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