]>
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> | |
1da177e4 | 12 | #include <linux/rbtree.h> |
22e2c507 | 13 | #include <linux/ioprio.h> |
7b679138 | 14 | #include <linux/blktrace_api.h> |
1da177e4 LT |
15 | |
16 | /* | |
17 | * tunables | |
18 | */ | |
fe094d98 JA |
19 | /* max queue in one round of service */ |
20 | static const int cfq_quantum = 4; | |
64100099 | 21 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
fe094d98 JA |
22 | /* maximum backwards seek, in KiB */ |
23 | static const int cfq_back_max = 16 * 1024; | |
24 | /* penalty of a backwards seek */ | |
25 | static const int cfq_back_penalty = 2; | |
64100099 | 26 | static const int cfq_slice_sync = HZ / 10; |
3b18152c | 27 | static int cfq_slice_async = HZ / 25; |
64100099 | 28 | static const int cfq_slice_async_rq = 2; |
caaa5f9f | 29 | static int cfq_slice_idle = HZ / 125; |
5db5d642 CZ |
30 | static const int cfq_target_latency = HZ * 3/10; /* 300 ms */ |
31 | static const int cfq_hist_divisor = 4; | |
22e2c507 | 32 | |
d9e7620e | 33 | /* |
0871714e | 34 | * offset from end of service tree |
d9e7620e | 35 | */ |
0871714e | 36 | #define CFQ_IDLE_DELAY (HZ / 5) |
d9e7620e JA |
37 | |
38 | /* | |
39 | * below this threshold, we consider thinktime immediate | |
40 | */ | |
41 | #define CFQ_MIN_TT (2) | |
42 | ||
e6c5bc73 JM |
43 | /* |
44 | * Allow merged cfqqs to perform this amount of seeky I/O before | |
45 | * deciding to break the queues up again. | |
46 | */ | |
47 | #define CFQQ_COOP_TOUT (HZ) | |
48 | ||
22e2c507 | 49 | #define CFQ_SLICE_SCALE (5) |
45333d5a | 50 | #define CFQ_HW_QUEUE_MIN (5) |
22e2c507 | 51 | |
fe094d98 JA |
52 | #define RQ_CIC(rq) \ |
53 | ((struct cfq_io_context *) (rq)->elevator_private) | |
7b679138 | 54 | #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2) |
1da177e4 | 55 | |
e18b890b CL |
56 | static struct kmem_cache *cfq_pool; |
57 | static struct kmem_cache *cfq_ioc_pool; | |
1da177e4 | 58 | |
245b2e70 | 59 | static DEFINE_PER_CPU(unsigned long, cfq_ioc_count); |
334e94de | 60 | static struct completion *ioc_gone; |
9a11b4ed | 61 | static DEFINE_SPINLOCK(ioc_gone_lock); |
334e94de | 62 | |
22e2c507 JA |
63 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR |
64 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
22e2c507 JA |
65 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) |
66 | ||
206dc69b JA |
67 | #define sample_valid(samples) ((samples) > 80) |
68 | ||
cc09e299 JA |
69 | /* |
70 | * Most of our rbtree usage is for sorting with min extraction, so | |
71 | * if we cache the leftmost node we don't have to walk down the tree | |
72 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
73 | * move this into the elevator for the rq sorting as well. | |
74 | */ | |
75 | struct cfq_rb_root { | |
76 | struct rb_root rb; | |
77 | struct rb_node *left; | |
aa6f6a3d | 78 | unsigned count; |
cc09e299 | 79 | }; |
aa6f6a3d | 80 | #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, } |
cc09e299 | 81 | |
6118b70b JA |
82 | /* |
83 | * Per process-grouping structure | |
84 | */ | |
85 | struct cfq_queue { | |
86 | /* reference count */ | |
87 | atomic_t ref; | |
88 | /* various state flags, see below */ | |
89 | unsigned int flags; | |
90 | /* parent cfq_data */ | |
91 | struct cfq_data *cfqd; | |
92 | /* service_tree member */ | |
93 | struct rb_node rb_node; | |
94 | /* service_tree key */ | |
95 | unsigned long rb_key; | |
96 | /* prio tree member */ | |
97 | struct rb_node p_node; | |
98 | /* prio tree root we belong to, if any */ | |
99 | struct rb_root *p_root; | |
100 | /* sorted list of pending requests */ | |
101 | struct rb_root sort_list; | |
102 | /* if fifo isn't expired, next request to serve */ | |
103 | struct request *next_rq; | |
104 | /* requests queued in sort_list */ | |
105 | int queued[2]; | |
106 | /* currently allocated requests */ | |
107 | int allocated[2]; | |
108 | /* fifo list of requests in sort_list */ | |
109 | struct list_head fifo; | |
110 | ||
111 | unsigned long slice_end; | |
112 | long slice_resid; | |
113 | unsigned int slice_dispatch; | |
114 | ||
115 | /* pending metadata requests */ | |
116 | int meta_pending; | |
117 | /* number of requests that are on the dispatch list or inside driver */ | |
118 | int dispatched; | |
119 | ||
120 | /* io prio of this group */ | |
121 | unsigned short ioprio, org_ioprio; | |
122 | unsigned short ioprio_class, org_ioprio_class; | |
123 | ||
b2c18e1e JM |
124 | unsigned int seek_samples; |
125 | u64 seek_total; | |
126 | sector_t seek_mean; | |
127 | sector_t last_request_pos; | |
e6c5bc73 | 128 | unsigned long seeky_start; |
b2c18e1e | 129 | |
6118b70b | 130 | pid_t pid; |
df5fe3e8 | 131 | |
aa6f6a3d | 132 | struct cfq_rb_root *service_tree; |
df5fe3e8 | 133 | struct cfq_queue *new_cfqq; |
6118b70b JA |
134 | }; |
135 | ||
c0324a02 CZ |
136 | /* |
137 | * Index in the service_trees. | |
138 | * IDLE is handled separately, so it has negative index | |
139 | */ | |
140 | enum wl_prio_t { | |
141 | IDLE_WORKLOAD = -1, | |
142 | BE_WORKLOAD = 0, | |
143 | RT_WORKLOAD = 1 | |
144 | }; | |
145 | ||
22e2c507 JA |
146 | /* |
147 | * Per block device queue structure | |
148 | */ | |
1da177e4 | 149 | struct cfq_data { |
165125e1 | 150 | struct request_queue *queue; |
22e2c507 JA |
151 | |
152 | /* | |
c0324a02 CZ |
153 | * rr lists of queues with requests, onle rr for each priority class. |
154 | * Counts are embedded in the cfq_rb_root | |
155 | */ | |
156 | struct cfq_rb_root service_trees[2]; | |
157 | struct cfq_rb_root service_tree_idle; | |
158 | /* | |
159 | * The priority currently being served | |
22e2c507 | 160 | */ |
c0324a02 | 161 | enum wl_prio_t serving_prio; |
a36e71f9 JA |
162 | |
163 | /* | |
164 | * Each priority tree is sorted by next_request position. These | |
165 | * trees are used when determining if two or more queues are | |
166 | * interleaving requests (see cfq_close_cooperator). | |
167 | */ | |
168 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
169 | ||
22e2c507 | 170 | unsigned int busy_queues; |
5db5d642 | 171 | unsigned int busy_queues_avg[2]; |
22e2c507 | 172 | |
5ad531db | 173 | int rq_in_driver[2]; |
3ed9a296 | 174 | int sync_flight; |
45333d5a AC |
175 | |
176 | /* | |
177 | * queue-depth detection | |
178 | */ | |
179 | int rq_queued; | |
25776e35 | 180 | int hw_tag; |
45333d5a AC |
181 | int hw_tag_samples; |
182 | int rq_in_driver_peak; | |
1da177e4 | 183 | |
22e2c507 JA |
184 | /* |
185 | * idle window management | |
186 | */ | |
187 | struct timer_list idle_slice_timer; | |
23e018a1 | 188 | struct work_struct unplug_work; |
1da177e4 | 189 | |
22e2c507 JA |
190 | struct cfq_queue *active_queue; |
191 | struct cfq_io_context *active_cic; | |
22e2c507 | 192 | |
c2dea2d1 VT |
193 | /* |
194 | * async queue for each priority case | |
195 | */ | |
196 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
197 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 198 | |
6d048f53 | 199 | sector_t last_position; |
1da177e4 | 200 | |
1da177e4 LT |
201 | /* |
202 | * tunables, see top of file | |
203 | */ | |
204 | unsigned int cfq_quantum; | |
22e2c507 | 205 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
206 | unsigned int cfq_back_penalty; |
207 | unsigned int cfq_back_max; | |
22e2c507 JA |
208 | unsigned int cfq_slice[2]; |
209 | unsigned int cfq_slice_async_rq; | |
210 | unsigned int cfq_slice_idle; | |
963b72fc | 211 | unsigned int cfq_latency; |
d9ff4187 AV |
212 | |
213 | struct list_head cic_list; | |
1da177e4 | 214 | |
6118b70b JA |
215 | /* |
216 | * Fallback dummy cfqq for extreme OOM conditions | |
217 | */ | |
218 | struct cfq_queue oom_cfqq; | |
365722bb VG |
219 | |
220 | unsigned long last_end_sync_rq; | |
1da177e4 LT |
221 | }; |
222 | ||
c0324a02 CZ |
223 | static struct cfq_rb_root *service_tree_for(enum wl_prio_t prio, |
224 | struct cfq_data *cfqd) | |
225 | { | |
226 | if (prio == IDLE_WORKLOAD) | |
227 | return &cfqd->service_tree_idle; | |
228 | ||
229 | return &cfqd->service_trees[prio]; | |
230 | } | |
231 | ||
3b18152c | 232 | enum cfqq_state_flags { |
b0b8d749 JA |
233 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
234 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 235 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 | 236 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ |
b0b8d749 JA |
237 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
238 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
239 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 240 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 241 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
b3b6d040 | 242 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ |
3b18152c JA |
243 | }; |
244 | ||
245 | #define CFQ_CFQQ_FNS(name) \ | |
246 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
247 | { \ | |
fe094d98 | 248 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
249 | } \ |
250 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
251 | { \ | |
fe094d98 | 252 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
253 | } \ |
254 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
255 | { \ | |
fe094d98 | 256 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
257 | } |
258 | ||
259 | CFQ_CFQQ_FNS(on_rr); | |
260 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 261 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c | 262 | CFQ_CFQQ_FNS(must_alloc_slice); |
3b18152c JA |
263 | CFQ_CFQQ_FNS(fifo_expire); |
264 | CFQ_CFQQ_FNS(idle_window); | |
265 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 266 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 267 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 268 | CFQ_CFQQ_FNS(coop); |
3b18152c JA |
269 | #undef CFQ_CFQQ_FNS |
270 | ||
7b679138 JA |
271 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
272 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
273 | #define cfq_log(cfqd, fmt, args...) \ | |
274 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
275 | ||
c0324a02 CZ |
276 | static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) |
277 | { | |
278 | if (cfq_class_idle(cfqq)) | |
279 | return IDLE_WORKLOAD; | |
280 | if (cfq_class_rt(cfqq)) | |
281 | return RT_WORKLOAD; | |
282 | return BE_WORKLOAD; | |
283 | } | |
284 | ||
285 | static inline int cfq_busy_queues_wl(enum wl_prio_t wl, struct cfq_data *cfqd) | |
286 | { | |
287 | if (wl == IDLE_WORKLOAD) | |
288 | return cfqd->service_tree_idle.count; | |
289 | ||
290 | return cfqd->service_trees[wl].count; | |
291 | } | |
292 | ||
165125e1 | 293 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
a6151c3a | 294 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, |
fd0928df | 295 | struct io_context *, gfp_t); |
4ac845a2 | 296 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
297 | struct io_context *); |
298 | ||
5ad531db JA |
299 | static inline int rq_in_driver(struct cfq_data *cfqd) |
300 | { | |
301 | return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1]; | |
302 | } | |
303 | ||
91fac317 | 304 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, |
a6151c3a | 305 | bool is_sync) |
91fac317 | 306 | { |
a6151c3a | 307 | return cic->cfqq[is_sync]; |
91fac317 VT |
308 | } |
309 | ||
310 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 311 | struct cfq_queue *cfqq, bool is_sync) |
91fac317 | 312 | { |
a6151c3a | 313 | cic->cfqq[is_sync] = cfqq; |
91fac317 VT |
314 | } |
315 | ||
316 | /* | |
317 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
318 | * set (in which case it could also be direct WRITE). | |
319 | */ | |
a6151c3a | 320 | static inline bool cfq_bio_sync(struct bio *bio) |
91fac317 | 321 | { |
a6151c3a | 322 | return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO); |
91fac317 | 323 | } |
1da177e4 | 324 | |
99f95e52 AM |
325 | /* |
326 | * scheduler run of queue, if there are requests pending and no one in the | |
327 | * driver that will restart queueing | |
328 | */ | |
23e018a1 | 329 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) |
99f95e52 | 330 | { |
7b679138 JA |
331 | if (cfqd->busy_queues) { |
332 | cfq_log(cfqd, "schedule dispatch"); | |
23e018a1 | 333 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 334 | } |
99f95e52 AM |
335 | } |
336 | ||
165125e1 | 337 | static int cfq_queue_empty(struct request_queue *q) |
99f95e52 AM |
338 | { |
339 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
340 | ||
b4878f24 | 341 | return !cfqd->busy_queues; |
99f95e52 AM |
342 | } |
343 | ||
44f7c160 JA |
344 | /* |
345 | * Scale schedule slice based on io priority. Use the sync time slice only | |
346 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
347 | * io only, should not get full sync slice length. | |
348 | */ | |
a6151c3a | 349 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, |
d9e7620e | 350 | unsigned short prio) |
44f7c160 | 351 | { |
d9e7620e | 352 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 353 | |
d9e7620e JA |
354 | WARN_ON(prio >= IOPRIO_BE_NR); |
355 | ||
356 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
357 | } | |
44f7c160 | 358 | |
d9e7620e JA |
359 | static inline int |
360 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
361 | { | |
362 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
363 | } |
364 | ||
5db5d642 CZ |
365 | /* |
366 | * get averaged number of queues of RT/BE priority. | |
367 | * average is updated, with a formula that gives more weight to higher numbers, | |
368 | * to quickly follows sudden increases and decrease slowly | |
369 | */ | |
370 | ||
371 | static inline unsigned | |
372 | cfq_get_avg_queues(struct cfq_data *cfqd, bool rt) { | |
373 | unsigned min_q, max_q; | |
374 | unsigned mult = cfq_hist_divisor - 1; | |
375 | unsigned round = cfq_hist_divisor / 2; | |
c0324a02 | 376 | unsigned busy = cfq_busy_queues_wl(rt, cfqd); |
5db5d642 CZ |
377 | |
378 | min_q = min(cfqd->busy_queues_avg[rt], busy); | |
379 | max_q = max(cfqd->busy_queues_avg[rt], busy); | |
380 | cfqd->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
381 | cfq_hist_divisor; | |
382 | return cfqd->busy_queues_avg[rt]; | |
383 | } | |
384 | ||
44f7c160 JA |
385 | static inline void |
386 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
387 | { | |
5db5d642 CZ |
388 | unsigned slice = cfq_prio_to_slice(cfqd, cfqq); |
389 | if (cfqd->cfq_latency) { | |
390 | /* interested queues (we consider only the ones with the same | |
391 | * priority class) */ | |
392 | unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq)); | |
393 | unsigned sync_slice = cfqd->cfq_slice[1]; | |
394 | unsigned expect_latency = sync_slice * iq; | |
395 | if (expect_latency > cfq_target_latency) { | |
396 | unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; | |
397 | /* scale low_slice according to IO priority | |
398 | * and sync vs async */ | |
399 | unsigned low_slice = | |
400 | min(slice, base_low_slice * slice / sync_slice); | |
401 | /* the adapted slice value is scaled to fit all iqs | |
402 | * into the target latency */ | |
403 | slice = max(slice * cfq_target_latency / expect_latency, | |
404 | low_slice); | |
405 | } | |
406 | } | |
407 | cfqq->slice_end = jiffies + slice; | |
7b679138 | 408 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
409 | } |
410 | ||
411 | /* | |
412 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
413 | * isn't valid until the first request from the dispatch is activated | |
414 | * and the slice time set. | |
415 | */ | |
a6151c3a | 416 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) |
44f7c160 JA |
417 | { |
418 | if (cfq_cfqq_slice_new(cfqq)) | |
419 | return 0; | |
420 | if (time_before(jiffies, cfqq->slice_end)) | |
421 | return 0; | |
422 | ||
423 | return 1; | |
424 | } | |
425 | ||
1da177e4 | 426 | /* |
5e705374 | 427 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 428 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 429 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 430 | */ |
5e705374 JA |
431 | static struct request * |
432 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2) | |
1da177e4 LT |
433 | { |
434 | sector_t last, s1, s2, d1 = 0, d2 = 0; | |
1da177e4 | 435 | unsigned long back_max; |
e8a99053 AM |
436 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
437 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
438 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 439 | |
5e705374 JA |
440 | if (rq1 == NULL || rq1 == rq2) |
441 | return rq2; | |
442 | if (rq2 == NULL) | |
443 | return rq1; | |
9c2c38a1 | 444 | |
5e705374 JA |
445 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
446 | return rq1; | |
447 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
448 | return rq2; | |
374f84ac JA |
449 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
450 | return rq1; | |
451 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
452 | return rq2; | |
1da177e4 | 453 | |
83096ebf TH |
454 | s1 = blk_rq_pos(rq1); |
455 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 456 | |
6d048f53 | 457 | last = cfqd->last_position; |
1da177e4 | 458 | |
1da177e4 LT |
459 | /* |
460 | * by definition, 1KiB is 2 sectors | |
461 | */ | |
462 | back_max = cfqd->cfq_back_max * 2; | |
463 | ||
464 | /* | |
465 | * Strict one way elevator _except_ in the case where we allow | |
466 | * short backward seeks which are biased as twice the cost of a | |
467 | * similar forward seek. | |
468 | */ | |
469 | if (s1 >= last) | |
470 | d1 = s1 - last; | |
471 | else if (s1 + back_max >= last) | |
472 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
473 | else | |
e8a99053 | 474 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
475 | |
476 | if (s2 >= last) | |
477 | d2 = s2 - last; | |
478 | else if (s2 + back_max >= last) | |
479 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
480 | else | |
e8a99053 | 481 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
482 | |
483 | /* Found required data */ | |
e8a99053 AM |
484 | |
485 | /* | |
486 | * By doing switch() on the bit mask "wrap" we avoid having to | |
487 | * check two variables for all permutations: --> faster! | |
488 | */ | |
489 | switch (wrap) { | |
5e705374 | 490 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 491 | if (d1 < d2) |
5e705374 | 492 | return rq1; |
e8a99053 | 493 | else if (d2 < d1) |
5e705374 | 494 | return rq2; |
e8a99053 AM |
495 | else { |
496 | if (s1 >= s2) | |
5e705374 | 497 | return rq1; |
e8a99053 | 498 | else |
5e705374 | 499 | return rq2; |
e8a99053 | 500 | } |
1da177e4 | 501 | |
e8a99053 | 502 | case CFQ_RQ2_WRAP: |
5e705374 | 503 | return rq1; |
e8a99053 | 504 | case CFQ_RQ1_WRAP: |
5e705374 JA |
505 | return rq2; |
506 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
507 | default: |
508 | /* | |
509 | * Since both rqs are wrapped, | |
510 | * start with the one that's further behind head | |
511 | * (--> only *one* back seek required), | |
512 | * since back seek takes more time than forward. | |
513 | */ | |
514 | if (s1 <= s2) | |
5e705374 | 515 | return rq1; |
1da177e4 | 516 | else |
5e705374 | 517 | return rq2; |
1da177e4 LT |
518 | } |
519 | } | |
520 | ||
498d3aa2 JA |
521 | /* |
522 | * The below is leftmost cache rbtree addon | |
523 | */ | |
0871714e | 524 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 JA |
525 | { |
526 | if (!root->left) | |
527 | root->left = rb_first(&root->rb); | |
528 | ||
0871714e JA |
529 | if (root->left) |
530 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
531 | ||
532 | return NULL; | |
cc09e299 JA |
533 | } |
534 | ||
a36e71f9 JA |
535 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
536 | { | |
537 | rb_erase(n, root); | |
538 | RB_CLEAR_NODE(n); | |
539 | } | |
540 | ||
cc09e299 JA |
541 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
542 | { | |
543 | if (root->left == n) | |
544 | root->left = NULL; | |
a36e71f9 | 545 | rb_erase_init(n, &root->rb); |
aa6f6a3d | 546 | --root->count; |
cc09e299 JA |
547 | } |
548 | ||
1da177e4 LT |
549 | /* |
550 | * would be nice to take fifo expire time into account as well | |
551 | */ | |
5e705374 JA |
552 | static struct request * |
553 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
554 | struct request *last) | |
1da177e4 | 555 | { |
21183b07 JA |
556 | struct rb_node *rbnext = rb_next(&last->rb_node); |
557 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 558 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 559 | |
21183b07 | 560 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
561 | |
562 | if (rbprev) | |
5e705374 | 563 | prev = rb_entry_rq(rbprev); |
1da177e4 | 564 | |
21183b07 | 565 | if (rbnext) |
5e705374 | 566 | next = rb_entry_rq(rbnext); |
21183b07 JA |
567 | else { |
568 | rbnext = rb_first(&cfqq->sort_list); | |
569 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 570 | next = rb_entry_rq(rbnext); |
21183b07 | 571 | } |
1da177e4 | 572 | |
21183b07 | 573 | return cfq_choose_req(cfqd, next, prev); |
1da177e4 LT |
574 | } |
575 | ||
d9e7620e JA |
576 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
577 | struct cfq_queue *cfqq) | |
1da177e4 | 578 | { |
d9e7620e JA |
579 | /* |
580 | * just an approximation, should be ok. | |
581 | */ | |
67e6b49e JA |
582 | return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
583 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); | |
d9e7620e JA |
584 | } |
585 | ||
498d3aa2 | 586 | /* |
c0324a02 | 587 | * The cfqd->service_trees holds all pending cfq_queue's that have |
498d3aa2 JA |
588 | * requests waiting to be processed. It is sorted in the order that |
589 | * we will service the queues. | |
590 | */ | |
a36e71f9 | 591 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 592 | bool add_front) |
d9e7620e | 593 | { |
0871714e JA |
594 | struct rb_node **p, *parent; |
595 | struct cfq_queue *__cfqq; | |
d9e7620e | 596 | unsigned long rb_key; |
c0324a02 | 597 | struct cfq_rb_root *service_tree; |
498d3aa2 | 598 | int left; |
d9e7620e | 599 | |
c0324a02 | 600 | service_tree = service_tree_for(cfqq_prio(cfqq), cfqd); |
0871714e JA |
601 | if (cfq_class_idle(cfqq)) { |
602 | rb_key = CFQ_IDLE_DELAY; | |
aa6f6a3d | 603 | parent = rb_last(&service_tree->rb); |
0871714e JA |
604 | if (parent && parent != &cfqq->rb_node) { |
605 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
606 | rb_key += __cfqq->rb_key; | |
607 | } else | |
608 | rb_key += jiffies; | |
609 | } else if (!add_front) { | |
b9c8946b JA |
610 | /* |
611 | * Get our rb key offset. Subtract any residual slice | |
612 | * value carried from last service. A negative resid | |
613 | * count indicates slice overrun, and this should position | |
614 | * the next service time further away in the tree. | |
615 | */ | |
edd75ffd | 616 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
b9c8946b | 617 | rb_key -= cfqq->slice_resid; |
edd75ffd | 618 | cfqq->slice_resid = 0; |
48e025e6 CZ |
619 | } else { |
620 | rb_key = -HZ; | |
aa6f6a3d | 621 | __cfqq = cfq_rb_first(service_tree); |
48e025e6 CZ |
622 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; |
623 | } | |
1da177e4 | 624 | |
d9e7620e | 625 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
99f9628a | 626 | /* |
d9e7620e | 627 | * same position, nothing more to do |
99f9628a | 628 | */ |
c0324a02 CZ |
629 | if (rb_key == cfqq->rb_key && |
630 | cfqq->service_tree == service_tree) | |
d9e7620e | 631 | return; |
1da177e4 | 632 | |
aa6f6a3d CZ |
633 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); |
634 | cfqq->service_tree = NULL; | |
1da177e4 | 635 | } |
d9e7620e | 636 | |
498d3aa2 | 637 | left = 1; |
0871714e | 638 | parent = NULL; |
aa6f6a3d CZ |
639 | cfqq->service_tree = service_tree; |
640 | p = &service_tree->rb.rb_node; | |
d9e7620e | 641 | while (*p) { |
67060e37 | 642 | struct rb_node **n; |
cc09e299 | 643 | |
d9e7620e JA |
644 | parent = *p; |
645 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
646 | ||
0c534e0a | 647 | /* |
c0324a02 | 648 | * sort by key, that represents service time. |
0c534e0a | 649 | */ |
c0324a02 | 650 | if (time_before(rb_key, __cfqq->rb_key)) |
67060e37 | 651 | n = &(*p)->rb_left; |
c0324a02 | 652 | else { |
67060e37 | 653 | n = &(*p)->rb_right; |
cc09e299 | 654 | left = 0; |
c0324a02 | 655 | } |
67060e37 JA |
656 | |
657 | p = n; | |
d9e7620e JA |
658 | } |
659 | ||
cc09e299 | 660 | if (left) |
aa6f6a3d | 661 | service_tree->left = &cfqq->rb_node; |
cc09e299 | 662 | |
d9e7620e JA |
663 | cfqq->rb_key = rb_key; |
664 | rb_link_node(&cfqq->rb_node, parent, p); | |
aa6f6a3d CZ |
665 | rb_insert_color(&cfqq->rb_node, &service_tree->rb); |
666 | service_tree->count++; | |
1da177e4 LT |
667 | } |
668 | ||
a36e71f9 | 669 | static struct cfq_queue * |
f2d1f0ae JA |
670 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
671 | sector_t sector, struct rb_node **ret_parent, | |
672 | struct rb_node ***rb_link) | |
a36e71f9 | 673 | { |
a36e71f9 JA |
674 | struct rb_node **p, *parent; |
675 | struct cfq_queue *cfqq = NULL; | |
676 | ||
677 | parent = NULL; | |
678 | p = &root->rb_node; | |
679 | while (*p) { | |
680 | struct rb_node **n; | |
681 | ||
682 | parent = *p; | |
683 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
684 | ||
685 | /* | |
686 | * Sort strictly based on sector. Smallest to the left, | |
687 | * largest to the right. | |
688 | */ | |
2e46e8b2 | 689 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 690 | n = &(*p)->rb_right; |
2e46e8b2 | 691 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
692 | n = &(*p)->rb_left; |
693 | else | |
694 | break; | |
695 | p = n; | |
3ac6c9f8 | 696 | cfqq = NULL; |
a36e71f9 JA |
697 | } |
698 | ||
699 | *ret_parent = parent; | |
700 | if (rb_link) | |
701 | *rb_link = p; | |
3ac6c9f8 | 702 | return cfqq; |
a36e71f9 JA |
703 | } |
704 | ||
705 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
706 | { | |
a36e71f9 JA |
707 | struct rb_node **p, *parent; |
708 | struct cfq_queue *__cfqq; | |
709 | ||
f2d1f0ae JA |
710 | if (cfqq->p_root) { |
711 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
712 | cfqq->p_root = NULL; | |
713 | } | |
a36e71f9 JA |
714 | |
715 | if (cfq_class_idle(cfqq)) | |
716 | return; | |
717 | if (!cfqq->next_rq) | |
718 | return; | |
719 | ||
f2d1f0ae | 720 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
721 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
722 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
723 | if (!__cfqq) { |
724 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
725 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
726 | } else | |
727 | cfqq->p_root = NULL; | |
a36e71f9 JA |
728 | } |
729 | ||
498d3aa2 JA |
730 | /* |
731 | * Update cfqq's position in the service tree. | |
732 | */ | |
edd75ffd | 733 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 734 | { |
6d048f53 JA |
735 | /* |
736 | * Resorting requires the cfqq to be on the RR list already. | |
737 | */ | |
a36e71f9 | 738 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 739 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
740 | cfq_prio_tree_add(cfqd, cfqq); |
741 | } | |
6d048f53 JA |
742 | } |
743 | ||
1da177e4 LT |
744 | /* |
745 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 746 | * the pending list according to last request service |
1da177e4 | 747 | */ |
febffd61 | 748 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 749 | { |
7b679138 | 750 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
751 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
752 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 | 753 | cfqd->busy_queues++; |
c0324a02 | 754 | |
edd75ffd | 755 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
756 | } |
757 | ||
498d3aa2 JA |
758 | /* |
759 | * Called when the cfqq no longer has requests pending, remove it from | |
760 | * the service tree. | |
761 | */ | |
febffd61 | 762 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 763 | { |
7b679138 | 764 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
765 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
766 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 767 | |
aa6f6a3d CZ |
768 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
769 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
770 | cfqq->service_tree = NULL; | |
771 | } | |
f2d1f0ae JA |
772 | if (cfqq->p_root) { |
773 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
774 | cfqq->p_root = NULL; | |
775 | } | |
d9e7620e | 776 | |
1da177e4 LT |
777 | BUG_ON(!cfqd->busy_queues); |
778 | cfqd->busy_queues--; | |
779 | } | |
780 | ||
781 | /* | |
782 | * rb tree support functions | |
783 | */ | |
febffd61 | 784 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 785 | { |
5e705374 | 786 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 787 | struct cfq_data *cfqd = cfqq->cfqd; |
5e705374 | 788 | const int sync = rq_is_sync(rq); |
1da177e4 | 789 | |
b4878f24 JA |
790 | BUG_ON(!cfqq->queued[sync]); |
791 | cfqq->queued[sync]--; | |
1da177e4 | 792 | |
5e705374 | 793 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 794 | |
dd67d051 | 795 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
b4878f24 | 796 | cfq_del_cfqq_rr(cfqd, cfqq); |
1da177e4 LT |
797 | } |
798 | ||
5e705374 | 799 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 800 | { |
5e705374 | 801 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 802 | struct cfq_data *cfqd = cfqq->cfqd; |
a36e71f9 | 803 | struct request *__alias, *prev; |
1da177e4 | 804 | |
5380a101 | 805 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
806 | |
807 | /* | |
808 | * looks a little odd, but the first insert might return an alias. | |
809 | * if that happens, put the alias on the dispatch list | |
810 | */ | |
21183b07 | 811 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 812 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
813 | |
814 | if (!cfq_cfqq_on_rr(cfqq)) | |
815 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
816 | |
817 | /* | |
818 | * check if this request is a better next-serve candidate | |
819 | */ | |
a36e71f9 | 820 | prev = cfqq->next_rq; |
5044eed4 | 821 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq); |
a36e71f9 JA |
822 | |
823 | /* | |
824 | * adjust priority tree position, if ->next_rq changes | |
825 | */ | |
826 | if (prev != cfqq->next_rq) | |
827 | cfq_prio_tree_add(cfqd, cfqq); | |
828 | ||
5044eed4 | 829 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
830 | } |
831 | ||
febffd61 | 832 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 833 | { |
5380a101 JA |
834 | elv_rb_del(&cfqq->sort_list, rq); |
835 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 836 | cfq_add_rq_rb(rq); |
1da177e4 LT |
837 | } |
838 | ||
206dc69b JA |
839 | static struct request * |
840 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 841 | { |
206dc69b | 842 | struct task_struct *tsk = current; |
91fac317 | 843 | struct cfq_io_context *cic; |
206dc69b | 844 | struct cfq_queue *cfqq; |
1da177e4 | 845 | |
4ac845a2 | 846 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
847 | if (!cic) |
848 | return NULL; | |
849 | ||
850 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
851 | if (cfqq) { |
852 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
853 | ||
21183b07 | 854 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 855 | } |
1da177e4 | 856 | |
1da177e4 LT |
857 | return NULL; |
858 | } | |
859 | ||
165125e1 | 860 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 861 | { |
22e2c507 | 862 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 863 | |
5ad531db | 864 | cfqd->rq_in_driver[rq_is_sync(rq)]++; |
7b679138 | 865 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
5ad531db | 866 | rq_in_driver(cfqd)); |
25776e35 | 867 | |
5b93629b | 868 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
869 | } |
870 | ||
165125e1 | 871 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 872 | { |
b4878f24 | 873 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5ad531db | 874 | const int sync = rq_is_sync(rq); |
b4878f24 | 875 | |
5ad531db JA |
876 | WARN_ON(!cfqd->rq_in_driver[sync]); |
877 | cfqd->rq_in_driver[sync]--; | |
7b679138 | 878 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
5ad531db | 879 | rq_in_driver(cfqd)); |
1da177e4 LT |
880 | } |
881 | ||
b4878f24 | 882 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 883 | { |
5e705374 | 884 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 885 | |
5e705374 JA |
886 | if (cfqq->next_rq == rq) |
887 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 888 | |
b4878f24 | 889 | list_del_init(&rq->queuelist); |
5e705374 | 890 | cfq_del_rq_rb(rq); |
374f84ac | 891 | |
45333d5a | 892 | cfqq->cfqd->rq_queued--; |
374f84ac JA |
893 | if (rq_is_meta(rq)) { |
894 | WARN_ON(!cfqq->meta_pending); | |
895 | cfqq->meta_pending--; | |
896 | } | |
1da177e4 LT |
897 | } |
898 | ||
165125e1 JA |
899 | static int cfq_merge(struct request_queue *q, struct request **req, |
900 | struct bio *bio) | |
1da177e4 LT |
901 | { |
902 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
903 | struct request *__rq; | |
1da177e4 | 904 | |
206dc69b | 905 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 906 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
907 | *req = __rq; |
908 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
909 | } |
910 | ||
911 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
912 | } |
913 | ||
165125e1 | 914 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 915 | int type) |
1da177e4 | 916 | { |
21183b07 | 917 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 918 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 919 | |
5e705374 | 920 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 921 | } |
1da177e4 LT |
922 | } |
923 | ||
924 | static void | |
165125e1 | 925 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
926 | struct request *next) |
927 | { | |
22e2c507 JA |
928 | /* |
929 | * reposition in fifo if next is older than rq | |
930 | */ | |
931 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
30996f40 | 932 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
22e2c507 | 933 | list_move(&rq->queuelist, &next->queuelist); |
30996f40 JA |
934 | rq_set_fifo_time(rq, rq_fifo_time(next)); |
935 | } | |
22e2c507 | 936 | |
b4878f24 | 937 | cfq_remove_request(next); |
22e2c507 JA |
938 | } |
939 | ||
165125e1 | 940 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
941 | struct bio *bio) |
942 | { | |
943 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 944 | struct cfq_io_context *cic; |
da775265 | 945 | struct cfq_queue *cfqq; |
da775265 JA |
946 | |
947 | /* | |
ec8acb69 | 948 | * Disallow merge of a sync bio into an async request. |
da775265 | 949 | */ |
91fac317 | 950 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
a6151c3a | 951 | return false; |
da775265 JA |
952 | |
953 | /* | |
719d3402 JA |
954 | * Lookup the cfqq that this bio will be queued with. Allow |
955 | * merge only if rq is queued there. | |
da775265 | 956 | */ |
4ac845a2 | 957 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 | 958 | if (!cic) |
a6151c3a | 959 | return false; |
719d3402 | 960 | |
91fac317 | 961 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
a6151c3a | 962 | return cfqq == RQ_CFQQ(rq); |
da775265 JA |
963 | } |
964 | ||
febffd61 JA |
965 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
966 | struct cfq_queue *cfqq) | |
22e2c507 JA |
967 | { |
968 | if (cfqq) { | |
7b679138 | 969 | cfq_log_cfqq(cfqd, cfqq, "set_active"); |
22e2c507 | 970 | cfqq->slice_end = 0; |
2f5cb738 JA |
971 | cfqq->slice_dispatch = 0; |
972 | ||
2f5cb738 | 973 | cfq_clear_cfqq_wait_request(cfqq); |
b029195d | 974 | cfq_clear_cfqq_must_dispatch(cfqq); |
3b18152c JA |
975 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
976 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 977 | cfq_mark_cfqq_slice_new(cfqq); |
2f5cb738 JA |
978 | |
979 | del_timer(&cfqd->idle_slice_timer); | |
22e2c507 JA |
980 | } |
981 | ||
982 | cfqd->active_queue = cfqq; | |
983 | } | |
984 | ||
7b14e3b5 JA |
985 | /* |
986 | * current cfqq expired its slice (or was too idle), select new one | |
987 | */ | |
988 | static void | |
989 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
a6151c3a | 990 | bool timed_out) |
7b14e3b5 | 991 | { |
7b679138 JA |
992 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
993 | ||
7b14e3b5 JA |
994 | if (cfq_cfqq_wait_request(cfqq)) |
995 | del_timer(&cfqd->idle_slice_timer); | |
996 | ||
7b14e3b5 JA |
997 | cfq_clear_cfqq_wait_request(cfqq); |
998 | ||
999 | /* | |
6084cdda | 1000 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 1001 | */ |
7b679138 | 1002 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) { |
c5b680f3 | 1003 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b679138 JA |
1004 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
1005 | } | |
7b14e3b5 | 1006 | |
edd75ffd | 1007 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
1008 | |
1009 | if (cfqq == cfqd->active_queue) | |
1010 | cfqd->active_queue = NULL; | |
1011 | ||
1012 | if (cfqd->active_cic) { | |
1013 | put_io_context(cfqd->active_cic->ioc); | |
1014 | cfqd->active_cic = NULL; | |
1015 | } | |
7b14e3b5 JA |
1016 | } |
1017 | ||
a6151c3a | 1018 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) |
7b14e3b5 JA |
1019 | { |
1020 | struct cfq_queue *cfqq = cfqd->active_queue; | |
1021 | ||
1022 | if (cfqq) | |
6084cdda | 1023 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
1024 | } |
1025 | ||
498d3aa2 JA |
1026 | /* |
1027 | * Get next queue for service. Unless we have a queue preemption, | |
1028 | * we'll simply select the first cfqq in the service tree. | |
1029 | */ | |
6d048f53 | 1030 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 1031 | { |
c0324a02 CZ |
1032 | struct cfq_rb_root *service_tree = |
1033 | service_tree_for(cfqd->serving_prio, cfqd); | |
d9e7620e | 1034 | |
c0324a02 CZ |
1035 | if (RB_EMPTY_ROOT(&service_tree->rb)) |
1036 | return NULL; | |
1037 | return cfq_rb_first(service_tree); | |
6d048f53 JA |
1038 | } |
1039 | ||
498d3aa2 JA |
1040 | /* |
1041 | * Get and set a new active queue for service. | |
1042 | */ | |
a36e71f9 JA |
1043 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
1044 | struct cfq_queue *cfqq) | |
6d048f53 | 1045 | { |
b3b6d040 | 1046 | if (!cfqq) |
a36e71f9 | 1047 | cfqq = cfq_get_next_queue(cfqd); |
6d048f53 | 1048 | |
22e2c507 | 1049 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 1050 | return cfqq; |
22e2c507 JA |
1051 | } |
1052 | ||
d9e7620e JA |
1053 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
1054 | struct request *rq) | |
1055 | { | |
83096ebf TH |
1056 | if (blk_rq_pos(rq) >= cfqd->last_position) |
1057 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 1058 | else |
83096ebf | 1059 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
1060 | } |
1061 | ||
b2c18e1e JM |
1062 | #define CFQQ_SEEK_THR 8 * 1024 |
1063 | #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR) | |
04dc6e71 | 1064 | |
b2c18e1e JM |
1065 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
1066 | struct request *rq) | |
6d048f53 | 1067 | { |
b2c18e1e | 1068 | sector_t sdist = cfqq->seek_mean; |
6d048f53 | 1069 | |
b2c18e1e JM |
1070 | if (!sample_valid(cfqq->seek_samples)) |
1071 | sdist = CFQQ_SEEK_THR; | |
6d048f53 | 1072 | |
04dc6e71 | 1073 | return cfq_dist_from_last(cfqd, rq) <= sdist; |
6d048f53 JA |
1074 | } |
1075 | ||
a36e71f9 JA |
1076 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
1077 | struct cfq_queue *cur_cfqq) | |
1078 | { | |
f2d1f0ae | 1079 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
1080 | struct rb_node *parent, *node; |
1081 | struct cfq_queue *__cfqq; | |
1082 | sector_t sector = cfqd->last_position; | |
1083 | ||
1084 | if (RB_EMPTY_ROOT(root)) | |
1085 | return NULL; | |
1086 | ||
1087 | /* | |
1088 | * First, if we find a request starting at the end of the last | |
1089 | * request, choose it. | |
1090 | */ | |
f2d1f0ae | 1091 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
1092 | if (__cfqq) |
1093 | return __cfqq; | |
1094 | ||
1095 | /* | |
1096 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1097 | * will contain the closest sector. | |
1098 | */ | |
1099 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
b2c18e1e | 1100 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1101 | return __cfqq; |
1102 | ||
2e46e8b2 | 1103 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1104 | node = rb_next(&__cfqq->p_node); |
1105 | else | |
1106 | node = rb_prev(&__cfqq->p_node); | |
1107 | if (!node) | |
1108 | return NULL; | |
1109 | ||
1110 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
b2c18e1e | 1111 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1112 | return __cfqq; |
1113 | ||
1114 | return NULL; | |
1115 | } | |
1116 | ||
1117 | /* | |
1118 | * cfqd - obvious | |
1119 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1120 | * closely cooperating with itself. | |
1121 | * | |
1122 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1123 | * one request, and that cfqd->last_position reflects a position on the disk | |
1124 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1125 | * assumption. | |
1126 | */ | |
1127 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
b3b6d040 | 1128 | struct cfq_queue *cur_cfqq) |
6d048f53 | 1129 | { |
a36e71f9 JA |
1130 | struct cfq_queue *cfqq; |
1131 | ||
e6c5bc73 JM |
1132 | if (!cfq_cfqq_sync(cur_cfqq)) |
1133 | return NULL; | |
1134 | if (CFQQ_SEEKY(cur_cfqq)) | |
1135 | return NULL; | |
1136 | ||
6d048f53 | 1137 | /* |
d9e7620e JA |
1138 | * We should notice if some of the queues are cooperating, eg |
1139 | * working closely on the same area of the disk. In that case, | |
1140 | * we can group them together and don't waste time idling. | |
6d048f53 | 1141 | */ |
a36e71f9 JA |
1142 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1143 | if (!cfqq) | |
1144 | return NULL; | |
1145 | ||
df5fe3e8 JM |
1146 | /* |
1147 | * It only makes sense to merge sync queues. | |
1148 | */ | |
1149 | if (!cfq_cfqq_sync(cfqq)) | |
1150 | return NULL; | |
e6c5bc73 JM |
1151 | if (CFQQ_SEEKY(cfqq)) |
1152 | return NULL; | |
df5fe3e8 | 1153 | |
c0324a02 CZ |
1154 | /* |
1155 | * Do not merge queues of different priority classes | |
1156 | */ | |
1157 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
1158 | return NULL; | |
1159 | ||
a36e71f9 | 1160 | return cfqq; |
6d048f53 JA |
1161 | } |
1162 | ||
a6d44e98 CZ |
1163 | /* |
1164 | * Determine whether we should enforce idle window for this queue. | |
1165 | */ | |
1166 | ||
1167 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1168 | { | |
1169 | enum wl_prio_t prio = cfqq_prio(cfqq); | |
1170 | struct cfq_rb_root *service_tree; | |
1171 | ||
1172 | /* We never do for idle class queues. */ | |
1173 | if (prio == IDLE_WORKLOAD) | |
1174 | return false; | |
1175 | ||
1176 | /* We do for queues that were marked with idle window flag. */ | |
1177 | if (cfq_cfqq_idle_window(cfqq)) | |
1178 | return true; | |
1179 | ||
1180 | /* | |
1181 | * Otherwise, we do only if they are the last ones | |
1182 | * in their service tree. | |
1183 | */ | |
1184 | service_tree = service_tree_for(prio, cfqd); | |
1185 | if (service_tree->count == 0) | |
1186 | return true; | |
1187 | ||
1188 | return (service_tree->count == 1 && cfq_rb_first(service_tree) == cfqq); | |
1189 | } | |
1190 | ||
6d048f53 | 1191 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1192 | { |
1792669c | 1193 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1194 | struct cfq_io_context *cic; |
7b14e3b5 JA |
1195 | unsigned long sl; |
1196 | ||
a68bbddb | 1197 | /* |
f7d7b7a7 JA |
1198 | * SSD device without seek penalty, disable idling. But only do so |
1199 | * for devices that support queuing, otherwise we still have a problem | |
1200 | * with sync vs async workloads. | |
a68bbddb | 1201 | */ |
f7d7b7a7 | 1202 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
1203 | return; |
1204 | ||
dd67d051 | 1205 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 1206 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
1207 | |
1208 | /* | |
1209 | * idle is disabled, either manually or by past process history | |
1210 | */ | |
a6d44e98 | 1211 | if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq)) |
6d048f53 JA |
1212 | return; |
1213 | ||
7b679138 JA |
1214 | /* |
1215 | * still requests with the driver, don't idle | |
1216 | */ | |
5ad531db | 1217 | if (rq_in_driver(cfqd)) |
7b679138 JA |
1218 | return; |
1219 | ||
22e2c507 JA |
1220 | /* |
1221 | * task has exited, don't wait | |
1222 | */ | |
206dc69b | 1223 | cic = cfqd->active_cic; |
66dac98e | 1224 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
1225 | return; |
1226 | ||
355b659c CZ |
1227 | /* |
1228 | * If our average think time is larger than the remaining time | |
1229 | * slice, then don't idle. This avoids overrunning the allotted | |
1230 | * time slice. | |
1231 | */ | |
1232 | if (sample_valid(cic->ttime_samples) && | |
1233 | (cfqq->slice_end - jiffies < cic->ttime_mean)) | |
1234 | return; | |
1235 | ||
3b18152c | 1236 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 1237 | |
206dc69b JA |
1238 | /* |
1239 | * we don't want to idle for seeks, but we do want to allow | |
1240 | * fair distribution of slice time for a process doing back-to-back | |
1241 | * seeks. so allow a little bit of time for him to submit a new rq | |
1242 | */ | |
6d048f53 | 1243 | sl = cfqd->cfq_slice_idle; |
b2c18e1e | 1244 | if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)) |
d9e7620e | 1245 | sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT)); |
206dc69b | 1246 | |
7b14e3b5 | 1247 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
9481ffdc | 1248 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); |
1da177e4 LT |
1249 | } |
1250 | ||
498d3aa2 JA |
1251 | /* |
1252 | * Move request from internal lists to the request queue dispatch list. | |
1253 | */ | |
165125e1 | 1254 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 1255 | { |
3ed9a296 | 1256 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1257 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 1258 | |
7b679138 JA |
1259 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
1260 | ||
06d21886 | 1261 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); |
5380a101 | 1262 | cfq_remove_request(rq); |
6d048f53 | 1263 | cfqq->dispatched++; |
5380a101 | 1264 | elv_dispatch_sort(q, rq); |
3ed9a296 JA |
1265 | |
1266 | if (cfq_cfqq_sync(cfqq)) | |
1267 | cfqd->sync_flight++; | |
1da177e4 LT |
1268 | } |
1269 | ||
1270 | /* | |
1271 | * return expired entry, or NULL to just start from scratch in rbtree | |
1272 | */ | |
febffd61 | 1273 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 | 1274 | { |
30996f40 | 1275 | struct request *rq = NULL; |
1da177e4 | 1276 | |
3b18152c | 1277 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 1278 | return NULL; |
cb887411 JA |
1279 | |
1280 | cfq_mark_cfqq_fifo_expire(cfqq); | |
1281 | ||
89850f7e JA |
1282 | if (list_empty(&cfqq->fifo)) |
1283 | return NULL; | |
1da177e4 | 1284 | |
89850f7e | 1285 | rq = rq_entry_fifo(cfqq->fifo.next); |
30996f40 | 1286 | if (time_before(jiffies, rq_fifo_time(rq))) |
7b679138 | 1287 | rq = NULL; |
1da177e4 | 1288 | |
30996f40 | 1289 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 1290 | return rq; |
1da177e4 LT |
1291 | } |
1292 | ||
22e2c507 JA |
1293 | static inline int |
1294 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1295 | { | |
1296 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 1297 | |
22e2c507 | 1298 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 1299 | |
22e2c507 | 1300 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
1301 | } |
1302 | ||
df5fe3e8 JM |
1303 | /* |
1304 | * Must be called with the queue_lock held. | |
1305 | */ | |
1306 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
1307 | { | |
1308 | int process_refs, io_refs; | |
1309 | ||
1310 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
1311 | process_refs = atomic_read(&cfqq->ref) - io_refs; | |
1312 | BUG_ON(process_refs < 0); | |
1313 | return process_refs; | |
1314 | } | |
1315 | ||
1316 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
1317 | { | |
e6c5bc73 | 1318 | int process_refs, new_process_refs; |
df5fe3e8 JM |
1319 | struct cfq_queue *__cfqq; |
1320 | ||
1321 | /* Avoid a circular list and skip interim queue merges */ | |
1322 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
1323 | if (__cfqq == cfqq) | |
1324 | return; | |
1325 | new_cfqq = __cfqq; | |
1326 | } | |
1327 | ||
1328 | process_refs = cfqq_process_refs(cfqq); | |
1329 | /* | |
1330 | * If the process for the cfqq has gone away, there is no | |
1331 | * sense in merging the queues. | |
1332 | */ | |
1333 | if (process_refs == 0) | |
1334 | return; | |
1335 | ||
e6c5bc73 JM |
1336 | /* |
1337 | * Merge in the direction of the lesser amount of work. | |
1338 | */ | |
1339 | new_process_refs = cfqq_process_refs(new_cfqq); | |
1340 | if (new_process_refs >= process_refs) { | |
1341 | cfqq->new_cfqq = new_cfqq; | |
1342 | atomic_add(process_refs, &new_cfqq->ref); | |
1343 | } else { | |
1344 | new_cfqq->new_cfqq = cfqq; | |
1345 | atomic_add(new_process_refs, &cfqq->ref); | |
1346 | } | |
df5fe3e8 JM |
1347 | } |
1348 | ||
22e2c507 | 1349 | /* |
498d3aa2 JA |
1350 | * Select a queue for service. If we have a current active queue, |
1351 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 1352 | */ |
1b5ed5e1 | 1353 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 1354 | { |
a36e71f9 | 1355 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 1356 | |
22e2c507 JA |
1357 | cfqq = cfqd->active_queue; |
1358 | if (!cfqq) | |
1359 | goto new_queue; | |
1da177e4 | 1360 | |
22e2c507 | 1361 | /* |
6d048f53 | 1362 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 1363 | */ |
b029195d | 1364 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) |
3b18152c | 1365 | goto expire; |
1da177e4 | 1366 | |
22e2c507 | 1367 | /* |
6d048f53 JA |
1368 | * The active queue has requests and isn't expired, allow it to |
1369 | * dispatch. | |
22e2c507 | 1370 | */ |
dd67d051 | 1371 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 1372 | goto keep_queue; |
6d048f53 | 1373 | |
a36e71f9 JA |
1374 | /* |
1375 | * If another queue has a request waiting within our mean seek | |
1376 | * distance, let it run. The expire code will check for close | |
1377 | * cooperators and put the close queue at the front of the service | |
df5fe3e8 | 1378 | * tree. If possible, merge the expiring queue with the new cfqq. |
a36e71f9 | 1379 | */ |
b3b6d040 | 1380 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); |
df5fe3e8 JM |
1381 | if (new_cfqq) { |
1382 | if (!cfqq->new_cfqq) | |
1383 | cfq_setup_merge(cfqq, new_cfqq); | |
a36e71f9 | 1384 | goto expire; |
df5fe3e8 | 1385 | } |
a36e71f9 | 1386 | |
6d048f53 JA |
1387 | /* |
1388 | * No requests pending. If the active queue still has requests in | |
1389 | * flight or is idling for a new request, allow either of these | |
1390 | * conditions to happen (or time out) before selecting a new queue. | |
1391 | */ | |
cc197479 | 1392 | if (timer_pending(&cfqd->idle_slice_timer) || |
a6d44e98 | 1393 | (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) { |
caaa5f9f JA |
1394 | cfqq = NULL; |
1395 | goto keep_queue; | |
22e2c507 JA |
1396 | } |
1397 | ||
3b18152c | 1398 | expire: |
6084cdda | 1399 | cfq_slice_expired(cfqd, 0); |
3b18152c | 1400 | new_queue: |
c0324a02 CZ |
1401 | if (!new_cfqq) { |
1402 | if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd)) | |
1403 | cfqd->serving_prio = RT_WORKLOAD; | |
1404 | else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd)) | |
1405 | cfqd->serving_prio = BE_WORKLOAD; | |
1406 | else | |
1407 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1408 | } | |
a36e71f9 | 1409 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 1410 | keep_queue: |
3b18152c | 1411 | return cfqq; |
22e2c507 JA |
1412 | } |
1413 | ||
febffd61 | 1414 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
1415 | { |
1416 | int dispatched = 0; | |
1417 | ||
1418 | while (cfqq->next_rq) { | |
1419 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
1420 | dispatched++; | |
1421 | } | |
1422 | ||
1423 | BUG_ON(!list_empty(&cfqq->fifo)); | |
1424 | return dispatched; | |
1425 | } | |
1426 | ||
498d3aa2 JA |
1427 | /* |
1428 | * Drain our current requests. Used for barriers and when switching | |
1429 | * io schedulers on-the-fly. | |
1430 | */ | |
d9e7620e | 1431 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 1432 | { |
0871714e | 1433 | struct cfq_queue *cfqq; |
d9e7620e | 1434 | int dispatched = 0; |
c0324a02 CZ |
1435 | int i; |
1436 | for (i = 0; i < 2; ++i) | |
1437 | while ((cfqq = cfq_rb_first(&cfqd->service_trees[i])) != NULL) | |
1438 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); | |
1b5ed5e1 | 1439 | |
c0324a02 | 1440 | while ((cfqq = cfq_rb_first(&cfqd->service_tree_idle)) != NULL) |
d9e7620e | 1441 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); |
1b5ed5e1 | 1442 | |
6084cdda | 1443 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
1444 | |
1445 | BUG_ON(cfqd->busy_queues); | |
1446 | ||
6923715a | 1447 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
1448 | return dispatched; |
1449 | } | |
1450 | ||
0b182d61 | 1451 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
2f5cb738 | 1452 | { |
2f5cb738 | 1453 | unsigned int max_dispatch; |
22e2c507 | 1454 | |
5ad531db JA |
1455 | /* |
1456 | * Drain async requests before we start sync IO | |
1457 | */ | |
a6d44e98 | 1458 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC]) |
0b182d61 | 1459 | return false; |
5ad531db | 1460 | |
2f5cb738 JA |
1461 | /* |
1462 | * If this is an async queue and we have sync IO in flight, let it wait | |
1463 | */ | |
1464 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) | |
0b182d61 | 1465 | return false; |
2f5cb738 JA |
1466 | |
1467 | max_dispatch = cfqd->cfq_quantum; | |
1468 | if (cfq_class_idle(cfqq)) | |
1469 | max_dispatch = 1; | |
b4878f24 | 1470 | |
2f5cb738 JA |
1471 | /* |
1472 | * Does this cfqq already have too much IO in flight? | |
1473 | */ | |
1474 | if (cfqq->dispatched >= max_dispatch) { | |
1475 | /* | |
1476 | * idle queue must always only have a single IO in flight | |
1477 | */ | |
3ed9a296 | 1478 | if (cfq_class_idle(cfqq)) |
0b182d61 | 1479 | return false; |
3ed9a296 | 1480 | |
2f5cb738 JA |
1481 | /* |
1482 | * We have other queues, don't allow more IO from this one | |
1483 | */ | |
1484 | if (cfqd->busy_queues > 1) | |
0b182d61 | 1485 | return false; |
9ede209e | 1486 | |
365722bb | 1487 | /* |
8e296755 | 1488 | * Sole queue user, allow bigger slice |
365722bb | 1489 | */ |
8e296755 JA |
1490 | max_dispatch *= 4; |
1491 | } | |
1492 | ||
1493 | /* | |
1494 | * Async queues must wait a bit before being allowed dispatch. | |
1495 | * We also ramp up the dispatch depth gradually for async IO, | |
1496 | * based on the last sync IO we serviced | |
1497 | */ | |
963b72fc | 1498 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { |
8e296755 JA |
1499 | unsigned long last_sync = jiffies - cfqd->last_end_sync_rq; |
1500 | unsigned int depth; | |
365722bb | 1501 | |
61f0c1dc | 1502 | depth = last_sync / cfqd->cfq_slice[1]; |
e00c54c3 JA |
1503 | if (!depth && !cfqq->dispatched) |
1504 | depth = 1; | |
8e296755 JA |
1505 | if (depth < max_dispatch) |
1506 | max_dispatch = depth; | |
2f5cb738 | 1507 | } |
3ed9a296 | 1508 | |
0b182d61 JA |
1509 | /* |
1510 | * If we're below the current max, allow a dispatch | |
1511 | */ | |
1512 | return cfqq->dispatched < max_dispatch; | |
1513 | } | |
1514 | ||
1515 | /* | |
1516 | * Dispatch a request from cfqq, moving them to the request queue | |
1517 | * dispatch list. | |
1518 | */ | |
1519 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1520 | { | |
1521 | struct request *rq; | |
1522 | ||
1523 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
1524 | ||
1525 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
1526 | return false; | |
1527 | ||
1528 | /* | |
1529 | * follow expired path, else get first next available | |
1530 | */ | |
1531 | rq = cfq_check_fifo(cfqq); | |
1532 | if (!rq) | |
1533 | rq = cfqq->next_rq; | |
1534 | ||
1535 | /* | |
1536 | * insert request into driver dispatch list | |
1537 | */ | |
1538 | cfq_dispatch_insert(cfqd->queue, rq); | |
1539 | ||
1540 | if (!cfqd->active_cic) { | |
1541 | struct cfq_io_context *cic = RQ_CIC(rq); | |
1542 | ||
1543 | atomic_long_inc(&cic->ioc->refcount); | |
1544 | cfqd->active_cic = cic; | |
1545 | } | |
1546 | ||
1547 | return true; | |
1548 | } | |
1549 | ||
1550 | /* | |
1551 | * Find the cfqq that we need to service and move a request from that to the | |
1552 | * dispatch list | |
1553 | */ | |
1554 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
1555 | { | |
1556 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1557 | struct cfq_queue *cfqq; | |
1558 | ||
1559 | if (!cfqd->busy_queues) | |
1560 | return 0; | |
1561 | ||
1562 | if (unlikely(force)) | |
1563 | return cfq_forced_dispatch(cfqd); | |
1564 | ||
1565 | cfqq = cfq_select_queue(cfqd); | |
1566 | if (!cfqq) | |
8e296755 JA |
1567 | return 0; |
1568 | ||
2f5cb738 | 1569 | /* |
0b182d61 | 1570 | * Dispatch a request from this cfqq, if it is allowed |
2f5cb738 | 1571 | */ |
0b182d61 JA |
1572 | if (!cfq_dispatch_request(cfqd, cfqq)) |
1573 | return 0; | |
1574 | ||
2f5cb738 | 1575 | cfqq->slice_dispatch++; |
b029195d | 1576 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 1577 | |
2f5cb738 JA |
1578 | /* |
1579 | * expire an async queue immediately if it has used up its slice. idle | |
1580 | * queue always expire after 1 dispatch round. | |
1581 | */ | |
1582 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
1583 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
1584 | cfq_class_idle(cfqq))) { | |
1585 | cfqq->slice_end = jiffies + 1; | |
1586 | cfq_slice_expired(cfqd, 0); | |
1da177e4 LT |
1587 | } |
1588 | ||
b217a903 | 1589 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); |
2f5cb738 | 1590 | return 1; |
1da177e4 LT |
1591 | } |
1592 | ||
1da177e4 | 1593 | /* |
5e705374 JA |
1594 | * task holds one reference to the queue, dropped when task exits. each rq |
1595 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 LT |
1596 | * |
1597 | * queue lock must be held here. | |
1598 | */ | |
1599 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
1600 | { | |
22e2c507 JA |
1601 | struct cfq_data *cfqd = cfqq->cfqd; |
1602 | ||
1603 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
1604 | |
1605 | if (!atomic_dec_and_test(&cfqq->ref)) | |
1606 | return; | |
1607 | ||
7b679138 | 1608 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 1609 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 1610 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
3b18152c | 1611 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 | 1612 | |
28f95cbc | 1613 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 1614 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 1615 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 1616 | } |
22e2c507 | 1617 | |
1da177e4 LT |
1618 | kmem_cache_free(cfq_pool, cfqq); |
1619 | } | |
1620 | ||
d6de8be7 JA |
1621 | /* |
1622 | * Must always be called with the rcu_read_lock() held | |
1623 | */ | |
07416d29 JA |
1624 | static void |
1625 | __call_for_each_cic(struct io_context *ioc, | |
1626 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1627 | { | |
1628 | struct cfq_io_context *cic; | |
1629 | struct hlist_node *n; | |
1630 | ||
1631 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) | |
1632 | func(ioc, cic); | |
1633 | } | |
1634 | ||
4ac845a2 | 1635 | /* |
34e6bbf2 | 1636 | * Call func for each cic attached to this ioc. |
4ac845a2 | 1637 | */ |
34e6bbf2 | 1638 | static void |
4ac845a2 JA |
1639 | call_for_each_cic(struct io_context *ioc, |
1640 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1da177e4 | 1641 | { |
4ac845a2 | 1642 | rcu_read_lock(); |
07416d29 | 1643 | __call_for_each_cic(ioc, func); |
4ac845a2 | 1644 | rcu_read_unlock(); |
34e6bbf2 FC |
1645 | } |
1646 | ||
1647 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
1648 | { | |
1649 | struct cfq_io_context *cic; | |
1650 | ||
1651 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
1652 | ||
1653 | kmem_cache_free(cfq_ioc_pool, cic); | |
245b2e70 | 1654 | elv_ioc_count_dec(cfq_ioc_count); |
34e6bbf2 | 1655 | |
9a11b4ed JA |
1656 | if (ioc_gone) { |
1657 | /* | |
1658 | * CFQ scheduler is exiting, grab exit lock and check | |
1659 | * the pending io context count. If it hits zero, | |
1660 | * complete ioc_gone and set it back to NULL | |
1661 | */ | |
1662 | spin_lock(&ioc_gone_lock); | |
245b2e70 | 1663 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { |
9a11b4ed JA |
1664 | complete(ioc_gone); |
1665 | ioc_gone = NULL; | |
1666 | } | |
1667 | spin_unlock(&ioc_gone_lock); | |
1668 | } | |
34e6bbf2 | 1669 | } |
4ac845a2 | 1670 | |
34e6bbf2 FC |
1671 | static void cfq_cic_free(struct cfq_io_context *cic) |
1672 | { | |
1673 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
1674 | } |
1675 | ||
1676 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
1677 | { | |
1678 | unsigned long flags; | |
1679 | ||
1680 | BUG_ON(!cic->dead_key); | |
1681 | ||
1682 | spin_lock_irqsave(&ioc->lock, flags); | |
1683 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
ffc4e759 | 1684 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
1685 | spin_unlock_irqrestore(&ioc->lock, flags); |
1686 | ||
34e6bbf2 | 1687 | cfq_cic_free(cic); |
4ac845a2 JA |
1688 | } |
1689 | ||
d6de8be7 JA |
1690 | /* |
1691 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
1692 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
1693 | * and ->trim() which is called with the task lock held | |
1694 | */ | |
4ac845a2 JA |
1695 | static void cfq_free_io_context(struct io_context *ioc) |
1696 | { | |
4ac845a2 | 1697 | /* |
34e6bbf2 FC |
1698 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
1699 | * so no more cic's are allowed to be linked into this ioc. So it | |
1700 | * should be ok to iterate over the known list, we will see all cic's | |
1701 | * since no new ones are added. | |
4ac845a2 | 1702 | */ |
07416d29 | 1703 | __call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
1704 | } |
1705 | ||
89850f7e | 1706 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1707 | { |
df5fe3e8 JM |
1708 | struct cfq_queue *__cfqq, *next; |
1709 | ||
28f95cbc | 1710 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 1711 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 1712 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 1713 | } |
22e2c507 | 1714 | |
df5fe3e8 JM |
1715 | /* |
1716 | * If this queue was scheduled to merge with another queue, be | |
1717 | * sure to drop the reference taken on that queue (and others in | |
1718 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
1719 | */ | |
1720 | __cfqq = cfqq->new_cfqq; | |
1721 | while (__cfqq) { | |
1722 | if (__cfqq == cfqq) { | |
1723 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
1724 | break; | |
1725 | } | |
1726 | next = __cfqq->new_cfqq; | |
1727 | cfq_put_queue(__cfqq); | |
1728 | __cfqq = next; | |
1729 | } | |
1730 | ||
89850f7e JA |
1731 | cfq_put_queue(cfqq); |
1732 | } | |
22e2c507 | 1733 | |
89850f7e JA |
1734 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
1735 | struct cfq_io_context *cic) | |
1736 | { | |
4faa3c81 FC |
1737 | struct io_context *ioc = cic->ioc; |
1738 | ||
fc46379d | 1739 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
1740 | |
1741 | /* | |
1742 | * Make sure key == NULL is seen for dead queues | |
1743 | */ | |
fc46379d | 1744 | smp_wmb(); |
4ac845a2 | 1745 | cic->dead_key = (unsigned long) cic->key; |
fc46379d JA |
1746 | cic->key = NULL; |
1747 | ||
4faa3c81 FC |
1748 | if (ioc->ioc_data == cic) |
1749 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
1750 | ||
ff6657c6 JA |
1751 | if (cic->cfqq[BLK_RW_ASYNC]) { |
1752 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
1753 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
1754 | } |
1755 | ||
ff6657c6 JA |
1756 | if (cic->cfqq[BLK_RW_SYNC]) { |
1757 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
1758 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 1759 | } |
89850f7e JA |
1760 | } |
1761 | ||
4ac845a2 JA |
1762 | static void cfq_exit_single_io_context(struct io_context *ioc, |
1763 | struct cfq_io_context *cic) | |
89850f7e JA |
1764 | { |
1765 | struct cfq_data *cfqd = cic->key; | |
1766 | ||
89850f7e | 1767 | if (cfqd) { |
165125e1 | 1768 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 1769 | unsigned long flags; |
89850f7e | 1770 | |
4ac845a2 | 1771 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
1772 | |
1773 | /* | |
1774 | * Ensure we get a fresh copy of the ->key to prevent | |
1775 | * race between exiting task and queue | |
1776 | */ | |
1777 | smp_read_barrier_depends(); | |
1778 | if (cic->key) | |
1779 | __cfq_exit_single_io_context(cfqd, cic); | |
1780 | ||
4ac845a2 | 1781 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 1782 | } |
1da177e4 LT |
1783 | } |
1784 | ||
498d3aa2 JA |
1785 | /* |
1786 | * The process that ioc belongs to has exited, we need to clean up | |
1787 | * and put the internal structures we have that belongs to that process. | |
1788 | */ | |
e2d74ac0 | 1789 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 1790 | { |
4ac845a2 | 1791 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
1792 | } |
1793 | ||
22e2c507 | 1794 | static struct cfq_io_context * |
8267e268 | 1795 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 1796 | { |
b5deef90 | 1797 | struct cfq_io_context *cic; |
1da177e4 | 1798 | |
94f6030c CL |
1799 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
1800 | cfqd->queue->node); | |
1da177e4 | 1801 | if (cic) { |
22e2c507 | 1802 | cic->last_end_request = jiffies; |
553698f9 | 1803 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 1804 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
1805 | cic->dtor = cfq_free_io_context; |
1806 | cic->exit = cfq_exit_io_context; | |
245b2e70 | 1807 | elv_ioc_count_inc(cfq_ioc_count); |
1da177e4 LT |
1808 | } |
1809 | ||
1810 | return cic; | |
1811 | } | |
1812 | ||
fd0928df | 1813 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
1814 | { |
1815 | struct task_struct *tsk = current; | |
1816 | int ioprio_class; | |
1817 | ||
3b18152c | 1818 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
1819 | return; |
1820 | ||
fd0928df | 1821 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 1822 | switch (ioprio_class) { |
fe094d98 JA |
1823 | default: |
1824 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
1825 | case IOPRIO_CLASS_NONE: | |
1826 | /* | |
6d63c275 | 1827 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
1828 | */ |
1829 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 1830 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
1831 | break; |
1832 | case IOPRIO_CLASS_RT: | |
1833 | cfqq->ioprio = task_ioprio(ioc); | |
1834 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
1835 | break; | |
1836 | case IOPRIO_CLASS_BE: | |
1837 | cfqq->ioprio = task_ioprio(ioc); | |
1838 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
1839 | break; | |
1840 | case IOPRIO_CLASS_IDLE: | |
1841 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
1842 | cfqq->ioprio = 7; | |
1843 | cfq_clear_cfqq_idle_window(cfqq); | |
1844 | break; | |
22e2c507 JA |
1845 | } |
1846 | ||
1847 | /* | |
1848 | * keep track of original prio settings in case we have to temporarily | |
1849 | * elevate the priority of this queue | |
1850 | */ | |
1851 | cfqq->org_ioprio = cfqq->ioprio; | |
1852 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 1853 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
1854 | } |
1855 | ||
febffd61 | 1856 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 1857 | { |
478a82b0 AV |
1858 | struct cfq_data *cfqd = cic->key; |
1859 | struct cfq_queue *cfqq; | |
c1b707d2 | 1860 | unsigned long flags; |
35e6077c | 1861 | |
caaa5f9f JA |
1862 | if (unlikely(!cfqd)) |
1863 | return; | |
1864 | ||
c1b707d2 | 1865 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 1866 | |
ff6657c6 | 1867 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
1868 | if (cfqq) { |
1869 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
1870 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
1871 | GFP_ATOMIC); | |
caaa5f9f | 1872 | if (new_cfqq) { |
ff6657c6 | 1873 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
1874 | cfq_put_queue(cfqq); |
1875 | } | |
22e2c507 | 1876 | } |
caaa5f9f | 1877 | |
ff6657c6 | 1878 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
1879 | if (cfqq) |
1880 | cfq_mark_cfqq_prio_changed(cfqq); | |
1881 | ||
c1b707d2 | 1882 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
1883 | } |
1884 | ||
fc46379d | 1885 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 1886 | { |
4ac845a2 | 1887 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 1888 | ioc->ioprio_changed = 0; |
22e2c507 JA |
1889 | } |
1890 | ||
d5036d77 | 1891 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 1892 | pid_t pid, bool is_sync) |
d5036d77 JA |
1893 | { |
1894 | RB_CLEAR_NODE(&cfqq->rb_node); | |
1895 | RB_CLEAR_NODE(&cfqq->p_node); | |
1896 | INIT_LIST_HEAD(&cfqq->fifo); | |
1897 | ||
1898 | atomic_set(&cfqq->ref, 0); | |
1899 | cfqq->cfqd = cfqd; | |
1900 | ||
1901 | cfq_mark_cfqq_prio_changed(cfqq); | |
1902 | ||
1903 | if (is_sync) { | |
1904 | if (!cfq_class_idle(cfqq)) | |
1905 | cfq_mark_cfqq_idle_window(cfqq); | |
1906 | cfq_mark_cfqq_sync(cfqq); | |
1907 | } | |
1908 | cfqq->pid = pid; | |
1909 | } | |
1910 | ||
22e2c507 | 1911 | static struct cfq_queue * |
a6151c3a | 1912 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, |
fd0928df | 1913 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 1914 | { |
22e2c507 | 1915 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 1916 | struct cfq_io_context *cic; |
22e2c507 JA |
1917 | |
1918 | retry: | |
4ac845a2 | 1919 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
1920 | /* cic always exists here */ |
1921 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 1922 | |
6118b70b JA |
1923 | /* |
1924 | * Always try a new alloc if we fell back to the OOM cfqq | |
1925 | * originally, since it should just be a temporary situation. | |
1926 | */ | |
1927 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
1928 | cfqq = NULL; | |
22e2c507 JA |
1929 | if (new_cfqq) { |
1930 | cfqq = new_cfqq; | |
1931 | new_cfqq = NULL; | |
1932 | } else if (gfp_mask & __GFP_WAIT) { | |
1933 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 1934 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 1935 | gfp_mask | __GFP_ZERO, |
94f6030c | 1936 | cfqd->queue->node); |
22e2c507 | 1937 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
1938 | if (new_cfqq) |
1939 | goto retry; | |
22e2c507 | 1940 | } else { |
94f6030c CL |
1941 | cfqq = kmem_cache_alloc_node(cfq_pool, |
1942 | gfp_mask | __GFP_ZERO, | |
1943 | cfqd->queue->node); | |
22e2c507 JA |
1944 | } |
1945 | ||
6118b70b JA |
1946 | if (cfqq) { |
1947 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
1948 | cfq_init_prio_data(cfqq, ioc); | |
1949 | cfq_log_cfqq(cfqd, cfqq, "alloced"); | |
1950 | } else | |
1951 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
1952 | } |
1953 | ||
1954 | if (new_cfqq) | |
1955 | kmem_cache_free(cfq_pool, new_cfqq); | |
1956 | ||
22e2c507 JA |
1957 | return cfqq; |
1958 | } | |
1959 | ||
c2dea2d1 VT |
1960 | static struct cfq_queue ** |
1961 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
1962 | { | |
fe094d98 | 1963 | switch (ioprio_class) { |
c2dea2d1 VT |
1964 | case IOPRIO_CLASS_RT: |
1965 | return &cfqd->async_cfqq[0][ioprio]; | |
1966 | case IOPRIO_CLASS_BE: | |
1967 | return &cfqd->async_cfqq[1][ioprio]; | |
1968 | case IOPRIO_CLASS_IDLE: | |
1969 | return &cfqd->async_idle_cfqq; | |
1970 | default: | |
1971 | BUG(); | |
1972 | } | |
1973 | } | |
1974 | ||
15c31be4 | 1975 | static struct cfq_queue * |
a6151c3a | 1976 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, |
15c31be4 JA |
1977 | gfp_t gfp_mask) |
1978 | { | |
fd0928df JA |
1979 | const int ioprio = task_ioprio(ioc); |
1980 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 1981 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
1982 | struct cfq_queue *cfqq = NULL; |
1983 | ||
c2dea2d1 VT |
1984 | if (!is_sync) { |
1985 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
1986 | cfqq = *async_cfqq; | |
1987 | } | |
1988 | ||
6118b70b | 1989 | if (!cfqq) |
fd0928df | 1990 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
1991 | |
1992 | /* | |
1993 | * pin the queue now that it's allocated, scheduler exit will prune it | |
1994 | */ | |
c2dea2d1 | 1995 | if (!is_sync && !(*async_cfqq)) { |
15c31be4 | 1996 | atomic_inc(&cfqq->ref); |
c2dea2d1 | 1997 | *async_cfqq = cfqq; |
15c31be4 JA |
1998 | } |
1999 | ||
2000 | atomic_inc(&cfqq->ref); | |
2001 | return cfqq; | |
2002 | } | |
2003 | ||
498d3aa2 JA |
2004 | /* |
2005 | * We drop cfq io contexts lazily, so we may find a dead one. | |
2006 | */ | |
dbecf3ab | 2007 | static void |
4ac845a2 JA |
2008 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
2009 | struct cfq_io_context *cic) | |
dbecf3ab | 2010 | { |
4ac845a2 JA |
2011 | unsigned long flags; |
2012 | ||
fc46379d | 2013 | WARN_ON(!list_empty(&cic->queue_list)); |
597bc485 | 2014 | |
4ac845a2 JA |
2015 | spin_lock_irqsave(&ioc->lock, flags); |
2016 | ||
4faa3c81 | 2017 | BUG_ON(ioc->ioc_data == cic); |
597bc485 | 2018 | |
4ac845a2 | 2019 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); |
ffc4e759 | 2020 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2021 | spin_unlock_irqrestore(&ioc->lock, flags); |
2022 | ||
2023 | cfq_cic_free(cic); | |
dbecf3ab OH |
2024 | } |
2025 | ||
e2d74ac0 | 2026 | static struct cfq_io_context * |
4ac845a2 | 2027 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 2028 | { |
e2d74ac0 | 2029 | struct cfq_io_context *cic; |
d6de8be7 | 2030 | unsigned long flags; |
4ac845a2 | 2031 | void *k; |
e2d74ac0 | 2032 | |
91fac317 VT |
2033 | if (unlikely(!ioc)) |
2034 | return NULL; | |
2035 | ||
d6de8be7 JA |
2036 | rcu_read_lock(); |
2037 | ||
597bc485 JA |
2038 | /* |
2039 | * we maintain a last-hit cache, to avoid browsing over the tree | |
2040 | */ | |
4ac845a2 | 2041 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
2042 | if (cic && cic->key == cfqd) { |
2043 | rcu_read_unlock(); | |
597bc485 | 2044 | return cic; |
d6de8be7 | 2045 | } |
597bc485 | 2046 | |
4ac845a2 | 2047 | do { |
4ac845a2 JA |
2048 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); |
2049 | rcu_read_unlock(); | |
2050 | if (!cic) | |
2051 | break; | |
be3b0753 OH |
2052 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
2053 | k = cic->key; | |
2054 | if (unlikely(!k)) { | |
4ac845a2 | 2055 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 2056 | rcu_read_lock(); |
4ac845a2 | 2057 | continue; |
dbecf3ab | 2058 | } |
e2d74ac0 | 2059 | |
d6de8be7 | 2060 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 2061 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 2062 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
2063 | break; |
2064 | } while (1); | |
e2d74ac0 | 2065 | |
4ac845a2 | 2066 | return cic; |
e2d74ac0 JA |
2067 | } |
2068 | ||
4ac845a2 JA |
2069 | /* |
2070 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
2071 | * the process specific cfq io context when entered from the block layer. | |
2072 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
2073 | */ | |
febffd61 JA |
2074 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
2075 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 2076 | { |
0261d688 | 2077 | unsigned long flags; |
4ac845a2 | 2078 | int ret; |
e2d74ac0 | 2079 | |
4ac845a2 JA |
2080 | ret = radix_tree_preload(gfp_mask); |
2081 | if (!ret) { | |
2082 | cic->ioc = ioc; | |
2083 | cic->key = cfqd; | |
e2d74ac0 | 2084 | |
4ac845a2 JA |
2085 | spin_lock_irqsave(&ioc->lock, flags); |
2086 | ret = radix_tree_insert(&ioc->radix_root, | |
2087 | (unsigned long) cfqd, cic); | |
ffc4e759 JA |
2088 | if (!ret) |
2089 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 2090 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 2091 | |
4ac845a2 JA |
2092 | radix_tree_preload_end(); |
2093 | ||
2094 | if (!ret) { | |
2095 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
2096 | list_add(&cic->queue_list, &cfqd->cic_list); | |
2097 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2098 | } | |
e2d74ac0 JA |
2099 | } |
2100 | ||
4ac845a2 JA |
2101 | if (ret) |
2102 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 2103 | |
4ac845a2 | 2104 | return ret; |
e2d74ac0 JA |
2105 | } |
2106 | ||
1da177e4 LT |
2107 | /* |
2108 | * Setup general io context and cfq io context. There can be several cfq | |
2109 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 2110 | * than one device managed by cfq. |
1da177e4 LT |
2111 | */ |
2112 | static struct cfq_io_context * | |
e2d74ac0 | 2113 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2114 | { |
22e2c507 | 2115 | struct io_context *ioc = NULL; |
1da177e4 | 2116 | struct cfq_io_context *cic; |
1da177e4 | 2117 | |
22e2c507 | 2118 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 2119 | |
b5deef90 | 2120 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
2121 | if (!ioc) |
2122 | return NULL; | |
2123 | ||
4ac845a2 | 2124 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
2125 | if (cic) |
2126 | goto out; | |
1da177e4 | 2127 | |
e2d74ac0 JA |
2128 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
2129 | if (cic == NULL) | |
2130 | goto err; | |
1da177e4 | 2131 | |
4ac845a2 JA |
2132 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
2133 | goto err_free; | |
2134 | ||
1da177e4 | 2135 | out: |
fc46379d JA |
2136 | smp_read_barrier_depends(); |
2137 | if (unlikely(ioc->ioprio_changed)) | |
2138 | cfq_ioc_set_ioprio(ioc); | |
2139 | ||
1da177e4 | 2140 | return cic; |
4ac845a2 JA |
2141 | err_free: |
2142 | cfq_cic_free(cic); | |
1da177e4 LT |
2143 | err: |
2144 | put_io_context(ioc); | |
2145 | return NULL; | |
2146 | } | |
2147 | ||
22e2c507 JA |
2148 | static void |
2149 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 2150 | { |
aaf1228d JA |
2151 | unsigned long elapsed = jiffies - cic->last_end_request; |
2152 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 2153 | |
22e2c507 JA |
2154 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
2155 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
2156 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
2157 | } | |
1da177e4 | 2158 | |
206dc69b | 2159 | static void |
b2c18e1e | 2160 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
6d048f53 | 2161 | struct request *rq) |
206dc69b JA |
2162 | { |
2163 | sector_t sdist; | |
2164 | u64 total; | |
2165 | ||
b2c18e1e | 2166 | if (!cfqq->last_request_pos) |
4d00aa47 | 2167 | sdist = 0; |
b2c18e1e JM |
2168 | else if (cfqq->last_request_pos < blk_rq_pos(rq)) |
2169 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
206dc69b | 2170 | else |
b2c18e1e | 2171 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); |
206dc69b JA |
2172 | |
2173 | /* | |
2174 | * Don't allow the seek distance to get too large from the | |
2175 | * odd fragment, pagein, etc | |
2176 | */ | |
b2c18e1e JM |
2177 | if (cfqq->seek_samples <= 60) /* second&third seek */ |
2178 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024); | |
206dc69b | 2179 | else |
b2c18e1e | 2180 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64); |
206dc69b | 2181 | |
b2c18e1e JM |
2182 | cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8; |
2183 | cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8; | |
2184 | total = cfqq->seek_total + (cfqq->seek_samples/2); | |
2185 | do_div(total, cfqq->seek_samples); | |
2186 | cfqq->seek_mean = (sector_t)total; | |
e6c5bc73 JM |
2187 | |
2188 | /* | |
2189 | * If this cfqq is shared between multiple processes, check to | |
2190 | * make sure that those processes are still issuing I/Os within | |
2191 | * the mean seek distance. If not, it may be time to break the | |
2192 | * queues apart again. | |
2193 | */ | |
2194 | if (cfq_cfqq_coop(cfqq)) { | |
2195 | if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start) | |
2196 | cfqq->seeky_start = jiffies; | |
2197 | else if (!CFQQ_SEEKY(cfqq)) | |
2198 | cfqq->seeky_start = 0; | |
2199 | } | |
206dc69b | 2200 | } |
1da177e4 | 2201 | |
22e2c507 JA |
2202 | /* |
2203 | * Disable idle window if the process thinks too long or seeks so much that | |
2204 | * it doesn't matter | |
2205 | */ | |
2206 | static void | |
2207 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2208 | struct cfq_io_context *cic) | |
2209 | { | |
7b679138 | 2210 | int old_idle, enable_idle; |
1be92f2f | 2211 | |
0871714e JA |
2212 | /* |
2213 | * Don't idle for async or idle io prio class | |
2214 | */ | |
2215 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
2216 | return; |
2217 | ||
c265a7f4 | 2218 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 2219 | |
66dac98e | 2220 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || |
b2c18e1e | 2221 | (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq))) |
22e2c507 JA |
2222 | enable_idle = 0; |
2223 | else if (sample_valid(cic->ttime_samples)) { | |
ec60e4f6 | 2224 | unsigned int slice_idle = cfqd->cfq_slice_idle; |
b2c18e1e | 2225 | if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)) |
ec60e4f6 CZ |
2226 | slice_idle = msecs_to_jiffies(CFQ_MIN_TT); |
2227 | if (cic->ttime_mean > slice_idle) | |
22e2c507 JA |
2228 | enable_idle = 0; |
2229 | else | |
2230 | enable_idle = 1; | |
1da177e4 LT |
2231 | } |
2232 | ||
7b679138 JA |
2233 | if (old_idle != enable_idle) { |
2234 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
2235 | if (enable_idle) | |
2236 | cfq_mark_cfqq_idle_window(cfqq); | |
2237 | else | |
2238 | cfq_clear_cfqq_idle_window(cfqq); | |
2239 | } | |
22e2c507 | 2240 | } |
1da177e4 | 2241 | |
22e2c507 JA |
2242 | /* |
2243 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
2244 | * no or if we aren't sure, a 1 will cause a preempt. | |
2245 | */ | |
a6151c3a | 2246 | static bool |
22e2c507 | 2247 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, |
5e705374 | 2248 | struct request *rq) |
22e2c507 | 2249 | { |
6d048f53 | 2250 | struct cfq_queue *cfqq; |
22e2c507 | 2251 | |
6d048f53 JA |
2252 | cfqq = cfqd->active_queue; |
2253 | if (!cfqq) | |
a6151c3a | 2254 | return false; |
22e2c507 | 2255 | |
6d048f53 | 2256 | if (cfq_slice_used(cfqq)) |
a6151c3a | 2257 | return true; |
6d048f53 JA |
2258 | |
2259 | if (cfq_class_idle(new_cfqq)) | |
a6151c3a | 2260 | return false; |
22e2c507 JA |
2261 | |
2262 | if (cfq_class_idle(cfqq)) | |
a6151c3a | 2263 | return true; |
1e3335de | 2264 | |
374f84ac JA |
2265 | /* |
2266 | * if the new request is sync, but the currently running queue is | |
2267 | * not, let the sync request have priority. | |
2268 | */ | |
5e705374 | 2269 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
a6151c3a | 2270 | return true; |
1e3335de | 2271 | |
374f84ac JA |
2272 | /* |
2273 | * So both queues are sync. Let the new request get disk time if | |
2274 | * it's a metadata request and the current queue is doing regular IO. | |
2275 | */ | |
2276 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
a6151c3a | 2277 | return false; |
22e2c507 | 2278 | |
3a9a3f6c DS |
2279 | /* |
2280 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
2281 | */ | |
2282 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
a6151c3a | 2283 | return true; |
3a9a3f6c | 2284 | |
1e3335de | 2285 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
a6151c3a | 2286 | return false; |
1e3335de JA |
2287 | |
2288 | /* | |
2289 | * if this request is as-good as one we would expect from the | |
2290 | * current cfqq, let it preempt | |
2291 | */ | |
b2c18e1e | 2292 | if (cfq_rq_close(cfqd, cfqq, rq)) |
a6151c3a | 2293 | return true; |
1e3335de | 2294 | |
a6151c3a | 2295 | return false; |
22e2c507 JA |
2296 | } |
2297 | ||
2298 | /* | |
2299 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
2300 | * let it have half of its nominal slice. | |
2301 | */ | |
2302 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2303 | { | |
7b679138 | 2304 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
6084cdda | 2305 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 2306 | |
bf572256 JA |
2307 | /* |
2308 | * Put the new queue at the front of the of the current list, | |
2309 | * so we know that it will be selected next. | |
2310 | */ | |
2311 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
2312 | |
2313 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 2314 | |
44f7c160 JA |
2315 | cfqq->slice_end = 0; |
2316 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
2317 | } |
2318 | ||
22e2c507 | 2319 | /* |
5e705374 | 2320 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
2321 | * something we should do about it |
2322 | */ | |
2323 | static void | |
5e705374 JA |
2324 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
2325 | struct request *rq) | |
22e2c507 | 2326 | { |
5e705374 | 2327 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 2328 | |
45333d5a | 2329 | cfqd->rq_queued++; |
374f84ac JA |
2330 | if (rq_is_meta(rq)) |
2331 | cfqq->meta_pending++; | |
2332 | ||
9c2c38a1 | 2333 | cfq_update_io_thinktime(cfqd, cic); |
b2c18e1e | 2334 | cfq_update_io_seektime(cfqd, cfqq, rq); |
9c2c38a1 JA |
2335 | cfq_update_idle_window(cfqd, cfqq, cic); |
2336 | ||
b2c18e1e | 2337 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
2338 | |
2339 | if (cfqq == cfqd->active_queue) { | |
2340 | /* | |
b029195d JA |
2341 | * Remember that we saw a request from this process, but |
2342 | * don't start queuing just yet. Otherwise we risk seeing lots | |
2343 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
2344 | * and merging. If the request is already larger than a single |
2345 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
2346 | * merging is already done. Ditto for a busy system that |
2347 | * has other work pending, don't risk delaying until the | |
2348 | * idle timer unplug to continue working. | |
22e2c507 | 2349 | */ |
d6ceb25e | 2350 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
2351 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
2352 | cfqd->busy_queues > 1) { | |
d6ceb25e | 2353 | del_timer(&cfqd->idle_slice_timer); |
a7f55792 | 2354 | __blk_run_queue(cfqd->queue); |
d6ceb25e | 2355 | } |
b029195d | 2356 | cfq_mark_cfqq_must_dispatch(cfqq); |
d6ceb25e | 2357 | } |
5e705374 | 2358 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
2359 | /* |
2360 | * not the active queue - expire current slice if it is | |
2361 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
2362 | * has some old slice time left and is of higher priority or |
2363 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
2364 | */ |
2365 | cfq_preempt_queue(cfqd, cfqq); | |
a7f55792 | 2366 | __blk_run_queue(cfqd->queue); |
22e2c507 | 2367 | } |
1da177e4 LT |
2368 | } |
2369 | ||
165125e1 | 2370 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2371 | { |
b4878f24 | 2372 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 2373 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 2374 | |
7b679138 | 2375 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 2376 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 2377 | |
30996f40 | 2378 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); |
22e2c507 | 2379 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
aa6f6a3d | 2380 | cfq_add_rq_rb(rq); |
22e2c507 | 2381 | |
5e705374 | 2382 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
2383 | } |
2384 | ||
45333d5a AC |
2385 | /* |
2386 | * Update hw_tag based on peak queue depth over 50 samples under | |
2387 | * sufficient load. | |
2388 | */ | |
2389 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
2390 | { | |
1a1238a7 SL |
2391 | struct cfq_queue *cfqq = cfqd->active_queue; |
2392 | ||
5ad531db JA |
2393 | if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak) |
2394 | cfqd->rq_in_driver_peak = rq_in_driver(cfqd); | |
45333d5a AC |
2395 | |
2396 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
5ad531db | 2397 | rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
2398 | return; |
2399 | ||
1a1238a7 SL |
2400 | /* |
2401 | * If active queue hasn't enough requests and can idle, cfq might not | |
2402 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
2403 | * case | |
2404 | */ | |
2405 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
2406 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
2407 | CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN) | |
2408 | return; | |
2409 | ||
45333d5a AC |
2410 | if (cfqd->hw_tag_samples++ < 50) |
2411 | return; | |
2412 | ||
2413 | if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN) | |
2414 | cfqd->hw_tag = 1; | |
2415 | else | |
2416 | cfqd->hw_tag = 0; | |
2417 | ||
2418 | cfqd->hw_tag_samples = 0; | |
2419 | cfqd->rq_in_driver_peak = 0; | |
2420 | } | |
2421 | ||
165125e1 | 2422 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 2423 | { |
5e705374 | 2424 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 2425 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 2426 | const int sync = rq_is_sync(rq); |
b4878f24 | 2427 | unsigned long now; |
1da177e4 | 2428 | |
b4878f24 | 2429 | now = jiffies; |
7b679138 | 2430 | cfq_log_cfqq(cfqd, cfqq, "complete"); |
1da177e4 | 2431 | |
45333d5a AC |
2432 | cfq_update_hw_tag(cfqd); |
2433 | ||
5ad531db | 2434 | WARN_ON(!cfqd->rq_in_driver[sync]); |
6d048f53 | 2435 | WARN_ON(!cfqq->dispatched); |
5ad531db | 2436 | cfqd->rq_in_driver[sync]--; |
6d048f53 | 2437 | cfqq->dispatched--; |
1da177e4 | 2438 | |
3ed9a296 JA |
2439 | if (cfq_cfqq_sync(cfqq)) |
2440 | cfqd->sync_flight--; | |
2441 | ||
365722bb | 2442 | if (sync) { |
5e705374 | 2443 | RQ_CIC(rq)->last_end_request = now; |
365722bb VG |
2444 | cfqd->last_end_sync_rq = now; |
2445 | } | |
caaa5f9f JA |
2446 | |
2447 | /* | |
2448 | * If this is the active queue, check if it needs to be expired, | |
2449 | * or if we want to idle in case it has no pending requests. | |
2450 | */ | |
2451 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
2452 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
2453 | ||
44f7c160 JA |
2454 | if (cfq_cfqq_slice_new(cfqq)) { |
2455 | cfq_set_prio_slice(cfqd, cfqq); | |
2456 | cfq_clear_cfqq_slice_new(cfqq); | |
2457 | } | |
a36e71f9 JA |
2458 | /* |
2459 | * If there are no requests waiting in this queue, and | |
2460 | * there are other queues ready to issue requests, AND | |
2461 | * those other queues are issuing requests within our | |
2462 | * mean seek distance, give them a chance to run instead | |
2463 | * of idling. | |
2464 | */ | |
0871714e | 2465 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
6084cdda | 2466 | cfq_slice_expired(cfqd, 1); |
b3b6d040 | 2467 | else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq) && |
a36e71f9 | 2468 | sync && !rq_noidle(rq)) |
6d048f53 | 2469 | cfq_arm_slice_timer(cfqd); |
caaa5f9f | 2470 | } |
6d048f53 | 2471 | |
5ad531db | 2472 | if (!rq_in_driver(cfqd)) |
23e018a1 | 2473 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
2474 | } |
2475 | ||
22e2c507 JA |
2476 | /* |
2477 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
2478 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
2479 | */ | |
2480 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 2481 | { |
22e2c507 JA |
2482 | if (has_fs_excl()) { |
2483 | /* | |
2484 | * boost idle prio on transactions that would lock out other | |
2485 | * users of the filesystem | |
2486 | */ | |
2487 | if (cfq_class_idle(cfqq)) | |
2488 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2489 | if (cfqq->ioprio > IOPRIO_NORM) | |
2490 | cfqq->ioprio = IOPRIO_NORM; | |
2491 | } else { | |
2492 | /* | |
2493 | * check if we need to unboost the queue | |
2494 | */ | |
2495 | if (cfqq->ioprio_class != cfqq->org_ioprio_class) | |
2496 | cfqq->ioprio_class = cfqq->org_ioprio_class; | |
2497 | if (cfqq->ioprio != cfqq->org_ioprio) | |
2498 | cfqq->ioprio = cfqq->org_ioprio; | |
2499 | } | |
22e2c507 | 2500 | } |
1da177e4 | 2501 | |
89850f7e | 2502 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 2503 | { |
1b379d8d | 2504 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 2505 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 2506 | return ELV_MQUEUE_MUST; |
3b18152c | 2507 | } |
1da177e4 | 2508 | |
22e2c507 | 2509 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
2510 | } |
2511 | ||
165125e1 | 2512 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
2513 | { |
2514 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2515 | struct task_struct *tsk = current; | |
91fac317 | 2516 | struct cfq_io_context *cic; |
22e2c507 JA |
2517 | struct cfq_queue *cfqq; |
2518 | ||
2519 | /* | |
2520 | * don't force setup of a queue from here, as a call to may_queue | |
2521 | * does not necessarily imply that a request actually will be queued. | |
2522 | * so just lookup a possibly existing queue, or return 'may queue' | |
2523 | * if that fails | |
2524 | */ | |
4ac845a2 | 2525 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
2526 | if (!cic) |
2527 | return ELV_MQUEUE_MAY; | |
2528 | ||
b0b78f81 | 2529 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 2530 | if (cfqq) { |
fd0928df | 2531 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
2532 | cfq_prio_boost(cfqq); |
2533 | ||
89850f7e | 2534 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
2535 | } |
2536 | ||
2537 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
2538 | } |
2539 | ||
1da177e4 LT |
2540 | /* |
2541 | * queue lock held here | |
2542 | */ | |
bb37b94c | 2543 | static void cfq_put_request(struct request *rq) |
1da177e4 | 2544 | { |
5e705374 | 2545 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 2546 | |
5e705374 | 2547 | if (cfqq) { |
22e2c507 | 2548 | const int rw = rq_data_dir(rq); |
1da177e4 | 2549 | |
22e2c507 JA |
2550 | BUG_ON(!cfqq->allocated[rw]); |
2551 | cfqq->allocated[rw]--; | |
1da177e4 | 2552 | |
5e705374 | 2553 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 2554 | |
1da177e4 | 2555 | rq->elevator_private = NULL; |
5e705374 | 2556 | rq->elevator_private2 = NULL; |
1da177e4 | 2557 | |
1da177e4 LT |
2558 | cfq_put_queue(cfqq); |
2559 | } | |
2560 | } | |
2561 | ||
df5fe3e8 JM |
2562 | static struct cfq_queue * |
2563 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, | |
2564 | struct cfq_queue *cfqq) | |
2565 | { | |
2566 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
2567 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
b3b6d040 | 2568 | cfq_mark_cfqq_coop(cfqq->new_cfqq); |
df5fe3e8 JM |
2569 | cfq_put_queue(cfqq); |
2570 | return cic_to_cfqq(cic, 1); | |
2571 | } | |
2572 | ||
e6c5bc73 JM |
2573 | static int should_split_cfqq(struct cfq_queue *cfqq) |
2574 | { | |
2575 | if (cfqq->seeky_start && | |
2576 | time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT)) | |
2577 | return 1; | |
2578 | return 0; | |
2579 | } | |
2580 | ||
2581 | /* | |
2582 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
2583 | * was the last process referring to said cfqq. | |
2584 | */ | |
2585 | static struct cfq_queue * | |
2586 | split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) | |
2587 | { | |
2588 | if (cfqq_process_refs(cfqq) == 1) { | |
2589 | cfqq->seeky_start = 0; | |
2590 | cfqq->pid = current->pid; | |
2591 | cfq_clear_cfqq_coop(cfqq); | |
2592 | return cfqq; | |
2593 | } | |
2594 | ||
2595 | cic_set_cfqq(cic, NULL, 1); | |
2596 | cfq_put_queue(cfqq); | |
2597 | return NULL; | |
2598 | } | |
1da177e4 | 2599 | /* |
22e2c507 | 2600 | * Allocate cfq data structures associated with this request. |
1da177e4 | 2601 | */ |
22e2c507 | 2602 | static int |
165125e1 | 2603 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
2604 | { |
2605 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2606 | struct cfq_io_context *cic; | |
2607 | const int rw = rq_data_dir(rq); | |
a6151c3a | 2608 | const bool is_sync = rq_is_sync(rq); |
22e2c507 | 2609 | struct cfq_queue *cfqq; |
1da177e4 LT |
2610 | unsigned long flags; |
2611 | ||
2612 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
2613 | ||
e2d74ac0 | 2614 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 2615 | |
1da177e4 LT |
2616 | spin_lock_irqsave(q->queue_lock, flags); |
2617 | ||
22e2c507 JA |
2618 | if (!cic) |
2619 | goto queue_fail; | |
2620 | ||
e6c5bc73 | 2621 | new_queue: |
91fac317 | 2622 | cfqq = cic_to_cfqq(cic, is_sync); |
32f2e807 | 2623 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { |
fd0928df | 2624 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 | 2625 | cic_set_cfqq(cic, cfqq, is_sync); |
df5fe3e8 | 2626 | } else { |
e6c5bc73 JM |
2627 | /* |
2628 | * If the queue was seeky for too long, break it apart. | |
2629 | */ | |
2630 | if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) { | |
2631 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); | |
2632 | cfqq = split_cfqq(cic, cfqq); | |
2633 | if (!cfqq) | |
2634 | goto new_queue; | |
2635 | } | |
2636 | ||
df5fe3e8 JM |
2637 | /* |
2638 | * Check to see if this queue is scheduled to merge with | |
2639 | * another, closely cooperating queue. The merging of | |
2640 | * queues happens here as it must be done in process context. | |
2641 | * The reference on new_cfqq was taken in merge_cfqqs. | |
2642 | */ | |
2643 | if (cfqq->new_cfqq) | |
2644 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
91fac317 | 2645 | } |
1da177e4 LT |
2646 | |
2647 | cfqq->allocated[rw]++; | |
22e2c507 | 2648 | atomic_inc(&cfqq->ref); |
1da177e4 | 2649 | |
5e705374 | 2650 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 2651 | |
5e705374 JA |
2652 | rq->elevator_private = cic; |
2653 | rq->elevator_private2 = cfqq; | |
2654 | return 0; | |
1da177e4 | 2655 | |
22e2c507 JA |
2656 | queue_fail: |
2657 | if (cic) | |
2658 | put_io_context(cic->ioc); | |
89850f7e | 2659 | |
23e018a1 | 2660 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 2661 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 2662 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
2663 | return 1; |
2664 | } | |
2665 | ||
65f27f38 | 2666 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 2667 | { |
65f27f38 | 2668 | struct cfq_data *cfqd = |
23e018a1 | 2669 | container_of(work, struct cfq_data, unplug_work); |
165125e1 | 2670 | struct request_queue *q = cfqd->queue; |
22e2c507 | 2671 | |
40bb54d1 | 2672 | spin_lock_irq(q->queue_lock); |
a7f55792 | 2673 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 2674 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
2675 | } |
2676 | ||
2677 | /* | |
2678 | * Timer running if the active_queue is currently idling inside its time slice | |
2679 | */ | |
2680 | static void cfq_idle_slice_timer(unsigned long data) | |
2681 | { | |
2682 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
2683 | struct cfq_queue *cfqq; | |
2684 | unsigned long flags; | |
3c6bd2f8 | 2685 | int timed_out = 1; |
22e2c507 | 2686 | |
7b679138 JA |
2687 | cfq_log(cfqd, "idle timer fired"); |
2688 | ||
22e2c507 JA |
2689 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
2690 | ||
fe094d98 JA |
2691 | cfqq = cfqd->active_queue; |
2692 | if (cfqq) { | |
3c6bd2f8 JA |
2693 | timed_out = 0; |
2694 | ||
b029195d JA |
2695 | /* |
2696 | * We saw a request before the queue expired, let it through | |
2697 | */ | |
2698 | if (cfq_cfqq_must_dispatch(cfqq)) | |
2699 | goto out_kick; | |
2700 | ||
22e2c507 JA |
2701 | /* |
2702 | * expired | |
2703 | */ | |
44f7c160 | 2704 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
2705 | goto expire; |
2706 | ||
2707 | /* | |
2708 | * only expire and reinvoke request handler, if there are | |
2709 | * other queues with pending requests | |
2710 | */ | |
caaa5f9f | 2711 | if (!cfqd->busy_queues) |
22e2c507 | 2712 | goto out_cont; |
22e2c507 JA |
2713 | |
2714 | /* | |
2715 | * not expired and it has a request pending, let it dispatch | |
2716 | */ | |
75e50984 | 2717 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 2718 | goto out_kick; |
22e2c507 JA |
2719 | } |
2720 | expire: | |
6084cdda | 2721 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 2722 | out_kick: |
23e018a1 | 2723 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
2724 | out_cont: |
2725 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2726 | } | |
2727 | ||
3b18152c JA |
2728 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
2729 | { | |
2730 | del_timer_sync(&cfqd->idle_slice_timer); | |
23e018a1 | 2731 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 2732 | } |
22e2c507 | 2733 | |
c2dea2d1 VT |
2734 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
2735 | { | |
2736 | int i; | |
2737 | ||
2738 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
2739 | if (cfqd->async_cfqq[0][i]) | |
2740 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
2741 | if (cfqd->async_cfqq[1][i]) | |
2742 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 2743 | } |
2389d1ef ON |
2744 | |
2745 | if (cfqd->async_idle_cfqq) | |
2746 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
2747 | } |
2748 | ||
b374d18a | 2749 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 2750 | { |
22e2c507 | 2751 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 2752 | struct request_queue *q = cfqd->queue; |
22e2c507 | 2753 | |
3b18152c | 2754 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 2755 | |
d9ff4187 | 2756 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 2757 | |
d9ff4187 | 2758 | if (cfqd->active_queue) |
6084cdda | 2759 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
2760 | |
2761 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
2762 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
2763 | struct cfq_io_context, | |
2764 | queue_list); | |
89850f7e JA |
2765 | |
2766 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 2767 | } |
e2d74ac0 | 2768 | |
c2dea2d1 | 2769 | cfq_put_async_queues(cfqd); |
15c31be4 | 2770 | |
d9ff4187 | 2771 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
2772 | |
2773 | cfq_shutdown_timer_wq(cfqd); | |
2774 | ||
a90d742e | 2775 | kfree(cfqd); |
1da177e4 LT |
2776 | } |
2777 | ||
165125e1 | 2778 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
2779 | { |
2780 | struct cfq_data *cfqd; | |
26a2ac00 | 2781 | int i; |
1da177e4 | 2782 | |
94f6030c | 2783 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 2784 | if (!cfqd) |
bc1c1169 | 2785 | return NULL; |
1da177e4 | 2786 | |
c0324a02 CZ |
2787 | for (i = 0; i < 2; ++i) |
2788 | cfqd->service_trees[i] = CFQ_RB_ROOT; | |
2789 | cfqd->service_tree_idle = CFQ_RB_ROOT; | |
26a2ac00 JA |
2790 | |
2791 | /* | |
2792 | * Not strictly needed (since RB_ROOT just clears the node and we | |
2793 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
2794 | * to add magic to the rb code | |
2795 | */ | |
2796 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
2797 | cfqd->prio_trees[i] = RB_ROOT; | |
2798 | ||
6118b70b JA |
2799 | /* |
2800 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
2801 | * Grab a permanent reference to it, so that the normal code flow | |
2802 | * will not attempt to free it. | |
2803 | */ | |
2804 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
2805 | atomic_inc(&cfqd->oom_cfqq.ref); | |
2806 | ||
d9ff4187 | 2807 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 2808 | |
1da177e4 | 2809 | cfqd->queue = q; |
1da177e4 | 2810 | |
22e2c507 JA |
2811 | init_timer(&cfqd->idle_slice_timer); |
2812 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
2813 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
2814 | ||
23e018a1 | 2815 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 2816 | |
1da177e4 | 2817 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
2818 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
2819 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
2820 | cfqd->cfq_back_max = cfq_back_max; |
2821 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
2822 | cfqd->cfq_slice[0] = cfq_slice_async; |
2823 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
2824 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
2825 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
963b72fc | 2826 | cfqd->cfq_latency = 1; |
45333d5a | 2827 | cfqd->hw_tag = 1; |
365722bb | 2828 | cfqd->last_end_sync_rq = jiffies; |
bc1c1169 | 2829 | return cfqd; |
1da177e4 LT |
2830 | } |
2831 | ||
2832 | static void cfq_slab_kill(void) | |
2833 | { | |
d6de8be7 JA |
2834 | /* |
2835 | * Caller already ensured that pending RCU callbacks are completed, | |
2836 | * so we should have no busy allocations at this point. | |
2837 | */ | |
1da177e4 LT |
2838 | if (cfq_pool) |
2839 | kmem_cache_destroy(cfq_pool); | |
2840 | if (cfq_ioc_pool) | |
2841 | kmem_cache_destroy(cfq_ioc_pool); | |
2842 | } | |
2843 | ||
2844 | static int __init cfq_slab_setup(void) | |
2845 | { | |
0a31bd5f | 2846 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
2847 | if (!cfq_pool) |
2848 | goto fail; | |
2849 | ||
34e6bbf2 | 2850 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
2851 | if (!cfq_ioc_pool) |
2852 | goto fail; | |
2853 | ||
2854 | return 0; | |
2855 | fail: | |
2856 | cfq_slab_kill(); | |
2857 | return -ENOMEM; | |
2858 | } | |
2859 | ||
1da177e4 LT |
2860 | /* |
2861 | * sysfs parts below --> | |
2862 | */ | |
1da177e4 LT |
2863 | static ssize_t |
2864 | cfq_var_show(unsigned int var, char *page) | |
2865 | { | |
2866 | return sprintf(page, "%d\n", var); | |
2867 | } | |
2868 | ||
2869 | static ssize_t | |
2870 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
2871 | { | |
2872 | char *p = (char *) page; | |
2873 | ||
2874 | *var = simple_strtoul(p, &p, 10); | |
2875 | return count; | |
2876 | } | |
2877 | ||
1da177e4 | 2878 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 2879 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 2880 | { \ |
3d1ab40f | 2881 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2882 | unsigned int __data = __VAR; \ |
2883 | if (__CONV) \ | |
2884 | __data = jiffies_to_msecs(__data); \ | |
2885 | return cfq_var_show(__data, (page)); \ | |
2886 | } | |
2887 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
2888 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
2889 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
2890 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
2891 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
2892 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
2893 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
2894 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
2895 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
963b72fc | 2896 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); |
1da177e4 LT |
2897 | #undef SHOW_FUNCTION |
2898 | ||
2899 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 2900 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 2901 | { \ |
3d1ab40f | 2902 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
2903 | unsigned int __data; \ |
2904 | int ret = cfq_var_store(&__data, (page), count); \ | |
2905 | if (__data < (MIN)) \ | |
2906 | __data = (MIN); \ | |
2907 | else if (__data > (MAX)) \ | |
2908 | __data = (MAX); \ | |
2909 | if (__CONV) \ | |
2910 | *(__PTR) = msecs_to_jiffies(__data); \ | |
2911 | else \ | |
2912 | *(__PTR) = __data; \ | |
2913 | return ret; \ | |
2914 | } | |
2915 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
2916 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
2917 | UINT_MAX, 1); | |
2918 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
2919 | UINT_MAX, 1); | |
e572ec7e | 2920 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
2921 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
2922 | UINT_MAX, 0); | |
22e2c507 JA |
2923 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
2924 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
2925 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
2926 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
2927 | UINT_MAX, 0); | |
963b72fc | 2928 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); |
1da177e4 LT |
2929 | #undef STORE_FUNCTION |
2930 | ||
e572ec7e AV |
2931 | #define CFQ_ATTR(name) \ |
2932 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
2933 | ||
2934 | static struct elv_fs_entry cfq_attrs[] = { | |
2935 | CFQ_ATTR(quantum), | |
e572ec7e AV |
2936 | CFQ_ATTR(fifo_expire_sync), |
2937 | CFQ_ATTR(fifo_expire_async), | |
2938 | CFQ_ATTR(back_seek_max), | |
2939 | CFQ_ATTR(back_seek_penalty), | |
2940 | CFQ_ATTR(slice_sync), | |
2941 | CFQ_ATTR(slice_async), | |
2942 | CFQ_ATTR(slice_async_rq), | |
2943 | CFQ_ATTR(slice_idle), | |
963b72fc | 2944 | CFQ_ATTR(low_latency), |
e572ec7e | 2945 | __ATTR_NULL |
1da177e4 LT |
2946 | }; |
2947 | ||
1da177e4 LT |
2948 | static struct elevator_type iosched_cfq = { |
2949 | .ops = { | |
2950 | .elevator_merge_fn = cfq_merge, | |
2951 | .elevator_merged_fn = cfq_merged_request, | |
2952 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 2953 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 2954 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 2955 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 2956 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
2957 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
2958 | .elevator_queue_empty_fn = cfq_queue_empty, | |
2959 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
2960 | .elevator_former_req_fn = elv_rb_former_request, |
2961 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
2962 | .elevator_set_req_fn = cfq_set_request, |
2963 | .elevator_put_req_fn = cfq_put_request, | |
2964 | .elevator_may_queue_fn = cfq_may_queue, | |
2965 | .elevator_init_fn = cfq_init_queue, | |
2966 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 2967 | .trim = cfq_free_io_context, |
1da177e4 | 2968 | }, |
3d1ab40f | 2969 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
2970 | .elevator_name = "cfq", |
2971 | .elevator_owner = THIS_MODULE, | |
2972 | }; | |
2973 | ||
2974 | static int __init cfq_init(void) | |
2975 | { | |
22e2c507 JA |
2976 | /* |
2977 | * could be 0 on HZ < 1000 setups | |
2978 | */ | |
2979 | if (!cfq_slice_async) | |
2980 | cfq_slice_async = 1; | |
2981 | if (!cfq_slice_idle) | |
2982 | cfq_slice_idle = 1; | |
2983 | ||
1da177e4 LT |
2984 | if (cfq_slab_setup()) |
2985 | return -ENOMEM; | |
2986 | ||
2fdd82bd | 2987 | elv_register(&iosched_cfq); |
1da177e4 | 2988 | |
2fdd82bd | 2989 | return 0; |
1da177e4 LT |
2990 | } |
2991 | ||
2992 | static void __exit cfq_exit(void) | |
2993 | { | |
6e9a4738 | 2994 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 2995 | elv_unregister(&iosched_cfq); |
334e94de | 2996 | ioc_gone = &all_gone; |
fba82272 OH |
2997 | /* ioc_gone's update must be visible before reading ioc_count */ |
2998 | smp_wmb(); | |
d6de8be7 JA |
2999 | |
3000 | /* | |
3001 | * this also protects us from entering cfq_slab_kill() with | |
3002 | * pending RCU callbacks | |
3003 | */ | |
245b2e70 | 3004 | if (elv_ioc_count_read(cfq_ioc_count)) |
9a11b4ed | 3005 | wait_for_completion(&all_gone); |
83521d3e | 3006 | cfq_slab_kill(); |
1da177e4 LT |
3007 | } |
3008 | ||
3009 | module_init(cfq_init); | |
3010 | module_exit(cfq_exit); | |
3011 | ||
3012 | MODULE_AUTHOR("Jens Axboe"); | |
3013 | MODULE_LICENSE("GPL"); | |
3014 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |