Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
1da177e4 LT |
2 | * CFQ, or complete fairness queueing, disk scheduler. |
3 | * | |
4 | * Based on ideas from a previously unfinished io | |
5 | * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli. | |
6 | * | |
0fe23479 | 7 | * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk> |
1da177e4 | 8 | */ |
1da177e4 | 9 | #include <linux/module.h> |
1cc9be68 AV |
10 | #include <linux/blkdev.h> |
11 | #include <linux/elevator.h> | |
ad5ebd2f | 12 | #include <linux/jiffies.h> |
1da177e4 | 13 | #include <linux/rbtree.h> |
22e2c507 | 14 | #include <linux/ioprio.h> |
7b679138 | 15 | #include <linux/blktrace_api.h> |
25bc6b07 | 16 | #include "blk-cgroup.h" |
1da177e4 LT |
17 | |
18 | /* | |
19 | * tunables | |
20 | */ | |
fe094d98 | 21 | /* max queue in one round of service */ |
abc3c744 | 22 | static const int cfq_quantum = 8; |
64100099 | 23 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
fe094d98 JA |
24 | /* maximum backwards seek, in KiB */ |
25 | static const int cfq_back_max = 16 * 1024; | |
26 | /* penalty of a backwards seek */ | |
27 | static const int cfq_back_penalty = 2; | |
64100099 | 28 | static const int cfq_slice_sync = HZ / 10; |
3b18152c | 29 | static int cfq_slice_async = HZ / 25; |
64100099 | 30 | static const int cfq_slice_async_rq = 2; |
caaa5f9f | 31 | static int cfq_slice_idle = HZ / 125; |
5db5d642 CZ |
32 | static const int cfq_target_latency = HZ * 3/10; /* 300 ms */ |
33 | static const int cfq_hist_divisor = 4; | |
22e2c507 | 34 | |
d9e7620e | 35 | /* |
0871714e | 36 | * offset from end of service tree |
d9e7620e | 37 | */ |
0871714e | 38 | #define CFQ_IDLE_DELAY (HZ / 5) |
d9e7620e JA |
39 | |
40 | /* | |
41 | * below this threshold, we consider thinktime immediate | |
42 | */ | |
43 | #define CFQ_MIN_TT (2) | |
44 | ||
22e2c507 | 45 | #define CFQ_SLICE_SCALE (5) |
45333d5a | 46 | #define CFQ_HW_QUEUE_MIN (5) |
25bc6b07 | 47 | #define CFQ_SERVICE_SHIFT 12 |
22e2c507 | 48 | |
3dde36dd | 49 | #define CFQQ_SEEK_THR (sector_t)(8 * 100) |
41647e7a | 50 | #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32) |
3dde36dd | 51 | #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8) |
ae54abed | 52 | |
fe094d98 JA |
53 | #define RQ_CIC(rq) \ |
54 | ((struct cfq_io_context *) (rq)->elevator_private) | |
7b679138 | 55 | #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2) |
1da177e4 | 56 | |
e18b890b CL |
57 | static struct kmem_cache *cfq_pool; |
58 | static struct kmem_cache *cfq_ioc_pool; | |
1da177e4 | 59 | |
245b2e70 | 60 | static DEFINE_PER_CPU(unsigned long, cfq_ioc_count); |
334e94de | 61 | static struct completion *ioc_gone; |
9a11b4ed | 62 | static DEFINE_SPINLOCK(ioc_gone_lock); |
334e94de | 63 | |
22e2c507 JA |
64 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR |
65 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
22e2c507 JA |
66 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) |
67 | ||
206dc69b | 68 | #define sample_valid(samples) ((samples) > 80) |
1fa8f6d6 | 69 | #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node) |
206dc69b | 70 | |
cc09e299 JA |
71 | /* |
72 | * Most of our rbtree usage is for sorting with min extraction, so | |
73 | * if we cache the leftmost node we don't have to walk down the tree | |
74 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
75 | * move this into the elevator for the rq sorting as well. | |
76 | */ | |
77 | struct cfq_rb_root { | |
78 | struct rb_root rb; | |
79 | struct rb_node *left; | |
aa6f6a3d | 80 | unsigned count; |
73e9ffdd | 81 | unsigned total_weight; |
1fa8f6d6 | 82 | u64 min_vdisktime; |
25bc6b07 | 83 | struct rb_node *active; |
cc09e299 | 84 | }; |
73e9ffdd RK |
85 | #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \ |
86 | .count = 0, .min_vdisktime = 0, } | |
cc09e299 | 87 | |
6118b70b JA |
88 | /* |
89 | * Per process-grouping structure | |
90 | */ | |
91 | struct cfq_queue { | |
92 | /* reference count */ | |
93 | atomic_t ref; | |
94 | /* various state flags, see below */ | |
95 | unsigned int flags; | |
96 | /* parent cfq_data */ | |
97 | struct cfq_data *cfqd; | |
98 | /* service_tree member */ | |
99 | struct rb_node rb_node; | |
100 | /* service_tree key */ | |
101 | unsigned long rb_key; | |
102 | /* prio tree member */ | |
103 | struct rb_node p_node; | |
104 | /* prio tree root we belong to, if any */ | |
105 | struct rb_root *p_root; | |
106 | /* sorted list of pending requests */ | |
107 | struct rb_root sort_list; | |
108 | /* if fifo isn't expired, next request to serve */ | |
109 | struct request *next_rq; | |
110 | /* requests queued in sort_list */ | |
111 | int queued[2]; | |
112 | /* currently allocated requests */ | |
113 | int allocated[2]; | |
114 | /* fifo list of requests in sort_list */ | |
115 | struct list_head fifo; | |
116 | ||
dae739eb VG |
117 | /* time when queue got scheduled in to dispatch first request. */ |
118 | unsigned long dispatch_start; | |
f75edf2d | 119 | unsigned int allocated_slice; |
c4081ba5 | 120 | unsigned int slice_dispatch; |
dae739eb VG |
121 | /* time when first request from queue completed and slice started. */ |
122 | unsigned long slice_start; | |
6118b70b JA |
123 | unsigned long slice_end; |
124 | long slice_resid; | |
6118b70b JA |
125 | |
126 | /* pending metadata requests */ | |
127 | int meta_pending; | |
128 | /* number of requests that are on the dispatch list or inside driver */ | |
129 | int dispatched; | |
130 | ||
131 | /* io prio of this group */ | |
132 | unsigned short ioprio, org_ioprio; | |
133 | unsigned short ioprio_class, org_ioprio_class; | |
134 | ||
c4081ba5 RK |
135 | pid_t pid; |
136 | ||
3dde36dd | 137 | u32 seek_history; |
b2c18e1e JM |
138 | sector_t last_request_pos; |
139 | ||
aa6f6a3d | 140 | struct cfq_rb_root *service_tree; |
df5fe3e8 | 141 | struct cfq_queue *new_cfqq; |
cdb16e8f | 142 | struct cfq_group *cfqg; |
ae30c286 | 143 | struct cfq_group *orig_cfqg; |
22084190 VG |
144 | /* Sectors dispatched in current dispatch round */ |
145 | unsigned long nr_sectors; | |
6118b70b JA |
146 | }; |
147 | ||
c0324a02 | 148 | /* |
718eee05 | 149 | * First index in the service_trees. |
c0324a02 CZ |
150 | * IDLE is handled separately, so it has negative index |
151 | */ | |
152 | enum wl_prio_t { | |
c0324a02 | 153 | BE_WORKLOAD = 0, |
615f0259 VG |
154 | RT_WORKLOAD = 1, |
155 | IDLE_WORKLOAD = 2, | |
c0324a02 CZ |
156 | }; |
157 | ||
718eee05 CZ |
158 | /* |
159 | * Second index in the service_trees. | |
160 | */ | |
161 | enum wl_type_t { | |
162 | ASYNC_WORKLOAD = 0, | |
163 | SYNC_NOIDLE_WORKLOAD = 1, | |
164 | SYNC_WORKLOAD = 2 | |
165 | }; | |
166 | ||
cdb16e8f VG |
167 | /* This is per cgroup per device grouping structure */ |
168 | struct cfq_group { | |
1fa8f6d6 VG |
169 | /* group service_tree member */ |
170 | struct rb_node rb_node; | |
171 | ||
172 | /* group service_tree key */ | |
173 | u64 vdisktime; | |
25bc6b07 | 174 | unsigned int weight; |
1fa8f6d6 VG |
175 | bool on_st; |
176 | ||
177 | /* number of cfqq currently on this group */ | |
178 | int nr_cfqq; | |
179 | ||
58ff82f3 VG |
180 | /* Per group busy queus average. Useful for workload slice calc. */ |
181 | unsigned int busy_queues_avg[2]; | |
cdb16e8f VG |
182 | /* |
183 | * rr lists of queues with requests, onle rr for each priority class. | |
184 | * Counts are embedded in the cfq_rb_root | |
185 | */ | |
186 | struct cfq_rb_root service_trees[2][3]; | |
187 | struct cfq_rb_root service_tree_idle; | |
dae739eb VG |
188 | |
189 | unsigned long saved_workload_slice; | |
190 | enum wl_type_t saved_workload; | |
191 | enum wl_prio_t saved_serving_prio; | |
25fb5169 VG |
192 | struct blkio_group blkg; |
193 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
194 | struct hlist_node cfqd_node; | |
b1c35769 | 195 | atomic_t ref; |
25fb5169 | 196 | #endif |
cdb16e8f | 197 | }; |
718eee05 | 198 | |
22e2c507 JA |
199 | /* |
200 | * Per block device queue structure | |
201 | */ | |
1da177e4 | 202 | struct cfq_data { |
165125e1 | 203 | struct request_queue *queue; |
1fa8f6d6 VG |
204 | /* Root service tree for cfq_groups */ |
205 | struct cfq_rb_root grp_service_tree; | |
cdb16e8f | 206 | struct cfq_group root_group; |
22e2c507 | 207 | |
c0324a02 CZ |
208 | /* |
209 | * The priority currently being served | |
22e2c507 | 210 | */ |
c0324a02 | 211 | enum wl_prio_t serving_prio; |
718eee05 CZ |
212 | enum wl_type_t serving_type; |
213 | unsigned long workload_expires; | |
cdb16e8f | 214 | struct cfq_group *serving_group; |
8e550632 | 215 | bool noidle_tree_requires_idle; |
a36e71f9 JA |
216 | |
217 | /* | |
218 | * Each priority tree is sorted by next_request position. These | |
219 | * trees are used when determining if two or more queues are | |
220 | * interleaving requests (see cfq_close_cooperator). | |
221 | */ | |
222 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
223 | ||
22e2c507 JA |
224 | unsigned int busy_queues; |
225 | ||
53c583d2 CZ |
226 | int rq_in_driver; |
227 | int rq_in_flight[2]; | |
45333d5a AC |
228 | |
229 | /* | |
230 | * queue-depth detection | |
231 | */ | |
232 | int rq_queued; | |
25776e35 | 233 | int hw_tag; |
e459dd08 CZ |
234 | /* |
235 | * hw_tag can be | |
236 | * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) | |
237 | * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) | |
238 | * 0 => no NCQ | |
239 | */ | |
240 | int hw_tag_est_depth; | |
241 | unsigned int hw_tag_samples; | |
1da177e4 | 242 | |
22e2c507 JA |
243 | /* |
244 | * idle window management | |
245 | */ | |
246 | struct timer_list idle_slice_timer; | |
23e018a1 | 247 | struct work_struct unplug_work; |
1da177e4 | 248 | |
22e2c507 JA |
249 | struct cfq_queue *active_queue; |
250 | struct cfq_io_context *active_cic; | |
22e2c507 | 251 | |
c2dea2d1 VT |
252 | /* |
253 | * async queue for each priority case | |
254 | */ | |
255 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
256 | struct cfq_queue *async_idle_cfqq; | |
15c31be4 | 257 | |
6d048f53 | 258 | sector_t last_position; |
1da177e4 | 259 | |
1da177e4 LT |
260 | /* |
261 | * tunables, see top of file | |
262 | */ | |
263 | unsigned int cfq_quantum; | |
22e2c507 | 264 | unsigned int cfq_fifo_expire[2]; |
1da177e4 LT |
265 | unsigned int cfq_back_penalty; |
266 | unsigned int cfq_back_max; | |
22e2c507 JA |
267 | unsigned int cfq_slice[2]; |
268 | unsigned int cfq_slice_async_rq; | |
269 | unsigned int cfq_slice_idle; | |
963b72fc | 270 | unsigned int cfq_latency; |
ae30c286 | 271 | unsigned int cfq_group_isolation; |
d9ff4187 AV |
272 | |
273 | struct list_head cic_list; | |
1da177e4 | 274 | |
6118b70b JA |
275 | /* |
276 | * Fallback dummy cfqq for extreme OOM conditions | |
277 | */ | |
278 | struct cfq_queue oom_cfqq; | |
365722bb | 279 | |
573412b2 | 280 | unsigned long last_delayed_sync; |
25fb5169 VG |
281 | |
282 | /* List of cfq groups being managed on this device*/ | |
283 | struct hlist_head cfqg_list; | |
bb729bc9 | 284 | struct rcu_head rcu; |
1da177e4 LT |
285 | }; |
286 | ||
25fb5169 VG |
287 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd); |
288 | ||
cdb16e8f VG |
289 | static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, |
290 | enum wl_prio_t prio, | |
65b32a57 | 291 | enum wl_type_t type) |
c0324a02 | 292 | { |
1fa8f6d6 VG |
293 | if (!cfqg) |
294 | return NULL; | |
295 | ||
c0324a02 | 296 | if (prio == IDLE_WORKLOAD) |
cdb16e8f | 297 | return &cfqg->service_tree_idle; |
c0324a02 | 298 | |
cdb16e8f | 299 | return &cfqg->service_trees[prio][type]; |
c0324a02 CZ |
300 | } |
301 | ||
3b18152c | 302 | enum cfqq_state_flags { |
b0b8d749 JA |
303 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ |
304 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
b029195d | 305 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ |
b0b8d749 | 306 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ |
b0b8d749 JA |
307 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
308 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
309 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
44f7c160 | 310 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ |
91fac317 | 311 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ |
b3b6d040 | 312 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ |
ae54abed | 313 | CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */ |
76280aff | 314 | CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ |
f75edf2d | 315 | CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */ |
3b18152c JA |
316 | }; |
317 | ||
318 | #define CFQ_CFQQ_FNS(name) \ | |
319 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
320 | { \ | |
fe094d98 | 321 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
322 | } \ |
323 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
324 | { \ | |
fe094d98 | 325 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ |
3b18152c JA |
326 | } \ |
327 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
328 | { \ | |
fe094d98 | 329 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ |
3b18152c JA |
330 | } |
331 | ||
332 | CFQ_CFQQ_FNS(on_rr); | |
333 | CFQ_CFQQ_FNS(wait_request); | |
b029195d | 334 | CFQ_CFQQ_FNS(must_dispatch); |
3b18152c | 335 | CFQ_CFQQ_FNS(must_alloc_slice); |
3b18152c JA |
336 | CFQ_CFQQ_FNS(fifo_expire); |
337 | CFQ_CFQQ_FNS(idle_window); | |
338 | CFQ_CFQQ_FNS(prio_changed); | |
44f7c160 | 339 | CFQ_CFQQ_FNS(slice_new); |
91fac317 | 340 | CFQ_CFQQ_FNS(sync); |
a36e71f9 | 341 | CFQ_CFQQ_FNS(coop); |
ae54abed | 342 | CFQ_CFQQ_FNS(split_coop); |
76280aff | 343 | CFQ_CFQQ_FNS(deep); |
f75edf2d | 344 | CFQ_CFQQ_FNS(wait_busy); |
3b18152c JA |
345 | #undef CFQ_CFQQ_FNS |
346 | ||
2868ef7b VG |
347 | #ifdef CONFIG_DEBUG_CFQ_IOSCHED |
348 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ | |
349 | blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \ | |
350 | cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \ | |
351 | blkg_path(&(cfqq)->cfqg->blkg), ##args); | |
352 | ||
353 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \ | |
354 | blk_add_trace_msg((cfqd)->queue, "%s " fmt, \ | |
355 | blkg_path(&(cfqg)->blkg), ##args); \ | |
356 | ||
357 | #else | |
7b679138 JA |
358 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ |
359 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
2868ef7b VG |
360 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0); |
361 | #endif | |
7b679138 JA |
362 | #define cfq_log(cfqd, fmt, args...) \ |
363 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
364 | ||
615f0259 VG |
365 | /* Traverses through cfq group service trees */ |
366 | #define for_each_cfqg_st(cfqg, i, j, st) \ | |
367 | for (i = 0; i <= IDLE_WORKLOAD; i++) \ | |
368 | for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ | |
369 | : &cfqg->service_tree_idle; \ | |
370 | (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ | |
371 | (i == IDLE_WORKLOAD && j == 0); \ | |
372 | j++, st = i < IDLE_WORKLOAD ? \ | |
373 | &cfqg->service_trees[i][j]: NULL) \ | |
374 | ||
375 | ||
c0324a02 CZ |
376 | static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) |
377 | { | |
378 | if (cfq_class_idle(cfqq)) | |
379 | return IDLE_WORKLOAD; | |
380 | if (cfq_class_rt(cfqq)) | |
381 | return RT_WORKLOAD; | |
382 | return BE_WORKLOAD; | |
383 | } | |
384 | ||
718eee05 CZ |
385 | |
386 | static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) | |
387 | { | |
388 | if (!cfq_cfqq_sync(cfqq)) | |
389 | return ASYNC_WORKLOAD; | |
390 | if (!cfq_cfqq_idle_window(cfqq)) | |
391 | return SYNC_NOIDLE_WORKLOAD; | |
392 | return SYNC_WORKLOAD; | |
393 | } | |
394 | ||
58ff82f3 VG |
395 | static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl, |
396 | struct cfq_data *cfqd, | |
397 | struct cfq_group *cfqg) | |
c0324a02 CZ |
398 | { |
399 | if (wl == IDLE_WORKLOAD) | |
cdb16e8f | 400 | return cfqg->service_tree_idle.count; |
c0324a02 | 401 | |
cdb16e8f VG |
402 | return cfqg->service_trees[wl][ASYNC_WORKLOAD].count |
403 | + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count | |
404 | + cfqg->service_trees[wl][SYNC_WORKLOAD].count; | |
c0324a02 CZ |
405 | } |
406 | ||
f26bd1f0 VG |
407 | static inline int cfqg_busy_async_queues(struct cfq_data *cfqd, |
408 | struct cfq_group *cfqg) | |
409 | { | |
410 | return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count | |
411 | + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count; | |
412 | } | |
413 | ||
165125e1 | 414 | static void cfq_dispatch_insert(struct request_queue *, struct request *); |
a6151c3a | 415 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, |
fd0928df | 416 | struct io_context *, gfp_t); |
4ac845a2 | 417 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, |
91fac317 VT |
418 | struct io_context *); |
419 | ||
420 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 421 | bool is_sync) |
91fac317 | 422 | { |
a6151c3a | 423 | return cic->cfqq[is_sync]; |
91fac317 VT |
424 | } |
425 | ||
426 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
a6151c3a | 427 | struct cfq_queue *cfqq, bool is_sync) |
91fac317 | 428 | { |
a6151c3a | 429 | cic->cfqq[is_sync] = cfqq; |
91fac317 VT |
430 | } |
431 | ||
432 | /* | |
433 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
434 | * set (in which case it could also be direct WRITE). | |
435 | */ | |
a6151c3a | 436 | static inline bool cfq_bio_sync(struct bio *bio) |
91fac317 | 437 | { |
a6151c3a | 438 | return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO); |
91fac317 | 439 | } |
1da177e4 | 440 | |
99f95e52 AM |
441 | /* |
442 | * scheduler run of queue, if there are requests pending and no one in the | |
443 | * driver that will restart queueing | |
444 | */ | |
23e018a1 | 445 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) |
99f95e52 | 446 | { |
7b679138 JA |
447 | if (cfqd->busy_queues) { |
448 | cfq_log(cfqd, "schedule dispatch"); | |
23e018a1 | 449 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); |
7b679138 | 450 | } |
99f95e52 AM |
451 | } |
452 | ||
165125e1 | 453 | static int cfq_queue_empty(struct request_queue *q) |
99f95e52 AM |
454 | { |
455 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
456 | ||
f04a6424 | 457 | return !cfqd->rq_queued; |
99f95e52 AM |
458 | } |
459 | ||
44f7c160 JA |
460 | /* |
461 | * Scale schedule slice based on io priority. Use the sync time slice only | |
462 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
463 | * io only, should not get full sync slice length. | |
464 | */ | |
a6151c3a | 465 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, |
d9e7620e | 466 | unsigned short prio) |
44f7c160 | 467 | { |
d9e7620e | 468 | const int base_slice = cfqd->cfq_slice[sync]; |
44f7c160 | 469 | |
d9e7620e JA |
470 | WARN_ON(prio >= IOPRIO_BE_NR); |
471 | ||
472 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
473 | } | |
44f7c160 | 474 | |
d9e7620e JA |
475 | static inline int |
476 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
477 | { | |
478 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
44f7c160 JA |
479 | } |
480 | ||
25bc6b07 VG |
481 | static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg) |
482 | { | |
483 | u64 d = delta << CFQ_SERVICE_SHIFT; | |
484 | ||
485 | d = d * BLKIO_WEIGHT_DEFAULT; | |
486 | do_div(d, cfqg->weight); | |
487 | return d; | |
488 | } | |
489 | ||
490 | static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
491 | { | |
492 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
493 | if (delta > 0) | |
494 | min_vdisktime = vdisktime; | |
495 | ||
496 | return min_vdisktime; | |
497 | } | |
498 | ||
499 | static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
500 | { | |
501 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
502 | if (delta < 0) | |
503 | min_vdisktime = vdisktime; | |
504 | ||
505 | return min_vdisktime; | |
506 | } | |
507 | ||
508 | static void update_min_vdisktime(struct cfq_rb_root *st) | |
509 | { | |
510 | u64 vdisktime = st->min_vdisktime; | |
511 | struct cfq_group *cfqg; | |
512 | ||
513 | if (st->active) { | |
514 | cfqg = rb_entry_cfqg(st->active); | |
515 | vdisktime = cfqg->vdisktime; | |
516 | } | |
517 | ||
518 | if (st->left) { | |
519 | cfqg = rb_entry_cfqg(st->left); | |
520 | vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime); | |
521 | } | |
522 | ||
523 | st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime); | |
524 | } | |
525 | ||
5db5d642 CZ |
526 | /* |
527 | * get averaged number of queues of RT/BE priority. | |
528 | * average is updated, with a formula that gives more weight to higher numbers, | |
529 | * to quickly follows sudden increases and decrease slowly | |
530 | */ | |
531 | ||
58ff82f3 VG |
532 | static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, |
533 | struct cfq_group *cfqg, bool rt) | |
5869619c | 534 | { |
5db5d642 CZ |
535 | unsigned min_q, max_q; |
536 | unsigned mult = cfq_hist_divisor - 1; | |
537 | unsigned round = cfq_hist_divisor / 2; | |
58ff82f3 | 538 | unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); |
5db5d642 | 539 | |
58ff82f3 VG |
540 | min_q = min(cfqg->busy_queues_avg[rt], busy); |
541 | max_q = max(cfqg->busy_queues_avg[rt], busy); | |
542 | cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
5db5d642 | 543 | cfq_hist_divisor; |
58ff82f3 VG |
544 | return cfqg->busy_queues_avg[rt]; |
545 | } | |
546 | ||
547 | static inline unsigned | |
548 | cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
549 | { | |
550 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
551 | ||
552 | return cfq_target_latency * cfqg->weight / st->total_weight; | |
5db5d642 CZ |
553 | } |
554 | ||
44f7c160 JA |
555 | static inline void |
556 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
557 | { | |
5db5d642 CZ |
558 | unsigned slice = cfq_prio_to_slice(cfqd, cfqq); |
559 | if (cfqd->cfq_latency) { | |
58ff82f3 VG |
560 | /* |
561 | * interested queues (we consider only the ones with the same | |
562 | * priority class in the cfq group) | |
563 | */ | |
564 | unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, | |
565 | cfq_class_rt(cfqq)); | |
5db5d642 CZ |
566 | unsigned sync_slice = cfqd->cfq_slice[1]; |
567 | unsigned expect_latency = sync_slice * iq; | |
58ff82f3 VG |
568 | unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg); |
569 | ||
570 | if (expect_latency > group_slice) { | |
5db5d642 CZ |
571 | unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; |
572 | /* scale low_slice according to IO priority | |
573 | * and sync vs async */ | |
574 | unsigned low_slice = | |
575 | min(slice, base_low_slice * slice / sync_slice); | |
576 | /* the adapted slice value is scaled to fit all iqs | |
577 | * into the target latency */ | |
58ff82f3 | 578 | slice = max(slice * group_slice / expect_latency, |
5db5d642 CZ |
579 | low_slice); |
580 | } | |
581 | } | |
dae739eb | 582 | cfqq->slice_start = jiffies; |
5db5d642 | 583 | cfqq->slice_end = jiffies + slice; |
f75edf2d | 584 | cfqq->allocated_slice = slice; |
7b679138 | 585 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); |
44f7c160 JA |
586 | } |
587 | ||
588 | /* | |
589 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
590 | * isn't valid until the first request from the dispatch is activated | |
591 | * and the slice time set. | |
592 | */ | |
a6151c3a | 593 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) |
44f7c160 JA |
594 | { |
595 | if (cfq_cfqq_slice_new(cfqq)) | |
596 | return 0; | |
597 | if (time_before(jiffies, cfqq->slice_end)) | |
598 | return 0; | |
599 | ||
600 | return 1; | |
601 | } | |
602 | ||
1da177e4 | 603 | /* |
5e705374 | 604 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1da177e4 | 605 | * We choose the request that is closest to the head right now. Distance |
e8a99053 | 606 | * behind the head is penalized and only allowed to a certain extent. |
1da177e4 | 607 | */ |
5e705374 | 608 | static struct request * |
cf7c25cf | 609 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) |
1da177e4 | 610 | { |
cf7c25cf | 611 | sector_t s1, s2, d1 = 0, d2 = 0; |
1da177e4 | 612 | unsigned long back_max; |
e8a99053 AM |
613 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
614 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
615 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1da177e4 | 616 | |
5e705374 JA |
617 | if (rq1 == NULL || rq1 == rq2) |
618 | return rq2; | |
619 | if (rq2 == NULL) | |
620 | return rq1; | |
9c2c38a1 | 621 | |
5e705374 JA |
622 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
623 | return rq1; | |
624 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
625 | return rq2; | |
374f84ac JA |
626 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) |
627 | return rq1; | |
628 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
629 | return rq2; | |
1da177e4 | 630 | |
83096ebf TH |
631 | s1 = blk_rq_pos(rq1); |
632 | s2 = blk_rq_pos(rq2); | |
1da177e4 | 633 | |
1da177e4 LT |
634 | /* |
635 | * by definition, 1KiB is 2 sectors | |
636 | */ | |
637 | back_max = cfqd->cfq_back_max * 2; | |
638 | ||
639 | /* | |
640 | * Strict one way elevator _except_ in the case where we allow | |
641 | * short backward seeks which are biased as twice the cost of a | |
642 | * similar forward seek. | |
643 | */ | |
644 | if (s1 >= last) | |
645 | d1 = s1 - last; | |
646 | else if (s1 + back_max >= last) | |
647 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
648 | else | |
e8a99053 | 649 | wrap |= CFQ_RQ1_WRAP; |
1da177e4 LT |
650 | |
651 | if (s2 >= last) | |
652 | d2 = s2 - last; | |
653 | else if (s2 + back_max >= last) | |
654 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
655 | else | |
e8a99053 | 656 | wrap |= CFQ_RQ2_WRAP; |
1da177e4 LT |
657 | |
658 | /* Found required data */ | |
e8a99053 AM |
659 | |
660 | /* | |
661 | * By doing switch() on the bit mask "wrap" we avoid having to | |
662 | * check two variables for all permutations: --> faster! | |
663 | */ | |
664 | switch (wrap) { | |
5e705374 | 665 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
e8a99053 | 666 | if (d1 < d2) |
5e705374 | 667 | return rq1; |
e8a99053 | 668 | else if (d2 < d1) |
5e705374 | 669 | return rq2; |
e8a99053 AM |
670 | else { |
671 | if (s1 >= s2) | |
5e705374 | 672 | return rq1; |
e8a99053 | 673 | else |
5e705374 | 674 | return rq2; |
e8a99053 | 675 | } |
1da177e4 | 676 | |
e8a99053 | 677 | case CFQ_RQ2_WRAP: |
5e705374 | 678 | return rq1; |
e8a99053 | 679 | case CFQ_RQ1_WRAP: |
5e705374 JA |
680 | return rq2; |
681 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
e8a99053 AM |
682 | default: |
683 | /* | |
684 | * Since both rqs are wrapped, | |
685 | * start with the one that's further behind head | |
686 | * (--> only *one* back seek required), | |
687 | * since back seek takes more time than forward. | |
688 | */ | |
689 | if (s1 <= s2) | |
5e705374 | 690 | return rq1; |
1da177e4 | 691 | else |
5e705374 | 692 | return rq2; |
1da177e4 LT |
693 | } |
694 | } | |
695 | ||
498d3aa2 JA |
696 | /* |
697 | * The below is leftmost cache rbtree addon | |
698 | */ | |
0871714e | 699 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) |
cc09e299 | 700 | { |
615f0259 VG |
701 | /* Service tree is empty */ |
702 | if (!root->count) | |
703 | return NULL; | |
704 | ||
cc09e299 JA |
705 | if (!root->left) |
706 | root->left = rb_first(&root->rb); | |
707 | ||
0871714e JA |
708 | if (root->left) |
709 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
710 | ||
711 | return NULL; | |
cc09e299 JA |
712 | } |
713 | ||
1fa8f6d6 VG |
714 | static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) |
715 | { | |
716 | if (!root->left) | |
717 | root->left = rb_first(&root->rb); | |
718 | ||
719 | if (root->left) | |
720 | return rb_entry_cfqg(root->left); | |
721 | ||
722 | return NULL; | |
723 | } | |
724 | ||
a36e71f9 JA |
725 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
726 | { | |
727 | rb_erase(n, root); | |
728 | RB_CLEAR_NODE(n); | |
729 | } | |
730 | ||
cc09e299 JA |
731 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) |
732 | { | |
733 | if (root->left == n) | |
734 | root->left = NULL; | |
a36e71f9 | 735 | rb_erase_init(n, &root->rb); |
aa6f6a3d | 736 | --root->count; |
cc09e299 JA |
737 | } |
738 | ||
1da177e4 LT |
739 | /* |
740 | * would be nice to take fifo expire time into account as well | |
741 | */ | |
5e705374 JA |
742 | static struct request * |
743 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
744 | struct request *last) | |
1da177e4 | 745 | { |
21183b07 JA |
746 | struct rb_node *rbnext = rb_next(&last->rb_node); |
747 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
5e705374 | 748 | struct request *next = NULL, *prev = NULL; |
1da177e4 | 749 | |
21183b07 | 750 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1da177e4 LT |
751 | |
752 | if (rbprev) | |
5e705374 | 753 | prev = rb_entry_rq(rbprev); |
1da177e4 | 754 | |
21183b07 | 755 | if (rbnext) |
5e705374 | 756 | next = rb_entry_rq(rbnext); |
21183b07 JA |
757 | else { |
758 | rbnext = rb_first(&cfqq->sort_list); | |
759 | if (rbnext && rbnext != &last->rb_node) | |
5e705374 | 760 | next = rb_entry_rq(rbnext); |
21183b07 | 761 | } |
1da177e4 | 762 | |
cf7c25cf | 763 | return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); |
1da177e4 LT |
764 | } |
765 | ||
d9e7620e JA |
766 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, |
767 | struct cfq_queue *cfqq) | |
1da177e4 | 768 | { |
d9e7620e JA |
769 | /* |
770 | * just an approximation, should be ok. | |
771 | */ | |
cdb16e8f | 772 | return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - |
464191c6 | 773 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); |
d9e7620e JA |
774 | } |
775 | ||
1fa8f6d6 VG |
776 | static inline s64 |
777 | cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
778 | { | |
779 | return cfqg->vdisktime - st->min_vdisktime; | |
780 | } | |
781 | ||
782 | static void | |
783 | __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
784 | { | |
785 | struct rb_node **node = &st->rb.rb_node; | |
786 | struct rb_node *parent = NULL; | |
787 | struct cfq_group *__cfqg; | |
788 | s64 key = cfqg_key(st, cfqg); | |
789 | int left = 1; | |
790 | ||
791 | while (*node != NULL) { | |
792 | parent = *node; | |
793 | __cfqg = rb_entry_cfqg(parent); | |
794 | ||
795 | if (key < cfqg_key(st, __cfqg)) | |
796 | node = &parent->rb_left; | |
797 | else { | |
798 | node = &parent->rb_right; | |
799 | left = 0; | |
800 | } | |
801 | } | |
802 | ||
803 | if (left) | |
804 | st->left = &cfqg->rb_node; | |
805 | ||
806 | rb_link_node(&cfqg->rb_node, parent, node); | |
807 | rb_insert_color(&cfqg->rb_node, &st->rb); | |
808 | } | |
809 | ||
810 | static void | |
811 | cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
812 | { | |
813 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
814 | struct cfq_group *__cfqg; | |
815 | struct rb_node *n; | |
816 | ||
817 | cfqg->nr_cfqq++; | |
818 | if (cfqg->on_st) | |
819 | return; | |
820 | ||
821 | /* | |
822 | * Currently put the group at the end. Later implement something | |
823 | * so that groups get lesser vtime based on their weights, so that | |
824 | * if group does not loose all if it was not continously backlogged. | |
825 | */ | |
826 | n = rb_last(&st->rb); | |
827 | if (n) { | |
828 | __cfqg = rb_entry_cfqg(n); | |
829 | cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; | |
830 | } else | |
831 | cfqg->vdisktime = st->min_vdisktime; | |
832 | ||
833 | __cfq_group_service_tree_add(st, cfqg); | |
834 | cfqg->on_st = true; | |
58ff82f3 | 835 | st->total_weight += cfqg->weight; |
1fa8f6d6 VG |
836 | } |
837 | ||
838 | static void | |
839 | cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
840 | { | |
841 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
842 | ||
25bc6b07 VG |
843 | if (st->active == &cfqg->rb_node) |
844 | st->active = NULL; | |
845 | ||
1fa8f6d6 VG |
846 | BUG_ON(cfqg->nr_cfqq < 1); |
847 | cfqg->nr_cfqq--; | |
25bc6b07 | 848 | |
1fa8f6d6 VG |
849 | /* If there are other cfq queues under this group, don't delete it */ |
850 | if (cfqg->nr_cfqq) | |
851 | return; | |
852 | ||
2868ef7b | 853 | cfq_log_cfqg(cfqd, cfqg, "del_from_rr group"); |
1fa8f6d6 | 854 | cfqg->on_st = false; |
58ff82f3 | 855 | st->total_weight -= cfqg->weight; |
1fa8f6d6 VG |
856 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) |
857 | cfq_rb_erase(&cfqg->rb_node, st); | |
dae739eb | 858 | cfqg->saved_workload_slice = 0; |
22084190 | 859 | blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1); |
dae739eb VG |
860 | } |
861 | ||
862 | static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq) | |
863 | { | |
f75edf2d | 864 | unsigned int slice_used; |
dae739eb VG |
865 | |
866 | /* | |
867 | * Queue got expired before even a single request completed or | |
868 | * got expired immediately after first request completion. | |
869 | */ | |
870 | if (!cfqq->slice_start || cfqq->slice_start == jiffies) { | |
871 | /* | |
872 | * Also charge the seek time incurred to the group, otherwise | |
873 | * if there are mutiple queues in the group, each can dispatch | |
874 | * a single request on seeky media and cause lots of seek time | |
875 | * and group will never know it. | |
876 | */ | |
877 | slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start), | |
878 | 1); | |
879 | } else { | |
880 | slice_used = jiffies - cfqq->slice_start; | |
f75edf2d VG |
881 | if (slice_used > cfqq->allocated_slice) |
882 | slice_used = cfqq->allocated_slice; | |
dae739eb VG |
883 | } |
884 | ||
22084190 VG |
885 | cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used, |
886 | cfqq->nr_sectors); | |
dae739eb VG |
887 | return slice_used; |
888 | } | |
889 | ||
890 | static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg, | |
891 | struct cfq_queue *cfqq) | |
892 | { | |
893 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
f26bd1f0 VG |
894 | unsigned int used_sl, charge_sl; |
895 | int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg) | |
896 | - cfqg->service_tree_idle.count; | |
897 | ||
898 | BUG_ON(nr_sync < 0); | |
899 | used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq); | |
dae739eb | 900 | |
f26bd1f0 VG |
901 | if (!cfq_cfqq_sync(cfqq) && !nr_sync) |
902 | charge_sl = cfqq->allocated_slice; | |
dae739eb VG |
903 | |
904 | /* Can't update vdisktime while group is on service tree */ | |
905 | cfq_rb_erase(&cfqg->rb_node, st); | |
f26bd1f0 | 906 | cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg); |
dae739eb VG |
907 | __cfq_group_service_tree_add(st, cfqg); |
908 | ||
909 | /* This group is being expired. Save the context */ | |
910 | if (time_after(cfqd->workload_expires, jiffies)) { | |
911 | cfqg->saved_workload_slice = cfqd->workload_expires | |
912 | - jiffies; | |
913 | cfqg->saved_workload = cfqd->serving_type; | |
914 | cfqg->saved_serving_prio = cfqd->serving_prio; | |
915 | } else | |
916 | cfqg->saved_workload_slice = 0; | |
2868ef7b VG |
917 | |
918 | cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime, | |
919 | st->min_vdisktime); | |
22084190 VG |
920 | blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl, |
921 | cfqq->nr_sectors); | |
1fa8f6d6 VG |
922 | } |
923 | ||
25fb5169 VG |
924 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
925 | static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg) | |
926 | { | |
927 | if (blkg) | |
928 | return container_of(blkg, struct cfq_group, blkg); | |
929 | return NULL; | |
930 | } | |
931 | ||
f8d461d6 VG |
932 | void |
933 | cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight) | |
934 | { | |
935 | cfqg_of_blkg(blkg)->weight = weight; | |
936 | } | |
937 | ||
25fb5169 VG |
938 | static struct cfq_group * |
939 | cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create) | |
940 | { | |
941 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup); | |
942 | struct cfq_group *cfqg = NULL; | |
943 | void *key = cfqd; | |
944 | int i, j; | |
945 | struct cfq_rb_root *st; | |
22084190 VG |
946 | struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; |
947 | unsigned int major, minor; | |
25fb5169 | 948 | |
25fb5169 VG |
949 | cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key)); |
950 | if (cfqg || !create) | |
951 | goto done; | |
952 | ||
953 | cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node); | |
954 | if (!cfqg) | |
955 | goto done; | |
956 | ||
957 | cfqg->weight = blkcg->weight; | |
958 | for_each_cfqg_st(cfqg, i, j, st) | |
959 | *st = CFQ_RB_ROOT; | |
960 | RB_CLEAR_NODE(&cfqg->rb_node); | |
961 | ||
b1c35769 VG |
962 | /* |
963 | * Take the initial reference that will be released on destroy | |
964 | * This can be thought of a joint reference by cgroup and | |
965 | * elevator which will be dropped by either elevator exit | |
966 | * or cgroup deletion path depending on who is exiting first. | |
967 | */ | |
968 | atomic_set(&cfqg->ref, 1); | |
969 | ||
25fb5169 | 970 | /* Add group onto cgroup list */ |
22084190 VG |
971 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); |
972 | blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd, | |
973 | MKDEV(major, minor)); | |
25fb5169 VG |
974 | |
975 | /* Add group on cfqd list */ | |
976 | hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); | |
977 | ||
978 | done: | |
25fb5169 VG |
979 | return cfqg; |
980 | } | |
981 | ||
982 | /* | |
983 | * Search for the cfq group current task belongs to. If create = 1, then also | |
984 | * create the cfq group if it does not exist. request_queue lock must be held. | |
985 | */ | |
986 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create) | |
987 | { | |
988 | struct cgroup *cgroup; | |
989 | struct cfq_group *cfqg = NULL; | |
990 | ||
991 | rcu_read_lock(); | |
992 | cgroup = task_cgroup(current, blkio_subsys_id); | |
993 | cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create); | |
994 | if (!cfqg && create) | |
995 | cfqg = &cfqd->root_group; | |
996 | rcu_read_unlock(); | |
997 | return cfqg; | |
998 | } | |
999 | ||
1000 | static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) | |
1001 | { | |
1002 | /* Currently, all async queues are mapped to root group */ | |
1003 | if (!cfq_cfqq_sync(cfqq)) | |
1004 | cfqg = &cfqq->cfqd->root_group; | |
1005 | ||
1006 | cfqq->cfqg = cfqg; | |
b1c35769 VG |
1007 | /* cfqq reference on cfqg */ |
1008 | atomic_inc(&cfqq->cfqg->ref); | |
1009 | } | |
1010 | ||
1011 | static void cfq_put_cfqg(struct cfq_group *cfqg) | |
1012 | { | |
1013 | struct cfq_rb_root *st; | |
1014 | int i, j; | |
1015 | ||
1016 | BUG_ON(atomic_read(&cfqg->ref) <= 0); | |
1017 | if (!atomic_dec_and_test(&cfqg->ref)) | |
1018 | return; | |
1019 | for_each_cfqg_st(cfqg, i, j, st) | |
1020 | BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL); | |
1021 | kfree(cfqg); | |
1022 | } | |
1023 | ||
1024 | static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1025 | { | |
1026 | /* Something wrong if we are trying to remove same group twice */ | |
1027 | BUG_ON(hlist_unhashed(&cfqg->cfqd_node)); | |
1028 | ||
1029 | hlist_del_init(&cfqg->cfqd_node); | |
1030 | ||
1031 | /* | |
1032 | * Put the reference taken at the time of creation so that when all | |
1033 | * queues are gone, group can be destroyed. | |
1034 | */ | |
1035 | cfq_put_cfqg(cfqg); | |
1036 | } | |
1037 | ||
1038 | static void cfq_release_cfq_groups(struct cfq_data *cfqd) | |
1039 | { | |
1040 | struct hlist_node *pos, *n; | |
1041 | struct cfq_group *cfqg; | |
1042 | ||
1043 | hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) { | |
1044 | /* | |
1045 | * If cgroup removal path got to blk_group first and removed | |
1046 | * it from cgroup list, then it will take care of destroying | |
1047 | * cfqg also. | |
1048 | */ | |
1049 | if (!blkiocg_del_blkio_group(&cfqg->blkg)) | |
1050 | cfq_destroy_cfqg(cfqd, cfqg); | |
1051 | } | |
25fb5169 | 1052 | } |
b1c35769 VG |
1053 | |
1054 | /* | |
1055 | * Blk cgroup controller notification saying that blkio_group object is being | |
1056 | * delinked as associated cgroup object is going away. That also means that | |
1057 | * no new IO will come in this group. So get rid of this group as soon as | |
1058 | * any pending IO in the group is finished. | |
1059 | * | |
1060 | * This function is called under rcu_read_lock(). key is the rcu protected | |
1061 | * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu | |
1062 | * read lock. | |
1063 | * | |
1064 | * "key" was fetched from blkio_group under blkio_cgroup->lock. That means | |
1065 | * it should not be NULL as even if elevator was exiting, cgroup deltion | |
1066 | * path got to it first. | |
1067 | */ | |
1068 | void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg) | |
1069 | { | |
1070 | unsigned long flags; | |
1071 | struct cfq_data *cfqd = key; | |
1072 | ||
1073 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
1074 | cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg)); | |
1075 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
1076 | } | |
1077 | ||
25fb5169 VG |
1078 | #else /* GROUP_IOSCHED */ |
1079 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create) | |
1080 | { | |
1081 | return &cfqd->root_group; | |
1082 | } | |
1083 | static inline void | |
1084 | cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) { | |
1085 | cfqq->cfqg = cfqg; | |
1086 | } | |
1087 | ||
b1c35769 VG |
1088 | static void cfq_release_cfq_groups(struct cfq_data *cfqd) {} |
1089 | static inline void cfq_put_cfqg(struct cfq_group *cfqg) {} | |
1090 | ||
25fb5169 VG |
1091 | #endif /* GROUP_IOSCHED */ |
1092 | ||
498d3aa2 | 1093 | /* |
c0324a02 | 1094 | * The cfqd->service_trees holds all pending cfq_queue's that have |
498d3aa2 JA |
1095 | * requests waiting to be processed. It is sorted in the order that |
1096 | * we will service the queues. | |
1097 | */ | |
a36e71f9 | 1098 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 1099 | bool add_front) |
d9e7620e | 1100 | { |
0871714e JA |
1101 | struct rb_node **p, *parent; |
1102 | struct cfq_queue *__cfqq; | |
d9e7620e | 1103 | unsigned long rb_key; |
c0324a02 | 1104 | struct cfq_rb_root *service_tree; |
498d3aa2 | 1105 | int left; |
dae739eb | 1106 | int new_cfqq = 1; |
ae30c286 VG |
1107 | int group_changed = 0; |
1108 | ||
1109 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
1110 | if (!cfqd->cfq_group_isolation | |
1111 | && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD | |
1112 | && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) { | |
1113 | /* Move this cfq to root group */ | |
1114 | cfq_log_cfqq(cfqd, cfqq, "moving to root group"); | |
1115 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) | |
1116 | cfq_group_service_tree_del(cfqd, cfqq->cfqg); | |
1117 | cfqq->orig_cfqg = cfqq->cfqg; | |
1118 | cfqq->cfqg = &cfqd->root_group; | |
1119 | atomic_inc(&cfqd->root_group.ref); | |
1120 | group_changed = 1; | |
1121 | } else if (!cfqd->cfq_group_isolation | |
1122 | && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) { | |
1123 | /* cfqq is sequential now needs to go to its original group */ | |
1124 | BUG_ON(cfqq->cfqg != &cfqd->root_group); | |
1125 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) | |
1126 | cfq_group_service_tree_del(cfqd, cfqq->cfqg); | |
1127 | cfq_put_cfqg(cfqq->cfqg); | |
1128 | cfqq->cfqg = cfqq->orig_cfqg; | |
1129 | cfqq->orig_cfqg = NULL; | |
1130 | group_changed = 1; | |
1131 | cfq_log_cfqq(cfqd, cfqq, "moved to origin group"); | |
1132 | } | |
1133 | #endif | |
d9e7620e | 1134 | |
cdb16e8f | 1135 | service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), |
65b32a57 | 1136 | cfqq_type(cfqq)); |
0871714e JA |
1137 | if (cfq_class_idle(cfqq)) { |
1138 | rb_key = CFQ_IDLE_DELAY; | |
aa6f6a3d | 1139 | parent = rb_last(&service_tree->rb); |
0871714e JA |
1140 | if (parent && parent != &cfqq->rb_node) { |
1141 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1142 | rb_key += __cfqq->rb_key; | |
1143 | } else | |
1144 | rb_key += jiffies; | |
1145 | } else if (!add_front) { | |
b9c8946b JA |
1146 | /* |
1147 | * Get our rb key offset. Subtract any residual slice | |
1148 | * value carried from last service. A negative resid | |
1149 | * count indicates slice overrun, and this should position | |
1150 | * the next service time further away in the tree. | |
1151 | */ | |
edd75ffd | 1152 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; |
b9c8946b | 1153 | rb_key -= cfqq->slice_resid; |
edd75ffd | 1154 | cfqq->slice_resid = 0; |
48e025e6 CZ |
1155 | } else { |
1156 | rb_key = -HZ; | |
aa6f6a3d | 1157 | __cfqq = cfq_rb_first(service_tree); |
48e025e6 CZ |
1158 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; |
1159 | } | |
1da177e4 | 1160 | |
d9e7620e | 1161 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
dae739eb | 1162 | new_cfqq = 0; |
99f9628a | 1163 | /* |
d9e7620e | 1164 | * same position, nothing more to do |
99f9628a | 1165 | */ |
c0324a02 CZ |
1166 | if (rb_key == cfqq->rb_key && |
1167 | cfqq->service_tree == service_tree) | |
d9e7620e | 1168 | return; |
1da177e4 | 1169 | |
aa6f6a3d CZ |
1170 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); |
1171 | cfqq->service_tree = NULL; | |
1da177e4 | 1172 | } |
d9e7620e | 1173 | |
498d3aa2 | 1174 | left = 1; |
0871714e | 1175 | parent = NULL; |
aa6f6a3d CZ |
1176 | cfqq->service_tree = service_tree; |
1177 | p = &service_tree->rb.rb_node; | |
d9e7620e | 1178 | while (*p) { |
67060e37 | 1179 | struct rb_node **n; |
cc09e299 | 1180 | |
d9e7620e JA |
1181 | parent = *p; |
1182 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1183 | ||
0c534e0a | 1184 | /* |
c0324a02 | 1185 | * sort by key, that represents service time. |
0c534e0a | 1186 | */ |
c0324a02 | 1187 | if (time_before(rb_key, __cfqq->rb_key)) |
67060e37 | 1188 | n = &(*p)->rb_left; |
c0324a02 | 1189 | else { |
67060e37 | 1190 | n = &(*p)->rb_right; |
cc09e299 | 1191 | left = 0; |
c0324a02 | 1192 | } |
67060e37 JA |
1193 | |
1194 | p = n; | |
d9e7620e JA |
1195 | } |
1196 | ||
cc09e299 | 1197 | if (left) |
aa6f6a3d | 1198 | service_tree->left = &cfqq->rb_node; |
cc09e299 | 1199 | |
d9e7620e JA |
1200 | cfqq->rb_key = rb_key; |
1201 | rb_link_node(&cfqq->rb_node, parent, p); | |
aa6f6a3d CZ |
1202 | rb_insert_color(&cfqq->rb_node, &service_tree->rb); |
1203 | service_tree->count++; | |
ae30c286 | 1204 | if ((add_front || !new_cfqq) && !group_changed) |
dae739eb | 1205 | return; |
1fa8f6d6 | 1206 | cfq_group_service_tree_add(cfqd, cfqq->cfqg); |
1da177e4 LT |
1207 | } |
1208 | ||
a36e71f9 | 1209 | static struct cfq_queue * |
f2d1f0ae JA |
1210 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, |
1211 | sector_t sector, struct rb_node **ret_parent, | |
1212 | struct rb_node ***rb_link) | |
a36e71f9 | 1213 | { |
a36e71f9 JA |
1214 | struct rb_node **p, *parent; |
1215 | struct cfq_queue *cfqq = NULL; | |
1216 | ||
1217 | parent = NULL; | |
1218 | p = &root->rb_node; | |
1219 | while (*p) { | |
1220 | struct rb_node **n; | |
1221 | ||
1222 | parent = *p; | |
1223 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
1224 | ||
1225 | /* | |
1226 | * Sort strictly based on sector. Smallest to the left, | |
1227 | * largest to the right. | |
1228 | */ | |
2e46e8b2 | 1229 | if (sector > blk_rq_pos(cfqq->next_rq)) |
a36e71f9 | 1230 | n = &(*p)->rb_right; |
2e46e8b2 | 1231 | else if (sector < blk_rq_pos(cfqq->next_rq)) |
a36e71f9 JA |
1232 | n = &(*p)->rb_left; |
1233 | else | |
1234 | break; | |
1235 | p = n; | |
3ac6c9f8 | 1236 | cfqq = NULL; |
a36e71f9 JA |
1237 | } |
1238 | ||
1239 | *ret_parent = parent; | |
1240 | if (rb_link) | |
1241 | *rb_link = p; | |
3ac6c9f8 | 1242 | return cfqq; |
a36e71f9 JA |
1243 | } |
1244 | ||
1245 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1246 | { | |
a36e71f9 JA |
1247 | struct rb_node **p, *parent; |
1248 | struct cfq_queue *__cfqq; | |
1249 | ||
f2d1f0ae JA |
1250 | if (cfqq->p_root) { |
1251 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1252 | cfqq->p_root = NULL; | |
1253 | } | |
a36e71f9 JA |
1254 | |
1255 | if (cfq_class_idle(cfqq)) | |
1256 | return; | |
1257 | if (!cfqq->next_rq) | |
1258 | return; | |
1259 | ||
f2d1f0ae | 1260 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; |
2e46e8b2 TH |
1261 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, |
1262 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
3ac6c9f8 JA |
1263 | if (!__cfqq) { |
1264 | rb_link_node(&cfqq->p_node, parent, p); | |
f2d1f0ae JA |
1265 | rb_insert_color(&cfqq->p_node, cfqq->p_root); |
1266 | } else | |
1267 | cfqq->p_root = NULL; | |
a36e71f9 JA |
1268 | } |
1269 | ||
498d3aa2 JA |
1270 | /* |
1271 | * Update cfqq's position in the service tree. | |
1272 | */ | |
edd75ffd | 1273 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
6d048f53 | 1274 | { |
6d048f53 JA |
1275 | /* |
1276 | * Resorting requires the cfqq to be on the RR list already. | |
1277 | */ | |
a36e71f9 | 1278 | if (cfq_cfqq_on_rr(cfqq)) { |
edd75ffd | 1279 | cfq_service_tree_add(cfqd, cfqq, 0); |
a36e71f9 JA |
1280 | cfq_prio_tree_add(cfqd, cfqq); |
1281 | } | |
6d048f53 JA |
1282 | } |
1283 | ||
1da177e4 LT |
1284 | /* |
1285 | * add to busy list of queues for service, trying to be fair in ordering | |
22e2c507 | 1286 | * the pending list according to last request service |
1da177e4 | 1287 | */ |
febffd61 | 1288 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1289 | { |
7b679138 | 1290 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); |
3b18152c JA |
1291 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1292 | cfq_mark_cfqq_on_rr(cfqq); | |
1da177e4 LT |
1293 | cfqd->busy_queues++; |
1294 | ||
edd75ffd | 1295 | cfq_resort_rr_list(cfqd, cfqq); |
1da177e4 LT |
1296 | } |
1297 | ||
498d3aa2 JA |
1298 | /* |
1299 | * Called when the cfqq no longer has requests pending, remove it from | |
1300 | * the service tree. | |
1301 | */ | |
febffd61 | 1302 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 1303 | { |
7b679138 | 1304 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); |
3b18152c JA |
1305 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); |
1306 | cfq_clear_cfqq_on_rr(cfqq); | |
1da177e4 | 1307 | |
aa6f6a3d CZ |
1308 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { |
1309 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
1310 | cfqq->service_tree = NULL; | |
1311 | } | |
f2d1f0ae JA |
1312 | if (cfqq->p_root) { |
1313 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1314 | cfqq->p_root = NULL; | |
1315 | } | |
d9e7620e | 1316 | |
1fa8f6d6 | 1317 | cfq_group_service_tree_del(cfqd, cfqq->cfqg); |
1da177e4 LT |
1318 | BUG_ON(!cfqd->busy_queues); |
1319 | cfqd->busy_queues--; | |
1320 | } | |
1321 | ||
1322 | /* | |
1323 | * rb tree support functions | |
1324 | */ | |
febffd61 | 1325 | static void cfq_del_rq_rb(struct request *rq) |
1da177e4 | 1326 | { |
5e705374 | 1327 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
5e705374 | 1328 | const int sync = rq_is_sync(rq); |
1da177e4 | 1329 | |
b4878f24 JA |
1330 | BUG_ON(!cfqq->queued[sync]); |
1331 | cfqq->queued[sync]--; | |
1da177e4 | 1332 | |
5e705374 | 1333 | elv_rb_del(&cfqq->sort_list, rq); |
1da177e4 | 1334 | |
f04a6424 VG |
1335 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { |
1336 | /* | |
1337 | * Queue will be deleted from service tree when we actually | |
1338 | * expire it later. Right now just remove it from prio tree | |
1339 | * as it is empty. | |
1340 | */ | |
1341 | if (cfqq->p_root) { | |
1342 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1343 | cfqq->p_root = NULL; | |
1344 | } | |
1345 | } | |
1da177e4 LT |
1346 | } |
1347 | ||
5e705374 | 1348 | static void cfq_add_rq_rb(struct request *rq) |
1da177e4 | 1349 | { |
5e705374 | 1350 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 1351 | struct cfq_data *cfqd = cfqq->cfqd; |
a36e71f9 | 1352 | struct request *__alias, *prev; |
1da177e4 | 1353 | |
5380a101 | 1354 | cfqq->queued[rq_is_sync(rq)]++; |
1da177e4 LT |
1355 | |
1356 | /* | |
1357 | * looks a little odd, but the first insert might return an alias. | |
1358 | * if that happens, put the alias on the dispatch list | |
1359 | */ | |
21183b07 | 1360 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) |
5e705374 | 1361 | cfq_dispatch_insert(cfqd->queue, __alias); |
5fccbf61 JA |
1362 | |
1363 | if (!cfq_cfqq_on_rr(cfqq)) | |
1364 | cfq_add_cfqq_rr(cfqd, cfqq); | |
5044eed4 JA |
1365 | |
1366 | /* | |
1367 | * check if this request is a better next-serve candidate | |
1368 | */ | |
a36e71f9 | 1369 | prev = cfqq->next_rq; |
cf7c25cf | 1370 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); |
a36e71f9 JA |
1371 | |
1372 | /* | |
1373 | * adjust priority tree position, if ->next_rq changes | |
1374 | */ | |
1375 | if (prev != cfqq->next_rq) | |
1376 | cfq_prio_tree_add(cfqd, cfqq); | |
1377 | ||
5044eed4 | 1378 | BUG_ON(!cfqq->next_rq); |
1da177e4 LT |
1379 | } |
1380 | ||
febffd61 | 1381 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) |
1da177e4 | 1382 | { |
5380a101 JA |
1383 | elv_rb_del(&cfqq->sort_list, rq); |
1384 | cfqq->queued[rq_is_sync(rq)]--; | |
5e705374 | 1385 | cfq_add_rq_rb(rq); |
1da177e4 LT |
1386 | } |
1387 | ||
206dc69b JA |
1388 | static struct request * |
1389 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1da177e4 | 1390 | { |
206dc69b | 1391 | struct task_struct *tsk = current; |
91fac317 | 1392 | struct cfq_io_context *cic; |
206dc69b | 1393 | struct cfq_queue *cfqq; |
1da177e4 | 1394 | |
4ac845a2 | 1395 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
1396 | if (!cic) |
1397 | return NULL; | |
1398 | ||
1399 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
89850f7e JA |
1400 | if (cfqq) { |
1401 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
1402 | ||
21183b07 | 1403 | return elv_rb_find(&cfqq->sort_list, sector); |
89850f7e | 1404 | } |
1da177e4 | 1405 | |
1da177e4 LT |
1406 | return NULL; |
1407 | } | |
1408 | ||
165125e1 | 1409 | static void cfq_activate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1410 | { |
22e2c507 | 1411 | struct cfq_data *cfqd = q->elevator->elevator_data; |
3b18152c | 1412 | |
53c583d2 | 1413 | cfqd->rq_in_driver++; |
7b679138 | 1414 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", |
53c583d2 | 1415 | cfqd->rq_in_driver); |
25776e35 | 1416 | |
5b93629b | 1417 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1da177e4 LT |
1418 | } |
1419 | ||
165125e1 | 1420 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) |
1da177e4 | 1421 | { |
b4878f24 JA |
1422 | struct cfq_data *cfqd = q->elevator->elevator_data; |
1423 | ||
53c583d2 CZ |
1424 | WARN_ON(!cfqd->rq_in_driver); |
1425 | cfqd->rq_in_driver--; | |
7b679138 | 1426 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", |
53c583d2 | 1427 | cfqd->rq_in_driver); |
1da177e4 LT |
1428 | } |
1429 | ||
b4878f24 | 1430 | static void cfq_remove_request(struct request *rq) |
1da177e4 | 1431 | { |
5e705374 | 1432 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
21183b07 | 1433 | |
5e705374 JA |
1434 | if (cfqq->next_rq == rq) |
1435 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1da177e4 | 1436 | |
b4878f24 | 1437 | list_del_init(&rq->queuelist); |
5e705374 | 1438 | cfq_del_rq_rb(rq); |
374f84ac | 1439 | |
45333d5a | 1440 | cfqq->cfqd->rq_queued--; |
374f84ac JA |
1441 | if (rq_is_meta(rq)) { |
1442 | WARN_ON(!cfqq->meta_pending); | |
1443 | cfqq->meta_pending--; | |
1444 | } | |
1da177e4 LT |
1445 | } |
1446 | ||
165125e1 JA |
1447 | static int cfq_merge(struct request_queue *q, struct request **req, |
1448 | struct bio *bio) | |
1da177e4 LT |
1449 | { |
1450 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1451 | struct request *__rq; | |
1da177e4 | 1452 | |
206dc69b | 1453 | __rq = cfq_find_rq_fmerge(cfqd, bio); |
22e2c507 | 1454 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
9817064b JA |
1455 | *req = __rq; |
1456 | return ELEVATOR_FRONT_MERGE; | |
1da177e4 LT |
1457 | } |
1458 | ||
1459 | return ELEVATOR_NO_MERGE; | |
1da177e4 LT |
1460 | } |
1461 | ||
165125e1 | 1462 | static void cfq_merged_request(struct request_queue *q, struct request *req, |
21183b07 | 1463 | int type) |
1da177e4 | 1464 | { |
21183b07 | 1465 | if (type == ELEVATOR_FRONT_MERGE) { |
5e705374 | 1466 | struct cfq_queue *cfqq = RQ_CFQQ(req); |
1da177e4 | 1467 | |
5e705374 | 1468 | cfq_reposition_rq_rb(cfqq, req); |
1da177e4 | 1469 | } |
1da177e4 LT |
1470 | } |
1471 | ||
1472 | static void | |
165125e1 | 1473 | cfq_merged_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
1474 | struct request *next) |
1475 | { | |
cf7c25cf | 1476 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 JA |
1477 | /* |
1478 | * reposition in fifo if next is older than rq | |
1479 | */ | |
1480 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
30996f40 | 1481 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
22e2c507 | 1482 | list_move(&rq->queuelist, &next->queuelist); |
30996f40 JA |
1483 | rq_set_fifo_time(rq, rq_fifo_time(next)); |
1484 | } | |
22e2c507 | 1485 | |
cf7c25cf CZ |
1486 | if (cfqq->next_rq == next) |
1487 | cfqq->next_rq = rq; | |
b4878f24 | 1488 | cfq_remove_request(next); |
22e2c507 JA |
1489 | } |
1490 | ||
165125e1 | 1491 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, |
da775265 JA |
1492 | struct bio *bio) |
1493 | { | |
1494 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
91fac317 | 1495 | struct cfq_io_context *cic; |
da775265 | 1496 | struct cfq_queue *cfqq; |
da775265 JA |
1497 | |
1498 | /* | |
ec8acb69 | 1499 | * Disallow merge of a sync bio into an async request. |
da775265 | 1500 | */ |
91fac317 | 1501 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) |
a6151c3a | 1502 | return false; |
da775265 JA |
1503 | |
1504 | /* | |
719d3402 JA |
1505 | * Lookup the cfqq that this bio will be queued with. Allow |
1506 | * merge only if rq is queued there. | |
da775265 | 1507 | */ |
4ac845a2 | 1508 | cic = cfq_cic_lookup(cfqd, current->io_context); |
91fac317 | 1509 | if (!cic) |
a6151c3a | 1510 | return false; |
719d3402 | 1511 | |
91fac317 | 1512 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); |
a6151c3a | 1513 | return cfqq == RQ_CFQQ(rq); |
da775265 JA |
1514 | } |
1515 | ||
febffd61 JA |
1516 | static void __cfq_set_active_queue(struct cfq_data *cfqd, |
1517 | struct cfq_queue *cfqq) | |
22e2c507 JA |
1518 | { |
1519 | if (cfqq) { | |
7b679138 | 1520 | cfq_log_cfqq(cfqd, cfqq, "set_active"); |
dae739eb VG |
1521 | cfqq->slice_start = 0; |
1522 | cfqq->dispatch_start = jiffies; | |
f75edf2d | 1523 | cfqq->allocated_slice = 0; |
22e2c507 | 1524 | cfqq->slice_end = 0; |
2f5cb738 | 1525 | cfqq->slice_dispatch = 0; |
22084190 | 1526 | cfqq->nr_sectors = 0; |
2f5cb738 | 1527 | |
2f5cb738 | 1528 | cfq_clear_cfqq_wait_request(cfqq); |
b029195d | 1529 | cfq_clear_cfqq_must_dispatch(cfqq); |
3b18152c JA |
1530 | cfq_clear_cfqq_must_alloc_slice(cfqq); |
1531 | cfq_clear_cfqq_fifo_expire(cfqq); | |
44f7c160 | 1532 | cfq_mark_cfqq_slice_new(cfqq); |
2f5cb738 JA |
1533 | |
1534 | del_timer(&cfqd->idle_slice_timer); | |
22e2c507 JA |
1535 | } |
1536 | ||
1537 | cfqd->active_queue = cfqq; | |
1538 | } | |
1539 | ||
7b14e3b5 JA |
1540 | /* |
1541 | * current cfqq expired its slice (or was too idle), select new one | |
1542 | */ | |
1543 | static void | |
1544 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
a6151c3a | 1545 | bool timed_out) |
7b14e3b5 | 1546 | { |
7b679138 JA |
1547 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); |
1548 | ||
7b14e3b5 JA |
1549 | if (cfq_cfqq_wait_request(cfqq)) |
1550 | del_timer(&cfqd->idle_slice_timer); | |
1551 | ||
7b14e3b5 | 1552 | cfq_clear_cfqq_wait_request(cfqq); |
f75edf2d | 1553 | cfq_clear_cfqq_wait_busy(cfqq); |
7b14e3b5 | 1554 | |
ae54abed SL |
1555 | /* |
1556 | * If this cfqq is shared between multiple processes, check to | |
1557 | * make sure that those processes are still issuing I/Os within | |
1558 | * the mean seek distance. If not, it may be time to break the | |
1559 | * queues apart again. | |
1560 | */ | |
1561 | if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq)) | |
1562 | cfq_mark_cfqq_split_coop(cfqq); | |
1563 | ||
7b14e3b5 | 1564 | /* |
6084cdda | 1565 | * store what was left of this slice, if the queue idled/timed out |
7b14e3b5 | 1566 | */ |
7b679138 | 1567 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) { |
c5b680f3 | 1568 | cfqq->slice_resid = cfqq->slice_end - jiffies; |
7b679138 JA |
1569 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); |
1570 | } | |
7b14e3b5 | 1571 | |
dae739eb VG |
1572 | cfq_group_served(cfqd, cfqq->cfqg, cfqq); |
1573 | ||
f04a6424 VG |
1574 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) |
1575 | cfq_del_cfqq_rr(cfqd, cfqq); | |
1576 | ||
edd75ffd | 1577 | cfq_resort_rr_list(cfqd, cfqq); |
7b14e3b5 JA |
1578 | |
1579 | if (cfqq == cfqd->active_queue) | |
1580 | cfqd->active_queue = NULL; | |
1581 | ||
dae739eb VG |
1582 | if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active) |
1583 | cfqd->grp_service_tree.active = NULL; | |
1584 | ||
7b14e3b5 JA |
1585 | if (cfqd->active_cic) { |
1586 | put_io_context(cfqd->active_cic->ioc); | |
1587 | cfqd->active_cic = NULL; | |
1588 | } | |
7b14e3b5 JA |
1589 | } |
1590 | ||
a6151c3a | 1591 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) |
7b14e3b5 JA |
1592 | { |
1593 | struct cfq_queue *cfqq = cfqd->active_queue; | |
1594 | ||
1595 | if (cfqq) | |
6084cdda | 1596 | __cfq_slice_expired(cfqd, cfqq, timed_out); |
7b14e3b5 JA |
1597 | } |
1598 | ||
498d3aa2 JA |
1599 | /* |
1600 | * Get next queue for service. Unless we have a queue preemption, | |
1601 | * we'll simply select the first cfqq in the service tree. | |
1602 | */ | |
6d048f53 | 1603 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) |
22e2c507 | 1604 | { |
c0324a02 | 1605 | struct cfq_rb_root *service_tree = |
cdb16e8f | 1606 | service_tree_for(cfqd->serving_group, cfqd->serving_prio, |
65b32a57 | 1607 | cfqd->serving_type); |
d9e7620e | 1608 | |
f04a6424 VG |
1609 | if (!cfqd->rq_queued) |
1610 | return NULL; | |
1611 | ||
1fa8f6d6 VG |
1612 | /* There is nothing to dispatch */ |
1613 | if (!service_tree) | |
1614 | return NULL; | |
c0324a02 CZ |
1615 | if (RB_EMPTY_ROOT(&service_tree->rb)) |
1616 | return NULL; | |
1617 | return cfq_rb_first(service_tree); | |
6d048f53 JA |
1618 | } |
1619 | ||
f04a6424 VG |
1620 | static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) |
1621 | { | |
25fb5169 | 1622 | struct cfq_group *cfqg; |
f04a6424 VG |
1623 | struct cfq_queue *cfqq; |
1624 | int i, j; | |
1625 | struct cfq_rb_root *st; | |
1626 | ||
1627 | if (!cfqd->rq_queued) | |
1628 | return NULL; | |
1629 | ||
25fb5169 VG |
1630 | cfqg = cfq_get_next_cfqg(cfqd); |
1631 | if (!cfqg) | |
1632 | return NULL; | |
1633 | ||
f04a6424 VG |
1634 | for_each_cfqg_st(cfqg, i, j, st) |
1635 | if ((cfqq = cfq_rb_first(st)) != NULL) | |
1636 | return cfqq; | |
1637 | return NULL; | |
1638 | } | |
1639 | ||
498d3aa2 JA |
1640 | /* |
1641 | * Get and set a new active queue for service. | |
1642 | */ | |
a36e71f9 JA |
1643 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, |
1644 | struct cfq_queue *cfqq) | |
6d048f53 | 1645 | { |
e00ef799 | 1646 | if (!cfqq) |
a36e71f9 | 1647 | cfqq = cfq_get_next_queue(cfqd); |
6d048f53 | 1648 | |
22e2c507 | 1649 | __cfq_set_active_queue(cfqd, cfqq); |
3b18152c | 1650 | return cfqq; |
22e2c507 JA |
1651 | } |
1652 | ||
d9e7620e JA |
1653 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, |
1654 | struct request *rq) | |
1655 | { | |
83096ebf TH |
1656 | if (blk_rq_pos(rq) >= cfqd->last_position) |
1657 | return blk_rq_pos(rq) - cfqd->last_position; | |
d9e7620e | 1658 | else |
83096ebf | 1659 | return cfqd->last_position - blk_rq_pos(rq); |
d9e7620e JA |
1660 | } |
1661 | ||
b2c18e1e | 1662 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
2f7a2d89 | 1663 | struct request *rq, bool for_preempt) |
6d048f53 | 1664 | { |
3dde36dd | 1665 | return cfq_dist_from_last(cfqd, rq) <= CFQQ_SEEK_THR; |
6d048f53 JA |
1666 | } |
1667 | ||
a36e71f9 JA |
1668 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, |
1669 | struct cfq_queue *cur_cfqq) | |
1670 | { | |
f2d1f0ae | 1671 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; |
a36e71f9 JA |
1672 | struct rb_node *parent, *node; |
1673 | struct cfq_queue *__cfqq; | |
1674 | sector_t sector = cfqd->last_position; | |
1675 | ||
1676 | if (RB_EMPTY_ROOT(root)) | |
1677 | return NULL; | |
1678 | ||
1679 | /* | |
1680 | * First, if we find a request starting at the end of the last | |
1681 | * request, choose it. | |
1682 | */ | |
f2d1f0ae | 1683 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); |
a36e71f9 JA |
1684 | if (__cfqq) |
1685 | return __cfqq; | |
1686 | ||
1687 | /* | |
1688 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1689 | * will contain the closest sector. | |
1690 | */ | |
1691 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
2f7a2d89 | 1692 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq, false)) |
a36e71f9 JA |
1693 | return __cfqq; |
1694 | ||
2e46e8b2 | 1695 | if (blk_rq_pos(__cfqq->next_rq) < sector) |
a36e71f9 JA |
1696 | node = rb_next(&__cfqq->p_node); |
1697 | else | |
1698 | node = rb_prev(&__cfqq->p_node); | |
1699 | if (!node) | |
1700 | return NULL; | |
1701 | ||
1702 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
2f7a2d89 | 1703 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq, false)) |
a36e71f9 JA |
1704 | return __cfqq; |
1705 | ||
1706 | return NULL; | |
1707 | } | |
1708 | ||
1709 | /* | |
1710 | * cfqd - obvious | |
1711 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1712 | * closely cooperating with itself. | |
1713 | * | |
1714 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1715 | * one request, and that cfqd->last_position reflects a position on the disk | |
1716 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1717 | * assumption. | |
1718 | */ | |
1719 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
b3b6d040 | 1720 | struct cfq_queue *cur_cfqq) |
6d048f53 | 1721 | { |
a36e71f9 JA |
1722 | struct cfq_queue *cfqq; |
1723 | ||
e6c5bc73 JM |
1724 | if (!cfq_cfqq_sync(cur_cfqq)) |
1725 | return NULL; | |
1726 | if (CFQQ_SEEKY(cur_cfqq)) | |
1727 | return NULL; | |
1728 | ||
b9d8f4c7 GJ |
1729 | /* |
1730 | * Don't search priority tree if it's the only queue in the group. | |
1731 | */ | |
1732 | if (cur_cfqq->cfqg->nr_cfqq == 1) | |
1733 | return NULL; | |
1734 | ||
6d048f53 | 1735 | /* |
d9e7620e JA |
1736 | * We should notice if some of the queues are cooperating, eg |
1737 | * working closely on the same area of the disk. In that case, | |
1738 | * we can group them together and don't waste time idling. | |
6d048f53 | 1739 | */ |
a36e71f9 JA |
1740 | cfqq = cfqq_close(cfqd, cur_cfqq); |
1741 | if (!cfqq) | |
1742 | return NULL; | |
1743 | ||
8682e1f1 VG |
1744 | /* If new queue belongs to different cfq_group, don't choose it */ |
1745 | if (cur_cfqq->cfqg != cfqq->cfqg) | |
1746 | return NULL; | |
1747 | ||
df5fe3e8 JM |
1748 | /* |
1749 | * It only makes sense to merge sync queues. | |
1750 | */ | |
1751 | if (!cfq_cfqq_sync(cfqq)) | |
1752 | return NULL; | |
e6c5bc73 JM |
1753 | if (CFQQ_SEEKY(cfqq)) |
1754 | return NULL; | |
df5fe3e8 | 1755 | |
c0324a02 CZ |
1756 | /* |
1757 | * Do not merge queues of different priority classes | |
1758 | */ | |
1759 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
1760 | return NULL; | |
1761 | ||
a36e71f9 | 1762 | return cfqq; |
6d048f53 JA |
1763 | } |
1764 | ||
a6d44e98 CZ |
1765 | /* |
1766 | * Determine whether we should enforce idle window for this queue. | |
1767 | */ | |
1768 | ||
1769 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1770 | { | |
1771 | enum wl_prio_t prio = cfqq_prio(cfqq); | |
718eee05 | 1772 | struct cfq_rb_root *service_tree = cfqq->service_tree; |
a6d44e98 | 1773 | |
f04a6424 VG |
1774 | BUG_ON(!service_tree); |
1775 | BUG_ON(!service_tree->count); | |
1776 | ||
a6d44e98 CZ |
1777 | /* We never do for idle class queues. */ |
1778 | if (prio == IDLE_WORKLOAD) | |
1779 | return false; | |
1780 | ||
1781 | /* We do for queues that were marked with idle window flag. */ | |
3c764b7a SL |
1782 | if (cfq_cfqq_idle_window(cfqq) && |
1783 | !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)) | |
a6d44e98 CZ |
1784 | return true; |
1785 | ||
1786 | /* | |
1787 | * Otherwise, we do only if they are the last ones | |
1788 | * in their service tree. | |
1789 | */ | |
1efe8fe1 | 1790 | return service_tree->count == 1 && cfq_cfqq_sync(cfqq); |
a6d44e98 CZ |
1791 | } |
1792 | ||
6d048f53 | 1793 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) |
22e2c507 | 1794 | { |
1792669c | 1795 | struct cfq_queue *cfqq = cfqd->active_queue; |
206dc69b | 1796 | struct cfq_io_context *cic; |
7b14e3b5 JA |
1797 | unsigned long sl; |
1798 | ||
a68bbddb | 1799 | /* |
f7d7b7a7 JA |
1800 | * SSD device without seek penalty, disable idling. But only do so |
1801 | * for devices that support queuing, otherwise we still have a problem | |
1802 | * with sync vs async workloads. | |
a68bbddb | 1803 | */ |
f7d7b7a7 | 1804 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) |
a68bbddb JA |
1805 | return; |
1806 | ||
dd67d051 | 1807 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); |
6d048f53 | 1808 | WARN_ON(cfq_cfqq_slice_new(cfqq)); |
22e2c507 JA |
1809 | |
1810 | /* | |
1811 | * idle is disabled, either manually or by past process history | |
1812 | */ | |
a6d44e98 | 1813 | if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq)) |
6d048f53 JA |
1814 | return; |
1815 | ||
7b679138 | 1816 | /* |
8e550632 | 1817 | * still active requests from this queue, don't idle |
7b679138 | 1818 | */ |
8e550632 | 1819 | if (cfqq->dispatched) |
7b679138 JA |
1820 | return; |
1821 | ||
22e2c507 JA |
1822 | /* |
1823 | * task has exited, don't wait | |
1824 | */ | |
206dc69b | 1825 | cic = cfqd->active_cic; |
66dac98e | 1826 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) |
6d048f53 JA |
1827 | return; |
1828 | ||
355b659c CZ |
1829 | /* |
1830 | * If our average think time is larger than the remaining time | |
1831 | * slice, then don't idle. This avoids overrunning the allotted | |
1832 | * time slice. | |
1833 | */ | |
1834 | if (sample_valid(cic->ttime_samples) && | |
1835 | (cfqq->slice_end - jiffies < cic->ttime_mean)) | |
1836 | return; | |
1837 | ||
3b18152c | 1838 | cfq_mark_cfqq_wait_request(cfqq); |
22e2c507 | 1839 | |
6d048f53 | 1840 | sl = cfqd->cfq_slice_idle; |
206dc69b | 1841 | |
7b14e3b5 | 1842 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); |
9481ffdc | 1843 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); |
1da177e4 LT |
1844 | } |
1845 | ||
498d3aa2 JA |
1846 | /* |
1847 | * Move request from internal lists to the request queue dispatch list. | |
1848 | */ | |
165125e1 | 1849 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) |
1da177e4 | 1850 | { |
3ed9a296 | 1851 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 1852 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 1853 | |
7b679138 JA |
1854 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); |
1855 | ||
06d21886 | 1856 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); |
5380a101 | 1857 | cfq_remove_request(rq); |
6d048f53 | 1858 | cfqq->dispatched++; |
5380a101 | 1859 | elv_dispatch_sort(q, rq); |
3ed9a296 | 1860 | |
53c583d2 | 1861 | cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++; |
22084190 | 1862 | cfqq->nr_sectors += blk_rq_sectors(rq); |
1da177e4 LT |
1863 | } |
1864 | ||
1865 | /* | |
1866 | * return expired entry, or NULL to just start from scratch in rbtree | |
1867 | */ | |
febffd61 | 1868 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) |
1da177e4 | 1869 | { |
30996f40 | 1870 | struct request *rq = NULL; |
1da177e4 | 1871 | |
3b18152c | 1872 | if (cfq_cfqq_fifo_expire(cfqq)) |
1da177e4 | 1873 | return NULL; |
cb887411 JA |
1874 | |
1875 | cfq_mark_cfqq_fifo_expire(cfqq); | |
1876 | ||
89850f7e JA |
1877 | if (list_empty(&cfqq->fifo)) |
1878 | return NULL; | |
1da177e4 | 1879 | |
89850f7e | 1880 | rq = rq_entry_fifo(cfqq->fifo.next); |
30996f40 | 1881 | if (time_before(jiffies, rq_fifo_time(rq))) |
7b679138 | 1882 | rq = NULL; |
1da177e4 | 1883 | |
30996f40 | 1884 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); |
6d048f53 | 1885 | return rq; |
1da177e4 LT |
1886 | } |
1887 | ||
22e2c507 JA |
1888 | static inline int |
1889 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1890 | { | |
1891 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1da177e4 | 1892 | |
22e2c507 | 1893 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); |
1da177e4 | 1894 | |
22e2c507 | 1895 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); |
1da177e4 LT |
1896 | } |
1897 | ||
df5fe3e8 JM |
1898 | /* |
1899 | * Must be called with the queue_lock held. | |
1900 | */ | |
1901 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
1902 | { | |
1903 | int process_refs, io_refs; | |
1904 | ||
1905 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
1906 | process_refs = atomic_read(&cfqq->ref) - io_refs; | |
1907 | BUG_ON(process_refs < 0); | |
1908 | return process_refs; | |
1909 | } | |
1910 | ||
1911 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
1912 | { | |
e6c5bc73 | 1913 | int process_refs, new_process_refs; |
df5fe3e8 JM |
1914 | struct cfq_queue *__cfqq; |
1915 | ||
1916 | /* Avoid a circular list and skip interim queue merges */ | |
1917 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
1918 | if (__cfqq == cfqq) | |
1919 | return; | |
1920 | new_cfqq = __cfqq; | |
1921 | } | |
1922 | ||
1923 | process_refs = cfqq_process_refs(cfqq); | |
1924 | /* | |
1925 | * If the process for the cfqq has gone away, there is no | |
1926 | * sense in merging the queues. | |
1927 | */ | |
1928 | if (process_refs == 0) | |
1929 | return; | |
1930 | ||
e6c5bc73 JM |
1931 | /* |
1932 | * Merge in the direction of the lesser amount of work. | |
1933 | */ | |
1934 | new_process_refs = cfqq_process_refs(new_cfqq); | |
1935 | if (new_process_refs >= process_refs) { | |
1936 | cfqq->new_cfqq = new_cfqq; | |
1937 | atomic_add(process_refs, &new_cfqq->ref); | |
1938 | } else { | |
1939 | new_cfqq->new_cfqq = cfqq; | |
1940 | atomic_add(new_process_refs, &cfqq->ref); | |
1941 | } | |
df5fe3e8 JM |
1942 | } |
1943 | ||
cdb16e8f | 1944 | static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, |
65b32a57 | 1945 | struct cfq_group *cfqg, enum wl_prio_t prio) |
718eee05 CZ |
1946 | { |
1947 | struct cfq_queue *queue; | |
1948 | int i; | |
1949 | bool key_valid = false; | |
1950 | unsigned long lowest_key = 0; | |
1951 | enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; | |
1952 | ||
65b32a57 VG |
1953 | for (i = 0; i <= SYNC_WORKLOAD; ++i) { |
1954 | /* select the one with lowest rb_key */ | |
1955 | queue = cfq_rb_first(service_tree_for(cfqg, prio, i)); | |
718eee05 CZ |
1956 | if (queue && |
1957 | (!key_valid || time_before(queue->rb_key, lowest_key))) { | |
1958 | lowest_key = queue->rb_key; | |
1959 | cur_best = i; | |
1960 | key_valid = true; | |
1961 | } | |
1962 | } | |
1963 | ||
1964 | return cur_best; | |
1965 | } | |
1966 | ||
cdb16e8f | 1967 | static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) |
718eee05 | 1968 | { |
718eee05 CZ |
1969 | unsigned slice; |
1970 | unsigned count; | |
cdb16e8f | 1971 | struct cfq_rb_root *st; |
58ff82f3 | 1972 | unsigned group_slice; |
718eee05 | 1973 | |
1fa8f6d6 VG |
1974 | if (!cfqg) { |
1975 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1976 | cfqd->workload_expires = jiffies + 1; | |
1977 | return; | |
1978 | } | |
1979 | ||
718eee05 | 1980 | /* Choose next priority. RT > BE > IDLE */ |
58ff82f3 | 1981 | if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) |
718eee05 | 1982 | cfqd->serving_prio = RT_WORKLOAD; |
58ff82f3 | 1983 | else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) |
718eee05 CZ |
1984 | cfqd->serving_prio = BE_WORKLOAD; |
1985 | else { | |
1986 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1987 | cfqd->workload_expires = jiffies + 1; | |
1988 | return; | |
1989 | } | |
1990 | ||
1991 | /* | |
1992 | * For RT and BE, we have to choose also the type | |
1993 | * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload | |
1994 | * expiration time | |
1995 | */ | |
65b32a57 | 1996 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); |
cdb16e8f | 1997 | count = st->count; |
718eee05 CZ |
1998 | |
1999 | /* | |
65b32a57 | 2000 | * check workload expiration, and that we still have other queues ready |
718eee05 | 2001 | */ |
65b32a57 | 2002 | if (count && !time_after(jiffies, cfqd->workload_expires)) |
718eee05 CZ |
2003 | return; |
2004 | ||
2005 | /* otherwise select new workload type */ | |
2006 | cfqd->serving_type = | |
65b32a57 VG |
2007 | cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio); |
2008 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); | |
cdb16e8f | 2009 | count = st->count; |
718eee05 CZ |
2010 | |
2011 | /* | |
2012 | * the workload slice is computed as a fraction of target latency | |
2013 | * proportional to the number of queues in that workload, over | |
2014 | * all the queues in the same priority class | |
2015 | */ | |
58ff82f3 VG |
2016 | group_slice = cfq_group_slice(cfqd, cfqg); |
2017 | ||
2018 | slice = group_slice * count / | |
2019 | max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], | |
2020 | cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); | |
718eee05 | 2021 | |
f26bd1f0 VG |
2022 | if (cfqd->serving_type == ASYNC_WORKLOAD) { |
2023 | unsigned int tmp; | |
2024 | ||
2025 | /* | |
2026 | * Async queues are currently system wide. Just taking | |
2027 | * proportion of queues with-in same group will lead to higher | |
2028 | * async ratio system wide as generally root group is going | |
2029 | * to have higher weight. A more accurate thing would be to | |
2030 | * calculate system wide asnc/sync ratio. | |
2031 | */ | |
2032 | tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg); | |
2033 | tmp = tmp/cfqd->busy_queues; | |
2034 | slice = min_t(unsigned, slice, tmp); | |
2035 | ||
718eee05 CZ |
2036 | /* async workload slice is scaled down according to |
2037 | * the sync/async slice ratio. */ | |
2038 | slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; | |
f26bd1f0 | 2039 | } else |
718eee05 CZ |
2040 | /* sync workload slice is at least 2 * cfq_slice_idle */ |
2041 | slice = max(slice, 2 * cfqd->cfq_slice_idle); | |
2042 | ||
2043 | slice = max_t(unsigned, slice, CFQ_MIN_TT); | |
2044 | cfqd->workload_expires = jiffies + slice; | |
8e550632 | 2045 | cfqd->noidle_tree_requires_idle = false; |
718eee05 CZ |
2046 | } |
2047 | ||
1fa8f6d6 VG |
2048 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) |
2049 | { | |
2050 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
25bc6b07 | 2051 | struct cfq_group *cfqg; |
1fa8f6d6 VG |
2052 | |
2053 | if (RB_EMPTY_ROOT(&st->rb)) | |
2054 | return NULL; | |
25bc6b07 VG |
2055 | cfqg = cfq_rb_first_group(st); |
2056 | st->active = &cfqg->rb_node; | |
2057 | update_min_vdisktime(st); | |
2058 | return cfqg; | |
1fa8f6d6 VG |
2059 | } |
2060 | ||
cdb16e8f VG |
2061 | static void cfq_choose_cfqg(struct cfq_data *cfqd) |
2062 | { | |
1fa8f6d6 VG |
2063 | struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); |
2064 | ||
2065 | cfqd->serving_group = cfqg; | |
dae739eb VG |
2066 | |
2067 | /* Restore the workload type data */ | |
2068 | if (cfqg->saved_workload_slice) { | |
2069 | cfqd->workload_expires = jiffies + cfqg->saved_workload_slice; | |
2070 | cfqd->serving_type = cfqg->saved_workload; | |
2071 | cfqd->serving_prio = cfqg->saved_serving_prio; | |
66ae2919 GJ |
2072 | } else |
2073 | cfqd->workload_expires = jiffies - 1; | |
2074 | ||
1fa8f6d6 | 2075 | choose_service_tree(cfqd, cfqg); |
cdb16e8f VG |
2076 | } |
2077 | ||
22e2c507 | 2078 | /* |
498d3aa2 JA |
2079 | * Select a queue for service. If we have a current active queue, |
2080 | * check whether to continue servicing it, or retrieve and set a new one. | |
22e2c507 | 2081 | */ |
1b5ed5e1 | 2082 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) |
1da177e4 | 2083 | { |
a36e71f9 | 2084 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
1da177e4 | 2085 | |
22e2c507 JA |
2086 | cfqq = cfqd->active_queue; |
2087 | if (!cfqq) | |
2088 | goto new_queue; | |
1da177e4 | 2089 | |
f04a6424 VG |
2090 | if (!cfqd->rq_queued) |
2091 | return NULL; | |
c244bb50 VG |
2092 | |
2093 | /* | |
2094 | * We were waiting for group to get backlogged. Expire the queue | |
2095 | */ | |
2096 | if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list)) | |
2097 | goto expire; | |
2098 | ||
22e2c507 | 2099 | /* |
6d048f53 | 2100 | * The active queue has run out of time, expire it and select new. |
22e2c507 | 2101 | */ |
7667aa06 VG |
2102 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) { |
2103 | /* | |
2104 | * If slice had not expired at the completion of last request | |
2105 | * we might not have turned on wait_busy flag. Don't expire | |
2106 | * the queue yet. Allow the group to get backlogged. | |
2107 | * | |
2108 | * The very fact that we have used the slice, that means we | |
2109 | * have been idling all along on this queue and it should be | |
2110 | * ok to wait for this request to complete. | |
2111 | */ | |
82bbbf28 VG |
2112 | if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list) |
2113 | && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { | |
2114 | cfqq = NULL; | |
7667aa06 | 2115 | goto keep_queue; |
82bbbf28 | 2116 | } else |
7667aa06 VG |
2117 | goto expire; |
2118 | } | |
1da177e4 | 2119 | |
22e2c507 | 2120 | /* |
6d048f53 JA |
2121 | * The active queue has requests and isn't expired, allow it to |
2122 | * dispatch. | |
22e2c507 | 2123 | */ |
dd67d051 | 2124 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 2125 | goto keep_queue; |
6d048f53 | 2126 | |
a36e71f9 JA |
2127 | /* |
2128 | * If another queue has a request waiting within our mean seek | |
2129 | * distance, let it run. The expire code will check for close | |
2130 | * cooperators and put the close queue at the front of the service | |
df5fe3e8 | 2131 | * tree. If possible, merge the expiring queue with the new cfqq. |
a36e71f9 | 2132 | */ |
b3b6d040 | 2133 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); |
df5fe3e8 JM |
2134 | if (new_cfqq) { |
2135 | if (!cfqq->new_cfqq) | |
2136 | cfq_setup_merge(cfqq, new_cfqq); | |
a36e71f9 | 2137 | goto expire; |
df5fe3e8 | 2138 | } |
a36e71f9 | 2139 | |
6d048f53 JA |
2140 | /* |
2141 | * No requests pending. If the active queue still has requests in | |
2142 | * flight or is idling for a new request, allow either of these | |
2143 | * conditions to happen (or time out) before selecting a new queue. | |
2144 | */ | |
cc197479 | 2145 | if (timer_pending(&cfqd->idle_slice_timer) || |
a6d44e98 | 2146 | (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) { |
caaa5f9f JA |
2147 | cfqq = NULL; |
2148 | goto keep_queue; | |
22e2c507 JA |
2149 | } |
2150 | ||
3b18152c | 2151 | expire: |
6084cdda | 2152 | cfq_slice_expired(cfqd, 0); |
3b18152c | 2153 | new_queue: |
718eee05 CZ |
2154 | /* |
2155 | * Current queue expired. Check if we have to switch to a new | |
2156 | * service tree | |
2157 | */ | |
2158 | if (!new_cfqq) | |
cdb16e8f | 2159 | cfq_choose_cfqg(cfqd); |
718eee05 | 2160 | |
a36e71f9 | 2161 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); |
22e2c507 | 2162 | keep_queue: |
3b18152c | 2163 | return cfqq; |
22e2c507 JA |
2164 | } |
2165 | ||
febffd61 | 2166 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) |
d9e7620e JA |
2167 | { |
2168 | int dispatched = 0; | |
2169 | ||
2170 | while (cfqq->next_rq) { | |
2171 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
2172 | dispatched++; | |
2173 | } | |
2174 | ||
2175 | BUG_ON(!list_empty(&cfqq->fifo)); | |
f04a6424 VG |
2176 | |
2177 | /* By default cfqq is not expired if it is empty. Do it explicitly */ | |
2178 | __cfq_slice_expired(cfqq->cfqd, cfqq, 0); | |
d9e7620e JA |
2179 | return dispatched; |
2180 | } | |
2181 | ||
498d3aa2 JA |
2182 | /* |
2183 | * Drain our current requests. Used for barriers and when switching | |
2184 | * io schedulers on-the-fly. | |
2185 | */ | |
d9e7620e | 2186 | static int cfq_forced_dispatch(struct cfq_data *cfqd) |
1b5ed5e1 | 2187 | { |
0871714e | 2188 | struct cfq_queue *cfqq; |
d9e7620e | 2189 | int dispatched = 0; |
cdb16e8f | 2190 | |
f04a6424 VG |
2191 | while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) |
2192 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); | |
1b5ed5e1 | 2193 | |
6084cdda | 2194 | cfq_slice_expired(cfqd, 0); |
1b5ed5e1 TH |
2195 | BUG_ON(cfqd->busy_queues); |
2196 | ||
6923715a | 2197 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); |
1b5ed5e1 TH |
2198 | return dispatched; |
2199 | } | |
2200 | ||
abc3c744 SL |
2201 | static inline bool cfq_slice_used_soon(struct cfq_data *cfqd, |
2202 | struct cfq_queue *cfqq) | |
2203 | { | |
2204 | /* the queue hasn't finished any request, can't estimate */ | |
2205 | if (cfq_cfqq_slice_new(cfqq)) | |
2206 | return 1; | |
2207 | if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched, | |
2208 | cfqq->slice_end)) | |
2209 | return 1; | |
2210 | ||
2211 | return 0; | |
2212 | } | |
2213 | ||
0b182d61 | 2214 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
2f5cb738 | 2215 | { |
2f5cb738 | 2216 | unsigned int max_dispatch; |
22e2c507 | 2217 | |
5ad531db JA |
2218 | /* |
2219 | * Drain async requests before we start sync IO | |
2220 | */ | |
53c583d2 | 2221 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC]) |
0b182d61 | 2222 | return false; |
5ad531db | 2223 | |
2f5cb738 JA |
2224 | /* |
2225 | * If this is an async queue and we have sync IO in flight, let it wait | |
2226 | */ | |
53c583d2 | 2227 | if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq)) |
0b182d61 | 2228 | return false; |
2f5cb738 | 2229 | |
abc3c744 | 2230 | max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1); |
2f5cb738 JA |
2231 | if (cfq_class_idle(cfqq)) |
2232 | max_dispatch = 1; | |
b4878f24 | 2233 | |
2f5cb738 JA |
2234 | /* |
2235 | * Does this cfqq already have too much IO in flight? | |
2236 | */ | |
2237 | if (cfqq->dispatched >= max_dispatch) { | |
2238 | /* | |
2239 | * idle queue must always only have a single IO in flight | |
2240 | */ | |
3ed9a296 | 2241 | if (cfq_class_idle(cfqq)) |
0b182d61 | 2242 | return false; |
3ed9a296 | 2243 | |
2f5cb738 JA |
2244 | /* |
2245 | * We have other queues, don't allow more IO from this one | |
2246 | */ | |
abc3c744 | 2247 | if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq)) |
0b182d61 | 2248 | return false; |
9ede209e | 2249 | |
365722bb | 2250 | /* |
474b18cc | 2251 | * Sole queue user, no limit |
365722bb | 2252 | */ |
abc3c744 SL |
2253 | if (cfqd->busy_queues == 1) |
2254 | max_dispatch = -1; | |
2255 | else | |
2256 | /* | |
2257 | * Normally we start throttling cfqq when cfq_quantum/2 | |
2258 | * requests have been dispatched. But we can drive | |
2259 | * deeper queue depths at the beginning of slice | |
2260 | * subjected to upper limit of cfq_quantum. | |
2261 | * */ | |
2262 | max_dispatch = cfqd->cfq_quantum; | |
8e296755 JA |
2263 | } |
2264 | ||
2265 | /* | |
2266 | * Async queues must wait a bit before being allowed dispatch. | |
2267 | * We also ramp up the dispatch depth gradually for async IO, | |
2268 | * based on the last sync IO we serviced | |
2269 | */ | |
963b72fc | 2270 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { |
573412b2 | 2271 | unsigned long last_sync = jiffies - cfqd->last_delayed_sync; |
8e296755 | 2272 | unsigned int depth; |
365722bb | 2273 | |
61f0c1dc | 2274 | depth = last_sync / cfqd->cfq_slice[1]; |
e00c54c3 JA |
2275 | if (!depth && !cfqq->dispatched) |
2276 | depth = 1; | |
8e296755 JA |
2277 | if (depth < max_dispatch) |
2278 | max_dispatch = depth; | |
2f5cb738 | 2279 | } |
3ed9a296 | 2280 | |
0b182d61 JA |
2281 | /* |
2282 | * If we're below the current max, allow a dispatch | |
2283 | */ | |
2284 | return cfqq->dispatched < max_dispatch; | |
2285 | } | |
2286 | ||
2287 | /* | |
2288 | * Dispatch a request from cfqq, moving them to the request queue | |
2289 | * dispatch list. | |
2290 | */ | |
2291 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2292 | { | |
2293 | struct request *rq; | |
2294 | ||
2295 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
2296 | ||
2297 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
2298 | return false; | |
2299 | ||
2300 | /* | |
2301 | * follow expired path, else get first next available | |
2302 | */ | |
2303 | rq = cfq_check_fifo(cfqq); | |
2304 | if (!rq) | |
2305 | rq = cfqq->next_rq; | |
2306 | ||
2307 | /* | |
2308 | * insert request into driver dispatch list | |
2309 | */ | |
2310 | cfq_dispatch_insert(cfqd->queue, rq); | |
2311 | ||
2312 | if (!cfqd->active_cic) { | |
2313 | struct cfq_io_context *cic = RQ_CIC(rq); | |
2314 | ||
2315 | atomic_long_inc(&cic->ioc->refcount); | |
2316 | cfqd->active_cic = cic; | |
2317 | } | |
2318 | ||
2319 | return true; | |
2320 | } | |
2321 | ||
2322 | /* | |
2323 | * Find the cfqq that we need to service and move a request from that to the | |
2324 | * dispatch list | |
2325 | */ | |
2326 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
2327 | { | |
2328 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2329 | struct cfq_queue *cfqq; | |
2330 | ||
2331 | if (!cfqd->busy_queues) | |
2332 | return 0; | |
2333 | ||
2334 | if (unlikely(force)) | |
2335 | return cfq_forced_dispatch(cfqd); | |
2336 | ||
2337 | cfqq = cfq_select_queue(cfqd); | |
2338 | if (!cfqq) | |
8e296755 JA |
2339 | return 0; |
2340 | ||
2f5cb738 | 2341 | /* |
0b182d61 | 2342 | * Dispatch a request from this cfqq, if it is allowed |
2f5cb738 | 2343 | */ |
0b182d61 JA |
2344 | if (!cfq_dispatch_request(cfqd, cfqq)) |
2345 | return 0; | |
2346 | ||
2f5cb738 | 2347 | cfqq->slice_dispatch++; |
b029195d | 2348 | cfq_clear_cfqq_must_dispatch(cfqq); |
22e2c507 | 2349 | |
2f5cb738 JA |
2350 | /* |
2351 | * expire an async queue immediately if it has used up its slice. idle | |
2352 | * queue always expire after 1 dispatch round. | |
2353 | */ | |
2354 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
2355 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
2356 | cfq_class_idle(cfqq))) { | |
2357 | cfqq->slice_end = jiffies + 1; | |
2358 | cfq_slice_expired(cfqd, 0); | |
1da177e4 LT |
2359 | } |
2360 | ||
b217a903 | 2361 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); |
2f5cb738 | 2362 | return 1; |
1da177e4 LT |
2363 | } |
2364 | ||
1da177e4 | 2365 | /* |
5e705374 JA |
2366 | * task holds one reference to the queue, dropped when task exits. each rq |
2367 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
1da177e4 | 2368 | * |
b1c35769 | 2369 | * Each cfq queue took a reference on the parent group. Drop it now. |
1da177e4 LT |
2370 | * queue lock must be held here. |
2371 | */ | |
2372 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
2373 | { | |
22e2c507 | 2374 | struct cfq_data *cfqd = cfqq->cfqd; |
878eaddd | 2375 | struct cfq_group *cfqg, *orig_cfqg; |
22e2c507 JA |
2376 | |
2377 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
1da177e4 LT |
2378 | |
2379 | if (!atomic_dec_and_test(&cfqq->ref)) | |
2380 | return; | |
2381 | ||
7b679138 | 2382 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); |
1da177e4 | 2383 | BUG_ON(rb_first(&cfqq->sort_list)); |
22e2c507 | 2384 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); |
b1c35769 | 2385 | cfqg = cfqq->cfqg; |
878eaddd | 2386 | orig_cfqg = cfqq->orig_cfqg; |
1da177e4 | 2387 | |
28f95cbc | 2388 | if (unlikely(cfqd->active_queue == cfqq)) { |
6084cdda | 2389 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 2390 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 2391 | } |
22e2c507 | 2392 | |
f04a6424 | 2393 | BUG_ON(cfq_cfqq_on_rr(cfqq)); |
1da177e4 | 2394 | kmem_cache_free(cfq_pool, cfqq); |
b1c35769 | 2395 | cfq_put_cfqg(cfqg); |
878eaddd VG |
2396 | if (orig_cfqg) |
2397 | cfq_put_cfqg(orig_cfqg); | |
1da177e4 LT |
2398 | } |
2399 | ||
d6de8be7 JA |
2400 | /* |
2401 | * Must always be called with the rcu_read_lock() held | |
2402 | */ | |
07416d29 JA |
2403 | static void |
2404 | __call_for_each_cic(struct io_context *ioc, | |
2405 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
2406 | { | |
2407 | struct cfq_io_context *cic; | |
2408 | struct hlist_node *n; | |
2409 | ||
2410 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) | |
2411 | func(ioc, cic); | |
2412 | } | |
2413 | ||
4ac845a2 | 2414 | /* |
34e6bbf2 | 2415 | * Call func for each cic attached to this ioc. |
4ac845a2 | 2416 | */ |
34e6bbf2 | 2417 | static void |
4ac845a2 JA |
2418 | call_for_each_cic(struct io_context *ioc, |
2419 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
1da177e4 | 2420 | { |
4ac845a2 | 2421 | rcu_read_lock(); |
07416d29 | 2422 | __call_for_each_cic(ioc, func); |
4ac845a2 | 2423 | rcu_read_unlock(); |
34e6bbf2 FC |
2424 | } |
2425 | ||
2426 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
2427 | { | |
2428 | struct cfq_io_context *cic; | |
2429 | ||
2430 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
2431 | ||
2432 | kmem_cache_free(cfq_ioc_pool, cic); | |
245b2e70 | 2433 | elv_ioc_count_dec(cfq_ioc_count); |
34e6bbf2 | 2434 | |
9a11b4ed JA |
2435 | if (ioc_gone) { |
2436 | /* | |
2437 | * CFQ scheduler is exiting, grab exit lock and check | |
2438 | * the pending io context count. If it hits zero, | |
2439 | * complete ioc_gone and set it back to NULL | |
2440 | */ | |
2441 | spin_lock(&ioc_gone_lock); | |
245b2e70 | 2442 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { |
9a11b4ed JA |
2443 | complete(ioc_gone); |
2444 | ioc_gone = NULL; | |
2445 | } | |
2446 | spin_unlock(&ioc_gone_lock); | |
2447 | } | |
34e6bbf2 | 2448 | } |
4ac845a2 | 2449 | |
34e6bbf2 FC |
2450 | static void cfq_cic_free(struct cfq_io_context *cic) |
2451 | { | |
2452 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
4ac845a2 JA |
2453 | } |
2454 | ||
2455 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
2456 | { | |
2457 | unsigned long flags; | |
2458 | ||
2459 | BUG_ON(!cic->dead_key); | |
2460 | ||
2461 | spin_lock_irqsave(&ioc->lock, flags); | |
2462 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
ffc4e759 | 2463 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2464 | spin_unlock_irqrestore(&ioc->lock, flags); |
2465 | ||
34e6bbf2 | 2466 | cfq_cic_free(cic); |
4ac845a2 JA |
2467 | } |
2468 | ||
d6de8be7 JA |
2469 | /* |
2470 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
2471 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
2472 | * and ->trim() which is called with the task lock held | |
2473 | */ | |
4ac845a2 JA |
2474 | static void cfq_free_io_context(struct io_context *ioc) |
2475 | { | |
4ac845a2 | 2476 | /* |
34e6bbf2 FC |
2477 | * ioc->refcount is zero here, or we are called from elv_unregister(), |
2478 | * so no more cic's are allowed to be linked into this ioc. So it | |
2479 | * should be ok to iterate over the known list, we will see all cic's | |
2480 | * since no new ones are added. | |
4ac845a2 | 2481 | */ |
07416d29 | 2482 | __call_for_each_cic(ioc, cic_free_func); |
1da177e4 LT |
2483 | } |
2484 | ||
89850f7e | 2485 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
1da177e4 | 2486 | { |
df5fe3e8 JM |
2487 | struct cfq_queue *__cfqq, *next; |
2488 | ||
28f95cbc | 2489 | if (unlikely(cfqq == cfqd->active_queue)) { |
6084cdda | 2490 | __cfq_slice_expired(cfqd, cfqq, 0); |
23e018a1 | 2491 | cfq_schedule_dispatch(cfqd); |
28f95cbc | 2492 | } |
22e2c507 | 2493 | |
df5fe3e8 JM |
2494 | /* |
2495 | * If this queue was scheduled to merge with another queue, be | |
2496 | * sure to drop the reference taken on that queue (and others in | |
2497 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
2498 | */ | |
2499 | __cfqq = cfqq->new_cfqq; | |
2500 | while (__cfqq) { | |
2501 | if (__cfqq == cfqq) { | |
2502 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
2503 | break; | |
2504 | } | |
2505 | next = __cfqq->new_cfqq; | |
2506 | cfq_put_queue(__cfqq); | |
2507 | __cfqq = next; | |
2508 | } | |
2509 | ||
89850f7e JA |
2510 | cfq_put_queue(cfqq); |
2511 | } | |
22e2c507 | 2512 | |
89850f7e JA |
2513 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, |
2514 | struct cfq_io_context *cic) | |
2515 | { | |
4faa3c81 FC |
2516 | struct io_context *ioc = cic->ioc; |
2517 | ||
fc46379d | 2518 | list_del_init(&cic->queue_list); |
4ac845a2 JA |
2519 | |
2520 | /* | |
2521 | * Make sure key == NULL is seen for dead queues | |
2522 | */ | |
fc46379d | 2523 | smp_wmb(); |
4ac845a2 | 2524 | cic->dead_key = (unsigned long) cic->key; |
fc46379d JA |
2525 | cic->key = NULL; |
2526 | ||
4faa3c81 FC |
2527 | if (ioc->ioc_data == cic) |
2528 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
2529 | ||
ff6657c6 JA |
2530 | if (cic->cfqq[BLK_RW_ASYNC]) { |
2531 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
2532 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
12a05732 AV |
2533 | } |
2534 | ||
ff6657c6 JA |
2535 | if (cic->cfqq[BLK_RW_SYNC]) { |
2536 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
2537 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
12a05732 | 2538 | } |
89850f7e JA |
2539 | } |
2540 | ||
4ac845a2 JA |
2541 | static void cfq_exit_single_io_context(struct io_context *ioc, |
2542 | struct cfq_io_context *cic) | |
89850f7e JA |
2543 | { |
2544 | struct cfq_data *cfqd = cic->key; | |
2545 | ||
89850f7e | 2546 | if (cfqd) { |
165125e1 | 2547 | struct request_queue *q = cfqd->queue; |
4ac845a2 | 2548 | unsigned long flags; |
89850f7e | 2549 | |
4ac845a2 | 2550 | spin_lock_irqsave(q->queue_lock, flags); |
62c1fe9d JA |
2551 | |
2552 | /* | |
2553 | * Ensure we get a fresh copy of the ->key to prevent | |
2554 | * race between exiting task and queue | |
2555 | */ | |
2556 | smp_read_barrier_depends(); | |
2557 | if (cic->key) | |
2558 | __cfq_exit_single_io_context(cfqd, cic); | |
2559 | ||
4ac845a2 | 2560 | spin_unlock_irqrestore(q->queue_lock, flags); |
89850f7e | 2561 | } |
1da177e4 LT |
2562 | } |
2563 | ||
498d3aa2 JA |
2564 | /* |
2565 | * The process that ioc belongs to has exited, we need to clean up | |
2566 | * and put the internal structures we have that belongs to that process. | |
2567 | */ | |
e2d74ac0 | 2568 | static void cfq_exit_io_context(struct io_context *ioc) |
1da177e4 | 2569 | { |
4ac845a2 | 2570 | call_for_each_cic(ioc, cfq_exit_single_io_context); |
1da177e4 LT |
2571 | } |
2572 | ||
22e2c507 | 2573 | static struct cfq_io_context * |
8267e268 | 2574 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2575 | { |
b5deef90 | 2576 | struct cfq_io_context *cic; |
1da177e4 | 2577 | |
94f6030c CL |
2578 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, |
2579 | cfqd->queue->node); | |
1da177e4 | 2580 | if (cic) { |
22e2c507 | 2581 | cic->last_end_request = jiffies; |
553698f9 | 2582 | INIT_LIST_HEAD(&cic->queue_list); |
ffc4e759 | 2583 | INIT_HLIST_NODE(&cic->cic_list); |
22e2c507 JA |
2584 | cic->dtor = cfq_free_io_context; |
2585 | cic->exit = cfq_exit_io_context; | |
245b2e70 | 2586 | elv_ioc_count_inc(cfq_ioc_count); |
1da177e4 LT |
2587 | } |
2588 | ||
2589 | return cic; | |
2590 | } | |
2591 | ||
fd0928df | 2592 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) |
22e2c507 JA |
2593 | { |
2594 | struct task_struct *tsk = current; | |
2595 | int ioprio_class; | |
2596 | ||
3b18152c | 2597 | if (!cfq_cfqq_prio_changed(cfqq)) |
22e2c507 JA |
2598 | return; |
2599 | ||
fd0928df | 2600 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); |
22e2c507 | 2601 | switch (ioprio_class) { |
fe094d98 JA |
2602 | default: |
2603 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
2604 | case IOPRIO_CLASS_NONE: | |
2605 | /* | |
6d63c275 | 2606 | * no prio set, inherit CPU scheduling settings |
fe094d98 JA |
2607 | */ |
2608 | cfqq->ioprio = task_nice_ioprio(tsk); | |
6d63c275 | 2609 | cfqq->ioprio_class = task_nice_ioclass(tsk); |
fe094d98 JA |
2610 | break; |
2611 | case IOPRIO_CLASS_RT: | |
2612 | cfqq->ioprio = task_ioprio(ioc); | |
2613 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
2614 | break; | |
2615 | case IOPRIO_CLASS_BE: | |
2616 | cfqq->ioprio = task_ioprio(ioc); | |
2617 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2618 | break; | |
2619 | case IOPRIO_CLASS_IDLE: | |
2620 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
2621 | cfqq->ioprio = 7; | |
2622 | cfq_clear_cfqq_idle_window(cfqq); | |
2623 | break; | |
22e2c507 JA |
2624 | } |
2625 | ||
2626 | /* | |
2627 | * keep track of original prio settings in case we have to temporarily | |
2628 | * elevate the priority of this queue | |
2629 | */ | |
2630 | cfqq->org_ioprio = cfqq->ioprio; | |
2631 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
3b18152c | 2632 | cfq_clear_cfqq_prio_changed(cfqq); |
22e2c507 JA |
2633 | } |
2634 | ||
febffd61 | 2635 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) |
22e2c507 | 2636 | { |
478a82b0 AV |
2637 | struct cfq_data *cfqd = cic->key; |
2638 | struct cfq_queue *cfqq; | |
c1b707d2 | 2639 | unsigned long flags; |
35e6077c | 2640 | |
caaa5f9f JA |
2641 | if (unlikely(!cfqd)) |
2642 | return; | |
2643 | ||
c1b707d2 | 2644 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
caaa5f9f | 2645 | |
ff6657c6 | 2646 | cfqq = cic->cfqq[BLK_RW_ASYNC]; |
caaa5f9f JA |
2647 | if (cfqq) { |
2648 | struct cfq_queue *new_cfqq; | |
ff6657c6 JA |
2649 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, |
2650 | GFP_ATOMIC); | |
caaa5f9f | 2651 | if (new_cfqq) { |
ff6657c6 | 2652 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; |
caaa5f9f JA |
2653 | cfq_put_queue(cfqq); |
2654 | } | |
22e2c507 | 2655 | } |
caaa5f9f | 2656 | |
ff6657c6 | 2657 | cfqq = cic->cfqq[BLK_RW_SYNC]; |
caaa5f9f JA |
2658 | if (cfqq) |
2659 | cfq_mark_cfqq_prio_changed(cfqq); | |
2660 | ||
c1b707d2 | 2661 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); |
22e2c507 JA |
2662 | } |
2663 | ||
fc46379d | 2664 | static void cfq_ioc_set_ioprio(struct io_context *ioc) |
22e2c507 | 2665 | { |
4ac845a2 | 2666 | call_for_each_cic(ioc, changed_ioprio); |
fc46379d | 2667 | ioc->ioprio_changed = 0; |
22e2c507 JA |
2668 | } |
2669 | ||
d5036d77 | 2670 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
a6151c3a | 2671 | pid_t pid, bool is_sync) |
d5036d77 JA |
2672 | { |
2673 | RB_CLEAR_NODE(&cfqq->rb_node); | |
2674 | RB_CLEAR_NODE(&cfqq->p_node); | |
2675 | INIT_LIST_HEAD(&cfqq->fifo); | |
2676 | ||
2677 | atomic_set(&cfqq->ref, 0); | |
2678 | cfqq->cfqd = cfqd; | |
2679 | ||
2680 | cfq_mark_cfqq_prio_changed(cfqq); | |
2681 | ||
2682 | if (is_sync) { | |
2683 | if (!cfq_class_idle(cfqq)) | |
2684 | cfq_mark_cfqq_idle_window(cfqq); | |
2685 | cfq_mark_cfqq_sync(cfqq); | |
2686 | } | |
2687 | cfqq->pid = pid; | |
2688 | } | |
2689 | ||
24610333 VG |
2690 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
2691 | static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic) | |
2692 | { | |
2693 | struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1); | |
2694 | struct cfq_data *cfqd = cic->key; | |
2695 | unsigned long flags; | |
2696 | struct request_queue *q; | |
2697 | ||
2698 | if (unlikely(!cfqd)) | |
2699 | return; | |
2700 | ||
2701 | q = cfqd->queue; | |
2702 | ||
2703 | spin_lock_irqsave(q->queue_lock, flags); | |
2704 | ||
2705 | if (sync_cfqq) { | |
2706 | /* | |
2707 | * Drop reference to sync queue. A new sync queue will be | |
2708 | * assigned in new group upon arrival of a fresh request. | |
2709 | */ | |
2710 | cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup"); | |
2711 | cic_set_cfqq(cic, NULL, 1); | |
2712 | cfq_put_queue(sync_cfqq); | |
2713 | } | |
2714 | ||
2715 | spin_unlock_irqrestore(q->queue_lock, flags); | |
2716 | } | |
2717 | ||
2718 | static void cfq_ioc_set_cgroup(struct io_context *ioc) | |
2719 | { | |
2720 | call_for_each_cic(ioc, changed_cgroup); | |
2721 | ioc->cgroup_changed = 0; | |
2722 | } | |
2723 | #endif /* CONFIG_CFQ_GROUP_IOSCHED */ | |
2724 | ||
22e2c507 | 2725 | static struct cfq_queue * |
a6151c3a | 2726 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, |
fd0928df | 2727 | struct io_context *ioc, gfp_t gfp_mask) |
22e2c507 | 2728 | { |
22e2c507 | 2729 | struct cfq_queue *cfqq, *new_cfqq = NULL; |
91fac317 | 2730 | struct cfq_io_context *cic; |
cdb16e8f | 2731 | struct cfq_group *cfqg; |
22e2c507 JA |
2732 | |
2733 | retry: | |
cdb16e8f | 2734 | cfqg = cfq_get_cfqg(cfqd, 1); |
4ac845a2 | 2735 | cic = cfq_cic_lookup(cfqd, ioc); |
91fac317 VT |
2736 | /* cic always exists here */ |
2737 | cfqq = cic_to_cfqq(cic, is_sync); | |
22e2c507 | 2738 | |
6118b70b JA |
2739 | /* |
2740 | * Always try a new alloc if we fell back to the OOM cfqq | |
2741 | * originally, since it should just be a temporary situation. | |
2742 | */ | |
2743 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
2744 | cfqq = NULL; | |
22e2c507 JA |
2745 | if (new_cfqq) { |
2746 | cfqq = new_cfqq; | |
2747 | new_cfqq = NULL; | |
2748 | } else if (gfp_mask & __GFP_WAIT) { | |
2749 | spin_unlock_irq(cfqd->queue->queue_lock); | |
94f6030c | 2750 | new_cfqq = kmem_cache_alloc_node(cfq_pool, |
6118b70b | 2751 | gfp_mask | __GFP_ZERO, |
94f6030c | 2752 | cfqd->queue->node); |
22e2c507 | 2753 | spin_lock_irq(cfqd->queue->queue_lock); |
6118b70b JA |
2754 | if (new_cfqq) |
2755 | goto retry; | |
22e2c507 | 2756 | } else { |
94f6030c CL |
2757 | cfqq = kmem_cache_alloc_node(cfq_pool, |
2758 | gfp_mask | __GFP_ZERO, | |
2759 | cfqd->queue->node); | |
22e2c507 JA |
2760 | } |
2761 | ||
6118b70b JA |
2762 | if (cfqq) { |
2763 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
2764 | cfq_init_prio_data(cfqq, ioc); | |
cdb16e8f | 2765 | cfq_link_cfqq_cfqg(cfqq, cfqg); |
6118b70b JA |
2766 | cfq_log_cfqq(cfqd, cfqq, "alloced"); |
2767 | } else | |
2768 | cfqq = &cfqd->oom_cfqq; | |
22e2c507 JA |
2769 | } |
2770 | ||
2771 | if (new_cfqq) | |
2772 | kmem_cache_free(cfq_pool, new_cfqq); | |
2773 | ||
22e2c507 JA |
2774 | return cfqq; |
2775 | } | |
2776 | ||
c2dea2d1 VT |
2777 | static struct cfq_queue ** |
2778 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
2779 | { | |
fe094d98 | 2780 | switch (ioprio_class) { |
c2dea2d1 VT |
2781 | case IOPRIO_CLASS_RT: |
2782 | return &cfqd->async_cfqq[0][ioprio]; | |
2783 | case IOPRIO_CLASS_BE: | |
2784 | return &cfqd->async_cfqq[1][ioprio]; | |
2785 | case IOPRIO_CLASS_IDLE: | |
2786 | return &cfqd->async_idle_cfqq; | |
2787 | default: | |
2788 | BUG(); | |
2789 | } | |
2790 | } | |
2791 | ||
15c31be4 | 2792 | static struct cfq_queue * |
a6151c3a | 2793 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, |
15c31be4 JA |
2794 | gfp_t gfp_mask) |
2795 | { | |
fd0928df JA |
2796 | const int ioprio = task_ioprio(ioc); |
2797 | const int ioprio_class = task_ioprio_class(ioc); | |
c2dea2d1 | 2798 | struct cfq_queue **async_cfqq = NULL; |
15c31be4 JA |
2799 | struct cfq_queue *cfqq = NULL; |
2800 | ||
c2dea2d1 VT |
2801 | if (!is_sync) { |
2802 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
2803 | cfqq = *async_cfqq; | |
2804 | } | |
2805 | ||
6118b70b | 2806 | if (!cfqq) |
fd0928df | 2807 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); |
15c31be4 JA |
2808 | |
2809 | /* | |
2810 | * pin the queue now that it's allocated, scheduler exit will prune it | |
2811 | */ | |
c2dea2d1 | 2812 | if (!is_sync && !(*async_cfqq)) { |
15c31be4 | 2813 | atomic_inc(&cfqq->ref); |
c2dea2d1 | 2814 | *async_cfqq = cfqq; |
15c31be4 JA |
2815 | } |
2816 | ||
2817 | atomic_inc(&cfqq->ref); | |
2818 | return cfqq; | |
2819 | } | |
2820 | ||
498d3aa2 JA |
2821 | /* |
2822 | * We drop cfq io contexts lazily, so we may find a dead one. | |
2823 | */ | |
dbecf3ab | 2824 | static void |
4ac845a2 JA |
2825 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, |
2826 | struct cfq_io_context *cic) | |
dbecf3ab | 2827 | { |
4ac845a2 JA |
2828 | unsigned long flags; |
2829 | ||
fc46379d | 2830 | WARN_ON(!list_empty(&cic->queue_list)); |
597bc485 | 2831 | |
4ac845a2 JA |
2832 | spin_lock_irqsave(&ioc->lock, flags); |
2833 | ||
4faa3c81 | 2834 | BUG_ON(ioc->ioc_data == cic); |
597bc485 | 2835 | |
4ac845a2 | 2836 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); |
ffc4e759 | 2837 | hlist_del_rcu(&cic->cic_list); |
4ac845a2 JA |
2838 | spin_unlock_irqrestore(&ioc->lock, flags); |
2839 | ||
2840 | cfq_cic_free(cic); | |
dbecf3ab OH |
2841 | } |
2842 | ||
e2d74ac0 | 2843 | static struct cfq_io_context * |
4ac845a2 | 2844 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) |
e2d74ac0 | 2845 | { |
e2d74ac0 | 2846 | struct cfq_io_context *cic; |
d6de8be7 | 2847 | unsigned long flags; |
4ac845a2 | 2848 | void *k; |
e2d74ac0 | 2849 | |
91fac317 VT |
2850 | if (unlikely(!ioc)) |
2851 | return NULL; | |
2852 | ||
d6de8be7 JA |
2853 | rcu_read_lock(); |
2854 | ||
597bc485 JA |
2855 | /* |
2856 | * we maintain a last-hit cache, to avoid browsing over the tree | |
2857 | */ | |
4ac845a2 | 2858 | cic = rcu_dereference(ioc->ioc_data); |
d6de8be7 JA |
2859 | if (cic && cic->key == cfqd) { |
2860 | rcu_read_unlock(); | |
597bc485 | 2861 | return cic; |
d6de8be7 | 2862 | } |
597bc485 | 2863 | |
4ac845a2 | 2864 | do { |
4ac845a2 JA |
2865 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); |
2866 | rcu_read_unlock(); | |
2867 | if (!cic) | |
2868 | break; | |
be3b0753 OH |
2869 | /* ->key must be copied to avoid race with cfq_exit_queue() */ |
2870 | k = cic->key; | |
2871 | if (unlikely(!k)) { | |
4ac845a2 | 2872 | cfq_drop_dead_cic(cfqd, ioc, cic); |
d6de8be7 | 2873 | rcu_read_lock(); |
4ac845a2 | 2874 | continue; |
dbecf3ab | 2875 | } |
e2d74ac0 | 2876 | |
d6de8be7 | 2877 | spin_lock_irqsave(&ioc->lock, flags); |
4ac845a2 | 2878 | rcu_assign_pointer(ioc->ioc_data, cic); |
d6de8be7 | 2879 | spin_unlock_irqrestore(&ioc->lock, flags); |
4ac845a2 JA |
2880 | break; |
2881 | } while (1); | |
e2d74ac0 | 2882 | |
4ac845a2 | 2883 | return cic; |
e2d74ac0 JA |
2884 | } |
2885 | ||
4ac845a2 JA |
2886 | /* |
2887 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
2888 | * the process specific cfq io context when entered from the block layer. | |
2889 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
2890 | */ | |
febffd61 JA |
2891 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, |
2892 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
e2d74ac0 | 2893 | { |
0261d688 | 2894 | unsigned long flags; |
4ac845a2 | 2895 | int ret; |
e2d74ac0 | 2896 | |
4ac845a2 JA |
2897 | ret = radix_tree_preload(gfp_mask); |
2898 | if (!ret) { | |
2899 | cic->ioc = ioc; | |
2900 | cic->key = cfqd; | |
e2d74ac0 | 2901 | |
4ac845a2 JA |
2902 | spin_lock_irqsave(&ioc->lock, flags); |
2903 | ret = radix_tree_insert(&ioc->radix_root, | |
2904 | (unsigned long) cfqd, cic); | |
ffc4e759 JA |
2905 | if (!ret) |
2906 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
4ac845a2 | 2907 | spin_unlock_irqrestore(&ioc->lock, flags); |
e2d74ac0 | 2908 | |
4ac845a2 JA |
2909 | radix_tree_preload_end(); |
2910 | ||
2911 | if (!ret) { | |
2912 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
2913 | list_add(&cic->queue_list, &cfqd->cic_list); | |
2914 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2915 | } | |
e2d74ac0 JA |
2916 | } |
2917 | ||
4ac845a2 JA |
2918 | if (ret) |
2919 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
fc46379d | 2920 | |
4ac845a2 | 2921 | return ret; |
e2d74ac0 JA |
2922 | } |
2923 | ||
1da177e4 LT |
2924 | /* |
2925 | * Setup general io context and cfq io context. There can be several cfq | |
2926 | * io contexts per general io context, if this process is doing io to more | |
e2d74ac0 | 2927 | * than one device managed by cfq. |
1da177e4 LT |
2928 | */ |
2929 | static struct cfq_io_context * | |
e2d74ac0 | 2930 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) |
1da177e4 | 2931 | { |
22e2c507 | 2932 | struct io_context *ioc = NULL; |
1da177e4 | 2933 | struct cfq_io_context *cic; |
1da177e4 | 2934 | |
22e2c507 | 2935 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1da177e4 | 2936 | |
b5deef90 | 2937 | ioc = get_io_context(gfp_mask, cfqd->queue->node); |
1da177e4 LT |
2938 | if (!ioc) |
2939 | return NULL; | |
2940 | ||
4ac845a2 | 2941 | cic = cfq_cic_lookup(cfqd, ioc); |
e2d74ac0 JA |
2942 | if (cic) |
2943 | goto out; | |
1da177e4 | 2944 | |
e2d74ac0 JA |
2945 | cic = cfq_alloc_io_context(cfqd, gfp_mask); |
2946 | if (cic == NULL) | |
2947 | goto err; | |
1da177e4 | 2948 | |
4ac845a2 JA |
2949 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) |
2950 | goto err_free; | |
2951 | ||
1da177e4 | 2952 | out: |
fc46379d JA |
2953 | smp_read_barrier_depends(); |
2954 | if (unlikely(ioc->ioprio_changed)) | |
2955 | cfq_ioc_set_ioprio(ioc); | |
2956 | ||
24610333 VG |
2957 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
2958 | if (unlikely(ioc->cgroup_changed)) | |
2959 | cfq_ioc_set_cgroup(ioc); | |
2960 | #endif | |
1da177e4 | 2961 | return cic; |
4ac845a2 JA |
2962 | err_free: |
2963 | cfq_cic_free(cic); | |
1da177e4 LT |
2964 | err: |
2965 | put_io_context(ioc); | |
2966 | return NULL; | |
2967 | } | |
2968 | ||
22e2c507 JA |
2969 | static void |
2970 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
1da177e4 | 2971 | { |
aaf1228d JA |
2972 | unsigned long elapsed = jiffies - cic->last_end_request; |
2973 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
db3b5848 | 2974 | |
22e2c507 JA |
2975 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; |
2976 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
2977 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
2978 | } | |
1da177e4 | 2979 | |
206dc69b | 2980 | static void |
b2c18e1e | 2981 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
6d048f53 | 2982 | struct request *rq) |
206dc69b | 2983 | { |
3dde36dd | 2984 | sector_t sdist = 0; |
41647e7a | 2985 | sector_t n_sec = blk_rq_sectors(rq); |
3dde36dd CZ |
2986 | if (cfqq->last_request_pos) { |
2987 | if (cfqq->last_request_pos < blk_rq_pos(rq)) | |
2988 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
2989 | else | |
2990 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); | |
2991 | } | |
206dc69b | 2992 | |
3dde36dd | 2993 | cfqq->seek_history <<= 1; |
41647e7a CZ |
2994 | if (blk_queue_nonrot(cfqd->queue)) |
2995 | cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT); | |
2996 | else | |
2997 | cfqq->seek_history |= (sdist > CFQQ_SEEK_THR); | |
206dc69b | 2998 | } |
1da177e4 | 2999 | |
22e2c507 JA |
3000 | /* |
3001 | * Disable idle window if the process thinks too long or seeks so much that | |
3002 | * it doesn't matter | |
3003 | */ | |
3004 | static void | |
3005 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3006 | struct cfq_io_context *cic) | |
3007 | { | |
7b679138 | 3008 | int old_idle, enable_idle; |
1be92f2f | 3009 | |
0871714e JA |
3010 | /* |
3011 | * Don't idle for async or idle io prio class | |
3012 | */ | |
3013 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
1be92f2f JA |
3014 | return; |
3015 | ||
c265a7f4 | 3016 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); |
1da177e4 | 3017 | |
76280aff CZ |
3018 | if (cfqq->queued[0] + cfqq->queued[1] >= 4) |
3019 | cfq_mark_cfqq_deep(cfqq); | |
3020 | ||
66dac98e | 3021 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || |
3dde36dd | 3022 | (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq))) |
22e2c507 JA |
3023 | enable_idle = 0; |
3024 | else if (sample_valid(cic->ttime_samples)) { | |
718eee05 | 3025 | if (cic->ttime_mean > cfqd->cfq_slice_idle) |
22e2c507 JA |
3026 | enable_idle = 0; |
3027 | else | |
3028 | enable_idle = 1; | |
1da177e4 LT |
3029 | } |
3030 | ||
7b679138 JA |
3031 | if (old_idle != enable_idle) { |
3032 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
3033 | if (enable_idle) | |
3034 | cfq_mark_cfqq_idle_window(cfqq); | |
3035 | else | |
3036 | cfq_clear_cfqq_idle_window(cfqq); | |
3037 | } | |
22e2c507 | 3038 | } |
1da177e4 | 3039 | |
22e2c507 JA |
3040 | /* |
3041 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
3042 | * no or if we aren't sure, a 1 will cause a preempt. | |
3043 | */ | |
a6151c3a | 3044 | static bool |
22e2c507 | 3045 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, |
5e705374 | 3046 | struct request *rq) |
22e2c507 | 3047 | { |
6d048f53 | 3048 | struct cfq_queue *cfqq; |
22e2c507 | 3049 | |
6d048f53 JA |
3050 | cfqq = cfqd->active_queue; |
3051 | if (!cfqq) | |
a6151c3a | 3052 | return false; |
22e2c507 | 3053 | |
6d048f53 | 3054 | if (cfq_class_idle(new_cfqq)) |
a6151c3a | 3055 | return false; |
22e2c507 JA |
3056 | |
3057 | if (cfq_class_idle(cfqq)) | |
a6151c3a | 3058 | return true; |
1e3335de | 3059 | |
875feb63 DS |
3060 | /* |
3061 | * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice. | |
3062 | */ | |
3063 | if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq)) | |
3064 | return false; | |
3065 | ||
374f84ac JA |
3066 | /* |
3067 | * if the new request is sync, but the currently running queue is | |
3068 | * not, let the sync request have priority. | |
3069 | */ | |
5e705374 | 3070 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) |
a6151c3a | 3071 | return true; |
1e3335de | 3072 | |
8682e1f1 VG |
3073 | if (new_cfqq->cfqg != cfqq->cfqg) |
3074 | return false; | |
3075 | ||
3076 | if (cfq_slice_used(cfqq)) | |
3077 | return true; | |
3078 | ||
3079 | /* Allow preemption only if we are idling on sync-noidle tree */ | |
3080 | if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && | |
3081 | cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && | |
3082 | new_cfqq->service_tree->count == 2 && | |
3083 | RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3084 | return true; | |
3085 | ||
374f84ac JA |
3086 | /* |
3087 | * So both queues are sync. Let the new request get disk time if | |
3088 | * it's a metadata request and the current queue is doing regular IO. | |
3089 | */ | |
3090 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
e6ec4fe2 | 3091 | return true; |
22e2c507 | 3092 | |
3a9a3f6c DS |
3093 | /* |
3094 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
3095 | */ | |
3096 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
a6151c3a | 3097 | return true; |
3a9a3f6c | 3098 | |
1e3335de | 3099 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) |
a6151c3a | 3100 | return false; |
1e3335de JA |
3101 | |
3102 | /* | |
3103 | * if this request is as-good as one we would expect from the | |
3104 | * current cfqq, let it preempt | |
3105 | */ | |
2f7a2d89 | 3106 | if (cfq_rq_close(cfqd, cfqq, rq, true)) |
a6151c3a | 3107 | return true; |
1e3335de | 3108 | |
a6151c3a | 3109 | return false; |
22e2c507 JA |
3110 | } |
3111 | ||
3112 | /* | |
3113 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
3114 | * let it have half of its nominal slice. | |
3115 | */ | |
3116 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3117 | { | |
7b679138 | 3118 | cfq_log_cfqq(cfqd, cfqq, "preempt"); |
6084cdda | 3119 | cfq_slice_expired(cfqd, 1); |
22e2c507 | 3120 | |
bf572256 JA |
3121 | /* |
3122 | * Put the new queue at the front of the of the current list, | |
3123 | * so we know that it will be selected next. | |
3124 | */ | |
3125 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
edd75ffd JA |
3126 | |
3127 | cfq_service_tree_add(cfqd, cfqq, 1); | |
bf572256 | 3128 | |
44f7c160 JA |
3129 | cfqq->slice_end = 0; |
3130 | cfq_mark_cfqq_slice_new(cfqq); | |
22e2c507 JA |
3131 | } |
3132 | ||
22e2c507 | 3133 | /* |
5e705374 | 3134 | * Called when a new fs request (rq) is added (to cfqq). Check if there's |
22e2c507 JA |
3135 | * something we should do about it |
3136 | */ | |
3137 | static void | |
5e705374 JA |
3138 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, |
3139 | struct request *rq) | |
22e2c507 | 3140 | { |
5e705374 | 3141 | struct cfq_io_context *cic = RQ_CIC(rq); |
12e9fddd | 3142 | |
45333d5a | 3143 | cfqd->rq_queued++; |
374f84ac JA |
3144 | if (rq_is_meta(rq)) |
3145 | cfqq->meta_pending++; | |
3146 | ||
9c2c38a1 | 3147 | cfq_update_io_thinktime(cfqd, cic); |
b2c18e1e | 3148 | cfq_update_io_seektime(cfqd, cfqq, rq); |
9c2c38a1 JA |
3149 | cfq_update_idle_window(cfqd, cfqq, cic); |
3150 | ||
b2c18e1e | 3151 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
22e2c507 JA |
3152 | |
3153 | if (cfqq == cfqd->active_queue) { | |
3154 | /* | |
b029195d JA |
3155 | * Remember that we saw a request from this process, but |
3156 | * don't start queuing just yet. Otherwise we risk seeing lots | |
3157 | * of tiny requests, because we disrupt the normal plugging | |
d6ceb25e JA |
3158 | * and merging. If the request is already larger than a single |
3159 | * page, let it rip immediately. For that case we assume that | |
2d870722 JA |
3160 | * merging is already done. Ditto for a busy system that |
3161 | * has other work pending, don't risk delaying until the | |
3162 | * idle timer unplug to continue working. | |
22e2c507 | 3163 | */ |
d6ceb25e | 3164 | if (cfq_cfqq_wait_request(cfqq)) { |
2d870722 JA |
3165 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || |
3166 | cfqd->busy_queues > 1) { | |
d6ceb25e | 3167 | del_timer(&cfqd->idle_slice_timer); |
554554f6 | 3168 | cfq_clear_cfqq_wait_request(cfqq); |
bf791937 VG |
3169 | __blk_run_queue(cfqd->queue); |
3170 | } else | |
3171 | cfq_mark_cfqq_must_dispatch(cfqq); | |
d6ceb25e | 3172 | } |
5e705374 | 3173 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { |
22e2c507 JA |
3174 | /* |
3175 | * not the active queue - expire current slice if it is | |
3176 | * idle and has expired it's mean thinktime or this new queue | |
3a9a3f6c DS |
3177 | * has some old slice time left and is of higher priority or |
3178 | * this new queue is RT and the current one is BE | |
22e2c507 JA |
3179 | */ |
3180 | cfq_preempt_queue(cfqd, cfqq); | |
a7f55792 | 3181 | __blk_run_queue(cfqd->queue); |
22e2c507 | 3182 | } |
1da177e4 LT |
3183 | } |
3184 | ||
165125e1 | 3185 | static void cfq_insert_request(struct request_queue *q, struct request *rq) |
1da177e4 | 3186 | { |
b4878f24 | 3187 | struct cfq_data *cfqd = q->elevator->elevator_data; |
5e705374 | 3188 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
22e2c507 | 3189 | |
7b679138 | 3190 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); |
fd0928df | 3191 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); |
1da177e4 | 3192 | |
30996f40 | 3193 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); |
22e2c507 | 3194 | list_add_tail(&rq->queuelist, &cfqq->fifo); |
aa6f6a3d | 3195 | cfq_add_rq_rb(rq); |
22e2c507 | 3196 | |
5e705374 | 3197 | cfq_rq_enqueued(cfqd, cfqq, rq); |
1da177e4 LT |
3198 | } |
3199 | ||
45333d5a AC |
3200 | /* |
3201 | * Update hw_tag based on peak queue depth over 50 samples under | |
3202 | * sufficient load. | |
3203 | */ | |
3204 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
3205 | { | |
1a1238a7 SL |
3206 | struct cfq_queue *cfqq = cfqd->active_queue; |
3207 | ||
53c583d2 CZ |
3208 | if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth) |
3209 | cfqd->hw_tag_est_depth = cfqd->rq_in_driver; | |
e459dd08 CZ |
3210 | |
3211 | if (cfqd->hw_tag == 1) | |
3212 | return; | |
45333d5a AC |
3213 | |
3214 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
53c583d2 | 3215 | cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
3216 | return; |
3217 | ||
1a1238a7 SL |
3218 | /* |
3219 | * If active queue hasn't enough requests and can idle, cfq might not | |
3220 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
3221 | * case | |
3222 | */ | |
3223 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
3224 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
53c583d2 | 3225 | CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN) |
1a1238a7 SL |
3226 | return; |
3227 | ||
45333d5a AC |
3228 | if (cfqd->hw_tag_samples++ < 50) |
3229 | return; | |
3230 | ||
e459dd08 | 3231 | if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) |
45333d5a AC |
3232 | cfqd->hw_tag = 1; |
3233 | else | |
3234 | cfqd->hw_tag = 0; | |
45333d5a AC |
3235 | } |
3236 | ||
7667aa06 VG |
3237 | static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq) |
3238 | { | |
3239 | struct cfq_io_context *cic = cfqd->active_cic; | |
3240 | ||
3241 | /* If there are other queues in the group, don't wait */ | |
3242 | if (cfqq->cfqg->nr_cfqq > 1) | |
3243 | return false; | |
3244 | ||
3245 | if (cfq_slice_used(cfqq)) | |
3246 | return true; | |
3247 | ||
3248 | /* if slice left is less than think time, wait busy */ | |
3249 | if (cic && sample_valid(cic->ttime_samples) | |
3250 | && (cfqq->slice_end - jiffies < cic->ttime_mean)) | |
3251 | return true; | |
3252 | ||
3253 | /* | |
3254 | * If think times is less than a jiffy than ttime_mean=0 and above | |
3255 | * will not be true. It might happen that slice has not expired yet | |
3256 | * but will expire soon (4-5 ns) during select_queue(). To cover the | |
3257 | * case where think time is less than a jiffy, mark the queue wait | |
3258 | * busy if only 1 jiffy is left in the slice. | |
3259 | */ | |
3260 | if (cfqq->slice_end - jiffies == 1) | |
3261 | return true; | |
3262 | ||
3263 | return false; | |
3264 | } | |
3265 | ||
165125e1 | 3266 | static void cfq_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 3267 | { |
5e705374 | 3268 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
b4878f24 | 3269 | struct cfq_data *cfqd = cfqq->cfqd; |
5380a101 | 3270 | const int sync = rq_is_sync(rq); |
b4878f24 | 3271 | unsigned long now; |
1da177e4 | 3272 | |
b4878f24 | 3273 | now = jiffies; |
2868ef7b | 3274 | cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq)); |
1da177e4 | 3275 | |
45333d5a AC |
3276 | cfq_update_hw_tag(cfqd); |
3277 | ||
53c583d2 | 3278 | WARN_ON(!cfqd->rq_in_driver); |
6d048f53 | 3279 | WARN_ON(!cfqq->dispatched); |
53c583d2 | 3280 | cfqd->rq_in_driver--; |
6d048f53 | 3281 | cfqq->dispatched--; |
1da177e4 | 3282 | |
53c583d2 | 3283 | cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--; |
3ed9a296 | 3284 | |
365722bb | 3285 | if (sync) { |
5e705374 | 3286 | RQ_CIC(rq)->last_end_request = now; |
573412b2 CZ |
3287 | if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now)) |
3288 | cfqd->last_delayed_sync = now; | |
365722bb | 3289 | } |
caaa5f9f JA |
3290 | |
3291 | /* | |
3292 | * If this is the active queue, check if it needs to be expired, | |
3293 | * or if we want to idle in case it has no pending requests. | |
3294 | */ | |
3295 | if (cfqd->active_queue == cfqq) { | |
a36e71f9 JA |
3296 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); |
3297 | ||
44f7c160 JA |
3298 | if (cfq_cfqq_slice_new(cfqq)) { |
3299 | cfq_set_prio_slice(cfqd, cfqq); | |
3300 | cfq_clear_cfqq_slice_new(cfqq); | |
3301 | } | |
f75edf2d VG |
3302 | |
3303 | /* | |
7667aa06 VG |
3304 | * Should we wait for next request to come in before we expire |
3305 | * the queue. | |
f75edf2d | 3306 | */ |
7667aa06 | 3307 | if (cfq_should_wait_busy(cfqd, cfqq)) { |
f75edf2d VG |
3308 | cfqq->slice_end = jiffies + cfqd->cfq_slice_idle; |
3309 | cfq_mark_cfqq_wait_busy(cfqq); | |
3310 | } | |
3311 | ||
a36e71f9 | 3312 | /* |
8e550632 CZ |
3313 | * Idling is not enabled on: |
3314 | * - expired queues | |
3315 | * - idle-priority queues | |
3316 | * - async queues | |
3317 | * - queues with still some requests queued | |
3318 | * - when there is a close cooperator | |
a36e71f9 | 3319 | */ |
0871714e | 3320 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) |
6084cdda | 3321 | cfq_slice_expired(cfqd, 1); |
8e550632 CZ |
3322 | else if (sync && cfqq_empty && |
3323 | !cfq_close_cooperator(cfqd, cfqq)) { | |
3324 | cfqd->noidle_tree_requires_idle |= !rq_noidle(rq); | |
3325 | /* | |
3326 | * Idling is enabled for SYNC_WORKLOAD. | |
3327 | * SYNC_NOIDLE_WORKLOAD idles at the end of the tree | |
3328 | * only if we processed at least one !rq_noidle request | |
3329 | */ | |
3330 | if (cfqd->serving_type == SYNC_WORKLOAD | |
c04645e5 VG |
3331 | || cfqd->noidle_tree_requires_idle |
3332 | || cfqq->cfqg->nr_cfqq == 1) | |
8e550632 CZ |
3333 | cfq_arm_slice_timer(cfqd); |
3334 | } | |
caaa5f9f | 3335 | } |
6d048f53 | 3336 | |
53c583d2 | 3337 | if (!cfqd->rq_in_driver) |
23e018a1 | 3338 | cfq_schedule_dispatch(cfqd); |
1da177e4 LT |
3339 | } |
3340 | ||
22e2c507 JA |
3341 | /* |
3342 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
3343 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
3344 | */ | |
3345 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
1da177e4 | 3346 | { |
22e2c507 JA |
3347 | if (has_fs_excl()) { |
3348 | /* | |
3349 | * boost idle prio on transactions that would lock out other | |
3350 | * users of the filesystem | |
3351 | */ | |
3352 | if (cfq_class_idle(cfqq)) | |
3353 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
3354 | if (cfqq->ioprio > IOPRIO_NORM) | |
3355 | cfqq->ioprio = IOPRIO_NORM; | |
3356 | } else { | |
3357 | /* | |
dddb7451 | 3358 | * unboost the queue (if needed) |
22e2c507 | 3359 | */ |
dddb7451 CZ |
3360 | cfqq->ioprio_class = cfqq->org_ioprio_class; |
3361 | cfqq->ioprio = cfqq->org_ioprio; | |
22e2c507 | 3362 | } |
22e2c507 | 3363 | } |
1da177e4 | 3364 | |
89850f7e | 3365 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) |
22e2c507 | 3366 | { |
1b379d8d | 3367 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { |
3b18152c | 3368 | cfq_mark_cfqq_must_alloc_slice(cfqq); |
22e2c507 | 3369 | return ELV_MQUEUE_MUST; |
3b18152c | 3370 | } |
1da177e4 | 3371 | |
22e2c507 | 3372 | return ELV_MQUEUE_MAY; |
22e2c507 JA |
3373 | } |
3374 | ||
165125e1 | 3375 | static int cfq_may_queue(struct request_queue *q, int rw) |
22e2c507 JA |
3376 | { |
3377 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3378 | struct task_struct *tsk = current; | |
91fac317 | 3379 | struct cfq_io_context *cic; |
22e2c507 JA |
3380 | struct cfq_queue *cfqq; |
3381 | ||
3382 | /* | |
3383 | * don't force setup of a queue from here, as a call to may_queue | |
3384 | * does not necessarily imply that a request actually will be queued. | |
3385 | * so just lookup a possibly existing queue, or return 'may queue' | |
3386 | * if that fails | |
3387 | */ | |
4ac845a2 | 3388 | cic = cfq_cic_lookup(cfqd, tsk->io_context); |
91fac317 VT |
3389 | if (!cic) |
3390 | return ELV_MQUEUE_MAY; | |
3391 | ||
b0b78f81 | 3392 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); |
22e2c507 | 3393 | if (cfqq) { |
fd0928df | 3394 | cfq_init_prio_data(cfqq, cic->ioc); |
22e2c507 JA |
3395 | cfq_prio_boost(cfqq); |
3396 | ||
89850f7e | 3397 | return __cfq_may_queue(cfqq); |
22e2c507 JA |
3398 | } |
3399 | ||
3400 | return ELV_MQUEUE_MAY; | |
1da177e4 LT |
3401 | } |
3402 | ||
1da177e4 LT |
3403 | /* |
3404 | * queue lock held here | |
3405 | */ | |
bb37b94c | 3406 | static void cfq_put_request(struct request *rq) |
1da177e4 | 3407 | { |
5e705374 | 3408 | struct cfq_queue *cfqq = RQ_CFQQ(rq); |
1da177e4 | 3409 | |
5e705374 | 3410 | if (cfqq) { |
22e2c507 | 3411 | const int rw = rq_data_dir(rq); |
1da177e4 | 3412 | |
22e2c507 JA |
3413 | BUG_ON(!cfqq->allocated[rw]); |
3414 | cfqq->allocated[rw]--; | |
1da177e4 | 3415 | |
5e705374 | 3416 | put_io_context(RQ_CIC(rq)->ioc); |
1da177e4 | 3417 | |
1da177e4 | 3418 | rq->elevator_private = NULL; |
5e705374 | 3419 | rq->elevator_private2 = NULL; |
1da177e4 | 3420 | |
1da177e4 LT |
3421 | cfq_put_queue(cfqq); |
3422 | } | |
3423 | } | |
3424 | ||
df5fe3e8 JM |
3425 | static struct cfq_queue * |
3426 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, | |
3427 | struct cfq_queue *cfqq) | |
3428 | { | |
3429 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
3430 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
b3b6d040 | 3431 | cfq_mark_cfqq_coop(cfqq->new_cfqq); |
df5fe3e8 JM |
3432 | cfq_put_queue(cfqq); |
3433 | return cic_to_cfqq(cic, 1); | |
3434 | } | |
3435 | ||
e6c5bc73 JM |
3436 | /* |
3437 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
3438 | * was the last process referring to said cfqq. | |
3439 | */ | |
3440 | static struct cfq_queue * | |
3441 | split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) | |
3442 | { | |
3443 | if (cfqq_process_refs(cfqq) == 1) { | |
e6c5bc73 JM |
3444 | cfqq->pid = current->pid; |
3445 | cfq_clear_cfqq_coop(cfqq); | |
ae54abed | 3446 | cfq_clear_cfqq_split_coop(cfqq); |
e6c5bc73 JM |
3447 | return cfqq; |
3448 | } | |
3449 | ||
3450 | cic_set_cfqq(cic, NULL, 1); | |
3451 | cfq_put_queue(cfqq); | |
3452 | return NULL; | |
3453 | } | |
1da177e4 | 3454 | /* |
22e2c507 | 3455 | * Allocate cfq data structures associated with this request. |
1da177e4 | 3456 | */ |
22e2c507 | 3457 | static int |
165125e1 | 3458 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
1da177e4 LT |
3459 | { |
3460 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3461 | struct cfq_io_context *cic; | |
3462 | const int rw = rq_data_dir(rq); | |
a6151c3a | 3463 | const bool is_sync = rq_is_sync(rq); |
22e2c507 | 3464 | struct cfq_queue *cfqq; |
1da177e4 LT |
3465 | unsigned long flags; |
3466 | ||
3467 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
3468 | ||
e2d74ac0 | 3469 | cic = cfq_get_io_context(cfqd, gfp_mask); |
22e2c507 | 3470 | |
1da177e4 LT |
3471 | spin_lock_irqsave(q->queue_lock, flags); |
3472 | ||
22e2c507 JA |
3473 | if (!cic) |
3474 | goto queue_fail; | |
3475 | ||
e6c5bc73 | 3476 | new_queue: |
91fac317 | 3477 | cfqq = cic_to_cfqq(cic, is_sync); |
32f2e807 | 3478 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { |
fd0928df | 3479 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); |
91fac317 | 3480 | cic_set_cfqq(cic, cfqq, is_sync); |
df5fe3e8 | 3481 | } else { |
e6c5bc73 JM |
3482 | /* |
3483 | * If the queue was seeky for too long, break it apart. | |
3484 | */ | |
ae54abed | 3485 | if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) { |
e6c5bc73 JM |
3486 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); |
3487 | cfqq = split_cfqq(cic, cfqq); | |
3488 | if (!cfqq) | |
3489 | goto new_queue; | |
3490 | } | |
3491 | ||
df5fe3e8 JM |
3492 | /* |
3493 | * Check to see if this queue is scheduled to merge with | |
3494 | * another, closely cooperating queue. The merging of | |
3495 | * queues happens here as it must be done in process context. | |
3496 | * The reference on new_cfqq was taken in merge_cfqqs. | |
3497 | */ | |
3498 | if (cfqq->new_cfqq) | |
3499 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
91fac317 | 3500 | } |
1da177e4 LT |
3501 | |
3502 | cfqq->allocated[rw]++; | |
22e2c507 | 3503 | atomic_inc(&cfqq->ref); |
1da177e4 | 3504 | |
5e705374 | 3505 | spin_unlock_irqrestore(q->queue_lock, flags); |
3b18152c | 3506 | |
5e705374 JA |
3507 | rq->elevator_private = cic; |
3508 | rq->elevator_private2 = cfqq; | |
3509 | return 0; | |
1da177e4 | 3510 | |
22e2c507 JA |
3511 | queue_fail: |
3512 | if (cic) | |
3513 | put_io_context(cic->ioc); | |
89850f7e | 3514 | |
23e018a1 | 3515 | cfq_schedule_dispatch(cfqd); |
1da177e4 | 3516 | spin_unlock_irqrestore(q->queue_lock, flags); |
7b679138 | 3517 | cfq_log(cfqd, "set_request fail"); |
1da177e4 LT |
3518 | return 1; |
3519 | } | |
3520 | ||
65f27f38 | 3521 | static void cfq_kick_queue(struct work_struct *work) |
22e2c507 | 3522 | { |
65f27f38 | 3523 | struct cfq_data *cfqd = |
23e018a1 | 3524 | container_of(work, struct cfq_data, unplug_work); |
165125e1 | 3525 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3526 | |
40bb54d1 | 3527 | spin_lock_irq(q->queue_lock); |
a7f55792 | 3528 | __blk_run_queue(cfqd->queue); |
40bb54d1 | 3529 | spin_unlock_irq(q->queue_lock); |
22e2c507 JA |
3530 | } |
3531 | ||
3532 | /* | |
3533 | * Timer running if the active_queue is currently idling inside its time slice | |
3534 | */ | |
3535 | static void cfq_idle_slice_timer(unsigned long data) | |
3536 | { | |
3537 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
3538 | struct cfq_queue *cfqq; | |
3539 | unsigned long flags; | |
3c6bd2f8 | 3540 | int timed_out = 1; |
22e2c507 | 3541 | |
7b679138 JA |
3542 | cfq_log(cfqd, "idle timer fired"); |
3543 | ||
22e2c507 JA |
3544 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); |
3545 | ||
fe094d98 JA |
3546 | cfqq = cfqd->active_queue; |
3547 | if (cfqq) { | |
3c6bd2f8 JA |
3548 | timed_out = 0; |
3549 | ||
b029195d JA |
3550 | /* |
3551 | * We saw a request before the queue expired, let it through | |
3552 | */ | |
3553 | if (cfq_cfqq_must_dispatch(cfqq)) | |
3554 | goto out_kick; | |
3555 | ||
22e2c507 JA |
3556 | /* |
3557 | * expired | |
3558 | */ | |
44f7c160 | 3559 | if (cfq_slice_used(cfqq)) |
22e2c507 JA |
3560 | goto expire; |
3561 | ||
3562 | /* | |
3563 | * only expire and reinvoke request handler, if there are | |
3564 | * other queues with pending requests | |
3565 | */ | |
caaa5f9f | 3566 | if (!cfqd->busy_queues) |
22e2c507 | 3567 | goto out_cont; |
22e2c507 JA |
3568 | |
3569 | /* | |
3570 | * not expired and it has a request pending, let it dispatch | |
3571 | */ | |
75e50984 | 3572 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) |
22e2c507 | 3573 | goto out_kick; |
76280aff CZ |
3574 | |
3575 | /* | |
3576 | * Queue depth flag is reset only when the idle didn't succeed | |
3577 | */ | |
3578 | cfq_clear_cfqq_deep(cfqq); | |
22e2c507 JA |
3579 | } |
3580 | expire: | |
6084cdda | 3581 | cfq_slice_expired(cfqd, timed_out); |
22e2c507 | 3582 | out_kick: |
23e018a1 | 3583 | cfq_schedule_dispatch(cfqd); |
22e2c507 JA |
3584 | out_cont: |
3585 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
3586 | } | |
3587 | ||
3b18152c JA |
3588 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) |
3589 | { | |
3590 | del_timer_sync(&cfqd->idle_slice_timer); | |
23e018a1 | 3591 | cancel_work_sync(&cfqd->unplug_work); |
3b18152c | 3592 | } |
22e2c507 | 3593 | |
c2dea2d1 VT |
3594 | static void cfq_put_async_queues(struct cfq_data *cfqd) |
3595 | { | |
3596 | int i; | |
3597 | ||
3598 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
3599 | if (cfqd->async_cfqq[0][i]) | |
3600 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
3601 | if (cfqd->async_cfqq[1][i]) | |
3602 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
c2dea2d1 | 3603 | } |
2389d1ef ON |
3604 | |
3605 | if (cfqd->async_idle_cfqq) | |
3606 | cfq_put_queue(cfqd->async_idle_cfqq); | |
c2dea2d1 VT |
3607 | } |
3608 | ||
bb729bc9 JA |
3609 | static void cfq_cfqd_free(struct rcu_head *head) |
3610 | { | |
3611 | kfree(container_of(head, struct cfq_data, rcu)); | |
3612 | } | |
3613 | ||
b374d18a | 3614 | static void cfq_exit_queue(struct elevator_queue *e) |
1da177e4 | 3615 | { |
22e2c507 | 3616 | struct cfq_data *cfqd = e->elevator_data; |
165125e1 | 3617 | struct request_queue *q = cfqd->queue; |
22e2c507 | 3618 | |
3b18152c | 3619 | cfq_shutdown_timer_wq(cfqd); |
e2d74ac0 | 3620 | |
d9ff4187 | 3621 | spin_lock_irq(q->queue_lock); |
e2d74ac0 | 3622 | |
d9ff4187 | 3623 | if (cfqd->active_queue) |
6084cdda | 3624 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); |
e2d74ac0 JA |
3625 | |
3626 | while (!list_empty(&cfqd->cic_list)) { | |
d9ff4187 AV |
3627 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, |
3628 | struct cfq_io_context, | |
3629 | queue_list); | |
89850f7e JA |
3630 | |
3631 | __cfq_exit_single_io_context(cfqd, cic); | |
d9ff4187 | 3632 | } |
e2d74ac0 | 3633 | |
c2dea2d1 | 3634 | cfq_put_async_queues(cfqd); |
b1c35769 VG |
3635 | cfq_release_cfq_groups(cfqd); |
3636 | blkiocg_del_blkio_group(&cfqd->root_group.blkg); | |
15c31be4 | 3637 | |
d9ff4187 | 3638 | spin_unlock_irq(q->queue_lock); |
a90d742e AV |
3639 | |
3640 | cfq_shutdown_timer_wq(cfqd); | |
3641 | ||
b1c35769 | 3642 | /* Wait for cfqg->blkg->key accessors to exit their grace periods. */ |
bb729bc9 | 3643 | call_rcu(&cfqd->rcu, cfq_cfqd_free); |
1da177e4 LT |
3644 | } |
3645 | ||
165125e1 | 3646 | static void *cfq_init_queue(struct request_queue *q) |
1da177e4 LT |
3647 | { |
3648 | struct cfq_data *cfqd; | |
718eee05 | 3649 | int i, j; |
cdb16e8f | 3650 | struct cfq_group *cfqg; |
615f0259 | 3651 | struct cfq_rb_root *st; |
1da177e4 | 3652 | |
94f6030c | 3653 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
1da177e4 | 3654 | if (!cfqd) |
bc1c1169 | 3655 | return NULL; |
1da177e4 | 3656 | |
1fa8f6d6 VG |
3657 | /* Init root service tree */ |
3658 | cfqd->grp_service_tree = CFQ_RB_ROOT; | |
3659 | ||
cdb16e8f VG |
3660 | /* Init root group */ |
3661 | cfqg = &cfqd->root_group; | |
615f0259 VG |
3662 | for_each_cfqg_st(cfqg, i, j, st) |
3663 | *st = CFQ_RB_ROOT; | |
1fa8f6d6 | 3664 | RB_CLEAR_NODE(&cfqg->rb_node); |
26a2ac00 | 3665 | |
25bc6b07 VG |
3666 | /* Give preference to root group over other groups */ |
3667 | cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT; | |
3668 | ||
25fb5169 | 3669 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
b1c35769 VG |
3670 | /* |
3671 | * Take a reference to root group which we never drop. This is just | |
3672 | * to make sure that cfq_put_cfqg() does not try to kfree root group | |
3673 | */ | |
3674 | atomic_set(&cfqg->ref, 1); | |
22084190 VG |
3675 | blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd, |
3676 | 0); | |
25fb5169 | 3677 | #endif |
26a2ac00 JA |
3678 | /* |
3679 | * Not strictly needed (since RB_ROOT just clears the node and we | |
3680 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
3681 | * to add magic to the rb code | |
3682 | */ | |
3683 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
3684 | cfqd->prio_trees[i] = RB_ROOT; | |
3685 | ||
6118b70b JA |
3686 | /* |
3687 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
3688 | * Grab a permanent reference to it, so that the normal code flow | |
3689 | * will not attempt to free it. | |
3690 | */ | |
3691 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
3692 | atomic_inc(&cfqd->oom_cfqq.ref); | |
cdb16e8f | 3693 | cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group); |
6118b70b | 3694 | |
d9ff4187 | 3695 | INIT_LIST_HEAD(&cfqd->cic_list); |
1da177e4 | 3696 | |
1da177e4 | 3697 | cfqd->queue = q; |
1da177e4 | 3698 | |
22e2c507 JA |
3699 | init_timer(&cfqd->idle_slice_timer); |
3700 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
3701 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
3702 | ||
23e018a1 | 3703 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); |
22e2c507 | 3704 | |
1da177e4 | 3705 | cfqd->cfq_quantum = cfq_quantum; |
22e2c507 JA |
3706 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; |
3707 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
1da177e4 LT |
3708 | cfqd->cfq_back_max = cfq_back_max; |
3709 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
22e2c507 JA |
3710 | cfqd->cfq_slice[0] = cfq_slice_async; |
3711 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
3712 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
3713 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
963b72fc | 3714 | cfqd->cfq_latency = 1; |
ae30c286 | 3715 | cfqd->cfq_group_isolation = 0; |
e459dd08 | 3716 | cfqd->hw_tag = -1; |
edc71131 CZ |
3717 | /* |
3718 | * we optimistically start assuming sync ops weren't delayed in last | |
3719 | * second, in order to have larger depth for async operations. | |
3720 | */ | |
573412b2 | 3721 | cfqd->last_delayed_sync = jiffies - HZ; |
bb729bc9 | 3722 | INIT_RCU_HEAD(&cfqd->rcu); |
bc1c1169 | 3723 | return cfqd; |
1da177e4 LT |
3724 | } |
3725 | ||
3726 | static void cfq_slab_kill(void) | |
3727 | { | |
d6de8be7 JA |
3728 | /* |
3729 | * Caller already ensured that pending RCU callbacks are completed, | |
3730 | * so we should have no busy allocations at this point. | |
3731 | */ | |
1da177e4 LT |
3732 | if (cfq_pool) |
3733 | kmem_cache_destroy(cfq_pool); | |
3734 | if (cfq_ioc_pool) | |
3735 | kmem_cache_destroy(cfq_ioc_pool); | |
3736 | } | |
3737 | ||
3738 | static int __init cfq_slab_setup(void) | |
3739 | { | |
0a31bd5f | 3740 | cfq_pool = KMEM_CACHE(cfq_queue, 0); |
1da177e4 LT |
3741 | if (!cfq_pool) |
3742 | goto fail; | |
3743 | ||
34e6bbf2 | 3744 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); |
1da177e4 LT |
3745 | if (!cfq_ioc_pool) |
3746 | goto fail; | |
3747 | ||
3748 | return 0; | |
3749 | fail: | |
3750 | cfq_slab_kill(); | |
3751 | return -ENOMEM; | |
3752 | } | |
3753 | ||
1da177e4 LT |
3754 | /* |
3755 | * sysfs parts below --> | |
3756 | */ | |
1da177e4 LT |
3757 | static ssize_t |
3758 | cfq_var_show(unsigned int var, char *page) | |
3759 | { | |
3760 | return sprintf(page, "%d\n", var); | |
3761 | } | |
3762 | ||
3763 | static ssize_t | |
3764 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
3765 | { | |
3766 | char *p = (char *) page; | |
3767 | ||
3768 | *var = simple_strtoul(p, &p, 10); | |
3769 | return count; | |
3770 | } | |
3771 | ||
1da177e4 | 3772 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
b374d18a | 3773 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
1da177e4 | 3774 | { \ |
3d1ab40f | 3775 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3776 | unsigned int __data = __VAR; \ |
3777 | if (__CONV) \ | |
3778 | __data = jiffies_to_msecs(__data); \ | |
3779 | return cfq_var_show(__data, (page)); \ | |
3780 | } | |
3781 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
22e2c507 JA |
3782 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); |
3783 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
e572ec7e AV |
3784 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); |
3785 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
22e2c507 JA |
3786 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); |
3787 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
3788 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
3789 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
963b72fc | 3790 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); |
ae30c286 | 3791 | SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0); |
1da177e4 LT |
3792 | #undef SHOW_FUNCTION |
3793 | ||
3794 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
b374d18a | 3795 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
1da177e4 | 3796 | { \ |
3d1ab40f | 3797 | struct cfq_data *cfqd = e->elevator_data; \ |
1da177e4 LT |
3798 | unsigned int __data; \ |
3799 | int ret = cfq_var_store(&__data, (page), count); \ | |
3800 | if (__data < (MIN)) \ | |
3801 | __data = (MIN); \ | |
3802 | else if (__data > (MAX)) \ | |
3803 | __data = (MAX); \ | |
3804 | if (__CONV) \ | |
3805 | *(__PTR) = msecs_to_jiffies(__data); \ | |
3806 | else \ | |
3807 | *(__PTR) = __data; \ | |
3808 | return ret; \ | |
3809 | } | |
3810 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
fe094d98 JA |
3811 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, |
3812 | UINT_MAX, 1); | |
3813 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
3814 | UINT_MAX, 1); | |
e572ec7e | 3815 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); |
fe094d98 JA |
3816 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, |
3817 | UINT_MAX, 0); | |
22e2c507 JA |
3818 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); |
3819 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
3820 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
fe094d98 JA |
3821 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, |
3822 | UINT_MAX, 0); | |
963b72fc | 3823 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); |
ae30c286 | 3824 | STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0); |
1da177e4 LT |
3825 | #undef STORE_FUNCTION |
3826 | ||
e572ec7e AV |
3827 | #define CFQ_ATTR(name) \ |
3828 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
3829 | ||
3830 | static struct elv_fs_entry cfq_attrs[] = { | |
3831 | CFQ_ATTR(quantum), | |
e572ec7e AV |
3832 | CFQ_ATTR(fifo_expire_sync), |
3833 | CFQ_ATTR(fifo_expire_async), | |
3834 | CFQ_ATTR(back_seek_max), | |
3835 | CFQ_ATTR(back_seek_penalty), | |
3836 | CFQ_ATTR(slice_sync), | |
3837 | CFQ_ATTR(slice_async), | |
3838 | CFQ_ATTR(slice_async_rq), | |
3839 | CFQ_ATTR(slice_idle), | |
963b72fc | 3840 | CFQ_ATTR(low_latency), |
ae30c286 | 3841 | CFQ_ATTR(group_isolation), |
e572ec7e | 3842 | __ATTR_NULL |
1da177e4 LT |
3843 | }; |
3844 | ||
1da177e4 LT |
3845 | static struct elevator_type iosched_cfq = { |
3846 | .ops = { | |
3847 | .elevator_merge_fn = cfq_merge, | |
3848 | .elevator_merged_fn = cfq_merged_request, | |
3849 | .elevator_merge_req_fn = cfq_merged_requests, | |
da775265 | 3850 | .elevator_allow_merge_fn = cfq_allow_merge, |
b4878f24 | 3851 | .elevator_dispatch_fn = cfq_dispatch_requests, |
1da177e4 | 3852 | .elevator_add_req_fn = cfq_insert_request, |
b4878f24 | 3853 | .elevator_activate_req_fn = cfq_activate_request, |
1da177e4 LT |
3854 | .elevator_deactivate_req_fn = cfq_deactivate_request, |
3855 | .elevator_queue_empty_fn = cfq_queue_empty, | |
3856 | .elevator_completed_req_fn = cfq_completed_request, | |
21183b07 JA |
3857 | .elevator_former_req_fn = elv_rb_former_request, |
3858 | .elevator_latter_req_fn = elv_rb_latter_request, | |
1da177e4 LT |
3859 | .elevator_set_req_fn = cfq_set_request, |
3860 | .elevator_put_req_fn = cfq_put_request, | |
3861 | .elevator_may_queue_fn = cfq_may_queue, | |
3862 | .elevator_init_fn = cfq_init_queue, | |
3863 | .elevator_exit_fn = cfq_exit_queue, | |
fc46379d | 3864 | .trim = cfq_free_io_context, |
1da177e4 | 3865 | }, |
3d1ab40f | 3866 | .elevator_attrs = cfq_attrs, |
1da177e4 LT |
3867 | .elevator_name = "cfq", |
3868 | .elevator_owner = THIS_MODULE, | |
3869 | }; | |
3870 | ||
3e252066 VG |
3871 | #ifdef CONFIG_CFQ_GROUP_IOSCHED |
3872 | static struct blkio_policy_type blkio_policy_cfq = { | |
3873 | .ops = { | |
3874 | .blkio_unlink_group_fn = cfq_unlink_blkio_group, | |
3875 | .blkio_update_group_weight_fn = cfq_update_blkio_group_weight, | |
3876 | }, | |
3877 | }; | |
3878 | #else | |
3879 | static struct blkio_policy_type blkio_policy_cfq; | |
3880 | #endif | |
3881 | ||
1da177e4 LT |
3882 | static int __init cfq_init(void) |
3883 | { | |
22e2c507 JA |
3884 | /* |
3885 | * could be 0 on HZ < 1000 setups | |
3886 | */ | |
3887 | if (!cfq_slice_async) | |
3888 | cfq_slice_async = 1; | |
3889 | if (!cfq_slice_idle) | |
3890 | cfq_slice_idle = 1; | |
3891 | ||
1da177e4 LT |
3892 | if (cfq_slab_setup()) |
3893 | return -ENOMEM; | |
3894 | ||
2fdd82bd | 3895 | elv_register(&iosched_cfq); |
3e252066 | 3896 | blkio_policy_register(&blkio_policy_cfq); |
1da177e4 | 3897 | |
2fdd82bd | 3898 | return 0; |
1da177e4 LT |
3899 | } |
3900 | ||
3901 | static void __exit cfq_exit(void) | |
3902 | { | |
6e9a4738 | 3903 | DECLARE_COMPLETION_ONSTACK(all_gone); |
3e252066 | 3904 | blkio_policy_unregister(&blkio_policy_cfq); |
1da177e4 | 3905 | elv_unregister(&iosched_cfq); |
334e94de | 3906 | ioc_gone = &all_gone; |
fba82272 OH |
3907 | /* ioc_gone's update must be visible before reading ioc_count */ |
3908 | smp_wmb(); | |
d6de8be7 JA |
3909 | |
3910 | /* | |
3911 | * this also protects us from entering cfq_slab_kill() with | |
3912 | * pending RCU callbacks | |
3913 | */ | |
245b2e70 | 3914 | if (elv_ioc_count_read(cfq_ioc_count)) |
9a11b4ed | 3915 | wait_for_completion(&all_gone); |
83521d3e | 3916 | cfq_slab_kill(); |
1da177e4 LT |
3917 | } |
3918 | ||
3919 | module_init(cfq_init); | |
3920 | module_exit(cfq_exit); | |
3921 | ||
3922 | MODULE_AUTHOR("Jens Axboe"); | |
3923 | MODULE_LICENSE("GPL"); | |
3924 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |