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> |
5a0e3ad6 | 10 | #include <linux/slab.h> |
1cc9be68 AV |
11 | #include <linux/blkdev.h> |
12 | #include <linux/elevator.h> | |
ad5ebd2f | 13 | #include <linux/jiffies.h> |
1da177e4 | 14 | #include <linux/rbtree.h> |
22e2c507 | 15 | #include <linux/ioprio.h> |
7b679138 | 16 | #include <linux/blktrace_api.h> |
e98ef89b | 17 | #include "cfq.h" |
1da177e4 LT |
18 | |
19 | /* | |
20 | * tunables | |
21 | */ | |
fe094d98 | 22 | /* max queue in one round of service */ |
abc3c744 | 23 | static const int cfq_quantum = 8; |
64100099 | 24 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
fe094d98 JA |
25 | /* maximum backwards seek, in KiB */ |
26 | static const int cfq_back_max = 16 * 1024; | |
27 | /* penalty of a backwards seek */ | |
28 | static const int cfq_back_penalty = 2; | |
64100099 | 29 | static const int cfq_slice_sync = HZ / 10; |
3b18152c | 30 | static int cfq_slice_async = HZ / 25; |
64100099 | 31 | static const int cfq_slice_async_rq = 2; |
caaa5f9f | 32 | static int cfq_slice_idle = HZ / 125; |
80bdf0c7 | 33 | static int cfq_group_idle = HZ / 125; |
5db5d642 CZ |
34 | static const int cfq_target_latency = HZ * 3/10; /* 300 ms */ |
35 | static const int cfq_hist_divisor = 4; | |
22e2c507 | 36 | |
d9e7620e | 37 | /* |
0871714e | 38 | * offset from end of service tree |
d9e7620e | 39 | */ |
0871714e | 40 | #define CFQ_IDLE_DELAY (HZ / 5) |
d9e7620e JA |
41 | |
42 | /* | |
43 | * below this threshold, we consider thinktime immediate | |
44 | */ | |
45 | #define CFQ_MIN_TT (2) | |
46 | ||
22e2c507 | 47 | #define CFQ_SLICE_SCALE (5) |
45333d5a | 48 | #define CFQ_HW_QUEUE_MIN (5) |
25bc6b07 | 49 | #define CFQ_SERVICE_SHIFT 12 |
22e2c507 | 50 | |
3dde36dd | 51 | #define CFQQ_SEEK_THR (sector_t)(8 * 100) |
e9ce335d | 52 | #define CFQQ_CLOSE_THR (sector_t)(8 * 1024) |
41647e7a | 53 | #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32) |
3dde36dd | 54 | #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8) |
ae54abed | 55 | |
fe094d98 | 56 | #define RQ_CIC(rq) \ |
c186794d MS |
57 | ((struct cfq_io_context *) (rq)->elevator_private[0]) |
58 | #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private[1]) | |
59 | #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private[2]) | |
1da177e4 | 60 | |
e18b890b CL |
61 | static struct kmem_cache *cfq_pool; |
62 | static struct kmem_cache *cfq_ioc_pool; | |
1da177e4 | 63 | |
245b2e70 | 64 | static DEFINE_PER_CPU(unsigned long, cfq_ioc_count); |
334e94de | 65 | static struct completion *ioc_gone; |
9a11b4ed | 66 | static DEFINE_SPINLOCK(ioc_gone_lock); |
334e94de | 67 | |
80b15c73 KK |
68 | static DEFINE_SPINLOCK(cic_index_lock); |
69 | static DEFINE_IDA(cic_index_ida); | |
70 | ||
22e2c507 JA |
71 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR |
72 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
22e2c507 JA |
73 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) |
74 | ||
206dc69b | 75 | #define sample_valid(samples) ((samples) > 80) |
1fa8f6d6 | 76 | #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node) |
206dc69b | 77 | |
cc09e299 JA |
78 | /* |
79 | * Most of our rbtree usage is for sorting with min extraction, so | |
80 | * if we cache the leftmost node we don't have to walk down the tree | |
81 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
82 | * move this into the elevator for the rq sorting as well. | |
83 | */ | |
84 | struct cfq_rb_root { | |
85 | struct rb_root rb; | |
86 | struct rb_node *left; | |
aa6f6a3d | 87 | unsigned count; |
73e9ffdd | 88 | unsigned total_weight; |
1fa8f6d6 | 89 | u64 min_vdisktime; |
f5f2b6ce | 90 | struct cfq_ttime ttime; |
cc09e299 | 91 | }; |
f5f2b6ce SL |
92 | #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \ |
93 | .ttime = {.last_end_request = jiffies,},} | |
cc09e299 | 94 | |
6118b70b JA |
95 | /* |
96 | * Per process-grouping structure | |
97 | */ | |
98 | struct cfq_queue { | |
99 | /* reference count */ | |
30d7b944 | 100 | int ref; |
6118b70b JA |
101 | /* various state flags, see below */ |
102 | unsigned int flags; | |
103 | /* parent cfq_data */ | |
104 | struct cfq_data *cfqd; | |
105 | /* service_tree member */ | |
106 | struct rb_node rb_node; | |
107 | /* service_tree key */ | |
108 | unsigned long rb_key; | |
109 | /* prio tree member */ | |
110 | struct rb_node p_node; | |
111 | /* prio tree root we belong to, if any */ | |
112 | struct rb_root *p_root; | |
113 | /* sorted list of pending requests */ | |
114 | struct rb_root sort_list; | |
115 | /* if fifo isn't expired, next request to serve */ | |
116 | struct request *next_rq; | |
117 | /* requests queued in sort_list */ | |
118 | int queued[2]; | |
119 | /* currently allocated requests */ | |
120 | int allocated[2]; | |
121 | /* fifo list of requests in sort_list */ | |
122 | struct list_head fifo; | |
123 | ||
dae739eb VG |
124 | /* time when queue got scheduled in to dispatch first request. */ |
125 | unsigned long dispatch_start; | |
f75edf2d | 126 | unsigned int allocated_slice; |
c4081ba5 | 127 | unsigned int slice_dispatch; |
dae739eb VG |
128 | /* time when first request from queue completed and slice started. */ |
129 | unsigned long slice_start; | |
6118b70b JA |
130 | unsigned long slice_end; |
131 | long slice_resid; | |
6118b70b | 132 | |
65299a3b CH |
133 | /* pending priority requests */ |
134 | int prio_pending; | |
6118b70b JA |
135 | /* number of requests that are on the dispatch list or inside driver */ |
136 | int dispatched; | |
137 | ||
138 | /* io prio of this group */ | |
139 | unsigned short ioprio, org_ioprio; | |
4aede84b | 140 | unsigned short ioprio_class; |
6118b70b | 141 | |
c4081ba5 RK |
142 | pid_t pid; |
143 | ||
3dde36dd | 144 | u32 seek_history; |
b2c18e1e JM |
145 | sector_t last_request_pos; |
146 | ||
aa6f6a3d | 147 | struct cfq_rb_root *service_tree; |
df5fe3e8 | 148 | struct cfq_queue *new_cfqq; |
cdb16e8f | 149 | struct cfq_group *cfqg; |
c4e7893e VG |
150 | /* Number of sectors dispatched from queue in single dispatch round */ |
151 | unsigned long nr_sectors; | |
6118b70b JA |
152 | }; |
153 | ||
c0324a02 | 154 | /* |
718eee05 | 155 | * First index in the service_trees. |
c0324a02 CZ |
156 | * IDLE is handled separately, so it has negative index |
157 | */ | |
158 | enum wl_prio_t { | |
c0324a02 | 159 | BE_WORKLOAD = 0, |
615f0259 VG |
160 | RT_WORKLOAD = 1, |
161 | IDLE_WORKLOAD = 2, | |
b4627321 | 162 | CFQ_PRIO_NR, |
c0324a02 CZ |
163 | }; |
164 | ||
718eee05 CZ |
165 | /* |
166 | * Second index in the service_trees. | |
167 | */ | |
168 | enum wl_type_t { | |
169 | ASYNC_WORKLOAD = 0, | |
170 | SYNC_NOIDLE_WORKLOAD = 1, | |
171 | SYNC_WORKLOAD = 2 | |
172 | }; | |
173 | ||
cdb16e8f VG |
174 | /* This is per cgroup per device grouping structure */ |
175 | struct cfq_group { | |
1fa8f6d6 VG |
176 | /* group service_tree member */ |
177 | struct rb_node rb_node; | |
178 | ||
179 | /* group service_tree key */ | |
180 | u64 vdisktime; | |
25bc6b07 | 181 | unsigned int weight; |
8184f93e JT |
182 | unsigned int new_weight; |
183 | bool needs_update; | |
1fa8f6d6 VG |
184 | |
185 | /* number of cfqq currently on this group */ | |
186 | int nr_cfqq; | |
187 | ||
cdb16e8f | 188 | /* |
4495a7d4 | 189 | * Per group busy queues average. Useful for workload slice calc. We |
b4627321 VG |
190 | * create the array for each prio class but at run time it is used |
191 | * only for RT and BE class and slot for IDLE class remains unused. | |
192 | * This is primarily done to avoid confusion and a gcc warning. | |
193 | */ | |
194 | unsigned int busy_queues_avg[CFQ_PRIO_NR]; | |
195 | /* | |
196 | * rr lists of queues with requests. We maintain service trees for | |
197 | * RT and BE classes. These trees are subdivided in subclasses | |
198 | * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE | |
199 | * class there is no subclassification and all the cfq queues go on | |
200 | * a single tree service_tree_idle. | |
cdb16e8f VG |
201 | * Counts are embedded in the cfq_rb_root |
202 | */ | |
203 | struct cfq_rb_root service_trees[2][3]; | |
204 | struct cfq_rb_root service_tree_idle; | |
dae739eb VG |
205 | |
206 | unsigned long saved_workload_slice; | |
207 | enum wl_type_t saved_workload; | |
208 | enum wl_prio_t saved_serving_prio; | |
25fb5169 VG |
209 | struct blkio_group blkg; |
210 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
211 | struct hlist_node cfqd_node; | |
329a6781 | 212 | int ref; |
25fb5169 | 213 | #endif |
80bdf0c7 VG |
214 | /* number of requests that are on the dispatch list or inside driver */ |
215 | int dispatched; | |
7700fc4f | 216 | struct cfq_ttime ttime; |
cdb16e8f | 217 | }; |
718eee05 | 218 | |
22e2c507 JA |
219 | /* |
220 | * Per block device queue structure | |
221 | */ | |
1da177e4 | 222 | struct cfq_data { |
165125e1 | 223 | struct request_queue *queue; |
1fa8f6d6 VG |
224 | /* Root service tree for cfq_groups */ |
225 | struct cfq_rb_root grp_service_tree; | |
cdb16e8f | 226 | struct cfq_group root_group; |
22e2c507 | 227 | |
c0324a02 CZ |
228 | /* |
229 | * The priority currently being served | |
22e2c507 | 230 | */ |
c0324a02 | 231 | enum wl_prio_t serving_prio; |
718eee05 CZ |
232 | enum wl_type_t serving_type; |
233 | unsigned long workload_expires; | |
cdb16e8f | 234 | struct cfq_group *serving_group; |
a36e71f9 JA |
235 | |
236 | /* | |
237 | * Each priority tree is sorted by next_request position. These | |
238 | * trees are used when determining if two or more queues are | |
239 | * interleaving requests (see cfq_close_cooperator). | |
240 | */ | |
241 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
242 | ||
22e2c507 | 243 | unsigned int busy_queues; |
ef8a41df | 244 | unsigned int busy_sync_queues; |
22e2c507 | 245 | |
53c583d2 CZ |
246 | int rq_in_driver; |
247 | int rq_in_flight[2]; | |
45333d5a AC |
248 | |
249 | /* | |
250 | * queue-depth detection | |
251 | */ | |
252 | int rq_queued; | |
25776e35 | 253 | int hw_tag; |
e459dd08 CZ |
254 | /* |
255 | * hw_tag can be | |
256 | * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) | |
257 | * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) | |
258 | * 0 => no NCQ | |
259 | */ | |
260 | int hw_tag_est_depth; | |
261 | unsigned int hw_tag_samples; | |
1da177e4 | 262 | |
22e2c507 JA |
263 | /* |
264 | * idle window management | |
265 | */ | |
266 | struct timer_list idle_slice_timer; | |
23e018a1 | 267 | struct work_struct unplug_work; |
1da177e4 | 268 | |
22e2c507 JA |
269 | struct cfq_queue *active_queue; |
270 | struct cfq_io_context *active_cic; | |
22e2c507 | 271 | |
c2dea2d1 VT |
272 | /* |
273 | * async queue for each priority case | |
274 | */ | |
275 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
276 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 277 | |
6d048f53 | 278 | sector_t last_position; |
1da177e4 | 279 | |
1da177e4 LT |
280 | /* |
281 | * tunables, see top of file | |
282 | */ | |
283 | unsigned int cfq_quantum; | |
22e2c507 | 284 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
285 | unsigned int cfq_back_penalty; |
286 | unsigned int cfq_back_max; | |
22e2c507 JA |
287 | unsigned int cfq_slice[2]; |
288 | unsigned int cfq_slice_async_rq; | |
289 | unsigned int cfq_slice_idle; | |
80bdf0c7 | 290 | unsigned int cfq_group_idle; |
963b72fc | 291 | unsigned int cfq_latency; |
d9ff4187 | 292 | |
80b15c73 | 293 | unsigned int cic_index; |
d9ff4187 | 294 | struct list_head cic_list; |
1da177e4 | 295 | |
6118b70b JA |
296 | /* |
297 | * Fallback dummy cfqq for extreme OOM conditions | |
298 | */ | |
299 | struct cfq_queue oom_cfqq; | |
365722bb | 300 | |
573412b2 | 301 | unsigned long last_delayed_sync; |
25fb5169 VG |
302 | |
303 | /* List of cfq groups being managed on this device*/ | |
304 | struct hlist_head cfqg_list; | |
56edf7d7 VG |
305 | |
306 | /* Number of groups which are on blkcg->blkg_list */ | |
307 | unsigned int nr_blkcg_linked_grps; | |
1da177e4 LT |
308 | }; |
309 | ||
25fb5169 VG |
310 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd); |
311 | ||
cdb16e8f VG |
312 | static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, |
313 | enum wl_prio_t prio, | |
65b32a57 | 314 | enum wl_type_t type) |
c0324a02 | 315 | { |
1fa8f6d6 VG |
316 | if (!cfqg) |
317 | return NULL; | |
318 | ||
c0324a02 | 319 | if (prio == IDLE_WORKLOAD) |
cdb16e8f | 320 | return &cfqg->service_tree_idle; |
c0324a02 | 321 | |
cdb16e8f | 322 | return &cfqg->service_trees[prio][type]; |
c0324a02 CZ |
323 | } |
324 | ||
3b18152c | 325 | enum cfqq_state_flags { |
b0b8d749 JA |
326 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
327 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 328 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 | 329 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ |
b0b8d749 JA |
330 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
331 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
332 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 333 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 334 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
b3b6d040 | 335 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ |
ae54abed | 336 | CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */ |
76280aff | 337 | CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ |
f75edf2d | 338 | CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */ |
3b18152c JA |
339 | }; |
340 | ||
341 | #define CFQ_CFQQ_FNS(name) \ | |
342 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
343 | { \ | |
fe094d98 | 344 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
345 | } \ |
346 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
347 | { \ | |
fe094d98 | 348 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
349 | } \ |
350 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
351 | { \ | |
fe094d98 | 352 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
353 | } |
354 | ||
355 | CFQ_CFQQ_FNS(on_rr); | |
356 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 357 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c | 358 | CFQ_CFQQ_FNS(must_alloc_slice); |
3b18152c JA |
359 | CFQ_CFQQ_FNS(fifo_expire); |
360 | CFQ_CFQQ_FNS(idle_window); | |
361 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 362 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 363 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 364 | CFQ_CFQQ_FNS(coop); |
ae54abed | 365 | CFQ_CFQQ_FNS(split_coop); |
76280aff | 366 | CFQ_CFQQ_FNS(deep); |
f75edf2d | 367 | CFQ_CFQQ_FNS(wait_busy); |
3b18152c JA |
368 | #undef CFQ_CFQQ_FNS |
369 | ||
afc24d49 | 370 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
2868ef7b VG |
371 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
372 | blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \ | |
373 | cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \ | |
4495a7d4 | 374 | blkg_path(&(cfqq)->cfqg->blkg), ##args) |
2868ef7b VG |
375 | |
376 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \ | |
377 | blk_add_trace_msg((cfqd)->queue, "%s " fmt, \ | |
4495a7d4 | 378 | blkg_path(&(cfqg)->blkg), ##args) \ |
2868ef7b VG |
379 | |
380 | #else | |
7b679138 JA |
381 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
382 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
4495a7d4 | 383 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0) |
2868ef7b | 384 | #endif |
7b679138 JA |
385 | #define cfq_log(cfqd, fmt, args...) \ |
386 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
387 | ||
615f0259 VG |
388 | /* Traverses through cfq group service trees */ |
389 | #define for_each_cfqg_st(cfqg, i, j, st) \ | |
390 | for (i = 0; i <= IDLE_WORKLOAD; i++) \ | |
391 | for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ | |
392 | : &cfqg->service_tree_idle; \ | |
393 | (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ | |
394 | (i == IDLE_WORKLOAD && j == 0); \ | |
395 | j++, st = i < IDLE_WORKLOAD ? \ | |
396 | &cfqg->service_trees[i][j]: NULL) \ | |
397 | ||
f5f2b6ce SL |
398 | static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd, |
399 | struct cfq_ttime *ttime, bool group_idle) | |
400 | { | |
401 | unsigned long slice; | |
402 | if (!sample_valid(ttime->ttime_samples)) | |
403 | return false; | |
404 | if (group_idle) | |
405 | slice = cfqd->cfq_group_idle; | |
406 | else | |
407 | slice = cfqd->cfq_slice_idle; | |
408 | return ttime->ttime_mean > slice; | |
409 | } | |
615f0259 | 410 | |
02b35081 VG |
411 | static inline bool iops_mode(struct cfq_data *cfqd) |
412 | { | |
413 | /* | |
414 | * If we are not idling on queues and it is a NCQ drive, parallel | |
415 | * execution of requests is on and measuring time is not possible | |
416 | * in most of the cases until and unless we drive shallower queue | |
417 | * depths and that becomes a performance bottleneck. In such cases | |
418 | * switch to start providing fairness in terms of number of IOs. | |
419 | */ | |
420 | if (!cfqd->cfq_slice_idle && cfqd->hw_tag) | |
421 | return true; | |
422 | else | |
423 | return false; | |
424 | } | |
425 | ||
c0324a02 CZ |
426 | static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) |
427 | { | |
428 | if (cfq_class_idle(cfqq)) | |
429 | return IDLE_WORKLOAD; | |
430 | if (cfq_class_rt(cfqq)) | |
431 | return RT_WORKLOAD; | |
432 | return BE_WORKLOAD; | |
433 | } | |
434 | ||
718eee05 CZ |
435 | |
436 | static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) | |
437 | { | |
438 | if (!cfq_cfqq_sync(cfqq)) | |
439 | return ASYNC_WORKLOAD; | |
440 | if (!cfq_cfqq_idle_window(cfqq)) | |
441 | return SYNC_NOIDLE_WORKLOAD; | |
442 | return SYNC_WORKLOAD; | |
443 | } | |
444 | ||
58ff82f3 VG |
445 | static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl, |
446 | struct cfq_data *cfqd, | |
447 | struct cfq_group *cfqg) | |
c0324a02 CZ |
448 | { |
449 | if (wl == IDLE_WORKLOAD) | |
cdb16e8f | 450 | return cfqg->service_tree_idle.count; |
c0324a02 | 451 | |
cdb16e8f VG |
452 | return cfqg->service_trees[wl][ASYNC_WORKLOAD].count |
453 | + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count | |
454 | + cfqg->service_trees[wl][SYNC_WORKLOAD].count; | |
c0324a02 CZ |
455 | } |
456 | ||
f26bd1f0 VG |
457 | static inline int cfqg_busy_async_queues(struct cfq_data *cfqd, |
458 | struct cfq_group *cfqg) | |
459 | { | |
460 | return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count | |
461 | + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count; | |
462 | } | |
463 | ||
165125e1 | 464 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
a6151c3a | 465 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, |
fd0928df | 466 | struct io_context *, gfp_t); |
4ac845a2 | 467 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
468 | struct io_context *); |
469 | ||
470 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 471 | bool is_sync) |
91fac317 | 472 | { |
a6151c3a | 473 | return cic->cfqq[is_sync]; |
91fac317 VT |
474 | } |
475 | ||
476 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 477 | struct cfq_queue *cfqq, bool is_sync) |
91fac317 | 478 | { |
a6151c3a | 479 | cic->cfqq[is_sync] = cfqq; |
91fac317 VT |
480 | } |
481 | ||
bca4b914 | 482 | #define CIC_DEAD_KEY 1ul |
80b15c73 | 483 | #define CIC_DEAD_INDEX_SHIFT 1 |
bca4b914 KK |
484 | |
485 | static inline void *cfqd_dead_key(struct cfq_data *cfqd) | |
486 | { | |
80b15c73 | 487 | return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY); |
bca4b914 KK |
488 | } |
489 | ||
490 | static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic) | |
491 | { | |
492 | struct cfq_data *cfqd = cic->key; | |
493 | ||
494 | if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY)) | |
495 | return NULL; | |
496 | ||
497 | return cfqd; | |
498 | } | |
499 | ||
91fac317 VT |
500 | /* |
501 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
502 | * set (in which case it could also be direct WRITE). | |
503 | */ | |
a6151c3a | 504 | static inline bool cfq_bio_sync(struct bio *bio) |
91fac317 | 505 | { |
7b6d91da | 506 | return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC); |
91fac317 | 507 | } |
1da177e4 | 508 | |
99f95e52 AM |
509 | /* |
510 | * scheduler run of queue, if there are requests pending and no one in the | |
511 | * driver that will restart queueing | |
512 | */ | |
23e018a1 | 513 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) |
99f95e52 | 514 | { |
7b679138 JA |
515 | if (cfqd->busy_queues) { |
516 | cfq_log(cfqd, "schedule dispatch"); | |
23e018a1 | 517 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 518 | } |
99f95e52 AM |
519 | } |
520 | ||
44f7c160 JA |
521 | /* |
522 | * Scale schedule slice based on io priority. Use the sync time slice only | |
523 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
524 | * io only, should not get full sync slice length. | |
525 | */ | |
a6151c3a | 526 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, |
d9e7620e | 527 | unsigned short prio) |
44f7c160 | 528 | { |
d9e7620e | 529 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 530 | |
d9e7620e JA |
531 | WARN_ON(prio >= IOPRIO_BE_NR); |
532 | ||
533 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
534 | } | |
44f7c160 | 535 | |
d9e7620e JA |
536 | static inline int |
537 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
538 | { | |
539 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
540 | } |
541 | ||
25bc6b07 VG |
542 | static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg) |
543 | { | |
544 | u64 d = delta << CFQ_SERVICE_SHIFT; | |
545 | ||
546 | d = d * BLKIO_WEIGHT_DEFAULT; | |
547 | do_div(d, cfqg->weight); | |
548 | return d; | |
549 | } | |
550 | ||
551 | static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
552 | { | |
553 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
554 | if (delta > 0) | |
555 | min_vdisktime = vdisktime; | |
556 | ||
557 | return min_vdisktime; | |
558 | } | |
559 | ||
560 | static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
561 | { | |
562 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
563 | if (delta < 0) | |
564 | min_vdisktime = vdisktime; | |
565 | ||
566 | return min_vdisktime; | |
567 | } | |
568 | ||
569 | static void update_min_vdisktime(struct cfq_rb_root *st) | |
570 | { | |
25bc6b07 VG |
571 | struct cfq_group *cfqg; |
572 | ||
25bc6b07 VG |
573 | if (st->left) { |
574 | cfqg = rb_entry_cfqg(st->left); | |
a6032710 GJ |
575 | st->min_vdisktime = max_vdisktime(st->min_vdisktime, |
576 | cfqg->vdisktime); | |
25bc6b07 | 577 | } |
25bc6b07 VG |
578 | } |
579 | ||
5db5d642 CZ |
580 | /* |
581 | * get averaged number of queues of RT/BE priority. | |
582 | * average is updated, with a formula that gives more weight to higher numbers, | |
583 | * to quickly follows sudden increases and decrease slowly | |
584 | */ | |
585 | ||
58ff82f3 VG |
586 | static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, |
587 | struct cfq_group *cfqg, bool rt) | |
5869619c | 588 | { |
5db5d642 CZ |
589 | unsigned min_q, max_q; |
590 | unsigned mult = cfq_hist_divisor - 1; | |
591 | unsigned round = cfq_hist_divisor / 2; | |
58ff82f3 | 592 | unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); |
5db5d642 | 593 | |
58ff82f3 VG |
594 | min_q = min(cfqg->busy_queues_avg[rt], busy); |
595 | max_q = max(cfqg->busy_queues_avg[rt], busy); | |
596 | cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
5db5d642 | 597 | cfq_hist_divisor; |
58ff82f3 VG |
598 | return cfqg->busy_queues_avg[rt]; |
599 | } | |
600 | ||
601 | static inline unsigned | |
602 | cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
603 | { | |
604 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
605 | ||
606 | return cfq_target_latency * cfqg->weight / st->total_weight; | |
5db5d642 CZ |
607 | } |
608 | ||
c553f8e3 | 609 | static inline unsigned |
ba5bd520 | 610 | cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
44f7c160 | 611 | { |
5db5d642 CZ |
612 | unsigned slice = cfq_prio_to_slice(cfqd, cfqq); |
613 | if (cfqd->cfq_latency) { | |
58ff82f3 VG |
614 | /* |
615 | * interested queues (we consider only the ones with the same | |
616 | * priority class in the cfq group) | |
617 | */ | |
618 | unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, | |
619 | cfq_class_rt(cfqq)); | |
5db5d642 CZ |
620 | unsigned sync_slice = cfqd->cfq_slice[1]; |
621 | unsigned expect_latency = sync_slice * iq; | |
58ff82f3 VG |
622 | unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg); |
623 | ||
624 | if (expect_latency > group_slice) { | |
5db5d642 CZ |
625 | unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; |
626 | /* scale low_slice according to IO priority | |
627 | * and sync vs async */ | |
628 | unsigned low_slice = | |
629 | min(slice, base_low_slice * slice / sync_slice); | |
630 | /* the adapted slice value is scaled to fit all iqs | |
631 | * into the target latency */ | |
58ff82f3 | 632 | slice = max(slice * group_slice / expect_latency, |
5db5d642 CZ |
633 | low_slice); |
634 | } | |
635 | } | |
c553f8e3 SL |
636 | return slice; |
637 | } | |
638 | ||
639 | static inline void | |
640 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
641 | { | |
ba5bd520 | 642 | unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq); |
c553f8e3 | 643 | |
dae739eb | 644 | cfqq->slice_start = jiffies; |
5db5d642 | 645 | cfqq->slice_end = jiffies + slice; |
f75edf2d | 646 | cfqq->allocated_slice = slice; |
7b679138 | 647 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
648 | } |
649 | ||
650 | /* | |
651 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
652 | * isn't valid until the first request from the dispatch is activated | |
653 | * and the slice time set. | |
654 | */ | |
a6151c3a | 655 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) |
44f7c160 JA |
656 | { |
657 | if (cfq_cfqq_slice_new(cfqq)) | |
c1e44756 | 658 | return false; |
44f7c160 | 659 | if (time_before(jiffies, cfqq->slice_end)) |
c1e44756 | 660 | return false; |
44f7c160 | 661 | |
c1e44756 | 662 | return true; |
44f7c160 JA |
663 | } |
664 | ||
1da177e4 | 665 | /* |
5e705374 | 666 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 667 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 668 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 669 | */ |
5e705374 | 670 | static struct request * |
cf7c25cf | 671 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) |
1da177e4 | 672 | { |
cf7c25cf | 673 | sector_t s1, s2, d1 = 0, d2 = 0; |
1da177e4 | 674 | unsigned long back_max; |
e8a99053 AM |
675 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
676 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
677 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 678 | |
5e705374 JA |
679 | if (rq1 == NULL || rq1 == rq2) |
680 | return rq2; | |
681 | if (rq2 == NULL) | |
682 | return rq1; | |
9c2c38a1 | 683 | |
229836bd NK |
684 | if (rq_is_sync(rq1) != rq_is_sync(rq2)) |
685 | return rq_is_sync(rq1) ? rq1 : rq2; | |
686 | ||
65299a3b CH |
687 | if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO) |
688 | return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2; | |
b53d1ed7 | 689 | |
83096ebf TH |
690 | s1 = blk_rq_pos(rq1); |
691 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 692 | |
1da177e4 LT |
693 | /* |
694 | * by definition, 1KiB is 2 sectors | |
695 | */ | |
696 | back_max = cfqd->cfq_back_max * 2; | |
697 | ||
698 | /* | |
699 | * Strict one way elevator _except_ in the case where we allow | |
700 | * short backward seeks which are biased as twice the cost of a | |
701 | * similar forward seek. | |
702 | */ | |
703 | if (s1 >= last) | |
704 | d1 = s1 - last; | |
705 | else if (s1 + back_max >= last) | |
706 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
707 | else | |
e8a99053 | 708 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
709 | |
710 | if (s2 >= last) | |
711 | d2 = s2 - last; | |
712 | else if (s2 + back_max >= last) | |
713 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
714 | else | |
e8a99053 | 715 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
716 | |
717 | /* Found required data */ | |
e8a99053 AM |
718 | |
719 | /* | |
720 | * By doing switch() on the bit mask "wrap" we avoid having to | |
721 | * check two variables for all permutations: --> faster! | |
722 | */ | |
723 | switch (wrap) { | |
5e705374 | 724 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 725 | if (d1 < d2) |
5e705374 | 726 | return rq1; |
e8a99053 | 727 | else if (d2 < d1) |
5e705374 | 728 | return rq2; |
e8a99053 AM |
729 | else { |
730 | if (s1 >= s2) | |
5e705374 | 731 | return rq1; |
e8a99053 | 732 | else |
5e705374 | 733 | return rq2; |
e8a99053 | 734 | } |
1da177e4 | 735 | |
e8a99053 | 736 | case CFQ_RQ2_WRAP: |
5e705374 | 737 | return rq1; |
e8a99053 | 738 | case CFQ_RQ1_WRAP: |
5e705374 JA |
739 | return rq2; |
740 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
741 | default: |
742 | /* | |
743 | * Since both rqs are wrapped, | |
744 | * start with the one that's further behind head | |
745 | * (--> only *one* back seek required), | |
746 | * since back seek takes more time than forward. | |
747 | */ | |
748 | if (s1 <= s2) | |
5e705374 | 749 | return rq1; |
1da177e4 | 750 | else |
5e705374 | 751 | return rq2; |
1da177e4 LT |
752 | } |
753 | } | |
754 | ||
498d3aa2 JA |
755 | /* |
756 | * The below is leftmost cache rbtree addon | |
757 | */ | |
0871714e | 758 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 | 759 | { |
615f0259 VG |
760 | /* Service tree is empty */ |
761 | if (!root->count) | |
762 | return NULL; | |
763 | ||
cc09e299 JA |
764 | if (!root->left) |
765 | root->left = rb_first(&root->rb); | |
766 | ||
0871714e JA |
767 | if (root->left) |
768 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
769 | ||
770 | return NULL; | |
cc09e299 JA |
771 | } |
772 | ||
1fa8f6d6 VG |
773 | static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) |
774 | { | |
775 | if (!root->left) | |
776 | root->left = rb_first(&root->rb); | |
777 | ||
778 | if (root->left) | |
779 | return rb_entry_cfqg(root->left); | |
780 | ||
781 | return NULL; | |
782 | } | |
783 | ||
a36e71f9 JA |
784 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
785 | { | |
786 | rb_erase(n, root); | |
787 | RB_CLEAR_NODE(n); | |
788 | } | |
789 | ||
cc09e299 JA |
790 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
791 | { | |
792 | if (root->left == n) | |
793 | root->left = NULL; | |
a36e71f9 | 794 | rb_erase_init(n, &root->rb); |
aa6f6a3d | 795 | --root->count; |
cc09e299 JA |
796 | } |
797 | ||
1da177e4 LT |
798 | /* |
799 | * would be nice to take fifo expire time into account as well | |
800 | */ | |
5e705374 JA |
801 | static struct request * |
802 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
803 | struct request *last) | |
1da177e4 | 804 | { |
21183b07 JA |
805 | struct rb_node *rbnext = rb_next(&last->rb_node); |
806 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 807 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 808 | |
21183b07 | 809 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
810 | |
811 | if (rbprev) | |
5e705374 | 812 | prev = rb_entry_rq(rbprev); |
1da177e4 | 813 | |
21183b07 | 814 | if (rbnext) |
5e705374 | 815 | next = rb_entry_rq(rbnext); |
21183b07 JA |
816 | else { |
817 | rbnext = rb_first(&cfqq->sort_list); | |
818 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 819 | next = rb_entry_rq(rbnext); |
21183b07 | 820 | } |
1da177e4 | 821 | |
cf7c25cf | 822 | return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); |
1da177e4 LT |
823 | } |
824 | ||
d9e7620e JA |
825 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
826 | struct cfq_queue *cfqq) | |
1da177e4 | 827 | { |
d9e7620e JA |
828 | /* |
829 | * just an approximation, should be ok. | |
830 | */ | |
cdb16e8f | 831 | return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
464191c6 | 832 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); |
d9e7620e JA |
833 | } |
834 | ||
1fa8f6d6 VG |
835 | static inline s64 |
836 | cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
837 | { | |
838 | return cfqg->vdisktime - st->min_vdisktime; | |
839 | } | |
840 | ||
841 | static void | |
842 | __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
843 | { | |
844 | struct rb_node **node = &st->rb.rb_node; | |
845 | struct rb_node *parent = NULL; | |
846 | struct cfq_group *__cfqg; | |
847 | s64 key = cfqg_key(st, cfqg); | |
848 | int left = 1; | |
849 | ||
850 | while (*node != NULL) { | |
851 | parent = *node; | |
852 | __cfqg = rb_entry_cfqg(parent); | |
853 | ||
854 | if (key < cfqg_key(st, __cfqg)) | |
855 | node = &parent->rb_left; | |
856 | else { | |
857 | node = &parent->rb_right; | |
858 | left = 0; | |
859 | } | |
860 | } | |
861 | ||
862 | if (left) | |
863 | st->left = &cfqg->rb_node; | |
864 | ||
865 | rb_link_node(&cfqg->rb_node, parent, node); | |
866 | rb_insert_color(&cfqg->rb_node, &st->rb); | |
867 | } | |
868 | ||
869 | static void | |
8184f93e JT |
870 | cfq_update_group_weight(struct cfq_group *cfqg) |
871 | { | |
872 | BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); | |
873 | if (cfqg->needs_update) { | |
874 | cfqg->weight = cfqg->new_weight; | |
875 | cfqg->needs_update = false; | |
876 | } | |
877 | } | |
878 | ||
879 | static void | |
880 | cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
881 | { | |
882 | BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); | |
883 | ||
884 | cfq_update_group_weight(cfqg); | |
885 | __cfq_group_service_tree_add(st, cfqg); | |
886 | st->total_weight += cfqg->weight; | |
887 | } | |
888 | ||
889 | static void | |
890 | cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1fa8f6d6 VG |
891 | { |
892 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
893 | struct cfq_group *__cfqg; | |
894 | struct rb_node *n; | |
895 | ||
896 | cfqg->nr_cfqq++; | |
760701bf | 897 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) |
1fa8f6d6 VG |
898 | return; |
899 | ||
900 | /* | |
901 | * Currently put the group at the end. Later implement something | |
902 | * so that groups get lesser vtime based on their weights, so that | |
25985edc | 903 | * if group does not loose all if it was not continuously backlogged. |
1fa8f6d6 VG |
904 | */ |
905 | n = rb_last(&st->rb); | |
906 | if (n) { | |
907 | __cfqg = rb_entry_cfqg(n); | |
908 | cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; | |
909 | } else | |
910 | cfqg->vdisktime = st->min_vdisktime; | |
8184f93e JT |
911 | cfq_group_service_tree_add(st, cfqg); |
912 | } | |
1fa8f6d6 | 913 | |
8184f93e JT |
914 | static void |
915 | cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
916 | { | |
917 | st->total_weight -= cfqg->weight; | |
918 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) | |
919 | cfq_rb_erase(&cfqg->rb_node, st); | |
1fa8f6d6 VG |
920 | } |
921 | ||
922 | static void | |
8184f93e | 923 | cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg) |
1fa8f6d6 VG |
924 | { |
925 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
926 | ||
927 | BUG_ON(cfqg->nr_cfqq < 1); | |
928 | cfqg->nr_cfqq--; | |
25bc6b07 | 929 | |
1fa8f6d6 VG |
930 | /* If there are other cfq queues under this group, don't delete it */ |
931 | if (cfqg->nr_cfqq) | |
932 | return; | |
933 | ||
2868ef7b | 934 | cfq_log_cfqg(cfqd, cfqg, "del_from_rr group"); |
8184f93e | 935 | cfq_group_service_tree_del(st, cfqg); |
dae739eb | 936 | cfqg->saved_workload_slice = 0; |
e98ef89b | 937 | cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1); |
dae739eb VG |
938 | } |
939 | ||
167400d3 JT |
940 | static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq, |
941 | unsigned int *unaccounted_time) | |
dae739eb | 942 | { |
f75edf2d | 943 | unsigned int slice_used; |
dae739eb VG |
944 | |
945 | /* | |
946 | * Queue got expired before even a single request completed or | |
947 | * got expired immediately after first request completion. | |
948 | */ | |
949 | if (!cfqq->slice_start || cfqq->slice_start == jiffies) { | |
950 | /* | |
951 | * Also charge the seek time incurred to the group, otherwise | |
952 | * if there are mutiple queues in the group, each can dispatch | |
953 | * a single request on seeky media and cause lots of seek time | |
954 | * and group will never know it. | |
955 | */ | |
956 | slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start), | |
957 | 1); | |
958 | } else { | |
959 | slice_used = jiffies - cfqq->slice_start; | |
167400d3 JT |
960 | if (slice_used > cfqq->allocated_slice) { |
961 | *unaccounted_time = slice_used - cfqq->allocated_slice; | |
f75edf2d | 962 | slice_used = cfqq->allocated_slice; |
167400d3 JT |
963 | } |
964 | if (time_after(cfqq->slice_start, cfqq->dispatch_start)) | |
965 | *unaccounted_time += cfqq->slice_start - | |
966 | cfqq->dispatch_start; | |
dae739eb VG |
967 | } |
968 | ||
dae739eb VG |
969 | return slice_used; |
970 | } | |
971 | ||
972 | static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg, | |
e5ff082e | 973 | struct cfq_queue *cfqq) |
dae739eb VG |
974 | { |
975 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
167400d3 | 976 | unsigned int used_sl, charge, unaccounted_sl = 0; |
f26bd1f0 VG |
977 | int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg) |
978 | - cfqg->service_tree_idle.count; | |
979 | ||
980 | BUG_ON(nr_sync < 0); | |
167400d3 | 981 | used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl); |
dae739eb | 982 | |
02b35081 VG |
983 | if (iops_mode(cfqd)) |
984 | charge = cfqq->slice_dispatch; | |
985 | else if (!cfq_cfqq_sync(cfqq) && !nr_sync) | |
986 | charge = cfqq->allocated_slice; | |
dae739eb VG |
987 | |
988 | /* Can't update vdisktime while group is on service tree */ | |
8184f93e | 989 | cfq_group_service_tree_del(st, cfqg); |
02b35081 | 990 | cfqg->vdisktime += cfq_scale_slice(charge, cfqg); |
8184f93e JT |
991 | /* If a new weight was requested, update now, off tree */ |
992 | cfq_group_service_tree_add(st, cfqg); | |
dae739eb VG |
993 | |
994 | /* This group is being expired. Save the context */ | |
995 | if (time_after(cfqd->workload_expires, jiffies)) { | |
996 | cfqg->saved_workload_slice = cfqd->workload_expires | |
997 | - jiffies; | |
998 | cfqg->saved_workload = cfqd->serving_type; | |
999 | cfqg->saved_serving_prio = cfqd->serving_prio; | |
1000 | } else | |
1001 | cfqg->saved_workload_slice = 0; | |
2868ef7b VG |
1002 | |
1003 | cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime, | |
1004 | st->min_vdisktime); | |
fd16d263 JP |
1005 | cfq_log_cfqq(cfqq->cfqd, cfqq, |
1006 | "sl_used=%u disp=%u charge=%u iops=%u sect=%lu", | |
1007 | used_sl, cfqq->slice_dispatch, charge, | |
1008 | iops_mode(cfqd), cfqq->nr_sectors); | |
167400d3 JT |
1009 | cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl, |
1010 | unaccounted_sl); | |
e98ef89b | 1011 | cfq_blkiocg_set_start_empty_time(&cfqg->blkg); |
1fa8f6d6 VG |
1012 | } |
1013 | ||
25fb5169 VG |
1014 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
1015 | static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg) | |
1016 | { | |
1017 | if (blkg) | |
1018 | return container_of(blkg, struct cfq_group, blkg); | |
1019 | return NULL; | |
1020 | } | |
1021 | ||
8aea4545 PB |
1022 | static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg, |
1023 | unsigned int weight) | |
f8d461d6 | 1024 | { |
8184f93e JT |
1025 | struct cfq_group *cfqg = cfqg_of_blkg(blkg); |
1026 | cfqg->new_weight = weight; | |
1027 | cfqg->needs_update = true; | |
f8d461d6 VG |
1028 | } |
1029 | ||
f469a7b4 VG |
1030 | static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd, |
1031 | struct cfq_group *cfqg, struct blkio_cgroup *blkcg) | |
25fb5169 | 1032 | { |
22084190 VG |
1033 | struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; |
1034 | unsigned int major, minor; | |
25fb5169 | 1035 | |
f469a7b4 VG |
1036 | /* |
1037 | * Add group onto cgroup list. It might happen that bdi->dev is | |
1038 | * not initialized yet. Initialize this new group without major | |
1039 | * and minor info and this info will be filled in once a new thread | |
1040 | * comes for IO. | |
1041 | */ | |
1042 | if (bdi->dev) { | |
a74b2ada | 1043 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); |
f469a7b4 VG |
1044 | cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, |
1045 | (void *)cfqd, MKDEV(major, minor)); | |
1046 | } else | |
1047 | cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, | |
1048 | (void *)cfqd, 0); | |
1049 | ||
1050 | cfqd->nr_blkcg_linked_grps++; | |
1051 | cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev); | |
1052 | ||
1053 | /* Add group on cfqd list */ | |
1054 | hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); | |
1055 | } | |
1056 | ||
1057 | /* | |
1058 | * Should be called from sleepable context. No request queue lock as per | |
1059 | * cpu stats are allocated dynamically and alloc_percpu needs to be called | |
1060 | * from sleepable context. | |
1061 | */ | |
1062 | static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd) | |
1063 | { | |
1064 | struct cfq_group *cfqg = NULL; | |
5624a4e4 | 1065 | int i, j, ret; |
f469a7b4 | 1066 | struct cfq_rb_root *st; |
25fb5169 VG |
1067 | |
1068 | cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node); | |
1069 | if (!cfqg) | |
f469a7b4 | 1070 | return NULL; |
25fb5169 | 1071 | |
25fb5169 VG |
1072 | for_each_cfqg_st(cfqg, i, j, st) |
1073 | *st = CFQ_RB_ROOT; | |
1074 | RB_CLEAR_NODE(&cfqg->rb_node); | |
1075 | ||
7700fc4f SL |
1076 | cfqg->ttime.last_end_request = jiffies; |
1077 | ||
b1c35769 VG |
1078 | /* |
1079 | * Take the initial reference that will be released on destroy | |
1080 | * This can be thought of a joint reference by cgroup and | |
1081 | * elevator which will be dropped by either elevator exit | |
1082 | * or cgroup deletion path depending on who is exiting first. | |
1083 | */ | |
329a6781 | 1084 | cfqg->ref = 1; |
5624a4e4 VG |
1085 | |
1086 | ret = blkio_alloc_blkg_stats(&cfqg->blkg); | |
1087 | if (ret) { | |
1088 | kfree(cfqg); | |
1089 | return NULL; | |
1090 | } | |
1091 | ||
f469a7b4 VG |
1092 | return cfqg; |
1093 | } | |
1094 | ||
1095 | static struct cfq_group * | |
1096 | cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg) | |
1097 | { | |
1098 | struct cfq_group *cfqg = NULL; | |
1099 | void *key = cfqd; | |
1100 | struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; | |
1101 | unsigned int major, minor; | |
b1c35769 | 1102 | |
180be2a0 | 1103 | /* |
f469a7b4 VG |
1104 | * This is the common case when there are no blkio cgroups. |
1105 | * Avoid lookup in this case | |
180be2a0 | 1106 | */ |
f469a7b4 VG |
1107 | if (blkcg == &blkio_root_cgroup) |
1108 | cfqg = &cfqd->root_group; | |
1109 | else | |
1110 | cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key)); | |
25fb5169 | 1111 | |
f469a7b4 VG |
1112 | if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) { |
1113 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); | |
1114 | cfqg->blkg.dev = MKDEV(major, minor); | |
1115 | } | |
25fb5169 | 1116 | |
25fb5169 VG |
1117 | return cfqg; |
1118 | } | |
1119 | ||
1120 | /* | |
3e59cf9d VG |
1121 | * Search for the cfq group current task belongs to. request_queue lock must |
1122 | * be held. | |
25fb5169 | 1123 | */ |
3e59cf9d | 1124 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd) |
25fb5169 | 1125 | { |
70087dc3 | 1126 | struct blkio_cgroup *blkcg; |
f469a7b4 VG |
1127 | struct cfq_group *cfqg = NULL, *__cfqg = NULL; |
1128 | struct request_queue *q = cfqd->queue; | |
25fb5169 VG |
1129 | |
1130 | rcu_read_lock(); | |
70087dc3 | 1131 | blkcg = task_blkio_cgroup(current); |
f469a7b4 VG |
1132 | cfqg = cfq_find_cfqg(cfqd, blkcg); |
1133 | if (cfqg) { | |
1134 | rcu_read_unlock(); | |
1135 | return cfqg; | |
1136 | } | |
1137 | ||
1138 | /* | |
1139 | * Need to allocate a group. Allocation of group also needs allocation | |
1140 | * of per cpu stats which in-turn takes a mutex() and can block. Hence | |
1141 | * we need to drop rcu lock and queue_lock before we call alloc. | |
1142 | * | |
1143 | * Not taking any queue reference here and assuming that queue is | |
1144 | * around by the time we return. CFQ queue allocation code does | |
1145 | * the same. It might be racy though. | |
1146 | */ | |
1147 | ||
1148 | rcu_read_unlock(); | |
1149 | spin_unlock_irq(q->queue_lock); | |
1150 | ||
1151 | cfqg = cfq_alloc_cfqg(cfqd); | |
1152 | ||
1153 | spin_lock_irq(q->queue_lock); | |
1154 | ||
1155 | rcu_read_lock(); | |
1156 | blkcg = task_blkio_cgroup(current); | |
1157 | ||
1158 | /* | |
1159 | * If some other thread already allocated the group while we were | |
1160 | * not holding queue lock, free up the group | |
1161 | */ | |
1162 | __cfqg = cfq_find_cfqg(cfqd, blkcg); | |
1163 | ||
1164 | if (__cfqg) { | |
1165 | kfree(cfqg); | |
1166 | rcu_read_unlock(); | |
1167 | return __cfqg; | |
1168 | } | |
1169 | ||
3e59cf9d | 1170 | if (!cfqg) |
25fb5169 | 1171 | cfqg = &cfqd->root_group; |
f469a7b4 VG |
1172 | |
1173 | cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg); | |
25fb5169 VG |
1174 | rcu_read_unlock(); |
1175 | return cfqg; | |
1176 | } | |
1177 | ||
7f1dc8a2 VG |
1178 | static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg) |
1179 | { | |
329a6781 | 1180 | cfqg->ref++; |
7f1dc8a2 VG |
1181 | return cfqg; |
1182 | } | |
1183 | ||
25fb5169 VG |
1184 | static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) |
1185 | { | |
1186 | /* Currently, all async queues are mapped to root group */ | |
1187 | if (!cfq_cfqq_sync(cfqq)) | |
1188 | cfqg = &cfqq->cfqd->root_group; | |
1189 | ||
1190 | cfqq->cfqg = cfqg; | |
b1c35769 | 1191 | /* cfqq reference on cfqg */ |
329a6781 | 1192 | cfqq->cfqg->ref++; |
b1c35769 VG |
1193 | } |
1194 | ||
1195 | static void cfq_put_cfqg(struct cfq_group *cfqg) | |
1196 | { | |
1197 | struct cfq_rb_root *st; | |
1198 | int i, j; | |
1199 | ||
329a6781 SL |
1200 | BUG_ON(cfqg->ref <= 0); |
1201 | cfqg->ref--; | |
1202 | if (cfqg->ref) | |
b1c35769 VG |
1203 | return; |
1204 | for_each_cfqg_st(cfqg, i, j, st) | |
b54ce60e | 1205 | BUG_ON(!RB_EMPTY_ROOT(&st->rb)); |
5624a4e4 | 1206 | free_percpu(cfqg->blkg.stats_cpu); |
b1c35769 VG |
1207 | kfree(cfqg); |
1208 | } | |
1209 | ||
1210 | static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1211 | { | |
1212 | /* Something wrong if we are trying to remove same group twice */ | |
1213 | BUG_ON(hlist_unhashed(&cfqg->cfqd_node)); | |
1214 | ||
1215 | hlist_del_init(&cfqg->cfqd_node); | |
1216 | ||
a5395b83 VG |
1217 | BUG_ON(cfqd->nr_blkcg_linked_grps <= 0); |
1218 | cfqd->nr_blkcg_linked_grps--; | |
1219 | ||
b1c35769 VG |
1220 | /* |
1221 | * Put the reference taken at the time of creation so that when all | |
1222 | * queues are gone, group can be destroyed. | |
1223 | */ | |
1224 | cfq_put_cfqg(cfqg); | |
1225 | } | |
1226 | ||
1227 | static void cfq_release_cfq_groups(struct cfq_data *cfqd) | |
1228 | { | |
1229 | struct hlist_node *pos, *n; | |
1230 | struct cfq_group *cfqg; | |
1231 | ||
1232 | hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) { | |
1233 | /* | |
1234 | * If cgroup removal path got to blk_group first and removed | |
1235 | * it from cgroup list, then it will take care of destroying | |
1236 | * cfqg also. | |
1237 | */ | |
e98ef89b | 1238 | if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg)) |
b1c35769 VG |
1239 | cfq_destroy_cfqg(cfqd, cfqg); |
1240 | } | |
25fb5169 | 1241 | } |
b1c35769 VG |
1242 | |
1243 | /* | |
1244 | * Blk cgroup controller notification saying that blkio_group object is being | |
1245 | * delinked as associated cgroup object is going away. That also means that | |
1246 | * no new IO will come in this group. So get rid of this group as soon as | |
1247 | * any pending IO in the group is finished. | |
1248 | * | |
1249 | * This function is called under rcu_read_lock(). key is the rcu protected | |
1250 | * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu | |
1251 | * read lock. | |
1252 | * | |
1253 | * "key" was fetched from blkio_group under blkio_cgroup->lock. That means | |
1254 | * it should not be NULL as even if elevator was exiting, cgroup deltion | |
1255 | * path got to it first. | |
1256 | */ | |
8aea4545 | 1257 | static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg) |
b1c35769 VG |
1258 | { |
1259 | unsigned long flags; | |
1260 | struct cfq_data *cfqd = key; | |
1261 | ||
1262 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
1263 | cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg)); | |
1264 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
1265 | } | |
1266 | ||
25fb5169 | 1267 | #else /* GROUP_IOSCHED */ |
3e59cf9d | 1268 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd) |
25fb5169 VG |
1269 | { |
1270 | return &cfqd->root_group; | |
1271 | } | |
7f1dc8a2 VG |
1272 | |
1273 | static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg) | |
1274 | { | |
50eaeb32 | 1275 | return cfqg; |
7f1dc8a2 VG |
1276 | } |
1277 | ||
25fb5169 VG |
1278 | static inline void |
1279 | cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) { | |
1280 | cfqq->cfqg = cfqg; | |
1281 | } | |
1282 | ||
b1c35769 VG |
1283 | static void cfq_release_cfq_groups(struct cfq_data *cfqd) {} |
1284 | static inline void cfq_put_cfqg(struct cfq_group *cfqg) {} | |
1285 | ||
25fb5169 VG |
1286 | #endif /* GROUP_IOSCHED */ |
1287 | ||
498d3aa2 | 1288 | /* |
c0324a02 | 1289 | * The cfqd->service_trees holds all pending cfq_queue's that have |
498d3aa2 JA |
1290 | * requests waiting to be processed. It is sorted in the order that |
1291 | * we will service the queues. | |
1292 | */ | |
a36e71f9 | 1293 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 1294 | bool add_front) |
d9e7620e | 1295 | { |
0871714e JA |
1296 | struct rb_node **p, *parent; |
1297 | struct cfq_queue *__cfqq; | |
d9e7620e | 1298 | unsigned long rb_key; |
c0324a02 | 1299 | struct cfq_rb_root *service_tree; |
498d3aa2 | 1300 | int left; |
dae739eb | 1301 | int new_cfqq = 1; |
ae30c286 | 1302 | |
cdb16e8f | 1303 | service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), |
65b32a57 | 1304 | cfqq_type(cfqq)); |
0871714e JA |
1305 | if (cfq_class_idle(cfqq)) { |
1306 | rb_key = CFQ_IDLE_DELAY; | |
aa6f6a3d | 1307 | parent = rb_last(&service_tree->rb); |
0871714e JA |
1308 | if (parent && parent != &cfqq->rb_node) { |
1309 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1310 | rb_key += __cfqq->rb_key; | |
1311 | } else | |
1312 | rb_key += jiffies; | |
1313 | } else if (!add_front) { | |
b9c8946b JA |
1314 | /* |
1315 | * Get our rb key offset. Subtract any residual slice | |
1316 | * value carried from last service. A negative resid | |
1317 | * count indicates slice overrun, and this should position | |
1318 | * the next service time further away in the tree. | |
1319 | */ | |
edd75ffd | 1320 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
b9c8946b | 1321 | rb_key -= cfqq->slice_resid; |
edd75ffd | 1322 | cfqq->slice_resid = 0; |
48e025e6 CZ |
1323 | } else { |
1324 | rb_key = -HZ; | |
aa6f6a3d | 1325 | __cfqq = cfq_rb_first(service_tree); |
48e025e6 CZ |
1326 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; |
1327 | } | |
1da177e4 | 1328 | |
d9e7620e | 1329 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
dae739eb | 1330 | new_cfqq = 0; |
99f9628a | 1331 | /* |
d9e7620e | 1332 | * same position, nothing more to do |
99f9628a | 1333 | */ |
c0324a02 CZ |
1334 | if (rb_key == cfqq->rb_key && |
1335 | cfqq->service_tree == service_tree) | |
d9e7620e | 1336 | return; |
1da177e4 | 1337 | |
aa6f6a3d CZ |
1338 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); |
1339 | cfqq->service_tree = NULL; | |
1da177e4 | 1340 | } |
d9e7620e | 1341 | |
498d3aa2 | 1342 | left = 1; |
0871714e | 1343 | parent = NULL; |
aa6f6a3d CZ |
1344 | cfqq->service_tree = service_tree; |
1345 | p = &service_tree->rb.rb_node; | |
d9e7620e | 1346 | while (*p) { |
67060e37 | 1347 | struct rb_node **n; |
cc09e299 | 1348 | |
d9e7620e JA |
1349 | parent = *p; |
1350 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1351 | ||
0c534e0a | 1352 | /* |
c0324a02 | 1353 | * sort by key, that represents service time. |
0c534e0a | 1354 | */ |
c0324a02 | 1355 | if (time_before(rb_key, __cfqq->rb_key)) |
67060e37 | 1356 | n = &(*p)->rb_left; |
c0324a02 | 1357 | else { |
67060e37 | 1358 | n = &(*p)->rb_right; |
cc09e299 | 1359 | left = 0; |
c0324a02 | 1360 | } |
67060e37 JA |
1361 | |
1362 | p = n; | |
d9e7620e JA |
1363 | } |
1364 | ||
cc09e299 | 1365 | if (left) |
aa6f6a3d | 1366 | service_tree->left = &cfqq->rb_node; |
cc09e299 | 1367 | |
d9e7620e JA |
1368 | cfqq->rb_key = rb_key; |
1369 | rb_link_node(&cfqq->rb_node, parent, p); | |
aa6f6a3d CZ |
1370 | rb_insert_color(&cfqq->rb_node, &service_tree->rb); |
1371 | service_tree->count++; | |
20359f27 | 1372 | if (add_front || !new_cfqq) |
dae739eb | 1373 | return; |
8184f93e | 1374 | cfq_group_notify_queue_add(cfqd, cfqq->cfqg); |
1da177e4 LT |
1375 | } |
1376 | ||
a36e71f9 | 1377 | static struct cfq_queue * |
f2d1f0ae JA |
1378 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
1379 | sector_t sector, struct rb_node **ret_parent, | |
1380 | struct rb_node ***rb_link) | |
a36e71f9 | 1381 | { |
a36e71f9 JA |
1382 | struct rb_node **p, *parent; |
1383 | struct cfq_queue *cfqq = NULL; | |
1384 | ||
1385 | parent = NULL; | |
1386 | p = &root->rb_node; | |
1387 | while (*p) { | |
1388 | struct rb_node **n; | |
1389 | ||
1390 | parent = *p; | |
1391 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
1392 | ||
1393 | /* | |
1394 | * Sort strictly based on sector. Smallest to the left, | |
1395 | * largest to the right. | |
1396 | */ | |
2e46e8b2 | 1397 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 1398 | n = &(*p)->rb_right; |
2e46e8b2 | 1399 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
1400 | n = &(*p)->rb_left; |
1401 | else | |
1402 | break; | |
1403 | p = n; | |
3ac6c9f8 | 1404 | cfqq = NULL; |
a36e71f9 JA |
1405 | } |
1406 | ||
1407 | *ret_parent = parent; | |
1408 | if (rb_link) | |
1409 | *rb_link = p; | |
3ac6c9f8 | 1410 | return cfqq; |
a36e71f9 JA |
1411 | } |
1412 | ||
1413 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1414 | { | |
a36e71f9 JA |
1415 | struct rb_node **p, *parent; |
1416 | struct cfq_queue *__cfqq; | |
1417 | ||
f2d1f0ae JA |
1418 | if (cfqq->p_root) { |
1419 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1420 | cfqq->p_root = NULL; | |
1421 | } | |
a36e71f9 JA |
1422 | |
1423 | if (cfq_class_idle(cfqq)) | |
1424 | return; | |
1425 | if (!cfqq->next_rq) | |
1426 | return; | |
1427 | ||
f2d1f0ae | 1428 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
1429 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
1430 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
1431 | if (!__cfqq) { |
1432 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
1433 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
1434 | } else | |
1435 | cfqq->p_root = NULL; | |
a36e71f9 JA |
1436 | } |
1437 | ||
498d3aa2 JA |
1438 | /* |
1439 | * Update cfqq's position in the service tree. | |
1440 | */ | |
edd75ffd | 1441 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 1442 | { |
6d048f53 JA |
1443 | /* |
1444 | * Resorting requires the cfqq to be on the RR list already. | |
1445 | */ | |
a36e71f9 | 1446 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 1447 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
1448 | cfq_prio_tree_add(cfqd, cfqq); |
1449 | } | |
6d048f53 JA |
1450 | } |
1451 | ||
1da177e4 LT |
1452 | /* |
1453 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 1454 | * the pending list according to last request service |
1da177e4 | 1455 | */ |
febffd61 | 1456 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1457 | { |
7b679138 | 1458 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
1459 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1460 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 | 1461 | cfqd->busy_queues++; |
ef8a41df SL |
1462 | if (cfq_cfqq_sync(cfqq)) |
1463 | cfqd->busy_sync_queues++; | |
1da177e4 | 1464 | |
edd75ffd | 1465 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
1466 | } |
1467 | ||
498d3aa2 JA |
1468 | /* |
1469 | * Called when the cfqq no longer has requests pending, remove it from | |
1470 | * the service tree. | |
1471 | */ | |
febffd61 | 1472 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1473 | { |
7b679138 | 1474 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
1475 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
1476 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 1477 | |
aa6f6a3d CZ |
1478 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
1479 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
1480 | cfqq->service_tree = NULL; | |
1481 | } | |
f2d1f0ae JA |
1482 | if (cfqq->p_root) { |
1483 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1484 | cfqq->p_root = NULL; | |
1485 | } | |
d9e7620e | 1486 | |
8184f93e | 1487 | cfq_group_notify_queue_del(cfqd, cfqq->cfqg); |
1da177e4 LT |
1488 | BUG_ON(!cfqd->busy_queues); |
1489 | cfqd->busy_queues--; | |
ef8a41df SL |
1490 | if (cfq_cfqq_sync(cfqq)) |
1491 | cfqd->busy_sync_queues--; | |
1da177e4 LT |
1492 | } |
1493 | ||
1494 | /* | |
1495 | * rb tree support functions | |
1496 | */ | |
febffd61 | 1497 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 1498 | { |
5e705374 | 1499 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
5e705374 | 1500 | const int sync = rq_is_sync(rq); |
1da177e4 | 1501 | |
b4878f24 JA |
1502 | BUG_ON(!cfqq->queued[sync]); |
1503 | cfqq->queued[sync]--; | |
1da177e4 | 1504 | |
5e705374 | 1505 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 1506 | |
f04a6424 VG |
1507 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { |
1508 | /* | |
1509 | * Queue will be deleted from service tree when we actually | |
1510 | * expire it later. Right now just remove it from prio tree | |
1511 | * as it is empty. | |
1512 | */ | |
1513 | if (cfqq->p_root) { | |
1514 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1515 | cfqq->p_root = NULL; | |
1516 | } | |
1517 | } | |
1da177e4 LT |
1518 | } |
1519 | ||
5e705374 | 1520 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 1521 | { |
5e705374 | 1522 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 1523 | struct cfq_data *cfqd = cfqq->cfqd; |
796d5116 | 1524 | struct request *prev; |
1da177e4 | 1525 | |
5380a101 | 1526 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 | 1527 | |
796d5116 | 1528 | elv_rb_add(&cfqq->sort_list, rq); |
5fccbf61 JA |
1529 | |
1530 | if (!cfq_cfqq_on_rr(cfqq)) | |
1531 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
1532 | |
1533 | /* | |
1534 | * check if this request is a better next-serve candidate | |
1535 | */ | |
a36e71f9 | 1536 | prev = cfqq->next_rq; |
cf7c25cf | 1537 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); |
a36e71f9 JA |
1538 | |
1539 | /* | |
1540 | * adjust priority tree position, if ->next_rq changes | |
1541 | */ | |
1542 | if (prev != cfqq->next_rq) | |
1543 | cfq_prio_tree_add(cfqd, cfqq); | |
1544 | ||
5044eed4 | 1545 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
1546 | } |
1547 | ||
febffd61 | 1548 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 1549 | { |
5380a101 JA |
1550 | elv_rb_del(&cfqq->sort_list, rq); |
1551 | cfqq->queued[rq_is_sync(rq)]--; | |
e98ef89b VG |
1552 | cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, |
1553 | rq_data_dir(rq), rq_is_sync(rq)); | |
5e705374 | 1554 | cfq_add_rq_rb(rq); |
e98ef89b | 1555 | cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg, |
7f1dc8a2 VG |
1556 | &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq), |
1557 | rq_is_sync(rq)); | |
1da177e4 LT |
1558 | } |
1559 | ||
206dc69b JA |
1560 | static struct request * |
1561 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 1562 | { |
206dc69b | 1563 | struct task_struct *tsk = current; |
91fac317 | 1564 | struct cfq_io_context *cic; |
206dc69b | 1565 | struct cfq_queue *cfqq; |
1da177e4 | 1566 | |
4ac845a2 | 1567 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
1568 | if (!cic) |
1569 | return NULL; | |
1570 | ||
1571 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
1572 | if (cfqq) { |
1573 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
1574 | ||
21183b07 | 1575 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 1576 | } |
1da177e4 | 1577 | |
1da177e4 LT |
1578 | return NULL; |
1579 | } | |
1580 | ||
165125e1 | 1581 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1582 | { |
22e2c507 | 1583 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 1584 | |
53c583d2 | 1585 | cfqd->rq_in_driver++; |
7b679138 | 1586 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
53c583d2 | 1587 | cfqd->rq_in_driver); |
25776e35 | 1588 | |
5b93629b | 1589 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
1590 | } |
1591 | ||
165125e1 | 1592 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1593 | { |
b4878f24 JA |
1594 | struct cfq_data *cfqd = q->elevator->elevator_data; |
1595 | ||
53c583d2 CZ |
1596 | WARN_ON(!cfqd->rq_in_driver); |
1597 | cfqd->rq_in_driver--; | |
7b679138 | 1598 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
53c583d2 | 1599 | cfqd->rq_in_driver); |
1da177e4 LT |
1600 | } |
1601 | ||
b4878f24 | 1602 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 1603 | { |
5e705374 | 1604 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 1605 | |
5e705374 JA |
1606 | if (cfqq->next_rq == rq) |
1607 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 1608 | |
b4878f24 | 1609 | list_del_init(&rq->queuelist); |
5e705374 | 1610 | cfq_del_rq_rb(rq); |
374f84ac | 1611 | |
45333d5a | 1612 | cfqq->cfqd->rq_queued--; |
e98ef89b VG |
1613 | cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, |
1614 | rq_data_dir(rq), rq_is_sync(rq)); | |
65299a3b CH |
1615 | if (rq->cmd_flags & REQ_PRIO) { |
1616 | WARN_ON(!cfqq->prio_pending); | |
1617 | cfqq->prio_pending--; | |
b53d1ed7 | 1618 | } |
1da177e4 LT |
1619 | } |
1620 | ||
165125e1 JA |
1621 | static int cfq_merge(struct request_queue *q, struct request **req, |
1622 | struct bio *bio) | |
1da177e4 LT |
1623 | { |
1624 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1625 | struct request *__rq; | |
1da177e4 | 1626 | |
206dc69b | 1627 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 1628 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
1629 | *req = __rq; |
1630 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1631 | } |
1632 | ||
1633 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1634 | } |
1635 | ||
165125e1 | 1636 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 1637 | int type) |
1da177e4 | 1638 | { |
21183b07 | 1639 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 1640 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 1641 | |
5e705374 | 1642 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 1643 | } |
1da177e4 LT |
1644 | } |
1645 | ||
812d4026 DS |
1646 | static void cfq_bio_merged(struct request_queue *q, struct request *req, |
1647 | struct bio *bio) | |
1648 | { | |
e98ef89b VG |
1649 | cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg, |
1650 | bio_data_dir(bio), cfq_bio_sync(bio)); | |
812d4026 DS |
1651 | } |
1652 | ||
1da177e4 | 1653 | static void |
165125e1 | 1654 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
1655 | struct request *next) |
1656 | { | |
cf7c25cf | 1657 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 JA |
1658 | /* |
1659 | * reposition in fifo if next is older than rq | |
1660 | */ | |
1661 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
30996f40 | 1662 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
22e2c507 | 1663 | list_move(&rq->queuelist, &next->queuelist); |
30996f40 JA |
1664 | rq_set_fifo_time(rq, rq_fifo_time(next)); |
1665 | } | |
22e2c507 | 1666 | |
cf7c25cf CZ |
1667 | if (cfqq->next_rq == next) |
1668 | cfqq->next_rq = rq; | |
b4878f24 | 1669 | cfq_remove_request(next); |
e98ef89b VG |
1670 | cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg, |
1671 | rq_data_dir(next), rq_is_sync(next)); | |
22e2c507 JA |
1672 | } |
1673 | ||
165125e1 | 1674 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
1675 | struct bio *bio) |
1676 | { | |
1677 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 1678 | struct cfq_io_context *cic; |
da775265 | 1679 | struct cfq_queue *cfqq; |
da775265 JA |
1680 | |
1681 | /* | |
ec8acb69 | 1682 | * Disallow merge of a sync bio into an async request. |
da775265 | 1683 | */ |
91fac317 | 1684 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
a6151c3a | 1685 | return false; |
da775265 JA |
1686 | |
1687 | /* | |
719d3402 JA |
1688 | * Lookup the cfqq that this bio will be queued with. Allow |
1689 | * merge only if rq is queued there. | |
da775265 | 1690 | */ |
4ac845a2 | 1691 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 | 1692 | if (!cic) |
a6151c3a | 1693 | return false; |
719d3402 | 1694 | |
91fac317 | 1695 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
a6151c3a | 1696 | return cfqq == RQ_CFQQ(rq); |
da775265 JA |
1697 | } |
1698 | ||
812df48d DS |
1699 | static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1700 | { | |
1701 | del_timer(&cfqd->idle_slice_timer); | |
e98ef89b | 1702 | cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg); |
812df48d DS |
1703 | } |
1704 | ||
febffd61 JA |
1705 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
1706 | struct cfq_queue *cfqq) | |
22e2c507 JA |
1707 | { |
1708 | if (cfqq) { | |
b1ffe737 DS |
1709 | cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d", |
1710 | cfqd->serving_prio, cfqd->serving_type); | |
62a37f6b JT |
1711 | cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg); |
1712 | cfqq->slice_start = 0; | |
1713 | cfqq->dispatch_start = jiffies; | |
1714 | cfqq->allocated_slice = 0; | |
1715 | cfqq->slice_end = 0; | |
1716 | cfqq->slice_dispatch = 0; | |
1717 | cfqq->nr_sectors = 0; | |
1718 | ||
1719 | cfq_clear_cfqq_wait_request(cfqq); | |
1720 | cfq_clear_cfqq_must_dispatch(cfqq); | |
1721 | cfq_clear_cfqq_must_alloc_slice(cfqq); | |
1722 | cfq_clear_cfqq_fifo_expire(cfqq); | |
1723 | cfq_mark_cfqq_slice_new(cfqq); | |
1724 | ||
1725 | cfq_del_timer(cfqd, cfqq); | |
22e2c507 JA |
1726 | } |
1727 | ||
1728 | cfqd->active_queue = cfqq; | |
1729 | } | |
1730 | ||
7b14e3b5 JA |
1731 | /* |
1732 | * current cfqq expired its slice (or was too idle), select new one | |
1733 | */ | |
1734 | static void | |
1735 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
e5ff082e | 1736 | bool timed_out) |
7b14e3b5 | 1737 | { |
7b679138 JA |
1738 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
1739 | ||
7b14e3b5 | 1740 | if (cfq_cfqq_wait_request(cfqq)) |
812df48d | 1741 | cfq_del_timer(cfqd, cfqq); |
7b14e3b5 | 1742 | |
7b14e3b5 | 1743 | cfq_clear_cfqq_wait_request(cfqq); |
f75edf2d | 1744 | cfq_clear_cfqq_wait_busy(cfqq); |
7b14e3b5 | 1745 | |
ae54abed SL |
1746 | /* |
1747 | * If this cfqq is shared between multiple processes, check to | |
1748 | * make sure that those processes are still issuing I/Os within | |
1749 | * the mean seek distance. If not, it may be time to break the | |
1750 | * queues apart again. | |
1751 | */ | |
1752 | if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq)) | |
1753 | cfq_mark_cfqq_split_coop(cfqq); | |
1754 | ||
7b14e3b5 | 1755 | /* |
6084cdda | 1756 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 1757 | */ |
c553f8e3 SL |
1758 | if (timed_out) { |
1759 | if (cfq_cfqq_slice_new(cfqq)) | |
ba5bd520 | 1760 | cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq); |
c553f8e3 SL |
1761 | else |
1762 | cfqq->slice_resid = cfqq->slice_end - jiffies; | |
7b679138 JA |
1763 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
1764 | } | |
7b14e3b5 | 1765 | |
e5ff082e | 1766 | cfq_group_served(cfqd, cfqq->cfqg, cfqq); |
dae739eb | 1767 | |
f04a6424 VG |
1768 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
1769 | cfq_del_cfqq_rr(cfqd, cfqq); | |
1770 | ||
edd75ffd | 1771 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
1772 | |
1773 | if (cfqq == cfqd->active_queue) | |
1774 | cfqd->active_queue = NULL; | |
1775 | ||
1776 | if (cfqd->active_cic) { | |
1777 | put_io_context(cfqd->active_cic->ioc); | |
1778 | cfqd->active_cic = NULL; | |
1779 | } | |
7b14e3b5 JA |
1780 | } |
1781 | ||
e5ff082e | 1782 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) |
7b14e3b5 JA |
1783 | { |
1784 | struct cfq_queue *cfqq = cfqd->active_queue; | |
1785 | ||
1786 | if (cfqq) | |
e5ff082e | 1787 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
1788 | } |
1789 | ||
498d3aa2 JA |
1790 | /* |
1791 | * Get next queue for service. Unless we have a queue preemption, | |
1792 | * we'll simply select the first cfqq in the service tree. | |
1793 | */ | |
6d048f53 | 1794 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 1795 | { |
c0324a02 | 1796 | struct cfq_rb_root *service_tree = |
cdb16e8f | 1797 | service_tree_for(cfqd->serving_group, cfqd->serving_prio, |
65b32a57 | 1798 | cfqd->serving_type); |
d9e7620e | 1799 | |
f04a6424 VG |
1800 | if (!cfqd->rq_queued) |
1801 | return NULL; | |
1802 | ||
1fa8f6d6 VG |
1803 | /* There is nothing to dispatch */ |
1804 | if (!service_tree) | |
1805 | return NULL; | |
c0324a02 CZ |
1806 | if (RB_EMPTY_ROOT(&service_tree->rb)) |
1807 | return NULL; | |
1808 | return cfq_rb_first(service_tree); | |
6d048f53 JA |
1809 | } |
1810 | ||
f04a6424 VG |
1811 | static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) |
1812 | { | |
25fb5169 | 1813 | struct cfq_group *cfqg; |
f04a6424 VG |
1814 | struct cfq_queue *cfqq; |
1815 | int i, j; | |
1816 | struct cfq_rb_root *st; | |
1817 | ||
1818 | if (!cfqd->rq_queued) | |
1819 | return NULL; | |
1820 | ||
25fb5169 VG |
1821 | cfqg = cfq_get_next_cfqg(cfqd); |
1822 | if (!cfqg) | |
1823 | return NULL; | |
1824 | ||
f04a6424 VG |
1825 | for_each_cfqg_st(cfqg, i, j, st) |
1826 | if ((cfqq = cfq_rb_first(st)) != NULL) | |
1827 | return cfqq; | |
1828 | return NULL; | |
1829 | } | |
1830 | ||
498d3aa2 JA |
1831 | /* |
1832 | * Get and set a new active queue for service. | |
1833 | */ | |
a36e71f9 JA |
1834 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
1835 | struct cfq_queue *cfqq) | |
6d048f53 | 1836 | { |
e00ef799 | 1837 | if (!cfqq) |
a36e71f9 | 1838 | cfqq = cfq_get_next_queue(cfqd); |
6d048f53 | 1839 | |
22e2c507 | 1840 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 1841 | return cfqq; |
22e2c507 JA |
1842 | } |
1843 | ||
d9e7620e JA |
1844 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
1845 | struct request *rq) | |
1846 | { | |
83096ebf TH |
1847 | if (blk_rq_pos(rq) >= cfqd->last_position) |
1848 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 1849 | else |
83096ebf | 1850 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
1851 | } |
1852 | ||
b2c18e1e | 1853 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
e9ce335d | 1854 | struct request *rq) |
6d048f53 | 1855 | { |
e9ce335d | 1856 | return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR; |
6d048f53 JA |
1857 | } |
1858 | ||
a36e71f9 JA |
1859 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
1860 | struct cfq_queue *cur_cfqq) | |
1861 | { | |
f2d1f0ae | 1862 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
1863 | struct rb_node *parent, *node; |
1864 | struct cfq_queue *__cfqq; | |
1865 | sector_t sector = cfqd->last_position; | |
1866 | ||
1867 | if (RB_EMPTY_ROOT(root)) | |
1868 | return NULL; | |
1869 | ||
1870 | /* | |
1871 | * First, if we find a request starting at the end of the last | |
1872 | * request, choose it. | |
1873 | */ | |
f2d1f0ae | 1874 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
1875 | if (__cfqq) |
1876 | return __cfqq; | |
1877 | ||
1878 | /* | |
1879 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1880 | * will contain the closest sector. | |
1881 | */ | |
1882 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
e9ce335d | 1883 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1884 | return __cfqq; |
1885 | ||
2e46e8b2 | 1886 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1887 | node = rb_next(&__cfqq->p_node); |
1888 | else | |
1889 | node = rb_prev(&__cfqq->p_node); | |
1890 | if (!node) | |
1891 | return NULL; | |
1892 | ||
1893 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
e9ce335d | 1894 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1895 | return __cfqq; |
1896 | ||
1897 | return NULL; | |
1898 | } | |
1899 | ||
1900 | /* | |
1901 | * cfqd - obvious | |
1902 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1903 | * closely cooperating with itself. | |
1904 | * | |
1905 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1906 | * one request, and that cfqd->last_position reflects a position on the disk | |
1907 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1908 | * assumption. | |
1909 | */ | |
1910 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
b3b6d040 | 1911 | struct cfq_queue *cur_cfqq) |
6d048f53 | 1912 | { |
a36e71f9 JA |
1913 | struct cfq_queue *cfqq; |
1914 | ||
39c01b21 DS |
1915 | if (cfq_class_idle(cur_cfqq)) |
1916 | return NULL; | |
e6c5bc73 JM |
1917 | if (!cfq_cfqq_sync(cur_cfqq)) |
1918 | return NULL; | |
1919 | if (CFQQ_SEEKY(cur_cfqq)) | |
1920 | return NULL; | |
1921 | ||
b9d8f4c7 GJ |
1922 | /* |
1923 | * Don't search priority tree if it's the only queue in the group. | |
1924 | */ | |
1925 | if (cur_cfqq->cfqg->nr_cfqq == 1) | |
1926 | return NULL; | |
1927 | ||
6d048f53 | 1928 | /* |
d9e7620e JA |
1929 | * We should notice if some of the queues are cooperating, eg |
1930 | * working closely on the same area of the disk. In that case, | |
1931 | * we can group them together and don't waste time idling. | |
6d048f53 | 1932 | */ |
a36e71f9 JA |
1933 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1934 | if (!cfqq) | |
1935 | return NULL; | |
1936 | ||
8682e1f1 VG |
1937 | /* If new queue belongs to different cfq_group, don't choose it */ |
1938 | if (cur_cfqq->cfqg != cfqq->cfqg) | |
1939 | return NULL; | |
1940 | ||
df5fe3e8 JM |
1941 | /* |
1942 | * It only makes sense to merge sync queues. | |
1943 | */ | |
1944 | if (!cfq_cfqq_sync(cfqq)) | |
1945 | return NULL; | |
e6c5bc73 JM |
1946 | if (CFQQ_SEEKY(cfqq)) |
1947 | return NULL; | |
df5fe3e8 | 1948 | |
c0324a02 CZ |
1949 | /* |
1950 | * Do not merge queues of different priority classes | |
1951 | */ | |
1952 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
1953 | return NULL; | |
1954 | ||
a36e71f9 | 1955 | return cfqq; |
6d048f53 JA |
1956 | } |
1957 | ||
a6d44e98 CZ |
1958 | /* |
1959 | * Determine whether we should enforce idle window for this queue. | |
1960 | */ | |
1961 | ||
1962 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1963 | { | |
1964 | enum wl_prio_t prio = cfqq_prio(cfqq); | |
718eee05 | 1965 | struct cfq_rb_root *service_tree = cfqq->service_tree; |
a6d44e98 | 1966 | |
f04a6424 VG |
1967 | BUG_ON(!service_tree); |
1968 | BUG_ON(!service_tree->count); | |
1969 | ||
b6508c16 VG |
1970 | if (!cfqd->cfq_slice_idle) |
1971 | return false; | |
1972 | ||
a6d44e98 CZ |
1973 | /* We never do for idle class queues. */ |
1974 | if (prio == IDLE_WORKLOAD) | |
1975 | return false; | |
1976 | ||
1977 | /* We do for queues that were marked with idle window flag. */ | |
3c764b7a SL |
1978 | if (cfq_cfqq_idle_window(cfqq) && |
1979 | !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)) | |
a6d44e98 CZ |
1980 | return true; |
1981 | ||
1982 | /* | |
1983 | * Otherwise, we do only if they are the last ones | |
1984 | * in their service tree. | |
1985 | */ | |
f5f2b6ce SL |
1986 | if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) && |
1987 | !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false)) | |
c1e44756 | 1988 | return true; |
b1ffe737 DS |
1989 | cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", |
1990 | service_tree->count); | |
c1e44756 | 1991 | return false; |
a6d44e98 CZ |
1992 | } |
1993 | ||
6d048f53 | 1994 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1995 | { |
1792669c | 1996 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1997 | struct cfq_io_context *cic; |
80bdf0c7 | 1998 | unsigned long sl, group_idle = 0; |
7b14e3b5 | 1999 | |
a68bbddb | 2000 | /* |
f7d7b7a7 JA |
2001 | * SSD device without seek penalty, disable idling. But only do so |
2002 | * for devices that support queuing, otherwise we still have a problem | |
2003 | * with sync vs async workloads. | |
a68bbddb | 2004 | */ |
f7d7b7a7 | 2005 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
2006 | return; |
2007 | ||
dd67d051 | 2008 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 2009 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
2010 | |
2011 | /* | |
2012 | * idle is disabled, either manually or by past process history | |
2013 | */ | |
80bdf0c7 VG |
2014 | if (!cfq_should_idle(cfqd, cfqq)) { |
2015 | /* no queue idling. Check for group idling */ | |
2016 | if (cfqd->cfq_group_idle) | |
2017 | group_idle = cfqd->cfq_group_idle; | |
2018 | else | |
2019 | return; | |
2020 | } | |
6d048f53 | 2021 | |
7b679138 | 2022 | /* |
8e550632 | 2023 | * still active requests from this queue, don't idle |
7b679138 | 2024 | */ |
8e550632 | 2025 | if (cfqq->dispatched) |
7b679138 JA |
2026 | return; |
2027 | ||
22e2c507 JA |
2028 | /* |
2029 | * task has exited, don't wait | |
2030 | */ | |
206dc69b | 2031 | cic = cfqd->active_cic; |
66dac98e | 2032 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
2033 | return; |
2034 | ||
355b659c CZ |
2035 | /* |
2036 | * If our average think time is larger than the remaining time | |
2037 | * slice, then don't idle. This avoids overrunning the allotted | |
2038 | * time slice. | |
2039 | */ | |
383cd721 SL |
2040 | if (sample_valid(cic->ttime.ttime_samples) && |
2041 | (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) { | |
fd16d263 | 2042 | cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu", |
383cd721 | 2043 | cic->ttime.ttime_mean); |
355b659c | 2044 | return; |
b1ffe737 | 2045 | } |
355b659c | 2046 | |
80bdf0c7 VG |
2047 | /* There are other queues in the group, don't do group idle */ |
2048 | if (group_idle && cfqq->cfqg->nr_cfqq > 1) | |
2049 | return; | |
2050 | ||
3b18152c | 2051 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 2052 | |
80bdf0c7 VG |
2053 | if (group_idle) |
2054 | sl = cfqd->cfq_group_idle; | |
2055 | else | |
2056 | sl = cfqd->cfq_slice_idle; | |
206dc69b | 2057 | |
7b14e3b5 | 2058 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
e98ef89b | 2059 | cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg); |
80bdf0c7 VG |
2060 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl, |
2061 | group_idle ? 1 : 0); | |
1da177e4 LT |
2062 | } |
2063 | ||
498d3aa2 JA |
2064 | /* |
2065 | * Move request from internal lists to the request queue dispatch list. | |
2066 | */ | |
165125e1 | 2067 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 2068 | { |
3ed9a296 | 2069 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 2070 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 2071 | |
7b679138 JA |
2072 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
2073 | ||
06d21886 | 2074 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); |
5380a101 | 2075 | cfq_remove_request(rq); |
6d048f53 | 2076 | cfqq->dispatched++; |
80bdf0c7 | 2077 | (RQ_CFQG(rq))->dispatched++; |
5380a101 | 2078 | elv_dispatch_sort(q, rq); |
3ed9a296 | 2079 | |
53c583d2 | 2080 | cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++; |
c4e7893e | 2081 | cfqq->nr_sectors += blk_rq_sectors(rq); |
e98ef89b | 2082 | cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq), |
84c124da | 2083 | rq_data_dir(rq), rq_is_sync(rq)); |
1da177e4 LT |
2084 | } |
2085 | ||
2086 | /* | |
2087 | * return expired entry, or NULL to just start from scratch in rbtree | |
2088 | */ | |
febffd61 | 2089 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 | 2090 | { |
30996f40 | 2091 | struct request *rq = NULL; |
1da177e4 | 2092 | |
3b18152c | 2093 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 2094 | return NULL; |
cb887411 JA |
2095 | |
2096 | cfq_mark_cfqq_fifo_expire(cfqq); | |
2097 | ||
89850f7e JA |
2098 | if (list_empty(&cfqq->fifo)) |
2099 | return NULL; | |
1da177e4 | 2100 | |
89850f7e | 2101 | rq = rq_entry_fifo(cfqq->fifo.next); |
30996f40 | 2102 | if (time_before(jiffies, rq_fifo_time(rq))) |
7b679138 | 2103 | rq = NULL; |
1da177e4 | 2104 | |
30996f40 | 2105 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 2106 | return rq; |
1da177e4 LT |
2107 | } |
2108 | ||
22e2c507 JA |
2109 | static inline int |
2110 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2111 | { | |
2112 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 2113 | |
22e2c507 | 2114 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 2115 | |
b9f8ce05 | 2116 | return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio); |
1da177e4 LT |
2117 | } |
2118 | ||
df5fe3e8 JM |
2119 | /* |
2120 | * Must be called with the queue_lock held. | |
2121 | */ | |
2122 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
2123 | { | |
2124 | int process_refs, io_refs; | |
2125 | ||
2126 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
30d7b944 | 2127 | process_refs = cfqq->ref - io_refs; |
df5fe3e8 JM |
2128 | BUG_ON(process_refs < 0); |
2129 | return process_refs; | |
2130 | } | |
2131 | ||
2132 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
2133 | { | |
e6c5bc73 | 2134 | int process_refs, new_process_refs; |
df5fe3e8 JM |
2135 | struct cfq_queue *__cfqq; |
2136 | ||
c10b61f0 JM |
2137 | /* |
2138 | * If there are no process references on the new_cfqq, then it is | |
2139 | * unsafe to follow the ->new_cfqq chain as other cfqq's in the | |
2140 | * chain may have dropped their last reference (not just their | |
2141 | * last process reference). | |
2142 | */ | |
2143 | if (!cfqq_process_refs(new_cfqq)) | |
2144 | return; | |
2145 | ||
df5fe3e8 JM |
2146 | /* Avoid a circular list and skip interim queue merges */ |
2147 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
2148 | if (__cfqq == cfqq) | |
2149 | return; | |
2150 | new_cfqq = __cfqq; | |
2151 | } | |
2152 | ||
2153 | process_refs = cfqq_process_refs(cfqq); | |
c10b61f0 | 2154 | new_process_refs = cfqq_process_refs(new_cfqq); |
df5fe3e8 JM |
2155 | /* |
2156 | * If the process for the cfqq has gone away, there is no | |
2157 | * sense in merging the queues. | |
2158 | */ | |
c10b61f0 | 2159 | if (process_refs == 0 || new_process_refs == 0) |
df5fe3e8 JM |
2160 | return; |
2161 | ||
e6c5bc73 JM |
2162 | /* |
2163 | * Merge in the direction of the lesser amount of work. | |
2164 | */ | |
e6c5bc73 JM |
2165 | if (new_process_refs >= process_refs) { |
2166 | cfqq->new_cfqq = new_cfqq; | |
30d7b944 | 2167 | new_cfqq->ref += process_refs; |
e6c5bc73 JM |
2168 | } else { |
2169 | new_cfqq->new_cfqq = cfqq; | |
30d7b944 | 2170 | cfqq->ref += new_process_refs; |
e6c5bc73 | 2171 | } |
df5fe3e8 JM |
2172 | } |
2173 | ||
cdb16e8f | 2174 | static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, |
65b32a57 | 2175 | struct cfq_group *cfqg, enum wl_prio_t prio) |
718eee05 CZ |
2176 | { |
2177 | struct cfq_queue *queue; | |
2178 | int i; | |
2179 | bool key_valid = false; | |
2180 | unsigned long lowest_key = 0; | |
2181 | enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; | |
2182 | ||
65b32a57 VG |
2183 | for (i = 0; i <= SYNC_WORKLOAD; ++i) { |
2184 | /* select the one with lowest rb_key */ | |
2185 | queue = cfq_rb_first(service_tree_for(cfqg, prio, i)); | |
718eee05 CZ |
2186 | if (queue && |
2187 | (!key_valid || time_before(queue->rb_key, lowest_key))) { | |
2188 | lowest_key = queue->rb_key; | |
2189 | cur_best = i; | |
2190 | key_valid = true; | |
2191 | } | |
2192 | } | |
2193 | ||
2194 | return cur_best; | |
2195 | } | |
2196 | ||
cdb16e8f | 2197 | static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) |
718eee05 | 2198 | { |
718eee05 CZ |
2199 | unsigned slice; |
2200 | unsigned count; | |
cdb16e8f | 2201 | struct cfq_rb_root *st; |
58ff82f3 | 2202 | unsigned group_slice; |
e4ea0c16 | 2203 | enum wl_prio_t original_prio = cfqd->serving_prio; |
1fa8f6d6 | 2204 | |
718eee05 | 2205 | /* Choose next priority. RT > BE > IDLE */ |
58ff82f3 | 2206 | if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) |
718eee05 | 2207 | cfqd->serving_prio = RT_WORKLOAD; |
58ff82f3 | 2208 | else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) |
718eee05 CZ |
2209 | cfqd->serving_prio = BE_WORKLOAD; |
2210 | else { | |
2211 | cfqd->serving_prio = IDLE_WORKLOAD; | |
2212 | cfqd->workload_expires = jiffies + 1; | |
2213 | return; | |
2214 | } | |
2215 | ||
e4ea0c16 SL |
2216 | if (original_prio != cfqd->serving_prio) |
2217 | goto new_workload; | |
2218 | ||
718eee05 CZ |
2219 | /* |
2220 | * For RT and BE, we have to choose also the type | |
2221 | * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload | |
2222 | * expiration time | |
2223 | */ | |
65b32a57 | 2224 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); |
cdb16e8f | 2225 | count = st->count; |
718eee05 CZ |
2226 | |
2227 | /* | |
65b32a57 | 2228 | * check workload expiration, and that we still have other queues ready |
718eee05 | 2229 | */ |
65b32a57 | 2230 | if (count && !time_after(jiffies, cfqd->workload_expires)) |
718eee05 CZ |
2231 | return; |
2232 | ||
e4ea0c16 | 2233 | new_workload: |
718eee05 CZ |
2234 | /* otherwise select new workload type */ |
2235 | cfqd->serving_type = | |
65b32a57 VG |
2236 | cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio); |
2237 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); | |
cdb16e8f | 2238 | count = st->count; |
718eee05 CZ |
2239 | |
2240 | /* | |
2241 | * the workload slice is computed as a fraction of target latency | |
2242 | * proportional to the number of queues in that workload, over | |
2243 | * all the queues in the same priority class | |
2244 | */ | |
58ff82f3 VG |
2245 | group_slice = cfq_group_slice(cfqd, cfqg); |
2246 | ||
2247 | slice = group_slice * count / | |
2248 | max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], | |
2249 | cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); | |
718eee05 | 2250 | |
f26bd1f0 VG |
2251 | if (cfqd->serving_type == ASYNC_WORKLOAD) { |
2252 | unsigned int tmp; | |
2253 | ||
2254 | /* | |
2255 | * Async queues are currently system wide. Just taking | |
2256 | * proportion of queues with-in same group will lead to higher | |
2257 | * async ratio system wide as generally root group is going | |
2258 | * to have higher weight. A more accurate thing would be to | |
2259 | * calculate system wide asnc/sync ratio. | |
2260 | */ | |
2261 | tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg); | |
2262 | tmp = tmp/cfqd->busy_queues; | |
2263 | slice = min_t(unsigned, slice, tmp); | |
2264 | ||
718eee05 CZ |
2265 | /* async workload slice is scaled down according to |
2266 | * the sync/async slice ratio. */ | |
2267 | slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; | |
f26bd1f0 | 2268 | } else |
718eee05 CZ |
2269 | /* sync workload slice is at least 2 * cfq_slice_idle */ |
2270 | slice = max(slice, 2 * cfqd->cfq_slice_idle); | |
2271 | ||
2272 | slice = max_t(unsigned, slice, CFQ_MIN_TT); | |
b1ffe737 | 2273 | cfq_log(cfqd, "workload slice:%d", slice); |
718eee05 CZ |
2274 | cfqd->workload_expires = jiffies + slice; |
2275 | } | |
2276 | ||
1fa8f6d6 VG |
2277 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) |
2278 | { | |
2279 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
25bc6b07 | 2280 | struct cfq_group *cfqg; |
1fa8f6d6 VG |
2281 | |
2282 | if (RB_EMPTY_ROOT(&st->rb)) | |
2283 | return NULL; | |
25bc6b07 | 2284 | cfqg = cfq_rb_first_group(st); |
25bc6b07 VG |
2285 | update_min_vdisktime(st); |
2286 | return cfqg; | |
1fa8f6d6 VG |
2287 | } |
2288 | ||
cdb16e8f VG |
2289 | static void cfq_choose_cfqg(struct cfq_data *cfqd) |
2290 | { | |
1fa8f6d6 VG |
2291 | struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); |
2292 | ||
2293 | cfqd->serving_group = cfqg; | |
dae739eb VG |
2294 | |
2295 | /* Restore the workload type data */ | |
2296 | if (cfqg->saved_workload_slice) { | |
2297 | cfqd->workload_expires = jiffies + cfqg->saved_workload_slice; | |
2298 | cfqd->serving_type = cfqg->saved_workload; | |
2299 | cfqd->serving_prio = cfqg->saved_serving_prio; | |
66ae2919 GJ |
2300 | } else |
2301 | cfqd->workload_expires = jiffies - 1; | |
2302 | ||
1fa8f6d6 | 2303 | choose_service_tree(cfqd, cfqg); |
cdb16e8f VG |
2304 | } |
2305 | ||
22e2c507 | 2306 | /* |
498d3aa2 JA |
2307 | * Select a queue for service. If we have a current active queue, |
2308 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 2309 | */ |
1b5ed5e1 | 2310 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 2311 | { |
a36e71f9 | 2312 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 2313 | |
22e2c507 JA |
2314 | cfqq = cfqd->active_queue; |
2315 | if (!cfqq) | |
2316 | goto new_queue; | |
1da177e4 | 2317 | |
f04a6424 VG |
2318 | if (!cfqd->rq_queued) |
2319 | return NULL; | |
c244bb50 VG |
2320 | |
2321 | /* | |
2322 | * We were waiting for group to get backlogged. Expire the queue | |
2323 | */ | |
2324 | if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list)) | |
2325 | goto expire; | |
2326 | ||
22e2c507 | 2327 | /* |
6d048f53 | 2328 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 2329 | */ |
7667aa06 VG |
2330 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) { |
2331 | /* | |
2332 | * If slice had not expired at the completion of last request | |
2333 | * we might not have turned on wait_busy flag. Don't expire | |
2334 | * the queue yet. Allow the group to get backlogged. | |
2335 | * | |
2336 | * The very fact that we have used the slice, that means we | |
2337 | * have been idling all along on this queue and it should be | |
2338 | * ok to wait for this request to complete. | |
2339 | */ | |
82bbbf28 VG |
2340 | if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list) |
2341 | && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { | |
2342 | cfqq = NULL; | |
7667aa06 | 2343 | goto keep_queue; |
82bbbf28 | 2344 | } else |
80bdf0c7 | 2345 | goto check_group_idle; |
7667aa06 | 2346 | } |
1da177e4 | 2347 | |
22e2c507 | 2348 | /* |
6d048f53 JA |
2349 | * The active queue has requests and isn't expired, allow it to |
2350 | * dispatch. | |
22e2c507 | 2351 | */ |
dd67d051 | 2352 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 2353 | goto keep_queue; |
6d048f53 | 2354 | |
a36e71f9 JA |
2355 | /* |
2356 | * If another queue has a request waiting within our mean seek | |
2357 | * distance, let it run. The expire code will check for close | |
2358 | * cooperators and put the close queue at the front of the service | |
df5fe3e8 | 2359 | * tree. If possible, merge the expiring queue with the new cfqq. |
a36e71f9 | 2360 | */ |
b3b6d040 | 2361 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); |
df5fe3e8 JM |
2362 | if (new_cfqq) { |
2363 | if (!cfqq->new_cfqq) | |
2364 | cfq_setup_merge(cfqq, new_cfqq); | |
a36e71f9 | 2365 | goto expire; |
df5fe3e8 | 2366 | } |
a36e71f9 | 2367 | |
6d048f53 JA |
2368 | /* |
2369 | * No requests pending. If the active queue still has requests in | |
2370 | * flight or is idling for a new request, allow either of these | |
2371 | * conditions to happen (or time out) before selecting a new queue. | |
2372 | */ | |
80bdf0c7 VG |
2373 | if (timer_pending(&cfqd->idle_slice_timer)) { |
2374 | cfqq = NULL; | |
2375 | goto keep_queue; | |
2376 | } | |
2377 | ||
8e1ac665 SL |
2378 | /* |
2379 | * This is a deep seek queue, but the device is much faster than | |
2380 | * the queue can deliver, don't idle | |
2381 | **/ | |
2382 | if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) && | |
2383 | (cfq_cfqq_slice_new(cfqq) || | |
2384 | (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) { | |
2385 | cfq_clear_cfqq_deep(cfqq); | |
2386 | cfq_clear_cfqq_idle_window(cfqq); | |
2387 | } | |
2388 | ||
80bdf0c7 VG |
2389 | if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { |
2390 | cfqq = NULL; | |
2391 | goto keep_queue; | |
2392 | } | |
2393 | ||
2394 | /* | |
2395 | * If group idle is enabled and there are requests dispatched from | |
2396 | * this group, wait for requests to complete. | |
2397 | */ | |
2398 | check_group_idle: | |
7700fc4f SL |
2399 | if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 && |
2400 | cfqq->cfqg->dispatched && | |
2401 | !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) { | |
caaa5f9f JA |
2402 | cfqq = NULL; |
2403 | goto keep_queue; | |
22e2c507 JA |
2404 | } |
2405 | ||
3b18152c | 2406 | expire: |
e5ff082e | 2407 | cfq_slice_expired(cfqd, 0); |
3b18152c | 2408 | new_queue: |
718eee05 CZ |
2409 | /* |
2410 | * Current queue expired. Check if we have to switch to a new | |
2411 | * service tree | |
2412 | */ | |
2413 | if (!new_cfqq) | |
cdb16e8f | 2414 | cfq_choose_cfqg(cfqd); |
718eee05 | 2415 | |
a36e71f9 | 2416 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 2417 | keep_queue: |
3b18152c | 2418 | return cfqq; |
22e2c507 JA |
2419 | } |
2420 | ||
febffd61 | 2421 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
2422 | { |
2423 | int dispatched = 0; | |
2424 | ||
2425 | while (cfqq->next_rq) { | |
2426 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
2427 | dispatched++; | |
2428 | } | |
2429 | ||
2430 | BUG_ON(!list_empty(&cfqq->fifo)); | |
f04a6424 VG |
2431 | |
2432 | /* By default cfqq is not expired if it is empty. Do it explicitly */ | |
e5ff082e | 2433 | __cfq_slice_expired(cfqq->cfqd, cfqq, 0); |
d9e7620e JA |
2434 | return dispatched; |
2435 | } | |
2436 | ||
498d3aa2 JA |
2437 | /* |
2438 | * Drain our current requests. Used for barriers and when switching | |
2439 | * io schedulers on-the-fly. | |
2440 | */ | |
d9e7620e | 2441 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 2442 | { |
0871714e | 2443 | struct cfq_queue *cfqq; |
d9e7620e | 2444 | int dispatched = 0; |
cdb16e8f | 2445 | |
3440c49f | 2446 | /* Expire the timeslice of the current active queue first */ |
e5ff082e | 2447 | cfq_slice_expired(cfqd, 0); |
3440c49f DS |
2448 | while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) { |
2449 | __cfq_set_active_queue(cfqd, cfqq); | |
f04a6424 | 2450 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); |
3440c49f | 2451 | } |
1b5ed5e1 | 2452 | |
1b5ed5e1 TH |
2453 | BUG_ON(cfqd->busy_queues); |
2454 | ||
6923715a | 2455 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
2456 | return dispatched; |
2457 | } | |
2458 | ||
abc3c744 SL |
2459 | static inline bool cfq_slice_used_soon(struct cfq_data *cfqd, |
2460 | struct cfq_queue *cfqq) | |
2461 | { | |
2462 | /* the queue hasn't finished any request, can't estimate */ | |
2463 | if (cfq_cfqq_slice_new(cfqq)) | |
c1e44756 | 2464 | return true; |
abc3c744 SL |
2465 | if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched, |
2466 | cfqq->slice_end)) | |
c1e44756 | 2467 | return true; |
abc3c744 | 2468 | |
c1e44756 | 2469 | return false; |
abc3c744 SL |
2470 | } |
2471 | ||
0b182d61 | 2472 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
2f5cb738 | 2473 | { |
2f5cb738 | 2474 | unsigned int max_dispatch; |
22e2c507 | 2475 | |
5ad531db JA |
2476 | /* |
2477 | * Drain async requests before we start sync IO | |
2478 | */ | |
53c583d2 | 2479 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC]) |
0b182d61 | 2480 | return false; |
5ad531db | 2481 | |
2f5cb738 JA |
2482 | /* |
2483 | * If this is an async queue and we have sync IO in flight, let it wait | |
2484 | */ | |
53c583d2 | 2485 | if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq)) |
0b182d61 | 2486 | return false; |
2f5cb738 | 2487 | |
abc3c744 | 2488 | max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1); |
2f5cb738 JA |
2489 | if (cfq_class_idle(cfqq)) |
2490 | max_dispatch = 1; | |
b4878f24 | 2491 | |
2f5cb738 JA |
2492 | /* |
2493 | * Does this cfqq already have too much IO in flight? | |
2494 | */ | |
2495 | if (cfqq->dispatched >= max_dispatch) { | |
ef8a41df | 2496 | bool promote_sync = false; |
2f5cb738 JA |
2497 | /* |
2498 | * idle queue must always only have a single IO in flight | |
2499 | */ | |
3ed9a296 | 2500 | if (cfq_class_idle(cfqq)) |
0b182d61 | 2501 | return false; |
3ed9a296 | 2502 | |
ef8a41df | 2503 | /* |
c4ade94f LS |
2504 | * If there is only one sync queue |
2505 | * we can ignore async queue here and give the sync | |
ef8a41df SL |
2506 | * queue no dispatch limit. The reason is a sync queue can |
2507 | * preempt async queue, limiting the sync queue doesn't make | |
2508 | * sense. This is useful for aiostress test. | |
2509 | */ | |
c4ade94f LS |
2510 | if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1) |
2511 | promote_sync = true; | |
ef8a41df | 2512 | |
2f5cb738 JA |
2513 | /* |
2514 | * We have other queues, don't allow more IO from this one | |
2515 | */ | |
ef8a41df SL |
2516 | if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) && |
2517 | !promote_sync) | |
0b182d61 | 2518 | return false; |
9ede209e | 2519 | |
365722bb | 2520 | /* |
474b18cc | 2521 | * Sole queue user, no limit |
365722bb | 2522 | */ |
ef8a41df | 2523 | if (cfqd->busy_queues == 1 || promote_sync) |
abc3c744 SL |
2524 | max_dispatch = -1; |
2525 | else | |
2526 | /* | |
2527 | * Normally we start throttling cfqq when cfq_quantum/2 | |
2528 | * requests have been dispatched. But we can drive | |
2529 | * deeper queue depths at the beginning of slice | |
2530 | * subjected to upper limit of cfq_quantum. | |
2531 | * */ | |
2532 | max_dispatch = cfqd->cfq_quantum; | |
8e296755 JA |
2533 | } |
2534 | ||
2535 | /* | |
2536 | * Async queues must wait a bit before being allowed dispatch. | |
2537 | * We also ramp up the dispatch depth gradually for async IO, | |
2538 | * based on the last sync IO we serviced | |
2539 | */ | |
963b72fc | 2540 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { |
573412b2 | 2541 | unsigned long last_sync = jiffies - cfqd->last_delayed_sync; |
8e296755 | 2542 | unsigned int depth; |
365722bb | 2543 | |
61f0c1dc | 2544 | depth = last_sync / cfqd->cfq_slice[1]; |
e00c54c3 JA |
2545 | if (!depth && !cfqq->dispatched) |
2546 | depth = 1; | |
8e296755 JA |
2547 | if (depth < max_dispatch) |
2548 | max_dispatch = depth; | |
2f5cb738 | 2549 | } |
3ed9a296 | 2550 | |
0b182d61 JA |
2551 | /* |
2552 | * If we're below the current max, allow a dispatch | |
2553 | */ | |
2554 | return cfqq->dispatched < max_dispatch; | |
2555 | } | |
2556 | ||
2557 | /* | |
2558 | * Dispatch a request from cfqq, moving them to the request queue | |
2559 | * dispatch list. | |
2560 | */ | |
2561 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2562 | { | |
2563 | struct request *rq; | |
2564 | ||
2565 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
2566 | ||
2567 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
2568 | return false; | |
2569 | ||
2570 | /* | |
2571 | * follow expired path, else get first next available | |
2572 | */ | |
2573 | rq = cfq_check_fifo(cfqq); | |
2574 | if (!rq) | |
2575 | rq = cfqq->next_rq; | |
2576 | ||
2577 | /* | |
2578 | * insert request into driver dispatch list | |
2579 | */ | |
2580 | cfq_dispatch_insert(cfqd->queue, rq); | |
2581 | ||
2582 | if (!cfqd->active_cic) { | |
2583 | struct cfq_io_context *cic = RQ_CIC(rq); | |
2584 | ||
2585 | atomic_long_inc(&cic->ioc->refcount); | |
2586 | cfqd->active_cic = cic; | |
2587 | } | |
2588 | ||
2589 | return true; | |
2590 | } | |
2591 | ||
2592 | /* | |
2593 | * Find the cfqq that we need to service and move a request from that to the | |
2594 | * dispatch list | |
2595 | */ | |
2596 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
2597 | { | |
2598 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2599 | struct cfq_queue *cfqq; | |
2600 | ||
2601 | if (!cfqd->busy_queues) | |
2602 | return 0; | |
2603 | ||
2604 | if (unlikely(force)) | |
2605 | return cfq_forced_dispatch(cfqd); | |
2606 | ||
2607 | cfqq = cfq_select_queue(cfqd); | |
2608 | if (!cfqq) | |
8e296755 JA |
2609 | return 0; |
2610 | ||
2f5cb738 | 2611 | /* |
0b182d61 | 2612 | * Dispatch a request from this cfqq, if it is allowed |
2f5cb738 | 2613 | */ |
0b182d61 JA |
2614 | if (!cfq_dispatch_request(cfqd, cfqq)) |
2615 | return 0; | |
2616 | ||
2f5cb738 | 2617 | cfqq->slice_dispatch++; |
b029195d | 2618 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 2619 | |
2f5cb738 JA |
2620 | /* |
2621 | * expire an async queue immediately if it has used up its slice. idle | |
2622 | * queue always expire after 1 dispatch round. | |
2623 | */ | |
2624 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
2625 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
2626 | cfq_class_idle(cfqq))) { | |
2627 | cfqq->slice_end = jiffies + 1; | |
e5ff082e | 2628 | cfq_slice_expired(cfqd, 0); |
1da177e4 LT |
2629 | } |
2630 | ||
b217a903 | 2631 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); |
2f5cb738 | 2632 | return 1; |
1da177e4 LT |
2633 | } |
2634 | ||
1da177e4 | 2635 | /* |
5e705374 JA |
2636 | * task holds one reference to the queue, dropped when task exits. each rq |
2637 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 | 2638 | * |
b1c35769 | 2639 | * Each cfq queue took a reference on the parent group. Drop it now. |
1da177e4 LT |
2640 | * queue lock must be held here. |
2641 | */ | |
2642 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
2643 | { | |
22e2c507 | 2644 | struct cfq_data *cfqd = cfqq->cfqd; |
0bbfeb83 | 2645 | struct cfq_group *cfqg; |
22e2c507 | 2646 | |
30d7b944 | 2647 | BUG_ON(cfqq->ref <= 0); |
1da177e4 | 2648 | |
30d7b944 SL |
2649 | cfqq->ref--; |
2650 | if (cfqq->ref) | |
1da177e4 LT |
2651 | return; |
2652 | ||
7b679138 | 2653 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 2654 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 2655 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
b1c35769 | 2656 | cfqg = cfqq->cfqg; |
1da177e4 | 2657 | |
28f95cbc | 2658 | if (unlikely(cfqd->active_queue == cfqq)) { |
e5ff082e | 2659 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 2660 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 2661 | } |
22e2c507 | 2662 | |
f04a6424 | 2663 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 | 2664 | kmem_cache_free(cfq_pool, cfqq); |
b1c35769 | 2665 | cfq_put_cfqg(cfqg); |
1da177e4 LT |
2666 | } |
2667 | ||
d6de8be7 | 2668 | /* |
5f45c695 | 2669 | * Call func for each cic attached to this ioc. |
d6de8be7 | 2670 | */ |
07416d29 | 2671 | static void |
5f45c695 JA |
2672 | call_for_each_cic(struct io_context *ioc, |
2673 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
07416d29 JA |
2674 | { |
2675 | struct cfq_io_context *cic; | |
2676 | struct hlist_node *n; | |
2677 | ||
5f45c695 JA |
2678 | rcu_read_lock(); |
2679 | ||
07416d29 JA |
2680 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) |
2681 | func(ioc, cic); | |
07416d29 | 2682 | |
4ac845a2 | 2683 | rcu_read_unlock(); |
34e6bbf2 FC |
2684 | } |
2685 | ||
2686 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
2687 | { | |
2688 | struct cfq_io_context *cic; | |
2689 | ||
2690 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
2691 | ||
2692 | kmem_cache_free(cfq_ioc_pool, cic); | |
245b2e70 | 2693 | elv_ioc_count_dec(cfq_ioc_count); |
34e6bbf2 | 2694 | |
9a11b4ed JA |
2695 | if (ioc_gone) { |
2696 | /* | |
2697 | * CFQ scheduler is exiting, grab exit lock and check | |
2698 | * the pending io context count. If it hits zero, | |
2699 | * complete ioc_gone and set it back to NULL | |
2700 | */ | |
2701 | spin_lock(&ioc_gone_lock); | |
245b2e70 | 2702 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { |
9a11b4ed JA |
2703 | complete(ioc_gone); |
2704 | ioc_gone = NULL; | |
2705 | } | |
2706 | spin_unlock(&ioc_gone_lock); | |
2707 | } | |
34e6bbf2 | 2708 | } |
4ac845a2 | 2709 | |
34e6bbf2 FC |
2710 | static void cfq_cic_free(struct cfq_io_context *cic) |
2711 | { | |
2712 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
2713 | } |
2714 | ||
2715 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
2716 | { | |
2717 | unsigned long flags; | |
bca4b914 | 2718 | unsigned long dead_key = (unsigned long) cic->key; |
4ac845a2 | 2719 | |
bca4b914 | 2720 | BUG_ON(!(dead_key & CIC_DEAD_KEY)); |
4ac845a2 JA |
2721 | |
2722 | spin_lock_irqsave(&ioc->lock, flags); | |
80b15c73 | 2723 | radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT); |
ffc4e759 | 2724 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2725 | spin_unlock_irqrestore(&ioc->lock, flags); |
2726 | ||
34e6bbf2 | 2727 | cfq_cic_free(cic); |
4ac845a2 JA |
2728 | } |
2729 | ||
d6de8be7 JA |
2730 | /* |
2731 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
2732 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
2733 | * and ->trim() which is called with the task lock held | |
2734 | */ | |
4ac845a2 JA |
2735 | static void cfq_free_io_context(struct io_context *ioc) |
2736 | { | |
4ac845a2 | 2737 | /* |
34e6bbf2 FC |
2738 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
2739 | * so no more cic's are allowed to be linked into this ioc. So it | |
2740 | * should be ok to iterate over the known list, we will see all cic's | |
2741 | * since no new ones are added. | |
4ac845a2 | 2742 | */ |
5f45c695 | 2743 | call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
2744 | } |
2745 | ||
d02a2c07 | 2746 | static void cfq_put_cooperator(struct cfq_queue *cfqq) |
1da177e4 | 2747 | { |
df5fe3e8 JM |
2748 | struct cfq_queue *__cfqq, *next; |
2749 | ||
df5fe3e8 JM |
2750 | /* |
2751 | * If this queue was scheduled to merge with another queue, be | |
2752 | * sure to drop the reference taken on that queue (and others in | |
2753 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
2754 | */ | |
2755 | __cfqq = cfqq->new_cfqq; | |
2756 | while (__cfqq) { | |
2757 | if (__cfqq == cfqq) { | |
2758 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
2759 | break; | |
2760 | } | |
2761 | next = __cfqq->new_cfqq; | |
2762 | cfq_put_queue(__cfqq); | |
2763 | __cfqq = next; | |
2764 | } | |
d02a2c07 SL |
2765 | } |
2766 | ||
2767 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2768 | { | |
2769 | if (unlikely(cfqq == cfqd->active_queue)) { | |
2770 | __cfq_slice_expired(cfqd, cfqq, 0); | |
2771 | cfq_schedule_dispatch(cfqd); | |
2772 | } | |
2773 | ||
2774 | cfq_put_cooperator(cfqq); | |
df5fe3e8 | 2775 | |
89850f7e JA |
2776 | cfq_put_queue(cfqq); |
2777 | } | |
22e2c507 | 2778 | |
89850f7e JA |
2779 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
2780 | struct cfq_io_context *cic) | |
2781 | { | |
4faa3c81 FC |
2782 | struct io_context *ioc = cic->ioc; |
2783 | ||
fc46379d | 2784 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
2785 | |
2786 | /* | |
bca4b914 | 2787 | * Make sure dead mark is seen for dead queues |
4ac845a2 | 2788 | */ |
fc46379d | 2789 | smp_wmb(); |
bca4b914 | 2790 | cic->key = cfqd_dead_key(cfqd); |
fc46379d | 2791 | |
3181faa8 | 2792 | rcu_read_lock(); |
9b50902d | 2793 | if (rcu_dereference(ioc->ioc_data) == cic) { |
3181faa8 | 2794 | rcu_read_unlock(); |
9b50902d | 2795 | spin_lock(&ioc->lock); |
4faa3c81 | 2796 | rcu_assign_pointer(ioc->ioc_data, NULL); |
9b50902d | 2797 | spin_unlock(&ioc->lock); |
3181faa8 SL |
2798 | } else |
2799 | rcu_read_unlock(); | |
4faa3c81 | 2800 | |
ff6657c6 JA |
2801 | if (cic->cfqq[BLK_RW_ASYNC]) { |
2802 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
2803 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
2804 | } |
2805 | ||
ff6657c6 JA |
2806 | if (cic->cfqq[BLK_RW_SYNC]) { |
2807 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
2808 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 2809 | } |
89850f7e JA |
2810 | } |
2811 | ||
4ac845a2 JA |
2812 | static void cfq_exit_single_io_context(struct io_context *ioc, |
2813 | struct cfq_io_context *cic) | |
89850f7e | 2814 | { |
bca4b914 | 2815 | struct cfq_data *cfqd = cic_to_cfqd(cic); |
89850f7e | 2816 | |
89850f7e | 2817 | if (cfqd) { |
165125e1 | 2818 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 2819 | unsigned long flags; |
89850f7e | 2820 | |
4ac845a2 | 2821 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
2822 | |
2823 | /* | |
2824 | * Ensure we get a fresh copy of the ->key to prevent | |
2825 | * race between exiting task and queue | |
2826 | */ | |
2827 | smp_read_barrier_depends(); | |
bca4b914 | 2828 | if (cic->key == cfqd) |
62c1fe9d JA |
2829 | __cfq_exit_single_io_context(cfqd, cic); |
2830 | ||
4ac845a2 | 2831 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 2832 | } |
1da177e4 LT |
2833 | } |
2834 | ||
498d3aa2 JA |
2835 | /* |
2836 | * The process that ioc belongs to has exited, we need to clean up | |
2837 | * and put the internal structures we have that belongs to that process. | |
2838 | */ | |
e2d74ac0 | 2839 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 2840 | { |
4ac845a2 | 2841 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
2842 | } |
2843 | ||
22e2c507 | 2844 | static struct cfq_io_context * |
8267e268 | 2845 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2846 | { |
b5deef90 | 2847 | struct cfq_io_context *cic; |
1da177e4 | 2848 | |
94f6030c CL |
2849 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
2850 | cfqd->queue->node); | |
1da177e4 | 2851 | if (cic) { |
383cd721 | 2852 | cic->ttime.last_end_request = jiffies; |
553698f9 | 2853 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 2854 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
2855 | cic->dtor = cfq_free_io_context; |
2856 | cic->exit = cfq_exit_io_context; | |
245b2e70 | 2857 | elv_ioc_count_inc(cfq_ioc_count); |
1da177e4 LT |
2858 | } |
2859 | ||
2860 | return cic; | |
2861 | } | |
2862 | ||
fd0928df | 2863 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
2864 | { |
2865 | struct task_struct *tsk = current; | |
2866 | int ioprio_class; | |
2867 | ||
3b18152c | 2868 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
2869 | return; |
2870 | ||
fd0928df | 2871 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 2872 | switch (ioprio_class) { |
fe094d98 JA |
2873 | default: |
2874 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
2875 | case IOPRIO_CLASS_NONE: | |
2876 | /* | |
6d63c275 | 2877 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
2878 | */ |
2879 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 2880 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
2881 | break; |
2882 | case IOPRIO_CLASS_RT: | |
2883 | cfqq->ioprio = task_ioprio(ioc); | |
2884 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
2885 | break; | |
2886 | case IOPRIO_CLASS_BE: | |
2887 | cfqq->ioprio = task_ioprio(ioc); | |
2888 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2889 | break; | |
2890 | case IOPRIO_CLASS_IDLE: | |
2891 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
2892 | cfqq->ioprio = 7; | |
2893 | cfq_clear_cfqq_idle_window(cfqq); | |
2894 | break; | |
22e2c507 JA |
2895 | } |
2896 | ||
2897 | /* | |
2898 | * keep track of original prio settings in case we have to temporarily | |
2899 | * elevate the priority of this queue | |
2900 | */ | |
2901 | cfqq->org_ioprio = cfqq->ioprio; | |
3b18152c | 2902 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
2903 | } |
2904 | ||
febffd61 | 2905 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 2906 | { |
bca4b914 | 2907 | struct cfq_data *cfqd = cic_to_cfqd(cic); |
478a82b0 | 2908 | struct cfq_queue *cfqq; |
c1b707d2 | 2909 | unsigned long flags; |
35e6077c | 2910 | |
caaa5f9f JA |
2911 | if (unlikely(!cfqd)) |
2912 | return; | |
2913 | ||
c1b707d2 | 2914 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 2915 | |
ff6657c6 | 2916 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
2917 | if (cfqq) { |
2918 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
2919 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
2920 | GFP_ATOMIC); | |
caaa5f9f | 2921 | if (new_cfqq) { |
ff6657c6 | 2922 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
2923 | cfq_put_queue(cfqq); |
2924 | } | |
22e2c507 | 2925 | } |
caaa5f9f | 2926 | |
ff6657c6 | 2927 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
2928 | if (cfqq) |
2929 | cfq_mark_cfqq_prio_changed(cfqq); | |
2930 | ||
c1b707d2 | 2931 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
2932 | } |
2933 | ||
fc46379d | 2934 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 2935 | { |
4ac845a2 | 2936 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 2937 | ioc->ioprio_changed = 0; |
22e2c507 JA |
2938 | } |
2939 | ||
d5036d77 | 2940 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 2941 | pid_t pid, bool is_sync) |
d5036d77 JA |
2942 | { |
2943 | RB_CLEAR_NODE(&cfqq->rb_node); | |
2944 | RB_CLEAR_NODE(&cfqq->p_node); | |
2945 | INIT_LIST_HEAD(&cfqq->fifo); | |
2946 | ||
30d7b944 | 2947 | cfqq->ref = 0; |
d5036d77 JA |
2948 | cfqq->cfqd = cfqd; |
2949 | ||
2950 | cfq_mark_cfqq_prio_changed(cfqq); | |
2951 | ||
2952 | if (is_sync) { | |
2953 | if (!cfq_class_idle(cfqq)) | |
2954 | cfq_mark_cfqq_idle_window(cfqq); | |
2955 | cfq_mark_cfqq_sync(cfqq); | |
2956 | } | |
2957 | cfqq->pid = pid; | |
2958 | } | |
2959 | ||
24610333 VG |
2960 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
2961 | static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic) | |
2962 | { | |
2963 | struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1); | |
bca4b914 | 2964 | struct cfq_data *cfqd = cic_to_cfqd(cic); |
24610333 VG |
2965 | unsigned long flags; |
2966 | struct request_queue *q; | |
2967 | ||
2968 | if (unlikely(!cfqd)) | |
2969 | return; | |
2970 | ||
2971 | q = cfqd->queue; | |
2972 | ||
2973 | spin_lock_irqsave(q->queue_lock, flags); | |
2974 | ||
2975 | if (sync_cfqq) { | |
2976 | /* | |
2977 | * Drop reference to sync queue. A new sync queue will be | |
2978 | * assigned in new group upon arrival of a fresh request. | |
2979 | */ | |
2980 | cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup"); | |
2981 | cic_set_cfqq(cic, NULL, 1); | |
2982 | cfq_put_queue(sync_cfqq); | |
2983 | } | |
2984 | ||
2985 | spin_unlock_irqrestore(q->queue_lock, flags); | |
2986 | } | |
2987 | ||
2988 | static void cfq_ioc_set_cgroup(struct io_context *ioc) | |
2989 | { | |
2990 | call_for_each_cic(ioc, changed_cgroup); | |
2991 | ioc->cgroup_changed = 0; | |
2992 | } | |
2993 | #endif /* CONFIG_CFQ_GROUP_IOSCHED */ | |
2994 | ||
22e2c507 | 2995 | static struct cfq_queue * |
a6151c3a | 2996 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, |
fd0928df | 2997 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 2998 | { |
22e2c507 | 2999 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 3000 | struct cfq_io_context *cic; |
cdb16e8f | 3001 | struct cfq_group *cfqg; |
22e2c507 JA |
3002 | |
3003 | retry: | |
3e59cf9d | 3004 | cfqg = cfq_get_cfqg(cfqd); |
4ac845a2 | 3005 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
3006 | /* cic always exists here */ |
3007 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 3008 | |
6118b70b JA |
3009 | /* |
3010 | * Always try a new alloc if we fell back to the OOM cfqq | |
3011 | * originally, since it should just be a temporary situation. | |
3012 | */ | |
3013 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
3014 | cfqq = NULL; | |
22e2c507 JA |
3015 | if (new_cfqq) { |
3016 | cfqq = new_cfqq; | |
3017 | new_cfqq = NULL; | |
3018 | } else if (gfp_mask & __GFP_WAIT) { | |
3019 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 3020 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 3021 | gfp_mask | __GFP_ZERO, |
94f6030c | 3022 | cfqd->queue->node); |
22e2c507 | 3023 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
3024 | if (new_cfqq) |
3025 | goto retry; | |
22e2c507 | 3026 | } else { |
94f6030c CL |
3027 | cfqq = kmem_cache_alloc_node(cfq_pool, |
3028 | gfp_mask | __GFP_ZERO, | |
3029 | cfqd->queue->node); | |
22e2c507 JA |
3030 | } |
3031 | ||
6118b70b JA |
3032 | if (cfqq) { |
3033 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
3034 | cfq_init_prio_data(cfqq, ioc); | |
cdb16e8f | 3035 | cfq_link_cfqq_cfqg(cfqq, cfqg); |
6118b70b JA |
3036 | cfq_log_cfqq(cfqd, cfqq, "alloced"); |
3037 | } else | |
3038 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
3039 | } |
3040 | ||
3041 | if (new_cfqq) | |
3042 | kmem_cache_free(cfq_pool, new_cfqq); | |
3043 | ||
22e2c507 JA |
3044 | return cfqq; |
3045 | } | |
3046 | ||
c2dea2d1 VT |
3047 | static struct cfq_queue ** |
3048 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
3049 | { | |
fe094d98 | 3050 | switch (ioprio_class) { |
c2dea2d1 VT |
3051 | case IOPRIO_CLASS_RT: |
3052 | return &cfqd->async_cfqq[0][ioprio]; | |
3053 | case IOPRIO_CLASS_BE: | |
3054 | return &cfqd->async_cfqq[1][ioprio]; | |
3055 | case IOPRIO_CLASS_IDLE: | |
3056 | return &cfqd->async_idle_cfqq; | |
3057 | default: | |
3058 | BUG(); | |
3059 | } | |
3060 | } | |
3061 | ||
15c31be4 | 3062 | static struct cfq_queue * |
a6151c3a | 3063 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, |
15c31be4 JA |
3064 | gfp_t gfp_mask) |
3065 | { | |
fd0928df JA |
3066 | const int ioprio = task_ioprio(ioc); |
3067 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 3068 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
3069 | struct cfq_queue *cfqq = NULL; |
3070 | ||
c2dea2d1 VT |
3071 | if (!is_sync) { |
3072 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
3073 | cfqq = *async_cfqq; | |
3074 | } | |
3075 | ||
6118b70b | 3076 | if (!cfqq) |
fd0928df | 3077 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
3078 | |
3079 | /* | |
3080 | * pin the queue now that it's allocated, scheduler exit will prune it | |
3081 | */ | |
c2dea2d1 | 3082 | if (!is_sync && !(*async_cfqq)) { |
30d7b944 | 3083 | cfqq->ref++; |
c2dea2d1 | 3084 | *async_cfqq = cfqq; |
15c31be4 JA |
3085 | } |
3086 | ||
30d7b944 | 3087 | cfqq->ref++; |
15c31be4 JA |
3088 | return cfqq; |
3089 | } | |
3090 | ||
498d3aa2 JA |
3091 | /* |
3092 | * We drop cfq io contexts lazily, so we may find a dead one. | |
3093 | */ | |
dbecf3ab | 3094 | static void |
4ac845a2 JA |
3095 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
3096 | struct cfq_io_context *cic) | |
dbecf3ab | 3097 | { |
4ac845a2 JA |
3098 | unsigned long flags; |
3099 | ||
fc46379d | 3100 | WARN_ON(!list_empty(&cic->queue_list)); |
bca4b914 | 3101 | BUG_ON(cic->key != cfqd_dead_key(cfqd)); |
597bc485 | 3102 | |
4ac845a2 JA |
3103 | spin_lock_irqsave(&ioc->lock, flags); |
3104 | ||
726e99ab SL |
3105 | BUG_ON(rcu_dereference_check(ioc->ioc_data, |
3106 | lockdep_is_held(&ioc->lock)) == cic); | |
597bc485 | 3107 | |
80b15c73 | 3108 | radix_tree_delete(&ioc->radix_root, cfqd->cic_index); |
ffc4e759 | 3109 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
3110 | spin_unlock_irqrestore(&ioc->lock, flags); |
3111 | ||
3112 | cfq_cic_free(cic); | |
dbecf3ab OH |
3113 | } |
3114 | ||
e2d74ac0 | 3115 | static struct cfq_io_context * |
4ac845a2 | 3116 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 3117 | { |
e2d74ac0 | 3118 | struct cfq_io_context *cic; |
d6de8be7 | 3119 | unsigned long flags; |
e2d74ac0 | 3120 | |
91fac317 VT |
3121 | if (unlikely(!ioc)) |
3122 | return NULL; | |
3123 | ||
d6de8be7 JA |
3124 | rcu_read_lock(); |
3125 | ||
597bc485 JA |
3126 | /* |
3127 | * we maintain a last-hit cache, to avoid browsing over the tree | |
3128 | */ | |
4ac845a2 | 3129 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
3130 | if (cic && cic->key == cfqd) { |
3131 | rcu_read_unlock(); | |
597bc485 | 3132 | return cic; |
d6de8be7 | 3133 | } |
597bc485 | 3134 | |
4ac845a2 | 3135 | do { |
80b15c73 | 3136 | cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index); |
4ac845a2 JA |
3137 | rcu_read_unlock(); |
3138 | if (!cic) | |
3139 | break; | |
bca4b914 | 3140 | if (unlikely(cic->key != cfqd)) { |
4ac845a2 | 3141 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 3142 | rcu_read_lock(); |
4ac845a2 | 3143 | continue; |
dbecf3ab | 3144 | } |
e2d74ac0 | 3145 | |
d6de8be7 | 3146 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 3147 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 3148 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
3149 | break; |
3150 | } while (1); | |
e2d74ac0 | 3151 | |
4ac845a2 | 3152 | return cic; |
e2d74ac0 JA |
3153 | } |
3154 | ||
4ac845a2 JA |
3155 | /* |
3156 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
3157 | * the process specific cfq io context when entered from the block layer. | |
3158 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
3159 | */ | |
febffd61 JA |
3160 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
3161 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 3162 | { |
0261d688 | 3163 | unsigned long flags; |
4ac845a2 | 3164 | int ret; |
e2d74ac0 | 3165 | |
4ac845a2 JA |
3166 | ret = radix_tree_preload(gfp_mask); |
3167 | if (!ret) { | |
3168 | cic->ioc = ioc; | |
3169 | cic->key = cfqd; | |
e2d74ac0 | 3170 | |
4ac845a2 JA |
3171 | spin_lock_irqsave(&ioc->lock, flags); |
3172 | ret = radix_tree_insert(&ioc->radix_root, | |
80b15c73 | 3173 | cfqd->cic_index, cic); |
ffc4e759 JA |
3174 | if (!ret) |
3175 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 3176 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 3177 | |
4ac845a2 JA |
3178 | radix_tree_preload_end(); |
3179 | ||
3180 | if (!ret) { | |
3181 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
3182 | list_add(&cic->queue_list, &cfqd->cic_list); | |
3183 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
3184 | } | |
e2d74ac0 JA |
3185 | } |
3186 | ||
4ac845a2 JA |
3187 | if (ret) |
3188 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 3189 | |
4ac845a2 | 3190 | return ret; |
e2d74ac0 JA |
3191 | } |
3192 | ||
1da177e4 LT |
3193 | /* |
3194 | * Setup general io context and cfq io context. There can be several cfq | |
3195 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 3196 | * than one device managed by cfq. |
1da177e4 LT |
3197 | */ |
3198 | static struct cfq_io_context * | |
e2d74ac0 | 3199 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 3200 | { |
22e2c507 | 3201 | struct io_context *ioc = NULL; |
1da177e4 | 3202 | struct cfq_io_context *cic; |
1da177e4 | 3203 | |
22e2c507 | 3204 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 3205 | |
b5deef90 | 3206 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
3207 | if (!ioc) |
3208 | return NULL; | |
3209 | ||
4ac845a2 | 3210 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
3211 | if (cic) |
3212 | goto out; | |
1da177e4 | 3213 | |
e2d74ac0 JA |
3214 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
3215 | if (cic == NULL) | |
3216 | goto err; | |
1da177e4 | 3217 | |
4ac845a2 JA |
3218 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
3219 | goto err_free; | |
3220 | ||
1da177e4 | 3221 | out: |
fc46379d JA |
3222 | smp_read_barrier_depends(); |
3223 | if (unlikely(ioc->ioprio_changed)) | |
3224 | cfq_ioc_set_ioprio(ioc); | |
3225 | ||
24610333 VG |
3226 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
3227 | if (unlikely(ioc->cgroup_changed)) | |
3228 | cfq_ioc_set_cgroup(ioc); | |
3229 | #endif | |
1da177e4 | 3230 | return cic; |
4ac845a2 JA |
3231 | err_free: |
3232 | cfq_cic_free(cic); | |
1da177e4 LT |
3233 | err: |
3234 | put_io_context(ioc); | |
3235 | return NULL; | |
3236 | } | |
3237 | ||
22e2c507 | 3238 | static void |
383cd721 | 3239 | __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle) |
1da177e4 | 3240 | { |
383cd721 SL |
3241 | unsigned long elapsed = jiffies - ttime->last_end_request; |
3242 | elapsed = min(elapsed, 2UL * slice_idle); | |
db3b5848 | 3243 | |
383cd721 SL |
3244 | ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8; |
3245 | ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8; | |
3246 | ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples; | |
3247 | } | |
3248 | ||
3249 | static void | |
3250 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3251 | struct cfq_io_context *cic) | |
3252 | { | |
f5f2b6ce | 3253 | if (cfq_cfqq_sync(cfqq)) { |
383cd721 | 3254 | __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle); |
f5f2b6ce SL |
3255 | __cfq_update_io_thinktime(&cfqq->service_tree->ttime, |
3256 | cfqd->cfq_slice_idle); | |
3257 | } | |
7700fc4f SL |
3258 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
3259 | __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle); | |
3260 | #endif | |
22e2c507 | 3261 | } |
1da177e4 | 3262 | |
206dc69b | 3263 | static void |
b2c18e1e | 3264 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
6d048f53 | 3265 | struct request *rq) |
206dc69b | 3266 | { |
3dde36dd | 3267 | sector_t sdist = 0; |
41647e7a | 3268 | sector_t n_sec = blk_rq_sectors(rq); |
3dde36dd CZ |
3269 | if (cfqq->last_request_pos) { |
3270 | if (cfqq->last_request_pos < blk_rq_pos(rq)) | |
3271 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
3272 | else | |
3273 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); | |
3274 | } | |
206dc69b | 3275 | |
3dde36dd | 3276 | cfqq->seek_history <<= 1; |
41647e7a CZ |
3277 | if (blk_queue_nonrot(cfqd->queue)) |
3278 | cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT); | |
3279 | else | |
3280 | cfqq->seek_history |= (sdist > CFQQ_SEEK_THR); | |
206dc69b | 3281 | } |
1da177e4 | 3282 | |
22e2c507 JA |
3283 | /* |
3284 | * Disable idle window if the process thinks too long or seeks so much that | |
3285 | * it doesn't matter | |
3286 | */ | |
3287 | static void | |
3288 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3289 | struct cfq_io_context *cic) | |
3290 | { | |
7b679138 | 3291 | int old_idle, enable_idle; |
1be92f2f | 3292 | |
0871714e JA |
3293 | /* |
3294 | * Don't idle for async or idle io prio class | |
3295 | */ | |
3296 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
3297 | return; |
3298 | ||
c265a7f4 | 3299 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 3300 | |
76280aff CZ |
3301 | if (cfqq->queued[0] + cfqq->queued[1] >= 4) |
3302 | cfq_mark_cfqq_deep(cfqq); | |
3303 | ||
749ef9f8 CZ |
3304 | if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE)) |
3305 | enable_idle = 0; | |
3306 | else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || | |
3dde36dd | 3307 | (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq))) |
22e2c507 | 3308 | enable_idle = 0; |
383cd721 SL |
3309 | else if (sample_valid(cic->ttime.ttime_samples)) { |
3310 | if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle) | |
22e2c507 JA |
3311 | enable_idle = 0; |
3312 | else | |
3313 | enable_idle = 1; | |
1da177e4 LT |
3314 | } |
3315 | ||
7b679138 JA |
3316 | if (old_idle != enable_idle) { |
3317 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
3318 | if (enable_idle) | |
3319 | cfq_mark_cfqq_idle_window(cfqq); | |
3320 | else | |
3321 | cfq_clear_cfqq_idle_window(cfqq); | |
3322 | } | |
22e2c507 | 3323 | } |
1da177e4 | 3324 | |
22e2c507 JA |
3325 | /* |
3326 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
3327 | * no or if we aren't sure, a 1 will cause a preempt. | |
3328 | */ | |
a6151c3a | 3329 | static bool |
22e2c507 | 3330 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, |
5e705374 | 3331 | struct request *rq) |
22e2c507 | 3332 | { |
6d048f53 | 3333 | struct cfq_queue *cfqq; |
22e2c507 | 3334 | |
6d048f53 JA |
3335 | cfqq = cfqd->active_queue; |
3336 | if (!cfqq) | |
a6151c3a | 3337 | return false; |
22e2c507 | 3338 | |
6d048f53 | 3339 | if (cfq_class_idle(new_cfqq)) |
a6151c3a | 3340 | return false; |
22e2c507 JA |
3341 | |
3342 | if (cfq_class_idle(cfqq)) | |
a6151c3a | 3343 | return true; |
1e3335de | 3344 | |
875feb63 DS |
3345 | /* |
3346 | * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice. | |
3347 | */ | |
3348 | if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq)) | |
3349 | return false; | |
3350 | ||
374f84ac JA |
3351 | /* |
3352 | * if the new request is sync, but the currently running queue is | |
3353 | * not, let the sync request have priority. | |
3354 | */ | |
5e705374 | 3355 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
a6151c3a | 3356 | return true; |
1e3335de | 3357 | |
8682e1f1 VG |
3358 | if (new_cfqq->cfqg != cfqq->cfqg) |
3359 | return false; | |
3360 | ||
3361 | if (cfq_slice_used(cfqq)) | |
3362 | return true; | |
3363 | ||
3364 | /* Allow preemption only if we are idling on sync-noidle tree */ | |
3365 | if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && | |
3366 | cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && | |
3367 | new_cfqq->service_tree->count == 2 && | |
3368 | RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3369 | return true; | |
3370 | ||
b53d1ed7 JA |
3371 | /* |
3372 | * So both queues are sync. Let the new request get disk time if | |
3373 | * it's a metadata request and the current queue is doing regular IO. | |
3374 | */ | |
65299a3b | 3375 | if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending) |
b53d1ed7 JA |
3376 | return true; |
3377 | ||
3a9a3f6c DS |
3378 | /* |
3379 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
3380 | */ | |
3381 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
a6151c3a | 3382 | return true; |
3a9a3f6c | 3383 | |
d2d59e18 SL |
3384 | /* An idle queue should not be idle now for some reason */ |
3385 | if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq)) | |
3386 | return true; | |
3387 | ||
1e3335de | 3388 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
a6151c3a | 3389 | return false; |
1e3335de JA |
3390 | |
3391 | /* | |
3392 | * if this request is as-good as one we would expect from the | |
3393 | * current cfqq, let it preempt | |
3394 | */ | |
e9ce335d | 3395 | if (cfq_rq_close(cfqd, cfqq, rq)) |
a6151c3a | 3396 | return true; |
1e3335de | 3397 | |
a6151c3a | 3398 | return false; |
22e2c507 JA |
3399 | } |
3400 | ||
3401 | /* | |
3402 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
3403 | * let it have half of its nominal slice. | |
3404 | */ | |
3405 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3406 | { | |
f8ae6e3e SL |
3407 | struct cfq_queue *old_cfqq = cfqd->active_queue; |
3408 | ||
7b679138 | 3409 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
e5ff082e | 3410 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 3411 | |
f8ae6e3e SL |
3412 | /* |
3413 | * workload type is changed, don't save slice, otherwise preempt | |
3414 | * doesn't happen | |
3415 | */ | |
3416 | if (cfqq_type(old_cfqq) != cfqq_type(cfqq)) | |
3417 | cfqq->cfqg->saved_workload_slice = 0; | |
3418 | ||
bf572256 JA |
3419 | /* |
3420 | * Put the new queue at the front of the of the current list, | |
3421 | * so we know that it will be selected next. | |
3422 | */ | |
3423 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
3424 | |
3425 | cfq_service_tree_add(cfqd, cfqq, 1); | |
eda5e0c9 | 3426 | |
62a37f6b JT |
3427 | cfqq->slice_end = 0; |
3428 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
3429 | } |
3430 | ||
22e2c507 | 3431 | /* |
5e705374 | 3432 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
3433 | * something we should do about it |
3434 | */ | |
3435 | static void | |
5e705374 JA |
3436 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
3437 | struct request *rq) | |
22e2c507 | 3438 | { |
5e705374 | 3439 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 3440 | |
45333d5a | 3441 | cfqd->rq_queued++; |
65299a3b CH |
3442 | if (rq->cmd_flags & REQ_PRIO) |
3443 | cfqq->prio_pending++; | |
374f84ac | 3444 | |
383cd721 | 3445 | cfq_update_io_thinktime(cfqd, cfqq, cic); |
b2c18e1e | 3446 | cfq_update_io_seektime(cfqd, cfqq, rq); |
9c2c38a1 JA |
3447 | cfq_update_idle_window(cfqd, cfqq, cic); |
3448 | ||
b2c18e1e | 3449 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
3450 | |
3451 | if (cfqq == cfqd->active_queue) { | |
3452 | /* | |
b029195d JA |
3453 | * Remember that we saw a request from this process, but |
3454 | * don't start queuing just yet. Otherwise we risk seeing lots | |
3455 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
3456 | * and merging. If the request is already larger than a single |
3457 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
3458 | * merging is already done. Ditto for a busy system that |
3459 | * has other work pending, don't risk delaying until the | |
3460 | * idle timer unplug to continue working. | |
22e2c507 | 3461 | */ |
d6ceb25e | 3462 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
3463 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
3464 | cfqd->busy_queues > 1) { | |
812df48d | 3465 | cfq_del_timer(cfqd, cfqq); |
554554f6 | 3466 | cfq_clear_cfqq_wait_request(cfqq); |
24ecfbe2 | 3467 | __blk_run_queue(cfqd->queue); |
a11cdaa7 | 3468 | } else { |
e98ef89b | 3469 | cfq_blkiocg_update_idle_time_stats( |
a11cdaa7 | 3470 | &cfqq->cfqg->blkg); |
bf791937 | 3471 | cfq_mark_cfqq_must_dispatch(cfqq); |
a11cdaa7 | 3472 | } |
d6ceb25e | 3473 | } |
5e705374 | 3474 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
3475 | /* |
3476 | * not the active queue - expire current slice if it is | |
3477 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
3478 | * has some old slice time left and is of higher priority or |
3479 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
3480 | */ |
3481 | cfq_preempt_queue(cfqd, cfqq); | |
24ecfbe2 | 3482 | __blk_run_queue(cfqd->queue); |
22e2c507 | 3483 | } |
1da177e4 LT |
3484 | } |
3485 | ||
165125e1 | 3486 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 3487 | { |
b4878f24 | 3488 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 3489 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 3490 | |
7b679138 | 3491 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 3492 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 3493 | |
30996f40 | 3494 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); |
22e2c507 | 3495 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
aa6f6a3d | 3496 | cfq_add_rq_rb(rq); |
e98ef89b | 3497 | cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg, |
cdc1184c DS |
3498 | &cfqd->serving_group->blkg, rq_data_dir(rq), |
3499 | rq_is_sync(rq)); | |
5e705374 | 3500 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
3501 | } |
3502 | ||
45333d5a AC |
3503 | /* |
3504 | * Update hw_tag based on peak queue depth over 50 samples under | |
3505 | * sufficient load. | |
3506 | */ | |
3507 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
3508 | { | |
1a1238a7 SL |
3509 | struct cfq_queue *cfqq = cfqd->active_queue; |
3510 | ||
53c583d2 CZ |
3511 | if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth) |
3512 | cfqd->hw_tag_est_depth = cfqd->rq_in_driver; | |
e459dd08 CZ |
3513 | |
3514 | if (cfqd->hw_tag == 1) | |
3515 | return; | |
45333d5a AC |
3516 | |
3517 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
53c583d2 | 3518 | cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
3519 | return; |
3520 | ||
1a1238a7 SL |
3521 | /* |
3522 | * If active queue hasn't enough requests and can idle, cfq might not | |
3523 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
3524 | * case | |
3525 | */ | |
3526 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
3527 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
53c583d2 | 3528 | CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN) |
1a1238a7 SL |
3529 | return; |
3530 | ||
45333d5a AC |
3531 | if (cfqd->hw_tag_samples++ < 50) |
3532 | return; | |
3533 | ||
e459dd08 | 3534 | if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
3535 | cfqd->hw_tag = 1; |
3536 | else | |
3537 | cfqd->hw_tag = 0; | |
45333d5a AC |
3538 | } |
3539 | ||
7667aa06 VG |
3540 | static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
3541 | { | |
3542 | struct cfq_io_context *cic = cfqd->active_cic; | |
3543 | ||
02a8f01b JT |
3544 | /* If the queue already has requests, don't wait */ |
3545 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3546 | return false; | |
3547 | ||
7667aa06 VG |
3548 | /* If there are other queues in the group, don't wait */ |
3549 | if (cfqq->cfqg->nr_cfqq > 1) | |
3550 | return false; | |
3551 | ||
7700fc4f SL |
3552 | /* the only queue in the group, but think time is big */ |
3553 | if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) | |
3554 | return false; | |
3555 | ||
7667aa06 VG |
3556 | if (cfq_slice_used(cfqq)) |
3557 | return true; | |
3558 | ||
3559 | /* if slice left is less than think time, wait busy */ | |
383cd721 SL |
3560 | if (cic && sample_valid(cic->ttime.ttime_samples) |
3561 | && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) | |
7667aa06 VG |
3562 | return true; |
3563 | ||
3564 | /* | |
3565 | * If think times is less than a jiffy than ttime_mean=0 and above | |
3566 | * will not be true. It might happen that slice has not expired yet | |
3567 | * but will expire soon (4-5 ns) during select_queue(). To cover the | |
3568 | * case where think time is less than a jiffy, mark the queue wait | |
3569 | * busy if only 1 jiffy is left in the slice. | |
3570 | */ | |
3571 | if (cfqq->slice_end - jiffies == 1) | |
3572 | return true; | |
3573 | ||
3574 | return false; | |
3575 | } | |
3576 | ||
165125e1 | 3577 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 3578 | { |
5e705374 | 3579 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 3580 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 3581 | const int sync = rq_is_sync(rq); |
b4878f24 | 3582 | unsigned long now; |
1da177e4 | 3583 | |
b4878f24 | 3584 | now = jiffies; |
33659ebb CH |
3585 | cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", |
3586 | !!(rq->cmd_flags & REQ_NOIDLE)); | |
1da177e4 | 3587 | |
45333d5a AC |
3588 | cfq_update_hw_tag(cfqd); |
3589 | ||
53c583d2 | 3590 | WARN_ON(!cfqd->rq_in_driver); |
6d048f53 | 3591 | WARN_ON(!cfqq->dispatched); |
53c583d2 | 3592 | cfqd->rq_in_driver--; |
6d048f53 | 3593 | cfqq->dispatched--; |
80bdf0c7 | 3594 | (RQ_CFQG(rq))->dispatched--; |
e98ef89b VG |
3595 | cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg, |
3596 | rq_start_time_ns(rq), rq_io_start_time_ns(rq), | |
3597 | rq_data_dir(rq), rq_is_sync(rq)); | |
1da177e4 | 3598 | |
53c583d2 | 3599 | cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--; |
3ed9a296 | 3600 | |
365722bb | 3601 | if (sync) { |
f5f2b6ce SL |
3602 | struct cfq_rb_root *service_tree; |
3603 | ||
383cd721 | 3604 | RQ_CIC(rq)->ttime.last_end_request = now; |
f5f2b6ce SL |
3605 | |
3606 | if (cfq_cfqq_on_rr(cfqq)) | |
3607 | service_tree = cfqq->service_tree; | |
3608 | else | |
3609 | service_tree = service_tree_for(cfqq->cfqg, | |
3610 | cfqq_prio(cfqq), cfqq_type(cfqq)); | |
3611 | service_tree->ttime.last_end_request = now; | |
573412b2 CZ |
3612 | if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now)) |
3613 | cfqd->last_delayed_sync = now; | |
365722bb | 3614 | } |
caaa5f9f | 3615 | |
7700fc4f SL |
3616 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
3617 | cfqq->cfqg->ttime.last_end_request = now; | |
3618 | #endif | |
3619 | ||
caaa5f9f JA |
3620 | /* |
3621 | * If this is the active queue, check if it needs to be expired, | |
3622 | * or if we want to idle in case it has no pending requests. | |
3623 | */ | |
3624 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
3625 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
3626 | ||
44f7c160 JA |
3627 | if (cfq_cfqq_slice_new(cfqq)) { |
3628 | cfq_set_prio_slice(cfqd, cfqq); | |
3629 | cfq_clear_cfqq_slice_new(cfqq); | |
3630 | } | |
f75edf2d VG |
3631 | |
3632 | /* | |
7667aa06 VG |
3633 | * Should we wait for next request to come in before we expire |
3634 | * the queue. | |
f75edf2d | 3635 | */ |
7667aa06 | 3636 | if (cfq_should_wait_busy(cfqd, cfqq)) { |
80bdf0c7 VG |
3637 | unsigned long extend_sl = cfqd->cfq_slice_idle; |
3638 | if (!cfqd->cfq_slice_idle) | |
3639 | extend_sl = cfqd->cfq_group_idle; | |
3640 | cfqq->slice_end = jiffies + extend_sl; | |
f75edf2d | 3641 | cfq_mark_cfqq_wait_busy(cfqq); |
b1ffe737 | 3642 | cfq_log_cfqq(cfqd, cfqq, "will busy wait"); |
f75edf2d VG |
3643 | } |
3644 | ||
a36e71f9 | 3645 | /* |
8e550632 CZ |
3646 | * Idling is not enabled on: |
3647 | * - expired queues | |
3648 | * - idle-priority queues | |
3649 | * - async queues | |
3650 | * - queues with still some requests queued | |
3651 | * - when there is a close cooperator | |
a36e71f9 | 3652 | */ |
0871714e | 3653 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
e5ff082e | 3654 | cfq_slice_expired(cfqd, 1); |
8e550632 CZ |
3655 | else if (sync && cfqq_empty && |
3656 | !cfq_close_cooperator(cfqd, cfqq)) { | |
749ef9f8 | 3657 | cfq_arm_slice_timer(cfqd); |
8e550632 | 3658 | } |
caaa5f9f | 3659 | } |
6d048f53 | 3660 | |
53c583d2 | 3661 | if (!cfqd->rq_in_driver) |
23e018a1 | 3662 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
3663 | } |
3664 | ||
89850f7e | 3665 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 3666 | { |
1b379d8d | 3667 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 3668 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 3669 | return ELV_MQUEUE_MUST; |
3b18152c | 3670 | } |
1da177e4 | 3671 | |
22e2c507 | 3672 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
3673 | } |
3674 | ||
165125e1 | 3675 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
3676 | { |
3677 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3678 | struct task_struct *tsk = current; | |
91fac317 | 3679 | struct cfq_io_context *cic; |
22e2c507 JA |
3680 | struct cfq_queue *cfqq; |
3681 | ||
3682 | /* | |
3683 | * don't force setup of a queue from here, as a call to may_queue | |
3684 | * does not necessarily imply that a request actually will be queued. | |
3685 | * so just lookup a possibly existing queue, or return 'may queue' | |
3686 | * if that fails | |
3687 | */ | |
4ac845a2 | 3688 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
3689 | if (!cic) |
3690 | return ELV_MQUEUE_MAY; | |
3691 | ||
b0b78f81 | 3692 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 3693 | if (cfqq) { |
fd0928df | 3694 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 | 3695 | |
89850f7e | 3696 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
3697 | } |
3698 | ||
3699 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
3700 | } |
3701 | ||
1da177e4 LT |
3702 | /* |
3703 | * queue lock held here | |
3704 | */ | |
bb37b94c | 3705 | static void cfq_put_request(struct request *rq) |
1da177e4 | 3706 | { |
5e705374 | 3707 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 3708 | |
5e705374 | 3709 | if (cfqq) { |
22e2c507 | 3710 | const int rw = rq_data_dir(rq); |
1da177e4 | 3711 | |
22e2c507 JA |
3712 | BUG_ON(!cfqq->allocated[rw]); |
3713 | cfqq->allocated[rw]--; | |
1da177e4 | 3714 | |
5e705374 | 3715 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 3716 | |
c186794d MS |
3717 | rq->elevator_private[0] = NULL; |
3718 | rq->elevator_private[1] = NULL; | |
1da177e4 | 3719 | |
7f1dc8a2 VG |
3720 | /* Put down rq reference on cfqg */ |
3721 | cfq_put_cfqg(RQ_CFQG(rq)); | |
c186794d | 3722 | rq->elevator_private[2] = NULL; |
7f1dc8a2 | 3723 | |
1da177e4 LT |
3724 | cfq_put_queue(cfqq); |
3725 | } | |
3726 | } | |
3727 | ||
df5fe3e8 JM |
3728 | static struct cfq_queue * |
3729 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, | |
3730 | struct cfq_queue *cfqq) | |
3731 | { | |
3732 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
3733 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
b3b6d040 | 3734 | cfq_mark_cfqq_coop(cfqq->new_cfqq); |
df5fe3e8 JM |
3735 | cfq_put_queue(cfqq); |
3736 | return cic_to_cfqq(cic, 1); | |
3737 | } | |
3738 | ||
e6c5bc73 JM |
3739 | /* |
3740 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
3741 | * was the last process referring to said cfqq. | |
3742 | */ | |
3743 | static struct cfq_queue * | |
3744 | split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) | |
3745 | { | |
3746 | if (cfqq_process_refs(cfqq) == 1) { | |
e6c5bc73 JM |
3747 | cfqq->pid = current->pid; |
3748 | cfq_clear_cfqq_coop(cfqq); | |
ae54abed | 3749 | cfq_clear_cfqq_split_coop(cfqq); |
e6c5bc73 JM |
3750 | return cfqq; |
3751 | } | |
3752 | ||
3753 | cic_set_cfqq(cic, NULL, 1); | |
d02a2c07 SL |
3754 | |
3755 | cfq_put_cooperator(cfqq); | |
3756 | ||
e6c5bc73 JM |
3757 | cfq_put_queue(cfqq); |
3758 | return NULL; | |
3759 | } | |
1da177e4 | 3760 | /* |
22e2c507 | 3761 | * Allocate cfq data structures associated with this request. |
1da177e4 | 3762 | */ |
22e2c507 | 3763 | static int |
165125e1 | 3764 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
3765 | { |
3766 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3767 | struct cfq_io_context *cic; | |
3768 | const int rw = rq_data_dir(rq); | |
a6151c3a | 3769 | const bool is_sync = rq_is_sync(rq); |
22e2c507 | 3770 | struct cfq_queue *cfqq; |
1da177e4 LT |
3771 | unsigned long flags; |
3772 | ||
3773 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
3774 | ||
e2d74ac0 | 3775 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 3776 | |
1da177e4 LT |
3777 | spin_lock_irqsave(q->queue_lock, flags); |
3778 | ||
22e2c507 JA |
3779 | if (!cic) |
3780 | goto queue_fail; | |
3781 | ||
e6c5bc73 | 3782 | new_queue: |
91fac317 | 3783 | cfqq = cic_to_cfqq(cic, is_sync); |
32f2e807 | 3784 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { |
fd0928df | 3785 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 | 3786 | cic_set_cfqq(cic, cfqq, is_sync); |
df5fe3e8 | 3787 | } else { |
e6c5bc73 JM |
3788 | /* |
3789 | * If the queue was seeky for too long, break it apart. | |
3790 | */ | |
ae54abed | 3791 | if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) { |
e6c5bc73 JM |
3792 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); |
3793 | cfqq = split_cfqq(cic, cfqq); | |
3794 | if (!cfqq) | |
3795 | goto new_queue; | |
3796 | } | |
3797 | ||
df5fe3e8 JM |
3798 | /* |
3799 | * Check to see if this queue is scheduled to merge with | |
3800 | * another, closely cooperating queue. The merging of | |
3801 | * queues happens here as it must be done in process context. | |
3802 | * The reference on new_cfqq was taken in merge_cfqqs. | |
3803 | */ | |
3804 | if (cfqq->new_cfqq) | |
3805 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
91fac317 | 3806 | } |
1da177e4 LT |
3807 | |
3808 | cfqq->allocated[rw]++; | |
1da177e4 | 3809 | |
6fae9c25 | 3810 | cfqq->ref++; |
c186794d MS |
3811 | rq->elevator_private[0] = cic; |
3812 | rq->elevator_private[1] = cfqq; | |
3813 | rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg); | |
93803e01 | 3814 | spin_unlock_irqrestore(q->queue_lock, flags); |
5e705374 | 3815 | return 0; |
1da177e4 | 3816 | |
22e2c507 | 3817 | queue_fail: |
23e018a1 | 3818 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 3819 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 3820 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
3821 | return 1; |
3822 | } | |
3823 | ||
65f27f38 | 3824 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 3825 | { |
65f27f38 | 3826 | struct cfq_data *cfqd = |
23e018a1 | 3827 | container_of(work, struct cfq_data, unplug_work); |
165125e1 | 3828 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3829 | |
40bb54d1 | 3830 | spin_lock_irq(q->queue_lock); |
24ecfbe2 | 3831 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 3832 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
3833 | } |
3834 | ||
3835 | /* | |
3836 | * Timer running if the active_queue is currently idling inside its time slice | |
3837 | */ | |
3838 | static void cfq_idle_slice_timer(unsigned long data) | |
3839 | { | |
3840 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
3841 | struct cfq_queue *cfqq; | |
3842 | unsigned long flags; | |
3c6bd2f8 | 3843 | int timed_out = 1; |
22e2c507 | 3844 | |
7b679138 JA |
3845 | cfq_log(cfqd, "idle timer fired"); |
3846 | ||
22e2c507 JA |
3847 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
3848 | ||
fe094d98 JA |
3849 | cfqq = cfqd->active_queue; |
3850 | if (cfqq) { | |
3c6bd2f8 JA |
3851 | timed_out = 0; |
3852 | ||
b029195d JA |
3853 | /* |
3854 | * We saw a request before the queue expired, let it through | |
3855 | */ | |
3856 | if (cfq_cfqq_must_dispatch(cfqq)) | |
3857 | goto out_kick; | |
3858 | ||
22e2c507 JA |
3859 | /* |
3860 | * expired | |
3861 | */ | |
44f7c160 | 3862 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
3863 | goto expire; |
3864 | ||
3865 | /* | |
3866 | * only expire and reinvoke request handler, if there are | |
3867 | * other queues with pending requests | |
3868 | */ | |
caaa5f9f | 3869 | if (!cfqd->busy_queues) |
22e2c507 | 3870 | goto out_cont; |
22e2c507 JA |
3871 | |
3872 | /* | |
3873 | * not expired and it has a request pending, let it dispatch | |
3874 | */ | |
75e50984 | 3875 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 3876 | goto out_kick; |
76280aff CZ |
3877 | |
3878 | /* | |
3879 | * Queue depth flag is reset only when the idle didn't succeed | |
3880 | */ | |
3881 | cfq_clear_cfqq_deep(cfqq); | |
22e2c507 JA |
3882 | } |
3883 | expire: | |
e5ff082e | 3884 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 3885 | out_kick: |
23e018a1 | 3886 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
3887 | out_cont: |
3888 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
3889 | } | |
3890 | ||
3b18152c JA |
3891 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
3892 | { | |
3893 | del_timer_sync(&cfqd->idle_slice_timer); | |
23e018a1 | 3894 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 3895 | } |
22e2c507 | 3896 | |
c2dea2d1 VT |
3897 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
3898 | { | |
3899 | int i; | |
3900 | ||
3901 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
3902 | if (cfqd->async_cfqq[0][i]) | |
3903 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
3904 | if (cfqd->async_cfqq[1][i]) | |
3905 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 3906 | } |
2389d1ef ON |
3907 | |
3908 | if (cfqd->async_idle_cfqq) | |
3909 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
3910 | } |
3911 | ||
b374d18a | 3912 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 3913 | { |
22e2c507 | 3914 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 3915 | struct request_queue *q = cfqd->queue; |
56edf7d7 | 3916 | bool wait = false; |
22e2c507 | 3917 | |
3b18152c | 3918 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 3919 | |
d9ff4187 | 3920 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 3921 | |
d9ff4187 | 3922 | if (cfqd->active_queue) |
e5ff082e | 3923 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
3924 | |
3925 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
3926 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
3927 | struct cfq_io_context, | |
3928 | queue_list); | |
89850f7e JA |
3929 | |
3930 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 3931 | } |
e2d74ac0 | 3932 | |
c2dea2d1 | 3933 | cfq_put_async_queues(cfqd); |
b1c35769 | 3934 | cfq_release_cfq_groups(cfqd); |
56edf7d7 VG |
3935 | |
3936 | /* | |
3937 | * If there are groups which we could not unlink from blkcg list, | |
3938 | * wait for a rcu period for them to be freed. | |
3939 | */ | |
3940 | if (cfqd->nr_blkcg_linked_grps) | |
3941 | wait = true; | |
15c31be4 | 3942 | |
d9ff4187 | 3943 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
3944 | |
3945 | cfq_shutdown_timer_wq(cfqd); | |
3946 | ||
80b15c73 KK |
3947 | spin_lock(&cic_index_lock); |
3948 | ida_remove(&cic_index_ida, cfqd->cic_index); | |
3949 | spin_unlock(&cic_index_lock); | |
3950 | ||
56edf7d7 VG |
3951 | /* |
3952 | * Wait for cfqg->blkg->key accessors to exit their grace periods. | |
3953 | * Do this wait only if there are other unlinked groups out | |
3954 | * there. This can happen if cgroup deletion path claimed the | |
3955 | * responsibility of cleaning up a group before queue cleanup code | |
3956 | * get to the group. | |
3957 | * | |
3958 | * Do not call synchronize_rcu() unconditionally as there are drivers | |
3959 | * which create/delete request queue hundreds of times during scan/boot | |
3960 | * and synchronize_rcu() can take significant time and slow down boot. | |
3961 | */ | |
3962 | if (wait) | |
3963 | synchronize_rcu(); | |
2abae55f VG |
3964 | |
3965 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
3966 | /* Free up per cpu stats for root group */ | |
3967 | free_percpu(cfqd->root_group.blkg.stats_cpu); | |
3968 | #endif | |
56edf7d7 | 3969 | kfree(cfqd); |
1da177e4 LT |
3970 | } |
3971 | ||
80b15c73 KK |
3972 | static int cfq_alloc_cic_index(void) |
3973 | { | |
3974 | int index, error; | |
3975 | ||
3976 | do { | |
3977 | if (!ida_pre_get(&cic_index_ida, GFP_KERNEL)) | |
3978 | return -ENOMEM; | |
3979 | ||
3980 | spin_lock(&cic_index_lock); | |
3981 | error = ida_get_new(&cic_index_ida, &index); | |
3982 | spin_unlock(&cic_index_lock); | |
3983 | if (error && error != -EAGAIN) | |
3984 | return error; | |
3985 | } while (error); | |
3986 | ||
3987 | return index; | |
3988 | } | |
3989 | ||
165125e1 | 3990 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
3991 | { |
3992 | struct cfq_data *cfqd; | |
718eee05 | 3993 | int i, j; |
cdb16e8f | 3994 | struct cfq_group *cfqg; |
615f0259 | 3995 | struct cfq_rb_root *st; |
1da177e4 | 3996 | |
80b15c73 KK |
3997 | i = cfq_alloc_cic_index(); |
3998 | if (i < 0) | |
3999 | return NULL; | |
4000 | ||
94f6030c | 4001 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1547010e NK |
4002 | if (!cfqd) { |
4003 | spin_lock(&cic_index_lock); | |
4004 | ida_remove(&cic_index_ida, i); | |
4005 | spin_unlock(&cic_index_lock); | |
bc1c1169 | 4006 | return NULL; |
1547010e | 4007 | } |
1da177e4 | 4008 | |
30d7b944 SL |
4009 | /* |
4010 | * Don't need take queue_lock in the routine, since we are | |
4011 | * initializing the ioscheduler, and nobody is using cfqd | |
4012 | */ | |
80b15c73 KK |
4013 | cfqd->cic_index = i; |
4014 | ||
1fa8f6d6 VG |
4015 | /* Init root service tree */ |
4016 | cfqd->grp_service_tree = CFQ_RB_ROOT; | |
4017 | ||
cdb16e8f VG |
4018 | /* Init root group */ |
4019 | cfqg = &cfqd->root_group; | |
615f0259 VG |
4020 | for_each_cfqg_st(cfqg, i, j, st) |
4021 | *st = CFQ_RB_ROOT; | |
1fa8f6d6 | 4022 | RB_CLEAR_NODE(&cfqg->rb_node); |
26a2ac00 | 4023 | |
25bc6b07 VG |
4024 | /* Give preference to root group over other groups */ |
4025 | cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT; | |
4026 | ||
25fb5169 | 4027 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
b1c35769 | 4028 | /* |
56edf7d7 VG |
4029 | * Set root group reference to 2. One reference will be dropped when |
4030 | * all groups on cfqd->cfqg_list are being deleted during queue exit. | |
4031 | * Other reference will remain there as we don't want to delete this | |
4032 | * group as it is statically allocated and gets destroyed when | |
4033 | * throtl_data goes away. | |
b1c35769 | 4034 | */ |
56edf7d7 | 4035 | cfqg->ref = 2; |
5624a4e4 VG |
4036 | |
4037 | if (blkio_alloc_blkg_stats(&cfqg->blkg)) { | |
4038 | kfree(cfqg); | |
4039 | kfree(cfqd); | |
4040 | return NULL; | |
4041 | } | |
4042 | ||
dcf097b2 | 4043 | rcu_read_lock(); |
5624a4e4 | 4044 | |
e98ef89b VG |
4045 | cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, |
4046 | (void *)cfqd, 0); | |
dcf097b2 | 4047 | rcu_read_unlock(); |
56edf7d7 VG |
4048 | cfqd->nr_blkcg_linked_grps++; |
4049 | ||
4050 | /* Add group on cfqd->cfqg_list */ | |
4051 | hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); | |
25fb5169 | 4052 | #endif |
26a2ac00 JA |
4053 | /* |
4054 | * Not strictly needed (since RB_ROOT just clears the node and we | |
4055 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
4056 | * to add magic to the rb code | |
4057 | */ | |
4058 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
4059 | cfqd->prio_trees[i] = RB_ROOT; | |
4060 | ||
6118b70b JA |
4061 | /* |
4062 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
4063 | * Grab a permanent reference to it, so that the normal code flow | |
4064 | * will not attempt to free it. | |
4065 | */ | |
4066 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
30d7b944 | 4067 | cfqd->oom_cfqq.ref++; |
cdb16e8f | 4068 | cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group); |
6118b70b | 4069 | |
d9ff4187 | 4070 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 4071 | |
1da177e4 | 4072 | cfqd->queue = q; |
1da177e4 | 4073 | |
22e2c507 JA |
4074 | init_timer(&cfqd->idle_slice_timer); |
4075 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
4076 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
4077 | ||
23e018a1 | 4078 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 4079 | |
1da177e4 | 4080 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
4081 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
4082 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
4083 | cfqd->cfq_back_max = cfq_back_max; |
4084 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
4085 | cfqd->cfq_slice[0] = cfq_slice_async; |
4086 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
4087 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
4088 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
80bdf0c7 | 4089 | cfqd->cfq_group_idle = cfq_group_idle; |
963b72fc | 4090 | cfqd->cfq_latency = 1; |
e459dd08 | 4091 | cfqd->hw_tag = -1; |
edc71131 CZ |
4092 | /* |
4093 | * we optimistically start assuming sync ops weren't delayed in last | |
4094 | * second, in order to have larger depth for async operations. | |
4095 | */ | |
573412b2 | 4096 | cfqd->last_delayed_sync = jiffies - HZ; |
bc1c1169 | 4097 | return cfqd; |
1da177e4 LT |
4098 | } |
4099 | ||
4100 | static void cfq_slab_kill(void) | |
4101 | { | |
d6de8be7 JA |
4102 | /* |
4103 | * Caller already ensured that pending RCU callbacks are completed, | |
4104 | * so we should have no busy allocations at this point. | |
4105 | */ | |
1da177e4 LT |
4106 | if (cfq_pool) |
4107 | kmem_cache_destroy(cfq_pool); | |
4108 | if (cfq_ioc_pool) | |
4109 | kmem_cache_destroy(cfq_ioc_pool); | |
4110 | } | |
4111 | ||
4112 | static int __init cfq_slab_setup(void) | |
4113 | { | |
0a31bd5f | 4114 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
4115 | if (!cfq_pool) |
4116 | goto fail; | |
4117 | ||
34e6bbf2 | 4118 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
4119 | if (!cfq_ioc_pool) |
4120 | goto fail; | |
4121 | ||
4122 | return 0; | |
4123 | fail: | |
4124 | cfq_slab_kill(); | |
4125 | return -ENOMEM; | |
4126 | } | |
4127 | ||
1da177e4 LT |
4128 | /* |
4129 | * sysfs parts below --> | |
4130 | */ | |
1da177e4 LT |
4131 | static ssize_t |
4132 | cfq_var_show(unsigned int var, char *page) | |
4133 | { | |
4134 | return sprintf(page, "%d\n", var); | |
4135 | } | |
4136 | ||
4137 | static ssize_t | |
4138 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
4139 | { | |
4140 | char *p = (char *) page; | |
4141 | ||
4142 | *var = simple_strtoul(p, &p, 10); | |
4143 | return count; | |
4144 | } | |
4145 | ||
1da177e4 | 4146 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 4147 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 4148 | { \ |
3d1ab40f | 4149 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
4150 | unsigned int __data = __VAR; \ |
4151 | if (__CONV) \ | |
4152 | __data = jiffies_to_msecs(__data); \ | |
4153 | return cfq_var_show(__data, (page)); \ | |
4154 | } | |
4155 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
4156 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
4157 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
4158 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
4159 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 | 4160 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
80bdf0c7 | 4161 | SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1); |
22e2c507 JA |
4162 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); |
4163 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
4164 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
963b72fc | 4165 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); |
1da177e4 LT |
4166 | #undef SHOW_FUNCTION |
4167 | ||
4168 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 4169 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 4170 | { \ |
3d1ab40f | 4171 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
4172 | unsigned int __data; \ |
4173 | int ret = cfq_var_store(&__data, (page), count); \ | |
4174 | if (__data < (MIN)) \ | |
4175 | __data = (MIN); \ | |
4176 | else if (__data > (MAX)) \ | |
4177 | __data = (MAX); \ | |
4178 | if (__CONV) \ | |
4179 | *(__PTR) = msecs_to_jiffies(__data); \ | |
4180 | else \ | |
4181 | *(__PTR) = __data; \ | |
4182 | return ret; \ | |
4183 | } | |
4184 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
4185 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
4186 | UINT_MAX, 1); | |
4187 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
4188 | UINT_MAX, 1); | |
e572ec7e | 4189 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
4190 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
4191 | UINT_MAX, 0); | |
22e2c507 | 4192 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
80bdf0c7 | 4193 | STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1); |
22e2c507 JA |
4194 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); |
4195 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
4196 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
4197 | UINT_MAX, 0); | |
963b72fc | 4198 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); |
1da177e4 LT |
4199 | #undef STORE_FUNCTION |
4200 | ||
e572ec7e AV |
4201 | #define CFQ_ATTR(name) \ |
4202 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
4203 | ||
4204 | static struct elv_fs_entry cfq_attrs[] = { | |
4205 | CFQ_ATTR(quantum), | |
e572ec7e AV |
4206 | CFQ_ATTR(fifo_expire_sync), |
4207 | CFQ_ATTR(fifo_expire_async), | |
4208 | CFQ_ATTR(back_seek_max), | |
4209 | CFQ_ATTR(back_seek_penalty), | |
4210 | CFQ_ATTR(slice_sync), | |
4211 | CFQ_ATTR(slice_async), | |
4212 | CFQ_ATTR(slice_async_rq), | |
4213 | CFQ_ATTR(slice_idle), | |
80bdf0c7 | 4214 | CFQ_ATTR(group_idle), |
963b72fc | 4215 | CFQ_ATTR(low_latency), |
e572ec7e | 4216 | __ATTR_NULL |
1da177e4 LT |
4217 | }; |
4218 | ||
1da177e4 LT |
4219 | static struct elevator_type iosched_cfq = { |
4220 | .ops = { | |
4221 | .elevator_merge_fn = cfq_merge, | |
4222 | .elevator_merged_fn = cfq_merged_request, | |
4223 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 4224 | .elevator_allow_merge_fn = cfq_allow_merge, |
812d4026 | 4225 | .elevator_bio_merged_fn = cfq_bio_merged, |
b4878f24 | 4226 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 4227 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 4228 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 | 4229 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
1da177e4 | 4230 | .elevator_completed_req_fn = cfq_completed_request, |
21183b07 JA |
4231 | .elevator_former_req_fn = elv_rb_former_request, |
4232 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
4233 | .elevator_set_req_fn = cfq_set_request, |
4234 | .elevator_put_req_fn = cfq_put_request, | |
4235 | .elevator_may_queue_fn = cfq_may_queue, | |
4236 | .elevator_init_fn = cfq_init_queue, | |
4237 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 4238 | .trim = cfq_free_io_context, |
1da177e4 | 4239 | }, |
3d1ab40f | 4240 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
4241 | .elevator_name = "cfq", |
4242 | .elevator_owner = THIS_MODULE, | |
4243 | }; | |
4244 | ||
3e252066 VG |
4245 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
4246 | static struct blkio_policy_type blkio_policy_cfq = { | |
4247 | .ops = { | |
4248 | .blkio_unlink_group_fn = cfq_unlink_blkio_group, | |
4249 | .blkio_update_group_weight_fn = cfq_update_blkio_group_weight, | |
4250 | }, | |
062a644d | 4251 | .plid = BLKIO_POLICY_PROP, |
3e252066 VG |
4252 | }; |
4253 | #else | |
4254 | static struct blkio_policy_type blkio_policy_cfq; | |
4255 | #endif | |
4256 | ||
1da177e4 LT |
4257 | static int __init cfq_init(void) |
4258 | { | |
22e2c507 JA |
4259 | /* |
4260 | * could be 0 on HZ < 1000 setups | |
4261 | */ | |
4262 | if (!cfq_slice_async) | |
4263 | cfq_slice_async = 1; | |
4264 | if (!cfq_slice_idle) | |
4265 | cfq_slice_idle = 1; | |
4266 | ||
80bdf0c7 VG |
4267 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
4268 | if (!cfq_group_idle) | |
4269 | cfq_group_idle = 1; | |
4270 | #else | |
4271 | cfq_group_idle = 0; | |
4272 | #endif | |
1da177e4 LT |
4273 | if (cfq_slab_setup()) |
4274 | return -ENOMEM; | |
4275 | ||
2fdd82bd | 4276 | elv_register(&iosched_cfq); |
3e252066 | 4277 | blkio_policy_register(&blkio_policy_cfq); |
1da177e4 | 4278 | |
2fdd82bd | 4279 | return 0; |
1da177e4 LT |
4280 | } |
4281 | ||
4282 | static void __exit cfq_exit(void) | |
4283 | { | |
6e9a4738 | 4284 | DECLARE_COMPLETION_ONSTACK(all_gone); |
3e252066 | 4285 | blkio_policy_unregister(&blkio_policy_cfq); |
1da177e4 | 4286 | elv_unregister(&iosched_cfq); |
334e94de | 4287 | ioc_gone = &all_gone; |
fba82272 OH |
4288 | /* ioc_gone's update must be visible before reading ioc_count */ |
4289 | smp_wmb(); | |
d6de8be7 JA |
4290 | |
4291 | /* | |
4292 | * this also protects us from entering cfq_slab_kill() with | |
4293 | * pending RCU callbacks | |
4294 | */ | |
245b2e70 | 4295 | if (elv_ioc_count_read(cfq_ioc_count)) |
9a11b4ed | 4296 | wait_for_completion(&all_gone); |
80b15c73 | 4297 | ida_destroy(&cic_index_ida); |
83521d3e | 4298 | cfq_slab_kill(); |
1da177e4 LT |
4299 | } |
4300 | ||
4301 | module_init(cfq_init); | |
4302 | module_exit(cfq_exit); | |
4303 | ||
4304 | MODULE_AUTHOR("Jens Axboe"); | |
4305 | MODULE_LICENSE("GPL"); | |
4306 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |