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