 | |
22e2c507 | 2534 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 2535 | |
b5deef90 | 2536 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
2537 | if (!ioc) |
2538 | return NULL; | |
2539 | ||
4ac845a2 | 2540 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
2541 | if (cic) |
2542 | goto out; | |
1da177e4 | 2543 | |
e2d74ac0 JA |
2544 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
2545 | if (cic == NULL) | |
2546 | goto err; | |
1da177e4 | 2547 | |
4ac845a2 JA |
2548 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
2549 | goto err_free; | |
2550 | ||
1da177e4 | 2551 | out: |
fc46379d JA |
2552 | smp_read_barrier_depends(); |
2553 | if (unlikely(ioc->ioprio_changed)) | |
2554 | cfq_ioc_set_ioprio(ioc); | |
2555 | ||
1da177e4 | 2556 | return cic; |
4ac845a2 JA |
2557 | err_free: |
2558 | cfq_cic_free(cic); | |
1da177e4 LT |
2559 | err: |
2560 | put_io_context(ioc); | |
2561 | return NULL; | |
2562 | } | |
2563 | ||
22e2c507 JA |
2564 | static void |
2565 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 2566 | { |
aaf1228d JA |
2567 | unsigned long elapsed = jiffies - cic->last_end_request; |
2568 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 2569 | |
22e2c507 JA |
2570 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
2571 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
2572 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
2573 | } | |
1da177e4 | 2574 | |
206dc69b | 2575 | static void |
b2c18e1e | 2576 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
6d048f53 | 2577 | struct request *rq) |
206dc69b JA |
2578 | { |
2579 | sector_t sdist; | |
2580 | u64 total; | |
2581 | ||
b2c18e1e | 2582 | if (!cfqq->last_request_pos) |
4d00aa47 | 2583 | sdist = 0; |
b2c18e1e JM |
2584 | else if (cfqq->last_request_pos < blk_rq_pos(rq)) |
2585 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
206dc69b | 2586 | else |
b2c18e1e | 2587 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); |
206dc69b JA |
2588 | |
2589 | /* | |
2590 | * Don't allow the seek distance to get too large from the | |
2591 | * odd fragment, pagein, etc | |
2592 | */ | |
b2c18e1e JM |
2593 | if (cfqq->seek_samples <= 60) /* second&third seek */ |
2594 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024); | |
206dc69b | 2595 | else |
b2c18e1e | 2596 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64); |
206dc69b | 2597 | |
b2c18e1e JM |
2598 | cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8; |
2599 | cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8; | |
2600 | total = cfqq->seek_total + (cfqq->seek_samples/2); | |
2601 | do_div(total, cfqq->seek_samples); | |
2602 | cfqq->seek_mean = (sector_t)total; | |
e6c5bc73 JM |
2603 | |
2604 | /* | |
2605 | * If this cfqq is shared between multiple processes, check to | |
2606 | * make sure that those processes are still issuing I/Os within | |
2607 | * the mean seek distance. If not, it may be time to break the | |
2608 | * queues apart again. | |
2609 | */ | |
2610 | if (cfq_cfqq_coop(cfqq)) { | |
2611 | if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start) | |
2612 | cfqq->seeky_start = jiffies; | |
2613 | else if (!CFQQ_SEEKY(cfqq)) | |
2614 | cfqq->seeky_start = 0; | |
2615 | } | |
206dc69b | 2616 | } |
1da177e4 | 2617 | |
22e2c507 JA |
2618 | /* |
2619 | * Disable idle window if the process thinks too long or seeks so much that | |
2620 | * it doesn't matter | |
2621 | */ | |
2622 | static void | |
2623 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2624 | struct cfq_io_context *cic) | |
2625 | { | |
7b679138 | 2626 | int old_idle, enable_idle; |
1be92f2f | 2627 | |
0871714e JA |
2628 | /* |
2629 | * Don't idle for async or idle io prio class | |
2630 | */ | |
2631 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
2632 | return; |
2633 | ||
c265a7f4 | 2634 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 2635 | |
76280aff CZ |
2636 | if (cfqq->queued[0] + cfqq->queued[1] >= 4) |
2637 | cfq_mark_cfqq_deep(cfqq); | |
2638 | ||
66dac98e | 2639 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || |
76280aff CZ |
2640 | (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples) |
2641 | && CFQQ_SEEKY(cfqq))) | |
22e2c507 JA |
2642 | enable_idle = 0; |
2643 | else if (sample_valid(cic->ttime_samples)) { | |
718eee05 | 2644 | if (cic->ttime_mean > cfqd->cfq_slice_idle) |
22e2c507 JA |
2645 | enable_idle = 0; |
2646 | else | |
2647 | enable_idle = 1; | |
1da177e4 LT |
2648 | } |
2649 | ||
7b679138 JA |
2650 | if (old_idle != enable_idle) { |
2651 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
2652 | if (enable_idle) | |
2653 | cfq_mark_cfqq_idle_window(cfqq); | |
2654 | else | |
2655 | cfq_clear_cfqq_idle_window(cfqq); | |
2656 | } | |
22e2c507 | 2657 | } |
1da177e4 | 2658 | |
22e2c507 JA |
2659 | /* |
2660 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
2661 | * no or if we aren't sure, a 1 will cause a preempt. | |
2662 | */ | |
a6151c3a | 2663 | static bool |
22e2c507 | 2664 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, |
5e705374 | 2665 | struct request *rq) |
22e2c507 | 2666 | { |
6d048f53 | 2667 | struct cfq_queue *cfqq; |
22e2c507 | 2668 | |
6d048f53 JA |
2669 | cfqq = cfqd->active_queue; |
2670 | if (!cfqq) | |
a6151c3a | 2671 | return false; |
22e2c507 | 2672 | |
6d048f53 | 2673 | if (cfq_slice_used(cfqq)) |
a6151c3a | 2674 | return true; |
6d048f53 JA |
2675 | |
2676 | if (cfq_class_idle(new_cfqq)) | |
a6151c3a | 2677 | return false; |
22e2c507 JA |
2678 | |
2679 | if (cfq_class_idle(cfqq)) | |
a6151c3a | 2680 | return true; |
1e3335de | 2681 | |
f04a6424 | 2682 | /* Allow preemption only if we are idling on sync-noidle tree */ |
e4a22919 CZ |
2683 | if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && |
2684 | cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && | |
f04a6424 VG |
2685 | new_cfqq->service_tree->count == 2 && |
2686 | RB_EMPTY_ROOT(&cfqq->sort_list)) | |
718eee05 CZ |
2687 | return true; |
2688 | ||
374f84ac JA |
2689 | /* |
2690 | * if the new request is sync, but the currently running queue is | |
2691 | * not, let the sync request have priority. | |
2692 | */ | |
5e705374 | 2693 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
a6151c3a | 2694 | return true; |
1e3335de | 2695 | |
374f84ac JA |
2696 | /* |
2697 | * So both queues are sync. Let the new request get disk time if | |
2698 | * it's a metadata request and the current queue is doing regular IO. | |
2699 | */ | |
2700 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
e6ec4fe2 | 2701 | return true; |
22e2c507 | 2702 | |
3a9a3f6c DS |
2703 | /* |
2704 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
2705 | */ | |
2706 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
a6151c3a | 2707 | return true; |
3a9a3f6c | 2708 | |
1e3335de | 2709 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
a6151c3a | 2710 | return false; |
1e3335de JA |
2711 | |
2712 | /* | |
2713 | * if this request is as-good as one we would expect from the | |
2714 | * current cfqq, let it preempt | |
2715 | */ | |
e00ef799 | 2716 | if (cfq_rq_close(cfqd, cfqq, rq)) |
a6151c3a | 2717 | return true; |
1e3335de | 2718 | |
a6151c3a | 2719 | return false; |
22e2c507 JA |
2720 | } |
2721 | ||
2722 | /* | |
2723 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
2724 | * let it have half of its nominal slice. | |
2725 | */ | |
2726 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2727 | { | |
7b679138 | 2728 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
6084cdda | 2729 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 2730 | |
bf572256 JA |
2731 | /* |
2732 | * Put the new queue at the front of the of the current list, | |
2733 | * so we know that it will be selected next. | |
2734 | */ | |
2735 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
2736 | |
2737 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 2738 | |
44f7c160 JA |
2739 | cfqq->slice_end = 0; |
2740 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
2741 | } |
2742 | ||
22e2c507 | 2743 | /* |
5e705374 | 2744 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
2745 | * something we should do about it |
2746 | */ | |
2747 | static void | |
5e705374 JA |
2748 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
2749 | struct request *rq) | |
22e2c507 | 2750 | { |
5e705374 | 2751 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 2752 | |
45333d5a | 2753 | cfqd->rq_queued++; |
374f84ac JA |
2754 | if (rq_is_meta(rq)) |
2755 | cfqq->meta_pending++; | |
2756 | ||
9c2c38a1 | 2757 | cfq_update_io_thinktime(cfqd, cic); |
b2c18e1e | 2758 | cfq_update_io_seektime(cfqd, cfqq, rq); |
9c2c38a1 JA |
2759 | cfq_update_idle_window(cfqd, cfqq, cic); |
2760 | ||
b2c18e1e | 2761 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
2762 | |
2763 | if (cfqq == cfqd->active_queue) { | |
2764 | /* | |
b029195d JA |
2765 | * Remember that we saw a request from this process, but |
2766 | * don't start queuing just yet. Otherwise we risk seeing lots | |
2767 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
2768 | * and merging. If the request is already larger than a single |
2769 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
2770 | * merging is already done. Ditto for a busy system that |
2771 | * has other work pending, don't risk delaying until the | |
2772 | * idle timer unplug to continue working. | |
22e2c507 | 2773 | */ |
d6ceb25e | 2774 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
2775 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
2776 | cfqd->busy_queues > 1) { | |
d6ceb25e | 2777 | del_timer(&cfqd->idle_slice_timer); |
bf791937 VG |
2778 | __blk_run_queue(cfqd->queue); |
2779 | } else | |
2780 | cfq_mark_cfqq_must_dispatch(cfqq); | |
d6ceb25e | 2781 | } |
5e705374 | 2782 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
2783 | /* |
2784 | * not the active queue - expire current slice if it is | |
2785 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
2786 | * has some old slice time left and is of higher priority or |
2787 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
2788 | */ |
2789 | cfq_preempt_queue(cfqd, cfqq); | |
a7f55792 | 2790 | __blk_run_queue(cfqd->queue); |
22e2c507 | 2791 | } |
1da177e4 LT |
2792 | } |
2793 | ||
165125e1 | 2794 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2795 | { |
b4878f24 | 2796 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 2797 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 2798 | |
7b679138 | 2799 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 2800 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 2801 | |
30996f40 | 2802 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); |
22e2c507 | 2803 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
aa6f6a3d | 2804 | cfq_add_rq_rb(rq); |
22e2c507 | 2805 | |
5e705374 | 2806 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
2807 | } |
2808 | ||
45333d5a AC |
2809 | /* |
2810 | * Update hw_tag based on peak queue depth over 50 samples under | |
2811 | * sufficient load. | |
2812 | */ | |
2813 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
2814 | { | |
1a1238a7 SL |
2815 | struct cfq_queue *cfqq = cfqd->active_queue; |
2816 | ||
e459dd08 CZ |
2817 | if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth) |
2818 | cfqd->hw_tag_est_depth = rq_in_driver(cfqd); | |
2819 | ||
2820 | if (cfqd->hw_tag == 1) | |
2821 | return; | |
45333d5a AC |
2822 | |
2823 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
5ad531db | 2824 | rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
2825 | return; |
2826 | ||
1a1238a7 SL |
2827 | /* |
2828 | * If active queue hasn't enough requests and can idle, cfq might not | |
2829 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
2830 | * case | |
2831 | */ | |
2832 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
2833 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
2834 | CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN) | |
2835 | return; | |
2836 | ||
45333d5a AC |
2837 | if (cfqd->hw_tag_samples++ < 50) |
2838 | return; | |
2839 | ||
e459dd08 | 2840 | if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
2841 | cfqd->hw_tag = 1; |
2842 | else | |
2843 | cfqd->hw_tag = 0; | |
45333d5a AC |
2844 | } |
2845 | ||
165125e1 | 2846 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2847 | { |
5e705374 | 2848 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 2849 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 2850 | const int sync = rq_is_sync(rq); |
b4878f24 | 2851 | unsigned long now; |
1da177e4 | 2852 | |
b4878f24 | 2853 | now = jiffies; |
7b679138 | 2854 | cfq_log_cfqq(cfqd, cfqq, "complete"); |
1da177e4 | 2855 | |
45333d5a AC |
2856 | cfq_update_hw_tag(cfqd); |
2857 | ||
5ad531db | 2858 | WARN_ON(!cfqd->rq_in_driver[sync]); |
6d048f53 | 2859 | WARN_ON(!cfqq->dispatched); |
5ad531db | 2860 | cfqd->rq_in_driver[sync]--; |
6d048f53 | 2861 | cfqq->dispatched--; |
1da177e4 | 2862 | |
3ed9a296 JA |
2863 | if (cfq_cfqq_sync(cfqq)) |
2864 | cfqd->sync_flight--; | |
2865 | ||
365722bb | 2866 | if (sync) { |
5e705374 | 2867 | RQ_CIC(rq)->last_end_request = now; |
365722bb VG |
2868 | cfqd->last_end_sync_rq = now; |
2869 | } | |
caaa5f9f JA |
2870 | |
2871 | /* | |
2872 | * If this is the active queue, check if it needs to be expired, | |
2873 | * or if we want to idle in case it has no pending requests. | |
2874 | */ | |
2875 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
2876 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
2877 | ||
44f7c160 JA |
2878 | if (cfq_cfqq_slice_new(cfqq)) { |
2879 | cfq_set_prio_slice(cfqd, cfqq); | |
2880 | cfq_clear_cfqq_slice_new(cfqq); | |
2881 | } | |
a36e71f9 | 2882 | /* |
8e550632 CZ |
2883 | * Idling is not enabled on: |
2884 | * - expired queues | |
2885 | * - idle-priority queues | |
2886 | * - async queues | |
2887 | * - queues with still some requests queued | |
2888 | * - when there is a close cooperator | |
a36e71f9 | 2889 | */ |
0871714e | 2890 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
6084cdda | 2891 | cfq_slice_expired(cfqd, 1); |
8e550632 CZ |
2892 | else if (sync && cfqq_empty && |
2893 | !cfq_close_cooperator(cfqd, cfqq)) { | |
2894 | cfqd->noidle_tree_requires_idle |= !rq_noidle(rq); | |
2895 | /* | |
2896 | * Idling is enabled for SYNC_WORKLOAD. | |
2897 | * SYNC_NOIDLE_WORKLOAD idles at the end of the tree | |
2898 | * only if we processed at least one !rq_noidle request | |
2899 | */ | |
2900 | if (cfqd->serving_type == SYNC_WORKLOAD | |
2901 | || cfqd->noidle_tree_requires_idle) | |
2902 | cfq_arm_slice_timer(cfqd); | |
2903 | } | |
caaa5f9f | 2904 | } |
6d048f53 | 2905 | |
5ad531db | 2906 | if (!rq_in_driver(cfqd)) |
23e018a1 | 2907 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
2908 | } |
2909 | ||
22e2c507 JA |
2910 | /* |
2911 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
2912 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
2913 | */ | |
2914 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 2915 | { |
22e2c507 JA |
2916 | if (has_fs_excl()) { |
2917 | /* | |
2918 | * boost idle prio on transactions that would lock out other | |
2919 | * users of the filesystem | |
2920 | */ | |
2921 | if (cfq_class_idle(cfqq)) | |
2922 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2923 | if (cfqq->ioprio > IOPRIO_NORM) | |
2924 | cfqq->ioprio = IOPRIO_NORM; | |
2925 | } else { | |
2926 | /* | |
dddb7451 | 2927 | * unboost the queue (if needed) |
22e2c507 | 2928 | */ |
dddb7451 CZ |
2929 | cfqq->ioprio_class = cfqq->org_ioprio_class; |
2930 | cfqq->ioprio = cfqq->org_ioprio; | |
22e2c507 | 2931 | } |
22e2c507 | 2932 | } |
1da177e4 | 2933 | |
89850f7e | 2934 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 2935 | { |
1b379d8d | 2936 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 2937 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 2938 | return ELV_MQUEUE_MUST; |
3b18152c | 2939 | } |
1da177e4 | 2940 | |
22e2c507 | 2941 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
2942 | } |
2943 | ||
165125e1 | 2944 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
2945 | { |
2946 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2947 | struct task_struct *tsk = current; | |
91fac317 | 2948 | struct cfq_io_context *cic; |
22e2c507 JA |
2949 | struct cfq_queue *cfqq; |
2950 | ||
2951 | /* | |
2952 | * don't force setup of a queue from here, as a call to may_queue | |
2953 | * does not necessarily imply that a request actually will be queued. | |
2954 | * so just lookup a possibly existing queue, or return 'may queue' | |
2955 | * if that fails | |
2956 | */ | |
4ac845a2 | 2957 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
2958 | if (!cic) |
2959 | return ELV_MQUEUE_MAY; | |
2960 | ||
b0b78f81 | 2961 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 2962 | if (cfqq) { |
fd0928df | 2963 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
2964 | cfq_prio_boost(cfqq); |
2965 | ||
89850f7e | 2966 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
2967 | } |
2968 | ||
2969 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
2970 | } |
2971 | ||
1da177e4 LT |
2972 | /* |
2973 | * queue lock held here | |
2974 | */ | |
bb37b94c | 2975 | static void cfq_put_request(struct request *rq) |
1da177e4 | 2976 | { |
5e705374 | 2977 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 2978 | |
5e705374 | 2979 | if (cfqq) { |
22e2c507 | 2980 | const int rw = rq_data_dir(rq); |
1da177e4 | 2981 | |
22e2c507 JA |
2982 | BUG_ON(!cfqq->allocated[rw]); |
2983 | cfqq->allocated[rw]--; | |
1da177e4 | 2984 | |
5e705374 | 2985 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 2986 | |
1da177e4 | 2987 | rq->elevator_private = NULL; |
5e705374 | 2988 | rq->elevator_private2 = NULL; |
1da177e4 | 2989 | |
1da177e4 LT |
2990 | cfq_put_queue(cfqq); |
2991 | } | |
2992 | } | |
2993 | ||
df5fe3e8 JM |
2994 | static struct cfq_queue * |
2995 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, | |
2996 | struct cfq_queue *cfqq) | |
2997 | { | |
2998 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
2999 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
b3b6d040 | 3000 | cfq_mark_cfqq_coop(cfqq->new_cfqq); |
df5fe3e8 JM |
3001 | cfq_put_queue(cfqq); |
3002 | return cic_to_cfqq(cic, 1); | |
3003 | } | |
3004 | ||
e6c5bc73 JM |
3005 | static int should_split_cfqq(struct cfq_queue *cfqq) |
3006 | { | |
3007 | if (cfqq->seeky_start && | |
3008 | time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT)) | |
3009 | return 1; | |
3010 | return 0; | |
3011 | } | |
3012 | ||
3013 | /* | |
3014 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
3015 | * was the last process referring to said cfqq. | |
3016 | */ | |
3017 | static struct cfq_queue * | |
3018 | split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) | |
3019 | { | |
3020 | if (cfqq_process_refs(cfqq) == 1) { | |
3021 | cfqq->seeky_start = 0; | |
3022 | cfqq->pid = current->pid; | |
3023 | cfq_clear_cfqq_coop(cfqq); | |
3024 | return cfqq; | |
3025 | } | |
3026 | ||
3027 | cic_set_cfqq(cic, NULL, 1); | |
3028 | cfq_put_queue(cfqq); | |
3029 | return NULL; | |
3030 | } | |
1da177e4 | 3031 | /* |
22e2c507 | 3032 | * Allocate cfq data structures associated with this request. |
1da177e4 | 3033 | */ |
22e2c507 | 3034 | static int |
165125e1 | 3035 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
3036 | { |
3037 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3038 | struct cfq_io_context *cic; | |
3039 | const int rw = rq_data_dir(rq); | |
a6151c3a | 3040 | const bool is_sync = rq_is_sync(rq); |
22e2c507 | 3041 | struct cfq_queue *cfqq; |
1da177e4 LT |
3042 | unsigned long flags; |
3043 | ||
3044 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
3045 | ||
e2d74ac0 | 3046 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 3047 | |
1da177e4 LT |
3048 | spin_lock_irqsave(q->queue_lock, flags); |
3049 | ||
22e2c507 JA |
3050 | if (!cic) |
3051 | goto queue_fail; | |
3052 | ||
e6c5bc73 | 3053 | new_queue: |
91fac317 | 3054 | cfqq = cic_to_cfqq(cic, is_sync); |
32f2e807 | 3055 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { |
fd0928df | 3056 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 | 3057 | cic_set_cfqq(cic, cfqq, is_sync); |
df5fe3e8 | 3058 | } else { |
e6c5bc73 JM |
3059 | /* |
3060 | * If the queue was seeky for too long, break it apart. | |
3061 | */ | |
3062 | if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) { | |
3063 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); | |
3064 | cfqq = split_cfqq(cic, cfqq); | |
3065 | if (!cfqq) | |
3066 | goto new_queue; | |
3067 | } | |
3068 | ||
df5fe3e8 JM |
3069 | /* |
3070 | * Check to see if this queue is scheduled to merge with | |
3071 | * another, closely cooperating queue. The merging of | |
3072 | * queues happens here as it must be done in process context. | |
3073 | * The reference on new_cfqq was taken in merge_cfqqs. | |
3074 | */ | |
3075 | if (cfqq->new_cfqq) | |
3076 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
91fac317 | 3077 | } |
1da177e4 LT |
3078 | |
3079 | cfqq->allocated[rw]++; | |
22e2c507 | 3080 | atomic_inc(&cfqq->ref); |
1da177e4 | 3081 | |
5e705374 | 3082 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 3083 | |
5e705374 JA |
3084 | rq->elevator_private = cic; |
3085 | rq->elevator_private2 = cfqq; | |
3086 | return 0; | |
1da177e4 | 3087 | |
22e2c507 JA |
3088 | queue_fail: |
3089 | if (cic) | |
3090 | put_io_context(cic->ioc); | |
89850f7e | 3091 | |
23e018a1 | 3092 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 3093 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 3094 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
3095 | return 1; |
3096 | } | |
3097 | ||
65f27f38 | 3098 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 3099 | { |
65f27f38 | 3100 | struct cfq_data *cfqd = |
23e018a1 | 3101 | container_of(work, struct cfq_data, unplug_work); |
165125e1 | 3102 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3103 | |
40bb54d1 | 3104 | spin_lock_irq(q->queue_lock); |
a7f55792 | 3105 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 3106 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
3107 | } |
3108 | ||
3109 | /* | |
3110 | * Timer running if the active_queue is currently idling inside its time slice | |
3111 | */ | |
3112 | static void cfq_idle_slice_timer(unsigned long data) | |
3113 | { | |
3114 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
3115 | struct cfq_queue *cfqq; | |
3116 | unsigned long flags; | |
3c6bd2f8 | 3117 | int timed_out = 1; |
22e2c507 | 3118 | |
7b679138 JA |
3119 | cfq_log(cfqd, "idle timer fired"); |
3120 | ||
22e2c507 JA |
3121 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
3122 | ||
fe094d98 JA |
3123 | cfqq = cfqd->active_queue; |
3124 | if (cfqq) { | |
3c6bd2f8 JA |
3125 | timed_out = 0; |
3126 | ||
b029195d JA |
3127 | /* |
3128 | * We saw a request before the queue expired, let it through | |
3129 | */ | |
3130 | if (cfq_cfqq_must_dispatch(cfqq)) | |
3131 | goto out_kick; | |
3132 | ||
22e2c507 JA |
3133 | /* |
3134 | * expired | |
3135 | */ | |
44f7c160 | 3136 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
3137 | goto expire; |
3138 | ||
3139 | /* | |
3140 | * only expire and reinvoke request handler, if there are | |
3141 | * other queues with pending requests | |
3142 | */ | |
caaa5f9f | 3143 | if (!cfqd->busy_queues) |
22e2c507 | 3144 | goto out_cont; |
22e2c507 JA |
3145 | |
3146 | /* | |
3147 | * not expired and it has a request pending, let it dispatch | |
3148 | */ | |
75e50984 | 3149 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 3150 | goto out_kick; |
76280aff CZ |
3151 | |
3152 | /* | |
3153 | * Queue depth flag is reset only when the idle didn't succeed | |
3154 | */ | |
3155 | cfq_clear_cfqq_deep(cfqq); | |
22e2c507 JA |
3156 | } |
3157 | expire: | |
6084cdda | 3158 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 3159 | out_kick: |
23e018a1 | 3160 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
3161 | out_cont: |
3162 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
3163 | } | |
3164 | ||
3b18152c JA |
3165 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
3166 | { | |
3167 | del_timer_sync(&cfqd->idle_slice_timer); | |
23e018a1 | 3168 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 3169 | } |
22e2c507 | 3170 | |
c2dea2d1 VT |
3171 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
3172 | { | |
3173 | int i; | |
3174 | ||
3175 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
3176 | if (cfqd->async_cfqq[0][i]) | |
3177 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
3178 | if (cfqd->async_cfqq[1][i]) | |
3179 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 3180 | } |
2389d1ef ON |
3181 | |
3182 | if (cfqd->async_idle_cfqq) | |
3183 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
3184 | } |
3185 | ||
b374d18a | 3186 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 3187 | { |
22e2c507 | 3188 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 3189 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3190 | |
3b18152c | 3191 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 3192 | |
d9ff4187 | 3193 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 3194 | |
d9ff4187 | 3195 | if (cfqd->active_queue) |
6084cdda | 3196 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
3197 | |
3198 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
3199 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
3200 | struct cfq_io_context, | |
3201 | queue_list); | |
89850f7e JA |
3202 | |
3203 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 3204 | } |
e2d74ac0 | 3205 | |
c2dea2d1 | 3206 | cfq_put_async_queues(cfqd); |
15c31be4 | 3207 | |
d9ff4187 | 3208 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
3209 | |
3210 | cfq_shutdown_timer_wq(cfqd); | |
3211 | ||
a90d742e | 3212 | kfree(cfqd); |
1da177e4 LT |
3213 | } |
3214 | ||
165125e1 | 3215 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
3216 | { |
3217 | struct cfq_data *cfqd; | |
718eee05 | 3218 | int i, j; |
cdb16e8f | 3219 | struct cfq_group *cfqg; |
615f0259 | 3220 | struct cfq_rb_root *st; |
1da177e4 | 3221 | |
94f6030c | 3222 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 3223 | if (!cfqd) |
bc1c1169 | 3224 | return NULL; |
1da177e4 | 3225 | |
1fa8f6d6 VG |
3226 | /* Init root service tree */ |
3227 | cfqd->grp_service_tree = CFQ_RB_ROOT; | |
3228 | ||
cdb16e8f VG |
3229 | /* Init root group */ |
3230 | cfqg = &cfqd->root_group; | |
615f0259 VG |
3231 | for_each_cfqg_st(cfqg, i, j, st) |
3232 | *st = CFQ_RB_ROOT; | |
1fa8f6d6 | 3233 | RB_CLEAR_NODE(&cfqg->rb_node); |
26a2ac00 | 3234 | |
25bc6b07 VG |
3235 | /* Give preference to root group over other groups */ |
3236 | cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT; | |
3237 | ||
26a2ac00 JA |
3238 | /* |
3239 | * Not strictly needed (since RB_ROOT just clears the node and we | |
3240 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
3241 | * to add magic to the rb code | |
3242 | */ | |
3243 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
3244 | cfqd->prio_trees[i] = RB_ROOT; | |
3245 | ||
6118b70b JA |
3246 | /* |
3247 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
3248 | * Grab a permanent reference to it, so that the normal code flow | |
3249 | * will not attempt to free it. | |
3250 | */ | |
3251 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
3252 | atomic_inc(&cfqd->oom_cfqq.ref); | |
cdb16e8f | 3253 | cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group); |
6118b70b | 3254 | |
d9ff4187 | 3255 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 3256 | |
1da177e4 | 3257 | cfqd->queue = q; |
1da177e4 | 3258 | |
22e2c507 JA |
3259 | init_timer(&cfqd->idle_slice_timer); |
3260 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
3261 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
3262 | ||
23e018a1 | 3263 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 3264 | |
1da177e4 | 3265 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
3266 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
3267 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
3268 | cfqd->cfq_back_max = cfq_back_max; |
3269 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
3270 | cfqd->cfq_slice[0] = cfq_slice_async; |
3271 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
3272 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
3273 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
963b72fc | 3274 | cfqd->cfq_latency = 1; |
e459dd08 | 3275 | cfqd->hw_tag = -1; |
365722bb | 3276 | cfqd->last_end_sync_rq = jiffies; |
bc1c1169 | 3277 | return cfqd; |
1da177e4 LT |
3278 | } |
3279 | ||
3280 | static void cfq_slab_kill(void) | |
3281 | { | |
d6de8be7 JA |
3282 | /* |
3283 | * Caller already ensured that pending RCU callbacks are completed, | |
3284 | * so we should have no busy allocations at this point. | |
3285 | */ | |
1da177e4 LT |
3286 | if (cfq_pool) |
3287 | kmem_cache_destroy(cfq_pool); | |
3288 | if (cfq_ioc_pool) | |
3289 | kmem_cache_destroy(cfq_ioc_pool); | |
3290 | } | |
3291 | ||
3292 | static int __init cfq_slab_setup(void) | |
3293 | { | |
0a31bd5f | 3294 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
3295 | if (!cfq_pool) |
3296 | goto fail; | |
3297 | ||
34e6bbf2 | 3298 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
3299 | if (!cfq_ioc_pool) |
3300 | goto fail; | |
3301 | ||
3302 | return 0; | |
3303 | fail: | |
3304 | cfq_slab_kill(); | |
3305 | return -ENOMEM; | |
3306 | } | |
3307 | ||
1da177e4 LT |
3308 | /* |
3309 | * sysfs parts below --> | |
3310 | */ | |
1da177e4 LT |
3311 | static ssize_t |
3312 | cfq_var_show(unsigned int var, char *page) | |
3313 | { | |
3314 | return sprintf(page, "%d\n", var); | |
3315 | } | |
3316 | ||
3317 | static ssize_t | |
3318 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
3319 | { | |
3320 | char *p = (char *) page; | |
3321 | ||
3322 | *var = simple_strtoul(p, &p, 10); | |
3323 | return count; | |
3324 | } | |
3325 | ||
1da177e4 | 3326 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 3327 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 3328 | { \ |
3d1ab40f | 3329 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3330 | unsigned int __data = __VAR; \ |
3331 | if (__CONV) \ | |
3332 | __data = jiffies_to_msecs(__data); \ | |
3333 | return cfq_var_show(__data, (page)); \ | |
3334 | } | |
3335 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
3336 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
3337 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
3338 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
3339 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
3340 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
3341 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
3342 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
3343 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
963b72fc | 3344 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); |
1da177e4 LT |
3345 | #undef SHOW_FUNCTION |
3346 | ||
3347 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 3348 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 3349 | { \ |
3d1ab40f | 3350 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3351 | unsigned int __data; \ |
3352 | int ret = cfq_var_store(&__data, (page), count); \ | |
3353 | if (__data < (MIN)) \ | |
3354 | __data = (MIN); \ | |
3355 | else if (__data > (MAX)) \ | |
3356 | __data = (MAX); \ | |
3357 | if (__CONV) \ | |
3358 | *(__PTR) = msecs_to_jiffies(__data); \ | |
3359 | else \ | |
3360 | *(__PTR) = __data; \ | |
3361 | return ret; \ | |
3362 | } | |
3363 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
3364 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
3365 | UINT_MAX, 1); | |
3366 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
3367 | UINT_MAX, 1); | |
e572ec7e | 3368 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
3369 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
3370 | UINT_MAX, 0); | |
22e2c507 JA |
3371 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
3372 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
3373 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
3374 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
3375 | UINT_MAX, 0); | |
963b72fc | 3376 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); |
1da177e4 LT |
3377 | #undef STORE_FUNCTION |
3378 | ||
e572ec7e AV |
3379 | #define CFQ_ATTR(name) \ |
3380 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
3381 | ||
3382 | static struct elv_fs_entry cfq_attrs[] = { | |
3383 | CFQ_ATTR(quantum), | |
e572ec7e AV |
3384 | CFQ_ATTR(fifo_expire_sync), |
3385 | CFQ_ATTR(fifo_expire_async), | |
3386 | CFQ_ATTR(back_seek_max), | |
3387 | CFQ_ATTR(back_seek_penalty), | |
3388 | CFQ_ATTR(slice_sync), | |
3389 | CFQ_ATTR(slice_async), | |
3390 | CFQ_ATTR(slice_async_rq), | |
3391 | CFQ_ATTR(slice_idle), | |
963b72fc | 3392 | CFQ_ATTR(low_latency), |
e572ec7e | 3393 | __ATTR_NULL |
1da177e4 LT |
3394 | }; |
3395 | ||
1da177e4 LT |
3396 | static struct elevator_type iosched_cfq = { |
3397 | .ops = { | |
3398 | .elevator_merge_fn = cfq_merge, | |
3399 | .elevator_merged_fn = cfq_merged_request, | |
3400 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 3401 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 3402 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 3403 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 3404 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
3405 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
3406 | .elevator_queue_empty_fn = cfq_queue_empty, | |
3407 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
3408 | .elevator_former_req_fn = elv_rb_former_request, |
3409 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
3410 | .elevator_set_req_fn = cfq_set_request, |
3411 | .elevator_put_req_fn = cfq_put_request, | |
3412 | .elevator_may_queue_fn = cfq_may_queue, | |
3413 | .elevator_init_fn = cfq_init_queue, | |
3414 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 3415 | .trim = cfq_free_io_context, |
1da177e4 | 3416 | }, |
3d1ab40f | 3417 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
3418 | .elevator_name = "cfq", |
3419 | .elevator_owner = THIS_MODULE, | |
3420 | }; | |
3421 | ||
3422 | static int __init cfq_init(void) | |
3423 | { | |
22e2c507 JA |
3424 | /* |
3425 | * could be 0 on HZ < 1000 setups | |
3426 | */ | |
3427 | if (!cfq_slice_async) | |
3428 | cfq_slice_async = 1; | |
3429 | if (!cfq_slice_idle) | |
3430 | cfq_slice_idle = 1; | |
3431 | ||
1da177e4 LT |
3432 | if (cfq_slab_setup()) |
3433 | return -ENOMEM; | |
3434 | ||
2fdd82bd | 3435 | elv_register(&iosched_cfq); |
1da177e4 | 3436 | |
2fdd82bd | 3437 | return 0; |
1da177e4 LT |
3438 | } |
3439 | ||
3440 | static void __exit cfq_exit(void) | |
3441 | { | |
6e9a4738 | 3442 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 3443 | elv_unregister(&iosched_cfq); |
334e94de | 3444 | ioc_gone = &all_gone; |
fba82272 OH |
3445 | /* ioc_gone's update must be visible before reading ioc_count */ |
3446 | smp_wmb(); | |
d6de8be7 JA |
3447 | |
3448 | /* | |
3449 | * this also protects us from entering cfq_slab_kill() with | |
3450 | * pending RCU callbacks | |
3451 | */ | |
245b2e70 | 3452 | if (elv_ioc_count_read(cfq_ioc_count)) |
9a11b4ed | 3453 | wait_for_completion(&all_gone); |
83521d3e | 3454 | cfq_slab_kill(); |
1da177e4 LT |
3455 | } |
3456 | ||
3457 | module_init(cfq_init); | |
3458 | module_exit(cfq_exit); | |
3459 | ||
3460 | MODULE_AUTHOR("Jens Axboe"); | |
3461 | MODULE_LICENSE("GPL"); | |
3462 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |