Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
1da177e4 LT |
2 | * CFQ, or complete fairness queueing, disk scheduler. |
3 | * | |
4 | * Based on ideas from a previously unfinished io | |
5 | * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli. | |
6 | * | |
0fe23479 | 7 | * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk> |
1da177e4 | 8 | */ |
1da177e4 | 9 | #include <linux/module.h> |
1cc9be68 AV |
10 | #include <linux/blkdev.h> |
11 | #include <linux/elevator.h> | |
ad5ebd2f | 12 | #include <linux/jiffies.h> |
1da177e4 | 13 | #include <linux/rbtree.h> |
22e2c507 | 14 | #include <linux/ioprio.h> |
7b679138 | 15 | #include <linux/blktrace_api.h> |
1da177e4 LT |
16 | |
17 | /* | |
18 | * tunables | |
19 | */ | |
fe094d98 JA |
20 | /* max queue in one round of service */ |
21 | static const int cfq_quantum = 4; | |
64100099 | 22 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
fe094d98 JA |
23 | /* maximum backwards seek, in KiB */ |
24 | static const int cfq_back_max = 16 * 1024; | |
25 | /* penalty of a backwards seek */ | |
26 | static const int cfq_back_penalty = 2; | |
64100099 | 27 | static const int cfq_slice_sync = HZ / 10; |
3b18152c | 28 | static int cfq_slice_async = HZ / 25; |
64100099 | 29 | static const int cfq_slice_async_rq = 2; |
caaa5f9f | 30 | static int cfq_slice_idle = HZ / 125; |
5db5d642 CZ |
31 | static const int cfq_target_latency = HZ * 3/10; /* 300 ms */ |
32 | static const int cfq_hist_divisor = 4; | |
22e2c507 | 33 | |
d9e7620e | 34 | /* |
0871714e | 35 | * offset from end of service tree |
d9e7620e | 36 | */ |
0871714e | 37 | #define CFQ_IDLE_DELAY (HZ / 5) |
d9e7620e JA |
38 | |
39 | /* | |
40 | * below this threshold, we consider thinktime immediate | |
41 | */ | |
42 | #define CFQ_MIN_TT (2) | |
43 | ||
e6c5bc73 JM |
44 | /* |
45 | * Allow merged cfqqs to perform this amount of seeky I/O before | |
46 | * deciding to break the queues up again. | |
47 | */ | |
48 | #define CFQQ_COOP_TOUT (HZ) | |
49 | ||
22e2c507 | 50 | #define CFQ_SLICE_SCALE (5) |
45333d5a | 51 | #define CFQ_HW_QUEUE_MIN (5) |
22e2c507 | 52 | |
fe094d98 JA |
53 | #define RQ_CIC(rq) \ |
54 | ((struct cfq_io_context *) (rq)->elevator_private) | |
7b679138 | 55 | #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2) |
1da177e4 | 56 | |
e18b890b CL |
57 | static struct kmem_cache *cfq_pool; |
58 | static struct kmem_cache *cfq_ioc_pool; | |
1da177e4 | 59 | |
245b2e70 | 60 | static DEFINE_PER_CPU(unsigned long, cfq_ioc_count); |
334e94de | 61 | static struct completion *ioc_gone; |
9a11b4ed | 62 | static DEFINE_SPINLOCK(ioc_gone_lock); |
334e94de | 63 | |
22e2c507 JA |
64 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR |
65 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
22e2c507 JA |
66 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) |
67 | ||
206dc69b JA |
68 | #define sample_valid(samples) ((samples) > 80) |
69 | ||
cc09e299 JA |
70 | /* |
71 | * Most of our rbtree usage is for sorting with min extraction, so | |
72 | * if we cache the leftmost node we don't have to walk down the tree | |
73 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
74 | * move this into the elevator for the rq sorting as well. | |
75 | */ | |
76 | struct cfq_rb_root { | |
77 | struct rb_root rb; | |
78 | struct rb_node *left; | |
aa6f6a3d | 79 | unsigned count; |
cc09e299 | 80 | }; |
aa6f6a3d | 81 | #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, } |
cc09e299 | 82 | |
6118b70b JA |
83 | /* |
84 | * Per process-grouping structure | |
85 | */ | |
86 | struct cfq_queue { | |
87 | /* reference count */ | |
88 | atomic_t ref; | |
89 | /* various state flags, see below */ | |
90 | unsigned int flags; | |
91 | /* parent cfq_data */ | |
92 | struct cfq_data *cfqd; | |
93 | /* service_tree member */ | |
94 | struct rb_node rb_node; | |
95 | /* service_tree key */ | |
96 | unsigned long rb_key; | |
97 | /* prio tree member */ | |
98 | struct rb_node p_node; | |
99 | /* prio tree root we belong to, if any */ | |
100 | struct rb_root *p_root; | |
101 | /* sorted list of pending requests */ | |
102 | struct rb_root sort_list; | |
103 | /* if fifo isn't expired, next request to serve */ | |
104 | struct request *next_rq; | |
105 | /* requests queued in sort_list */ | |
106 | int queued[2]; | |
107 | /* currently allocated requests */ | |
108 | int allocated[2]; | |
109 | /* fifo list of requests in sort_list */ | |
110 | struct list_head fifo; | |
111 | ||
112 | unsigned long slice_end; | |
113 | long slice_resid; | |
114 | unsigned int slice_dispatch; | |
115 | ||
116 | /* pending metadata requests */ | |
117 | int meta_pending; | |
118 | /* number of requests that are on the dispatch list or inside driver */ | |
119 | int dispatched; | |
120 | ||
121 | /* io prio of this group */ | |
122 | unsigned short ioprio, org_ioprio; | |
123 | unsigned short ioprio_class, org_ioprio_class; | |
124 | ||
b2c18e1e JM |
125 | unsigned int seek_samples; |
126 | u64 seek_total; | |
127 | sector_t seek_mean; | |
128 | sector_t last_request_pos; | |
e6c5bc73 | 129 | unsigned long seeky_start; |
b2c18e1e | 130 | |
6118b70b | 131 | pid_t pid; |
df5fe3e8 | 132 | |
aa6f6a3d | 133 | struct cfq_rb_root *service_tree; |
df5fe3e8 | 134 | struct cfq_queue *new_cfqq; |
6118b70b JA |
135 | }; |
136 | ||
c0324a02 | 137 | /* |
718eee05 | 138 | * First index in the service_trees. |
c0324a02 CZ |
139 | * IDLE is handled separately, so it has negative index |
140 | */ | |
141 | enum wl_prio_t { | |
142 | IDLE_WORKLOAD = -1, | |
143 | BE_WORKLOAD = 0, | |
144 | RT_WORKLOAD = 1 | |
145 | }; | |
146 | ||
718eee05 CZ |
147 | /* |
148 | * Second index in the service_trees. | |
149 | */ | |
150 | enum wl_type_t { | |
151 | ASYNC_WORKLOAD = 0, | |
152 | SYNC_NOIDLE_WORKLOAD = 1, | |
153 | SYNC_WORKLOAD = 2 | |
154 | }; | |
155 | ||
156 | ||
22e2c507 JA |
157 | /* |
158 | * Per block device queue structure | |
159 | */ | |
1da177e4 | 160 | struct cfq_data { |
165125e1 | 161 | struct request_queue *queue; |
22e2c507 JA |
162 | |
163 | /* | |
c0324a02 CZ |
164 | * rr lists of queues with requests, onle rr for each priority class. |
165 | * Counts are embedded in the cfq_rb_root | |
166 | */ | |
718eee05 | 167 | struct cfq_rb_root service_trees[2][3]; |
c0324a02 CZ |
168 | struct cfq_rb_root service_tree_idle; |
169 | /* | |
170 | * The priority currently being served | |
22e2c507 | 171 | */ |
c0324a02 | 172 | enum wl_prio_t serving_prio; |
718eee05 CZ |
173 | enum wl_type_t serving_type; |
174 | unsigned long workload_expires; | |
a36e71f9 JA |
175 | |
176 | /* | |
177 | * Each priority tree is sorted by next_request position. These | |
178 | * trees are used when determining if two or more queues are | |
179 | * interleaving requests (see cfq_close_cooperator). | |
180 | */ | |
181 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
182 | ||
22e2c507 | 183 | unsigned int busy_queues; |
5db5d642 | 184 | unsigned int busy_queues_avg[2]; |
22e2c507 | 185 | |
5ad531db | 186 | int rq_in_driver[2]; |
3ed9a296 | 187 | int sync_flight; |
45333d5a AC |
188 | |
189 | /* | |
190 | * queue-depth detection | |
191 | */ | |
192 | int rq_queued; | |
25776e35 | 193 | int hw_tag; |
e459dd08 CZ |
194 | /* |
195 | * hw_tag can be | |
196 | * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) | |
197 | * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) | |
198 | * 0 => no NCQ | |
199 | */ | |
200 | int hw_tag_est_depth; | |
201 | unsigned int hw_tag_samples; | |
1da177e4 | 202 | |
22e2c507 JA |
203 | /* |
204 | * idle window management | |
205 | */ | |
206 | struct timer_list idle_slice_timer; | |
23e018a1 | 207 | struct work_struct unplug_work; |
1da177e4 | 208 | |
22e2c507 JA |
209 | struct cfq_queue *active_queue; |
210 | struct cfq_io_context *active_cic; | |
22e2c507 | 211 | |
c2dea2d1 VT |
212 | /* |
213 | * async queue for each priority case | |
214 | */ | |
215 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
216 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 217 | |
6d048f53 | 218 | sector_t last_position; |
1da177e4 | 219 | |
1da177e4 LT |
220 | /* |
221 | * tunables, see top of file | |
222 | */ | |
223 | unsigned int cfq_quantum; | |
22e2c507 | 224 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
225 | unsigned int cfq_back_penalty; |
226 | unsigned int cfq_back_max; | |
22e2c507 JA |
227 | unsigned int cfq_slice[2]; |
228 | unsigned int cfq_slice_async_rq; | |
229 | unsigned int cfq_slice_idle; | |
963b72fc | 230 | unsigned int cfq_latency; |
d9ff4187 AV |
231 | |
232 | struct list_head cic_list; | |
1da177e4 | 233 | |
6118b70b JA |
234 | /* |
235 | * Fallback dummy cfqq for extreme OOM conditions | |
236 | */ | |
237 | struct cfq_queue oom_cfqq; | |
365722bb VG |
238 | |
239 | unsigned long last_end_sync_rq; | |
1da177e4 LT |
240 | }; |
241 | ||
c0324a02 | 242 | static struct cfq_rb_root *service_tree_for(enum wl_prio_t prio, |
718eee05 | 243 | enum wl_type_t type, |
c0324a02 CZ |
244 | struct cfq_data *cfqd) |
245 | { | |
246 | if (prio == IDLE_WORKLOAD) | |
247 | return &cfqd->service_tree_idle; | |
248 | ||
718eee05 | 249 | return &cfqd->service_trees[prio][type]; |
c0324a02 CZ |
250 | } |
251 | ||
3b18152c | 252 | enum cfqq_state_flags { |
b0b8d749 JA |
253 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
254 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 255 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 | 256 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ |
b0b8d749 JA |
257 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
258 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
259 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 260 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 261 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
b3b6d040 | 262 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ |
3b18152c JA |
263 | }; |
264 | ||
265 | #define CFQ_CFQQ_FNS(name) \ | |
266 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
267 | { \ | |
fe094d98 | 268 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
269 | } \ |
270 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
271 | { \ | |
fe094d98 | 272 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
273 | } \ |
274 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
275 | { \ | |
fe094d98 | 276 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
277 | } |
278 | ||
279 | CFQ_CFQQ_FNS(on_rr); | |
280 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 281 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c | 282 | CFQ_CFQQ_FNS(must_alloc_slice); |
3b18152c JA |
283 | CFQ_CFQQ_FNS(fifo_expire); |
284 | CFQ_CFQQ_FNS(idle_window); | |
285 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 286 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 287 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 288 | CFQ_CFQQ_FNS(coop); |
3b18152c JA |
289 | #undef CFQ_CFQQ_FNS |
290 | ||
7b679138 JA |
291 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
292 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
293 | #define cfq_log(cfqd, fmt, args...) \ | |
294 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
295 | ||
c0324a02 CZ |
296 | static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) |
297 | { | |
298 | if (cfq_class_idle(cfqq)) | |
299 | return IDLE_WORKLOAD; | |
300 | if (cfq_class_rt(cfqq)) | |
301 | return RT_WORKLOAD; | |
302 | return BE_WORKLOAD; | |
303 | } | |
304 | ||
718eee05 CZ |
305 | |
306 | static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) | |
307 | { | |
308 | if (!cfq_cfqq_sync(cfqq)) | |
309 | return ASYNC_WORKLOAD; | |
310 | if (!cfq_cfqq_idle_window(cfqq)) | |
311 | return SYNC_NOIDLE_WORKLOAD; | |
312 | return SYNC_WORKLOAD; | |
313 | } | |
314 | ||
c0324a02 CZ |
315 | static inline int cfq_busy_queues_wl(enum wl_prio_t wl, struct cfq_data *cfqd) |
316 | { | |
317 | if (wl == IDLE_WORKLOAD) | |
318 | return cfqd->service_tree_idle.count; | |
319 | ||
718eee05 CZ |
320 | return cfqd->service_trees[wl][ASYNC_WORKLOAD].count |
321 | + cfqd->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count | |
322 | + cfqd->service_trees[wl][SYNC_WORKLOAD].count; | |
c0324a02 CZ |
323 | } |
324 | ||
165125e1 | 325 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
a6151c3a | 326 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, |
fd0928df | 327 | struct io_context *, gfp_t); |
4ac845a2 | 328 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
329 | struct io_context *); |
330 | ||
5ad531db JA |
331 | static inline int rq_in_driver(struct cfq_data *cfqd) |
332 | { | |
333 | return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1]; | |
334 | } | |
335 | ||
91fac317 | 336 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, |
a6151c3a | 337 | bool is_sync) |
91fac317 | 338 | { |
a6151c3a | 339 | return cic->cfqq[is_sync]; |
91fac317 VT |
340 | } |
341 | ||
342 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 343 | struct cfq_queue *cfqq, bool is_sync) |
91fac317 | 344 | { |
a6151c3a | 345 | cic->cfqq[is_sync] = cfqq; |
91fac317 VT |
346 | } |
347 | ||
348 | /* | |
349 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
350 | * set (in which case it could also be direct WRITE). | |
351 | */ | |
a6151c3a | 352 | static inline bool cfq_bio_sync(struct bio *bio) |
91fac317 | 353 | { |
a6151c3a | 354 | return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO); |
91fac317 | 355 | } |
1da177e4 | 356 | |
99f95e52 AM |
357 | /* |
358 | * scheduler run of queue, if there are requests pending and no one in the | |
359 | * driver that will restart queueing | |
360 | */ | |
23e018a1 | 361 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) |
99f95e52 | 362 | { |
7b679138 JA |
363 | if (cfqd->busy_queues) { |
364 | cfq_log(cfqd, "schedule dispatch"); | |
23e018a1 | 365 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 366 | } |
99f95e52 AM |
367 | } |
368 | ||
165125e1 | 369 | static int cfq_queue_empty(struct request_queue *q) |
99f95e52 AM |
370 | { |
371 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
372 | ||
b4878f24 | 373 | return !cfqd->busy_queues; |
99f95e52 AM |
374 | } |
375 | ||
44f7c160 JA |
376 | /* |
377 | * Scale schedule slice based on io priority. Use the sync time slice only | |
378 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
379 | * io only, should not get full sync slice length. | |
380 | */ | |
a6151c3a | 381 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, |
d9e7620e | 382 | unsigned short prio) |
44f7c160 | 383 | { |
d9e7620e | 384 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 385 | |
d9e7620e JA |
386 | WARN_ON(prio >= IOPRIO_BE_NR); |
387 | ||
388 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
389 | } | |
44f7c160 | 390 | |
d9e7620e JA |
391 | static inline int |
392 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
393 | { | |
394 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
395 | } |
396 | ||
5db5d642 CZ |
397 | /* |
398 | * get averaged number of queues of RT/BE priority. | |
399 | * average is updated, with a formula that gives more weight to higher numbers, | |
400 | * to quickly follows sudden increases and decrease slowly | |
401 | */ | |
402 | ||
5869619c JA |
403 | static inline unsigned cfq_get_avg_queues(struct cfq_data *cfqd, bool rt) |
404 | { | |
5db5d642 CZ |
405 | unsigned min_q, max_q; |
406 | unsigned mult = cfq_hist_divisor - 1; | |
407 | unsigned round = cfq_hist_divisor / 2; | |
c0324a02 | 408 | unsigned busy = cfq_busy_queues_wl(rt, cfqd); |
5db5d642 CZ |
409 | |
410 | min_q = min(cfqd->busy_queues_avg[rt], busy); | |
411 | max_q = max(cfqd->busy_queues_avg[rt], busy); | |
412 | cfqd->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
413 | cfq_hist_divisor; | |
414 | return cfqd->busy_queues_avg[rt]; | |
415 | } | |
416 | ||
44f7c160 JA |
417 | static inline void |
418 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
419 | { | |
5db5d642 CZ |
420 | unsigned slice = cfq_prio_to_slice(cfqd, cfqq); |
421 | if (cfqd->cfq_latency) { | |
422 | /* interested queues (we consider only the ones with the same | |
423 | * priority class) */ | |
424 | unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq)); | |
425 | unsigned sync_slice = cfqd->cfq_slice[1]; | |
426 | unsigned expect_latency = sync_slice * iq; | |
427 | if (expect_latency > cfq_target_latency) { | |
428 | unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; | |
429 | /* scale low_slice according to IO priority | |
430 | * and sync vs async */ | |
431 | unsigned low_slice = | |
432 | min(slice, base_low_slice * slice / sync_slice); | |
433 | /* the adapted slice value is scaled to fit all iqs | |
434 | * into the target latency */ | |
435 | slice = max(slice * cfq_target_latency / expect_latency, | |
436 | low_slice); | |
437 | } | |
438 | } | |
439 | cfqq->slice_end = jiffies + slice; | |
7b679138 | 440 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
441 | } |
442 | ||
443 | /* | |
444 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
445 | * isn't valid until the first request from the dispatch is activated | |
446 | * and the slice time set. | |
447 | */ | |
a6151c3a | 448 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) |
44f7c160 JA |
449 | { |
450 | if (cfq_cfqq_slice_new(cfqq)) | |
451 | return 0; | |
452 | if (time_before(jiffies, cfqq->slice_end)) | |
453 | return 0; | |
454 | ||
455 | return 1; | |
456 | } | |
457 | ||
1da177e4 | 458 | /* |
5e705374 | 459 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 460 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 461 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 462 | */ |
5e705374 | 463 | static struct request * |
cf7c25cf | 464 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) |
1da177e4 | 465 | { |
cf7c25cf | 466 | sector_t s1, s2, d1 = 0, d2 = 0; |
1da177e4 | 467 | unsigned long back_max; |
e8a99053 AM |
468 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
469 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
470 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 471 | |
5e705374 JA |
472 | if (rq1 == NULL || rq1 == rq2) |
473 | return rq2; | |
474 | if (rq2 == NULL) | |
475 | return rq1; | |
9c2c38a1 | 476 | |
5e705374 JA |
477 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
478 | return rq1; | |
479 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
480 | return rq2; | |
374f84ac JA |
481 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
482 | return rq1; | |
483 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
484 | return rq2; | |
1da177e4 | 485 | |
83096ebf TH |
486 | s1 = blk_rq_pos(rq1); |
487 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 488 | |
1da177e4 LT |
489 | /* |
490 | * by definition, 1KiB is 2 sectors | |
491 | */ | |
492 | back_max = cfqd->cfq_back_max * 2; | |
493 | ||
494 | /* | |
495 | * Strict one way elevator _except_ in the case where we allow | |
496 | * short backward seeks which are biased as twice the cost of a | |
497 | * similar forward seek. | |
498 | */ | |
499 | if (s1 >= last) | |
500 | d1 = s1 - last; | |
501 | else if (s1 + back_max >= last) | |
502 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
503 | else | |
e8a99053 | 504 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
505 | |
506 | if (s2 >= last) | |
507 | d2 = s2 - last; | |
508 | else if (s2 + back_max >= last) | |
509 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
510 | else | |
e8a99053 | 511 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
512 | |
513 | /* Found required data */ | |
e8a99053 AM |
514 | |
515 | /* | |
516 | * By doing switch() on the bit mask "wrap" we avoid having to | |
517 | * check two variables for all permutations: --> faster! | |
518 | */ | |
519 | switch (wrap) { | |
5e705374 | 520 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 521 | if (d1 < d2) |
5e705374 | 522 | return rq1; |
e8a99053 | 523 | else if (d2 < d1) |
5e705374 | 524 | return rq2; |
e8a99053 AM |
525 | else { |
526 | if (s1 >= s2) | |
5e705374 | 527 | return rq1; |
e8a99053 | 528 | else |
5e705374 | 529 | return rq2; |
e8a99053 | 530 | } |
1da177e4 | 531 | |
e8a99053 | 532 | case CFQ_RQ2_WRAP: |
5e705374 | 533 | return rq1; |
e8a99053 | 534 | case CFQ_RQ1_WRAP: |
5e705374 JA |
535 | return rq2; |
536 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
537 | default: |
538 | /* | |
539 | * Since both rqs are wrapped, | |
540 | * start with the one that's further behind head | |
541 | * (--> only *one* back seek required), | |
542 | * since back seek takes more time than forward. | |
543 | */ | |
544 | if (s1 <= s2) | |
5e705374 | 545 | return rq1; |
1da177e4 | 546 | else |
5e705374 | 547 | return rq2; |
1da177e4 LT |
548 | } |
549 | } | |
550 | ||
498d3aa2 JA |
551 | /* |
552 | * The below is leftmost cache rbtree addon | |
553 | */ | |
0871714e | 554 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 JA |
555 | { |
556 | if (!root->left) | |
557 | root->left = rb_first(&root->rb); | |
558 | ||
0871714e JA |
559 | if (root->left) |
560 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
561 | ||
562 | return NULL; | |
cc09e299 JA |
563 | } |
564 | ||
a36e71f9 JA |
565 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
566 | { | |
567 | rb_erase(n, root); | |
568 | RB_CLEAR_NODE(n); | |
569 | } | |
570 | ||
cc09e299 JA |
571 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
572 | { | |
573 | if (root->left == n) | |
574 | root->left = NULL; | |
a36e71f9 | 575 | rb_erase_init(n, &root->rb); |
aa6f6a3d | 576 | --root->count; |
cc09e299 JA |
577 | } |
578 | ||
1da177e4 LT |
579 | /* |
580 | * would be nice to take fifo expire time into account as well | |
581 | */ | |
5e705374 JA |
582 | static struct request * |
583 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
584 | struct request *last) | |
1da177e4 | 585 | { |
21183b07 JA |
586 | struct rb_node *rbnext = rb_next(&last->rb_node); |
587 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 588 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 589 | |
21183b07 | 590 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
591 | |
592 | if (rbprev) | |
5e705374 | 593 | prev = rb_entry_rq(rbprev); |
1da177e4 | 594 | |
21183b07 | 595 | if (rbnext) |
5e705374 | 596 | next = rb_entry_rq(rbnext); |
21183b07 JA |
597 | else { |
598 | rbnext = rb_first(&cfqq->sort_list); | |
599 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 600 | next = rb_entry_rq(rbnext); |
21183b07 | 601 | } |
1da177e4 | 602 | |
cf7c25cf | 603 | return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); |
1da177e4 LT |
604 | } |
605 | ||
d9e7620e JA |
606 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
607 | struct cfq_queue *cfqq) | |
1da177e4 | 608 | { |
3586e917 GJ |
609 | struct cfq_rb_root *service_tree; |
610 | ||
611 | service_tree = service_tree_for(cfqq_prio(cfqq), cfqq_type(cfqq), cfqd); | |
612 | ||
d9e7620e JA |
613 | /* |
614 | * just an approximation, should be ok. | |
615 | */ | |
3586e917 GJ |
616 | return service_tree->count * (cfq_prio_slice(cfqd, 1, 0) - |
617 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); | |
d9e7620e JA |
618 | } |
619 | ||
498d3aa2 | 620 | /* |
c0324a02 | 621 | * The cfqd->service_trees holds all pending cfq_queue's that have |
498d3aa2 JA |
622 | * requests waiting to be processed. It is sorted in the order that |
623 | * we will service the queues. | |
624 | */ | |
a36e71f9 | 625 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 626 | bool add_front) |
d9e7620e | 627 | { |
0871714e JA |
628 | struct rb_node **p, *parent; |
629 | struct cfq_queue *__cfqq; | |
d9e7620e | 630 | unsigned long rb_key; |
c0324a02 | 631 | struct cfq_rb_root *service_tree; |
498d3aa2 | 632 | int left; |
d9e7620e | 633 | |
718eee05 | 634 | service_tree = service_tree_for(cfqq_prio(cfqq), cfqq_type(cfqq), cfqd); |
0871714e JA |
635 | if (cfq_class_idle(cfqq)) { |
636 | rb_key = CFQ_IDLE_DELAY; | |
aa6f6a3d | 637 | parent = rb_last(&service_tree->rb); |
0871714e JA |
638 | if (parent && parent != &cfqq->rb_node) { |
639 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
640 | rb_key += __cfqq->rb_key; | |
641 | } else | |
642 | rb_key += jiffies; | |
643 | } else if (!add_front) { | |
b9c8946b JA |
644 | /* |
645 | * Get our rb key offset. Subtract any residual slice | |
646 | * value carried from last service. A negative resid | |
647 | * count indicates slice overrun, and this should position | |
648 | * the next service time further away in the tree. | |
649 | */ | |
edd75ffd | 650 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
b9c8946b | 651 | rb_key -= cfqq->slice_resid; |
edd75ffd | 652 | cfqq->slice_resid = 0; |
48e025e6 CZ |
653 | } else { |
654 | rb_key = -HZ; | |
aa6f6a3d | 655 | __cfqq = cfq_rb_first(service_tree); |
48e025e6 CZ |
656 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; |
657 | } | |
1da177e4 | 658 | |
d9e7620e | 659 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
99f9628a | 660 | /* |
d9e7620e | 661 | * same position, nothing more to do |
99f9628a | 662 | */ |
c0324a02 CZ |
663 | if (rb_key == cfqq->rb_key && |
664 | cfqq->service_tree == service_tree) | |
d9e7620e | 665 | return; |
1da177e4 | 666 | |
aa6f6a3d CZ |
667 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); |
668 | cfqq->service_tree = NULL; | |
1da177e4 | 669 | } |
d9e7620e | 670 | |
498d3aa2 | 671 | left = 1; |
0871714e | 672 | parent = NULL; |
aa6f6a3d CZ |
673 | cfqq->service_tree = service_tree; |
674 | p = &service_tree->rb.rb_node; | |
d9e7620e | 675 | while (*p) { |
67060e37 | 676 | struct rb_node **n; |
cc09e299 | 677 | |
d9e7620e JA |
678 | parent = *p; |
679 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
680 | ||
0c534e0a | 681 | /* |
c0324a02 | 682 | * sort by key, that represents service time. |
0c534e0a | 683 | */ |
c0324a02 | 684 | if (time_before(rb_key, __cfqq->rb_key)) |
67060e37 | 685 | n = &(*p)->rb_left; |
c0324a02 | 686 | else { |
67060e37 | 687 | n = &(*p)->rb_right; |
cc09e299 | 688 | left = 0; |
c0324a02 | 689 | } |
67060e37 JA |
690 | |
691 | p = n; | |
d9e7620e JA |
692 | } |
693 | ||
cc09e299 | 694 | if (left) |
aa6f6a3d | 695 | service_tree->left = &cfqq->rb_node; |
cc09e299 | 696 | |
d9e7620e JA |
697 | cfqq->rb_key = rb_key; |
698 | rb_link_node(&cfqq->rb_node, parent, p); | |
aa6f6a3d CZ |
699 | rb_insert_color(&cfqq->rb_node, &service_tree->rb); |
700 | service_tree->count++; | |
1da177e4 LT |
701 | } |
702 | ||
a36e71f9 | 703 | static struct cfq_queue * |
f2d1f0ae JA |
704 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
705 | sector_t sector, struct rb_node **ret_parent, | |
706 | struct rb_node ***rb_link) | |
a36e71f9 | 707 | { |
a36e71f9 JA |
708 | struct rb_node **p, *parent; |
709 | struct cfq_queue *cfqq = NULL; | |
710 | ||
711 | parent = NULL; | |
712 | p = &root->rb_node; | |
713 | while (*p) { | |
714 | struct rb_node **n; | |
715 | ||
716 | parent = *p; | |
717 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
718 | ||
719 | /* | |
720 | * Sort strictly based on sector. Smallest to the left, | |
721 | * largest to the right. | |
722 | */ | |
2e46e8b2 | 723 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 724 | n = &(*p)->rb_right; |
2e46e8b2 | 725 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
726 | n = &(*p)->rb_left; |
727 | else | |
728 | break; | |
729 | p = n; | |
3ac6c9f8 | 730 | cfqq = NULL; |
a36e71f9 JA |
731 | } |
732 | ||
733 | *ret_parent = parent; | |
734 | if (rb_link) | |
735 | *rb_link = p; | |
3ac6c9f8 | 736 | return cfqq; |
a36e71f9 JA |
737 | } |
738 | ||
739 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
740 | { | |
a36e71f9 JA |
741 | struct rb_node **p, *parent; |
742 | struct cfq_queue *__cfqq; | |
743 | ||
f2d1f0ae JA |
744 | if (cfqq->p_root) { |
745 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
746 | cfqq->p_root = NULL; | |
747 | } | |
a36e71f9 JA |
748 | |
749 | if (cfq_class_idle(cfqq)) | |
750 | return; | |
751 | if (!cfqq->next_rq) | |
752 | return; | |
753 | ||
f2d1f0ae | 754 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
755 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
756 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
757 | if (!__cfqq) { |
758 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
759 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
760 | } else | |
761 | cfqq->p_root = NULL; | |
a36e71f9 JA |
762 | } |
763 | ||
498d3aa2 JA |
764 | /* |
765 | * Update cfqq's position in the service tree. | |
766 | */ | |
edd75ffd | 767 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 768 | { |
6d048f53 JA |
769 | /* |
770 | * Resorting requires the cfqq to be on the RR list already. | |
771 | */ | |
a36e71f9 | 772 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 773 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
774 | cfq_prio_tree_add(cfqd, cfqq); |
775 | } | |
6d048f53 JA |
776 | } |
777 | ||
1da177e4 LT |
778 | /* |
779 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 780 | * the pending list according to last request service |
1da177e4 | 781 | */ |
febffd61 | 782 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 783 | { |
7b679138 | 784 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
785 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
786 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 LT |
787 | cfqd->busy_queues++; |
788 | ||
edd75ffd | 789 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
790 | } |
791 | ||
498d3aa2 JA |
792 | /* |
793 | * Called when the cfqq no longer has requests pending, remove it from | |
794 | * the service tree. | |
795 | */ | |
febffd61 | 796 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 797 | { |
7b679138 | 798 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
799 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
800 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 801 | |
aa6f6a3d CZ |
802 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
803 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
804 | cfqq->service_tree = NULL; | |
805 | } | |
f2d1f0ae JA |
806 | if (cfqq->p_root) { |
807 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
808 | cfqq->p_root = NULL; | |
809 | } | |
d9e7620e | 810 | |
1da177e4 LT |
811 | BUG_ON(!cfqd->busy_queues); |
812 | cfqd->busy_queues--; | |
813 | } | |
814 | ||
815 | /* | |
816 | * rb tree support functions | |
817 | */ | |
febffd61 | 818 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 819 | { |
5e705374 | 820 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 821 | struct cfq_data *cfqd = cfqq->cfqd; |
5e705374 | 822 | const int sync = rq_is_sync(rq); |
1da177e4 | 823 | |
b4878f24 JA |
824 | BUG_ON(!cfqq->queued[sync]); |
825 | cfqq->queued[sync]--; | |
1da177e4 | 826 | |
5e705374 | 827 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 828 | |
dd67d051 | 829 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
b4878f24 | 830 | cfq_del_cfqq_rr(cfqd, cfqq); |
1da177e4 LT |
831 | } |
832 | ||
5e705374 | 833 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 834 | { |
5e705374 | 835 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 836 | struct cfq_data *cfqd = cfqq->cfqd; |
a36e71f9 | 837 | struct request *__alias, *prev; |
1da177e4 | 838 | |
5380a101 | 839 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
840 | |
841 | /* | |
842 | * looks a little odd, but the first insert might return an alias. | |
843 | * if that happens, put the alias on the dispatch list | |
844 | */ | |
21183b07 | 845 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 846 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
847 | |
848 | if (!cfq_cfqq_on_rr(cfqq)) | |
849 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
850 | |
851 | /* | |
852 | * check if this request is a better next-serve candidate | |
853 | */ | |
a36e71f9 | 854 | prev = cfqq->next_rq; |
cf7c25cf | 855 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); |
a36e71f9 JA |
856 | |
857 | /* | |
858 | * adjust priority tree position, if ->next_rq changes | |
859 | */ | |
860 | if (prev != cfqq->next_rq) | |
861 | cfq_prio_tree_add(cfqd, cfqq); | |
862 | ||
5044eed4 | 863 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
864 | } |
865 | ||
febffd61 | 866 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 867 | { |
5380a101 JA |
868 | elv_rb_del(&cfqq->sort_list, rq); |
869 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 870 | cfq_add_rq_rb(rq); |
1da177e4 LT |
871 | } |
872 | ||
206dc69b JA |
873 | static struct request * |
874 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 875 | { |
206dc69b | 876 | struct task_struct *tsk = current; |
91fac317 | 877 | struct cfq_io_context *cic; |
206dc69b | 878 | struct cfq_queue *cfqq; |
1da177e4 | 879 | |
4ac845a2 | 880 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
881 | if (!cic) |
882 | return NULL; | |
883 | ||
884 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
885 | if (cfqq) { |
886 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
887 | ||
21183b07 | 888 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 889 | } |
1da177e4 | 890 | |
1da177e4 LT |
891 | return NULL; |
892 | } | |
893 | ||
165125e1 | 894 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 895 | { |
22e2c507 | 896 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 897 | |
5ad531db | 898 | cfqd->rq_in_driver[rq_is_sync(rq)]++; |
7b679138 | 899 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
5ad531db | 900 | rq_in_driver(cfqd)); |
25776e35 | 901 | |
5b93629b | 902 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
903 | } |
904 | ||
165125e1 | 905 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 906 | { |
b4878f24 | 907 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5ad531db | 908 | const int sync = rq_is_sync(rq); |
b4878f24 | 909 | |
5ad531db JA |
910 | WARN_ON(!cfqd->rq_in_driver[sync]); |
911 | cfqd->rq_in_driver[sync]--; | |
7b679138 | 912 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
5ad531db | 913 | rq_in_driver(cfqd)); |
1da177e4 LT |
914 | } |
915 | ||
b4878f24 | 916 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 917 | { |
5e705374 | 918 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 919 | |
5e705374 JA |
920 | if (cfqq->next_rq == rq) |
921 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 922 | |
b4878f24 | 923 | list_del_init(&rq->queuelist); |
5e705374 | 924 | cfq_del_rq_rb(rq); |
374f84ac | 925 | |
45333d5a | 926 | cfqq->cfqd->rq_queued--; |
374f84ac JA |
927 | if (rq_is_meta(rq)) { |
928 | WARN_ON(!cfqq->meta_pending); | |
929 | cfqq->meta_pending--; | |
930 | } | |
1da177e4 LT |
931 | } |
932 | ||
165125e1 JA |
933 | static int cfq_merge(struct request_queue *q, struct request **req, |
934 | struct bio *bio) | |
1da177e4 LT |
935 | { |
936 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
937 | struct request *__rq; | |
1da177e4 | 938 | |
206dc69b | 939 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 940 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
941 | *req = __rq; |
942 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
943 | } |
944 | ||
945 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
946 | } |
947 | ||
165125e1 | 948 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 949 | int type) |
1da177e4 | 950 | { |
21183b07 | 951 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 952 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 953 | |
5e705374 | 954 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 955 | } |
1da177e4 LT |
956 | } |
957 | ||
958 | static void | |
165125e1 | 959 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
960 | struct request *next) |
961 | { | |
cf7c25cf | 962 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 JA |
963 | /* |
964 | * reposition in fifo if next is older than rq | |
965 | */ | |
966 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
30996f40 | 967 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
22e2c507 | 968 | list_move(&rq->queuelist, &next->queuelist); |
30996f40 JA |
969 | rq_set_fifo_time(rq, rq_fifo_time(next)); |
970 | } | |
22e2c507 | 971 | |
cf7c25cf CZ |
972 | if (cfqq->next_rq == next) |
973 | cfqq->next_rq = rq; | |
b4878f24 | 974 | cfq_remove_request(next); |
22e2c507 JA |
975 | } |
976 | ||
165125e1 | 977 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
978 | struct bio *bio) |
979 | { | |
980 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 981 | struct cfq_io_context *cic; |
da775265 | 982 | struct cfq_queue *cfqq; |
da775265 JA |
983 | |
984 | /* | |
ec8acb69 | 985 | * Disallow merge of a sync bio into an async request. |
da775265 | 986 | */ |
91fac317 | 987 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
a6151c3a | 988 | return false; |
da775265 JA |
989 | |
990 | /* | |
719d3402 JA |
991 | * Lookup the cfqq that this bio will be queued with. Allow |
992 | * merge only if rq is queued there. | |
da775265 | 993 | */ |
4ac845a2 | 994 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 | 995 | if (!cic) |
a6151c3a | 996 | return false; |
719d3402 | 997 | |
91fac317 | 998 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
a6151c3a | 999 | return cfqq == RQ_CFQQ(rq); |
da775265 JA |
1000 | } |
1001 | ||
febffd61 JA |
1002 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
1003 | struct cfq_queue *cfqq) | |
22e2c507 JA |
1004 | { |
1005 | if (cfqq) { | |
7b679138 | 1006 | cfq_log_cfqq(cfqd, cfqq, "set_active"); |
22e2c507 | 1007 | cfqq->slice_end = 0; |
2f5cb738 JA |
1008 | cfqq->slice_dispatch = 0; |
1009 | ||
2f5cb738 | 1010 | cfq_clear_cfqq_wait_request(cfqq); |
b029195d | 1011 | cfq_clear_cfqq_must_dispatch(cfqq); |
3b18152c JA |
1012 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
1013 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 1014 | cfq_mark_cfqq_slice_new(cfqq); |
2f5cb738 JA |
1015 | |
1016 | del_timer(&cfqd->idle_slice_timer); | |
22e2c507 JA |
1017 | } |
1018 | ||
1019 | cfqd->active_queue = cfqq; | |
1020 | } | |
1021 | ||
7b14e3b5 JA |
1022 | /* |
1023 | * current cfqq expired its slice (or was too idle), select new one | |
1024 | */ | |
1025 | static void | |
1026 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
a6151c3a | 1027 | bool timed_out) |
7b14e3b5 | 1028 | { |
7b679138 JA |
1029 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
1030 | ||
7b14e3b5 JA |
1031 | if (cfq_cfqq_wait_request(cfqq)) |
1032 | del_timer(&cfqd->idle_slice_timer); | |
1033 | ||
7b14e3b5 JA |
1034 | cfq_clear_cfqq_wait_request(cfqq); |
1035 | ||
1036 | /* | |
6084cdda | 1037 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 1038 | */ |
7b679138 | 1039 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) { |
c5b680f3 | 1040 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b679138 JA |
1041 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
1042 | } | |
7b14e3b5 | 1043 | |
edd75ffd | 1044 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
1045 | |
1046 | if (cfqq == cfqd->active_queue) | |
1047 | cfqd->active_queue = NULL; | |
1048 | ||
1049 | if (cfqd->active_cic) { | |
1050 | put_io_context(cfqd->active_cic->ioc); | |
1051 | cfqd->active_cic = NULL; | |
1052 | } | |
7b14e3b5 JA |
1053 | } |
1054 | ||
a6151c3a | 1055 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) |
7b14e3b5 JA |
1056 | { |
1057 | struct cfq_queue *cfqq = cfqd->active_queue; | |
1058 | ||
1059 | if (cfqq) | |
6084cdda | 1060 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
1061 | } |
1062 | ||
498d3aa2 JA |
1063 | /* |
1064 | * Get next queue for service. Unless we have a queue preemption, | |
1065 | * we'll simply select the first cfqq in the service tree. | |
1066 | */ | |
6d048f53 | 1067 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 1068 | { |
c0324a02 | 1069 | struct cfq_rb_root *service_tree = |
718eee05 | 1070 | service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd); |
d9e7620e | 1071 | |
c0324a02 CZ |
1072 | if (RB_EMPTY_ROOT(&service_tree->rb)) |
1073 | return NULL; | |
1074 | return cfq_rb_first(service_tree); | |
6d048f53 JA |
1075 | } |
1076 | ||
498d3aa2 JA |
1077 | /* |
1078 | * Get and set a new active queue for service. | |
1079 | */ | |
a36e71f9 JA |
1080 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
1081 | struct cfq_queue *cfqq) | |
6d048f53 | 1082 | { |
e00ef799 | 1083 | if (!cfqq) |
a36e71f9 | 1084 | cfqq = cfq_get_next_queue(cfqd); |
6d048f53 | 1085 | |
22e2c507 | 1086 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 1087 | return cfqq; |
22e2c507 JA |
1088 | } |
1089 | ||
d9e7620e JA |
1090 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
1091 | struct request *rq) | |
1092 | { | |
83096ebf TH |
1093 | if (blk_rq_pos(rq) >= cfqd->last_position) |
1094 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 1095 | else |
83096ebf | 1096 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
1097 | } |
1098 | ||
b2c18e1e JM |
1099 | #define CFQQ_SEEK_THR 8 * 1024 |
1100 | #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR) | |
04dc6e71 | 1101 | |
b2c18e1e JM |
1102 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
1103 | struct request *rq) | |
6d048f53 | 1104 | { |
b2c18e1e | 1105 | sector_t sdist = cfqq->seek_mean; |
6d048f53 | 1106 | |
b2c18e1e JM |
1107 | if (!sample_valid(cfqq->seek_samples)) |
1108 | sdist = CFQQ_SEEK_THR; | |
6d048f53 | 1109 | |
04dc6e71 | 1110 | return cfq_dist_from_last(cfqd, rq) <= sdist; |
6d048f53 JA |
1111 | } |
1112 | ||
a36e71f9 JA |
1113 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
1114 | struct cfq_queue *cur_cfqq) | |
1115 | { | |
f2d1f0ae | 1116 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
1117 | struct rb_node *parent, *node; |
1118 | struct cfq_queue *__cfqq; | |
1119 | sector_t sector = cfqd->last_position; | |
1120 | ||
1121 | if (RB_EMPTY_ROOT(root)) | |
1122 | return NULL; | |
1123 | ||
1124 | /* | |
1125 | * First, if we find a request starting at the end of the last | |
1126 | * request, choose it. | |
1127 | */ | |
f2d1f0ae | 1128 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
1129 | if (__cfqq) |
1130 | return __cfqq; | |
1131 | ||
1132 | /* | |
1133 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1134 | * will contain the closest sector. | |
1135 | */ | |
1136 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
b2c18e1e | 1137 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1138 | return __cfqq; |
1139 | ||
2e46e8b2 | 1140 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1141 | node = rb_next(&__cfqq->p_node); |
1142 | else | |
1143 | node = rb_prev(&__cfqq->p_node); | |
1144 | if (!node) | |
1145 | return NULL; | |
1146 | ||
1147 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
b2c18e1e | 1148 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1149 | return __cfqq; |
1150 | ||
1151 | return NULL; | |
1152 | } | |
1153 | ||
1154 | /* | |
1155 | * cfqd - obvious | |
1156 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1157 | * closely cooperating with itself. | |
1158 | * | |
1159 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1160 | * one request, and that cfqd->last_position reflects a position on the disk | |
1161 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1162 | * assumption. | |
1163 | */ | |
1164 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
b3b6d040 | 1165 | struct cfq_queue *cur_cfqq) |
6d048f53 | 1166 | { |
a36e71f9 JA |
1167 | struct cfq_queue *cfqq; |
1168 | ||
e6c5bc73 JM |
1169 | if (!cfq_cfqq_sync(cur_cfqq)) |
1170 | return NULL; | |
1171 | if (CFQQ_SEEKY(cur_cfqq)) | |
1172 | return NULL; | |
1173 | ||
6d048f53 | 1174 | /* |
d9e7620e JA |
1175 | * We should notice if some of the queues are cooperating, eg |
1176 | * working closely on the same area of the disk. In that case, | |
1177 | * we can group them together and don't waste time idling. | |
6d048f53 | 1178 | */ |
a36e71f9 JA |
1179 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1180 | if (!cfqq) | |
1181 | return NULL; | |
1182 | ||
df5fe3e8 JM |
1183 | /* |
1184 | * It only makes sense to merge sync queues. | |
1185 | */ | |
1186 | if (!cfq_cfqq_sync(cfqq)) | |
1187 | return NULL; | |
e6c5bc73 JM |
1188 | if (CFQQ_SEEKY(cfqq)) |
1189 | return NULL; | |
df5fe3e8 | 1190 | |
c0324a02 CZ |
1191 | /* |
1192 | * Do not merge queues of different priority classes | |
1193 | */ | |
1194 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
1195 | return NULL; | |
1196 | ||
a36e71f9 | 1197 | return cfqq; |
6d048f53 JA |
1198 | } |
1199 | ||
a6d44e98 CZ |
1200 | /* |
1201 | * Determine whether we should enforce idle window for this queue. | |
1202 | */ | |
1203 | ||
1204 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1205 | { | |
1206 | enum wl_prio_t prio = cfqq_prio(cfqq); | |
718eee05 | 1207 | struct cfq_rb_root *service_tree = cfqq->service_tree; |
a6d44e98 CZ |
1208 | |
1209 | /* We never do for idle class queues. */ | |
1210 | if (prio == IDLE_WORKLOAD) | |
1211 | return false; | |
1212 | ||
1213 | /* We do for queues that were marked with idle window flag. */ | |
1214 | if (cfq_cfqq_idle_window(cfqq)) | |
1215 | return true; | |
1216 | ||
1217 | /* | |
1218 | * Otherwise, we do only if they are the last ones | |
1219 | * in their service tree. | |
1220 | */ | |
718eee05 CZ |
1221 | if (!service_tree) |
1222 | service_tree = service_tree_for(prio, cfqq_type(cfqq), cfqd); | |
1223 | ||
a6d44e98 CZ |
1224 | if (service_tree->count == 0) |
1225 | return true; | |
1226 | ||
1227 | return (service_tree->count == 1 && cfq_rb_first(service_tree) == cfqq); | |
1228 | } | |
1229 | ||
6d048f53 | 1230 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1231 | { |
1792669c | 1232 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1233 | struct cfq_io_context *cic; |
7b14e3b5 JA |
1234 | unsigned long sl; |
1235 | ||
a68bbddb | 1236 | /* |
f7d7b7a7 JA |
1237 | * SSD device without seek penalty, disable idling. But only do so |
1238 | * for devices that support queuing, otherwise we still have a problem | |
1239 | * with sync vs async workloads. | |
a68bbddb | 1240 | */ |
f7d7b7a7 | 1241 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
1242 | return; |
1243 | ||
dd67d051 | 1244 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 1245 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
1246 | |
1247 | /* | |
1248 | * idle is disabled, either manually or by past process history | |
1249 | */ | |
a6d44e98 | 1250 | if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq)) |
6d048f53 JA |
1251 | return; |
1252 | ||
7b679138 JA |
1253 | /* |
1254 | * still requests with the driver, don't idle | |
1255 | */ | |
5ad531db | 1256 | if (rq_in_driver(cfqd)) |
7b679138 JA |
1257 | return; |
1258 | ||
22e2c507 JA |
1259 | /* |
1260 | * task has exited, don't wait | |
1261 | */ | |
206dc69b | 1262 | cic = cfqd->active_cic; |
66dac98e | 1263 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
1264 | return; |
1265 | ||
355b659c CZ |
1266 | /* |
1267 | * If our average think time is larger than the remaining time | |
1268 | * slice, then don't idle. This avoids overrunning the allotted | |
1269 | * time slice. | |
1270 | */ | |
1271 | if (sample_valid(cic->ttime_samples) && | |
1272 | (cfqq->slice_end - jiffies < cic->ttime_mean)) | |
1273 | return; | |
1274 | ||
3b18152c | 1275 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 1276 | |
6d048f53 | 1277 | sl = cfqd->cfq_slice_idle; |
206dc69b | 1278 | |
7b14e3b5 | 1279 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
9481ffdc | 1280 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); |
1da177e4 LT |
1281 | } |
1282 | ||
498d3aa2 JA |
1283 | /* |
1284 | * Move request from internal lists to the request queue dispatch list. | |
1285 | */ | |
165125e1 | 1286 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 1287 | { |
3ed9a296 | 1288 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1289 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 1290 | |
7b679138 JA |
1291 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
1292 | ||
06d21886 | 1293 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); |
5380a101 | 1294 | cfq_remove_request(rq); |
6d048f53 | 1295 | cfqq->dispatched++; |
5380a101 | 1296 | elv_dispatch_sort(q, rq); |
3ed9a296 JA |
1297 | |
1298 | if (cfq_cfqq_sync(cfqq)) | |
1299 | cfqd->sync_flight++; | |
1da177e4 LT |
1300 | } |
1301 | ||
1302 | /* | |
1303 | * return expired entry, or NULL to just start from scratch in rbtree | |
1304 | */ | |
febffd61 | 1305 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 | 1306 | { |
30996f40 | 1307 | struct request *rq = NULL; |
1da177e4 | 1308 | |
3b18152c | 1309 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 1310 | return NULL; |
cb887411 JA |
1311 | |
1312 | cfq_mark_cfqq_fifo_expire(cfqq); | |
1313 | ||
89850f7e JA |
1314 | if (list_empty(&cfqq->fifo)) |
1315 | return NULL; | |
1da177e4 | 1316 | |
89850f7e | 1317 | rq = rq_entry_fifo(cfqq->fifo.next); |
30996f40 | 1318 | if (time_before(jiffies, rq_fifo_time(rq))) |
7b679138 | 1319 | rq = NULL; |
1da177e4 | 1320 | |
30996f40 | 1321 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 1322 | return rq; |
1da177e4 LT |
1323 | } |
1324 | ||
22e2c507 JA |
1325 | static inline int |
1326 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1327 | { | |
1328 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 1329 | |
22e2c507 | 1330 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 1331 | |
22e2c507 | 1332 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
1333 | } |
1334 | ||
df5fe3e8 JM |
1335 | /* |
1336 | * Must be called with the queue_lock held. | |
1337 | */ | |
1338 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
1339 | { | |
1340 | int process_refs, io_refs; | |
1341 | ||
1342 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
1343 | process_refs = atomic_read(&cfqq->ref) - io_refs; | |
1344 | BUG_ON(process_refs < 0); | |
1345 | return process_refs; | |
1346 | } | |
1347 | ||
1348 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
1349 | { | |
e6c5bc73 | 1350 | int process_refs, new_process_refs; |
df5fe3e8 JM |
1351 | struct cfq_queue *__cfqq; |
1352 | ||
1353 | /* Avoid a circular list and skip interim queue merges */ | |
1354 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
1355 | if (__cfqq == cfqq) | |
1356 | return; | |
1357 | new_cfqq = __cfqq; | |
1358 | } | |
1359 | ||
1360 | process_refs = cfqq_process_refs(cfqq); | |
1361 | /* | |
1362 | * If the process for the cfqq has gone away, there is no | |
1363 | * sense in merging the queues. | |
1364 | */ | |
1365 | if (process_refs == 0) | |
1366 | return; | |
1367 | ||
e6c5bc73 JM |
1368 | /* |
1369 | * Merge in the direction of the lesser amount of work. | |
1370 | */ | |
1371 | new_process_refs = cfqq_process_refs(new_cfqq); | |
1372 | if (new_process_refs >= process_refs) { | |
1373 | cfqq->new_cfqq = new_cfqq; | |
1374 | atomic_add(process_refs, &new_cfqq->ref); | |
1375 | } else { | |
1376 | new_cfqq->new_cfqq = cfqq; | |
1377 | atomic_add(new_process_refs, &cfqq->ref); | |
1378 | } | |
df5fe3e8 JM |
1379 | } |
1380 | ||
718eee05 CZ |
1381 | static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, enum wl_prio_t prio, |
1382 | bool prio_changed) | |
1383 | { | |
1384 | struct cfq_queue *queue; | |
1385 | int i; | |
1386 | bool key_valid = false; | |
1387 | unsigned long lowest_key = 0; | |
1388 | enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; | |
1389 | ||
1390 | if (prio_changed) { | |
1391 | /* | |
1392 | * When priorities switched, we prefer starting | |
1393 | * from SYNC_NOIDLE (first choice), or just SYNC | |
1394 | * over ASYNC | |
1395 | */ | |
1396 | if (service_tree_for(prio, cur_best, cfqd)->count) | |
1397 | return cur_best; | |
1398 | cur_best = SYNC_WORKLOAD; | |
1399 | if (service_tree_for(prio, cur_best, cfqd)->count) | |
1400 | return cur_best; | |
1401 | ||
1402 | return ASYNC_WORKLOAD; | |
1403 | } | |
1404 | ||
1405 | for (i = 0; i < 3; ++i) { | |
1406 | /* otherwise, select the one with lowest rb_key */ | |
1407 | queue = cfq_rb_first(service_tree_for(prio, i, cfqd)); | |
1408 | if (queue && | |
1409 | (!key_valid || time_before(queue->rb_key, lowest_key))) { | |
1410 | lowest_key = queue->rb_key; | |
1411 | cur_best = i; | |
1412 | key_valid = true; | |
1413 | } | |
1414 | } | |
1415 | ||
1416 | return cur_best; | |
1417 | } | |
1418 | ||
1419 | static void choose_service_tree(struct cfq_data *cfqd) | |
1420 | { | |
1421 | enum wl_prio_t previous_prio = cfqd->serving_prio; | |
1422 | bool prio_changed; | |
1423 | unsigned slice; | |
1424 | unsigned count; | |
1425 | ||
1426 | /* Choose next priority. RT > BE > IDLE */ | |
1427 | if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd)) | |
1428 | cfqd->serving_prio = RT_WORKLOAD; | |
1429 | else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd)) | |
1430 | cfqd->serving_prio = BE_WORKLOAD; | |
1431 | else { | |
1432 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1433 | cfqd->workload_expires = jiffies + 1; | |
1434 | return; | |
1435 | } | |
1436 | ||
1437 | /* | |
1438 | * For RT and BE, we have to choose also the type | |
1439 | * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload | |
1440 | * expiration time | |
1441 | */ | |
1442 | prio_changed = (cfqd->serving_prio != previous_prio); | |
1443 | count = service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd) | |
1444 | ->count; | |
1445 | ||
1446 | /* | |
1447 | * If priority didn't change, check workload expiration, | |
1448 | * and that we still have other queues ready | |
1449 | */ | |
1450 | if (!prio_changed && count && | |
1451 | !time_after(jiffies, cfqd->workload_expires)) | |
1452 | return; | |
1453 | ||
1454 | /* otherwise select new workload type */ | |
1455 | cfqd->serving_type = | |
1456 | cfq_choose_wl(cfqd, cfqd->serving_prio, prio_changed); | |
1457 | count = service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd) | |
1458 | ->count; | |
1459 | ||
1460 | /* | |
1461 | * the workload slice is computed as a fraction of target latency | |
1462 | * proportional to the number of queues in that workload, over | |
1463 | * all the queues in the same priority class | |
1464 | */ | |
1465 | slice = cfq_target_latency * count / | |
1466 | max_t(unsigned, cfqd->busy_queues_avg[cfqd->serving_prio], | |
1467 | cfq_busy_queues_wl(cfqd->serving_prio, cfqd)); | |
1468 | ||
1469 | if (cfqd->serving_type == ASYNC_WORKLOAD) | |
1470 | /* async workload slice is scaled down according to | |
1471 | * the sync/async slice ratio. */ | |
1472 | slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; | |
1473 | else | |
1474 | /* sync workload slice is at least 2 * cfq_slice_idle */ | |
1475 | slice = max(slice, 2 * cfqd->cfq_slice_idle); | |
1476 | ||
1477 | slice = max_t(unsigned, slice, CFQ_MIN_TT); | |
1478 | cfqd->workload_expires = jiffies + slice; | |
1479 | } | |
1480 | ||
22e2c507 | 1481 | /* |
498d3aa2 JA |
1482 | * Select a queue for service. If we have a current active queue, |
1483 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 1484 | */ |
1b5ed5e1 | 1485 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 1486 | { |
a36e71f9 | 1487 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 1488 | |
22e2c507 JA |
1489 | cfqq = cfqd->active_queue; |
1490 | if (!cfqq) | |
1491 | goto new_queue; | |
1da177e4 | 1492 | |
22e2c507 | 1493 | /* |
6d048f53 | 1494 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 1495 | */ |
b029195d | 1496 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) |
3b18152c | 1497 | goto expire; |
1da177e4 | 1498 | |
22e2c507 | 1499 | /* |
6d048f53 JA |
1500 | * The active queue has requests and isn't expired, allow it to |
1501 | * dispatch. | |
22e2c507 | 1502 | */ |
dd67d051 | 1503 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 1504 | goto keep_queue; |
6d048f53 | 1505 | |
a36e71f9 JA |
1506 | /* |
1507 | * If another queue has a request waiting within our mean seek | |
1508 | * distance, let it run. The expire code will check for close | |
1509 | * cooperators and put the close queue at the front of the service | |
df5fe3e8 | 1510 | * tree. If possible, merge the expiring queue with the new cfqq. |
a36e71f9 | 1511 | */ |
b3b6d040 | 1512 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); |
df5fe3e8 JM |
1513 | if (new_cfqq) { |
1514 | if (!cfqq->new_cfqq) | |
1515 | cfq_setup_merge(cfqq, new_cfqq); | |
a36e71f9 | 1516 | goto expire; |
df5fe3e8 | 1517 | } |
a36e71f9 | 1518 | |
6d048f53 JA |
1519 | /* |
1520 | * No requests pending. If the active queue still has requests in | |
1521 | * flight or is idling for a new request, allow either of these | |
1522 | * conditions to happen (or time out) before selecting a new queue. | |
1523 | */ | |
cc197479 | 1524 | if (timer_pending(&cfqd->idle_slice_timer) || |
a6d44e98 | 1525 | (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) { |
caaa5f9f JA |
1526 | cfqq = NULL; |
1527 | goto keep_queue; | |
22e2c507 JA |
1528 | } |
1529 | ||
3b18152c | 1530 | expire: |
6084cdda | 1531 | cfq_slice_expired(cfqd, 0); |
3b18152c | 1532 | new_queue: |
718eee05 CZ |
1533 | /* |
1534 | * Current queue expired. Check if we have to switch to a new | |
1535 | * service tree | |
1536 | */ | |
1537 | if (!new_cfqq) | |
1538 | choose_service_tree(cfqd); | |
1539 | ||
a36e71f9 | 1540 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 1541 | keep_queue: |
3b18152c | 1542 | return cfqq; |
22e2c507 JA |
1543 | } |
1544 | ||
febffd61 | 1545 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
1546 | { |
1547 | int dispatched = 0; | |
1548 | ||
1549 | while (cfqq->next_rq) { | |
1550 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
1551 | dispatched++; | |
1552 | } | |
1553 | ||
1554 | BUG_ON(!list_empty(&cfqq->fifo)); | |
1555 | return dispatched; | |
1556 | } | |
1557 | ||
498d3aa2 JA |
1558 | /* |
1559 | * Drain our current requests. Used for barriers and when switching | |
1560 | * io schedulers on-the-fly. | |
1561 | */ | |
d9e7620e | 1562 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 1563 | { |
0871714e | 1564 | struct cfq_queue *cfqq; |
d9e7620e | 1565 | int dispatched = 0; |
718eee05 | 1566 | int i, j; |
c0324a02 | 1567 | for (i = 0; i < 2; ++i) |
718eee05 CZ |
1568 | for (j = 0; j < 3; ++j) |
1569 | while ((cfqq = cfq_rb_first(&cfqd->service_trees[i][j])) | |
1570 | != NULL) | |
1571 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); | |
1b5ed5e1 | 1572 | |
c0324a02 | 1573 | while ((cfqq = cfq_rb_first(&cfqd->service_tree_idle)) != NULL) |
d9e7620e | 1574 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); |
1b5ed5e1 | 1575 | |
6084cdda | 1576 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
1577 | |
1578 | BUG_ON(cfqd->busy_queues); | |
1579 | ||
6923715a | 1580 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
1581 | return dispatched; |
1582 | } | |
1583 | ||
0b182d61 | 1584 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
2f5cb738 | 1585 | { |
2f5cb738 | 1586 | unsigned int max_dispatch; |
22e2c507 | 1587 | |
5ad531db JA |
1588 | /* |
1589 | * Drain async requests before we start sync IO | |
1590 | */ | |
a6d44e98 | 1591 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC]) |
0b182d61 | 1592 | return false; |
5ad531db | 1593 | |
2f5cb738 JA |
1594 | /* |
1595 | * If this is an async queue and we have sync IO in flight, let it wait | |
1596 | */ | |
1597 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) | |
0b182d61 | 1598 | return false; |
2f5cb738 JA |
1599 | |
1600 | max_dispatch = cfqd->cfq_quantum; | |
1601 | if (cfq_class_idle(cfqq)) | |
1602 | max_dispatch = 1; | |
b4878f24 | 1603 | |
2f5cb738 JA |
1604 | /* |
1605 | * Does this cfqq already have too much IO in flight? | |
1606 | */ | |
1607 | if (cfqq->dispatched >= max_dispatch) { | |
1608 | /* | |
1609 | * idle queue must always only have a single IO in flight | |
1610 | */ | |
3ed9a296 | 1611 | if (cfq_class_idle(cfqq)) |
0b182d61 | 1612 | return false; |
3ed9a296 | 1613 | |
2f5cb738 JA |
1614 | /* |
1615 | * We have other queues, don't allow more IO from this one | |
1616 | */ | |
1617 | if (cfqd->busy_queues > 1) | |
0b182d61 | 1618 | return false; |
9ede209e | 1619 | |
365722bb | 1620 | /* |
8e296755 | 1621 | * Sole queue user, allow bigger slice |
365722bb | 1622 | */ |
8e296755 JA |
1623 | max_dispatch *= 4; |
1624 | } | |
1625 | ||
1626 | /* | |
1627 | * Async queues must wait a bit before being allowed dispatch. | |
1628 | * We also ramp up the dispatch depth gradually for async IO, | |
1629 | * based on the last sync IO we serviced | |
1630 | */ | |
963b72fc | 1631 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { |
8e296755 JA |
1632 | unsigned long last_sync = jiffies - cfqd->last_end_sync_rq; |
1633 | unsigned int depth; | |
365722bb | 1634 | |
61f0c1dc | 1635 | depth = last_sync / cfqd->cfq_slice[1]; |
e00c54c3 JA |
1636 | if (!depth && !cfqq->dispatched) |
1637 | depth = 1; | |
8e296755 JA |
1638 | if (depth < max_dispatch) |
1639 | max_dispatch = depth; | |
2f5cb738 | 1640 | } |
3ed9a296 | 1641 | |
0b182d61 JA |
1642 | /* |
1643 | * If we're below the current max, allow a dispatch | |
1644 | */ | |
1645 | return cfqq->dispatched < max_dispatch; | |
1646 | } | |
1647 | ||
1648 | /* | |
1649 | * Dispatch a request from cfqq, moving them to the request queue | |
1650 | * dispatch list. | |
1651 | */ | |
1652 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1653 | { | |
1654 | struct request *rq; | |
1655 | ||
1656 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
1657 | ||
1658 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
1659 | return false; | |
1660 | ||
1661 | /* | |
1662 | * follow expired path, else get first next available | |
1663 | */ | |
1664 | rq = cfq_check_fifo(cfqq); | |
1665 | if (!rq) | |
1666 | rq = cfqq->next_rq; | |
1667 | ||
1668 | /* | |
1669 | * insert request into driver dispatch list | |
1670 | */ | |
1671 | cfq_dispatch_insert(cfqd->queue, rq); | |
1672 | ||
1673 | if (!cfqd->active_cic) { | |
1674 | struct cfq_io_context *cic = RQ_CIC(rq); | |
1675 | ||
1676 | atomic_long_inc(&cic->ioc->refcount); | |
1677 | cfqd->active_cic = cic; | |
1678 | } | |
1679 | ||
1680 | return true; | |
1681 | } | |
1682 | ||
1683 | /* | |
1684 | * Find the cfqq that we need to service and move a request from that to the | |
1685 | * dispatch list | |
1686 | */ | |
1687 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
1688 | { | |
1689 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1690 | struct cfq_queue *cfqq; | |
1691 | ||
1692 | if (!cfqd->busy_queues) | |
1693 | return 0; | |
1694 | ||
1695 | if (unlikely(force)) | |
1696 | return cfq_forced_dispatch(cfqd); | |
1697 | ||
1698 | cfqq = cfq_select_queue(cfqd); | |
1699 | if (!cfqq) | |
8e296755 JA |
1700 | return 0; |
1701 | ||
2f5cb738 | 1702 | /* |
0b182d61 | 1703 | * Dispatch a request from this cfqq, if it is allowed |
2f5cb738 | 1704 | */ |
0b182d61 JA |
1705 | if (!cfq_dispatch_request(cfqd, cfqq)) |
1706 | return 0; | |
1707 | ||
2f5cb738 | 1708 | cfqq->slice_dispatch++; |
b029195d | 1709 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 1710 | |
2f5cb738 JA |
1711 | /* |
1712 | * expire an async queue immediately if it has used up its slice. idle | |
1713 | * queue always expire after 1 dispatch round. | |
1714 | */ | |
1715 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
1716 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
1717 | cfq_class_idle(cfqq))) { | |
1718 | cfqq->slice_end = jiffies + 1; | |
1719 | cfq_slice_expired(cfqd, 0); | |
1da177e4 LT |
1720 | } |
1721 | ||
b217a903 | 1722 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); |
2f5cb738 | 1723 | return 1; |
1da177e4 LT |
1724 | } |
1725 | ||
1da177e4 | 1726 | /* |
5e705374 JA |
1727 | * task holds one reference to the queue, dropped when task exits. each rq |
1728 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 LT |
1729 | * |
1730 | * queue lock must be held here. | |
1731 | */ | |
1732 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
1733 | { | |
22e2c507 JA |
1734 | struct cfq_data *cfqd = cfqq->cfqd; |
1735 | ||
1736 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
1737 | |
1738 | if (!atomic_dec_and_test(&cfqq->ref)) | |
1739 | return; | |
1740 | ||
7b679138 | 1741 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 1742 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 1743 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
3b18152c | 1744 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 | 1745 | |
28f95cbc | 1746 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 1747 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 1748 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 1749 | } |
22e2c507 | 1750 | |
1da177e4 LT |
1751 | kmem_cache_free(cfq_pool, cfqq); |
1752 | } | |
1753 | ||
d6de8be7 JA |
1754 | /* |
1755 | * Must always be called with the rcu_read_lock() held | |
1756 | */ | |
07416d29 JA |
1757 | static void |
1758 | __call_for_each_cic(struct io_context *ioc, | |
1759 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1760 | { | |
1761 | struct cfq_io_context *cic; | |
1762 | struct hlist_node *n; | |
1763 | ||
1764 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) | |
1765 | func(ioc, cic); | |
1766 | } | |
1767 | ||
4ac845a2 | 1768 | /* |
34e6bbf2 | 1769 | * Call func for each cic attached to this ioc. |
4ac845a2 | 1770 | */ |
34e6bbf2 | 1771 | static void |
4ac845a2 JA |
1772 | call_for_each_cic(struct io_context *ioc, |
1773 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1da177e4 | 1774 | { |
4ac845a2 | 1775 | rcu_read_lock(); |
07416d29 | 1776 | __call_for_each_cic(ioc, func); |
4ac845a2 | 1777 | rcu_read_unlock(); |
34e6bbf2 FC |
1778 | } |
1779 | ||
1780 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
1781 | { | |
1782 | struct cfq_io_context *cic; | |
1783 | ||
1784 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
1785 | ||
1786 | kmem_cache_free(cfq_ioc_pool, cic); | |
245b2e70 | 1787 | elv_ioc_count_dec(cfq_ioc_count); |
34e6bbf2 | 1788 | |
9a11b4ed JA |
1789 | if (ioc_gone) { |
1790 | /* | |
1791 | * CFQ scheduler is exiting, grab exit lock and check | |
1792 | * the pending io context count. If it hits zero, | |
1793 | * complete ioc_gone and set it back to NULL | |
1794 | */ | |
1795 | spin_lock(&ioc_gone_lock); | |
245b2e70 | 1796 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { |
9a11b4ed JA |
1797 | complete(ioc_gone); |
1798 | ioc_gone = NULL; | |
1799 | } | |
1800 | spin_unlock(&ioc_gone_lock); | |
1801 | } | |
34e6bbf2 | 1802 | } |
4ac845a2 | 1803 | |
34e6bbf2 FC |
1804 | static void cfq_cic_free(struct cfq_io_context *cic) |
1805 | { | |
1806 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
1807 | } |
1808 | ||
1809 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
1810 | { | |
1811 | unsigned long flags; | |
1812 | ||
1813 | BUG_ON(!cic->dead_key); | |
1814 | ||
1815 | spin_lock_irqsave(&ioc->lock, flags); | |
1816 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
ffc4e759 | 1817 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
1818 | spin_unlock_irqrestore(&ioc->lock, flags); |
1819 | ||
34e6bbf2 | 1820 | cfq_cic_free(cic); |
4ac845a2 JA |
1821 | } |
1822 | ||
d6de8be7 JA |
1823 | /* |
1824 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
1825 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
1826 | * and ->trim() which is called with the task lock held | |
1827 | */ | |
4ac845a2 JA |
1828 | static void cfq_free_io_context(struct io_context *ioc) |
1829 | { | |
4ac845a2 | 1830 | /* |
34e6bbf2 FC |
1831 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
1832 | * so no more cic's are allowed to be linked into this ioc. So it | |
1833 | * should be ok to iterate over the known list, we will see all cic's | |
1834 | * since no new ones are added. | |
4ac845a2 | 1835 | */ |
07416d29 | 1836 | __call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
1837 | } |
1838 | ||
89850f7e | 1839 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1840 | { |
df5fe3e8 JM |
1841 | struct cfq_queue *__cfqq, *next; |
1842 | ||
28f95cbc | 1843 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 1844 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 1845 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 1846 | } |
22e2c507 | 1847 | |
df5fe3e8 JM |
1848 | /* |
1849 | * If this queue was scheduled to merge with another queue, be | |
1850 | * sure to drop the reference taken on that queue (and others in | |
1851 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
1852 | */ | |
1853 | __cfqq = cfqq->new_cfqq; | |
1854 | while (__cfqq) { | |
1855 | if (__cfqq == cfqq) { | |
1856 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
1857 | break; | |
1858 | } | |
1859 | next = __cfqq->new_cfqq; | |
1860 | cfq_put_queue(__cfqq); | |
1861 | __cfqq = next; | |
1862 | } | |
1863 | ||
89850f7e JA |
1864 | cfq_put_queue(cfqq); |
1865 | } | |
22e2c507 | 1866 | |
89850f7e JA |
1867 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
1868 | struct cfq_io_context *cic) | |
1869 | { | |
4faa3c81 FC |
1870 | struct io_context *ioc = cic->ioc; |
1871 | ||
fc46379d | 1872 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
1873 | |
1874 | /* | |
1875 | * Make sure key == NULL is seen for dead queues | |
1876 | */ | |
fc46379d | 1877 | smp_wmb(); |
4ac845a2 | 1878 | cic->dead_key = (unsigned long) cic->key; |
fc46379d JA |
1879 | cic->key = NULL; |
1880 | ||
4faa3c81 FC |
1881 | if (ioc->ioc_data == cic) |
1882 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
1883 | ||
ff6657c6 JA |
1884 | if (cic->cfqq[BLK_RW_ASYNC]) { |
1885 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
1886 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
1887 | } |
1888 | ||
ff6657c6 JA |
1889 | if (cic->cfqq[BLK_RW_SYNC]) { |
1890 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
1891 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 1892 | } |
89850f7e JA |
1893 | } |
1894 | ||
4ac845a2 JA |
1895 | static void cfq_exit_single_io_context(struct io_context *ioc, |
1896 | struct cfq_io_context *cic) | |
89850f7e JA |
1897 | { |
1898 | struct cfq_data *cfqd = cic->key; | |
1899 | ||
89850f7e | 1900 | if (cfqd) { |
165125e1 | 1901 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 1902 | unsigned long flags; |
89850f7e | 1903 | |
4ac845a2 | 1904 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
1905 | |
1906 | /* | |
1907 | * Ensure we get a fresh copy of the ->key to prevent | |
1908 | * race between exiting task and queue | |
1909 | */ | |
1910 | smp_read_barrier_depends(); | |
1911 | if (cic->key) | |
1912 | __cfq_exit_single_io_context(cfqd, cic); | |
1913 | ||
4ac845a2 | 1914 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 1915 | } |
1da177e4 LT |
1916 | } |
1917 | ||
498d3aa2 JA |
1918 | /* |
1919 | * The process that ioc belongs to has exited, we need to clean up | |
1920 | * and put the internal structures we have that belongs to that process. | |
1921 | */ | |
e2d74ac0 | 1922 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 1923 | { |
4ac845a2 | 1924 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
1925 | } |
1926 | ||
22e2c507 | 1927 | static struct cfq_io_context * |
8267e268 | 1928 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1929 | { |
b5deef90 | 1930 | struct cfq_io_context *cic; |
1da177e4 | 1931 | |
94f6030c CL |
1932 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
1933 | cfqd->queue->node); | |
1da177e4 | 1934 | if (cic) { |
22e2c507 | 1935 | cic->last_end_request = jiffies; |
553698f9 | 1936 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 1937 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
1938 | cic->dtor = cfq_free_io_context; |
1939 | cic->exit = cfq_exit_io_context; | |
245b2e70 | 1940 | elv_ioc_count_inc(cfq_ioc_count); |
1da177e4 LT |
1941 | } |
1942 | ||
1943 | return cic; | |
1944 | } | |
1945 | ||
fd0928df | 1946 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
1947 | { |
1948 | struct task_struct *tsk = current; | |
1949 | int ioprio_class; | |
1950 | ||
3b18152c | 1951 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
1952 | return; |
1953 | ||
fd0928df | 1954 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 1955 | switch (ioprio_class) { |
fe094d98 JA |
1956 | default: |
1957 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
1958 | case IOPRIO_CLASS_NONE: | |
1959 | /* | |
6d63c275 | 1960 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
1961 | */ |
1962 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 1963 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
1964 | break; |
1965 | case IOPRIO_CLASS_RT: | |
1966 | cfqq->ioprio = task_ioprio(ioc); | |
1967 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
1968 | break; | |
1969 | case IOPRIO_CLASS_BE: | |
1970 | cfqq->ioprio = task_ioprio(ioc); | |
1971 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
1972 | break; | |
1973 | case IOPRIO_CLASS_IDLE: | |
1974 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
1975 | cfqq->ioprio = 7; | |
1976 | cfq_clear_cfqq_idle_window(cfqq); | |
1977 | break; | |
22e2c507 JA |
1978 | } |
1979 | ||
1980 | /* | |
1981 | * keep track of original prio settings in case we have to temporarily | |
1982 | * elevate the priority of this queue | |
1983 | */ | |
1984 | cfqq->org_ioprio = cfqq->ioprio; | |
1985 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 1986 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
1987 | } |
1988 | ||
febffd61 | 1989 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 1990 | { |
478a82b0 AV |
1991 | struct cfq_data *cfqd = cic->key; |
1992 | struct cfq_queue *cfqq; | |
c1b707d2 | 1993 | unsigned long flags; |
35e6077c | 1994 | |
caaa5f9f JA |
1995 | if (unlikely(!cfqd)) |
1996 | return; | |
1997 | ||
c1b707d2 | 1998 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 1999 | |
ff6657c6 | 2000 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
2001 | if (cfqq) { |
2002 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
2003 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
2004 | GFP_ATOMIC); | |
caaa5f9f | 2005 | if (new_cfqq) { |
ff6657c6 | 2006 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
2007 | cfq_put_queue(cfqq); |
2008 | } | |
22e2c507 | 2009 | } |
caaa5f9f | 2010 | |
ff6657c6 | 2011 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
2012 | if (cfqq) |
2013 | cfq_mark_cfqq_prio_changed(cfqq); | |
2014 | ||
c1b707d2 | 2015 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
2016 | } |
2017 | ||
fc46379d | 2018 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 2019 | { |
4ac845a2 | 2020 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 2021 | ioc->ioprio_changed = 0; |
22e2c507 JA |
2022 | } |
2023 | ||
d5036d77 | 2024 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 2025 | pid_t pid, bool is_sync) |
d5036d77 JA |
2026 | { |
2027 | RB_CLEAR_NODE(&cfqq->rb_node); | |
2028 | RB_CLEAR_NODE(&cfqq->p_node); | |
2029 | INIT_LIST_HEAD(&cfqq->fifo); | |
2030 | ||
2031 | atomic_set(&cfqq->ref, 0); | |
2032 | cfqq->cfqd = cfqd; | |
2033 | ||
2034 | cfq_mark_cfqq_prio_changed(cfqq); | |
2035 | ||
2036 | if (is_sync) { | |
2037 | if (!cfq_class_idle(cfqq)) | |
2038 | cfq_mark_cfqq_idle_window(cfqq); | |
2039 | cfq_mark_cfqq_sync(cfqq); | |
2040 | } | |
2041 | cfqq->pid = pid; | |
2042 | } | |
2043 | ||
22e2c507 | 2044 | static struct cfq_queue * |
a6151c3a | 2045 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, |
fd0928df | 2046 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 2047 | { |
22e2c507 | 2048 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 2049 | struct cfq_io_context *cic; |
22e2c507 JA |
2050 | |
2051 | retry: | |
4ac845a2 | 2052 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
2053 | /* cic always exists here */ |
2054 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 2055 | |
6118b70b JA |
2056 | /* |
2057 | * Always try a new alloc if we fell back to the OOM cfqq | |
2058 | * originally, since it should just be a temporary situation. | |
2059 | */ | |
2060 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
2061 | cfqq = NULL; | |
22e2c507 JA |
2062 | if (new_cfqq) { |
2063 | cfqq = new_cfqq; | |
2064 | new_cfqq = NULL; | |
2065 | } else if (gfp_mask & __GFP_WAIT) { | |
2066 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 2067 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 2068 | gfp_mask | __GFP_ZERO, |
94f6030c | 2069 | cfqd->queue->node); |
22e2c507 | 2070 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
2071 | if (new_cfqq) |
2072 | goto retry; | |
22e2c507 | 2073 | } else { |
94f6030c CL |
2074 | cfqq = kmem_cache_alloc_node(cfq_pool, |
2075 | gfp_mask | __GFP_ZERO, | |
2076 | cfqd->queue->node); | |
22e2c507 JA |
2077 | } |
2078 | ||
6118b70b JA |
2079 | if (cfqq) { |
2080 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
2081 | cfq_init_prio_data(cfqq, ioc); | |
2082 | cfq_log_cfqq(cfqd, cfqq, "alloced"); | |
2083 | } else | |
2084 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
2085 | } |
2086 | ||
2087 | if (new_cfqq) | |
2088 | kmem_cache_free(cfq_pool, new_cfqq); | |
2089 | ||
22e2c507 JA |
2090 | return cfqq; |
2091 | } | |
2092 | ||
c2dea2d1 VT |
2093 | static struct cfq_queue ** |
2094 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
2095 | { | |
fe094d98 | 2096 | switch (ioprio_class) { |
c2dea2d1 VT |
2097 | case IOPRIO_CLASS_RT: |
2098 | return &cfqd->async_cfqq[0][ioprio]; | |
2099 | case IOPRIO_CLASS_BE: | |
2100 | return &cfqd->async_cfqq[1][ioprio]; | |
2101 | case IOPRIO_CLASS_IDLE: | |
2102 | return &cfqd->async_idle_cfqq; | |
2103 | default: | |
2104 | BUG(); | |
2105 | } | |
2106 | } | |
2107 | ||
15c31be4 | 2108 | static struct cfq_queue * |
a6151c3a | 2109 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, |
15c31be4 JA |
2110 | gfp_t gfp_mask) |
2111 | { | |
fd0928df JA |
2112 | const int ioprio = task_ioprio(ioc); |
2113 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 2114 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
2115 | struct cfq_queue *cfqq = NULL; |
2116 | ||
c2dea2d1 VT |
2117 | if (!is_sync) { |
2118 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
2119 | cfqq = *async_cfqq; | |
2120 | } | |
2121 | ||
6118b70b | 2122 | if (!cfqq) |
fd0928df | 2123 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
2124 | |
2125 | /* | |
2126 | * pin the queue now that it's allocated, scheduler exit will prune it | |
2127 | */ | |
c2dea2d1 | 2128 | if (!is_sync && !(*async_cfqq)) { |
15c31be4 | 2129 | atomic_inc(&cfqq->ref); |
c2dea2d1 | 2130 | *async_cfqq = cfqq; |
15c31be4 JA |
2131 | } |
2132 | ||
2133 | atomic_inc(&cfqq->ref); | |
2134 | return cfqq; | |
2135 | } | |
2136 | ||
498d3aa2 JA |
2137 | /* |
2138 | * We drop cfq io contexts lazily, so we may find a dead one. | |
2139 | */ | |
dbecf3ab | 2140 | static void |
4ac845a2 JA |
2141 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
2142 | struct cfq_io_context *cic) | |
dbecf3ab | 2143 | { |
4ac845a2 JA |
2144 | unsigned long flags; |
2145 | ||
fc46379d | 2146 | WARN_ON(!list_empty(&cic->queue_list)); |
597bc485 | 2147 | |
4ac845a2 JA |
2148 | spin_lock_irqsave(&ioc->lock, flags); |
2149 | ||
4faa3c81 | 2150 | BUG_ON(ioc->ioc_data == cic); |
597bc485 | 2151 | |
4ac845a2 | 2152 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); |
ffc4e759 | 2153 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2154 | spin_unlock_irqrestore(&ioc->lock, flags); |
2155 | ||
2156 | cfq_cic_free(cic); | |
dbecf3ab OH |
2157 | } |
2158 | ||
e2d74ac0 | 2159 | static struct cfq_io_context * |
4ac845a2 | 2160 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 2161 | { |
e2d74ac0 | 2162 | struct cfq_io_context *cic; |
d6de8be7 | 2163 | unsigned long flags; |
4ac845a2 | 2164 | void *k; |
e2d74ac0 | 2165 | |
91fac317 VT |
2166 | if (unlikely(!ioc)) |
2167 | return NULL; | |
2168 | ||
d6de8be7 JA |
2169 | rcu_read_lock(); |
2170 | ||
597bc485 JA |
2171 | /* |
2172 | * we maintain a last-hit cache, to avoid browsing over the tree | |
2173 | */ | |
4ac845a2 | 2174 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
2175 | if (cic && cic->key == cfqd) { |
2176 | rcu_read_unlock(); | |
597bc485 | 2177 | return cic; |
d6de8be7 | 2178 | } |
597bc485 | 2179 | |
4ac845a2 | 2180 | do { |
4ac845a2 JA |
2181 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); |
2182 | rcu_read_unlock(); | |
2183 | if (!cic) | |
2184 | break; | |
be3b0753 OH |
2185 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
2186 | k = cic->key; | |
2187 | if (unlikely(!k)) { | |
4ac845a2 | 2188 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 2189 | rcu_read_lock(); |
4ac845a2 | 2190 | continue; |
dbecf3ab | 2191 | } |
e2d74ac0 | 2192 | |
d6de8be7 | 2193 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 2194 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 2195 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
2196 | break; |
2197 | } while (1); | |
e2d74ac0 | 2198 | |
4ac845a2 | 2199 | return cic; |
e2d74ac0 JA |
2200 | } |
2201 | ||
4ac845a2 JA |
2202 | /* |
2203 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
2204 | * the process specific cfq io context when entered from the block layer. | |
2205 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
2206 | */ | |
febffd61 JA |
2207 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
2208 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 2209 | { |
0261d688 | 2210 | unsigned long flags; |
4ac845a2 | 2211 | int ret; |
e2d74ac0 | 2212 | |
4ac845a2 JA |
2213 | ret = radix_tree_preload(gfp_mask); |
2214 | if (!ret) { | |
2215 | cic->ioc = ioc; | |
2216 | cic->key = cfqd; | |
e2d74ac0 | 2217 | |
4ac845a2 JA |
2218 | spin_lock_irqsave(&ioc->lock, flags); |
2219 | ret = radix_tree_insert(&ioc->radix_root, | |
2220 | (unsigned long) cfqd, cic); | |
ffc4e759 JA |
2221 | if (!ret) |
2222 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 2223 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 2224 | |
4ac845a2 JA |
2225 | radix_tree_preload_end(); |
2226 | ||
2227 | if (!ret) { | |
2228 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
2229 | list_add(&cic->queue_list, &cfqd->cic_list); | |
2230 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2231 | } | |
e2d74ac0 JA |
2232 | } |
2233 | ||
4ac845a2 JA |
2234 | if (ret) |
2235 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 2236 | |
4ac845a2 | 2237 | return ret; |
e2d74ac0 JA |
2238 | } |
2239 | ||
1da177e4 LT |
2240 | /* |
2241 | * Setup general io context and cfq io context. There can be several cfq | |
2242 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 2243 | * than one device managed by cfq. |
1da177e4 LT |
2244 | */ |
2245 | static struct cfq_io_context * | |
e2d74ac0 | 2246 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2247 | { |
22e2c507 | 2248 | struct io_context *ioc = NULL; |
1da177e4 | 2249 | struct cfq_io_context *cic; |
1da177e4 | 2250 | |
22e2c507 | 2251 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 2252 | |
b5deef90 | 2253 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
2254 | if (!ioc) |
2255 | return NULL; | |
2256 | ||
4ac845a2 | 2257 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
2258 | if (cic) |
2259 | goto out; | |
1da177e4 | 2260 | |
e2d74ac0 JA |
2261 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
2262 | if (cic == NULL) | |
2263 | goto err; | |
1da177e4 | 2264 | |
4ac845a2 JA |
2265 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
2266 | goto err_free; | |
2267 | ||
1da177e4 | 2268 | out: |
fc46379d JA |
2269 | smp_read_barrier_depends(); |
2270 | if (unlikely(ioc->ioprio_changed)) | |
2271 | cfq_ioc_set_ioprio(ioc); | |
2272 | ||
1da177e4 | 2273 | return cic; |
4ac845a2 JA |
2274 | err_free: |
2275 | cfq_cic_free(cic); | |
1da177e4 LT |
2276 | err: |
2277 | put_io_context(ioc); | |
2278 | return NULL; | |
2279 | } | |
2280 | ||
22e2c507 JA |
2281 | static void |
2282 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 2283 | { |
aaf1228d JA |
2284 | unsigned long elapsed = jiffies - cic->last_end_request; |
2285 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 2286 | |
22e2c507 JA |
2287 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
2288 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
2289 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
2290 | } | |
1da177e4 | 2291 | |
206dc69b | 2292 | static void |
b2c18e1e | 2293 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
6d048f53 | 2294 | struct request *rq) |
206dc69b JA |
2295 | { |
2296 | sector_t sdist; | |
2297 | u64 total; | |
2298 | ||
b2c18e1e | 2299 | if (!cfqq->last_request_pos) |
4d00aa47 | 2300 | sdist = 0; |
b2c18e1e JM |
2301 | else if (cfqq->last_request_pos < blk_rq_pos(rq)) |
2302 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
206dc69b | 2303 | else |
b2c18e1e | 2304 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); |
206dc69b JA |
2305 | |
2306 | /* | |
2307 | * Don't allow the seek distance to get too large from the | |
2308 | * odd fragment, pagein, etc | |
2309 | */ | |
b2c18e1e JM |
2310 | if (cfqq->seek_samples <= 60) /* second&third seek */ |
2311 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024); | |
206dc69b | 2312 | else |
b2c18e1e | 2313 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64); |
206dc69b | 2314 | |
b2c18e1e JM |
2315 | cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8; |
2316 | cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8; | |
2317 | total = cfqq->seek_total + (cfqq->seek_samples/2); | |
2318 | do_div(total, cfqq->seek_samples); | |
2319 | cfqq->seek_mean = (sector_t)total; | |
e6c5bc73 JM |
2320 | |
2321 | /* | |
2322 | * If this cfqq is shared between multiple processes, check to | |
2323 | * make sure that those processes are still issuing I/Os within | |
2324 | * the mean seek distance. If not, it may be time to break the | |
2325 | * queues apart again. | |
2326 | */ | |
2327 | if (cfq_cfqq_coop(cfqq)) { | |
2328 | if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start) | |
2329 | cfqq->seeky_start = jiffies; | |
2330 | else if (!CFQQ_SEEKY(cfqq)) | |
2331 | cfqq->seeky_start = 0; | |
2332 | } | |
206dc69b | 2333 | } |
1da177e4 | 2334 | |
22e2c507 JA |
2335 | /* |
2336 | * Disable idle window if the process thinks too long or seeks so much that | |
2337 | * it doesn't matter | |
2338 | */ | |
2339 | static void | |
2340 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2341 | struct cfq_io_context *cic) | |
2342 | { | |
7b679138 | 2343 | int old_idle, enable_idle; |
1be92f2f | 2344 | |
0871714e JA |
2345 | /* |
2346 | * Don't idle for async or idle io prio class | |
2347 | */ | |
2348 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
2349 | return; |
2350 | ||
c265a7f4 | 2351 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 2352 | |
66dac98e | 2353 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || |
718eee05 | 2354 | (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))) |
22e2c507 JA |
2355 | enable_idle = 0; |
2356 | else if (sample_valid(cic->ttime_samples)) { | |
718eee05 | 2357 | if (cic->ttime_mean > cfqd->cfq_slice_idle) |
22e2c507 JA |
2358 | enable_idle = 0; |
2359 | else | |
2360 | enable_idle = 1; | |
1da177e4 LT |
2361 | } |
2362 | ||
7b679138 JA |
2363 | if (old_idle != enable_idle) { |
2364 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
2365 | if (enable_idle) | |
2366 | cfq_mark_cfqq_idle_window(cfqq); | |
2367 | else | |
2368 | cfq_clear_cfqq_idle_window(cfqq); | |
2369 | } | |
22e2c507 | 2370 | } |
1da177e4 | 2371 | |
22e2c507 JA |
2372 | /* |
2373 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
2374 | * no or if we aren't sure, a 1 will cause a preempt. | |
2375 | */ | |
a6151c3a | 2376 | static bool |
22e2c507 | 2377 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, |
5e705374 | 2378 | struct request *rq) |
22e2c507 | 2379 | { |
6d048f53 | 2380 | struct cfq_queue *cfqq; |
22e2c507 | 2381 | |
6d048f53 JA |
2382 | cfqq = cfqd->active_queue; |
2383 | if (!cfqq) | |
a6151c3a | 2384 | return false; |
22e2c507 | 2385 | |
6d048f53 | 2386 | if (cfq_slice_used(cfqq)) |
a6151c3a | 2387 | return true; |
6d048f53 JA |
2388 | |
2389 | if (cfq_class_idle(new_cfqq)) | |
a6151c3a | 2390 | return false; |
22e2c507 JA |
2391 | |
2392 | if (cfq_class_idle(cfqq)) | |
a6151c3a | 2393 | return true; |
1e3335de | 2394 | |
718eee05 CZ |
2395 | if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD |
2396 | && new_cfqq->service_tree == cfqq->service_tree) | |
2397 | return true; | |
2398 | ||
374f84ac JA |
2399 | /* |
2400 | * if the new request is sync, but the currently running queue is | |
2401 | * not, let the sync request have priority. | |
2402 | */ | |
5e705374 | 2403 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
a6151c3a | 2404 | return true; |
1e3335de | 2405 | |
374f84ac JA |
2406 | /* |
2407 | * So both queues are sync. Let the new request get disk time if | |
2408 | * it's a metadata request and the current queue is doing regular IO. | |
2409 | */ | |
2410 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
e6ec4fe2 | 2411 | return true; |
22e2c507 | 2412 | |
3a9a3f6c DS |
2413 | /* |
2414 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
2415 | */ | |
2416 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
a6151c3a | 2417 | return true; |
3a9a3f6c | 2418 | |
1e3335de | 2419 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
a6151c3a | 2420 | return false; |
1e3335de JA |
2421 | |
2422 | /* | |
2423 | * if this request is as-good as one we would expect from the | |
2424 | * current cfqq, let it preempt | |
2425 | */ | |
e00ef799 | 2426 | if (cfq_rq_close(cfqd, cfqq, rq)) |
a6151c3a | 2427 | return true; |
1e3335de | 2428 | |
a6151c3a | 2429 | return false; |
22e2c507 JA |
2430 | } |
2431 | ||
2432 | /* | |
2433 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
2434 | * let it have half of its nominal slice. | |
2435 | */ | |
2436 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2437 | { | |
7b679138 | 2438 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
6084cdda | 2439 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 2440 | |
bf572256 JA |
2441 | /* |
2442 | * Put the new queue at the front of the of the current list, | |
2443 | * so we know that it will be selected next. | |
2444 | */ | |
2445 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
2446 | |
2447 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 2448 | |
44f7c160 JA |
2449 | cfqq->slice_end = 0; |
2450 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
2451 | } |
2452 | ||
22e2c507 | 2453 | /* |
5e705374 | 2454 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
2455 | * something we should do about it |
2456 | */ | |
2457 | static void | |
5e705374 JA |
2458 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
2459 | struct request *rq) | |
22e2c507 | 2460 | { |
5e705374 | 2461 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 2462 | |
45333d5a | 2463 | cfqd->rq_queued++; |
374f84ac JA |
2464 | if (rq_is_meta(rq)) |
2465 | cfqq->meta_pending++; | |
2466 | ||
9c2c38a1 | 2467 | cfq_update_io_thinktime(cfqd, cic); |
b2c18e1e | 2468 | cfq_update_io_seektime(cfqd, cfqq, rq); |
9c2c38a1 JA |
2469 | cfq_update_idle_window(cfqd, cfqq, cic); |
2470 | ||
b2c18e1e | 2471 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
2472 | |
2473 | if (cfqq == cfqd->active_queue) { | |
2474 | /* | |
b029195d JA |
2475 | * Remember that we saw a request from this process, but |
2476 | * don't start queuing just yet. Otherwise we risk seeing lots | |
2477 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
2478 | * and merging. If the request is already larger than a single |
2479 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
2480 | * merging is already done. Ditto for a busy system that |
2481 | * has other work pending, don't risk delaying until the | |
2482 | * idle timer unplug to continue working. | |
22e2c507 | 2483 | */ |
d6ceb25e | 2484 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
2485 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
2486 | cfqd->busy_queues > 1) { | |
d6ceb25e | 2487 | del_timer(&cfqd->idle_slice_timer); |
a7f55792 | 2488 | __blk_run_queue(cfqd->queue); |
d6ceb25e | 2489 | } |
b029195d | 2490 | cfq_mark_cfqq_must_dispatch(cfqq); |
d6ceb25e | 2491 | } |
5e705374 | 2492 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
2493 | /* |
2494 | * not the active queue - expire current slice if it is | |
2495 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
2496 | * has some old slice time left and is of higher priority or |
2497 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
2498 | */ |
2499 | cfq_preempt_queue(cfqd, cfqq); | |
a7f55792 | 2500 | __blk_run_queue(cfqd->queue); |
22e2c507 | 2501 | } |
1da177e4 LT |
2502 | } |
2503 | ||
165125e1 | 2504 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2505 | { |
b4878f24 | 2506 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 2507 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 2508 | |
7b679138 | 2509 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 2510 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 2511 | |
30996f40 | 2512 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); |
22e2c507 | 2513 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
aa6f6a3d | 2514 | cfq_add_rq_rb(rq); |
22e2c507 | 2515 | |
5e705374 | 2516 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
2517 | } |
2518 | ||
45333d5a AC |
2519 | /* |
2520 | * Update hw_tag based on peak queue depth over 50 samples under | |
2521 | * sufficient load. | |
2522 | */ | |
2523 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
2524 | { | |
1a1238a7 SL |
2525 | struct cfq_queue *cfqq = cfqd->active_queue; |
2526 | ||
e459dd08 CZ |
2527 | if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth) |
2528 | cfqd->hw_tag_est_depth = rq_in_driver(cfqd); | |
2529 | ||
2530 | if (cfqd->hw_tag == 1) | |
2531 | return; | |
45333d5a AC |
2532 | |
2533 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
5ad531db | 2534 | rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
2535 | return; |
2536 | ||
1a1238a7 SL |
2537 | /* |
2538 | * If active queue hasn't enough requests and can idle, cfq might not | |
2539 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
2540 | * case | |
2541 | */ | |
2542 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
2543 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
2544 | CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN) | |
2545 | return; | |
2546 | ||
45333d5a AC |
2547 | if (cfqd->hw_tag_samples++ < 50) |
2548 | return; | |
2549 | ||
e459dd08 | 2550 | if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
2551 | cfqd->hw_tag = 1; |
2552 | else | |
2553 | cfqd->hw_tag = 0; | |
45333d5a AC |
2554 | } |
2555 | ||
165125e1 | 2556 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2557 | { |
5e705374 | 2558 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 2559 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 2560 | const int sync = rq_is_sync(rq); |
b4878f24 | 2561 | unsigned long now; |
1da177e4 | 2562 | |
b4878f24 | 2563 | now = jiffies; |
7b679138 | 2564 | cfq_log_cfqq(cfqd, cfqq, "complete"); |
1da177e4 | 2565 | |
45333d5a AC |
2566 | cfq_update_hw_tag(cfqd); |
2567 | ||
5ad531db | 2568 | WARN_ON(!cfqd->rq_in_driver[sync]); |
6d048f53 | 2569 | WARN_ON(!cfqq->dispatched); |
5ad531db | 2570 | cfqd->rq_in_driver[sync]--; |
6d048f53 | 2571 | cfqq->dispatched--; |
1da177e4 | 2572 | |
3ed9a296 JA |
2573 | if (cfq_cfqq_sync(cfqq)) |
2574 | cfqd->sync_flight--; | |
2575 | ||
365722bb | 2576 | if (sync) { |
5e705374 | 2577 | RQ_CIC(rq)->last_end_request = now; |
365722bb VG |
2578 | cfqd->last_end_sync_rq = now; |
2579 | } | |
caaa5f9f JA |
2580 | |
2581 | /* | |
2582 | * If this is the active queue, check if it needs to be expired, | |
2583 | * or if we want to idle in case it has no pending requests. | |
2584 | */ | |
2585 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
2586 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
2587 | ||
44f7c160 JA |
2588 | if (cfq_cfqq_slice_new(cfqq)) { |
2589 | cfq_set_prio_slice(cfqd, cfqq); | |
2590 | cfq_clear_cfqq_slice_new(cfqq); | |
2591 | } | |
a36e71f9 JA |
2592 | /* |
2593 | * If there are no requests waiting in this queue, and | |
2594 | * there are other queues ready to issue requests, AND | |
2595 | * those other queues are issuing requests within our | |
2596 | * mean seek distance, give them a chance to run instead | |
2597 | * of idling. | |
2598 | */ | |
0871714e | 2599 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
6084cdda | 2600 | cfq_slice_expired(cfqd, 1); |
b3b6d040 | 2601 | else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq) && |
a36e71f9 | 2602 | sync && !rq_noidle(rq)) |
6d048f53 | 2603 | cfq_arm_slice_timer(cfqd); |
caaa5f9f | 2604 | } |
6d048f53 | 2605 | |
5ad531db | 2606 | if (!rq_in_driver(cfqd)) |
23e018a1 | 2607 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
2608 | } |
2609 | ||
22e2c507 JA |
2610 | /* |
2611 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
2612 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
2613 | */ | |
2614 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 2615 | { |
22e2c507 JA |
2616 | if (has_fs_excl()) { |
2617 | /* | |
2618 | * boost idle prio on transactions that would lock out other | |
2619 | * users of the filesystem | |
2620 | */ | |
2621 | if (cfq_class_idle(cfqq)) | |
2622 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2623 | if (cfqq->ioprio > IOPRIO_NORM) | |
2624 | cfqq->ioprio = IOPRIO_NORM; | |
2625 | } else { | |
2626 | /* | |
dddb7451 | 2627 | * unboost the queue (if needed) |
22e2c507 | 2628 | */ |
dddb7451 CZ |
2629 | cfqq->ioprio_class = cfqq->org_ioprio_class; |
2630 | cfqq->ioprio = cfqq->org_ioprio; | |
22e2c507 | 2631 | } |
22e2c507 | 2632 | } |
1da177e4 | 2633 | |
89850f7e | 2634 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 2635 | { |
1b379d8d | 2636 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 2637 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 2638 | return ELV_MQUEUE_MUST; |
3b18152c | 2639 | } |
1da177e4 | 2640 | |
22e2c507 | 2641 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
2642 | } |
2643 | ||
165125e1 | 2644 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
2645 | { |
2646 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2647 | struct task_struct *tsk = current; | |
91fac317 | 2648 | struct cfq_io_context *cic; |
22e2c507 JA |
2649 | struct cfq_queue *cfqq; |
2650 | ||
2651 | /* | |
2652 | * don't force setup of a queue from here, as a call to may_queue | |
2653 | * does not necessarily imply that a request actually will be queued. | |
2654 | * so just lookup a possibly existing queue, or return 'may queue' | |
2655 | * if that fails | |
2656 | */ | |
4ac845a2 | 2657 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
2658 | if (!cic) |
2659 | return ELV_MQUEUE_MAY; | |
2660 | ||
b0b78f81 | 2661 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 2662 | if (cfqq) { |
fd0928df | 2663 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
2664 | cfq_prio_boost(cfqq); |
2665 | ||
89850f7e | 2666 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
2667 | } |
2668 | ||
2669 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
2670 | } |
2671 | ||
1da177e4 LT |
2672 | /* |
2673 | * queue lock held here | |
2674 | */ | |
bb37b94c | 2675 | static void cfq_put_request(struct request *rq) |
1da177e4 | 2676 | { |
5e705374 | 2677 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 2678 | |
5e705374 | 2679 | if (cfqq) { |
22e2c507 | 2680 | const int rw = rq_data_dir(rq); |
1da177e4 | 2681 | |
22e2c507 JA |
2682 | BUG_ON(!cfqq->allocated[rw]); |
2683 | cfqq->allocated[rw]--; | |
1da177e4 | 2684 | |
5e705374 | 2685 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 2686 | |
1da177e4 | 2687 | rq->elevator_private = NULL; |
5e705374 | 2688 | rq->elevator_private2 = NULL; |
1da177e4 | 2689 | |
1da177e4 LT |
2690 | cfq_put_queue(cfqq); |
2691 | } | |
2692 | } | |
2693 | ||
df5fe3e8 JM |
2694 | static struct cfq_queue * |
2695 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, | |
2696 | struct cfq_queue *cfqq) | |
2697 | { | |
2698 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
2699 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
b3b6d040 | 2700 | cfq_mark_cfqq_coop(cfqq->new_cfqq); |
df5fe3e8 JM |
2701 | cfq_put_queue(cfqq); |
2702 | return cic_to_cfqq(cic, 1); | |
2703 | } | |
2704 | ||
e6c5bc73 JM |
2705 | static int should_split_cfqq(struct cfq_queue *cfqq) |
2706 | { | |
2707 | if (cfqq->seeky_start && | |
2708 | time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT)) | |
2709 | return 1; | |
2710 | return 0; | |
2711 | } | |
2712 | ||
2713 | /* | |
2714 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
2715 | * was the last process referring to said cfqq. | |
2716 | */ | |
2717 | static struct cfq_queue * | |
2718 | split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) | |
2719 | { | |
2720 | if (cfqq_process_refs(cfqq) == 1) { | |
2721 | cfqq->seeky_start = 0; | |
2722 | cfqq->pid = current->pid; | |
2723 | cfq_clear_cfqq_coop(cfqq); | |
2724 | return cfqq; | |
2725 | } | |
2726 | ||
2727 | cic_set_cfqq(cic, NULL, 1); | |
2728 | cfq_put_queue(cfqq); | |
2729 | return NULL; | |
2730 | } | |
1da177e4 | 2731 | /* |
22e2c507 | 2732 | * Allocate cfq data structures associated with this request. |
1da177e4 | 2733 | */ |
22e2c507 | 2734 | static int |
165125e1 | 2735 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
2736 | { |
2737 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2738 | struct cfq_io_context *cic; | |
2739 | const int rw = rq_data_dir(rq); | |
a6151c3a | 2740 | const bool is_sync = rq_is_sync(rq); |
22e2c507 | 2741 | struct cfq_queue *cfqq; |
1da177e4 LT |
2742 | unsigned long flags; |
2743 | ||
2744 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
2745 | ||
e2d74ac0 | 2746 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 2747 | |
1da177e4 LT |
2748 | spin_lock_irqsave(q->queue_lock, flags); |
2749 | ||
22e2c507 JA |
2750 | if (!cic) |
2751 | goto queue_fail; | |
2752 | ||
e6c5bc73 | 2753 | new_queue: |
91fac317 | 2754 | cfqq = cic_to_cfqq(cic, is_sync); |
32f2e807 | 2755 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { |
fd0928df | 2756 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 | 2757 | cic_set_cfqq(cic, cfqq, is_sync); |
df5fe3e8 | 2758 | } else { |
e6c5bc73 JM |
2759 | /* |
2760 | * If the queue was seeky for too long, break it apart. | |
2761 | */ | |
2762 | if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) { | |
2763 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); | |
2764 | cfqq = split_cfqq(cic, cfqq); | |
2765 | if (!cfqq) | |
2766 | goto new_queue; | |
2767 | } | |
2768 | ||
df5fe3e8 JM |
2769 | /* |
2770 | * Check to see if this queue is scheduled to merge with | |
2771 | * another, closely cooperating queue. The merging of | |
2772 | * queues happens here as it must be done in process context. | |
2773 | * The reference on new_cfqq was taken in merge_cfqqs. | |
2774 | */ | |
2775 | if (cfqq->new_cfqq) | |
2776 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
91fac317 | 2777 | } |
1da177e4 LT |
2778 | |
2779 | cfqq->allocated[rw]++; | |
22e2c507 | 2780 | atomic_inc(&cfqq->ref); |
1da177e4 | 2781 | |
5e705374 | 2782 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 2783 | |
5e705374 JA |
2784 | rq->elevator_private = cic; |
2785 | rq->elevator_private2 = cfqq; | |
2786 | return 0; | |
1da177e4 | 2787 | |
22e2c507 JA |
2788 | queue_fail: |
2789 | if (cic) | |
2790 | put_io_context(cic->ioc); | |
89850f7e | 2791 | |
23e018a1 | 2792 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 2793 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 2794 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
2795 | return 1; |
2796 | } | |
2797 | ||
65f27f38 | 2798 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 2799 | { |
65f27f38 | 2800 | struct cfq_data *cfqd = |
23e018a1 | 2801 | container_of(work, struct cfq_data, unplug_work); |
165125e1 | 2802 | struct request_queue *q = cfqd->queue; |
22e2c507 | 2803 | |
40bb54d1 | 2804 | spin_lock_irq(q->queue_lock); |
a7f55792 | 2805 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 2806 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
2807 | } |
2808 | ||
2809 | /* | |
2810 | * Timer running if the active_queue is currently idling inside its time slice | |
2811 | */ | |
2812 | static void cfq_idle_slice_timer(unsigned long data) | |
2813 | { | |
2814 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
2815 | struct cfq_queue *cfqq; | |
2816 | unsigned long flags; | |
3c6bd2f8 | 2817 | int timed_out = 1; |
22e2c507 | 2818 | |
7b679138 JA |
2819 | cfq_log(cfqd, "idle timer fired"); |
2820 | ||
22e2c507 JA |
2821 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
2822 | ||
fe094d98 JA |
2823 | cfqq = cfqd->active_queue; |
2824 | if (cfqq) { | |
3c6bd2f8 JA |
2825 | timed_out = 0; |
2826 | ||
b029195d JA |
2827 | /* |
2828 | * We saw a request before the queue expired, let it through | |
2829 | */ | |
2830 | if (cfq_cfqq_must_dispatch(cfqq)) | |
2831 | goto out_kick; | |
2832 | ||
22e2c507 JA |
2833 | /* |
2834 | * expired | |
2835 | */ | |
44f7c160 | 2836 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
2837 | goto expire; |
2838 | ||
2839 | /* | |
2840 | * only expire and reinvoke request handler, if there are | |
2841 | * other queues with pending requests | |
2842 | */ | |
caaa5f9f | 2843 | if (!cfqd->busy_queues) |
22e2c507 | 2844 | goto out_cont; |
22e2c507 JA |
2845 | |
2846 | /* | |
2847 | * not expired and it has a request pending, let it dispatch | |
2848 | */ | |
75e50984 | 2849 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 2850 | goto out_kick; |
22e2c507 JA |
2851 | } |
2852 | expire: | |
6084cdda | 2853 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 2854 | out_kick: |
23e018a1 | 2855 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
2856 | out_cont: |
2857 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2858 | } | |
2859 | ||
3b18152c JA |
2860 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
2861 | { | |
2862 | del_timer_sync(&cfqd->idle_slice_timer); | |
23e018a1 | 2863 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 2864 | } |
22e2c507 | 2865 | |
c2dea2d1 VT |
2866 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
2867 | { | |
2868 | int i; | |
2869 | ||
2870 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
2871 | if (cfqd->async_cfqq[0][i]) | |
2872 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
2873 | if (cfqd->async_cfqq[1][i]) | |
2874 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 2875 | } |
2389d1ef ON |
2876 | |
2877 | if (cfqd->async_idle_cfqq) | |
2878 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
2879 | } |
2880 | ||
b374d18a | 2881 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 2882 | { |
22e2c507 | 2883 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 2884 | struct request_queue *q = cfqd->queue; |
22e2c507 | 2885 | |
3b18152c | 2886 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 2887 | |
d9ff4187 | 2888 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 2889 | |
d9ff4187 | 2890 | if (cfqd->active_queue) |
6084cdda | 2891 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
2892 | |
2893 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
2894 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
2895 | struct cfq_io_context, | |
2896 | queue_list); | |
89850f7e JA |
2897 | |
2898 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 2899 | } |
e2d74ac0 | 2900 | |
c2dea2d1 | 2901 | cfq_put_async_queues(cfqd); |
15c31be4 | 2902 | |
d9ff4187 | 2903 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
2904 | |
2905 | cfq_shutdown_timer_wq(cfqd); | |
2906 | ||
a90d742e | 2907 | kfree(cfqd); |
1da177e4 LT |
2908 | } |
2909 | ||
165125e1 | 2910 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
2911 | { |
2912 | struct cfq_data *cfqd; | |
718eee05 | 2913 | int i, j; |
1da177e4 | 2914 | |
94f6030c | 2915 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 2916 | if (!cfqd) |
bc1c1169 | 2917 | return NULL; |
1da177e4 | 2918 | |
c0324a02 | 2919 | for (i = 0; i < 2; ++i) |
718eee05 CZ |
2920 | for (j = 0; j < 3; ++j) |
2921 | cfqd->service_trees[i][j] = CFQ_RB_ROOT; | |
c0324a02 | 2922 | cfqd->service_tree_idle = CFQ_RB_ROOT; |
26a2ac00 JA |
2923 | |
2924 | /* | |
2925 | * Not strictly needed (since RB_ROOT just clears the node and we | |
2926 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
2927 | * to add magic to the rb code | |
2928 | */ | |
2929 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
2930 | cfqd->prio_trees[i] = RB_ROOT; | |
2931 | ||
6118b70b JA |
2932 | /* |
2933 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
2934 | * Grab a permanent reference to it, so that the normal code flow | |
2935 | * will not attempt to free it. | |
2936 | */ | |
2937 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
2938 | atomic_inc(&cfqd->oom_cfqq.ref); | |
2939 | ||
d9ff4187 | 2940 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 2941 | |
1da177e4 | 2942 | cfqd->queue = q; |
1da177e4 | 2943 | |
22e2c507 JA |
2944 | init_timer(&cfqd->idle_slice_timer); |
2945 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
2946 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
2947 | ||
23e018a1 | 2948 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 2949 | |
1da177e4 | 2950 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
2951 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
2952 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
2953 | cfqd->cfq_back_max = cfq_back_max; |
2954 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
2955 | cfqd->cfq_slice[0] = cfq_slice_async; |
2956 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
2957 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
2958 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
963b72fc | 2959 | cfqd->cfq_latency = 1; |
e459dd08 | 2960 | cfqd->hw_tag = -1; |
365722bb | 2961 | cfqd->last_end_sync_rq = jiffies; |
bc1c1169 | 2962 | return cfqd; |
1da177e4 LT |
2963 | } |
2964 | ||
2965 | static void cfq_slab_kill(void) | |
2966 | { | |
d6de8be7 JA |
2967 | /* |
2968 | * Caller already ensured that pending RCU callbacks are completed, | |
2969 | * so we should have no busy allocations at this point. | |
2970 | */ | |
1da177e4 LT |
2971 | if (cfq_pool) |
2972 | kmem_cache_destroy(cfq_pool); | |
2973 | if (cfq_ioc_pool) | |
2974 | kmem_cache_destroy(cfq_ioc_pool); | |
2975 | } | |
2976 | ||
2977 | static int __init cfq_slab_setup(void) | |
2978 | { | |
0a31bd5f | 2979 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
2980 | if (!cfq_pool) |
2981 | goto fail; | |
2982 | ||
34e6bbf2 | 2983 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
2984 | if (!cfq_ioc_pool) |
2985 | goto fail; | |
2986 | ||
2987 | return 0; | |
2988 | fail: | |
2989 | cfq_slab_kill(); | |
2990 | return -ENOMEM; | |
2991 | } | |
2992 | ||
1da177e4 LT |
2993 | /* |
2994 | * sysfs parts below --> | |
2995 | */ | |
1da177e4 LT |
2996 | static ssize_t |
2997 | cfq_var_show(unsigned int var, char *page) | |
2998 | { | |
2999 | return sprintf(page, "%d\n", var); | |
3000 | } | |
3001 | ||
3002 | static ssize_t | |
3003 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
3004 | { | |
3005 | char *p = (char *) page; | |
3006 | ||
3007 | *var = simple_strtoul(p, &p, 10); | |
3008 | return count; | |
3009 | } | |
3010 | ||
1da177e4 | 3011 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 3012 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 3013 | { \ |
3d1ab40f | 3014 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3015 | unsigned int __data = __VAR; \ |
3016 | if (__CONV) \ | |
3017 | __data = jiffies_to_msecs(__data); \ | |
3018 | return cfq_var_show(__data, (page)); \ | |
3019 | } | |
3020 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
3021 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
3022 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
3023 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
3024 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
3025 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
3026 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
3027 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
3028 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
963b72fc | 3029 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); |
1da177e4 LT |
3030 | #undef SHOW_FUNCTION |
3031 | ||
3032 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 3033 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 3034 | { \ |
3d1ab40f | 3035 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3036 | unsigned int __data; \ |
3037 | int ret = cfq_var_store(&__data, (page), count); \ | |
3038 | if (__data < (MIN)) \ | |
3039 | __data = (MIN); \ | |
3040 | else if (__data > (MAX)) \ | |
3041 | __data = (MAX); \ | |
3042 | if (__CONV) \ | |
3043 | *(__PTR) = msecs_to_jiffies(__data); \ | |
3044 | else \ | |
3045 | *(__PTR) = __data; \ | |
3046 | return ret; \ | |
3047 | } | |
3048 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
3049 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
3050 | UINT_MAX, 1); | |
3051 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
3052 | UINT_MAX, 1); | |
e572ec7e | 3053 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
3054 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
3055 | UINT_MAX, 0); | |
22e2c507 JA |
3056 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
3057 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
3058 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
3059 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
3060 | UINT_MAX, 0); | |
963b72fc | 3061 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); |
1da177e4 LT |
3062 | #undef STORE_FUNCTION |
3063 | ||
e572ec7e AV |
3064 | #define CFQ_ATTR(name) \ |
3065 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
3066 | ||
3067 | static struct elv_fs_entry cfq_attrs[] = { | |
3068 | CFQ_ATTR(quantum), | |
e572ec7e AV |
3069 | CFQ_ATTR(fifo_expire_sync), |
3070 | CFQ_ATTR(fifo_expire_async), | |
3071 | CFQ_ATTR(back_seek_max), | |
3072 | CFQ_ATTR(back_seek_penalty), | |
3073 | CFQ_ATTR(slice_sync), | |
3074 | CFQ_ATTR(slice_async), | |
3075 | CFQ_ATTR(slice_async_rq), | |
3076 | CFQ_ATTR(slice_idle), | |
963b72fc | 3077 | CFQ_ATTR(low_latency), |
e572ec7e | 3078 | __ATTR_NULL |
1da177e4 LT |
3079 | }; |
3080 | ||
1da177e4 LT |
3081 | static struct elevator_type iosched_cfq = { |
3082 | .ops = { | |
3083 | .elevator_merge_fn = cfq_merge, | |
3084 | .elevator_merged_fn = cfq_merged_request, | |
3085 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 3086 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 3087 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 3088 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 3089 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
3090 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
3091 | .elevator_queue_empty_fn = cfq_queue_empty, | |
3092 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
3093 | .elevator_former_req_fn = elv_rb_former_request, |
3094 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
3095 | .elevator_set_req_fn = cfq_set_request, |
3096 | .elevator_put_req_fn = cfq_put_request, | |
3097 | .elevator_may_queue_fn = cfq_may_queue, | |
3098 | .elevator_init_fn = cfq_init_queue, | |
3099 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 3100 | .trim = cfq_free_io_context, |
1da177e4 | 3101 | }, |
3d1ab40f | 3102 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
3103 | .elevator_name = "cfq", |
3104 | .elevator_owner = THIS_MODULE, | |
3105 | }; | |
3106 | ||
3107 | static int __init cfq_init(void) | |
3108 | { | |
22e2c507 JA |
3109 | /* |
3110 | * could be 0 on HZ < 1000 setups | |
3111 | */ | |
3112 | if (!cfq_slice_async) | |
3113 | cfq_slice_async = 1; | |
3114 | if (!cfq_slice_idle) | |
3115 | cfq_slice_idle = 1; | |
3116 | ||
1da177e4 LT |
3117 | if (cfq_slab_setup()) |
3118 | return -ENOMEM; | |
3119 | ||
2fdd82bd | 3120 | elv_register(&iosched_cfq); |
1da177e4 | 3121 | |
2fdd82bd | 3122 | return 0; |
1da177e4 LT |
3123 | } |
3124 | ||
3125 | static void __exit cfq_exit(void) | |
3126 | { | |
6e9a4738 | 3127 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 3128 | elv_unregister(&iosched_cfq); |
334e94de | 3129 | ioc_gone = &all_gone; |
fba82272 OH |
3130 | /* ioc_gone's update must be visible before reading ioc_count */ |
3131 | smp_wmb(); | |
d6de8be7 JA |
3132 | |
3133 | /* | |
3134 | * this also protects us from entering cfq_slab_kill() with | |
3135 | * pending RCU callbacks | |
3136 | */ | |
245b2e70 | 3137 | if (elv_ioc_count_read(cfq_ioc_count)) |
9a11b4ed | 3138 | wait_for_completion(&all_gone); |
83521d3e | 3139 | cfq_slab_kill(); |
1da177e4 LT |
3140 | } |
3141 | ||
3142 | module_init(cfq_init); | |
3143 | module_exit(cfq_exit); | |
3144 | ||
3145 | MODULE_AUTHOR("Jens Axboe"); | |
3146 | MODULE_LICENSE("GPL"); | |
3147 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |