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 | ||
dae739eb VG |
118 | /* time when queue got scheduled in to dispatch first request. */ |
119 | unsigned long dispatch_start; | |
120 | /* time when first request from queue completed and slice started. */ | |
121 | unsigned long slice_start; | |
6118b70b JA |
122 | unsigned long slice_end; |
123 | long slice_resid; | |
124 | unsigned int slice_dispatch; | |
125 | ||
126 | /* pending metadata requests */ | |
127 | int meta_pending; | |
128 | /* number of requests that are on the dispatch list or inside driver */ | |
129 | int dispatched; | |
130 | ||
131 | /* io prio of this group */ | |
132 | unsigned short ioprio, org_ioprio; | |
133 | unsigned short ioprio_class, org_ioprio_class; | |
134 | ||
b2c18e1e JM |
135 | unsigned int seek_samples; |
136 | u64 seek_total; | |
137 | sector_t seek_mean; | |
138 | sector_t last_request_pos; | |
e6c5bc73 | 139 | unsigned long seeky_start; |
b2c18e1e | 140 | |
6118b70b | 141 | pid_t pid; |
df5fe3e8 | 142 | |
aa6f6a3d | 143 | struct cfq_rb_root *service_tree; |
df5fe3e8 | 144 | struct cfq_queue *new_cfqq; |
cdb16e8f | 145 | struct cfq_group *cfqg; |
6118b70b JA |
146 | }; |
147 | ||
c0324a02 | 148 | /* |
718eee05 | 149 | * First index in the service_trees. |
c0324a02 CZ |
150 | * IDLE is handled separately, so it has negative index |
151 | */ | |
152 | enum wl_prio_t { | |
c0324a02 | 153 | BE_WORKLOAD = 0, |
615f0259 VG |
154 | RT_WORKLOAD = 1, |
155 | IDLE_WORKLOAD = 2, | |
c0324a02 CZ |
156 | }; |
157 | ||
718eee05 CZ |
158 | /* |
159 | * Second index in the service_trees. | |
160 | */ | |
161 | enum wl_type_t { | |
162 | ASYNC_WORKLOAD = 0, | |
163 | SYNC_NOIDLE_WORKLOAD = 1, | |
164 | SYNC_WORKLOAD = 2 | |
165 | }; | |
166 | ||
cdb16e8f VG |
167 | /* This is per cgroup per device grouping structure */ |
168 | struct cfq_group { | |
1fa8f6d6 VG |
169 | /* group service_tree member */ |
170 | struct rb_node rb_node; | |
171 | ||
172 | /* group service_tree key */ | |
173 | u64 vdisktime; | |
25bc6b07 | 174 | unsigned int weight; |
1fa8f6d6 VG |
175 | bool on_st; |
176 | ||
177 | /* number of cfqq currently on this group */ | |
178 | int nr_cfqq; | |
179 | ||
58ff82f3 VG |
180 | /* Per group busy queus average. Useful for workload slice calc. */ |
181 | unsigned int busy_queues_avg[2]; | |
cdb16e8f VG |
182 | /* |
183 | * rr lists of queues with requests, onle rr for each priority class. | |
184 | * Counts are embedded in the cfq_rb_root | |
185 | */ | |
186 | struct cfq_rb_root service_trees[2][3]; | |
187 | struct cfq_rb_root service_tree_idle; | |
dae739eb VG |
188 | |
189 | unsigned long saved_workload_slice; | |
190 | enum wl_type_t saved_workload; | |
191 | enum wl_prio_t saved_serving_prio; | |
cdb16e8f | 192 | }; |
718eee05 | 193 | |
22e2c507 JA |
194 | /* |
195 | * Per block device queue structure | |
196 | */ | |
1da177e4 | 197 | struct cfq_data { |
165125e1 | 198 | struct request_queue *queue; |
1fa8f6d6 VG |
199 | /* Root service tree for cfq_groups */ |
200 | struct cfq_rb_root grp_service_tree; | |
cdb16e8f | 201 | struct cfq_group root_group; |
58ff82f3 VG |
202 | /* Number of active cfq groups on group service tree */ |
203 | int nr_groups; | |
22e2c507 | 204 | |
c0324a02 CZ |
205 | /* |
206 | * The priority currently being served | |
22e2c507 | 207 | */ |
c0324a02 | 208 | enum wl_prio_t serving_prio; |
718eee05 CZ |
209 | enum wl_type_t serving_type; |
210 | unsigned long workload_expires; | |
cdb16e8f | 211 | struct cfq_group *serving_group; |
8e550632 | 212 | bool noidle_tree_requires_idle; |
a36e71f9 JA |
213 | |
214 | /* | |
215 | * Each priority tree is sorted by next_request position. These | |
216 | * trees are used when determining if two or more queues are | |
217 | * interleaving requests (see cfq_close_cooperator). | |
218 | */ | |
219 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
220 | ||
22e2c507 JA |
221 | unsigned int busy_queues; |
222 | ||
5ad531db | 223 | int rq_in_driver[2]; |
3ed9a296 | 224 | int sync_flight; |
45333d5a AC |
225 | |
226 | /* | |
227 | * queue-depth detection | |
228 | */ | |
229 | int rq_queued; | |
25776e35 | 230 | int hw_tag; |
e459dd08 CZ |
231 | /* |
232 | * hw_tag can be | |
233 | * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) | |
234 | * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) | |
235 | * 0 => no NCQ | |
236 | */ | |
237 | int hw_tag_est_depth; | |
238 | unsigned int hw_tag_samples; | |
1da177e4 | 239 | |
22e2c507 JA |
240 | /* |
241 | * idle window management | |
242 | */ | |
243 | struct timer_list idle_slice_timer; | |
23e018a1 | 244 | struct work_struct unplug_work; |
1da177e4 | 245 | |
22e2c507 JA |
246 | struct cfq_queue *active_queue; |
247 | struct cfq_io_context *active_cic; | |
22e2c507 | 248 | |
c2dea2d1 VT |
249 | /* |
250 | * async queue for each priority case | |
251 | */ | |
252 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
253 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 254 | |
6d048f53 | 255 | sector_t last_position; |
1da177e4 | 256 | |
1da177e4 LT |
257 | /* |
258 | * tunables, see top of file | |
259 | */ | |
260 | unsigned int cfq_quantum; | |
22e2c507 | 261 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
262 | unsigned int cfq_back_penalty; |
263 | unsigned int cfq_back_max; | |
22e2c507 JA |
264 | unsigned int cfq_slice[2]; |
265 | unsigned int cfq_slice_async_rq; | |
266 | unsigned int cfq_slice_idle; | |
963b72fc | 267 | unsigned int cfq_latency; |
d9ff4187 AV |
268 | |
269 | struct list_head cic_list; | |
1da177e4 | 270 | |
6118b70b JA |
271 | /* |
272 | * Fallback dummy cfqq for extreme OOM conditions | |
273 | */ | |
274 | struct cfq_queue oom_cfqq; | |
365722bb VG |
275 | |
276 | unsigned long last_end_sync_rq; | |
1da177e4 LT |
277 | }; |
278 | ||
cdb16e8f VG |
279 | static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, |
280 | enum wl_prio_t prio, | |
718eee05 | 281 | enum wl_type_t type, |
c0324a02 CZ |
282 | struct cfq_data *cfqd) |
283 | { | |
1fa8f6d6 VG |
284 | if (!cfqg) |
285 | return NULL; | |
286 | ||
c0324a02 | 287 | if (prio == IDLE_WORKLOAD) |
cdb16e8f | 288 | return &cfqg->service_tree_idle; |
c0324a02 | 289 | |
cdb16e8f | 290 | return &cfqg->service_trees[prio][type]; |
c0324a02 CZ |
291 | } |
292 | ||
3b18152c | 293 | enum cfqq_state_flags { |
b0b8d749 JA |
294 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
295 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 296 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 | 297 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ |
b0b8d749 JA |
298 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
299 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
300 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 301 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 302 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
b3b6d040 | 303 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ |
76280aff | 304 | CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ |
3b18152c JA |
305 | }; |
306 | ||
307 | #define CFQ_CFQQ_FNS(name) \ | |
308 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
309 | { \ | |
fe094d98 | 310 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
311 | } \ |
312 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
313 | { \ | |
fe094d98 | 314 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
315 | } \ |
316 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
317 | { \ | |
fe094d98 | 318 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
319 | } |
320 | ||
321 | CFQ_CFQQ_FNS(on_rr); | |
322 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 323 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c | 324 | CFQ_CFQQ_FNS(must_alloc_slice); |
3b18152c JA |
325 | CFQ_CFQQ_FNS(fifo_expire); |
326 | CFQ_CFQQ_FNS(idle_window); | |
327 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 328 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 329 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 330 | CFQ_CFQQ_FNS(coop); |
76280aff | 331 | CFQ_CFQQ_FNS(deep); |
3b18152c JA |
332 | #undef CFQ_CFQQ_FNS |
333 | ||
7b679138 JA |
334 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
335 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
336 | #define cfq_log(cfqd, fmt, args...) \ | |
337 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
338 | ||
615f0259 VG |
339 | /* Traverses through cfq group service trees */ |
340 | #define for_each_cfqg_st(cfqg, i, j, st) \ | |
341 | for (i = 0; i <= IDLE_WORKLOAD; i++) \ | |
342 | for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ | |
343 | : &cfqg->service_tree_idle; \ | |
344 | (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ | |
345 | (i == IDLE_WORKLOAD && j == 0); \ | |
346 | j++, st = i < IDLE_WORKLOAD ? \ | |
347 | &cfqg->service_trees[i][j]: NULL) \ | |
348 | ||
349 | ||
c0324a02 CZ |
350 | static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) |
351 | { | |
352 | if (cfq_class_idle(cfqq)) | |
353 | return IDLE_WORKLOAD; | |
354 | if (cfq_class_rt(cfqq)) | |
355 | return RT_WORKLOAD; | |
356 | return BE_WORKLOAD; | |
357 | } | |
358 | ||
718eee05 CZ |
359 | |
360 | static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) | |
361 | { | |
362 | if (!cfq_cfqq_sync(cfqq)) | |
363 | return ASYNC_WORKLOAD; | |
364 | if (!cfq_cfqq_idle_window(cfqq)) | |
365 | return SYNC_NOIDLE_WORKLOAD; | |
366 | return SYNC_WORKLOAD; | |
367 | } | |
368 | ||
58ff82f3 VG |
369 | static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl, |
370 | struct cfq_data *cfqd, | |
371 | struct cfq_group *cfqg) | |
c0324a02 CZ |
372 | { |
373 | if (wl == IDLE_WORKLOAD) | |
cdb16e8f | 374 | return cfqg->service_tree_idle.count; |
c0324a02 | 375 | |
cdb16e8f VG |
376 | return cfqg->service_trees[wl][ASYNC_WORKLOAD].count |
377 | + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count | |
378 | + cfqg->service_trees[wl][SYNC_WORKLOAD].count; | |
c0324a02 CZ |
379 | } |
380 | ||
165125e1 | 381 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
a6151c3a | 382 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, |
fd0928df | 383 | struct io_context *, gfp_t); |
4ac845a2 | 384 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
385 | struct io_context *); |
386 | ||
5ad531db JA |
387 | static inline int rq_in_driver(struct cfq_data *cfqd) |
388 | { | |
389 | return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1]; | |
390 | } | |
391 | ||
91fac317 | 392 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, |
a6151c3a | 393 | bool is_sync) |
91fac317 | 394 | { |
a6151c3a | 395 | return cic->cfqq[is_sync]; |
91fac317 VT |
396 | } |
397 | ||
398 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 399 | struct cfq_queue *cfqq, bool is_sync) |
91fac317 | 400 | { |
a6151c3a | 401 | cic->cfqq[is_sync] = cfqq; |
91fac317 VT |
402 | } |
403 | ||
404 | /* | |
405 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
406 | * set (in which case it could also be direct WRITE). | |
407 | */ | |
a6151c3a | 408 | static inline bool cfq_bio_sync(struct bio *bio) |
91fac317 | 409 | { |
a6151c3a | 410 | return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO); |
91fac317 | 411 | } |
1da177e4 | 412 | |
99f95e52 AM |
413 | /* |
414 | * scheduler run of queue, if there are requests pending and no one in the | |
415 | * driver that will restart queueing | |
416 | */ | |
23e018a1 | 417 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) |
99f95e52 | 418 | { |
7b679138 JA |
419 | if (cfqd->busy_queues) { |
420 | cfq_log(cfqd, "schedule dispatch"); | |
23e018a1 | 421 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 422 | } |
99f95e52 AM |
423 | } |
424 | ||
165125e1 | 425 | static int cfq_queue_empty(struct request_queue *q) |
99f95e52 AM |
426 | { |
427 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
428 | ||
f04a6424 | 429 | return !cfqd->rq_queued; |
99f95e52 AM |
430 | } |
431 | ||
44f7c160 JA |
432 | /* |
433 | * Scale schedule slice based on io priority. Use the sync time slice only | |
434 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
435 | * io only, should not get full sync slice length. | |
436 | */ | |
a6151c3a | 437 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, |
d9e7620e | 438 | unsigned short prio) |
44f7c160 | 439 | { |
d9e7620e | 440 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 441 | |
d9e7620e JA |
442 | WARN_ON(prio >= IOPRIO_BE_NR); |
443 | ||
444 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
445 | } | |
44f7c160 | 446 | |
d9e7620e JA |
447 | static inline int |
448 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
449 | { | |
450 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
451 | } |
452 | ||
25bc6b07 VG |
453 | static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg) |
454 | { | |
455 | u64 d = delta << CFQ_SERVICE_SHIFT; | |
456 | ||
457 | d = d * BLKIO_WEIGHT_DEFAULT; | |
458 | do_div(d, cfqg->weight); | |
459 | return d; | |
460 | } | |
461 | ||
462 | static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
463 | { | |
464 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
465 | if (delta > 0) | |
466 | min_vdisktime = vdisktime; | |
467 | ||
468 | return min_vdisktime; | |
469 | } | |
470 | ||
471 | static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
472 | { | |
473 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
474 | if (delta < 0) | |
475 | min_vdisktime = vdisktime; | |
476 | ||
477 | return min_vdisktime; | |
478 | } | |
479 | ||
480 | static void update_min_vdisktime(struct cfq_rb_root *st) | |
481 | { | |
482 | u64 vdisktime = st->min_vdisktime; | |
483 | struct cfq_group *cfqg; | |
484 | ||
485 | if (st->active) { | |
486 | cfqg = rb_entry_cfqg(st->active); | |
487 | vdisktime = cfqg->vdisktime; | |
488 | } | |
489 | ||
490 | if (st->left) { | |
491 | cfqg = rb_entry_cfqg(st->left); | |
492 | vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime); | |
493 | } | |
494 | ||
495 | st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime); | |
496 | } | |
497 | ||
5db5d642 CZ |
498 | /* |
499 | * get averaged number of queues of RT/BE priority. | |
500 | * average is updated, with a formula that gives more weight to higher numbers, | |
501 | * to quickly follows sudden increases and decrease slowly | |
502 | */ | |
503 | ||
58ff82f3 VG |
504 | static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, |
505 | struct cfq_group *cfqg, bool rt) | |
5869619c | 506 | { |
5db5d642 CZ |
507 | unsigned min_q, max_q; |
508 | unsigned mult = cfq_hist_divisor - 1; | |
509 | unsigned round = cfq_hist_divisor / 2; | |
58ff82f3 | 510 | unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); |
5db5d642 | 511 | |
58ff82f3 VG |
512 | min_q = min(cfqg->busy_queues_avg[rt], busy); |
513 | max_q = max(cfqg->busy_queues_avg[rt], busy); | |
514 | cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
5db5d642 | 515 | cfq_hist_divisor; |
58ff82f3 VG |
516 | return cfqg->busy_queues_avg[rt]; |
517 | } | |
518 | ||
519 | static inline unsigned | |
520 | cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
521 | { | |
522 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
523 | ||
524 | return cfq_target_latency * cfqg->weight / st->total_weight; | |
5db5d642 CZ |
525 | } |
526 | ||
44f7c160 JA |
527 | static inline void |
528 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
529 | { | |
5db5d642 CZ |
530 | unsigned slice = cfq_prio_to_slice(cfqd, cfqq); |
531 | if (cfqd->cfq_latency) { | |
58ff82f3 VG |
532 | /* |
533 | * interested queues (we consider only the ones with the same | |
534 | * priority class in the cfq group) | |
535 | */ | |
536 | unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, | |
537 | cfq_class_rt(cfqq)); | |
5db5d642 CZ |
538 | unsigned sync_slice = cfqd->cfq_slice[1]; |
539 | unsigned expect_latency = sync_slice * iq; | |
58ff82f3 VG |
540 | unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg); |
541 | ||
542 | if (expect_latency > group_slice) { | |
5db5d642 CZ |
543 | unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; |
544 | /* scale low_slice according to IO priority | |
545 | * and sync vs async */ | |
546 | unsigned low_slice = | |
547 | min(slice, base_low_slice * slice / sync_slice); | |
548 | /* the adapted slice value is scaled to fit all iqs | |
549 | * into the target latency */ | |
58ff82f3 | 550 | slice = max(slice * group_slice / expect_latency, |
5db5d642 CZ |
551 | low_slice); |
552 | } | |
553 | } | |
dae739eb | 554 | cfqq->slice_start = jiffies; |
5db5d642 | 555 | cfqq->slice_end = jiffies + slice; |
7b679138 | 556 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
557 | } |
558 | ||
559 | /* | |
560 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
561 | * isn't valid until the first request from the dispatch is activated | |
562 | * and the slice time set. | |
563 | */ | |
a6151c3a | 564 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) |
44f7c160 JA |
565 | { |
566 | if (cfq_cfqq_slice_new(cfqq)) | |
567 | return 0; | |
568 | if (time_before(jiffies, cfqq->slice_end)) | |
569 | return 0; | |
570 | ||
571 | return 1; | |
572 | } | |
573 | ||
1da177e4 | 574 | /* |
5e705374 | 575 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 576 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 577 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 578 | */ |
5e705374 | 579 | static struct request * |
cf7c25cf | 580 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) |
1da177e4 | 581 | { |
cf7c25cf | 582 | sector_t s1, s2, d1 = 0, d2 = 0; |
1da177e4 | 583 | unsigned long back_max; |
e8a99053 AM |
584 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
585 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
586 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 587 | |
5e705374 JA |
588 | if (rq1 == NULL || rq1 == rq2) |
589 | return rq2; | |
590 | if (rq2 == NULL) | |
591 | return rq1; | |
9c2c38a1 | 592 | |
5e705374 JA |
593 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
594 | return rq1; | |
595 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
596 | return rq2; | |
374f84ac JA |
597 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
598 | return rq1; | |
599 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
600 | return rq2; | |
1da177e4 | 601 | |
83096ebf TH |
602 | s1 = blk_rq_pos(rq1); |
603 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 604 | |
1da177e4 LT |
605 | /* |
606 | * by definition, 1KiB is 2 sectors | |
607 | */ | |
608 | back_max = cfqd->cfq_back_max * 2; | |
609 | ||
610 | /* | |
611 | * Strict one way elevator _except_ in the case where we allow | |
612 | * short backward seeks which are biased as twice the cost of a | |
613 | * similar forward seek. | |
614 | */ | |
615 | if (s1 >= last) | |
616 | d1 = s1 - last; | |
617 | else if (s1 + back_max >= last) | |
618 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
619 | else | |
e8a99053 | 620 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
621 | |
622 | if (s2 >= last) | |
623 | d2 = s2 - last; | |
624 | else if (s2 + back_max >= last) | |
625 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
626 | else | |
e8a99053 | 627 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
628 | |
629 | /* Found required data */ | |
e8a99053 AM |
630 | |
631 | /* | |
632 | * By doing switch() on the bit mask "wrap" we avoid having to | |
633 | * check two variables for all permutations: --> faster! | |
634 | */ | |
635 | switch (wrap) { | |
5e705374 | 636 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 637 | if (d1 < d2) |
5e705374 | 638 | return rq1; |
e8a99053 | 639 | else if (d2 < d1) |
5e705374 | 640 | return rq2; |
e8a99053 AM |
641 | else { |
642 | if (s1 >= s2) | |
5e705374 | 643 | return rq1; |
e8a99053 | 644 | else |
5e705374 | 645 | return rq2; |
e8a99053 | 646 | } |
1da177e4 | 647 | |
e8a99053 | 648 | case CFQ_RQ2_WRAP: |
5e705374 | 649 | return rq1; |
e8a99053 | 650 | case CFQ_RQ1_WRAP: |
5e705374 JA |
651 | return rq2; |
652 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
653 | default: |
654 | /* | |
655 | * Since both rqs are wrapped, | |
656 | * start with the one that's further behind head | |
657 | * (--> only *one* back seek required), | |
658 | * since back seek takes more time than forward. | |
659 | */ | |
660 | if (s1 <= s2) | |
5e705374 | 661 | return rq1; |
1da177e4 | 662 | else |
5e705374 | 663 | return rq2; |
1da177e4 LT |
664 | } |
665 | } | |
666 | ||
498d3aa2 JA |
667 | /* |
668 | * The below is leftmost cache rbtree addon | |
669 | */ | |
0871714e | 670 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 | 671 | { |
615f0259 VG |
672 | /* Service tree is empty */ |
673 | if (!root->count) | |
674 | return NULL; | |
675 | ||
cc09e299 JA |
676 | if (!root->left) |
677 | root->left = rb_first(&root->rb); | |
678 | ||
0871714e JA |
679 | if (root->left) |
680 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
681 | ||
682 | return NULL; | |
cc09e299 JA |
683 | } |
684 | ||
1fa8f6d6 VG |
685 | static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) |
686 | { | |
687 | if (!root->left) | |
688 | root->left = rb_first(&root->rb); | |
689 | ||
690 | if (root->left) | |
691 | return rb_entry_cfqg(root->left); | |
692 | ||
693 | return NULL; | |
694 | } | |
695 | ||
a36e71f9 JA |
696 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
697 | { | |
698 | rb_erase(n, root); | |
699 | RB_CLEAR_NODE(n); | |
700 | } | |
701 | ||
cc09e299 JA |
702 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
703 | { | |
704 | if (root->left == n) | |
705 | root->left = NULL; | |
a36e71f9 | 706 | rb_erase_init(n, &root->rb); |
aa6f6a3d | 707 | --root->count; |
cc09e299 JA |
708 | } |
709 | ||
1da177e4 LT |
710 | /* |
711 | * would be nice to take fifo expire time into account as well | |
712 | */ | |
5e705374 JA |
713 | static struct request * |
714 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
715 | struct request *last) | |
1da177e4 | 716 | { |
21183b07 JA |
717 | struct rb_node *rbnext = rb_next(&last->rb_node); |
718 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 719 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 720 | |
21183b07 | 721 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
722 | |
723 | if (rbprev) | |
5e705374 | 724 | prev = rb_entry_rq(rbprev); |
1da177e4 | 725 | |
21183b07 | 726 | if (rbnext) |
5e705374 | 727 | next = rb_entry_rq(rbnext); |
21183b07 JA |
728 | else { |
729 | rbnext = rb_first(&cfqq->sort_list); | |
730 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 731 | next = rb_entry_rq(rbnext); |
21183b07 | 732 | } |
1da177e4 | 733 | |
cf7c25cf | 734 | return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); |
1da177e4 LT |
735 | } |
736 | ||
d9e7620e JA |
737 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
738 | struct cfq_queue *cfqq) | |
1da177e4 | 739 | { |
d9e7620e JA |
740 | /* |
741 | * just an approximation, should be ok. | |
742 | */ | |
cdb16e8f | 743 | return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
464191c6 | 744 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); |
d9e7620e JA |
745 | } |
746 | ||
1fa8f6d6 VG |
747 | static inline s64 |
748 | cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
749 | { | |
750 | return cfqg->vdisktime - st->min_vdisktime; | |
751 | } | |
752 | ||
753 | static void | |
754 | __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
755 | { | |
756 | struct rb_node **node = &st->rb.rb_node; | |
757 | struct rb_node *parent = NULL; | |
758 | struct cfq_group *__cfqg; | |
759 | s64 key = cfqg_key(st, cfqg); | |
760 | int left = 1; | |
761 | ||
762 | while (*node != NULL) { | |
763 | parent = *node; | |
764 | __cfqg = rb_entry_cfqg(parent); | |
765 | ||
766 | if (key < cfqg_key(st, __cfqg)) | |
767 | node = &parent->rb_left; | |
768 | else { | |
769 | node = &parent->rb_right; | |
770 | left = 0; | |
771 | } | |
772 | } | |
773 | ||
774 | if (left) | |
775 | st->left = &cfqg->rb_node; | |
776 | ||
777 | rb_link_node(&cfqg->rb_node, parent, node); | |
778 | rb_insert_color(&cfqg->rb_node, &st->rb); | |
779 | } | |
780 | ||
781 | static void | |
782 | cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
783 | { | |
784 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
785 | struct cfq_group *__cfqg; | |
786 | struct rb_node *n; | |
787 | ||
788 | cfqg->nr_cfqq++; | |
789 | if (cfqg->on_st) | |
790 | return; | |
791 | ||
792 | /* | |
793 | * Currently put the group at the end. Later implement something | |
794 | * so that groups get lesser vtime based on their weights, so that | |
795 | * if group does not loose all if it was not continously backlogged. | |
796 | */ | |
797 | n = rb_last(&st->rb); | |
798 | if (n) { | |
799 | __cfqg = rb_entry_cfqg(n); | |
800 | cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; | |
801 | } else | |
802 | cfqg->vdisktime = st->min_vdisktime; | |
803 | ||
804 | __cfq_group_service_tree_add(st, cfqg); | |
805 | cfqg->on_st = true; | |
58ff82f3 VG |
806 | cfqd->nr_groups++; |
807 | st->total_weight += cfqg->weight; | |
1fa8f6d6 VG |
808 | } |
809 | ||
810 | static void | |
811 | cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
812 | { | |
813 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
814 | ||
25bc6b07 VG |
815 | if (st->active == &cfqg->rb_node) |
816 | st->active = NULL; | |
817 | ||
1fa8f6d6 VG |
818 | BUG_ON(cfqg->nr_cfqq < 1); |
819 | cfqg->nr_cfqq--; | |
25bc6b07 | 820 | |
1fa8f6d6 VG |
821 | /* If there are other cfq queues under this group, don't delete it */ |
822 | if (cfqg->nr_cfqq) | |
823 | return; | |
824 | ||
825 | cfqg->on_st = false; | |
58ff82f3 VG |
826 | cfqd->nr_groups--; |
827 | st->total_weight -= cfqg->weight; | |
1fa8f6d6 VG |
828 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) |
829 | cfq_rb_erase(&cfqg->rb_node, st); | |
dae739eb VG |
830 | cfqg->saved_workload_slice = 0; |
831 | } | |
832 | ||
833 | static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq) | |
834 | { | |
835 | unsigned int slice_used, allocated_slice; | |
836 | ||
837 | /* | |
838 | * Queue got expired before even a single request completed or | |
839 | * got expired immediately after first request completion. | |
840 | */ | |
841 | if (!cfqq->slice_start || cfqq->slice_start == jiffies) { | |
842 | /* | |
843 | * Also charge the seek time incurred to the group, otherwise | |
844 | * if there are mutiple queues in the group, each can dispatch | |
845 | * a single request on seeky media and cause lots of seek time | |
846 | * and group will never know it. | |
847 | */ | |
848 | slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start), | |
849 | 1); | |
850 | } else { | |
851 | slice_used = jiffies - cfqq->slice_start; | |
852 | allocated_slice = cfqq->slice_end - cfqq->slice_start; | |
853 | if (slice_used > allocated_slice) | |
854 | slice_used = allocated_slice; | |
855 | } | |
856 | ||
857 | cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used); | |
858 | return slice_used; | |
859 | } | |
860 | ||
861 | static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg, | |
862 | struct cfq_queue *cfqq) | |
863 | { | |
864 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
865 | unsigned int used_sl; | |
866 | ||
867 | used_sl = cfq_cfqq_slice_usage(cfqq); | |
868 | ||
869 | /* Can't update vdisktime while group is on service tree */ | |
870 | cfq_rb_erase(&cfqg->rb_node, st); | |
871 | cfqg->vdisktime += cfq_scale_slice(used_sl, cfqg); | |
872 | __cfq_group_service_tree_add(st, cfqg); | |
873 | ||
874 | /* This group is being expired. Save the context */ | |
875 | if (time_after(cfqd->workload_expires, jiffies)) { | |
876 | cfqg->saved_workload_slice = cfqd->workload_expires | |
877 | - jiffies; | |
878 | cfqg->saved_workload = cfqd->serving_type; | |
879 | cfqg->saved_serving_prio = cfqd->serving_prio; | |
880 | } else | |
881 | cfqg->saved_workload_slice = 0; | |
1fa8f6d6 VG |
882 | } |
883 | ||
498d3aa2 | 884 | /* |
c0324a02 | 885 | * The cfqd->service_trees holds all pending cfq_queue's that have |
498d3aa2 JA |
886 | * requests waiting to be processed. It is sorted in the order that |
887 | * we will service the queues. | |
888 | */ | |
a36e71f9 | 889 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 890 | bool add_front) |
d9e7620e | 891 | { |
0871714e JA |
892 | struct rb_node **p, *parent; |
893 | struct cfq_queue *__cfqq; | |
d9e7620e | 894 | unsigned long rb_key; |
c0324a02 | 895 | struct cfq_rb_root *service_tree; |
498d3aa2 | 896 | int left; |
dae739eb | 897 | int new_cfqq = 1; |
d9e7620e | 898 | |
cdb16e8f VG |
899 | service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), |
900 | cfqq_type(cfqq), cfqd); | |
0871714e JA |
901 | if (cfq_class_idle(cfqq)) { |
902 | rb_key = CFQ_IDLE_DELAY; | |
aa6f6a3d | 903 | parent = rb_last(&service_tree->rb); |
0871714e JA |
904 | if (parent && parent != &cfqq->rb_node) { |
905 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
906 | rb_key += __cfqq->rb_key; | |
907 | } else | |
908 | rb_key += jiffies; | |
909 | } else if (!add_front) { | |
b9c8946b JA |
910 | /* |
911 | * Get our rb key offset. Subtract any residual slice | |
912 | * value carried from last service. A negative resid | |
913 | * count indicates slice overrun, and this should position | |
914 | * the next service time further away in the tree. | |
915 | */ | |
edd75ffd | 916 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
b9c8946b | 917 | rb_key -= cfqq->slice_resid; |
edd75ffd | 918 | cfqq->slice_resid = 0; |
48e025e6 CZ |
919 | } else { |
920 | rb_key = -HZ; | |
aa6f6a3d | 921 | __cfqq = cfq_rb_first(service_tree); |
48e025e6 CZ |
922 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; |
923 | } | |
1da177e4 | 924 | |
d9e7620e | 925 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
dae739eb | 926 | new_cfqq = 0; |
99f9628a | 927 | /* |
d9e7620e | 928 | * same position, nothing more to do |
99f9628a | 929 | */ |
c0324a02 CZ |
930 | if (rb_key == cfqq->rb_key && |
931 | cfqq->service_tree == service_tree) | |
d9e7620e | 932 | return; |
1da177e4 | 933 | |
aa6f6a3d CZ |
934 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); |
935 | cfqq->service_tree = NULL; | |
1da177e4 | 936 | } |
d9e7620e | 937 | |
498d3aa2 | 938 | left = 1; |
0871714e | 939 | parent = NULL; |
aa6f6a3d CZ |
940 | cfqq->service_tree = service_tree; |
941 | p = &service_tree->rb.rb_node; | |
d9e7620e | 942 | while (*p) { |
67060e37 | 943 | struct rb_node **n; |
cc09e299 | 944 | |
d9e7620e JA |
945 | parent = *p; |
946 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
947 | ||
0c534e0a | 948 | /* |
c0324a02 | 949 | * sort by key, that represents service time. |
0c534e0a | 950 | */ |
c0324a02 | 951 | if (time_before(rb_key, __cfqq->rb_key)) |
67060e37 | 952 | n = &(*p)->rb_left; |
c0324a02 | 953 | else { |
67060e37 | 954 | n = &(*p)->rb_right; |
cc09e299 | 955 | left = 0; |
c0324a02 | 956 | } |
67060e37 JA |
957 | |
958 | p = n; | |
d9e7620e JA |
959 | } |
960 | ||
cc09e299 | 961 | if (left) |
aa6f6a3d | 962 | service_tree->left = &cfqq->rb_node; |
cc09e299 | 963 | |
d9e7620e JA |
964 | cfqq->rb_key = rb_key; |
965 | rb_link_node(&cfqq->rb_node, parent, p); | |
aa6f6a3d CZ |
966 | rb_insert_color(&cfqq->rb_node, &service_tree->rb); |
967 | service_tree->count++; | |
dae739eb VG |
968 | if (add_front || !new_cfqq) |
969 | return; | |
1fa8f6d6 | 970 | cfq_group_service_tree_add(cfqd, cfqq->cfqg); |
1da177e4 LT |
971 | } |
972 | ||
a36e71f9 | 973 | static struct cfq_queue * |
f2d1f0ae JA |
974 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
975 | sector_t sector, struct rb_node **ret_parent, | |
976 | struct rb_node ***rb_link) | |
a36e71f9 | 977 | { |
a36e71f9 JA |
978 | struct rb_node **p, *parent; |
979 | struct cfq_queue *cfqq = NULL; | |
980 | ||
981 | parent = NULL; | |
982 | p = &root->rb_node; | |
983 | while (*p) { | |
984 | struct rb_node **n; | |
985 | ||
986 | parent = *p; | |
987 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
988 | ||
989 | /* | |
990 | * Sort strictly based on sector. Smallest to the left, | |
991 | * largest to the right. | |
992 | */ | |
2e46e8b2 | 993 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 994 | n = &(*p)->rb_right; |
2e46e8b2 | 995 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
996 | n = &(*p)->rb_left; |
997 | else | |
998 | break; | |
999 | p = n; | |
3ac6c9f8 | 1000 | cfqq = NULL; |
a36e71f9 JA |
1001 | } |
1002 | ||
1003 | *ret_parent = parent; | |
1004 | if (rb_link) | |
1005 | *rb_link = p; | |
3ac6c9f8 | 1006 | return cfqq; |
a36e71f9 JA |
1007 | } |
1008 | ||
1009 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1010 | { | |
a36e71f9 JA |
1011 | struct rb_node **p, *parent; |
1012 | struct cfq_queue *__cfqq; | |
1013 | ||
f2d1f0ae JA |
1014 | if (cfqq->p_root) { |
1015 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1016 | cfqq->p_root = NULL; | |
1017 | } | |
a36e71f9 JA |
1018 | |
1019 | if (cfq_class_idle(cfqq)) | |
1020 | return; | |
1021 | if (!cfqq->next_rq) | |
1022 | return; | |
1023 | ||
f2d1f0ae | 1024 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
1025 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
1026 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
1027 | if (!__cfqq) { |
1028 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
1029 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
1030 | } else | |
1031 | cfqq->p_root = NULL; | |
a36e71f9 JA |
1032 | } |
1033 | ||
498d3aa2 JA |
1034 | /* |
1035 | * Update cfqq's position in the service tree. | |
1036 | */ | |
edd75ffd | 1037 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 1038 | { |
6d048f53 JA |
1039 | /* |
1040 | * Resorting requires the cfqq to be on the RR list already. | |
1041 | */ | |
a36e71f9 | 1042 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 1043 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
1044 | cfq_prio_tree_add(cfqd, cfqq); |
1045 | } | |
6d048f53 JA |
1046 | } |
1047 | ||
1da177e4 LT |
1048 | /* |
1049 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 1050 | * the pending list according to last request service |
1da177e4 | 1051 | */ |
febffd61 | 1052 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1053 | { |
7b679138 | 1054 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
1055 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1056 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 LT |
1057 | cfqd->busy_queues++; |
1058 | ||
edd75ffd | 1059 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
1060 | } |
1061 | ||
498d3aa2 JA |
1062 | /* |
1063 | * Called when the cfqq no longer has requests pending, remove it from | |
1064 | * the service tree. | |
1065 | */ | |
febffd61 | 1066 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1067 | { |
7b679138 | 1068 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
1069 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
1070 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 1071 | |
aa6f6a3d CZ |
1072 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
1073 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
1074 | cfqq->service_tree = NULL; | |
1075 | } | |
f2d1f0ae JA |
1076 | if (cfqq->p_root) { |
1077 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1078 | cfqq->p_root = NULL; | |
1079 | } | |
d9e7620e | 1080 | |
1fa8f6d6 | 1081 | cfq_group_service_tree_del(cfqd, cfqq->cfqg); |
1da177e4 LT |
1082 | BUG_ON(!cfqd->busy_queues); |
1083 | cfqd->busy_queues--; | |
1084 | } | |
1085 | ||
1086 | /* | |
1087 | * rb tree support functions | |
1088 | */ | |
febffd61 | 1089 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 1090 | { |
5e705374 | 1091 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
5e705374 | 1092 | const int sync = rq_is_sync(rq); |
1da177e4 | 1093 | |
b4878f24 JA |
1094 | BUG_ON(!cfqq->queued[sync]); |
1095 | cfqq->queued[sync]--; | |
1da177e4 | 1096 | |
5e705374 | 1097 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 1098 | |
f04a6424 VG |
1099 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { |
1100 | /* | |
1101 | * Queue will be deleted from service tree when we actually | |
1102 | * expire it later. Right now just remove it from prio tree | |
1103 | * as it is empty. | |
1104 | */ | |
1105 | if (cfqq->p_root) { | |
1106 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1107 | cfqq->p_root = NULL; | |
1108 | } | |
1109 | } | |
1da177e4 LT |
1110 | } |
1111 | ||
5e705374 | 1112 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 1113 | { |
5e705374 | 1114 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 1115 | struct cfq_data *cfqd = cfqq->cfqd; |
a36e71f9 | 1116 | struct request *__alias, *prev; |
1da177e4 | 1117 | |
5380a101 | 1118 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
1119 | |
1120 | /* | |
1121 | * looks a little odd, but the first insert might return an alias. | |
1122 | * if that happens, put the alias on the dispatch list | |
1123 | */ | |
21183b07 | 1124 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 1125 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
1126 | |
1127 | if (!cfq_cfqq_on_rr(cfqq)) | |
1128 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
1129 | |
1130 | /* | |
1131 | * check if this request is a better next-serve candidate | |
1132 | */ | |
a36e71f9 | 1133 | prev = cfqq->next_rq; |
cf7c25cf | 1134 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); |
a36e71f9 JA |
1135 | |
1136 | /* | |
1137 | * adjust priority tree position, if ->next_rq changes | |
1138 | */ | |
1139 | if (prev != cfqq->next_rq) | |
1140 | cfq_prio_tree_add(cfqd, cfqq); | |
1141 | ||
5044eed4 | 1142 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
1143 | } |
1144 | ||
febffd61 | 1145 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 1146 | { |
5380a101 JA |
1147 | elv_rb_del(&cfqq->sort_list, rq); |
1148 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 1149 | cfq_add_rq_rb(rq); |
1da177e4 LT |
1150 | } |
1151 | ||
206dc69b JA |
1152 | static struct request * |
1153 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 1154 | { |
206dc69b | 1155 | struct task_struct *tsk = current; |
91fac317 | 1156 | struct cfq_io_context *cic; |
206dc69b | 1157 | struct cfq_queue *cfqq; |
1da177e4 | 1158 | |
4ac845a2 | 1159 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
1160 | if (!cic) |
1161 | return NULL; | |
1162 | ||
1163 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
1164 | if (cfqq) { |
1165 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
1166 | ||
21183b07 | 1167 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 1168 | } |
1da177e4 | 1169 | |
1da177e4 LT |
1170 | return NULL; |
1171 | } | |
1172 | ||
165125e1 | 1173 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1174 | { |
22e2c507 | 1175 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 1176 | |
5ad531db | 1177 | cfqd->rq_in_driver[rq_is_sync(rq)]++; |
7b679138 | 1178 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
5ad531db | 1179 | rq_in_driver(cfqd)); |
25776e35 | 1180 | |
5b93629b | 1181 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
1182 | } |
1183 | ||
165125e1 | 1184 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1185 | { |
b4878f24 | 1186 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5ad531db | 1187 | const int sync = rq_is_sync(rq); |
b4878f24 | 1188 | |
5ad531db JA |
1189 | WARN_ON(!cfqd->rq_in_driver[sync]); |
1190 | cfqd->rq_in_driver[sync]--; | |
7b679138 | 1191 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
5ad531db | 1192 | rq_in_driver(cfqd)); |
1da177e4 LT |
1193 | } |
1194 | ||
b4878f24 | 1195 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 1196 | { |
5e705374 | 1197 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 1198 | |
5e705374 JA |
1199 | if (cfqq->next_rq == rq) |
1200 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 1201 | |
b4878f24 | 1202 | list_del_init(&rq->queuelist); |
5e705374 | 1203 | cfq_del_rq_rb(rq); |
374f84ac | 1204 | |
45333d5a | 1205 | cfqq->cfqd->rq_queued--; |
374f84ac JA |
1206 | if (rq_is_meta(rq)) { |
1207 | WARN_ON(!cfqq->meta_pending); | |
1208 | cfqq->meta_pending--; | |
1209 | } | |
1da177e4 LT |
1210 | } |
1211 | ||
165125e1 JA |
1212 | static int cfq_merge(struct request_queue *q, struct request **req, |
1213 | struct bio *bio) | |
1da177e4 LT |
1214 | { |
1215 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1216 | struct request *__rq; | |
1da177e4 | 1217 | |
206dc69b | 1218 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 1219 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
1220 | *req = __rq; |
1221 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1222 | } |
1223 | ||
1224 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1225 | } |
1226 | ||
165125e1 | 1227 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 1228 | int type) |
1da177e4 | 1229 | { |
21183b07 | 1230 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 1231 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 1232 | |
5e705374 | 1233 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 1234 | } |
1da177e4 LT |
1235 | } |
1236 | ||
1237 | static void | |
165125e1 | 1238 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
1239 | struct request *next) |
1240 | { | |
cf7c25cf | 1241 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 JA |
1242 | /* |
1243 | * reposition in fifo if next is older than rq | |
1244 | */ | |
1245 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
30996f40 | 1246 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
22e2c507 | 1247 | list_move(&rq->queuelist, &next->queuelist); |
30996f40 JA |
1248 | rq_set_fifo_time(rq, rq_fifo_time(next)); |
1249 | } | |
22e2c507 | 1250 | |
cf7c25cf CZ |
1251 | if (cfqq->next_rq == next) |
1252 | cfqq->next_rq = rq; | |
b4878f24 | 1253 | cfq_remove_request(next); |
22e2c507 JA |
1254 | } |
1255 | ||
165125e1 | 1256 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
1257 | struct bio *bio) |
1258 | { | |
1259 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 1260 | struct cfq_io_context *cic; |
da775265 | 1261 | struct cfq_queue *cfqq; |
da775265 JA |
1262 | |
1263 | /* | |
ec8acb69 | 1264 | * Disallow merge of a sync bio into an async request. |
da775265 | 1265 | */ |
91fac317 | 1266 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
a6151c3a | 1267 | return false; |
da775265 JA |
1268 | |
1269 | /* | |
719d3402 JA |
1270 | * Lookup the cfqq that this bio will be queued with. Allow |
1271 | * merge only if rq is queued there. | |
da775265 | 1272 | */ |
4ac845a2 | 1273 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 | 1274 | if (!cic) |
a6151c3a | 1275 | return false; |
719d3402 | 1276 | |
91fac317 | 1277 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
a6151c3a | 1278 | return cfqq == RQ_CFQQ(rq); |
da775265 JA |
1279 | } |
1280 | ||
febffd61 JA |
1281 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
1282 | struct cfq_queue *cfqq) | |
22e2c507 JA |
1283 | { |
1284 | if (cfqq) { | |
7b679138 | 1285 | cfq_log_cfqq(cfqd, cfqq, "set_active"); |
dae739eb VG |
1286 | cfqq->slice_start = 0; |
1287 | cfqq->dispatch_start = jiffies; | |
22e2c507 | 1288 | cfqq->slice_end = 0; |
2f5cb738 JA |
1289 | cfqq->slice_dispatch = 0; |
1290 | ||
2f5cb738 | 1291 | cfq_clear_cfqq_wait_request(cfqq); |
b029195d | 1292 | cfq_clear_cfqq_must_dispatch(cfqq); |
3b18152c JA |
1293 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
1294 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 1295 | cfq_mark_cfqq_slice_new(cfqq); |
2f5cb738 JA |
1296 | |
1297 | del_timer(&cfqd->idle_slice_timer); | |
22e2c507 JA |
1298 | } |
1299 | ||
1300 | cfqd->active_queue = cfqq; | |
1301 | } | |
1302 | ||
7b14e3b5 JA |
1303 | /* |
1304 | * current cfqq expired its slice (or was too idle), select new one | |
1305 | */ | |
1306 | static void | |
1307 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
a6151c3a | 1308 | bool timed_out) |
7b14e3b5 | 1309 | { |
7b679138 JA |
1310 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
1311 | ||
7b14e3b5 JA |
1312 | if (cfq_cfqq_wait_request(cfqq)) |
1313 | del_timer(&cfqd->idle_slice_timer); | |
1314 | ||
7b14e3b5 JA |
1315 | cfq_clear_cfqq_wait_request(cfqq); |
1316 | ||
1317 | /* | |
6084cdda | 1318 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 1319 | */ |
7b679138 | 1320 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) { |
c5b680f3 | 1321 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b679138 JA |
1322 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
1323 | } | |
7b14e3b5 | 1324 | |
dae739eb VG |
1325 | cfq_group_served(cfqd, cfqq->cfqg, cfqq); |
1326 | ||
f04a6424 VG |
1327 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
1328 | cfq_del_cfqq_rr(cfqd, cfqq); | |
1329 | ||
edd75ffd | 1330 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
1331 | |
1332 | if (cfqq == cfqd->active_queue) | |
1333 | cfqd->active_queue = NULL; | |
1334 | ||
dae739eb VG |
1335 | if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active) |
1336 | cfqd->grp_service_tree.active = NULL; | |
1337 | ||
7b14e3b5 JA |
1338 | if (cfqd->active_cic) { |
1339 | put_io_context(cfqd->active_cic->ioc); | |
1340 | cfqd->active_cic = NULL; | |
1341 | } | |
7b14e3b5 JA |
1342 | } |
1343 | ||
a6151c3a | 1344 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) |
7b14e3b5 JA |
1345 | { |
1346 | struct cfq_queue *cfqq = cfqd->active_queue; | |
1347 | ||
1348 | if (cfqq) | |
6084cdda | 1349 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
1350 | } |
1351 | ||
498d3aa2 JA |
1352 | /* |
1353 | * Get next queue for service. Unless we have a queue preemption, | |
1354 | * we'll simply select the first cfqq in the service tree. | |
1355 | */ | |
6d048f53 | 1356 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 1357 | { |
c0324a02 | 1358 | struct cfq_rb_root *service_tree = |
cdb16e8f VG |
1359 | service_tree_for(cfqd->serving_group, cfqd->serving_prio, |
1360 | cfqd->serving_type, cfqd); | |
d9e7620e | 1361 | |
f04a6424 VG |
1362 | if (!cfqd->rq_queued) |
1363 | return NULL; | |
1364 | ||
1fa8f6d6 VG |
1365 | /* There is nothing to dispatch */ |
1366 | if (!service_tree) | |
1367 | return NULL; | |
c0324a02 CZ |
1368 | if (RB_EMPTY_ROOT(&service_tree->rb)) |
1369 | return NULL; | |
1370 | return cfq_rb_first(service_tree); | |
6d048f53 JA |
1371 | } |
1372 | ||
f04a6424 VG |
1373 | static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) |
1374 | { | |
1375 | struct cfq_group *cfqg = &cfqd->root_group; | |
1376 | struct cfq_queue *cfqq; | |
1377 | int i, j; | |
1378 | struct cfq_rb_root *st; | |
1379 | ||
1380 | if (!cfqd->rq_queued) | |
1381 | return NULL; | |
1382 | ||
1383 | for_each_cfqg_st(cfqg, i, j, st) | |
1384 | if ((cfqq = cfq_rb_first(st)) != NULL) | |
1385 | return cfqq; | |
1386 | return NULL; | |
1387 | } | |
1388 | ||
498d3aa2 JA |
1389 | /* |
1390 | * Get and set a new active queue for service. | |
1391 | */ | |
a36e71f9 JA |
1392 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
1393 | struct cfq_queue *cfqq) | |
6d048f53 | 1394 | { |
e00ef799 | 1395 | if (!cfqq) |
a36e71f9 | 1396 | cfqq = cfq_get_next_queue(cfqd); |
6d048f53 | 1397 | |
22e2c507 | 1398 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 1399 | return cfqq; |
22e2c507 JA |
1400 | } |
1401 | ||
d9e7620e JA |
1402 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
1403 | struct request *rq) | |
1404 | { | |
83096ebf TH |
1405 | if (blk_rq_pos(rq) >= cfqd->last_position) |
1406 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 1407 | else |
83096ebf | 1408 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
1409 | } |
1410 | ||
b2c18e1e JM |
1411 | #define CFQQ_SEEK_THR 8 * 1024 |
1412 | #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR) | |
04dc6e71 | 1413 | |
b2c18e1e JM |
1414 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
1415 | struct request *rq) | |
6d048f53 | 1416 | { |
b2c18e1e | 1417 | sector_t sdist = cfqq->seek_mean; |
6d048f53 | 1418 | |
b2c18e1e JM |
1419 | if (!sample_valid(cfqq->seek_samples)) |
1420 | sdist = CFQQ_SEEK_THR; | |
6d048f53 | 1421 | |
04dc6e71 | 1422 | return cfq_dist_from_last(cfqd, rq) <= sdist; |
6d048f53 JA |
1423 | } |
1424 | ||
a36e71f9 JA |
1425 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
1426 | struct cfq_queue *cur_cfqq) | |
1427 | { | |
f2d1f0ae | 1428 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
1429 | struct rb_node *parent, *node; |
1430 | struct cfq_queue *__cfqq; | |
1431 | sector_t sector = cfqd->last_position; | |
1432 | ||
1433 | if (RB_EMPTY_ROOT(root)) | |
1434 | return NULL; | |
1435 | ||
1436 | /* | |
1437 | * First, if we find a request starting at the end of the last | |
1438 | * request, choose it. | |
1439 | */ | |
f2d1f0ae | 1440 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
1441 | if (__cfqq) |
1442 | return __cfqq; | |
1443 | ||
1444 | /* | |
1445 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1446 | * will contain the closest sector. | |
1447 | */ | |
1448 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
b2c18e1e | 1449 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1450 | return __cfqq; |
1451 | ||
2e46e8b2 | 1452 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1453 | node = rb_next(&__cfqq->p_node); |
1454 | else | |
1455 | node = rb_prev(&__cfqq->p_node); | |
1456 | if (!node) | |
1457 | return NULL; | |
1458 | ||
1459 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
b2c18e1e | 1460 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1461 | return __cfqq; |
1462 | ||
1463 | return NULL; | |
1464 | } | |
1465 | ||
1466 | /* | |
1467 | * cfqd - obvious | |
1468 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1469 | * closely cooperating with itself. | |
1470 | * | |
1471 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1472 | * one request, and that cfqd->last_position reflects a position on the disk | |
1473 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1474 | * assumption. | |
1475 | */ | |
1476 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
b3b6d040 | 1477 | struct cfq_queue *cur_cfqq) |
6d048f53 | 1478 | { |
a36e71f9 JA |
1479 | struct cfq_queue *cfqq; |
1480 | ||
e6c5bc73 JM |
1481 | if (!cfq_cfqq_sync(cur_cfqq)) |
1482 | return NULL; | |
1483 | if (CFQQ_SEEKY(cur_cfqq)) | |
1484 | return NULL; | |
1485 | ||
6d048f53 | 1486 | /* |
d9e7620e JA |
1487 | * We should notice if some of the queues are cooperating, eg |
1488 | * working closely on the same area of the disk. In that case, | |
1489 | * we can group them together and don't waste time idling. | |
6d048f53 | 1490 | */ |
a36e71f9 JA |
1491 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1492 | if (!cfqq) | |
1493 | return NULL; | |
1494 | ||
df5fe3e8 JM |
1495 | /* |
1496 | * It only makes sense to merge sync queues. | |
1497 | */ | |
1498 | if (!cfq_cfqq_sync(cfqq)) | |
1499 | return NULL; | |
e6c5bc73 JM |
1500 | if (CFQQ_SEEKY(cfqq)) |
1501 | return NULL; | |
df5fe3e8 | 1502 | |
c0324a02 CZ |
1503 | /* |
1504 | * Do not merge queues of different priority classes | |
1505 | */ | |
1506 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
1507 | return NULL; | |
1508 | ||
a36e71f9 | 1509 | return cfqq; |
6d048f53 JA |
1510 | } |
1511 | ||
a6d44e98 CZ |
1512 | /* |
1513 | * Determine whether we should enforce idle window for this queue. | |
1514 | */ | |
1515 | ||
1516 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1517 | { | |
1518 | enum wl_prio_t prio = cfqq_prio(cfqq); | |
718eee05 | 1519 | struct cfq_rb_root *service_tree = cfqq->service_tree; |
a6d44e98 | 1520 | |
f04a6424 VG |
1521 | BUG_ON(!service_tree); |
1522 | BUG_ON(!service_tree->count); | |
1523 | ||
a6d44e98 CZ |
1524 | /* We never do for idle class queues. */ |
1525 | if (prio == IDLE_WORKLOAD) | |
1526 | return false; | |
1527 | ||
1528 | /* We do for queues that were marked with idle window flag. */ | |
1529 | if (cfq_cfqq_idle_window(cfqq)) | |
1530 | return true; | |
1531 | ||
1532 | /* | |
1533 | * Otherwise, we do only if they are the last ones | |
1534 | * in their service tree. | |
1535 | */ | |
f04a6424 | 1536 | return service_tree->count == 1; |
a6d44e98 CZ |
1537 | } |
1538 | ||
6d048f53 | 1539 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1540 | { |
1792669c | 1541 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1542 | struct cfq_io_context *cic; |
7b14e3b5 JA |
1543 | unsigned long sl; |
1544 | ||
a68bbddb | 1545 | /* |
f7d7b7a7 JA |
1546 | * SSD device without seek penalty, disable idling. But only do so |
1547 | * for devices that support queuing, otherwise we still have a problem | |
1548 | * with sync vs async workloads. | |
a68bbddb | 1549 | */ |
f7d7b7a7 | 1550 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
1551 | return; |
1552 | ||
dd67d051 | 1553 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 1554 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
1555 | |
1556 | /* | |
1557 | * idle is disabled, either manually or by past process history | |
1558 | */ | |
a6d44e98 | 1559 | if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq)) |
6d048f53 JA |
1560 | return; |
1561 | ||
7b679138 | 1562 | /* |
8e550632 | 1563 | * still active requests from this queue, don't idle |
7b679138 | 1564 | */ |
8e550632 | 1565 | if (cfqq->dispatched) |
7b679138 JA |
1566 | return; |
1567 | ||
22e2c507 JA |
1568 | /* |
1569 | * task has exited, don't wait | |
1570 | */ | |
206dc69b | 1571 | cic = cfqd->active_cic; |
66dac98e | 1572 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
1573 | return; |
1574 | ||
355b659c CZ |
1575 | /* |
1576 | * If our average think time is larger than the remaining time | |
1577 | * slice, then don't idle. This avoids overrunning the allotted | |
1578 | * time slice. | |
1579 | */ | |
1580 | if (sample_valid(cic->ttime_samples) && | |
1581 | (cfqq->slice_end - jiffies < cic->ttime_mean)) | |
1582 | return; | |
1583 | ||
3b18152c | 1584 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 1585 | |
6d048f53 | 1586 | sl = cfqd->cfq_slice_idle; |
206dc69b | 1587 | |
7b14e3b5 | 1588 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
9481ffdc | 1589 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); |
1da177e4 LT |
1590 | } |
1591 | ||
498d3aa2 JA |
1592 | /* |
1593 | * Move request from internal lists to the request queue dispatch list. | |
1594 | */ | |
165125e1 | 1595 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 1596 | { |
3ed9a296 | 1597 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1598 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 1599 | |
7b679138 JA |
1600 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
1601 | ||
06d21886 | 1602 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); |
5380a101 | 1603 | cfq_remove_request(rq); |
6d048f53 | 1604 | cfqq->dispatched++; |
5380a101 | 1605 | elv_dispatch_sort(q, rq); |
3ed9a296 JA |
1606 | |
1607 | if (cfq_cfqq_sync(cfqq)) | |
1608 | cfqd->sync_flight++; | |
1da177e4 LT |
1609 | } |
1610 | ||
1611 | /* | |
1612 | * return expired entry, or NULL to just start from scratch in rbtree | |
1613 | */ | |
febffd61 | 1614 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 | 1615 | { |
30996f40 | 1616 | struct request *rq = NULL; |
1da177e4 | 1617 | |
3b18152c | 1618 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 1619 | return NULL; |
cb887411 JA |
1620 | |
1621 | cfq_mark_cfqq_fifo_expire(cfqq); | |
1622 | ||
89850f7e JA |
1623 | if (list_empty(&cfqq->fifo)) |
1624 | return NULL; | |
1da177e4 | 1625 | |
89850f7e | 1626 | rq = rq_entry_fifo(cfqq->fifo.next); |
30996f40 | 1627 | if (time_before(jiffies, rq_fifo_time(rq))) |
7b679138 | 1628 | rq = NULL; |
1da177e4 | 1629 | |
30996f40 | 1630 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 1631 | return rq; |
1da177e4 LT |
1632 | } |
1633 | ||
22e2c507 JA |
1634 | static inline int |
1635 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1636 | { | |
1637 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 1638 | |
22e2c507 | 1639 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 1640 | |
22e2c507 | 1641 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
1642 | } |
1643 | ||
df5fe3e8 JM |
1644 | /* |
1645 | * Must be called with the queue_lock held. | |
1646 | */ | |
1647 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
1648 | { | |
1649 | int process_refs, io_refs; | |
1650 | ||
1651 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
1652 | process_refs = atomic_read(&cfqq->ref) - io_refs; | |
1653 | BUG_ON(process_refs < 0); | |
1654 | return process_refs; | |
1655 | } | |
1656 | ||
1657 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
1658 | { | |
e6c5bc73 | 1659 | int process_refs, new_process_refs; |
df5fe3e8 JM |
1660 | struct cfq_queue *__cfqq; |
1661 | ||
1662 | /* Avoid a circular list and skip interim queue merges */ | |
1663 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
1664 | if (__cfqq == cfqq) | |
1665 | return; | |
1666 | new_cfqq = __cfqq; | |
1667 | } | |
1668 | ||
1669 | process_refs = cfqq_process_refs(cfqq); | |
1670 | /* | |
1671 | * If the process for the cfqq has gone away, there is no | |
1672 | * sense in merging the queues. | |
1673 | */ | |
1674 | if (process_refs == 0) | |
1675 | return; | |
1676 | ||
e6c5bc73 JM |
1677 | /* |
1678 | * Merge in the direction of the lesser amount of work. | |
1679 | */ | |
1680 | new_process_refs = cfqq_process_refs(new_cfqq); | |
1681 | if (new_process_refs >= process_refs) { | |
1682 | cfqq->new_cfqq = new_cfqq; | |
1683 | atomic_add(process_refs, &new_cfqq->ref); | |
1684 | } else { | |
1685 | new_cfqq->new_cfqq = cfqq; | |
1686 | atomic_add(new_process_refs, &cfqq->ref); | |
1687 | } | |
df5fe3e8 JM |
1688 | } |
1689 | ||
cdb16e8f VG |
1690 | static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, |
1691 | struct cfq_group *cfqg, enum wl_prio_t prio, | |
1692 | bool prio_changed) | |
718eee05 CZ |
1693 | { |
1694 | struct cfq_queue *queue; | |
1695 | int i; | |
1696 | bool key_valid = false; | |
1697 | unsigned long lowest_key = 0; | |
1698 | enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; | |
1699 | ||
1700 | if (prio_changed) { | |
1701 | /* | |
1702 | * When priorities switched, we prefer starting | |
1703 | * from SYNC_NOIDLE (first choice), or just SYNC | |
1704 | * over ASYNC | |
1705 | */ | |
cdb16e8f | 1706 | if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) |
718eee05 CZ |
1707 | return cur_best; |
1708 | cur_best = SYNC_WORKLOAD; | |
cdb16e8f | 1709 | if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) |
718eee05 CZ |
1710 | return cur_best; |
1711 | ||
1712 | return ASYNC_WORKLOAD; | |
1713 | } | |
1714 | ||
1715 | for (i = 0; i < 3; ++i) { | |
1716 | /* otherwise, select the one with lowest rb_key */ | |
cdb16e8f | 1717 | queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd)); |
718eee05 CZ |
1718 | if (queue && |
1719 | (!key_valid || time_before(queue->rb_key, lowest_key))) { | |
1720 | lowest_key = queue->rb_key; | |
1721 | cur_best = i; | |
1722 | key_valid = true; | |
1723 | } | |
1724 | } | |
1725 | ||
1726 | return cur_best; | |
1727 | } | |
1728 | ||
cdb16e8f | 1729 | static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) |
718eee05 CZ |
1730 | { |
1731 | enum wl_prio_t previous_prio = cfqd->serving_prio; | |
1732 | bool prio_changed; | |
1733 | unsigned slice; | |
1734 | unsigned count; | |
cdb16e8f | 1735 | struct cfq_rb_root *st; |
58ff82f3 | 1736 | unsigned group_slice; |
718eee05 | 1737 | |
1fa8f6d6 VG |
1738 | if (!cfqg) { |
1739 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1740 | cfqd->workload_expires = jiffies + 1; | |
1741 | return; | |
1742 | } | |
1743 | ||
718eee05 | 1744 | /* Choose next priority. RT > BE > IDLE */ |
58ff82f3 | 1745 | if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) |
718eee05 | 1746 | cfqd->serving_prio = RT_WORKLOAD; |
58ff82f3 | 1747 | else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) |
718eee05 CZ |
1748 | cfqd->serving_prio = BE_WORKLOAD; |
1749 | else { | |
1750 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1751 | cfqd->workload_expires = jiffies + 1; | |
1752 | return; | |
1753 | } | |
1754 | ||
1755 | /* | |
1756 | * For RT and BE, we have to choose also the type | |
1757 | * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload | |
1758 | * expiration time | |
1759 | */ | |
1760 | prio_changed = (cfqd->serving_prio != previous_prio); | |
cdb16e8f VG |
1761 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, |
1762 | cfqd); | |
1763 | count = st->count; | |
718eee05 CZ |
1764 | |
1765 | /* | |
1766 | * If priority didn't change, check workload expiration, | |
1767 | * and that we still have other queues ready | |
1768 | */ | |
1769 | if (!prio_changed && count && | |
1770 | !time_after(jiffies, cfqd->workload_expires)) | |
1771 | return; | |
1772 | ||
1773 | /* otherwise select new workload type */ | |
1774 | cfqd->serving_type = | |
cdb16e8f VG |
1775 | cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed); |
1776 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, | |
1777 | cfqd); | |
1778 | count = st->count; | |
718eee05 CZ |
1779 | |
1780 | /* | |
1781 | * the workload slice is computed as a fraction of target latency | |
1782 | * proportional to the number of queues in that workload, over | |
1783 | * all the queues in the same priority class | |
1784 | */ | |
58ff82f3 VG |
1785 | group_slice = cfq_group_slice(cfqd, cfqg); |
1786 | ||
1787 | slice = group_slice * count / | |
1788 | max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], | |
1789 | cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); | |
718eee05 CZ |
1790 | |
1791 | if (cfqd->serving_type == ASYNC_WORKLOAD) | |
1792 | /* async workload slice is scaled down according to | |
1793 | * the sync/async slice ratio. */ | |
1794 | slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; | |
1795 | else | |
1796 | /* sync workload slice is at least 2 * cfq_slice_idle */ | |
1797 | slice = max(slice, 2 * cfqd->cfq_slice_idle); | |
1798 | ||
1799 | slice = max_t(unsigned, slice, CFQ_MIN_TT); | |
1800 | cfqd->workload_expires = jiffies + slice; | |
8e550632 | 1801 | cfqd->noidle_tree_requires_idle = false; |
718eee05 CZ |
1802 | } |
1803 | ||
1fa8f6d6 VG |
1804 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) |
1805 | { | |
1806 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
25bc6b07 | 1807 | struct cfq_group *cfqg; |
1fa8f6d6 VG |
1808 | |
1809 | if (RB_EMPTY_ROOT(&st->rb)) | |
1810 | return NULL; | |
25bc6b07 VG |
1811 | cfqg = cfq_rb_first_group(st); |
1812 | st->active = &cfqg->rb_node; | |
1813 | update_min_vdisktime(st); | |
1814 | return cfqg; | |
1fa8f6d6 VG |
1815 | } |
1816 | ||
cdb16e8f VG |
1817 | static void cfq_choose_cfqg(struct cfq_data *cfqd) |
1818 | { | |
1fa8f6d6 VG |
1819 | struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); |
1820 | ||
1821 | cfqd->serving_group = cfqg; | |
dae739eb VG |
1822 | |
1823 | /* Restore the workload type data */ | |
1824 | if (cfqg->saved_workload_slice) { | |
1825 | cfqd->workload_expires = jiffies + cfqg->saved_workload_slice; | |
1826 | cfqd->serving_type = cfqg->saved_workload; | |
1827 | cfqd->serving_prio = cfqg->saved_serving_prio; | |
1828 | } | |
1fa8f6d6 | 1829 | choose_service_tree(cfqd, cfqg); |
cdb16e8f VG |
1830 | } |
1831 | ||
22e2c507 | 1832 | /* |
498d3aa2 JA |
1833 | * Select a queue for service. If we have a current active queue, |
1834 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 1835 | */ |
1b5ed5e1 | 1836 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 1837 | { |
a36e71f9 | 1838 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 1839 | |
22e2c507 JA |
1840 | cfqq = cfqd->active_queue; |
1841 | if (!cfqq) | |
1842 | goto new_queue; | |
1da177e4 | 1843 | |
f04a6424 VG |
1844 | if (!cfqd->rq_queued) |
1845 | return NULL; | |
22e2c507 | 1846 | /* |
6d048f53 | 1847 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 1848 | */ |
b029195d | 1849 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) |
3b18152c | 1850 | goto expire; |
1da177e4 | 1851 | |
22e2c507 | 1852 | /* |
6d048f53 JA |
1853 | * The active queue has requests and isn't expired, allow it to |
1854 | * dispatch. | |
22e2c507 | 1855 | */ |
dd67d051 | 1856 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 1857 | goto keep_queue; |
6d048f53 | 1858 | |
a36e71f9 JA |
1859 | /* |
1860 | * If another queue has a request waiting within our mean seek | |
1861 | * distance, let it run. The expire code will check for close | |
1862 | * cooperators and put the close queue at the front of the service | |
df5fe3e8 | 1863 | * tree. If possible, merge the expiring queue with the new cfqq. |
a36e71f9 | 1864 | */ |
b3b6d040 | 1865 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); |
df5fe3e8 JM |
1866 | if (new_cfqq) { |
1867 | if (!cfqq->new_cfqq) | |
1868 | cfq_setup_merge(cfqq, new_cfqq); | |
a36e71f9 | 1869 | goto expire; |
df5fe3e8 | 1870 | } |
a36e71f9 | 1871 | |
6d048f53 JA |
1872 | /* |
1873 | * No requests pending. If the active queue still has requests in | |
1874 | * flight or is idling for a new request, allow either of these | |
1875 | * conditions to happen (or time out) before selecting a new queue. | |
1876 | */ | |
cc197479 | 1877 | if (timer_pending(&cfqd->idle_slice_timer) || |
a6d44e98 | 1878 | (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) { |
caaa5f9f JA |
1879 | cfqq = NULL; |
1880 | goto keep_queue; | |
22e2c507 JA |
1881 | } |
1882 | ||
3b18152c | 1883 | expire: |
6084cdda | 1884 | cfq_slice_expired(cfqd, 0); |
3b18152c | 1885 | new_queue: |
718eee05 CZ |
1886 | /* |
1887 | * Current queue expired. Check if we have to switch to a new | |
1888 | * service tree | |
1889 | */ | |
1890 | if (!new_cfqq) | |
cdb16e8f | 1891 | cfq_choose_cfqg(cfqd); |
718eee05 | 1892 | |
a36e71f9 | 1893 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 1894 | keep_queue: |
3b18152c | 1895 | return cfqq; |
22e2c507 JA |
1896 | } |
1897 | ||
febffd61 | 1898 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
1899 | { |
1900 | int dispatched = 0; | |
1901 | ||
1902 | while (cfqq->next_rq) { | |
1903 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
1904 | dispatched++; | |
1905 | } | |
1906 | ||
1907 | BUG_ON(!list_empty(&cfqq->fifo)); | |
f04a6424 VG |
1908 | |
1909 | /* By default cfqq is not expired if it is empty. Do it explicitly */ | |
1910 | __cfq_slice_expired(cfqq->cfqd, cfqq, 0); | |
d9e7620e JA |
1911 | return dispatched; |
1912 | } | |
1913 | ||
498d3aa2 JA |
1914 | /* |
1915 | * Drain our current requests. Used for barriers and when switching | |
1916 | * io schedulers on-the-fly. | |
1917 | */ | |
d9e7620e | 1918 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 1919 | { |
0871714e | 1920 | struct cfq_queue *cfqq; |
d9e7620e | 1921 | int dispatched = 0; |
cdb16e8f | 1922 | |
f04a6424 VG |
1923 | while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) |
1924 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); | |
1b5ed5e1 | 1925 | |
6084cdda | 1926 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
1927 | BUG_ON(cfqd->busy_queues); |
1928 | ||
6923715a | 1929 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
1930 | return dispatched; |
1931 | } | |
1932 | ||
0b182d61 | 1933 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
2f5cb738 | 1934 | { |
2f5cb738 | 1935 | unsigned int max_dispatch; |
22e2c507 | 1936 | |
5ad531db JA |
1937 | /* |
1938 | * Drain async requests before we start sync IO | |
1939 | */ | |
a6d44e98 | 1940 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC]) |
0b182d61 | 1941 | return false; |
5ad531db | 1942 | |
2f5cb738 JA |
1943 | /* |
1944 | * If this is an async queue and we have sync IO in flight, let it wait | |
1945 | */ | |
1946 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) | |
0b182d61 | 1947 | return false; |
2f5cb738 JA |
1948 | |
1949 | max_dispatch = cfqd->cfq_quantum; | |
1950 | if (cfq_class_idle(cfqq)) | |
1951 | max_dispatch = 1; | |
b4878f24 | 1952 | |
2f5cb738 JA |
1953 | /* |
1954 | * Does this cfqq already have too much IO in flight? | |
1955 | */ | |
1956 | if (cfqq->dispatched >= max_dispatch) { | |
1957 | /* | |
1958 | * idle queue must always only have a single IO in flight | |
1959 | */ | |
3ed9a296 | 1960 | if (cfq_class_idle(cfqq)) |
0b182d61 | 1961 | return false; |
3ed9a296 | 1962 | |
2f5cb738 JA |
1963 | /* |
1964 | * We have other queues, don't allow more IO from this one | |
1965 | */ | |
1966 | if (cfqd->busy_queues > 1) | |
0b182d61 | 1967 | return false; |
9ede209e | 1968 | |
365722bb | 1969 | /* |
474b18cc | 1970 | * Sole queue user, no limit |
365722bb | 1971 | */ |
474b18cc | 1972 | max_dispatch = -1; |
8e296755 JA |
1973 | } |
1974 | ||
1975 | /* | |
1976 | * Async queues must wait a bit before being allowed dispatch. | |
1977 | * We also ramp up the dispatch depth gradually for async IO, | |
1978 | * based on the last sync IO we serviced | |
1979 | */ | |
963b72fc | 1980 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { |
8e296755 JA |
1981 | unsigned long last_sync = jiffies - cfqd->last_end_sync_rq; |
1982 | unsigned int depth; | |
365722bb | 1983 | |
61f0c1dc | 1984 | depth = last_sync / cfqd->cfq_slice[1]; |
e00c54c3 JA |
1985 | if (!depth && !cfqq->dispatched) |
1986 | depth = 1; | |
8e296755 JA |
1987 | if (depth < max_dispatch) |
1988 | max_dispatch = depth; | |
2f5cb738 | 1989 | } |
3ed9a296 | 1990 | |
0b182d61 JA |
1991 | /* |
1992 | * If we're below the current max, allow a dispatch | |
1993 | */ | |
1994 | return cfqq->dispatched < max_dispatch; | |
1995 | } | |
1996 | ||
1997 | /* | |
1998 | * Dispatch a request from cfqq, moving them to the request queue | |
1999 | * dispatch list. | |
2000 | */ | |
2001 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2002 | { | |
2003 | struct request *rq; | |
2004 | ||
2005 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
2006 | ||
2007 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
2008 | return false; | |
2009 | ||
2010 | /* | |
2011 | * follow expired path, else get first next available | |
2012 | */ | |
2013 | rq = cfq_check_fifo(cfqq); | |
2014 | if (!rq) | |
2015 | rq = cfqq->next_rq; | |
2016 | ||
2017 | /* | |
2018 | * insert request into driver dispatch list | |
2019 | */ | |
2020 | cfq_dispatch_insert(cfqd->queue, rq); | |
2021 | ||
2022 | if (!cfqd->active_cic) { | |
2023 | struct cfq_io_context *cic = RQ_CIC(rq); | |
2024 | ||
2025 | atomic_long_inc(&cic->ioc->refcount); | |
2026 | cfqd->active_cic = cic; | |
2027 | } | |
2028 | ||
2029 | return true; | |
2030 | } | |
2031 | ||
2032 | /* | |
2033 | * Find the cfqq that we need to service and move a request from that to the | |
2034 | * dispatch list | |
2035 | */ | |
2036 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
2037 | { | |
2038 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2039 | struct cfq_queue *cfqq; | |
2040 | ||
2041 | if (!cfqd->busy_queues) | |
2042 | return 0; | |
2043 | ||
2044 | if (unlikely(force)) | |
2045 | return cfq_forced_dispatch(cfqd); | |
2046 | ||
2047 | cfqq = cfq_select_queue(cfqd); | |
2048 | if (!cfqq) | |
8e296755 JA |
2049 | return 0; |
2050 | ||
2f5cb738 | 2051 | /* |
0b182d61 | 2052 | * Dispatch a request from this cfqq, if it is allowed |
2f5cb738 | 2053 | */ |
0b182d61 JA |
2054 | if (!cfq_dispatch_request(cfqd, cfqq)) |
2055 | return 0; | |
2056 | ||
2f5cb738 | 2057 | cfqq->slice_dispatch++; |
b029195d | 2058 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 2059 | |
2f5cb738 JA |
2060 | /* |
2061 | * expire an async queue immediately if it has used up its slice. idle | |
2062 | * queue always expire after 1 dispatch round. | |
2063 | */ | |
2064 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
2065 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
2066 | cfq_class_idle(cfqq))) { | |
2067 | cfqq->slice_end = jiffies + 1; | |
2068 | cfq_slice_expired(cfqd, 0); | |
1da177e4 LT |
2069 | } |
2070 | ||
b217a903 | 2071 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); |
2f5cb738 | 2072 | return 1; |
1da177e4 LT |
2073 | } |
2074 | ||
1da177e4 | 2075 | /* |
5e705374 JA |
2076 | * task holds one reference to the queue, dropped when task exits. each rq |
2077 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 LT |
2078 | * |
2079 | * queue lock must be held here. | |
2080 | */ | |
2081 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
2082 | { | |
22e2c507 JA |
2083 | struct cfq_data *cfqd = cfqq->cfqd; |
2084 | ||
2085 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
2086 | |
2087 | if (!atomic_dec_and_test(&cfqq->ref)) | |
2088 | return; | |
2089 | ||
7b679138 | 2090 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 2091 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 2092 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
1da177e4 | 2093 | |
28f95cbc | 2094 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 2095 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 2096 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 2097 | } |
22e2c507 | 2098 | |
f04a6424 | 2099 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 LT |
2100 | kmem_cache_free(cfq_pool, cfqq); |
2101 | } | |
2102 | ||
d6de8be7 JA |
2103 | /* |
2104 | * Must always be called with the rcu_read_lock() held | |
2105 | */ | |
07416d29 JA |
2106 | static void |
2107 | __call_for_each_cic(struct io_context *ioc, | |
2108 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
2109 | { | |
2110 | struct cfq_io_context *cic; | |
2111 | struct hlist_node *n; | |
2112 | ||
2113 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) | |
2114 | func(ioc, cic); | |
2115 | } | |
2116 | ||
4ac845a2 | 2117 | /* |
34e6bbf2 | 2118 | * Call func for each cic attached to this ioc. |
4ac845a2 | 2119 | */ |
34e6bbf2 | 2120 | static void |
4ac845a2 JA |
2121 | call_for_each_cic(struct io_context *ioc, |
2122 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1da177e4 | 2123 | { |
4ac845a2 | 2124 | rcu_read_lock(); |
07416d29 | 2125 | __call_for_each_cic(ioc, func); |
4ac845a2 | 2126 | rcu_read_unlock(); |
34e6bbf2 FC |
2127 | } |
2128 | ||
2129 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
2130 | { | |
2131 | struct cfq_io_context *cic; | |
2132 | ||
2133 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
2134 | ||
2135 | kmem_cache_free(cfq_ioc_pool, cic); | |
245b2e70 | 2136 | elv_ioc_count_dec(cfq_ioc_count); |
34e6bbf2 | 2137 | |
9a11b4ed JA |
2138 | if (ioc_gone) { |
2139 | /* | |
2140 | * CFQ scheduler is exiting, grab exit lock and check | |
2141 | * the pending io context count. If it hits zero, | |
2142 | * complete ioc_gone and set it back to NULL | |
2143 | */ | |
2144 | spin_lock(&ioc_gone_lock); | |
245b2e70 | 2145 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { |
9a11b4ed JA |
2146 | complete(ioc_gone); |
2147 | ioc_gone = NULL; | |
2148 | } | |
2149 | spin_unlock(&ioc_gone_lock); | |
2150 | } | |
34e6bbf2 | 2151 | } |
4ac845a2 | 2152 | |
34e6bbf2 FC |
2153 | static void cfq_cic_free(struct cfq_io_context *cic) |
2154 | { | |
2155 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
2156 | } |
2157 | ||
2158 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
2159 | { | |
2160 | unsigned long flags; | |
2161 | ||
2162 | BUG_ON(!cic->dead_key); | |
2163 | ||
2164 | spin_lock_irqsave(&ioc->lock, flags); | |
2165 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
ffc4e759 | 2166 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2167 | spin_unlock_irqrestore(&ioc->lock, flags); |
2168 | ||
34e6bbf2 | 2169 | cfq_cic_free(cic); |
4ac845a2 JA |
2170 | } |
2171 | ||
d6de8be7 JA |
2172 | /* |
2173 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
2174 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
2175 | * and ->trim() which is called with the task lock held | |
2176 | */ | |
4ac845a2 JA |
2177 | static void cfq_free_io_context(struct io_context *ioc) |
2178 | { | |
4ac845a2 | 2179 | /* |
34e6bbf2 FC |
2180 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
2181 | * so no more cic's are allowed to be linked into this ioc. So it | |
2182 | * should be ok to iterate over the known list, we will see all cic's | |
2183 | * since no new ones are added. | |
4ac845a2 | 2184 | */ |
07416d29 | 2185 | __call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
2186 | } |
2187 | ||
89850f7e | 2188 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 2189 | { |
df5fe3e8 JM |
2190 | struct cfq_queue *__cfqq, *next; |
2191 | ||
28f95cbc | 2192 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 2193 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 2194 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 2195 | } |
22e2c507 | 2196 | |
df5fe3e8 JM |
2197 | /* |
2198 | * If this queue was scheduled to merge with another queue, be | |
2199 | * sure to drop the reference taken on that queue (and others in | |
2200 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
2201 | */ | |
2202 | __cfqq = cfqq->new_cfqq; | |
2203 | while (__cfqq) { | |
2204 | if (__cfqq == cfqq) { | |
2205 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
2206 | break; | |
2207 | } | |
2208 | next = __cfqq->new_cfqq; | |
2209 | cfq_put_queue(__cfqq); | |
2210 | __cfqq = next; | |
2211 | } | |
2212 | ||
89850f7e JA |
2213 | cfq_put_queue(cfqq); |
2214 | } | |
22e2c507 | 2215 | |
89850f7e JA |
2216 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
2217 | struct cfq_io_context *cic) | |
2218 | { | |
4faa3c81 FC |
2219 | struct io_context *ioc = cic->ioc; |
2220 | ||
fc46379d | 2221 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
2222 | |
2223 | /* | |
2224 | * Make sure key == NULL is seen for dead queues | |
2225 | */ | |
fc46379d | 2226 | smp_wmb(); |
4ac845a2 | 2227 | cic->dead_key = (unsigned long) cic->key; |
fc46379d JA |
2228 | cic->key = NULL; |
2229 | ||
4faa3c81 FC |
2230 | if (ioc->ioc_data == cic) |
2231 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
2232 | ||
ff6657c6 JA |
2233 | if (cic->cfqq[BLK_RW_ASYNC]) { |
2234 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
2235 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
2236 | } |
2237 | ||
ff6657c6 JA |
2238 | if (cic->cfqq[BLK_RW_SYNC]) { |
2239 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
2240 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 2241 | } |
89850f7e JA |
2242 | } |
2243 | ||
4ac845a2 JA |
2244 | static void cfq_exit_single_io_context(struct io_context *ioc, |
2245 | struct cfq_io_context *cic) | |
89850f7e JA |
2246 | { |
2247 | struct cfq_data *cfqd = cic->key; | |
2248 | ||
89850f7e | 2249 | if (cfqd) { |
165125e1 | 2250 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 2251 | unsigned long flags; |
89850f7e | 2252 | |
4ac845a2 | 2253 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
2254 | |
2255 | /* | |
2256 | * Ensure we get a fresh copy of the ->key to prevent | |
2257 | * race between exiting task and queue | |
2258 | */ | |
2259 | smp_read_barrier_depends(); | |
2260 | if (cic->key) | |
2261 | __cfq_exit_single_io_context(cfqd, cic); | |
2262 | ||
4ac845a2 | 2263 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 2264 | } |
1da177e4 LT |
2265 | } |
2266 | ||
498d3aa2 JA |
2267 | /* |
2268 | * The process that ioc belongs to has exited, we need to clean up | |
2269 | * and put the internal structures we have that belongs to that process. | |
2270 | */ | |
e2d74ac0 | 2271 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 2272 | { |
4ac845a2 | 2273 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
2274 | } |
2275 | ||
22e2c507 | 2276 | static struct cfq_io_context * |
8267e268 | 2277 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2278 | { |
b5deef90 | 2279 | struct cfq_io_context *cic; |
1da177e4 | 2280 | |
94f6030c CL |
2281 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
2282 | cfqd->queue->node); | |
1da177e4 | 2283 | if (cic) { |
22e2c507 | 2284 | cic->last_end_request = jiffies; |
553698f9 | 2285 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 2286 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
2287 | cic->dtor = cfq_free_io_context; |
2288 | cic->exit = cfq_exit_io_context; | |
245b2e70 | 2289 | elv_ioc_count_inc(cfq_ioc_count); |
1da177e4 LT |
2290 | } |
2291 | ||
2292 | return cic; | |
2293 | } | |
2294 | ||
fd0928df | 2295 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
2296 | { |
2297 | struct task_struct *tsk = current; | |
2298 | int ioprio_class; | |
2299 | ||
3b18152c | 2300 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
2301 | return; |
2302 | ||
fd0928df | 2303 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 2304 | switch (ioprio_class) { |
fe094d98 JA |
2305 | default: |
2306 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
2307 | case IOPRIO_CLASS_NONE: | |
2308 | /* | |
6d63c275 | 2309 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
2310 | */ |
2311 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 2312 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
2313 | break; |
2314 | case IOPRIO_CLASS_RT: | |
2315 | cfqq->ioprio = task_ioprio(ioc); | |
2316 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
2317 | break; | |
2318 | case IOPRIO_CLASS_BE: | |
2319 | cfqq->ioprio = task_ioprio(ioc); | |
2320 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2321 | break; | |
2322 | case IOPRIO_CLASS_IDLE: | |
2323 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
2324 | cfqq->ioprio = 7; | |
2325 | cfq_clear_cfqq_idle_window(cfqq); | |
2326 | break; | |
22e2c507 JA |
2327 | } |
2328 | ||
2329 | /* | |
2330 | * keep track of original prio settings in case we have to temporarily | |
2331 | * elevate the priority of this queue | |
2332 | */ | |
2333 | cfqq->org_ioprio = cfqq->ioprio; | |
2334 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 2335 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
2336 | } |
2337 | ||
febffd61 | 2338 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 2339 | { |
478a82b0 AV |
2340 | struct cfq_data *cfqd = cic->key; |
2341 | struct cfq_queue *cfqq; | |
c1b707d2 | 2342 | unsigned long flags; |
35e6077c | 2343 | |
caaa5f9f JA |
2344 | if (unlikely(!cfqd)) |
2345 | return; | |
2346 | ||
c1b707d2 | 2347 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 2348 | |
ff6657c6 | 2349 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
2350 | if (cfqq) { |
2351 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
2352 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
2353 | GFP_ATOMIC); | |
caaa5f9f | 2354 | if (new_cfqq) { |
ff6657c6 | 2355 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
2356 | cfq_put_queue(cfqq); |
2357 | } | |
22e2c507 | 2358 | } |
caaa5f9f | 2359 | |
ff6657c6 | 2360 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
2361 | if (cfqq) |
2362 | cfq_mark_cfqq_prio_changed(cfqq); | |
2363 | ||
c1b707d2 | 2364 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
2365 | } |
2366 | ||
fc46379d | 2367 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 2368 | { |
4ac845a2 | 2369 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 2370 | ioc->ioprio_changed = 0; |
22e2c507 JA |
2371 | } |
2372 | ||
d5036d77 | 2373 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 2374 | pid_t pid, bool is_sync) |
d5036d77 JA |
2375 | { |
2376 | RB_CLEAR_NODE(&cfqq->rb_node); | |
2377 | RB_CLEAR_NODE(&cfqq->p_node); | |
2378 | INIT_LIST_HEAD(&cfqq->fifo); | |
2379 | ||
2380 | atomic_set(&cfqq->ref, 0); | |
2381 | cfqq->cfqd = cfqd; | |
2382 | ||
2383 | cfq_mark_cfqq_prio_changed(cfqq); | |
2384 | ||
2385 | if (is_sync) { | |
2386 | if (!cfq_class_idle(cfqq)) | |
2387 | cfq_mark_cfqq_idle_window(cfqq); | |
2388 | cfq_mark_cfqq_sync(cfqq); | |
2389 | } | |
2390 | cfqq->pid = pid; | |
2391 | } | |
2392 | ||
cdb16e8f VG |
2393 | static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) |
2394 | { | |
2395 | cfqq->cfqg = cfqg; | |
2396 | } | |
2397 | ||
2398 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create) | |
2399 | { | |
2400 | return &cfqd->root_group; | |
2401 | } | |
2402 | ||
22e2c507 | 2403 | static struct cfq_queue * |
a6151c3a | 2404 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, |
fd0928df | 2405 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 2406 | { |
22e2c507 | 2407 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 2408 | struct cfq_io_context *cic; |
cdb16e8f | 2409 | struct cfq_group *cfqg; |
22e2c507 JA |
2410 | |
2411 | retry: | |
cdb16e8f | 2412 | cfqg = cfq_get_cfqg(cfqd, 1); |
4ac845a2 | 2413 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
2414 | /* cic always exists here */ |
2415 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 2416 | |
6118b70b JA |
2417 | /* |
2418 | * Always try a new alloc if we fell back to the OOM cfqq | |
2419 | * originally, since it should just be a temporary situation. | |
2420 | */ | |
2421 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
2422 | cfqq = NULL; | |
22e2c507 JA |
2423 | if (new_cfqq) { |
2424 | cfqq = new_cfqq; | |
2425 | new_cfqq = NULL; | |
2426 | } else if (gfp_mask & __GFP_WAIT) { | |
2427 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 2428 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 2429 | gfp_mask | __GFP_ZERO, |
94f6030c | 2430 | cfqd->queue->node); |
22e2c507 | 2431 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
2432 | if (new_cfqq) |
2433 | goto retry; | |
22e2c507 | 2434 | } else { |
94f6030c CL |
2435 | cfqq = kmem_cache_alloc_node(cfq_pool, |
2436 | gfp_mask | __GFP_ZERO, | |
2437 | cfqd->queue->node); | |
22e2c507 JA |
2438 | } |
2439 | ||
6118b70b JA |
2440 | if (cfqq) { |
2441 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
2442 | cfq_init_prio_data(cfqq, ioc); | |
cdb16e8f | 2443 | cfq_link_cfqq_cfqg(cfqq, cfqg); |
6118b70b JA |
2444 | cfq_log_cfqq(cfqd, cfqq, "alloced"); |
2445 | } else | |
2446 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
2447 | } |
2448 | ||
2449 | if (new_cfqq) | |
2450 | kmem_cache_free(cfq_pool, new_cfqq); | |
2451 | ||
22e2c507 JA |
2452 | return cfqq; |
2453 | } | |
2454 | ||
c2dea2d1 VT |
2455 | static struct cfq_queue ** |
2456 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
2457 | { | |
fe094d98 | 2458 | switch (ioprio_class) { |
c2dea2d1 VT |
2459 | case IOPRIO_CLASS_RT: |
2460 | return &cfqd->async_cfqq[0][ioprio]; | |
2461 | case IOPRIO_CLASS_BE: | |
2462 | return &cfqd->async_cfqq[1][ioprio]; | |
2463 | case IOPRIO_CLASS_IDLE: | |
2464 | return &cfqd->async_idle_cfqq; | |
2465 | default: | |
2466 | BUG(); | |
2467 | } | |
2468 | } | |
2469 | ||
15c31be4 | 2470 | static struct cfq_queue * |
a6151c3a | 2471 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, |
15c31be4 JA |
2472 | gfp_t gfp_mask) |
2473 | { | |
fd0928df JA |
2474 | const int ioprio = task_ioprio(ioc); |
2475 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 2476 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
2477 | struct cfq_queue *cfqq = NULL; |
2478 | ||
c2dea2d1 VT |
2479 | if (!is_sync) { |
2480 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
2481 | cfqq = *async_cfqq; | |
2482 | } | |
2483 | ||
6118b70b | 2484 | if (!cfqq) |
fd0928df | 2485 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
2486 | |
2487 | /* | |
2488 | * pin the queue now that it's allocated, scheduler exit will prune it | |
2489 | */ | |
c2dea2d1 | 2490 | if (!is_sync && !(*async_cfqq)) { |
15c31be4 | 2491 | atomic_inc(&cfqq->ref); |
c2dea2d1 | 2492 | *async_cfqq = cfqq; |
15c31be4 JA |
2493 | } |
2494 | ||
2495 | atomic_inc(&cfqq->ref); | |
2496 | return cfqq; | |
2497 | } | |
2498 | ||
498d3aa2 JA |
2499 | /* |
2500 | * We drop cfq io contexts lazily, so we may find a dead one. | |
2501 | */ | |
dbecf3ab | 2502 | static void |
4ac845a2 JA |
2503 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
2504 | struct cfq_io_context *cic) | |
dbecf3ab | 2505 | { |
4ac845a2 JA |
2506 | unsigned long flags; |
2507 | ||
fc46379d | 2508 | WARN_ON(!list_empty(&cic->queue_list)); |
597bc485 | 2509 | |
4ac845a2 JA |
2510 | spin_lock_irqsave(&ioc->lock, flags); |
2511 | ||
4faa3c81 | 2512 | BUG_ON(ioc->ioc_data == cic); |
597bc485 | 2513 | |
4ac845a2 | 2514 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); |
ffc4e759 | 2515 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2516 | spin_unlock_irqrestore(&ioc->lock, flags); |
2517 | ||
2518 | cfq_cic_free(cic); | |
dbecf3ab OH |
2519 | } |
2520 | ||
e2d74ac0 | 2521 | static struct cfq_io_context * |
4ac845a2 | 2522 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 2523 | { |
e2d74ac0 | 2524 | struct cfq_io_context *cic; |
d6de8be7 | 2525 | unsigned long flags; |
4ac845a2 | 2526 | void *k; |
e2d74ac0 | 2527 | |
91fac317 VT |
2528 | if (unlikely(!ioc)) |
2529 | return NULL; | |
2530 | ||
d6de8be7 JA |
2531 | rcu_read_lock(); |
2532 | ||
597bc485 JA |
2533 | /* |
2534 | * we maintain a last-hit cache, to avoid browsing over the tree | |
2535 | */ | |
4ac845a2 | 2536 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
2537 | if (cic && cic->key == cfqd) { |
2538 | rcu_read_unlock(); | |
597bc485 | 2539 | return cic; |
d6de8be7 | 2540 | } |
597bc485 | 2541 | |
4ac845a2 | 2542 | do { |
4ac845a2 JA |
2543 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); |
2544 | rcu_read_unlock(); | |
2545 | if (!cic) | |
2546 | break; | |
be3b0753 OH |
2547 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
2548 | k = cic->key; | |
2549 | if (unlikely(!k)) { | |
4ac845a2 | 2550 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 2551 | rcu_read_lock(); |
4ac845a2 | 2552 | continue; |
dbecf3ab | 2553 | } |
e2d74ac0 | 2554 | |
d6de8be7 | 2555 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 2556 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 2557 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
2558 | break; |
2559 | } while (1); | |
e2d74ac0 | 2560 | |
4ac845a2 | 2561 | return cic; |
e2d74ac0 JA |
2562 | } |
2563 | ||
4ac845a2 JA |
2564 | /* |
2565 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
2566 | * the process specific cfq io context when entered from the block layer. | |
2567 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
2568 | */ | |
febffd61 JA |
2569 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
2570 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 2571 | { |
0261d688 | 2572 | unsigned long flags; |
4ac845a2 | 2573 | int ret; |
e2d74ac0 | 2574 | |
4ac845a2 JA |
2575 | ret = radix_tree_preload(gfp_mask); |
2576 | if (!ret) { | |
2577 | cic->ioc = ioc; | |
2578 | cic->key = cfqd; | |
e2d74ac0 | 2579 | |
4ac845a2 JA |
2580 | spin_lock_irqsave(&ioc->lock, flags); |
2581 | ret = radix_tree_insert(&ioc->radix_root, | |
2582 | (unsigned long) cfqd, cic); | |
ffc4e759 JA |
2583 | if (!ret) |
2584 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 2585 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 2586 | |
4ac845a2 JA |
2587 | radix_tree_preload_end(); |
2588 | ||
2589 | if (!ret) { | |
2590 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
2591 | list_add(&cic->queue_list, &cfqd->cic_list); | |
2592 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2593 | } | |
e2d74ac0 JA |
2594 | } |
2595 | ||
4ac845a2 JA |
2596 | if (ret) |
2597 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 2598 | |
4ac845a2 | 2599 | return ret; |
e2d74ac0 JA |
2600 | } |
2601 | ||
1da177e4 LT |
2602 | /* |
2603 | * Setup general io context and cfq io context. There can be several cfq | |
2604 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 2605 | * than one device managed by cfq. |
1da177e4 LT |
2606 | */ |
2607 | static struct cfq_io_context * | |
e2d74ac0 | 2608 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2609 | { |
22e2c507 | 2610 | struct io_context *ioc = NULL; |
1da177e4 | 2611 | struct cfq_io_context *cic; |
1da177e4 | 2612 | |
22e2c507 | 2613 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 2614 | |
b5deef90 | 2615 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
2616 | if (!ioc) |
2617 | return NULL; | |
2618 | ||
4ac845a2 | 2619 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
2620 | if (cic) |
2621 | goto out; | |
1da177e4 | 2622 | |
e2d74ac0 JA |
2623 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
2624 | if (cic == NULL) | |
2625 | goto err; | |
1da177e4 | 2626 | |
4ac845a2 JA |
2627 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
2628 | goto err_free; | |
2629 | ||
1da177e4 | 2630 | out: |
fc46379d JA |
2631 | smp_read_barrier_depends(); |
2632 | if (unlikely(ioc->ioprio_changed)) | |
2633 | cfq_ioc_set_ioprio(ioc); | |
2634 | ||
1da177e4 | 2635 | return cic; |
4ac845a2 JA |
2636 | err_free: |
2637 | cfq_cic_free(cic); | |
1da177e4 LT |
2638 | err: |
2639 | put_io_context(ioc); | |
2640 | return NULL; | |
2641 | } | |
2642 | ||
22e2c507 JA |
2643 | static void |
2644 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 2645 | { |
aaf1228d JA |
2646 | unsigned long elapsed = jiffies - cic->last_end_request; |
2647 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 2648 | |
22e2c507 JA |
2649 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
2650 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
2651 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
2652 | } | |
1da177e4 | 2653 | |
206dc69b | 2654 | static void |
b2c18e1e | 2655 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
6d048f53 | 2656 | struct request *rq) |
206dc69b JA |
2657 | { |
2658 | sector_t sdist; | |
2659 | u64 total; | |
2660 | ||
b2c18e1e | 2661 | if (!cfqq->last_request_pos) |
4d00aa47 | 2662 | sdist = 0; |
b2c18e1e JM |
2663 | else if (cfqq->last_request_pos < blk_rq_pos(rq)) |
2664 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
206dc69b | 2665 | else |
b2c18e1e | 2666 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); |
206dc69b JA |
2667 | |
2668 | /* | |
2669 | * Don't allow the seek distance to get too large from the | |
2670 | * odd fragment, pagein, etc | |
2671 | */ | |
b2c18e1e JM |
2672 | if (cfqq->seek_samples <= 60) /* second&third seek */ |
2673 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024); | |
206dc69b | 2674 | else |
b2c18e1e | 2675 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64); |
206dc69b | 2676 | |
b2c18e1e JM |
2677 | cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8; |
2678 | cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8; | |
2679 | total = cfqq->seek_total + (cfqq->seek_samples/2); | |
2680 | do_div(total, cfqq->seek_samples); | |
2681 | cfqq->seek_mean = (sector_t)total; | |
e6c5bc73 JM |
2682 | |
2683 | /* | |
2684 | * If this cfqq is shared between multiple processes, check to | |
2685 | * make sure that those processes are still issuing I/Os within | |
2686 | * the mean seek distance. If not, it may be time to break the | |
2687 | * queues apart again. | |
2688 | */ | |
2689 | if (cfq_cfqq_coop(cfqq)) { | |
2690 | if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start) | |
2691 | cfqq->seeky_start = jiffies; | |
2692 | else if (!CFQQ_SEEKY(cfqq)) | |
2693 | cfqq->seeky_start = 0; | |
2694 | } | |
206dc69b | 2695 | } |
1da177e4 | 2696 | |
22e2c507 JA |
2697 | /* |
2698 | * Disable idle window if the process thinks too long or seeks so much that | |
2699 | * it doesn't matter | |
2700 | */ | |
2701 | static void | |
2702 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2703 | struct cfq_io_context *cic) | |
2704 | { | |
7b679138 | 2705 | int old_idle, enable_idle; |
1be92f2f | 2706 | |
0871714e JA |
2707 | /* |
2708 | * Don't idle for async or idle io prio class | |
2709 | */ | |
2710 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
2711 | return; |
2712 | ||
c265a7f4 | 2713 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 2714 | |
76280aff CZ |
2715 | if (cfqq->queued[0] + cfqq->queued[1] >= 4) |
2716 | cfq_mark_cfqq_deep(cfqq); | |
2717 | ||
66dac98e | 2718 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || |
76280aff CZ |
2719 | (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples) |
2720 | && CFQQ_SEEKY(cfqq))) | |
22e2c507 JA |
2721 | enable_idle = 0; |
2722 | else if (sample_valid(cic->ttime_samples)) { | |
718eee05 | 2723 | if (cic->ttime_mean > cfqd->cfq_slice_idle) |
22e2c507 JA |
2724 | enable_idle = 0; |
2725 | else | |
2726 | enable_idle = 1; | |
1da177e4 LT |
2727 | } |
2728 | ||
7b679138 JA |
2729 | if (old_idle != enable_idle) { |
2730 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
2731 | if (enable_idle) | |
2732 | cfq_mark_cfqq_idle_window(cfqq); | |
2733 | else | |
2734 | cfq_clear_cfqq_idle_window(cfqq); | |
2735 | } | |
22e2c507 | 2736 | } |
1da177e4 | 2737 | |
22e2c507 JA |
2738 | /* |
2739 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
2740 | * no or if we aren't sure, a 1 will cause a preempt. | |
2741 | */ | |
a6151c3a | 2742 | static bool |
22e2c507 | 2743 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, |
5e705374 | 2744 | struct request *rq) |
22e2c507 | 2745 | { |
6d048f53 | 2746 | struct cfq_queue *cfqq; |
22e2c507 | 2747 | |
6d048f53 JA |
2748 | cfqq = cfqd->active_queue; |
2749 | if (!cfqq) | |
a6151c3a | 2750 | return false; |
22e2c507 | 2751 | |
6d048f53 | 2752 | if (cfq_slice_used(cfqq)) |
a6151c3a | 2753 | return true; |
6d048f53 JA |
2754 | |
2755 | if (cfq_class_idle(new_cfqq)) | |
a6151c3a | 2756 | return false; |
22e2c507 JA |
2757 | |
2758 | if (cfq_class_idle(cfqq)) | |
a6151c3a | 2759 | return true; |
1e3335de | 2760 | |
f04a6424 | 2761 | /* Allow preemption only if we are idling on sync-noidle tree */ |
e4a22919 CZ |
2762 | if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && |
2763 | cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && | |
f04a6424 VG |
2764 | new_cfqq->service_tree->count == 2 && |
2765 | RB_EMPTY_ROOT(&cfqq->sort_list)) | |
718eee05 CZ |
2766 | return true; |
2767 | ||
374f84ac JA |
2768 | /* |
2769 | * if the new request is sync, but the currently running queue is | |
2770 | * not, let the sync request have priority. | |
2771 | */ | |
5e705374 | 2772 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
a6151c3a | 2773 | return true; |
1e3335de | 2774 | |
374f84ac JA |
2775 | /* |
2776 | * So both queues are sync. Let the new request get disk time if | |
2777 | * it's a metadata request and the current queue is doing regular IO. | |
2778 | */ | |
2779 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
e6ec4fe2 | 2780 | return true; |
22e2c507 | 2781 | |
3a9a3f6c DS |
2782 | /* |
2783 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
2784 | */ | |
2785 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
a6151c3a | 2786 | return true; |
3a9a3f6c | 2787 | |
1e3335de | 2788 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
a6151c3a | 2789 | return false; |
1e3335de JA |
2790 | |
2791 | /* | |
2792 | * if this request is as-good as one we would expect from the | |
2793 | * current cfqq, let it preempt | |
2794 | */ | |
e00ef799 | 2795 | if (cfq_rq_close(cfqd, cfqq, rq)) |
a6151c3a | 2796 | return true; |
1e3335de | 2797 | |
a6151c3a | 2798 | return false; |
22e2c507 JA |
2799 | } |
2800 | ||
2801 | /* | |
2802 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
2803 | * let it have half of its nominal slice. | |
2804 | */ | |
2805 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2806 | { | |
7b679138 | 2807 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
6084cdda | 2808 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 2809 | |
bf572256 JA |
2810 | /* |
2811 | * Put the new queue at the front of the of the current list, | |
2812 | * so we know that it will be selected next. | |
2813 | */ | |
2814 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
2815 | |
2816 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 2817 | |
44f7c160 JA |
2818 | cfqq->slice_end = 0; |
2819 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
2820 | } |
2821 | ||
22e2c507 | 2822 | /* |
5e705374 | 2823 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
2824 | * something we should do about it |
2825 | */ | |
2826 | static void | |
5e705374 JA |
2827 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
2828 | struct request *rq) | |
22e2c507 | 2829 | { |
5e705374 | 2830 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 2831 | |
45333d5a | 2832 | cfqd->rq_queued++; |
374f84ac JA |
2833 | if (rq_is_meta(rq)) |
2834 | cfqq->meta_pending++; | |
2835 | ||
9c2c38a1 | 2836 | cfq_update_io_thinktime(cfqd, cic); |
b2c18e1e | 2837 | cfq_update_io_seektime(cfqd, cfqq, rq); |
9c2c38a1 JA |
2838 | cfq_update_idle_window(cfqd, cfqq, cic); |
2839 | ||
b2c18e1e | 2840 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
2841 | |
2842 | if (cfqq == cfqd->active_queue) { | |
2843 | /* | |
b029195d JA |
2844 | * Remember that we saw a request from this process, but |
2845 | * don't start queuing just yet. Otherwise we risk seeing lots | |
2846 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
2847 | * and merging. If the request is already larger than a single |
2848 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
2849 | * merging is already done. Ditto for a busy system that |
2850 | * has other work pending, don't risk delaying until the | |
2851 | * idle timer unplug to continue working. | |
22e2c507 | 2852 | */ |
d6ceb25e | 2853 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
2854 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
2855 | cfqd->busy_queues > 1) { | |
d6ceb25e | 2856 | del_timer(&cfqd->idle_slice_timer); |
bf791937 VG |
2857 | __blk_run_queue(cfqd->queue); |
2858 | } else | |
2859 | cfq_mark_cfqq_must_dispatch(cfqq); | |
d6ceb25e | 2860 | } |
5e705374 | 2861 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
2862 | /* |
2863 | * not the active queue - expire current slice if it is | |
2864 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
2865 | * has some old slice time left and is of higher priority or |
2866 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
2867 | */ |
2868 | cfq_preempt_queue(cfqd, cfqq); | |
a7f55792 | 2869 | __blk_run_queue(cfqd->queue); |
22e2c507 | 2870 | } |
1da177e4 LT |
2871 | } |
2872 | ||
165125e1 | 2873 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2874 | { |
b4878f24 | 2875 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 2876 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 2877 | |
7b679138 | 2878 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 2879 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 2880 | |
30996f40 | 2881 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); |
22e2c507 | 2882 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
aa6f6a3d | 2883 | cfq_add_rq_rb(rq); |
22e2c507 | 2884 | |
5e705374 | 2885 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
2886 | } |
2887 | ||
45333d5a AC |
2888 | /* |
2889 | * Update hw_tag based on peak queue depth over 50 samples under | |
2890 | * sufficient load. | |
2891 | */ | |
2892 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
2893 | { | |
1a1238a7 SL |
2894 | struct cfq_queue *cfqq = cfqd->active_queue; |
2895 | ||
e459dd08 CZ |
2896 | if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth) |
2897 | cfqd->hw_tag_est_depth = rq_in_driver(cfqd); | |
2898 | ||
2899 | if (cfqd->hw_tag == 1) | |
2900 | return; | |
45333d5a AC |
2901 | |
2902 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
5ad531db | 2903 | rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
2904 | return; |
2905 | ||
1a1238a7 SL |
2906 | /* |
2907 | * If active queue hasn't enough requests and can idle, cfq might not | |
2908 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
2909 | * case | |
2910 | */ | |
2911 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
2912 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
2913 | CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN) | |
2914 | return; | |
2915 | ||
45333d5a AC |
2916 | if (cfqd->hw_tag_samples++ < 50) |
2917 | return; | |
2918 | ||
e459dd08 | 2919 | if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
2920 | cfqd->hw_tag = 1; |
2921 | else | |
2922 | cfqd->hw_tag = 0; | |
45333d5a AC |
2923 | } |
2924 | ||
165125e1 | 2925 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2926 | { |
5e705374 | 2927 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 2928 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 2929 | const int sync = rq_is_sync(rq); |
b4878f24 | 2930 | unsigned long now; |
1da177e4 | 2931 | |
b4878f24 | 2932 | now = jiffies; |
7b679138 | 2933 | cfq_log_cfqq(cfqd, cfqq, "complete"); |
1da177e4 | 2934 | |
45333d5a AC |
2935 | cfq_update_hw_tag(cfqd); |
2936 | ||
5ad531db | 2937 | WARN_ON(!cfqd->rq_in_driver[sync]); |
6d048f53 | 2938 | WARN_ON(!cfqq->dispatched); |
5ad531db | 2939 | cfqd->rq_in_driver[sync]--; |
6d048f53 | 2940 | cfqq->dispatched--; |
1da177e4 | 2941 | |
3ed9a296 JA |
2942 | if (cfq_cfqq_sync(cfqq)) |
2943 | cfqd->sync_flight--; | |
2944 | ||
365722bb | 2945 | if (sync) { |
5e705374 | 2946 | RQ_CIC(rq)->last_end_request = now; |
365722bb VG |
2947 | cfqd->last_end_sync_rq = now; |
2948 | } | |
caaa5f9f JA |
2949 | |
2950 | /* | |
2951 | * If this is the active queue, check if it needs to be expired, | |
2952 | * or if we want to idle in case it has no pending requests. | |
2953 | */ | |
2954 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
2955 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
2956 | ||
44f7c160 JA |
2957 | if (cfq_cfqq_slice_new(cfqq)) { |
2958 | cfq_set_prio_slice(cfqd, cfqq); | |
2959 | cfq_clear_cfqq_slice_new(cfqq); | |
2960 | } | |
a36e71f9 | 2961 | /* |
8e550632 CZ |
2962 | * Idling is not enabled on: |
2963 | * - expired queues | |
2964 | * - idle-priority queues | |
2965 | * - async queues | |
2966 | * - queues with still some requests queued | |
2967 | * - when there is a close cooperator | |
a36e71f9 | 2968 | */ |
0871714e | 2969 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
6084cdda | 2970 | cfq_slice_expired(cfqd, 1); |
8e550632 CZ |
2971 | else if (sync && cfqq_empty && |
2972 | !cfq_close_cooperator(cfqd, cfqq)) { | |
2973 | cfqd->noidle_tree_requires_idle |= !rq_noidle(rq); | |
2974 | /* | |
2975 | * Idling is enabled for SYNC_WORKLOAD. | |
2976 | * SYNC_NOIDLE_WORKLOAD idles at the end of the tree | |
2977 | * only if we processed at least one !rq_noidle request | |
2978 | */ | |
2979 | if (cfqd->serving_type == SYNC_WORKLOAD | |
2980 | || cfqd->noidle_tree_requires_idle) | |
2981 | cfq_arm_slice_timer(cfqd); | |
2982 | } | |
caaa5f9f | 2983 | } |
6d048f53 | 2984 | |
5ad531db | 2985 | if (!rq_in_driver(cfqd)) |
23e018a1 | 2986 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
2987 | } |
2988 | ||
22e2c507 JA |
2989 | /* |
2990 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
2991 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
2992 | */ | |
2993 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 2994 | { |
22e2c507 JA |
2995 | if (has_fs_excl()) { |
2996 | /* | |
2997 | * boost idle prio on transactions that would lock out other | |
2998 | * users of the filesystem | |
2999 | */ | |
3000 | if (cfq_class_idle(cfqq)) | |
3001 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
3002 | if (cfqq->ioprio > IOPRIO_NORM) | |
3003 | cfqq->ioprio = IOPRIO_NORM; | |
3004 | } else { | |
3005 | /* | |
dddb7451 | 3006 | * unboost the queue (if needed) |
22e2c507 | 3007 | */ |
dddb7451 CZ |
3008 | cfqq->ioprio_class = cfqq->org_ioprio_class; |
3009 | cfqq->ioprio = cfqq->org_ioprio; | |
22e2c507 | 3010 | } |
22e2c507 | 3011 | } |
1da177e4 | 3012 | |
89850f7e | 3013 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 3014 | { |
1b379d8d | 3015 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 3016 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 3017 | return ELV_MQUEUE_MUST; |
3b18152c | 3018 | } |
1da177e4 | 3019 | |
22e2c507 | 3020 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
3021 | } |
3022 | ||
165125e1 | 3023 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
3024 | { |
3025 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3026 | struct task_struct *tsk = current; | |
91fac317 | 3027 | struct cfq_io_context *cic; |
22e2c507 JA |
3028 | struct cfq_queue *cfqq; |
3029 | ||
3030 | /* | |
3031 | * don't force setup of a queue from here, as a call to may_queue | |
3032 | * does not necessarily imply that a request actually will be queued. | |
3033 | * so just lookup a possibly existing queue, or return 'may queue' | |
3034 | * if that fails | |
3035 | */ | |
4ac845a2 | 3036 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
3037 | if (!cic) |
3038 | return ELV_MQUEUE_MAY; | |
3039 | ||
b0b78f81 | 3040 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 3041 | if (cfqq) { |
fd0928df | 3042 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
3043 | cfq_prio_boost(cfqq); |
3044 | ||
89850f7e | 3045 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
3046 | } |
3047 | ||
3048 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
3049 | } |
3050 | ||
1da177e4 LT |
3051 | /* |
3052 | * queue lock held here | |
3053 | */ | |
bb37b94c | 3054 | static void cfq_put_request(struct request *rq) |
1da177e4 | 3055 | { |
5e705374 | 3056 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 3057 | |
5e705374 | 3058 | if (cfqq) { |
22e2c507 | 3059 | const int rw = rq_data_dir(rq); |
1da177e4 | 3060 | |
22e2c507 JA |
3061 | BUG_ON(!cfqq->allocated[rw]); |
3062 | cfqq->allocated[rw]--; | |
1da177e4 | 3063 | |
5e705374 | 3064 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 3065 | |
1da177e4 | 3066 | rq->elevator_private = NULL; |
5e705374 | 3067 | rq->elevator_private2 = NULL; |
1da177e4 | 3068 | |
1da177e4 LT |
3069 | cfq_put_queue(cfqq); |
3070 | } | |
3071 | } | |
3072 | ||
df5fe3e8 JM |
3073 | static struct cfq_queue * |
3074 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, | |
3075 | struct cfq_queue *cfqq) | |
3076 | { | |
3077 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
3078 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
b3b6d040 | 3079 | cfq_mark_cfqq_coop(cfqq->new_cfqq); |
df5fe3e8 JM |
3080 | cfq_put_queue(cfqq); |
3081 | return cic_to_cfqq(cic, 1); | |
3082 | } | |
3083 | ||
e6c5bc73 JM |
3084 | static int should_split_cfqq(struct cfq_queue *cfqq) |
3085 | { | |
3086 | if (cfqq->seeky_start && | |
3087 | time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT)) | |
3088 | return 1; | |
3089 | return 0; | |
3090 | } | |
3091 | ||
3092 | /* | |
3093 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
3094 | * was the last process referring to said cfqq. | |
3095 | */ | |
3096 | static struct cfq_queue * | |
3097 | split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) | |
3098 | { | |
3099 | if (cfqq_process_refs(cfqq) == 1) { | |
3100 | cfqq->seeky_start = 0; | |
3101 | cfqq->pid = current->pid; | |
3102 | cfq_clear_cfqq_coop(cfqq); | |
3103 | return cfqq; | |
3104 | } | |
3105 | ||
3106 | cic_set_cfqq(cic, NULL, 1); | |
3107 | cfq_put_queue(cfqq); | |
3108 | return NULL; | |
3109 | } | |
1da177e4 | 3110 | /* |
22e2c507 | 3111 | * Allocate cfq data structures associated with this request. |
1da177e4 | 3112 | */ |
22e2c507 | 3113 | static int |
165125e1 | 3114 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
3115 | { |
3116 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3117 | struct cfq_io_context *cic; | |
3118 | const int rw = rq_data_dir(rq); | |
a6151c3a | 3119 | const bool is_sync = rq_is_sync(rq); |
22e2c507 | 3120 | struct cfq_queue *cfqq; |
1da177e4 LT |
3121 | unsigned long flags; |
3122 | ||
3123 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
3124 | ||
e2d74ac0 | 3125 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 3126 | |
1da177e4 LT |
3127 | spin_lock_irqsave(q->queue_lock, flags); |
3128 | ||
22e2c507 JA |
3129 | if (!cic) |
3130 | goto queue_fail; | |
3131 | ||
e6c5bc73 | 3132 | new_queue: |
91fac317 | 3133 | cfqq = cic_to_cfqq(cic, is_sync); |
32f2e807 | 3134 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { |
fd0928df | 3135 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 | 3136 | cic_set_cfqq(cic, cfqq, is_sync); |
df5fe3e8 | 3137 | } else { |
e6c5bc73 JM |
3138 | /* |
3139 | * If the queue was seeky for too long, break it apart. | |
3140 | */ | |
3141 | if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) { | |
3142 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); | |
3143 | cfqq = split_cfqq(cic, cfqq); | |
3144 | if (!cfqq) | |
3145 | goto new_queue; | |
3146 | } | |
3147 | ||
df5fe3e8 JM |
3148 | /* |
3149 | * Check to see if this queue is scheduled to merge with | |
3150 | * another, closely cooperating queue. The merging of | |
3151 | * queues happens here as it must be done in process context. | |
3152 | * The reference on new_cfqq was taken in merge_cfqqs. | |
3153 | */ | |
3154 | if (cfqq->new_cfqq) | |
3155 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
91fac317 | 3156 | } |
1da177e4 LT |
3157 | |
3158 | cfqq->allocated[rw]++; | |
22e2c507 | 3159 | atomic_inc(&cfqq->ref); |
1da177e4 | 3160 | |
5e705374 | 3161 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 3162 | |
5e705374 JA |
3163 | rq->elevator_private = cic; |
3164 | rq->elevator_private2 = cfqq; | |
3165 | return 0; | |
1da177e4 | 3166 | |
22e2c507 JA |
3167 | queue_fail: |
3168 | if (cic) | |
3169 | put_io_context(cic->ioc); | |
89850f7e | 3170 | |
23e018a1 | 3171 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 3172 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 3173 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
3174 | return 1; |
3175 | } | |
3176 | ||
65f27f38 | 3177 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 3178 | { |
65f27f38 | 3179 | struct cfq_data *cfqd = |
23e018a1 | 3180 | container_of(work, struct cfq_data, unplug_work); |
165125e1 | 3181 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3182 | |
40bb54d1 | 3183 | spin_lock_irq(q->queue_lock); |
a7f55792 | 3184 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 3185 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
3186 | } |
3187 | ||
3188 | /* | |
3189 | * Timer running if the active_queue is currently idling inside its time slice | |
3190 | */ | |
3191 | static void cfq_idle_slice_timer(unsigned long data) | |
3192 | { | |
3193 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
3194 | struct cfq_queue *cfqq; | |
3195 | unsigned long flags; | |
3c6bd2f8 | 3196 | int timed_out = 1; |
22e2c507 | 3197 | |
7b679138 JA |
3198 | cfq_log(cfqd, "idle timer fired"); |
3199 | ||
22e2c507 JA |
3200 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
3201 | ||
fe094d98 JA |
3202 | cfqq = cfqd->active_queue; |
3203 | if (cfqq) { | |
3c6bd2f8 JA |
3204 | timed_out = 0; |
3205 | ||
b029195d JA |
3206 | /* |
3207 | * We saw a request before the queue expired, let it through | |
3208 | */ | |
3209 | if (cfq_cfqq_must_dispatch(cfqq)) | |
3210 | goto out_kick; | |
3211 | ||
22e2c507 JA |
3212 | /* |
3213 | * expired | |
3214 | */ | |
44f7c160 | 3215 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
3216 | goto expire; |
3217 | ||
3218 | /* | |
3219 | * only expire and reinvoke request handler, if there are | |
3220 | * other queues with pending requests | |
3221 | */ | |
caaa5f9f | 3222 | if (!cfqd->busy_queues) |
22e2c507 | 3223 | goto out_cont; |
22e2c507 JA |
3224 | |
3225 | /* | |
3226 | * not expired and it has a request pending, let it dispatch | |
3227 | */ | |
75e50984 | 3228 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 3229 | goto out_kick; |
76280aff CZ |
3230 | |
3231 | /* | |
3232 | * Queue depth flag is reset only when the idle didn't succeed | |
3233 | */ | |
3234 | cfq_clear_cfqq_deep(cfqq); | |
22e2c507 JA |
3235 | } |
3236 | expire: | |
6084cdda | 3237 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 3238 | out_kick: |
23e018a1 | 3239 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
3240 | out_cont: |
3241 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
3242 | } | |
3243 | ||
3b18152c JA |
3244 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
3245 | { | |
3246 | del_timer_sync(&cfqd->idle_slice_timer); | |
23e018a1 | 3247 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 3248 | } |
22e2c507 | 3249 | |
c2dea2d1 VT |
3250 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
3251 | { | |
3252 | int i; | |
3253 | ||
3254 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
3255 | if (cfqd->async_cfqq[0][i]) | |
3256 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
3257 | if (cfqd->async_cfqq[1][i]) | |
3258 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 3259 | } |
2389d1ef ON |
3260 | |
3261 | if (cfqd->async_idle_cfqq) | |
3262 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
3263 | } |
3264 | ||
b374d18a | 3265 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 3266 | { |
22e2c507 | 3267 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 3268 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3269 | |
3b18152c | 3270 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 3271 | |
d9ff4187 | 3272 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 3273 | |
d9ff4187 | 3274 | if (cfqd->active_queue) |
6084cdda | 3275 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
3276 | |
3277 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
3278 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
3279 | struct cfq_io_context, | |
3280 | queue_list); | |
89850f7e JA |
3281 | |
3282 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 3283 | } |
e2d74ac0 | 3284 | |
c2dea2d1 | 3285 | cfq_put_async_queues(cfqd); |
15c31be4 | 3286 | |
d9ff4187 | 3287 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
3288 | |
3289 | cfq_shutdown_timer_wq(cfqd); | |
3290 | ||
a90d742e | 3291 | kfree(cfqd); |
1da177e4 LT |
3292 | } |
3293 | ||
165125e1 | 3294 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
3295 | { |
3296 | struct cfq_data *cfqd; | |
718eee05 | 3297 | int i, j; |
cdb16e8f | 3298 | struct cfq_group *cfqg; |
615f0259 | 3299 | struct cfq_rb_root *st; |
1da177e4 | 3300 | |
94f6030c | 3301 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 3302 | if (!cfqd) |
bc1c1169 | 3303 | return NULL; |
1da177e4 | 3304 | |
1fa8f6d6 VG |
3305 | /* Init root service tree */ |
3306 | cfqd->grp_service_tree = CFQ_RB_ROOT; | |
3307 | ||
cdb16e8f VG |
3308 | /* Init root group */ |
3309 | cfqg = &cfqd->root_group; | |
615f0259 VG |
3310 | for_each_cfqg_st(cfqg, i, j, st) |
3311 | *st = CFQ_RB_ROOT; | |
1fa8f6d6 | 3312 | RB_CLEAR_NODE(&cfqg->rb_node); |
26a2ac00 | 3313 | |
25bc6b07 VG |
3314 | /* Give preference to root group over other groups */ |
3315 | cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT; | |
3316 | ||
26a2ac00 JA |
3317 | /* |
3318 | * Not strictly needed (since RB_ROOT just clears the node and we | |
3319 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
3320 | * to add magic to the rb code | |
3321 | */ | |
3322 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
3323 | cfqd->prio_trees[i] = RB_ROOT; | |
3324 | ||
6118b70b JA |
3325 | /* |
3326 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
3327 | * Grab a permanent reference to it, so that the normal code flow | |
3328 | * will not attempt to free it. | |
3329 | */ | |
3330 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
3331 | atomic_inc(&cfqd->oom_cfqq.ref); | |
cdb16e8f | 3332 | cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group); |
6118b70b | 3333 | |
d9ff4187 | 3334 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 3335 | |
1da177e4 | 3336 | cfqd->queue = q; |
1da177e4 | 3337 | |
22e2c507 JA |
3338 | init_timer(&cfqd->idle_slice_timer); |
3339 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
3340 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
3341 | ||
23e018a1 | 3342 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 3343 | |
1da177e4 | 3344 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
3345 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
3346 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
3347 | cfqd->cfq_back_max = cfq_back_max; |
3348 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
3349 | cfqd->cfq_slice[0] = cfq_slice_async; |
3350 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
3351 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
3352 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
963b72fc | 3353 | cfqd->cfq_latency = 1; |
e459dd08 | 3354 | cfqd->hw_tag = -1; |
365722bb | 3355 | cfqd->last_end_sync_rq = jiffies; |
bc1c1169 | 3356 | return cfqd; |
1da177e4 LT |
3357 | } |
3358 | ||
3359 | static void cfq_slab_kill(void) | |
3360 | { | |
d6de8be7 JA |
3361 | /* |
3362 | * Caller already ensured that pending RCU callbacks are completed, | |
3363 | * so we should have no busy allocations at this point. | |
3364 | */ | |
1da177e4 LT |
3365 | if (cfq_pool) |
3366 | kmem_cache_destroy(cfq_pool); | |
3367 | if (cfq_ioc_pool) | |
3368 | kmem_cache_destroy(cfq_ioc_pool); | |
3369 | } | |
3370 | ||
3371 | static int __init cfq_slab_setup(void) | |
3372 | { | |
0a31bd5f | 3373 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
3374 | if (!cfq_pool) |
3375 | goto fail; | |
3376 | ||
34e6bbf2 | 3377 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
3378 | if (!cfq_ioc_pool) |
3379 | goto fail; | |
3380 | ||
3381 | return 0; | |
3382 | fail: | |
3383 | cfq_slab_kill(); | |
3384 | return -ENOMEM; | |
3385 | } | |
3386 | ||
1da177e4 LT |
3387 | /* |
3388 | * sysfs parts below --> | |
3389 | */ | |
1da177e4 LT |
3390 | static ssize_t |
3391 | cfq_var_show(unsigned int var, char *page) | |
3392 | { | |
3393 | return sprintf(page, "%d\n", var); | |
3394 | } | |
3395 | ||
3396 | static ssize_t | |
3397 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
3398 | { | |
3399 | char *p = (char *) page; | |
3400 | ||
3401 | *var = simple_strtoul(p, &p, 10); | |
3402 | return count; | |
3403 | } | |
3404 | ||
1da177e4 | 3405 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 3406 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 3407 | { \ |
3d1ab40f | 3408 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3409 | unsigned int __data = __VAR; \ |
3410 | if (__CONV) \ | |
3411 | __data = jiffies_to_msecs(__data); \ | |
3412 | return cfq_var_show(__data, (page)); \ | |
3413 | } | |
3414 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
3415 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
3416 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
3417 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
3418 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
3419 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
3420 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
3421 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
3422 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
963b72fc | 3423 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); |
1da177e4 LT |
3424 | #undef SHOW_FUNCTION |
3425 | ||
3426 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 3427 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 3428 | { \ |
3d1ab40f | 3429 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3430 | unsigned int __data; \ |
3431 | int ret = cfq_var_store(&__data, (page), count); \ | |
3432 | if (__data < (MIN)) \ | |
3433 | __data = (MIN); \ | |
3434 | else if (__data > (MAX)) \ | |
3435 | __data = (MAX); \ | |
3436 | if (__CONV) \ | |
3437 | *(__PTR) = msecs_to_jiffies(__data); \ | |
3438 | else \ | |
3439 | *(__PTR) = __data; \ | |
3440 | return ret; \ | |
3441 | } | |
3442 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
3443 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
3444 | UINT_MAX, 1); | |
3445 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
3446 | UINT_MAX, 1); | |
e572ec7e | 3447 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
3448 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
3449 | UINT_MAX, 0); | |
22e2c507 JA |
3450 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
3451 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
3452 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
3453 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
3454 | UINT_MAX, 0); | |
963b72fc | 3455 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); |
1da177e4 LT |
3456 | #undef STORE_FUNCTION |
3457 | ||
e572ec7e AV |
3458 | #define CFQ_ATTR(name) \ |
3459 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
3460 | ||
3461 | static struct elv_fs_entry cfq_attrs[] = { | |
3462 | CFQ_ATTR(quantum), | |
e572ec7e AV |
3463 | CFQ_ATTR(fifo_expire_sync), |
3464 | CFQ_ATTR(fifo_expire_async), | |
3465 | CFQ_ATTR(back_seek_max), | |
3466 | CFQ_ATTR(back_seek_penalty), | |
3467 | CFQ_ATTR(slice_sync), | |
3468 | CFQ_ATTR(slice_async), | |
3469 | CFQ_ATTR(slice_async_rq), | |
3470 | CFQ_ATTR(slice_idle), | |
963b72fc | 3471 | CFQ_ATTR(low_latency), |
e572ec7e | 3472 | __ATTR_NULL |
1da177e4 LT |
3473 | }; |
3474 | ||
1da177e4 LT |
3475 | static struct elevator_type iosched_cfq = { |
3476 | .ops = { | |
3477 | .elevator_merge_fn = cfq_merge, | |
3478 | .elevator_merged_fn = cfq_merged_request, | |
3479 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 3480 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 3481 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 3482 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 3483 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
3484 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
3485 | .elevator_queue_empty_fn = cfq_queue_empty, | |
3486 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
3487 | .elevator_former_req_fn = elv_rb_former_request, |
3488 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
3489 | .elevator_set_req_fn = cfq_set_request, |
3490 | .elevator_put_req_fn = cfq_put_request, | |
3491 | .elevator_may_queue_fn = cfq_may_queue, | |
3492 | .elevator_init_fn = cfq_init_queue, | |
3493 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 3494 | .trim = cfq_free_io_context, |
1da177e4 | 3495 | }, |
3d1ab40f | 3496 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
3497 | .elevator_name = "cfq", |
3498 | .elevator_owner = THIS_MODULE, | |
3499 | }; | |
3500 | ||
3501 | static int __init cfq_init(void) | |
3502 | { | |
22e2c507 JA |
3503 | /* |
3504 | * could be 0 on HZ < 1000 setups | |
3505 | */ | |
3506 | if (!cfq_slice_async) | |
3507 | cfq_slice_async = 1; | |
3508 | if (!cfq_slice_idle) | |
3509 | cfq_slice_idle = 1; | |
3510 | ||
1da177e4 LT |
3511 | if (cfq_slab_setup()) |
3512 | return -ENOMEM; | |
3513 | ||
2fdd82bd | 3514 | elv_register(&iosched_cfq); |
1da177e4 | 3515 | |
2fdd82bd | 3516 | return 0; |
1da177e4 LT |
3517 | } |
3518 | ||
3519 | static void __exit cfq_exit(void) | |
3520 | { | |
6e9a4738 | 3521 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 3522 | elv_unregister(&iosched_cfq); |
334e94de | 3523 | ioc_gone = &all_gone; |
fba82272 OH |
3524 | /* ioc_gone's update must be visible before reading ioc_count */ |
3525 | smp_wmb(); | |
d6de8be7 JA |
3526 | |
3527 | /* | |
3528 | * this also protects us from entering cfq_slab_kill() with | |
3529 | * pending RCU callbacks | |
3530 | */ | |
245b2e70 | 3531 | if (elv_ioc_count_read(cfq_ioc_count)) |
9a11b4ed | 3532 | wait_for_completion(&all_gone); |
83521d3e | 3533 | cfq_slab_kill(); |
1da177e4 LT |
3534 | } |
3535 | ||
3536 | module_init(cfq_init); | |
3537 | module_exit(cfq_exit); | |
3538 | ||
3539 | MODULE_AUTHOR("Jens Axboe"); | |
3540 | MODULE_LICENSE("GPL"); | |
3541 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |