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 | { | |
637f5085 | 15 | cpu_set(rq->cpu, rq->rd->rto_mask); |
4fd29176 SR |
16 | /* |
17 | * Make sure the mask is visible before we set | |
18 | * the overload count. That is checked to determine | |
19 | * if we should look at the mask. It would be a shame | |
20 | * if we looked at the mask, but the mask was not | |
21 | * updated yet. | |
22 | */ | |
23 | wmb(); | |
637f5085 | 24 | atomic_inc(&rq->rd->rto_count); |
4fd29176 | 25 | } |
84de4274 | 26 | |
4fd29176 SR |
27 | static inline void rt_clear_overload(struct rq *rq) |
28 | { | |
29 | /* the order here really doesn't matter */ | |
637f5085 GH |
30 | atomic_dec(&rq->rd->rto_count); |
31 | cpu_clear(rq->cpu, rq->rd->rto_mask); | |
4fd29176 | 32 | } |
73fe6aae GH |
33 | |
34 | static void update_rt_migration(struct rq *rq) | |
35 | { | |
637f5085 | 36 | if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) { |
cdc8eb98 GH |
37 | if (!rq->rt.overloaded) { |
38 | rt_set_overload(rq); | |
39 | rq->rt.overloaded = 1; | |
40 | } | |
41 | } else if (rq->rt.overloaded) { | |
73fe6aae | 42 | rt_clear_overload(rq); |
637f5085 GH |
43 | rq->rt.overloaded = 0; |
44 | } | |
73fe6aae | 45 | } |
4fd29176 SR |
46 | #endif /* CONFIG_SMP */ |
47 | ||
6f505b16 | 48 | static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) |
fa85ae24 | 49 | { |
6f505b16 PZ |
50 | return container_of(rt_se, struct task_struct, rt); |
51 | } | |
52 | ||
53 | static inline int on_rt_rq(struct sched_rt_entity *rt_se) | |
54 | { | |
55 | return !list_empty(&rt_se->run_list); | |
56 | } | |
57 | ||
052f1dc7 | 58 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 59 | |
9f0c1e56 | 60 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) |
6f505b16 PZ |
61 | { |
62 | if (!rt_rq->tg) | |
9f0c1e56 | 63 | return RUNTIME_INF; |
6f505b16 | 64 | |
ac086bc2 PZ |
65 | return rt_rq->rt_runtime; |
66 | } | |
67 | ||
68 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) | |
69 | { | |
70 | return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period); | |
6f505b16 PZ |
71 | } |
72 | ||
73 | #define for_each_leaf_rt_rq(rt_rq, rq) \ | |
74 | list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list) | |
75 | ||
76 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) | |
77 | { | |
78 | return rt_rq->rq; | |
79 | } | |
80 | ||
81 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) | |
82 | { | |
83 | return rt_se->rt_rq; | |
84 | } | |
85 | ||
86 | #define for_each_sched_rt_entity(rt_se) \ | |
87 | for (; rt_se; rt_se = rt_se->parent) | |
88 | ||
89 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) | |
90 | { | |
91 | return rt_se->my_q; | |
92 | } | |
93 | ||
94 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se); | |
95 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se); | |
96 | ||
9f0c1e56 | 97 | static void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
6f505b16 PZ |
98 | { |
99 | struct sched_rt_entity *rt_se = rt_rq->rt_se; | |
100 | ||
101 | if (rt_se && !on_rt_rq(rt_se) && rt_rq->rt_nr_running) { | |
1020387f PZ |
102 | struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr; |
103 | ||
6f505b16 | 104 | enqueue_rt_entity(rt_se); |
1020387f PZ |
105 | if (rt_rq->highest_prio < curr->prio) |
106 | resched_task(curr); | |
6f505b16 PZ |
107 | } |
108 | } | |
109 | ||
9f0c1e56 | 110 | static void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
6f505b16 PZ |
111 | { |
112 | struct sched_rt_entity *rt_se = rt_rq->rt_se; | |
113 | ||
114 | if (rt_se && on_rt_rq(rt_se)) | |
115 | dequeue_rt_entity(rt_se); | |
116 | } | |
117 | ||
23b0fdfc PZ |
118 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
119 | { | |
120 | return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted; | |
121 | } | |
122 | ||
123 | static int rt_se_boosted(struct sched_rt_entity *rt_se) | |
124 | { | |
125 | struct rt_rq *rt_rq = group_rt_rq(rt_se); | |
126 | struct task_struct *p; | |
127 | ||
128 | if (rt_rq) | |
129 | return !!rt_rq->rt_nr_boosted; | |
130 | ||
131 | p = rt_task_of(rt_se); | |
132 | return p->prio != p->normal_prio; | |
133 | } | |
134 | ||
d0b27fa7 PZ |
135 | #ifdef CONFIG_SMP |
136 | static inline cpumask_t sched_rt_period_mask(void) | |
137 | { | |
138 | return cpu_rq(smp_processor_id())->rd->span; | |
139 | } | |
6f505b16 | 140 | #else |
d0b27fa7 PZ |
141 | static inline cpumask_t sched_rt_period_mask(void) |
142 | { | |
143 | return cpu_online_map; | |
144 | } | |
145 | #endif | |
6f505b16 | 146 | |
d0b27fa7 PZ |
147 | static inline |
148 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) | |
6f505b16 | 149 | { |
d0b27fa7 PZ |
150 | return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu]; |
151 | } | |
9f0c1e56 | 152 | |
ac086bc2 PZ |
153 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
154 | { | |
155 | return &rt_rq->tg->rt_bandwidth; | |
156 | } | |
157 | ||
d0b27fa7 PZ |
158 | #else |
159 | ||
160 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) | |
161 | { | |
ac086bc2 PZ |
162 | return rt_rq->rt_runtime; |
163 | } | |
164 | ||
165 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) | |
166 | { | |
167 | return ktime_to_ns(def_rt_bandwidth.rt_period); | |
6f505b16 PZ |
168 | } |
169 | ||
170 | #define for_each_leaf_rt_rq(rt_rq, rq) \ | |
171 | for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL) | |
172 | ||
173 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) | |
174 | { | |
175 | return container_of(rt_rq, struct rq, rt); | |
176 | } | |
177 | ||
178 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) | |
179 | { | |
180 | struct task_struct *p = rt_task_of(rt_se); | |
181 | struct rq *rq = task_rq(p); | |
182 | ||
183 | return &rq->rt; | |
184 | } | |
185 | ||
186 | #define for_each_sched_rt_entity(rt_se) \ | |
187 | for (; rt_se; rt_se = NULL) | |
188 | ||
189 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) | |
190 | { | |
191 | return NULL; | |
192 | } | |
193 | ||
9f0c1e56 | 194 | static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
6f505b16 PZ |
195 | { |
196 | } | |
197 | ||
9f0c1e56 | 198 | static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
6f505b16 PZ |
199 | { |
200 | } | |
201 | ||
23b0fdfc PZ |
202 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
203 | { | |
204 | return rt_rq->rt_throttled; | |
205 | } | |
d0b27fa7 PZ |
206 | |
207 | static inline cpumask_t sched_rt_period_mask(void) | |
208 | { | |
209 | return cpu_online_map; | |
210 | } | |
211 | ||
212 | static inline | |
213 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) | |
214 | { | |
215 | return &cpu_rq(cpu)->rt; | |
216 | } | |
217 | ||
ac086bc2 PZ |
218 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
219 | { | |
220 | return &def_rt_bandwidth; | |
221 | } | |
222 | ||
6f505b16 PZ |
223 | #endif |
224 | ||
d0b27fa7 PZ |
225 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun) |
226 | { | |
227 | int i, idle = 1; | |
228 | cpumask_t span; | |
229 | ||
230 | if (rt_b->rt_runtime == RUNTIME_INF) | |
231 | return 1; | |
232 | ||
233 | span = sched_rt_period_mask(); | |
234 | for_each_cpu_mask(i, span) { | |
235 | int enqueue = 0; | |
236 | struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i); | |
237 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
238 | ||
239 | spin_lock(&rq->lock); | |
240 | if (rt_rq->rt_time) { | |
ac086bc2 | 241 | u64 runtime; |
d0b27fa7 | 242 | |
ac086bc2 PZ |
243 | spin_lock(&rt_rq->rt_runtime_lock); |
244 | runtime = rt_rq->rt_runtime; | |
d0b27fa7 PZ |
245 | rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime); |
246 | if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) { | |
247 | rt_rq->rt_throttled = 0; | |
248 | enqueue = 1; | |
249 | } | |
250 | if (rt_rq->rt_time || rt_rq->rt_nr_running) | |
251 | idle = 0; | |
ac086bc2 | 252 | spin_unlock(&rt_rq->rt_runtime_lock); |
d0b27fa7 PZ |
253 | } |
254 | ||
255 | if (enqueue) | |
256 | sched_rt_rq_enqueue(rt_rq); | |
257 | spin_unlock(&rq->lock); | |
258 | } | |
259 | ||
260 | return idle; | |
261 | } | |
262 | ||
ac086bc2 PZ |
263 | #ifdef CONFIG_SMP |
264 | static int balance_runtime(struct rt_rq *rt_rq) | |
265 | { | |
266 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); | |
267 | struct root_domain *rd = cpu_rq(smp_processor_id())->rd; | |
268 | int i, weight, more = 0; | |
269 | u64 rt_period; | |
270 | ||
271 | weight = cpus_weight(rd->span); | |
272 | ||
273 | spin_lock(&rt_b->rt_runtime_lock); | |
274 | rt_period = ktime_to_ns(rt_b->rt_period); | |
275 | for_each_cpu_mask(i, rd->span) { | |
276 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); | |
277 | s64 diff; | |
278 | ||
279 | if (iter == rt_rq) | |
280 | continue; | |
281 | ||
282 | spin_lock(&iter->rt_runtime_lock); | |
283 | diff = iter->rt_runtime - iter->rt_time; | |
284 | if (diff > 0) { | |
285 | do_div(diff, weight); | |
286 | if (rt_rq->rt_runtime + diff > rt_period) | |
287 | diff = rt_period - rt_rq->rt_runtime; | |
288 | iter->rt_runtime -= diff; | |
289 | rt_rq->rt_runtime += diff; | |
290 | more = 1; | |
291 | if (rt_rq->rt_runtime == rt_period) { | |
292 | spin_unlock(&iter->rt_runtime_lock); | |
293 | break; | |
294 | } | |
295 | } | |
296 | spin_unlock(&iter->rt_runtime_lock); | |
297 | } | |
298 | spin_unlock(&rt_b->rt_runtime_lock); | |
299 | ||
300 | return more; | |
301 | } | |
302 | #endif | |
303 | ||
6f505b16 PZ |
304 | static inline int rt_se_prio(struct sched_rt_entity *rt_se) |
305 | { | |
052f1dc7 | 306 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 PZ |
307 | struct rt_rq *rt_rq = group_rt_rq(rt_se); |
308 | ||
309 | if (rt_rq) | |
310 | return rt_rq->highest_prio; | |
311 | #endif | |
312 | ||
313 | return rt_task_of(rt_se)->prio; | |
314 | } | |
315 | ||
9f0c1e56 | 316 | static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq) |
6f505b16 | 317 | { |
9f0c1e56 | 318 | u64 runtime = sched_rt_runtime(rt_rq); |
fa85ae24 | 319 | |
9f0c1e56 | 320 | if (runtime == RUNTIME_INF) |
fa85ae24 PZ |
321 | return 0; |
322 | ||
323 | if (rt_rq->rt_throttled) | |
23b0fdfc | 324 | return rt_rq_throttled(rt_rq); |
fa85ae24 | 325 | |
ac086bc2 PZ |
326 | if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq)) |
327 | return 0; | |
328 | ||
329 | #ifdef CONFIG_SMP | |
330 | if (rt_rq->rt_time > runtime) { | |
331 | int more; | |
332 | ||
333 | spin_unlock(&rt_rq->rt_runtime_lock); | |
334 | more = balance_runtime(rt_rq); | |
335 | spin_lock(&rt_rq->rt_runtime_lock); | |
336 | ||
337 | if (more) | |
338 | runtime = sched_rt_runtime(rt_rq); | |
339 | } | |
340 | #endif | |
341 | ||
9f0c1e56 | 342 | if (rt_rq->rt_time > runtime) { |
6f505b16 | 343 | rt_rq->rt_throttled = 1; |
23b0fdfc | 344 | if (rt_rq_throttled(rt_rq)) { |
9f0c1e56 | 345 | sched_rt_rq_dequeue(rt_rq); |
23b0fdfc PZ |
346 | return 1; |
347 | } | |
fa85ae24 PZ |
348 | } |
349 | ||
350 | return 0; | |
351 | } | |
352 | ||
bb44e5d1 IM |
353 | /* |
354 | * Update the current task's runtime statistics. Skip current tasks that | |
355 | * are not in our scheduling class. | |
356 | */ | |
a9957449 | 357 | static void update_curr_rt(struct rq *rq) |
bb44e5d1 IM |
358 | { |
359 | struct task_struct *curr = rq->curr; | |
6f505b16 PZ |
360 | struct sched_rt_entity *rt_se = &curr->rt; |
361 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); | |
bb44e5d1 IM |
362 | u64 delta_exec; |
363 | ||
364 | if (!task_has_rt_policy(curr)) | |
365 | return; | |
366 | ||
d281918d | 367 | delta_exec = rq->clock - curr->se.exec_start; |
bb44e5d1 IM |
368 | if (unlikely((s64)delta_exec < 0)) |
369 | delta_exec = 0; | |
6cfb0d5d IM |
370 | |
371 | schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec)); | |
bb44e5d1 IM |
372 | |
373 | curr->se.sum_exec_runtime += delta_exec; | |
d281918d | 374 | curr->se.exec_start = rq->clock; |
d842de87 | 375 | cpuacct_charge(curr, delta_exec); |
fa85ae24 | 376 | |
354d60c2 DG |
377 | for_each_sched_rt_entity(rt_se) { |
378 | rt_rq = rt_rq_of_se(rt_se); | |
379 | ||
380 | spin_lock(&rt_rq->rt_runtime_lock); | |
381 | rt_rq->rt_time += delta_exec; | |
382 | if (sched_rt_runtime_exceeded(rt_rq)) | |
383 | resched_task(curr); | |
384 | spin_unlock(&rt_rq->rt_runtime_lock); | |
385 | } | |
bb44e5d1 IM |
386 | } |
387 | ||
6f505b16 PZ |
388 | static inline |
389 | void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
63489e45 | 390 | { |
6f505b16 PZ |
391 | WARN_ON(!rt_prio(rt_se_prio(rt_se))); |
392 | rt_rq->rt_nr_running++; | |
052f1dc7 | 393 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
6f505b16 PZ |
394 | if (rt_se_prio(rt_se) < rt_rq->highest_prio) |
395 | rt_rq->highest_prio = rt_se_prio(rt_se); | |
396 | #endif | |
764a9d6f | 397 | #ifdef CONFIG_SMP |
6f505b16 PZ |
398 | if (rt_se->nr_cpus_allowed > 1) { |
399 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
73fe6aae | 400 | rq->rt.rt_nr_migratory++; |
6f505b16 | 401 | } |
73fe6aae | 402 | |
6f505b16 PZ |
403 | update_rt_migration(rq_of_rt_rq(rt_rq)); |
404 | #endif | |
052f1dc7 | 405 | #ifdef CONFIG_RT_GROUP_SCHED |
23b0fdfc PZ |
406 | if (rt_se_boosted(rt_se)) |
407 | rt_rq->rt_nr_boosted++; | |
d0b27fa7 PZ |
408 | |
409 | if (rt_rq->tg) | |
410 | start_rt_bandwidth(&rt_rq->tg->rt_bandwidth); | |
411 | #else | |
412 | start_rt_bandwidth(&def_rt_bandwidth); | |
23b0fdfc | 413 | #endif |
63489e45 SR |
414 | } |
415 | ||
6f505b16 PZ |
416 | static inline |
417 | void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
63489e45 | 418 | { |
6f505b16 PZ |
419 | WARN_ON(!rt_prio(rt_se_prio(rt_se))); |
420 | WARN_ON(!rt_rq->rt_nr_running); | |
421 | rt_rq->rt_nr_running--; | |
052f1dc7 | 422 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
6f505b16 | 423 | if (rt_rq->rt_nr_running) { |
764a9d6f SR |
424 | struct rt_prio_array *array; |
425 | ||
6f505b16 PZ |
426 | WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio); |
427 | if (rt_se_prio(rt_se) == rt_rq->highest_prio) { | |
764a9d6f | 428 | /* recalculate */ |
6f505b16 PZ |
429 | array = &rt_rq->active; |
430 | rt_rq->highest_prio = | |
764a9d6f SR |
431 | sched_find_first_bit(array->bitmap); |
432 | } /* otherwise leave rq->highest prio alone */ | |
433 | } else | |
6f505b16 PZ |
434 | rt_rq->highest_prio = MAX_RT_PRIO; |
435 | #endif | |
436 | #ifdef CONFIG_SMP | |
437 | if (rt_se->nr_cpus_allowed > 1) { | |
438 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
73fe6aae | 439 | rq->rt.rt_nr_migratory--; |
6f505b16 | 440 | } |
73fe6aae | 441 | |
6f505b16 | 442 | update_rt_migration(rq_of_rt_rq(rt_rq)); |
764a9d6f | 443 | #endif /* CONFIG_SMP */ |
052f1dc7 | 444 | #ifdef CONFIG_RT_GROUP_SCHED |
23b0fdfc PZ |
445 | if (rt_se_boosted(rt_se)) |
446 | rt_rq->rt_nr_boosted--; | |
447 | ||
448 | WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted); | |
449 | #endif | |
63489e45 SR |
450 | } |
451 | ||
6f505b16 | 452 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se) |
bb44e5d1 | 453 | { |
6f505b16 PZ |
454 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); |
455 | struct rt_prio_array *array = &rt_rq->active; | |
456 | struct rt_rq *group_rq = group_rt_rq(rt_se); | |
bb44e5d1 | 457 | |
23b0fdfc | 458 | if (group_rq && rt_rq_throttled(group_rq)) |
6f505b16 | 459 | return; |
63489e45 | 460 | |
6f505b16 PZ |
461 | list_add_tail(&rt_se->run_list, array->queue + rt_se_prio(rt_se)); |
462 | __set_bit(rt_se_prio(rt_se), array->bitmap); | |
78f2c7db | 463 | |
6f505b16 PZ |
464 | inc_rt_tasks(rt_se, rt_rq); |
465 | } | |
466 | ||
467 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se) | |
468 | { | |
469 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); | |
470 | struct rt_prio_array *array = &rt_rq->active; | |
471 | ||
472 | list_del_init(&rt_se->run_list); | |
473 | if (list_empty(array->queue + rt_se_prio(rt_se))) | |
474 | __clear_bit(rt_se_prio(rt_se), array->bitmap); | |
475 | ||
476 | dec_rt_tasks(rt_se, rt_rq); | |
477 | } | |
478 | ||
479 | /* | |
480 | * Because the prio of an upper entry depends on the lower | |
481 | * entries, we must remove entries top - down. | |
6f505b16 PZ |
482 | */ |
483 | static void dequeue_rt_stack(struct task_struct *p) | |
484 | { | |
58d6c2d7 | 485 | struct sched_rt_entity *rt_se, *back = NULL; |
6f505b16 | 486 | |
58d6c2d7 PZ |
487 | rt_se = &p->rt; |
488 | for_each_sched_rt_entity(rt_se) { | |
489 | rt_se->back = back; | |
490 | back = rt_se; | |
491 | } | |
492 | ||
493 | for (rt_se = back; rt_se; rt_se = rt_se->back) { | |
494 | if (on_rt_rq(rt_se)) | |
495 | dequeue_rt_entity(rt_se); | |
496 | } | |
bb44e5d1 IM |
497 | } |
498 | ||
499 | /* | |
500 | * Adding/removing a task to/from a priority array: | |
501 | */ | |
6f505b16 PZ |
502 | static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup) |
503 | { | |
504 | struct sched_rt_entity *rt_se = &p->rt; | |
505 | ||
506 | if (wakeup) | |
507 | rt_se->timeout = 0; | |
508 | ||
509 | dequeue_rt_stack(p); | |
510 | ||
511 | /* | |
512 | * enqueue everybody, bottom - up. | |
513 | */ | |
514 | for_each_sched_rt_entity(rt_se) | |
515 | enqueue_rt_entity(rt_se); | |
6f505b16 PZ |
516 | } |
517 | ||
f02231e5 | 518 | static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep) |
bb44e5d1 | 519 | { |
6f505b16 PZ |
520 | struct sched_rt_entity *rt_se = &p->rt; |
521 | struct rt_rq *rt_rq; | |
bb44e5d1 | 522 | |
f1e14ef6 | 523 | update_curr_rt(rq); |
bb44e5d1 | 524 | |
6f505b16 PZ |
525 | dequeue_rt_stack(p); |
526 | ||
527 | /* | |
528 | * re-enqueue all non-empty rt_rq entities. | |
529 | */ | |
530 | for_each_sched_rt_entity(rt_se) { | |
531 | rt_rq = group_rt_rq(rt_se); | |
532 | if (rt_rq && rt_rq->rt_nr_running) | |
533 | enqueue_rt_entity(rt_se); | |
534 | } | |
bb44e5d1 IM |
535 | } |
536 | ||
537 | /* | |
538 | * Put task to the end of the run list without the overhead of dequeue | |
539 | * followed by enqueue. | |
540 | */ | |
6f505b16 PZ |
541 | static |
542 | void requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se) | |
543 | { | |
544 | struct rt_prio_array *array = &rt_rq->active; | |
545 | ||
546 | list_move_tail(&rt_se->run_list, array->queue + rt_se_prio(rt_se)); | |
547 | } | |
548 | ||
bb44e5d1 IM |
549 | static void requeue_task_rt(struct rq *rq, struct task_struct *p) |
550 | { | |
6f505b16 PZ |
551 | struct sched_rt_entity *rt_se = &p->rt; |
552 | struct rt_rq *rt_rq; | |
bb44e5d1 | 553 | |
6f505b16 PZ |
554 | for_each_sched_rt_entity(rt_se) { |
555 | rt_rq = rt_rq_of_se(rt_se); | |
556 | requeue_rt_entity(rt_rq, rt_se); | |
557 | } | |
bb44e5d1 IM |
558 | } |
559 | ||
6f505b16 | 560 | static void yield_task_rt(struct rq *rq) |
bb44e5d1 | 561 | { |
4530d7ab | 562 | requeue_task_rt(rq, rq->curr); |
bb44e5d1 IM |
563 | } |
564 | ||
e7693a36 | 565 | #ifdef CONFIG_SMP |
318e0893 GH |
566 | static int find_lowest_rq(struct task_struct *task); |
567 | ||
e7693a36 GH |
568 | static int select_task_rq_rt(struct task_struct *p, int sync) |
569 | { | |
318e0893 GH |
570 | struct rq *rq = task_rq(p); |
571 | ||
572 | /* | |
e1f47d89 SR |
573 | * If the current task is an RT task, then |
574 | * try to see if we can wake this RT task up on another | |
575 | * runqueue. Otherwise simply start this RT task | |
576 | * on its current runqueue. | |
577 | * | |
578 | * We want to avoid overloading runqueues. Even if | |
579 | * the RT task is of higher priority than the current RT task. | |
580 | * RT tasks behave differently than other tasks. If | |
581 | * one gets preempted, we try to push it off to another queue. | |
582 | * So trying to keep a preempting RT task on the same | |
583 | * cache hot CPU will force the running RT task to | |
584 | * a cold CPU. So we waste all the cache for the lower | |
585 | * RT task in hopes of saving some of a RT task | |
586 | * that is just being woken and probably will have | |
587 | * cold cache anyway. | |
318e0893 | 588 | */ |
17b3279b | 589 | if (unlikely(rt_task(rq->curr)) && |
6f505b16 | 590 | (p->rt.nr_cpus_allowed > 1)) { |
318e0893 GH |
591 | int cpu = find_lowest_rq(p); |
592 | ||
593 | return (cpu == -1) ? task_cpu(p) : cpu; | |
594 | } | |
595 | ||
596 | /* | |
597 | * Otherwise, just let it ride on the affined RQ and the | |
598 | * post-schedule router will push the preempted task away | |
599 | */ | |
e7693a36 GH |
600 | return task_cpu(p); |
601 | } | |
602 | #endif /* CONFIG_SMP */ | |
603 | ||
bb44e5d1 IM |
604 | /* |
605 | * Preempt the current task with a newly woken task if needed: | |
606 | */ | |
607 | static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p) | |
608 | { | |
609 | if (p->prio < rq->curr->prio) | |
610 | resched_task(rq->curr); | |
611 | } | |
612 | ||
6f505b16 PZ |
613 | static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq, |
614 | struct rt_rq *rt_rq) | |
bb44e5d1 | 615 | { |
6f505b16 PZ |
616 | struct rt_prio_array *array = &rt_rq->active; |
617 | struct sched_rt_entity *next = NULL; | |
bb44e5d1 IM |
618 | struct list_head *queue; |
619 | int idx; | |
620 | ||
621 | idx = sched_find_first_bit(array->bitmap); | |
6f505b16 | 622 | BUG_ON(idx >= MAX_RT_PRIO); |
bb44e5d1 IM |
623 | |
624 | queue = array->queue + idx; | |
6f505b16 | 625 | next = list_entry(queue->next, struct sched_rt_entity, run_list); |
326587b8 | 626 | |
6f505b16 PZ |
627 | return next; |
628 | } | |
bb44e5d1 | 629 | |
6f505b16 PZ |
630 | static struct task_struct *pick_next_task_rt(struct rq *rq) |
631 | { | |
632 | struct sched_rt_entity *rt_se; | |
633 | struct task_struct *p; | |
634 | struct rt_rq *rt_rq; | |
bb44e5d1 | 635 | |
6f505b16 PZ |
636 | rt_rq = &rq->rt; |
637 | ||
638 | if (unlikely(!rt_rq->rt_nr_running)) | |
639 | return NULL; | |
640 | ||
23b0fdfc | 641 | if (rt_rq_throttled(rt_rq)) |
6f505b16 PZ |
642 | return NULL; |
643 | ||
644 | do { | |
645 | rt_se = pick_next_rt_entity(rq, rt_rq); | |
326587b8 | 646 | BUG_ON(!rt_se); |
6f505b16 PZ |
647 | rt_rq = group_rt_rq(rt_se); |
648 | } while (rt_rq); | |
649 | ||
650 | p = rt_task_of(rt_se); | |
651 | p->se.exec_start = rq->clock; | |
652 | return p; | |
bb44e5d1 IM |
653 | } |
654 | ||
31ee529c | 655 | static void put_prev_task_rt(struct rq *rq, struct task_struct *p) |
bb44e5d1 | 656 | { |
f1e14ef6 | 657 | update_curr_rt(rq); |
bb44e5d1 IM |
658 | p->se.exec_start = 0; |
659 | } | |
660 | ||
681f3e68 | 661 | #ifdef CONFIG_SMP |
6f505b16 | 662 | |
e8fa1362 SR |
663 | /* Only try algorithms three times */ |
664 | #define RT_MAX_TRIES 3 | |
665 | ||
666 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest); | |
667 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep); | |
668 | ||
f65eda4f SR |
669 | static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) |
670 | { | |
671 | if (!task_running(rq, p) && | |
73fe6aae | 672 | (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) && |
6f505b16 | 673 | (p->rt.nr_cpus_allowed > 1)) |
f65eda4f SR |
674 | return 1; |
675 | return 0; | |
676 | } | |
677 | ||
e8fa1362 | 678 | /* Return the second highest RT task, NULL otherwise */ |
79064fbf | 679 | static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu) |
e8fa1362 | 680 | { |
6f505b16 PZ |
681 | struct task_struct *next = NULL; |
682 | struct sched_rt_entity *rt_se; | |
683 | struct rt_prio_array *array; | |
684 | struct rt_rq *rt_rq; | |
e8fa1362 SR |
685 | int idx; |
686 | ||
6f505b16 PZ |
687 | for_each_leaf_rt_rq(rt_rq, rq) { |
688 | array = &rt_rq->active; | |
689 | idx = sched_find_first_bit(array->bitmap); | |
690 | next_idx: | |
691 | if (idx >= MAX_RT_PRIO) | |
692 | continue; | |
693 | if (next && next->prio < idx) | |
694 | continue; | |
695 | list_for_each_entry(rt_se, array->queue + idx, run_list) { | |
696 | struct task_struct *p = rt_task_of(rt_se); | |
697 | if (pick_rt_task(rq, p, cpu)) { | |
698 | next = p; | |
699 | break; | |
700 | } | |
701 | } | |
702 | if (!next) { | |
703 | idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1); | |
704 | goto next_idx; | |
705 | } | |
f65eda4f SR |
706 | } |
707 | ||
e8fa1362 SR |
708 | return next; |
709 | } | |
710 | ||
711 | static DEFINE_PER_CPU(cpumask_t, local_cpu_mask); | |
712 | ||
6e1254d2 | 713 | static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask) |
e8fa1362 | 714 | { |
6e1254d2 | 715 | int lowest_prio = -1; |
610bf056 | 716 | int lowest_cpu = -1; |
06f90dbd | 717 | int count = 0; |
610bf056 | 718 | int cpu; |
e8fa1362 | 719 | |
637f5085 | 720 | cpus_and(*lowest_mask, task_rq(task)->rd->online, task->cpus_allowed); |
e8fa1362 | 721 | |
07b4032c GH |
722 | /* |
723 | * Scan each rq for the lowest prio. | |
724 | */ | |
610bf056 | 725 | for_each_cpu_mask(cpu, *lowest_mask) { |
07b4032c | 726 | struct rq *rq = cpu_rq(cpu); |
e8fa1362 | 727 | |
07b4032c GH |
728 | /* We look for lowest RT prio or non-rt CPU */ |
729 | if (rq->rt.highest_prio >= MAX_RT_PRIO) { | |
610bf056 SR |
730 | /* |
731 | * if we already found a low RT queue | |
732 | * and now we found this non-rt queue | |
733 | * clear the mask and set our bit. | |
734 | * Otherwise just return the queue as is | |
735 | * and the count==1 will cause the algorithm | |
736 | * to use the first bit found. | |
737 | */ | |
738 | if (lowest_cpu != -1) { | |
6e1254d2 | 739 | cpus_clear(*lowest_mask); |
610bf056 SR |
740 | cpu_set(rq->cpu, *lowest_mask); |
741 | } | |
6e1254d2 | 742 | return 1; |
07b4032c GH |
743 | } |
744 | ||
745 | /* no locking for now */ | |
6e1254d2 GH |
746 | if ((rq->rt.highest_prio > task->prio) |
747 | && (rq->rt.highest_prio >= lowest_prio)) { | |
748 | if (rq->rt.highest_prio > lowest_prio) { | |
749 | /* new low - clear old data */ | |
750 | lowest_prio = rq->rt.highest_prio; | |
610bf056 SR |
751 | lowest_cpu = cpu; |
752 | count = 0; | |
6e1254d2 | 753 | } |
06f90dbd | 754 | count++; |
610bf056 SR |
755 | } else |
756 | cpu_clear(cpu, *lowest_mask); | |
757 | } | |
758 | ||
759 | /* | |
760 | * Clear out all the set bits that represent | |
761 | * runqueues that were of higher prio than | |
762 | * the lowest_prio. | |
763 | */ | |
764 | if (lowest_cpu > 0) { | |
765 | /* | |
766 | * Perhaps we could add another cpumask op to | |
767 | * zero out bits. Like cpu_zero_bits(cpumask, nrbits); | |
768 | * Then that could be optimized to use memset and such. | |
769 | */ | |
770 | for_each_cpu_mask(cpu, *lowest_mask) { | |
771 | if (cpu >= lowest_cpu) | |
772 | break; | |
773 | cpu_clear(cpu, *lowest_mask); | |
e8fa1362 | 774 | } |
07b4032c GH |
775 | } |
776 | ||
06f90dbd | 777 | return count; |
6e1254d2 GH |
778 | } |
779 | ||
780 | static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask) | |
781 | { | |
782 | int first; | |
783 | ||
784 | /* "this_cpu" is cheaper to preempt than a remote processor */ | |
785 | if ((this_cpu != -1) && cpu_isset(this_cpu, *mask)) | |
786 | return this_cpu; | |
787 | ||
788 | first = first_cpu(*mask); | |
789 | if (first != NR_CPUS) | |
790 | return first; | |
791 | ||
792 | return -1; | |
793 | } | |
794 | ||
795 | static int find_lowest_rq(struct task_struct *task) | |
796 | { | |
797 | struct sched_domain *sd; | |
798 | cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask); | |
799 | int this_cpu = smp_processor_id(); | |
800 | int cpu = task_cpu(task); | |
06f90dbd GH |
801 | int count = find_lowest_cpus(task, lowest_mask); |
802 | ||
803 | if (!count) | |
804 | return -1; /* No targets found */ | |
6e1254d2 | 805 | |
06f90dbd GH |
806 | /* |
807 | * There is no sense in performing an optimal search if only one | |
808 | * target is found. | |
809 | */ | |
810 | if (count == 1) | |
811 | return first_cpu(*lowest_mask); | |
6e1254d2 GH |
812 | |
813 | /* | |
814 | * At this point we have built a mask of cpus representing the | |
815 | * lowest priority tasks in the system. Now we want to elect | |
816 | * the best one based on our affinity and topology. | |
817 | * | |
818 | * We prioritize the last cpu that the task executed on since | |
819 | * it is most likely cache-hot in that location. | |
820 | */ | |
821 | if (cpu_isset(cpu, *lowest_mask)) | |
822 | return cpu; | |
823 | ||
824 | /* | |
825 | * Otherwise, we consult the sched_domains span maps to figure | |
826 | * out which cpu is logically closest to our hot cache data. | |
827 | */ | |
828 | if (this_cpu == cpu) | |
829 | this_cpu = -1; /* Skip this_cpu opt if the same */ | |
830 | ||
831 | for_each_domain(cpu, sd) { | |
832 | if (sd->flags & SD_WAKE_AFFINE) { | |
833 | cpumask_t domain_mask; | |
834 | int best_cpu; | |
835 | ||
836 | cpus_and(domain_mask, sd->span, *lowest_mask); | |
837 | ||
838 | best_cpu = pick_optimal_cpu(this_cpu, | |
839 | &domain_mask); | |
840 | if (best_cpu != -1) | |
841 | return best_cpu; | |
842 | } | |
843 | } | |
844 | ||
845 | /* | |
846 | * And finally, if there were no matches within the domains | |
847 | * just give the caller *something* to work with from the compatible | |
848 | * locations. | |
849 | */ | |
850 | return pick_optimal_cpu(this_cpu, lowest_mask); | |
07b4032c GH |
851 | } |
852 | ||
853 | /* Will lock the rq it finds */ | |
4df64c0b | 854 | static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) |
07b4032c GH |
855 | { |
856 | struct rq *lowest_rq = NULL; | |
07b4032c | 857 | int tries; |
4df64c0b | 858 | int cpu; |
e8fa1362 | 859 | |
07b4032c GH |
860 | for (tries = 0; tries < RT_MAX_TRIES; tries++) { |
861 | cpu = find_lowest_rq(task); | |
862 | ||
2de0b463 | 863 | if ((cpu == -1) || (cpu == rq->cpu)) |
e8fa1362 SR |
864 | break; |
865 | ||
07b4032c GH |
866 | lowest_rq = cpu_rq(cpu); |
867 | ||
e8fa1362 | 868 | /* if the prio of this runqueue changed, try again */ |
07b4032c | 869 | if (double_lock_balance(rq, lowest_rq)) { |
e8fa1362 SR |
870 | /* |
871 | * We had to unlock the run queue. In | |
872 | * the mean time, task could have | |
873 | * migrated already or had its affinity changed. | |
874 | * Also make sure that it wasn't scheduled on its rq. | |
875 | */ | |
07b4032c | 876 | if (unlikely(task_rq(task) != rq || |
4df64c0b IM |
877 | !cpu_isset(lowest_rq->cpu, |
878 | task->cpus_allowed) || | |
07b4032c | 879 | task_running(rq, task) || |
e8fa1362 | 880 | !task->se.on_rq)) { |
4df64c0b | 881 | |
e8fa1362 SR |
882 | spin_unlock(&lowest_rq->lock); |
883 | lowest_rq = NULL; | |
884 | break; | |
885 | } | |
886 | } | |
887 | ||
888 | /* If this rq is still suitable use it. */ | |
889 | if (lowest_rq->rt.highest_prio > task->prio) | |
890 | break; | |
891 | ||
892 | /* try again */ | |
893 | spin_unlock(&lowest_rq->lock); | |
894 | lowest_rq = NULL; | |
895 | } | |
896 | ||
897 | return lowest_rq; | |
898 | } | |
899 | ||
900 | /* | |
901 | * If the current CPU has more than one RT task, see if the non | |
902 | * running task can migrate over to a CPU that is running a task | |
903 | * of lesser priority. | |
904 | */ | |
697f0a48 | 905 | static int push_rt_task(struct rq *rq) |
e8fa1362 SR |
906 | { |
907 | struct task_struct *next_task; | |
908 | struct rq *lowest_rq; | |
909 | int ret = 0; | |
910 | int paranoid = RT_MAX_TRIES; | |
911 | ||
a22d7fc1 GH |
912 | if (!rq->rt.overloaded) |
913 | return 0; | |
914 | ||
697f0a48 | 915 | next_task = pick_next_highest_task_rt(rq, -1); |
e8fa1362 SR |
916 | if (!next_task) |
917 | return 0; | |
918 | ||
919 | retry: | |
697f0a48 | 920 | if (unlikely(next_task == rq->curr)) { |
f65eda4f | 921 | WARN_ON(1); |
e8fa1362 | 922 | return 0; |
f65eda4f | 923 | } |
e8fa1362 SR |
924 | |
925 | /* | |
926 | * It's possible that the next_task slipped in of | |
927 | * higher priority than current. If that's the case | |
928 | * just reschedule current. | |
929 | */ | |
697f0a48 GH |
930 | if (unlikely(next_task->prio < rq->curr->prio)) { |
931 | resched_task(rq->curr); | |
e8fa1362 SR |
932 | return 0; |
933 | } | |
934 | ||
697f0a48 | 935 | /* We might release rq lock */ |
e8fa1362 SR |
936 | get_task_struct(next_task); |
937 | ||
938 | /* find_lock_lowest_rq locks the rq if found */ | |
697f0a48 | 939 | lowest_rq = find_lock_lowest_rq(next_task, rq); |
e8fa1362 SR |
940 | if (!lowest_rq) { |
941 | struct task_struct *task; | |
942 | /* | |
697f0a48 | 943 | * find lock_lowest_rq releases rq->lock |
e8fa1362 SR |
944 | * so it is possible that next_task has changed. |
945 | * If it has, then try again. | |
946 | */ | |
697f0a48 | 947 | task = pick_next_highest_task_rt(rq, -1); |
e8fa1362 SR |
948 | if (unlikely(task != next_task) && task && paranoid--) { |
949 | put_task_struct(next_task); | |
950 | next_task = task; | |
951 | goto retry; | |
952 | } | |
953 | goto out; | |
954 | } | |
955 | ||
697f0a48 | 956 | deactivate_task(rq, next_task, 0); |
e8fa1362 SR |
957 | set_task_cpu(next_task, lowest_rq->cpu); |
958 | activate_task(lowest_rq, next_task, 0); | |
959 | ||
960 | resched_task(lowest_rq->curr); | |
961 | ||
962 | spin_unlock(&lowest_rq->lock); | |
963 | ||
964 | ret = 1; | |
965 | out: | |
966 | put_task_struct(next_task); | |
967 | ||
968 | return ret; | |
969 | } | |
970 | ||
971 | /* | |
972 | * TODO: Currently we just use the second highest prio task on | |
973 | * the queue, and stop when it can't migrate (or there's | |
974 | * no more RT tasks). There may be a case where a lower | |
975 | * priority RT task has a different affinity than the | |
976 | * higher RT task. In this case the lower RT task could | |
977 | * possibly be able to migrate where as the higher priority | |
978 | * RT task could not. We currently ignore this issue. | |
979 | * Enhancements are welcome! | |
980 | */ | |
981 | static void push_rt_tasks(struct rq *rq) | |
982 | { | |
983 | /* push_rt_task will return true if it moved an RT */ | |
984 | while (push_rt_task(rq)) | |
985 | ; | |
986 | } | |
987 | ||
f65eda4f SR |
988 | static int pull_rt_task(struct rq *this_rq) |
989 | { | |
80bf3171 IM |
990 | int this_cpu = this_rq->cpu, ret = 0, cpu; |
991 | struct task_struct *p, *next; | |
f65eda4f | 992 | struct rq *src_rq; |
f65eda4f | 993 | |
637f5085 | 994 | if (likely(!rt_overloaded(this_rq))) |
f65eda4f SR |
995 | return 0; |
996 | ||
997 | next = pick_next_task_rt(this_rq); | |
998 | ||
637f5085 | 999 | for_each_cpu_mask(cpu, this_rq->rd->rto_mask) { |
f65eda4f SR |
1000 | if (this_cpu == cpu) |
1001 | continue; | |
1002 | ||
1003 | src_rq = cpu_rq(cpu); | |
f65eda4f SR |
1004 | /* |
1005 | * We can potentially drop this_rq's lock in | |
1006 | * double_lock_balance, and another CPU could | |
1007 | * steal our next task - hence we must cause | |
1008 | * the caller to recalculate the next task | |
1009 | * in that case: | |
1010 | */ | |
1011 | if (double_lock_balance(this_rq, src_rq)) { | |
1012 | struct task_struct *old_next = next; | |
80bf3171 | 1013 | |
f65eda4f SR |
1014 | next = pick_next_task_rt(this_rq); |
1015 | if (next != old_next) | |
1016 | ret = 1; | |
1017 | } | |
1018 | ||
1019 | /* | |
1020 | * Are there still pullable RT tasks? | |
1021 | */ | |
614ee1f6 MG |
1022 | if (src_rq->rt.rt_nr_running <= 1) |
1023 | goto skip; | |
f65eda4f | 1024 | |
f65eda4f SR |
1025 | p = pick_next_highest_task_rt(src_rq, this_cpu); |
1026 | ||
1027 | /* | |
1028 | * Do we have an RT task that preempts | |
1029 | * the to-be-scheduled task? | |
1030 | */ | |
1031 | if (p && (!next || (p->prio < next->prio))) { | |
1032 | WARN_ON(p == src_rq->curr); | |
1033 | WARN_ON(!p->se.on_rq); | |
1034 | ||
1035 | /* | |
1036 | * There's a chance that p is higher in priority | |
1037 | * than what's currently running on its cpu. | |
1038 | * This is just that p is wakeing up and hasn't | |
1039 | * had a chance to schedule. We only pull | |
1040 | * p if it is lower in priority than the | |
1041 | * current task on the run queue or | |
1042 | * this_rq next task is lower in prio than | |
1043 | * the current task on that rq. | |
1044 | */ | |
1045 | if (p->prio < src_rq->curr->prio || | |
1046 | (next && next->prio < src_rq->curr->prio)) | |
614ee1f6 | 1047 | goto skip; |
f65eda4f SR |
1048 | |
1049 | ret = 1; | |
1050 | ||
1051 | deactivate_task(src_rq, p, 0); | |
1052 | set_task_cpu(p, this_cpu); | |
1053 | activate_task(this_rq, p, 0); | |
1054 | /* | |
1055 | * We continue with the search, just in | |
1056 | * case there's an even higher prio task | |
1057 | * in another runqueue. (low likelyhood | |
1058 | * but possible) | |
80bf3171 | 1059 | * |
f65eda4f SR |
1060 | * Update next so that we won't pick a task |
1061 | * on another cpu with a priority lower (or equal) | |
1062 | * than the one we just picked. | |
1063 | */ | |
1064 | next = p; | |
1065 | ||
1066 | } | |
614ee1f6 | 1067 | skip: |
f65eda4f SR |
1068 | spin_unlock(&src_rq->lock); |
1069 | } | |
1070 | ||
1071 | return ret; | |
1072 | } | |
1073 | ||
9a897c5a | 1074 | static void pre_schedule_rt(struct rq *rq, struct task_struct *prev) |
f65eda4f SR |
1075 | { |
1076 | /* Try to pull RT tasks here if we lower this rq's prio */ | |
7f51f298 | 1077 | if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio) |
f65eda4f SR |
1078 | pull_rt_task(rq); |
1079 | } | |
1080 | ||
9a897c5a | 1081 | static void post_schedule_rt(struct rq *rq) |
e8fa1362 SR |
1082 | { |
1083 | /* | |
1084 | * If we have more than one rt_task queued, then | |
1085 | * see if we can push the other rt_tasks off to other CPUS. | |
1086 | * Note we may release the rq lock, and since | |
1087 | * the lock was owned by prev, we need to release it | |
1088 | * first via finish_lock_switch and then reaquire it here. | |
1089 | */ | |
a22d7fc1 | 1090 | if (unlikely(rq->rt.overloaded)) { |
e8fa1362 SR |
1091 | spin_lock_irq(&rq->lock); |
1092 | push_rt_tasks(rq); | |
1093 | spin_unlock_irq(&rq->lock); | |
1094 | } | |
1095 | } | |
1096 | ||
8ae121ac GH |
1097 | /* |
1098 | * If we are not running and we are not going to reschedule soon, we should | |
1099 | * try to push tasks away now | |
1100 | */ | |
9a897c5a | 1101 | static void task_wake_up_rt(struct rq *rq, struct task_struct *p) |
4642dafd | 1102 | { |
9a897c5a | 1103 | if (!task_running(rq, p) && |
8ae121ac | 1104 | !test_tsk_need_resched(rq->curr) && |
a22d7fc1 | 1105 | rq->rt.overloaded) |
4642dafd SR |
1106 | push_rt_tasks(rq); |
1107 | } | |
1108 | ||
43010659 | 1109 | static unsigned long |
bb44e5d1 | 1110 | load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, |
e1d1484f PW |
1111 | unsigned long max_load_move, |
1112 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1113 | int *all_pinned, int *this_best_prio) | |
bb44e5d1 | 1114 | { |
c7a1e46a SR |
1115 | /* don't touch RT tasks */ |
1116 | return 0; | |
e1d1484f PW |
1117 | } |
1118 | ||
1119 | static int | |
1120 | move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1121 | struct sched_domain *sd, enum cpu_idle_type idle) | |
1122 | { | |
c7a1e46a SR |
1123 | /* don't touch RT tasks */ |
1124 | return 0; | |
bb44e5d1 | 1125 | } |
deeeccd4 | 1126 | |
cd8ba7cd MT |
1127 | static void set_cpus_allowed_rt(struct task_struct *p, |
1128 | const cpumask_t *new_mask) | |
73fe6aae GH |
1129 | { |
1130 | int weight = cpus_weight(*new_mask); | |
1131 | ||
1132 | BUG_ON(!rt_task(p)); | |
1133 | ||
1134 | /* | |
1135 | * Update the migration status of the RQ if we have an RT task | |
1136 | * which is running AND changing its weight value. | |
1137 | */ | |
6f505b16 | 1138 | if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) { |
73fe6aae GH |
1139 | struct rq *rq = task_rq(p); |
1140 | ||
6f505b16 | 1141 | if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) { |
73fe6aae | 1142 | rq->rt.rt_nr_migratory++; |
6f505b16 | 1143 | } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) { |
73fe6aae GH |
1144 | BUG_ON(!rq->rt.rt_nr_migratory); |
1145 | rq->rt.rt_nr_migratory--; | |
1146 | } | |
1147 | ||
1148 | update_rt_migration(rq); | |
1149 | } | |
1150 | ||
1151 | p->cpus_allowed = *new_mask; | |
6f505b16 | 1152 | p->rt.nr_cpus_allowed = weight; |
73fe6aae | 1153 | } |
deeeccd4 | 1154 | |
bdd7c81b IM |
1155 | /* Assumes rq->lock is held */ |
1156 | static void join_domain_rt(struct rq *rq) | |
1157 | { | |
1158 | if (rq->rt.overloaded) | |
1159 | rt_set_overload(rq); | |
1160 | } | |
1161 | ||
1162 | /* Assumes rq->lock is held */ | |
1163 | static void leave_domain_rt(struct rq *rq) | |
1164 | { | |
1165 | if (rq->rt.overloaded) | |
1166 | rt_clear_overload(rq); | |
1167 | } | |
cb469845 SR |
1168 | |
1169 | /* | |
1170 | * When switch from the rt queue, we bring ourselves to a position | |
1171 | * that we might want to pull RT tasks from other runqueues. | |
1172 | */ | |
1173 | static void switched_from_rt(struct rq *rq, struct task_struct *p, | |
1174 | int running) | |
1175 | { | |
1176 | /* | |
1177 | * If there are other RT tasks then we will reschedule | |
1178 | * and the scheduling of the other RT tasks will handle | |
1179 | * the balancing. But if we are the last RT task | |
1180 | * we may need to handle the pulling of RT tasks | |
1181 | * now. | |
1182 | */ | |
1183 | if (!rq->rt.rt_nr_running) | |
1184 | pull_rt_task(rq); | |
1185 | } | |
1186 | #endif /* CONFIG_SMP */ | |
1187 | ||
1188 | /* | |
1189 | * When switching a task to RT, we may overload the runqueue | |
1190 | * with RT tasks. In this case we try to push them off to | |
1191 | * other runqueues. | |
1192 | */ | |
1193 | static void switched_to_rt(struct rq *rq, struct task_struct *p, | |
1194 | int running) | |
1195 | { | |
1196 | int check_resched = 1; | |
1197 | ||
1198 | /* | |
1199 | * If we are already running, then there's nothing | |
1200 | * that needs to be done. But if we are not running | |
1201 | * we may need to preempt the current running task. | |
1202 | * If that current running task is also an RT task | |
1203 | * then see if we can move to another run queue. | |
1204 | */ | |
1205 | if (!running) { | |
1206 | #ifdef CONFIG_SMP | |
1207 | if (rq->rt.overloaded && push_rt_task(rq) && | |
1208 | /* Don't resched if we changed runqueues */ | |
1209 | rq != task_rq(p)) | |
1210 | check_resched = 0; | |
1211 | #endif /* CONFIG_SMP */ | |
1212 | if (check_resched && p->prio < rq->curr->prio) | |
1213 | resched_task(rq->curr); | |
1214 | } | |
1215 | } | |
1216 | ||
1217 | /* | |
1218 | * Priority of the task has changed. This may cause | |
1219 | * us to initiate a push or pull. | |
1220 | */ | |
1221 | static void prio_changed_rt(struct rq *rq, struct task_struct *p, | |
1222 | int oldprio, int running) | |
1223 | { | |
1224 | if (running) { | |
1225 | #ifdef CONFIG_SMP | |
1226 | /* | |
1227 | * If our priority decreases while running, we | |
1228 | * may need to pull tasks to this runqueue. | |
1229 | */ | |
1230 | if (oldprio < p->prio) | |
1231 | pull_rt_task(rq); | |
1232 | /* | |
1233 | * If there's a higher priority task waiting to run | |
6fa46fa5 SR |
1234 | * then reschedule. Note, the above pull_rt_task |
1235 | * can release the rq lock and p could migrate. | |
1236 | * Only reschedule if p is still on the same runqueue. | |
cb469845 | 1237 | */ |
6fa46fa5 | 1238 | if (p->prio > rq->rt.highest_prio && rq->curr == p) |
cb469845 SR |
1239 | resched_task(p); |
1240 | #else | |
1241 | /* For UP simply resched on drop of prio */ | |
1242 | if (oldprio < p->prio) | |
1243 | resched_task(p); | |
e8fa1362 | 1244 | #endif /* CONFIG_SMP */ |
cb469845 SR |
1245 | } else { |
1246 | /* | |
1247 | * This task is not running, but if it is | |
1248 | * greater than the current running task | |
1249 | * then reschedule. | |
1250 | */ | |
1251 | if (p->prio < rq->curr->prio) | |
1252 | resched_task(rq->curr); | |
1253 | } | |
1254 | } | |
1255 | ||
78f2c7db PZ |
1256 | static void watchdog(struct rq *rq, struct task_struct *p) |
1257 | { | |
1258 | unsigned long soft, hard; | |
1259 | ||
1260 | if (!p->signal) | |
1261 | return; | |
1262 | ||
1263 | soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur; | |
1264 | hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max; | |
1265 | ||
1266 | if (soft != RLIM_INFINITY) { | |
1267 | unsigned long next; | |
1268 | ||
1269 | p->rt.timeout++; | |
1270 | next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ); | |
5a52dd50 | 1271 | if (p->rt.timeout > next) |
78f2c7db PZ |
1272 | p->it_sched_expires = p->se.sum_exec_runtime; |
1273 | } | |
1274 | } | |
bb44e5d1 | 1275 | |
8f4d37ec | 1276 | static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued) |
bb44e5d1 | 1277 | { |
67e2be02 PZ |
1278 | update_curr_rt(rq); |
1279 | ||
78f2c7db PZ |
1280 | watchdog(rq, p); |
1281 | ||
bb44e5d1 IM |
1282 | /* |
1283 | * RR tasks need a special form of timeslice management. | |
1284 | * FIFO tasks have no timeslices. | |
1285 | */ | |
1286 | if (p->policy != SCHED_RR) | |
1287 | return; | |
1288 | ||
fa717060 | 1289 | if (--p->rt.time_slice) |
bb44e5d1 IM |
1290 | return; |
1291 | ||
fa717060 | 1292 | p->rt.time_slice = DEF_TIMESLICE; |
bb44e5d1 | 1293 | |
98fbc798 DA |
1294 | /* |
1295 | * Requeue to the end of queue if we are not the only element | |
1296 | * on the queue: | |
1297 | */ | |
fa717060 | 1298 | if (p->rt.run_list.prev != p->rt.run_list.next) { |
98fbc798 DA |
1299 | requeue_task_rt(rq, p); |
1300 | set_tsk_need_resched(p); | |
1301 | } | |
bb44e5d1 IM |
1302 | } |
1303 | ||
83b699ed SV |
1304 | static void set_curr_task_rt(struct rq *rq) |
1305 | { | |
1306 | struct task_struct *p = rq->curr; | |
1307 | ||
1308 | p->se.exec_start = rq->clock; | |
1309 | } | |
1310 | ||
2abdad0a | 1311 | static const struct sched_class rt_sched_class = { |
5522d5d5 | 1312 | .next = &fair_sched_class, |
bb44e5d1 IM |
1313 | .enqueue_task = enqueue_task_rt, |
1314 | .dequeue_task = dequeue_task_rt, | |
1315 | .yield_task = yield_task_rt, | |
e7693a36 GH |
1316 | #ifdef CONFIG_SMP |
1317 | .select_task_rq = select_task_rq_rt, | |
1318 | #endif /* CONFIG_SMP */ | |
bb44e5d1 IM |
1319 | |
1320 | .check_preempt_curr = check_preempt_curr_rt, | |
1321 | ||
1322 | .pick_next_task = pick_next_task_rt, | |
1323 | .put_prev_task = put_prev_task_rt, | |
1324 | ||
681f3e68 | 1325 | #ifdef CONFIG_SMP |
bb44e5d1 | 1326 | .load_balance = load_balance_rt, |
e1d1484f | 1327 | .move_one_task = move_one_task_rt, |
73fe6aae | 1328 | .set_cpus_allowed = set_cpus_allowed_rt, |
bdd7c81b IM |
1329 | .join_domain = join_domain_rt, |
1330 | .leave_domain = leave_domain_rt, | |
9a897c5a SR |
1331 | .pre_schedule = pre_schedule_rt, |
1332 | .post_schedule = post_schedule_rt, | |
1333 | .task_wake_up = task_wake_up_rt, | |
cb469845 | 1334 | .switched_from = switched_from_rt, |
681f3e68 | 1335 | #endif |
bb44e5d1 | 1336 | |
83b699ed | 1337 | .set_curr_task = set_curr_task_rt, |
bb44e5d1 | 1338 | .task_tick = task_tick_rt, |
cb469845 SR |
1339 | |
1340 | .prio_changed = prio_changed_rt, | |
1341 | .switched_to = switched_to_rt, | |
bb44e5d1 | 1342 | }; |