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; |
cc09e299 | 90 | }; |
73e9ffdd RK |
91 | #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \ |
92 | .count = 0, .min_vdisktime = 0, } | |
cc09e299 | 93 | |
6118b70b JA |
94 | /* |
95 | * Per process-grouping structure | |
96 | */ | |
97 | struct cfq_queue { | |
98 | /* reference count */ | |
30d7b944 | 99 | int ref; |
6118b70b JA |
100 | /* various state flags, see below */ |
101 | unsigned int flags; | |
102 | /* parent cfq_data */ | |
103 | struct cfq_data *cfqd; | |
104 | /* service_tree member */ | |
105 | struct rb_node rb_node; | |
106 | /* service_tree key */ | |
107 | unsigned long rb_key; | |
108 | /* prio tree member */ | |
109 | struct rb_node p_node; | |
110 | /* prio tree root we belong to, if any */ | |
111 | struct rb_root *p_root; | |
112 | /* sorted list of pending requests */ | |
113 | struct rb_root sort_list; | |
114 | /* if fifo isn't expired, next request to serve */ | |
115 | struct request *next_rq; | |
116 | /* requests queued in sort_list */ | |
117 | int queued[2]; | |
118 | /* currently allocated requests */ | |
119 | int allocated[2]; | |
120 | /* fifo list of requests in sort_list */ | |
121 | struct list_head fifo; | |
122 | ||
dae739eb VG |
123 | /* time when queue got scheduled in to dispatch first request. */ |
124 | unsigned long dispatch_start; | |
f75edf2d | 125 | unsigned int allocated_slice; |
c4081ba5 | 126 | unsigned int slice_dispatch; |
dae739eb VG |
127 | /* time when first request from queue completed and slice started. */ |
128 | unsigned long slice_start; | |
6118b70b JA |
129 | unsigned long slice_end; |
130 | long slice_resid; | |
6118b70b JA |
131 | |
132 | /* pending metadata requests */ | |
133 | int meta_pending; | |
134 | /* number of requests that are on the dispatch list or inside driver */ | |
135 | int dispatched; | |
136 | ||
137 | /* io prio of this group */ | |
138 | unsigned short ioprio, org_ioprio; | |
139 | unsigned short ioprio_class, org_ioprio_class; | |
140 | ||
c4081ba5 RK |
141 | pid_t pid; |
142 | ||
3dde36dd | 143 | u32 seek_history; |
b2c18e1e JM |
144 | sector_t last_request_pos; |
145 | ||
aa6f6a3d | 146 | struct cfq_rb_root *service_tree; |
df5fe3e8 | 147 | struct cfq_queue *new_cfqq; |
cdb16e8f | 148 | struct cfq_group *cfqg; |
c4e7893e VG |
149 | /* Number of sectors dispatched from queue in single dispatch round */ |
150 | unsigned long nr_sectors; | |
6118b70b JA |
151 | }; |
152 | ||
c0324a02 | 153 | /* |
718eee05 | 154 | * First index in the service_trees. |
c0324a02 CZ |
155 | * IDLE is handled separately, so it has negative index |
156 | */ | |
157 | enum wl_prio_t { | |
c0324a02 | 158 | BE_WORKLOAD = 0, |
615f0259 VG |
159 | RT_WORKLOAD = 1, |
160 | IDLE_WORKLOAD = 2, | |
b4627321 | 161 | CFQ_PRIO_NR, |
c0324a02 CZ |
162 | }; |
163 | ||
718eee05 CZ |
164 | /* |
165 | * Second index in the service_trees. | |
166 | */ | |
167 | enum wl_type_t { | |
168 | ASYNC_WORKLOAD = 0, | |
169 | SYNC_NOIDLE_WORKLOAD = 1, | |
170 | SYNC_WORKLOAD = 2 | |
171 | }; | |
172 | ||
cdb16e8f VG |
173 | /* This is per cgroup per device grouping structure */ |
174 | struct cfq_group { | |
1fa8f6d6 VG |
175 | /* group service_tree member */ |
176 | struct rb_node rb_node; | |
177 | ||
178 | /* group service_tree key */ | |
179 | u64 vdisktime; | |
25bc6b07 | 180 | unsigned int weight; |
8184f93e JT |
181 | unsigned int new_weight; |
182 | bool needs_update; | |
1fa8f6d6 VG |
183 | |
184 | /* number of cfqq currently on this group */ | |
185 | int nr_cfqq; | |
186 | ||
cdb16e8f | 187 | /* |
b4627321 VG |
188 | * Per group busy queus average. Useful for workload slice calc. We |
189 | * create the array for each prio class but at run time it is used | |
190 | * only for RT and BE class and slot for IDLE class remains unused. | |
191 | * This is primarily done to avoid confusion and a gcc warning. | |
192 | */ | |
193 | unsigned int busy_queues_avg[CFQ_PRIO_NR]; | |
194 | /* | |
195 | * rr lists of queues with requests. We maintain service trees for | |
196 | * RT and BE classes. These trees are subdivided in subclasses | |
197 | * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE | |
198 | * class there is no subclassification and all the cfq queues go on | |
199 | * a single tree service_tree_idle. | |
cdb16e8f VG |
200 | * Counts are embedded in the cfq_rb_root |
201 | */ | |
202 | struct cfq_rb_root service_trees[2][3]; | |
203 | struct cfq_rb_root service_tree_idle; | |
dae739eb VG |
204 | |
205 | unsigned long saved_workload_slice; | |
206 | enum wl_type_t saved_workload; | |
207 | enum wl_prio_t saved_serving_prio; | |
25fb5169 VG |
208 | struct blkio_group blkg; |
209 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
210 | struct hlist_node cfqd_node; | |
329a6781 | 211 | int ref; |
25fb5169 | 212 | #endif |
80bdf0c7 VG |
213 | /* number of requests that are on the dispatch list or inside driver */ |
214 | int dispatched; | |
cdb16e8f | 215 | }; |
718eee05 | 216 | |
22e2c507 JA |
217 | /* |
218 | * Per block device queue structure | |
219 | */ | |
1da177e4 | 220 | struct cfq_data { |
165125e1 | 221 | struct request_queue *queue; |
1fa8f6d6 VG |
222 | /* Root service tree for cfq_groups */ |
223 | struct cfq_rb_root grp_service_tree; | |
cdb16e8f | 224 | struct cfq_group root_group; |
22e2c507 | 225 | |
c0324a02 CZ |
226 | /* |
227 | * The priority currently being served | |
22e2c507 | 228 | */ |
c0324a02 | 229 | enum wl_prio_t serving_prio; |
718eee05 CZ |
230 | enum wl_type_t serving_type; |
231 | unsigned long workload_expires; | |
cdb16e8f | 232 | struct cfq_group *serving_group; |
a36e71f9 JA |
233 | |
234 | /* | |
235 | * Each priority tree is sorted by next_request position. These | |
236 | * trees are used when determining if two or more queues are | |
237 | * interleaving requests (see cfq_close_cooperator). | |
238 | */ | |
239 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
240 | ||
22e2c507 | 241 | unsigned int busy_queues; |
ef8a41df | 242 | unsigned int busy_sync_queues; |
22e2c507 | 243 | |
53c583d2 CZ |
244 | int rq_in_driver; |
245 | int rq_in_flight[2]; | |
45333d5a AC |
246 | |
247 | /* | |
248 | * queue-depth detection | |
249 | */ | |
250 | int rq_queued; | |
25776e35 | 251 | int hw_tag; |
e459dd08 CZ |
252 | /* |
253 | * hw_tag can be | |
254 | * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) | |
255 | * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) | |
256 | * 0 => no NCQ | |
257 | */ | |
258 | int hw_tag_est_depth; | |
259 | unsigned int hw_tag_samples; | |
1da177e4 | 260 | |
22e2c507 JA |
261 | /* |
262 | * idle window management | |
263 | */ | |
264 | struct timer_list idle_slice_timer; | |
23e018a1 | 265 | struct work_struct unplug_work; |
1da177e4 | 266 | |
22e2c507 JA |
267 | struct cfq_queue *active_queue; |
268 | struct cfq_io_context *active_cic; | |
22e2c507 | 269 | |
c2dea2d1 VT |
270 | /* |
271 | * async queue for each priority case | |
272 | */ | |
273 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
274 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 275 | |
6d048f53 | 276 | sector_t last_position; |
1da177e4 | 277 | |
1da177e4 LT |
278 | /* |
279 | * tunables, see top of file | |
280 | */ | |
281 | unsigned int cfq_quantum; | |
22e2c507 | 282 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
283 | unsigned int cfq_back_penalty; |
284 | unsigned int cfq_back_max; | |
22e2c507 JA |
285 | unsigned int cfq_slice[2]; |
286 | unsigned int cfq_slice_async_rq; | |
287 | unsigned int cfq_slice_idle; | |
80bdf0c7 | 288 | unsigned int cfq_group_idle; |
963b72fc | 289 | unsigned int cfq_latency; |
d9ff4187 | 290 | |
80b15c73 | 291 | unsigned int cic_index; |
d9ff4187 | 292 | struct list_head cic_list; |
1da177e4 | 293 | |
6118b70b JA |
294 | /* |
295 | * Fallback dummy cfqq for extreme OOM conditions | |
296 | */ | |
297 | struct cfq_queue oom_cfqq; | |
365722bb | 298 | |
573412b2 | 299 | unsigned long last_delayed_sync; |
25fb5169 VG |
300 | |
301 | /* List of cfq groups being managed on this device*/ | |
302 | struct hlist_head cfqg_list; | |
56edf7d7 VG |
303 | |
304 | /* Number of groups which are on blkcg->blkg_list */ | |
305 | unsigned int nr_blkcg_linked_grps; | |
1da177e4 LT |
306 | }; |
307 | ||
25fb5169 VG |
308 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd); |
309 | ||
cdb16e8f VG |
310 | static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, |
311 | enum wl_prio_t prio, | |
65b32a57 | 312 | enum wl_type_t type) |
c0324a02 | 313 | { |
1fa8f6d6 VG |
314 | if (!cfqg) |
315 | return NULL; | |
316 | ||
c0324a02 | 317 | if (prio == IDLE_WORKLOAD) |
cdb16e8f | 318 | return &cfqg->service_tree_idle; |
c0324a02 | 319 | |
cdb16e8f | 320 | return &cfqg->service_trees[prio][type]; |
c0324a02 CZ |
321 | } |
322 | ||
3b18152c | 323 | enum cfqq_state_flags { |
b0b8d749 JA |
324 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
325 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 326 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 | 327 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ |
b0b8d749 JA |
328 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
329 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
330 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 331 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 332 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
b3b6d040 | 333 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ |
ae54abed | 334 | CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */ |
76280aff | 335 | CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ |
f75edf2d | 336 | CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */ |
3b18152c JA |
337 | }; |
338 | ||
339 | #define CFQ_CFQQ_FNS(name) \ | |
340 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
341 | { \ | |
fe094d98 | 342 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
343 | } \ |
344 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
345 | { \ | |
fe094d98 | 346 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
347 | } \ |
348 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
349 | { \ | |
fe094d98 | 350 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
351 | } |
352 | ||
353 | CFQ_CFQQ_FNS(on_rr); | |
354 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 355 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c | 356 | CFQ_CFQQ_FNS(must_alloc_slice); |
3b18152c JA |
357 | CFQ_CFQQ_FNS(fifo_expire); |
358 | CFQ_CFQQ_FNS(idle_window); | |
359 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 360 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 361 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 362 | CFQ_CFQQ_FNS(coop); |
ae54abed | 363 | CFQ_CFQQ_FNS(split_coop); |
76280aff | 364 | CFQ_CFQQ_FNS(deep); |
f75edf2d | 365 | CFQ_CFQQ_FNS(wait_busy); |
3b18152c JA |
366 | #undef CFQ_CFQQ_FNS |
367 | ||
afc24d49 | 368 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
2868ef7b VG |
369 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
370 | blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \ | |
371 | cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \ | |
372 | blkg_path(&(cfqq)->cfqg->blkg), ##args); | |
373 | ||
374 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \ | |
375 | blk_add_trace_msg((cfqd)->queue, "%s " fmt, \ | |
376 | blkg_path(&(cfqg)->blkg), ##args); \ | |
377 | ||
378 | #else | |
7b679138 JA |
379 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
380 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
2868ef7b VG |
381 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0); |
382 | #endif | |
7b679138 JA |
383 | #define cfq_log(cfqd, fmt, args...) \ |
384 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
385 | ||
615f0259 VG |
386 | /* Traverses through cfq group service trees */ |
387 | #define for_each_cfqg_st(cfqg, i, j, st) \ | |
388 | for (i = 0; i <= IDLE_WORKLOAD; i++) \ | |
389 | for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ | |
390 | : &cfqg->service_tree_idle; \ | |
391 | (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ | |
392 | (i == IDLE_WORKLOAD && j == 0); \ | |
393 | j++, st = i < IDLE_WORKLOAD ? \ | |
394 | &cfqg->service_trees[i][j]: NULL) \ | |
395 | ||
396 | ||
02b35081 VG |
397 | static inline bool iops_mode(struct cfq_data *cfqd) |
398 | { | |
399 | /* | |
400 | * If we are not idling on queues and it is a NCQ drive, parallel | |
401 | * execution of requests is on and measuring time is not possible | |
402 | * in most of the cases until and unless we drive shallower queue | |
403 | * depths and that becomes a performance bottleneck. In such cases | |
404 | * switch to start providing fairness in terms of number of IOs. | |
405 | */ | |
406 | if (!cfqd->cfq_slice_idle && cfqd->hw_tag) | |
407 | return true; | |
408 | else | |
409 | return false; | |
410 | } | |
411 | ||
c0324a02 CZ |
412 | static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) |
413 | { | |
414 | if (cfq_class_idle(cfqq)) | |
415 | return IDLE_WORKLOAD; | |
416 | if (cfq_class_rt(cfqq)) | |
417 | return RT_WORKLOAD; | |
418 | return BE_WORKLOAD; | |
419 | } | |
420 | ||
718eee05 CZ |
421 | |
422 | static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) | |
423 | { | |
424 | if (!cfq_cfqq_sync(cfqq)) | |
425 | return ASYNC_WORKLOAD; | |
426 | if (!cfq_cfqq_idle_window(cfqq)) | |
427 | return SYNC_NOIDLE_WORKLOAD; | |
428 | return SYNC_WORKLOAD; | |
429 | } | |
430 | ||
58ff82f3 VG |
431 | static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl, |
432 | struct cfq_data *cfqd, | |
433 | struct cfq_group *cfqg) | |
c0324a02 CZ |
434 | { |
435 | if (wl == IDLE_WORKLOAD) | |
cdb16e8f | 436 | return cfqg->service_tree_idle.count; |
c0324a02 | 437 | |
cdb16e8f VG |
438 | return cfqg->service_trees[wl][ASYNC_WORKLOAD].count |
439 | + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count | |
440 | + cfqg->service_trees[wl][SYNC_WORKLOAD].count; | |
c0324a02 CZ |
441 | } |
442 | ||
f26bd1f0 VG |
443 | static inline int cfqg_busy_async_queues(struct cfq_data *cfqd, |
444 | struct cfq_group *cfqg) | |
445 | { | |
446 | return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count | |
447 | + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count; | |
448 | } | |
449 | ||
165125e1 | 450 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
a6151c3a | 451 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, |
fd0928df | 452 | struct io_context *, gfp_t); |
4ac845a2 | 453 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
454 | struct io_context *); |
455 | ||
456 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 457 | bool is_sync) |
91fac317 | 458 | { |
a6151c3a | 459 | return cic->cfqq[is_sync]; |
91fac317 VT |
460 | } |
461 | ||
462 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 463 | struct cfq_queue *cfqq, bool is_sync) |
91fac317 | 464 | { |
a6151c3a | 465 | cic->cfqq[is_sync] = cfqq; |
91fac317 VT |
466 | } |
467 | ||
bca4b914 | 468 | #define CIC_DEAD_KEY 1ul |
80b15c73 | 469 | #define CIC_DEAD_INDEX_SHIFT 1 |
bca4b914 KK |
470 | |
471 | static inline void *cfqd_dead_key(struct cfq_data *cfqd) | |
472 | { | |
80b15c73 | 473 | return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY); |
bca4b914 KK |
474 | } |
475 | ||
476 | static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic) | |
477 | { | |
478 | struct cfq_data *cfqd = cic->key; | |
479 | ||
480 | if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY)) | |
481 | return NULL; | |
482 | ||
483 | return cfqd; | |
484 | } | |
485 | ||
91fac317 VT |
486 | /* |
487 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
488 | * set (in which case it could also be direct WRITE). | |
489 | */ | |
a6151c3a | 490 | static inline bool cfq_bio_sync(struct bio *bio) |
91fac317 | 491 | { |
7b6d91da | 492 | return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC); |
91fac317 | 493 | } |
1da177e4 | 494 | |
99f95e52 AM |
495 | /* |
496 | * scheduler run of queue, if there are requests pending and no one in the | |
497 | * driver that will restart queueing | |
498 | */ | |
23e018a1 | 499 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) |
99f95e52 | 500 | { |
7b679138 JA |
501 | if (cfqd->busy_queues) { |
502 | cfq_log(cfqd, "schedule dispatch"); | |
23e018a1 | 503 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 504 | } |
99f95e52 AM |
505 | } |
506 | ||
44f7c160 JA |
507 | /* |
508 | * Scale schedule slice based on io priority. Use the sync time slice only | |
509 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
510 | * io only, should not get full sync slice length. | |
511 | */ | |
a6151c3a | 512 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, |
d9e7620e | 513 | unsigned short prio) |
44f7c160 | 514 | { |
d9e7620e | 515 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 516 | |
d9e7620e JA |
517 | WARN_ON(prio >= IOPRIO_BE_NR); |
518 | ||
519 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
520 | } | |
44f7c160 | 521 | |
d9e7620e JA |
522 | static inline int |
523 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
524 | { | |
525 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
526 | } |
527 | ||
25bc6b07 VG |
528 | static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg) |
529 | { | |
530 | u64 d = delta << CFQ_SERVICE_SHIFT; | |
531 | ||
532 | d = d * BLKIO_WEIGHT_DEFAULT; | |
533 | do_div(d, cfqg->weight); | |
534 | return d; | |
535 | } | |
536 | ||
537 | static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
538 | { | |
539 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
540 | if (delta > 0) | |
541 | min_vdisktime = vdisktime; | |
542 | ||
543 | return min_vdisktime; | |
544 | } | |
545 | ||
546 | static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
547 | { | |
548 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
549 | if (delta < 0) | |
550 | min_vdisktime = vdisktime; | |
551 | ||
552 | return min_vdisktime; | |
553 | } | |
554 | ||
555 | static void update_min_vdisktime(struct cfq_rb_root *st) | |
556 | { | |
25bc6b07 VG |
557 | struct cfq_group *cfqg; |
558 | ||
25bc6b07 VG |
559 | if (st->left) { |
560 | cfqg = rb_entry_cfqg(st->left); | |
a6032710 GJ |
561 | st->min_vdisktime = max_vdisktime(st->min_vdisktime, |
562 | cfqg->vdisktime); | |
25bc6b07 | 563 | } |
25bc6b07 VG |
564 | } |
565 | ||
5db5d642 CZ |
566 | /* |
567 | * get averaged number of queues of RT/BE priority. | |
568 | * average is updated, with a formula that gives more weight to higher numbers, | |
569 | * to quickly follows sudden increases and decrease slowly | |
570 | */ | |
571 | ||
58ff82f3 VG |
572 | static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, |
573 | struct cfq_group *cfqg, bool rt) | |
5869619c | 574 | { |
5db5d642 CZ |
575 | unsigned min_q, max_q; |
576 | unsigned mult = cfq_hist_divisor - 1; | |
577 | unsigned round = cfq_hist_divisor / 2; | |
58ff82f3 | 578 | unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); |
5db5d642 | 579 | |
58ff82f3 VG |
580 | min_q = min(cfqg->busy_queues_avg[rt], busy); |
581 | max_q = max(cfqg->busy_queues_avg[rt], busy); | |
582 | cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
5db5d642 | 583 | cfq_hist_divisor; |
58ff82f3 VG |
584 | return cfqg->busy_queues_avg[rt]; |
585 | } | |
586 | ||
587 | static inline unsigned | |
588 | cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
589 | { | |
590 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
591 | ||
592 | return cfq_target_latency * cfqg->weight / st->total_weight; | |
5db5d642 CZ |
593 | } |
594 | ||
c553f8e3 | 595 | static inline unsigned |
ba5bd520 | 596 | cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
44f7c160 | 597 | { |
5db5d642 CZ |
598 | unsigned slice = cfq_prio_to_slice(cfqd, cfqq); |
599 | if (cfqd->cfq_latency) { | |
58ff82f3 VG |
600 | /* |
601 | * interested queues (we consider only the ones with the same | |
602 | * priority class in the cfq group) | |
603 | */ | |
604 | unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, | |
605 | cfq_class_rt(cfqq)); | |
5db5d642 CZ |
606 | unsigned sync_slice = cfqd->cfq_slice[1]; |
607 | unsigned expect_latency = sync_slice * iq; | |
58ff82f3 VG |
608 | unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg); |
609 | ||
610 | if (expect_latency > group_slice) { | |
5db5d642 CZ |
611 | unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; |
612 | /* scale low_slice according to IO priority | |
613 | * and sync vs async */ | |
614 | unsigned low_slice = | |
615 | min(slice, base_low_slice * slice / sync_slice); | |
616 | /* the adapted slice value is scaled to fit all iqs | |
617 | * into the target latency */ | |
58ff82f3 | 618 | slice = max(slice * group_slice / expect_latency, |
5db5d642 CZ |
619 | low_slice); |
620 | } | |
621 | } | |
c553f8e3 SL |
622 | return slice; |
623 | } | |
624 | ||
625 | static inline void | |
626 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
627 | { | |
ba5bd520 | 628 | unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq); |
c553f8e3 | 629 | |
dae739eb | 630 | cfqq->slice_start = jiffies; |
5db5d642 | 631 | cfqq->slice_end = jiffies + slice; |
f75edf2d | 632 | cfqq->allocated_slice = slice; |
7b679138 | 633 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
634 | } |
635 | ||
636 | /* | |
637 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
638 | * isn't valid until the first request from the dispatch is activated | |
639 | * and the slice time set. | |
640 | */ | |
a6151c3a | 641 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) |
44f7c160 JA |
642 | { |
643 | if (cfq_cfqq_slice_new(cfqq)) | |
c1e44756 | 644 | return false; |
44f7c160 | 645 | if (time_before(jiffies, cfqq->slice_end)) |
c1e44756 | 646 | return false; |
44f7c160 | 647 | |
c1e44756 | 648 | return true; |
44f7c160 JA |
649 | } |
650 | ||
1da177e4 | 651 | /* |
5e705374 | 652 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 653 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 654 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 655 | */ |
5e705374 | 656 | static struct request * |
cf7c25cf | 657 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) |
1da177e4 | 658 | { |
cf7c25cf | 659 | sector_t s1, s2, d1 = 0, d2 = 0; |
1da177e4 | 660 | unsigned long back_max; |
e8a99053 AM |
661 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
662 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
663 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 664 | |
5e705374 JA |
665 | if (rq1 == NULL || rq1 == rq2) |
666 | return rq2; | |
667 | if (rq2 == NULL) | |
668 | return rq1; | |
9c2c38a1 | 669 | |
229836bd NK |
670 | if (rq_is_sync(rq1) != rq_is_sync(rq2)) |
671 | return rq_is_sync(rq1) ? rq1 : rq2; | |
672 | ||
673 | if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_META) | |
674 | return rq1->cmd_flags & REQ_META ? rq1 : rq2; | |
1da177e4 | 675 | |
83096ebf TH |
676 | s1 = blk_rq_pos(rq1); |
677 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 678 | |
1da177e4 LT |
679 | /* |
680 | * by definition, 1KiB is 2 sectors | |
681 | */ | |
682 | back_max = cfqd->cfq_back_max * 2; | |
683 | ||
684 | /* | |
685 | * Strict one way elevator _except_ in the case where we allow | |
686 | * short backward seeks which are biased as twice the cost of a | |
687 | * similar forward seek. | |
688 | */ | |
689 | if (s1 >= last) | |
690 | d1 = s1 - last; | |
691 | else if (s1 + back_max >= last) | |
692 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
693 | else | |
e8a99053 | 694 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
695 | |
696 | if (s2 >= last) | |
697 | d2 = s2 - last; | |
698 | else if (s2 + back_max >= last) | |
699 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
700 | else | |
e8a99053 | 701 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
702 | |
703 | /* Found required data */ | |
e8a99053 AM |
704 | |
705 | /* | |
706 | * By doing switch() on the bit mask "wrap" we avoid having to | |
707 | * check two variables for all permutations: --> faster! | |
708 | */ | |
709 | switch (wrap) { | |
5e705374 | 710 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 711 | if (d1 < d2) |
5e705374 | 712 | return rq1; |
e8a99053 | 713 | else if (d2 < d1) |
5e705374 | 714 | return rq2; |
e8a99053 AM |
715 | else { |
716 | if (s1 >= s2) | |
5e705374 | 717 | return rq1; |
e8a99053 | 718 | else |
5e705374 | 719 | return rq2; |
e8a99053 | 720 | } |
1da177e4 | 721 | |
e8a99053 | 722 | case CFQ_RQ2_WRAP: |
5e705374 | 723 | return rq1; |
e8a99053 | 724 | case CFQ_RQ1_WRAP: |
5e705374 JA |
725 | return rq2; |
726 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
727 | default: |
728 | /* | |
729 | * Since both rqs are wrapped, | |
730 | * start with the one that's further behind head | |
731 | * (--> only *one* back seek required), | |
732 | * since back seek takes more time than forward. | |
733 | */ | |
734 | if (s1 <= s2) | |
5e705374 | 735 | return rq1; |
1da177e4 | 736 | else |
5e705374 | 737 | return rq2; |
1da177e4 LT |
738 | } |
739 | } | |
740 | ||
498d3aa2 JA |
741 | /* |
742 | * The below is leftmost cache rbtree addon | |
743 | */ | |
0871714e | 744 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 | 745 | { |
615f0259 VG |
746 | /* Service tree is empty */ |
747 | if (!root->count) | |
748 | return NULL; | |
749 | ||
cc09e299 JA |
750 | if (!root->left) |
751 | root->left = rb_first(&root->rb); | |
752 | ||
0871714e JA |
753 | if (root->left) |
754 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
755 | ||
756 | return NULL; | |
cc09e299 JA |
757 | } |
758 | ||
1fa8f6d6 VG |
759 | static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) |
760 | { | |
761 | if (!root->left) | |
762 | root->left = rb_first(&root->rb); | |
763 | ||
764 | if (root->left) | |
765 | return rb_entry_cfqg(root->left); | |
766 | ||
767 | return NULL; | |
768 | } | |
769 | ||
a36e71f9 JA |
770 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
771 | { | |
772 | rb_erase(n, root); | |
773 | RB_CLEAR_NODE(n); | |
774 | } | |
775 | ||
cc09e299 JA |
776 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
777 | { | |
778 | if (root->left == n) | |
779 | root->left = NULL; | |
a36e71f9 | 780 | rb_erase_init(n, &root->rb); |
aa6f6a3d | 781 | --root->count; |
cc09e299 JA |
782 | } |
783 | ||
1da177e4 LT |
784 | /* |
785 | * would be nice to take fifo expire time into account as well | |
786 | */ | |
5e705374 JA |
787 | static struct request * |
788 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
789 | struct request *last) | |
1da177e4 | 790 | { |
21183b07 JA |
791 | struct rb_node *rbnext = rb_next(&last->rb_node); |
792 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 793 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 794 | |
21183b07 | 795 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
796 | |
797 | if (rbprev) | |
5e705374 | 798 | prev = rb_entry_rq(rbprev); |
1da177e4 | 799 | |
21183b07 | 800 | if (rbnext) |
5e705374 | 801 | next = rb_entry_rq(rbnext); |
21183b07 JA |
802 | else { |
803 | rbnext = rb_first(&cfqq->sort_list); | |
804 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 805 | next = rb_entry_rq(rbnext); |
21183b07 | 806 | } |
1da177e4 | 807 | |
cf7c25cf | 808 | return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); |
1da177e4 LT |
809 | } |
810 | ||
d9e7620e JA |
811 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
812 | struct cfq_queue *cfqq) | |
1da177e4 | 813 | { |
d9e7620e JA |
814 | /* |
815 | * just an approximation, should be ok. | |
816 | */ | |
cdb16e8f | 817 | return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
464191c6 | 818 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); |
d9e7620e JA |
819 | } |
820 | ||
1fa8f6d6 VG |
821 | static inline s64 |
822 | cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
823 | { | |
824 | return cfqg->vdisktime - st->min_vdisktime; | |
825 | } | |
826 | ||
827 | static void | |
828 | __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
829 | { | |
830 | struct rb_node **node = &st->rb.rb_node; | |
831 | struct rb_node *parent = NULL; | |
832 | struct cfq_group *__cfqg; | |
833 | s64 key = cfqg_key(st, cfqg); | |
834 | int left = 1; | |
835 | ||
836 | while (*node != NULL) { | |
837 | parent = *node; | |
838 | __cfqg = rb_entry_cfqg(parent); | |
839 | ||
840 | if (key < cfqg_key(st, __cfqg)) | |
841 | node = &parent->rb_left; | |
842 | else { | |
843 | node = &parent->rb_right; | |
844 | left = 0; | |
845 | } | |
846 | } | |
847 | ||
848 | if (left) | |
849 | st->left = &cfqg->rb_node; | |
850 | ||
851 | rb_link_node(&cfqg->rb_node, parent, node); | |
852 | rb_insert_color(&cfqg->rb_node, &st->rb); | |
853 | } | |
854 | ||
855 | static void | |
8184f93e JT |
856 | cfq_update_group_weight(struct cfq_group *cfqg) |
857 | { | |
858 | BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); | |
859 | if (cfqg->needs_update) { | |
860 | cfqg->weight = cfqg->new_weight; | |
861 | cfqg->needs_update = false; | |
862 | } | |
863 | } | |
864 | ||
865 | static void | |
866 | cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
867 | { | |
868 | BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); | |
869 | ||
870 | cfq_update_group_weight(cfqg); | |
871 | __cfq_group_service_tree_add(st, cfqg); | |
872 | st->total_weight += cfqg->weight; | |
873 | } | |
874 | ||
875 | static void | |
876 | cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1fa8f6d6 VG |
877 | { |
878 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
879 | struct cfq_group *__cfqg; | |
880 | struct rb_node *n; | |
881 | ||
882 | cfqg->nr_cfqq++; | |
760701bf | 883 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) |
1fa8f6d6 VG |
884 | return; |
885 | ||
886 | /* | |
887 | * Currently put the group at the end. Later implement something | |
888 | * so that groups get lesser vtime based on their weights, so that | |
25985edc | 889 | * if group does not loose all if it was not continuously backlogged. |
1fa8f6d6 VG |
890 | */ |
891 | n = rb_last(&st->rb); | |
892 | if (n) { | |
893 | __cfqg = rb_entry_cfqg(n); | |
894 | cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; | |
895 | } else | |
896 | cfqg->vdisktime = st->min_vdisktime; | |
8184f93e JT |
897 | cfq_group_service_tree_add(st, cfqg); |
898 | } | |
1fa8f6d6 | 899 | |
8184f93e JT |
900 | static void |
901 | cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
902 | { | |
903 | st->total_weight -= cfqg->weight; | |
904 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) | |
905 | cfq_rb_erase(&cfqg->rb_node, st); | |
1fa8f6d6 VG |
906 | } |
907 | ||
908 | static void | |
8184f93e | 909 | cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg) |
1fa8f6d6 VG |
910 | { |
911 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
912 | ||
913 | BUG_ON(cfqg->nr_cfqq < 1); | |
914 | cfqg->nr_cfqq--; | |
25bc6b07 | 915 | |
1fa8f6d6 VG |
916 | /* If there are other cfq queues under this group, don't delete it */ |
917 | if (cfqg->nr_cfqq) | |
918 | return; | |
919 | ||
2868ef7b | 920 | cfq_log_cfqg(cfqd, cfqg, "del_from_rr group"); |
8184f93e | 921 | cfq_group_service_tree_del(st, cfqg); |
dae739eb | 922 | cfqg->saved_workload_slice = 0; |
e98ef89b | 923 | cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1); |
dae739eb VG |
924 | } |
925 | ||
167400d3 JT |
926 | static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq, |
927 | unsigned int *unaccounted_time) | |
dae739eb | 928 | { |
f75edf2d | 929 | unsigned int slice_used; |
dae739eb VG |
930 | |
931 | /* | |
932 | * Queue got expired before even a single request completed or | |
933 | * got expired immediately after first request completion. | |
934 | */ | |
935 | if (!cfqq->slice_start || cfqq->slice_start == jiffies) { | |
936 | /* | |
937 | * Also charge the seek time incurred to the group, otherwise | |
938 | * if there are mutiple queues in the group, each can dispatch | |
939 | * a single request on seeky media and cause lots of seek time | |
940 | * and group will never know it. | |
941 | */ | |
942 | slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start), | |
943 | 1); | |
944 | } else { | |
945 | slice_used = jiffies - cfqq->slice_start; | |
167400d3 JT |
946 | if (slice_used > cfqq->allocated_slice) { |
947 | *unaccounted_time = slice_used - cfqq->allocated_slice; | |
f75edf2d | 948 | slice_used = cfqq->allocated_slice; |
167400d3 JT |
949 | } |
950 | if (time_after(cfqq->slice_start, cfqq->dispatch_start)) | |
951 | *unaccounted_time += cfqq->slice_start - | |
952 | cfqq->dispatch_start; | |
dae739eb VG |
953 | } |
954 | ||
dae739eb VG |
955 | return slice_used; |
956 | } | |
957 | ||
958 | static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg, | |
e5ff082e | 959 | struct cfq_queue *cfqq) |
dae739eb VG |
960 | { |
961 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
167400d3 | 962 | unsigned int used_sl, charge, unaccounted_sl = 0; |
f26bd1f0 VG |
963 | int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg) |
964 | - cfqg->service_tree_idle.count; | |
965 | ||
966 | BUG_ON(nr_sync < 0); | |
167400d3 | 967 | used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl); |
dae739eb | 968 | |
02b35081 VG |
969 | if (iops_mode(cfqd)) |
970 | charge = cfqq->slice_dispatch; | |
971 | else if (!cfq_cfqq_sync(cfqq) && !nr_sync) | |
972 | charge = cfqq->allocated_slice; | |
dae739eb VG |
973 | |
974 | /* Can't update vdisktime while group is on service tree */ | |
8184f93e | 975 | cfq_group_service_tree_del(st, cfqg); |
02b35081 | 976 | cfqg->vdisktime += cfq_scale_slice(charge, cfqg); |
8184f93e JT |
977 | /* If a new weight was requested, update now, off tree */ |
978 | cfq_group_service_tree_add(st, cfqg); | |
dae739eb VG |
979 | |
980 | /* This group is being expired. Save the context */ | |
981 | if (time_after(cfqd->workload_expires, jiffies)) { | |
982 | cfqg->saved_workload_slice = cfqd->workload_expires | |
983 | - jiffies; | |
984 | cfqg->saved_workload = cfqd->serving_type; | |
985 | cfqg->saved_serving_prio = cfqd->serving_prio; | |
986 | } else | |
987 | cfqg->saved_workload_slice = 0; | |
2868ef7b VG |
988 | |
989 | cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime, | |
990 | st->min_vdisktime); | |
c4e7893e VG |
991 | cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u" |
992 | " sect=%u", used_sl, cfqq->slice_dispatch, charge, | |
993 | iops_mode(cfqd), cfqq->nr_sectors); | |
167400d3 JT |
994 | cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl, |
995 | unaccounted_sl); | |
e98ef89b | 996 | cfq_blkiocg_set_start_empty_time(&cfqg->blkg); |
1fa8f6d6 VG |
997 | } |
998 | ||
25fb5169 VG |
999 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
1000 | static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg) | |
1001 | { | |
1002 | if (blkg) | |
1003 | return container_of(blkg, struct cfq_group, blkg); | |
1004 | return NULL; | |
1005 | } | |
1006 | ||
fe071437 VG |
1007 | void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg, |
1008 | unsigned int weight) | |
f8d461d6 | 1009 | { |
8184f93e JT |
1010 | struct cfq_group *cfqg = cfqg_of_blkg(blkg); |
1011 | cfqg->new_weight = weight; | |
1012 | cfqg->needs_update = true; | |
f8d461d6 VG |
1013 | } |
1014 | ||
f469a7b4 VG |
1015 | static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd, |
1016 | struct cfq_group *cfqg, struct blkio_cgroup *blkcg) | |
25fb5169 | 1017 | { |
22084190 VG |
1018 | struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; |
1019 | unsigned int major, minor; | |
25fb5169 | 1020 | |
f469a7b4 VG |
1021 | /* |
1022 | * Add group onto cgroup list. It might happen that bdi->dev is | |
1023 | * not initialized yet. Initialize this new group without major | |
1024 | * and minor info and this info will be filled in once a new thread | |
1025 | * comes for IO. | |
1026 | */ | |
1027 | if (bdi->dev) { | |
a74b2ada | 1028 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); |
f469a7b4 VG |
1029 | cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, |
1030 | (void *)cfqd, MKDEV(major, minor)); | |
1031 | } else | |
1032 | cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, | |
1033 | (void *)cfqd, 0); | |
1034 | ||
1035 | cfqd->nr_blkcg_linked_grps++; | |
1036 | cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev); | |
1037 | ||
1038 | /* Add group on cfqd list */ | |
1039 | hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); | |
1040 | } | |
1041 | ||
1042 | /* | |
1043 | * Should be called from sleepable context. No request queue lock as per | |
1044 | * cpu stats are allocated dynamically and alloc_percpu needs to be called | |
1045 | * from sleepable context. | |
1046 | */ | |
1047 | static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd) | |
1048 | { | |
1049 | struct cfq_group *cfqg = NULL; | |
5624a4e4 | 1050 | int i, j, ret; |
f469a7b4 | 1051 | struct cfq_rb_root *st; |
25fb5169 VG |
1052 | |
1053 | cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node); | |
1054 | if (!cfqg) | |
f469a7b4 | 1055 | return NULL; |
25fb5169 | 1056 | |
25fb5169 VG |
1057 | for_each_cfqg_st(cfqg, i, j, st) |
1058 | *st = CFQ_RB_ROOT; | |
1059 | RB_CLEAR_NODE(&cfqg->rb_node); | |
1060 | ||
b1c35769 VG |
1061 | /* |
1062 | * Take the initial reference that will be released on destroy | |
1063 | * This can be thought of a joint reference by cgroup and | |
1064 | * elevator which will be dropped by either elevator exit | |
1065 | * or cgroup deletion path depending on who is exiting first. | |
1066 | */ | |
329a6781 | 1067 | cfqg->ref = 1; |
5624a4e4 VG |
1068 | |
1069 | ret = blkio_alloc_blkg_stats(&cfqg->blkg); | |
1070 | if (ret) { | |
1071 | kfree(cfqg); | |
1072 | return NULL; | |
1073 | } | |
1074 | ||
f469a7b4 VG |
1075 | return cfqg; |
1076 | } | |
1077 | ||
1078 | static struct cfq_group * | |
1079 | cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg) | |
1080 | { | |
1081 | struct cfq_group *cfqg = NULL; | |
1082 | void *key = cfqd; | |
1083 | struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; | |
1084 | unsigned int major, minor; | |
b1c35769 | 1085 | |
180be2a0 | 1086 | /* |
f469a7b4 VG |
1087 | * This is the common case when there are no blkio cgroups. |
1088 | * Avoid lookup in this case | |
180be2a0 | 1089 | */ |
f469a7b4 VG |
1090 | if (blkcg == &blkio_root_cgroup) |
1091 | cfqg = &cfqd->root_group; | |
1092 | else | |
1093 | cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key)); | |
25fb5169 | 1094 | |
f469a7b4 VG |
1095 | if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) { |
1096 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); | |
1097 | cfqg->blkg.dev = MKDEV(major, minor); | |
1098 | } | |
25fb5169 | 1099 | |
25fb5169 VG |
1100 | return cfqg; |
1101 | } | |
1102 | ||
1103 | /* | |
3e59cf9d VG |
1104 | * Search for the cfq group current task belongs to. request_queue lock must |
1105 | * be held. | |
25fb5169 | 1106 | */ |
3e59cf9d | 1107 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd) |
25fb5169 | 1108 | { |
70087dc3 | 1109 | struct blkio_cgroup *blkcg; |
f469a7b4 VG |
1110 | struct cfq_group *cfqg = NULL, *__cfqg = NULL; |
1111 | struct request_queue *q = cfqd->queue; | |
25fb5169 VG |
1112 | |
1113 | rcu_read_lock(); | |
70087dc3 | 1114 | blkcg = task_blkio_cgroup(current); |
f469a7b4 VG |
1115 | cfqg = cfq_find_cfqg(cfqd, blkcg); |
1116 | if (cfqg) { | |
1117 | rcu_read_unlock(); | |
1118 | return cfqg; | |
1119 | } | |
1120 | ||
1121 | /* | |
1122 | * Need to allocate a group. Allocation of group also needs allocation | |
1123 | * of per cpu stats which in-turn takes a mutex() and can block. Hence | |
1124 | * we need to drop rcu lock and queue_lock before we call alloc. | |
1125 | * | |
1126 | * Not taking any queue reference here and assuming that queue is | |
1127 | * around by the time we return. CFQ queue allocation code does | |
1128 | * the same. It might be racy though. | |
1129 | */ | |
1130 | ||
1131 | rcu_read_unlock(); | |
1132 | spin_unlock_irq(q->queue_lock); | |
1133 | ||
1134 | cfqg = cfq_alloc_cfqg(cfqd); | |
1135 | ||
1136 | spin_lock_irq(q->queue_lock); | |
1137 | ||
1138 | rcu_read_lock(); | |
1139 | blkcg = task_blkio_cgroup(current); | |
1140 | ||
1141 | /* | |
1142 | * If some other thread already allocated the group while we were | |
1143 | * not holding queue lock, free up the group | |
1144 | */ | |
1145 | __cfqg = cfq_find_cfqg(cfqd, blkcg); | |
1146 | ||
1147 | if (__cfqg) { | |
1148 | kfree(cfqg); | |
1149 | rcu_read_unlock(); | |
1150 | return __cfqg; | |
1151 | } | |
1152 | ||
3e59cf9d | 1153 | if (!cfqg) |
25fb5169 | 1154 | cfqg = &cfqd->root_group; |
f469a7b4 VG |
1155 | |
1156 | cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg); | |
25fb5169 VG |
1157 | rcu_read_unlock(); |
1158 | return cfqg; | |
1159 | } | |
1160 | ||
7f1dc8a2 VG |
1161 | static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg) |
1162 | { | |
329a6781 | 1163 | cfqg->ref++; |
7f1dc8a2 VG |
1164 | return cfqg; |
1165 | } | |
1166 | ||
25fb5169 VG |
1167 | static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) |
1168 | { | |
1169 | /* Currently, all async queues are mapped to root group */ | |
1170 | if (!cfq_cfqq_sync(cfqq)) | |
1171 | cfqg = &cfqq->cfqd->root_group; | |
1172 | ||
1173 | cfqq->cfqg = cfqg; | |
b1c35769 | 1174 | /* cfqq reference on cfqg */ |
329a6781 | 1175 | cfqq->cfqg->ref++; |
b1c35769 VG |
1176 | } |
1177 | ||
1178 | static void cfq_put_cfqg(struct cfq_group *cfqg) | |
1179 | { | |
1180 | struct cfq_rb_root *st; | |
1181 | int i, j; | |
1182 | ||
329a6781 SL |
1183 | BUG_ON(cfqg->ref <= 0); |
1184 | cfqg->ref--; | |
1185 | if (cfqg->ref) | |
b1c35769 VG |
1186 | return; |
1187 | for_each_cfqg_st(cfqg, i, j, st) | |
b54ce60e | 1188 | BUG_ON(!RB_EMPTY_ROOT(&st->rb)); |
5624a4e4 | 1189 | free_percpu(cfqg->blkg.stats_cpu); |
b1c35769 VG |
1190 | kfree(cfqg); |
1191 | } | |
1192 | ||
1193 | static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1194 | { | |
1195 | /* Something wrong if we are trying to remove same group twice */ | |
1196 | BUG_ON(hlist_unhashed(&cfqg->cfqd_node)); | |
1197 | ||
1198 | hlist_del_init(&cfqg->cfqd_node); | |
1199 | ||
1200 | /* | |
1201 | * Put the reference taken at the time of creation so that when all | |
1202 | * queues are gone, group can be destroyed. | |
1203 | */ | |
1204 | cfq_put_cfqg(cfqg); | |
1205 | } | |
1206 | ||
1207 | static void cfq_release_cfq_groups(struct cfq_data *cfqd) | |
1208 | { | |
1209 | struct hlist_node *pos, *n; | |
1210 | struct cfq_group *cfqg; | |
1211 | ||
1212 | hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) { | |
1213 | /* | |
1214 | * If cgroup removal path got to blk_group first and removed | |
1215 | * it from cgroup list, then it will take care of destroying | |
1216 | * cfqg also. | |
1217 | */ | |
e98ef89b | 1218 | if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg)) |
b1c35769 VG |
1219 | cfq_destroy_cfqg(cfqd, cfqg); |
1220 | } | |
25fb5169 | 1221 | } |
b1c35769 VG |
1222 | |
1223 | /* | |
1224 | * Blk cgroup controller notification saying that blkio_group object is being | |
1225 | * delinked as associated cgroup object is going away. That also means that | |
1226 | * no new IO will come in this group. So get rid of this group as soon as | |
1227 | * any pending IO in the group is finished. | |
1228 | * | |
1229 | * This function is called under rcu_read_lock(). key is the rcu protected | |
1230 | * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu | |
1231 | * read lock. | |
1232 | * | |
1233 | * "key" was fetched from blkio_group under blkio_cgroup->lock. That means | |
1234 | * it should not be NULL as even if elevator was exiting, cgroup deltion | |
1235 | * path got to it first. | |
1236 | */ | |
1237 | void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg) | |
1238 | { | |
1239 | unsigned long flags; | |
1240 | struct cfq_data *cfqd = key; | |
1241 | ||
1242 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
1243 | cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg)); | |
1244 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
1245 | } | |
1246 | ||
25fb5169 | 1247 | #else /* GROUP_IOSCHED */ |
3e59cf9d | 1248 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd) |
25fb5169 VG |
1249 | { |
1250 | return &cfqd->root_group; | |
1251 | } | |
7f1dc8a2 VG |
1252 | |
1253 | static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg) | |
1254 | { | |
50eaeb32 | 1255 | return cfqg; |
7f1dc8a2 VG |
1256 | } |
1257 | ||
25fb5169 VG |
1258 | static inline void |
1259 | cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) { | |
1260 | cfqq->cfqg = cfqg; | |
1261 | } | |
1262 | ||
b1c35769 VG |
1263 | static void cfq_release_cfq_groups(struct cfq_data *cfqd) {} |
1264 | static inline void cfq_put_cfqg(struct cfq_group *cfqg) {} | |
1265 | ||
25fb5169 VG |
1266 | #endif /* GROUP_IOSCHED */ |
1267 | ||
498d3aa2 | 1268 | /* |
c0324a02 | 1269 | * The cfqd->service_trees holds all pending cfq_queue's that have |
498d3aa2 JA |
1270 | * requests waiting to be processed. It is sorted in the order that |
1271 | * we will service the queues. | |
1272 | */ | |
a36e71f9 | 1273 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 1274 | bool add_front) |
d9e7620e | 1275 | { |
0871714e JA |
1276 | struct rb_node **p, *parent; |
1277 | struct cfq_queue *__cfqq; | |
d9e7620e | 1278 | unsigned long rb_key; |
c0324a02 | 1279 | struct cfq_rb_root *service_tree; |
498d3aa2 | 1280 | int left; |
dae739eb | 1281 | int new_cfqq = 1; |
ae30c286 | 1282 | |
cdb16e8f | 1283 | service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), |
65b32a57 | 1284 | cfqq_type(cfqq)); |
0871714e JA |
1285 | if (cfq_class_idle(cfqq)) { |
1286 | rb_key = CFQ_IDLE_DELAY; | |
aa6f6a3d | 1287 | parent = rb_last(&service_tree->rb); |
0871714e JA |
1288 | if (parent && parent != &cfqq->rb_node) { |
1289 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1290 | rb_key += __cfqq->rb_key; | |
1291 | } else | |
1292 | rb_key += jiffies; | |
1293 | } else if (!add_front) { | |
b9c8946b JA |
1294 | /* |
1295 | * Get our rb key offset. Subtract any residual slice | |
1296 | * value carried from last service. A negative resid | |
1297 | * count indicates slice overrun, and this should position | |
1298 | * the next service time further away in the tree. | |
1299 | */ | |
edd75ffd | 1300 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
b9c8946b | 1301 | rb_key -= cfqq->slice_resid; |
edd75ffd | 1302 | cfqq->slice_resid = 0; |
48e025e6 CZ |
1303 | } else { |
1304 | rb_key = -HZ; | |
aa6f6a3d | 1305 | __cfqq = cfq_rb_first(service_tree); |
48e025e6 CZ |
1306 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; |
1307 | } | |
1da177e4 | 1308 | |
d9e7620e | 1309 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
dae739eb | 1310 | new_cfqq = 0; |
99f9628a | 1311 | /* |
d9e7620e | 1312 | * same position, nothing more to do |
99f9628a | 1313 | */ |
c0324a02 CZ |
1314 | if (rb_key == cfqq->rb_key && |
1315 | cfqq->service_tree == service_tree) | |
d9e7620e | 1316 | return; |
1da177e4 | 1317 | |
aa6f6a3d CZ |
1318 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); |
1319 | cfqq->service_tree = NULL; | |
1da177e4 | 1320 | } |
d9e7620e | 1321 | |
498d3aa2 | 1322 | left = 1; |
0871714e | 1323 | parent = NULL; |
aa6f6a3d CZ |
1324 | cfqq->service_tree = service_tree; |
1325 | p = &service_tree->rb.rb_node; | |
d9e7620e | 1326 | while (*p) { |
67060e37 | 1327 | struct rb_node **n; |
cc09e299 | 1328 | |
d9e7620e JA |
1329 | parent = *p; |
1330 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1331 | ||
0c534e0a | 1332 | /* |
c0324a02 | 1333 | * sort by key, that represents service time. |
0c534e0a | 1334 | */ |
c0324a02 | 1335 | if (time_before(rb_key, __cfqq->rb_key)) |
67060e37 | 1336 | n = &(*p)->rb_left; |
c0324a02 | 1337 | else { |
67060e37 | 1338 | n = &(*p)->rb_right; |
cc09e299 | 1339 | left = 0; |
c0324a02 | 1340 | } |
67060e37 JA |
1341 | |
1342 | p = n; | |
d9e7620e JA |
1343 | } |
1344 | ||
cc09e299 | 1345 | if (left) |
aa6f6a3d | 1346 | service_tree->left = &cfqq->rb_node; |
cc09e299 | 1347 | |
d9e7620e JA |
1348 | cfqq->rb_key = rb_key; |
1349 | rb_link_node(&cfqq->rb_node, parent, p); | |
aa6f6a3d CZ |
1350 | rb_insert_color(&cfqq->rb_node, &service_tree->rb); |
1351 | service_tree->count++; | |
20359f27 | 1352 | if (add_front || !new_cfqq) |
dae739eb | 1353 | return; |
8184f93e | 1354 | cfq_group_notify_queue_add(cfqd, cfqq->cfqg); |
1da177e4 LT |
1355 | } |
1356 | ||
a36e71f9 | 1357 | static struct cfq_queue * |
f2d1f0ae JA |
1358 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
1359 | sector_t sector, struct rb_node **ret_parent, | |
1360 | struct rb_node ***rb_link) | |
a36e71f9 | 1361 | { |
a36e71f9 JA |
1362 | struct rb_node **p, *parent; |
1363 | struct cfq_queue *cfqq = NULL; | |
1364 | ||
1365 | parent = NULL; | |
1366 | p = &root->rb_node; | |
1367 | while (*p) { | |
1368 | struct rb_node **n; | |
1369 | ||
1370 | parent = *p; | |
1371 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
1372 | ||
1373 | /* | |
1374 | * Sort strictly based on sector. Smallest to the left, | |
1375 | * largest to the right. | |
1376 | */ | |
2e46e8b2 | 1377 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 1378 | n = &(*p)->rb_right; |
2e46e8b2 | 1379 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
1380 | n = &(*p)->rb_left; |
1381 | else | |
1382 | break; | |
1383 | p = n; | |
3ac6c9f8 | 1384 | cfqq = NULL; |
a36e71f9 JA |
1385 | } |
1386 | ||
1387 | *ret_parent = parent; | |
1388 | if (rb_link) | |
1389 | *rb_link = p; | |
3ac6c9f8 | 1390 | return cfqq; |
a36e71f9 JA |
1391 | } |
1392 | ||
1393 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1394 | { | |
a36e71f9 JA |
1395 | struct rb_node **p, *parent; |
1396 | struct cfq_queue *__cfqq; | |
1397 | ||
f2d1f0ae JA |
1398 | if (cfqq->p_root) { |
1399 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1400 | cfqq->p_root = NULL; | |
1401 | } | |
a36e71f9 JA |
1402 | |
1403 | if (cfq_class_idle(cfqq)) | |
1404 | return; | |
1405 | if (!cfqq->next_rq) | |
1406 | return; | |
1407 | ||
f2d1f0ae | 1408 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
1409 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
1410 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
1411 | if (!__cfqq) { |
1412 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
1413 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
1414 | } else | |
1415 | cfqq->p_root = NULL; | |
a36e71f9 JA |
1416 | } |
1417 | ||
498d3aa2 JA |
1418 | /* |
1419 | * Update cfqq's position in the service tree. | |
1420 | */ | |
edd75ffd | 1421 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 1422 | { |
6d048f53 JA |
1423 | /* |
1424 | * Resorting requires the cfqq to be on the RR list already. | |
1425 | */ | |
a36e71f9 | 1426 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 1427 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
1428 | cfq_prio_tree_add(cfqd, cfqq); |
1429 | } | |
6d048f53 JA |
1430 | } |
1431 | ||
1da177e4 LT |
1432 | /* |
1433 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 1434 | * the pending list according to last request service |
1da177e4 | 1435 | */ |
febffd61 | 1436 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1437 | { |
7b679138 | 1438 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
1439 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1440 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 | 1441 | cfqd->busy_queues++; |
ef8a41df SL |
1442 | if (cfq_cfqq_sync(cfqq)) |
1443 | cfqd->busy_sync_queues++; | |
1da177e4 | 1444 | |
edd75ffd | 1445 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
1446 | } |
1447 | ||
498d3aa2 JA |
1448 | /* |
1449 | * Called when the cfqq no longer has requests pending, remove it from | |
1450 | * the service tree. | |
1451 | */ | |
febffd61 | 1452 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1453 | { |
7b679138 | 1454 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
1455 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
1456 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 1457 | |
aa6f6a3d CZ |
1458 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
1459 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
1460 | cfqq->service_tree = NULL; | |
1461 | } | |
f2d1f0ae JA |
1462 | if (cfqq->p_root) { |
1463 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1464 | cfqq->p_root = NULL; | |
1465 | } | |
d9e7620e | 1466 | |
8184f93e | 1467 | cfq_group_notify_queue_del(cfqd, cfqq->cfqg); |
1da177e4 LT |
1468 | BUG_ON(!cfqd->busy_queues); |
1469 | cfqd->busy_queues--; | |
ef8a41df SL |
1470 | if (cfq_cfqq_sync(cfqq)) |
1471 | cfqd->busy_sync_queues--; | |
1da177e4 LT |
1472 | } |
1473 | ||
1474 | /* | |
1475 | * rb tree support functions | |
1476 | */ | |
febffd61 | 1477 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 1478 | { |
5e705374 | 1479 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
5e705374 | 1480 | const int sync = rq_is_sync(rq); |
1da177e4 | 1481 | |
b4878f24 JA |
1482 | BUG_ON(!cfqq->queued[sync]); |
1483 | cfqq->queued[sync]--; | |
1da177e4 | 1484 | |
5e705374 | 1485 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 1486 | |
f04a6424 VG |
1487 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { |
1488 | /* | |
1489 | * Queue will be deleted from service tree when we actually | |
1490 | * expire it later. Right now just remove it from prio tree | |
1491 | * as it is empty. | |
1492 | */ | |
1493 | if (cfqq->p_root) { | |
1494 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1495 | cfqq->p_root = NULL; | |
1496 | } | |
1497 | } | |
1da177e4 LT |
1498 | } |
1499 | ||
5e705374 | 1500 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 1501 | { |
5e705374 | 1502 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 1503 | struct cfq_data *cfqd = cfqq->cfqd; |
a36e71f9 | 1504 | struct request *__alias, *prev; |
1da177e4 | 1505 | |
5380a101 | 1506 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
1507 | |
1508 | /* | |
1509 | * looks a little odd, but the first insert might return an alias. | |
1510 | * if that happens, put the alias on the dispatch list | |
1511 | */ | |
21183b07 | 1512 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 1513 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
1514 | |
1515 | if (!cfq_cfqq_on_rr(cfqq)) | |
1516 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
1517 | |
1518 | /* | |
1519 | * check if this request is a better next-serve candidate | |
1520 | */ | |
a36e71f9 | 1521 | prev = cfqq->next_rq; |
cf7c25cf | 1522 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); |
a36e71f9 JA |
1523 | |
1524 | /* | |
1525 | * adjust priority tree position, if ->next_rq changes | |
1526 | */ | |
1527 | if (prev != cfqq->next_rq) | |
1528 | cfq_prio_tree_add(cfqd, cfqq); | |
1529 | ||
5044eed4 | 1530 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
1531 | } |
1532 | ||
febffd61 | 1533 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 1534 | { |
5380a101 JA |
1535 | elv_rb_del(&cfqq->sort_list, rq); |
1536 | cfqq->queued[rq_is_sync(rq)]--; | |
e98ef89b VG |
1537 | cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, |
1538 | rq_data_dir(rq), rq_is_sync(rq)); | |
5e705374 | 1539 | cfq_add_rq_rb(rq); |
e98ef89b | 1540 | cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg, |
7f1dc8a2 VG |
1541 | &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq), |
1542 | rq_is_sync(rq)); | |
1da177e4 LT |
1543 | } |
1544 | ||
206dc69b JA |
1545 | static struct request * |
1546 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 1547 | { |
206dc69b | 1548 | struct task_struct *tsk = current; |
91fac317 | 1549 | struct cfq_io_context *cic; |
206dc69b | 1550 | struct cfq_queue *cfqq; |
1da177e4 | 1551 | |
4ac845a2 | 1552 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
1553 | if (!cic) |
1554 | return NULL; | |
1555 | ||
1556 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
1557 | if (cfqq) { |
1558 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
1559 | ||
21183b07 | 1560 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 1561 | } |
1da177e4 | 1562 | |
1da177e4 LT |
1563 | return NULL; |
1564 | } | |
1565 | ||
165125e1 | 1566 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1567 | { |
22e2c507 | 1568 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 1569 | |
53c583d2 | 1570 | cfqd->rq_in_driver++; |
7b679138 | 1571 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
53c583d2 | 1572 | cfqd->rq_in_driver); |
25776e35 | 1573 | |
5b93629b | 1574 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
1575 | } |
1576 | ||
165125e1 | 1577 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1578 | { |
b4878f24 JA |
1579 | struct cfq_data *cfqd = q->elevator->elevator_data; |
1580 | ||
53c583d2 CZ |
1581 | WARN_ON(!cfqd->rq_in_driver); |
1582 | cfqd->rq_in_driver--; | |
7b679138 | 1583 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
53c583d2 | 1584 | cfqd->rq_in_driver); |
1da177e4 LT |
1585 | } |
1586 | ||
b4878f24 | 1587 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 1588 | { |
5e705374 | 1589 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 1590 | |
5e705374 JA |
1591 | if (cfqq->next_rq == rq) |
1592 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 1593 | |
b4878f24 | 1594 | list_del_init(&rq->queuelist); |
5e705374 | 1595 | cfq_del_rq_rb(rq); |
374f84ac | 1596 | |
45333d5a | 1597 | cfqq->cfqd->rq_queued--; |
e98ef89b VG |
1598 | cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, |
1599 | rq_data_dir(rq), rq_is_sync(rq)); | |
7b6d91da | 1600 | if (rq->cmd_flags & REQ_META) { |
374f84ac JA |
1601 | WARN_ON(!cfqq->meta_pending); |
1602 | cfqq->meta_pending--; | |
1603 | } | |
1da177e4 LT |
1604 | } |
1605 | ||
165125e1 JA |
1606 | static int cfq_merge(struct request_queue *q, struct request **req, |
1607 | struct bio *bio) | |
1da177e4 LT |
1608 | { |
1609 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1610 | struct request *__rq; | |
1da177e4 | 1611 | |
206dc69b | 1612 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 1613 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
1614 | *req = __rq; |
1615 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1616 | } |
1617 | ||
1618 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1619 | } |
1620 | ||
165125e1 | 1621 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 1622 | int type) |
1da177e4 | 1623 | { |
21183b07 | 1624 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 1625 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 1626 | |
5e705374 | 1627 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 1628 | } |
1da177e4 LT |
1629 | } |
1630 | ||
812d4026 DS |
1631 | static void cfq_bio_merged(struct request_queue *q, struct request *req, |
1632 | struct bio *bio) | |
1633 | { | |
e98ef89b VG |
1634 | cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg, |
1635 | bio_data_dir(bio), cfq_bio_sync(bio)); | |
812d4026 DS |
1636 | } |
1637 | ||
1da177e4 | 1638 | static void |
165125e1 | 1639 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
1640 | struct request *next) |
1641 | { | |
cf7c25cf | 1642 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 JA |
1643 | /* |
1644 | * reposition in fifo if next is older than rq | |
1645 | */ | |
1646 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
30996f40 | 1647 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
22e2c507 | 1648 | list_move(&rq->queuelist, &next->queuelist); |
30996f40 JA |
1649 | rq_set_fifo_time(rq, rq_fifo_time(next)); |
1650 | } | |
22e2c507 | 1651 | |
cf7c25cf CZ |
1652 | if (cfqq->next_rq == next) |
1653 | cfqq->next_rq = rq; | |
b4878f24 | 1654 | cfq_remove_request(next); |
e98ef89b VG |
1655 | cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg, |
1656 | rq_data_dir(next), rq_is_sync(next)); | |
22e2c507 JA |
1657 | } |
1658 | ||
165125e1 | 1659 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
1660 | struct bio *bio) |
1661 | { | |
1662 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 1663 | struct cfq_io_context *cic; |
da775265 | 1664 | struct cfq_queue *cfqq; |
da775265 JA |
1665 | |
1666 | /* | |
ec8acb69 | 1667 | * Disallow merge of a sync bio into an async request. |
da775265 | 1668 | */ |
91fac317 | 1669 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
a6151c3a | 1670 | return false; |
da775265 JA |
1671 | |
1672 | /* | |
719d3402 JA |
1673 | * Lookup the cfqq that this bio will be queued with. Allow |
1674 | * merge only if rq is queued there. | |
da775265 | 1675 | */ |
4ac845a2 | 1676 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 | 1677 | if (!cic) |
a6151c3a | 1678 | return false; |
719d3402 | 1679 | |
91fac317 | 1680 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
a6151c3a | 1681 | return cfqq == RQ_CFQQ(rq); |
da775265 JA |
1682 | } |
1683 | ||
812df48d DS |
1684 | static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1685 | { | |
1686 | del_timer(&cfqd->idle_slice_timer); | |
e98ef89b | 1687 | cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg); |
812df48d DS |
1688 | } |
1689 | ||
febffd61 JA |
1690 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
1691 | struct cfq_queue *cfqq) | |
22e2c507 JA |
1692 | { |
1693 | if (cfqq) { | |
b1ffe737 DS |
1694 | cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d", |
1695 | cfqd->serving_prio, cfqd->serving_type); | |
62a37f6b JT |
1696 | cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg); |
1697 | cfqq->slice_start = 0; | |
1698 | cfqq->dispatch_start = jiffies; | |
1699 | cfqq->allocated_slice = 0; | |
1700 | cfqq->slice_end = 0; | |
1701 | cfqq->slice_dispatch = 0; | |
1702 | cfqq->nr_sectors = 0; | |
1703 | ||
1704 | cfq_clear_cfqq_wait_request(cfqq); | |
1705 | cfq_clear_cfqq_must_dispatch(cfqq); | |
1706 | cfq_clear_cfqq_must_alloc_slice(cfqq); | |
1707 | cfq_clear_cfqq_fifo_expire(cfqq); | |
1708 | cfq_mark_cfqq_slice_new(cfqq); | |
1709 | ||
1710 | cfq_del_timer(cfqd, cfqq); | |
22e2c507 JA |
1711 | } |
1712 | ||
1713 | cfqd->active_queue = cfqq; | |
1714 | } | |
1715 | ||
7b14e3b5 JA |
1716 | /* |
1717 | * current cfqq expired its slice (or was too idle), select new one | |
1718 | */ | |
1719 | static void | |
1720 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
e5ff082e | 1721 | bool timed_out) |
7b14e3b5 | 1722 | { |
7b679138 JA |
1723 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
1724 | ||
7b14e3b5 | 1725 | if (cfq_cfqq_wait_request(cfqq)) |
812df48d | 1726 | cfq_del_timer(cfqd, cfqq); |
7b14e3b5 | 1727 | |
7b14e3b5 | 1728 | cfq_clear_cfqq_wait_request(cfqq); |
f75edf2d | 1729 | cfq_clear_cfqq_wait_busy(cfqq); |
7b14e3b5 | 1730 | |
ae54abed SL |
1731 | /* |
1732 | * If this cfqq is shared between multiple processes, check to | |
1733 | * make sure that those processes are still issuing I/Os within | |
1734 | * the mean seek distance. If not, it may be time to break the | |
1735 | * queues apart again. | |
1736 | */ | |
1737 | if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq)) | |
1738 | cfq_mark_cfqq_split_coop(cfqq); | |
1739 | ||
7b14e3b5 | 1740 | /* |
6084cdda | 1741 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 1742 | */ |
c553f8e3 SL |
1743 | if (timed_out) { |
1744 | if (cfq_cfqq_slice_new(cfqq)) | |
ba5bd520 | 1745 | cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq); |
c553f8e3 SL |
1746 | else |
1747 | cfqq->slice_resid = cfqq->slice_end - jiffies; | |
7b679138 JA |
1748 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
1749 | } | |
7b14e3b5 | 1750 | |
e5ff082e | 1751 | cfq_group_served(cfqd, cfqq->cfqg, cfqq); |
dae739eb | 1752 | |
f04a6424 VG |
1753 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
1754 | cfq_del_cfqq_rr(cfqd, cfqq); | |
1755 | ||
edd75ffd | 1756 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
1757 | |
1758 | if (cfqq == cfqd->active_queue) | |
1759 | cfqd->active_queue = NULL; | |
1760 | ||
1761 | if (cfqd->active_cic) { | |
1762 | put_io_context(cfqd->active_cic->ioc); | |
1763 | cfqd->active_cic = NULL; | |
1764 | } | |
7b14e3b5 JA |
1765 | } |
1766 | ||
e5ff082e | 1767 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) |
7b14e3b5 JA |
1768 | { |
1769 | struct cfq_queue *cfqq = cfqd->active_queue; | |
1770 | ||
1771 | if (cfqq) | |
e5ff082e | 1772 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
1773 | } |
1774 | ||
498d3aa2 JA |
1775 | /* |
1776 | * Get next queue for service. Unless we have a queue preemption, | |
1777 | * we'll simply select the first cfqq in the service tree. | |
1778 | */ | |
6d048f53 | 1779 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 1780 | { |
c0324a02 | 1781 | struct cfq_rb_root *service_tree = |
cdb16e8f | 1782 | service_tree_for(cfqd->serving_group, cfqd->serving_prio, |
65b32a57 | 1783 | cfqd->serving_type); |
d9e7620e | 1784 | |
f04a6424 VG |
1785 | if (!cfqd->rq_queued) |
1786 | return NULL; | |
1787 | ||
1fa8f6d6 VG |
1788 | /* There is nothing to dispatch */ |
1789 | if (!service_tree) | |
1790 | return NULL; | |
c0324a02 CZ |
1791 | if (RB_EMPTY_ROOT(&service_tree->rb)) |
1792 | return NULL; | |
1793 | return cfq_rb_first(service_tree); | |
6d048f53 JA |
1794 | } |
1795 | ||
f04a6424 VG |
1796 | static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) |
1797 | { | |
25fb5169 | 1798 | struct cfq_group *cfqg; |
f04a6424 VG |
1799 | struct cfq_queue *cfqq; |
1800 | int i, j; | |
1801 | struct cfq_rb_root *st; | |
1802 | ||
1803 | if (!cfqd->rq_queued) | |
1804 | return NULL; | |
1805 | ||
25fb5169 VG |
1806 | cfqg = cfq_get_next_cfqg(cfqd); |
1807 | if (!cfqg) | |
1808 | return NULL; | |
1809 | ||
f04a6424 VG |
1810 | for_each_cfqg_st(cfqg, i, j, st) |
1811 | if ((cfqq = cfq_rb_first(st)) != NULL) | |
1812 | return cfqq; | |
1813 | return NULL; | |
1814 | } | |
1815 | ||
498d3aa2 JA |
1816 | /* |
1817 | * Get and set a new active queue for service. | |
1818 | */ | |
a36e71f9 JA |
1819 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
1820 | struct cfq_queue *cfqq) | |
6d048f53 | 1821 | { |
e00ef799 | 1822 | if (!cfqq) |
a36e71f9 | 1823 | cfqq = cfq_get_next_queue(cfqd); |
6d048f53 | 1824 | |
22e2c507 | 1825 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 1826 | return cfqq; |
22e2c507 JA |
1827 | } |
1828 | ||
d9e7620e JA |
1829 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
1830 | struct request *rq) | |
1831 | { | |
83096ebf TH |
1832 | if (blk_rq_pos(rq) >= cfqd->last_position) |
1833 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 1834 | else |
83096ebf | 1835 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
1836 | } |
1837 | ||
b2c18e1e | 1838 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
e9ce335d | 1839 | struct request *rq) |
6d048f53 | 1840 | { |
e9ce335d | 1841 | return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR; |
6d048f53 JA |
1842 | } |
1843 | ||
a36e71f9 JA |
1844 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
1845 | struct cfq_queue *cur_cfqq) | |
1846 | { | |
f2d1f0ae | 1847 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
1848 | struct rb_node *parent, *node; |
1849 | struct cfq_queue *__cfqq; | |
1850 | sector_t sector = cfqd->last_position; | |
1851 | ||
1852 | if (RB_EMPTY_ROOT(root)) | |
1853 | return NULL; | |
1854 | ||
1855 | /* | |
1856 | * First, if we find a request starting at the end of the last | |
1857 | * request, choose it. | |
1858 | */ | |
f2d1f0ae | 1859 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
1860 | if (__cfqq) |
1861 | return __cfqq; | |
1862 | ||
1863 | /* | |
1864 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1865 | * will contain the closest sector. | |
1866 | */ | |
1867 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
e9ce335d | 1868 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1869 | return __cfqq; |
1870 | ||
2e46e8b2 | 1871 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1872 | node = rb_next(&__cfqq->p_node); |
1873 | else | |
1874 | node = rb_prev(&__cfqq->p_node); | |
1875 | if (!node) | |
1876 | return NULL; | |
1877 | ||
1878 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
e9ce335d | 1879 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) |
a36e71f9 JA |
1880 | return __cfqq; |
1881 | ||
1882 | return NULL; | |
1883 | } | |
1884 | ||
1885 | /* | |
1886 | * cfqd - obvious | |
1887 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1888 | * closely cooperating with itself. | |
1889 | * | |
1890 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1891 | * one request, and that cfqd->last_position reflects a position on the disk | |
1892 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1893 | * assumption. | |
1894 | */ | |
1895 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
b3b6d040 | 1896 | struct cfq_queue *cur_cfqq) |
6d048f53 | 1897 | { |
a36e71f9 JA |
1898 | struct cfq_queue *cfqq; |
1899 | ||
39c01b21 DS |
1900 | if (cfq_class_idle(cur_cfqq)) |
1901 | return NULL; | |
e6c5bc73 JM |
1902 | if (!cfq_cfqq_sync(cur_cfqq)) |
1903 | return NULL; | |
1904 | if (CFQQ_SEEKY(cur_cfqq)) | |
1905 | return NULL; | |
1906 | ||
b9d8f4c7 GJ |
1907 | /* |
1908 | * Don't search priority tree if it's the only queue in the group. | |
1909 | */ | |
1910 | if (cur_cfqq->cfqg->nr_cfqq == 1) | |
1911 | return NULL; | |
1912 | ||
6d048f53 | 1913 | /* |
d9e7620e JA |
1914 | * We should notice if some of the queues are cooperating, eg |
1915 | * working closely on the same area of the disk. In that case, | |
1916 | * we can group them together and don't waste time idling. | |
6d048f53 | 1917 | */ |
a36e71f9 JA |
1918 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1919 | if (!cfqq) | |
1920 | return NULL; | |
1921 | ||
8682e1f1 VG |
1922 | /* If new queue belongs to different cfq_group, don't choose it */ |
1923 | if (cur_cfqq->cfqg != cfqq->cfqg) | |
1924 | return NULL; | |
1925 | ||
df5fe3e8 JM |
1926 | /* |
1927 | * It only makes sense to merge sync queues. | |
1928 | */ | |
1929 | if (!cfq_cfqq_sync(cfqq)) | |
1930 | return NULL; | |
e6c5bc73 JM |
1931 | if (CFQQ_SEEKY(cfqq)) |
1932 | return NULL; | |
df5fe3e8 | 1933 | |
c0324a02 CZ |
1934 | /* |
1935 | * Do not merge queues of different priority classes | |
1936 | */ | |
1937 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
1938 | return NULL; | |
1939 | ||
a36e71f9 | 1940 | return cfqq; |
6d048f53 JA |
1941 | } |
1942 | ||
a6d44e98 CZ |
1943 | /* |
1944 | * Determine whether we should enforce idle window for this queue. | |
1945 | */ | |
1946 | ||
1947 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1948 | { | |
1949 | enum wl_prio_t prio = cfqq_prio(cfqq); | |
718eee05 | 1950 | struct cfq_rb_root *service_tree = cfqq->service_tree; |
a6d44e98 | 1951 | |
f04a6424 VG |
1952 | BUG_ON(!service_tree); |
1953 | BUG_ON(!service_tree->count); | |
1954 | ||
b6508c16 VG |
1955 | if (!cfqd->cfq_slice_idle) |
1956 | return false; | |
1957 | ||
a6d44e98 CZ |
1958 | /* We never do for idle class queues. */ |
1959 | if (prio == IDLE_WORKLOAD) | |
1960 | return false; | |
1961 | ||
1962 | /* We do for queues that were marked with idle window flag. */ | |
3c764b7a SL |
1963 | if (cfq_cfqq_idle_window(cfqq) && |
1964 | !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)) | |
a6d44e98 CZ |
1965 | return true; |
1966 | ||
1967 | /* | |
1968 | * Otherwise, we do only if they are the last ones | |
1969 | * in their service tree. | |
1970 | */ | |
b1ffe737 | 1971 | if (service_tree->count == 1 && cfq_cfqq_sync(cfqq)) |
c1e44756 | 1972 | return true; |
b1ffe737 DS |
1973 | cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", |
1974 | service_tree->count); | |
c1e44756 | 1975 | return false; |
a6d44e98 CZ |
1976 | } |
1977 | ||
6d048f53 | 1978 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1979 | { |
1792669c | 1980 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1981 | struct cfq_io_context *cic; |
80bdf0c7 | 1982 | unsigned long sl, group_idle = 0; |
7b14e3b5 | 1983 | |
a68bbddb | 1984 | /* |
f7d7b7a7 JA |
1985 | * SSD device without seek penalty, disable idling. But only do so |
1986 | * for devices that support queuing, otherwise we still have a problem | |
1987 | * with sync vs async workloads. | |
a68bbddb | 1988 | */ |
f7d7b7a7 | 1989 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
1990 | return; |
1991 | ||
dd67d051 | 1992 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 1993 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
1994 | |
1995 | /* | |
1996 | * idle is disabled, either manually or by past process history | |
1997 | */ | |
80bdf0c7 VG |
1998 | if (!cfq_should_idle(cfqd, cfqq)) { |
1999 | /* no queue idling. Check for group idling */ | |
2000 | if (cfqd->cfq_group_idle) | |
2001 | group_idle = cfqd->cfq_group_idle; | |
2002 | else | |
2003 | return; | |
2004 | } | |
6d048f53 | 2005 | |
7b679138 | 2006 | /* |
8e550632 | 2007 | * still active requests from this queue, don't idle |
7b679138 | 2008 | */ |
8e550632 | 2009 | if (cfqq->dispatched) |
7b679138 JA |
2010 | return; |
2011 | ||
22e2c507 JA |
2012 | /* |
2013 | * task has exited, don't wait | |
2014 | */ | |
206dc69b | 2015 | cic = cfqd->active_cic; |
66dac98e | 2016 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
2017 | return; |
2018 | ||
355b659c CZ |
2019 | /* |
2020 | * If our average think time is larger than the remaining time | |
2021 | * slice, then don't idle. This avoids overrunning the allotted | |
2022 | * time slice. | |
2023 | */ | |
2024 | if (sample_valid(cic->ttime_samples) && | |
b1ffe737 DS |
2025 | (cfqq->slice_end - jiffies < cic->ttime_mean)) { |
2026 | cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d", | |
2027 | cic->ttime_mean); | |
355b659c | 2028 | return; |
b1ffe737 | 2029 | } |
355b659c | 2030 | |
80bdf0c7 VG |
2031 | /* There are other queues in the group, don't do group idle */ |
2032 | if (group_idle && cfqq->cfqg->nr_cfqq > 1) | |
2033 | return; | |
2034 | ||
3b18152c | 2035 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 2036 | |
80bdf0c7 VG |
2037 | if (group_idle) |
2038 | sl = cfqd->cfq_group_idle; | |
2039 | else | |
2040 | sl = cfqd->cfq_slice_idle; | |
206dc69b | 2041 | |
7b14e3b5 | 2042 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
e98ef89b | 2043 | cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg); |
80bdf0c7 VG |
2044 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl, |
2045 | group_idle ? 1 : 0); | |
1da177e4 LT |
2046 | } |
2047 | ||
498d3aa2 JA |
2048 | /* |
2049 | * Move request from internal lists to the request queue dispatch list. | |
2050 | */ | |
165125e1 | 2051 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 2052 | { |
3ed9a296 | 2053 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 2054 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 2055 | |
7b679138 JA |
2056 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
2057 | ||
06d21886 | 2058 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); |
5380a101 | 2059 | cfq_remove_request(rq); |
6d048f53 | 2060 | cfqq->dispatched++; |
80bdf0c7 | 2061 | (RQ_CFQG(rq))->dispatched++; |
5380a101 | 2062 | elv_dispatch_sort(q, rq); |
3ed9a296 | 2063 | |
53c583d2 | 2064 | cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++; |
c4e7893e | 2065 | cfqq->nr_sectors += blk_rq_sectors(rq); |
e98ef89b | 2066 | cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq), |
84c124da | 2067 | rq_data_dir(rq), rq_is_sync(rq)); |
1da177e4 LT |
2068 | } |
2069 | ||
2070 | /* | |
2071 | * return expired entry, or NULL to just start from scratch in rbtree | |
2072 | */ | |
febffd61 | 2073 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 | 2074 | { |
30996f40 | 2075 | struct request *rq = NULL; |
1da177e4 | 2076 | |
3b18152c | 2077 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 2078 | return NULL; |
cb887411 JA |
2079 | |
2080 | cfq_mark_cfqq_fifo_expire(cfqq); | |
2081 | ||
89850f7e JA |
2082 | if (list_empty(&cfqq->fifo)) |
2083 | return NULL; | |
1da177e4 | 2084 | |
89850f7e | 2085 | rq = rq_entry_fifo(cfqq->fifo.next); |
30996f40 | 2086 | if (time_before(jiffies, rq_fifo_time(rq))) |
7b679138 | 2087 | rq = NULL; |
1da177e4 | 2088 | |
30996f40 | 2089 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 2090 | return rq; |
1da177e4 LT |
2091 | } |
2092 | ||
22e2c507 JA |
2093 | static inline int |
2094 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2095 | { | |
2096 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 2097 | |
22e2c507 | 2098 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 2099 | |
b9f8ce05 | 2100 | return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio); |
1da177e4 LT |
2101 | } |
2102 | ||
df5fe3e8 JM |
2103 | /* |
2104 | * Must be called with the queue_lock held. | |
2105 | */ | |
2106 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
2107 | { | |
2108 | int process_refs, io_refs; | |
2109 | ||
2110 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
30d7b944 | 2111 | process_refs = cfqq->ref - io_refs; |
df5fe3e8 JM |
2112 | BUG_ON(process_refs < 0); |
2113 | return process_refs; | |
2114 | } | |
2115 | ||
2116 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
2117 | { | |
e6c5bc73 | 2118 | int process_refs, new_process_refs; |
df5fe3e8 JM |
2119 | struct cfq_queue *__cfqq; |
2120 | ||
c10b61f0 JM |
2121 | /* |
2122 | * If there are no process references on the new_cfqq, then it is | |
2123 | * unsafe to follow the ->new_cfqq chain as other cfqq's in the | |
2124 | * chain may have dropped their last reference (not just their | |
2125 | * last process reference). | |
2126 | */ | |
2127 | if (!cfqq_process_refs(new_cfqq)) | |
2128 | return; | |
2129 | ||
df5fe3e8 JM |
2130 | /* Avoid a circular list and skip interim queue merges */ |
2131 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
2132 | if (__cfqq == cfqq) | |
2133 | return; | |
2134 | new_cfqq = __cfqq; | |
2135 | } | |
2136 | ||
2137 | process_refs = cfqq_process_refs(cfqq); | |
c10b61f0 | 2138 | new_process_refs = cfqq_process_refs(new_cfqq); |
df5fe3e8 JM |
2139 | /* |
2140 | * If the process for the cfqq has gone away, there is no | |
2141 | * sense in merging the queues. | |
2142 | */ | |
c10b61f0 | 2143 | if (process_refs == 0 || new_process_refs == 0) |
df5fe3e8 JM |
2144 | return; |
2145 | ||
e6c5bc73 JM |
2146 | /* |
2147 | * Merge in the direction of the lesser amount of work. | |
2148 | */ | |
e6c5bc73 JM |
2149 | if (new_process_refs >= process_refs) { |
2150 | cfqq->new_cfqq = new_cfqq; | |
30d7b944 | 2151 | new_cfqq->ref += process_refs; |
e6c5bc73 JM |
2152 | } else { |
2153 | new_cfqq->new_cfqq = cfqq; | |
30d7b944 | 2154 | cfqq->ref += new_process_refs; |
e6c5bc73 | 2155 | } |
df5fe3e8 JM |
2156 | } |
2157 | ||
cdb16e8f | 2158 | static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, |
65b32a57 | 2159 | struct cfq_group *cfqg, enum wl_prio_t prio) |
718eee05 CZ |
2160 | { |
2161 | struct cfq_queue *queue; | |
2162 | int i; | |
2163 | bool key_valid = false; | |
2164 | unsigned long lowest_key = 0; | |
2165 | enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; | |
2166 | ||
65b32a57 VG |
2167 | for (i = 0; i <= SYNC_WORKLOAD; ++i) { |
2168 | /* select the one with lowest rb_key */ | |
2169 | queue = cfq_rb_first(service_tree_for(cfqg, prio, i)); | |
718eee05 CZ |
2170 | if (queue && |
2171 | (!key_valid || time_before(queue->rb_key, lowest_key))) { | |
2172 | lowest_key = queue->rb_key; | |
2173 | cur_best = i; | |
2174 | key_valid = true; | |
2175 | } | |
2176 | } | |
2177 | ||
2178 | return cur_best; | |
2179 | } | |
2180 | ||
cdb16e8f | 2181 | static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) |
718eee05 | 2182 | { |
718eee05 CZ |
2183 | unsigned slice; |
2184 | unsigned count; | |
cdb16e8f | 2185 | struct cfq_rb_root *st; |
58ff82f3 | 2186 | unsigned group_slice; |
e4ea0c16 | 2187 | enum wl_prio_t original_prio = cfqd->serving_prio; |
1fa8f6d6 | 2188 | |
718eee05 | 2189 | /* Choose next priority. RT > BE > IDLE */ |
58ff82f3 | 2190 | if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) |
718eee05 | 2191 | cfqd->serving_prio = RT_WORKLOAD; |
58ff82f3 | 2192 | else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) |
718eee05 CZ |
2193 | cfqd->serving_prio = BE_WORKLOAD; |
2194 | else { | |
2195 | cfqd->serving_prio = IDLE_WORKLOAD; | |
2196 | cfqd->workload_expires = jiffies + 1; | |
2197 | return; | |
2198 | } | |
2199 | ||
e4ea0c16 SL |
2200 | if (original_prio != cfqd->serving_prio) |
2201 | goto new_workload; | |
2202 | ||
718eee05 CZ |
2203 | /* |
2204 | * For RT and BE, we have to choose also the type | |
2205 | * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload | |
2206 | * expiration time | |
2207 | */ | |
65b32a57 | 2208 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); |
cdb16e8f | 2209 | count = st->count; |
718eee05 CZ |
2210 | |
2211 | /* | |
65b32a57 | 2212 | * check workload expiration, and that we still have other queues ready |
718eee05 | 2213 | */ |
65b32a57 | 2214 | if (count && !time_after(jiffies, cfqd->workload_expires)) |
718eee05 CZ |
2215 | return; |
2216 | ||
e4ea0c16 | 2217 | new_workload: |
718eee05 CZ |
2218 | /* otherwise select new workload type */ |
2219 | cfqd->serving_type = | |
65b32a57 VG |
2220 | cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio); |
2221 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); | |
cdb16e8f | 2222 | count = st->count; |
718eee05 CZ |
2223 | |
2224 | /* | |
2225 | * the workload slice is computed as a fraction of target latency | |
2226 | * proportional to the number of queues in that workload, over | |
2227 | * all the queues in the same priority class | |
2228 | */ | |
58ff82f3 VG |
2229 | group_slice = cfq_group_slice(cfqd, cfqg); |
2230 | ||
2231 | slice = group_slice * count / | |
2232 | max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], | |
2233 | cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); | |
718eee05 | 2234 | |
f26bd1f0 VG |
2235 | if (cfqd->serving_type == ASYNC_WORKLOAD) { |
2236 | unsigned int tmp; | |
2237 | ||
2238 | /* | |
2239 | * Async queues are currently system wide. Just taking | |
2240 | * proportion of queues with-in same group will lead to higher | |
2241 | * async ratio system wide as generally root group is going | |
2242 | * to have higher weight. A more accurate thing would be to | |
2243 | * calculate system wide asnc/sync ratio. | |
2244 | */ | |
2245 | tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg); | |
2246 | tmp = tmp/cfqd->busy_queues; | |
2247 | slice = min_t(unsigned, slice, tmp); | |
2248 | ||
718eee05 CZ |
2249 | /* async workload slice is scaled down according to |
2250 | * the sync/async slice ratio. */ | |
2251 | slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; | |
f26bd1f0 | 2252 | } else |
718eee05 CZ |
2253 | /* sync workload slice is at least 2 * cfq_slice_idle */ |
2254 | slice = max(slice, 2 * cfqd->cfq_slice_idle); | |
2255 | ||
2256 | slice = max_t(unsigned, slice, CFQ_MIN_TT); | |
b1ffe737 | 2257 | cfq_log(cfqd, "workload slice:%d", slice); |
718eee05 CZ |
2258 | cfqd->workload_expires = jiffies + slice; |
2259 | } | |
2260 | ||
1fa8f6d6 VG |
2261 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) |
2262 | { | |
2263 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
25bc6b07 | 2264 | struct cfq_group *cfqg; |
1fa8f6d6 VG |
2265 | |
2266 | if (RB_EMPTY_ROOT(&st->rb)) | |
2267 | return NULL; | |
25bc6b07 | 2268 | cfqg = cfq_rb_first_group(st); |
25bc6b07 VG |
2269 | update_min_vdisktime(st); |
2270 | return cfqg; | |
1fa8f6d6 VG |
2271 | } |
2272 | ||
cdb16e8f VG |
2273 | static void cfq_choose_cfqg(struct cfq_data *cfqd) |
2274 | { | |
1fa8f6d6 VG |
2275 | struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); |
2276 | ||
2277 | cfqd->serving_group = cfqg; | |
dae739eb VG |
2278 | |
2279 | /* Restore the workload type data */ | |
2280 | if (cfqg->saved_workload_slice) { | |
2281 | cfqd->workload_expires = jiffies + cfqg->saved_workload_slice; | |
2282 | cfqd->serving_type = cfqg->saved_workload; | |
2283 | cfqd->serving_prio = cfqg->saved_serving_prio; | |
66ae2919 GJ |
2284 | } else |
2285 | cfqd->workload_expires = jiffies - 1; | |
2286 | ||
1fa8f6d6 | 2287 | choose_service_tree(cfqd, cfqg); |
cdb16e8f VG |
2288 | } |
2289 | ||
22e2c507 | 2290 | /* |
498d3aa2 JA |
2291 | * Select a queue for service. If we have a current active queue, |
2292 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 2293 | */ |
1b5ed5e1 | 2294 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 2295 | { |
a36e71f9 | 2296 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 2297 | |
22e2c507 JA |
2298 | cfqq = cfqd->active_queue; |
2299 | if (!cfqq) | |
2300 | goto new_queue; | |
1da177e4 | 2301 | |
f04a6424 VG |
2302 | if (!cfqd->rq_queued) |
2303 | return NULL; | |
c244bb50 VG |
2304 | |
2305 | /* | |
2306 | * We were waiting for group to get backlogged. Expire the queue | |
2307 | */ | |
2308 | if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list)) | |
2309 | goto expire; | |
2310 | ||
22e2c507 | 2311 | /* |
6d048f53 | 2312 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 2313 | */ |
7667aa06 VG |
2314 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) { |
2315 | /* | |
2316 | * If slice had not expired at the completion of last request | |
2317 | * we might not have turned on wait_busy flag. Don't expire | |
2318 | * the queue yet. Allow the group to get backlogged. | |
2319 | * | |
2320 | * The very fact that we have used the slice, that means we | |
2321 | * have been idling all along on this queue and it should be | |
2322 | * ok to wait for this request to complete. | |
2323 | */ | |
82bbbf28 VG |
2324 | if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list) |
2325 | && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { | |
2326 | cfqq = NULL; | |
7667aa06 | 2327 | goto keep_queue; |
82bbbf28 | 2328 | } else |
80bdf0c7 | 2329 | goto check_group_idle; |
7667aa06 | 2330 | } |
1da177e4 | 2331 | |
22e2c507 | 2332 | /* |
6d048f53 JA |
2333 | * The active queue has requests and isn't expired, allow it to |
2334 | * dispatch. | |
22e2c507 | 2335 | */ |
dd67d051 | 2336 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 2337 | goto keep_queue; |
6d048f53 | 2338 | |
a36e71f9 JA |
2339 | /* |
2340 | * If another queue has a request waiting within our mean seek | |
2341 | * distance, let it run. The expire code will check for close | |
2342 | * cooperators and put the close queue at the front of the service | |
df5fe3e8 | 2343 | * tree. If possible, merge the expiring queue with the new cfqq. |
a36e71f9 | 2344 | */ |
b3b6d040 | 2345 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); |
df5fe3e8 JM |
2346 | if (new_cfqq) { |
2347 | if (!cfqq->new_cfqq) | |
2348 | cfq_setup_merge(cfqq, new_cfqq); | |
a36e71f9 | 2349 | goto expire; |
df5fe3e8 | 2350 | } |
a36e71f9 | 2351 | |
6d048f53 JA |
2352 | /* |
2353 | * No requests pending. If the active queue still has requests in | |
2354 | * flight or is idling for a new request, allow either of these | |
2355 | * conditions to happen (or time out) before selecting a new queue. | |
2356 | */ | |
80bdf0c7 VG |
2357 | if (timer_pending(&cfqd->idle_slice_timer)) { |
2358 | cfqq = NULL; | |
2359 | goto keep_queue; | |
2360 | } | |
2361 | ||
8e1ac665 SL |
2362 | /* |
2363 | * This is a deep seek queue, but the device is much faster than | |
2364 | * the queue can deliver, don't idle | |
2365 | **/ | |
2366 | if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) && | |
2367 | (cfq_cfqq_slice_new(cfqq) || | |
2368 | (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) { | |
2369 | cfq_clear_cfqq_deep(cfqq); | |
2370 | cfq_clear_cfqq_idle_window(cfqq); | |
2371 | } | |
2372 | ||
80bdf0c7 VG |
2373 | if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { |
2374 | cfqq = NULL; | |
2375 | goto keep_queue; | |
2376 | } | |
2377 | ||
2378 | /* | |
2379 | * If group idle is enabled and there are requests dispatched from | |
2380 | * this group, wait for requests to complete. | |
2381 | */ | |
2382 | check_group_idle: | |
2383 | if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 | |
2384 | && cfqq->cfqg->dispatched) { | |
caaa5f9f JA |
2385 | cfqq = NULL; |
2386 | goto keep_queue; | |
22e2c507 JA |
2387 | } |
2388 | ||
3b18152c | 2389 | expire: |
e5ff082e | 2390 | cfq_slice_expired(cfqd, 0); |
3b18152c | 2391 | new_queue: |
718eee05 CZ |
2392 | /* |
2393 | * Current queue expired. Check if we have to switch to a new | |
2394 | * service tree | |
2395 | */ | |
2396 | if (!new_cfqq) | |
cdb16e8f | 2397 | cfq_choose_cfqg(cfqd); |
718eee05 | 2398 | |
a36e71f9 | 2399 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 2400 | keep_queue: |
3b18152c | 2401 | return cfqq; |
22e2c507 JA |
2402 | } |
2403 | ||
febffd61 | 2404 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
2405 | { |
2406 | int dispatched = 0; | |
2407 | ||
2408 | while (cfqq->next_rq) { | |
2409 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
2410 | dispatched++; | |
2411 | } | |
2412 | ||
2413 | BUG_ON(!list_empty(&cfqq->fifo)); | |
f04a6424 VG |
2414 | |
2415 | /* By default cfqq is not expired if it is empty. Do it explicitly */ | |
e5ff082e | 2416 | __cfq_slice_expired(cfqq->cfqd, cfqq, 0); |
d9e7620e JA |
2417 | return dispatched; |
2418 | } | |
2419 | ||
498d3aa2 JA |
2420 | /* |
2421 | * Drain our current requests. Used for barriers and when switching | |
2422 | * io schedulers on-the-fly. | |
2423 | */ | |
d9e7620e | 2424 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 2425 | { |
0871714e | 2426 | struct cfq_queue *cfqq; |
d9e7620e | 2427 | int dispatched = 0; |
cdb16e8f | 2428 | |
3440c49f | 2429 | /* Expire the timeslice of the current active queue first */ |
e5ff082e | 2430 | cfq_slice_expired(cfqd, 0); |
3440c49f DS |
2431 | while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) { |
2432 | __cfq_set_active_queue(cfqd, cfqq); | |
f04a6424 | 2433 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); |
3440c49f | 2434 | } |
1b5ed5e1 | 2435 | |
1b5ed5e1 TH |
2436 | BUG_ON(cfqd->busy_queues); |
2437 | ||
6923715a | 2438 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
2439 | return dispatched; |
2440 | } | |
2441 | ||
abc3c744 SL |
2442 | static inline bool cfq_slice_used_soon(struct cfq_data *cfqd, |
2443 | struct cfq_queue *cfqq) | |
2444 | { | |
2445 | /* the queue hasn't finished any request, can't estimate */ | |
2446 | if (cfq_cfqq_slice_new(cfqq)) | |
c1e44756 | 2447 | return true; |
abc3c744 SL |
2448 | if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched, |
2449 | cfqq->slice_end)) | |
c1e44756 | 2450 | return true; |
abc3c744 | 2451 | |
c1e44756 | 2452 | return false; |
abc3c744 SL |
2453 | } |
2454 | ||
0b182d61 | 2455 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
2f5cb738 | 2456 | { |
2f5cb738 | 2457 | unsigned int max_dispatch; |
22e2c507 | 2458 | |
5ad531db JA |
2459 | /* |
2460 | * Drain async requests before we start sync IO | |
2461 | */ | |
53c583d2 | 2462 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC]) |
0b182d61 | 2463 | return false; |
5ad531db | 2464 | |
2f5cb738 JA |
2465 | /* |
2466 | * If this is an async queue and we have sync IO in flight, let it wait | |
2467 | */ | |
53c583d2 | 2468 | if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq)) |
0b182d61 | 2469 | return false; |
2f5cb738 | 2470 | |
abc3c744 | 2471 | max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1); |
2f5cb738 JA |
2472 | if (cfq_class_idle(cfqq)) |
2473 | max_dispatch = 1; | |
b4878f24 | 2474 | |
2f5cb738 JA |
2475 | /* |
2476 | * Does this cfqq already have too much IO in flight? | |
2477 | */ | |
2478 | if (cfqq->dispatched >= max_dispatch) { | |
ef8a41df | 2479 | bool promote_sync = false; |
2f5cb738 JA |
2480 | /* |
2481 | * idle queue must always only have a single IO in flight | |
2482 | */ | |
3ed9a296 | 2483 | if (cfq_class_idle(cfqq)) |
0b182d61 | 2484 | return false; |
3ed9a296 | 2485 | |
ef8a41df | 2486 | /* |
c4ade94f LS |
2487 | * If there is only one sync queue |
2488 | * we can ignore async queue here and give the sync | |
ef8a41df SL |
2489 | * queue no dispatch limit. The reason is a sync queue can |
2490 | * preempt async queue, limiting the sync queue doesn't make | |
2491 | * sense. This is useful for aiostress test. | |
2492 | */ | |
c4ade94f LS |
2493 | if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1) |
2494 | promote_sync = true; | |
ef8a41df | 2495 | |
2f5cb738 JA |
2496 | /* |
2497 | * We have other queues, don't allow more IO from this one | |
2498 | */ | |
ef8a41df SL |
2499 | if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) && |
2500 | !promote_sync) | |
0b182d61 | 2501 | return false; |
9ede209e | 2502 | |
365722bb | 2503 | /* |
474b18cc | 2504 | * Sole queue user, no limit |
365722bb | 2505 | */ |
ef8a41df | 2506 | if (cfqd->busy_queues == 1 || promote_sync) |
abc3c744 SL |
2507 | max_dispatch = -1; |
2508 | else | |
2509 | /* | |
2510 | * Normally we start throttling cfqq when cfq_quantum/2 | |
2511 | * requests have been dispatched. But we can drive | |
2512 | * deeper queue depths at the beginning of slice | |
2513 | * subjected to upper limit of cfq_quantum. | |
2514 | * */ | |
2515 | max_dispatch = cfqd->cfq_quantum; | |
8e296755 JA |
2516 | } |
2517 | ||
2518 | /* | |
2519 | * Async queues must wait a bit before being allowed dispatch. | |
2520 | * We also ramp up the dispatch depth gradually for async IO, | |
2521 | * based on the last sync IO we serviced | |
2522 | */ | |
963b72fc | 2523 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { |
573412b2 | 2524 | unsigned long last_sync = jiffies - cfqd->last_delayed_sync; |
8e296755 | 2525 | unsigned int depth; |
365722bb | 2526 | |
61f0c1dc | 2527 | depth = last_sync / cfqd->cfq_slice[1]; |
e00c54c3 JA |
2528 | if (!depth && !cfqq->dispatched) |
2529 | depth = 1; | |
8e296755 JA |
2530 | if (depth < max_dispatch) |
2531 | max_dispatch = depth; | |
2f5cb738 | 2532 | } |
3ed9a296 | 2533 | |
0b182d61 JA |
2534 | /* |
2535 | * If we're below the current max, allow a dispatch | |
2536 | */ | |
2537 | return cfqq->dispatched < max_dispatch; | |
2538 | } | |
2539 | ||
2540 | /* | |
2541 | * Dispatch a request from cfqq, moving them to the request queue | |
2542 | * dispatch list. | |
2543 | */ | |
2544 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2545 | { | |
2546 | struct request *rq; | |
2547 | ||
2548 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
2549 | ||
2550 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
2551 | return false; | |
2552 | ||
2553 | /* | |
2554 | * follow expired path, else get first next available | |
2555 | */ | |
2556 | rq = cfq_check_fifo(cfqq); | |
2557 | if (!rq) | |
2558 | rq = cfqq->next_rq; | |
2559 | ||
2560 | /* | |
2561 | * insert request into driver dispatch list | |
2562 | */ | |
2563 | cfq_dispatch_insert(cfqd->queue, rq); | |
2564 | ||
2565 | if (!cfqd->active_cic) { | |
2566 | struct cfq_io_context *cic = RQ_CIC(rq); | |
2567 | ||
2568 | atomic_long_inc(&cic->ioc->refcount); | |
2569 | cfqd->active_cic = cic; | |
2570 | } | |
2571 | ||
2572 | return true; | |
2573 | } | |
2574 | ||
2575 | /* | |
2576 | * Find the cfqq that we need to service and move a request from that to the | |
2577 | * dispatch list | |
2578 | */ | |
2579 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
2580 | { | |
2581 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2582 | struct cfq_queue *cfqq; | |
2583 | ||
2584 | if (!cfqd->busy_queues) | |
2585 | return 0; | |
2586 | ||
2587 | if (unlikely(force)) | |
2588 | return cfq_forced_dispatch(cfqd); | |
2589 | ||
2590 | cfqq = cfq_select_queue(cfqd); | |
2591 | if (!cfqq) | |
8e296755 JA |
2592 | return 0; |
2593 | ||
2f5cb738 | 2594 | /* |
0b182d61 | 2595 | * Dispatch a request from this cfqq, if it is allowed |
2f5cb738 | 2596 | */ |
0b182d61 JA |
2597 | if (!cfq_dispatch_request(cfqd, cfqq)) |
2598 | return 0; | |
2599 | ||
2f5cb738 | 2600 | cfqq->slice_dispatch++; |
b029195d | 2601 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 2602 | |
2f5cb738 JA |
2603 | /* |
2604 | * expire an async queue immediately if it has used up its slice. idle | |
2605 | * queue always expire after 1 dispatch round. | |
2606 | */ | |
2607 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
2608 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
2609 | cfq_class_idle(cfqq))) { | |
2610 | cfqq->slice_end = jiffies + 1; | |
e5ff082e | 2611 | cfq_slice_expired(cfqd, 0); |
1da177e4 LT |
2612 | } |
2613 | ||
b217a903 | 2614 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); |
2f5cb738 | 2615 | return 1; |
1da177e4 LT |
2616 | } |
2617 | ||
1da177e4 | 2618 | /* |
5e705374 JA |
2619 | * task holds one reference to the queue, dropped when task exits. each rq |
2620 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 | 2621 | * |
b1c35769 | 2622 | * Each cfq queue took a reference on the parent group. Drop it now. |
1da177e4 LT |
2623 | * queue lock must be held here. |
2624 | */ | |
2625 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
2626 | { | |
22e2c507 | 2627 | struct cfq_data *cfqd = cfqq->cfqd; |
0bbfeb83 | 2628 | struct cfq_group *cfqg; |
22e2c507 | 2629 | |
30d7b944 | 2630 | BUG_ON(cfqq->ref <= 0); |
1da177e4 | 2631 | |
30d7b944 SL |
2632 | cfqq->ref--; |
2633 | if (cfqq->ref) | |
1da177e4 LT |
2634 | return; |
2635 | ||
7b679138 | 2636 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 2637 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 2638 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
b1c35769 | 2639 | cfqg = cfqq->cfqg; |
1da177e4 | 2640 | |
28f95cbc | 2641 | if (unlikely(cfqd->active_queue == cfqq)) { |
e5ff082e | 2642 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 2643 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 2644 | } |
22e2c507 | 2645 | |
f04a6424 | 2646 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 | 2647 | kmem_cache_free(cfq_pool, cfqq); |
b1c35769 | 2648 | cfq_put_cfqg(cfqg); |
1da177e4 LT |
2649 | } |
2650 | ||
d6de8be7 | 2651 | /* |
5f45c695 | 2652 | * Call func for each cic attached to this ioc. |
d6de8be7 | 2653 | */ |
07416d29 | 2654 | static void |
5f45c695 JA |
2655 | call_for_each_cic(struct io_context *ioc, |
2656 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
07416d29 JA |
2657 | { |
2658 | struct cfq_io_context *cic; | |
2659 | struct hlist_node *n; | |
2660 | ||
5f45c695 JA |
2661 | rcu_read_lock(); |
2662 | ||
07416d29 JA |
2663 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) |
2664 | func(ioc, cic); | |
07416d29 | 2665 | |
4ac845a2 | 2666 | rcu_read_unlock(); |
34e6bbf2 FC |
2667 | } |
2668 | ||
2669 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
2670 | { | |
2671 | struct cfq_io_context *cic; | |
2672 | ||
2673 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
2674 | ||
2675 | kmem_cache_free(cfq_ioc_pool, cic); | |
245b2e70 | 2676 | elv_ioc_count_dec(cfq_ioc_count); |
34e6bbf2 | 2677 | |
9a11b4ed JA |
2678 | if (ioc_gone) { |
2679 | /* | |
2680 | * CFQ scheduler is exiting, grab exit lock and check | |
2681 | * the pending io context count. If it hits zero, | |
2682 | * complete ioc_gone and set it back to NULL | |
2683 | */ | |
2684 | spin_lock(&ioc_gone_lock); | |
245b2e70 | 2685 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { |
9a11b4ed JA |
2686 | complete(ioc_gone); |
2687 | ioc_gone = NULL; | |
2688 | } | |
2689 | spin_unlock(&ioc_gone_lock); | |
2690 | } | |
34e6bbf2 | 2691 | } |
4ac845a2 | 2692 | |
34e6bbf2 FC |
2693 | static void cfq_cic_free(struct cfq_io_context *cic) |
2694 | { | |
2695 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
2696 | } |
2697 | ||
2698 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
2699 | { | |
2700 | unsigned long flags; | |
bca4b914 | 2701 | unsigned long dead_key = (unsigned long) cic->key; |
4ac845a2 | 2702 | |
bca4b914 | 2703 | BUG_ON(!(dead_key & CIC_DEAD_KEY)); |
4ac845a2 JA |
2704 | |
2705 | spin_lock_irqsave(&ioc->lock, flags); | |
80b15c73 | 2706 | radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT); |
ffc4e759 | 2707 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2708 | spin_unlock_irqrestore(&ioc->lock, flags); |
2709 | ||
34e6bbf2 | 2710 | cfq_cic_free(cic); |
4ac845a2 JA |
2711 | } |
2712 | ||
d6de8be7 JA |
2713 | /* |
2714 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
2715 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
2716 | * and ->trim() which is called with the task lock held | |
2717 | */ | |
4ac845a2 JA |
2718 | static void cfq_free_io_context(struct io_context *ioc) |
2719 | { | |
4ac845a2 | 2720 | /* |
34e6bbf2 FC |
2721 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
2722 | * so no more cic's are allowed to be linked into this ioc. So it | |
2723 | * should be ok to iterate over the known list, we will see all cic's | |
2724 | * since no new ones are added. | |
4ac845a2 | 2725 | */ |
5f45c695 | 2726 | call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
2727 | } |
2728 | ||
d02a2c07 | 2729 | static void cfq_put_cooperator(struct cfq_queue *cfqq) |
1da177e4 | 2730 | { |
df5fe3e8 JM |
2731 | struct cfq_queue *__cfqq, *next; |
2732 | ||
df5fe3e8 JM |
2733 | /* |
2734 | * If this queue was scheduled to merge with another queue, be | |
2735 | * sure to drop the reference taken on that queue (and others in | |
2736 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
2737 | */ | |
2738 | __cfqq = cfqq->new_cfqq; | |
2739 | while (__cfqq) { | |
2740 | if (__cfqq == cfqq) { | |
2741 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
2742 | break; | |
2743 | } | |
2744 | next = __cfqq->new_cfqq; | |
2745 | cfq_put_queue(__cfqq); | |
2746 | __cfqq = next; | |
2747 | } | |
d02a2c07 SL |
2748 | } |
2749 | ||
2750 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2751 | { | |
2752 | if (unlikely(cfqq == cfqd->active_queue)) { | |
2753 | __cfq_slice_expired(cfqd, cfqq, 0); | |
2754 | cfq_schedule_dispatch(cfqd); | |
2755 | } | |
2756 | ||
2757 | cfq_put_cooperator(cfqq); | |
df5fe3e8 | 2758 | |
89850f7e JA |
2759 | cfq_put_queue(cfqq); |
2760 | } | |
22e2c507 | 2761 | |
89850f7e JA |
2762 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
2763 | struct cfq_io_context *cic) | |
2764 | { | |
4faa3c81 FC |
2765 | struct io_context *ioc = cic->ioc; |
2766 | ||
fc46379d | 2767 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
2768 | |
2769 | /* | |
bca4b914 | 2770 | * Make sure dead mark is seen for dead queues |
4ac845a2 | 2771 | */ |
fc46379d | 2772 | smp_wmb(); |
bca4b914 | 2773 | cic->key = cfqd_dead_key(cfqd); |
fc46379d | 2774 | |
4faa3c81 FC |
2775 | if (ioc->ioc_data == cic) |
2776 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
2777 | ||
ff6657c6 JA |
2778 | if (cic->cfqq[BLK_RW_ASYNC]) { |
2779 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
2780 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
2781 | } |
2782 | ||
ff6657c6 JA |
2783 | if (cic->cfqq[BLK_RW_SYNC]) { |
2784 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
2785 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 2786 | } |
89850f7e JA |
2787 | } |
2788 | ||
4ac845a2 JA |
2789 | static void cfq_exit_single_io_context(struct io_context *ioc, |
2790 | struct cfq_io_context *cic) | |
89850f7e | 2791 | { |
bca4b914 | 2792 | struct cfq_data *cfqd = cic_to_cfqd(cic); |
89850f7e | 2793 | |
89850f7e | 2794 | if (cfqd) { |
165125e1 | 2795 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 2796 | unsigned long flags; |
89850f7e | 2797 | |
4ac845a2 | 2798 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
2799 | |
2800 | /* | |
2801 | * Ensure we get a fresh copy of the ->key to prevent | |
2802 | * race between exiting task and queue | |
2803 | */ | |
2804 | smp_read_barrier_depends(); | |
bca4b914 | 2805 | if (cic->key == cfqd) |
62c1fe9d JA |
2806 | __cfq_exit_single_io_context(cfqd, cic); |
2807 | ||
4ac845a2 | 2808 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 2809 | } |
1da177e4 LT |
2810 | } |
2811 | ||
498d3aa2 JA |
2812 | /* |
2813 | * The process that ioc belongs to has exited, we need to clean up | |
2814 | * and put the internal structures we have that belongs to that process. | |
2815 | */ | |
e2d74ac0 | 2816 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 2817 | { |
4ac845a2 | 2818 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
2819 | } |
2820 | ||
22e2c507 | 2821 | static struct cfq_io_context * |
8267e268 | 2822 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2823 | { |
b5deef90 | 2824 | struct cfq_io_context *cic; |
1da177e4 | 2825 | |
94f6030c CL |
2826 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
2827 | cfqd->queue->node); | |
1da177e4 | 2828 | if (cic) { |
22e2c507 | 2829 | cic->last_end_request = jiffies; |
553698f9 | 2830 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 2831 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
2832 | cic->dtor = cfq_free_io_context; |
2833 | cic->exit = cfq_exit_io_context; | |
245b2e70 | 2834 | elv_ioc_count_inc(cfq_ioc_count); |
1da177e4 LT |
2835 | } |
2836 | ||
2837 | return cic; | |
2838 | } | |
2839 | ||
fd0928df | 2840 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
2841 | { |
2842 | struct task_struct *tsk = current; | |
2843 | int ioprio_class; | |
2844 | ||
3b18152c | 2845 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
2846 | return; |
2847 | ||
fd0928df | 2848 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 2849 | switch (ioprio_class) { |
fe094d98 JA |
2850 | default: |
2851 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
2852 | case IOPRIO_CLASS_NONE: | |
2853 | /* | |
6d63c275 | 2854 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
2855 | */ |
2856 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 2857 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
2858 | break; |
2859 | case IOPRIO_CLASS_RT: | |
2860 | cfqq->ioprio = task_ioprio(ioc); | |
2861 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
2862 | break; | |
2863 | case IOPRIO_CLASS_BE: | |
2864 | cfqq->ioprio = task_ioprio(ioc); | |
2865 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2866 | break; | |
2867 | case IOPRIO_CLASS_IDLE: | |
2868 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
2869 | cfqq->ioprio = 7; | |
2870 | cfq_clear_cfqq_idle_window(cfqq); | |
2871 | break; | |
22e2c507 JA |
2872 | } |
2873 | ||
2874 | /* | |
2875 | * keep track of original prio settings in case we have to temporarily | |
2876 | * elevate the priority of this queue | |
2877 | */ | |
2878 | cfqq->org_ioprio = cfqq->ioprio; | |
2879 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 2880 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
2881 | } |
2882 | ||
febffd61 | 2883 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 2884 | { |
bca4b914 | 2885 | struct cfq_data *cfqd = cic_to_cfqd(cic); |
478a82b0 | 2886 | struct cfq_queue *cfqq; |
c1b707d2 | 2887 | unsigned long flags; |
35e6077c | 2888 | |
caaa5f9f JA |
2889 | if (unlikely(!cfqd)) |
2890 | return; | |
2891 | ||
c1b707d2 | 2892 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 2893 | |
ff6657c6 | 2894 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
2895 | if (cfqq) { |
2896 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
2897 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
2898 | GFP_ATOMIC); | |
caaa5f9f | 2899 | if (new_cfqq) { |
ff6657c6 | 2900 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
2901 | cfq_put_queue(cfqq); |
2902 | } | |
22e2c507 | 2903 | } |
caaa5f9f | 2904 | |
ff6657c6 | 2905 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
2906 | if (cfqq) |
2907 | cfq_mark_cfqq_prio_changed(cfqq); | |
2908 | ||
c1b707d2 | 2909 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
2910 | } |
2911 | ||
fc46379d | 2912 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 2913 | { |
4ac845a2 | 2914 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 2915 | ioc->ioprio_changed = 0; |
22e2c507 JA |
2916 | } |
2917 | ||
d5036d77 | 2918 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 2919 | pid_t pid, bool is_sync) |
d5036d77 JA |
2920 | { |
2921 | RB_CLEAR_NODE(&cfqq->rb_node); | |
2922 | RB_CLEAR_NODE(&cfqq->p_node); | |
2923 | INIT_LIST_HEAD(&cfqq->fifo); | |
2924 | ||
30d7b944 | 2925 | cfqq->ref = 0; |
d5036d77 JA |
2926 | cfqq->cfqd = cfqd; |
2927 | ||
2928 | cfq_mark_cfqq_prio_changed(cfqq); | |
2929 | ||
2930 | if (is_sync) { | |
2931 | if (!cfq_class_idle(cfqq)) | |
2932 | cfq_mark_cfqq_idle_window(cfqq); | |
2933 | cfq_mark_cfqq_sync(cfqq); | |
2934 | } | |
2935 | cfqq->pid = pid; | |
2936 | } | |
2937 | ||
24610333 VG |
2938 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
2939 | static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic) | |
2940 | { | |
2941 | struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1); | |
bca4b914 | 2942 | struct cfq_data *cfqd = cic_to_cfqd(cic); |
24610333 VG |
2943 | unsigned long flags; |
2944 | struct request_queue *q; | |
2945 | ||
2946 | if (unlikely(!cfqd)) | |
2947 | return; | |
2948 | ||
2949 | q = cfqd->queue; | |
2950 | ||
2951 | spin_lock_irqsave(q->queue_lock, flags); | |
2952 | ||
2953 | if (sync_cfqq) { | |
2954 | /* | |
2955 | * Drop reference to sync queue. A new sync queue will be | |
2956 | * assigned in new group upon arrival of a fresh request. | |
2957 | */ | |
2958 | cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup"); | |
2959 | cic_set_cfqq(cic, NULL, 1); | |
2960 | cfq_put_queue(sync_cfqq); | |
2961 | } | |
2962 | ||
2963 | spin_unlock_irqrestore(q->queue_lock, flags); | |
2964 | } | |
2965 | ||
2966 | static void cfq_ioc_set_cgroup(struct io_context *ioc) | |
2967 | { | |
2968 | call_for_each_cic(ioc, changed_cgroup); | |
2969 | ioc->cgroup_changed = 0; | |
2970 | } | |
2971 | #endif /* CONFIG_CFQ_GROUP_IOSCHED */ | |
2972 | ||
22e2c507 | 2973 | static struct cfq_queue * |
a6151c3a | 2974 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, |
fd0928df | 2975 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 2976 | { |
22e2c507 | 2977 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 2978 | struct cfq_io_context *cic; |
cdb16e8f | 2979 | struct cfq_group *cfqg; |
22e2c507 JA |
2980 | |
2981 | retry: | |
3e59cf9d | 2982 | cfqg = cfq_get_cfqg(cfqd); |
4ac845a2 | 2983 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
2984 | /* cic always exists here */ |
2985 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 2986 | |
6118b70b JA |
2987 | /* |
2988 | * Always try a new alloc if we fell back to the OOM cfqq | |
2989 | * originally, since it should just be a temporary situation. | |
2990 | */ | |
2991 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
2992 | cfqq = NULL; | |
22e2c507 JA |
2993 | if (new_cfqq) { |
2994 | cfqq = new_cfqq; | |
2995 | new_cfqq = NULL; | |
2996 | } else if (gfp_mask & __GFP_WAIT) { | |
2997 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 2998 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 2999 | gfp_mask | __GFP_ZERO, |
94f6030c | 3000 | cfqd->queue->node); |
22e2c507 | 3001 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
3002 | if (new_cfqq) |
3003 | goto retry; | |
22e2c507 | 3004 | } else { |
94f6030c CL |
3005 | cfqq = kmem_cache_alloc_node(cfq_pool, |
3006 | gfp_mask | __GFP_ZERO, | |
3007 | cfqd->queue->node); | |
22e2c507 JA |
3008 | } |
3009 | ||
6118b70b JA |
3010 | if (cfqq) { |
3011 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
3012 | cfq_init_prio_data(cfqq, ioc); | |
cdb16e8f | 3013 | cfq_link_cfqq_cfqg(cfqq, cfqg); |
6118b70b JA |
3014 | cfq_log_cfqq(cfqd, cfqq, "alloced"); |
3015 | } else | |
3016 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
3017 | } |
3018 | ||
3019 | if (new_cfqq) | |
3020 | kmem_cache_free(cfq_pool, new_cfqq); | |
3021 | ||
22e2c507 JA |
3022 | return cfqq; |
3023 | } | |
3024 | ||
c2dea2d1 VT |
3025 | static struct cfq_queue ** |
3026 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
3027 | { | |
fe094d98 | 3028 | switch (ioprio_class) { |
c2dea2d1 VT |
3029 | case IOPRIO_CLASS_RT: |
3030 | return &cfqd->async_cfqq[0][ioprio]; | |
3031 | case IOPRIO_CLASS_BE: | |
3032 | return &cfqd->async_cfqq[1][ioprio]; | |
3033 | case IOPRIO_CLASS_IDLE: | |
3034 | return &cfqd->async_idle_cfqq; | |
3035 | default: | |
3036 | BUG(); | |
3037 | } | |
3038 | } | |
3039 | ||
15c31be4 | 3040 | static struct cfq_queue * |
a6151c3a | 3041 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, |
15c31be4 JA |
3042 | gfp_t gfp_mask) |
3043 | { | |
fd0928df JA |
3044 | const int ioprio = task_ioprio(ioc); |
3045 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 3046 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
3047 | struct cfq_queue *cfqq = NULL; |
3048 | ||
c2dea2d1 VT |
3049 | if (!is_sync) { |
3050 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
3051 | cfqq = *async_cfqq; | |
3052 | } | |
3053 | ||
6118b70b | 3054 | if (!cfqq) |
fd0928df | 3055 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
3056 | |
3057 | /* | |
3058 | * pin the queue now that it's allocated, scheduler exit will prune it | |
3059 | */ | |
c2dea2d1 | 3060 | if (!is_sync && !(*async_cfqq)) { |
30d7b944 | 3061 | cfqq->ref++; |
c2dea2d1 | 3062 | *async_cfqq = cfqq; |
15c31be4 JA |
3063 | } |
3064 | ||
30d7b944 | 3065 | cfqq->ref++; |
15c31be4 JA |
3066 | return cfqq; |
3067 | } | |
3068 | ||
498d3aa2 JA |
3069 | /* |
3070 | * We drop cfq io contexts lazily, so we may find a dead one. | |
3071 | */ | |
dbecf3ab | 3072 | static void |
4ac845a2 JA |
3073 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
3074 | struct cfq_io_context *cic) | |
dbecf3ab | 3075 | { |
4ac845a2 JA |
3076 | unsigned long flags; |
3077 | ||
fc46379d | 3078 | WARN_ON(!list_empty(&cic->queue_list)); |
bca4b914 | 3079 | BUG_ON(cic->key != cfqd_dead_key(cfqd)); |
597bc485 | 3080 | |
4ac845a2 JA |
3081 | spin_lock_irqsave(&ioc->lock, flags); |
3082 | ||
4faa3c81 | 3083 | BUG_ON(ioc->ioc_data == cic); |
597bc485 | 3084 | |
80b15c73 | 3085 | radix_tree_delete(&ioc->radix_root, cfqd->cic_index); |
ffc4e759 | 3086 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
3087 | spin_unlock_irqrestore(&ioc->lock, flags); |
3088 | ||
3089 | cfq_cic_free(cic); | |
dbecf3ab OH |
3090 | } |
3091 | ||
e2d74ac0 | 3092 | static struct cfq_io_context * |
4ac845a2 | 3093 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 3094 | { |
e2d74ac0 | 3095 | struct cfq_io_context *cic; |
d6de8be7 | 3096 | unsigned long flags; |
e2d74ac0 | 3097 | |
91fac317 VT |
3098 | if (unlikely(!ioc)) |
3099 | return NULL; | |
3100 | ||
d6de8be7 JA |
3101 | rcu_read_lock(); |
3102 | ||
597bc485 JA |
3103 | /* |
3104 | * we maintain a last-hit cache, to avoid browsing over the tree | |
3105 | */ | |
4ac845a2 | 3106 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
3107 | if (cic && cic->key == cfqd) { |
3108 | rcu_read_unlock(); | |
597bc485 | 3109 | return cic; |
d6de8be7 | 3110 | } |
597bc485 | 3111 | |
4ac845a2 | 3112 | do { |
80b15c73 | 3113 | cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index); |
4ac845a2 JA |
3114 | rcu_read_unlock(); |
3115 | if (!cic) | |
3116 | break; | |
bca4b914 | 3117 | if (unlikely(cic->key != cfqd)) { |
4ac845a2 | 3118 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 3119 | rcu_read_lock(); |
4ac845a2 | 3120 | continue; |
dbecf3ab | 3121 | } |
e2d74ac0 | 3122 | |
d6de8be7 | 3123 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 3124 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 3125 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
3126 | break; |
3127 | } while (1); | |
e2d74ac0 | 3128 | |
4ac845a2 | 3129 | return cic; |
e2d74ac0 JA |
3130 | } |
3131 | ||
4ac845a2 JA |
3132 | /* |
3133 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
3134 | * the process specific cfq io context when entered from the block layer. | |
3135 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
3136 | */ | |
febffd61 JA |
3137 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
3138 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 3139 | { |
0261d688 | 3140 | unsigned long flags; |
4ac845a2 | 3141 | int ret; |
e2d74ac0 | 3142 | |
4ac845a2 JA |
3143 | ret = radix_tree_preload(gfp_mask); |
3144 | if (!ret) { | |
3145 | cic->ioc = ioc; | |
3146 | cic->key = cfqd; | |
e2d74ac0 | 3147 | |
4ac845a2 JA |
3148 | spin_lock_irqsave(&ioc->lock, flags); |
3149 | ret = radix_tree_insert(&ioc->radix_root, | |
80b15c73 | 3150 | cfqd->cic_index, cic); |
ffc4e759 JA |
3151 | if (!ret) |
3152 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 3153 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 3154 | |
4ac845a2 JA |
3155 | radix_tree_preload_end(); |
3156 | ||
3157 | if (!ret) { | |
3158 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
3159 | list_add(&cic->queue_list, &cfqd->cic_list); | |
3160 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
3161 | } | |
e2d74ac0 JA |
3162 | } |
3163 | ||
4ac845a2 JA |
3164 | if (ret) |
3165 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 3166 | |
4ac845a2 | 3167 | return ret; |
e2d74ac0 JA |
3168 | } |
3169 | ||
1da177e4 LT |
3170 | /* |
3171 | * Setup general io context and cfq io context. There can be several cfq | |
3172 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 3173 | * than one device managed by cfq. |
1da177e4 LT |
3174 | */ |
3175 | static struct cfq_io_context * | |
e2d74ac0 | 3176 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 3177 | { |
22e2c507 | 3178 | struct io_context *ioc = NULL; |
1da177e4 | 3179 | struct cfq_io_context *cic; |
1da177e4 | 3180 | |
22e2c507 | 3181 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 3182 | |
b5deef90 | 3183 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
3184 | if (!ioc) |
3185 | return NULL; | |
3186 | ||
4ac845a2 | 3187 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
3188 | if (cic) |
3189 | goto out; | |
1da177e4 | 3190 | |
e2d74ac0 JA |
3191 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
3192 | if (cic == NULL) | |
3193 | goto err; | |
1da177e4 | 3194 | |
4ac845a2 JA |
3195 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
3196 | goto err_free; | |
3197 | ||
1da177e4 | 3198 | out: |
fc46379d JA |
3199 | smp_read_barrier_depends(); |
3200 | if (unlikely(ioc->ioprio_changed)) | |
3201 | cfq_ioc_set_ioprio(ioc); | |
3202 | ||
24610333 VG |
3203 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
3204 | if (unlikely(ioc->cgroup_changed)) | |
3205 | cfq_ioc_set_cgroup(ioc); | |
3206 | #endif | |
1da177e4 | 3207 | return cic; |
4ac845a2 JA |
3208 | err_free: |
3209 | cfq_cic_free(cic); | |
1da177e4 LT |
3210 | err: |
3211 | put_io_context(ioc); | |
3212 | return NULL; | |
3213 | } | |
3214 | ||
22e2c507 JA |
3215 | static void |
3216 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 3217 | { |
aaf1228d JA |
3218 | unsigned long elapsed = jiffies - cic->last_end_request; |
3219 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 3220 | |
22e2c507 JA |
3221 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
3222 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
3223 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
3224 | } | |
1da177e4 | 3225 | |
206dc69b | 3226 | static void |
b2c18e1e | 3227 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
6d048f53 | 3228 | struct request *rq) |
206dc69b | 3229 | { |
3dde36dd | 3230 | sector_t sdist = 0; |
41647e7a | 3231 | sector_t n_sec = blk_rq_sectors(rq); |
3dde36dd CZ |
3232 | if (cfqq->last_request_pos) { |
3233 | if (cfqq->last_request_pos < blk_rq_pos(rq)) | |
3234 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
3235 | else | |
3236 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); | |
3237 | } | |
206dc69b | 3238 | |
3dde36dd | 3239 | cfqq->seek_history <<= 1; |
41647e7a CZ |
3240 | if (blk_queue_nonrot(cfqd->queue)) |
3241 | cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT); | |
3242 | else | |
3243 | cfqq->seek_history |= (sdist > CFQQ_SEEK_THR); | |
206dc69b | 3244 | } |
1da177e4 | 3245 | |
22e2c507 JA |
3246 | /* |
3247 | * Disable idle window if the process thinks too long or seeks so much that | |
3248 | * it doesn't matter | |
3249 | */ | |
3250 | static void | |
3251 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3252 | struct cfq_io_context *cic) | |
3253 | { | |
7b679138 | 3254 | int old_idle, enable_idle; |
1be92f2f | 3255 | |
0871714e JA |
3256 | /* |
3257 | * Don't idle for async or idle io prio class | |
3258 | */ | |
3259 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
3260 | return; |
3261 | ||
c265a7f4 | 3262 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 3263 | |
76280aff CZ |
3264 | if (cfqq->queued[0] + cfqq->queued[1] >= 4) |
3265 | cfq_mark_cfqq_deep(cfqq); | |
3266 | ||
749ef9f8 CZ |
3267 | if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE)) |
3268 | enable_idle = 0; | |
3269 | else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || | |
3dde36dd | 3270 | (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq))) |
22e2c507 JA |
3271 | enable_idle = 0; |
3272 | else if (sample_valid(cic->ttime_samples)) { | |
718eee05 | 3273 | if (cic->ttime_mean > cfqd->cfq_slice_idle) |
22e2c507 JA |
3274 | enable_idle = 0; |
3275 | else | |
3276 | enable_idle = 1; | |
1da177e4 LT |
3277 | } |
3278 | ||
7b679138 JA |
3279 | if (old_idle != enable_idle) { |
3280 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
3281 | if (enable_idle) | |
3282 | cfq_mark_cfqq_idle_window(cfqq); | |
3283 | else | |
3284 | cfq_clear_cfqq_idle_window(cfqq); | |
3285 | } | |
22e2c507 | 3286 | } |
1da177e4 | 3287 | |
22e2c507 JA |
3288 | /* |
3289 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
3290 | * no or if we aren't sure, a 1 will cause a preempt. | |
3291 | */ | |
a6151c3a | 3292 | static bool |
22e2c507 | 3293 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, |
5e705374 | 3294 | struct request *rq) |
22e2c507 | 3295 | { |
6d048f53 | 3296 | struct cfq_queue *cfqq; |
22e2c507 | 3297 | |
6d048f53 JA |
3298 | cfqq = cfqd->active_queue; |
3299 | if (!cfqq) | |
a6151c3a | 3300 | return false; |
22e2c507 | 3301 | |
6d048f53 | 3302 | if (cfq_class_idle(new_cfqq)) |
a6151c3a | 3303 | return false; |
22e2c507 JA |
3304 | |
3305 | if (cfq_class_idle(cfqq)) | |
a6151c3a | 3306 | return true; |
1e3335de | 3307 | |
875feb63 DS |
3308 | /* |
3309 | * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice. | |
3310 | */ | |
3311 | if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq)) | |
3312 | return false; | |
3313 | ||
374f84ac JA |
3314 | /* |
3315 | * if the new request is sync, but the currently running queue is | |
3316 | * not, let the sync request have priority. | |
3317 | */ | |
5e705374 | 3318 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
a6151c3a | 3319 | return true; |
1e3335de | 3320 | |
8682e1f1 VG |
3321 | if (new_cfqq->cfqg != cfqq->cfqg) |
3322 | return false; | |
3323 | ||
3324 | if (cfq_slice_used(cfqq)) | |
3325 | return true; | |
3326 | ||
3327 | /* Allow preemption only if we are idling on sync-noidle tree */ | |
3328 | if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && | |
3329 | cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && | |
3330 | new_cfqq->service_tree->count == 2 && | |
3331 | RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3332 | return true; | |
3333 | ||
374f84ac JA |
3334 | /* |
3335 | * So both queues are sync. Let the new request get disk time if | |
3336 | * it's a metadata request and the current queue is doing regular IO. | |
3337 | */ | |
7b6d91da | 3338 | if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending) |
e6ec4fe2 | 3339 | return true; |
22e2c507 | 3340 | |
3a9a3f6c DS |
3341 | /* |
3342 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
3343 | */ | |
3344 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
a6151c3a | 3345 | return true; |
3a9a3f6c | 3346 | |
d2d59e18 SL |
3347 | /* An idle queue should not be idle now for some reason */ |
3348 | if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq)) | |
3349 | return true; | |
3350 | ||
1e3335de | 3351 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
a6151c3a | 3352 | return false; |
1e3335de JA |
3353 | |
3354 | /* | |
3355 | * if this request is as-good as one we would expect from the | |
3356 | * current cfqq, let it preempt | |
3357 | */ | |
e9ce335d | 3358 | if (cfq_rq_close(cfqd, cfqq, rq)) |
a6151c3a | 3359 | return true; |
1e3335de | 3360 | |
a6151c3a | 3361 | return false; |
22e2c507 JA |
3362 | } |
3363 | ||
3364 | /* | |
3365 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
3366 | * let it have half of its nominal slice. | |
3367 | */ | |
3368 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3369 | { | |
f8ae6e3e SL |
3370 | struct cfq_queue *old_cfqq = cfqd->active_queue; |
3371 | ||
7b679138 | 3372 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
e5ff082e | 3373 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 3374 | |
f8ae6e3e SL |
3375 | /* |
3376 | * workload type is changed, don't save slice, otherwise preempt | |
3377 | * doesn't happen | |
3378 | */ | |
3379 | if (cfqq_type(old_cfqq) != cfqq_type(cfqq)) | |
3380 | cfqq->cfqg->saved_workload_slice = 0; | |
3381 | ||
bf572256 JA |
3382 | /* |
3383 | * Put the new queue at the front of the of the current list, | |
3384 | * so we know that it will be selected next. | |
3385 | */ | |
3386 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
3387 | |
3388 | cfq_service_tree_add(cfqd, cfqq, 1); | |
eda5e0c9 | 3389 | |
62a37f6b JT |
3390 | cfqq->slice_end = 0; |
3391 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
3392 | } |
3393 | ||
22e2c507 | 3394 | /* |
5e705374 | 3395 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
3396 | * something we should do about it |
3397 | */ | |
3398 | static void | |
5e705374 JA |
3399 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
3400 | struct request *rq) | |
22e2c507 | 3401 | { |
5e705374 | 3402 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 3403 | |
45333d5a | 3404 | cfqd->rq_queued++; |
7b6d91da | 3405 | if (rq->cmd_flags & REQ_META) |
374f84ac JA |
3406 | cfqq->meta_pending++; |
3407 | ||
9c2c38a1 | 3408 | cfq_update_io_thinktime(cfqd, cic); |
b2c18e1e | 3409 | cfq_update_io_seektime(cfqd, cfqq, rq); |
9c2c38a1 JA |
3410 | cfq_update_idle_window(cfqd, cfqq, cic); |
3411 | ||
b2c18e1e | 3412 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
3413 | |
3414 | if (cfqq == cfqd->active_queue) { | |
3415 | /* | |
b029195d JA |
3416 | * Remember that we saw a request from this process, but |
3417 | * don't start queuing just yet. Otherwise we risk seeing lots | |
3418 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
3419 | * and merging. If the request is already larger than a single |
3420 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
3421 | * merging is already done. Ditto for a busy system that |
3422 | * has other work pending, don't risk delaying until the | |
3423 | * idle timer unplug to continue working. | |
22e2c507 | 3424 | */ |
d6ceb25e | 3425 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
3426 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
3427 | cfqd->busy_queues > 1) { | |
812df48d | 3428 | cfq_del_timer(cfqd, cfqq); |
554554f6 | 3429 | cfq_clear_cfqq_wait_request(cfqq); |
24ecfbe2 | 3430 | __blk_run_queue(cfqd->queue); |
a11cdaa7 | 3431 | } else { |
e98ef89b | 3432 | cfq_blkiocg_update_idle_time_stats( |
a11cdaa7 | 3433 | &cfqq->cfqg->blkg); |
bf791937 | 3434 | cfq_mark_cfqq_must_dispatch(cfqq); |
a11cdaa7 | 3435 | } |
d6ceb25e | 3436 | } |
5e705374 | 3437 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
3438 | /* |
3439 | * not the active queue - expire current slice if it is | |
3440 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
3441 | * has some old slice time left and is of higher priority or |
3442 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
3443 | */ |
3444 | cfq_preempt_queue(cfqd, cfqq); | |
24ecfbe2 | 3445 | __blk_run_queue(cfqd->queue); |
22e2c507 | 3446 | } |
1da177e4 LT |
3447 | } |
3448 | ||
165125e1 | 3449 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 3450 | { |
b4878f24 | 3451 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 3452 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 3453 | |
7b679138 | 3454 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 3455 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 3456 | |
30996f40 | 3457 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); |
22e2c507 | 3458 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
aa6f6a3d | 3459 | cfq_add_rq_rb(rq); |
e98ef89b | 3460 | cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg, |
cdc1184c DS |
3461 | &cfqd->serving_group->blkg, rq_data_dir(rq), |
3462 | rq_is_sync(rq)); | |
5e705374 | 3463 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
3464 | } |
3465 | ||
45333d5a AC |
3466 | /* |
3467 | * Update hw_tag based on peak queue depth over 50 samples under | |
3468 | * sufficient load. | |
3469 | */ | |
3470 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
3471 | { | |
1a1238a7 SL |
3472 | struct cfq_queue *cfqq = cfqd->active_queue; |
3473 | ||
53c583d2 CZ |
3474 | if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth) |
3475 | cfqd->hw_tag_est_depth = cfqd->rq_in_driver; | |
e459dd08 CZ |
3476 | |
3477 | if (cfqd->hw_tag == 1) | |
3478 | return; | |
45333d5a AC |
3479 | |
3480 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
53c583d2 | 3481 | cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
3482 | return; |
3483 | ||
1a1238a7 SL |
3484 | /* |
3485 | * If active queue hasn't enough requests and can idle, cfq might not | |
3486 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
3487 | * case | |
3488 | */ | |
3489 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
3490 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
53c583d2 | 3491 | CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN) |
1a1238a7 SL |
3492 | return; |
3493 | ||
45333d5a AC |
3494 | if (cfqd->hw_tag_samples++ < 50) |
3495 | return; | |
3496 | ||
e459dd08 | 3497 | if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
3498 | cfqd->hw_tag = 1; |
3499 | else | |
3500 | cfqd->hw_tag = 0; | |
45333d5a AC |
3501 | } |
3502 | ||
7667aa06 VG |
3503 | static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
3504 | { | |
3505 | struct cfq_io_context *cic = cfqd->active_cic; | |
3506 | ||
02a8f01b JT |
3507 | /* If the queue already has requests, don't wait */ |
3508 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3509 | return false; | |
3510 | ||
7667aa06 VG |
3511 | /* If there are other queues in the group, don't wait */ |
3512 | if (cfqq->cfqg->nr_cfqq > 1) | |
3513 | return false; | |
3514 | ||
3515 | if (cfq_slice_used(cfqq)) | |
3516 | return true; | |
3517 | ||
3518 | /* if slice left is less than think time, wait busy */ | |
3519 | if (cic && sample_valid(cic->ttime_samples) | |
3520 | && (cfqq->slice_end - jiffies < cic->ttime_mean)) | |
3521 | return true; | |
3522 | ||
3523 | /* | |
3524 | * If think times is less than a jiffy than ttime_mean=0 and above | |
3525 | * will not be true. It might happen that slice has not expired yet | |
3526 | * but will expire soon (4-5 ns) during select_queue(). To cover the | |
3527 | * case where think time is less than a jiffy, mark the queue wait | |
3528 | * busy if only 1 jiffy is left in the slice. | |
3529 | */ | |
3530 | if (cfqq->slice_end - jiffies == 1) | |
3531 | return true; | |
3532 | ||
3533 | return false; | |
3534 | } | |
3535 | ||
165125e1 | 3536 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 3537 | { |
5e705374 | 3538 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 3539 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 3540 | const int sync = rq_is_sync(rq); |
b4878f24 | 3541 | unsigned long now; |
1da177e4 | 3542 | |
b4878f24 | 3543 | now = jiffies; |
33659ebb CH |
3544 | cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", |
3545 | !!(rq->cmd_flags & REQ_NOIDLE)); | |
1da177e4 | 3546 | |
45333d5a AC |
3547 | cfq_update_hw_tag(cfqd); |
3548 | ||
53c583d2 | 3549 | WARN_ON(!cfqd->rq_in_driver); |
6d048f53 | 3550 | WARN_ON(!cfqq->dispatched); |
53c583d2 | 3551 | cfqd->rq_in_driver--; |
6d048f53 | 3552 | cfqq->dispatched--; |
80bdf0c7 | 3553 | (RQ_CFQG(rq))->dispatched--; |
e98ef89b VG |
3554 | cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg, |
3555 | rq_start_time_ns(rq), rq_io_start_time_ns(rq), | |
3556 | rq_data_dir(rq), rq_is_sync(rq)); | |
1da177e4 | 3557 | |
53c583d2 | 3558 | cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--; |
3ed9a296 | 3559 | |
365722bb | 3560 | if (sync) { |
5e705374 | 3561 | RQ_CIC(rq)->last_end_request = now; |
573412b2 CZ |
3562 | if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now)) |
3563 | cfqd->last_delayed_sync = now; | |
365722bb | 3564 | } |
caaa5f9f JA |
3565 | |
3566 | /* | |
3567 | * If this is the active queue, check if it needs to be expired, | |
3568 | * or if we want to idle in case it has no pending requests. | |
3569 | */ | |
3570 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
3571 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
3572 | ||
44f7c160 JA |
3573 | if (cfq_cfqq_slice_new(cfqq)) { |
3574 | cfq_set_prio_slice(cfqd, cfqq); | |
3575 | cfq_clear_cfqq_slice_new(cfqq); | |
3576 | } | |
f75edf2d VG |
3577 | |
3578 | /* | |
7667aa06 VG |
3579 | * Should we wait for next request to come in before we expire |
3580 | * the queue. | |
f75edf2d | 3581 | */ |
7667aa06 | 3582 | if (cfq_should_wait_busy(cfqd, cfqq)) { |
80bdf0c7 VG |
3583 | unsigned long extend_sl = cfqd->cfq_slice_idle; |
3584 | if (!cfqd->cfq_slice_idle) | |
3585 | extend_sl = cfqd->cfq_group_idle; | |
3586 | cfqq->slice_end = jiffies + extend_sl; | |
f75edf2d | 3587 | cfq_mark_cfqq_wait_busy(cfqq); |
b1ffe737 | 3588 | cfq_log_cfqq(cfqd, cfqq, "will busy wait"); |
f75edf2d VG |
3589 | } |
3590 | ||
a36e71f9 | 3591 | /* |
8e550632 CZ |
3592 | * Idling is not enabled on: |
3593 | * - expired queues | |
3594 | * - idle-priority queues | |
3595 | * - async queues | |
3596 | * - queues with still some requests queued | |
3597 | * - when there is a close cooperator | |
a36e71f9 | 3598 | */ |
0871714e | 3599 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
e5ff082e | 3600 | cfq_slice_expired(cfqd, 1); |
8e550632 CZ |
3601 | else if (sync && cfqq_empty && |
3602 | !cfq_close_cooperator(cfqd, cfqq)) { | |
749ef9f8 | 3603 | cfq_arm_slice_timer(cfqd); |
8e550632 | 3604 | } |
caaa5f9f | 3605 | } |
6d048f53 | 3606 | |
53c583d2 | 3607 | if (!cfqd->rq_in_driver) |
23e018a1 | 3608 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
3609 | } |
3610 | ||
22e2c507 JA |
3611 | /* |
3612 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
3613 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
3614 | */ | |
3615 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 3616 | { |
22e2c507 JA |
3617 | if (has_fs_excl()) { |
3618 | /* | |
3619 | * boost idle prio on transactions that would lock out other | |
3620 | * users of the filesystem | |
3621 | */ | |
3622 | if (cfq_class_idle(cfqq)) | |
3623 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
3624 | if (cfqq->ioprio > IOPRIO_NORM) | |
3625 | cfqq->ioprio = IOPRIO_NORM; | |
3626 | } else { | |
3627 | /* | |
dddb7451 | 3628 | * unboost the queue (if needed) |
22e2c507 | 3629 | */ |
dddb7451 CZ |
3630 | cfqq->ioprio_class = cfqq->org_ioprio_class; |
3631 | cfqq->ioprio = cfqq->org_ioprio; | |
22e2c507 | 3632 | } |
22e2c507 | 3633 | } |
1da177e4 | 3634 | |
89850f7e | 3635 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 3636 | { |
1b379d8d | 3637 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 3638 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 3639 | return ELV_MQUEUE_MUST; |
3b18152c | 3640 | } |
1da177e4 | 3641 | |
22e2c507 | 3642 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
3643 | } |
3644 | ||
165125e1 | 3645 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
3646 | { |
3647 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3648 | struct task_struct *tsk = current; | |
91fac317 | 3649 | struct cfq_io_context *cic; |
22e2c507 JA |
3650 | struct cfq_queue *cfqq; |
3651 | ||
3652 | /* | |
3653 | * don't force setup of a queue from here, as a call to may_queue | |
3654 | * does not necessarily imply that a request actually will be queued. | |
3655 | * so just lookup a possibly existing queue, or return 'may queue' | |
3656 | * if that fails | |
3657 | */ | |
4ac845a2 | 3658 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
3659 | if (!cic) |
3660 | return ELV_MQUEUE_MAY; | |
3661 | ||
b0b78f81 | 3662 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 3663 | if (cfqq) { |
fd0928df | 3664 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
3665 | cfq_prio_boost(cfqq); |
3666 | ||
89850f7e | 3667 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
3668 | } |
3669 | ||
3670 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
3671 | } |
3672 | ||
1da177e4 LT |
3673 | /* |
3674 | * queue lock held here | |
3675 | */ | |
bb37b94c | 3676 | static void cfq_put_request(struct request *rq) |
1da177e4 | 3677 | { |
5e705374 | 3678 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 3679 | |
5e705374 | 3680 | if (cfqq) { |
22e2c507 | 3681 | const int rw = rq_data_dir(rq); |
1da177e4 | 3682 | |
22e2c507 JA |
3683 | BUG_ON(!cfqq->allocated[rw]); |
3684 | cfqq->allocated[rw]--; | |
1da177e4 | 3685 | |
5e705374 | 3686 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 3687 | |
c186794d MS |
3688 | rq->elevator_private[0] = NULL; |
3689 | rq->elevator_private[1] = NULL; | |
1da177e4 | 3690 | |
7f1dc8a2 VG |
3691 | /* Put down rq reference on cfqg */ |
3692 | cfq_put_cfqg(RQ_CFQG(rq)); | |
c186794d | 3693 | rq->elevator_private[2] = NULL; |
7f1dc8a2 | 3694 | |
1da177e4 LT |
3695 | cfq_put_queue(cfqq); |
3696 | } | |
3697 | } | |
3698 | ||
df5fe3e8 JM |
3699 | static struct cfq_queue * |
3700 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, | |
3701 | struct cfq_queue *cfqq) | |
3702 | { | |
3703 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
3704 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
b3b6d040 | 3705 | cfq_mark_cfqq_coop(cfqq->new_cfqq); |
df5fe3e8 JM |
3706 | cfq_put_queue(cfqq); |
3707 | return cic_to_cfqq(cic, 1); | |
3708 | } | |
3709 | ||
e6c5bc73 JM |
3710 | /* |
3711 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
3712 | * was the last process referring to said cfqq. | |
3713 | */ | |
3714 | static struct cfq_queue * | |
3715 | split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) | |
3716 | { | |
3717 | if (cfqq_process_refs(cfqq) == 1) { | |
e6c5bc73 JM |
3718 | cfqq->pid = current->pid; |
3719 | cfq_clear_cfqq_coop(cfqq); | |
ae54abed | 3720 | cfq_clear_cfqq_split_coop(cfqq); |
e6c5bc73 JM |
3721 | return cfqq; |
3722 | } | |
3723 | ||
3724 | cic_set_cfqq(cic, NULL, 1); | |
d02a2c07 SL |
3725 | |
3726 | cfq_put_cooperator(cfqq); | |
3727 | ||
e6c5bc73 JM |
3728 | cfq_put_queue(cfqq); |
3729 | return NULL; | |
3730 | } | |
1da177e4 | 3731 | /* |
22e2c507 | 3732 | * Allocate cfq data structures associated with this request. |
1da177e4 | 3733 | */ |
22e2c507 | 3734 | static int |
165125e1 | 3735 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
3736 | { |
3737 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3738 | struct cfq_io_context *cic; | |
3739 | const int rw = rq_data_dir(rq); | |
a6151c3a | 3740 | const bool is_sync = rq_is_sync(rq); |
22e2c507 | 3741 | struct cfq_queue *cfqq; |
1da177e4 LT |
3742 | unsigned long flags; |
3743 | ||
3744 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
3745 | ||
e2d74ac0 | 3746 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 3747 | |
1da177e4 LT |
3748 | spin_lock_irqsave(q->queue_lock, flags); |
3749 | ||
22e2c507 JA |
3750 | if (!cic) |
3751 | goto queue_fail; | |
3752 | ||
e6c5bc73 | 3753 | new_queue: |
91fac317 | 3754 | cfqq = cic_to_cfqq(cic, is_sync); |
32f2e807 | 3755 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { |
fd0928df | 3756 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 | 3757 | cic_set_cfqq(cic, cfqq, is_sync); |
df5fe3e8 | 3758 | } else { |
e6c5bc73 JM |
3759 | /* |
3760 | * If the queue was seeky for too long, break it apart. | |
3761 | */ | |
ae54abed | 3762 | if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) { |
e6c5bc73 JM |
3763 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); |
3764 | cfqq = split_cfqq(cic, cfqq); | |
3765 | if (!cfqq) | |
3766 | goto new_queue; | |
3767 | } | |
3768 | ||
df5fe3e8 JM |
3769 | /* |
3770 | * Check to see if this queue is scheduled to merge with | |
3771 | * another, closely cooperating queue. The merging of | |
3772 | * queues happens here as it must be done in process context. | |
3773 | * The reference on new_cfqq was taken in merge_cfqqs. | |
3774 | */ | |
3775 | if (cfqq->new_cfqq) | |
3776 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
91fac317 | 3777 | } |
1da177e4 LT |
3778 | |
3779 | cfqq->allocated[rw]++; | |
1da177e4 | 3780 | |
6fae9c25 | 3781 | cfqq->ref++; |
c186794d MS |
3782 | rq->elevator_private[0] = cic; |
3783 | rq->elevator_private[1] = cfqq; | |
3784 | rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg); | |
93803e01 | 3785 | spin_unlock_irqrestore(q->queue_lock, flags); |
5e705374 | 3786 | return 0; |
1da177e4 | 3787 | |
22e2c507 JA |
3788 | queue_fail: |
3789 | if (cic) | |
3790 | put_io_context(cic->ioc); | |
89850f7e | 3791 | |
23e018a1 | 3792 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 3793 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 3794 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
3795 | return 1; |
3796 | } | |
3797 | ||
65f27f38 | 3798 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 3799 | { |
65f27f38 | 3800 | struct cfq_data *cfqd = |
23e018a1 | 3801 | container_of(work, struct cfq_data, unplug_work); |
165125e1 | 3802 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3803 | |
40bb54d1 | 3804 | spin_lock_irq(q->queue_lock); |
24ecfbe2 | 3805 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 3806 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
3807 | } |
3808 | ||
3809 | /* | |
3810 | * Timer running if the active_queue is currently idling inside its time slice | |
3811 | */ | |
3812 | static void cfq_idle_slice_timer(unsigned long data) | |
3813 | { | |
3814 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
3815 | struct cfq_queue *cfqq; | |
3816 | unsigned long flags; | |
3c6bd2f8 | 3817 | int timed_out = 1; |
22e2c507 | 3818 | |
7b679138 JA |
3819 | cfq_log(cfqd, "idle timer fired"); |
3820 | ||
22e2c507 JA |
3821 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
3822 | ||
fe094d98 JA |
3823 | cfqq = cfqd->active_queue; |
3824 | if (cfqq) { | |
3c6bd2f8 JA |
3825 | timed_out = 0; |
3826 | ||
b029195d JA |
3827 | /* |
3828 | * We saw a request before the queue expired, let it through | |
3829 | */ | |
3830 | if (cfq_cfqq_must_dispatch(cfqq)) | |
3831 | goto out_kick; | |
3832 | ||
22e2c507 JA |
3833 | /* |
3834 | * expired | |
3835 | */ | |
44f7c160 | 3836 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
3837 | goto expire; |
3838 | ||
3839 | /* | |
3840 | * only expire and reinvoke request handler, if there are | |
3841 | * other queues with pending requests | |
3842 | */ | |
caaa5f9f | 3843 | if (!cfqd->busy_queues) |
22e2c507 | 3844 | goto out_cont; |
22e2c507 JA |
3845 | |
3846 | /* | |
3847 | * not expired and it has a request pending, let it dispatch | |
3848 | */ | |
75e50984 | 3849 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 3850 | goto out_kick; |
76280aff CZ |
3851 | |
3852 | /* | |
3853 | * Queue depth flag is reset only when the idle didn't succeed | |
3854 | */ | |
3855 | cfq_clear_cfqq_deep(cfqq); | |
22e2c507 JA |
3856 | } |
3857 | expire: | |
e5ff082e | 3858 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 3859 | out_kick: |
23e018a1 | 3860 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
3861 | out_cont: |
3862 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
3863 | } | |
3864 | ||
3b18152c JA |
3865 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
3866 | { | |
3867 | del_timer_sync(&cfqd->idle_slice_timer); | |
23e018a1 | 3868 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 3869 | } |
22e2c507 | 3870 | |
c2dea2d1 VT |
3871 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
3872 | { | |
3873 | int i; | |
3874 | ||
3875 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
3876 | if (cfqd->async_cfqq[0][i]) | |
3877 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
3878 | if (cfqd->async_cfqq[1][i]) | |
3879 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 3880 | } |
2389d1ef ON |
3881 | |
3882 | if (cfqd->async_idle_cfqq) | |
3883 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
3884 | } |
3885 | ||
b374d18a | 3886 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 3887 | { |
22e2c507 | 3888 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 3889 | struct request_queue *q = cfqd->queue; |
56edf7d7 | 3890 | bool wait = false; |
22e2c507 | 3891 | |
3b18152c | 3892 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 3893 | |
d9ff4187 | 3894 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 3895 | |
d9ff4187 | 3896 | if (cfqd->active_queue) |
e5ff082e | 3897 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
3898 | |
3899 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
3900 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
3901 | struct cfq_io_context, | |
3902 | queue_list); | |
89850f7e JA |
3903 | |
3904 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 3905 | } |
e2d74ac0 | 3906 | |
c2dea2d1 | 3907 | cfq_put_async_queues(cfqd); |
b1c35769 | 3908 | cfq_release_cfq_groups(cfqd); |
56edf7d7 VG |
3909 | |
3910 | /* | |
3911 | * If there are groups which we could not unlink from blkcg list, | |
3912 | * wait for a rcu period for them to be freed. | |
3913 | */ | |
3914 | if (cfqd->nr_blkcg_linked_grps) | |
3915 | wait = true; | |
15c31be4 | 3916 | |
d9ff4187 | 3917 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
3918 | |
3919 | cfq_shutdown_timer_wq(cfqd); | |
3920 | ||
80b15c73 KK |
3921 | spin_lock(&cic_index_lock); |
3922 | ida_remove(&cic_index_ida, cfqd->cic_index); | |
3923 | spin_unlock(&cic_index_lock); | |
3924 | ||
56edf7d7 VG |
3925 | /* |
3926 | * Wait for cfqg->blkg->key accessors to exit their grace periods. | |
3927 | * Do this wait only if there are other unlinked groups out | |
3928 | * there. This can happen if cgroup deletion path claimed the | |
3929 | * responsibility of cleaning up a group before queue cleanup code | |
3930 | * get to the group. | |
3931 | * | |
3932 | * Do not call synchronize_rcu() unconditionally as there are drivers | |
3933 | * which create/delete request queue hundreds of times during scan/boot | |
3934 | * and synchronize_rcu() can take significant time and slow down boot. | |
3935 | */ | |
3936 | if (wait) | |
3937 | synchronize_rcu(); | |
2abae55f VG |
3938 | |
3939 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
3940 | /* Free up per cpu stats for root group */ | |
3941 | free_percpu(cfqd->root_group.blkg.stats_cpu); | |
3942 | #endif | |
56edf7d7 | 3943 | kfree(cfqd); |
1da177e4 LT |
3944 | } |
3945 | ||
80b15c73 KK |
3946 | static int cfq_alloc_cic_index(void) |
3947 | { | |
3948 | int index, error; | |
3949 | ||
3950 | do { | |
3951 | if (!ida_pre_get(&cic_index_ida, GFP_KERNEL)) | |
3952 | return -ENOMEM; | |
3953 | ||
3954 | spin_lock(&cic_index_lock); | |
3955 | error = ida_get_new(&cic_index_ida, &index); | |
3956 | spin_unlock(&cic_index_lock); | |
3957 | if (error && error != -EAGAIN) | |
3958 | return error; | |
3959 | } while (error); | |
3960 | ||
3961 | return index; | |
3962 | } | |
3963 | ||
165125e1 | 3964 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
3965 | { |
3966 | struct cfq_data *cfqd; | |
718eee05 | 3967 | int i, j; |
cdb16e8f | 3968 | struct cfq_group *cfqg; |
615f0259 | 3969 | struct cfq_rb_root *st; |
1da177e4 | 3970 | |
80b15c73 KK |
3971 | i = cfq_alloc_cic_index(); |
3972 | if (i < 0) | |
3973 | return NULL; | |
3974 | ||
94f6030c | 3975 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1547010e NK |
3976 | if (!cfqd) { |
3977 | spin_lock(&cic_index_lock); | |
3978 | ida_remove(&cic_index_ida, i); | |
3979 | spin_unlock(&cic_index_lock); | |
bc1c1169 | 3980 | return NULL; |
1547010e | 3981 | } |
1da177e4 | 3982 | |
30d7b944 SL |
3983 | /* |
3984 | * Don't need take queue_lock in the routine, since we are | |
3985 | * initializing the ioscheduler, and nobody is using cfqd | |
3986 | */ | |
80b15c73 KK |
3987 | cfqd->cic_index = i; |
3988 | ||
1fa8f6d6 VG |
3989 | /* Init root service tree */ |
3990 | cfqd->grp_service_tree = CFQ_RB_ROOT; | |
3991 | ||
cdb16e8f VG |
3992 | /* Init root group */ |
3993 | cfqg = &cfqd->root_group; | |
615f0259 VG |
3994 | for_each_cfqg_st(cfqg, i, j, st) |
3995 | *st = CFQ_RB_ROOT; | |
1fa8f6d6 | 3996 | RB_CLEAR_NODE(&cfqg->rb_node); |
26a2ac00 | 3997 | |
25bc6b07 VG |
3998 | /* Give preference to root group over other groups */ |
3999 | cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT; | |
4000 | ||
25fb5169 | 4001 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
b1c35769 | 4002 | /* |
56edf7d7 VG |
4003 | * Set root group reference to 2. One reference will be dropped when |
4004 | * all groups on cfqd->cfqg_list are being deleted during queue exit. | |
4005 | * Other reference will remain there as we don't want to delete this | |
4006 | * group as it is statically allocated and gets destroyed when | |
4007 | * throtl_data goes away. | |
b1c35769 | 4008 | */ |
56edf7d7 | 4009 | cfqg->ref = 2; |
5624a4e4 VG |
4010 | |
4011 | if (blkio_alloc_blkg_stats(&cfqg->blkg)) { | |
4012 | kfree(cfqg); | |
4013 | kfree(cfqd); | |
4014 | return NULL; | |
4015 | } | |
4016 | ||
dcf097b2 | 4017 | rcu_read_lock(); |
5624a4e4 | 4018 | |
e98ef89b VG |
4019 | cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, |
4020 | (void *)cfqd, 0); | |
dcf097b2 | 4021 | rcu_read_unlock(); |
56edf7d7 VG |
4022 | cfqd->nr_blkcg_linked_grps++; |
4023 | ||
4024 | /* Add group on cfqd->cfqg_list */ | |
4025 | hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); | |
25fb5169 | 4026 | #endif |
26a2ac00 JA |
4027 | /* |
4028 | * Not strictly needed (since RB_ROOT just clears the node and we | |
4029 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
4030 | * to add magic to the rb code | |
4031 | */ | |
4032 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
4033 | cfqd->prio_trees[i] = RB_ROOT; | |
4034 | ||
6118b70b JA |
4035 | /* |
4036 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
4037 | * Grab a permanent reference to it, so that the normal code flow | |
4038 | * will not attempt to free it. | |
4039 | */ | |
4040 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
30d7b944 | 4041 | cfqd->oom_cfqq.ref++; |
cdb16e8f | 4042 | cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group); |
6118b70b | 4043 | |
d9ff4187 | 4044 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 4045 | |
1da177e4 | 4046 | cfqd->queue = q; |
1da177e4 | 4047 | |
22e2c507 JA |
4048 | init_timer(&cfqd->idle_slice_timer); |
4049 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
4050 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
4051 | ||
23e018a1 | 4052 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 4053 | |
1da177e4 | 4054 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
4055 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
4056 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
4057 | cfqd->cfq_back_max = cfq_back_max; |
4058 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
4059 | cfqd->cfq_slice[0] = cfq_slice_async; |
4060 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
4061 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
4062 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
80bdf0c7 | 4063 | cfqd->cfq_group_idle = cfq_group_idle; |
963b72fc | 4064 | cfqd->cfq_latency = 1; |
e459dd08 | 4065 | cfqd->hw_tag = -1; |
edc71131 CZ |
4066 | /* |
4067 | * we optimistically start assuming sync ops weren't delayed in last | |
4068 | * second, in order to have larger depth for async operations. | |
4069 | */ | |
573412b2 | 4070 | cfqd->last_delayed_sync = jiffies - HZ; |
bc1c1169 | 4071 | return cfqd; |
1da177e4 LT |
4072 | } |
4073 | ||
4074 | static void cfq_slab_kill(void) | |
4075 | { | |
d6de8be7 JA |
4076 | /* |
4077 | * Caller already ensured that pending RCU callbacks are completed, | |
4078 | * so we should have no busy allocations at this point. | |
4079 | */ | |
1da177e4 LT |
4080 | if (cfq_pool) |
4081 | kmem_cache_destroy(cfq_pool); | |
4082 | if (cfq_ioc_pool) | |
4083 | kmem_cache_destroy(cfq_ioc_pool); | |
4084 | } | |
4085 | ||
4086 | static int __init cfq_slab_setup(void) | |
4087 | { | |
0a31bd5f | 4088 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
4089 | if (!cfq_pool) |
4090 | goto fail; | |
4091 | ||
34e6bbf2 | 4092 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
4093 | if (!cfq_ioc_pool) |
4094 | goto fail; | |
4095 | ||
4096 | return 0; | |
4097 | fail: | |
4098 | cfq_slab_kill(); | |
4099 | return -ENOMEM; | |
4100 | } | |
4101 | ||
1da177e4 LT |
4102 | /* |
4103 | * sysfs parts below --> | |
4104 | */ | |
1da177e4 LT |
4105 | static ssize_t |
4106 | cfq_var_show(unsigned int var, char *page) | |
4107 | { | |
4108 | return sprintf(page, "%d\n", var); | |
4109 | } | |
4110 | ||
4111 | static ssize_t | |
4112 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
4113 | { | |
4114 | char *p = (char *) page; | |
4115 | ||
4116 | *var = simple_strtoul(p, &p, 10); | |
4117 | return count; | |
4118 | } | |
4119 | ||
1da177e4 | 4120 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 4121 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 4122 | { \ |
3d1ab40f | 4123 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
4124 | unsigned int __data = __VAR; \ |
4125 | if (__CONV) \ | |
4126 | __data = jiffies_to_msecs(__data); \ | |
4127 | return cfq_var_show(__data, (page)); \ | |
4128 | } | |
4129 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
4130 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
4131 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
4132 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
4133 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 | 4134 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
80bdf0c7 | 4135 | SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1); |
22e2c507 JA |
4136 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); |
4137 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
4138 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
963b72fc | 4139 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); |
1da177e4 LT |
4140 | #undef SHOW_FUNCTION |
4141 | ||
4142 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 4143 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 4144 | { \ |
3d1ab40f | 4145 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
4146 | unsigned int __data; \ |
4147 | int ret = cfq_var_store(&__data, (page), count); \ | |
4148 | if (__data < (MIN)) \ | |
4149 | __data = (MIN); \ | |
4150 | else if (__data > (MAX)) \ | |
4151 | __data = (MAX); \ | |
4152 | if (__CONV) \ | |
4153 | *(__PTR) = msecs_to_jiffies(__data); \ | |
4154 | else \ | |
4155 | *(__PTR) = __data; \ | |
4156 | return ret; \ | |
4157 | } | |
4158 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
4159 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
4160 | UINT_MAX, 1); | |
4161 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
4162 | UINT_MAX, 1); | |
e572ec7e | 4163 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
4164 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
4165 | UINT_MAX, 0); | |
22e2c507 | 4166 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
80bdf0c7 | 4167 | STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1); |
22e2c507 JA |
4168 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); |
4169 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
4170 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
4171 | UINT_MAX, 0); | |
963b72fc | 4172 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); |
1da177e4 LT |
4173 | #undef STORE_FUNCTION |
4174 | ||
e572ec7e AV |
4175 | #define CFQ_ATTR(name) \ |
4176 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
4177 | ||
4178 | static struct elv_fs_entry cfq_attrs[] = { | |
4179 | CFQ_ATTR(quantum), | |
e572ec7e AV |
4180 | CFQ_ATTR(fifo_expire_sync), |
4181 | CFQ_ATTR(fifo_expire_async), | |
4182 | CFQ_ATTR(back_seek_max), | |
4183 | CFQ_ATTR(back_seek_penalty), | |
4184 | CFQ_ATTR(slice_sync), | |
4185 | CFQ_ATTR(slice_async), | |
4186 | CFQ_ATTR(slice_async_rq), | |
4187 | CFQ_ATTR(slice_idle), | |
80bdf0c7 | 4188 | CFQ_ATTR(group_idle), |
963b72fc | 4189 | CFQ_ATTR(low_latency), |
e572ec7e | 4190 | __ATTR_NULL |
1da177e4 LT |
4191 | }; |
4192 | ||
1da177e4 LT |
4193 | static struct elevator_type iosched_cfq = { |
4194 | .ops = { | |
4195 | .elevator_merge_fn = cfq_merge, | |
4196 | .elevator_merged_fn = cfq_merged_request, | |
4197 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 4198 | .elevator_allow_merge_fn = cfq_allow_merge, |
812d4026 | 4199 | .elevator_bio_merged_fn = cfq_bio_merged, |
b4878f24 | 4200 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 4201 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 4202 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 | 4203 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
1da177e4 | 4204 | .elevator_completed_req_fn = cfq_completed_request, |
21183b07 JA |
4205 | .elevator_former_req_fn = elv_rb_former_request, |
4206 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
4207 | .elevator_set_req_fn = cfq_set_request, |
4208 | .elevator_put_req_fn = cfq_put_request, | |
4209 | .elevator_may_queue_fn = cfq_may_queue, | |
4210 | .elevator_init_fn = cfq_init_queue, | |
4211 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 4212 | .trim = cfq_free_io_context, |
1da177e4 | 4213 | }, |
3d1ab40f | 4214 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
4215 | .elevator_name = "cfq", |
4216 | .elevator_owner = THIS_MODULE, | |
4217 | }; | |
4218 | ||
3e252066 VG |
4219 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
4220 | static struct blkio_policy_type blkio_policy_cfq = { | |
4221 | .ops = { | |
4222 | .blkio_unlink_group_fn = cfq_unlink_blkio_group, | |
4223 | .blkio_update_group_weight_fn = cfq_update_blkio_group_weight, | |
4224 | }, | |
062a644d | 4225 | .plid = BLKIO_POLICY_PROP, |
3e252066 VG |
4226 | }; |
4227 | #else | |
4228 | static struct blkio_policy_type blkio_policy_cfq; | |
4229 | #endif | |
4230 | ||
1da177e4 LT |
4231 | static int __init cfq_init(void) |
4232 | { | |
22e2c507 JA |
4233 | /* |
4234 | * could be 0 on HZ < 1000 setups | |
4235 | */ | |
4236 | if (!cfq_slice_async) | |
4237 | cfq_slice_async = 1; | |
4238 | if (!cfq_slice_idle) | |
4239 | cfq_slice_idle = 1; | |
4240 | ||
80bdf0c7 VG |
4241 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
4242 | if (!cfq_group_idle) | |
4243 | cfq_group_idle = 1; | |
4244 | #else | |
4245 | cfq_group_idle = 0; | |
4246 | #endif | |
1da177e4 LT |
4247 | if (cfq_slab_setup()) |
4248 | return -ENOMEM; | |
4249 | ||
2fdd82bd | 4250 | elv_register(&iosched_cfq); |
3e252066 | 4251 | blkio_policy_register(&blkio_policy_cfq); |
1da177e4 | 4252 | |
2fdd82bd | 4253 | return 0; |
1da177e4 LT |
4254 | } |
4255 | ||
4256 | static void __exit cfq_exit(void) | |
4257 | { | |
6e9a4738 | 4258 | DECLARE_COMPLETION_ONSTACK(all_gone); |
3e252066 | 4259 | blkio_policy_unregister(&blkio_policy_cfq); |
1da177e4 | 4260 | elv_unregister(&iosched_cfq); |
334e94de | 4261 | ioc_gone = &all_gone; |
fba82272 OH |
4262 | /* ioc_gone's update must be visible before reading ioc_count */ |
4263 | smp_wmb(); | |
d6de8be7 JA |
4264 | |
4265 | /* | |
4266 | * this also protects us from entering cfq_slab_kill() with | |
4267 | * pending RCU callbacks | |
4268 | */ | |
245b2e70 | 4269 | if (elv_ioc_count_read(cfq_ioc_count)) |
9a11b4ed | 4270 | wait_for_completion(&all_gone); |
80b15c73 | 4271 | ida_destroy(&cic_index_ida); |
83521d3e | 4272 | cfq_slab_kill(); |
1da177e4 LT |
4273 | } |
4274 | ||
4275 | module_init(cfq_init); | |
4276 | module_exit(cfq_exit); | |
4277 | ||
4278 | MODULE_AUTHOR("Jens Axboe"); | |
4279 | MODULE_LICENSE("GPL"); | |
4280 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |