]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
bb44e5d1 IM |
2 | /* |
3 | * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR | |
4 | * policies) | |
5 | */ | |
6 | ||
029632fb PZ |
7 | #include "sched.h" |
8 | ||
9 | #include <linux/slab.h> | |
b6366f04 | 10 | #include <linux/irq_work.h> |
029632fb | 11 | |
ce0dbbbb | 12 | int sched_rr_timeslice = RR_TIMESLICE; |
975e155e | 13 | int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE; |
ce0dbbbb | 14 | |
029632fb PZ |
15 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); |
16 | ||
17 | struct rt_bandwidth def_rt_bandwidth; | |
18 | ||
19 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) | |
20 | { | |
21 | struct rt_bandwidth *rt_b = | |
22 | container_of(timer, struct rt_bandwidth, rt_period_timer); | |
029632fb | 23 | int idle = 0; |
77a4d1a1 | 24 | int overrun; |
029632fb | 25 | |
77a4d1a1 | 26 | raw_spin_lock(&rt_b->rt_runtime_lock); |
029632fb | 27 | for (;;) { |
77a4d1a1 | 28 | overrun = hrtimer_forward_now(timer, rt_b->rt_period); |
029632fb PZ |
29 | if (!overrun) |
30 | break; | |
31 | ||
77a4d1a1 | 32 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
029632fb | 33 | idle = do_sched_rt_period_timer(rt_b, overrun); |
77a4d1a1 | 34 | raw_spin_lock(&rt_b->rt_runtime_lock); |
029632fb | 35 | } |
4cfafd30 PZ |
36 | if (idle) |
37 | rt_b->rt_period_active = 0; | |
77a4d1a1 | 38 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
029632fb PZ |
39 | |
40 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
41 | } | |
42 | ||
43 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) | |
44 | { | |
45 | rt_b->rt_period = ns_to_ktime(period); | |
46 | rt_b->rt_runtime = runtime; | |
47 | ||
48 | raw_spin_lock_init(&rt_b->rt_runtime_lock); | |
49 | ||
50 | hrtimer_init(&rt_b->rt_period_timer, | |
51 | CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
52 | rt_b->rt_period_timer.function = sched_rt_period_timer; | |
53 | } | |
54 | ||
55 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b) | |
56 | { | |
57 | if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF) | |
58 | return; | |
59 | ||
029632fb | 60 | raw_spin_lock(&rt_b->rt_runtime_lock); |
4cfafd30 PZ |
61 | if (!rt_b->rt_period_active) { |
62 | rt_b->rt_period_active = 1; | |
c3a990dc SR |
63 | /* |
64 | * SCHED_DEADLINE updates the bandwidth, as a run away | |
65 | * RT task with a DL task could hog a CPU. But DL does | |
66 | * not reset the period. If a deadline task was running | |
67 | * without an RT task running, it can cause RT tasks to | |
68 | * throttle when they start up. Kick the timer right away | |
69 | * to update the period. | |
70 | */ | |
71 | hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0)); | |
4cfafd30 PZ |
72 | hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED); |
73 | } | |
029632fb PZ |
74 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
75 | } | |
76 | ||
07c54f7a | 77 | void init_rt_rq(struct rt_rq *rt_rq) |
029632fb PZ |
78 | { |
79 | struct rt_prio_array *array; | |
80 | int i; | |
81 | ||
82 | array = &rt_rq->active; | |
83 | for (i = 0; i < MAX_RT_PRIO; i++) { | |
84 | INIT_LIST_HEAD(array->queue + i); | |
85 | __clear_bit(i, array->bitmap); | |
86 | } | |
87 | /* delimiter for bitsearch: */ | |
88 | __set_bit(MAX_RT_PRIO, array->bitmap); | |
89 | ||
90 | #if defined CONFIG_SMP | |
91 | rt_rq->highest_prio.curr = MAX_RT_PRIO; | |
92 | rt_rq->highest_prio.next = MAX_RT_PRIO; | |
93 | rt_rq->rt_nr_migratory = 0; | |
94 | rt_rq->overloaded = 0; | |
95 | plist_head_init(&rt_rq->pushable_tasks); | |
b6366f04 | 96 | #endif /* CONFIG_SMP */ |
f4ebcbc0 KT |
97 | /* We start is dequeued state, because no RT tasks are queued */ |
98 | rt_rq->rt_queued = 0; | |
029632fb PZ |
99 | |
100 | rt_rq->rt_time = 0; | |
101 | rt_rq->rt_throttled = 0; | |
102 | rt_rq->rt_runtime = 0; | |
103 | raw_spin_lock_init(&rt_rq->rt_runtime_lock); | |
104 | } | |
105 | ||
8f48894f | 106 | #ifdef CONFIG_RT_GROUP_SCHED |
029632fb PZ |
107 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) |
108 | { | |
109 | hrtimer_cancel(&rt_b->rt_period_timer); | |
110 | } | |
8f48894f PZ |
111 | |
112 | #define rt_entity_is_task(rt_se) (!(rt_se)->my_q) | |
113 | ||
398a153b GH |
114 | static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) |
115 | { | |
8f48894f PZ |
116 | #ifdef CONFIG_SCHED_DEBUG |
117 | WARN_ON_ONCE(!rt_entity_is_task(rt_se)); | |
118 | #endif | |
398a153b GH |
119 | return container_of(rt_se, struct task_struct, rt); |
120 | } | |
121 | ||
398a153b GH |
122 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) |
123 | { | |
124 | return rt_rq->rq; | |
125 | } | |
126 | ||
127 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) | |
128 | { | |
129 | return rt_se->rt_rq; | |
130 | } | |
131 | ||
653d07a6 KT |
132 | static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se) |
133 | { | |
134 | struct rt_rq *rt_rq = rt_se->rt_rq; | |
135 | ||
136 | return rt_rq->rq; | |
137 | } | |
138 | ||
029632fb PZ |
139 | void free_rt_sched_group(struct task_group *tg) |
140 | { | |
141 | int i; | |
142 | ||
143 | if (tg->rt_se) | |
144 | destroy_rt_bandwidth(&tg->rt_bandwidth); | |
145 | ||
146 | for_each_possible_cpu(i) { | |
147 | if (tg->rt_rq) | |
148 | kfree(tg->rt_rq[i]); | |
149 | if (tg->rt_se) | |
150 | kfree(tg->rt_se[i]); | |
151 | } | |
152 | ||
153 | kfree(tg->rt_rq); | |
154 | kfree(tg->rt_se); | |
155 | } | |
156 | ||
157 | void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, | |
158 | struct sched_rt_entity *rt_se, int cpu, | |
159 | struct sched_rt_entity *parent) | |
160 | { | |
161 | struct rq *rq = cpu_rq(cpu); | |
162 | ||
163 | rt_rq->highest_prio.curr = MAX_RT_PRIO; | |
164 | rt_rq->rt_nr_boosted = 0; | |
165 | rt_rq->rq = rq; | |
166 | rt_rq->tg = tg; | |
167 | ||
168 | tg->rt_rq[cpu] = rt_rq; | |
169 | tg->rt_se[cpu] = rt_se; | |
170 | ||
171 | if (!rt_se) | |
172 | return; | |
173 | ||
174 | if (!parent) | |
175 | rt_se->rt_rq = &rq->rt; | |
176 | else | |
177 | rt_se->rt_rq = parent->my_q; | |
178 | ||
179 | rt_se->my_q = rt_rq; | |
180 | rt_se->parent = parent; | |
181 | INIT_LIST_HEAD(&rt_se->run_list); | |
182 | } | |
183 | ||
184 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
185 | { | |
186 | struct rt_rq *rt_rq; | |
187 | struct sched_rt_entity *rt_se; | |
188 | int i; | |
189 | ||
190 | tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); | |
191 | if (!tg->rt_rq) | |
192 | goto err; | |
193 | tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); | |
194 | if (!tg->rt_se) | |
195 | goto err; | |
196 | ||
197 | init_rt_bandwidth(&tg->rt_bandwidth, | |
198 | ktime_to_ns(def_rt_bandwidth.rt_period), 0); | |
199 | ||
200 | for_each_possible_cpu(i) { | |
201 | rt_rq = kzalloc_node(sizeof(struct rt_rq), | |
202 | GFP_KERNEL, cpu_to_node(i)); | |
203 | if (!rt_rq) | |
204 | goto err; | |
205 | ||
206 | rt_se = kzalloc_node(sizeof(struct sched_rt_entity), | |
207 | GFP_KERNEL, cpu_to_node(i)); | |
208 | if (!rt_se) | |
209 | goto err_free_rq; | |
210 | ||
07c54f7a | 211 | init_rt_rq(rt_rq); |
029632fb PZ |
212 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; |
213 | init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]); | |
214 | } | |
215 | ||
216 | return 1; | |
217 | ||
218 | err_free_rq: | |
219 | kfree(rt_rq); | |
220 | err: | |
221 | return 0; | |
222 | } | |
223 | ||
398a153b GH |
224 | #else /* CONFIG_RT_GROUP_SCHED */ |
225 | ||
a1ba4d8b PZ |
226 | #define rt_entity_is_task(rt_se) (1) |
227 | ||
8f48894f PZ |
228 | static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) |
229 | { | |
230 | return container_of(rt_se, struct task_struct, rt); | |
231 | } | |
232 | ||
398a153b GH |
233 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) |
234 | { | |
235 | return container_of(rt_rq, struct rq, rt); | |
236 | } | |
237 | ||
653d07a6 | 238 | static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se) |
398a153b GH |
239 | { |
240 | struct task_struct *p = rt_task_of(rt_se); | |
653d07a6 KT |
241 | |
242 | return task_rq(p); | |
243 | } | |
244 | ||
245 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) | |
246 | { | |
247 | struct rq *rq = rq_of_rt_se(rt_se); | |
398a153b GH |
248 | |
249 | return &rq->rt; | |
250 | } | |
251 | ||
029632fb PZ |
252 | void free_rt_sched_group(struct task_group *tg) { } |
253 | ||
254 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
255 | { | |
256 | return 1; | |
257 | } | |
398a153b GH |
258 | #endif /* CONFIG_RT_GROUP_SCHED */ |
259 | ||
4fd29176 | 260 | #ifdef CONFIG_SMP |
84de4274 | 261 | |
8046d680 | 262 | static void pull_rt_task(struct rq *this_rq); |
38033c37 | 263 | |
dc877341 PZ |
264 | static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev) |
265 | { | |
266 | /* Try to pull RT tasks here if we lower this rq's prio */ | |
267 | return rq->rt.highest_prio.curr > prev->prio; | |
268 | } | |
269 | ||
637f5085 | 270 | static inline int rt_overloaded(struct rq *rq) |
4fd29176 | 271 | { |
637f5085 | 272 | return atomic_read(&rq->rd->rto_count); |
4fd29176 | 273 | } |
84de4274 | 274 | |
4fd29176 SR |
275 | static inline void rt_set_overload(struct rq *rq) |
276 | { | |
1f11eb6a GH |
277 | if (!rq->online) |
278 | return; | |
279 | ||
c6c4927b | 280 | cpumask_set_cpu(rq->cpu, rq->rd->rto_mask); |
4fd29176 SR |
281 | /* |
282 | * Make sure the mask is visible before we set | |
283 | * the overload count. That is checked to determine | |
284 | * if we should look at the mask. It would be a shame | |
285 | * if we looked at the mask, but the mask was not | |
286 | * updated yet. | |
7c3f2ab7 PZ |
287 | * |
288 | * Matched by the barrier in pull_rt_task(). | |
4fd29176 | 289 | */ |
7c3f2ab7 | 290 | smp_wmb(); |
637f5085 | 291 | atomic_inc(&rq->rd->rto_count); |
4fd29176 | 292 | } |
84de4274 | 293 | |
4fd29176 SR |
294 | static inline void rt_clear_overload(struct rq *rq) |
295 | { | |
1f11eb6a GH |
296 | if (!rq->online) |
297 | return; | |
298 | ||
4fd29176 | 299 | /* the order here really doesn't matter */ |
637f5085 | 300 | atomic_dec(&rq->rd->rto_count); |
c6c4927b | 301 | cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask); |
4fd29176 | 302 | } |
73fe6aae | 303 | |
398a153b | 304 | static void update_rt_migration(struct rt_rq *rt_rq) |
73fe6aae | 305 | { |
a1ba4d8b | 306 | if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) { |
398a153b GH |
307 | if (!rt_rq->overloaded) { |
308 | rt_set_overload(rq_of_rt_rq(rt_rq)); | |
309 | rt_rq->overloaded = 1; | |
cdc8eb98 | 310 | } |
398a153b GH |
311 | } else if (rt_rq->overloaded) { |
312 | rt_clear_overload(rq_of_rt_rq(rt_rq)); | |
313 | rt_rq->overloaded = 0; | |
637f5085 | 314 | } |
73fe6aae | 315 | } |
4fd29176 | 316 | |
398a153b GH |
317 | static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
318 | { | |
29baa747 PZ |
319 | struct task_struct *p; |
320 | ||
a1ba4d8b PZ |
321 | if (!rt_entity_is_task(rt_se)) |
322 | return; | |
323 | ||
29baa747 | 324 | p = rt_task_of(rt_se); |
a1ba4d8b PZ |
325 | rt_rq = &rq_of_rt_rq(rt_rq)->rt; |
326 | ||
327 | rt_rq->rt_nr_total++; | |
4b53a341 | 328 | if (p->nr_cpus_allowed > 1) |
398a153b GH |
329 | rt_rq->rt_nr_migratory++; |
330 | ||
331 | update_rt_migration(rt_rq); | |
332 | } | |
333 | ||
334 | static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
335 | { | |
29baa747 PZ |
336 | struct task_struct *p; |
337 | ||
a1ba4d8b PZ |
338 | if (!rt_entity_is_task(rt_se)) |
339 | return; | |
340 | ||
29baa747 | 341 | p = rt_task_of(rt_se); |
a1ba4d8b PZ |
342 | rt_rq = &rq_of_rt_rq(rt_rq)->rt; |
343 | ||
344 | rt_rq->rt_nr_total--; | |
4b53a341 | 345 | if (p->nr_cpus_allowed > 1) |
398a153b GH |
346 | rt_rq->rt_nr_migratory--; |
347 | ||
348 | update_rt_migration(rt_rq); | |
349 | } | |
350 | ||
5181f4a4 SR |
351 | static inline int has_pushable_tasks(struct rq *rq) |
352 | { | |
353 | return !plist_head_empty(&rq->rt.pushable_tasks); | |
354 | } | |
355 | ||
fd7a4bed PZ |
356 | static DEFINE_PER_CPU(struct callback_head, rt_push_head); |
357 | static DEFINE_PER_CPU(struct callback_head, rt_pull_head); | |
e3fca9e7 PZ |
358 | |
359 | static void push_rt_tasks(struct rq *); | |
fd7a4bed | 360 | static void pull_rt_task(struct rq *); |
e3fca9e7 PZ |
361 | |
362 | static inline void queue_push_tasks(struct rq *rq) | |
dc877341 | 363 | { |
e3fca9e7 PZ |
364 | if (!has_pushable_tasks(rq)) |
365 | return; | |
366 | ||
fd7a4bed PZ |
367 | queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks); |
368 | } | |
369 | ||
370 | static inline void queue_pull_task(struct rq *rq) | |
371 | { | |
372 | queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task); | |
dc877341 PZ |
373 | } |
374 | ||
917b627d GH |
375 | static void enqueue_pushable_task(struct rq *rq, struct task_struct *p) |
376 | { | |
377 | plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); | |
378 | plist_node_init(&p->pushable_tasks, p->prio); | |
379 | plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks); | |
5181f4a4 SR |
380 | |
381 | /* Update the highest prio pushable task */ | |
382 | if (p->prio < rq->rt.highest_prio.next) | |
383 | rq->rt.highest_prio.next = p->prio; | |
917b627d GH |
384 | } |
385 | ||
386 | static void dequeue_pushable_task(struct rq *rq, struct task_struct *p) | |
387 | { | |
388 | plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); | |
917b627d | 389 | |
5181f4a4 SR |
390 | /* Update the new highest prio pushable task */ |
391 | if (has_pushable_tasks(rq)) { | |
392 | p = plist_first_entry(&rq->rt.pushable_tasks, | |
393 | struct task_struct, pushable_tasks); | |
394 | rq->rt.highest_prio.next = p->prio; | |
395 | } else | |
396 | rq->rt.highest_prio.next = MAX_RT_PRIO; | |
bcf08df3 IM |
397 | } |
398 | ||
917b627d GH |
399 | #else |
400 | ||
ceacc2c1 | 401 | static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p) |
fa85ae24 | 402 | { |
6f505b16 PZ |
403 | } |
404 | ||
ceacc2c1 PZ |
405 | static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p) |
406 | { | |
407 | } | |
408 | ||
b07430ac | 409 | static inline |
ceacc2c1 PZ |
410 | void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
411 | { | |
412 | } | |
413 | ||
398a153b | 414 | static inline |
ceacc2c1 PZ |
415 | void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
416 | { | |
417 | } | |
917b627d | 418 | |
dc877341 PZ |
419 | static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev) |
420 | { | |
421 | return false; | |
422 | } | |
423 | ||
8046d680 | 424 | static inline void pull_rt_task(struct rq *this_rq) |
dc877341 | 425 | { |
dc877341 PZ |
426 | } |
427 | ||
e3fca9e7 | 428 | static inline void queue_push_tasks(struct rq *rq) |
dc877341 PZ |
429 | { |
430 | } | |
4fd29176 SR |
431 | #endif /* CONFIG_SMP */ |
432 | ||
f4ebcbc0 KT |
433 | static void enqueue_top_rt_rq(struct rt_rq *rt_rq); |
434 | static void dequeue_top_rt_rq(struct rt_rq *rt_rq); | |
435 | ||
6f505b16 PZ |
436 | static inline int on_rt_rq(struct sched_rt_entity *rt_se) |
437 | { | |
ff77e468 | 438 | return rt_se->on_rq; |
6f505b16 PZ |
439 | } |
440 | ||
052f1dc7 | 441 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 442 | |
9f0c1e56 | 443 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) |
6f505b16 PZ |
444 | { |
445 | if (!rt_rq->tg) | |
9f0c1e56 | 446 | return RUNTIME_INF; |
6f505b16 | 447 | |
ac086bc2 PZ |
448 | return rt_rq->rt_runtime; |
449 | } | |
450 | ||
451 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) | |
452 | { | |
453 | return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period); | |
6f505b16 PZ |
454 | } |
455 | ||
ec514c48 CX |
456 | typedef struct task_group *rt_rq_iter_t; |
457 | ||
1c09ab0d YZ |
458 | static inline struct task_group *next_task_group(struct task_group *tg) |
459 | { | |
460 | do { | |
461 | tg = list_entry_rcu(tg->list.next, | |
462 | typeof(struct task_group), list); | |
463 | } while (&tg->list != &task_groups && task_group_is_autogroup(tg)); | |
464 | ||
465 | if (&tg->list == &task_groups) | |
466 | tg = NULL; | |
467 | ||
468 | return tg; | |
469 | } | |
470 | ||
471 | #define for_each_rt_rq(rt_rq, iter, rq) \ | |
472 | for (iter = container_of(&task_groups, typeof(*iter), list); \ | |
473 | (iter = next_task_group(iter)) && \ | |
474 | (rt_rq = iter->rt_rq[cpu_of(rq)]);) | |
ec514c48 | 475 | |
6f505b16 PZ |
476 | #define for_each_sched_rt_entity(rt_se) \ |
477 | for (; rt_se; rt_se = rt_se->parent) | |
478 | ||
479 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) | |
480 | { | |
481 | return rt_se->my_q; | |
482 | } | |
483 | ||
ff77e468 PZ |
484 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags); |
485 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags); | |
6f505b16 | 486 | |
9f0c1e56 | 487 | static void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
6f505b16 | 488 | { |
f6121f4f | 489 | struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr; |
8875125e | 490 | struct rq *rq = rq_of_rt_rq(rt_rq); |
74b7eb58 YZ |
491 | struct sched_rt_entity *rt_se; |
492 | ||
8875125e | 493 | int cpu = cpu_of(rq); |
0c3b9168 BS |
494 | |
495 | rt_se = rt_rq->tg->rt_se[cpu]; | |
6f505b16 | 496 | |
f6121f4f | 497 | if (rt_rq->rt_nr_running) { |
f4ebcbc0 KT |
498 | if (!rt_se) |
499 | enqueue_top_rt_rq(rt_rq); | |
500 | else if (!on_rt_rq(rt_se)) | |
ff77e468 | 501 | enqueue_rt_entity(rt_se, 0); |
f4ebcbc0 | 502 | |
e864c499 | 503 | if (rt_rq->highest_prio.curr < curr->prio) |
8875125e | 504 | resched_curr(rq); |
6f505b16 PZ |
505 | } |
506 | } | |
507 | ||
9f0c1e56 | 508 | static void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
6f505b16 | 509 | { |
74b7eb58 | 510 | struct sched_rt_entity *rt_se; |
0c3b9168 | 511 | int cpu = cpu_of(rq_of_rt_rq(rt_rq)); |
74b7eb58 | 512 | |
0c3b9168 | 513 | rt_se = rt_rq->tg->rt_se[cpu]; |
6f505b16 | 514 | |
f4ebcbc0 KT |
515 | if (!rt_se) |
516 | dequeue_top_rt_rq(rt_rq); | |
517 | else if (on_rt_rq(rt_se)) | |
ff77e468 | 518 | dequeue_rt_entity(rt_se, 0); |
6f505b16 PZ |
519 | } |
520 | ||
46383648 KT |
521 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
522 | { | |
523 | return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted; | |
524 | } | |
525 | ||
23b0fdfc PZ |
526 | static int rt_se_boosted(struct sched_rt_entity *rt_se) |
527 | { | |
528 | struct rt_rq *rt_rq = group_rt_rq(rt_se); | |
529 | struct task_struct *p; | |
530 | ||
531 | if (rt_rq) | |
532 | return !!rt_rq->rt_nr_boosted; | |
533 | ||
534 | p = rt_task_of(rt_se); | |
535 | return p->prio != p->normal_prio; | |
536 | } | |
537 | ||
d0b27fa7 | 538 | #ifdef CONFIG_SMP |
c6c4927b | 539 | static inline const struct cpumask *sched_rt_period_mask(void) |
d0b27fa7 | 540 | { |
424c93fe | 541 | return this_rq()->rd->span; |
d0b27fa7 | 542 | } |
6f505b16 | 543 | #else |
c6c4927b | 544 | static inline const struct cpumask *sched_rt_period_mask(void) |
d0b27fa7 | 545 | { |
c6c4927b | 546 | return cpu_online_mask; |
d0b27fa7 PZ |
547 | } |
548 | #endif | |
6f505b16 | 549 | |
d0b27fa7 PZ |
550 | static inline |
551 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) | |
6f505b16 | 552 | { |
d0b27fa7 PZ |
553 | return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu]; |
554 | } | |
9f0c1e56 | 555 | |
ac086bc2 PZ |
556 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
557 | { | |
558 | return &rt_rq->tg->rt_bandwidth; | |
559 | } | |
560 | ||
55e12e5e | 561 | #else /* !CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 PZ |
562 | |
563 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) | |
564 | { | |
ac086bc2 PZ |
565 | return rt_rq->rt_runtime; |
566 | } | |
567 | ||
568 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) | |
569 | { | |
570 | return ktime_to_ns(def_rt_bandwidth.rt_period); | |
6f505b16 PZ |
571 | } |
572 | ||
ec514c48 CX |
573 | typedef struct rt_rq *rt_rq_iter_t; |
574 | ||
575 | #define for_each_rt_rq(rt_rq, iter, rq) \ | |
576 | for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL) | |
577 | ||
6f505b16 PZ |
578 | #define for_each_sched_rt_entity(rt_se) \ |
579 | for (; rt_se; rt_se = NULL) | |
580 | ||
581 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) | |
582 | { | |
583 | return NULL; | |
584 | } | |
585 | ||
9f0c1e56 | 586 | static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
6f505b16 | 587 | { |
f4ebcbc0 KT |
588 | struct rq *rq = rq_of_rt_rq(rt_rq); |
589 | ||
590 | if (!rt_rq->rt_nr_running) | |
591 | return; | |
592 | ||
593 | enqueue_top_rt_rq(rt_rq); | |
8875125e | 594 | resched_curr(rq); |
6f505b16 PZ |
595 | } |
596 | ||
9f0c1e56 | 597 | static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
6f505b16 | 598 | { |
f4ebcbc0 | 599 | dequeue_top_rt_rq(rt_rq); |
6f505b16 PZ |
600 | } |
601 | ||
46383648 KT |
602 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
603 | { | |
604 | return rt_rq->rt_throttled; | |
605 | } | |
606 | ||
c6c4927b | 607 | static inline const struct cpumask *sched_rt_period_mask(void) |
d0b27fa7 | 608 | { |
c6c4927b | 609 | return cpu_online_mask; |
d0b27fa7 PZ |
610 | } |
611 | ||
612 | static inline | |
613 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) | |
614 | { | |
615 | return &cpu_rq(cpu)->rt; | |
616 | } | |
617 | ||
ac086bc2 PZ |
618 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
619 | { | |
620 | return &def_rt_bandwidth; | |
621 | } | |
622 | ||
55e12e5e | 623 | #endif /* CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 | 624 | |
faa59937 JL |
625 | bool sched_rt_bandwidth_account(struct rt_rq *rt_rq) |
626 | { | |
627 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); | |
628 | ||
629 | return (hrtimer_active(&rt_b->rt_period_timer) || | |
630 | rt_rq->rt_time < rt_b->rt_runtime); | |
631 | } | |
632 | ||
ac086bc2 | 633 | #ifdef CONFIG_SMP |
78333cdd PZ |
634 | /* |
635 | * We ran out of runtime, see if we can borrow some from our neighbours. | |
636 | */ | |
269b26a5 | 637 | static void do_balance_runtime(struct rt_rq *rt_rq) |
ac086bc2 PZ |
638 | { |
639 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); | |
aa7f6730 | 640 | struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd; |
269b26a5 | 641 | int i, weight; |
ac086bc2 PZ |
642 | u64 rt_period; |
643 | ||
c6c4927b | 644 | weight = cpumask_weight(rd->span); |
ac086bc2 | 645 | |
0986b11b | 646 | raw_spin_lock(&rt_b->rt_runtime_lock); |
ac086bc2 | 647 | rt_period = ktime_to_ns(rt_b->rt_period); |
c6c4927b | 648 | for_each_cpu(i, rd->span) { |
ac086bc2 PZ |
649 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); |
650 | s64 diff; | |
651 | ||
652 | if (iter == rt_rq) | |
653 | continue; | |
654 | ||
0986b11b | 655 | raw_spin_lock(&iter->rt_runtime_lock); |
78333cdd PZ |
656 | /* |
657 | * Either all rqs have inf runtime and there's nothing to steal | |
658 | * or __disable_runtime() below sets a specific rq to inf to | |
659 | * indicate its been disabled and disalow stealing. | |
660 | */ | |
7def2be1 PZ |
661 | if (iter->rt_runtime == RUNTIME_INF) |
662 | goto next; | |
663 | ||
78333cdd PZ |
664 | /* |
665 | * From runqueues with spare time, take 1/n part of their | |
666 | * spare time, but no more than our period. | |
667 | */ | |
ac086bc2 PZ |
668 | diff = iter->rt_runtime - iter->rt_time; |
669 | if (diff > 0) { | |
58838cf3 | 670 | diff = div_u64((u64)diff, weight); |
ac086bc2 PZ |
671 | if (rt_rq->rt_runtime + diff > rt_period) |
672 | diff = rt_period - rt_rq->rt_runtime; | |
673 | iter->rt_runtime -= diff; | |
674 | rt_rq->rt_runtime += diff; | |
ac086bc2 | 675 | if (rt_rq->rt_runtime == rt_period) { |
0986b11b | 676 | raw_spin_unlock(&iter->rt_runtime_lock); |
ac086bc2 PZ |
677 | break; |
678 | } | |
679 | } | |
7def2be1 | 680 | next: |
0986b11b | 681 | raw_spin_unlock(&iter->rt_runtime_lock); |
ac086bc2 | 682 | } |
0986b11b | 683 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
ac086bc2 | 684 | } |
7def2be1 | 685 | |
78333cdd PZ |
686 | /* |
687 | * Ensure this RQ takes back all the runtime it lend to its neighbours. | |
688 | */ | |
7def2be1 PZ |
689 | static void __disable_runtime(struct rq *rq) |
690 | { | |
691 | struct root_domain *rd = rq->rd; | |
ec514c48 | 692 | rt_rq_iter_t iter; |
7def2be1 PZ |
693 | struct rt_rq *rt_rq; |
694 | ||
695 | if (unlikely(!scheduler_running)) | |
696 | return; | |
697 | ||
ec514c48 | 698 | for_each_rt_rq(rt_rq, iter, rq) { |
7def2be1 PZ |
699 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
700 | s64 want; | |
701 | int i; | |
702 | ||
0986b11b TG |
703 | raw_spin_lock(&rt_b->rt_runtime_lock); |
704 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
78333cdd PZ |
705 | /* |
706 | * Either we're all inf and nobody needs to borrow, or we're | |
707 | * already disabled and thus have nothing to do, or we have | |
708 | * exactly the right amount of runtime to take out. | |
709 | */ | |
7def2be1 PZ |
710 | if (rt_rq->rt_runtime == RUNTIME_INF || |
711 | rt_rq->rt_runtime == rt_b->rt_runtime) | |
712 | goto balanced; | |
0986b11b | 713 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
7def2be1 | 714 | |
78333cdd PZ |
715 | /* |
716 | * Calculate the difference between what we started out with | |
717 | * and what we current have, that's the amount of runtime | |
718 | * we lend and now have to reclaim. | |
719 | */ | |
7def2be1 PZ |
720 | want = rt_b->rt_runtime - rt_rq->rt_runtime; |
721 | ||
78333cdd PZ |
722 | /* |
723 | * Greedy reclaim, take back as much as we can. | |
724 | */ | |
c6c4927b | 725 | for_each_cpu(i, rd->span) { |
7def2be1 PZ |
726 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); |
727 | s64 diff; | |
728 | ||
78333cdd PZ |
729 | /* |
730 | * Can't reclaim from ourselves or disabled runqueues. | |
731 | */ | |
f1679d08 | 732 | if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF) |
7def2be1 PZ |
733 | continue; |
734 | ||
0986b11b | 735 | raw_spin_lock(&iter->rt_runtime_lock); |
7def2be1 PZ |
736 | if (want > 0) { |
737 | diff = min_t(s64, iter->rt_runtime, want); | |
738 | iter->rt_runtime -= diff; | |
739 | want -= diff; | |
740 | } else { | |
741 | iter->rt_runtime -= want; | |
742 | want -= want; | |
743 | } | |
0986b11b | 744 | raw_spin_unlock(&iter->rt_runtime_lock); |
7def2be1 PZ |
745 | |
746 | if (!want) | |
747 | break; | |
748 | } | |
749 | ||
0986b11b | 750 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
78333cdd PZ |
751 | /* |
752 | * We cannot be left wanting - that would mean some runtime | |
753 | * leaked out of the system. | |
754 | */ | |
7def2be1 PZ |
755 | BUG_ON(want); |
756 | balanced: | |
78333cdd PZ |
757 | /* |
758 | * Disable all the borrow logic by pretending we have inf | |
759 | * runtime - in which case borrowing doesn't make sense. | |
760 | */ | |
7def2be1 | 761 | rt_rq->rt_runtime = RUNTIME_INF; |
a4c96ae3 | 762 | rt_rq->rt_throttled = 0; |
0986b11b TG |
763 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
764 | raw_spin_unlock(&rt_b->rt_runtime_lock); | |
99b62567 KT |
765 | |
766 | /* Make rt_rq available for pick_next_task() */ | |
767 | sched_rt_rq_enqueue(rt_rq); | |
7def2be1 PZ |
768 | } |
769 | } | |
770 | ||
7def2be1 PZ |
771 | static void __enable_runtime(struct rq *rq) |
772 | { | |
ec514c48 | 773 | rt_rq_iter_t iter; |
7def2be1 PZ |
774 | struct rt_rq *rt_rq; |
775 | ||
776 | if (unlikely(!scheduler_running)) | |
777 | return; | |
778 | ||
78333cdd PZ |
779 | /* |
780 | * Reset each runqueue's bandwidth settings | |
781 | */ | |
ec514c48 | 782 | for_each_rt_rq(rt_rq, iter, rq) { |
7def2be1 PZ |
783 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
784 | ||
0986b11b TG |
785 | raw_spin_lock(&rt_b->rt_runtime_lock); |
786 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
7def2be1 PZ |
787 | rt_rq->rt_runtime = rt_b->rt_runtime; |
788 | rt_rq->rt_time = 0; | |
baf25731 | 789 | rt_rq->rt_throttled = 0; |
0986b11b TG |
790 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
791 | raw_spin_unlock(&rt_b->rt_runtime_lock); | |
7def2be1 PZ |
792 | } |
793 | } | |
794 | ||
269b26a5 | 795 | static void balance_runtime(struct rt_rq *rt_rq) |
eff6549b | 796 | { |
4a6184ce | 797 | if (!sched_feat(RT_RUNTIME_SHARE)) |
269b26a5 | 798 | return; |
4a6184ce | 799 | |
eff6549b | 800 | if (rt_rq->rt_time > rt_rq->rt_runtime) { |
0986b11b | 801 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
269b26a5 | 802 | do_balance_runtime(rt_rq); |
0986b11b | 803 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
eff6549b | 804 | } |
eff6549b | 805 | } |
55e12e5e | 806 | #else /* !CONFIG_SMP */ |
269b26a5 | 807 | static inline void balance_runtime(struct rt_rq *rt_rq) {} |
55e12e5e | 808 | #endif /* CONFIG_SMP */ |
ac086bc2 | 809 | |
eff6549b PZ |
810 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun) |
811 | { | |
42c62a58 | 812 | int i, idle = 1, throttled = 0; |
c6c4927b | 813 | const struct cpumask *span; |
eff6549b | 814 | |
eff6549b | 815 | span = sched_rt_period_mask(); |
e221d028 MG |
816 | #ifdef CONFIG_RT_GROUP_SCHED |
817 | /* | |
818 | * FIXME: isolated CPUs should really leave the root task group, | |
819 | * whether they are isolcpus or were isolated via cpusets, lest | |
820 | * the timer run on a CPU which does not service all runqueues, | |
821 | * potentially leaving other CPUs indefinitely throttled. If | |
822 | * isolation is really required, the user will turn the throttle | |
823 | * off to kill the perturbations it causes anyway. Meanwhile, | |
824 | * this maintains functionality for boot and/or troubleshooting. | |
825 | */ | |
826 | if (rt_b == &root_task_group.rt_bandwidth) | |
827 | span = cpu_online_mask; | |
828 | #endif | |
c6c4927b | 829 | for_each_cpu(i, span) { |
eff6549b PZ |
830 | int enqueue = 0; |
831 | struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i); | |
832 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
c249f255 DK |
833 | int skip; |
834 | ||
835 | /* | |
836 | * When span == cpu_online_mask, taking each rq->lock | |
837 | * can be time-consuming. Try to avoid it when possible. | |
838 | */ | |
839 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
840 | skip = !rt_rq->rt_time && !rt_rq->rt_nr_running; | |
841 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
842 | if (skip) | |
843 | continue; | |
eff6549b | 844 | |
05fa785c | 845 | raw_spin_lock(&rq->lock); |
3f774942 DB |
846 | update_rq_clock(rq); |
847 | ||
eff6549b PZ |
848 | if (rt_rq->rt_time) { |
849 | u64 runtime; | |
850 | ||
0986b11b | 851 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
eff6549b PZ |
852 | if (rt_rq->rt_throttled) |
853 | balance_runtime(rt_rq); | |
854 | runtime = rt_rq->rt_runtime; | |
855 | rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime); | |
856 | if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) { | |
857 | rt_rq->rt_throttled = 0; | |
858 | enqueue = 1; | |
61eadef6 MG |
859 | |
860 | /* | |
9edfbfed PZ |
861 | * When we're idle and a woken (rt) task is |
862 | * throttled check_preempt_curr() will set | |
863 | * skip_update and the time between the wakeup | |
864 | * and this unthrottle will get accounted as | |
865 | * 'runtime'. | |
61eadef6 MG |
866 | */ |
867 | if (rt_rq->rt_nr_running && rq->curr == rq->idle) | |
9edfbfed | 868 | rq_clock_skip_update(rq, false); |
eff6549b PZ |
869 | } |
870 | if (rt_rq->rt_time || rt_rq->rt_nr_running) | |
871 | idle = 0; | |
0986b11b | 872 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
0c3b9168 | 873 | } else if (rt_rq->rt_nr_running) { |
6c3df255 | 874 | idle = 0; |
0c3b9168 BS |
875 | if (!rt_rq_throttled(rt_rq)) |
876 | enqueue = 1; | |
877 | } | |
42c62a58 PZ |
878 | if (rt_rq->rt_throttled) |
879 | throttled = 1; | |
eff6549b PZ |
880 | |
881 | if (enqueue) | |
882 | sched_rt_rq_enqueue(rt_rq); | |
05fa785c | 883 | raw_spin_unlock(&rq->lock); |
eff6549b PZ |
884 | } |
885 | ||
42c62a58 PZ |
886 | if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)) |
887 | return 1; | |
888 | ||
eff6549b PZ |
889 | return idle; |
890 | } | |
ac086bc2 | 891 | |
6f505b16 PZ |
892 | static inline int rt_se_prio(struct sched_rt_entity *rt_se) |
893 | { | |
052f1dc7 | 894 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 PZ |
895 | struct rt_rq *rt_rq = group_rt_rq(rt_se); |
896 | ||
897 | if (rt_rq) | |
e864c499 | 898 | return rt_rq->highest_prio.curr; |
6f505b16 PZ |
899 | #endif |
900 | ||
901 | return rt_task_of(rt_se)->prio; | |
902 | } | |
903 | ||
9f0c1e56 | 904 | static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq) |
6f505b16 | 905 | { |
9f0c1e56 | 906 | u64 runtime = sched_rt_runtime(rt_rq); |
fa85ae24 | 907 | |
fa85ae24 | 908 | if (rt_rq->rt_throttled) |
23b0fdfc | 909 | return rt_rq_throttled(rt_rq); |
fa85ae24 | 910 | |
5b680fd6 | 911 | if (runtime >= sched_rt_period(rt_rq)) |
ac086bc2 PZ |
912 | return 0; |
913 | ||
b79f3833 PZ |
914 | balance_runtime(rt_rq); |
915 | runtime = sched_rt_runtime(rt_rq); | |
916 | if (runtime == RUNTIME_INF) | |
917 | return 0; | |
ac086bc2 | 918 | |
9f0c1e56 | 919 | if (rt_rq->rt_time > runtime) { |
7abc63b1 PZ |
920 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
921 | ||
922 | /* | |
923 | * Don't actually throttle groups that have no runtime assigned | |
924 | * but accrue some time due to boosting. | |
925 | */ | |
926 | if (likely(rt_b->rt_runtime)) { | |
927 | rt_rq->rt_throttled = 1; | |
c224815d | 928 | printk_deferred_once("sched: RT throttling activated\n"); |
7abc63b1 PZ |
929 | } else { |
930 | /* | |
931 | * In case we did anyway, make it go away, | |
932 | * replenishment is a joke, since it will replenish us | |
933 | * with exactly 0 ns. | |
934 | */ | |
935 | rt_rq->rt_time = 0; | |
936 | } | |
937 | ||
23b0fdfc | 938 | if (rt_rq_throttled(rt_rq)) { |
9f0c1e56 | 939 | sched_rt_rq_dequeue(rt_rq); |
23b0fdfc PZ |
940 | return 1; |
941 | } | |
fa85ae24 PZ |
942 | } |
943 | ||
944 | return 0; | |
945 | } | |
946 | ||
bb44e5d1 IM |
947 | /* |
948 | * Update the current task's runtime statistics. Skip current tasks that | |
949 | * are not in our scheduling class. | |
950 | */ | |
a9957449 | 951 | static void update_curr_rt(struct rq *rq) |
bb44e5d1 IM |
952 | { |
953 | struct task_struct *curr = rq->curr; | |
6f505b16 | 954 | struct sched_rt_entity *rt_se = &curr->rt; |
bb44e5d1 IM |
955 | u64 delta_exec; |
956 | ||
06c3bc65 | 957 | if (curr->sched_class != &rt_sched_class) |
bb44e5d1 IM |
958 | return; |
959 | ||
78becc27 | 960 | delta_exec = rq_clock_task(rq) - curr->se.exec_start; |
fc79e240 KT |
961 | if (unlikely((s64)delta_exec <= 0)) |
962 | return; | |
6cfb0d5d | 963 | |
58919e83 | 964 | /* Kick cpufreq (see the comment in kernel/sched/sched.h). */ |
674e7541 | 965 | cpufreq_update_util(rq, SCHED_CPUFREQ_RT); |
594dd290 | 966 | |
42c62a58 PZ |
967 | schedstat_set(curr->se.statistics.exec_max, |
968 | max(curr->se.statistics.exec_max, delta_exec)); | |
bb44e5d1 IM |
969 | |
970 | curr->se.sum_exec_runtime += delta_exec; | |
f06febc9 FM |
971 | account_group_exec_runtime(curr, delta_exec); |
972 | ||
78becc27 | 973 | curr->se.exec_start = rq_clock_task(rq); |
d2cc5ed6 | 974 | cgroup_account_cputime(curr, delta_exec); |
fa85ae24 | 975 | |
e9e9250b PZ |
976 | sched_rt_avg_update(rq, delta_exec); |
977 | ||
0b148fa0 PZ |
978 | if (!rt_bandwidth_enabled()) |
979 | return; | |
980 | ||
354d60c2 | 981 | for_each_sched_rt_entity(rt_se) { |
0b07939c | 982 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); |
354d60c2 | 983 | |
cc2991cf | 984 | if (sched_rt_runtime(rt_rq) != RUNTIME_INF) { |
0986b11b | 985 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
cc2991cf PZ |
986 | rt_rq->rt_time += delta_exec; |
987 | if (sched_rt_runtime_exceeded(rt_rq)) | |
8875125e | 988 | resched_curr(rq); |
0986b11b | 989 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
cc2991cf | 990 | } |
354d60c2 | 991 | } |
bb44e5d1 IM |
992 | } |
993 | ||
f4ebcbc0 KT |
994 | static void |
995 | dequeue_top_rt_rq(struct rt_rq *rt_rq) | |
996 | { | |
997 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
998 | ||
999 | BUG_ON(&rq->rt != rt_rq); | |
1000 | ||
1001 | if (!rt_rq->rt_queued) | |
1002 | return; | |
1003 | ||
1004 | BUG_ON(!rq->nr_running); | |
1005 | ||
72465447 | 1006 | sub_nr_running(rq, rt_rq->rt_nr_running); |
f4ebcbc0 KT |
1007 | rt_rq->rt_queued = 0; |
1008 | } | |
1009 | ||
1010 | static void | |
1011 | enqueue_top_rt_rq(struct rt_rq *rt_rq) | |
1012 | { | |
1013 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
1014 | ||
1015 | BUG_ON(&rq->rt != rt_rq); | |
1016 | ||
1017 | if (rt_rq->rt_queued) | |
1018 | return; | |
1019 | if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running) | |
1020 | return; | |
1021 | ||
72465447 | 1022 | add_nr_running(rq, rt_rq->rt_nr_running); |
f4ebcbc0 KT |
1023 | rt_rq->rt_queued = 1; |
1024 | } | |
1025 | ||
398a153b | 1026 | #if defined CONFIG_SMP |
e864c499 | 1027 | |
398a153b GH |
1028 | static void |
1029 | inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) | |
63489e45 | 1030 | { |
4d984277 | 1031 | struct rq *rq = rq_of_rt_rq(rt_rq); |
1f11eb6a | 1032 | |
757dfcaa KT |
1033 | #ifdef CONFIG_RT_GROUP_SCHED |
1034 | /* | |
1035 | * Change rq's cpupri only if rt_rq is the top queue. | |
1036 | */ | |
1037 | if (&rq->rt != rt_rq) | |
1038 | return; | |
1039 | #endif | |
5181f4a4 SR |
1040 | if (rq->online && prio < prev_prio) |
1041 | cpupri_set(&rq->rd->cpupri, rq->cpu, prio); | |
398a153b | 1042 | } |
73fe6aae | 1043 | |
398a153b GH |
1044 | static void |
1045 | dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) | |
1046 | { | |
1047 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
d0b27fa7 | 1048 | |
757dfcaa KT |
1049 | #ifdef CONFIG_RT_GROUP_SCHED |
1050 | /* | |
1051 | * Change rq's cpupri only if rt_rq is the top queue. | |
1052 | */ | |
1053 | if (&rq->rt != rt_rq) | |
1054 | return; | |
1055 | #endif | |
398a153b GH |
1056 | if (rq->online && rt_rq->highest_prio.curr != prev_prio) |
1057 | cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr); | |
63489e45 SR |
1058 | } |
1059 | ||
398a153b GH |
1060 | #else /* CONFIG_SMP */ |
1061 | ||
6f505b16 | 1062 | static inline |
398a153b GH |
1063 | void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} |
1064 | static inline | |
1065 | void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} | |
1066 | ||
1067 | #endif /* CONFIG_SMP */ | |
6e0534f2 | 1068 | |
052f1dc7 | 1069 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
398a153b GH |
1070 | static void |
1071 | inc_rt_prio(struct rt_rq *rt_rq, int prio) | |
1072 | { | |
1073 | int prev_prio = rt_rq->highest_prio.curr; | |
1074 | ||
1075 | if (prio < prev_prio) | |
1076 | rt_rq->highest_prio.curr = prio; | |
1077 | ||
1078 | inc_rt_prio_smp(rt_rq, prio, prev_prio); | |
1079 | } | |
1080 | ||
1081 | static void | |
1082 | dec_rt_prio(struct rt_rq *rt_rq, int prio) | |
1083 | { | |
1084 | int prev_prio = rt_rq->highest_prio.curr; | |
1085 | ||
6f505b16 | 1086 | if (rt_rq->rt_nr_running) { |
764a9d6f | 1087 | |
398a153b | 1088 | WARN_ON(prio < prev_prio); |
764a9d6f | 1089 | |
e864c499 | 1090 | /* |
398a153b GH |
1091 | * This may have been our highest task, and therefore |
1092 | * we may have some recomputation to do | |
e864c499 | 1093 | */ |
398a153b | 1094 | if (prio == prev_prio) { |
e864c499 GH |
1095 | struct rt_prio_array *array = &rt_rq->active; |
1096 | ||
1097 | rt_rq->highest_prio.curr = | |
764a9d6f | 1098 | sched_find_first_bit(array->bitmap); |
e864c499 GH |
1099 | } |
1100 | ||
764a9d6f | 1101 | } else |
e864c499 | 1102 | rt_rq->highest_prio.curr = MAX_RT_PRIO; |
73fe6aae | 1103 | |
398a153b GH |
1104 | dec_rt_prio_smp(rt_rq, prio, prev_prio); |
1105 | } | |
1f11eb6a | 1106 | |
398a153b GH |
1107 | #else |
1108 | ||
1109 | static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {} | |
1110 | static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {} | |
1111 | ||
1112 | #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */ | |
6e0534f2 | 1113 | |
052f1dc7 | 1114 | #ifdef CONFIG_RT_GROUP_SCHED |
398a153b GH |
1115 | |
1116 | static void | |
1117 | inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1118 | { | |
1119 | if (rt_se_boosted(rt_se)) | |
1120 | rt_rq->rt_nr_boosted++; | |
1121 | ||
1122 | if (rt_rq->tg) | |
1123 | start_rt_bandwidth(&rt_rq->tg->rt_bandwidth); | |
1124 | } | |
1125 | ||
1126 | static void | |
1127 | dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1128 | { | |
23b0fdfc PZ |
1129 | if (rt_se_boosted(rt_se)) |
1130 | rt_rq->rt_nr_boosted--; | |
1131 | ||
1132 | WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted); | |
398a153b GH |
1133 | } |
1134 | ||
1135 | #else /* CONFIG_RT_GROUP_SCHED */ | |
1136 | ||
1137 | static void | |
1138 | inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1139 | { | |
1140 | start_rt_bandwidth(&def_rt_bandwidth); | |
1141 | } | |
1142 | ||
1143 | static inline | |
1144 | void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {} | |
1145 | ||
1146 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
1147 | ||
22abdef3 KT |
1148 | static inline |
1149 | unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se) | |
1150 | { | |
1151 | struct rt_rq *group_rq = group_rt_rq(rt_se); | |
1152 | ||
1153 | if (group_rq) | |
1154 | return group_rq->rt_nr_running; | |
1155 | else | |
1156 | return 1; | |
1157 | } | |
1158 | ||
01d36d0a FW |
1159 | static inline |
1160 | unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se) | |
1161 | { | |
1162 | struct rt_rq *group_rq = group_rt_rq(rt_se); | |
1163 | struct task_struct *tsk; | |
1164 | ||
1165 | if (group_rq) | |
1166 | return group_rq->rr_nr_running; | |
1167 | ||
1168 | tsk = rt_task_of(rt_se); | |
1169 | ||
1170 | return (tsk->policy == SCHED_RR) ? 1 : 0; | |
1171 | } | |
1172 | ||
398a153b GH |
1173 | static inline |
1174 | void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1175 | { | |
1176 | int prio = rt_se_prio(rt_se); | |
1177 | ||
1178 | WARN_ON(!rt_prio(prio)); | |
22abdef3 | 1179 | rt_rq->rt_nr_running += rt_se_nr_running(rt_se); |
01d36d0a | 1180 | rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se); |
398a153b GH |
1181 | |
1182 | inc_rt_prio(rt_rq, prio); | |
1183 | inc_rt_migration(rt_se, rt_rq); | |
1184 | inc_rt_group(rt_se, rt_rq); | |
1185 | } | |
1186 | ||
1187 | static inline | |
1188 | void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1189 | { | |
1190 | WARN_ON(!rt_prio(rt_se_prio(rt_se))); | |
1191 | WARN_ON(!rt_rq->rt_nr_running); | |
22abdef3 | 1192 | rt_rq->rt_nr_running -= rt_se_nr_running(rt_se); |
01d36d0a | 1193 | rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se); |
398a153b GH |
1194 | |
1195 | dec_rt_prio(rt_rq, rt_se_prio(rt_se)); | |
1196 | dec_rt_migration(rt_se, rt_rq); | |
1197 | dec_rt_group(rt_se, rt_rq); | |
63489e45 SR |
1198 | } |
1199 | ||
ff77e468 PZ |
1200 | /* |
1201 | * Change rt_se->run_list location unless SAVE && !MOVE | |
1202 | * | |
1203 | * assumes ENQUEUE/DEQUEUE flags match | |
1204 | */ | |
1205 | static inline bool move_entity(unsigned int flags) | |
1206 | { | |
1207 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE) | |
1208 | return false; | |
1209 | ||
1210 | return true; | |
1211 | } | |
1212 | ||
1213 | static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array) | |
1214 | { | |
1215 | list_del_init(&rt_se->run_list); | |
1216 | ||
1217 | if (list_empty(array->queue + rt_se_prio(rt_se))) | |
1218 | __clear_bit(rt_se_prio(rt_se), array->bitmap); | |
1219 | ||
1220 | rt_se->on_list = 0; | |
1221 | } | |
1222 | ||
1223 | static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) | |
bb44e5d1 | 1224 | { |
6f505b16 PZ |
1225 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); |
1226 | struct rt_prio_array *array = &rt_rq->active; | |
1227 | struct rt_rq *group_rq = group_rt_rq(rt_se); | |
20b6331b | 1228 | struct list_head *queue = array->queue + rt_se_prio(rt_se); |
bb44e5d1 | 1229 | |
ad2a3f13 PZ |
1230 | /* |
1231 | * Don't enqueue the group if its throttled, or when empty. | |
1232 | * The latter is a consequence of the former when a child group | |
1233 | * get throttled and the current group doesn't have any other | |
1234 | * active members. | |
1235 | */ | |
ff77e468 PZ |
1236 | if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) { |
1237 | if (rt_se->on_list) | |
1238 | __delist_rt_entity(rt_se, array); | |
6f505b16 | 1239 | return; |
ff77e468 | 1240 | } |
63489e45 | 1241 | |
ff77e468 PZ |
1242 | if (move_entity(flags)) { |
1243 | WARN_ON_ONCE(rt_se->on_list); | |
1244 | if (flags & ENQUEUE_HEAD) | |
1245 | list_add(&rt_se->run_list, queue); | |
1246 | else | |
1247 | list_add_tail(&rt_se->run_list, queue); | |
1248 | ||
1249 | __set_bit(rt_se_prio(rt_se), array->bitmap); | |
1250 | rt_se->on_list = 1; | |
1251 | } | |
1252 | rt_se->on_rq = 1; | |
78f2c7db | 1253 | |
6f505b16 PZ |
1254 | inc_rt_tasks(rt_se, rt_rq); |
1255 | } | |
1256 | ||
ff77e468 | 1257 | static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) |
6f505b16 PZ |
1258 | { |
1259 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); | |
1260 | struct rt_prio_array *array = &rt_rq->active; | |
1261 | ||
ff77e468 PZ |
1262 | if (move_entity(flags)) { |
1263 | WARN_ON_ONCE(!rt_se->on_list); | |
1264 | __delist_rt_entity(rt_se, array); | |
1265 | } | |
1266 | rt_se->on_rq = 0; | |
6f505b16 PZ |
1267 | |
1268 | dec_rt_tasks(rt_se, rt_rq); | |
1269 | } | |
1270 | ||
1271 | /* | |
1272 | * Because the prio of an upper entry depends on the lower | |
1273 | * entries, we must remove entries top - down. | |
6f505b16 | 1274 | */ |
ff77e468 | 1275 | static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags) |
6f505b16 | 1276 | { |
ad2a3f13 | 1277 | struct sched_rt_entity *back = NULL; |
6f505b16 | 1278 | |
58d6c2d7 PZ |
1279 | for_each_sched_rt_entity(rt_se) { |
1280 | rt_se->back = back; | |
1281 | back = rt_se; | |
1282 | } | |
1283 | ||
f4ebcbc0 KT |
1284 | dequeue_top_rt_rq(rt_rq_of_se(back)); |
1285 | ||
58d6c2d7 PZ |
1286 | for (rt_se = back; rt_se; rt_se = rt_se->back) { |
1287 | if (on_rt_rq(rt_se)) | |
ff77e468 | 1288 | __dequeue_rt_entity(rt_se, flags); |
ad2a3f13 PZ |
1289 | } |
1290 | } | |
1291 | ||
ff77e468 | 1292 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) |
ad2a3f13 | 1293 | { |
f4ebcbc0 KT |
1294 | struct rq *rq = rq_of_rt_se(rt_se); |
1295 | ||
ff77e468 | 1296 | dequeue_rt_stack(rt_se, flags); |
ad2a3f13 | 1297 | for_each_sched_rt_entity(rt_se) |
ff77e468 | 1298 | __enqueue_rt_entity(rt_se, flags); |
f4ebcbc0 | 1299 | enqueue_top_rt_rq(&rq->rt); |
ad2a3f13 PZ |
1300 | } |
1301 | ||
ff77e468 | 1302 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) |
ad2a3f13 | 1303 | { |
f4ebcbc0 KT |
1304 | struct rq *rq = rq_of_rt_se(rt_se); |
1305 | ||
ff77e468 | 1306 | dequeue_rt_stack(rt_se, flags); |
ad2a3f13 PZ |
1307 | |
1308 | for_each_sched_rt_entity(rt_se) { | |
1309 | struct rt_rq *rt_rq = group_rt_rq(rt_se); | |
1310 | ||
1311 | if (rt_rq && rt_rq->rt_nr_running) | |
ff77e468 | 1312 | __enqueue_rt_entity(rt_se, flags); |
58d6c2d7 | 1313 | } |
f4ebcbc0 | 1314 | enqueue_top_rt_rq(&rq->rt); |
bb44e5d1 IM |
1315 | } |
1316 | ||
1317 | /* | |
1318 | * Adding/removing a task to/from a priority array: | |
1319 | */ | |
ea87bb78 | 1320 | static void |
371fd7e7 | 1321 | enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags) |
6f505b16 PZ |
1322 | { |
1323 | struct sched_rt_entity *rt_se = &p->rt; | |
1324 | ||
371fd7e7 | 1325 | if (flags & ENQUEUE_WAKEUP) |
6f505b16 PZ |
1326 | rt_se->timeout = 0; |
1327 | ||
ff77e468 | 1328 | enqueue_rt_entity(rt_se, flags); |
c09595f6 | 1329 | |
4b53a341 | 1330 | if (!task_current(rq, p) && p->nr_cpus_allowed > 1) |
917b627d | 1331 | enqueue_pushable_task(rq, p); |
6f505b16 PZ |
1332 | } |
1333 | ||
371fd7e7 | 1334 | static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags) |
bb44e5d1 | 1335 | { |
6f505b16 | 1336 | struct sched_rt_entity *rt_se = &p->rt; |
bb44e5d1 | 1337 | |
f1e14ef6 | 1338 | update_curr_rt(rq); |
ff77e468 | 1339 | dequeue_rt_entity(rt_se, flags); |
c09595f6 | 1340 | |
917b627d | 1341 | dequeue_pushable_task(rq, p); |
bb44e5d1 IM |
1342 | } |
1343 | ||
1344 | /* | |
60686317 RW |
1345 | * Put task to the head or the end of the run list without the overhead of |
1346 | * dequeue followed by enqueue. | |
bb44e5d1 | 1347 | */ |
7ebefa8c DA |
1348 | static void |
1349 | requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head) | |
6f505b16 | 1350 | { |
1cdad715 | 1351 | if (on_rt_rq(rt_se)) { |
7ebefa8c DA |
1352 | struct rt_prio_array *array = &rt_rq->active; |
1353 | struct list_head *queue = array->queue + rt_se_prio(rt_se); | |
1354 | ||
1355 | if (head) | |
1356 | list_move(&rt_se->run_list, queue); | |
1357 | else | |
1358 | list_move_tail(&rt_se->run_list, queue); | |
1cdad715 | 1359 | } |
6f505b16 PZ |
1360 | } |
1361 | ||
7ebefa8c | 1362 | static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head) |
bb44e5d1 | 1363 | { |
6f505b16 PZ |
1364 | struct sched_rt_entity *rt_se = &p->rt; |
1365 | struct rt_rq *rt_rq; | |
bb44e5d1 | 1366 | |
6f505b16 PZ |
1367 | for_each_sched_rt_entity(rt_se) { |
1368 | rt_rq = rt_rq_of_se(rt_se); | |
7ebefa8c | 1369 | requeue_rt_entity(rt_rq, rt_se, head); |
6f505b16 | 1370 | } |
bb44e5d1 IM |
1371 | } |
1372 | ||
6f505b16 | 1373 | static void yield_task_rt(struct rq *rq) |
bb44e5d1 | 1374 | { |
7ebefa8c | 1375 | requeue_task_rt(rq, rq->curr, 0); |
bb44e5d1 IM |
1376 | } |
1377 | ||
e7693a36 | 1378 | #ifdef CONFIG_SMP |
318e0893 GH |
1379 | static int find_lowest_rq(struct task_struct *task); |
1380 | ||
0017d735 | 1381 | static int |
ac66f547 | 1382 | select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags) |
e7693a36 | 1383 | { |
7608dec2 PZ |
1384 | struct task_struct *curr; |
1385 | struct rq *rq; | |
c37495fd SR |
1386 | |
1387 | /* For anything but wake ups, just return the task_cpu */ | |
1388 | if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK) | |
1389 | goto out; | |
1390 | ||
7608dec2 PZ |
1391 | rq = cpu_rq(cpu); |
1392 | ||
1393 | rcu_read_lock(); | |
316c1608 | 1394 | curr = READ_ONCE(rq->curr); /* unlocked access */ |
7608dec2 | 1395 | |
318e0893 | 1396 | /* |
7608dec2 | 1397 | * If the current task on @p's runqueue is an RT task, then |
e1f47d89 SR |
1398 | * try to see if we can wake this RT task up on another |
1399 | * runqueue. Otherwise simply start this RT task | |
1400 | * on its current runqueue. | |
1401 | * | |
43fa5460 SR |
1402 | * We want to avoid overloading runqueues. If the woken |
1403 | * task is a higher priority, then it will stay on this CPU | |
1404 | * and the lower prio task should be moved to another CPU. | |
1405 | * Even though this will probably make the lower prio task | |
1406 | * lose its cache, we do not want to bounce a higher task | |
1407 | * around just because it gave up its CPU, perhaps for a | |
1408 | * lock? | |
1409 | * | |
1410 | * For equal prio tasks, we just let the scheduler sort it out. | |
7608dec2 PZ |
1411 | * |
1412 | * Otherwise, just let it ride on the affined RQ and the | |
1413 | * post-schedule router will push the preempted task away | |
1414 | * | |
1415 | * This test is optimistic, if we get it wrong the load-balancer | |
1416 | * will have to sort it out. | |
318e0893 | 1417 | */ |
7608dec2 | 1418 | if (curr && unlikely(rt_task(curr)) && |
4b53a341 | 1419 | (curr->nr_cpus_allowed < 2 || |
6bfa687c | 1420 | curr->prio <= p->prio)) { |
7608dec2 | 1421 | int target = find_lowest_rq(p); |
318e0893 | 1422 | |
80e3d87b TC |
1423 | /* |
1424 | * Don't bother moving it if the destination CPU is | |
1425 | * not running a lower priority task. | |
1426 | */ | |
1427 | if (target != -1 && | |
1428 | p->prio < cpu_rq(target)->rt.highest_prio.curr) | |
7608dec2 | 1429 | cpu = target; |
318e0893 | 1430 | } |
7608dec2 | 1431 | rcu_read_unlock(); |
318e0893 | 1432 | |
c37495fd | 1433 | out: |
7608dec2 | 1434 | return cpu; |
e7693a36 | 1435 | } |
7ebefa8c DA |
1436 | |
1437 | static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p) | |
1438 | { | |
308a623a WL |
1439 | /* |
1440 | * Current can't be migrated, useless to reschedule, | |
1441 | * let's hope p can move out. | |
1442 | */ | |
4b53a341 | 1443 | if (rq->curr->nr_cpus_allowed == 1 || |
308a623a | 1444 | !cpupri_find(&rq->rd->cpupri, rq->curr, NULL)) |
7ebefa8c DA |
1445 | return; |
1446 | ||
308a623a WL |
1447 | /* |
1448 | * p is migratable, so let's not schedule it and | |
1449 | * see if it is pushed or pulled somewhere else. | |
1450 | */ | |
4b53a341 | 1451 | if (p->nr_cpus_allowed != 1 |
13b8bd0a RR |
1452 | && cpupri_find(&rq->rd->cpupri, p, NULL)) |
1453 | return; | |
24600ce8 | 1454 | |
7ebefa8c DA |
1455 | /* |
1456 | * There appears to be other cpus that can accept | |
1457 | * current and none to run 'p', so lets reschedule | |
1458 | * to try and push current away: | |
1459 | */ | |
1460 | requeue_task_rt(rq, p, 1); | |
8875125e | 1461 | resched_curr(rq); |
7ebefa8c DA |
1462 | } |
1463 | ||
e7693a36 GH |
1464 | #endif /* CONFIG_SMP */ |
1465 | ||
bb44e5d1 IM |
1466 | /* |
1467 | * Preempt the current task with a newly woken task if needed: | |
1468 | */ | |
7d478721 | 1469 | static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags) |
bb44e5d1 | 1470 | { |
45c01e82 | 1471 | if (p->prio < rq->curr->prio) { |
8875125e | 1472 | resched_curr(rq); |
45c01e82 GH |
1473 | return; |
1474 | } | |
1475 | ||
1476 | #ifdef CONFIG_SMP | |
1477 | /* | |
1478 | * If: | |
1479 | * | |
1480 | * - the newly woken task is of equal priority to the current task | |
1481 | * - the newly woken task is non-migratable while current is migratable | |
1482 | * - current will be preempted on the next reschedule | |
1483 | * | |
1484 | * we should check to see if current can readily move to a different | |
1485 | * cpu. If so, we will reschedule to allow the push logic to try | |
1486 | * to move current somewhere else, making room for our non-migratable | |
1487 | * task. | |
1488 | */ | |
8dd0de8b | 1489 | if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr)) |
7ebefa8c | 1490 | check_preempt_equal_prio(rq, p); |
45c01e82 | 1491 | #endif |
bb44e5d1 IM |
1492 | } |
1493 | ||
6f505b16 PZ |
1494 | static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq, |
1495 | struct rt_rq *rt_rq) | |
bb44e5d1 | 1496 | { |
6f505b16 PZ |
1497 | struct rt_prio_array *array = &rt_rq->active; |
1498 | struct sched_rt_entity *next = NULL; | |
bb44e5d1 IM |
1499 | struct list_head *queue; |
1500 | int idx; | |
1501 | ||
1502 | idx = sched_find_first_bit(array->bitmap); | |
6f505b16 | 1503 | BUG_ON(idx >= MAX_RT_PRIO); |
bb44e5d1 IM |
1504 | |
1505 | queue = array->queue + idx; | |
6f505b16 | 1506 | next = list_entry(queue->next, struct sched_rt_entity, run_list); |
326587b8 | 1507 | |
6f505b16 PZ |
1508 | return next; |
1509 | } | |
bb44e5d1 | 1510 | |
917b627d | 1511 | static struct task_struct *_pick_next_task_rt(struct rq *rq) |
6f505b16 PZ |
1512 | { |
1513 | struct sched_rt_entity *rt_se; | |
1514 | struct task_struct *p; | |
606dba2e | 1515 | struct rt_rq *rt_rq = &rq->rt; |
6f505b16 PZ |
1516 | |
1517 | do { | |
1518 | rt_se = pick_next_rt_entity(rq, rt_rq); | |
326587b8 | 1519 | BUG_ON(!rt_se); |
6f505b16 PZ |
1520 | rt_rq = group_rt_rq(rt_se); |
1521 | } while (rt_rq); | |
1522 | ||
1523 | p = rt_task_of(rt_se); | |
78becc27 | 1524 | p->se.exec_start = rq_clock_task(rq); |
917b627d GH |
1525 | |
1526 | return p; | |
1527 | } | |
1528 | ||
606dba2e | 1529 | static struct task_struct * |
d8ac8971 | 1530 | pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
917b627d | 1531 | { |
606dba2e PZ |
1532 | struct task_struct *p; |
1533 | struct rt_rq *rt_rq = &rq->rt; | |
1534 | ||
37e117c0 | 1535 | if (need_pull_rt_task(rq, prev)) { |
cbce1a68 PZ |
1536 | /* |
1537 | * This is OK, because current is on_cpu, which avoids it being | |
1538 | * picked for load-balance and preemption/IRQs are still | |
1539 | * disabled avoiding further scheduler activity on it and we're | |
1540 | * being very careful to re-start the picking loop. | |
1541 | */ | |
d8ac8971 | 1542 | rq_unpin_lock(rq, rf); |
38033c37 | 1543 | pull_rt_task(rq); |
d8ac8971 | 1544 | rq_repin_lock(rq, rf); |
37e117c0 PZ |
1545 | /* |
1546 | * pull_rt_task() can drop (and re-acquire) rq->lock; this | |
a1d9a323 KT |
1547 | * means a dl or stop task can slip in, in which case we need |
1548 | * to re-start task selection. | |
37e117c0 | 1549 | */ |
da0c1e65 | 1550 | if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) || |
a1d9a323 | 1551 | rq->dl.dl_nr_running)) |
37e117c0 PZ |
1552 | return RETRY_TASK; |
1553 | } | |
38033c37 | 1554 | |
734ff2a7 KT |
1555 | /* |
1556 | * We may dequeue prev's rt_rq in put_prev_task(). | |
1557 | * So, we update time before rt_nr_running check. | |
1558 | */ | |
1559 | if (prev->sched_class == &rt_sched_class) | |
1560 | update_curr_rt(rq); | |
1561 | ||
f4ebcbc0 | 1562 | if (!rt_rq->rt_queued) |
606dba2e PZ |
1563 | return NULL; |
1564 | ||
3f1d2a31 | 1565 | put_prev_task(rq, prev); |
606dba2e PZ |
1566 | |
1567 | p = _pick_next_task_rt(rq); | |
917b627d GH |
1568 | |
1569 | /* The running task is never eligible for pushing */ | |
f3f1768f | 1570 | dequeue_pushable_task(rq, p); |
917b627d | 1571 | |
e3fca9e7 | 1572 | queue_push_tasks(rq); |
3f029d3c | 1573 | |
6f505b16 | 1574 | return p; |
bb44e5d1 IM |
1575 | } |
1576 | ||
31ee529c | 1577 | static void put_prev_task_rt(struct rq *rq, struct task_struct *p) |
bb44e5d1 | 1578 | { |
f1e14ef6 | 1579 | update_curr_rt(rq); |
917b627d GH |
1580 | |
1581 | /* | |
1582 | * The previous task needs to be made eligible for pushing | |
1583 | * if it is still active | |
1584 | */ | |
4b53a341 | 1585 | if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1) |
917b627d | 1586 | enqueue_pushable_task(rq, p); |
bb44e5d1 IM |
1587 | } |
1588 | ||
681f3e68 | 1589 | #ifdef CONFIG_SMP |
6f505b16 | 1590 | |
e8fa1362 SR |
1591 | /* Only try algorithms three times */ |
1592 | #define RT_MAX_TRIES 3 | |
1593 | ||
f65eda4f SR |
1594 | static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) |
1595 | { | |
1596 | if (!task_running(rq, p) && | |
0c98d344 | 1597 | cpumask_test_cpu(cpu, &p->cpus_allowed)) |
f65eda4f SR |
1598 | return 1; |
1599 | return 0; | |
1600 | } | |
1601 | ||
e23ee747 KT |
1602 | /* |
1603 | * Return the highest pushable rq's task, which is suitable to be executed | |
1604 | * on the cpu, NULL otherwise | |
1605 | */ | |
1606 | static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu) | |
e8fa1362 | 1607 | { |
e23ee747 KT |
1608 | struct plist_head *head = &rq->rt.pushable_tasks; |
1609 | struct task_struct *p; | |
3d07467b | 1610 | |
e23ee747 KT |
1611 | if (!has_pushable_tasks(rq)) |
1612 | return NULL; | |
3d07467b | 1613 | |
e23ee747 KT |
1614 | plist_for_each_entry(p, head, pushable_tasks) { |
1615 | if (pick_rt_task(rq, p, cpu)) | |
1616 | return p; | |
f65eda4f SR |
1617 | } |
1618 | ||
e23ee747 | 1619 | return NULL; |
e8fa1362 SR |
1620 | } |
1621 | ||
0e3900e6 | 1622 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask); |
e8fa1362 | 1623 | |
6e1254d2 GH |
1624 | static int find_lowest_rq(struct task_struct *task) |
1625 | { | |
1626 | struct sched_domain *sd; | |
4ba29684 | 1627 | struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask); |
6e1254d2 GH |
1628 | int this_cpu = smp_processor_id(); |
1629 | int cpu = task_cpu(task); | |
06f90dbd | 1630 | |
0da938c4 SR |
1631 | /* Make sure the mask is initialized first */ |
1632 | if (unlikely(!lowest_mask)) | |
1633 | return -1; | |
1634 | ||
4b53a341 | 1635 | if (task->nr_cpus_allowed == 1) |
6e0534f2 | 1636 | return -1; /* No other targets possible */ |
6e1254d2 | 1637 | |
6e0534f2 GH |
1638 | if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask)) |
1639 | return -1; /* No targets found */ | |
6e1254d2 GH |
1640 | |
1641 | /* | |
1642 | * At this point we have built a mask of cpus representing the | |
1643 | * lowest priority tasks in the system. Now we want to elect | |
1644 | * the best one based on our affinity and topology. | |
1645 | * | |
1646 | * We prioritize the last cpu that the task executed on since | |
1647 | * it is most likely cache-hot in that location. | |
1648 | */ | |
96f874e2 | 1649 | if (cpumask_test_cpu(cpu, lowest_mask)) |
6e1254d2 GH |
1650 | return cpu; |
1651 | ||
1652 | /* | |
1653 | * Otherwise, we consult the sched_domains span maps to figure | |
1654 | * out which cpu is logically closest to our hot cache data. | |
1655 | */ | |
e2c88063 RR |
1656 | if (!cpumask_test_cpu(this_cpu, lowest_mask)) |
1657 | this_cpu = -1; /* Skip this_cpu opt if not among lowest */ | |
6e1254d2 | 1658 | |
cd4ae6ad | 1659 | rcu_read_lock(); |
e2c88063 RR |
1660 | for_each_domain(cpu, sd) { |
1661 | if (sd->flags & SD_WAKE_AFFINE) { | |
1662 | int best_cpu; | |
6e1254d2 | 1663 | |
e2c88063 RR |
1664 | /* |
1665 | * "this_cpu" is cheaper to preempt than a | |
1666 | * remote processor. | |
1667 | */ | |
1668 | if (this_cpu != -1 && | |
cd4ae6ad XF |
1669 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { |
1670 | rcu_read_unlock(); | |
e2c88063 | 1671 | return this_cpu; |
cd4ae6ad | 1672 | } |
e2c88063 RR |
1673 | |
1674 | best_cpu = cpumask_first_and(lowest_mask, | |
1675 | sched_domain_span(sd)); | |
cd4ae6ad XF |
1676 | if (best_cpu < nr_cpu_ids) { |
1677 | rcu_read_unlock(); | |
e2c88063 | 1678 | return best_cpu; |
cd4ae6ad | 1679 | } |
6e1254d2 GH |
1680 | } |
1681 | } | |
cd4ae6ad | 1682 | rcu_read_unlock(); |
6e1254d2 GH |
1683 | |
1684 | /* | |
1685 | * And finally, if there were no matches within the domains | |
1686 | * just give the caller *something* to work with from the compatible | |
1687 | * locations. | |
1688 | */ | |
e2c88063 RR |
1689 | if (this_cpu != -1) |
1690 | return this_cpu; | |
1691 | ||
1692 | cpu = cpumask_any(lowest_mask); | |
1693 | if (cpu < nr_cpu_ids) | |
1694 | return cpu; | |
1695 | return -1; | |
07b4032c GH |
1696 | } |
1697 | ||
1698 | /* Will lock the rq it finds */ | |
4df64c0b | 1699 | static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) |
07b4032c GH |
1700 | { |
1701 | struct rq *lowest_rq = NULL; | |
07b4032c | 1702 | int tries; |
4df64c0b | 1703 | int cpu; |
e8fa1362 | 1704 | |
07b4032c GH |
1705 | for (tries = 0; tries < RT_MAX_TRIES; tries++) { |
1706 | cpu = find_lowest_rq(task); | |
1707 | ||
2de0b463 | 1708 | if ((cpu == -1) || (cpu == rq->cpu)) |
e8fa1362 SR |
1709 | break; |
1710 | ||
07b4032c GH |
1711 | lowest_rq = cpu_rq(cpu); |
1712 | ||
80e3d87b TC |
1713 | if (lowest_rq->rt.highest_prio.curr <= task->prio) { |
1714 | /* | |
1715 | * Target rq has tasks of equal or higher priority, | |
1716 | * retrying does not release any lock and is unlikely | |
1717 | * to yield a different result. | |
1718 | */ | |
1719 | lowest_rq = NULL; | |
1720 | break; | |
1721 | } | |
1722 | ||
e8fa1362 | 1723 | /* if the prio of this runqueue changed, try again */ |
07b4032c | 1724 | if (double_lock_balance(rq, lowest_rq)) { |
e8fa1362 SR |
1725 | /* |
1726 | * We had to unlock the run queue. In | |
1727 | * the mean time, task could have | |
1728 | * migrated already or had its affinity changed. | |
1729 | * Also make sure that it wasn't scheduled on its rq. | |
1730 | */ | |
07b4032c | 1731 | if (unlikely(task_rq(task) != rq || |
0c98d344 | 1732 | !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_allowed) || |
07b4032c | 1733 | task_running(rq, task) || |
13b5ab02 | 1734 | !rt_task(task) || |
da0c1e65 | 1735 | !task_on_rq_queued(task))) { |
4df64c0b | 1736 | |
7f1b4393 | 1737 | double_unlock_balance(rq, lowest_rq); |
e8fa1362 SR |
1738 | lowest_rq = NULL; |
1739 | break; | |
1740 | } | |
1741 | } | |
1742 | ||
1743 | /* If this rq is still suitable use it. */ | |
e864c499 | 1744 | if (lowest_rq->rt.highest_prio.curr > task->prio) |
e8fa1362 SR |
1745 | break; |
1746 | ||
1747 | /* try again */ | |
1b12bbc7 | 1748 | double_unlock_balance(rq, lowest_rq); |
e8fa1362 SR |
1749 | lowest_rq = NULL; |
1750 | } | |
1751 | ||
1752 | return lowest_rq; | |
1753 | } | |
1754 | ||
917b627d GH |
1755 | static struct task_struct *pick_next_pushable_task(struct rq *rq) |
1756 | { | |
1757 | struct task_struct *p; | |
1758 | ||
1759 | if (!has_pushable_tasks(rq)) | |
1760 | return NULL; | |
1761 | ||
1762 | p = plist_first_entry(&rq->rt.pushable_tasks, | |
1763 | struct task_struct, pushable_tasks); | |
1764 | ||
1765 | BUG_ON(rq->cpu != task_cpu(p)); | |
1766 | BUG_ON(task_current(rq, p)); | |
4b53a341 | 1767 | BUG_ON(p->nr_cpus_allowed <= 1); |
917b627d | 1768 | |
da0c1e65 | 1769 | BUG_ON(!task_on_rq_queued(p)); |
917b627d GH |
1770 | BUG_ON(!rt_task(p)); |
1771 | ||
1772 | return p; | |
1773 | } | |
1774 | ||
e8fa1362 SR |
1775 | /* |
1776 | * If the current CPU has more than one RT task, see if the non | |
1777 | * running task can migrate over to a CPU that is running a task | |
1778 | * of lesser priority. | |
1779 | */ | |
697f0a48 | 1780 | static int push_rt_task(struct rq *rq) |
e8fa1362 SR |
1781 | { |
1782 | struct task_struct *next_task; | |
1783 | struct rq *lowest_rq; | |
311e800e | 1784 | int ret = 0; |
e8fa1362 | 1785 | |
a22d7fc1 GH |
1786 | if (!rq->rt.overloaded) |
1787 | return 0; | |
1788 | ||
917b627d | 1789 | next_task = pick_next_pushable_task(rq); |
e8fa1362 SR |
1790 | if (!next_task) |
1791 | return 0; | |
1792 | ||
49246274 | 1793 | retry: |
697f0a48 | 1794 | if (unlikely(next_task == rq->curr)) { |
f65eda4f | 1795 | WARN_ON(1); |
e8fa1362 | 1796 | return 0; |
f65eda4f | 1797 | } |
e8fa1362 SR |
1798 | |
1799 | /* | |
1800 | * It's possible that the next_task slipped in of | |
1801 | * higher priority than current. If that's the case | |
1802 | * just reschedule current. | |
1803 | */ | |
697f0a48 | 1804 | if (unlikely(next_task->prio < rq->curr->prio)) { |
8875125e | 1805 | resched_curr(rq); |
e8fa1362 SR |
1806 | return 0; |
1807 | } | |
1808 | ||
697f0a48 | 1809 | /* We might release rq lock */ |
e8fa1362 SR |
1810 | get_task_struct(next_task); |
1811 | ||
1812 | /* find_lock_lowest_rq locks the rq if found */ | |
697f0a48 | 1813 | lowest_rq = find_lock_lowest_rq(next_task, rq); |
e8fa1362 SR |
1814 | if (!lowest_rq) { |
1815 | struct task_struct *task; | |
1816 | /* | |
311e800e | 1817 | * find_lock_lowest_rq releases rq->lock |
1563513d GH |
1818 | * so it is possible that next_task has migrated. |
1819 | * | |
1820 | * We need to make sure that the task is still on the same | |
1821 | * run-queue and is also still the next task eligible for | |
1822 | * pushing. | |
e8fa1362 | 1823 | */ |
917b627d | 1824 | task = pick_next_pushable_task(rq); |
de16b91e | 1825 | if (task == next_task) { |
1563513d | 1826 | /* |
311e800e HD |
1827 | * The task hasn't migrated, and is still the next |
1828 | * eligible task, but we failed to find a run-queue | |
1829 | * to push it to. Do not retry in this case, since | |
1830 | * other cpus will pull from us when ready. | |
1563513d | 1831 | */ |
1563513d | 1832 | goto out; |
e8fa1362 | 1833 | } |
917b627d | 1834 | |
1563513d GH |
1835 | if (!task) |
1836 | /* No more tasks, just exit */ | |
1837 | goto out; | |
1838 | ||
917b627d | 1839 | /* |
1563513d | 1840 | * Something has shifted, try again. |
917b627d | 1841 | */ |
1563513d GH |
1842 | put_task_struct(next_task); |
1843 | next_task = task; | |
1844 | goto retry; | |
e8fa1362 SR |
1845 | } |
1846 | ||
697f0a48 | 1847 | deactivate_task(rq, next_task, 0); |
e8fa1362 SR |
1848 | set_task_cpu(next_task, lowest_rq->cpu); |
1849 | activate_task(lowest_rq, next_task, 0); | |
311e800e | 1850 | ret = 1; |
e8fa1362 | 1851 | |
8875125e | 1852 | resched_curr(lowest_rq); |
e8fa1362 | 1853 | |
1b12bbc7 | 1854 | double_unlock_balance(rq, lowest_rq); |
e8fa1362 | 1855 | |
e8fa1362 SR |
1856 | out: |
1857 | put_task_struct(next_task); | |
1858 | ||
311e800e | 1859 | return ret; |
e8fa1362 SR |
1860 | } |
1861 | ||
e8fa1362 SR |
1862 | static void push_rt_tasks(struct rq *rq) |
1863 | { | |
1864 | /* push_rt_task will return true if it moved an RT */ | |
1865 | while (push_rt_task(rq)) | |
1866 | ; | |
1867 | } | |
1868 | ||
b6366f04 | 1869 | #ifdef HAVE_RT_PUSH_IPI |
4bdced5c | 1870 | |
b6366f04 | 1871 | /* |
4bdced5c SRRH |
1872 | * When a high priority task schedules out from a CPU and a lower priority |
1873 | * task is scheduled in, a check is made to see if there's any RT tasks | |
1874 | * on other CPUs that are waiting to run because a higher priority RT task | |
1875 | * is currently running on its CPU. In this case, the CPU with multiple RT | |
1876 | * tasks queued on it (overloaded) needs to be notified that a CPU has opened | |
1877 | * up that may be able to run one of its non-running queued RT tasks. | |
1878 | * | |
1879 | * All CPUs with overloaded RT tasks need to be notified as there is currently | |
1880 | * no way to know which of these CPUs have the highest priority task waiting | |
1881 | * to run. Instead of trying to take a spinlock on each of these CPUs, | |
1882 | * which has shown to cause large latency when done on machines with many | |
1883 | * CPUs, sending an IPI to the CPUs to have them push off the overloaded | |
1884 | * RT tasks waiting to run. | |
1885 | * | |
1886 | * Just sending an IPI to each of the CPUs is also an issue, as on large | |
1887 | * count CPU machines, this can cause an IPI storm on a CPU, especially | |
1888 | * if its the only CPU with multiple RT tasks queued, and a large number | |
1889 | * of CPUs scheduling a lower priority task at the same time. | |
1890 | * | |
1891 | * Each root domain has its own irq work function that can iterate over | |
1892 | * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT | |
1893 | * tassk must be checked if there's one or many CPUs that are lowering | |
1894 | * their priority, there's a single irq work iterator that will try to | |
1895 | * push off RT tasks that are waiting to run. | |
1896 | * | |
1897 | * When a CPU schedules a lower priority task, it will kick off the | |
1898 | * irq work iterator that will jump to each CPU with overloaded RT tasks. | |
1899 | * As it only takes the first CPU that schedules a lower priority task | |
1900 | * to start the process, the rto_start variable is incremented and if | |
1901 | * the atomic result is one, then that CPU will try to take the rto_lock. | |
1902 | * This prevents high contention on the lock as the process handles all | |
1903 | * CPUs scheduling lower priority tasks. | |
1904 | * | |
1905 | * All CPUs that are scheduling a lower priority task will increment the | |
1906 | * rt_loop_next variable. This will make sure that the irq work iterator | |
1907 | * checks all RT overloaded CPUs whenever a CPU schedules a new lower | |
1908 | * priority task, even if the iterator is in the middle of a scan. Incrementing | |
1909 | * the rt_loop_next will cause the iterator to perform another scan. | |
b6366f04 | 1910 | * |
b6366f04 | 1911 | */ |
189c78df | 1912 | static int rto_next_cpu(struct root_domain *rd) |
b6366f04 | 1913 | { |
4bdced5c | 1914 | int next; |
b6366f04 SR |
1915 | int cpu; |
1916 | ||
b6366f04 | 1917 | /* |
4bdced5c SRRH |
1918 | * When starting the IPI RT pushing, the rto_cpu is set to -1, |
1919 | * rt_next_cpu() will simply return the first CPU found in | |
1920 | * the rto_mask. | |
1921 | * | |
1922 | * If rto_next_cpu() is called with rto_cpu is a valid cpu, it | |
1923 | * will return the next CPU found in the rto_mask. | |
1924 | * | |
1925 | * If there are no more CPUs left in the rto_mask, then a check is made | |
1926 | * against rto_loop and rto_loop_next. rto_loop is only updated with | |
1927 | * the rto_lock held, but any CPU may increment the rto_loop_next | |
1928 | * without any locking. | |
b6366f04 | 1929 | */ |
4bdced5c | 1930 | for (;;) { |
b6366f04 | 1931 | |
4bdced5c SRRH |
1932 | /* When rto_cpu is -1 this acts like cpumask_first() */ |
1933 | cpu = cpumask_next(rd->rto_cpu, rd->rto_mask); | |
b6366f04 | 1934 | |
4bdced5c | 1935 | rd->rto_cpu = cpu; |
b6366f04 | 1936 | |
4bdced5c SRRH |
1937 | if (cpu < nr_cpu_ids) |
1938 | return cpu; | |
b6366f04 | 1939 | |
4bdced5c SRRH |
1940 | rd->rto_cpu = -1; |
1941 | ||
1942 | /* | |
1943 | * ACQUIRE ensures we see the @rto_mask changes | |
1944 | * made prior to the @next value observed. | |
1945 | * | |
1946 | * Matches WMB in rt_set_overload(). | |
1947 | */ | |
1948 | next = atomic_read_acquire(&rd->rto_loop_next); | |
b6366f04 | 1949 | |
4bdced5c | 1950 | if (rd->rto_loop == next) |
b6366f04 | 1951 | break; |
4bdced5c SRRH |
1952 | |
1953 | rd->rto_loop = next; | |
b6366f04 SR |
1954 | } |
1955 | ||
4bdced5c | 1956 | return -1; |
b6366f04 SR |
1957 | } |
1958 | ||
4bdced5c SRRH |
1959 | static inline bool rto_start_trylock(atomic_t *v) |
1960 | { | |
1961 | return !atomic_cmpxchg_acquire(v, 0, 1); | |
1962 | } | |
b6366f04 | 1963 | |
4bdced5c | 1964 | static inline void rto_start_unlock(atomic_t *v) |
b6366f04 | 1965 | { |
4bdced5c SRRH |
1966 | atomic_set_release(v, 0); |
1967 | } | |
b6366f04 | 1968 | |
4bdced5c SRRH |
1969 | static void tell_cpu_to_push(struct rq *rq) |
1970 | { | |
1971 | int cpu = -1; | |
b6366f04 | 1972 | |
4bdced5c SRRH |
1973 | /* Keep the loop going if the IPI is currently active */ |
1974 | atomic_inc(&rq->rd->rto_loop_next); | |
b6366f04 | 1975 | |
4bdced5c SRRH |
1976 | /* Only one CPU can initiate a loop at a time */ |
1977 | if (!rto_start_trylock(&rq->rd->rto_loop_start)) | |
b6366f04 SR |
1978 | return; |
1979 | ||
4bdced5c | 1980 | raw_spin_lock(&rq->rd->rto_lock); |
b6366f04 | 1981 | |
4bdced5c SRRH |
1982 | /* |
1983 | * The rto_cpu is updated under the lock, if it has a valid cpu | |
1984 | * then the IPI is still running and will continue due to the | |
1985 | * update to loop_next, and nothing needs to be done here. | |
1986 | * Otherwise it is finishing up and an ipi needs to be sent. | |
1987 | */ | |
1988 | if (rq->rd->rto_cpu < 0) | |
189c78df | 1989 | cpu = rto_next_cpu(rq->rd); |
4bdced5c SRRH |
1990 | |
1991 | raw_spin_unlock(&rq->rd->rto_lock); | |
1992 | ||
1993 | rto_start_unlock(&rq->rd->rto_loop_start); | |
1994 | ||
3add5ce3 SRV |
1995 | if (cpu >= 0) { |
1996 | /* Make sure the rd does not get freed while pushing */ | |
1997 | sched_get_rd(rq->rd); | |
4bdced5c | 1998 | irq_work_queue_on(&rq->rd->rto_push_work, cpu); |
3add5ce3 | 1999 | } |
b6366f04 SR |
2000 | } |
2001 | ||
2002 | /* Called from hardirq context */ | |
4bdced5c | 2003 | void rto_push_irq_work_func(struct irq_work *work) |
b6366f04 | 2004 | { |
189c78df SRV |
2005 | struct root_domain *rd = |
2006 | container_of(work, struct root_domain, rto_push_work); | |
4bdced5c | 2007 | struct rq *rq; |
b6366f04 SR |
2008 | int cpu; |
2009 | ||
4bdced5c | 2010 | rq = this_rq(); |
b6366f04 | 2011 | |
4bdced5c SRRH |
2012 | /* |
2013 | * We do not need to grab the lock to check for has_pushable_tasks. | |
2014 | * When it gets updated, a check is made if a push is possible. | |
2015 | */ | |
b6366f04 SR |
2016 | if (has_pushable_tasks(rq)) { |
2017 | raw_spin_lock(&rq->lock); | |
4bdced5c | 2018 | push_rt_tasks(rq); |
b6366f04 SR |
2019 | raw_spin_unlock(&rq->lock); |
2020 | } | |
2021 | ||
189c78df | 2022 | raw_spin_lock(&rd->rto_lock); |
b6366f04 | 2023 | |
4bdced5c | 2024 | /* Pass the IPI to the next rt overloaded queue */ |
189c78df | 2025 | cpu = rto_next_cpu(rd); |
b6366f04 | 2026 | |
189c78df | 2027 | raw_spin_unlock(&rd->rto_lock); |
b6366f04 | 2028 | |
3add5ce3 SRV |
2029 | if (cpu < 0) { |
2030 | sched_put_rd(rd); | |
b6366f04 | 2031 | return; |
3add5ce3 | 2032 | } |
b6366f04 | 2033 | |
b6366f04 | 2034 | /* Try the next RT overloaded CPU */ |
189c78df | 2035 | irq_work_queue_on(&rd->rto_push_work, cpu); |
b6366f04 SR |
2036 | } |
2037 | #endif /* HAVE_RT_PUSH_IPI */ | |
2038 | ||
8046d680 | 2039 | static void pull_rt_task(struct rq *this_rq) |
f65eda4f | 2040 | { |
8046d680 PZ |
2041 | int this_cpu = this_rq->cpu, cpu; |
2042 | bool resched = false; | |
a8728944 | 2043 | struct task_struct *p; |
f65eda4f | 2044 | struct rq *src_rq; |
f73c52a5 | 2045 | int rt_overload_count = rt_overloaded(this_rq); |
f65eda4f | 2046 | |
f73c52a5 | 2047 | if (likely(!rt_overload_count)) |
8046d680 | 2048 | return; |
f65eda4f | 2049 | |
7c3f2ab7 PZ |
2050 | /* |
2051 | * Match the barrier from rt_set_overloaded; this guarantees that if we | |
2052 | * see overloaded we must also see the rto_mask bit. | |
2053 | */ | |
2054 | smp_rmb(); | |
2055 | ||
f73c52a5 SR |
2056 | /* If we are the only overloaded CPU do nothing */ |
2057 | if (rt_overload_count == 1 && | |
2058 | cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask)) | |
2059 | return; | |
2060 | ||
b6366f04 SR |
2061 | #ifdef HAVE_RT_PUSH_IPI |
2062 | if (sched_feat(RT_PUSH_IPI)) { | |
2063 | tell_cpu_to_push(this_rq); | |
8046d680 | 2064 | return; |
b6366f04 SR |
2065 | } |
2066 | #endif | |
2067 | ||
c6c4927b | 2068 | for_each_cpu(cpu, this_rq->rd->rto_mask) { |
f65eda4f SR |
2069 | if (this_cpu == cpu) |
2070 | continue; | |
2071 | ||
2072 | src_rq = cpu_rq(cpu); | |
74ab8e4f GH |
2073 | |
2074 | /* | |
2075 | * Don't bother taking the src_rq->lock if the next highest | |
2076 | * task is known to be lower-priority than our current task. | |
2077 | * This may look racy, but if this value is about to go | |
2078 | * logically higher, the src_rq will push this task away. | |
2079 | * And if its going logically lower, we do not care | |
2080 | */ | |
2081 | if (src_rq->rt.highest_prio.next >= | |
2082 | this_rq->rt.highest_prio.curr) | |
2083 | continue; | |
2084 | ||
f65eda4f SR |
2085 | /* |
2086 | * We can potentially drop this_rq's lock in | |
2087 | * double_lock_balance, and another CPU could | |
a8728944 | 2088 | * alter this_rq |
f65eda4f | 2089 | */ |
a8728944 | 2090 | double_lock_balance(this_rq, src_rq); |
f65eda4f SR |
2091 | |
2092 | /* | |
e23ee747 KT |
2093 | * We can pull only a task, which is pushable |
2094 | * on its rq, and no others. | |
f65eda4f | 2095 | */ |
e23ee747 | 2096 | p = pick_highest_pushable_task(src_rq, this_cpu); |
f65eda4f SR |
2097 | |
2098 | /* | |
2099 | * Do we have an RT task that preempts | |
2100 | * the to-be-scheduled task? | |
2101 | */ | |
a8728944 | 2102 | if (p && (p->prio < this_rq->rt.highest_prio.curr)) { |
f65eda4f | 2103 | WARN_ON(p == src_rq->curr); |
da0c1e65 | 2104 | WARN_ON(!task_on_rq_queued(p)); |
f65eda4f SR |
2105 | |
2106 | /* | |
2107 | * There's a chance that p is higher in priority | |
2108 | * than what's currently running on its cpu. | |
2109 | * This is just that p is wakeing up and hasn't | |
2110 | * had a chance to schedule. We only pull | |
2111 | * p if it is lower in priority than the | |
a8728944 | 2112 | * current task on the run queue |
f65eda4f | 2113 | */ |
a8728944 | 2114 | if (p->prio < src_rq->curr->prio) |
614ee1f6 | 2115 | goto skip; |
f65eda4f | 2116 | |
8046d680 | 2117 | resched = true; |
f65eda4f SR |
2118 | |
2119 | deactivate_task(src_rq, p, 0); | |
2120 | set_task_cpu(p, this_cpu); | |
2121 | activate_task(this_rq, p, 0); | |
2122 | /* | |
2123 | * We continue with the search, just in | |
2124 | * case there's an even higher prio task | |
25985edc | 2125 | * in another runqueue. (low likelihood |
f65eda4f | 2126 | * but possible) |
f65eda4f | 2127 | */ |
f65eda4f | 2128 | } |
49246274 | 2129 | skip: |
1b12bbc7 | 2130 | double_unlock_balance(this_rq, src_rq); |
f65eda4f SR |
2131 | } |
2132 | ||
8046d680 PZ |
2133 | if (resched) |
2134 | resched_curr(this_rq); | |
f65eda4f SR |
2135 | } |
2136 | ||
8ae121ac GH |
2137 | /* |
2138 | * If we are not running and we are not going to reschedule soon, we should | |
2139 | * try to push tasks away now | |
2140 | */ | |
efbbd05a | 2141 | static void task_woken_rt(struct rq *rq, struct task_struct *p) |
4642dafd | 2142 | { |
9a897c5a | 2143 | if (!task_running(rq, p) && |
8ae121ac | 2144 | !test_tsk_need_resched(rq->curr) && |
4b53a341 | 2145 | p->nr_cpus_allowed > 1 && |
1baca4ce | 2146 | (dl_task(rq->curr) || rt_task(rq->curr)) && |
4b53a341 | 2147 | (rq->curr->nr_cpus_allowed < 2 || |
3be209a8 | 2148 | rq->curr->prio <= p->prio)) |
4642dafd SR |
2149 | push_rt_tasks(rq); |
2150 | } | |
2151 | ||
bdd7c81b | 2152 | /* Assumes rq->lock is held */ |
1f11eb6a | 2153 | static void rq_online_rt(struct rq *rq) |
bdd7c81b IM |
2154 | { |
2155 | if (rq->rt.overloaded) | |
2156 | rt_set_overload(rq); | |
6e0534f2 | 2157 | |
7def2be1 PZ |
2158 | __enable_runtime(rq); |
2159 | ||
e864c499 | 2160 | cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr); |
bdd7c81b IM |
2161 | } |
2162 | ||
2163 | /* Assumes rq->lock is held */ | |
1f11eb6a | 2164 | static void rq_offline_rt(struct rq *rq) |
bdd7c81b IM |
2165 | { |
2166 | if (rq->rt.overloaded) | |
2167 | rt_clear_overload(rq); | |
6e0534f2 | 2168 | |
7def2be1 PZ |
2169 | __disable_runtime(rq); |
2170 | ||
6e0534f2 | 2171 | cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID); |
bdd7c81b | 2172 | } |
cb469845 SR |
2173 | |
2174 | /* | |
2175 | * When switch from the rt queue, we bring ourselves to a position | |
2176 | * that we might want to pull RT tasks from other runqueues. | |
2177 | */ | |
da7a735e | 2178 | static void switched_from_rt(struct rq *rq, struct task_struct *p) |
cb469845 SR |
2179 | { |
2180 | /* | |
2181 | * If there are other RT tasks then we will reschedule | |
2182 | * and the scheduling of the other RT tasks will handle | |
2183 | * the balancing. But if we are the last RT task | |
2184 | * we may need to handle the pulling of RT tasks | |
2185 | * now. | |
2186 | */ | |
da0c1e65 | 2187 | if (!task_on_rq_queued(p) || rq->rt.rt_nr_running) |
1158ddb5 KT |
2188 | return; |
2189 | ||
fd7a4bed | 2190 | queue_pull_task(rq); |
cb469845 | 2191 | } |
3d8cbdf8 | 2192 | |
11c785b7 | 2193 | void __init init_sched_rt_class(void) |
3d8cbdf8 RR |
2194 | { |
2195 | unsigned int i; | |
2196 | ||
029632fb | 2197 | for_each_possible_cpu(i) { |
eaa95840 | 2198 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i), |
6ca09dfc | 2199 | GFP_KERNEL, cpu_to_node(i)); |
029632fb | 2200 | } |
3d8cbdf8 | 2201 | } |
cb469845 SR |
2202 | #endif /* CONFIG_SMP */ |
2203 | ||
2204 | /* | |
2205 | * When switching a task to RT, we may overload the runqueue | |
2206 | * with RT tasks. In this case we try to push them off to | |
2207 | * other runqueues. | |
2208 | */ | |
da7a735e | 2209 | static void switched_to_rt(struct rq *rq, struct task_struct *p) |
cb469845 | 2210 | { |
cb469845 SR |
2211 | /* |
2212 | * If we are already running, then there's nothing | |
2213 | * that needs to be done. But if we are not running | |
2214 | * we may need to preempt the current running task. | |
2215 | * If that current running task is also an RT task | |
2216 | * then see if we can move to another run queue. | |
2217 | */ | |
da0c1e65 | 2218 | if (task_on_rq_queued(p) && rq->curr != p) { |
cb469845 | 2219 | #ifdef CONFIG_SMP |
4b53a341 | 2220 | if (p->nr_cpus_allowed > 1 && rq->rt.overloaded) |
fd7a4bed | 2221 | queue_push_tasks(rq); |
619bd4a7 | 2222 | #endif /* CONFIG_SMP */ |
79d3dd40 | 2223 | if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq))) |
8875125e | 2224 | resched_curr(rq); |
cb469845 SR |
2225 | } |
2226 | } | |
2227 | ||
2228 | /* | |
2229 | * Priority of the task has changed. This may cause | |
2230 | * us to initiate a push or pull. | |
2231 | */ | |
da7a735e PZ |
2232 | static void |
2233 | prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 2234 | { |
da0c1e65 | 2235 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
2236 | return; |
2237 | ||
2238 | if (rq->curr == p) { | |
cb469845 SR |
2239 | #ifdef CONFIG_SMP |
2240 | /* | |
2241 | * If our priority decreases while running, we | |
2242 | * may need to pull tasks to this runqueue. | |
2243 | */ | |
2244 | if (oldprio < p->prio) | |
fd7a4bed PZ |
2245 | queue_pull_task(rq); |
2246 | ||
cb469845 SR |
2247 | /* |
2248 | * If there's a higher priority task waiting to run | |
fd7a4bed | 2249 | * then reschedule. |
cb469845 | 2250 | */ |
fd7a4bed | 2251 | if (p->prio > rq->rt.highest_prio.curr) |
8875125e | 2252 | resched_curr(rq); |
cb469845 SR |
2253 | #else |
2254 | /* For UP simply resched on drop of prio */ | |
2255 | if (oldprio < p->prio) | |
8875125e | 2256 | resched_curr(rq); |
e8fa1362 | 2257 | #endif /* CONFIG_SMP */ |
cb469845 SR |
2258 | } else { |
2259 | /* | |
2260 | * This task is not running, but if it is | |
2261 | * greater than the current running task | |
2262 | * then reschedule. | |
2263 | */ | |
2264 | if (p->prio < rq->curr->prio) | |
8875125e | 2265 | resched_curr(rq); |
cb469845 SR |
2266 | } |
2267 | } | |
2268 | ||
b18b6a9c | 2269 | #ifdef CONFIG_POSIX_TIMERS |
78f2c7db PZ |
2270 | static void watchdog(struct rq *rq, struct task_struct *p) |
2271 | { | |
2272 | unsigned long soft, hard; | |
2273 | ||
78d7d407 JS |
2274 | /* max may change after cur was read, this will be fixed next tick */ |
2275 | soft = task_rlimit(p, RLIMIT_RTTIME); | |
2276 | hard = task_rlimit_max(p, RLIMIT_RTTIME); | |
78f2c7db PZ |
2277 | |
2278 | if (soft != RLIM_INFINITY) { | |
2279 | unsigned long next; | |
2280 | ||
57d2aa00 YX |
2281 | if (p->rt.watchdog_stamp != jiffies) { |
2282 | p->rt.timeout++; | |
2283 | p->rt.watchdog_stamp = jiffies; | |
2284 | } | |
2285 | ||
78f2c7db | 2286 | next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ); |
5a52dd50 | 2287 | if (p->rt.timeout > next) |
f06febc9 | 2288 | p->cputime_expires.sched_exp = p->se.sum_exec_runtime; |
78f2c7db PZ |
2289 | } |
2290 | } | |
b18b6a9c NP |
2291 | #else |
2292 | static inline void watchdog(struct rq *rq, struct task_struct *p) { } | |
2293 | #endif | |
bb44e5d1 | 2294 | |
8f4d37ec | 2295 | static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued) |
bb44e5d1 | 2296 | { |
454c7999 CC |
2297 | struct sched_rt_entity *rt_se = &p->rt; |
2298 | ||
67e2be02 PZ |
2299 | update_curr_rt(rq); |
2300 | ||
78f2c7db PZ |
2301 | watchdog(rq, p); |
2302 | ||
bb44e5d1 IM |
2303 | /* |
2304 | * RR tasks need a special form of timeslice management. | |
2305 | * FIFO tasks have no timeslices. | |
2306 | */ | |
2307 | if (p->policy != SCHED_RR) | |
2308 | return; | |
2309 | ||
fa717060 | 2310 | if (--p->rt.time_slice) |
bb44e5d1 IM |
2311 | return; |
2312 | ||
ce0dbbbb | 2313 | p->rt.time_slice = sched_rr_timeslice; |
bb44e5d1 | 2314 | |
98fbc798 | 2315 | /* |
e9aa39bb LB |
2316 | * Requeue to the end of queue if we (and all of our ancestors) are not |
2317 | * the only element on the queue | |
98fbc798 | 2318 | */ |
454c7999 CC |
2319 | for_each_sched_rt_entity(rt_se) { |
2320 | if (rt_se->run_list.prev != rt_se->run_list.next) { | |
2321 | requeue_task_rt(rq, p, 0); | |
8aa6f0eb | 2322 | resched_curr(rq); |
454c7999 CC |
2323 | return; |
2324 | } | |
98fbc798 | 2325 | } |
bb44e5d1 IM |
2326 | } |
2327 | ||
83b699ed SV |
2328 | static void set_curr_task_rt(struct rq *rq) |
2329 | { | |
2330 | struct task_struct *p = rq->curr; | |
2331 | ||
78becc27 | 2332 | p->se.exec_start = rq_clock_task(rq); |
917b627d GH |
2333 | |
2334 | /* The running task is never eligible for pushing */ | |
2335 | dequeue_pushable_task(rq, p); | |
83b699ed SV |
2336 | } |
2337 | ||
6d686f45 | 2338 | static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task) |
0d721cea PW |
2339 | { |
2340 | /* | |
2341 | * Time slice is 0 for SCHED_FIFO tasks | |
2342 | */ | |
2343 | if (task->policy == SCHED_RR) | |
ce0dbbbb | 2344 | return sched_rr_timeslice; |
0d721cea PW |
2345 | else |
2346 | return 0; | |
2347 | } | |
2348 | ||
029632fb | 2349 | const struct sched_class rt_sched_class = { |
5522d5d5 | 2350 | .next = &fair_sched_class, |
bb44e5d1 IM |
2351 | .enqueue_task = enqueue_task_rt, |
2352 | .dequeue_task = dequeue_task_rt, | |
2353 | .yield_task = yield_task_rt, | |
2354 | ||
2355 | .check_preempt_curr = check_preempt_curr_rt, | |
2356 | ||
2357 | .pick_next_task = pick_next_task_rt, | |
2358 | .put_prev_task = put_prev_task_rt, | |
2359 | ||
681f3e68 | 2360 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
2361 | .select_task_rq = select_task_rq_rt, |
2362 | ||
6c37067e | 2363 | .set_cpus_allowed = set_cpus_allowed_common, |
1f11eb6a GH |
2364 | .rq_online = rq_online_rt, |
2365 | .rq_offline = rq_offline_rt, | |
efbbd05a | 2366 | .task_woken = task_woken_rt, |
cb469845 | 2367 | .switched_from = switched_from_rt, |
681f3e68 | 2368 | #endif |
bb44e5d1 | 2369 | |
83b699ed | 2370 | .set_curr_task = set_curr_task_rt, |
bb44e5d1 | 2371 | .task_tick = task_tick_rt, |
cb469845 | 2372 | |
0d721cea PW |
2373 | .get_rr_interval = get_rr_interval_rt, |
2374 | ||
cb469845 SR |
2375 | .prio_changed = prio_changed_rt, |
2376 | .switched_to = switched_to_rt, | |
6e998916 SG |
2377 | |
2378 | .update_curr = update_curr_rt, | |
bb44e5d1 | 2379 | }; |
ada18de2 | 2380 | |
8887cd99 NP |
2381 | #ifdef CONFIG_RT_GROUP_SCHED |
2382 | /* | |
2383 | * Ensure that the real time constraints are schedulable. | |
2384 | */ | |
2385 | static DEFINE_MUTEX(rt_constraints_mutex); | |
2386 | ||
2387 | /* Must be called with tasklist_lock held */ | |
2388 | static inline int tg_has_rt_tasks(struct task_group *tg) | |
2389 | { | |
2390 | struct task_struct *g, *p; | |
2391 | ||
2392 | /* | |
2393 | * Autogroups do not have RT tasks; see autogroup_create(). | |
2394 | */ | |
2395 | if (task_group_is_autogroup(tg)) | |
2396 | return 0; | |
2397 | ||
2398 | for_each_process_thread(g, p) { | |
2399 | if (rt_task(p) && task_group(p) == tg) | |
2400 | return 1; | |
2401 | } | |
2402 | ||
2403 | return 0; | |
2404 | } | |
2405 | ||
2406 | struct rt_schedulable_data { | |
2407 | struct task_group *tg; | |
2408 | u64 rt_period; | |
2409 | u64 rt_runtime; | |
2410 | }; | |
2411 | ||
2412 | static int tg_rt_schedulable(struct task_group *tg, void *data) | |
2413 | { | |
2414 | struct rt_schedulable_data *d = data; | |
2415 | struct task_group *child; | |
2416 | unsigned long total, sum = 0; | |
2417 | u64 period, runtime; | |
2418 | ||
2419 | period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
2420 | runtime = tg->rt_bandwidth.rt_runtime; | |
2421 | ||
2422 | if (tg == d->tg) { | |
2423 | period = d->rt_period; | |
2424 | runtime = d->rt_runtime; | |
2425 | } | |
2426 | ||
2427 | /* | |
2428 | * Cannot have more runtime than the period. | |
2429 | */ | |
2430 | if (runtime > period && runtime != RUNTIME_INF) | |
2431 | return -EINVAL; | |
2432 | ||
2433 | /* | |
2434 | * Ensure we don't starve existing RT tasks. | |
2435 | */ | |
2436 | if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) | |
2437 | return -EBUSY; | |
2438 | ||
2439 | total = to_ratio(period, runtime); | |
2440 | ||
2441 | /* | |
2442 | * Nobody can have more than the global setting allows. | |
2443 | */ | |
2444 | if (total > to_ratio(global_rt_period(), global_rt_runtime())) | |
2445 | return -EINVAL; | |
2446 | ||
2447 | /* | |
2448 | * The sum of our children's runtime should not exceed our own. | |
2449 | */ | |
2450 | list_for_each_entry_rcu(child, &tg->children, siblings) { | |
2451 | period = ktime_to_ns(child->rt_bandwidth.rt_period); | |
2452 | runtime = child->rt_bandwidth.rt_runtime; | |
2453 | ||
2454 | if (child == d->tg) { | |
2455 | period = d->rt_period; | |
2456 | runtime = d->rt_runtime; | |
2457 | } | |
2458 | ||
2459 | sum += to_ratio(period, runtime); | |
2460 | } | |
2461 | ||
2462 | if (sum > total) | |
2463 | return -EINVAL; | |
2464 | ||
2465 | return 0; | |
2466 | } | |
2467 | ||
2468 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) | |
2469 | { | |
2470 | int ret; | |
2471 | ||
2472 | struct rt_schedulable_data data = { | |
2473 | .tg = tg, | |
2474 | .rt_period = period, | |
2475 | .rt_runtime = runtime, | |
2476 | }; | |
2477 | ||
2478 | rcu_read_lock(); | |
2479 | ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); | |
2480 | rcu_read_unlock(); | |
2481 | ||
2482 | return ret; | |
2483 | } | |
2484 | ||
2485 | static int tg_set_rt_bandwidth(struct task_group *tg, | |
2486 | u64 rt_period, u64 rt_runtime) | |
2487 | { | |
2488 | int i, err = 0; | |
2489 | ||
2490 | /* | |
2491 | * Disallowing the root group RT runtime is BAD, it would disallow the | |
2492 | * kernel creating (and or operating) RT threads. | |
2493 | */ | |
2494 | if (tg == &root_task_group && rt_runtime == 0) | |
2495 | return -EINVAL; | |
2496 | ||
2497 | /* No period doesn't make any sense. */ | |
2498 | if (rt_period == 0) | |
2499 | return -EINVAL; | |
2500 | ||
2501 | mutex_lock(&rt_constraints_mutex); | |
2502 | read_lock(&tasklist_lock); | |
2503 | err = __rt_schedulable(tg, rt_period, rt_runtime); | |
2504 | if (err) | |
2505 | goto unlock; | |
2506 | ||
2507 | raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
2508 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); | |
2509 | tg->rt_bandwidth.rt_runtime = rt_runtime; | |
2510 | ||
2511 | for_each_possible_cpu(i) { | |
2512 | struct rt_rq *rt_rq = tg->rt_rq[i]; | |
2513 | ||
2514 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
2515 | rt_rq->rt_runtime = rt_runtime; | |
2516 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
2517 | } | |
2518 | raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
2519 | unlock: | |
2520 | read_unlock(&tasklist_lock); | |
2521 | mutex_unlock(&rt_constraints_mutex); | |
2522 | ||
2523 | return err; | |
2524 | } | |
2525 | ||
2526 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) | |
2527 | { | |
2528 | u64 rt_runtime, rt_period; | |
2529 | ||
2530 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
2531 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; | |
2532 | if (rt_runtime_us < 0) | |
2533 | rt_runtime = RUNTIME_INF; | |
2534 | ||
2535 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); | |
2536 | } | |
2537 | ||
2538 | long sched_group_rt_runtime(struct task_group *tg) | |
2539 | { | |
2540 | u64 rt_runtime_us; | |
2541 | ||
2542 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) | |
2543 | return -1; | |
2544 | ||
2545 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; | |
2546 | do_div(rt_runtime_us, NSEC_PER_USEC); | |
2547 | return rt_runtime_us; | |
2548 | } | |
2549 | ||
2550 | int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us) | |
2551 | { | |
2552 | u64 rt_runtime, rt_period; | |
2553 | ||
2554 | rt_period = rt_period_us * NSEC_PER_USEC; | |
2555 | rt_runtime = tg->rt_bandwidth.rt_runtime; | |
2556 | ||
2557 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); | |
2558 | } | |
2559 | ||
2560 | long sched_group_rt_period(struct task_group *tg) | |
2561 | { | |
2562 | u64 rt_period_us; | |
2563 | ||
2564 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
2565 | do_div(rt_period_us, NSEC_PER_USEC); | |
2566 | return rt_period_us; | |
2567 | } | |
2568 | ||
2569 | static int sched_rt_global_constraints(void) | |
2570 | { | |
2571 | int ret = 0; | |
2572 | ||
2573 | mutex_lock(&rt_constraints_mutex); | |
2574 | read_lock(&tasklist_lock); | |
2575 | ret = __rt_schedulable(NULL, 0, 0); | |
2576 | read_unlock(&tasklist_lock); | |
2577 | mutex_unlock(&rt_constraints_mutex); | |
2578 | ||
2579 | return ret; | |
2580 | } | |
2581 | ||
2582 | int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) | |
2583 | { | |
2584 | /* Don't accept realtime tasks when there is no way for them to run */ | |
2585 | if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) | |
2586 | return 0; | |
2587 | ||
2588 | return 1; | |
2589 | } | |
2590 | ||
2591 | #else /* !CONFIG_RT_GROUP_SCHED */ | |
2592 | static int sched_rt_global_constraints(void) | |
2593 | { | |
2594 | unsigned long flags; | |
2595 | int i; | |
2596 | ||
2597 | raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); | |
2598 | for_each_possible_cpu(i) { | |
2599 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; | |
2600 | ||
2601 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
2602 | rt_rq->rt_runtime = global_rt_runtime(); | |
2603 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
2604 | } | |
2605 | raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); | |
2606 | ||
2607 | return 0; | |
2608 | } | |
2609 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
2610 | ||
2611 | static int sched_rt_global_validate(void) | |
2612 | { | |
2613 | if (sysctl_sched_rt_period <= 0) | |
2614 | return -EINVAL; | |
2615 | ||
2616 | if ((sysctl_sched_rt_runtime != RUNTIME_INF) && | |
2617 | (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) | |
2618 | return -EINVAL; | |
2619 | ||
2620 | return 0; | |
2621 | } | |
2622 | ||
2623 | static void sched_rt_do_global(void) | |
2624 | { | |
2625 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); | |
2626 | def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); | |
2627 | } | |
2628 | ||
2629 | int sched_rt_handler(struct ctl_table *table, int write, | |
2630 | void __user *buffer, size_t *lenp, | |
2631 | loff_t *ppos) | |
2632 | { | |
2633 | int old_period, old_runtime; | |
2634 | static DEFINE_MUTEX(mutex); | |
2635 | int ret; | |
2636 | ||
2637 | mutex_lock(&mutex); | |
2638 | old_period = sysctl_sched_rt_period; | |
2639 | old_runtime = sysctl_sched_rt_runtime; | |
2640 | ||
2641 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
2642 | ||
2643 | if (!ret && write) { | |
2644 | ret = sched_rt_global_validate(); | |
2645 | if (ret) | |
2646 | goto undo; | |
2647 | ||
2648 | ret = sched_dl_global_validate(); | |
2649 | if (ret) | |
2650 | goto undo; | |
2651 | ||
2652 | ret = sched_rt_global_constraints(); | |
2653 | if (ret) | |
2654 | goto undo; | |
2655 | ||
2656 | sched_rt_do_global(); | |
2657 | sched_dl_do_global(); | |
2658 | } | |
2659 | if (0) { | |
2660 | undo: | |
2661 | sysctl_sched_rt_period = old_period; | |
2662 | sysctl_sched_rt_runtime = old_runtime; | |
2663 | } | |
2664 | mutex_unlock(&mutex); | |
2665 | ||
2666 | return ret; | |
2667 | } | |
2668 | ||
2669 | int sched_rr_handler(struct ctl_table *table, int write, | |
2670 | void __user *buffer, size_t *lenp, | |
2671 | loff_t *ppos) | |
2672 | { | |
2673 | int ret; | |
2674 | static DEFINE_MUTEX(mutex); | |
2675 | ||
2676 | mutex_lock(&mutex); | |
2677 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
2678 | /* | |
2679 | * Make sure that internally we keep jiffies. | |
2680 | * Also, writing zero resets the timeslice to default: | |
2681 | */ | |
2682 | if (!ret && write) { | |
2683 | sched_rr_timeslice = | |
2684 | sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE : | |
2685 | msecs_to_jiffies(sysctl_sched_rr_timeslice); | |
2686 | } | |
2687 | mutex_unlock(&mutex); | |
2688 | return ret; | |
2689 | } | |
2690 | ||
ada18de2 | 2691 | #ifdef CONFIG_SCHED_DEBUG |
029632fb | 2692 | void print_rt_stats(struct seq_file *m, int cpu) |
ada18de2 | 2693 | { |
ec514c48 | 2694 | rt_rq_iter_t iter; |
ada18de2 PZ |
2695 | struct rt_rq *rt_rq; |
2696 | ||
2697 | rcu_read_lock(); | |
ec514c48 | 2698 | for_each_rt_rq(rt_rq, iter, cpu_rq(cpu)) |
ada18de2 PZ |
2699 | print_rt_rq(m, cpu, rt_rq); |
2700 | rcu_read_unlock(); | |
2701 | } | |
55e12e5e | 2702 | #endif /* CONFIG_SCHED_DEBUG */ |