Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
1da177e4 LT |
2 | * CFQ, or complete fairness queueing, disk scheduler. |
3 | * | |
4 | * Based on ideas from a previously unfinished io | |
5 | * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli. | |
6 | * | |
0fe23479 | 7 | * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk> |
1da177e4 | 8 | */ |
1da177e4 | 9 | #include <linux/module.h> |
1cc9be68 AV |
10 | #include <linux/blkdev.h> |
11 | #include <linux/elevator.h> | |
ad5ebd2f | 12 | #include <linux/jiffies.h> |
1da177e4 | 13 | #include <linux/rbtree.h> |
22e2c507 | 14 | #include <linux/ioprio.h> |
7b679138 | 15 | #include <linux/blktrace_api.h> |
25bc6b07 | 16 | #include "blk-cgroup.h" |
1da177e4 LT |
17 | |
18 | /* | |
19 | * tunables | |
20 | */ | |
fe094d98 JA |
21 | /* max queue in one round of service */ |
22 | static const int cfq_quantum = 4; | |
64100099 | 23 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
fe094d98 JA |
24 | /* maximum backwards seek, in KiB */ |
25 | static const int cfq_back_max = 16 * 1024; | |
26 | /* penalty of a backwards seek */ | |
27 | static const int cfq_back_penalty = 2; | |
64100099 | 28 | static const int cfq_slice_sync = HZ / 10; |
3b18152c | 29 | static int cfq_slice_async = HZ / 25; |
64100099 | 30 | static const int cfq_slice_async_rq = 2; |
caaa5f9f | 31 | static int cfq_slice_idle = HZ / 125; |
5db5d642 CZ |
32 | static const int cfq_target_latency = HZ * 3/10; /* 300 ms */ |
33 | static const int cfq_hist_divisor = 4; | |
22e2c507 | 34 | |
d9e7620e | 35 | /* |
0871714e | 36 | * offset from end of service tree |
d9e7620e | 37 | */ |
0871714e | 38 | #define CFQ_IDLE_DELAY (HZ / 5) |
d9e7620e JA |
39 | |
40 | /* | |
41 | * below this threshold, we consider thinktime immediate | |
42 | */ | |
43 | #define CFQ_MIN_TT (2) | |
44 | ||
e6c5bc73 JM |
45 | /* |
46 | * Allow merged cfqqs to perform this amount of seeky I/O before | |
47 | * deciding to break the queues up again. | |
48 | */ | |
49 | #define CFQQ_COOP_TOUT (HZ) | |
50 | ||
22e2c507 | 51 | #define CFQ_SLICE_SCALE (5) |
45333d5a | 52 | #define CFQ_HW_QUEUE_MIN (5) |
25bc6b07 | 53 | #define CFQ_SERVICE_SHIFT 12 |
22e2c507 | 54 | |
fe094d98 JA |
55 | #define RQ_CIC(rq) \ |
56 | ((struct cfq_io_context *) (rq)->elevator_private) | |
7b679138 | 57 | #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2) |
1da177e4 | 58 | |
e18b890b CL |
59 | static struct kmem_cache *cfq_pool; |
60 | static struct kmem_cache *cfq_ioc_pool; | |
1da177e4 | 61 | |
245b2e70 | 62 | static DEFINE_PER_CPU(unsigned long, cfq_ioc_count); |
334e94de | 63 | static struct completion *ioc_gone; |
9a11b4ed | 64 | static DEFINE_SPINLOCK(ioc_gone_lock); |
334e94de | 65 | |
22e2c507 JA |
66 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR |
67 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
22e2c507 JA |
68 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) |
69 | ||
206dc69b | 70 | #define sample_valid(samples) ((samples) > 80) |
1fa8f6d6 | 71 | #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node) |
206dc69b | 72 | |
cc09e299 JA |
73 | /* |
74 | * Most of our rbtree usage is for sorting with min extraction, so | |
75 | * if we cache the leftmost node we don't have to walk down the tree | |
76 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
77 | * move this into the elevator for the rq sorting as well. | |
78 | */ | |
79 | struct cfq_rb_root { | |
80 | struct rb_root rb; | |
81 | struct rb_node *left; | |
aa6f6a3d | 82 | unsigned count; |
1fa8f6d6 | 83 | u64 min_vdisktime; |
25bc6b07 | 84 | struct rb_node *active; |
58ff82f3 | 85 | unsigned total_weight; |
cc09e299 | 86 | }; |
1fa8f6d6 | 87 | #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, } |
cc09e299 | 88 | |
6118b70b JA |
89 | /* |
90 | * Per process-grouping structure | |
91 | */ | |
92 | struct cfq_queue { | |
93 | /* reference count */ | |
94 | atomic_t ref; | |
95 | /* various state flags, see below */ | |
96 | unsigned int flags; | |
97 | /* parent cfq_data */ | |
98 | struct cfq_data *cfqd; | |
99 | /* service_tree member */ | |
100 | struct rb_node rb_node; | |
101 | /* service_tree key */ | |
102 | unsigned long rb_key; | |
103 | /* prio tree member */ | |
104 | struct rb_node p_node; | |
105 | /* prio tree root we belong to, if any */ | |
106 | struct rb_root *p_root; | |
107 | /* sorted list of pending requests */ | |
108 | struct rb_root sort_list; | |
109 | /* if fifo isn't expired, next request to serve */ | |
110 | struct request *next_rq; | |
111 | /* requests queued in sort_list */ | |
112 | int queued[2]; | |
113 | /* currently allocated requests */ | |
114 | int allocated[2]; | |
115 | /* fifo list of requests in sort_list */ | |
116 | struct list_head fifo; | |
117 | ||
dae739eb VG |
118 | /* time when queue got scheduled in to dispatch first request. */ |
119 | unsigned long dispatch_start; | |
f75edf2d | 120 | unsigned int allocated_slice; |
dae739eb VG |
121 | /* time when first request from queue completed and slice started. */ |
122 | unsigned long slice_start; | |
6118b70b JA |
123 | unsigned long slice_end; |
124 | long slice_resid; | |
125 | unsigned int slice_dispatch; | |
126 | ||
127 | /* pending metadata requests */ | |
128 | int meta_pending; | |
129 | /* number of requests that are on the dispatch list or inside driver */ | |
130 | int dispatched; | |
131 | ||
132 | /* io prio of this group */ | |
133 | unsigned short ioprio, org_ioprio; | |
134 | unsigned short ioprio_class, org_ioprio_class; | |
135 | ||
b2c18e1e JM |
136 | unsigned int seek_samples; |
137 | u64 seek_total; | |
138 | sector_t seek_mean; | |
139 | sector_t last_request_pos; | |
e6c5bc73 | 140 | unsigned long seeky_start; |
b2c18e1e | 141 | |
6118b70b | 142 | pid_t pid; |
df5fe3e8 | 143 | |
aa6f6a3d | 144 | struct cfq_rb_root *service_tree; |
df5fe3e8 | 145 | struct cfq_queue *new_cfqq; |
cdb16e8f | 146 | struct cfq_group *cfqg; |
ae30c286 | 147 | struct cfq_group *orig_cfqg; |
22084190 VG |
148 | /* Sectors dispatched in current dispatch round */ |
149 | unsigned long nr_sectors; | |
6118b70b JA |
150 | }; |
151 | ||
c0324a02 | 152 | /* |
718eee05 | 153 | * First index in the service_trees. |
c0324a02 CZ |
154 | * IDLE is handled separately, so it has negative index |
155 | */ | |
156 | enum wl_prio_t { | |
c0324a02 | 157 | BE_WORKLOAD = 0, |
615f0259 VG |
158 | RT_WORKLOAD = 1, |
159 | IDLE_WORKLOAD = 2, | |
c0324a02 CZ |
160 | }; |
161 | ||
718eee05 CZ |
162 | /* |
163 | * Second index in the service_trees. | |
164 | */ | |
165 | enum wl_type_t { | |
166 | ASYNC_WORKLOAD = 0, | |
167 | SYNC_NOIDLE_WORKLOAD = 1, | |
168 | SYNC_WORKLOAD = 2 | |
169 | }; | |
170 | ||
cdb16e8f VG |
171 | /* This is per cgroup per device grouping structure */ |
172 | struct cfq_group { | |
1fa8f6d6 VG |
173 | /* group service_tree member */ |
174 | struct rb_node rb_node; | |
175 | ||
176 | /* group service_tree key */ | |
177 | u64 vdisktime; | |
25bc6b07 | 178 | unsigned int weight; |
1fa8f6d6 VG |
179 | bool on_st; |
180 | ||
181 | /* number of cfqq currently on this group */ | |
182 | int nr_cfqq; | |
183 | ||
58ff82f3 VG |
184 | /* Per group busy queus average. Useful for workload slice calc. */ |
185 | unsigned int busy_queues_avg[2]; | |
cdb16e8f VG |
186 | /* |
187 | * rr lists of queues with requests, onle rr for each priority class. | |
188 | * Counts are embedded in the cfq_rb_root | |
189 | */ | |
190 | struct cfq_rb_root service_trees[2][3]; | |
191 | struct cfq_rb_root service_tree_idle; | |
dae739eb VG |
192 | |
193 | unsigned long saved_workload_slice; | |
194 | enum wl_type_t saved_workload; | |
195 | enum wl_prio_t saved_serving_prio; | |
25fb5169 VG |
196 | struct blkio_group blkg; |
197 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
198 | struct hlist_node cfqd_node; | |
b1c35769 | 199 | atomic_t ref; |
25fb5169 | 200 | #endif |
cdb16e8f | 201 | }; |
718eee05 | 202 | |
22e2c507 JA |
203 | /* |
204 | * Per block device queue structure | |
205 | */ | |
1da177e4 | 206 | struct cfq_data { |
165125e1 | 207 | struct request_queue *queue; |
1fa8f6d6 VG |
208 | /* Root service tree for cfq_groups */ |
209 | struct cfq_rb_root grp_service_tree; | |
cdb16e8f | 210 | struct cfq_group root_group; |
58ff82f3 VG |
211 | /* Number of active cfq groups on group service tree */ |
212 | int nr_groups; | |
22e2c507 | 213 | |
c0324a02 CZ |
214 | /* |
215 | * The priority currently being served | |
22e2c507 | 216 | */ |
c0324a02 | 217 | enum wl_prio_t serving_prio; |
718eee05 CZ |
218 | enum wl_type_t serving_type; |
219 | unsigned long workload_expires; | |
cdb16e8f | 220 | struct cfq_group *serving_group; |
8e550632 | 221 | bool noidle_tree_requires_idle; |
a36e71f9 JA |
222 | |
223 | /* | |
224 | * Each priority tree is sorted by next_request position. These | |
225 | * trees are used when determining if two or more queues are | |
226 | * interleaving requests (see cfq_close_cooperator). | |
227 | */ | |
228 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
229 | ||
22e2c507 JA |
230 | unsigned int busy_queues; |
231 | ||
5ad531db | 232 | int rq_in_driver[2]; |
3ed9a296 | 233 | int sync_flight; |
45333d5a AC |
234 | |
235 | /* | |
236 | * queue-depth detection | |
237 | */ | |
238 | int rq_queued; | |
25776e35 | 239 | int hw_tag; |
e459dd08 CZ |
240 | /* |
241 | * hw_tag can be | |
242 | * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) | |
243 | * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) | |
244 | * 0 => no NCQ | |
245 | */ | |
246 | int hw_tag_est_depth; | |
247 | unsigned int hw_tag_samples; | |
1da177e4 | 248 | |
22e2c507 JA |
249 | /* |
250 | * idle window management | |
251 | */ | |
252 | struct timer_list idle_slice_timer; | |
23e018a1 | 253 | struct work_struct unplug_work; |
1da177e4 | 254 | |
22e2c507 JA |
255 | struct cfq_queue *active_queue; |
256 | struct cfq_io_context *active_cic; | |
22e2c507 | 257 | |
c2dea2d1 VT |
258 | /* |
259 | * async queue for each priority case | |
260 | */ | |
261 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
262 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 263 | |
6d048f53 | 264 | sector_t last_position; |
1da177e4 | 265 | |
1da177e4 LT |
266 | /* |
267 | * tunables, see top of file | |
268 | */ | |
269 | unsigned int cfq_quantum; | |
22e2c507 | 270 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
271 | unsigned int cfq_back_penalty; |
272 | unsigned int cfq_back_max; | |
22e2c507 JA |
273 | unsigned int cfq_slice[2]; |
274 | unsigned int cfq_slice_async_rq; | |
275 | unsigned int cfq_slice_idle; | |
963b72fc | 276 | unsigned int cfq_latency; |
ae30c286 | 277 | unsigned int cfq_group_isolation; |
d9ff4187 AV |
278 | |
279 | struct list_head cic_list; | |
1da177e4 | 280 | |
6118b70b JA |
281 | /* |
282 | * Fallback dummy cfqq for extreme OOM conditions | |
283 | */ | |
284 | struct cfq_queue oom_cfqq; | |
365722bb VG |
285 | |
286 | unsigned long last_end_sync_rq; | |
25fb5169 VG |
287 | |
288 | /* List of cfq groups being managed on this device*/ | |
289 | struct hlist_head cfqg_list; | |
1da177e4 LT |
290 | }; |
291 | ||
25fb5169 VG |
292 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd); |
293 | ||
cdb16e8f VG |
294 | static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, |
295 | enum wl_prio_t prio, | |
718eee05 | 296 | enum wl_type_t type, |
c0324a02 CZ |
297 | struct cfq_data *cfqd) |
298 | { | |
1fa8f6d6 VG |
299 | if (!cfqg) |
300 | return NULL; | |
301 | ||
c0324a02 | 302 | if (prio == IDLE_WORKLOAD) |
cdb16e8f | 303 | return &cfqg->service_tree_idle; |
c0324a02 | 304 | |
cdb16e8f | 305 | return &cfqg->service_trees[prio][type]; |
c0324a02 CZ |
306 | } |
307 | ||
3b18152c | 308 | enum cfqq_state_flags { |
b0b8d749 JA |
309 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
310 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 311 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 | 312 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ |
b0b8d749 JA |
313 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
314 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
315 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 316 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 317 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
b3b6d040 | 318 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ |
76280aff | 319 | CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ |
f75edf2d VG |
320 | CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */ |
321 | CFQ_CFQQ_FLAG_wait_busy_done, /* Got new request. Expire the queue */ | |
3b18152c JA |
322 | }; |
323 | ||
324 | #define CFQ_CFQQ_FNS(name) \ | |
325 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
326 | { \ | |
fe094d98 | 327 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
328 | } \ |
329 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
330 | { \ | |
fe094d98 | 331 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
332 | } \ |
333 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
334 | { \ | |
fe094d98 | 335 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
336 | } |
337 | ||
338 | CFQ_CFQQ_FNS(on_rr); | |
339 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 340 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c | 341 | CFQ_CFQQ_FNS(must_alloc_slice); |
3b18152c JA |
342 | CFQ_CFQQ_FNS(fifo_expire); |
343 | CFQ_CFQQ_FNS(idle_window); | |
344 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 345 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 346 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 347 | CFQ_CFQQ_FNS(coop); |
76280aff | 348 | CFQ_CFQQ_FNS(deep); |
f75edf2d VG |
349 | CFQ_CFQQ_FNS(wait_busy); |
350 | CFQ_CFQQ_FNS(wait_busy_done); | |
3b18152c JA |
351 | #undef CFQ_CFQQ_FNS |
352 | ||
2868ef7b VG |
353 | #ifdef CONFIG_DEBUG_CFQ_IOSCHED |
354 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ | |
355 | blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \ | |
356 | cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \ | |
357 | blkg_path(&(cfqq)->cfqg->blkg), ##args); | |
358 | ||
359 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \ | |
360 | blk_add_trace_msg((cfqd)->queue, "%s " fmt, \ | |
361 | blkg_path(&(cfqg)->blkg), ##args); \ | |
362 | ||
363 | #else | |
7b679138 JA |
364 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
365 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
2868ef7b VG |
366 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0); |
367 | #endif | |
7b679138 JA |
368 | #define cfq_log(cfqd, fmt, args...) \ |
369 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
370 | ||
615f0259 VG |
371 | /* Traverses through cfq group service trees */ |
372 | #define for_each_cfqg_st(cfqg, i, j, st) \ | |
373 | for (i = 0; i <= IDLE_WORKLOAD; i++) \ | |
374 | for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ | |
375 | : &cfqg->service_tree_idle; \ | |
376 | (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ | |
377 | (i == IDLE_WORKLOAD && j == 0); \ | |
378 | j++, st = i < IDLE_WORKLOAD ? \ | |
379 | &cfqg->service_trees[i][j]: NULL) \ | |
380 | ||
381 | ||
c0324a02 CZ |
382 | static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) |
383 | { | |
384 | if (cfq_class_idle(cfqq)) | |
385 | return IDLE_WORKLOAD; | |
386 | if (cfq_class_rt(cfqq)) | |
387 | return RT_WORKLOAD; | |
388 | return BE_WORKLOAD; | |
389 | } | |
390 | ||
718eee05 CZ |
391 | |
392 | static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) | |
393 | { | |
394 | if (!cfq_cfqq_sync(cfqq)) | |
395 | return ASYNC_WORKLOAD; | |
396 | if (!cfq_cfqq_idle_window(cfqq)) | |
397 | return SYNC_NOIDLE_WORKLOAD; | |
398 | return SYNC_WORKLOAD; | |
399 | } | |
400 | ||
58ff82f3 VG |
401 | static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl, |
402 | struct cfq_data *cfqd, | |
403 | struct cfq_group *cfqg) | |
c0324a02 CZ |
404 | { |
405 | if (wl == IDLE_WORKLOAD) | |
cdb16e8f | 406 | return cfqg->service_tree_idle.count; |
c0324a02 | 407 | |
cdb16e8f VG |
408 | return cfqg->service_trees[wl][ASYNC_WORKLOAD].count |
409 | + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count | |
410 | + cfqg->service_trees[wl][SYNC_WORKLOAD].count; | |
c0324a02 CZ |
411 | } |
412 | ||
f26bd1f0 VG |
413 | static inline int cfqg_busy_async_queues(struct cfq_data *cfqd, |
414 | struct cfq_group *cfqg) | |
415 | { | |
416 | return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count | |
417 | + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count; | |
418 | } | |
419 | ||
165125e1 | 420 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
a6151c3a | 421 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, |
fd0928df | 422 | struct io_context *, gfp_t); |
4ac845a2 | 423 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
424 | struct io_context *); |
425 | ||
5ad531db JA |
426 | static inline int rq_in_driver(struct cfq_data *cfqd) |
427 | { | |
428 | return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1]; | |
429 | } | |
430 | ||
91fac317 | 431 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, |
a6151c3a | 432 | bool is_sync) |
91fac317 | 433 | { |
a6151c3a | 434 | return cic->cfqq[is_sync]; |
91fac317 VT |
435 | } |
436 | ||
437 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 438 | struct cfq_queue *cfqq, bool is_sync) |
91fac317 | 439 | { |
a6151c3a | 440 | cic->cfqq[is_sync] = cfqq; |
91fac317 VT |
441 | } |
442 | ||
443 | /* | |
444 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
445 | * set (in which case it could also be direct WRITE). | |
446 | */ | |
a6151c3a | 447 | static inline bool cfq_bio_sync(struct bio *bio) |
91fac317 | 448 | { |
a6151c3a | 449 | return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO); |
91fac317 | 450 | } |
1da177e4 | 451 | |
99f95e52 AM |
452 | /* |
453 | * scheduler run of queue, if there are requests pending and no one in the | |
454 | * driver that will restart queueing | |
455 | */ | |
23e018a1 | 456 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) |
99f95e52 | 457 | { |
7b679138 JA |
458 | if (cfqd->busy_queues) { |
459 | cfq_log(cfqd, "schedule dispatch"); | |
23e018a1 | 460 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 461 | } |
99f95e52 AM |
462 | } |
463 | ||
165125e1 | 464 | static int cfq_queue_empty(struct request_queue *q) |
99f95e52 AM |
465 | { |
466 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
467 | ||
f04a6424 | 468 | return !cfqd->rq_queued; |
99f95e52 AM |
469 | } |
470 | ||
44f7c160 JA |
471 | /* |
472 | * Scale schedule slice based on io priority. Use the sync time slice only | |
473 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
474 | * io only, should not get full sync slice length. | |
475 | */ | |
a6151c3a | 476 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, |
d9e7620e | 477 | unsigned short prio) |
44f7c160 | 478 | { |
d9e7620e | 479 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 480 | |
d9e7620e JA |
481 | WARN_ON(prio >= IOPRIO_BE_NR); |
482 | ||
483 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
484 | } | |
44f7c160 | 485 | |
d9e7620e JA |
486 | static inline int |
487 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
488 | { | |
489 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
490 | } |
491 | ||
25bc6b07 VG |
492 | static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg) |
493 | { | |
494 | u64 d = delta << CFQ_SERVICE_SHIFT; | |
495 | ||
496 | d = d * BLKIO_WEIGHT_DEFAULT; | |
497 | do_div(d, cfqg->weight); | |
498 | return d; | |
499 | } | |
500 | ||
501 | static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
502 | { | |
503 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
504 | if (delta > 0) | |
505 | min_vdisktime = vdisktime; | |
506 | ||
507 | return min_vdisktime; | |
508 | } | |
509 | ||
510 | static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
511 | { | |
512 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
513 | if (delta < 0) | |
514 | min_vdisktime = vdisktime; | |
515 | ||
516 | return min_vdisktime; | |
517 | } | |
518 | ||
519 | static void update_min_vdisktime(struct cfq_rb_root *st) | |
520 | { | |
521 | u64 vdisktime = st->min_vdisktime; | |
522 | struct cfq_group *cfqg; | |
523 | ||
524 | if (st->active) { | |
525 | cfqg = rb_entry_cfqg(st->active); | |
526 | vdisktime = cfqg->vdisktime; | |
527 | } | |
528 | ||
529 | if (st->left) { | |
530 | cfqg = rb_entry_cfqg(st->left); | |
531 | vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime); | |
532 | } | |
533 | ||
534 | st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime); | |
535 | } | |
536 | ||
5db5d642 CZ |
537 | /* |
538 | * get averaged number of queues of RT/BE priority. | |
539 | * average is updated, with a formula that gives more weight to higher numbers, | |
540 | * to quickly follows sudden increases and decrease slowly | |
541 | */ | |
542 | ||
58ff82f3 VG |
543 | static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, |
544 | struct cfq_group *cfqg, bool rt) | |
5869619c | 545 | { |
5db5d642 CZ |
546 | unsigned min_q, max_q; |
547 | unsigned mult = cfq_hist_divisor - 1; | |
548 | unsigned round = cfq_hist_divisor / 2; | |
58ff82f3 | 549 | unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); |
5db5d642 | 550 | |
58ff82f3 VG |
551 | min_q = min(cfqg->busy_queues_avg[rt], busy); |
552 | max_q = max(cfqg->busy_queues_avg[rt], busy); | |
553 | cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
5db5d642 | 554 | cfq_hist_divisor; |
58ff82f3 VG |
555 | return cfqg->busy_queues_avg[rt]; |
556 | } | |
557 | ||
558 | static inline unsigned | |
559 | cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
560 | { | |
561 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
562 | ||
563 | return cfq_target_latency * cfqg->weight / st->total_weight; | |
5db5d642 CZ |
564 | } |
565 | ||
44f7c160 JA |
566 | static inline void |
567 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
568 | { | |
5db5d642 CZ |
569 | unsigned slice = cfq_prio_to_slice(cfqd, cfqq); |
570 | if (cfqd->cfq_latency) { | |
58ff82f3 VG |
571 | /* |
572 | * interested queues (we consider only the ones with the same | |
573 | * priority class in the cfq group) | |
574 | */ | |
575 | unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, | |
576 | cfq_class_rt(cfqq)); | |
5db5d642 CZ |
577 | unsigned sync_slice = cfqd->cfq_slice[1]; |
578 | unsigned expect_latency = sync_slice * iq; | |
58ff82f3 VG |
579 | unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg); |
580 | ||
581 | if (expect_latency > group_slice) { | |
5db5d642 CZ |
582 | unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; |
583 | /* scale low_slice according to IO priority | |
584 | * and sync vs async */ | |
585 | unsigned low_slice = | |
586 | min(slice, base_low_slice * slice / sync_slice); | |
587 | /* the adapted slice value is scaled to fit all iqs | |
588 | * into the target latency */ | |
58ff82f3 | 589 | slice = max(slice * group_slice / expect_latency, |
5db5d642 CZ |
590 | low_slice); |
591 | } | |
592 | } | |
dae739eb | 593 | cfqq->slice_start = jiffies; |
5db5d642 | 594 | cfqq->slice_end = jiffies + slice; |
f75edf2d | 595 | cfqq->allocated_slice = slice; |
7b679138 | 596 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
597 | } |
598 | ||
599 | /* | |
600 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
601 | * isn't valid until the first request from the dispatch is activated | |
602 | * and the slice time set. | |
603 | */ | |
a6151c3a | 604 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) |
44f7c160 JA |
605 | { |
606 | if (cfq_cfqq_slice_new(cfqq)) | |
607 | return 0; | |
608 | if (time_before(jiffies, cfqq->slice_end)) | |
609 | return 0; | |
610 | ||
611 | return 1; | |
612 | } | |
613 | ||
1da177e4 | 614 | /* |
5e705374 | 615 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 616 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 617 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 618 | */ |
5e705374 | 619 | static struct request * |
cf7c25cf | 620 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) |
1da177e4 | 621 | { |
cf7c25cf | 622 | sector_t s1, s2, d1 = 0, d2 = 0; |
1da177e4 | 623 | unsigned long back_max; |
e8a99053 AM |
624 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
625 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
626 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 627 | |
5e705374 JA |
628 | if (rq1 == NULL || rq1 == rq2) |
629 | return rq2; | |
630 | if (rq2 == NULL) | |
631 | return rq1; | |
9c2c38a1 | 632 | |
5e705374 JA |
633 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
634 | return rq1; | |
635 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
636 | return rq2; | |
374f84ac JA |
637 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
638 | return rq1; | |
639 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
640 | return rq2; | |
1da177e4 | 641 | |
83096ebf TH |
642 | s1 = blk_rq_pos(rq1); |
643 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 644 | |
1da177e4 LT |
645 | /* |
646 | * by definition, 1KiB is 2 sectors | |
647 | */ | |
648 | back_max = cfqd->cfq_back_max * 2; | |
649 | ||
650 | /* | |
651 | * Strict one way elevator _except_ in the case where we allow | |
652 | * short backward seeks which are biased as twice the cost of a | |
653 | * similar forward seek. | |
654 | */ | |
655 | if (s1 >= last) | |
656 | d1 = s1 - last; | |
657 | else if (s1 + back_max >= last) | |
658 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
659 | else | |
e8a99053 | 660 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
661 | |
662 | if (s2 >= last) | |
663 | d2 = s2 - last; | |
664 | else if (s2 + back_max >= last) | |
665 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
666 | else | |
e8a99053 | 667 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
668 | |
669 | /* Found required data */ | |
e8a99053 AM |
670 | |
671 | /* | |
672 | * By doing switch() on the bit mask "wrap" we avoid having to | |
673 | * check two variables for all permutations: --> faster! | |
674 | */ | |
675 | switch (wrap) { | |
5e705374 | 676 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 677 | if (d1 < d2) |
5e705374 | 678 | return rq1; |
e8a99053 | 679 | else if (d2 < d1) |
5e705374 | 680 | return rq2; |
e8a99053 AM |
681 | else { |
682 | if (s1 >= s2) | |
5e705374 | 683 | return rq1; |
e8a99053 | 684 | else |
5e705374 | 685 | return rq2; |
e8a99053 | 686 | } |
1da177e4 | 687 | |
e8a99053 | 688 | case CFQ_RQ2_WRAP: |
5e705374 | 689 | return rq1; |
e8a99053 | 690 | case CFQ_RQ1_WRAP: |
5e705374 JA |
691 | return rq2; |
692 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
693 | default: |
694 | /* | |
695 | * Since both rqs are wrapped, | |
696 | * start with the one that's further behind head | |
697 | * (--> only *one* back seek required), | |
698 | * since back seek takes more time than forward. | |
699 | */ | |
700 | if (s1 <= s2) | |
5e705374 | 701 | return rq1; |
1da177e4 | 702 | else |
5e705374 | 703 | return rq2; |
1da177e4 LT |
704 | } |
705 | } | |
706 | ||
498d3aa2 JA |
707 | /* |
708 | * The below is leftmost cache rbtree addon | |
709 | */ | |
0871714e | 710 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 | 711 | { |
615f0259 VG |
712 | /* Service tree is empty */ |
713 | if (!root->count) | |
714 | return NULL; | |
715 | ||
cc09e299 JA |
716 | if (!root->left) |
717 | root->left = rb_first(&root->rb); | |
718 | ||
0871714e JA |
719 | if (root->left) |
720 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
721 | ||
722 | return NULL; | |
cc09e299 JA |
723 | } |
724 | ||
1fa8f6d6 VG |
725 | static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) |
726 | { | |
727 | if (!root->left) | |
728 | root->left = rb_first(&root->rb); | |
729 | ||
730 | if (root->left) | |
731 | return rb_entry_cfqg(root->left); | |
732 | ||
733 | return NULL; | |
734 | } | |
735 | ||
a36e71f9 JA |
736 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
737 | { | |
738 | rb_erase(n, root); | |
739 | RB_CLEAR_NODE(n); | |
740 | } | |
741 | ||
cc09e299 JA |
742 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
743 | { | |
744 | if (root->left == n) | |
745 | root->left = NULL; | |
a36e71f9 | 746 | rb_erase_init(n, &root->rb); |
aa6f6a3d | 747 | --root->count; |
cc09e299 JA |
748 | } |
749 | ||
1da177e4 LT |
750 | /* |
751 | * would be nice to take fifo expire time into account as well | |
752 | */ | |
5e705374 JA |
753 | static struct request * |
754 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
755 | struct request *last) | |
1da177e4 | 756 | { |
21183b07 JA |
757 | struct rb_node *rbnext = rb_next(&last->rb_node); |
758 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 759 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 760 | |
21183b07 | 761 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
762 | |
763 | if (rbprev) | |
5e705374 | 764 | prev = rb_entry_rq(rbprev); |
1da177e4 | 765 | |
21183b07 | 766 | if (rbnext) |
5e705374 | 767 | next = rb_entry_rq(rbnext); |
21183b07 JA |
768 | else { |
769 | rbnext = rb_first(&cfqq->sort_list); | |
770 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 771 | next = rb_entry_rq(rbnext); |
21183b07 | 772 | } |
1da177e4 | 773 | |
cf7c25cf | 774 | return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); |
1da177e4 LT |
775 | } |
776 | ||
d9e7620e JA |
777 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
778 | struct cfq_queue *cfqq) | |
1da177e4 | 779 | { |
d9e7620e JA |
780 | /* |
781 | * just an approximation, should be ok. | |
782 | */ | |
cdb16e8f | 783 | return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
464191c6 | 784 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); |
d9e7620e JA |
785 | } |
786 | ||
1fa8f6d6 VG |
787 | static inline s64 |
788 | cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
789 | { | |
790 | return cfqg->vdisktime - st->min_vdisktime; | |
791 | } | |
792 | ||
793 | static void | |
794 | __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
795 | { | |
796 | struct rb_node **node = &st->rb.rb_node; | |
797 | struct rb_node *parent = NULL; | |
798 | struct cfq_group *__cfqg; | |
799 | s64 key = cfqg_key(st, cfqg); | |
800 | int left = 1; | |
801 | ||
802 | while (*node != NULL) { | |
803 | parent = *node; | |
804 | __cfqg = rb_entry_cfqg(parent); | |
805 | ||
806 | if (key < cfqg_key(st, __cfqg)) | |
807 | node = &parent->rb_left; | |
808 | else { | |
809 | node = &parent->rb_right; | |
810 | left = 0; | |
811 | } | |
812 | } | |
813 | ||
814 | if (left) | |
815 | st->left = &cfqg->rb_node; | |
816 | ||
817 | rb_link_node(&cfqg->rb_node, parent, node); | |
818 | rb_insert_color(&cfqg->rb_node, &st->rb); | |
819 | } | |
820 | ||
821 | static void | |
822 | cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
823 | { | |
824 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
825 | struct cfq_group *__cfqg; | |
826 | struct rb_node *n; | |
827 | ||
828 | cfqg->nr_cfqq++; | |
829 | if (cfqg->on_st) | |
830 | return; | |
831 | ||
832 | /* | |
833 | * Currently put the group at the end. Later implement something | |
834 | * so that groups get lesser vtime based on their weights, so that | |
835 | * if group does not loose all if it was not continously backlogged. | |
836 | */ | |
837 | n = rb_last(&st->rb); | |
838 | if (n) { | |
839 | __cfqg = rb_entry_cfqg(n); | |
840 | cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; | |
841 | } else | |
842 | cfqg->vdisktime = st->min_vdisktime; | |
843 | ||
844 | __cfq_group_service_tree_add(st, cfqg); | |
845 | cfqg->on_st = true; | |
58ff82f3 VG |
846 | cfqd->nr_groups++; |
847 | st->total_weight += cfqg->weight; | |
1fa8f6d6 VG |
848 | } |
849 | ||
850 | static void | |
851 | cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
852 | { | |
853 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
854 | ||
25bc6b07 VG |
855 | if (st->active == &cfqg->rb_node) |
856 | st->active = NULL; | |
857 | ||
1fa8f6d6 VG |
858 | BUG_ON(cfqg->nr_cfqq < 1); |
859 | cfqg->nr_cfqq--; | |
25bc6b07 | 860 | |
1fa8f6d6 VG |
861 | /* If there are other cfq queues under this group, don't delete it */ |
862 | if (cfqg->nr_cfqq) | |
863 | return; | |
864 | ||
2868ef7b | 865 | cfq_log_cfqg(cfqd, cfqg, "del_from_rr group"); |
1fa8f6d6 | 866 | cfqg->on_st = false; |
58ff82f3 VG |
867 | cfqd->nr_groups--; |
868 | st->total_weight -= cfqg->weight; | |
1fa8f6d6 VG |
869 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) |
870 | cfq_rb_erase(&cfqg->rb_node, st); | |
dae739eb | 871 | cfqg->saved_workload_slice = 0; |
22084190 | 872 | blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1); |
dae739eb VG |
873 | } |
874 | ||
875 | static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq) | |
876 | { | |
f75edf2d | 877 | unsigned int slice_used; |
dae739eb VG |
878 | |
879 | /* | |
880 | * Queue got expired before even a single request completed or | |
881 | * got expired immediately after first request completion. | |
882 | */ | |
883 | if (!cfqq->slice_start || cfqq->slice_start == jiffies) { | |
884 | /* | |
885 | * Also charge the seek time incurred to the group, otherwise | |
886 | * if there are mutiple queues in the group, each can dispatch | |
887 | * a single request on seeky media and cause lots of seek time | |
888 | * and group will never know it. | |
889 | */ | |
890 | slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start), | |
891 | 1); | |
892 | } else { | |
893 | slice_used = jiffies - cfqq->slice_start; | |
f75edf2d VG |
894 | if (slice_used > cfqq->allocated_slice) |
895 | slice_used = cfqq->allocated_slice; | |
dae739eb VG |
896 | } |
897 | ||
22084190 VG |
898 | cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used, |
899 | cfqq->nr_sectors); | |
dae739eb VG |
900 | return slice_used; |
901 | } | |
902 | ||
903 | static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg, | |
904 | struct cfq_queue *cfqq) | |
905 | { | |
906 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
f26bd1f0 VG |
907 | unsigned int used_sl, charge_sl; |
908 | int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg) | |
909 | - cfqg->service_tree_idle.count; | |
910 | ||
911 | BUG_ON(nr_sync < 0); | |
912 | used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq); | |
dae739eb | 913 | |
f26bd1f0 VG |
914 | if (!cfq_cfqq_sync(cfqq) && !nr_sync) |
915 | charge_sl = cfqq->allocated_slice; | |
dae739eb VG |
916 | |
917 | /* Can't update vdisktime while group is on service tree */ | |
918 | cfq_rb_erase(&cfqg->rb_node, st); | |
f26bd1f0 | 919 | cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg); |
dae739eb VG |
920 | __cfq_group_service_tree_add(st, cfqg); |
921 | ||
922 | /* This group is being expired. Save the context */ | |
923 | if (time_after(cfqd->workload_expires, jiffies)) { | |
924 | cfqg->saved_workload_slice = cfqd->workload_expires | |
925 | - jiffies; | |
926 | cfqg->saved_workload = cfqd->serving_type; | |
927 | cfqg->saved_serving_prio = cfqd->serving_prio; | |
928 | } else | |
929 | cfqg->saved_workload_slice = 0; | |
2868ef7b VG |
930 | |
931 | cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime, | |
932 | st->min_vdisktime); | |
22084190 VG |
933 | blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl, |
934 | cfqq->nr_sectors); | |
1fa8f6d6 VG |
935 | } |
936 | ||
25fb5169 VG |
937 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
938 | static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg) | |
939 | { | |
940 | if (blkg) | |
941 | return container_of(blkg, struct cfq_group, blkg); | |
942 | return NULL; | |
943 | } | |
944 | ||
f8d461d6 VG |
945 | void |
946 | cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight) | |
947 | { | |
948 | cfqg_of_blkg(blkg)->weight = weight; | |
949 | } | |
950 | ||
25fb5169 VG |
951 | static struct cfq_group * |
952 | cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create) | |
953 | { | |
954 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup); | |
955 | struct cfq_group *cfqg = NULL; | |
956 | void *key = cfqd; | |
957 | int i, j; | |
958 | struct cfq_rb_root *st; | |
22084190 VG |
959 | struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; |
960 | unsigned int major, minor; | |
25fb5169 VG |
961 | |
962 | /* Do we need to take this reference */ | |
963 | if (!css_tryget(&blkcg->css)) | |
964 | return NULL;; | |
965 | ||
966 | cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key)); | |
967 | if (cfqg || !create) | |
968 | goto done; | |
969 | ||
970 | cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node); | |
971 | if (!cfqg) | |
972 | goto done; | |
973 | ||
974 | cfqg->weight = blkcg->weight; | |
975 | for_each_cfqg_st(cfqg, i, j, st) | |
976 | *st = CFQ_RB_ROOT; | |
977 | RB_CLEAR_NODE(&cfqg->rb_node); | |
978 | ||
b1c35769 VG |
979 | /* |
980 | * Take the initial reference that will be released on destroy | |
981 | * This can be thought of a joint reference by cgroup and | |
982 | * elevator which will be dropped by either elevator exit | |
983 | * or cgroup deletion path depending on who is exiting first. | |
984 | */ | |
985 | atomic_set(&cfqg->ref, 1); | |
986 | ||
25fb5169 | 987 | /* Add group onto cgroup list */ |
22084190 VG |
988 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); |
989 | blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd, | |
990 | MKDEV(major, minor)); | |
25fb5169 VG |
991 | |
992 | /* Add group on cfqd list */ | |
993 | hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); | |
994 | ||
995 | done: | |
996 | css_put(&blkcg->css); | |
997 | return cfqg; | |
998 | } | |
999 | ||
1000 | /* | |
1001 | * Search for the cfq group current task belongs to. If create = 1, then also | |
1002 | * create the cfq group if it does not exist. request_queue lock must be held. | |
1003 | */ | |
1004 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create) | |
1005 | { | |
1006 | struct cgroup *cgroup; | |
1007 | struct cfq_group *cfqg = NULL; | |
1008 | ||
1009 | rcu_read_lock(); | |
1010 | cgroup = task_cgroup(current, blkio_subsys_id); | |
1011 | cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create); | |
1012 | if (!cfqg && create) | |
1013 | cfqg = &cfqd->root_group; | |
1014 | rcu_read_unlock(); | |
1015 | return cfqg; | |
1016 | } | |
1017 | ||
1018 | static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) | |
1019 | { | |
1020 | /* Currently, all async queues are mapped to root group */ | |
1021 | if (!cfq_cfqq_sync(cfqq)) | |
1022 | cfqg = &cfqq->cfqd->root_group; | |
1023 | ||
1024 | cfqq->cfqg = cfqg; | |
b1c35769 VG |
1025 | /* cfqq reference on cfqg */ |
1026 | atomic_inc(&cfqq->cfqg->ref); | |
1027 | } | |
1028 | ||
1029 | static void cfq_put_cfqg(struct cfq_group *cfqg) | |
1030 | { | |
1031 | struct cfq_rb_root *st; | |
1032 | int i, j; | |
1033 | ||
1034 | BUG_ON(atomic_read(&cfqg->ref) <= 0); | |
1035 | if (!atomic_dec_and_test(&cfqg->ref)) | |
1036 | return; | |
1037 | for_each_cfqg_st(cfqg, i, j, st) | |
1038 | BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL); | |
1039 | kfree(cfqg); | |
1040 | } | |
1041 | ||
1042 | static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1043 | { | |
1044 | /* Something wrong if we are trying to remove same group twice */ | |
1045 | BUG_ON(hlist_unhashed(&cfqg->cfqd_node)); | |
1046 | ||
1047 | hlist_del_init(&cfqg->cfqd_node); | |
1048 | ||
1049 | /* | |
1050 | * Put the reference taken at the time of creation so that when all | |
1051 | * queues are gone, group can be destroyed. | |
1052 | */ | |
1053 | cfq_put_cfqg(cfqg); | |
1054 | } | |
1055 | ||
1056 | static void cfq_release_cfq_groups(struct cfq_data *cfqd) | |
1057 | { | |
1058 | struct hlist_node *pos, *n; | |
1059 | struct cfq_group *cfqg; | |
1060 | ||
1061 | hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) { | |
1062 | /* | |
1063 | * If cgroup removal path got to blk_group first and removed | |
1064 | * it from cgroup list, then it will take care of destroying | |
1065 | * cfqg also. | |
1066 | */ | |
1067 | if (!blkiocg_del_blkio_group(&cfqg->blkg)) | |
1068 | cfq_destroy_cfqg(cfqd, cfqg); | |
1069 | } | |
25fb5169 | 1070 | } |
b1c35769 VG |
1071 | |
1072 | /* | |
1073 | * Blk cgroup controller notification saying that blkio_group object is being | |
1074 | * delinked as associated cgroup object is going away. That also means that | |
1075 | * no new IO will come in this group. So get rid of this group as soon as | |
1076 | * any pending IO in the group is finished. | |
1077 | * | |
1078 | * This function is called under rcu_read_lock(). key is the rcu protected | |
1079 | * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu | |
1080 | * read lock. | |
1081 | * | |
1082 | * "key" was fetched from blkio_group under blkio_cgroup->lock. That means | |
1083 | * it should not be NULL as even if elevator was exiting, cgroup deltion | |
1084 | * path got to it first. | |
1085 | */ | |
1086 | void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg) | |
1087 | { | |
1088 | unsigned long flags; | |
1089 | struct cfq_data *cfqd = key; | |
1090 | ||
1091 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
1092 | cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg)); | |
1093 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
1094 | } | |
1095 | ||
25fb5169 VG |
1096 | #else /* GROUP_IOSCHED */ |
1097 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create) | |
1098 | { | |
1099 | return &cfqd->root_group; | |
1100 | } | |
1101 | static inline void | |
1102 | cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) { | |
1103 | cfqq->cfqg = cfqg; | |
1104 | } | |
1105 | ||
b1c35769 VG |
1106 | static void cfq_release_cfq_groups(struct cfq_data *cfqd) {} |
1107 | static inline void cfq_put_cfqg(struct cfq_group *cfqg) {} | |
1108 | ||
25fb5169 VG |
1109 | #endif /* GROUP_IOSCHED */ |
1110 | ||
498d3aa2 | 1111 | /* |
c0324a02 | 1112 | * The cfqd->service_trees holds all pending cfq_queue's that have |
498d3aa2 JA |
1113 | * requests waiting to be processed. It is sorted in the order that |
1114 | * we will service the queues. | |
1115 | */ | |
a36e71f9 | 1116 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 1117 | bool add_front) |
d9e7620e | 1118 | { |
0871714e JA |
1119 | struct rb_node **p, *parent; |
1120 | struct cfq_queue *__cfqq; | |
d9e7620e | 1121 | unsigned long rb_key; |
c0324a02 | 1122 | struct cfq_rb_root *service_tree; |
498d3aa2 | 1123 | int left; |
dae739eb | 1124 | int new_cfqq = 1; |
ae30c286 VG |
1125 | int group_changed = 0; |
1126 | ||
1127 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
1128 | if (!cfqd->cfq_group_isolation | |
1129 | && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD | |
1130 | && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) { | |
1131 | /* Move this cfq to root group */ | |
1132 | cfq_log_cfqq(cfqd, cfqq, "moving to root group"); | |
1133 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) | |
1134 | cfq_group_service_tree_del(cfqd, cfqq->cfqg); | |
1135 | cfqq->orig_cfqg = cfqq->cfqg; | |
1136 | cfqq->cfqg = &cfqd->root_group; | |
1137 | atomic_inc(&cfqd->root_group.ref); | |
1138 | group_changed = 1; | |
1139 | } else if (!cfqd->cfq_group_isolation | |
1140 | && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) { | |
1141 | /* cfqq is sequential now needs to go to its original group */ | |
1142 | BUG_ON(cfqq->cfqg != &cfqd->root_group); | |
1143 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) | |
1144 | cfq_group_service_tree_del(cfqd, cfqq->cfqg); | |
1145 | cfq_put_cfqg(cfqq->cfqg); | |
1146 | cfqq->cfqg = cfqq->orig_cfqg; | |
1147 | cfqq->orig_cfqg = NULL; | |
1148 | group_changed = 1; | |
1149 | cfq_log_cfqq(cfqd, cfqq, "moved to origin group"); | |
1150 | } | |
1151 | #endif | |
d9e7620e | 1152 | |
cdb16e8f VG |
1153 | service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), |
1154 | cfqq_type(cfqq), cfqd); | |
0871714e JA |
1155 | if (cfq_class_idle(cfqq)) { |
1156 | rb_key = CFQ_IDLE_DELAY; | |
aa6f6a3d | 1157 | parent = rb_last(&service_tree->rb); |
0871714e JA |
1158 | if (parent && parent != &cfqq->rb_node) { |
1159 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1160 | rb_key += __cfqq->rb_key; | |
1161 | } else | |
1162 | rb_key += jiffies; | |
1163 | } else if (!add_front) { | |
b9c8946b JA |
1164 | /* |
1165 | * Get our rb key offset. Subtract any residual slice | |
1166 | * value carried from last service. A negative resid | |
1167 | * count indicates slice overrun, and this should position | |
1168 | * the next service time further away in the tree. | |
1169 | */ | |
edd75ffd | 1170 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
b9c8946b | 1171 | rb_key -= cfqq->slice_resid; |
edd75ffd | 1172 | cfqq->slice_resid = 0; |
48e025e6 CZ |
1173 | } else { |
1174 | rb_key = -HZ; | |
aa6f6a3d | 1175 | __cfqq = cfq_rb_first(service_tree); |
48e025e6 CZ |
1176 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; |
1177 | } | |
1da177e4 | 1178 | |
d9e7620e | 1179 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
dae739eb | 1180 | new_cfqq = 0; |
99f9628a | 1181 | /* |
d9e7620e | 1182 | * same position, nothing more to do |
99f9628a | 1183 | */ |
c0324a02 CZ |
1184 | if (rb_key == cfqq->rb_key && |
1185 | cfqq->service_tree == service_tree) | |
d9e7620e | 1186 | return; |
1da177e4 | 1187 | |
aa6f6a3d CZ |
1188 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); |
1189 | cfqq->service_tree = NULL; | |
1da177e4 | 1190 | } |
d9e7620e | 1191 | |
498d3aa2 | 1192 | left = 1; |
0871714e | 1193 | parent = NULL; |
aa6f6a3d CZ |
1194 | cfqq->service_tree = service_tree; |
1195 | p = &service_tree->rb.rb_node; | |
d9e7620e | 1196 | while (*p) { |
67060e37 | 1197 | struct rb_node **n; |
cc09e299 | 1198 | |
d9e7620e JA |
1199 | parent = *p; |
1200 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1201 | ||
0c534e0a | 1202 | /* |
c0324a02 | 1203 | * sort by key, that represents service time. |
0c534e0a | 1204 | */ |
c0324a02 | 1205 | if (time_before(rb_key, __cfqq->rb_key)) |
67060e37 | 1206 | n = &(*p)->rb_left; |
c0324a02 | 1207 | else { |
67060e37 | 1208 | n = &(*p)->rb_right; |
cc09e299 | 1209 | left = 0; |
c0324a02 | 1210 | } |
67060e37 JA |
1211 | |
1212 | p = n; | |
d9e7620e JA |
1213 | } |
1214 | ||
cc09e299 | 1215 | if (left) |
aa6f6a3d | 1216 | service_tree->left = &cfqq->rb_node; |
cc09e299 | 1217 | |
d9e7620e JA |
1218 | cfqq->rb_key = rb_key; |
1219 | rb_link_node(&cfqq->rb_node, parent, p); | |
aa6f6a3d CZ |
1220 | rb_insert_color(&cfqq->rb_node, &service_tree->rb); |
1221 | service_tree->count++; | |
ae30c286 | 1222 | if ((add_front || !new_cfqq) && !group_changed) |
dae739eb | 1223 | return; |
1fa8f6d6 | 1224 | cfq_group_service_tree_add(cfqd, cfqq->cfqg); |
1da177e4 LT |
1225 | } |
1226 | ||
a36e71f9 | 1227 | static struct cfq_queue * |
f2d1f0ae JA |
1228 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
1229 | sector_t sector, struct rb_node **ret_parent, | |
1230 | struct rb_node ***rb_link) | |
a36e71f9 | 1231 | { |
a36e71f9 JA |
1232 | struct rb_node **p, *parent; |
1233 | struct cfq_queue *cfqq = NULL; | |
1234 | ||
1235 | parent = NULL; | |
1236 | p = &root->rb_node; | |
1237 | while (*p) { | |
1238 | struct rb_node **n; | |
1239 | ||
1240 | parent = *p; | |
1241 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
1242 | ||
1243 | /* | |
1244 | * Sort strictly based on sector. Smallest to the left, | |
1245 | * largest to the right. | |
1246 | */ | |
2e46e8b2 | 1247 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 1248 | n = &(*p)->rb_right; |
2e46e8b2 | 1249 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
1250 | n = &(*p)->rb_left; |
1251 | else | |
1252 | break; | |
1253 | p = n; | |
3ac6c9f8 | 1254 | cfqq = NULL; |
a36e71f9 JA |
1255 | } |
1256 | ||
1257 | *ret_parent = parent; | |
1258 | if (rb_link) | |
1259 | *rb_link = p; | |
3ac6c9f8 | 1260 | return cfqq; |
a36e71f9 JA |
1261 | } |
1262 | ||
1263 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1264 | { | |
a36e71f9 JA |
1265 | struct rb_node **p, *parent; |
1266 | struct cfq_queue *__cfqq; | |
1267 | ||
f2d1f0ae JA |
1268 | if (cfqq->p_root) { |
1269 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1270 | cfqq->p_root = NULL; | |
1271 | } | |
a36e71f9 JA |
1272 | |
1273 | if (cfq_class_idle(cfqq)) | |
1274 | return; | |
1275 | if (!cfqq->next_rq) | |
1276 | return; | |
1277 | ||
f2d1f0ae | 1278 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
1279 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
1280 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
1281 | if (!__cfqq) { |
1282 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
1283 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
1284 | } else | |
1285 | cfqq->p_root = NULL; | |
a36e71f9 JA |
1286 | } |
1287 | ||
498d3aa2 JA |
1288 | /* |
1289 | * Update cfqq's position in the service tree. | |
1290 | */ | |
edd75ffd | 1291 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 1292 | { |
6d048f53 JA |
1293 | /* |
1294 | * Resorting requires the cfqq to be on the RR list already. | |
1295 | */ | |
a36e71f9 | 1296 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 1297 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
1298 | cfq_prio_tree_add(cfqd, cfqq); |
1299 | } | |
6d048f53 JA |
1300 | } |
1301 | ||
1da177e4 LT |
1302 | /* |
1303 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 1304 | * the pending list according to last request service |
1da177e4 | 1305 | */ |
febffd61 | 1306 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1307 | { |
7b679138 | 1308 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
1309 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1310 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 LT |
1311 | cfqd->busy_queues++; |
1312 | ||
edd75ffd | 1313 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
1314 | } |
1315 | ||
498d3aa2 JA |
1316 | /* |
1317 | * Called when the cfqq no longer has requests pending, remove it from | |
1318 | * the service tree. | |
1319 | */ | |
febffd61 | 1320 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1321 | { |
7b679138 | 1322 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
1323 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
1324 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 1325 | |
aa6f6a3d CZ |
1326 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
1327 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
1328 | cfqq->service_tree = NULL; | |
1329 | } | |
f2d1f0ae JA |
1330 | if (cfqq->p_root) { |
1331 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1332 | cfqq->p_root = NULL; | |
1333 | } | |
d9e7620e | 1334 | |
1fa8f6d6 | 1335 | cfq_group_service_tree_del(cfqd, cfqq->cfqg); |
1da177e4 LT |
1336 | BUG_ON(!cfqd->busy_queues); |
1337 | cfqd->busy_queues--; | |
1338 | } | |
1339 | ||
1340 | /* | |
1341 | * rb tree support functions | |
1342 | */ | |
febffd61 | 1343 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 1344 | { |
5e705374 | 1345 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
5e705374 | 1346 | const int sync = rq_is_sync(rq); |
1da177e4 | 1347 | |
b4878f24 JA |
1348 | BUG_ON(!cfqq->queued[sync]); |
1349 | cfqq->queued[sync]--; | |
1da177e4 | 1350 | |
5e705374 | 1351 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 1352 | |
f04a6424 VG |
1353 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { |
1354 | /* | |
1355 | * Queue will be deleted from service tree when we actually | |
1356 | * expire it later. Right now just remove it from prio tree | |
1357 | * as it is empty. | |
1358 | */ | |
1359 | if (cfqq->p_root) { | |
1360 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1361 | cfqq->p_root = NULL; | |
1362 | } | |
1363 | } | |
1da177e4 LT |
1364 | } |
1365 | ||
5e705374 | 1366 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 1367 | { |
5e705374 | 1368 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 1369 | struct cfq_data *cfqd = cfqq->cfqd; |
a36e71f9 | 1370 | struct request *__alias, *prev; |
1da177e4 | 1371 | |
5380a101 | 1372 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
1373 | |
1374 | /* | |
1375 | * looks a little odd, but the first insert might return an alias. | |
1376 | * if that happens, put the alias on the dispatch list | |
1377 | */ | |
21183b07 | 1378 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 1379 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
1380 | |
1381 | if (!cfq_cfqq_on_rr(cfqq)) | |
1382 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
1383 | |
1384 | /* | |
1385 | * check if this request is a better next-serve candidate | |
1386 | */ | |
a36e71f9 | 1387 | prev = cfqq->next_rq; |
cf7c25cf | 1388 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); |
a36e71f9 JA |
1389 | |
1390 | /* | |
1391 | * adjust priority tree position, if ->next_rq changes | |
1392 | */ | |
1393 | if (prev != cfqq->next_rq) | |
1394 | cfq_prio_tree_add(cfqd, cfqq); | |
1395 | ||
5044eed4 | 1396 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
1397 | } |
1398 | ||
febffd61 | 1399 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 1400 | { |
5380a101 JA |
1401 | elv_rb_del(&cfqq->sort_list, rq); |
1402 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 1403 | cfq_add_rq_rb(rq); |
1da177e4 LT |
1404 | } |
1405 | ||
206dc69b JA |
1406 | static struct request * |
1407 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 1408 | { |
206dc69b | 1409 | struct task_struct *tsk = current; |
91fac317 | 1410 | struct cfq_io_context *cic; |
206dc69b | 1411 | struct cfq_queue *cfqq; |
1da177e4 | 1412 | |
4ac845a2 | 1413 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
1414 | if (!cic) |
1415 | return NULL; | |
1416 | ||
1417 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
1418 | if (cfqq) { |
1419 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
1420 | ||
21183b07 | 1421 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 1422 | } |
1da177e4 | 1423 | |
1da177e4 LT |
1424 | return NULL; |
1425 | } | |
1426 | ||
165125e1 | 1427 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1428 | { |
22e2c507 | 1429 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 1430 | |
5ad531db | 1431 | cfqd->rq_in_driver[rq_is_sync(rq)]++; |
7b679138 | 1432 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
5ad531db | 1433 | rq_in_driver(cfqd)); |
25776e35 | 1434 | |
5b93629b | 1435 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
1436 | } |
1437 | ||
165125e1 | 1438 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1439 | { |
b4878f24 | 1440 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5ad531db | 1441 | const int sync = rq_is_sync(rq); |
b4878f24 | 1442 | |
5ad531db JA |
1443 | WARN_ON(!cfqd->rq_in_driver[sync]); |
1444 | cfqd->rq_in_driver[sync]--; | |
7b679138 | 1445 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
5ad531db | 1446 | rq_in_driver(cfqd)); |
1da177e4 LT |
1447 | } |
1448 | ||
b4878f24 | 1449 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 1450 | { |
5e705374 | 1451 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 1452 | |
5e705374 JA |
1453 | if (cfqq->next_rq == rq) |
1454 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 1455 | |
b4878f24 | 1456 | list_del_init(&rq->queuelist); |
5e705374 | 1457 | cfq_del_rq_rb(rq); |
374f84ac | 1458 | |
45333d5a | 1459 | cfqq->cfqd->rq_queued--; |
374f84ac JA |
1460 | if (rq_is_meta(rq)) { |
1461 | WARN_ON(!cfqq->meta_pending); | |
1462 | cfqq->meta_pending--; | |
1463 | } | |
1da177e4 LT |
1464 | } |
1465 | ||
165125e1 JA |
1466 | static int cfq_merge(struct request_queue *q, struct request **req, |
1467 | struct bio *bio) | |
1da177e4 LT |
1468 | { |
1469 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1470 | struct request *__rq; | |
1da177e4 | 1471 | |
206dc69b | 1472 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 1473 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
1474 | *req = __rq; |
1475 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1476 | } |
1477 | ||
1478 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1479 | } |
1480 | ||
165125e1 | 1481 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 1482 | int type) |
1da177e4 | 1483 | { |
21183b07 | 1484 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 1485 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 1486 | |
5e705374 | 1487 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 1488 | } |
1da177e4 LT |
1489 | } |
1490 | ||
1491 | static void | |
165125e1 | 1492 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
1493 | struct request *next) |
1494 | { | |
cf7c25cf | 1495 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 JA |
1496 | /* |
1497 | * reposition in fifo if next is older than rq | |
1498 | */ | |
1499 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
30996f40 | 1500 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
22e2c507 | 1501 | list_move(&rq->queuelist, &next->queuelist); |
30996f40 JA |
1502 | rq_set_fifo_time(rq, rq_fifo_time(next)); |
1503 | } | |
22e2c507 | 1504 | |
cf7c25cf CZ |
1505 | if (cfqq->next_rq == next) |
1506 | cfqq->next_rq = rq; | |
b4878f24 | 1507 | cfq_remove_request(next); |
22e2c507 JA |
1508 | } |
1509 | ||
165125e1 | 1510 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
1511 | struct bio *bio) |
1512 | { | |
1513 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 1514 | struct cfq_io_context *cic; |
da775265 | 1515 | struct cfq_queue *cfqq; |
da775265 | 1516 | |
8682e1f1 VG |
1517 | /* Deny merge if bio and rq don't belong to same cfq group */ |
1518 | if ((RQ_CFQQ(rq))->cfqg != cfq_get_cfqg(cfqd, 0)) | |
1519 | return false; | |
da775265 | 1520 | /* |
ec8acb69 | 1521 | * Disallow merge of a sync bio into an async request. |
da775265 | 1522 | */ |
91fac317 | 1523 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
a6151c3a | 1524 | return false; |
da775265 JA |
1525 | |
1526 | /* | |
719d3402 JA |
1527 | * Lookup the cfqq that this bio will be queued with. Allow |
1528 | * merge only if rq is queued there. | |
da775265 | 1529 | */ |
4ac845a2 | 1530 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 | 1531 | if (!cic) |
a6151c3a | 1532 | return false; |
719d3402 | 1533 | |
91fac317 | 1534 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
a6151c3a | 1535 | return cfqq == RQ_CFQQ(rq); |
da775265 JA |
1536 | } |
1537 | ||
febffd61 JA |
1538 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
1539 | struct cfq_queue *cfqq) | |
22e2c507 JA |
1540 | { |
1541 | if (cfqq) { | |
7b679138 | 1542 | cfq_log_cfqq(cfqd, cfqq, "set_active"); |
dae739eb VG |
1543 | cfqq->slice_start = 0; |
1544 | cfqq->dispatch_start = jiffies; | |
f75edf2d | 1545 | cfqq->allocated_slice = 0; |
22e2c507 | 1546 | cfqq->slice_end = 0; |
2f5cb738 | 1547 | cfqq->slice_dispatch = 0; |
22084190 | 1548 | cfqq->nr_sectors = 0; |
2f5cb738 | 1549 | |
2f5cb738 | 1550 | cfq_clear_cfqq_wait_request(cfqq); |
b029195d | 1551 | cfq_clear_cfqq_must_dispatch(cfqq); |
3b18152c JA |
1552 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
1553 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 1554 | cfq_mark_cfqq_slice_new(cfqq); |
2f5cb738 JA |
1555 | |
1556 | del_timer(&cfqd->idle_slice_timer); | |
22e2c507 JA |
1557 | } |
1558 | ||
1559 | cfqd->active_queue = cfqq; | |
1560 | } | |
1561 | ||
7b14e3b5 JA |
1562 | /* |
1563 | * current cfqq expired its slice (or was too idle), select new one | |
1564 | */ | |
1565 | static void | |
1566 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
a6151c3a | 1567 | bool timed_out) |
7b14e3b5 | 1568 | { |
7b679138 JA |
1569 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
1570 | ||
7b14e3b5 JA |
1571 | if (cfq_cfqq_wait_request(cfqq)) |
1572 | del_timer(&cfqd->idle_slice_timer); | |
1573 | ||
7b14e3b5 | 1574 | cfq_clear_cfqq_wait_request(cfqq); |
f75edf2d VG |
1575 | cfq_clear_cfqq_wait_busy(cfqq); |
1576 | cfq_clear_cfqq_wait_busy_done(cfqq); | |
7b14e3b5 JA |
1577 | |
1578 | /* | |
6084cdda | 1579 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 1580 | */ |
7b679138 | 1581 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) { |
c5b680f3 | 1582 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b679138 JA |
1583 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
1584 | } | |
7b14e3b5 | 1585 | |
dae739eb VG |
1586 | cfq_group_served(cfqd, cfqq->cfqg, cfqq); |
1587 | ||
f04a6424 VG |
1588 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
1589 | cfq_del_cfqq_rr(cfqd, cfqq); | |
1590 | ||
edd75ffd | 1591 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
1592 | |
1593 | if (cfqq == cfqd->active_queue) | |
1594 | cfqd->active_queue = NULL; | |
1595 | ||
dae739eb VG |
1596 | if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active) |
1597 | cfqd->grp_service_tree.active = NULL; | |
1598 | ||
7b14e3b5 JA |
1599 | if (cfqd->active_cic) { |
1600 | put_io_context(cfqd->active_cic->ioc); | |
1601 | cfqd->active_cic = NULL; | |
1602 | } | |
7b14e3b5 JA |
1603 | } |
1604 | ||
a6151c3a | 1605 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) |
7b14e3b5 JA |
1606 | { |
1607 | struct cfq_queue *cfqq = cfqd->active_queue; | |
1608 | ||
1609 | if (cfqq) | |
6084cdda | 1610 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
1611 | } |
1612 | ||
498d3aa2 JA |
1613 | /* |
1614 | * Get next queue for service. Unless we have a queue preemption, | |
1615 | * we'll simply select the first cfqq in the service tree. | |
1616 | */ | |
6d048f53 | 1617 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 1618 | { |
c0324a02 | 1619 | struct cfq_rb_root *service_tree = |
cdb16e8f VG |
1620 | service_tree_for(cfqd->serving_group, cfqd->serving_prio, |
1621 | cfqd->serving_type, cfqd); | |
d9e7620e | 1622 | |
f04a6424 VG |
1623 | if (!cfqd->rq_queued) |
1624 | return NULL; | |
1625 | ||
1fa8f6d6 VG |
1626 | /* There is nothing to dispatch */ |
1627 | if (!service_tree) | |
1628 | return NULL; | |
c0324a02 CZ |
1629 | if (RB_EMPTY_ROOT(&service_tree->rb)) |
1630 | return NULL; | |
1631 | return cfq_rb_first(service_tree); | |
6d048f53 JA |
1632 | } |
1633 | ||
f04a6424 VG |
1634 | static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) |
1635 | { | |
25fb5169 | 1636 | struct cfq_group *cfqg; |
f04a6424 VG |
1637 | struct cfq_queue *cfqq; |
1638 | int i, j; | |
1639 | struct cfq_rb_root *st; | |
1640 | ||
1641 | if (!cfqd->rq_queued) | |
1642 | return NULL; | |
1643 | ||
25fb5169 VG |
1644 | cfqg = cfq_get_next_cfqg(cfqd); |
1645 | if (!cfqg) | |
1646 | return NULL; | |
1647 | ||
f04a6424 VG |
1648 | for_each_cfqg_st(cfqg, i, j, st) |
1649 | if ((cfqq = cfq_rb_first(st)) != NULL) | |
1650 | return cfqq; | |
1651 | return NULL; | |
1652 | } | |
1653 | ||
498d3aa2 JA |
1654 | /* |
1655 | * Get and set a new active queue for service. | |
1656 | */ | |
a36e71f9 JA |
1657 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
1658 | struct cfq_queue *cfqq) | |
6d048f53 | 1659 | { |
e00ef799 | 1660 | if (!cfqq) |
a36e71f9 | 1661 | cfqq = cfq_get_next_queue(cfqd); |
6d048f53 | 1662 | |
22e2c507 | 1663 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 1664 | return cfqq; |
22e2c507 JA |
1665 | } |
1666 | ||
d9e7620e JA |
1667 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
1668 | struct request *rq) | |
1669 | { | |
83096ebf TH |
1670 | if (blk_rq_pos(rq) >= cfqd->last_position) |
1671 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 1672 | else |
83096ebf | 1673 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
1674 | } |
1675 | ||
b2c18e1e JM |
1676 | #define CFQQ_SEEK_THR 8 * 1024 |
1677 | #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR) | |
04dc6e71 | 1678 | |
b2c18e1e JM |
1679 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
1680 | struct request *rq) | |
6d048f53 | 1681 | { |
b2c18e1e | 1682 | sector_t sdist = cfqq->seek_mean; |
6d048f53 | 1683 | |
b2c18e1e JM |
1684 | if (!sample_valid(cfqq->seek_samples)) |
1685 | sdist = CFQQ_SEEK_THR; | |
6d048f53 | 1686 | |
04dc6e71 | 1687 | return cfq_dist_from_last(cfqd, rq) <= sdist; |
6d048f53 JA |
1688 | } |
1689 | ||
a36e71f9 JA |
1690 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
1691 | struct cfq_queue *cur_cfqq) | |
1692 | { | |
f2d1f0ae | 1693 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
1694 | struct rb_node *parent, *node; |
1695 | struct cfq_queue *__cfqq; | |
1696 | sector_t sector = cfqd->last_position; | |
1697 | ||
1698 | if (RB_EMPTY_ROOT(root)) | |
1699 | return NULL; | |
1700 | ||
1701 | /* | |
1702 | * First, if we find a request starting at the end of the last | |
1703 | * request, choose it. | |
1704 | */ | |
f2d1f0ae | 1705 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
1706 | if (__cfqq) |
1707 | return __cfqq; | |
1708 | ||
1709 | /* | |
1710 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1711 | * will contain the closest sector. | |
1712 | */ | |
1713 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
b2c18e1e | 1714 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1715 | return __cfqq; |
1716 | ||
2e46e8b2 | 1717 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1718 | node = rb_next(&__cfqq->p_node); |
1719 | else | |
1720 | node = rb_prev(&__cfqq->p_node); | |
1721 | if (!node) | |
1722 | return NULL; | |
1723 | ||
1724 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
b2c18e1e | 1725 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1726 | return __cfqq; |
1727 | ||
1728 | return NULL; | |
1729 | } | |
1730 | ||
1731 | /* | |
1732 | * cfqd - obvious | |
1733 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1734 | * closely cooperating with itself. | |
1735 | * | |
1736 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1737 | * one request, and that cfqd->last_position reflects a position on the disk | |
1738 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1739 | * assumption. | |
1740 | */ | |
1741 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
b3b6d040 | 1742 | struct cfq_queue *cur_cfqq) |
6d048f53 | 1743 | { |
a36e71f9 JA |
1744 | struct cfq_queue *cfqq; |
1745 | ||
e6c5bc73 JM |
1746 | if (!cfq_cfqq_sync(cur_cfqq)) |
1747 | return NULL; | |
1748 | if (CFQQ_SEEKY(cur_cfqq)) | |
1749 | return NULL; | |
1750 | ||
6d048f53 | 1751 | /* |
d9e7620e JA |
1752 | * We should notice if some of the queues are cooperating, eg |
1753 | * working closely on the same area of the disk. In that case, | |
1754 | * we can group them together and don't waste time idling. | |
6d048f53 | 1755 | */ |
a36e71f9 JA |
1756 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1757 | if (!cfqq) | |
1758 | return NULL; | |
1759 | ||
8682e1f1 VG |
1760 | /* If new queue belongs to different cfq_group, don't choose it */ |
1761 | if (cur_cfqq->cfqg != cfqq->cfqg) | |
1762 | return NULL; | |
1763 | ||
df5fe3e8 JM |
1764 | /* |
1765 | * It only makes sense to merge sync queues. | |
1766 | */ | |
1767 | if (!cfq_cfqq_sync(cfqq)) | |
1768 | return NULL; | |
e6c5bc73 JM |
1769 | if (CFQQ_SEEKY(cfqq)) |
1770 | return NULL; | |
df5fe3e8 | 1771 | |
c0324a02 CZ |
1772 | /* |
1773 | * Do not merge queues of different priority classes | |
1774 | */ | |
1775 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
1776 | return NULL; | |
1777 | ||
a36e71f9 | 1778 | return cfqq; |
6d048f53 JA |
1779 | } |
1780 | ||
a6d44e98 CZ |
1781 | /* |
1782 | * Determine whether we should enforce idle window for this queue. | |
1783 | */ | |
1784 | ||
1785 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1786 | { | |
1787 | enum wl_prio_t prio = cfqq_prio(cfqq); | |
718eee05 | 1788 | struct cfq_rb_root *service_tree = cfqq->service_tree; |
a6d44e98 | 1789 | |
f04a6424 VG |
1790 | BUG_ON(!service_tree); |
1791 | BUG_ON(!service_tree->count); | |
1792 | ||
a6d44e98 CZ |
1793 | /* We never do for idle class queues. */ |
1794 | if (prio == IDLE_WORKLOAD) | |
1795 | return false; | |
1796 | ||
1797 | /* We do for queues that were marked with idle window flag. */ | |
1798 | if (cfq_cfqq_idle_window(cfqq)) | |
1799 | return true; | |
1800 | ||
1801 | /* | |
1802 | * Otherwise, we do only if they are the last ones | |
1803 | * in their service tree. | |
1804 | */ | |
f04a6424 | 1805 | return service_tree->count == 1; |
a6d44e98 CZ |
1806 | } |
1807 | ||
6d048f53 | 1808 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1809 | { |
1792669c | 1810 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1811 | struct cfq_io_context *cic; |
7b14e3b5 JA |
1812 | unsigned long sl; |
1813 | ||
a68bbddb | 1814 | /* |
f7d7b7a7 JA |
1815 | * SSD device without seek penalty, disable idling. But only do so |
1816 | * for devices that support queuing, otherwise we still have a problem | |
1817 | * with sync vs async workloads. | |
a68bbddb | 1818 | */ |
f7d7b7a7 | 1819 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
1820 | return; |
1821 | ||
dd67d051 | 1822 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 1823 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
1824 | |
1825 | /* | |
1826 | * idle is disabled, either manually or by past process history | |
1827 | */ | |
a6d44e98 | 1828 | if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq)) |
6d048f53 JA |
1829 | return; |
1830 | ||
7b679138 | 1831 | /* |
8e550632 | 1832 | * still active requests from this queue, don't idle |
7b679138 | 1833 | */ |
8e550632 | 1834 | if (cfqq->dispatched) |
7b679138 JA |
1835 | return; |
1836 | ||
22e2c507 JA |
1837 | /* |
1838 | * task has exited, don't wait | |
1839 | */ | |
206dc69b | 1840 | cic = cfqd->active_cic; |
66dac98e | 1841 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
1842 | return; |
1843 | ||
355b659c CZ |
1844 | /* |
1845 | * If our average think time is larger than the remaining time | |
1846 | * slice, then don't idle. This avoids overrunning the allotted | |
1847 | * time slice. | |
1848 | */ | |
1849 | if (sample_valid(cic->ttime_samples) && | |
1850 | (cfqq->slice_end - jiffies < cic->ttime_mean)) | |
1851 | return; | |
1852 | ||
3b18152c | 1853 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 1854 | |
6d048f53 | 1855 | sl = cfqd->cfq_slice_idle; |
206dc69b | 1856 | |
7b14e3b5 | 1857 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
9481ffdc | 1858 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); |
1da177e4 LT |
1859 | } |
1860 | ||
498d3aa2 JA |
1861 | /* |
1862 | * Move request from internal lists to the request queue dispatch list. | |
1863 | */ | |
165125e1 | 1864 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 1865 | { |
3ed9a296 | 1866 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1867 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 1868 | |
7b679138 JA |
1869 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
1870 | ||
06d21886 | 1871 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); |
5380a101 | 1872 | cfq_remove_request(rq); |
6d048f53 | 1873 | cfqq->dispatched++; |
5380a101 | 1874 | elv_dispatch_sort(q, rq); |
3ed9a296 JA |
1875 | |
1876 | if (cfq_cfqq_sync(cfqq)) | |
1877 | cfqd->sync_flight++; | |
22084190 | 1878 | cfqq->nr_sectors += blk_rq_sectors(rq); |
1da177e4 LT |
1879 | } |
1880 | ||
1881 | /* | |
1882 | * return expired entry, or NULL to just start from scratch in rbtree | |
1883 | */ | |
febffd61 | 1884 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 | 1885 | { |
30996f40 | 1886 | struct request *rq = NULL; |
1da177e4 | 1887 | |
3b18152c | 1888 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 1889 | return NULL; |
cb887411 JA |
1890 | |
1891 | cfq_mark_cfqq_fifo_expire(cfqq); | |
1892 | ||
89850f7e JA |
1893 | if (list_empty(&cfqq->fifo)) |
1894 | return NULL; | |
1da177e4 | 1895 | |
89850f7e | 1896 | rq = rq_entry_fifo(cfqq->fifo.next); |
30996f40 | 1897 | if (time_before(jiffies, rq_fifo_time(rq))) |
7b679138 | 1898 | rq = NULL; |
1da177e4 | 1899 | |
30996f40 | 1900 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 1901 | return rq; |
1da177e4 LT |
1902 | } |
1903 | ||
22e2c507 JA |
1904 | static inline int |
1905 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1906 | { | |
1907 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 1908 | |
22e2c507 | 1909 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 1910 | |
22e2c507 | 1911 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
1912 | } |
1913 | ||
df5fe3e8 JM |
1914 | /* |
1915 | * Must be called with the queue_lock held. | |
1916 | */ | |
1917 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
1918 | { | |
1919 | int process_refs, io_refs; | |
1920 | ||
1921 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
1922 | process_refs = atomic_read(&cfqq->ref) - io_refs; | |
1923 | BUG_ON(process_refs < 0); | |
1924 | return process_refs; | |
1925 | } | |
1926 | ||
1927 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
1928 | { | |
e6c5bc73 | 1929 | int process_refs, new_process_refs; |
df5fe3e8 JM |
1930 | struct cfq_queue *__cfqq; |
1931 | ||
1932 | /* Avoid a circular list and skip interim queue merges */ | |
1933 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
1934 | if (__cfqq == cfqq) | |
1935 | return; | |
1936 | new_cfqq = __cfqq; | |
1937 | } | |
1938 | ||
1939 | process_refs = cfqq_process_refs(cfqq); | |
1940 | /* | |
1941 | * If the process for the cfqq has gone away, there is no | |
1942 | * sense in merging the queues. | |
1943 | */ | |
1944 | if (process_refs == 0) | |
1945 | return; | |
1946 | ||
e6c5bc73 JM |
1947 | /* |
1948 | * Merge in the direction of the lesser amount of work. | |
1949 | */ | |
1950 | new_process_refs = cfqq_process_refs(new_cfqq); | |
1951 | if (new_process_refs >= process_refs) { | |
1952 | cfqq->new_cfqq = new_cfqq; | |
1953 | atomic_add(process_refs, &new_cfqq->ref); | |
1954 | } else { | |
1955 | new_cfqq->new_cfqq = cfqq; | |
1956 | atomic_add(new_process_refs, &cfqq->ref); | |
1957 | } | |
df5fe3e8 JM |
1958 | } |
1959 | ||
cdb16e8f VG |
1960 | static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, |
1961 | struct cfq_group *cfqg, enum wl_prio_t prio, | |
1962 | bool prio_changed) | |
718eee05 CZ |
1963 | { |
1964 | struct cfq_queue *queue; | |
1965 | int i; | |
1966 | bool key_valid = false; | |
1967 | unsigned long lowest_key = 0; | |
1968 | enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; | |
1969 | ||
1970 | if (prio_changed) { | |
1971 | /* | |
1972 | * When priorities switched, we prefer starting | |
1973 | * from SYNC_NOIDLE (first choice), or just SYNC | |
1974 | * over ASYNC | |
1975 | */ | |
cdb16e8f | 1976 | if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) |
718eee05 CZ |
1977 | return cur_best; |
1978 | cur_best = SYNC_WORKLOAD; | |
cdb16e8f | 1979 | if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) |
718eee05 CZ |
1980 | return cur_best; |
1981 | ||
1982 | return ASYNC_WORKLOAD; | |
1983 | } | |
1984 | ||
1985 | for (i = 0; i < 3; ++i) { | |
1986 | /* otherwise, select the one with lowest rb_key */ | |
cdb16e8f | 1987 | queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd)); |
718eee05 CZ |
1988 | if (queue && |
1989 | (!key_valid || time_before(queue->rb_key, lowest_key))) { | |
1990 | lowest_key = queue->rb_key; | |
1991 | cur_best = i; | |
1992 | key_valid = true; | |
1993 | } | |
1994 | } | |
1995 | ||
1996 | return cur_best; | |
1997 | } | |
1998 | ||
cdb16e8f | 1999 | static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) |
718eee05 CZ |
2000 | { |
2001 | enum wl_prio_t previous_prio = cfqd->serving_prio; | |
2002 | bool prio_changed; | |
2003 | unsigned slice; | |
2004 | unsigned count; | |
cdb16e8f | 2005 | struct cfq_rb_root *st; |
58ff82f3 | 2006 | unsigned group_slice; |
718eee05 | 2007 | |
1fa8f6d6 VG |
2008 | if (!cfqg) { |
2009 | cfqd->serving_prio = IDLE_WORKLOAD; | |
2010 | cfqd->workload_expires = jiffies + 1; | |
2011 | return; | |
2012 | } | |
2013 | ||
718eee05 | 2014 | /* Choose next priority. RT > BE > IDLE */ |
58ff82f3 | 2015 | if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) |
718eee05 | 2016 | cfqd->serving_prio = RT_WORKLOAD; |
58ff82f3 | 2017 | else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) |
718eee05 CZ |
2018 | cfqd->serving_prio = BE_WORKLOAD; |
2019 | else { | |
2020 | cfqd->serving_prio = IDLE_WORKLOAD; | |
2021 | cfqd->workload_expires = jiffies + 1; | |
2022 | return; | |
2023 | } | |
2024 | ||
2025 | /* | |
2026 | * For RT and BE, we have to choose also the type | |
2027 | * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload | |
2028 | * expiration time | |
2029 | */ | |
2030 | prio_changed = (cfqd->serving_prio != previous_prio); | |
cdb16e8f VG |
2031 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, |
2032 | cfqd); | |
2033 | count = st->count; | |
718eee05 CZ |
2034 | |
2035 | /* | |
2036 | * If priority didn't change, check workload expiration, | |
2037 | * and that we still have other queues ready | |
2038 | */ | |
2039 | if (!prio_changed && count && | |
2040 | !time_after(jiffies, cfqd->workload_expires)) | |
2041 | return; | |
2042 | ||
2043 | /* otherwise select new workload type */ | |
2044 | cfqd->serving_type = | |
cdb16e8f VG |
2045 | cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed); |
2046 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, | |
2047 | cfqd); | |
2048 | count = st->count; | |
718eee05 CZ |
2049 | |
2050 | /* | |
2051 | * the workload slice is computed as a fraction of target latency | |
2052 | * proportional to the number of queues in that workload, over | |
2053 | * all the queues in the same priority class | |
2054 | */ | |
58ff82f3 VG |
2055 | group_slice = cfq_group_slice(cfqd, cfqg); |
2056 | ||
2057 | slice = group_slice * count / | |
2058 | max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], | |
2059 | cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); | |
718eee05 | 2060 | |
f26bd1f0 VG |
2061 | if (cfqd->serving_type == ASYNC_WORKLOAD) { |
2062 | unsigned int tmp; | |
2063 | ||
2064 | /* | |
2065 | * Async queues are currently system wide. Just taking | |
2066 | * proportion of queues with-in same group will lead to higher | |
2067 | * async ratio system wide as generally root group is going | |
2068 | * to have higher weight. A more accurate thing would be to | |
2069 | * calculate system wide asnc/sync ratio. | |
2070 | */ | |
2071 | tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg); | |
2072 | tmp = tmp/cfqd->busy_queues; | |
2073 | slice = min_t(unsigned, slice, tmp); | |
2074 | ||
718eee05 CZ |
2075 | /* async workload slice is scaled down according to |
2076 | * the sync/async slice ratio. */ | |
2077 | slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; | |
f26bd1f0 | 2078 | } else |
718eee05 CZ |
2079 | /* sync workload slice is at least 2 * cfq_slice_idle */ |
2080 | slice = max(slice, 2 * cfqd->cfq_slice_idle); | |
2081 | ||
2082 | slice = max_t(unsigned, slice, CFQ_MIN_TT); | |
2083 | cfqd->workload_expires = jiffies + slice; | |
8e550632 | 2084 | cfqd->noidle_tree_requires_idle = false; |
718eee05 CZ |
2085 | } |
2086 | ||
1fa8f6d6 VG |
2087 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) |
2088 | { | |
2089 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
25bc6b07 | 2090 | struct cfq_group *cfqg; |
1fa8f6d6 VG |
2091 | |
2092 | if (RB_EMPTY_ROOT(&st->rb)) | |
2093 | return NULL; | |
25bc6b07 VG |
2094 | cfqg = cfq_rb_first_group(st); |
2095 | st->active = &cfqg->rb_node; | |
2096 | update_min_vdisktime(st); | |
2097 | return cfqg; | |
1fa8f6d6 VG |
2098 | } |
2099 | ||
cdb16e8f VG |
2100 | static void cfq_choose_cfqg(struct cfq_data *cfqd) |
2101 | { | |
1fa8f6d6 VG |
2102 | struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); |
2103 | ||
2104 | cfqd->serving_group = cfqg; | |
dae739eb VG |
2105 | |
2106 | /* Restore the workload type data */ | |
2107 | if (cfqg->saved_workload_slice) { | |
2108 | cfqd->workload_expires = jiffies + cfqg->saved_workload_slice; | |
2109 | cfqd->serving_type = cfqg->saved_workload; | |
2110 | cfqd->serving_prio = cfqg->saved_serving_prio; | |
2111 | } | |
1fa8f6d6 | 2112 | choose_service_tree(cfqd, cfqg); |
cdb16e8f VG |
2113 | } |
2114 | ||
22e2c507 | 2115 | /* |
498d3aa2 JA |
2116 | * Select a queue for service. If we have a current active queue, |
2117 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 2118 | */ |
1b5ed5e1 | 2119 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 2120 | { |
a36e71f9 | 2121 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 2122 | |
22e2c507 JA |
2123 | cfqq = cfqd->active_queue; |
2124 | if (!cfqq) | |
2125 | goto new_queue; | |
1da177e4 | 2126 | |
f04a6424 VG |
2127 | if (!cfqd->rq_queued) |
2128 | return NULL; | |
22e2c507 | 2129 | /* |
6d048f53 | 2130 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 2131 | */ |
f75edf2d VG |
2132 | if ((cfq_slice_used(cfqq) || cfq_cfqq_wait_busy_done(cfqq)) |
2133 | && !cfq_cfqq_must_dispatch(cfqq)) | |
3b18152c | 2134 | goto expire; |
1da177e4 | 2135 | |
22e2c507 | 2136 | /* |
6d048f53 JA |
2137 | * The active queue has requests and isn't expired, allow it to |
2138 | * dispatch. | |
22e2c507 | 2139 | */ |
dd67d051 | 2140 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 2141 | goto keep_queue; |
6d048f53 | 2142 | |
a36e71f9 JA |
2143 | /* |
2144 | * If another queue has a request waiting within our mean seek | |
2145 | * distance, let it run. The expire code will check for close | |
2146 | * cooperators and put the close queue at the front of the service | |
df5fe3e8 | 2147 | * tree. If possible, merge the expiring queue with the new cfqq. |
a36e71f9 | 2148 | */ |
b3b6d040 | 2149 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); |
df5fe3e8 JM |
2150 | if (new_cfqq) { |
2151 | if (!cfqq->new_cfqq) | |
2152 | cfq_setup_merge(cfqq, new_cfqq); | |
a36e71f9 | 2153 | goto expire; |
df5fe3e8 | 2154 | } |
a36e71f9 | 2155 | |
6d048f53 JA |
2156 | /* |
2157 | * No requests pending. If the active queue still has requests in | |
2158 | * flight or is idling for a new request, allow either of these | |
2159 | * conditions to happen (or time out) before selecting a new queue. | |
2160 | */ | |
cc197479 | 2161 | if (timer_pending(&cfqd->idle_slice_timer) || |
a6d44e98 | 2162 | (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) { |
caaa5f9f JA |
2163 | cfqq = NULL; |
2164 | goto keep_queue; | |
22e2c507 JA |
2165 | } |
2166 | ||
3b18152c | 2167 | expire: |
6084cdda | 2168 | cfq_slice_expired(cfqd, 0); |
3b18152c | 2169 | new_queue: |
718eee05 CZ |
2170 | /* |
2171 | * Current queue expired. Check if we have to switch to a new | |
2172 | * service tree | |
2173 | */ | |
2174 | if (!new_cfqq) | |
cdb16e8f | 2175 | cfq_choose_cfqg(cfqd); |
718eee05 | 2176 | |
a36e71f9 | 2177 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 2178 | keep_queue: |
3b18152c | 2179 | return cfqq; |
22e2c507 JA |
2180 | } |
2181 | ||
febffd61 | 2182 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
2183 | { |
2184 | int dispatched = 0; | |
2185 | ||
2186 | while (cfqq->next_rq) { | |
2187 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
2188 | dispatched++; | |
2189 | } | |
2190 | ||
2191 | BUG_ON(!list_empty(&cfqq->fifo)); | |
f04a6424 VG |
2192 | |
2193 | /* By default cfqq is not expired if it is empty. Do it explicitly */ | |
2194 | __cfq_slice_expired(cfqq->cfqd, cfqq, 0); | |
d9e7620e JA |
2195 | return dispatched; |
2196 | } | |
2197 | ||
498d3aa2 JA |
2198 | /* |
2199 | * Drain our current requests. Used for barriers and when switching | |
2200 | * io schedulers on-the-fly. | |
2201 | */ | |
d9e7620e | 2202 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 2203 | { |
0871714e | 2204 | struct cfq_queue *cfqq; |
d9e7620e | 2205 | int dispatched = 0; |
cdb16e8f | 2206 | |
f04a6424 VG |
2207 | while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) |
2208 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); | |
1b5ed5e1 | 2209 | |
6084cdda | 2210 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
2211 | BUG_ON(cfqd->busy_queues); |
2212 | ||
6923715a | 2213 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
2214 | return dispatched; |
2215 | } | |
2216 | ||
0b182d61 | 2217 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
2f5cb738 | 2218 | { |
2f5cb738 | 2219 | unsigned int max_dispatch; |
22e2c507 | 2220 | |
5ad531db JA |
2221 | /* |
2222 | * Drain async requests before we start sync IO | |
2223 | */ | |
a6d44e98 | 2224 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC]) |
0b182d61 | 2225 | return false; |
5ad531db | 2226 | |
2f5cb738 JA |
2227 | /* |
2228 | * If this is an async queue and we have sync IO in flight, let it wait | |
2229 | */ | |
2230 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) | |
0b182d61 | 2231 | return false; |
2f5cb738 JA |
2232 | |
2233 | max_dispatch = cfqd->cfq_quantum; | |
2234 | if (cfq_class_idle(cfqq)) | |
2235 | max_dispatch = 1; | |
b4878f24 | 2236 | |
2f5cb738 JA |
2237 | /* |
2238 | * Does this cfqq already have too much IO in flight? | |
2239 | */ | |
2240 | if (cfqq->dispatched >= max_dispatch) { | |
2241 | /* | |
2242 | * idle queue must always only have a single IO in flight | |
2243 | */ | |
3ed9a296 | 2244 | if (cfq_class_idle(cfqq)) |
0b182d61 | 2245 | return false; |
3ed9a296 | 2246 | |
2f5cb738 JA |
2247 | /* |
2248 | * We have other queues, don't allow more IO from this one | |
2249 | */ | |
2250 | if (cfqd->busy_queues > 1) | |
0b182d61 | 2251 | return false; |
9ede209e | 2252 | |
365722bb | 2253 | /* |
474b18cc | 2254 | * Sole queue user, no limit |
365722bb | 2255 | */ |
474b18cc | 2256 | max_dispatch = -1; |
8e296755 JA |
2257 | } |
2258 | ||
2259 | /* | |
2260 | * Async queues must wait a bit before being allowed dispatch. | |
2261 | * We also ramp up the dispatch depth gradually for async IO, | |
2262 | * based on the last sync IO we serviced | |
2263 | */ | |
963b72fc | 2264 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { |
8e296755 JA |
2265 | unsigned long last_sync = jiffies - cfqd->last_end_sync_rq; |
2266 | unsigned int depth; | |
365722bb | 2267 | |
61f0c1dc | 2268 | depth = last_sync / cfqd->cfq_slice[1]; |
e00c54c3 JA |
2269 | if (!depth && !cfqq->dispatched) |
2270 | depth = 1; | |
8e296755 JA |
2271 | if (depth < max_dispatch) |
2272 | max_dispatch = depth; | |
2f5cb738 | 2273 | } |
3ed9a296 | 2274 | |
0b182d61 JA |
2275 | /* |
2276 | * If we're below the current max, allow a dispatch | |
2277 | */ | |
2278 | return cfqq->dispatched < max_dispatch; | |
2279 | } | |
2280 | ||
2281 | /* | |
2282 | * Dispatch a request from cfqq, moving them to the request queue | |
2283 | * dispatch list. | |
2284 | */ | |
2285 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2286 | { | |
2287 | struct request *rq; | |
2288 | ||
2289 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
2290 | ||
2291 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
2292 | return false; | |
2293 | ||
2294 | /* | |
2295 | * follow expired path, else get first next available | |
2296 | */ | |
2297 | rq = cfq_check_fifo(cfqq); | |
2298 | if (!rq) | |
2299 | rq = cfqq->next_rq; | |
2300 | ||
2301 | /* | |
2302 | * insert request into driver dispatch list | |
2303 | */ | |
2304 | cfq_dispatch_insert(cfqd->queue, rq); | |
2305 | ||
2306 | if (!cfqd->active_cic) { | |
2307 | struct cfq_io_context *cic = RQ_CIC(rq); | |
2308 | ||
2309 | atomic_long_inc(&cic->ioc->refcount); | |
2310 | cfqd->active_cic = cic; | |
2311 | } | |
2312 | ||
2313 | return true; | |
2314 | } | |
2315 | ||
2316 | /* | |
2317 | * Find the cfqq that we need to service and move a request from that to the | |
2318 | * dispatch list | |
2319 | */ | |
2320 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
2321 | { | |
2322 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2323 | struct cfq_queue *cfqq; | |
2324 | ||
2325 | if (!cfqd->busy_queues) | |
2326 | return 0; | |
2327 | ||
2328 | if (unlikely(force)) | |
2329 | return cfq_forced_dispatch(cfqd); | |
2330 | ||
2331 | cfqq = cfq_select_queue(cfqd); | |
2332 | if (!cfqq) | |
8e296755 JA |
2333 | return 0; |
2334 | ||
2f5cb738 | 2335 | /* |
0b182d61 | 2336 | * Dispatch a request from this cfqq, if it is allowed |
2f5cb738 | 2337 | */ |
0b182d61 JA |
2338 | if (!cfq_dispatch_request(cfqd, cfqq)) |
2339 | return 0; | |
2340 | ||
2f5cb738 | 2341 | cfqq->slice_dispatch++; |
b029195d | 2342 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 2343 | |
2f5cb738 JA |
2344 | /* |
2345 | * expire an async queue immediately if it has used up its slice. idle | |
2346 | * queue always expire after 1 dispatch round. | |
2347 | */ | |
2348 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
2349 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
2350 | cfq_class_idle(cfqq))) { | |
2351 | cfqq->slice_end = jiffies + 1; | |
2352 | cfq_slice_expired(cfqd, 0); | |
1da177e4 LT |
2353 | } |
2354 | ||
b217a903 | 2355 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); |
2f5cb738 | 2356 | return 1; |
1da177e4 LT |
2357 | } |
2358 | ||
1da177e4 | 2359 | /* |
5e705374 JA |
2360 | * task holds one reference to the queue, dropped when task exits. each rq |
2361 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 | 2362 | * |
b1c35769 | 2363 | * Each cfq queue took a reference on the parent group. Drop it now. |
1da177e4 LT |
2364 | * queue lock must be held here. |
2365 | */ | |
2366 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
2367 | { | |
22e2c507 | 2368 | struct cfq_data *cfqd = cfqq->cfqd; |
b1c35769 | 2369 | struct cfq_group *cfqg; |
22e2c507 JA |
2370 | |
2371 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
2372 | |
2373 | if (!atomic_dec_and_test(&cfqq->ref)) | |
2374 | return; | |
2375 | ||
7b679138 | 2376 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 2377 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 2378 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
b1c35769 | 2379 | cfqg = cfqq->cfqg; |
1da177e4 | 2380 | |
28f95cbc | 2381 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 2382 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 2383 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 2384 | } |
22e2c507 | 2385 | |
f04a6424 | 2386 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 | 2387 | kmem_cache_free(cfq_pool, cfqq); |
b1c35769 | 2388 | cfq_put_cfqg(cfqg); |
ae30c286 VG |
2389 | if (cfqq->orig_cfqg) |
2390 | cfq_put_cfqg(cfqq->orig_cfqg); | |
1da177e4 LT |
2391 | } |
2392 | ||
d6de8be7 JA |
2393 | /* |
2394 | * Must always be called with the rcu_read_lock() held | |
2395 | */ | |
07416d29 JA |
2396 | static void |
2397 | __call_for_each_cic(struct io_context *ioc, | |
2398 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
2399 | { | |
2400 | struct cfq_io_context *cic; | |
2401 | struct hlist_node *n; | |
2402 | ||
2403 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) | |
2404 | func(ioc, cic); | |
2405 | } | |
2406 | ||
4ac845a2 | 2407 | /* |
34e6bbf2 | 2408 | * Call func for each cic attached to this ioc. |
4ac845a2 | 2409 | */ |
34e6bbf2 | 2410 | static void |
4ac845a2 JA |
2411 | call_for_each_cic(struct io_context *ioc, |
2412 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1da177e4 | 2413 | { |
4ac845a2 | 2414 | rcu_read_lock(); |
07416d29 | 2415 | __call_for_each_cic(ioc, func); |
4ac845a2 | 2416 | rcu_read_unlock(); |
34e6bbf2 FC |
2417 | } |
2418 | ||
2419 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
2420 | { | |
2421 | struct cfq_io_context *cic; | |
2422 | ||
2423 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
2424 | ||
2425 | kmem_cache_free(cfq_ioc_pool, cic); | |
245b2e70 | 2426 | elv_ioc_count_dec(cfq_ioc_count); |
34e6bbf2 | 2427 | |
9a11b4ed JA |
2428 | if (ioc_gone) { |
2429 | /* | |
2430 | * CFQ scheduler is exiting, grab exit lock and check | |
2431 | * the pending io context count. If it hits zero, | |
2432 | * complete ioc_gone and set it back to NULL | |
2433 | */ | |
2434 | spin_lock(&ioc_gone_lock); | |
245b2e70 | 2435 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { |
9a11b4ed JA |
2436 | complete(ioc_gone); |
2437 | ioc_gone = NULL; | |
2438 | } | |
2439 | spin_unlock(&ioc_gone_lock); | |
2440 | } | |
34e6bbf2 | 2441 | } |
4ac845a2 | 2442 | |
34e6bbf2 FC |
2443 | static void cfq_cic_free(struct cfq_io_context *cic) |
2444 | { | |
2445 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
2446 | } |
2447 | ||
2448 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
2449 | { | |
2450 | unsigned long flags; | |
2451 | ||
2452 | BUG_ON(!cic->dead_key); | |
2453 | ||
2454 | spin_lock_irqsave(&ioc->lock, flags); | |
2455 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
ffc4e759 | 2456 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2457 | spin_unlock_irqrestore(&ioc->lock, flags); |
2458 | ||
34e6bbf2 | 2459 | cfq_cic_free(cic); |
4ac845a2 JA |
2460 | } |
2461 | ||
d6de8be7 JA |
2462 | /* |
2463 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
2464 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
2465 | * and ->trim() which is called with the task lock held | |
2466 | */ | |
4ac845a2 JA |
2467 | static void cfq_free_io_context(struct io_context *ioc) |
2468 | { | |
4ac845a2 | 2469 | /* |
34e6bbf2 FC |
2470 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
2471 | * so no more cic's are allowed to be linked into this ioc. So it | |
2472 | * should be ok to iterate over the known list, we will see all cic's | |
2473 | * since no new ones are added. | |
4ac845a2 | 2474 | */ |
07416d29 | 2475 | __call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
2476 | } |
2477 | ||
89850f7e | 2478 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 2479 | { |
df5fe3e8 JM |
2480 | struct cfq_queue *__cfqq, *next; |
2481 | ||
28f95cbc | 2482 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 2483 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 2484 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 2485 | } |
22e2c507 | 2486 | |
df5fe3e8 JM |
2487 | /* |
2488 | * If this queue was scheduled to merge with another queue, be | |
2489 | * sure to drop the reference taken on that queue (and others in | |
2490 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
2491 | */ | |
2492 | __cfqq = cfqq->new_cfqq; | |
2493 | while (__cfqq) { | |
2494 | if (__cfqq == cfqq) { | |
2495 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
2496 | break; | |
2497 | } | |
2498 | next = __cfqq->new_cfqq; | |
2499 | cfq_put_queue(__cfqq); | |
2500 | __cfqq = next; | |
2501 | } | |
2502 | ||
89850f7e JA |
2503 | cfq_put_queue(cfqq); |
2504 | } | |
22e2c507 | 2505 | |
89850f7e JA |
2506 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
2507 | struct cfq_io_context *cic) | |
2508 | { | |
4faa3c81 FC |
2509 | struct io_context *ioc = cic->ioc; |
2510 | ||
fc46379d | 2511 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
2512 | |
2513 | /* | |
2514 | * Make sure key == NULL is seen for dead queues | |
2515 | */ | |
fc46379d | 2516 | smp_wmb(); |
4ac845a2 | 2517 | cic->dead_key = (unsigned long) cic->key; |
fc46379d JA |
2518 | cic->key = NULL; |
2519 | ||
4faa3c81 FC |
2520 | if (ioc->ioc_data == cic) |
2521 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
2522 | ||
ff6657c6 JA |
2523 | if (cic->cfqq[BLK_RW_ASYNC]) { |
2524 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
2525 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
2526 | } |
2527 | ||
ff6657c6 JA |
2528 | if (cic->cfqq[BLK_RW_SYNC]) { |
2529 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
2530 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 2531 | } |
89850f7e JA |
2532 | } |
2533 | ||
4ac845a2 JA |
2534 | static void cfq_exit_single_io_context(struct io_context *ioc, |
2535 | struct cfq_io_context *cic) | |
89850f7e JA |
2536 | { |
2537 | struct cfq_data *cfqd = cic->key; | |
2538 | ||
89850f7e | 2539 | if (cfqd) { |
165125e1 | 2540 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 2541 | unsigned long flags; |
89850f7e | 2542 | |
4ac845a2 | 2543 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
2544 | |
2545 | /* | |
2546 | * Ensure we get a fresh copy of the ->key to prevent | |
2547 | * race between exiting task and queue | |
2548 | */ | |
2549 | smp_read_barrier_depends(); | |
2550 | if (cic->key) | |
2551 | __cfq_exit_single_io_context(cfqd, cic); | |
2552 | ||
4ac845a2 | 2553 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 2554 | } |
1da177e4 LT |
2555 | } |
2556 | ||
498d3aa2 JA |
2557 | /* |
2558 | * The process that ioc belongs to has exited, we need to clean up | |
2559 | * and put the internal structures we have that belongs to that process. | |
2560 | */ | |
e2d74ac0 | 2561 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 2562 | { |
4ac845a2 | 2563 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
2564 | } |
2565 | ||
22e2c507 | 2566 | static struct cfq_io_context * |
8267e268 | 2567 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2568 | { |
b5deef90 | 2569 | struct cfq_io_context *cic; |
1da177e4 | 2570 | |
94f6030c CL |
2571 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
2572 | cfqd->queue->node); | |
1da177e4 | 2573 | if (cic) { |
22e2c507 | 2574 | cic->last_end_request = jiffies; |
553698f9 | 2575 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 2576 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
2577 | cic->dtor = cfq_free_io_context; |
2578 | cic->exit = cfq_exit_io_context; | |
245b2e70 | 2579 | elv_ioc_count_inc(cfq_ioc_count); |
1da177e4 LT |
2580 | } |
2581 | ||
2582 | return cic; | |
2583 | } | |
2584 | ||
fd0928df | 2585 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
2586 | { |
2587 | struct task_struct *tsk = current; | |
2588 | int ioprio_class; | |
2589 | ||
3b18152c | 2590 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
2591 | return; |
2592 | ||
fd0928df | 2593 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 2594 | switch (ioprio_class) { |
fe094d98 JA |
2595 | default: |
2596 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
2597 | case IOPRIO_CLASS_NONE: | |
2598 | /* | |
6d63c275 | 2599 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
2600 | */ |
2601 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 2602 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
2603 | break; |
2604 | case IOPRIO_CLASS_RT: | |
2605 | cfqq->ioprio = task_ioprio(ioc); | |
2606 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
2607 | break; | |
2608 | case IOPRIO_CLASS_BE: | |
2609 | cfqq->ioprio = task_ioprio(ioc); | |
2610 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2611 | break; | |
2612 | case IOPRIO_CLASS_IDLE: | |
2613 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
2614 | cfqq->ioprio = 7; | |
2615 | cfq_clear_cfqq_idle_window(cfqq); | |
2616 | break; | |
22e2c507 JA |
2617 | } |
2618 | ||
2619 | /* | |
2620 | * keep track of original prio settings in case we have to temporarily | |
2621 | * elevate the priority of this queue | |
2622 | */ | |
2623 | cfqq->org_ioprio = cfqq->ioprio; | |
2624 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 2625 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
2626 | } |
2627 | ||
febffd61 | 2628 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 2629 | { |
478a82b0 AV |
2630 | struct cfq_data *cfqd = cic->key; |
2631 | struct cfq_queue *cfqq; | |
c1b707d2 | 2632 | unsigned long flags; |
35e6077c | 2633 | |
caaa5f9f JA |
2634 | if (unlikely(!cfqd)) |
2635 | return; | |
2636 | ||
c1b707d2 | 2637 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 2638 | |
ff6657c6 | 2639 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
2640 | if (cfqq) { |
2641 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
2642 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
2643 | GFP_ATOMIC); | |
caaa5f9f | 2644 | if (new_cfqq) { |
ff6657c6 | 2645 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
2646 | cfq_put_queue(cfqq); |
2647 | } | |
22e2c507 | 2648 | } |
caaa5f9f | 2649 | |
ff6657c6 | 2650 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
2651 | if (cfqq) |
2652 | cfq_mark_cfqq_prio_changed(cfqq); | |
2653 | ||
c1b707d2 | 2654 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
2655 | } |
2656 | ||
fc46379d | 2657 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 2658 | { |
4ac845a2 | 2659 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 2660 | ioc->ioprio_changed = 0; |
22e2c507 JA |
2661 | } |
2662 | ||
d5036d77 | 2663 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 2664 | pid_t pid, bool is_sync) |
d5036d77 JA |
2665 | { |
2666 | RB_CLEAR_NODE(&cfqq->rb_node); | |
2667 | RB_CLEAR_NODE(&cfqq->p_node); | |
2668 | INIT_LIST_HEAD(&cfqq->fifo); | |
2669 | ||
2670 | atomic_set(&cfqq->ref, 0); | |
2671 | cfqq->cfqd = cfqd; | |
2672 | ||
2673 | cfq_mark_cfqq_prio_changed(cfqq); | |
2674 | ||
2675 | if (is_sync) { | |
2676 | if (!cfq_class_idle(cfqq)) | |
2677 | cfq_mark_cfqq_idle_window(cfqq); | |
2678 | cfq_mark_cfqq_sync(cfqq); | |
2679 | } | |
2680 | cfqq->pid = pid; | |
2681 | } | |
2682 | ||
24610333 VG |
2683 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
2684 | static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic) | |
2685 | { | |
2686 | struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1); | |
2687 | struct cfq_data *cfqd = cic->key; | |
2688 | unsigned long flags; | |
2689 | struct request_queue *q; | |
2690 | ||
2691 | if (unlikely(!cfqd)) | |
2692 | return; | |
2693 | ||
2694 | q = cfqd->queue; | |
2695 | ||
2696 | spin_lock_irqsave(q->queue_lock, flags); | |
2697 | ||
2698 | if (sync_cfqq) { | |
2699 | /* | |
2700 | * Drop reference to sync queue. A new sync queue will be | |
2701 | * assigned in new group upon arrival of a fresh request. | |
2702 | */ | |
2703 | cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup"); | |
2704 | cic_set_cfqq(cic, NULL, 1); | |
2705 | cfq_put_queue(sync_cfqq); | |
2706 | } | |
2707 | ||
2708 | spin_unlock_irqrestore(q->queue_lock, flags); | |
2709 | } | |
2710 | ||
2711 | static void cfq_ioc_set_cgroup(struct io_context *ioc) | |
2712 | { | |
2713 | call_for_each_cic(ioc, changed_cgroup); | |
2714 | ioc->cgroup_changed = 0; | |
2715 | } | |
2716 | #endif /* CONFIG_CFQ_GROUP_IOSCHED */ | |
2717 | ||
22e2c507 | 2718 | static struct cfq_queue * |
a6151c3a | 2719 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, |
fd0928df | 2720 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 2721 | { |
22e2c507 | 2722 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 2723 | struct cfq_io_context *cic; |
cdb16e8f | 2724 | struct cfq_group *cfqg; |
22e2c507 JA |
2725 | |
2726 | retry: | |
cdb16e8f | 2727 | cfqg = cfq_get_cfqg(cfqd, 1); |
4ac845a2 | 2728 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
2729 | /* cic always exists here */ |
2730 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 2731 | |
6118b70b JA |
2732 | /* |
2733 | * Always try a new alloc if we fell back to the OOM cfqq | |
2734 | * originally, since it should just be a temporary situation. | |
2735 | */ | |
2736 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
2737 | cfqq = NULL; | |
22e2c507 JA |
2738 | if (new_cfqq) { |
2739 | cfqq = new_cfqq; | |
2740 | new_cfqq = NULL; | |
2741 | } else if (gfp_mask & __GFP_WAIT) { | |
2742 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 2743 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 2744 | gfp_mask | __GFP_ZERO, |
94f6030c | 2745 | cfqd->queue->node); |
22e2c507 | 2746 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
2747 | if (new_cfqq) |
2748 | goto retry; | |
22e2c507 | 2749 | } else { |
94f6030c CL |
2750 | cfqq = kmem_cache_alloc_node(cfq_pool, |
2751 | gfp_mask | __GFP_ZERO, | |
2752 | cfqd->queue->node); | |
22e2c507 JA |
2753 | } |
2754 | ||
6118b70b JA |
2755 | if (cfqq) { |
2756 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
2757 | cfq_init_prio_data(cfqq, ioc); | |
cdb16e8f | 2758 | cfq_link_cfqq_cfqg(cfqq, cfqg); |
6118b70b JA |
2759 | cfq_log_cfqq(cfqd, cfqq, "alloced"); |
2760 | } else | |
2761 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
2762 | } |
2763 | ||
2764 | if (new_cfqq) | |
2765 | kmem_cache_free(cfq_pool, new_cfqq); | |
2766 | ||
22e2c507 JA |
2767 | return cfqq; |
2768 | } | |
2769 | ||
c2dea2d1 VT |
2770 | static struct cfq_queue ** |
2771 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
2772 | { | |
fe094d98 | 2773 | switch (ioprio_class) { |
c2dea2d1 VT |
2774 | case IOPRIO_CLASS_RT: |
2775 | return &cfqd->async_cfqq[0][ioprio]; | |
2776 | case IOPRIO_CLASS_BE: | |
2777 | return &cfqd->async_cfqq[1][ioprio]; | |
2778 | case IOPRIO_CLASS_IDLE: | |
2779 | return &cfqd->async_idle_cfqq; | |
2780 | default: | |
2781 | BUG(); | |
2782 | } | |
2783 | } | |
2784 | ||
15c31be4 | 2785 | static struct cfq_queue * |
a6151c3a | 2786 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, |
15c31be4 JA |
2787 | gfp_t gfp_mask) |
2788 | { | |
fd0928df JA |
2789 | const int ioprio = task_ioprio(ioc); |
2790 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 2791 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
2792 | struct cfq_queue *cfqq = NULL; |
2793 | ||
c2dea2d1 VT |
2794 | if (!is_sync) { |
2795 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
2796 | cfqq = *async_cfqq; | |
2797 | } | |
2798 | ||
6118b70b | 2799 | if (!cfqq) |
fd0928df | 2800 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
2801 | |
2802 | /* | |
2803 | * pin the queue now that it's allocated, scheduler exit will prune it | |
2804 | */ | |
c2dea2d1 | 2805 | if (!is_sync && !(*async_cfqq)) { |
15c31be4 | 2806 | atomic_inc(&cfqq->ref); |
c2dea2d1 | 2807 | *async_cfqq = cfqq; |
15c31be4 JA |
2808 | } |
2809 | ||
2810 | atomic_inc(&cfqq->ref); | |
2811 | return cfqq; | |
2812 | } | |
2813 | ||
498d3aa2 JA |
2814 | /* |
2815 | * We drop cfq io contexts lazily, so we may find a dead one. | |
2816 | */ | |
dbecf3ab | 2817 | static void |
4ac845a2 JA |
2818 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
2819 | struct cfq_io_context *cic) | |
dbecf3ab | 2820 | { |
4ac845a2 JA |
2821 | unsigned long flags; |
2822 | ||
fc46379d | 2823 | WARN_ON(!list_empty(&cic->queue_list)); |
597bc485 | 2824 | |
4ac845a2 JA |
2825 | spin_lock_irqsave(&ioc->lock, flags); |
2826 | ||
4faa3c81 | 2827 | BUG_ON(ioc->ioc_data == cic); |
597bc485 | 2828 | |
4ac845a2 | 2829 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); |
ffc4e759 | 2830 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2831 | spin_unlock_irqrestore(&ioc->lock, flags); |
2832 | ||
2833 | cfq_cic_free(cic); | |
dbecf3ab OH |
2834 | } |
2835 | ||
e2d74ac0 | 2836 | static struct cfq_io_context * |
4ac845a2 | 2837 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 2838 | { |
e2d74ac0 | 2839 | struct cfq_io_context *cic; |
d6de8be7 | 2840 | unsigned long flags; |
4ac845a2 | 2841 | void *k; |
e2d74ac0 | 2842 | |
91fac317 VT |
2843 | if (unlikely(!ioc)) |
2844 | return NULL; | |
2845 | ||
d6de8be7 JA |
2846 | rcu_read_lock(); |
2847 | ||
597bc485 JA |
2848 | /* |
2849 | * we maintain a last-hit cache, to avoid browsing over the tree | |
2850 | */ | |
4ac845a2 | 2851 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
2852 | if (cic && cic->key == cfqd) { |
2853 | rcu_read_unlock(); | |
597bc485 | 2854 | return cic; |
d6de8be7 | 2855 | } |
597bc485 | 2856 | |
4ac845a2 | 2857 | do { |
4ac845a2 JA |
2858 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); |
2859 | rcu_read_unlock(); | |
2860 | if (!cic) | |
2861 | break; | |
be3b0753 OH |
2862 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
2863 | k = cic->key; | |
2864 | if (unlikely(!k)) { | |
4ac845a2 | 2865 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 2866 | rcu_read_lock(); |
4ac845a2 | 2867 | continue; |
dbecf3ab | 2868 | } |
e2d74ac0 | 2869 | |
d6de8be7 | 2870 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 2871 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 2872 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
2873 | break; |
2874 | } while (1); | |
e2d74ac0 | 2875 | |
4ac845a2 | 2876 | return cic; |
e2d74ac0 JA |
2877 | } |
2878 | ||
4ac845a2 JA |
2879 | /* |
2880 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
2881 | * the process specific cfq io context when entered from the block layer. | |
2882 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
2883 | */ | |
febffd61 JA |
2884 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
2885 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 2886 | { |
0261d688 | 2887 | unsigned long flags; |
4ac845a2 | 2888 | int ret; |
e2d74ac0 | 2889 | |
4ac845a2 JA |
2890 | ret = radix_tree_preload(gfp_mask); |
2891 | if (!ret) { | |
2892 | cic->ioc = ioc; | |
2893 | cic->key = cfqd; | |
e2d74ac0 | 2894 | |
4ac845a2 JA |
2895 | spin_lock_irqsave(&ioc->lock, flags); |
2896 | ret = radix_tree_insert(&ioc->radix_root, | |
2897 | (unsigned long) cfqd, cic); | |
ffc4e759 JA |
2898 | if (!ret) |
2899 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 2900 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 2901 | |
4ac845a2 JA |
2902 | radix_tree_preload_end(); |
2903 | ||
2904 | if (!ret) { | |
2905 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
2906 | list_add(&cic->queue_list, &cfqd->cic_list); | |
2907 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2908 | } | |
e2d74ac0 JA |
2909 | } |
2910 | ||
4ac845a2 JA |
2911 | if (ret) |
2912 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 2913 | |
4ac845a2 | 2914 | return ret; |
e2d74ac0 JA |
2915 | } |
2916 | ||
1da177e4 LT |
2917 | /* |
2918 | * Setup general io context and cfq io context. There can be several cfq | |
2919 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 2920 | * than one device managed by cfq. |
1da177e4 LT |
2921 | */ |
2922 | static struct cfq_io_context * | |
e2d74ac0 | 2923 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2924 | { |
22e2c507 | 2925 | struct io_context *ioc = NULL; |
1da177e4 | 2926 | struct cfq_io_context *cic; |
1da177e4 | 2927 | |
22e2c507 | 2928 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 2929 | |
b5deef90 | 2930 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
2931 | if (!ioc) |
2932 | return NULL; | |
2933 | ||
4ac845a2 | 2934 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
2935 | if (cic) |
2936 | goto out; | |
1da177e4 | 2937 | |
e2d74ac0 JA |
2938 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
2939 | if (cic == NULL) | |
2940 | goto err; | |
1da177e4 | 2941 | |
4ac845a2 JA |
2942 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
2943 | goto err_free; | |
2944 | ||
1da177e4 | 2945 | out: |
fc46379d JA |
2946 | smp_read_barrier_depends(); |
2947 | if (unlikely(ioc->ioprio_changed)) | |
2948 | cfq_ioc_set_ioprio(ioc); | |
2949 | ||
24610333 VG |
2950 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
2951 | if (unlikely(ioc->cgroup_changed)) | |
2952 | cfq_ioc_set_cgroup(ioc); | |
2953 | #endif | |
1da177e4 | 2954 | return cic; |
4ac845a2 JA |
2955 | err_free: |
2956 | cfq_cic_free(cic); | |
1da177e4 LT |
2957 | err: |
2958 | put_io_context(ioc); | |
2959 | return NULL; | |
2960 | } | |
2961 | ||
22e2c507 JA |
2962 | static void |
2963 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 2964 | { |
aaf1228d JA |
2965 | unsigned long elapsed = jiffies - cic->last_end_request; |
2966 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 2967 | |
22e2c507 JA |
2968 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
2969 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
2970 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
2971 | } | |
1da177e4 | 2972 | |
206dc69b | 2973 | static void |
b2c18e1e | 2974 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
6d048f53 | 2975 | struct request *rq) |
206dc69b JA |
2976 | { |
2977 | sector_t sdist; | |
2978 | u64 total; | |
2979 | ||
b2c18e1e | 2980 | if (!cfqq->last_request_pos) |
4d00aa47 | 2981 | sdist = 0; |
b2c18e1e JM |
2982 | else if (cfqq->last_request_pos < blk_rq_pos(rq)) |
2983 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
206dc69b | 2984 | else |
b2c18e1e | 2985 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); |
206dc69b JA |
2986 | |
2987 | /* | |
2988 | * Don't allow the seek distance to get too large from the | |
2989 | * odd fragment, pagein, etc | |
2990 | */ | |
b2c18e1e JM |
2991 | if (cfqq->seek_samples <= 60) /* second&third seek */ |
2992 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024); | |
206dc69b | 2993 | else |
b2c18e1e | 2994 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64); |
206dc69b | 2995 | |
b2c18e1e JM |
2996 | cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8; |
2997 | cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8; | |
2998 | total = cfqq->seek_total + (cfqq->seek_samples/2); | |
2999 | do_div(total, cfqq->seek_samples); | |
3000 | cfqq->seek_mean = (sector_t)total; | |
e6c5bc73 JM |
3001 | |
3002 | /* | |
3003 | * If this cfqq is shared between multiple processes, check to | |
3004 | * make sure that those processes are still issuing I/Os within | |
3005 | * the mean seek distance. If not, it may be time to break the | |
3006 | * queues apart again. | |
3007 | */ | |
3008 | if (cfq_cfqq_coop(cfqq)) { | |
3009 | if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start) | |
3010 | cfqq->seeky_start = jiffies; | |
3011 | else if (!CFQQ_SEEKY(cfqq)) | |
3012 | cfqq->seeky_start = 0; | |
3013 | } | |
206dc69b | 3014 | } |
1da177e4 | 3015 | |
22e2c507 JA |
3016 | /* |
3017 | * Disable idle window if the process thinks too long or seeks so much that | |
3018 | * it doesn't matter | |
3019 | */ | |
3020 | static void | |
3021 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3022 | struct cfq_io_context *cic) | |
3023 | { | |
7b679138 | 3024 | int old_idle, enable_idle; |
1be92f2f | 3025 | |
0871714e JA |
3026 | /* |
3027 | * Don't idle for async or idle io prio class | |
3028 | */ | |
3029 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
3030 | return; |
3031 | ||
c265a7f4 | 3032 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 3033 | |
76280aff CZ |
3034 | if (cfqq->queued[0] + cfqq->queued[1] >= 4) |
3035 | cfq_mark_cfqq_deep(cfqq); | |
3036 | ||
66dac98e | 3037 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || |
76280aff CZ |
3038 | (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples) |
3039 | && CFQQ_SEEKY(cfqq))) | |
22e2c507 JA |
3040 | enable_idle = 0; |
3041 | else if (sample_valid(cic->ttime_samples)) { | |
718eee05 | 3042 | if (cic->ttime_mean > cfqd->cfq_slice_idle) |
22e2c507 JA |
3043 | enable_idle = 0; |
3044 | else | |
3045 | enable_idle = 1; | |
1da177e4 LT |
3046 | } |
3047 | ||
7b679138 JA |
3048 | if (old_idle != enable_idle) { |
3049 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
3050 | if (enable_idle) | |
3051 | cfq_mark_cfqq_idle_window(cfqq); | |
3052 | else | |
3053 | cfq_clear_cfqq_idle_window(cfqq); | |
3054 | } | |
22e2c507 | 3055 | } |
1da177e4 | 3056 | |
22e2c507 JA |
3057 | /* |
3058 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
3059 | * no or if we aren't sure, a 1 will cause a preempt. | |
3060 | */ | |
a6151c3a | 3061 | static bool |
22e2c507 | 3062 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, |
5e705374 | 3063 | struct request *rq) |
22e2c507 | 3064 | { |
6d048f53 | 3065 | struct cfq_queue *cfqq; |
22e2c507 | 3066 | |
6d048f53 JA |
3067 | cfqq = cfqd->active_queue; |
3068 | if (!cfqq) | |
a6151c3a | 3069 | return false; |
22e2c507 | 3070 | |
6d048f53 | 3071 | if (cfq_class_idle(new_cfqq)) |
a6151c3a | 3072 | return false; |
22e2c507 JA |
3073 | |
3074 | if (cfq_class_idle(cfqq)) | |
a6151c3a | 3075 | return true; |
1e3335de | 3076 | |
374f84ac JA |
3077 | /* |
3078 | * if the new request is sync, but the currently running queue is | |
3079 | * not, let the sync request have priority. | |
3080 | */ | |
5e705374 | 3081 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
a6151c3a | 3082 | return true; |
1e3335de | 3083 | |
8682e1f1 VG |
3084 | if (new_cfqq->cfqg != cfqq->cfqg) |
3085 | return false; | |
3086 | ||
3087 | if (cfq_slice_used(cfqq)) | |
3088 | return true; | |
3089 | ||
3090 | /* Allow preemption only if we are idling on sync-noidle tree */ | |
3091 | if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && | |
3092 | cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && | |
3093 | new_cfqq->service_tree->count == 2 && | |
3094 | RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3095 | return true; | |
3096 | ||
374f84ac JA |
3097 | /* |
3098 | * So both queues are sync. Let the new request get disk time if | |
3099 | * it's a metadata request and the current queue is doing regular IO. | |
3100 | */ | |
3101 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
e6ec4fe2 | 3102 | return true; |
22e2c507 | 3103 | |
3a9a3f6c DS |
3104 | /* |
3105 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
3106 | */ | |
3107 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
a6151c3a | 3108 | return true; |
3a9a3f6c | 3109 | |
1e3335de | 3110 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
a6151c3a | 3111 | return false; |
1e3335de JA |
3112 | |
3113 | /* | |
3114 | * if this request is as-good as one we would expect from the | |
3115 | * current cfqq, let it preempt | |
3116 | */ | |
e00ef799 | 3117 | if (cfq_rq_close(cfqd, cfqq, rq)) |
a6151c3a | 3118 | return true; |
1e3335de | 3119 | |
a6151c3a | 3120 | return false; |
22e2c507 JA |
3121 | } |
3122 | ||
3123 | /* | |
3124 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
3125 | * let it have half of its nominal slice. | |
3126 | */ | |
3127 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3128 | { | |
7b679138 | 3129 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
6084cdda | 3130 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 3131 | |
bf572256 JA |
3132 | /* |
3133 | * Put the new queue at the front of the of the current list, | |
3134 | * so we know that it will be selected next. | |
3135 | */ | |
3136 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
3137 | |
3138 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 3139 | |
44f7c160 JA |
3140 | cfqq->slice_end = 0; |
3141 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
3142 | } |
3143 | ||
22e2c507 | 3144 | /* |
5e705374 | 3145 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
3146 | * something we should do about it |
3147 | */ | |
3148 | static void | |
5e705374 JA |
3149 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
3150 | struct request *rq) | |
22e2c507 | 3151 | { |
5e705374 | 3152 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 3153 | |
45333d5a | 3154 | cfqd->rq_queued++; |
374f84ac JA |
3155 | if (rq_is_meta(rq)) |
3156 | cfqq->meta_pending++; | |
3157 | ||
9c2c38a1 | 3158 | cfq_update_io_thinktime(cfqd, cic); |
b2c18e1e | 3159 | cfq_update_io_seektime(cfqd, cfqq, rq); |
9c2c38a1 JA |
3160 | cfq_update_idle_window(cfqd, cfqq, cic); |
3161 | ||
b2c18e1e | 3162 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
3163 | |
3164 | if (cfqq == cfqd->active_queue) { | |
f75edf2d VG |
3165 | if (cfq_cfqq_wait_busy(cfqq)) { |
3166 | cfq_clear_cfqq_wait_busy(cfqq); | |
3167 | cfq_mark_cfqq_wait_busy_done(cfqq); | |
3168 | } | |
22e2c507 | 3169 | /* |
b029195d JA |
3170 | * Remember that we saw a request from this process, but |
3171 | * don't start queuing just yet. Otherwise we risk seeing lots | |
3172 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
3173 | * and merging. If the request is already larger than a single |
3174 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
3175 | * merging is already done. Ditto for a busy system that |
3176 | * has other work pending, don't risk delaying until the | |
3177 | * idle timer unplug to continue working. | |
22e2c507 | 3178 | */ |
d6ceb25e | 3179 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
3180 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
3181 | cfqd->busy_queues > 1) { | |
d6ceb25e | 3182 | del_timer(&cfqd->idle_slice_timer); |
bf791937 VG |
3183 | __blk_run_queue(cfqd->queue); |
3184 | } else | |
3185 | cfq_mark_cfqq_must_dispatch(cfqq); | |
d6ceb25e | 3186 | } |
5e705374 | 3187 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
3188 | /* |
3189 | * not the active queue - expire current slice if it is | |
3190 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
3191 | * has some old slice time left and is of higher priority or |
3192 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
3193 | */ |
3194 | cfq_preempt_queue(cfqd, cfqq); | |
a7f55792 | 3195 | __blk_run_queue(cfqd->queue); |
22e2c507 | 3196 | } |
1da177e4 LT |
3197 | } |
3198 | ||
165125e1 | 3199 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 3200 | { |
b4878f24 | 3201 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 3202 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 3203 | |
7b679138 | 3204 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 3205 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 3206 | |
30996f40 | 3207 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); |
22e2c507 | 3208 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
aa6f6a3d | 3209 | cfq_add_rq_rb(rq); |
22e2c507 | 3210 | |
5e705374 | 3211 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
3212 | } |
3213 | ||
45333d5a AC |
3214 | /* |
3215 | * Update hw_tag based on peak queue depth over 50 samples under | |
3216 | * sufficient load. | |
3217 | */ | |
3218 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
3219 | { | |
1a1238a7 SL |
3220 | struct cfq_queue *cfqq = cfqd->active_queue; |
3221 | ||
e459dd08 CZ |
3222 | if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth) |
3223 | cfqd->hw_tag_est_depth = rq_in_driver(cfqd); | |
3224 | ||
3225 | if (cfqd->hw_tag == 1) | |
3226 | return; | |
45333d5a AC |
3227 | |
3228 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
5ad531db | 3229 | rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
3230 | return; |
3231 | ||
1a1238a7 SL |
3232 | /* |
3233 | * If active queue hasn't enough requests and can idle, cfq might not | |
3234 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
3235 | * case | |
3236 | */ | |
3237 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
3238 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
3239 | CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN) | |
3240 | return; | |
3241 | ||
45333d5a AC |
3242 | if (cfqd->hw_tag_samples++ < 50) |
3243 | return; | |
3244 | ||
e459dd08 | 3245 | if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
3246 | cfqd->hw_tag = 1; |
3247 | else | |
3248 | cfqd->hw_tag = 0; | |
45333d5a AC |
3249 | } |
3250 | ||
165125e1 | 3251 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 3252 | { |
5e705374 | 3253 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 3254 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 3255 | const int sync = rq_is_sync(rq); |
b4878f24 | 3256 | unsigned long now; |
1da177e4 | 3257 | |
b4878f24 | 3258 | now = jiffies; |
2868ef7b | 3259 | cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq)); |
1da177e4 | 3260 | |
45333d5a AC |
3261 | cfq_update_hw_tag(cfqd); |
3262 | ||
5ad531db | 3263 | WARN_ON(!cfqd->rq_in_driver[sync]); |
6d048f53 | 3264 | WARN_ON(!cfqq->dispatched); |
5ad531db | 3265 | cfqd->rq_in_driver[sync]--; |
6d048f53 | 3266 | cfqq->dispatched--; |
1da177e4 | 3267 | |
3ed9a296 JA |
3268 | if (cfq_cfqq_sync(cfqq)) |
3269 | cfqd->sync_flight--; | |
3270 | ||
365722bb | 3271 | if (sync) { |
5e705374 | 3272 | RQ_CIC(rq)->last_end_request = now; |
365722bb VG |
3273 | cfqd->last_end_sync_rq = now; |
3274 | } | |
caaa5f9f JA |
3275 | |
3276 | /* | |
3277 | * If this is the active queue, check if it needs to be expired, | |
3278 | * or if we want to idle in case it has no pending requests. | |
3279 | */ | |
3280 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
3281 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
3282 | ||
44f7c160 JA |
3283 | if (cfq_cfqq_slice_new(cfqq)) { |
3284 | cfq_set_prio_slice(cfqd, cfqq); | |
3285 | cfq_clear_cfqq_slice_new(cfqq); | |
3286 | } | |
f75edf2d VG |
3287 | |
3288 | /* | |
3289 | * If this queue consumed its slice and this is last queue | |
3290 | * in the group, wait for next request before we expire | |
3291 | * the queue | |
3292 | */ | |
3293 | if (cfq_slice_used(cfqq) && cfqq->cfqg->nr_cfqq == 1) { | |
3294 | cfqq->slice_end = jiffies + cfqd->cfq_slice_idle; | |
3295 | cfq_mark_cfqq_wait_busy(cfqq); | |
3296 | } | |
3297 | ||
a36e71f9 | 3298 | /* |
8e550632 CZ |
3299 | * Idling is not enabled on: |
3300 | * - expired queues | |
3301 | * - idle-priority queues | |
3302 | * - async queues | |
3303 | * - queues with still some requests queued | |
3304 | * - when there is a close cooperator | |
a36e71f9 | 3305 | */ |
0871714e | 3306 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
6084cdda | 3307 | cfq_slice_expired(cfqd, 1); |
8e550632 CZ |
3308 | else if (sync && cfqq_empty && |
3309 | !cfq_close_cooperator(cfqd, cfqq)) { | |
3310 | cfqd->noidle_tree_requires_idle |= !rq_noidle(rq); | |
3311 | /* | |
3312 | * Idling is enabled for SYNC_WORKLOAD. | |
3313 | * SYNC_NOIDLE_WORKLOAD idles at the end of the tree | |
3314 | * only if we processed at least one !rq_noidle request | |
3315 | */ | |
3316 | if (cfqd->serving_type == SYNC_WORKLOAD | |
3317 | || cfqd->noidle_tree_requires_idle) | |
3318 | cfq_arm_slice_timer(cfqd); | |
3319 | } | |
caaa5f9f | 3320 | } |
6d048f53 | 3321 | |
5ad531db | 3322 | if (!rq_in_driver(cfqd)) |
23e018a1 | 3323 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
3324 | } |
3325 | ||
22e2c507 JA |
3326 | /* |
3327 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
3328 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
3329 | */ | |
3330 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 3331 | { |
22e2c507 JA |
3332 | if (has_fs_excl()) { |
3333 | /* | |
3334 | * boost idle prio on transactions that would lock out other | |
3335 | * users of the filesystem | |
3336 | */ | |
3337 | if (cfq_class_idle(cfqq)) | |
3338 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
3339 | if (cfqq->ioprio > IOPRIO_NORM) | |
3340 | cfqq->ioprio = IOPRIO_NORM; | |
3341 | } else { | |
3342 | /* | |
dddb7451 | 3343 | * unboost the queue (if needed) |
22e2c507 | 3344 | */ |
dddb7451 CZ |
3345 | cfqq->ioprio_class = cfqq->org_ioprio_class; |
3346 | cfqq->ioprio = cfqq->org_ioprio; | |
22e2c507 | 3347 | } |
22e2c507 | 3348 | } |
1da177e4 | 3349 | |
89850f7e | 3350 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 3351 | { |
1b379d8d | 3352 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 3353 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 3354 | return ELV_MQUEUE_MUST; |
3b18152c | 3355 | } |
1da177e4 | 3356 | |
22e2c507 | 3357 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
3358 | } |
3359 | ||
165125e1 | 3360 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
3361 | { |
3362 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3363 | struct task_struct *tsk = current; | |
91fac317 | 3364 | struct cfq_io_context *cic; |
22e2c507 JA |
3365 | struct cfq_queue *cfqq; |
3366 | ||
3367 | /* | |
3368 | * don't force setup of a queue from here, as a call to may_queue | |
3369 | * does not necessarily imply that a request actually will be queued. | |
3370 | * so just lookup a possibly existing queue, or return 'may queue' | |
3371 | * if that fails | |
3372 | */ | |
4ac845a2 | 3373 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
3374 | if (!cic) |
3375 | return ELV_MQUEUE_MAY; | |
3376 | ||
b0b78f81 | 3377 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 3378 | if (cfqq) { |
fd0928df | 3379 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
3380 | cfq_prio_boost(cfqq); |
3381 | ||
89850f7e | 3382 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
3383 | } |
3384 | ||
3385 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
3386 | } |
3387 | ||
1da177e4 LT |
3388 | /* |
3389 | * queue lock held here | |
3390 | */ | |
bb37b94c | 3391 | static void cfq_put_request(struct request *rq) |
1da177e4 | 3392 | { |
5e705374 | 3393 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 3394 | |
5e705374 | 3395 | if (cfqq) { |
22e2c507 | 3396 | const int rw = rq_data_dir(rq); |
1da177e4 | 3397 | |
22e2c507 JA |
3398 | BUG_ON(!cfqq->allocated[rw]); |
3399 | cfqq->allocated[rw]--; | |
1da177e4 | 3400 | |
5e705374 | 3401 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 3402 | |
1da177e4 | 3403 | rq->elevator_private = NULL; |
5e705374 | 3404 | rq->elevator_private2 = NULL; |
1da177e4 | 3405 | |
1da177e4 LT |
3406 | cfq_put_queue(cfqq); |
3407 | } | |
3408 | } | |
3409 | ||
df5fe3e8 JM |
3410 | static struct cfq_queue * |
3411 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, | |
3412 | struct cfq_queue *cfqq) | |
3413 | { | |
3414 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
3415 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
b3b6d040 | 3416 | cfq_mark_cfqq_coop(cfqq->new_cfqq); |
df5fe3e8 JM |
3417 | cfq_put_queue(cfqq); |
3418 | return cic_to_cfqq(cic, 1); | |
3419 | } | |
3420 | ||
e6c5bc73 JM |
3421 | static int should_split_cfqq(struct cfq_queue *cfqq) |
3422 | { | |
3423 | if (cfqq->seeky_start && | |
3424 | time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT)) | |
3425 | return 1; | |
3426 | return 0; | |
3427 | } | |
3428 | ||
3429 | /* | |
3430 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
3431 | * was the last process referring to said cfqq. | |
3432 | */ | |
3433 | static struct cfq_queue * | |
3434 | split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) | |
3435 | { | |
3436 | if (cfqq_process_refs(cfqq) == 1) { | |
3437 | cfqq->seeky_start = 0; | |
3438 | cfqq->pid = current->pid; | |
3439 | cfq_clear_cfqq_coop(cfqq); | |
3440 | return cfqq; | |
3441 | } | |
3442 | ||
3443 | cic_set_cfqq(cic, NULL, 1); | |
3444 | cfq_put_queue(cfqq); | |
3445 | return NULL; | |
3446 | } | |
1da177e4 | 3447 | /* |
22e2c507 | 3448 | * Allocate cfq data structures associated with this request. |
1da177e4 | 3449 | */ |
22e2c507 | 3450 | static int |
165125e1 | 3451 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
3452 | { |
3453 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3454 | struct cfq_io_context *cic; | |
3455 | const int rw = rq_data_dir(rq); | |
a6151c3a | 3456 | const bool is_sync = rq_is_sync(rq); |
22e2c507 | 3457 | struct cfq_queue *cfqq; |
1da177e4 LT |
3458 | unsigned long flags; |
3459 | ||
3460 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
3461 | ||
e2d74ac0 | 3462 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 3463 | |
1da177e4 LT |
3464 | spin_lock_irqsave(q->queue_lock, flags); |
3465 | ||
22e2c507 JA |
3466 | if (!cic) |
3467 | goto queue_fail; | |
3468 | ||
e6c5bc73 | 3469 | new_queue: |
91fac317 | 3470 | cfqq = cic_to_cfqq(cic, is_sync); |
32f2e807 | 3471 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { |
fd0928df | 3472 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 | 3473 | cic_set_cfqq(cic, cfqq, is_sync); |
df5fe3e8 | 3474 | } else { |
e6c5bc73 JM |
3475 | /* |
3476 | * If the queue was seeky for too long, break it apart. | |
3477 | */ | |
3478 | if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) { | |
3479 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); | |
3480 | cfqq = split_cfqq(cic, cfqq); | |
3481 | if (!cfqq) | |
3482 | goto new_queue; | |
3483 | } | |
3484 | ||
df5fe3e8 JM |
3485 | /* |
3486 | * Check to see if this queue is scheduled to merge with | |
3487 | * another, closely cooperating queue. The merging of | |
3488 | * queues happens here as it must be done in process context. | |
3489 | * The reference on new_cfqq was taken in merge_cfqqs. | |
3490 | */ | |
3491 | if (cfqq->new_cfqq) | |
3492 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
91fac317 | 3493 | } |
1da177e4 LT |
3494 | |
3495 | cfqq->allocated[rw]++; | |
22e2c507 | 3496 | atomic_inc(&cfqq->ref); |
1da177e4 | 3497 | |
5e705374 | 3498 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 3499 | |
5e705374 JA |
3500 | rq->elevator_private = cic; |
3501 | rq->elevator_private2 = cfqq; | |
3502 | return 0; | |
1da177e4 | 3503 | |
22e2c507 JA |
3504 | queue_fail: |
3505 | if (cic) | |
3506 | put_io_context(cic->ioc); | |
89850f7e | 3507 | |
23e018a1 | 3508 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 3509 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 3510 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
3511 | return 1; |
3512 | } | |
3513 | ||
65f27f38 | 3514 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 3515 | { |
65f27f38 | 3516 | struct cfq_data *cfqd = |
23e018a1 | 3517 | container_of(work, struct cfq_data, unplug_work); |
165125e1 | 3518 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3519 | |
40bb54d1 | 3520 | spin_lock_irq(q->queue_lock); |
a7f55792 | 3521 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 3522 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
3523 | } |
3524 | ||
3525 | /* | |
3526 | * Timer running if the active_queue is currently idling inside its time slice | |
3527 | */ | |
3528 | static void cfq_idle_slice_timer(unsigned long data) | |
3529 | { | |
3530 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
3531 | struct cfq_queue *cfqq; | |
3532 | unsigned long flags; | |
3c6bd2f8 | 3533 | int timed_out = 1; |
22e2c507 | 3534 | |
7b679138 JA |
3535 | cfq_log(cfqd, "idle timer fired"); |
3536 | ||
22e2c507 JA |
3537 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
3538 | ||
fe094d98 JA |
3539 | cfqq = cfqd->active_queue; |
3540 | if (cfqq) { | |
3c6bd2f8 JA |
3541 | timed_out = 0; |
3542 | ||
b029195d JA |
3543 | /* |
3544 | * We saw a request before the queue expired, let it through | |
3545 | */ | |
3546 | if (cfq_cfqq_must_dispatch(cfqq)) | |
3547 | goto out_kick; | |
3548 | ||
22e2c507 JA |
3549 | /* |
3550 | * expired | |
3551 | */ | |
44f7c160 | 3552 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
3553 | goto expire; |
3554 | ||
3555 | /* | |
3556 | * only expire and reinvoke request handler, if there are | |
3557 | * other queues with pending requests | |
3558 | */ | |
caaa5f9f | 3559 | if (!cfqd->busy_queues) |
22e2c507 | 3560 | goto out_cont; |
22e2c507 JA |
3561 | |
3562 | /* | |
3563 | * not expired and it has a request pending, let it dispatch | |
3564 | */ | |
75e50984 | 3565 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 3566 | goto out_kick; |
76280aff CZ |
3567 | |
3568 | /* | |
3569 | * Queue depth flag is reset only when the idle didn't succeed | |
3570 | */ | |
3571 | cfq_clear_cfqq_deep(cfqq); | |
22e2c507 JA |
3572 | } |
3573 | expire: | |
6084cdda | 3574 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 3575 | out_kick: |
23e018a1 | 3576 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
3577 | out_cont: |
3578 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
3579 | } | |
3580 | ||
3b18152c JA |
3581 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
3582 | { | |
3583 | del_timer_sync(&cfqd->idle_slice_timer); | |
23e018a1 | 3584 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 3585 | } |
22e2c507 | 3586 | |
c2dea2d1 VT |
3587 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
3588 | { | |
3589 | int i; | |
3590 | ||
3591 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
3592 | if (cfqd->async_cfqq[0][i]) | |
3593 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
3594 | if (cfqd->async_cfqq[1][i]) | |
3595 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 3596 | } |
2389d1ef ON |
3597 | |
3598 | if (cfqd->async_idle_cfqq) | |
3599 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
3600 | } |
3601 | ||
b374d18a | 3602 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 3603 | { |
22e2c507 | 3604 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 3605 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3606 | |
3b18152c | 3607 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 3608 | |
d9ff4187 | 3609 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 3610 | |
d9ff4187 | 3611 | if (cfqd->active_queue) |
6084cdda | 3612 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
3613 | |
3614 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
3615 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
3616 | struct cfq_io_context, | |
3617 | queue_list); | |
89850f7e JA |
3618 | |
3619 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 3620 | } |
e2d74ac0 | 3621 | |
c2dea2d1 | 3622 | cfq_put_async_queues(cfqd); |
b1c35769 VG |
3623 | cfq_release_cfq_groups(cfqd); |
3624 | blkiocg_del_blkio_group(&cfqd->root_group.blkg); | |
15c31be4 | 3625 | |
d9ff4187 | 3626 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
3627 | |
3628 | cfq_shutdown_timer_wq(cfqd); | |
3629 | ||
b1c35769 VG |
3630 | /* Wait for cfqg->blkg->key accessors to exit their grace periods. */ |
3631 | synchronize_rcu(); | |
a90d742e | 3632 | kfree(cfqd); |
1da177e4 LT |
3633 | } |
3634 | ||
165125e1 | 3635 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
3636 | { |
3637 | struct cfq_data *cfqd; | |
718eee05 | 3638 | int i, j; |
cdb16e8f | 3639 | struct cfq_group *cfqg; |
615f0259 | 3640 | struct cfq_rb_root *st; |
1da177e4 | 3641 | |
94f6030c | 3642 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 3643 | if (!cfqd) |
bc1c1169 | 3644 | return NULL; |
1da177e4 | 3645 | |
1fa8f6d6 VG |
3646 | /* Init root service tree */ |
3647 | cfqd->grp_service_tree = CFQ_RB_ROOT; | |
3648 | ||
cdb16e8f VG |
3649 | /* Init root group */ |
3650 | cfqg = &cfqd->root_group; | |
615f0259 VG |
3651 | for_each_cfqg_st(cfqg, i, j, st) |
3652 | *st = CFQ_RB_ROOT; | |
1fa8f6d6 | 3653 | RB_CLEAR_NODE(&cfqg->rb_node); |
26a2ac00 | 3654 | |
25bc6b07 VG |
3655 | /* Give preference to root group over other groups */ |
3656 | cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT; | |
3657 | ||
25fb5169 | 3658 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
b1c35769 VG |
3659 | /* |
3660 | * Take a reference to root group which we never drop. This is just | |
3661 | * to make sure that cfq_put_cfqg() does not try to kfree root group | |
3662 | */ | |
3663 | atomic_set(&cfqg->ref, 1); | |
22084190 VG |
3664 | blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd, |
3665 | 0); | |
25fb5169 | 3666 | #endif |
26a2ac00 JA |
3667 | /* |
3668 | * Not strictly needed (since RB_ROOT just clears the node and we | |
3669 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
3670 | * to add magic to the rb code | |
3671 | */ | |
3672 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
3673 | cfqd->prio_trees[i] = RB_ROOT; | |
3674 | ||
6118b70b JA |
3675 | /* |
3676 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
3677 | * Grab a permanent reference to it, so that the normal code flow | |
3678 | * will not attempt to free it. | |
3679 | */ | |
3680 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
3681 | atomic_inc(&cfqd->oom_cfqq.ref); | |
cdb16e8f | 3682 | cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group); |
6118b70b | 3683 | |
d9ff4187 | 3684 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 3685 | |
1da177e4 | 3686 | cfqd->queue = q; |
1da177e4 | 3687 | |
22e2c507 JA |
3688 | init_timer(&cfqd->idle_slice_timer); |
3689 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
3690 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
3691 | ||
23e018a1 | 3692 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 3693 | |
1da177e4 | 3694 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
3695 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
3696 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
3697 | cfqd->cfq_back_max = cfq_back_max; |
3698 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
3699 | cfqd->cfq_slice[0] = cfq_slice_async; |
3700 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
3701 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
3702 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
963b72fc | 3703 | cfqd->cfq_latency = 1; |
ae30c286 | 3704 | cfqd->cfq_group_isolation = 0; |
e459dd08 | 3705 | cfqd->hw_tag = -1; |
365722bb | 3706 | cfqd->last_end_sync_rq = jiffies; |
bc1c1169 | 3707 | return cfqd; |
1da177e4 LT |
3708 | } |
3709 | ||
3710 | static void cfq_slab_kill(void) | |
3711 | { | |
d6de8be7 JA |
3712 | /* |
3713 | * Caller already ensured that pending RCU callbacks are completed, | |
3714 | * so we should have no busy allocations at this point. | |
3715 | */ | |
1da177e4 LT |
3716 | if (cfq_pool) |
3717 | kmem_cache_destroy(cfq_pool); | |
3718 | if (cfq_ioc_pool) | |
3719 | kmem_cache_destroy(cfq_ioc_pool); | |
3720 | } | |
3721 | ||
3722 | static int __init cfq_slab_setup(void) | |
3723 | { | |
0a31bd5f | 3724 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
3725 | if (!cfq_pool) |
3726 | goto fail; | |
3727 | ||
34e6bbf2 | 3728 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
3729 | if (!cfq_ioc_pool) |
3730 | goto fail; | |
3731 | ||
3732 | return 0; | |
3733 | fail: | |
3734 | cfq_slab_kill(); | |
3735 | return -ENOMEM; | |
3736 | } | |
3737 | ||
1da177e4 LT |
3738 | /* |
3739 | * sysfs parts below --> | |
3740 | */ | |
1da177e4 LT |
3741 | static ssize_t |
3742 | cfq_var_show(unsigned int var, char *page) | |
3743 | { | |
3744 | return sprintf(page, "%d\n", var); | |
3745 | } | |
3746 | ||
3747 | static ssize_t | |
3748 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
3749 | { | |
3750 | char *p = (char *) page; | |
3751 | ||
3752 | *var = simple_strtoul(p, &p, 10); | |
3753 | return count; | |
3754 | } | |
3755 | ||
1da177e4 | 3756 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 3757 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 3758 | { \ |
3d1ab40f | 3759 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3760 | unsigned int __data = __VAR; \ |
3761 | if (__CONV) \ | |
3762 | __data = jiffies_to_msecs(__data); \ | |
3763 | return cfq_var_show(__data, (page)); \ | |
3764 | } | |
3765 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
3766 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
3767 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
3768 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
3769 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
3770 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
3771 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
3772 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
3773 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
963b72fc | 3774 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); |
ae30c286 | 3775 | SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0); |
1da177e4 LT |
3776 | #undef SHOW_FUNCTION |
3777 | ||
3778 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 3779 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 3780 | { \ |
3d1ab40f | 3781 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3782 | unsigned int __data; \ |
3783 | int ret = cfq_var_store(&__data, (page), count); \ | |
3784 | if (__data < (MIN)) \ | |
3785 | __data = (MIN); \ | |
3786 | else if (__data > (MAX)) \ | |
3787 | __data = (MAX); \ | |
3788 | if (__CONV) \ | |
3789 | *(__PTR) = msecs_to_jiffies(__data); \ | |
3790 | else \ | |
3791 | *(__PTR) = __data; \ | |
3792 | return ret; \ | |
3793 | } | |
3794 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
3795 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
3796 | UINT_MAX, 1); | |
3797 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
3798 | UINT_MAX, 1); | |
e572ec7e | 3799 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
3800 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
3801 | UINT_MAX, 0); | |
22e2c507 JA |
3802 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
3803 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
3804 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
3805 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
3806 | UINT_MAX, 0); | |
963b72fc | 3807 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); |
ae30c286 | 3808 | STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0); |
1da177e4 LT |
3809 | #undef STORE_FUNCTION |
3810 | ||
e572ec7e AV |
3811 | #define CFQ_ATTR(name) \ |
3812 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
3813 | ||
3814 | static struct elv_fs_entry cfq_attrs[] = { | |
3815 | CFQ_ATTR(quantum), | |
e572ec7e AV |
3816 | CFQ_ATTR(fifo_expire_sync), |
3817 | CFQ_ATTR(fifo_expire_async), | |
3818 | CFQ_ATTR(back_seek_max), | |
3819 | CFQ_ATTR(back_seek_penalty), | |
3820 | CFQ_ATTR(slice_sync), | |
3821 | CFQ_ATTR(slice_async), | |
3822 | CFQ_ATTR(slice_async_rq), | |
3823 | CFQ_ATTR(slice_idle), | |
963b72fc | 3824 | CFQ_ATTR(low_latency), |
ae30c286 | 3825 | CFQ_ATTR(group_isolation), |
e572ec7e | 3826 | __ATTR_NULL |
1da177e4 LT |
3827 | }; |
3828 | ||
1da177e4 LT |
3829 | static struct elevator_type iosched_cfq = { |
3830 | .ops = { | |
3831 | .elevator_merge_fn = cfq_merge, | |
3832 | .elevator_merged_fn = cfq_merged_request, | |
3833 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 3834 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 3835 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 3836 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 3837 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
3838 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
3839 | .elevator_queue_empty_fn = cfq_queue_empty, | |
3840 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
3841 | .elevator_former_req_fn = elv_rb_former_request, |
3842 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
3843 | .elevator_set_req_fn = cfq_set_request, |
3844 | .elevator_put_req_fn = cfq_put_request, | |
3845 | .elevator_may_queue_fn = cfq_may_queue, | |
3846 | .elevator_init_fn = cfq_init_queue, | |
3847 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 3848 | .trim = cfq_free_io_context, |
1da177e4 | 3849 | }, |
3d1ab40f | 3850 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
3851 | .elevator_name = "cfq", |
3852 | .elevator_owner = THIS_MODULE, | |
3853 | }; | |
3854 | ||
3855 | static int __init cfq_init(void) | |
3856 | { | |
22e2c507 JA |
3857 | /* |
3858 | * could be 0 on HZ < 1000 setups | |
3859 | */ | |
3860 | if (!cfq_slice_async) | |
3861 | cfq_slice_async = 1; | |
3862 | if (!cfq_slice_idle) | |
3863 | cfq_slice_idle = 1; | |
3864 | ||
1da177e4 LT |
3865 | if (cfq_slab_setup()) |
3866 | return -ENOMEM; | |
3867 | ||
2fdd82bd | 3868 | elv_register(&iosched_cfq); |
1da177e4 | 3869 | |
2fdd82bd | 3870 | return 0; |
1da177e4 LT |
3871 | } |
3872 | ||
3873 | static void __exit cfq_exit(void) | |
3874 | { | |
6e9a4738 | 3875 | DECLARE_COMPLETION_ONSTACK(all_gone); |
1da177e4 | 3876 | elv_unregister(&iosched_cfq); |
334e94de | 3877 | ioc_gone = &all_gone; |
fba82272 OH |
3878 | /* ioc_gone's update must be visible before reading ioc_count */ |
3879 | smp_wmb(); | |
d6de8be7 JA |
3880 | |
3881 | /* | |
3882 | * this also protects us from entering cfq_slab_kill() with | |
3883 | * pending RCU callbacks | |
3884 | */ | |
245b2e70 | 3885 | if (elv_ioc_count_read(cfq_ioc_count)) |
9a11b4ed | 3886 | wait_for_completion(&all_gone); |
83521d3e | 3887 | cfq_slab_kill(); |
1da177e4 LT |
3888 | } |
3889 | ||
3890 | module_init(cfq_init); | |
3891 | module_exit(cfq_exit); | |
3892 | ||
3893 | MODULE_AUTHOR("Jens Axboe"); | |
3894 | MODULE_LICENSE("GPL"); | |
3895 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |