Commit | Line | Data |
---|---|---|
bb44e5d1 IM |
1 | /* |
2 | * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR | |
3 | * policies) | |
4 | */ | |
5 | ||
4fd29176 | 6 | #ifdef CONFIG_SMP |
84de4274 | 7 | |
637f5085 | 8 | static inline int rt_overloaded(struct rq *rq) |
4fd29176 | 9 | { |
637f5085 | 10 | return atomic_read(&rq->rd->rto_count); |
4fd29176 | 11 | } |
84de4274 | 12 | |
4fd29176 SR |
13 | static inline void rt_set_overload(struct rq *rq) |
14 | { | |
1f11eb6a GH |
15 | if (!rq->online) |
16 | return; | |
17 | ||
637f5085 | 18 | cpu_set(rq->cpu, rq->rd->rto_mask); |
4fd29176 SR |
19 | /* |
20 | * Make sure the mask is visible before we set | |
21 | * the overload count. That is checked to determine | |
22 | * if we should look at the mask. It would be a shame | |
23 | * if we looked at the mask, but the mask was not | |
24 | * updated yet. | |
25 | */ | |
26 | wmb(); | |
637f5085 | 27 | atomic_inc(&rq->rd->rto_count); |
4fd29176 | 28 | } |
84de4274 | 29 | |
4fd29176 SR |
30 | static inline void rt_clear_overload(struct rq *rq) |
31 | { | |
1f11eb6a GH |
32 | if (!rq->online) |
33 | return; | |
34 | ||
4fd29176 | 35 | /* the order here really doesn't matter */ |
637f5085 GH |
36 | atomic_dec(&rq->rd->rto_count); |
37 | cpu_clear(rq->cpu, rq->rd->rto_mask); | |
4fd29176 | 38 | } |
73fe6aae GH |
39 | |
40 | static void update_rt_migration(struct rq *rq) | |
41 | { | |
637f5085 | 42 | if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) { |
cdc8eb98 GH |
43 | if (!rq->rt.overloaded) { |
44 | rt_set_overload(rq); | |
45 | rq->rt.overloaded = 1; | |
46 | } | |
47 | } else if (rq->rt.overloaded) { | |
73fe6aae | 48 | rt_clear_overload(rq); |
637f5085 GH |
49 | rq->rt.overloaded = 0; |
50 | } | |
73fe6aae | 51 | } |
4fd29176 SR |
52 | #endif /* CONFIG_SMP */ |
53 | ||
6f505b16 | 54 | static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) |
fa85ae24 | 55 | { |
6f505b16 PZ |
56 | return container_of(rt_se, struct task_struct, rt); |
57 | } | |
58 | ||
59 | static inline int on_rt_rq(struct sched_rt_entity *rt_se) | |
60 | { | |
61 | return !list_empty(&rt_se->run_list); | |
62 | } | |
63 | ||
052f1dc7 | 64 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 65 | |
9f0c1e56 | 66 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) |
6f505b16 PZ |
67 | { |
68 | if (!rt_rq->tg) | |
9f0c1e56 | 69 | return RUNTIME_INF; |
6f505b16 | 70 | |
ac086bc2 PZ |
71 | return rt_rq->rt_runtime; |
72 | } | |
73 | ||
74 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) | |
75 | { | |
76 | return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period); | |
6f505b16 PZ |
77 | } |
78 | ||
79 | #define for_each_leaf_rt_rq(rt_rq, rq) \ | |
80 | list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list) | |
81 | ||
82 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) | |
83 | { | |
84 | return rt_rq->rq; | |
85 | } | |
86 | ||
87 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) | |
88 | { | |
89 | return rt_se->rt_rq; | |
90 | } | |
91 | ||
92 | #define for_each_sched_rt_entity(rt_se) \ | |
93 | for (; rt_se; rt_se = rt_se->parent) | |
94 | ||
95 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) | |
96 | { | |
97 | return rt_se->my_q; | |
98 | } | |
99 | ||
100 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se); | |
101 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se); | |
102 | ||
9f0c1e56 | 103 | static void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
6f505b16 | 104 | { |
f6121f4f | 105 | struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr; |
6f505b16 PZ |
106 | struct sched_rt_entity *rt_se = rt_rq->rt_se; |
107 | ||
f6121f4f DF |
108 | if (rt_rq->rt_nr_running) { |
109 | if (rt_se && !on_rt_rq(rt_se)) | |
110 | enqueue_rt_entity(rt_se); | |
1020387f PZ |
111 | if (rt_rq->highest_prio < curr->prio) |
112 | resched_task(curr); | |
6f505b16 PZ |
113 | } |
114 | } | |
115 | ||
9f0c1e56 | 116 | static void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
6f505b16 PZ |
117 | { |
118 | struct sched_rt_entity *rt_se = rt_rq->rt_se; | |
119 | ||
120 | if (rt_se && on_rt_rq(rt_se)) | |
121 | dequeue_rt_entity(rt_se); | |
122 | } | |
123 | ||
23b0fdfc PZ |
124 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
125 | { | |
126 | return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted; | |
127 | } | |
128 | ||
129 | static int rt_se_boosted(struct sched_rt_entity *rt_se) | |
130 | { | |
131 | struct rt_rq *rt_rq = group_rt_rq(rt_se); | |
132 | struct task_struct *p; | |
133 | ||
134 | if (rt_rq) | |
135 | return !!rt_rq->rt_nr_boosted; | |
136 | ||
137 | p = rt_task_of(rt_se); | |
138 | return p->prio != p->normal_prio; | |
139 | } | |
140 | ||
d0b27fa7 PZ |
141 | #ifdef CONFIG_SMP |
142 | static inline cpumask_t sched_rt_period_mask(void) | |
143 | { | |
144 | return cpu_rq(smp_processor_id())->rd->span; | |
145 | } | |
6f505b16 | 146 | #else |
d0b27fa7 PZ |
147 | static inline cpumask_t sched_rt_period_mask(void) |
148 | { | |
149 | return cpu_online_map; | |
150 | } | |
151 | #endif | |
6f505b16 | 152 | |
d0b27fa7 PZ |
153 | static inline |
154 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) | |
6f505b16 | 155 | { |
d0b27fa7 PZ |
156 | return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu]; |
157 | } | |
9f0c1e56 | 158 | |
ac086bc2 PZ |
159 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
160 | { | |
161 | return &rt_rq->tg->rt_bandwidth; | |
162 | } | |
163 | ||
55e12e5e | 164 | #else /* !CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 PZ |
165 | |
166 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) | |
167 | { | |
ac086bc2 PZ |
168 | return rt_rq->rt_runtime; |
169 | } | |
170 | ||
171 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) | |
172 | { | |
173 | return ktime_to_ns(def_rt_bandwidth.rt_period); | |
6f505b16 PZ |
174 | } |
175 | ||
176 | #define for_each_leaf_rt_rq(rt_rq, rq) \ | |
177 | for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL) | |
178 | ||
179 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) | |
180 | { | |
181 | return container_of(rt_rq, struct rq, rt); | |
182 | } | |
183 | ||
184 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) | |
185 | { | |
186 | struct task_struct *p = rt_task_of(rt_se); | |
187 | struct rq *rq = task_rq(p); | |
188 | ||
189 | return &rq->rt; | |
190 | } | |
191 | ||
192 | #define for_each_sched_rt_entity(rt_se) \ | |
193 | for (; rt_se; rt_se = NULL) | |
194 | ||
195 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) | |
196 | { | |
197 | return NULL; | |
198 | } | |
199 | ||
9f0c1e56 | 200 | static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
6f505b16 | 201 | { |
f3ade837 JB |
202 | if (rt_rq->rt_nr_running) |
203 | resched_task(rq_of_rt_rq(rt_rq)->curr); | |
6f505b16 PZ |
204 | } |
205 | ||
9f0c1e56 | 206 | static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
6f505b16 PZ |
207 | { |
208 | } | |
209 | ||
23b0fdfc PZ |
210 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
211 | { | |
212 | return rt_rq->rt_throttled; | |
213 | } | |
d0b27fa7 PZ |
214 | |
215 | static inline cpumask_t sched_rt_period_mask(void) | |
216 | { | |
217 | return cpu_online_map; | |
218 | } | |
219 | ||
220 | static inline | |
221 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) | |
222 | { | |
223 | return &cpu_rq(cpu)->rt; | |
224 | } | |
225 | ||
ac086bc2 PZ |
226 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
227 | { | |
228 | return &def_rt_bandwidth; | |
229 | } | |
230 | ||
55e12e5e | 231 | #endif /* CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 | 232 | |
ac086bc2 | 233 | #ifdef CONFIG_SMP |
78333cdd PZ |
234 | /* |
235 | * We ran out of runtime, see if we can borrow some from our neighbours. | |
236 | */ | |
b79f3833 | 237 | static int do_balance_runtime(struct rt_rq *rt_rq) |
ac086bc2 PZ |
238 | { |
239 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); | |
240 | struct root_domain *rd = cpu_rq(smp_processor_id())->rd; | |
241 | int i, weight, more = 0; | |
242 | u64 rt_period; | |
243 | ||
244 | weight = cpus_weight(rd->span); | |
245 | ||
246 | spin_lock(&rt_b->rt_runtime_lock); | |
247 | rt_period = ktime_to_ns(rt_b->rt_period); | |
363ab6f1 | 248 | for_each_cpu_mask_nr(i, rd->span) { |
ac086bc2 PZ |
249 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); |
250 | s64 diff; | |
251 | ||
252 | if (iter == rt_rq) | |
253 | continue; | |
254 | ||
255 | spin_lock(&iter->rt_runtime_lock); | |
78333cdd PZ |
256 | /* |
257 | * Either all rqs have inf runtime and there's nothing to steal | |
258 | * or __disable_runtime() below sets a specific rq to inf to | |
259 | * indicate its been disabled and disalow stealing. | |
260 | */ | |
7def2be1 PZ |
261 | if (iter->rt_runtime == RUNTIME_INF) |
262 | goto next; | |
263 | ||
78333cdd PZ |
264 | /* |
265 | * From runqueues with spare time, take 1/n part of their | |
266 | * spare time, but no more than our period. | |
267 | */ | |
ac086bc2 PZ |
268 | diff = iter->rt_runtime - iter->rt_time; |
269 | if (diff > 0) { | |
58838cf3 | 270 | diff = div_u64((u64)diff, weight); |
ac086bc2 PZ |
271 | if (rt_rq->rt_runtime + diff > rt_period) |
272 | diff = rt_period - rt_rq->rt_runtime; | |
273 | iter->rt_runtime -= diff; | |
274 | rt_rq->rt_runtime += diff; | |
275 | more = 1; | |
276 | if (rt_rq->rt_runtime == rt_period) { | |
277 | spin_unlock(&iter->rt_runtime_lock); | |
278 | break; | |
279 | } | |
280 | } | |
7def2be1 | 281 | next: |
ac086bc2 PZ |
282 | spin_unlock(&iter->rt_runtime_lock); |
283 | } | |
284 | spin_unlock(&rt_b->rt_runtime_lock); | |
285 | ||
286 | return more; | |
287 | } | |
7def2be1 | 288 | |
78333cdd PZ |
289 | /* |
290 | * Ensure this RQ takes back all the runtime it lend to its neighbours. | |
291 | */ | |
7def2be1 PZ |
292 | static void __disable_runtime(struct rq *rq) |
293 | { | |
294 | struct root_domain *rd = rq->rd; | |
295 | struct rt_rq *rt_rq; | |
296 | ||
297 | if (unlikely(!scheduler_running)) | |
298 | return; | |
299 | ||
300 | for_each_leaf_rt_rq(rt_rq, rq) { | |
301 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); | |
302 | s64 want; | |
303 | int i; | |
304 | ||
305 | spin_lock(&rt_b->rt_runtime_lock); | |
306 | spin_lock(&rt_rq->rt_runtime_lock); | |
78333cdd PZ |
307 | /* |
308 | * Either we're all inf and nobody needs to borrow, or we're | |
309 | * already disabled and thus have nothing to do, or we have | |
310 | * exactly the right amount of runtime to take out. | |
311 | */ | |
7def2be1 PZ |
312 | if (rt_rq->rt_runtime == RUNTIME_INF || |
313 | rt_rq->rt_runtime == rt_b->rt_runtime) | |
314 | goto balanced; | |
315 | spin_unlock(&rt_rq->rt_runtime_lock); | |
316 | ||
78333cdd PZ |
317 | /* |
318 | * Calculate the difference between what we started out with | |
319 | * and what we current have, that's the amount of runtime | |
320 | * we lend and now have to reclaim. | |
321 | */ | |
7def2be1 PZ |
322 | want = rt_b->rt_runtime - rt_rq->rt_runtime; |
323 | ||
78333cdd PZ |
324 | /* |
325 | * Greedy reclaim, take back as much as we can. | |
326 | */ | |
7def2be1 PZ |
327 | for_each_cpu_mask(i, rd->span) { |
328 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); | |
329 | s64 diff; | |
330 | ||
78333cdd PZ |
331 | /* |
332 | * Can't reclaim from ourselves or disabled runqueues. | |
333 | */ | |
f1679d08 | 334 | if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF) |
7def2be1 PZ |
335 | continue; |
336 | ||
337 | spin_lock(&iter->rt_runtime_lock); | |
338 | if (want > 0) { | |
339 | diff = min_t(s64, iter->rt_runtime, want); | |
340 | iter->rt_runtime -= diff; | |
341 | want -= diff; | |
342 | } else { | |
343 | iter->rt_runtime -= want; | |
344 | want -= want; | |
345 | } | |
346 | spin_unlock(&iter->rt_runtime_lock); | |
347 | ||
348 | if (!want) | |
349 | break; | |
350 | } | |
351 | ||
352 | spin_lock(&rt_rq->rt_runtime_lock); | |
78333cdd PZ |
353 | /* |
354 | * We cannot be left wanting - that would mean some runtime | |
355 | * leaked out of the system. | |
356 | */ | |
7def2be1 PZ |
357 | BUG_ON(want); |
358 | balanced: | |
78333cdd PZ |
359 | /* |
360 | * Disable all the borrow logic by pretending we have inf | |
361 | * runtime - in which case borrowing doesn't make sense. | |
362 | */ | |
7def2be1 PZ |
363 | rt_rq->rt_runtime = RUNTIME_INF; |
364 | spin_unlock(&rt_rq->rt_runtime_lock); | |
365 | spin_unlock(&rt_b->rt_runtime_lock); | |
366 | } | |
367 | } | |
368 | ||
369 | static void disable_runtime(struct rq *rq) | |
370 | { | |
371 | unsigned long flags; | |
372 | ||
373 | spin_lock_irqsave(&rq->lock, flags); | |
374 | __disable_runtime(rq); | |
375 | spin_unlock_irqrestore(&rq->lock, flags); | |
376 | } | |
377 | ||
378 | static void __enable_runtime(struct rq *rq) | |
379 | { | |
7def2be1 PZ |
380 | struct rt_rq *rt_rq; |
381 | ||
382 | if (unlikely(!scheduler_running)) | |
383 | return; | |
384 | ||
78333cdd PZ |
385 | /* |
386 | * Reset each runqueue's bandwidth settings | |
387 | */ | |
7def2be1 PZ |
388 | for_each_leaf_rt_rq(rt_rq, rq) { |
389 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); | |
390 | ||
391 | spin_lock(&rt_b->rt_runtime_lock); | |
392 | spin_lock(&rt_rq->rt_runtime_lock); | |
393 | rt_rq->rt_runtime = rt_b->rt_runtime; | |
394 | rt_rq->rt_time = 0; | |
baf25731 | 395 | rt_rq->rt_throttled = 0; |
7def2be1 PZ |
396 | spin_unlock(&rt_rq->rt_runtime_lock); |
397 | spin_unlock(&rt_b->rt_runtime_lock); | |
398 | } | |
399 | } | |
400 | ||
401 | static void enable_runtime(struct rq *rq) | |
402 | { | |
403 | unsigned long flags; | |
404 | ||
405 | spin_lock_irqsave(&rq->lock, flags); | |
406 | __enable_runtime(rq); | |
407 | spin_unlock_irqrestore(&rq->lock, flags); | |
408 | } | |
409 | ||
eff6549b PZ |
410 | static int balance_runtime(struct rt_rq *rt_rq) |
411 | { | |
412 | int more = 0; | |
413 | ||
414 | if (rt_rq->rt_time > rt_rq->rt_runtime) { | |
415 | spin_unlock(&rt_rq->rt_runtime_lock); | |
416 | more = do_balance_runtime(rt_rq); | |
417 | spin_lock(&rt_rq->rt_runtime_lock); | |
418 | } | |
419 | ||
420 | return more; | |
421 | } | |
55e12e5e | 422 | #else /* !CONFIG_SMP */ |
eff6549b PZ |
423 | static inline int balance_runtime(struct rt_rq *rt_rq) |
424 | { | |
425 | return 0; | |
426 | } | |
55e12e5e | 427 | #endif /* CONFIG_SMP */ |
ac086bc2 | 428 | |
eff6549b PZ |
429 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun) |
430 | { | |
431 | int i, idle = 1; | |
432 | cpumask_t span; | |
433 | ||
0b148fa0 | 434 | if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF) |
eff6549b PZ |
435 | return 1; |
436 | ||
437 | span = sched_rt_period_mask(); | |
438 | for_each_cpu_mask(i, span) { | |
439 | int enqueue = 0; | |
440 | struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i); | |
441 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
442 | ||
443 | spin_lock(&rq->lock); | |
444 | if (rt_rq->rt_time) { | |
445 | u64 runtime; | |
446 | ||
447 | spin_lock(&rt_rq->rt_runtime_lock); | |
448 | if (rt_rq->rt_throttled) | |
449 | balance_runtime(rt_rq); | |
450 | runtime = rt_rq->rt_runtime; | |
451 | rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime); | |
452 | if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) { | |
453 | rt_rq->rt_throttled = 0; | |
454 | enqueue = 1; | |
455 | } | |
456 | if (rt_rq->rt_time || rt_rq->rt_nr_running) | |
457 | idle = 0; | |
458 | spin_unlock(&rt_rq->rt_runtime_lock); | |
6c3df255 PZ |
459 | } else if (rt_rq->rt_nr_running) |
460 | idle = 0; | |
eff6549b PZ |
461 | |
462 | if (enqueue) | |
463 | sched_rt_rq_enqueue(rt_rq); | |
464 | spin_unlock(&rq->lock); | |
465 | } | |
466 | ||
467 | return idle; | |
468 | } | |
ac086bc2 | 469 | |
6f505b16 PZ |
470 | static inline int rt_se_prio(struct sched_rt_entity *rt_se) |
471 | { | |
052f1dc7 | 472 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 PZ |
473 | struct rt_rq *rt_rq = group_rt_rq(rt_se); |
474 | ||
475 | if (rt_rq) | |
476 | return rt_rq->highest_prio; | |
477 | #endif | |
478 | ||
479 | return rt_task_of(rt_se)->prio; | |
480 | } | |
481 | ||
9f0c1e56 | 482 | static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq) |
6f505b16 | 483 | { |
9f0c1e56 | 484 | u64 runtime = sched_rt_runtime(rt_rq); |
fa85ae24 | 485 | |
fa85ae24 | 486 | if (rt_rq->rt_throttled) |
23b0fdfc | 487 | return rt_rq_throttled(rt_rq); |
fa85ae24 | 488 | |
ac086bc2 PZ |
489 | if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq)) |
490 | return 0; | |
491 | ||
b79f3833 PZ |
492 | balance_runtime(rt_rq); |
493 | runtime = sched_rt_runtime(rt_rq); | |
494 | if (runtime == RUNTIME_INF) | |
495 | return 0; | |
ac086bc2 | 496 | |
9f0c1e56 | 497 | if (rt_rq->rt_time > runtime) { |
6f505b16 | 498 | rt_rq->rt_throttled = 1; |
23b0fdfc | 499 | if (rt_rq_throttled(rt_rq)) { |
9f0c1e56 | 500 | sched_rt_rq_dequeue(rt_rq); |
23b0fdfc PZ |
501 | return 1; |
502 | } | |
fa85ae24 PZ |
503 | } |
504 | ||
505 | return 0; | |
506 | } | |
507 | ||
bb44e5d1 IM |
508 | /* |
509 | * Update the current task's runtime statistics. Skip current tasks that | |
510 | * are not in our scheduling class. | |
511 | */ | |
a9957449 | 512 | static void update_curr_rt(struct rq *rq) |
bb44e5d1 IM |
513 | { |
514 | struct task_struct *curr = rq->curr; | |
6f505b16 PZ |
515 | struct sched_rt_entity *rt_se = &curr->rt; |
516 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); | |
bb44e5d1 IM |
517 | u64 delta_exec; |
518 | ||
519 | if (!task_has_rt_policy(curr)) | |
520 | return; | |
521 | ||
d281918d | 522 | delta_exec = rq->clock - curr->se.exec_start; |
bb44e5d1 IM |
523 | if (unlikely((s64)delta_exec < 0)) |
524 | delta_exec = 0; | |
6cfb0d5d IM |
525 | |
526 | schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec)); | |
bb44e5d1 IM |
527 | |
528 | curr->se.sum_exec_runtime += delta_exec; | |
f06febc9 FM |
529 | account_group_exec_runtime(curr, delta_exec); |
530 | ||
d281918d | 531 | curr->se.exec_start = rq->clock; |
d842de87 | 532 | cpuacct_charge(curr, delta_exec); |
fa85ae24 | 533 | |
0b148fa0 PZ |
534 | if (!rt_bandwidth_enabled()) |
535 | return; | |
536 | ||
354d60c2 DG |
537 | for_each_sched_rt_entity(rt_se) { |
538 | rt_rq = rt_rq_of_se(rt_se); | |
539 | ||
540 | spin_lock(&rt_rq->rt_runtime_lock); | |
cc2991cf PZ |
541 | if (sched_rt_runtime(rt_rq) != RUNTIME_INF) { |
542 | rt_rq->rt_time += delta_exec; | |
543 | if (sched_rt_runtime_exceeded(rt_rq)) | |
544 | resched_task(curr); | |
545 | } | |
354d60c2 DG |
546 | spin_unlock(&rt_rq->rt_runtime_lock); |
547 | } | |
bb44e5d1 IM |
548 | } |
549 | ||
6f505b16 PZ |
550 | static inline |
551 | void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
63489e45 | 552 | { |
6f505b16 PZ |
553 | WARN_ON(!rt_prio(rt_se_prio(rt_se))); |
554 | rt_rq->rt_nr_running++; | |
052f1dc7 | 555 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
6e0534f2 | 556 | if (rt_se_prio(rt_se) < rt_rq->highest_prio) { |
577b4a58 | 557 | #ifdef CONFIG_SMP |
6e0534f2 | 558 | struct rq *rq = rq_of_rt_rq(rt_rq); |
577b4a58 | 559 | #endif |
1f11eb6a | 560 | |
6f505b16 | 561 | rt_rq->highest_prio = rt_se_prio(rt_se); |
1100ac91 | 562 | #ifdef CONFIG_SMP |
1f11eb6a GH |
563 | if (rq->online) |
564 | cpupri_set(&rq->rd->cpupri, rq->cpu, | |
565 | rt_se_prio(rt_se)); | |
1100ac91 | 566 | #endif |
6e0534f2 | 567 | } |
6f505b16 | 568 | #endif |
764a9d6f | 569 | #ifdef CONFIG_SMP |
6f505b16 PZ |
570 | if (rt_se->nr_cpus_allowed > 1) { |
571 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
1100ac91 | 572 | |
73fe6aae | 573 | rq->rt.rt_nr_migratory++; |
6f505b16 | 574 | } |
73fe6aae | 575 | |
6f505b16 PZ |
576 | update_rt_migration(rq_of_rt_rq(rt_rq)); |
577 | #endif | |
052f1dc7 | 578 | #ifdef CONFIG_RT_GROUP_SCHED |
23b0fdfc PZ |
579 | if (rt_se_boosted(rt_se)) |
580 | rt_rq->rt_nr_boosted++; | |
d0b27fa7 PZ |
581 | |
582 | if (rt_rq->tg) | |
583 | start_rt_bandwidth(&rt_rq->tg->rt_bandwidth); | |
584 | #else | |
585 | start_rt_bandwidth(&def_rt_bandwidth); | |
23b0fdfc | 586 | #endif |
63489e45 SR |
587 | } |
588 | ||
6f505b16 PZ |
589 | static inline |
590 | void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
63489e45 | 591 | { |
6e0534f2 GH |
592 | #ifdef CONFIG_SMP |
593 | int highest_prio = rt_rq->highest_prio; | |
594 | #endif | |
595 | ||
6f505b16 PZ |
596 | WARN_ON(!rt_prio(rt_se_prio(rt_se))); |
597 | WARN_ON(!rt_rq->rt_nr_running); | |
598 | rt_rq->rt_nr_running--; | |
052f1dc7 | 599 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
6f505b16 | 600 | if (rt_rq->rt_nr_running) { |
764a9d6f SR |
601 | struct rt_prio_array *array; |
602 | ||
6f505b16 PZ |
603 | WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio); |
604 | if (rt_se_prio(rt_se) == rt_rq->highest_prio) { | |
764a9d6f | 605 | /* recalculate */ |
6f505b16 PZ |
606 | array = &rt_rq->active; |
607 | rt_rq->highest_prio = | |
764a9d6f SR |
608 | sched_find_first_bit(array->bitmap); |
609 | } /* otherwise leave rq->highest prio alone */ | |
610 | } else | |
6f505b16 PZ |
611 | rt_rq->highest_prio = MAX_RT_PRIO; |
612 | #endif | |
613 | #ifdef CONFIG_SMP | |
614 | if (rt_se->nr_cpus_allowed > 1) { | |
615 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
73fe6aae | 616 | rq->rt.rt_nr_migratory--; |
6f505b16 | 617 | } |
73fe6aae | 618 | |
6e0534f2 GH |
619 | if (rt_rq->highest_prio != highest_prio) { |
620 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
1f11eb6a GH |
621 | |
622 | if (rq->online) | |
623 | cpupri_set(&rq->rd->cpupri, rq->cpu, | |
624 | rt_rq->highest_prio); | |
6e0534f2 GH |
625 | } |
626 | ||
6f505b16 | 627 | update_rt_migration(rq_of_rt_rq(rt_rq)); |
764a9d6f | 628 | #endif /* CONFIG_SMP */ |
052f1dc7 | 629 | #ifdef CONFIG_RT_GROUP_SCHED |
23b0fdfc PZ |
630 | if (rt_se_boosted(rt_se)) |
631 | rt_rq->rt_nr_boosted--; | |
632 | ||
633 | WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted); | |
634 | #endif | |
63489e45 SR |
635 | } |
636 | ||
ad2a3f13 | 637 | static void __enqueue_rt_entity(struct sched_rt_entity *rt_se) |
bb44e5d1 | 638 | { |
6f505b16 PZ |
639 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); |
640 | struct rt_prio_array *array = &rt_rq->active; | |
641 | struct rt_rq *group_rq = group_rt_rq(rt_se); | |
20b6331b | 642 | struct list_head *queue = array->queue + rt_se_prio(rt_se); |
bb44e5d1 | 643 | |
ad2a3f13 PZ |
644 | /* |
645 | * Don't enqueue the group if its throttled, or when empty. | |
646 | * The latter is a consequence of the former when a child group | |
647 | * get throttled and the current group doesn't have any other | |
648 | * active members. | |
649 | */ | |
650 | if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) | |
6f505b16 | 651 | return; |
63489e45 | 652 | |
7ebefa8c | 653 | list_add_tail(&rt_se->run_list, queue); |
6f505b16 | 654 | __set_bit(rt_se_prio(rt_se), array->bitmap); |
78f2c7db | 655 | |
6f505b16 PZ |
656 | inc_rt_tasks(rt_se, rt_rq); |
657 | } | |
658 | ||
ad2a3f13 | 659 | static void __dequeue_rt_entity(struct sched_rt_entity *rt_se) |
6f505b16 PZ |
660 | { |
661 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); | |
662 | struct rt_prio_array *array = &rt_rq->active; | |
663 | ||
664 | list_del_init(&rt_se->run_list); | |
665 | if (list_empty(array->queue + rt_se_prio(rt_se))) | |
666 | __clear_bit(rt_se_prio(rt_se), array->bitmap); | |
667 | ||
668 | dec_rt_tasks(rt_se, rt_rq); | |
669 | } | |
670 | ||
671 | /* | |
672 | * Because the prio of an upper entry depends on the lower | |
673 | * entries, we must remove entries top - down. | |
6f505b16 | 674 | */ |
ad2a3f13 | 675 | static void dequeue_rt_stack(struct sched_rt_entity *rt_se) |
6f505b16 | 676 | { |
ad2a3f13 | 677 | struct sched_rt_entity *back = NULL; |
6f505b16 | 678 | |
58d6c2d7 PZ |
679 | for_each_sched_rt_entity(rt_se) { |
680 | rt_se->back = back; | |
681 | back = rt_se; | |
682 | } | |
683 | ||
684 | for (rt_se = back; rt_se; rt_se = rt_se->back) { | |
685 | if (on_rt_rq(rt_se)) | |
ad2a3f13 PZ |
686 | __dequeue_rt_entity(rt_se); |
687 | } | |
688 | } | |
689 | ||
690 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se) | |
691 | { | |
692 | dequeue_rt_stack(rt_se); | |
693 | for_each_sched_rt_entity(rt_se) | |
694 | __enqueue_rt_entity(rt_se); | |
695 | } | |
696 | ||
697 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se) | |
698 | { | |
699 | dequeue_rt_stack(rt_se); | |
700 | ||
701 | for_each_sched_rt_entity(rt_se) { | |
702 | struct rt_rq *rt_rq = group_rt_rq(rt_se); | |
703 | ||
704 | if (rt_rq && rt_rq->rt_nr_running) | |
705 | __enqueue_rt_entity(rt_se); | |
58d6c2d7 | 706 | } |
bb44e5d1 IM |
707 | } |
708 | ||
709 | /* | |
710 | * Adding/removing a task to/from a priority array: | |
711 | */ | |
6f505b16 PZ |
712 | static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup) |
713 | { | |
714 | struct sched_rt_entity *rt_se = &p->rt; | |
715 | ||
716 | if (wakeup) | |
717 | rt_se->timeout = 0; | |
718 | ||
ad2a3f13 | 719 | enqueue_rt_entity(rt_se); |
c09595f6 PZ |
720 | |
721 | inc_cpu_load(rq, p->se.load.weight); | |
6f505b16 PZ |
722 | } |
723 | ||
f02231e5 | 724 | static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep) |
bb44e5d1 | 725 | { |
6f505b16 | 726 | struct sched_rt_entity *rt_se = &p->rt; |
bb44e5d1 | 727 | |
f1e14ef6 | 728 | update_curr_rt(rq); |
ad2a3f13 | 729 | dequeue_rt_entity(rt_se); |
c09595f6 PZ |
730 | |
731 | dec_cpu_load(rq, p->se.load.weight); | |
bb44e5d1 IM |
732 | } |
733 | ||
734 | /* | |
735 | * Put task to the end of the run list without the overhead of dequeue | |
736 | * followed by enqueue. | |
737 | */ | |
7ebefa8c DA |
738 | static void |
739 | requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head) | |
6f505b16 | 740 | { |
1cdad715 | 741 | if (on_rt_rq(rt_se)) { |
7ebefa8c DA |
742 | struct rt_prio_array *array = &rt_rq->active; |
743 | struct list_head *queue = array->queue + rt_se_prio(rt_se); | |
744 | ||
745 | if (head) | |
746 | list_move(&rt_se->run_list, queue); | |
747 | else | |
748 | list_move_tail(&rt_se->run_list, queue); | |
1cdad715 | 749 | } |
6f505b16 PZ |
750 | } |
751 | ||
7ebefa8c | 752 | static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head) |
bb44e5d1 | 753 | { |
6f505b16 PZ |
754 | struct sched_rt_entity *rt_se = &p->rt; |
755 | struct rt_rq *rt_rq; | |
bb44e5d1 | 756 | |
6f505b16 PZ |
757 | for_each_sched_rt_entity(rt_se) { |
758 | rt_rq = rt_rq_of_se(rt_se); | |
7ebefa8c | 759 | requeue_rt_entity(rt_rq, rt_se, head); |
6f505b16 | 760 | } |
bb44e5d1 IM |
761 | } |
762 | ||
6f505b16 | 763 | static void yield_task_rt(struct rq *rq) |
bb44e5d1 | 764 | { |
7ebefa8c | 765 | requeue_task_rt(rq, rq->curr, 0); |
bb44e5d1 IM |
766 | } |
767 | ||
e7693a36 | 768 | #ifdef CONFIG_SMP |
318e0893 GH |
769 | static int find_lowest_rq(struct task_struct *task); |
770 | ||
e7693a36 GH |
771 | static int select_task_rq_rt(struct task_struct *p, int sync) |
772 | { | |
318e0893 GH |
773 | struct rq *rq = task_rq(p); |
774 | ||
775 | /* | |
e1f47d89 SR |
776 | * If the current task is an RT task, then |
777 | * try to see if we can wake this RT task up on another | |
778 | * runqueue. Otherwise simply start this RT task | |
779 | * on its current runqueue. | |
780 | * | |
781 | * We want to avoid overloading runqueues. Even if | |
782 | * the RT task is of higher priority than the current RT task. | |
783 | * RT tasks behave differently than other tasks. If | |
784 | * one gets preempted, we try to push it off to another queue. | |
785 | * So trying to keep a preempting RT task on the same | |
786 | * cache hot CPU will force the running RT task to | |
787 | * a cold CPU. So we waste all the cache for the lower | |
788 | * RT task in hopes of saving some of a RT task | |
789 | * that is just being woken and probably will have | |
790 | * cold cache anyway. | |
318e0893 | 791 | */ |
17b3279b | 792 | if (unlikely(rt_task(rq->curr)) && |
6f505b16 | 793 | (p->rt.nr_cpus_allowed > 1)) { |
318e0893 GH |
794 | int cpu = find_lowest_rq(p); |
795 | ||
796 | return (cpu == -1) ? task_cpu(p) : cpu; | |
797 | } | |
798 | ||
799 | /* | |
800 | * Otherwise, just let it ride on the affined RQ and the | |
801 | * post-schedule router will push the preempted task away | |
802 | */ | |
e7693a36 GH |
803 | return task_cpu(p); |
804 | } | |
7ebefa8c DA |
805 | |
806 | static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p) | |
807 | { | |
808 | cpumask_t mask; | |
809 | ||
810 | if (rq->curr->rt.nr_cpus_allowed == 1) | |
811 | return; | |
812 | ||
813 | if (p->rt.nr_cpus_allowed != 1 | |
814 | && cpupri_find(&rq->rd->cpupri, p, &mask)) | |
815 | return; | |
816 | ||
817 | if (!cpupri_find(&rq->rd->cpupri, rq->curr, &mask)) | |
818 | return; | |
819 | ||
820 | /* | |
821 | * There appears to be other cpus that can accept | |
822 | * current and none to run 'p', so lets reschedule | |
823 | * to try and push current away: | |
824 | */ | |
825 | requeue_task_rt(rq, p, 1); | |
826 | resched_task(rq->curr); | |
827 | } | |
828 | ||
e7693a36 GH |
829 | #endif /* CONFIG_SMP */ |
830 | ||
bb44e5d1 IM |
831 | /* |
832 | * Preempt the current task with a newly woken task if needed: | |
833 | */ | |
15afe09b | 834 | static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync) |
bb44e5d1 | 835 | { |
45c01e82 | 836 | if (p->prio < rq->curr->prio) { |
bb44e5d1 | 837 | resched_task(rq->curr); |
45c01e82 GH |
838 | return; |
839 | } | |
840 | ||
841 | #ifdef CONFIG_SMP | |
842 | /* | |
843 | * If: | |
844 | * | |
845 | * - the newly woken task is of equal priority to the current task | |
846 | * - the newly woken task is non-migratable while current is migratable | |
847 | * - current will be preempted on the next reschedule | |
848 | * | |
849 | * we should check to see if current can readily move to a different | |
850 | * cpu. If so, we will reschedule to allow the push logic to try | |
851 | * to move current somewhere else, making room for our non-migratable | |
852 | * task. | |
853 | */ | |
7ebefa8c DA |
854 | if (p->prio == rq->curr->prio && !need_resched()) |
855 | check_preempt_equal_prio(rq, p); | |
45c01e82 | 856 | #endif |
bb44e5d1 IM |
857 | } |
858 | ||
6f505b16 PZ |
859 | static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq, |
860 | struct rt_rq *rt_rq) | |
bb44e5d1 | 861 | { |
6f505b16 PZ |
862 | struct rt_prio_array *array = &rt_rq->active; |
863 | struct sched_rt_entity *next = NULL; | |
bb44e5d1 IM |
864 | struct list_head *queue; |
865 | int idx; | |
866 | ||
867 | idx = sched_find_first_bit(array->bitmap); | |
6f505b16 | 868 | BUG_ON(idx >= MAX_RT_PRIO); |
bb44e5d1 IM |
869 | |
870 | queue = array->queue + idx; | |
6f505b16 | 871 | next = list_entry(queue->next, struct sched_rt_entity, run_list); |
326587b8 | 872 | |
6f505b16 PZ |
873 | return next; |
874 | } | |
bb44e5d1 | 875 | |
6f505b16 PZ |
876 | static struct task_struct *pick_next_task_rt(struct rq *rq) |
877 | { | |
878 | struct sched_rt_entity *rt_se; | |
879 | struct task_struct *p; | |
880 | struct rt_rq *rt_rq; | |
bb44e5d1 | 881 | |
6f505b16 PZ |
882 | rt_rq = &rq->rt; |
883 | ||
884 | if (unlikely(!rt_rq->rt_nr_running)) | |
885 | return NULL; | |
886 | ||
23b0fdfc | 887 | if (rt_rq_throttled(rt_rq)) |
6f505b16 PZ |
888 | return NULL; |
889 | ||
890 | do { | |
891 | rt_se = pick_next_rt_entity(rq, rt_rq); | |
326587b8 | 892 | BUG_ON(!rt_se); |
6f505b16 PZ |
893 | rt_rq = group_rt_rq(rt_se); |
894 | } while (rt_rq); | |
895 | ||
896 | p = rt_task_of(rt_se); | |
897 | p->se.exec_start = rq->clock; | |
898 | return p; | |
bb44e5d1 IM |
899 | } |
900 | ||
31ee529c | 901 | static void put_prev_task_rt(struct rq *rq, struct task_struct *p) |
bb44e5d1 | 902 | { |
f1e14ef6 | 903 | update_curr_rt(rq); |
bb44e5d1 IM |
904 | p->se.exec_start = 0; |
905 | } | |
906 | ||
681f3e68 | 907 | #ifdef CONFIG_SMP |
6f505b16 | 908 | |
e8fa1362 SR |
909 | /* Only try algorithms three times */ |
910 | #define RT_MAX_TRIES 3 | |
911 | ||
912 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest); | |
1b12bbc7 PZ |
913 | static void double_unlock_balance(struct rq *this_rq, struct rq *busiest); |
914 | ||
e8fa1362 SR |
915 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep); |
916 | ||
f65eda4f SR |
917 | static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) |
918 | { | |
919 | if (!task_running(rq, p) && | |
73fe6aae | 920 | (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) && |
6f505b16 | 921 | (p->rt.nr_cpus_allowed > 1)) |
f65eda4f SR |
922 | return 1; |
923 | return 0; | |
924 | } | |
925 | ||
e8fa1362 | 926 | /* Return the second highest RT task, NULL otherwise */ |
79064fbf | 927 | static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu) |
e8fa1362 | 928 | { |
6f505b16 PZ |
929 | struct task_struct *next = NULL; |
930 | struct sched_rt_entity *rt_se; | |
931 | struct rt_prio_array *array; | |
932 | struct rt_rq *rt_rq; | |
e8fa1362 SR |
933 | int idx; |
934 | ||
6f505b16 PZ |
935 | for_each_leaf_rt_rq(rt_rq, rq) { |
936 | array = &rt_rq->active; | |
937 | idx = sched_find_first_bit(array->bitmap); | |
938 | next_idx: | |
939 | if (idx >= MAX_RT_PRIO) | |
940 | continue; | |
941 | if (next && next->prio < idx) | |
942 | continue; | |
943 | list_for_each_entry(rt_se, array->queue + idx, run_list) { | |
944 | struct task_struct *p = rt_task_of(rt_se); | |
945 | if (pick_rt_task(rq, p, cpu)) { | |
946 | next = p; | |
947 | break; | |
948 | } | |
949 | } | |
950 | if (!next) { | |
951 | idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1); | |
952 | goto next_idx; | |
953 | } | |
f65eda4f SR |
954 | } |
955 | ||
e8fa1362 SR |
956 | return next; |
957 | } | |
958 | ||
959 | static DEFINE_PER_CPU(cpumask_t, local_cpu_mask); | |
960 | ||
6e1254d2 GH |
961 | static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask) |
962 | { | |
963 | int first; | |
964 | ||
965 | /* "this_cpu" is cheaper to preempt than a remote processor */ | |
966 | if ((this_cpu != -1) && cpu_isset(this_cpu, *mask)) | |
967 | return this_cpu; | |
968 | ||
969 | first = first_cpu(*mask); | |
970 | if (first != NR_CPUS) | |
971 | return first; | |
972 | ||
973 | return -1; | |
974 | } | |
975 | ||
976 | static int find_lowest_rq(struct task_struct *task) | |
977 | { | |
978 | struct sched_domain *sd; | |
979 | cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask); | |
980 | int this_cpu = smp_processor_id(); | |
981 | int cpu = task_cpu(task); | |
06f90dbd | 982 | |
6e0534f2 GH |
983 | if (task->rt.nr_cpus_allowed == 1) |
984 | return -1; /* No other targets possible */ | |
6e1254d2 | 985 | |
6e0534f2 GH |
986 | if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask)) |
987 | return -1; /* No targets found */ | |
6e1254d2 | 988 | |
e761b772 MK |
989 | /* |
990 | * Only consider CPUs that are usable for migration. | |
991 | * I guess we might want to change cpupri_find() to ignore those | |
992 | * in the first place. | |
993 | */ | |
994 | cpus_and(*lowest_mask, *lowest_mask, cpu_active_map); | |
995 | ||
6e1254d2 GH |
996 | /* |
997 | * At this point we have built a mask of cpus representing the | |
998 | * lowest priority tasks in the system. Now we want to elect | |
999 | * the best one based on our affinity and topology. | |
1000 | * | |
1001 | * We prioritize the last cpu that the task executed on since | |
1002 | * it is most likely cache-hot in that location. | |
1003 | */ | |
1004 | if (cpu_isset(cpu, *lowest_mask)) | |
1005 | return cpu; | |
1006 | ||
1007 | /* | |
1008 | * Otherwise, we consult the sched_domains span maps to figure | |
1009 | * out which cpu is logically closest to our hot cache data. | |
1010 | */ | |
1011 | if (this_cpu == cpu) | |
1012 | this_cpu = -1; /* Skip this_cpu opt if the same */ | |
1013 | ||
1014 | for_each_domain(cpu, sd) { | |
1015 | if (sd->flags & SD_WAKE_AFFINE) { | |
1016 | cpumask_t domain_mask; | |
1017 | int best_cpu; | |
1018 | ||
1019 | cpus_and(domain_mask, sd->span, *lowest_mask); | |
1020 | ||
1021 | best_cpu = pick_optimal_cpu(this_cpu, | |
1022 | &domain_mask); | |
1023 | if (best_cpu != -1) | |
1024 | return best_cpu; | |
1025 | } | |
1026 | } | |
1027 | ||
1028 | /* | |
1029 | * And finally, if there were no matches within the domains | |
1030 | * just give the caller *something* to work with from the compatible | |
1031 | * locations. | |
1032 | */ | |
1033 | return pick_optimal_cpu(this_cpu, lowest_mask); | |
07b4032c GH |
1034 | } |
1035 | ||
1036 | /* Will lock the rq it finds */ | |
4df64c0b | 1037 | static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) |
07b4032c GH |
1038 | { |
1039 | struct rq *lowest_rq = NULL; | |
07b4032c | 1040 | int tries; |
4df64c0b | 1041 | int cpu; |
e8fa1362 | 1042 | |
07b4032c GH |
1043 | for (tries = 0; tries < RT_MAX_TRIES; tries++) { |
1044 | cpu = find_lowest_rq(task); | |
1045 | ||
2de0b463 | 1046 | if ((cpu == -1) || (cpu == rq->cpu)) |
e8fa1362 SR |
1047 | break; |
1048 | ||
07b4032c GH |
1049 | lowest_rq = cpu_rq(cpu); |
1050 | ||
e8fa1362 | 1051 | /* if the prio of this runqueue changed, try again */ |
07b4032c | 1052 | if (double_lock_balance(rq, lowest_rq)) { |
e8fa1362 SR |
1053 | /* |
1054 | * We had to unlock the run queue. In | |
1055 | * the mean time, task could have | |
1056 | * migrated already or had its affinity changed. | |
1057 | * Also make sure that it wasn't scheduled on its rq. | |
1058 | */ | |
07b4032c | 1059 | if (unlikely(task_rq(task) != rq || |
4df64c0b IM |
1060 | !cpu_isset(lowest_rq->cpu, |
1061 | task->cpus_allowed) || | |
07b4032c | 1062 | task_running(rq, task) || |
e8fa1362 | 1063 | !task->se.on_rq)) { |
4df64c0b | 1064 | |
e8fa1362 SR |
1065 | spin_unlock(&lowest_rq->lock); |
1066 | lowest_rq = NULL; | |
1067 | break; | |
1068 | } | |
1069 | } | |
1070 | ||
1071 | /* If this rq is still suitable use it. */ | |
1072 | if (lowest_rq->rt.highest_prio > task->prio) | |
1073 | break; | |
1074 | ||
1075 | /* try again */ | |
1b12bbc7 | 1076 | double_unlock_balance(rq, lowest_rq); |
e8fa1362 SR |
1077 | lowest_rq = NULL; |
1078 | } | |
1079 | ||
1080 | return lowest_rq; | |
1081 | } | |
1082 | ||
1083 | /* | |
1084 | * If the current CPU has more than one RT task, see if the non | |
1085 | * running task can migrate over to a CPU that is running a task | |
1086 | * of lesser priority. | |
1087 | */ | |
697f0a48 | 1088 | static int push_rt_task(struct rq *rq) |
e8fa1362 SR |
1089 | { |
1090 | struct task_struct *next_task; | |
1091 | struct rq *lowest_rq; | |
1092 | int ret = 0; | |
1093 | int paranoid = RT_MAX_TRIES; | |
1094 | ||
a22d7fc1 GH |
1095 | if (!rq->rt.overloaded) |
1096 | return 0; | |
1097 | ||
697f0a48 | 1098 | next_task = pick_next_highest_task_rt(rq, -1); |
e8fa1362 SR |
1099 | if (!next_task) |
1100 | return 0; | |
1101 | ||
1102 | retry: | |
697f0a48 | 1103 | if (unlikely(next_task == rq->curr)) { |
f65eda4f | 1104 | WARN_ON(1); |
e8fa1362 | 1105 | return 0; |
f65eda4f | 1106 | } |
e8fa1362 SR |
1107 | |
1108 | /* | |
1109 | * It's possible that the next_task slipped in of | |
1110 | * higher priority than current. If that's the case | |
1111 | * just reschedule current. | |
1112 | */ | |
697f0a48 GH |
1113 | if (unlikely(next_task->prio < rq->curr->prio)) { |
1114 | resched_task(rq->curr); | |
e8fa1362 SR |
1115 | return 0; |
1116 | } | |
1117 | ||
697f0a48 | 1118 | /* We might release rq lock */ |
e8fa1362 SR |
1119 | get_task_struct(next_task); |
1120 | ||
1121 | /* find_lock_lowest_rq locks the rq if found */ | |
697f0a48 | 1122 | lowest_rq = find_lock_lowest_rq(next_task, rq); |
e8fa1362 SR |
1123 | if (!lowest_rq) { |
1124 | struct task_struct *task; | |
1125 | /* | |
697f0a48 | 1126 | * find lock_lowest_rq releases rq->lock |
e8fa1362 SR |
1127 | * so it is possible that next_task has changed. |
1128 | * If it has, then try again. | |
1129 | */ | |
697f0a48 | 1130 | task = pick_next_highest_task_rt(rq, -1); |
e8fa1362 SR |
1131 | if (unlikely(task != next_task) && task && paranoid--) { |
1132 | put_task_struct(next_task); | |
1133 | next_task = task; | |
1134 | goto retry; | |
1135 | } | |
1136 | goto out; | |
1137 | } | |
1138 | ||
697f0a48 | 1139 | deactivate_task(rq, next_task, 0); |
e8fa1362 SR |
1140 | set_task_cpu(next_task, lowest_rq->cpu); |
1141 | activate_task(lowest_rq, next_task, 0); | |
1142 | ||
1143 | resched_task(lowest_rq->curr); | |
1144 | ||
1b12bbc7 | 1145 | double_unlock_balance(rq, lowest_rq); |
e8fa1362 SR |
1146 | |
1147 | ret = 1; | |
1148 | out: | |
1149 | put_task_struct(next_task); | |
1150 | ||
1151 | return ret; | |
1152 | } | |
1153 | ||
1154 | /* | |
1155 | * TODO: Currently we just use the second highest prio task on | |
1156 | * the queue, and stop when it can't migrate (or there's | |
1157 | * no more RT tasks). There may be a case where a lower | |
1158 | * priority RT task has a different affinity than the | |
1159 | * higher RT task. In this case the lower RT task could | |
1160 | * possibly be able to migrate where as the higher priority | |
1161 | * RT task could not. We currently ignore this issue. | |
1162 | * Enhancements are welcome! | |
1163 | */ | |
1164 | static void push_rt_tasks(struct rq *rq) | |
1165 | { | |
1166 | /* push_rt_task will return true if it moved an RT */ | |
1167 | while (push_rt_task(rq)) | |
1168 | ; | |
1169 | } | |
1170 | ||
f65eda4f SR |
1171 | static int pull_rt_task(struct rq *this_rq) |
1172 | { | |
80bf3171 IM |
1173 | int this_cpu = this_rq->cpu, ret = 0, cpu; |
1174 | struct task_struct *p, *next; | |
f65eda4f | 1175 | struct rq *src_rq; |
f65eda4f | 1176 | |
637f5085 | 1177 | if (likely(!rt_overloaded(this_rq))) |
f65eda4f SR |
1178 | return 0; |
1179 | ||
1180 | next = pick_next_task_rt(this_rq); | |
1181 | ||
363ab6f1 | 1182 | for_each_cpu_mask_nr(cpu, this_rq->rd->rto_mask) { |
f65eda4f SR |
1183 | if (this_cpu == cpu) |
1184 | continue; | |
1185 | ||
1186 | src_rq = cpu_rq(cpu); | |
f65eda4f SR |
1187 | /* |
1188 | * We can potentially drop this_rq's lock in | |
1189 | * double_lock_balance, and another CPU could | |
1190 | * steal our next task - hence we must cause | |
1191 | * the caller to recalculate the next task | |
1192 | * in that case: | |
1193 | */ | |
1194 | if (double_lock_balance(this_rq, src_rq)) { | |
1195 | struct task_struct *old_next = next; | |
80bf3171 | 1196 | |
f65eda4f SR |
1197 | next = pick_next_task_rt(this_rq); |
1198 | if (next != old_next) | |
1199 | ret = 1; | |
1200 | } | |
1201 | ||
1202 | /* | |
1203 | * Are there still pullable RT tasks? | |
1204 | */ | |
614ee1f6 MG |
1205 | if (src_rq->rt.rt_nr_running <= 1) |
1206 | goto skip; | |
f65eda4f | 1207 | |
f65eda4f SR |
1208 | p = pick_next_highest_task_rt(src_rq, this_cpu); |
1209 | ||
1210 | /* | |
1211 | * Do we have an RT task that preempts | |
1212 | * the to-be-scheduled task? | |
1213 | */ | |
1214 | if (p && (!next || (p->prio < next->prio))) { | |
1215 | WARN_ON(p == src_rq->curr); | |
1216 | WARN_ON(!p->se.on_rq); | |
1217 | ||
1218 | /* | |
1219 | * There's a chance that p is higher in priority | |
1220 | * than what's currently running on its cpu. | |
1221 | * This is just that p is wakeing up and hasn't | |
1222 | * had a chance to schedule. We only pull | |
1223 | * p if it is lower in priority than the | |
1224 | * current task on the run queue or | |
1225 | * this_rq next task is lower in prio than | |
1226 | * the current task on that rq. | |
1227 | */ | |
1228 | if (p->prio < src_rq->curr->prio || | |
1229 | (next && next->prio < src_rq->curr->prio)) | |
614ee1f6 | 1230 | goto skip; |
f65eda4f SR |
1231 | |
1232 | ret = 1; | |
1233 | ||
1234 | deactivate_task(src_rq, p, 0); | |
1235 | set_task_cpu(p, this_cpu); | |
1236 | activate_task(this_rq, p, 0); | |
1237 | /* | |
1238 | * We continue with the search, just in | |
1239 | * case there's an even higher prio task | |
1240 | * in another runqueue. (low likelyhood | |
1241 | * but possible) | |
80bf3171 | 1242 | * |
f65eda4f SR |
1243 | * Update next so that we won't pick a task |
1244 | * on another cpu with a priority lower (or equal) | |
1245 | * than the one we just picked. | |
1246 | */ | |
1247 | next = p; | |
1248 | ||
1249 | } | |
614ee1f6 | 1250 | skip: |
1b12bbc7 | 1251 | double_unlock_balance(this_rq, src_rq); |
f65eda4f SR |
1252 | } |
1253 | ||
1254 | return ret; | |
1255 | } | |
1256 | ||
9a897c5a | 1257 | static void pre_schedule_rt(struct rq *rq, struct task_struct *prev) |
f65eda4f SR |
1258 | { |
1259 | /* Try to pull RT tasks here if we lower this rq's prio */ | |
7f51f298 | 1260 | if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio) |
f65eda4f SR |
1261 | pull_rt_task(rq); |
1262 | } | |
1263 | ||
9a897c5a | 1264 | static void post_schedule_rt(struct rq *rq) |
e8fa1362 SR |
1265 | { |
1266 | /* | |
1267 | * If we have more than one rt_task queued, then | |
1268 | * see if we can push the other rt_tasks off to other CPUS. | |
1269 | * Note we may release the rq lock, and since | |
1270 | * the lock was owned by prev, we need to release it | |
1271 | * first via finish_lock_switch and then reaquire it here. | |
1272 | */ | |
a22d7fc1 | 1273 | if (unlikely(rq->rt.overloaded)) { |
e8fa1362 SR |
1274 | spin_lock_irq(&rq->lock); |
1275 | push_rt_tasks(rq); | |
1276 | spin_unlock_irq(&rq->lock); | |
1277 | } | |
1278 | } | |
1279 | ||
8ae121ac GH |
1280 | /* |
1281 | * If we are not running and we are not going to reschedule soon, we should | |
1282 | * try to push tasks away now | |
1283 | */ | |
9a897c5a | 1284 | static void task_wake_up_rt(struct rq *rq, struct task_struct *p) |
4642dafd | 1285 | { |
9a897c5a | 1286 | if (!task_running(rq, p) && |
8ae121ac | 1287 | !test_tsk_need_resched(rq->curr) && |
a22d7fc1 | 1288 | rq->rt.overloaded) |
4642dafd SR |
1289 | push_rt_tasks(rq); |
1290 | } | |
1291 | ||
43010659 | 1292 | static unsigned long |
bb44e5d1 | 1293 | load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f PW |
1294 | unsigned long max_load_move, |
1295 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1296 | int *all_pinned, int *this_best_prio) | |
bb44e5d1 | 1297 | { |
c7a1e46a SR |
1298 | /* don't touch RT tasks */ |
1299 | return 0; | |
e1d1484f PW |
1300 | } |
1301 | ||
1302 | static int | |
1303 | move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1304 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1305 | { | |
c7a1e46a SR |
1306 | /* don't touch RT tasks */ |
1307 | return 0; | |
bb44e5d1 | 1308 | } |
deeeccd4 | 1309 | |
cd8ba7cd MT |
1310 | static void set_cpus_allowed_rt(struct task_struct *p, |
1311 | const cpumask_t *new_mask) | |
73fe6aae GH |
1312 | { |
1313 | int weight = cpus_weight(*new_mask); | |
1314 | ||
1315 | BUG_ON(!rt_task(p)); | |
1316 | ||
1317 | /* | |
1318 | * Update the migration status of the RQ if we have an RT task | |
1319 | * which is running AND changing its weight value. | |
1320 | */ | |
6f505b16 | 1321 | if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) { |
73fe6aae GH |
1322 | struct rq *rq = task_rq(p); |
1323 | ||
6f505b16 | 1324 | if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) { |
73fe6aae | 1325 | rq->rt.rt_nr_migratory++; |
6f505b16 | 1326 | } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) { |
73fe6aae GH |
1327 | BUG_ON(!rq->rt.rt_nr_migratory); |
1328 | rq->rt.rt_nr_migratory--; | |
1329 | } | |
1330 | ||
1331 | update_rt_migration(rq); | |
1332 | } | |
1333 | ||
1334 | p->cpus_allowed = *new_mask; | |
6f505b16 | 1335 | p->rt.nr_cpus_allowed = weight; |
73fe6aae | 1336 | } |
deeeccd4 | 1337 | |
bdd7c81b | 1338 | /* Assumes rq->lock is held */ |
1f11eb6a | 1339 | static void rq_online_rt(struct rq *rq) |
bdd7c81b IM |
1340 | { |
1341 | if (rq->rt.overloaded) | |
1342 | rt_set_overload(rq); | |
6e0534f2 | 1343 | |
7def2be1 PZ |
1344 | __enable_runtime(rq); |
1345 | ||
6e0534f2 | 1346 | cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio); |
bdd7c81b IM |
1347 | } |
1348 | ||
1349 | /* Assumes rq->lock is held */ | |
1f11eb6a | 1350 | static void rq_offline_rt(struct rq *rq) |
bdd7c81b IM |
1351 | { |
1352 | if (rq->rt.overloaded) | |
1353 | rt_clear_overload(rq); | |
6e0534f2 | 1354 | |
7def2be1 PZ |
1355 | __disable_runtime(rq); |
1356 | ||
6e0534f2 | 1357 | cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID); |
bdd7c81b | 1358 | } |
cb469845 SR |
1359 | |
1360 | /* | |
1361 | * When switch from the rt queue, we bring ourselves to a position | |
1362 | * that we might want to pull RT tasks from other runqueues. | |
1363 | */ | |
1364 | static void switched_from_rt(struct rq *rq, struct task_struct *p, | |
1365 | int running) | |
1366 | { | |
1367 | /* | |
1368 | * If there are other RT tasks then we will reschedule | |
1369 | * and the scheduling of the other RT tasks will handle | |
1370 | * the balancing. But if we are the last RT task | |
1371 | * we may need to handle the pulling of RT tasks | |
1372 | * now. | |
1373 | */ | |
1374 | if (!rq->rt.rt_nr_running) | |
1375 | pull_rt_task(rq); | |
1376 | } | |
1377 | #endif /* CONFIG_SMP */ | |
1378 | ||
1379 | /* | |
1380 | * When switching a task to RT, we may overload the runqueue | |
1381 | * with RT tasks. In this case we try to push them off to | |
1382 | * other runqueues. | |
1383 | */ | |
1384 | static void switched_to_rt(struct rq *rq, struct task_struct *p, | |
1385 | int running) | |
1386 | { | |
1387 | int check_resched = 1; | |
1388 | ||
1389 | /* | |
1390 | * If we are already running, then there's nothing | |
1391 | * that needs to be done. But if we are not running | |
1392 | * we may need to preempt the current running task. | |
1393 | * If that current running task is also an RT task | |
1394 | * then see if we can move to another run queue. | |
1395 | */ | |
1396 | if (!running) { | |
1397 | #ifdef CONFIG_SMP | |
1398 | if (rq->rt.overloaded && push_rt_task(rq) && | |
1399 | /* Don't resched if we changed runqueues */ | |
1400 | rq != task_rq(p)) | |
1401 | check_resched = 0; | |
1402 | #endif /* CONFIG_SMP */ | |
1403 | if (check_resched && p->prio < rq->curr->prio) | |
1404 | resched_task(rq->curr); | |
1405 | } | |
1406 | } | |
1407 | ||
1408 | /* | |
1409 | * Priority of the task has changed. This may cause | |
1410 | * us to initiate a push or pull. | |
1411 | */ | |
1412 | static void prio_changed_rt(struct rq *rq, struct task_struct *p, | |
1413 | int oldprio, int running) | |
1414 | { | |
1415 | if (running) { | |
1416 | #ifdef CONFIG_SMP | |
1417 | /* | |
1418 | * If our priority decreases while running, we | |
1419 | * may need to pull tasks to this runqueue. | |
1420 | */ | |
1421 | if (oldprio < p->prio) | |
1422 | pull_rt_task(rq); | |
1423 | /* | |
1424 | * If there's a higher priority task waiting to run | |
6fa46fa5 SR |
1425 | * then reschedule. Note, the above pull_rt_task |
1426 | * can release the rq lock and p could migrate. | |
1427 | * Only reschedule if p is still on the same runqueue. | |
cb469845 | 1428 | */ |
6fa46fa5 | 1429 | if (p->prio > rq->rt.highest_prio && rq->curr == p) |
cb469845 SR |
1430 | resched_task(p); |
1431 | #else | |
1432 | /* For UP simply resched on drop of prio */ | |
1433 | if (oldprio < p->prio) | |
1434 | resched_task(p); | |
e8fa1362 | 1435 | #endif /* CONFIG_SMP */ |
cb469845 SR |
1436 | } else { |
1437 | /* | |
1438 | * This task is not running, but if it is | |
1439 | * greater than the current running task | |
1440 | * then reschedule. | |
1441 | */ | |
1442 | if (p->prio < rq->curr->prio) | |
1443 | resched_task(rq->curr); | |
1444 | } | |
1445 | } | |
1446 | ||
78f2c7db PZ |
1447 | static void watchdog(struct rq *rq, struct task_struct *p) |
1448 | { | |
1449 | unsigned long soft, hard; | |
1450 | ||
1451 | if (!p->signal) | |
1452 | return; | |
1453 | ||
1454 | soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur; | |
1455 | hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max; | |
1456 | ||
1457 | if (soft != RLIM_INFINITY) { | |
1458 | unsigned long next; | |
1459 | ||
1460 | p->rt.timeout++; | |
1461 | next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ); | |
5a52dd50 | 1462 | if (p->rt.timeout > next) |
f06febc9 | 1463 | p->cputime_expires.sched_exp = p->se.sum_exec_runtime; |
78f2c7db PZ |
1464 | } |
1465 | } | |
bb44e5d1 | 1466 | |
8f4d37ec | 1467 | static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued) |
bb44e5d1 | 1468 | { |
67e2be02 PZ |
1469 | update_curr_rt(rq); |
1470 | ||
78f2c7db PZ |
1471 | watchdog(rq, p); |
1472 | ||
bb44e5d1 IM |
1473 | /* |
1474 | * RR tasks need a special form of timeslice management. | |
1475 | * FIFO tasks have no timeslices. | |
1476 | */ | |
1477 | if (p->policy != SCHED_RR) | |
1478 | return; | |
1479 | ||
fa717060 | 1480 | if (--p->rt.time_slice) |
bb44e5d1 IM |
1481 | return; |
1482 | ||
fa717060 | 1483 | p->rt.time_slice = DEF_TIMESLICE; |
bb44e5d1 | 1484 | |
98fbc798 DA |
1485 | /* |
1486 | * Requeue to the end of queue if we are not the only element | |
1487 | * on the queue: | |
1488 | */ | |
fa717060 | 1489 | if (p->rt.run_list.prev != p->rt.run_list.next) { |
7ebefa8c | 1490 | requeue_task_rt(rq, p, 0); |
98fbc798 DA |
1491 | set_tsk_need_resched(p); |
1492 | } | |
bb44e5d1 IM |
1493 | } |
1494 | ||
83b699ed SV |
1495 | static void set_curr_task_rt(struct rq *rq) |
1496 | { | |
1497 | struct task_struct *p = rq->curr; | |
1498 | ||
1499 | p->se.exec_start = rq->clock; | |
1500 | } | |
1501 | ||
2abdad0a | 1502 | static const struct sched_class rt_sched_class = { |
5522d5d5 | 1503 | .next = &fair_sched_class, |
bb44e5d1 IM |
1504 | .enqueue_task = enqueue_task_rt, |
1505 | .dequeue_task = dequeue_task_rt, | |
1506 | .yield_task = yield_task_rt, | |
1507 | ||
1508 | .check_preempt_curr = check_preempt_curr_rt, | |
1509 | ||
1510 | .pick_next_task = pick_next_task_rt, | |
1511 | .put_prev_task = put_prev_task_rt, | |
1512 | ||
681f3e68 | 1513 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
1514 | .select_task_rq = select_task_rq_rt, |
1515 | ||
bb44e5d1 | 1516 | .load_balance = load_balance_rt, |
e1d1484f | 1517 | .move_one_task = move_one_task_rt, |
73fe6aae | 1518 | .set_cpus_allowed = set_cpus_allowed_rt, |
1f11eb6a GH |
1519 | .rq_online = rq_online_rt, |
1520 | .rq_offline = rq_offline_rt, | |
9a897c5a SR |
1521 | .pre_schedule = pre_schedule_rt, |
1522 | .post_schedule = post_schedule_rt, | |
1523 | .task_wake_up = task_wake_up_rt, | |
cb469845 | 1524 | .switched_from = switched_from_rt, |
681f3e68 | 1525 | #endif |
bb44e5d1 | 1526 | |
83b699ed | 1527 | .set_curr_task = set_curr_task_rt, |
bb44e5d1 | 1528 | .task_tick = task_tick_rt, |
cb469845 SR |
1529 | |
1530 | .prio_changed = prio_changed_rt, | |
1531 | .switched_to = switched_to_rt, | |
bb44e5d1 | 1532 | }; |
ada18de2 PZ |
1533 | |
1534 | #ifdef CONFIG_SCHED_DEBUG | |
1535 | extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq); | |
1536 | ||
1537 | static void print_rt_stats(struct seq_file *m, int cpu) | |
1538 | { | |
1539 | struct rt_rq *rt_rq; | |
1540 | ||
1541 | rcu_read_lock(); | |
1542 | for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu)) | |
1543 | print_rt_rq(m, cpu, rt_rq); | |
1544 | rcu_read_unlock(); | |
1545 | } | |
55e12e5e | 1546 | #endif /* CONFIG_SCHED_DEBUG */ |